diff --git a/Dockerfile b/Dockerfile index 8655c51f34c..bde3412ed7f 100644 --- a/Dockerfile +++ b/Dockerfile @@ -66,9 +66,11 @@ RUN npm install --prefer-offline --no-audit && \ # frontend stats the readme path during dep resolution, so we `touch` an # empty placeholder — the real README is restored by `COPY . .` below. # -# `uv sync --frozen --no-install-project --extra all` installs only the -# deps reachable through the composite `[all]` extra (handpicked set -# intended for the production image). We do NOT use `--all-extras`: +# `uv sync --frozen --no-install-project --extra all --extra messaging` +# installs the deps reachable through the composite `[all]` extra +# (handpicked set intended for the production image), plus gateway +# messaging adapters that should work in the published image without a +# first-boot lazy install. We do NOT use `--all-extras`: # that would pull in `[rl]` (atroposlib + tinker + torch + wandb from # git), `[yc-bench]` (another git dep), and `[termux-all]` (Android # redundancy), none of which belong in the published container. @@ -76,7 +78,7 @@ RUN npm install --prefer-offline --no-audit && \ # The editable link is created after the source copy below. COPY pyproject.toml uv.lock ./ RUN touch ./README.md -RUN uv sync --frozen --no-install-project --extra all +RUN uv sync --frozen --no-install-project --extra all --extra messaging # ---------- Source code ---------- # .dockerignore excludes node_modules, so the installs above survive. @@ -94,10 +96,10 @@ RUN cd web && npm run build && \ # hermes_cli/main.py succeeds (see #18800). /opt/hermes/web is build-time # only (HERMES_WEB_DIST points at hermes_cli/web_dist) and is intentionally # not chowned here. -# The .venv MUST be hermes-writable so lazy_deps.py can install platform -# packages (discord.py, telegram, slack, etc.) at first gateway boot. -# Without this, `uv pip install` fails with EACCES and all messaging -# adapters silently fail to load. See tools/lazy_deps.py. +# The .venv MUST remain hermes-writable so lazy_deps.py can install +# remaining optional platform packages and future pin bumps at first use. +# Without this, `uv pip install` fails with EACCES and adapters silently +# fail to load. See tools/lazy_deps.py. USER root RUN chmod -R a+rX /opt/hermes && \ chown -R hermes:hermes /opt/hermes/.venv /opt/hermes/ui-tui /opt/hermes/node_modules diff --git a/acp_adapter/tools.py b/acp_adapter/tools.py index 31ae943a056..77a62e243bc 100644 --- a/acp_adapter/tools.py +++ b/acp_adapter/tools.py @@ -1123,7 +1123,6 @@ def build_tool_start( ) # Generic fallback - import json try: args_text = json.dumps(arguments, indent=2, default=str) except (TypeError, ValueError): diff --git a/acp_registry/agent.json b/acp_registry/agent.json index b94a48e089f..b23d1642a94 100644 --- a/acp_registry/agent.json +++ b/acp_registry/agent.json @@ -1,7 +1,7 @@ { "id": "hermes-agent", "name": "Hermes Agent", - "version": "0.13.0", + "version": "0.14.0", "description": "Self-improving open-source AI agent by Nous Research with ACP editor integration, persistent memory, skills, and rich tool support.", "repository": "https://github.com/NousResearch/hermes-agent", "website": "https://hermes-agent.nousresearch.com/docs/user-guide/features/acp", @@ -9,7 +9,7 @@ "license": "MIT", "distribution": { "uvx": { - "package": "hermes-agent[acp]==0.13.0", + "package": "hermes-agent[acp]==0.14.0", "args": ["hermes-acp"] } } diff --git a/agent/agent_init.py b/agent/agent_init.py new file mode 100644 index 00000000000..df8fe229e7b --- /dev/null +++ b/agent/agent_init.py @@ -0,0 +1,1469 @@ +"""Implementation of :meth:`AIAgent.__init__` — extracted as a module function. + +``AIAgent.__init__`` is one of the longest methods in the codebase (60+ +parameters, ~1,400 lines of attribute initialization, provider +auto-detection, credential resolution, context-engine bootstrap, etc.). +Keeping it in ``run_agent.py`` bloats that file with code that's mostly +"setup state, then forget". + +After this extraction the body lives here as ``init_agent(agent, ...)`` +and :meth:`AIAgent.__init__` is a thin wrapper that calls +``init_agent(self, ...)``. All imports the body needs at module-load +time are listed below; the body also performs many lazy imports inside +its own scope that come along unchanged. + +Symbols that tests patch on ``run_agent.*`` (``OpenAI``, ``cleanup_vm``, +etc.) are resolved through :func:`_ra` so the patch contract is +preserved. +""" + +from __future__ import annotations + +import logging +import os +import re +import sys +import threading +import time +import uuid +from datetime import datetime +from pathlib import Path +from typing import Any, Dict, List, Optional +from urllib.parse import urlparse, parse_qs, urlunparse + +from agent.context_compressor import ContextCompressor +from agent.iteration_budget import IterationBudget +from agent.memory_manager import StreamingContextScrubber +from agent.model_metadata import ( + MINIMUM_CONTEXT_LENGTH, + fetch_model_metadata, + get_model_context_length, + is_local_endpoint, + query_ollama_num_ctx, +) +from agent.process_bootstrap import _install_safe_stdio +from agent.subdirectory_hints import SubdirectoryHintTracker +from agent.think_scrubber import StreamingThinkScrubber +from agent.tool_guardrails import ( + ToolCallGuardrailConfig, + ToolCallGuardrailController, + ToolGuardrailDecision, +) +from hermes_cli.config import cfg_get +from hermes_cli.timeouts import get_provider_request_timeout +from hermes_constants import get_hermes_home +from model_tools import check_toolset_requirements, get_tool_definitions +from utils import base_url_host_matches + +# Use the same logger name as run_agent so tests patching ``run_agent.logger`` +# capture our warnings. (run_agent.py also does +# ``logger = logging.getLogger(__name__)``, which resolves to "run_agent" +# from inside that module.) +logger = logging.getLogger("run_agent") + + +def _ra(): + """Lazy reference to ``run_agent`` so callers can patch + ``run_agent.OpenAI`` / ``run_agent.cleanup_vm`` / ... and have those + patches reach this code path. + """ + import run_agent + return run_agent + + +def init_agent( + agent, + base_url: str = None, + api_key: str = None, + provider: str = None, + api_mode: str = None, + acp_command: str = None, + acp_args: list[str] | None = None, + command: str = None, + args: list[str] | None = None, + model: str = "", + max_iterations: int = 90, # Default tool-calling iterations (shared with subagents) + tool_delay: float = 1.0, + enabled_toolsets: List[str] = None, + disabled_toolsets: List[str] = None, + save_trajectories: bool = False, + verbose_logging: bool = False, + quiet_mode: bool = False, + ephemeral_system_prompt: str = None, + log_prefix_chars: int = 100, + log_prefix: str = "", + providers_allowed: List[str] = None, + providers_ignored: List[str] = None, + providers_order: List[str] = None, + provider_sort: str = None, + provider_require_parameters: bool = False, + provider_data_collection: str = None, + openrouter_min_coding_score: Optional[float] = None, + session_id: str = None, + tool_progress_callback: callable = None, + tool_start_callback: callable = None, + tool_complete_callback: callable = None, + thinking_callback: callable = None, + reasoning_callback: callable = None, + clarify_callback: callable = None, + step_callback: callable = None, + stream_delta_callback: callable = None, + interim_assistant_callback: callable = None, + tool_gen_callback: callable = None, + status_callback: callable = None, + max_tokens: int = None, + reasoning_config: Dict[str, Any] = None, + service_tier: str = None, + request_overrides: Dict[str, Any] = None, + prefill_messages: List[Dict[str, Any]] = None, + platform: str = None, + user_id: str = None, + user_name: str = None, + chat_id: str = None, + chat_name: str = None, + chat_type: str = None, + thread_id: str = None, + gateway_session_key: str = None, + skip_context_files: bool = False, + load_soul_identity: bool = False, + skip_memory: bool = False, + session_db=None, + parent_session_id: str = None, + iteration_budget: "IterationBudget" = None, + fallback_model: Dict[str, Any] = None, + credential_pool=None, + checkpoints_enabled: bool = False, + checkpoint_max_snapshots: int = 20, + checkpoint_max_total_size_mb: int = 500, + checkpoint_max_file_size_mb: int = 10, + pass_session_id: bool = False, +): + """ + Initialize the AI Agent. + + Args: + base_url (str): Base URL for the model API (optional) + api_key (str): API key for authentication (optional, uses env var if not provided) + provider (str): Provider identifier (optional; used for telemetry/routing hints) + api_mode (str): API mode override: "chat_completions" or "codex_responses" + model (str): Model name to use (default: "anthropic/claude-opus-4.6") + max_iterations (int): Maximum number of tool calling iterations (default: 90) + tool_delay (float): Delay between tool calls in seconds (default: 1.0) + enabled_toolsets (List[str]): Only enable tools from these toolsets (optional) + disabled_toolsets (List[str]): Disable tools from these toolsets (optional) + save_trajectories (bool): Whether to save conversation trajectories to JSONL files (default: False) + verbose_logging (bool): Enable verbose logging for debugging (default: False) + quiet_mode (bool): Suppress progress output for clean CLI experience (default: False) + ephemeral_system_prompt (str): System prompt used during agent execution but NOT saved to trajectories (optional) + log_prefix_chars (int): Number of characters to show in log previews for tool calls/responses (default: 100) + log_prefix (str): Prefix to add to all log messages for identification in parallel processing (default: "") + providers_allowed (List[str]): OpenRouter providers to allow (optional) + providers_ignored (List[str]): OpenRouter providers to ignore (optional) + providers_order (List[str]): OpenRouter providers to try in order (optional) + provider_sort (str): Sort providers by price/throughput/latency (optional) + openrouter_min_coding_score (float): Coding-score floor (0.0-1.0) for the + openrouter/pareto-code router. Only applied when model == "openrouter/pareto-code". + None or empty = let OpenRouter pick the strongest available coder. + session_id (str): Pre-generated session ID for logging (optional, auto-generated if not provided) + tool_progress_callback (callable): Callback function(tool_name, args_preview) for progress notifications + clarify_callback (callable): Callback function(question, choices) -> str for interactive user questions. + Provided by the platform layer (CLI or gateway). If None, the clarify tool returns an error. + max_tokens (int): Maximum tokens for model responses (optional, uses model default if not set) + reasoning_config (Dict): OpenRouter reasoning configuration override (e.g. {"effort": "none"} to disable thinking). + If None, defaults to {"enabled": True, "effort": "medium"} for OpenRouter. Set to disable/customize reasoning. + prefill_messages (List[Dict]): Messages to prepend to conversation history as prefilled context. + Useful for injecting a few-shot example or priming the model's response style. + Example: [{"role": "user", "content": "Hi!"}, {"role": "assistant", "content": "Hello!"}] + NOTE: Anthropic Sonnet 4.6+ and Opus 4.6+ reject a conversation that ends on an + assistant-role message (400 error). For those models use structured outputs or + output_config.format instead of a trailing-assistant prefill. + platform (str): The interface platform the user is on (e.g. "cli", "telegram", "discord", "whatsapp"). + Used to inject platform-specific formatting hints into the system prompt. + skip_context_files (bool): If True, skip auto-injection of SOUL.md, AGENTS.md, and .cursorrules + into the system prompt. Use this for batch processing and data generation to avoid + polluting trajectories with user-specific persona or project instructions. + load_soul_identity (bool): If True, still use ~/.hermes/SOUL.md as the primary + identity even when skip_context_files=True. Project context files from the cwd + remain skipped. + """ + _install_safe_stdio() + + agent.model = model + agent.max_iterations = max_iterations + # Shared iteration budget — parent creates, children inherit. + # Consumed by every LLM turn across parent + all subagents. + agent.iteration_budget = iteration_budget or IterationBudget(max_iterations) + agent.tool_delay = tool_delay + agent.save_trajectories = save_trajectories + agent.verbose_logging = verbose_logging + agent.quiet_mode = quiet_mode + agent.ephemeral_system_prompt = ephemeral_system_prompt + agent.platform = platform # "cli", "telegram", "discord", "whatsapp", etc. + agent._user_id = user_id # Platform user identifier (gateway sessions) + agent._user_name = user_name + agent._chat_id = chat_id + agent._chat_name = chat_name + agent._chat_type = chat_type + agent._thread_id = thread_id + agent._gateway_session_key = gateway_session_key # Stable per-chat key (e.g. agent:main:telegram:dm:123) + # Pluggable print function — CLI replaces this with _cprint so that + # raw ANSI status lines are routed through prompt_toolkit's renderer + # instead of going directly to stdout where patch_stdout's StdoutProxy + # would mangle the escape sequences. None = use builtins.print. + agent._print_fn = None + agent.background_review_callback = None # Optional sync callback for gateway delivery + agent.skip_context_files = skip_context_files + agent.load_soul_identity = load_soul_identity + agent.pass_session_id = pass_session_id + agent._credential_pool = credential_pool + agent.log_prefix_chars = log_prefix_chars + agent.log_prefix = f"{log_prefix} " if log_prefix else "" + # Store effective base URL for feature detection (prompt caching, reasoning, etc.) + agent.base_url = base_url or "" + provider_name = provider.strip().lower() if isinstance(provider, str) and provider.strip() else None + agent.provider = provider_name or "" + agent.acp_command = acp_command or command + agent.acp_args = list(acp_args or args or []) + if api_mode in {"chat_completions", "codex_responses", "anthropic_messages", "bedrock_converse", "codex_app_server"}: + agent.api_mode = api_mode + elif agent.provider == "openai-codex": + agent.api_mode = "codex_responses" + elif agent.provider in {"xai", "xai-oauth"}: + agent.api_mode = "codex_responses" + elif (provider_name is None) and ( + agent._base_url_hostname == "chatgpt.com" + and "/backend-api/codex" in agent._base_url_lower + ): + agent.api_mode = "codex_responses" + agent.provider = "openai-codex" + elif (provider_name is None) and agent._base_url_hostname == "api.x.ai": + agent.api_mode = "codex_responses" + agent.provider = "xai" + elif agent.provider == "anthropic" or (provider_name is None and agent._base_url_hostname == "api.anthropic.com"): + agent.api_mode = "anthropic_messages" + agent.provider = "anthropic" + elif agent._base_url_lower.rstrip("/").endswith("/anthropic"): + # Third-party Anthropic-compatible endpoints (e.g. MiniMax, DashScope) + # use a URL convention ending in /anthropic. Auto-detect these so the + # Anthropic Messages API adapter is used instead of chat completions. + agent.api_mode = "anthropic_messages" + elif agent.provider == "bedrock" or ( + agent._base_url_hostname.startswith("bedrock-runtime.") + and base_url_host_matches(agent._base_url_lower, "amazonaws.com") + ): + # AWS Bedrock — auto-detect from provider name or base URL + # (bedrock-runtime..amazonaws.com). + agent.api_mode = "bedrock_converse" + else: + agent.api_mode = "chat_completions" + + # Eagerly warm the transport cache so import errors surface at init, + # not mid-conversation. Also validates the api_mode is registered. + try: + agent._get_transport() + except Exception: + pass # Non-fatal — transport may not exist for all modes yet + + try: + from hermes_cli.model_normalize import ( + _AGGREGATOR_PROVIDERS, + normalize_model_for_provider, + ) + + if agent.provider not in _AGGREGATOR_PROVIDERS: + agent.model = normalize_model_for_provider(agent.model, agent.provider) + except Exception: + pass + + # GPT-5.x models usually require the Responses API path, but some + # providers have exceptions (for example Copilot's gpt-5-mini still + # uses chat completions). Also auto-upgrade for direct OpenAI URLs + # (api.openai.com) since all newer tool-calling models prefer + # Responses there. ACP runtimes are excluded: CopilotACPClient + # handles its own routing and does not implement the Responses API + # surface. + # When api_mode was explicitly provided, respect it — the user + # knows what their endpoint supports (#10473). + # Exception: Azure OpenAI serves gpt-5.x on /chat/completions and + # does NOT support the Responses API — skip the upgrade for Azure + # (openai.azure.com), even though it looks OpenAI-compatible. + if ( + api_mode is None + and agent.api_mode == "chat_completions" + and agent.provider != "copilot-acp" + and not str(agent.base_url or "").lower().startswith("acp://copilot") + and not str(agent.base_url or "").lower().startswith("acp+tcp://") + and not agent._is_azure_openai_url() + and ( + agent._is_direct_openai_url() + or agent._provider_model_requires_responses_api( + agent.model, + provider=agent.provider, + ) + ) + ): + agent.api_mode = "codex_responses" + # Invalidate the eager-warmed transport cache — api_mode changed + # from chat_completions to codex_responses after the warm at __init__. + if hasattr(agent, "_transport_cache"): + agent._transport_cache.clear() + + # Pre-warm OpenRouter model metadata cache in a background thread. + # fetch_model_metadata() is cached for 1 hour; this avoids a blocking + # HTTP request on the first API response when pricing is estimated. + # Use a process-level Event so this thread is only spawned once — a new + # AIAgent is created for every gateway request, so without the guard + # each message leaks one OS thread and the process eventually exhausts + # the system thread limit (RuntimeError: can't start new thread). + if (agent.provider == "openrouter" or agent._is_openrouter_url()) and \ + not _ra()._openrouter_prewarm_done.is_set(): + _ra()._openrouter_prewarm_done.set() + threading.Thread( + target=fetch_model_metadata, + daemon=True, + name="openrouter-prewarm", + ).start() + + agent.tool_progress_callback = tool_progress_callback + agent.tool_start_callback = tool_start_callback + agent.tool_complete_callback = tool_complete_callback + agent.suppress_status_output = False + agent.thinking_callback = thinking_callback + agent.reasoning_callback = reasoning_callback + agent.clarify_callback = clarify_callback + agent.step_callback = step_callback + agent.stream_delta_callback = stream_delta_callback + agent.interim_assistant_callback = interim_assistant_callback + agent.status_callback = status_callback + agent.tool_gen_callback = tool_gen_callback + + + # Tool execution state — allows _vprint during tool execution + # even when stream consumers are registered (no tokens streaming then) + agent._executing_tools = False + agent._tool_guardrails = ToolCallGuardrailController() + agent._tool_guardrail_halt_decision: ToolGuardrailDecision | None = None + + # Interrupt mechanism for breaking out of tool loops + agent._interrupt_requested = False + agent._interrupt_message = None # Optional message that triggered interrupt + agent._execution_thread_id: int | None = None # Set at run_conversation() start + agent._interrupt_thread_signal_pending = False + agent._client_lock = threading.RLock() + + # /steer mechanism — inject a user note into the next tool result + # without interrupting the agent. Unlike interrupt(), steer() does + # NOT set _interrupt_requested; it waits for the current tool batch + # to finish naturally, then the drain hook appends the text to the + # last tool result's content so the model sees it on its next + # iteration. Message-role alternation is preserved (we modify an + # existing tool message rather than inserting a new user turn). + agent._pending_steer: Optional[str] = None + agent._pending_steer_lock = threading.Lock() + + # Concurrent-tool worker thread tracking. `_execute_tool_calls_concurrent` + # runs each tool on its own ThreadPoolExecutor worker — those worker + # threads have tids distinct from `_execution_thread_id`, so + # `_set_interrupt(True, _execution_thread_id)` alone does NOT cause + # `is_interrupted()` inside the worker to return True. Track the + # workers here so `interrupt()` / `clear_interrupt()` can fan out to + # their tids explicitly. + agent._tool_worker_threads: set[int] = set() + agent._tool_worker_threads_lock = threading.Lock() + + # Subagent delegation state + agent._delegate_depth = 0 # 0 = top-level agent, incremented for children + agent._active_children = [] # Running child AIAgents (for interrupt propagation) + agent._active_children_lock = threading.Lock() + + # Store OpenRouter provider preferences + agent.providers_allowed = providers_allowed + agent.providers_ignored = providers_ignored + agent.providers_order = providers_order + agent.provider_sort = provider_sort + agent.provider_require_parameters = provider_require_parameters + agent.provider_data_collection = provider_data_collection + agent.openrouter_min_coding_score = openrouter_min_coding_score + + # Store toolset filtering options + agent.enabled_toolsets = enabled_toolsets + agent.disabled_toolsets = disabled_toolsets + + # Model response configuration + agent.max_tokens = max_tokens # None = use model default + agent.reasoning_config = reasoning_config # None = use default (medium for OpenRouter) + agent.service_tier = service_tier + agent.request_overrides = dict(request_overrides or {}) + agent.prefill_messages = prefill_messages or [] # Prefilled conversation turns + agent._force_ascii_payload = False + + # Anthropic prompt caching: auto-enabled for Claude models on native + # Anthropic, OpenRouter, and third-party gateways that speak the + # Anthropic protocol (``api_mode == 'anthropic_messages'``). Reduces + # input costs by ~75% on multi-turn conversations. Uses system_and_3 + # strategy (4 breakpoints). See ``_anthropic_prompt_cache_policy`` + # for the layout-vs-transport decision. + agent._use_prompt_caching, agent._use_native_cache_layout = ( + agent._anthropic_prompt_cache_policy() + ) + # Anthropic supports "5m" (default) and "1h" cache TTL tiers. Read from + # config.yaml under prompt_caching.cache_ttl; unknown values keep "5m". + # 1h tier costs 2x on write vs 1.25x for 5m, but amortizes across long + # sessions with >5-minute pauses between turns (#14971). + agent._cache_ttl = "5m" + try: + from hermes_cli.config import load_config as _load_pc_cfg + + _pc_cfg = _load_pc_cfg().get("prompt_caching", {}) or {} + _ttl = _pc_cfg.get("cache_ttl", "5m") + if _ttl in {"5m", "1h"}: + agent._cache_ttl = _ttl + except Exception: + pass + + # Iteration budget: the LLM is only notified when it actually exhausts + # the iteration budget (api_call_count >= max_iterations). At that + # point we inject ONE message, allow one final API call, and if the + # model doesn't produce a text response, force a user-message asking + # it to summarise. No intermediate pressure warnings — they caused + # models to "give up" prematurely on complex tasks (#7915). + agent._budget_exhausted_injected = False + agent._budget_grace_call = False + + # Activity tracking — updated on each API call, tool execution, and + # stream chunk. Used by the gateway timeout handler to report what the + # agent was doing when it was killed, and by the "still working" + # notifications to show progress. + agent._last_activity_ts: float = time.time() + agent._last_activity_desc: str = "initializing" + agent._current_tool: str | None = None + agent._api_call_count: int = 0 + + # Rate limit tracking — updated from x-ratelimit-* response headers + # after each API call. Accessed by /usage slash command. + agent._rate_limit_state: Optional["RateLimitState"] = None + + # OpenRouter response cache hit counter — incremented when + # X-OpenRouter-Cache-Status: HIT is seen in streaming response headers. + agent._or_cache_hits: int = 0 + + # Centralized logging — agent.log (INFO+) and errors.log (WARNING+) + # both live under ~/.hermes/logs/. Idempotent, so gateway mode + # (which creates a new AIAgent per message) won't duplicate handlers. + from hermes_logging import setup_logging, setup_verbose_logging + setup_logging(hermes_home=_ra()._hermes_home) + + if agent.verbose_logging: + setup_verbose_logging() + _ra().logger.info("Verbose logging enabled (third-party library logs suppressed)") + elif agent.quiet_mode: + # In quiet mode (CLI default), keep console output clean — + # but DO NOT raise per-logger levels. Doing so prevents the + # root logger's file handlers (agent.log, errors.log) from + # ever seeing the records, because Python checks + # logger.isEnabledFor() before handler propagation. We rely + # on the fact that hermes_logging.setup_logging() does not + # install a console StreamHandler in quiet mode — so INFO + # records flow to the file handlers but never reach a + # console. Any future noise reduction belongs at the + # handler level inside hermes_logging.py, not here. + pass + + # Internal stream callback (set during streaming TTS). + # Initialized here so _vprint can reference it before run_conversation. + agent._stream_callback = None + # Deferred paragraph break flag — set after tool iterations so a + # single "\n\n" is prepended to the next real text delta. + agent._stream_needs_break = False + # Stateful scrubber for spans split across stream + # deltas (#5719). sanitize_context() alone can't survive chunk + # boundaries because the block regex needs both tags in one string. + agent._stream_context_scrubber = StreamingContextScrubber() + # Stateful scrubber for reasoning/thinking tags in streamed deltas + # (#17924). Replaces the per-delta _strip_think_blocks regex that + # destroyed downstream state (e.g. MiniMax-M2.7 streaming + # '' as delta1 and 'Let me check' as delta2 — the regex + # erased delta1, so downstream state machines never learned a + # block was open and leaked delta2 as content). + agent._stream_think_scrubber = StreamingThinkScrubber() + # Visible assistant text already delivered through live token callbacks + # during the current model response. Used to avoid re-sending the same + # commentary when the provider later returns it as a completed interim + # assistant message. + agent._current_streamed_assistant_text = "" + + # Optional current-turn user-message override used when the API-facing + # user message intentionally differs from the persisted transcript + # (e.g. CLI voice mode adds a temporary prefix for the live call only). + agent._persist_user_message_idx = None + agent._persist_user_message_override = None + + # Cache anthropic image-to-text fallbacks per image payload/URL so a + # single tool loop does not repeatedly re-run auxiliary vision on the + # same image history. + agent._anthropic_image_fallback_cache: Dict[str, str] = {} + + # Initialize LLM client via centralized provider router. + # The router handles auth resolution, base URL, headers, and + # Codex/Anthropic wrapping for all known providers. + # raw_codex=True because the main agent needs direct responses.stream() + # access for Codex Responses API streaming. + agent._anthropic_client = None + agent._is_anthropic_oauth = False + + # Resolve per-provider / per-model request timeout once up front so + # every client construction path below (Anthropic native, OpenAI-wire, + # router-based implicit auth) can apply it consistently. Bedrock + # Claude uses its own timeout path and is not covered here. + _provider_timeout = get_provider_request_timeout(agent.provider, agent.model) + + if agent.api_mode == "anthropic_messages": + from agent.anthropic_adapter import build_anthropic_client, resolve_anthropic_token + # Bedrock + Claude → use AnthropicBedrock SDK for full feature parity + # (prompt caching, thinking budgets, adaptive thinking). + _is_bedrock_anthropic = agent.provider == "bedrock" + if _is_bedrock_anthropic: + from agent.anthropic_adapter import build_anthropic_bedrock_client + _region_match = re.search(r"bedrock-runtime\.([a-z0-9-]+)\.", base_url or "") + _br_region = _region_match.group(1) if _region_match else "us-east-1" + agent._bedrock_region = _br_region + agent._anthropic_client = build_anthropic_bedrock_client(_br_region) + agent._anthropic_api_key = "aws-sdk" + agent._anthropic_base_url = base_url + agent._is_anthropic_oauth = False + agent.api_key = "aws-sdk" + agent.client = None + agent._client_kwargs = {} + if not agent.quiet_mode: + print(f"🤖 AI Agent initialized with model: {agent.model} (AWS Bedrock + AnthropicBedrock SDK, {_br_region})") + else: + # Only fall back to ANTHROPIC_TOKEN when the provider is actually Anthropic. + # Other anthropic_messages providers (MiniMax, Alibaba, etc.) must use their own API key. + # Falling back would send Anthropic credentials to third-party endpoints (Fixes #1739, #minimax-401). + _is_native_anthropic = agent.provider == "anthropic" + effective_key = (api_key or resolve_anthropic_token() or "") if _is_native_anthropic else (api_key or "") + agent.api_key = effective_key + agent._anthropic_api_key = effective_key + agent._anthropic_base_url = base_url + # Only mark the session as OAuth-authenticated when the token + # genuinely belongs to native Anthropic. Third-party providers + # (MiniMax, Kimi, GLM, LiteLLM proxies) that accept the + # Anthropic protocol must never trip OAuth code paths — doing + # so injects Claude-Code identity headers and system prompts + # that cause 401/403 on their endpoints. Guards #1739 and + # the third-party identity-injection bug. + from agent.anthropic_adapter import _is_oauth_token as _is_oat + agent._is_anthropic_oauth = _is_oat(effective_key) if _is_native_anthropic else False + agent._anthropic_client = build_anthropic_client(effective_key, base_url, timeout=_provider_timeout) + # No OpenAI client needed for Anthropic mode + agent.client = None + agent._client_kwargs = {} + if not agent.quiet_mode: + print(f"🤖 AI Agent initialized with model: {agent.model} (Anthropic native)") + if effective_key and len(effective_key) > 12: + print(f"🔑 Using token: {effective_key[:8]}...{effective_key[-4:]}") + elif agent.api_mode == "bedrock_converse": + # AWS Bedrock — uses boto3 directly, no OpenAI client needed. + # Region is extracted from the base_url or defaults to us-east-1. + _region_match = re.search(r"bedrock-runtime\.([a-z0-9-]+)\.", base_url or "") + agent._bedrock_region = _region_match.group(1) if _region_match else "us-east-1" + # Guardrail config — read from config.yaml at init time. + agent._bedrock_guardrail_config = None + try: + from hermes_cli.config import load_config as _load_br_cfg + _gr = _load_br_cfg().get("bedrock", {}).get("guardrail", {}) + if _gr.get("guardrail_identifier") and _gr.get("guardrail_version"): + agent._bedrock_guardrail_config = { + "guardrailIdentifier": _gr["guardrail_identifier"], + "guardrailVersion": _gr["guardrail_version"], + } + if _gr.get("stream_processing_mode"): + agent._bedrock_guardrail_config["streamProcessingMode"] = _gr["stream_processing_mode"] + if _gr.get("trace"): + agent._bedrock_guardrail_config["trace"] = _gr["trace"] + except Exception: + pass + agent.client = None + agent._client_kwargs = {} + if not agent.quiet_mode: + _gr_label = " + Guardrails" if agent._bedrock_guardrail_config else "" + print(f"🤖 AI Agent initialized with model: {agent.model} (AWS Bedrock, {agent._bedrock_region}{_gr_label})") + else: + if api_key and base_url: + # Explicit credentials from CLI/gateway — construct directly. + # The runtime provider resolver already handled auth for us. + # Extract query params (e.g. Azure api-version) from base_url + # and pass via default_query to prevent loss during SDK URL + # joining (httpx drops query string when joining paths). + _parsed_url = urlparse(base_url) + if _parsed_url.query: + _clean_url = urlunparse(_parsed_url._replace(query="")) + _query_params = { + k: v[0] for k, v in parse_qs(_parsed_url.query).items() + } + client_kwargs = { + "api_key": api_key, + "base_url": _clean_url, + "default_query": _query_params, + } + else: + client_kwargs = {"api_key": api_key, "base_url": base_url} + if _provider_timeout is not None: + client_kwargs["timeout"] = _provider_timeout + if agent.provider == "copilot-acp": + client_kwargs["command"] = agent.acp_command + client_kwargs["args"] = agent.acp_args + effective_base = base_url + if base_url_host_matches(effective_base, "openrouter.ai"): + from agent.auxiliary_client import build_or_headers + client_kwargs["default_headers"] = build_or_headers() + elif base_url_host_matches(effective_base, "integrate.api.nvidia.com"): + from agent.auxiliary_client import build_nvidia_nim_headers + client_kwargs["default_headers"] = build_nvidia_nim_headers(effective_base) + elif base_url_host_matches(effective_base, "api.routermint.com"): + client_kwargs["default_headers"] = _ra()._routermint_headers() + elif base_url_host_matches(effective_base, "api.githubcopilot.com"): + from hermes_cli.models import copilot_default_headers + + client_kwargs["default_headers"] = copilot_default_headers() + elif base_url_host_matches(effective_base, "api.kimi.com"): + client_kwargs["default_headers"] = { + "User-Agent": "claude-code/0.1.0", + } + elif base_url_host_matches(effective_base, "portal.qwen.ai"): + client_kwargs["default_headers"] = _ra()._qwen_portal_headers() + elif base_url_host_matches(effective_base, "chatgpt.com"): + from agent.auxiliary_client import _codex_cloudflare_headers + client_kwargs["default_headers"] = _codex_cloudflare_headers(api_key) + elif "default_headers" not in client_kwargs: + # Fall back to profile.default_headers for providers that + # declare custom headers (e.g. Vercel AI Gateway attribution, + # Kimi User-Agent on non-kimi.com endpoints). + try: + from providers import get_provider_profile as _gpf + _ph = _gpf(agent.provider) + if _ph and _ph.default_headers: + client_kwargs["default_headers"] = dict(_ph.default_headers) + except Exception: + pass + else: + # No explicit creds — use the centralized provider router + from agent.auxiliary_client import resolve_provider_client + _routed_client, _ = resolve_provider_client( + agent.provider or "auto", model=agent.model, raw_codex=True) + if _routed_client is not None: + client_kwargs = { + "api_key": _routed_client.api_key, + "base_url": str(_routed_client.base_url), + } + if _provider_timeout is not None: + client_kwargs["timeout"] = _provider_timeout + # Preserve provider-specific headers the router set. The + # OpenAI SDK stores caller-provided default_headers in + # _custom_headers; older/mocked clients may expose + # _default_headers instead. + _routed_headers = getattr(_routed_client, "_custom_headers", None) + if not _routed_headers: + _routed_headers = getattr(_routed_client, "_default_headers", None) + if _routed_headers: + client_kwargs["default_headers"] = dict(_routed_headers) + else: + # When the user explicitly chose a non-OpenRouter provider + # but no credentials were found, fail fast with a clear + # message instead of silently routing through OpenRouter. + _explicit = (agent.provider or "").strip().lower() + if _explicit and _explicit not in {"auto", "openrouter", "custom"}: + # Look up the actual env var name from the provider + # config — some providers use non-standard names + # (e.g. alibaba → DASHSCOPE_API_KEY, not ALIBABA_API_KEY). + _env_hint = f"{_explicit.upper()}_API_KEY" + try: + from hermes_cli.auth import PROVIDER_REGISTRY + _pcfg = PROVIDER_REGISTRY.get(_explicit) + if _pcfg and _pcfg.api_key_env_vars: + _env_hint = _pcfg.api_key_env_vars[0] + except Exception: + pass + # --- Init-time fallback (#17929) --- + _fb_entries = [] + if isinstance(fallback_model, list): + _fb_entries = [ + f for f in fallback_model + if isinstance(f, dict) and f.get("provider") and f.get("model") + ] + elif isinstance(fallback_model, dict) and fallback_model.get("provider") and fallback_model.get("model"): + _fb_entries = [fallback_model] + _fb_resolved = False + for _fb in _fb_entries: + _fb_explicit_key = (_fb.get("api_key") or "").strip() or None + if not _fb_explicit_key: + _fb_key_env = (_fb.get("key_env") or _fb.get("api_key_env") or "").strip() + if _fb_key_env: + _fb_explicit_key = os.getenv(_fb_key_env, "").strip() or None + _fb_client, _fb_model = resolve_provider_client( + _fb["provider"], model=_fb["model"], raw_codex=True, + explicit_base_url=_fb.get("base_url"), + explicit_api_key=_fb_explicit_key, + ) + if _fb_client is not None: + agent.provider = _fb["provider"] + agent.model = _fb_model or _fb["model"] + agent._fallback_activated = True + client_kwargs = { + "api_key": _fb_client.api_key, + "base_url": str(_fb_client.base_url), + } + if _provider_timeout is not None: + client_kwargs["timeout"] = _provider_timeout + _fb_headers = getattr(_fb_client, "_custom_headers", None) + if not _fb_headers: + _fb_headers = getattr(_fb_client, "_default_headers", None) + if _fb_headers: + client_kwargs["default_headers"] = dict(_fb_headers) + _fb_resolved = True + break + if not _fb_resolved: + raise RuntimeError( + f"Provider '{_explicit}' is set in config.yaml but no API key " + f"was found. Set the {_env_hint} environment " + f"variable, or switch to a different provider with `hermes model`." + ) + if not getattr(agent, "_fallback_activated", False): + # No provider configured — reject with a clear message. + raise RuntimeError( + "No LLM provider configured. Run `hermes model` to " + "select a provider, or run `hermes setup` for first-time " + "configuration." + ) + + agent._client_kwargs = client_kwargs # stored for rebuilding after interrupt + + # Enable fine-grained tool streaming for Claude on OpenRouter. + # Without this, Anthropic buffers the entire tool call and goes + # silent for minutes while thinking — OpenRouter's upstream proxy + # times out during the silence. The beta header makes Anthropic + # stream tool call arguments token-by-token, keeping the + # connection alive. + _effective_base = str(client_kwargs.get("base_url", "")).lower() + if base_url_host_matches(_effective_base, "openrouter.ai") and "claude" in (agent.model or "").lower(): + headers = client_kwargs.get("default_headers") or {} + existing_beta = headers.get("x-anthropic-beta", "") + _FINE_GRAINED = "fine-grained-tool-streaming-2025-05-14" + if _FINE_GRAINED not in existing_beta: + if existing_beta: + headers["x-anthropic-beta"] = f"{existing_beta},{_FINE_GRAINED}" + else: + headers["x-anthropic-beta"] = _FINE_GRAINED + client_kwargs["default_headers"] = headers + + agent.api_key = client_kwargs.get("api_key", "") + agent.base_url = client_kwargs.get("base_url", agent.base_url) + try: + agent.client = agent._create_openai_client(client_kwargs, reason="agent_init", shared=True) + if not agent.quiet_mode: + print(f"🤖 AI Agent initialized with model: {agent.model}") + if base_url: + print(f"🔗 Using custom base URL: {base_url}") + # Always show API key info (masked) for debugging auth issues + key_used = client_kwargs.get("api_key", "none") + if key_used and key_used != "dummy-key" and len(key_used) > 12: + print(f"🔑 Using API key: {key_used[:8]}...{key_used[-4:]}") + else: + print(f"⚠️ Warning: API key appears invalid or missing (got: '{key_used[:20] if key_used else 'none'}...')") + except Exception as e: + raise RuntimeError(f"Failed to initialize OpenAI client: {e}") + + # Provider fallback chain — ordered list of backup providers tried + # when the primary is exhausted (rate-limit, overload, connection + # failure). Supports both legacy single-dict ``fallback_model`` and + # new list ``fallback_providers`` format. + if isinstance(fallback_model, list): + agent._fallback_chain = [ + f for f in fallback_model + if isinstance(f, dict) and f.get("provider") and f.get("model") + ] + elif isinstance(fallback_model, dict) and fallback_model.get("provider") and fallback_model.get("model"): + agent._fallback_chain = [fallback_model] + else: + agent._fallback_chain = [] + agent._fallback_index = 0 + agent._fallback_activated = getattr(agent, "_fallback_activated", False) + # Legacy attribute kept for backward compat (tests, external callers) + agent._fallback_model = agent._fallback_chain[0] if agent._fallback_chain else None + if agent._fallback_chain and not agent.quiet_mode: + if len(agent._fallback_chain) == 1: + fb = agent._fallback_chain[0] + print(f"🔄 Fallback model: {fb['model']} ({fb['provider']})") + else: + print(f"🔄 Fallback chain ({len(agent._fallback_chain)} providers): " + + " → ".join(f"{f['model']} ({f['provider']})" for f in agent._fallback_chain)) + + # Get available tools with filtering + agent.tools = _ra().get_tool_definitions( + enabled_toolsets=enabled_toolsets, + disabled_toolsets=disabled_toolsets, + quiet_mode=agent.quiet_mode, + ) + + # Show tool configuration and store valid tool names for validation + agent.valid_tool_names = set() + if agent.tools: + agent.valid_tool_names = {tool["function"]["name"] for tool in agent.tools} + tool_names = sorted(agent.valid_tool_names) + if not agent.quiet_mode: + print(f"🛠️ Loaded {len(agent.tools)} tools: {', '.join(tool_names)}") + + # Show filtering info if applied + if enabled_toolsets: + print(f" ✅ Enabled toolsets: {', '.join(enabled_toolsets)}") + if disabled_toolsets: + print(f" ❌ Disabled toolsets: {', '.join(disabled_toolsets)}") + elif not agent.quiet_mode: + print("🛠️ No tools loaded (all tools filtered out or unavailable)") + + # Check tool requirements + if agent.tools and not agent.quiet_mode: + requirements = _ra().check_toolset_requirements() + missing_reqs = [name for name, available in requirements.items() if not available] + if missing_reqs: + print(f"⚠️ Some tools may not work due to missing requirements: {missing_reqs}") + + # Show trajectory saving status + if agent.save_trajectories and not agent.quiet_mode: + print("📝 Trajectory saving enabled") + + # Show ephemeral system prompt status + if agent.ephemeral_system_prompt and not agent.quiet_mode: + prompt_preview = agent.ephemeral_system_prompt[:60] + "..." if len(agent.ephemeral_system_prompt) > 60 else agent.ephemeral_system_prompt + print(f"🔒 Ephemeral system prompt: '{prompt_preview}' (not saved to trajectories)") + + # Show prompt caching status + if agent._use_prompt_caching and not agent.quiet_mode: + if agent._use_native_cache_layout and agent.provider == "anthropic": + source = "native Anthropic" + elif agent._use_native_cache_layout: + source = "Anthropic-compatible endpoint" + else: + source = "Claude via OpenRouter" + print(f"💾 Prompt caching: ENABLED ({source}, {agent._cache_ttl} TTL)") + + # Session logging setup - auto-save conversation trajectories for debugging + agent.session_start = datetime.now() + if session_id: + # Use provided session ID (e.g., from CLI) + agent.session_id = session_id + else: + # Generate a new session ID + timestamp_str = agent.session_start.strftime("%Y%m%d_%H%M%S") + short_uuid = uuid.uuid4().hex[:6] + agent.session_id = f"{timestamp_str}_{short_uuid}" + + # Expose session ID to tools (terminal, execute_code) so agents can + # reference their own session for --resume commands, cross-session + # coordination, and logging. Uses the ContextVar system from + # session_context.py for concurrency safety (gateway runs multiple + # sessions in one process). Also writes os.environ as fallback for + # CLI mode where ContextVars aren't used. + os.environ["HERMES_SESSION_ID"] = agent.session_id + try: + from gateway.session_context import _SESSION_ID + _SESSION_ID.set(agent.session_id) + except Exception: + pass # CLI/test mode — ContextVar not needed + + # Session logs go into ~/.hermes/sessions/ alongside gateway sessions + hermes_home = get_hermes_home() + agent.logs_dir = hermes_home / "sessions" + agent.logs_dir.mkdir(parents=True, exist_ok=True) + agent.session_log_file = agent.logs_dir / f"session_{agent.session_id}.json" + + # Track conversation messages for session logging + agent._session_messages: List[Dict[str, Any]] = [] + agent._memory_write_origin = "assistant_tool" + agent._memory_write_context = "foreground" + + # Cached system prompt -- built once per session, only rebuilt on compression + agent._cached_system_prompt: Optional[str] = None + + # Filesystem checkpoint manager (transparent — not a tool) + from tools.checkpoint_manager import CheckpointManager + agent._checkpoint_mgr = CheckpointManager( + enabled=checkpoints_enabled, + max_snapshots=checkpoint_max_snapshots, + max_total_size_mb=checkpoint_max_total_size_mb, + max_file_size_mb=checkpoint_max_file_size_mb, + ) + + # SQLite session store (optional -- provided by CLI or gateway) + agent._session_db = session_db + agent._parent_session_id = parent_session_id + agent._last_flushed_db_idx = 0 # tracks DB-write cursor to prevent duplicate writes + agent._session_db_created = False # DB row deferred to run_conversation() + agent._session_init_model_config = { + "max_iterations": agent.max_iterations, + "reasoning_config": reasoning_config, + "max_tokens": max_tokens, + } + + # In-memory todo list for task planning (one per agent/session) + from tools.todo_tool import TodoStore + agent._todo_store = TodoStore() + + # Load config once for memory, skills, and compression sections + try: + from hermes_cli.config import load_config as _load_agent_config + _agent_cfg = _load_agent_config() + except Exception: + _agent_cfg = {} + try: + agent._tool_guardrails = ToolCallGuardrailController( + ToolCallGuardrailConfig.from_mapping( + _agent_cfg.get("tool_loop_guardrails", {}) + ) + ) + except Exception as _tlg_err: + _ra().logger.warning("Tool loop guardrail config ignored: %s", _tlg_err) + # Cache only the derived auxiliary compression context override that is + # needed later by the startup feasibility check. Avoid exposing a + # broad pseudo-public config object on the agent instance. + agent._aux_compression_context_length_config = None + + # Persistent memory (MEMORY.md + USER.md) -- loaded from disk + agent._memory_store = None + agent._memory_enabled = False + agent._user_profile_enabled = False + agent._memory_nudge_interval = 10 + agent._turns_since_memory = 0 + agent._iters_since_skill = 0 + if not skip_memory: + try: + mem_config = _agent_cfg.get("memory", {}) + agent._memory_enabled = mem_config.get("memory_enabled", False) + agent._user_profile_enabled = mem_config.get("user_profile_enabled", False) + agent._memory_nudge_interval = int(mem_config.get("nudge_interval", 10)) + if agent._memory_enabled or agent._user_profile_enabled: + from tools.memory_tool import MemoryStore + agent._memory_store = MemoryStore( + memory_char_limit=mem_config.get("memory_char_limit", 2200), + user_char_limit=mem_config.get("user_char_limit", 1375), + ) + agent._memory_store.load_from_disk() + except Exception: + pass # Memory is optional -- don't break agent init + + + + # Memory provider plugin (external — one at a time, alongside built-in) + # Reads memory.provider from config to select which plugin to activate. + agent._memory_manager = None + if not skip_memory: + try: + _mem_provider_name = mem_config.get("provider", "") if mem_config else "" + + if _mem_provider_name and _mem_provider_name.strip(): + from agent.memory_manager import MemoryManager as _MemoryManager + from plugins.memory import load_memory_provider as _load_mem + agent._memory_manager = _MemoryManager() + _mp = _load_mem(_mem_provider_name) + if _mp and _mp.is_available(): + agent._memory_manager.add_provider(_mp) + if agent._memory_manager.providers: + _init_kwargs = { + "session_id": agent.session_id, + "platform": platform or "cli", + "hermes_home": str(get_hermes_home()), + "agent_context": "primary", + } + # Thread session title for memory provider scoping + # (e.g. honcho uses this to derive chat-scoped session keys) + if agent._session_db: + try: + _st = agent._session_db.get_session_title(agent.session_id) + if _st: + _init_kwargs["session_title"] = _st + except Exception: + pass + # Thread gateway user identity for per-user memory scoping + if agent._user_id: + _init_kwargs["user_id"] = agent._user_id + if agent._user_name: + _init_kwargs["user_name"] = agent._user_name + if agent._chat_id: + _init_kwargs["chat_id"] = agent._chat_id + if agent._chat_name: + _init_kwargs["chat_name"] = agent._chat_name + if agent._chat_type: + _init_kwargs["chat_type"] = agent._chat_type + if agent._thread_id: + _init_kwargs["thread_id"] = agent._thread_id + # Thread gateway session key for stable per-chat Honcho session isolation + if agent._gateway_session_key: + _init_kwargs["gateway_session_key"] = agent._gateway_session_key + # Profile identity for per-profile provider scoping + try: + from hermes_cli.profiles import get_active_profile_name + _profile = get_active_profile_name() + _init_kwargs["agent_identity"] = _profile + _init_kwargs["agent_workspace"] = "hermes" + except Exception: + pass + agent._memory_manager.initialize_all(**_init_kwargs) + _ra().logger.info("Memory provider '%s' activated", _mem_provider_name) + else: + _ra().logger.debug("Memory provider '%s' not found or not available", _mem_provider_name) + agent._memory_manager = None + except Exception as _mpe: + _ra().logger.warning("Memory provider plugin init failed: %s", _mpe) + agent._memory_manager = None + + # Inject memory provider tool schemas into the tool surface. + # Skip tools whose names already exist (plugins may register the + # same tools via ctx.register_tool(), which lands in agent.tools + # through _ra().get_tool_definitions()). Duplicate function names cause + # 400 errors on providers that enforce unique names (e.g. Xiaomi + # MiMo via Nous Portal). + if agent._memory_manager and agent.tools is not None: + _existing_tool_names = { + t.get("function", {}).get("name") + for t in agent.tools + if isinstance(t, dict) + } + for _schema in agent._memory_manager.get_all_tool_schemas(): + _tname = _schema.get("name", "") + if _tname and _tname in _existing_tool_names: + continue # already registered via plugin path + _wrapped = {"type": "function", "function": _schema} + agent.tools.append(_wrapped) + if _tname: + agent.valid_tool_names.add(_tname) + _existing_tool_names.add(_tname) + + # Skills config: nudge interval for skill creation reminders + agent._skill_nudge_interval = 10 + try: + skills_config = _agent_cfg.get("skills", {}) + agent._skill_nudge_interval = int(skills_config.get("creation_nudge_interval", 10)) + except Exception: + pass + + # Tool-use enforcement config: "auto" (default — matches hardcoded + # model list), true (always), false (never), or list of substrings. + _agent_section = _agent_cfg.get("agent", {}) + if not isinstance(_agent_section, dict): + _agent_section = {} + agent._tool_use_enforcement = _agent_section.get("tool_use_enforcement", "auto") + + # App-level API retry count (wraps each model API call). Default 3, + # overridable via agent.api_max_retries in config.yaml. See #11616. + try: + _raw_api_retries = _agent_section.get("api_max_retries", 3) + _api_retries = int(_raw_api_retries) + _api_retries = max(_api_retries, 1) # 1 = no retry (single attempt) + except (TypeError, ValueError): + _api_retries = 3 + agent._api_max_retries = _api_retries + + # Initialize context compressor for automatic context management + # Compresses conversation when approaching model's context limit + # Configuration via config.yaml (compression section) + _compression_cfg = _agent_cfg.get("compression", {}) + if not isinstance(_compression_cfg, dict): + _compression_cfg = {} + compression_threshold = float(_compression_cfg.get("threshold", 0.50)) + try: + from agent.auxiliary_client import _compression_threshold_for_model as _cthresh_fn + _model_cthresh = _cthresh_fn(agent.model) + if _model_cthresh is not None: + compression_threshold = _model_cthresh + except Exception: + pass + compression_enabled = str(_compression_cfg.get("enabled", True)).lower() in {"true", "1", "yes"} + compression_target_ratio = float(_compression_cfg.get("target_ratio", 0.20)) + compression_protect_last = int(_compression_cfg.get("protect_last_n", 20)) + # protect_first_n is the number of non-system messages to protect at + # the head, in addition to the system prompt (which is always + # implicitly protected by the compressor). Floor at 0 — a value of + # 0 means "preserve only the system prompt + summary + tail", which + # is a legitimate (and common) configuration for long-running + # rolling-compaction sessions. + compression_protect_first = max( + 0, int(_compression_cfg.get("protect_first_n", 3)) + ) + + # Read optional explicit context_length override for the auxiliary + # compression model. Custom endpoints often cannot report this via + # /models, so the startup feasibility check needs the config hint. + try: + _aux_cfg = cfg_get(_agent_cfg, "auxiliary", "compression", default={}) + except Exception: + _aux_cfg = {} + if isinstance(_aux_cfg, dict): + _aux_context_config = _aux_cfg.get("context_length") + else: + _aux_context_config = None + if _aux_context_config is not None: + try: + _aux_context_config = int(_aux_context_config) + except (TypeError, ValueError): + _aux_context_config = None + agent._aux_compression_context_length_config = _aux_context_config + + # Read explicit model output-token override from config when the + # caller did not pass one directly. + _model_cfg = _agent_cfg.get("model", {}) + if agent.max_tokens is None and isinstance(_model_cfg, dict): + _config_max_tokens = _model_cfg.get("max_tokens") + if _config_max_tokens is not None: + try: + if isinstance(_config_max_tokens, bool): + raise ValueError + _parsed_max_tokens = int(_config_max_tokens) + if _parsed_max_tokens <= 0: + raise ValueError + agent.max_tokens = _parsed_max_tokens + except (TypeError, ValueError): + _ra().logger.warning( + "Invalid model.max_tokens in config.yaml: %r — " + "must be a positive integer (e.g. 4096). " + "Falling back to provider default.", + _config_max_tokens, + ) + print( + f"\n⚠ Invalid model.max_tokens in config.yaml: {_config_max_tokens!r}\n" + f" Must be a positive integer (e.g. 4096).\n" + f" Falling back to provider default.\n", + file=sys.stderr, + ) + agent._session_init_model_config["max_tokens"] = agent.max_tokens + + # Read explicit context_length override from model config + if isinstance(_model_cfg, dict): + _config_context_length = _model_cfg.get("context_length") + else: + _config_context_length = None + if _config_context_length is not None: + try: + _config_context_length = int(_config_context_length) + except (TypeError, ValueError): + _ra().logger.warning( + "Invalid model.context_length in config.yaml: %r — " + "must be a plain integer (e.g. 256000, not '256K'). " + "Falling back to auto-detection.", + _config_context_length, + ) + print( + f"\n⚠ Invalid model.context_length in config.yaml: {_config_context_length!r}\n" + f" Must be a plain integer (e.g. 256000, not '256K').\n" + f" Falling back to auto-detected context window.\n", + file=sys.stderr, + ) + _config_context_length = None + + # Resolve custom_providers list once for reuse below (startup + # context-length override and plugin context-engine init). + try: + from hermes_cli.config import get_compatible_custom_providers + _custom_providers = get_compatible_custom_providers(_agent_cfg) + except Exception: + _custom_providers = _agent_cfg.get("custom_providers") + if not isinstance(_custom_providers, list): + _custom_providers = [] + + # Store for reuse by _check_compression_model_feasibility (auxiliary + # compression model context-length detection needs the same list). + agent._custom_providers = _custom_providers + + # Check custom_providers per-model context_length + if _config_context_length is None and _custom_providers: + try: + from hermes_cli.config import get_custom_provider_context_length + _cp_ctx_resolved = get_custom_provider_context_length( + model=agent.model, + base_url=agent.base_url, + custom_providers=_custom_providers, + ) + if _cp_ctx_resolved: + _config_context_length = int(_cp_ctx_resolved) + except Exception: + _cp_ctx_resolved = None + + # Surface a clear warning if the user set a context_length but it + # wasn't a valid positive int — the helper silently skips those. + if _config_context_length is None: + _target = agent.base_url.rstrip("/") if agent.base_url else "" + for _cp_entry in _custom_providers: + if not isinstance(_cp_entry, dict): + continue + _cp_url = (_cp_entry.get("base_url") or "").rstrip("/") + if _target and _cp_url == _target: + _cp_models = _cp_entry.get("models", {}) + if isinstance(_cp_models, dict): + _cp_model_cfg = _cp_models.get(agent.model, {}) + if isinstance(_cp_model_cfg, dict): + _cp_ctx = _cp_model_cfg.get("context_length") + if _cp_ctx is not None: + try: + _parsed = int(_cp_ctx) + if _parsed <= 0: + raise ValueError + except (TypeError, ValueError): + _ra().logger.warning( + "Invalid context_length for model %r in " + "custom_providers: %r — must be a positive " + "integer (e.g. 256000, not '256K'). " + "Falling back to auto-detection.", + agent.model, _cp_ctx, + ) + print( + f"\n⚠ Invalid context_length for model {agent.model!r} in custom_providers: {_cp_ctx!r}\n" + f" Must be a positive integer (e.g. 256000, not '256K').\n" + f" Falling back to auto-detected context window.\n", + file=sys.stderr, + ) + break + + # Persist for reuse on switch_model / fallback activation. Must come + # AFTER the custom_providers branch so per-model overrides aren't lost. + agent._config_context_length = _config_context_length + + agent._ensure_lmstudio_runtime_loaded(_config_context_length) + + + + # Select context engine: config-driven (like memory providers). + # 1. Check config.yaml context.engine setting + # 2. Check plugins/context_engine// directory (repo-shipped) + # 3. Check general plugin system (user-installed plugins) + # 4. Fall back to built-in ContextCompressor + _selected_engine = None + _engine_name = "compressor" # default + try: + _ctx_cfg = _agent_cfg.get("context", {}) if isinstance(_agent_cfg, dict) else {} + _engine_name = _ctx_cfg.get("engine", "compressor") or "compressor" + except Exception: + pass + + if _engine_name != "compressor": + # Try loading from plugins/context_engine// + try: + from plugins.context_engine import load_context_engine + _selected_engine = load_context_engine(_engine_name) + except Exception as _ce_load_err: + _ra().logger.debug("Context engine load from plugins/context_engine/: %s", _ce_load_err) + + # Try general plugin system as fallback + if _selected_engine is None: + try: + from hermes_cli.plugins import get_plugin_context_engine + _candidate = get_plugin_context_engine() + if _candidate and _candidate.name == _engine_name: + _selected_engine = _candidate + except Exception: + pass + + if _selected_engine is None: + _ra().logger.warning( + "Context engine '%s' not found — falling back to built-in compressor", + _engine_name, + ) + # else: config says "compressor" — use built-in, don't auto-activate plugins + + if _selected_engine is not None: + agent.context_compressor = _selected_engine + # Resolve context_length for plugin engines — mirrors switch_model() path + from agent.model_metadata import get_model_context_length + _plugin_ctx_len = get_model_context_length( + agent.model, + base_url=agent.base_url, + api_key=getattr(agent, "api_key", ""), + config_context_length=_config_context_length, + provider=agent.provider, + custom_providers=_custom_providers, + ) + agent.context_compressor.update_model( + model=agent.model, + context_length=_plugin_ctx_len, + base_url=agent.base_url, + api_key=getattr(agent, "api_key", ""), + provider=agent.provider, + ) + if not agent.quiet_mode: + _ra().logger.info("Using context engine: %s", _selected_engine.name) + else: + agent.context_compressor = ContextCompressor( + model=agent.model, + threshold_percent=compression_threshold, + protect_first_n=compression_protect_first, + protect_last_n=compression_protect_last, + summary_target_ratio=compression_target_ratio, + summary_model_override=None, + quiet_mode=agent.quiet_mode, + base_url=agent.base_url, + api_key=getattr(agent, "api_key", ""), + config_context_length=_config_context_length, + provider=agent.provider, + api_mode=agent.api_mode, + ) + agent.compression_enabled = compression_enabled + + # Reject models whose context window is below the minimum required + # for reliable tool-calling workflows (64K tokens). + from agent.model_metadata import MINIMUM_CONTEXT_LENGTH + _ctx = getattr(agent.context_compressor, "context_length", 0) + if _ctx and _ctx < MINIMUM_CONTEXT_LENGTH: + raise ValueError( + f"Model {agent.model} has a context window of {_ctx:,} tokens, " + f"which is below the minimum {MINIMUM_CONTEXT_LENGTH:,} required " + f"by Hermes Agent. Choose a model with at least " + f"{MINIMUM_CONTEXT_LENGTH // 1000}K context, or set " + f"model.context_length in config.yaml to override." + ) + + # Inject context engine tool schemas (e.g. lcm_grep, lcm_describe, lcm_expand). + # Skip names that are already present — the _ra().get_tool_definitions() + # quiet_mode cache returned a shared list pre-#17335, so a stray + # mutation here would poison subsequent agent inits in the same + # Gateway process and trip provider-side 'duplicate tool name' + # errors. Even with the cache fix, dedup is the right defense + # against plugin paths that may register the same schemas via + # ctx.register_tool(). Mirrors the memory tools dedup above. + agent._context_engine_tool_names: set = set() + if hasattr(agent, "context_compressor") and agent.context_compressor and agent.tools is not None: + _existing_tool_names = { + t.get("function", {}).get("name") + for t in agent.tools + if isinstance(t, dict) + } + for _schema in agent.context_compressor.get_tool_schemas(): + _tname = _schema.get("name", "") + if _tname and _tname in _existing_tool_names: + continue # already registered via plugin/cache path + _wrapped = {"type": "function", "function": _schema} + agent.tools.append(_wrapped) + if _tname: + agent.valid_tool_names.add(_tname) + agent._context_engine_tool_names.add(_tname) + _existing_tool_names.add(_tname) + + # Notify context engine of session start + if hasattr(agent, "context_compressor") and agent.context_compressor: + try: + agent.context_compressor.on_session_start( + agent.session_id, + hermes_home=str(get_hermes_home()), + platform=agent.platform or "cli", + model=agent.model, + context_length=getattr(agent.context_compressor, "context_length", 0), + ) + except Exception as _ce_err: + _ra().logger.debug("Context engine on_session_start: %s", _ce_err) + + agent._subdirectory_hints = SubdirectoryHintTracker( + working_dir=os.getenv("TERMINAL_CWD") or None, + ) + agent._user_turn_count = 0 + + # Cumulative token usage for the session + agent.session_prompt_tokens = 0 + agent.session_completion_tokens = 0 + agent.session_total_tokens = 0 + agent.session_api_calls = 0 + agent.session_input_tokens = 0 + agent.session_output_tokens = 0 + agent.session_cache_read_tokens = 0 + agent.session_cache_write_tokens = 0 + agent.session_reasoning_tokens = 0 + agent.session_estimated_cost_usd = 0.0 + agent.session_cost_status = "unknown" + agent.session_cost_source = "none" + + # ── Ollama num_ctx injection ── + # Ollama defaults to 2048 context regardless of the model's capabilities. + # When running against an Ollama server, detect the model's max context + # and pass num_ctx on every chat request so the full window is used. + # User override: set model.ollama_num_ctx in config.yaml to cap VRAM use. + # If model.context_length is set, it caps num_ctx so the user's VRAM + # budget is respected even when GGUF metadata advertises a larger window. + agent._ollama_num_ctx: int | None = None + _ollama_num_ctx_override = None + if isinstance(_model_cfg, dict): + _ollama_num_ctx_override = _model_cfg.get("ollama_num_ctx") + if _ollama_num_ctx_override is not None: + try: + agent._ollama_num_ctx = int(_ollama_num_ctx_override) + except (TypeError, ValueError): + _ra().logger.debug("Invalid ollama_num_ctx config value: %r", _ollama_num_ctx_override) + if agent._ollama_num_ctx is None and agent.base_url and is_local_endpoint(agent.base_url): + try: + _detected = query_ollama_num_ctx(agent.model, agent.base_url, api_key=agent.api_key or "") + if _detected and _detected > 0: + agent._ollama_num_ctx = _detected + except Exception as exc: + _ra().logger.debug("Ollama num_ctx detection failed: %s", exc) + # Cap auto-detected ollama_num_ctx to the user's explicit context_length. + # Without this, GGUF metadata can advertise 256K+ which Ollama honours + # by allocating that much VRAM — blowing up small GPUs even though the + # user explicitly set a smaller context_length in config.yaml. + if ( + agent._ollama_num_ctx + and _config_context_length + and _ollama_num_ctx_override is None # don't override explicit ollama_num_ctx + and agent._ollama_num_ctx > _config_context_length + ): + _ra().logger.info( + "Ollama num_ctx capped: %d -> %d (model.context_length override)", + agent._ollama_num_ctx, _config_context_length, + ) + agent._ollama_num_ctx = _config_context_length + if agent._ollama_num_ctx and not agent.quiet_mode: + _ra().logger.info( + "Ollama num_ctx: will request %d tokens (model max from /api/show)", + agent._ollama_num_ctx, + ) + + if not agent.quiet_mode: + if compression_enabled: + print(f"📊 Context limit: {agent.context_compressor.context_length:,} tokens (compress at {int(compression_threshold*100)}% = {agent.context_compressor.threshold_tokens:,})") + else: + print(f"📊 Context limit: {agent.context_compressor.context_length:,} tokens (auto-compression disabled)") + + # Check immediately so CLI users see the warning at startup. + # Gateway status_callback is not yet wired, so any warning is stored + # in _compression_warning and replayed in the first run_conversation(). + agent._compression_warning = None + agent._check_compression_model_feasibility() + + # Snapshot primary runtime for per-turn restoration. When fallback + # activates during a turn, the next turn restores these values so the + # preferred model gets a fresh attempt each time. Uses a single dict + # so new state fields are easy to add without N individual attributes. + _cc = agent.context_compressor + agent._primary_runtime = { + "model": agent.model, + "provider": agent.provider, + "base_url": agent.base_url, + "api_mode": agent.api_mode, + "api_key": getattr(agent, "api_key", ""), + "client_kwargs": dict(agent._client_kwargs), + "use_prompt_caching": agent._use_prompt_caching, + "use_native_cache_layout": agent._use_native_cache_layout, + # Context engine state that _try_activate_fallback() overwrites. + # Use getattr for model/base_url/api_key/provider since plugin + # engines may not have these (they're ContextCompressor-specific). + "compressor_model": getattr(_cc, "model", agent.model), + "compressor_base_url": getattr(_cc, "base_url", agent.base_url), + "compressor_api_key": getattr(_cc, "api_key", ""), + "compressor_provider": getattr(_cc, "provider", agent.provider), + "compressor_context_length": _cc.context_length, + "compressor_threshold_tokens": _cc.threshold_tokens, + } + if agent.api_mode == "anthropic_messages": + agent._primary_runtime.update({ + "anthropic_api_key": agent._anthropic_api_key, + "anthropic_base_url": agent._anthropic_base_url, + "is_anthropic_oauth": agent._is_anthropic_oauth, + }) + + + +__all__ = ["init_agent"] diff --git a/agent/agent_runtime_helpers.py b/agent/agent_runtime_helpers.py new file mode 100644 index 00000000000..b5c70392946 --- /dev/null +++ b/agent/agent_runtime_helpers.py @@ -0,0 +1,2134 @@ +"""Assorted AIAgent runtime helpers — moved out of run_agent.py for clarity. + +Each function takes the parent ``AIAgent`` as its first argument +(``agent``) except for the static helpers (``sanitize_tool_call_arguments``, +``drop_thinking_only_and_merge_users``) which are stateless. AIAgent +keeps thin forwarders for backward compatibility. + +Methods covered: +* ``convert_to_trajectory_format`` — internal -> trajectory-file format +* ``sanitize_tool_call_arguments`` — repair corrupted JSON in tool_calls +* ``repair_message_sequence`` — enforce alternation invariants +* ``strip_think_blocks`` — remove inline reasoning from stored content +* ``recover_with_credential_pool`` — rotate pool entries on 429 +* ``try_recover_primary_transport`` — re-create OpenAI client after rate-limit +* ``drop_thinking_only_and_merge_users`` — Anthropic-style cleanup +* ``restore_primary_runtime`` — un-do fallback activation +* ``extract_reasoning`` — pull reasoning fields out of API responses +* ``dump_api_request_debug`` — write request body for post-mortem +* ``anthropic_prompt_cache_policy`` — compute cache_control breakpoints +* ``create_openai_client`` — build the per-agent OpenAI SDK client +""" + +from __future__ import annotations + +import copy +import json +import logging +import os +import re +import threading +import time +import uuid +from datetime import datetime +from pathlib import Path +from typing import Any, Dict, List, Optional, Tuple + +from hermes_cli.timeouts import get_provider_request_timeout +from agent.message_sanitization import ( + _repair_tool_call_arguments, + _sanitize_surrogates, +) +from agent.tool_dispatch_helpers import _trajectory_normalize_msg +from agent.trajectory import convert_scratchpad_to_think +from agent.error_classifier import classify_api_error, FailoverReason +from utils import base_url_host_matches, base_url_hostname, env_var_enabled, atomic_json_write + +logger = logging.getLogger(__name__) + + +def _ra(): + """Lazy ``run_agent`` reference for test-patch routing.""" + import run_agent + return run_agent + + + +def convert_to_trajectory_format(agent, messages: List[Dict[str, Any]], user_query: str, completed: bool) -> List[Dict[str, Any]]: + """ + Convert internal message format to trajectory format for saving. + + Args: + messages (List[Dict]): Internal message history + user_query (str): Original user query + completed (bool): Whether the conversation completed successfully + + Returns: + List[Dict]: Messages in trajectory format + """ + # Normalize multimodal tool results — trajectories are text-only, so + # replace image-bearing tool messages with their text_summary to avoid + # embedding ~1MB base64 blobs into every saved trajectory. + messages = [_trajectory_normalize_msg(m) for m in messages] + trajectory = [] + + # Add system message with tool definitions + system_msg = ( + "You are a function calling AI model. You are provided with function signatures within XML tags. " + "You may call one or more functions to assist with the user query. If available tools are not relevant in assisting " + "with user query, just respond in natural conversational language. Don't make assumptions about what values to plug " + "into functions. After calling & executing the functions, you will be provided with function results within " + " XML tags. Here are the available tools:\n" + f"\n{agent._format_tools_for_system_message()}\n\n" + "For each function call return a JSON object, with the following pydantic model json schema for each:\n" + "{'title': 'FunctionCall', 'type': 'object', 'properties': {'name': {'title': 'Name', 'type': 'string'}, " + "'arguments': {'title': 'Arguments', 'type': 'object'}}, 'required': ['name', 'arguments']}\n" + "Each function call should be enclosed within XML tags.\n" + "Example:\n\n{'name': ,'arguments': }\n" + ) + + trajectory.append({ + "from": "system", + "value": system_msg + }) + + # Add the actual user prompt (from the dataset) as the first human message + trajectory.append({ + "from": "human", + "value": user_query + }) + + # Skip the first message (the user query) since we already added it above. + # Prefill messages are injected at API-call time only (not in the messages + # list), so no offset adjustment is needed here. + i = 1 + + while i < len(messages): + msg = messages[i] + + if msg["role"] == "assistant": + # Check if this message has tool calls + if "tool_calls" in msg and msg["tool_calls"]: + # Format assistant message with tool calls + # Add tags around reasoning for trajectory storage + content = "" + + # Prepend reasoning in tags if available (native thinking tokens) + if msg.get("reasoning") and msg["reasoning"].strip(): + content = f"\n{msg['reasoning']}\n\n" + + if msg.get("content") and msg["content"].strip(): + # Convert any tags to tags + # (used when native thinking is disabled and model reasons via XML) + content += convert_scratchpad_to_think(msg["content"]) + "\n" + + # Add tool calls wrapped in XML tags + for tool_call in msg["tool_calls"]: + if not tool_call or not isinstance(tool_call, dict): continue + # Parse arguments - should always succeed since we validate during conversation + # but keep try-except as safety net + try: + arguments = json.loads(tool_call["function"]["arguments"]) if isinstance(tool_call["function"]["arguments"], str) else tool_call["function"]["arguments"] + except json.JSONDecodeError: + # This shouldn't happen since we validate and retry during conversation, + # but if it does, log warning and use empty dict + logging.warning(f"Unexpected invalid JSON in trajectory conversion: {tool_call['function']['arguments'][:100]}") + arguments = {} + + tool_call_json = { + "name": tool_call["function"]["name"], + "arguments": arguments + } + content += f"\n{json.dumps(tool_call_json, ensure_ascii=False)}\n\n" + + # Ensure every gpt turn has a block (empty if no reasoning) + # so the format is consistent for training data + if "" not in content: + content = "\n\n" + content + + trajectory.append({ + "from": "gpt", + "value": content.rstrip() + }) + + # Collect all subsequent tool responses + tool_responses = [] + j = i + 1 + while j < len(messages) and messages[j]["role"] == "tool": + tool_msg = messages[j] + # Format tool response with XML tags + tool_response = "\n" + + # Try to parse tool content as JSON if it looks like JSON + tool_content = tool_msg["content"] + try: + if tool_content.strip().startswith(("{", "[")): + tool_content = json.loads(tool_content) + except (json.JSONDecodeError, AttributeError): + pass # Keep as string if not valid JSON + + tool_index = len(tool_responses) + tool_name = ( + msg["tool_calls"][tool_index]["function"]["name"] + if tool_index < len(msg["tool_calls"]) + else "unknown" + ) + tool_response += json.dumps({ + "tool_call_id": tool_msg.get("tool_call_id", ""), + "name": tool_name, + "content": tool_content + }, ensure_ascii=False) + tool_response += "\n" + tool_responses.append(tool_response) + j += 1 + + # Add all tool responses as a single message + if tool_responses: + trajectory.append({ + "from": "tool", + "value": "\n".join(tool_responses) + }) + i = j - 1 # Skip the tool messages we just processed + + else: + # Regular assistant message without tool calls + # Add tags around reasoning for trajectory storage + content = "" + + # Prepend reasoning in tags if available (native thinking tokens) + if msg.get("reasoning") and msg["reasoning"].strip(): + content = f"\n{msg['reasoning']}\n\n" + + # Convert any tags to tags + # (used when native thinking is disabled and model reasons via XML) + raw_content = msg["content"] or "" + content += convert_scratchpad_to_think(raw_content) + + # Ensure every gpt turn has a block (empty if no reasoning) + if "" not in content: + content = "\n\n" + content + + trajectory.append({ + "from": "gpt", + "value": content.strip() + }) + + elif msg["role"] == "user": + trajectory.append({ + "from": "human", + "value": msg["content"] + }) + + i += 1 + + return trajectory + + + +def sanitize_tool_call_arguments( + messages: list, + *, + logger=None, + session_id: str = None, +) -> int: + """Repair corrupted assistant tool-call argument JSON in-place.""" + log = logger or logging.getLogger(__name__) + if not isinstance(messages, list): + return 0 + + repaired = 0 + marker = _ra().AIAgent._TOOL_CALL_ARGUMENTS_CORRUPTION_MARKER + + def _prepend_marker(tool_msg: dict) -> None: + existing = tool_msg.get("content") + if isinstance(existing, str): + if not existing: + tool_msg["content"] = marker + elif not existing.startswith(marker): + tool_msg["content"] = f"{marker}\n{existing}" + return + if existing is None: + tool_msg["content"] = marker + return + try: + existing_text = json.dumps(existing) + except TypeError: + existing_text = str(existing) + tool_msg["content"] = f"{marker}\n{existing_text}" + + message_index = 0 + while message_index < len(messages): + msg = messages[message_index] + if not isinstance(msg, dict) or msg.get("role") != "assistant": + message_index += 1 + continue + + tool_calls = msg.get("tool_calls") + if not isinstance(tool_calls, list) or not tool_calls: + message_index += 1 + continue + + insert_at = message_index + 1 + for tool_call in tool_calls: + if not isinstance(tool_call, dict): + continue + function = tool_call.get("function") + if not isinstance(function, dict): + continue + + arguments = function.get("arguments") + if arguments is None or arguments == "": + function["arguments"] = "{}" + continue + if isinstance(arguments, str) and not arguments.strip(): + function["arguments"] = "{}" + continue + if not isinstance(arguments, str): + continue + + try: + json.loads(arguments) + except json.JSONDecodeError: + tool_call_id = tool_call.get("id") + function_name = function.get("name", "?") + preview = arguments[:80] + log.warning( + "Corrupted tool_call arguments repaired before request " + "(session=%s, message_index=%s, tool_call_id=%s, function=%s, preview=%r)", + session_id or "-", + message_index, + tool_call_id or "-", + function_name, + preview, + ) + function["arguments"] = "{}" + + existing_tool_msg = None + scan_index = message_index + 1 + while scan_index < len(messages): + candidate = messages[scan_index] + if not isinstance(candidate, dict) or candidate.get("role") != "tool": + break + if candidate.get("tool_call_id") == tool_call_id: + existing_tool_msg = candidate + break + scan_index += 1 + + if existing_tool_msg is None: + messages.insert( + insert_at, + { + "role": "tool", + "name": function_name if function_name != "?" else "", + "tool_call_id": tool_call_id, + "content": marker, + }, + ) + insert_at += 1 + else: + _prepend_marker(existing_tool_msg) + + repaired += 1 + + message_index += 1 + + return repaired + + + +def repair_message_sequence(agent, messages: List[Dict]) -> int: + """Collapse malformed role-alternation left in the live history. + + Providers (OpenAI, OpenRouter, Anthropic) expect strict alternation: + after the system message, user/tool alternates with assistant, with + no two consecutive user messages and no tool-result that doesn't + follow an assistant-with-tool_calls. Violations cause silent empty + responses on most providers, which triggers the empty-retry loop. + + This runs right before the API call as a defensive belt — by the + time it fires, the scaffolding strip should already have prevented + most shapes, but external callers (gateway multi-queue replay, + session resume, cron, explicit conversation_history passed in by + host code) can feed in already-broken histories. + + Repairs applied: + 1. Stray ``tool`` messages whose ``tool_call_id`` doesn't match + any preceding assistant tool_call — dropped. + 2. Consecutive ``user`` messages — merged with newline separator + so no user input is lost. + + Deliberately does NOT rewind orphan ``assistant(tool_calls)+tool`` + pairs that precede a user message — that pattern IS valid when the + previous turn completed normally and the user jumped in to redirect + before the model got a continuation turn (the ongoing dialog + pattern). The empty-response scaffolding stripper handles the + genuinely-broken variant via its flag-gated rewind. + + Returns the number of repairs made (for logging/telemetry). + """ + if not messages: + return 0 + + repairs = 0 + + # Pass 1: drop stray tool messages that don't follow a known + # assistant tool_call_id. Uses a rolling set of known ids refreshed + # on each assistant message. + known_tool_ids: set = set() + filtered: List[Dict] = [] + for msg in messages: + if not isinstance(msg, dict): + filtered.append(msg) + continue + role = msg.get("role") + if role == "assistant": + known_tool_ids = set() + for tc in (msg.get("tool_calls") or []): + tc_id = tc.get("id") if isinstance(tc, dict) else None + if tc_id: + known_tool_ids.add(tc_id) + filtered.append(msg) + elif role == "tool": + tc_id = msg.get("tool_call_id") + if tc_id and tc_id in known_tool_ids: + filtered.append(msg) + else: + repairs += 1 + else: + if role == "user": + # A user turn closes the tool-result run; subsequent + # tool messages without a fresh assistant tool_call + # are orphans. + known_tool_ids = set() + filtered.append(msg) + + # Pass 2: merge consecutive user messages. Preserves all user input + # so nothing the user typed is lost. + merged: List[Dict] = [] + for msg in filtered: + if ( + merged + and isinstance(msg, dict) + and msg.get("role") == "user" + and isinstance(merged[-1], dict) + and merged[-1].get("role") == "user" + ): + prev = merged[-1] + prev_content = prev.get("content", "") + new_content = msg.get("content", "") + # Only merge plain-text content; leave multimodal (list) + # content alone — collapsing image/audio blocks risks + # mangling the attachment structure. + if isinstance(prev_content, str) and isinstance(new_content, str): + prev["content"] = ( + (prev_content + "\n\n" + new_content) + if prev_content and new_content + else (prev_content or new_content) + ) + repairs += 1 + continue + merged.append(msg) + + if repairs > 0: + # Rewrite in place so downstream paths (persistence, return + # value, session DB flush) see the repaired sequence. + messages[:] = merged + + return repairs + + + +def strip_think_blocks(agent, content: str) -> str: + """Remove reasoning/thinking blocks from content, returning only visible text. + + Handles four cases: + 1. Closed tag pairs (``…``) — the common path when + the provider emits complete reasoning blocks. + 2. Unterminated open tag at a block boundary (start of text or + after a newline) — e.g. MiniMax M2.7 / NIM endpoints where the + closing tag is dropped. Everything from the open tag to end + of string is stripped. The block-boundary check mirrors + ``gateway/stream_consumer.py``'s filter so models that mention + ```` in prose aren't over-stripped. + 3. Stray orphan open/close tags that slip through. + 4. Tag variants: ````, ````, ````, + ````, ```` (Gemma 4), all + case-insensitive. + + Additionally strips standalone tool-call XML blocks that some open + models (notably Gemma variants on OpenRouter) emit inside assistant + content instead of via the structured ``tool_calls`` field: + * ``…`` + * ``…`` + * ``…`` + * ``…`` + * ``…`` + * ``…`` (Gemma style) + Ported from openclaw/openclaw#67318. The ```` variant is + boundary-gated (only strips when the tag sits at start-of-line or + after punctuation and carries a ``name="..."`` attribute) so prose + mentions like "Use in JavaScript" are preserved. + """ + if not content: + return "" + # 1. Closed tag pairs — case-insensitive for all variants so + # mixed-case tags (, ) don't slip through to + # the unterminated-tag pass and take trailing content with them. + content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) + content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) + content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) + content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) + content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) + # 1b. Tool-call XML blocks (openclaw/openclaw#67318). Handle the + # generic tag names first — they have no attribute gating since + # a literal in prose is already vanishingly rare. + for _tc_name in ("tool_call", "tool_calls", "tool_result", + "function_call", "function_calls"): + content = re.sub( + rf'<{_tc_name}\b[^>]*>.*?', + '', + content, + flags=re.DOTALL | re.IGNORECASE, + ) + # 1c. ... — Gemma-style standalone + # tool call. Only strip when the tag sits at a block boundary + # (start of text, after a newline, or after sentence-ending + # punctuation) AND carries a name="..." attribute. This keeps + # prose mentions like "Use to declare" safe. + content = re.sub( + r'(?:(?<=^)|(?<=[\n\r.!?:]))[ \t]*' + r']*\bname\s*=[^>]*>' + r'(?:(?:(?!).)*)', + '', + content, + flags=re.DOTALL | re.IGNORECASE, + ) + # 2. Unterminated reasoning block — open tag at a block boundary + # (start of text, or after a newline) with no matching close. + # Strip from the tag to end of string. Fixes #8878 / #9568 + # (MiniMax M2.7 leaking raw reasoning into assistant content). + content = re.sub( + r'(?:^|\n)[ \t]*<(?:think|thinking|reasoning|thought|REASONING_SCRATCHPAD)\b[^>]*>.*$', + '', + content, + flags=re.DOTALL | re.IGNORECASE, + ) + # 3. Stray orphan open/close tags that slipped through. + content = re.sub( + r'\s*', + '', + content, + flags=re.IGNORECASE, + ) + # 3b. Stray tool-call closers. (We do NOT strip bare or + # unterminated because a truncated tail + # during streaming may still be valuable to the user; matches + # OpenClaw's intentional asymmetry.) + content = re.sub( + r'\s*', + '', + content, + flags=re.IGNORECASE, + ) + return content + + + +def recover_with_credential_pool( + agent, + *, + status_code: Optional[int], + has_retried_429: bool, + classified_reason: Optional[FailoverReason] = None, + error_context: Optional[Dict[str, Any]] = None, +) -> tuple[bool, bool]: + """Attempt credential recovery via pool rotation. + + Returns (recovered, has_retried_429). + On rate limits: first occurrence retries same credential (sets flag True). + second consecutive failure rotates to next credential. + On billing exhaustion: immediately rotates. + On auth failures: attempts token refresh before rotating. + + `classified_reason` lets the recovery path honor the structured error + classifier instead of relying only on raw HTTP codes. This matters for + providers that surface billing/rate-limit/auth conditions under a + different status code, such as Anthropic returning HTTP 400 for + "out of extra usage". + """ + pool = agent._credential_pool + if pool is None: + return False, has_retried_429 + + effective_reason = classified_reason + if effective_reason is None: + if status_code == 402: + effective_reason = FailoverReason.billing + elif status_code == 429: + effective_reason = FailoverReason.rate_limit + elif status_code in {401, 403}: + effective_reason = FailoverReason.auth + + if effective_reason == FailoverReason.billing: + rotate_status = status_code if status_code is not None else 402 + next_entry = pool.mark_exhausted_and_rotate(status_code=rotate_status, error_context=error_context) + if next_entry is not None: + _ra().logger.info( + "Credential %s (billing) — rotated to pool entry %s", + rotate_status, + getattr(next_entry, "id", "?"), + ) + agent._swap_credential(next_entry) + return True, False + return False, has_retried_429 + + if effective_reason == FailoverReason.rate_limit: + usage_limit_reached = False + if error_context: + context_reason = str(error_context.get("reason") or "").lower() + context_message = str(error_context.get("message") or "").lower() + usage_limit_reached = ( + "usage_limit_reached" in context_reason + or "usage limit has been reached" in context_message + ) + if not has_retried_429 and not usage_limit_reached: + return False, True + rotate_status = status_code if status_code is not None else 429 + next_entry = pool.mark_exhausted_and_rotate(status_code=rotate_status, error_context=error_context) + if next_entry is not None: + _ra().logger.info( + "Credential %s (rate limit) — rotated to pool entry %s", + rotate_status, + getattr(next_entry, "id", "?"), + ) + agent._swap_credential(next_entry) + return True, False + return False, True + + if effective_reason == FailoverReason.auth: + if agent._is_entitlement_failure(error_context, status_code): + _ra().logger.info( + "Credential %s — entitlement-shaped 403 from %s; " + "skipping pool refresh (account lacks subscription, " + "not a transient auth failure).", + status_code if status_code is not None else "auth", + agent.provider or "provider", + ) + return False, has_retried_429 + refreshed = pool.try_refresh_current() + if refreshed is not None: + _ra().logger.info(f"Credential auth failure — refreshed pool entry {getattr(refreshed, 'id', '?')}") + agent._swap_credential(refreshed) + return True, has_retried_429 + # Refresh failed — rotate to next credential instead of giving up. + # The failed entry is already marked exhausted by try_refresh_current(). + rotate_status = status_code if status_code is not None else 401 + next_entry = pool.mark_exhausted_and_rotate(status_code=rotate_status, error_context=error_context) + if next_entry is not None: + _ra().logger.info( + "Credential %s (auth refresh failed) — rotated to pool entry %s", + rotate_status, + getattr(next_entry, "id", "?"), + ) + agent._swap_credential(next_entry) + return True, False + + return False, has_retried_429 + + + +def try_recover_primary_transport( + agent, api_error: Exception, *, retry_count: int, max_retries: int, +) -> bool: + """Attempt one extra primary-provider recovery cycle for transient transport failures. + + After ``max_retries`` exhaust, rebuild the primary client (clearing + stale connection pools) and give it one more attempt before falling + back. This is most useful for direct endpoints (custom, Z.AI, + Anthropic, OpenAI, local models) where a TCP-level hiccup does not + mean the provider is down. + + Skipped for proxy/aggregator providers (OpenRouter, Nous) which + already manage connection pools and retries server-side — if our + retries through them are exhausted, one more rebuilt client won't help. + """ + if agent._fallback_activated: + return False + + # Only for transient transport errors + error_type = type(api_error).__name__ + if error_type not in _TRANSIENT_TRANSPORT_ERRORS: + return False + + # Skip for aggregator providers — they manage their own retry infra + if agent._is_openrouter_url(): + return False + provider_lower = (agent.provider or "").strip().lower() + if provider_lower in {"nous", "nous-research"}: + return False + + try: + # Close existing client to release stale connections + if getattr(agent, "client", None) is not None: + try: + agent._close_openai_client( + agent.client, reason="primary_recovery", shared=True, + ) + except Exception: + pass + + # Rebuild from primary snapshot + rt = agent._primary_runtime + agent._client_kwargs = dict(rt["client_kwargs"]) + agent.model = rt["model"] + agent.provider = rt["provider"] + agent.base_url = rt["base_url"] + agent.api_mode = rt["api_mode"] + if hasattr(agent, "_transport_cache"): + agent._transport_cache.clear() + agent.api_key = rt["api_key"] + + if agent.api_mode == "anthropic_messages": + from agent.anthropic_adapter import build_anthropic_client + agent._anthropic_api_key = rt["anthropic_api_key"] + agent._anthropic_base_url = rt["anthropic_base_url"] + agent._anthropic_client = build_anthropic_client( + rt["anthropic_api_key"], rt["anthropic_base_url"], + timeout=get_provider_request_timeout(agent.provider, agent.model), + ) + agent._is_anthropic_oauth = rt["is_anthropic_oauth"] + agent.client = None + else: + agent.client = agent._create_openai_client( + dict(rt["client_kwargs"]), + reason="primary_recovery", + shared=True, + ) + + wait_time = min(3 + retry_count, 8) + agent._vprint( + f"{agent.log_prefix}🔁 Transient {error_type} on {agent.provider} — " + f"rebuilt client, waiting {wait_time}s before one last primary attempt.", + force=True, + ) + time.sleep(wait_time) + return True + except Exception as e: + logging.warning("Primary transport recovery failed: %s", e) + return False + +# ── End provider fallback ────────────────────────────────────────────── + + + +def drop_thinking_only_and_merge_users( + messages: List[Dict[str, Any]], +) -> List[Dict[str, Any]]: + """Drop thinking-only assistant turns; merge any adjacent user messages left behind. + + Runs on the per-call ``api_messages`` copy only. The stored + conversation history (``agent.messages``) is never mutated, so the + user still sees the thinking block in the CLI/gateway transcript and + session persistence keeps the full trace. Only the wire copy sent to + the provider is cleaned. + + Why drop-and-merge rather than inject stub text: + - Fabricating ``"."`` / ``"(continued)"`` text lies in the history + and makes future turns see model output the model didn't emit. + - Dropping the turn preserves honesty; merging adjacent user messages + preserves the provider's role-alternation invariant. + - This is the pattern used by Claude Code's ``normalizeMessagesForAPI`` + (filterOrphanedThinkingOnlyMessages + mergeAdjacentUserMessages). + """ + if not messages: + return messages + + # Pass 1: drop thinking-only assistant turns. + kept = [m for m in messages if not _ra().AIAgent._is_thinking_only_assistant(m)] + dropped = len(messages) - len(kept) + if dropped == 0: + return messages + + # Pass 2: merge any newly-adjacent user messages. + merged: List[Dict[str, Any]] = [] + merges = 0 + for m in kept: + prev = merged[-1] if merged else None + if ( + prev is not None + and prev.get("role") == "user" + and m.get("role") == "user" + ): + prev_content = prev.get("content", "") + cur_content = m.get("content", "") + # Work on a copy of ``prev`` so the caller's input dicts are + # never mutated. ``_sanitize_api_messages`` upstream already + # hands us per-call copies, but staying pure here means we + # can be called safely from anywhere (tests, other loops). + prev_copy = dict(prev) + # Only string-content merge is meaningful for role-alternation + # purposes. If either side is a list (multimodal), append as a + # separate block rather than collapsing. + if isinstance(prev_content, str) and isinstance(cur_content, str): + sep = "\n\n" if prev_content and cur_content else "" + prev_copy["content"] = prev_content + sep + cur_content + elif isinstance(prev_content, list) and isinstance(cur_content, list): + prev_copy["content"] = list(prev_content) + list(cur_content) + elif isinstance(prev_content, list) and isinstance(cur_content, str): + if cur_content: + prev_copy["content"] = list(prev_content) + [ + {"type": "text", "text": cur_content} + ] + else: + prev_copy["content"] = list(prev_content) + elif isinstance(prev_content, str) and isinstance(cur_content, list): + new_blocks: List[Dict[str, Any]] = [] + if prev_content: + new_blocks.append({"type": "text", "text": prev_content}) + new_blocks.extend(cur_content) + prev_copy["content"] = new_blocks + else: + # Unknown content shape — fall back to appending separately + # (violates alternation, but safer than raising in a hot path). + merged.append(m) + continue + merged[-1] = prev_copy + merges += 1 + else: + merged.append(m) + + _ra().logger.debug( + "Pre-call sanitizer: dropped %d thinking-only assistant turn(s), " + "merged %d adjacent user message(s)", + dropped, + merges, + ) + return merged + + + +def restore_primary_runtime(agent) -> bool: + """Restore the primary runtime at the start of a new turn. + + In long-lived CLI sessions a single AIAgent instance spans multiple + turns. Without restoration, one transient failure pins the session + to the fallback provider for every subsequent turn. Calling this at + the top of ``run_conversation()`` makes fallback turn-scoped. + + The gateway caches agents across messages (``_agent_cache`` in + ``gateway/run.py``), so this restoration IS needed there too. + """ + if not agent._fallback_activated: + # Reset the chain index even when no fallback was activated this + # turn. Without this, a turn where _try_activate_fallback() was + # called but returned False (chain exhausted or provider not + # configured) leaves _fallback_index >= len(_fallback_chain) while + # _fallback_activated stays False. The next turn skips this block + # entirely, stranding the index and silently blocking all future + # fallback attempts for the session. Fixes #20465. + agent._fallback_index = 0 + return False + + if getattr(agent, "_rate_limited_until", 0) > time.monotonic(): + return False # primary still in rate-limit cooldown, stay on fallback + + rt = agent._primary_runtime + try: + # ── Core runtime state ── + agent.model = rt["model"] + agent.provider = rt["provider"] + agent.base_url = rt["base_url"] # setter updates _base_url_lower + agent.api_mode = rt["api_mode"] + if hasattr(agent, "_transport_cache"): + agent._transport_cache.clear() + agent.api_key = rt["api_key"] + agent._client_kwargs = dict(rt["client_kwargs"]) + agent._use_prompt_caching = rt["use_prompt_caching"] + # Default to native layout when the restored snapshot predates the + # native-vs-proxy split (older sessions saved before this PR). + agent._use_native_cache_layout = rt.get( + "use_native_cache_layout", + agent.api_mode == "anthropic_messages" and agent.provider == "anthropic", + ) + + # ── Rebuild client for the primary provider ── + if agent.api_mode == "anthropic_messages": + from agent.anthropic_adapter import build_anthropic_client + agent._anthropic_api_key = rt["anthropic_api_key"] + agent._anthropic_base_url = rt["anthropic_base_url"] + agent._anthropic_client = build_anthropic_client( + rt["anthropic_api_key"], rt["anthropic_base_url"], + timeout=get_provider_request_timeout(agent.provider, agent.model), + ) + agent._is_anthropic_oauth = rt["is_anthropic_oauth"] + agent.client = None + else: + agent.client = agent._create_openai_client( + dict(rt["client_kwargs"]), + reason="restore_primary", + shared=True, + ) + + # ── Restore context engine state ── + cc = agent.context_compressor + cc.update_model( + model=rt["compressor_model"], + context_length=rt["compressor_context_length"], + base_url=rt["compressor_base_url"], + api_key=rt["compressor_api_key"], + provider=rt["compressor_provider"], + ) + + # ── Reset fallback chain for the new turn ── + agent._fallback_activated = False + agent._fallback_index = 0 + + logging.info( + "Primary runtime restored for new turn: %s (%s)", + agent.model, agent.provider, + ) + return True + except Exception as e: + logging.warning("Failed to restore primary runtime: %s", e) + return False + +# Which error types indicate a transient transport failure worth +# one more attempt with a rebuilt client / connection pool. +_TRANSIENT_TRANSPORT_ERRORS = frozenset({ + "ReadTimeout", "ConnectTimeout", "PoolTimeout", + "ConnectError", "RemoteProtocolError", + "APIConnectionError", "APITimeoutError", +}) + + + +def extract_reasoning(agent, assistant_message) -> Optional[str]: + """ + Extract reasoning/thinking content from an assistant message. + + OpenRouter and various providers can return reasoning in multiple formats: + 1. message.reasoning - Direct reasoning field (DeepSeek, Qwen, etc.) + 2. message.reasoning_content - Alternative field (Moonshot AI, Novita, etc.) + 3. message.reasoning_details - Array of {type, summary, ...} objects (OpenRouter unified) + + Args: + assistant_message: The assistant message object from the API response + + Returns: + Combined reasoning text, or None if no reasoning found + """ + reasoning_parts = [] + + # Check direct reasoning field + if hasattr(assistant_message, 'reasoning') and assistant_message.reasoning: + reasoning_parts.append(assistant_message.reasoning) + + # Check reasoning_content field (alternative name used by some providers) + if hasattr(assistant_message, 'reasoning_content') and assistant_message.reasoning_content: + # Don't duplicate if same as reasoning + if assistant_message.reasoning_content not in reasoning_parts: + reasoning_parts.append(assistant_message.reasoning_content) + + # Check reasoning_details array (OpenRouter unified format) + # Format: [{"type": "reasoning.summary", "summary": "...", ...}, ...] + if hasattr(assistant_message, 'reasoning_details') and assistant_message.reasoning_details: + for detail in assistant_message.reasoning_details: + if isinstance(detail, dict): + # Extract summary from reasoning detail object + summary = ( + detail.get('summary') + or detail.get('thinking') + or detail.get('content') + or detail.get('text') + ) + if summary and summary not in reasoning_parts: + reasoning_parts.append(summary) + + # Some providers embed reasoning directly inside assistant content + # instead of returning structured reasoning fields. Only fall back + # to inline extraction when no structured reasoning was found. + content = getattr(assistant_message, "content", None) + if not reasoning_parts and isinstance(content, list): + # DeepSeek V4 Pro (and compatible providers) return content as a + # list of typed blocks, e.g.: + # [{"type": "thinking", "thinking": "..."}, {"type": "output", ...}] + # Without this branch the thinking text is silently dropped and the + # next turn fails with HTTP 400 ("thinking must be passed back"). + # Refs #21944. + for block in content: + if isinstance(block, dict) and block.get("type") == "thinking": + thinking_text = block.get("thinking") or block.get("text") or "" + thinking_text = thinking_text.strip() + if thinking_text and thinking_text not in reasoning_parts: + reasoning_parts.append(thinking_text) + if not reasoning_parts and isinstance(content, str) and content: + inline_patterns = ( + r"(.*?)", + r"(.*?)", + r"(.*?)", + r"(.*?)", + r"(.*?)", + ) + for pattern in inline_patterns: + flags = re.DOTALL | re.IGNORECASE + for block in re.findall(pattern, content, flags=flags): + cleaned = block.strip() + if cleaned and cleaned not in reasoning_parts: + reasoning_parts.append(cleaned) + + # Combine all reasoning parts + if reasoning_parts: + return "\n\n".join(reasoning_parts) + + return None + + + +def dump_api_request_debug( + agent, + api_kwargs: Dict[str, Any], + *, + reason: str, + error: Optional[Exception] = None, +) -> Optional[Path]: + """ + Dump a debug-friendly HTTP request record for the active inference API. + + Captures the request body from api_kwargs (excluding transport-only keys + like timeout). Intended for debugging provider-side 4xx failures where + retries are not useful. + """ + try: + body = copy.deepcopy(api_kwargs) + body.pop("timeout", None) + body = {k: v for k, v in body.items() if v is not None} + + api_key = None + try: + api_key = getattr(agent.client, "api_key", None) + except Exception as e: + _ra().logger.debug("Could not extract API key for debug dump: %s", e) + + dump_payload: Dict[str, Any] = { + "timestamp": datetime.now().isoformat(), + "session_id": agent.session_id, + "reason": reason, + "request": { + "method": "POST", + "url": f"{agent.base_url.rstrip('/')}{'/responses' if agent.api_mode == 'codex_responses' else '/chat/completions'}", + "headers": { + "Authorization": f"Bearer {agent._mask_api_key_for_logs(api_key)}", + "Content-Type": "application/json", + }, + "body": body, + }, + } + + if error is not None: + error_info: Dict[str, Any] = { + "type": type(error).__name__, + "message": str(error), + } + for attr_name in ("status_code", "request_id", "code", "param", "type"): + attr_value = getattr(error, attr_name, None) + if attr_value is not None: + error_info[attr_name] = attr_value + + body_attr = getattr(error, "body", None) + if body_attr is not None: + error_info["body"] = body_attr + + response_obj = getattr(error, "response", None) + if response_obj is not None: + try: + error_info["response_status"] = getattr(response_obj, "status_code", None) + error_info["response_text"] = response_obj.text + except Exception as e: + _ra().logger.debug("Could not extract error response details: %s", e) + + dump_payload["error"] = error_info + + timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f") + dump_file = agent.logs_dir / f"request_dump_{agent.session_id}_{timestamp}.json" + dump_file.write_text( + json.dumps(dump_payload, ensure_ascii=False, indent=2, default=str), + encoding="utf-8", + ) + + agent._vprint(f"{agent.log_prefix}🧾 Request debug dump written to: {dump_file}") + + if env_var_enabled("HERMES_DUMP_REQUEST_STDOUT"): + print(json.dumps(dump_payload, ensure_ascii=False, indent=2, default=str)) + + return dump_file + except Exception as dump_error: + if agent.verbose_logging: + logging.warning(f"Failed to dump API request debug payload: {dump_error}") + return None + + + +def anthropic_prompt_cache_policy( + agent, + *, + provider: Optional[str] = None, + base_url: Optional[str] = None, + api_mode: Optional[str] = None, + model: Optional[str] = None, +) -> tuple[bool, bool]: + """Decide whether to apply Anthropic prompt caching and which layout to use. + + Returns ``(should_cache, use_native_layout)``: + * ``should_cache`` — inject ``cache_control`` breakpoints for this + request (applies to OpenRouter Claude, native Anthropic, and + third-party gateways that speak the native Anthropic protocol). + * ``use_native_layout`` — place markers on the *inner* content + blocks (native Anthropic accepts and requires this layout); + when False markers go on the message envelope (OpenRouter and + OpenAI-wire proxies expect the looser layout). + + Third-party providers using the native Anthropic transport + (``api_mode == 'anthropic_messages'`` + Claude-named model) get + caching with the native layout so they benefit from the same + cost reduction as direct Anthropic callers, provided their + gateway implements the Anthropic cache_control contract + (MiniMax, Zhipu GLM, LiteLLM's Anthropic proxy mode all do). + + Qwen / Alibaba-family models on OpenCode, OpenCode Go, and direct + Alibaba (DashScope) also honour Anthropic-style ``cache_control`` + markers on OpenAI-wire chat completions. Upstream pi-mono #3392 / + pi #3393 documented this for opencode-go Qwen. Without markers + these providers serve zero cache hits, re-billing the full prompt + on every turn. + """ + eff_provider = (provider if provider is not None else agent.provider) or "" + eff_base_url = base_url if base_url is not None else (agent.base_url or "") + eff_api_mode = api_mode if api_mode is not None else (agent.api_mode or "") + eff_model = (model if model is not None else agent.model) or "" + + model_lower = eff_model.lower() + provider_lower = eff_provider.lower() + is_claude = "claude" in model_lower + is_openrouter = base_url_host_matches(eff_base_url, "openrouter.ai") + # Nous Portal proxies to OpenRouter behind the scenes — identical + # OpenAI-wire envelope cache_control semantics. Treat it as an + # OpenRouter-equivalent endpoint for caching layout purposes. + is_nous_portal = "nousresearch" in eff_base_url.lower() + is_anthropic_wire = eff_api_mode == "anthropic_messages" + is_native_anthropic = ( + is_anthropic_wire + and (eff_provider == "anthropic" or base_url_hostname(eff_base_url) == "api.anthropic.com") + ) + + if is_native_anthropic: + return True, True + if (is_openrouter or is_nous_portal) and is_claude: + return True, False + # Nous Portal Qwen (e.g. qwen3.6-plus) takes the same envelope-layout + # cache_control path as Portal Claude. Portal proxies to OpenRouter + # and the upstream Qwen route accepts cache_control markers; without + # this branch the alibaba-family check below only matches + # provider=opencode/alibaba and Portal traffic falls through to + # (False, False), serving 0% cache hits and re-billing the full + # prompt on every turn. + if is_nous_portal and "qwen" in model_lower: + return True, False + if is_anthropic_wire and is_claude: + # Third-party Anthropic-compatible gateway. + return True, True + + # MiniMax on its Anthropic-compatible endpoint serves its own + # model family (MiniMax-M2.7, M2.5, M2.1, M2) with documented + # cache_control support (0.1× read pricing, 5-minute TTL). The + # blanket is_claude gate above excludes these — opt them in + # explicitly via provider id or host match so users on + # provider=minimax / minimax-cn (or custom endpoints pointing at + # api.minimax.io/anthropic / api.minimaxi.com/anthropic) get the + # same cost reduction as Claude traffic. + # Docs: https://platform.minimax.io/docs/api-reference/anthropic-api-compatible-cache + if is_anthropic_wire: + is_minimax_provider = provider_lower in {"minimax", "minimax-cn"} + is_minimax_host = ( + base_url_host_matches(eff_base_url, "api.minimax.io") + or base_url_host_matches(eff_base_url, "api.minimaxi.com") + ) + if is_minimax_provider or is_minimax_host: + return True, True + + # Qwen/Alibaba on OpenCode (Zen/Go) and native DashScope: OpenAI-wire + # transport that accepts Anthropic-style cache_control markers and + # rewards them with real cache hits. Without this branch + # qwen3.6-plus on opencode-go reports 0% cached tokens and burns + # through the subscription on every turn. + model_is_qwen = "qwen" in model_lower + provider_is_alibaba_family = provider_lower in { + "opencode", "opencode-zen", "opencode-go", "alibaba", + } + if provider_is_alibaba_family and model_is_qwen: + # Envelope layout (native_anthropic=False): markers on inner + # content parts, not top-level tool messages. Matches + # pi-mono's "alibaba" cacheControlFormat. + return True, False + + return False, False + + + +def create_openai_client(agent, client_kwargs: dict, *, reason: str, shared: bool) -> Any: + from agent.auxiliary_client import _validate_base_url, _validate_proxy_env_urls + # Treat client_kwargs as read-only. Callers pass agent._client_kwargs (or shallow + # copies of it) in; any in-place mutation leaks back into the stored dict and is + # reused on subsequent requests. #10933 hit this by injecting an httpx.Client + # transport that was torn down after the first request, so the next request + # wrapped a closed transport and raised "Cannot send a request, as the client + # has been closed" on every retry. The revert resolved that specific path; this + # copy locks the contract so future transport/keepalive work can't reintroduce + # the same class of bug. + client_kwargs = dict(client_kwargs) + _validate_proxy_env_urls() + _validate_base_url(client_kwargs.get("base_url")) + if agent.provider == "copilot-acp" or str(client_kwargs.get("base_url", "")).startswith("acp://copilot"): + from agent.copilot_acp_client import CopilotACPClient + + client = CopilotACPClient(**client_kwargs) + _ra().logger.info( + "Copilot ACP client created (%s, shared=%s) %s", + reason, + shared, + agent._client_log_context(), + ) + return client + if agent.provider == "google-gemini-cli" or str(client_kwargs.get("base_url", "")).startswith("cloudcode-pa://"): + from agent.gemini_cloudcode_adapter import GeminiCloudCodeClient + + # Strip OpenAI-specific kwargs the Gemini client doesn't accept + safe_kwargs = { + k: v for k, v in client_kwargs.items() + if k in {"api_key", "base_url", "default_headers", "project_id", "timeout"} + } + client = GeminiCloudCodeClient(**safe_kwargs) + _ra().logger.info( + "Gemini Cloud Code Assist client created (%s, shared=%s) %s", + reason, + shared, + agent._client_log_context(), + ) + return client + if agent.provider == "gemini": + from agent.gemini_native_adapter import GeminiNativeClient, is_native_gemini_base_url + + base_url = str(client_kwargs.get("base_url", "") or "") + if is_native_gemini_base_url(base_url): + safe_kwargs = { + k: v for k, v in client_kwargs.items() + if k in {"api_key", "base_url", "default_headers", "timeout", "http_client"} + } + if "http_client" not in safe_kwargs: + keepalive_http = agent._build_keepalive_http_client(base_url) + if keepalive_http is not None: + safe_kwargs["http_client"] = keepalive_http + client = GeminiNativeClient(**safe_kwargs) + _ra().logger.info( + "Gemini native client created (%s, shared=%s) %s", + reason, + shared, + agent._client_log_context(), + ) + return client + # Inject TCP keepalives so the kernel detects dead provider connections + # instead of letting them sit silently in CLOSE-WAIT (#10324). Without + # this, a peer that drops mid-stream leaves the socket in a state where + # epoll_wait never fires, ``httpx`` read timeout may not trigger, and + # the agent hangs until manually killed. Probes after 30s idle, retry + # every 10s, give up after 3 → dead peer detected within ~60s. + # + # Safety against #10933: the ``client_kwargs = dict(client_kwargs)`` + # above means this injection only lands in the local per-call copy, + # never back into ``agent._client_kwargs``. Each ``_create_openai_client`` + # invocation therefore gets its OWN fresh ``httpx.Client`` whose + # lifetime is tied to the OpenAI client it is passed to. When the + # OpenAI client is closed (rebuild, teardown, credential rotation), + # the paired ``httpx.Client`` closes with it, and the next call + # constructs a fresh one — no stale closed transport can be reused. + # Tests in ``tests/run_agent/test_create_openai_client_reuse.py`` and + # ``tests/run_agent/test_sequential_chats_live.py`` pin this invariant. + if "http_client" not in client_kwargs: + keepalive_http = agent._build_keepalive_http_client(client_kwargs.get("base_url", "")) + if keepalive_http is not None: + client_kwargs["http_client"] = keepalive_http + # Uses the module-level `OpenAI` name, resolved lazily on first + # access via __getattr__ below. Tests patch via `run_agent.OpenAI`. + client = _ra().OpenAI(**client_kwargs) + _ra().logger.info( + "OpenAI client created (%s, shared=%s) %s", + reason, + shared, + agent._client_log_context(), + ) + return client + + +def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mode=''): + """Switch the model/provider in-place for a live agent. + + Called by the /model command handlers (CLI and gateway) after + ``model_switch.switch_model()`` has resolved credentials and + validated the model. This method performs the actual runtime + swap: rebuilding clients, updating caching flags, and refreshing + the context compressor. + + The implementation mirrors ``_try_activate_fallback()`` for the + client-swap logic but also updates ``_primary_runtime`` so the + change persists across turns (unlike fallback which is + turn-scoped). + """ + from hermes_cli.providers import determine_api_mode + + # ── Determine api_mode if not provided ── + if not api_mode: + api_mode = determine_api_mode(new_provider, base_url) + + # Defense-in-depth: ensure OpenCode base_url doesn't carry a trailing + # /v1 into the anthropic_messages client, which would cause the SDK to + # hit /v1/v1/messages. `model_switch.switch_model()` already strips + # this, but we guard here so any direct callers (future code paths, + # tests) can't reintroduce the double-/v1 404 bug. + if ( + api_mode == "anthropic_messages" + and new_provider in {"opencode-zen", "opencode-go"} + and isinstance(base_url, str) + and base_url + ): + base_url = re.sub(r"/v1/?$", "", base_url) + + old_model = agent.model + old_provider = agent.provider + + # Clear the per-config context_length override so the new model's + # actual context window is resolved via get_model_context_length() + # instead of inheriting the stale value from the previous model. + agent._config_context_length = None + + # ── Swap core runtime fields ── + agent.model = new_model + agent.provider = new_provider + # Use new base_url when provided; only fall back to current when the + # new provider genuinely has no endpoint (e.g. native SDK providers). + # Without this guard the old provider's URL (e.g. Ollama's localhost + # address) would persist silently after switching to a cloud provider + # that returns an empty base_url string. + if base_url: + agent.base_url = base_url + agent.api_mode = api_mode + # Invalidate transport cache — new api_mode may need a different transport + if hasattr(agent, "_transport_cache"): + agent._transport_cache.clear() + if api_key: + agent.api_key = api_key + + # ── Build new client ── + if api_mode == "anthropic_messages": + from agent.anthropic_adapter import ( + build_anthropic_client, + resolve_anthropic_token, + _is_oauth_token, + ) + # Only fall back to ANTHROPIC_TOKEN when the provider is actually Anthropic. + # Other anthropic_messages providers (MiniMax, Alibaba, etc.) must use their own + # API key — falling back would send Anthropic credentials to third-party endpoints. + _is_native_anthropic = new_provider == "anthropic" + effective_key = (api_key or agent.api_key or resolve_anthropic_token() or "") if _is_native_anthropic else (api_key or agent.api_key or "") + agent.api_key = effective_key + agent._anthropic_api_key = effective_key + agent._anthropic_base_url = base_url or getattr(agent, "_anthropic_base_url", None) + agent._anthropic_client = build_anthropic_client( + effective_key, agent._anthropic_base_url, + timeout=get_provider_request_timeout(agent.provider, agent.model), + ) + agent._is_anthropic_oauth = _is_oauth_token(effective_key) if _is_native_anthropic else False + agent.client = None + agent._client_kwargs = {} + else: + effective_key = api_key or agent.api_key + effective_base = base_url or agent.base_url + agent._client_kwargs = { + "api_key": effective_key, + "base_url": effective_base, + } + _sm_timeout = get_provider_request_timeout(agent.provider, agent.model) + if _sm_timeout is not None: + agent._client_kwargs["timeout"] = _sm_timeout + agent.client = agent._create_openai_client( + dict(agent._client_kwargs), + reason="switch_model", + shared=True, + ) + + # ── Re-evaluate prompt caching ── + agent._use_prompt_caching, agent._use_native_cache_layout = ( + agent._anthropic_prompt_cache_policy( + provider=new_provider, + base_url=agent.base_url, + api_mode=api_mode, + model=new_model, + ) + ) + + # ── LM Studio: preload before probing context length ── + agent._ensure_lmstudio_runtime_loaded() + + # ── Update context compressor ── + if hasattr(agent, "context_compressor") and agent.context_compressor: + from agent.model_metadata import get_model_context_length + # Re-read custom_providers from live config so per-model + # context_length overrides are honored when switching to a + # custom provider mid-session (closes #15779). + _sm_custom_providers = None + try: + from hermes_cli.config import load_config, get_compatible_custom_providers + _sm_cfg = load_config() + _sm_custom_providers = get_compatible_custom_providers(_sm_cfg) + except Exception: + _sm_custom_providers = None + new_context_length = get_model_context_length( + agent.model, + base_url=agent.base_url, + api_key=agent.api_key, + provider=agent.provider, + config_context_length=getattr(agent, "_config_context_length", None), + custom_providers=_sm_custom_providers, + ) + agent.context_compressor.update_model( + model=agent.model, + context_length=new_context_length, + base_url=agent.base_url, + api_key=getattr(agent, "api_key", ""), + provider=agent.provider, + api_mode=agent.api_mode, + ) + + # ── Invalidate cached system prompt so it rebuilds next turn ── + agent._cached_system_prompt = None + + # ── Update _primary_runtime so the change persists across turns ── + _cc = agent.context_compressor if hasattr(agent, "context_compressor") and agent.context_compressor else None + agent._primary_runtime = { + "model": agent.model, + "provider": agent.provider, + "base_url": agent.base_url, + "api_mode": agent.api_mode, + "api_key": getattr(agent, "api_key", ""), + "client_kwargs": dict(agent._client_kwargs), + "use_prompt_caching": agent._use_prompt_caching, + "use_native_cache_layout": agent._use_native_cache_layout, + "compressor_model": getattr(_cc, "model", agent.model) if _cc else agent.model, + "compressor_base_url": getattr(_cc, "base_url", agent.base_url) if _cc else agent.base_url, + "compressor_api_key": getattr(_cc, "api_key", "") if _cc else "", + "compressor_provider": getattr(_cc, "provider", agent.provider) if _cc else agent.provider, + "compressor_context_length": _cc.context_length if _cc else 0, + "compressor_threshold_tokens": _cc.threshold_tokens if _cc else 0, + } + if api_mode == "anthropic_messages": + agent._primary_runtime.update({ + "anthropic_api_key": agent._anthropic_api_key, + "anthropic_base_url": agent._anthropic_base_url, + "is_anthropic_oauth": agent._is_anthropic_oauth, + }) + + # ── Reset fallback state ── + agent._fallback_activated = False + agent._fallback_index = 0 + + # When the user deliberately swaps primary providers (e.g. openrouter + # → anthropic), drop any fallback entries that target the OLD primary + # or the NEW one. The chain was seeded from config at agent init for + # the original provider — without pruning, a failed turn on the new + # primary silently re-activates the provider the user just rejected, + # which is exactly what was reported during TUI v2 blitz testing + # ("switched to anthropic, tui keeps trying openrouter"). + old_norm = (old_provider or "").strip().lower() + new_norm = (new_provider or "").strip().lower() + fallback_chain = list(getattr(agent, "_fallback_chain", []) or []) + if old_norm and new_norm and old_norm != new_norm: + fallback_chain = [ + entry for entry in fallback_chain + if (entry.get("provider") or "").strip().lower() not in {old_norm, new_norm} + ] + agent._fallback_chain = fallback_chain + agent._fallback_model = fallback_chain[0] if fallback_chain else None + + logging.info( + "Model switched in-place: %s (%s) -> %s (%s)", + old_model, old_provider, new_model, new_provider, + ) + + + +def invoke_tool(agent, function_name: str, function_args: dict, effective_task_id: str, + tool_call_id: Optional[str] = None, messages: list = None, + pre_tool_block_checked: bool = False) -> str: + """Invoke a single tool and return the result string. No display logic. + + Handles both agent-level tools (todo, memory, etc.) and registry-dispatched + tools. Used by the concurrent execution path; the sequential path retains + its own inline invocation for backward-compatible display handling. + """ + # Check plugin hooks for a block directive before executing anything. + block_message: Optional[str] = None + if not pre_tool_block_checked: + try: + from hermes_cli.plugins import get_pre_tool_call_block_message + block_message = get_pre_tool_call_block_message( + function_name, function_args, task_id=effective_task_id or "", + ) + except Exception: + pass + if block_message is not None: + return json.dumps({"error": block_message}, ensure_ascii=False) + + if function_name == "todo": + from tools.todo_tool import todo_tool as _todo_tool + return _todo_tool( + todos=function_args.get("todos"), + merge=function_args.get("merge", False), + store=agent._todo_store, + ) + elif function_name == "session_search": + session_db = agent._get_session_db_for_recall() + if not session_db: + from hermes_state import format_session_db_unavailable + return json.dumps({"success": False, "error": format_session_db_unavailable()}) + from tools.session_search_tool import session_search as _session_search + return _session_search( + query=function_args.get("query", ""), + role_filter=function_args.get("role_filter"), + limit=function_args.get("limit", 3), + db=session_db, + current_session_id=agent.session_id, + ) + elif function_name == "memory": + target = function_args.get("target", "memory") + from tools.memory_tool import memory_tool as _memory_tool + result = _memory_tool( + action=function_args.get("action"), + target=target, + content=function_args.get("content"), + old_text=function_args.get("old_text"), + store=agent._memory_store, + ) + # Bridge: notify external memory provider of built-in memory writes + if agent._memory_manager and function_args.get("action") in {"add", "replace"}: + try: + agent._memory_manager.on_memory_write( + function_args.get("action", ""), + target, + function_args.get("content", ""), + metadata=agent._build_memory_write_metadata( + task_id=effective_task_id, + tool_call_id=tool_call_id, + ), + ) + except Exception: + pass + return result + elif agent._memory_manager and agent._memory_manager.has_tool(function_name): + return agent._memory_manager.handle_tool_call(function_name, function_args) + elif function_name == "clarify": + from tools.clarify_tool import clarify_tool as _clarify_tool + return _clarify_tool( + question=function_args.get("question", ""), + choices=function_args.get("choices"), + callback=agent.clarify_callback, + ) + elif function_name == "delegate_task": + return agent._dispatch_delegate_task(function_args) + else: + return _ra().handle_function_call( + function_name, function_args, effective_task_id, + tool_call_id=tool_call_id, + session_id=agent.session_id or "", + enabled_tools=list(agent.valid_tool_names) if agent.valid_tool_names else None, + skip_pre_tool_call_hook=True, + ) + + + +def repair_tool_call(agent, tool_name: str) -> str | None: + """Attempt to repair a mismatched tool name before aborting. + + Models sometimes emit variants of a tool name that differ only + in casing, separators, or class-like suffixes. Normalize + aggressively before falling back to fuzzy match: + + 1. Lowercase direct match. + 2. Lowercase + hyphens/spaces -> underscores. + 3. CamelCase -> snake_case (TodoTool -> todo_tool). + 4. Strip trailing ``_tool`` / ``-tool`` / ``tool`` suffix that + Claude-style models sometimes tack on (TodoTool_tool -> + TodoTool -> Todo -> todo). Applied twice so double-tacked + suffixes like ``TodoTool_tool`` reduce all the way. + 5. Fuzzy match (difflib, cutoff=0.7). + + See #14784 for the original reports (TodoTool_tool, Patch_tool, + BrowserClick_tool were all returning "Unknown tool" before). + + Returns the repaired name if found in valid_tool_names, else None. + """ + import re + from difflib import get_close_matches + + if not tool_name: + return None + + def _norm(s: str) -> str: + return s.lower().replace("-", "_").replace(" ", "_") + + def _camel_snake(s: str) -> str: + return re.sub(r"(? str | None: + lc = s.lower() + for suffix in ("_tool", "-tool", "tool"): + if lc.endswith(suffix): + return s[: -len(suffix)].rstrip("_-") + return None + + # Cheap fast-paths first — these cover the common case. + lowered = tool_name.lower() + if lowered in agent.valid_tool_names: + return lowered + normalized = _norm(tool_name) + if normalized in agent.valid_tool_names: + return normalized + + # Build the full candidate set for class-like emissions. + cands: set[str] = {tool_name, lowered, normalized, _camel_snake(tool_name)} + # Strip trailing tool-suffix up to twice — TodoTool_tool needs it. + for _ in range(2): + extra: set[str] = set() + for c in cands: + stripped = _strip_tool_suffix(c) + if stripped: + extra.add(stripped) + extra.add(_norm(stripped)) + extra.add(_camel_snake(stripped)) + cands |= extra + + for c in cands: + if c and c in agent.valid_tool_names: + return c + + # Fuzzy match as last resort. + matches = get_close_matches(lowered, agent.valid_tool_names, n=1, cutoff=0.7) + if matches: + return matches[0] + + return None + + + +def sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]: + """Fix orphaned tool_call / tool_result pairs before every LLM call. + + Runs unconditionally — not gated on whether the context compressor + is present — so orphans from session loading or manual message + manipulation are always caught. + """ + # --- Role allowlist: drop messages with roles the API won't accept --- + filtered = [] + for msg in messages: + role = msg.get("role") + if role not in _ra().AIAgent._VALID_API_ROLES: + _ra().logger.debug( + "Pre-call sanitizer: dropping message with invalid role %r", + role, + ) + continue + filtered.append(msg) + messages = filtered + + surviving_call_ids: set = set() + for msg in messages: + if msg.get("role") == "assistant": + for tc in msg.get("tool_calls") or []: + cid = _ra().AIAgent._get_tool_call_id_static(tc) + if cid: + surviving_call_ids.add(cid) + + result_call_ids: set = set() + for msg in messages: + if msg.get("role") == "tool": + cid = msg.get("tool_call_id") + if cid: + result_call_ids.add(cid) + + # 1. Drop tool results with no matching assistant call + orphaned_results = result_call_ids - surviving_call_ids + if orphaned_results: + messages = [ + m for m in messages + if not (m.get("role") == "tool" and m.get("tool_call_id") in orphaned_results) + ] + _ra().logger.debug( + "Pre-call sanitizer: removed %d orphaned tool result(s)", + len(orphaned_results), + ) + + # 2. Inject stub results for calls whose result was dropped + missing_results = surviving_call_ids - result_call_ids + if missing_results: + patched: List[Dict[str, Any]] = [] + for msg in messages: + patched.append(msg) + if msg.get("role") == "assistant": + for tc in msg.get("tool_calls") or []: + cid = _ra().AIAgent._get_tool_call_id_static(tc) + if cid in missing_results: + patched.append({ + "role": "tool", + "name": _ra().AIAgent._get_tool_call_name_static(tc), + "content": "[Result unavailable — see context summary above]", + "tool_call_id": cid, + }) + messages = patched + _ra().logger.debug( + "Pre-call sanitizer: added %d stub tool result(s)", + len(missing_results), + ) + return messages + + + +def looks_like_codex_intermediate_ack( + agent, + user_message: str, + assistant_content: str, + messages: List[Dict[str, Any]], +) -> bool: + """Detect a planning/ack message that should continue instead of ending the turn.""" + if any(isinstance(msg, dict) and msg.get("role") == "tool" for msg in messages): + return False + + assistant_text = agent._strip_think_blocks(assistant_content or "").strip().lower() + if not assistant_text: + return False + if len(assistant_text) > 1200: + return False + + has_future_ack = bool( + re.search(r"\b(i['’]ll|i will|let me|i can do that|i can help with that)\b", assistant_text) + ) + if not has_future_ack: + return False + + action_markers = ( + "look into", + "look at", + "inspect", + "scan", + "check", + "analyz", + "review", + "explore", + "read", + "open", + "run", + "test", + "fix", + "debug", + "search", + "find", + "walkthrough", + "report back", + "summarize", + ) + workspace_markers = ( + "directory", + "current directory", + "current dir", + "cwd", + "repo", + "repository", + "codebase", + "project", + "folder", + "filesystem", + "file tree", + "files", + "path", + ) + + user_text = (user_message or "").strip().lower() + user_targets_workspace = ( + any(marker in user_text for marker in workspace_markers) + or "~/" in user_text + or "/" in user_text + ) + assistant_mentions_action = any(marker in assistant_text for marker in action_markers) + assistant_targets_workspace = any( + marker in assistant_text for marker in workspace_markers + ) + return (user_targets_workspace or assistant_targets_workspace) and assistant_mentions_action + + + + +def copy_reasoning_content_for_api(agent, source_msg: dict, api_msg: dict) -> None: + """Copy provider-facing reasoning fields onto an API replay message.""" + if source_msg.get("role") != "assistant": + return + + # 1. Explicit reasoning_content already set — preserve it verbatim + # (includes DeepSeek/Kimi's own space-placeholder written at creation + # time, and any valid reasoning content from the same provider). + # + # Exception: sessions persisted BEFORE #17341 have empty-string + # placeholders pinned at creation time. DeepSeek V4 Pro rejects + # those with HTTP 400. When the active provider enforces the + # thinking-mode echo, upgrade "" → " " on replay so stale history + # doesn't 400 the user on the next turn. + existing = source_msg.get("reasoning_content") + if isinstance(existing, str): + if existing == "" and agent._needs_thinking_reasoning_pad(): + api_msg["reasoning_content"] = " " + else: + api_msg["reasoning_content"] = existing + return + + needs_thinking_pad = agent._needs_thinking_reasoning_pad() + + # 2. Cross-provider poisoned history (#15748): on DeepSeek/Kimi, + # if the source turn has tool_calls AND a 'reasoning' field but no + # 'reasoning_content' key, the 'reasoning' text was written by a + # prior provider (e.g. MiniMax) — DeepSeek's own _build_assistant_message + # pins reasoning_content at creation time for tool-call turns, so the + # shape (reasoning set, reasoning_content absent, tool_calls present) + # is unreachable from same-provider DeepSeek history after this fix. + # Inject a single space to satisfy the API without leaking another + # provider's chain of thought to DeepSeek/Kimi. Space (not "") + # because DeepSeek V4 Pro rejects empty-string reasoning_content + # in thinking mode (refs #17341). + normalized_reasoning = source_msg.get("reasoning") + if ( + needs_thinking_pad + and source_msg.get("tool_calls") + and isinstance(normalized_reasoning, str) + and normalized_reasoning + ): + api_msg["reasoning_content"] = " " + return + + # 3. Healthy session: promote 'reasoning' field to 'reasoning_content' + # for providers that use the internal 'reasoning' key. + # This must happen before the unconditional empty-string fallback so + # genuine reasoning content is not overwritten (#15812 regression in + # PR #15478). + if isinstance(normalized_reasoning, str) and normalized_reasoning: + api_msg["reasoning_content"] = normalized_reasoning + return + + # 4. DeepSeek / Kimi thinking mode: all assistant messages need + # reasoning_content. Inject a single space to satisfy the provider's + # requirement when no explicit reasoning content is present. Covers + # both tool-call turns (already-poisoned history with no reasoning + # at all) and plain text turns. Space (not "") because DeepSeek V4 + # Pro tightened validation and rejects empty string with HTTP 400 + # ("The reasoning content in the thinking mode must be passed back + # to the API"). Refs #17341. + if needs_thinking_pad: + api_msg["reasoning_content"] = " " + return + + # 5. reasoning_content was present but not a string (e.g. None after + # context compaction). Don't pass null to the API. + api_msg.pop("reasoning_content", None) + + + +def cleanup_dead_connections(agent) -> bool: + """Detect and clean up dead TCP connections on the primary client. + + Inspects the httpx connection pool for sockets in unhealthy states + (CLOSE-WAIT, errors). If any are found, force-closes all sockets + and rebuilds the primary client from scratch. + + Returns True if dead connections were found and cleaned up. + """ + client = getattr(agent, "client", None) + if client is None: + return False + try: + http_client = getattr(client, "_client", None) + if http_client is None: + return False + transport = getattr(http_client, "_transport", None) + if transport is None: + return False + pool = getattr(transport, "_pool", None) + if pool is None: + return False + connections = ( + getattr(pool, "_connections", None) + or getattr(pool, "_pool", None) + or [] + ) + dead_count = 0 + for conn in list(connections): + # Check for connections that are idle but have closed sockets + stream = ( + getattr(conn, "_network_stream", None) + or getattr(conn, "_stream", None) + ) + if stream is None: + continue + sock = getattr(stream, "_sock", None) + if sock is None: + sock = getattr(stream, "stream", None) + if sock is not None: + sock = getattr(sock, "_sock", None) + if sock is None: + continue + # Probe socket health with a non-blocking recv peek + import socket as _socket + try: + sock.setblocking(False) + data = sock.recv(1, _socket.MSG_PEEK | _socket.MSG_DONTWAIT) + if data == b"": + dead_count += 1 + except BlockingIOError: + pass # No data available — socket is healthy + except OSError: + dead_count += 1 + finally: + try: + sock.setblocking(True) + except OSError: + pass + if dead_count > 0: + _ra().logger.warning( + "Found %d dead connection(s) in client pool — rebuilding client", + dead_count, + ) + agent._replace_primary_openai_client(reason="dead_connection_cleanup") + return True + except Exception as exc: + _ra().logger.debug("Dead connection check error: %s", exc) + return False + + + +def extract_api_error_context(error: Exception) -> Dict[str, Any]: + """Extract structured rate-limit details from provider errors.""" + context: Dict[str, Any] = {} + + body = getattr(error, "body", None) + payload = None + if isinstance(body, dict): + payload = body.get("error") if isinstance(body.get("error"), dict) else body + if isinstance(payload, dict): + reason = payload.get("code") or payload.get("type") or payload.get("error") + if isinstance(reason, str) and reason.strip(): + context["reason"] = reason.strip() + message = payload.get("message") or payload.get("error_description") + if isinstance(message, str) and message.strip(): + context["message"] = message.strip() + for key in ("resets_at", "reset_at"): + value = payload.get(key) + if value not in {None, ""}: + context["reset_at"] = value + break + retry_after = payload.get("retry_after") + if retry_after not in {None, ""} and "reset_at" not in context: + try: + context["reset_at"] = time.time() + float(retry_after) + except (TypeError, ValueError): + pass + + response = getattr(error, "response", None) + headers = getattr(response, "headers", None) + if headers: + retry_after = headers.get("retry-after") or headers.get("Retry-After") + if retry_after and "reset_at" not in context: + try: + context["reset_at"] = time.time() + float(retry_after) + except (TypeError, ValueError): + pass + ratelimit_reset = headers.get("x-ratelimit-reset") + if ratelimit_reset and "reset_at" not in context: + context["reset_at"] = ratelimit_reset + + if "message" not in context: + raw_message = str(error).strip() + if raw_message: + context["message"] = raw_message[:500] + + if "reset_at" not in context: + message = context.get("message") or "" + if isinstance(message, str): + delay_match = re.search(r"quotaResetDelay[:\s\"]+(\\d+(?:\\.\\d+)?)(ms|s)", message, re.IGNORECASE) + if delay_match: + value = float(delay_match.group(1)) + seconds = value / 1000.0 if delay_match.group(2).lower() == "ms" else value + context["reset_at"] = time.time() + seconds + else: + sec_match = re.search( + r"retry\s+(?:after\s+)?(\d+(?:\.\d+)?)\s*(?:sec|secs|seconds|s\b)", + message, + re.IGNORECASE, + ) + if sec_match: + context["reset_at"] = time.time() + float(sec_match.group(1)) + + return context + + + +def apply_pending_steer_to_tool_results(agent, messages: list, num_tool_msgs: int) -> None: + """Append any pending /steer text to the last tool result in this turn. + + Called at the end of a tool-call batch, before the next API call. + The steer is appended to the last ``role:"tool"`` message's content + with a clear marker so the model understands it came from the user + and NOT from the tool itself. Role alternation is preserved — + nothing new is inserted, we only modify existing content. + + Args: + messages: The running messages list. + num_tool_msgs: Number of tool results appended in this batch; + used to locate the tail slice safely. + """ + if num_tool_msgs <= 0 or not messages: + return + steer_text = agent._drain_pending_steer() + if not steer_text: + return + # Find the last tool-role message in the recent tail. Skipping + # non-tool messages defends against future code appending + # something else at the boundary. + target_idx = None + for j in range(len(messages) - 1, max(len(messages) - num_tool_msgs - 1, -1), -1): + msg = messages[j] + if isinstance(msg, dict) and msg.get("role") == "tool": + target_idx = j + break + if target_idx is None: + # No tool result in this batch (e.g. all skipped by interrupt); + # put the steer back so the caller's fallback path can deliver + # it as a normal next-turn user message. + _lock = getattr(agent, "_pending_steer_lock", None) + if _lock is not None: + with _lock: + if agent._pending_steer: + agent._pending_steer = agent._pending_steer + "\n" + steer_text + else: + agent._pending_steer = steer_text + else: + existing = getattr(agent, "_pending_steer", None) + agent._pending_steer = (existing + "\n" + steer_text) if existing else steer_text + return + marker = f"\n\nUser guidance: {steer_text}" + existing_content = messages[target_idx].get("content", "") + if not isinstance(existing_content, str): + # Anthropic multimodal content blocks — preserve them and append + # a text block at the end. + try: + blocks = list(existing_content) if existing_content else [] + blocks.append({"type": "text", "text": marker.lstrip()}) + messages[target_idx]["content"] = blocks + except Exception: + # Fall back to string replacement if content shape is unexpected. + messages[target_idx]["content"] = f"{existing_content}{marker}" + else: + messages[target_idx]["content"] = existing_content + marker + _ra().logger.info( + "Delivered /steer to agent after tool batch (%d chars): %s", + len(steer_text), + steer_text[:120] + ("..." if len(steer_text) > 120 else ""), + ) + + + +def force_close_tcp_sockets(client: Any) -> int: + """Force-close underlying TCP sockets to prevent CLOSE-WAIT accumulation. + + When a provider drops a connection mid-stream, httpx's ``client.close()`` + performs a graceful shutdown which leaves sockets in CLOSE-WAIT until the + OS times them out (often minutes). This method walks the httpx transport + pool and issues ``socket.shutdown(SHUT_RDWR)`` + ``socket.close()`` to + force an immediate TCP RST, freeing the file descriptors. + + Returns the number of sockets force-closed. + """ + import socket as _socket + + closed = 0 + try: + http_client = getattr(client, "_client", None) + if http_client is None: + return 0 + transport = getattr(http_client, "_transport", None) + if transport is None: + return 0 + pool = getattr(transport, "_pool", None) + if pool is None: + return 0 + # httpx uses httpcore connection pools; connections live in + # _connections (list) or _pool (list) depending on version. + connections = ( + getattr(pool, "_connections", None) + or getattr(pool, "_pool", None) + or [] + ) + for conn in list(connections): + stream = ( + getattr(conn, "_network_stream", None) + or getattr(conn, "_stream", None) + ) + if stream is None: + continue + sock = getattr(stream, "_sock", None) + if sock is None: + sock = getattr(stream, "stream", None) + if sock is not None: + sock = getattr(sock, "_sock", None) + if sock is None: + continue + try: + sock.shutdown(_socket.SHUT_RDWR) + except OSError: + pass + try: + sock.close() + except OSError: + pass + closed += 1 + except Exception as exc: + _ra().logger.debug("Force-close TCP sockets sweep error: %s", exc) + return closed + + + +__all__ = [ + "convert_to_trajectory_format", + "sanitize_tool_call_arguments", + "repair_message_sequence", + "strip_think_blocks", + "recover_with_credential_pool", + "try_recover_primary_transport", + "drop_thinking_only_and_merge_users", + "restore_primary_runtime", + "extract_reasoning", + "dump_api_request_debug", + "anthropic_prompt_cache_policy", + "create_openai_client", + "switch_model", + "invoke_tool", + "repair_tool_call", + "sanitize_api_messages", + "looks_like_codex_intermediate_ack", + "copy_reasoning_content_for_api", + "cleanup_dead_connections", + "extract_api_error_context", + "apply_pending_steer_to_tool_results", + "force_close_tcp_sockets", +] diff --git a/agent/auxiliary_client.py b/agent/auxiliary_client.py index cfc44e5f2a6..ba78833248e 100644 --- a/agent/auxiliary_client.py +++ b/agent/auxiliary_client.py @@ -755,7 +755,8 @@ class _CodexCompletionsAdapter: def _check_cancelled() -> None: if deadline is not None and time.monotonic() >= deadline: - timed_out.set() + if not timed_out.is_set(): + _close_client_on_timeout() raise TimeoutError(_timeout_message()) try: from tools.interrupt import is_interrupted @@ -1233,7 +1234,7 @@ def _read_nous_auth() -> Optional[dict]: def _nous_api_key(provider: dict) -> str: - """Extract the best API key from a Nous provider state dict.""" + """Extract the Nous runtime credential from the compatibility field.""" return provider.get("agent_key") or provider.get("access_token", "") @@ -1246,17 +1247,25 @@ def _resolve_nous_runtime_api(*, force_refresh: bool = False) -> Optional[tuple[ """Return fresh Nous runtime credentials when available. This mirrors the main agent's 401 recovery path and keeps auxiliary - clients aligned with the singleton auth store + mint flow instead of + clients aligned with the singleton auth store + JWT/mint flow instead of relying only on whatever raw tokens happen to be sitting in auth.json or the credential pool. """ try: - from hermes_cli.auth import resolve_nous_runtime_credentials + from hermes_cli.auth import ( + NOUS_INFERENCE_AUTH_MODE_AUTO, + NOUS_INFERENCE_AUTH_MODE_LEGACY, + resolve_nous_runtime_credentials, + ) creds = resolve_nous_runtime_credentials( min_key_ttl_seconds=max(60, int(os.getenv("HERMES_NOUS_MIN_KEY_TTL_SECONDS", "1800"))), timeout_seconds=float(os.getenv("HERMES_NOUS_TIMEOUT_SECONDS", "15")), - force_mint=force_refresh, + inference_auth_mode=( + NOUS_INFERENCE_AUTH_MODE_LEGACY + if force_refresh + else NOUS_INFERENCE_AUTH_MODE_AUTO + ), ) except Exception as exc: logger.debug("Auxiliary Nous runtime credential resolution failed: %s", exc) @@ -1473,7 +1482,7 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]: -def _try_openrouter(explicit_api_key: str = None) -> Tuple[Optional[OpenAI], Optional[str]]: +def _try_openrouter(explicit_api_key: str = None, model: str = None) -> Tuple[Optional[OpenAI], Optional[str]]: pool_present, entry = _select_pool_entry("openrouter") if pool_present: or_key = explicit_api_key or _pool_runtime_api_key(entry) @@ -1483,7 +1492,7 @@ def _try_openrouter(explicit_api_key: str = None) -> Tuple[Optional[OpenAI], Opt base_url = _pool_runtime_base_url(entry, OPENROUTER_BASE_URL) or OPENROUTER_BASE_URL logger.debug("Auxiliary client: OpenRouter via pool") return OpenAI(api_key=or_key, base_url=base_url, - default_headers=build_or_headers()), _OPENROUTER_MODEL + default_headers=build_or_headers()), model or _OPENROUTER_MODEL or_key = explicit_api_key or os.getenv("OPENROUTER_API_KEY") if not or_key: @@ -1491,7 +1500,7 @@ def _try_openrouter(explicit_api_key: str = None) -> Tuple[Optional[OpenAI], Opt return None, None logger.debug("Auxiliary client: OpenRouter") return OpenAI(api_key=or_key, base_url=OPENROUTER_BASE_URL, - default_headers=build_or_headers()), _OPENROUTER_MODEL + default_headers=build_or_headers()), model or _OPENROUTER_MODEL def _describe_openrouter_unavailable() -> str: @@ -2087,7 +2096,13 @@ def _is_payment_error(exc: Exception) -> bool: """Detect payment/credit/quota exhaustion errors. Returns True for HTTP 402 (Payment Required) and for 429/other errors - whose message indicates billing exhaustion rather than rate limiting. + whose message indicates billing exhaustion or daily quota exhaustion + rather than transient rate limiting. + + Daily token quota errors (e.g. Bedrock "Too many tokens per day", + Vertex AI "quota exceeded") are functionally equivalent to credit + exhaustion — the provider cannot serve the request until the quota + resets — and should trigger the same provider-fallback logic. """ status = getattr(exc, "status_code", None) if status == 402: @@ -2095,10 +2110,19 @@ def _is_payment_error(exc: Exception) -> bool: err_lower = str(exc).lower() # OpenRouter and other providers include "credits" or "afford" in 402 bodies, # but sometimes wrap them in 429 or other codes. + # Daily quota exhaustion from Bedrock, Vertex AI, and similar providers + # uses different language but is semantically identical to credit exhaustion. if status in {402, 429, None}: - if any(kw in err_lower for kw in ("credits", "insufficient funds", - "can only afford", "billing", - "payment required")): + if any(kw in err_lower for kw in ( + "credits", "insufficient funds", + "can only afford", "billing", + "payment required", + # Daily / monthly quota exhaustion keywords + "quota exceeded", "quota_exceeded", + "too many tokens per day", "daily limit", + "tokens per day", "daily quota", + "resource exhausted", # Vertex AI / gRPC quota errors + )): return True return False @@ -2500,12 +2524,15 @@ def _refresh_provider_credentials(provider: str) -> bool: _evict_cached_clients(normalized) return True if normalized == "nous": - from hermes_cli.auth import resolve_nous_runtime_credentials + from hermes_cli.auth import ( + NOUS_INFERENCE_AUTH_MODE_LEGACY, + resolve_nous_runtime_credentials, + ) creds = resolve_nous_runtime_credentials( min_key_ttl_seconds=max(60, int(os.getenv("HERMES_NOUS_MIN_KEY_TTL_SECONDS", "1800"))), timeout_seconds=float(os.getenv("HERMES_NOUS_TIMEOUT_SECONDS", "15")), - force_mint=True, + inference_auth_mode=NOUS_INFERENCE_AUTH_MODE_LEGACY, ) if not str(creds.get("api_key", "") or "").strip(): return False @@ -2579,6 +2606,133 @@ def _try_payment_fallback( return None, None, "" +def _try_main_agent_model_fallback( + failed_provider: str, + task: str = None, + reason: str = "error", +) -> Tuple[Optional[Any], Optional[str], str]: + """Last-resort fallback to the user's main agent provider + model. + + Used after the configured fallback_chain is exhausted (or empty) for + users with an explicit auxiliary provider. This is the "safety net" + layer: if nothing the user asked for can serve the request, try the + main chat model before giving up. + + Skips when the failed provider already IS the main provider (no point + retrying the same backend that just failed). + + Returns: + (client, model, provider_label) or (None, None, "") if no fallback. + """ + main_provider = (_read_main_provider() or "").strip() + main_model = (_read_main_model() or "").strip() + if not main_provider or not main_model or main_provider.lower() in {"auto", ""}: + return None, None, "" + + skip = (failed_provider or "").lower().strip() + if main_provider.lower() == skip: + # The thing that failed IS the main model — nothing to fall back to. + return None, None, "" + if _is_provider_unhealthy(main_provider): + _log_skip_unhealthy(main_provider, task) + return None, None, "" + + try: + client, resolved_model = resolve_provider_client( + provider=main_provider, model=main_model, + ) + except Exception: + client, resolved_model = None, None + + if client is None: + return None, None, "" + + label = f"main-agent({main_provider})" + logger.info( + "Auxiliary %s: %s on %s — falling back to main agent model %s (%s)", + task or "call", reason, failed_provider, label, resolved_model or main_model, + ) + return client, resolved_model or main_model, label + + +def _try_configured_fallback_chain( + task: str, + failed_provider: str, + reason: str = "error", +) -> Tuple[Optional[Any], Optional[str], str]: + """Try user-configured fallback_chain for a specific auxiliary task. + + Reads auxiliary..fallback_chain from config.yaml and tries each + entry in order. Each entry must have at least ``provider``; ``model``, + ``base_url``, and ``api_key`` are optional. + + Returns: + (client, model, provider_label) or (None, None, "") if no fallback. + """ + if not task: + return None, None, "" + + task_config = _get_auxiliary_task_config(task) + chain = task_config.get("fallback_chain") + if not chain or not isinstance(chain, list): + return None, None, "" + + skip = failed_provider.lower().strip() + tried = [] + + for i, entry in enumerate(chain): + if not isinstance(entry, dict): + continue + fb_provider = str(entry.get("provider", "")).strip() + if not fb_provider or fb_provider.lower() == skip: + continue + fb_model = str(entry.get("model", "")).strip() or None + fb_base_url = str(entry.get("base_url", "")).strip() or None + fb_api_key = str(entry.get("api_key", "")).strip() or None + + label = f"fallback_chain[{i}]({fb_provider})" + + try: + fb_client = _resolve_single_provider( + fb_provider, fb_model, fb_base_url, fb_api_key) + except Exception: + fb_client = None + + if fb_client is not None: + logger.info( + "Auxiliary %s: %s on %s — configured fallback to %s (%s)", + task, reason, failed_provider, label, fb_model or "default", + ) + return fb_client, fb_model, label + tried.append(label) + + if tried: + logger.debug( + "Auxiliary %s: configured fallback_chain exhausted (tried: %s)", + task, ", ".join(tried), + ) + return None, None, "" + + +def _resolve_single_provider( + provider: str, + model: Optional[str] = None, + base_url: Optional[str] = None, + api_key: Optional[str] = None, +) -> Optional[Any]: + """Resolve a single provider entry from fallback_chain to an OpenAI client. + + Uses the existing provider resolution infrastructure where possible. + """ + # Reuse resolve_provider_client which handles provider→client mapping + client, resolved_model = resolve_provider_client( + provider=provider, + model=model, + base_url=base_url, + api_key=api_key, + ) + return client + def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Optional[OpenAI], Optional[str]]: """Full auto-detection chain. @@ -3049,10 +3203,17 @@ def resolve_provider_client( if custom_entry: custom_base = custom_entry.get("base_url", "").strip() custom_key = custom_entry.get("api_key", "").strip() - custom_key_env = custom_entry.get("key_env", "").strip() + custom_key_env = (custom_entry.get("key_env") or custom_entry.get("api_key_env") or "").strip() if not custom_key and custom_key_env: custom_key = os.getenv(custom_key_env, "").strip() custom_key = custom_key or "no-key-required" + if custom_key == "no-key-required": + logger.warning( + "resolve_provider_client: named custom provider %r has no resolvable " + "api_key — request will be sent with placeholder no-key-required " + "and will 401 on auth-required endpoints", + custom_entry.get("name") or provider, + ) # An explicit per-task api_mode override (from _resolve_task_provider_model) # wins; otherwise fall back to what the provider entry declared. entry_api_mode = (api_mode or custom_entry.get("api_mode") or "").strip() @@ -3400,7 +3561,7 @@ def _resolve_strict_vision_backend( if provider == "copilot": return resolve_provider_client("copilot", model, is_vision=True) if provider == "openrouter": - return _try_openrouter() + return _try_openrouter(model=model) if provider == "nous": return _try_nous(vision=True) if provider == "openai-codex": @@ -4519,11 +4680,17 @@ def call_llm( or _is_connection_error(first_err) or _is_rate_limit_error(first_err) ) - # Only try alternative providers when the user didn't explicitly - # configure this task's provider. Explicit provider = hard constraint; - # auto (the default) = best-effort fallback chain. (#7559) + # Respect explicit provider choice for transient errors (auth, request + # validation, etc.) but allow fallback when the provider clearly cannot + # serve the request due to capacity: payment/quota exhaustion and + # connection failures are capacity problems, not request constraints. + # See #26803: daily token quota (429 + "too many tokens per day") must + # fall back just like a 402 credit error. is_auto = resolved_provider in {"auto", "", None} - if should_fallback and is_auto: + # Capacity errors bypass the explicit-provider gate: the provider + # literally cannot serve this request regardless of user intent. + is_capacity_error = _is_payment_error(first_err) or _is_connection_error(first_err) + if should_fallback and (is_auto or is_capacity_error): if _is_payment_error(first_err): reason = "payment error" # Resolve the actual provider label (resolved_provider may be @@ -4539,8 +4706,24 @@ def call_llm( reason = "connection error" logger.info("Auxiliary %s: %s on %s (%s), trying fallback", task or "call", reason, resolved_provider, first_err) - fb_client, fb_model, fb_label = _try_payment_fallback( - resolved_provider, task, reason=reason) + + # Fallback order (#26882, #26803): + # 1. User-configured fallback_chain (per-task) if set + # 2. Main agent model (last-resort safety net) + # For auto users (no explicit aux provider), use the full + # auto-detection chain instead — its Step 1 IS the main agent + # model, so users on `auto` already get main-model fallback. + fb_client, fb_model, fb_label = (None, None, "") + if is_auto: + fb_client, fb_model, fb_label = _try_payment_fallback( + resolved_provider, task, reason=reason) + else: + fb_client, fb_model, fb_label = _try_configured_fallback_chain( + task, resolved_provider or "auto", reason=reason) + if fb_client is None: + fb_client, fb_model, fb_label = _try_main_agent_model_fallback( + resolved_provider, task, reason=reason) + if fb_client is not None: fb_kwargs = _build_call_kwargs( fb_label, fb_model, messages, @@ -4550,6 +4733,14 @@ def call_llm( base_url=str(getattr(fb_client, "base_url", "") or "")) return _validate_llm_response( fb_client.chat.completions.create(**fb_kwargs), task) + # All fallback layers exhausted — emit a single user-visible + # warning so the operator knows aux task is about to fail. + # (#26882) The error itself is re-raised below. + logger.warning( + "Auxiliary %s: %s on %s and all fallbacks exhausted " + "(fallback_chain + main agent model). Raising original error.", + task or "call", reason, resolved_provider, + ) # Connection/timeout errors leave the cached client poisoned (closed # httpx transport, half-read stream, dead async loop). Drop it from # the cache regardless of whether we found a fallback above so the @@ -4851,8 +5042,12 @@ async def async_call_llm( or _is_connection_error(first_err) or _is_rate_limit_error(first_err) ) + # Capacity errors (payment/quota/connection) bypass the explicit-provider + # gate — the provider cannot serve the request regardless of user intent. + # See #26803: daily token quota must fall back like a 402 credit error. is_auto = resolved_provider in {"auto", "", None} - if should_fallback and is_auto: + is_capacity_error = _is_payment_error(first_err) or _is_connection_error(first_err) + if should_fallback and (is_auto or is_capacity_error): if _is_payment_error(first_err): reason = "payment error" _mark_provider_unhealthy( @@ -4864,8 +5059,23 @@ async def async_call_llm( reason = "connection error" logger.info("Auxiliary %s (async): %s on %s (%s), trying fallback", task or "call", reason, resolved_provider, first_err) - fb_client, fb_model, fb_label = _try_payment_fallback( - resolved_provider, task, reason=reason) + + # Fallback order (#26882, #26803): + # 1. User-configured fallback_chain (per-task) if set + # 2. Main agent model (last-resort safety net) + # Auto users get the full auto-detection chain instead — its + # Step 1 IS the main agent model. + fb_client, fb_model, fb_label = (None, None, "") + if is_auto: + fb_client, fb_model, fb_label = _try_payment_fallback( + resolved_provider, task, reason=reason) + else: + fb_client, fb_model, fb_label = _try_configured_fallback_chain( + task, resolved_provider or "auto", reason=reason) + if fb_client is None: + fb_client, fb_model, fb_label = _try_main_agent_model_fallback( + resolved_provider, task, reason=reason) + if fb_client is not None: fb_kwargs = _build_call_kwargs( fb_label, fb_model, messages, @@ -4881,6 +5091,12 @@ async def async_call_llm( fb_kwargs["model"] = async_fb_model return _validate_llm_response( await async_fb.chat.completions.create(**fb_kwargs), task) + # All fallback layers exhausted — warn before re-raising. (#26882) + logger.warning( + "Auxiliary %s (async): %s on %s and all fallbacks exhausted " + "(fallback_chain + main agent model). Raising original error.", + task or "call", reason, resolved_provider, + ) # Mirror the sync path: drop poisoned clients on connection/timeout # so the next aux call rebuilds. See issue #23432. if _is_connection_error(first_err): diff --git a/agent/background_review.py b/agent/background_review.py new file mode 100644 index 00000000000..83292029c6c --- /dev/null +++ b/agent/background_review.py @@ -0,0 +1,570 @@ +"""Background memory/skill review — fork the agent to evaluate the turn. + +After every turn, ``AIAgent.run_conversation`` may call +:func:`spawn_background_review` to fire off a daemon thread that replays +the conversation snapshot in a forked :class:`AIAgent` and asks itself +"should any skill/memory be saved or updated?". Writes go straight to +the memory + skill stores. Main conversation and prompt cache are never +touched. + +The fork inherits the parent's live runtime (provider, model, base_url, +credentials, cached system prompt) so it hits the same prefix cache and +uses the same auth. It runs with a tool whitelist limited to memory and +skill management tools; everything else is denied at runtime. + +See the ``hermes-agent-dev`` skill (``references/self-improvement-loop.md``) +for invariants and PR review criteria. +""" + +from __future__ import annotations + +import contextlib +import json +import logging +import os +from typing import Any, Dict, List, Optional + +logger = logging.getLogger(__name__) + + +# Review-prompt strings — used by ``spawn_background_review_thread`` to build +# the user-message that the forked review agent receives. AIAgent exposes +# them as class attributes (``_MEMORY_REVIEW_PROMPT`` etc.) for back-compat; +# the actual text lives here so future edits are one-place. +_MEMORY_REVIEW_PROMPT = ( + "Review the conversation above and consider saving to memory if appropriate.\n\n" + "Focus on:\n" + "1. Has the user revealed things about themselves — their persona, desires, " + "preferences, or personal details worth remembering?\n" + "2. Has the user expressed expectations about how you should behave, their work " + "style, or ways they want you to operate?\n\n" + "If something stands out, save it using the memory tool. " + "If nothing is worth saving, just say 'Nothing to save.' and stop." +) + +_SKILL_REVIEW_PROMPT = ( + "Review the conversation above and update the skill library. Be " + "ACTIVE — most sessions produce at least one skill update, even if " + "small. A pass that does nothing is a missed learning opportunity, " + "not a neutral outcome.\n\n" + "Target shape of the library: CLASS-LEVEL skills, each with a rich " + "SKILL.md and a `references/` directory for session-specific detail. " + "Not a long flat list of narrow one-session-one-skill entries. This " + "shapes HOW you update, not WHETHER you update.\n\n" + "Signals to look for (any one of these warrants action):\n" + " • User corrected your style, tone, format, legibility, or " + "verbosity. Frustration signals like 'stop doing X', 'this is too " + "verbose', 'don't format like this', 'why are you explaining', " + "'just give me the answer', 'you always do Y and I hate it', or an " + "explicit 'remember this' are FIRST-CLASS skill signals, not just " + "memory signals. Update the relevant skill(s) to embed the " + "preference so the next session starts already knowing.\n" + " • User corrected your workflow, approach, or sequence of steps. " + "Encode the correction as a pitfall or explicit step in the skill " + "that governs that class of task.\n" + " • Non-trivial technique, fix, workaround, debugging path, or " + "tool-usage pattern emerged that a future session would benefit " + "from. Capture it.\n" + " • A skill that got loaded or consulted this session turned out " + "to be wrong, missing a step, or outdated. Patch it NOW.\n\n" + "Preference order — prefer the earliest action that fits, but do " + "pick one when a signal above fired:\n" + " 1. UPDATE A CURRENTLY-LOADED SKILL. Look back through the " + "conversation for skills the user loaded via /skill-name or you " + "read via skill_view. If any of them covers the territory of the " + "new learning, PATCH that one first. It is the skill that was in " + "play, so it's the right one to extend.\n" + " 2. UPDATE AN EXISTING UMBRELLA (via skills_list + skill_view). " + "If no loaded skill fits but an existing class-level skill does, " + "patch it. Add a subsection, a pitfall, or broaden a trigger.\n" + " 3. ADD A SUPPORT FILE under an existing umbrella. Skills can be " + "packaged with three kinds of support files — use the right " + "directory per kind:\n" + " • `references/.md` — session-specific detail (error " + "transcripts, reproduction recipes, provider quirks) AND " + "condensed knowledge banks: quoted research, API docs, external " + "authoritative excerpts, or domain notes you found while working " + "on the problem. Write it concise and for the value of the task, " + "not as a full mirror of upstream docs.\n" + " • `templates/.` — starter files meant to be " + "copied and modified (boilerplate configs, scaffolding, a " + "known-good example the agent can `reproduce with modifications`).\n" + " • `scripts/.` — statically re-runnable actions " + "the skill can invoke directly (verification scripts, fixture " + "generators, deterministic probes, anything the agent should run " + "rather than hand-type each time).\n" + " Add support files via skill_manage action=write_file with " + "file_path starting 'references/', 'templates/', or 'scripts/'. " + "The umbrella's SKILL.md should gain a one-line pointer to any " + "new support file so future agents know it exists.\n" + " 4. CREATE A NEW CLASS-LEVEL UMBRELLA SKILL when no existing " + "skill covers the class. The name MUST be at the class level. " + "The name MUST NOT be a specific PR number, error string, feature " + "codename, library-alone name, or 'fix-X / debug-Y / audit-Z-today' " + "session artifact. If the proposed name only makes sense for " + "today's task, it's wrong — fall back to (1), (2), or (3).\n\n" + "User-preference embedding (important): when the user expressed a " + "style/format/workflow preference, the update belongs in the " + "SKILL.md body, not just in memory. Memory captures 'who the user " + "is and what the current situation and state of your operations " + "are'; skills capture 'how to do this class of task for this " + "user'. When they complain about how you handled a task, the " + "skill that governs that task needs to carry the lesson.\n\n" + "If you notice two existing skills that overlap, note it in your " + "reply — the background curator handles consolidation at scale.\n\n" + "Do NOT capture (these become persistent self-imposed constraints " + "that bite you later when the environment changes):\n" + " • Environment-dependent failures: missing binaries, fresh-install " + "errors, post-migration path mismatches, 'command not found', " + "unconfigured credentials, uninstalled packages. The user can fix " + "these — they are not durable rules.\n" + " • Negative claims about tools or features ('browser tools do not " + "work', 'X tool is broken', 'cannot use Y from execute_code'). These " + "harden into refusals the agent cites against itself for months " + "after the actual problem was fixed.\n" + " • Session-specific transient errors that resolved before the " + "conversation ended. If retrying worked, the lesson is the retry " + "pattern, not the original failure.\n" + " • One-off task narratives. A user asking 'summarize today's " + "market' or 'analyze this PR' is not a class of work that warrants " + "a skill.\n\n" + "If a tool failed because of setup state, capture the FIX (install " + "command, config step, env var to set) under an existing setup or " + "troubleshooting skill — never 'this tool does not work' as a " + "standalone constraint.\n\n" + "'Nothing to save.' is a real option but should NOT be the " + "default. If the session ran smoothly with no corrections and " + "produced no new technique, just say 'Nothing to save.' and stop. " + "Otherwise, act." +) + +_COMBINED_REVIEW_PROMPT = ( + "Review the conversation above and update two things:\n\n" + "**Memory**: who the user is. Did the user reveal persona, " + "desires, preferences, personal details, or expectations about " + "how you should behave? Save facts about the user and durable " + "preferences with the memory tool.\n\n" + "**Skills**: how to do this class of task. Be ACTIVE — most " + "sessions produce at least one skill update. A pass that does " + "nothing is a missed learning opportunity, not a neutral outcome.\n\n" + "Target shape of the skill library: CLASS-LEVEL skills with a rich " + "SKILL.md and a `references/` directory for session-specific detail. " + "Not a long flat list of narrow one-session-one-skill entries.\n\n" + "Signals that warrant a skill update (any one is enough):\n" + " • User corrected your style, tone, format, legibility, " + "verbosity, or approach. Frustration is a FIRST-CLASS skill " + "signal, not just a memory signal. 'stop doing X', 'don't format " + "like this', 'I hate when you Y' — embed the lesson in the skill " + "that governs that task so the next session starts fixed.\n" + " • Non-trivial technique, fix, workaround, or debugging path " + "emerged.\n" + " • A skill that was loaded or consulted turned out wrong, " + "missing, or outdated — patch it now.\n\n" + "Preference order for skills — pick the earliest that fits:\n" + " 1. UPDATE A CURRENTLY-LOADED SKILL. Check what skills were " + "loaded via /skill-name or skill_view in the conversation. If one " + "of them covers the learning, PATCH it first. It was in play; " + "it's the right place.\n" + " 2. UPDATE AN EXISTING UMBRELLA (skills_list + skill_view to " + "find the right one). Patch it.\n" + " 3. ADD A SUPPORT FILE under an existing umbrella via " + "skill_manage action=write_file. Three kinds: " + "`references/.md` for session-specific detail OR condensed " + "knowledge banks (quoted research, API docs excerpts, domain " + "notes) written concise and task-focused; `templates/.` " + "for starter files meant to be copied and modified; " + "`scripts/.` for statically re-runnable actions " + "(verification, fixture generators, probes). Add a one-line " + "pointer in SKILL.md so future agents find them.\n" + " 4. CREATE A NEW CLASS-LEVEL UMBRELLA when nothing exists. " + "Name at the class level — NOT a PR number, error string, " + "codename, library-alone name, or 'fix-X / debug-Y' session " + "artifact. If the name only fits today's task, fall back to (1), " + "(2), or (3).\n\n" + "User-preference embedding: when the user complains about how " + "you handled a task, update the skill that governs that task — " + "memory alone isn't enough. Memory says 'who the user is and " + "what the current situation and state of your operations are'; " + "skills say 'how to do this class of task for this user'. Both " + "should carry user-preference lessons when relevant.\n\n" + "If you notice overlapping existing skills, mention it — the " + "background curator handles consolidation.\n\n" + "Do NOT capture as skills (these become persistent self-imposed " + "constraints that bite you later when the environment changes):\n" + " • Environment-dependent failures: missing binaries, fresh-install " + "errors, post-migration path mismatches, 'command not found', " + "unconfigured credentials, uninstalled packages. The user can fix " + "these — they are not durable rules.\n" + " • Negative claims about tools or features ('browser tools do not " + "work', 'X tool is broken', 'cannot use Y from execute_code'). These " + "harden into refusals the agent cites against itself for months " + "after the actual problem was fixed.\n" + " • Session-specific transient errors that resolved before the " + "conversation ended. If retrying worked, the lesson is the retry " + "pattern, not the original failure.\n" + " • One-off task narratives. A user asking 'summarize today's " + "market' or 'analyze this PR' is not a class of work that warrants " + "a skill.\n\n" + "If a tool failed because of setup state, capture the FIX (install " + "command, config step, env var to set) under an existing setup or " + "troubleshooting skill — never 'this tool does not work' as a " + "standalone constraint.\n\n" + "Act on whichever of the two dimensions has real signal. If " + "genuinely nothing stands out on either, say 'Nothing to save.' " + "and stop — but don't reach for that conclusion as a default." +) + + + +def summarize_background_review_actions( + review_messages: List[Dict], + prior_snapshot: List[Dict], +) -> List[str]: + """Build the human-facing action summary for a background review pass. + + Walks the review agent's session messages and collects "successful tool + action" descriptions to surface to the user (e.g. "Memory updated"). + Tool messages already present in ``prior_snapshot`` are skipped so we + don't re-surface stale results from the prior conversation that the + review agent inherited via ``conversation_history`` (issue #14944). + + Matching is by ``tool_call_id`` when available, with a content-equality + fallback for tool messages that lack one. + """ + existing_tool_call_ids = set() + existing_tool_contents = set() + for prior in prior_snapshot or []: + if not isinstance(prior, dict) or prior.get("role") != "tool": + continue + tcid = prior.get("tool_call_id") + if tcid: + existing_tool_call_ids.add(tcid) + else: + content = prior.get("content") + if isinstance(content, str): + existing_tool_contents.add(content) + + actions: List[str] = [] + for msg in review_messages or []: + if not isinstance(msg, dict) or msg.get("role") != "tool": + continue + tcid = msg.get("tool_call_id") + if tcid and tcid in existing_tool_call_ids: + continue + if not tcid: + content_str = msg.get("content") + if isinstance(content_str, str) and content_str in existing_tool_contents: + continue + try: + data = json.loads(msg.get("content", "{}")) + except (json.JSONDecodeError, TypeError): + continue + if not isinstance(data, dict) or not data.get("success"): + continue + message = data.get("message", "") + target = data.get("target", "") + if "created" in message.lower(): + actions.append(message) + elif "updated" in message.lower(): + actions.append(message) + elif "added" in message.lower() or (target and "add" in message.lower()): + label = "Memory" if target == "memory" else "User profile" if target == "user" else target + actions.append(f"{label} updated") + elif "Entry added" in message: + label = "Memory" if target == "memory" else "User profile" if target == "user" else target + actions.append(f"{label} updated") + elif "removed" in message.lower() or "replaced" in message.lower(): + label = "Memory" if target == "memory" else "User profile" if target == "user" else target + actions.append(f"{label} updated") + return actions + + +def build_memory_write_metadata( + agent: Any, + *, + write_origin: Optional[str] = None, + execution_context: Optional[str] = None, + task_id: Optional[str] = None, + tool_call_id: Optional[str] = None, +) -> Dict[str, Any]: + """Build provenance metadata for external memory-provider mirrors.""" + metadata: Dict[str, Any] = { + "write_origin": write_origin or getattr(agent, "_memory_write_origin", "assistant_tool"), + "execution_context": ( + execution_context + or getattr(agent, "_memory_write_context", "foreground") + ), + "session_id": agent.session_id or "", + "parent_session_id": agent._parent_session_id or "", + "platform": agent.platform or os.environ.get("HERMES_SESSION_SOURCE", "cli"), + "tool_name": "memory", + } + if task_id: + metadata["task_id"] = task_id + if tool_call_id: + metadata["tool_call_id"] = tool_call_id + return {k: v for k, v in metadata.items() if v not in {None, ""}} + + +def _run_review_in_thread( + agent: Any, + messages_snapshot: List[Dict], + prompt: str, +) -> None: + """Worker function executed in the background-review daemon thread. + + Spawns a forked ``AIAgent`` inheriting the parent's runtime, runs the + review prompt, and surfaces a compact action summary back to the user + via ``agent._safe_print`` and ``agent.background_review_callback``. + """ + # Local import to avoid a hard circular dep at module load. + from run_agent import AIAgent + from tools.terminal_tool import set_approval_callback as _set_approval_callback + + # Install a non-interactive approval callback on this worker + # thread so any dangerous-command guard the review agent trips + # resolves to "deny" instead of falling back to input() -- which + # deadlocks against the parent's prompt_toolkit TUI (#15216). + # Same pattern as _subagent_auto_deny in tools/delegate_tool.py. + def _bg_review_auto_deny(command, description, **kwargs): + logger.warning( + "Background review auto-denied dangerous command: %s (%s)", + command, description, + ) + return "deny" + try: + _set_approval_callback(_bg_review_auto_deny) + except Exception: + pass + + review_agent = None + review_messages: List[Dict] = [] + try: + with open(os.devnull, "w", encoding="utf-8") as _devnull, \ + contextlib.redirect_stdout(_devnull), \ + contextlib.redirect_stderr(_devnull): + # Inherit the parent agent's live runtime (provider, model, + # base_url, api_key, api_mode) so the fork uses the exact + # same credentials the main turn is using. Without this, + # AIAgent.__init__ re-runs auto-resolution from env vars, + # which fails for OAuth-only providers, session-scoped + # creds, or credential-pool setups where the resolver can't + # reconstruct auth from scratch -- producing the spurious + # "No LLM provider configured" warning at end of turn. + _parent_runtime = agent._current_main_runtime() + _parent_api_mode = _parent_runtime.get("api_mode") or None + # The review fork needs to call agent-loop tools (memory, + # skill_manage). Those tools require Hermes' own dispatch, + # which the codex_app_server runtime bypasses entirely + # (it runs the turn inside codex's subprocess). So when + # the parent is on codex_app_server, downgrade the review + # fork to codex_responses — same auth/credentials, but + # talks to the OpenAI Responses API directly so Hermes + # owns the loop and the agent-loop tools dispatch. + if _parent_api_mode == "codex_app_server": + _parent_api_mode = "codex_responses" + # skip_memory=True keeps the review fork from + # touching external memory plugins (honcho, mem0, + # supermemory, etc.). Without it, the fork's + # __init__ rebuilds its own _memory_manager from + # config, scoped to the parent's session_id, and + # run_conversation() then leaks the harness prompt + # into the user's real memory namespace via three + # ingestion sites: on_turn_start (cadence + turn + # message), prefetch_all (recall query), and + # sync_all (harness prompt + review output recorded + # as a (user, assistant) turn pair). Built-in + # MEMORY.md / USER.md state is re-bound from the + # parent below so memory(action="add") writes from + # the review still land on disk; the review just + # has zero side effects on external providers. + review_agent = AIAgent( + model=agent.model, + max_iterations=16, + quiet_mode=True, + platform=agent.platform, + provider=agent.provider, + api_mode=_parent_api_mode, + base_url=_parent_runtime.get("base_url") or None, + api_key=_parent_runtime.get("api_key") or None, + credential_pool=getattr(agent, "_credential_pool", None), + parent_session_id=agent.session_id, + skip_memory=True, + ) + review_agent._memory_write_origin = "background_review" + review_agent._memory_write_context = "background_review" + review_agent._memory_store = agent._memory_store + review_agent._memory_enabled = agent._memory_enabled + review_agent._user_profile_enabled = agent._user_profile_enabled + review_agent._memory_nudge_interval = 0 + review_agent._skill_nudge_interval = 0 + # Suppress all status/warning emits from the fork so the + # user only sees the final successful-action summary. + # Without this, mid-review "Iteration budget exhausted", + # rate-limit retries, compression warnings, and other + # lifecycle messages bubble up through _emit_status -> + # _vprint and leak past the stdout redirect (they go via + # _print_fn/status_callback, which bypass sys.stdout). + review_agent.suppress_status_output = True + # Inherit the parent's cached system prompt verbatim so + # the review fork's outbound HTTP request hits the same + # Anthropic/OpenRouter prefix cache the parent warmed. + # Without this, the fork rebuilds the system prompt from + # scratch (fresh _hermes_now() timestamp, fresh + # session_id, narrower toolset → different skills_prompt) + # and the byte-exact prefix-cache key misses. See + # issue #25322 and PR #17276 for the full analysis + + # measured impact (~26% end-to-end cost reduction on + # Sonnet 4.5). + review_agent._cached_system_prompt = agent._cached_system_prompt + # Defensive: pin session_start + session_id to the + # parent's so any code path that re-renders parts of + # the system prompt (compression, plugin hooks) still + # produces byte-identical output. The cached-prompt + # assignment above already short-circuits the normal + # rebuild path, but these pins guarantee parity even + # if a future code path bypasses the cache. + review_agent.session_start = agent.session_start + review_agent.session_id = agent.session_id + + from model_tools import get_tool_definitions + from hermes_cli.plugins import ( + set_thread_tool_whitelist, + clear_thread_tool_whitelist, + ) + + review_whitelist = { + t["function"]["name"] + for t in get_tool_definitions( + enabled_toolsets=["memory", "skills"], + quiet_mode=True, + ) + } + set_thread_tool_whitelist( + review_whitelist, + deny_msg_fmt=( + "Background review denied non-whitelisted tool: " + "{tool_name}. Only memory/skill tools are allowed." + ), + ) + try: + review_agent.run_conversation( + user_message=( + prompt + + "\n\nYou can only call memory and skill " + "management tools. Other tools will be denied " + "at runtime — do not attempt them." + ), + conversation_history=messages_snapshot, + ) + finally: + clear_thread_tool_whitelist() + + # Tear down memory providers while stdout is still + # redirected so background thread teardown (Honcho flush, + # Hindsight sync, etc.) stays silent. The finally block + # below is a safety net for the exception path. + try: + review_agent.shutdown_memory_provider() + except Exception: + pass + try: + review_agent.close() + except Exception: + pass + review_messages = list(getattr(review_agent, "_session_messages", [])) + review_agent = None + + # Scan the review agent's messages for successful tool actions + # and surface a compact summary to the user. Tool messages + # already present in messages_snapshot must be skipped, since + # the review agent inherits that history and would otherwise + # re-surface stale "created"/"updated" messages from the prior + # conversation as if they just happened (issue #14944). + actions = summarize_background_review_actions( + review_messages, + messages_snapshot, + ) + + if actions: + summary = " · ".join(dict.fromkeys(actions)) + agent._safe_print( + f" 💾 Self-improvement review: {summary}" + ) + _bg_cb = agent.background_review_callback + if _bg_cb: + try: + _bg_cb( + f"💾 Self-improvement review: {summary}" + ) + except Exception: + pass + + except Exception as e: + logger.warning("Background memory/skill review failed: %s", e) + agent._emit_auxiliary_failure("background review", e) + finally: + # Safety-net cleanup for the exception path. Normal + # completion already shut down inside redirect_stdout above. + # Re-open devnull here so any teardown output (Honcho flush, + # Hindsight sync, background thread joins) stays silent even + # on the exception path where redirect_stdout already exited. + if review_agent is not None: + try: + with open(os.devnull, "w", encoding="utf-8") as _fn, \ + contextlib.redirect_stdout(_fn), \ + contextlib.redirect_stderr(_fn): + try: + review_agent.shutdown_memory_provider() + except Exception: + pass + try: + review_agent.close() + except Exception: + pass + except Exception: + pass + # Clear the approval callback on this bg-review thread so a + # recycled thread-id doesn't inherit a stale reference. + try: + _set_approval_callback(None) + except Exception: + pass + + +def spawn_background_review_thread( + agent: Any, + messages_snapshot: List[Dict], + review_memory: bool = False, + review_skills: bool = False, +): + """Build the review thread target and prompt for a background review. + + Returns a ``(target, prompt)`` tuple. The caller (``AIAgent._spawn_background_review``) + owns the actual ``threading.Thread`` construction so test-level patches + of ``run_agent.threading.Thread`` keep working. + """ + # Pick the right prompt based on which triggers fired. Allow per-agent + # override (the prompts moved to module-level constants but old code paths + # that set agent._MEMORY_REVIEW_PROMPT etc. directly keep working). + if review_memory and review_skills: + prompt = getattr(agent, "_COMBINED_REVIEW_PROMPT", _COMBINED_REVIEW_PROMPT) + elif review_memory: + prompt = getattr(agent, "_MEMORY_REVIEW_PROMPT", _MEMORY_REVIEW_PROMPT) + else: + prompt = getattr(agent, "_SKILL_REVIEW_PROMPT", _SKILL_REVIEW_PROMPT) + + def _target() -> None: + _run_review_in_thread(agent, messages_snapshot, prompt) + + return _target, prompt + + +__all__ = [ + "_MEMORY_REVIEW_PROMPT", + "_SKILL_REVIEW_PROMPT", + "_COMBINED_REVIEW_PROMPT", + "spawn_background_review_thread", + "summarize_background_review_actions", + "build_memory_write_metadata", +] diff --git a/agent/bedrock_adapter.py b/agent/bedrock_adapter.py index 34eebd73ba8..620d1c99785 100644 --- a/agent/bedrock_adapter.py +++ b/agent/bedrock_adapter.py @@ -36,6 +36,19 @@ from typing import Any, Dict, List, Optional, Tuple logger = logging.getLogger(__name__) +# --------------------------------------------------------------------------- +# Ensure boto3/botocore are installed before any code in this module runs. +# Upstream removed boto3 from [all] extras (PRs #24220, #24515); lazy_deps +# handles on-demand installation so the Bedrock provider still works in the +# EKS deployment without baking boto3 into the base image. +# --------------------------------------------------------------------------- +try: + from tools.lazy_deps import ensure + ensure("provider.bedrock", prompt=False) +except Exception: + pass # lazy_deps unavailable or install failed — let downstream imports surface the real error + + # --------------------------------------------------------------------------- # Lazy boto3 import — only loaded when the Bedrock provider is actually used. # This keeps startup fast for users who don't use Bedrock. diff --git a/agent/browser_provider.py b/agent/browser_provider.py new file mode 100644 index 00000000000..75e88e584f3 --- /dev/null +++ b/agent/browser_provider.py @@ -0,0 +1,175 @@ +""" +Browser Provider ABC +==================== + +Defines the pluggable-backend interface for cloud browser providers +(Browserbase, Browser Use, Firecrawl, …). Providers register instances via +:meth:`PluginContext.register_browser_provider`; the active one (selected via +``browser.cloud_provider`` in ``config.yaml``) services every cloud-mode +``browser_*`` tool call. + +Providers live in ``/plugins/browser//`` (built-in, auto-loaded as +``kind: backend``) or ``~/.hermes/plugins/browser//`` (user, opt-in via +``plugins.enabled``). + +This ABC mirrors :class:`agent.web_search_provider.WebSearchProvider` (PR +#25182) — same shape, same registration flow, same picker integration. The +legacy in-tree ``tools.browser_providers.base.CloudBrowserProvider`` ABC was +deleted in PR #25214 (this work) along with the per-vendor inline modules in +``tools/browser_providers/``; the lifecycle contract documented below is +preserved bit-for-bit so the tool wrapper (:mod:`tools.browser_tool`) does +not have to translate. + +Session metadata contract (preserved from the legacy ``CloudBrowserProvider``):: + + { + "session_name": str, # unique name for agent-browser --session + "bb_session_id": str, # provider session ID (for close/cleanup) + "cdp_url": str, # CDP websocket URL + "features": dict, # feature flags that were enabled + "external_call_id": str, # optional, managed-gateway billing key + } + +``bb_session_id`` is a legacy key name kept verbatim for backward compat with +:mod:`tools.browser_tool` — it holds the provider's session ID regardless of +which provider is in use. +""" + +from __future__ import annotations + +import abc +from typing import Any, Dict + + +# --------------------------------------------------------------------------- +# ABC +# --------------------------------------------------------------------------- + + +class BrowserProvider(abc.ABC): + """Abstract base class for a cloud browser backend. + + Subclasses must implement :meth:`name`, :meth:`is_available`, and the + three lifecycle methods: :meth:`create_session`, :meth:`close_session`, + :meth:`emergency_cleanup`. + + The lifecycle shape preserves the legacy ``CloudBrowserProvider`` contract + bit-for-bit so the dispatcher in :mod:`tools.browser_tool` is a pure + registry lookup — no per-provider conditionals, no shape translation. + """ + + @property + @abc.abstractmethod + def name(self) -> str: + """Stable short identifier used in the ``browser.cloud_provider`` + config key. + + Lowercase, hyphens permitted to preserve existing user-visible names. + Examples: ``browserbase``, ``browser-use``, ``firecrawl``. + """ + + @property + def display_name(self) -> str: + """Human-readable label shown in ``hermes tools``. Defaults to ``name``.""" + return self.name + + @abc.abstractmethod + def is_available(self) -> bool: + """Return True when this provider can service calls. + + Typically a cheap check (env var present, managed-gateway token + readable, optional Python dep importable). Must NOT make network + calls — this runs at tool-registration time and on every + ``hermes tools`` paint. + + Mirrors the legacy ``CloudBrowserProvider.is_configured()`` method; + renamed for parity with :class:`agent.web_search_provider.WebSearchProvider`. + """ + + @abc.abstractmethod + def create_session(self, task_id: str) -> Dict[str, object]: + """Create a cloud browser session and return session metadata. + + Must return a dict with at least:: + + { + "session_name": str, # unique name for agent-browser --session + "bb_session_id": str, # provider session ID (for close/cleanup) + "cdp_url": str, # CDP websocket URL + "features": dict, # feature flags that were enabled + } + + ``bb_session_id`` is a legacy key name kept for backward compat with + the rest of :mod:`tools.browser_tool` — it holds the provider's + session ID regardless of which provider is in use. + + May raise ``ValueError`` (missing credentials) or ``RuntimeError`` + (network / API failure); the dispatcher surfaces these to the user. + """ + + @abc.abstractmethod + def close_session(self, session_id: str) -> bool: + """Release / terminate a cloud session by its provider session ID. + + Returns True on success, False on failure. Should not raise — log and + return False on any exception so the dispatcher's cleanup loop keeps + moving across sessions. + """ + + @abc.abstractmethod + def emergency_cleanup(self, session_id: str) -> None: + """Best-effort session teardown during process exit. + + Called from atexit / signal handlers. Must tolerate missing + credentials, network errors, etc. — log and move on. Must not raise. + """ + + def get_setup_schema(self) -> Dict[str, Any]: + """Return provider metadata for the ``hermes tools`` picker. + + Used by :mod:`hermes_cli.tools_config` to inject this provider as a + row in the Browser Automation picker. Shape mirrors the existing + hardcoded entries in ``TOOL_CATEGORIES["browser"]``:: + + { + "name": "Browserbase", + "badge": "paid", + "tag": "Cloud browser with stealth and proxies", + "env_vars": [ + {"key": "BROWSERBASE_API_KEY", + "prompt": "Browserbase API key", + "url": "https://browserbase.com"}, + ], + "post_setup": "agent_browser", + } + + Default: minimal entry derived from :attr:`display_name`. Override to + expose API key prompts, badges, managed-Nous gating, and the + ``post_setup`` install hook. + """ + return { + "name": self.display_name, + "badge": "", + "tag": "", + "env_vars": [], + } + + # ------------------------------------------------------------------ + # Backward-compat shims for the legacy CloudBrowserProvider API + # ------------------------------------------------------------------ + # + # The pre-PR-#25214 ABC exposed ``is_configured()`` and ``provider_name()``; + # ``tools.browser_tool`` has ~6 callers that still use those names. Rather + # than churn every callsite (and break out-of-tree downstream code that + # subclassed CloudBrowserProvider), we expose the old names as thin + # delegations to the new API. Subclasses MUST implement :meth:`is_available` + # and :attr:`name`; they may override ``is_configured`` / ``provider_name`` + # for compatibility with the legacy ABC but it is not required. + + def is_configured(self) -> bool: + """Backward-compat alias for :meth:`is_available`.""" + return self.is_available() + + def provider_name(self) -> str: + """Backward-compat alias returning :attr:`display_name`.""" + return self.display_name diff --git a/agent/browser_registry.py b/agent/browser_registry.py new file mode 100644 index 00000000000..db608744b34 --- /dev/null +++ b/agent/browser_registry.py @@ -0,0 +1,223 @@ +""" +Browser Provider Registry +========================= + +Central map of registered cloud browser providers. Populated by plugins at +import-time via :meth:`PluginContext.register_browser_provider`; consumed by +:func:`tools.browser_tool._get_cloud_provider` to route each cloud-mode +``browser_*`` tool call to the active backend. + +Active selection +---------------- +The active provider is chosen by configuration with this precedence: + +1. ``browser.cloud_provider`` in ``config.yaml`` (explicit override). +2. Legacy preference order — ``browser-use`` → ``browserbase`` — filtered by + availability. Matches the historic auto-detect order in + :func:`tools.browser_tool._get_cloud_provider` (Browser Use checked first + because it covers both the managed Nous gateway and direct API key path; + Browserbase as the older direct-credentials fallback). ``firecrawl`` is + intentionally NOT in the legacy walk — users only get Firecrawl as a + cloud browser when they explicitly set ``browser.cloud_provider: + firecrawl``, matching pre-migration behaviour where Firecrawl was never + auto-selected. +3. Otherwise ``None`` — the dispatcher falls back to local browser mode. + +The explicit-config branch (rule 1) intentionally ignores ``is_available()`` +so the dispatcher surfaces a typed "X_API_KEY is not set" error to the user +instead of silently switching backends. Matches the legacy +:func:`tools.browser_tool._get_cloud_provider` behaviour for configured names. + +Note: there is no "capability" split here (unlike the web subsystem, which +has search/extract/crawl). Every browser provider implements the full +:class:`agent.browser_provider.BrowserProvider` lifecycle; the registry's +job is purely selection, not capability routing. +""" + +from __future__ import annotations + +import logging +import threading +from typing import Dict, List, Optional + +from agent.browser_provider import BrowserProvider + +logger = logging.getLogger(__name__) + + +_providers: Dict[str, BrowserProvider] = {} +_lock = threading.Lock() + + +def register_provider(provider: BrowserProvider) -> None: + """Register a cloud browser provider. + + Re-registration (same ``name``) overwrites the previous entry and logs + a debug message — makes hot-reload scenarios (tests, dev loops) behave + predictably. + """ + if not isinstance(provider, BrowserProvider): + raise TypeError( + f"register_provider() expects a BrowserProvider instance, " + f"got {type(provider).__name__}" + ) + name = provider.name + if not isinstance(name, str) or not name.strip(): + raise ValueError("Browser provider .name must be a non-empty string") + with _lock: + existing = _providers.get(name) + _providers[name] = provider + if existing is not None: + logger.debug( + "Browser provider '%s' re-registered (was %r)", + name, type(existing).__name__, + ) + else: + logger.debug( + "Registered browser provider '%s' (%s)", + name, type(provider).__name__, + ) + + +def list_providers() -> List[BrowserProvider]: + """Return all registered providers, sorted by name.""" + with _lock: + items = list(_providers.values()) + return sorted(items, key=lambda p: p.name) + + +def get_provider(name: str) -> Optional[BrowserProvider]: + """Return the provider registered under *name*, or None.""" + if not isinstance(name, str): + return None + with _lock: + return _providers.get(name.strip()) + + +# --------------------------------------------------------------------------- +# Active-provider resolution +# --------------------------------------------------------------------------- + + +# Legacy auto-detect order — used when no ``browser.cloud_provider`` is set. +# Matches the pre-migration walk in :func:`tools.browser_tool._get_cloud_provider`. +# Firecrawl is intentionally absent so users with ``FIRECRAWL_API_KEY`` set +# for web-extract don't get silently routed to a paid cloud browser. See +# :func:`_resolve` for the full rationale. +_LEGACY_PREFERENCE = ( + "browser-use", + "browserbase", +) + + +def _resolve(configured: Optional[str]) -> Optional[BrowserProvider]: + """Resolve the active browser provider. + + Resolution rules (in order): + + 1. **Explicit "local".** Returns None — the dispatcher disables cloud + mode entirely. Mirrors legacy short-circuit in + :func:`tools.browser_tool._get_cloud_provider`. + 2. **Explicit config wins, ignoring availability.** If ``configured`` + names a registered provider, return it even if its + :meth:`is_available` returns False — the dispatcher will surface a + precise "X_API_KEY is not set" error instead of silently routing + somewhere else. + 3. **Legacy preference walk, filtered by availability.** Walk + :data:`_LEGACY_PREFERENCE` (``browser-use`` → ``browserbase``) looking + for a provider whose ``is_available()`` is True. + + There is intentionally NO "single-eligible shortcut" rule here (unlike + :func:`agent.web_search_registry._resolve`). Pre-migration, the + auto-detect branch in ``tools.browser_tool._get_cloud_provider`` only + considered Browser Use and Browserbase; Firecrawl was reachable only + via an explicit ``browser.cloud_provider: firecrawl`` config key. + Preserving that gate matters because Firecrawl shares its API key with + the *web* extract plugin (``plugins/web/firecrawl/``), so users who set + ``FIRECRAWL_API_KEY`` for web extract must NOT get silently routed to a + paid cloud browser on a fresh install. Third-party browser-provider + plugins added under ``~/.hermes/plugins/browser//`` are subject + to the same gate — they must be explicitly configured to take effect. + + Returns None when no provider is configured AND no available provider + matches the legacy preference; the dispatcher then falls back to local + browser mode. + """ + with _lock: + snapshot = dict(_providers) + + def _is_available_safe(p: BrowserProvider) -> bool: + """Wrap ``is_available()`` so a buggy provider doesn't kill resolution.""" + try: + return bool(p.is_available()) + except Exception as exc: # noqa: BLE001 + logger.warning( + "Browser provider %s.is_available() raised %s — treating as unavailable", + p.name, exc, exc_info=True, + ) + return False + + # 1. Explicit "local" short-circuit. + if configured == "local": + return None + + # 2. Explicit config wins — return regardless of is_available() so the + # user gets a precise downstream error message rather than a silent + # backend switch. Matches _get_cloud_provider() in browser_tool.py. + if configured: + provider = snapshot.get(configured) + if provider is not None: + return provider + logger.debug( + "browser cloud_provider '%s' configured but not registered; " + "falling back to auto-detect", + configured, + ) + + # 3. Legacy preference walk — only providers in _LEGACY_PREFERENCE are + # auto-eligible. Filtered by availability so we don't surface a + # provider the user has no credentials for. See docstring for why + # we do NOT fall back to "any single-eligible registered provider". + for legacy in _LEGACY_PREFERENCE: + provider = snapshot.get(legacy) + if provider is not None and _is_available_safe(provider): + return provider + + return None + + +def get_active_browser_provider() -> Optional[BrowserProvider]: + """Resolve the currently-active cloud browser provider. + + Reads ``browser.cloud_provider`` from config.yaml; falls back per the + module docstring. Returns None for local mode or when no provider is + available. + """ + try: + from hermes_cli.config import read_raw_config + + cfg = read_raw_config() + browser_cfg = cfg.get("browser", {}) + except Exception as exc: + logger.debug("Could not read browser config: %s", exc) + browser_cfg = {} + + configured: Optional[str] = None + if isinstance(browser_cfg, dict) and "cloud_provider" in browser_cfg: + try: + from tools.tool_backend_helpers import normalize_browser_cloud_provider + + configured = normalize_browser_cloud_provider( + browser_cfg.get("cloud_provider") + ) + except Exception as exc: + logger.debug("normalize_browser_cloud_provider failed: %s", exc) + configured = None + + return _resolve(configured) + + +def _reset_for_tests() -> None: + """Clear the registry. **Test-only.**""" + with _lock: + _providers.clear() diff --git a/agent/chat_completion_helpers.py b/agent/chat_completion_helpers.py new file mode 100644 index 00000000000..ee5b957bf2f --- /dev/null +++ b/agent/chat_completion_helpers.py @@ -0,0 +1,2066 @@ +"""Helper functions for the chat-completions code path. + +Extracted from :class:`AIAgent` for cleanliness — bodies of the +non-streaming API call, request kwargs builder, assistant-message +materializer, provider-fallback activator, max-iterations handler, +and per-turn resource cleanup. + +Each function takes the parent ``AIAgent`` as its first argument +(``agent``). :class:`AIAgent` keeps thin forwarder methods so call +sites unchanged. Symbols that tests patch on ``run_agent`` (e.g. +``cleanup_vm`` / ``cleanup_browser`` in +``test_zombie_process_cleanup.py``) are resolved through +:func:`_ra` so the patch contract is preserved. +""" + +from __future__ import annotations + +import concurrent.futures +import contextvars +import copy +import json +import logging +import os +import random +import re +import sys +import threading +import time +import uuid +from datetime import datetime +from pathlib import Path +from types import SimpleNamespace +from typing import Any, Dict, List, Optional, Tuple +from urllib.parse import urlparse, parse_qs, urlunparse + +from hermes_cli.timeouts import get_provider_request_timeout, get_provider_stale_timeout +from agent.error_classifier import classify_api_error, FailoverReason +from agent.model_metadata import is_local_endpoint +from agent.message_sanitization import ( + _sanitize_surrogates, + _sanitize_messages_surrogates, + _sanitize_structure_surrogates, + _sanitize_messages_non_ascii, + _sanitize_tools_non_ascii, + _sanitize_structure_non_ascii, + _strip_images_from_messages, + _strip_non_ascii, + _repair_tool_call_arguments, + _escape_invalid_chars_in_json_strings, +) +from agent.tool_dispatch_helpers import ( + _is_multimodal_tool_result, + _multimodal_text_summary, +) +from agent.retry_utils import jittered_backoff +from agent.tool_guardrails import ( + ToolGuardrailDecision, + append_toolguard_guidance, + toolguard_synthetic_result, +) +from tools.terminal_tool import is_persistent_env +from utils import base_url_host_matches, base_url_hostname + +logger = logging.getLogger(__name__) + + +def _ra(): + """Lazy ``run_agent`` reference. + + Used to honor test patches like + ``patch("run_agent.cleanup_vm")`` / ``patch("run_agent.cleanup_browser")`` + that target symbols imported into ``run_agent``'s namespace. + """ + import run_agent + return run_agent + + + +def interruptible_api_call(agent, api_kwargs: dict): + """ + Run the API call in a background thread so the main conversation loop + can detect interrupts without waiting for the full HTTP round-trip. + + Each worker thread gets its own OpenAI client instance. Interrupts only + close that worker-local client, so retries and other requests never + inherit a closed transport. + + Includes a stale-call detector: if no response arrives within the + configured timeout, the connection is killed and an error raised so + the main retry loop can try again with backoff / credential rotation / + provider fallback. + """ + result = {"response": None, "error": None} + request_client_holder = {"client": None} + + def _call(): + try: + if agent.api_mode == "codex_responses": + request_client_holder["client"] = agent._create_request_openai_client( + reason="codex_stream_request", + api_kwargs=api_kwargs, + ) + result["response"] = agent._run_codex_stream( + api_kwargs, + client=request_client_holder["client"], + on_first_delta=getattr(agent, "_codex_on_first_delta", None), + ) + elif agent.api_mode == "anthropic_messages": + result["response"] = agent._anthropic_messages_create(api_kwargs) + elif agent.api_mode == "bedrock_converse": + # Bedrock uses boto3 directly — no OpenAI client needed. + # normalize_converse_response produces an OpenAI-compatible + # SimpleNamespace so the rest of the agent loop can treat + # bedrock responses like chat_completions responses. + from agent.bedrock_adapter import ( + _get_bedrock_runtime_client, + invalidate_runtime_client, + is_stale_connection_error, + normalize_converse_response, + ) + region = api_kwargs.pop("__bedrock_region__", "us-east-1") + api_kwargs.pop("__bedrock_converse__", None) + client = _get_bedrock_runtime_client(region) + try: + raw_response = client.converse(**api_kwargs) + except Exception as _bedrock_exc: + # Evict the cached client on stale-connection failures + # so the outer retry loop builds a fresh client/pool. + if is_stale_connection_error(_bedrock_exc): + invalidate_runtime_client(region) + raise + result["response"] = normalize_converse_response(raw_response) + else: + request_client_holder["client"] = agent._create_request_openai_client( + reason="chat_completion_request", + api_kwargs=api_kwargs, + ) + result["response"] = request_client_holder["client"].chat.completions.create(**api_kwargs) + except Exception as e: + result["error"] = e + finally: + request_client = request_client_holder.get("client") + if request_client is not None: + agent._close_request_openai_client(request_client, reason="request_complete") + + # ── Stale-call timeout (mirrors streaming stale detector) ──────── + # Non-streaming calls return nothing until the full response is + # ready. Without this, a hung provider can block for the full + # httpx timeout (default 1800s) with zero feedback. The stale + # detector kills the connection early so the main retry loop can + # apply richer recovery (credential rotation, provider fallback). + _stale_timeout = agent._compute_non_stream_stale_timeout( + api_kwargs.get("messages", []) + ) + + _call_start = time.time() + agent._touch_activity("waiting for non-streaming API response") + + t = threading.Thread(target=_call, daemon=True) + t.start() + _poll_count = 0 + while t.is_alive(): + t.join(timeout=0.3) + _poll_count += 1 + + # Touch activity every ~30s so the gateway's inactivity + # monitor knows we're alive while waiting for the response. + if _poll_count % 100 == 0: # 100 × 0.3s = 30s + _elapsed = time.time() - _call_start + agent._touch_activity( + f"waiting for non-streaming response ({int(_elapsed)}s elapsed)" + ) + + # Stale-call detector: kill the connection if no response + # arrives within the configured timeout. + _elapsed = time.time() - _call_start + if _elapsed > _stale_timeout: + _est_ctx = sum(len(str(v)) for v in api_kwargs.get("messages", [])) // 4 + logger.warning( + "Non-streaming API call stale for %.0fs (threshold %.0fs). " + "model=%s context=~%s tokens. Killing connection.", + _elapsed, _stale_timeout, + api_kwargs.get("model", "unknown"), f"{_est_ctx:,}", + ) + agent._emit_status( + f"⚠️ No response from provider for {int(_elapsed)}s " + f"(non-streaming, model: {api_kwargs.get('model', 'unknown')}). " + f"Aborting call." + ) + try: + if agent.api_mode == "anthropic_messages": + agent._anthropic_client.close() + agent._rebuild_anthropic_client() + else: + rc = request_client_holder.get("client") + if rc is not None: + agent._close_request_openai_client(rc, reason="stale_call_kill") + except Exception: + pass + agent._touch_activity( + f"stale non-streaming call killed after {int(_elapsed)}s" + ) + # Wait briefly for the thread to notice the closed connection. + t.join(timeout=2.0) + if result["error"] is None and result["response"] is None: + result["error"] = TimeoutError( + f"Non-streaming API call timed out after {int(_elapsed)}s " + f"with no response (threshold: {int(_stale_timeout)}s)" + ) + break + + if agent._interrupt_requested: + # Force-close the in-flight worker-local HTTP connection to stop + # token generation without poisoning the shared client used to + # seed future retries. + try: + if agent.api_mode == "anthropic_messages": + agent._anthropic_client.close() + agent._rebuild_anthropic_client() + else: + request_client = request_client_holder.get("client") + if request_client is not None: + agent._close_request_openai_client(request_client, reason="interrupt_abort") + except Exception: + pass + raise InterruptedError("Agent interrupted during API call") + if result["error"] is not None: + raise result["error"] + return result["response"] + + + +def build_api_kwargs(agent, api_messages: list) -> dict: + """Build the keyword arguments dict for the active API mode.""" + tools_for_api = agent.tools + + if agent.api_mode == "anthropic_messages": + _transport = agent._get_transport() + anthropic_messages = agent._prepare_anthropic_messages_for_api(api_messages) + ctx_len = getattr(agent, "context_compressor", None) + ctx_len = ctx_len.context_length if ctx_len else None + ephemeral_out = getattr(agent, "_ephemeral_max_output_tokens", None) + if ephemeral_out is not None: + agent._ephemeral_max_output_tokens = None # consume immediately + return _transport.build_kwargs( + model=agent.model, + messages=anthropic_messages, + tools=tools_for_api, + max_tokens=ephemeral_out if ephemeral_out is not None else agent.max_tokens, + reasoning_config=agent.reasoning_config, + is_oauth=agent._is_anthropic_oauth, + preserve_dots=agent._anthropic_preserve_dots(), + context_length=ctx_len, + base_url=getattr(agent, "_anthropic_base_url", None), + fast_mode=(agent.request_overrides or {}).get("speed") == "fast", + drop_context_1m_beta=bool(getattr(agent, "_oauth_1m_beta_disabled", False)), + ) + + # AWS Bedrock native Converse API — bypasses the OpenAI client entirely. + # The adapter handles message/tool conversion and boto3 calls directly. + if agent.api_mode == "bedrock_converse": + _bt = agent._get_transport() + region = getattr(agent, "_bedrock_region", None) or "us-east-1" + guardrail = getattr(agent, "_bedrock_guardrail_config", None) + return _bt.build_kwargs( + model=agent.model, + messages=api_messages, + tools=tools_for_api, + max_tokens=agent.max_tokens or 4096, + region=region, + guardrail_config=guardrail, + ) + + if agent.api_mode == "codex_responses": + _ct = agent._get_transport() + is_github_responses = ( + base_url_host_matches(agent.base_url, "models.github.ai") + or base_url_host_matches(agent.base_url, "api.githubcopilot.com") + ) + is_codex_backend = ( + agent.provider == "openai-codex" + or ( + agent._base_url_hostname == "chatgpt.com" + and "/backend-api/codex" in agent._base_url_lower + ) + ) + is_xai_responses = agent.provider in {"xai", "xai-oauth"} or agent._base_url_hostname == "api.x.ai" + _msgs_for_codex = agent._prepare_messages_for_non_vision_model(api_messages) + + # xAI's /responses endpoint rejects ``pattern`` and ``format`` keywords + # in tool schemas (HTTP 400 "Invalid arguments passed to the model"). + # Most commonly hit when MCP-derived tools carry JSON Schema validation + # keywords through. Strip them before building kwargs. See #27197. + if is_xai_responses: + try: + from tools.schema_sanitizer import strip_pattern_and_format + tools_for_api, _ = strip_pattern_and_format(tools_for_api) + except Exception as exc: + logger.warning( + "%s⚠️ Failed to sanitize tool schemas for xAI: %s", + getattr(agent, "log_prefix", ""), exc, + ) + + return _ct.build_kwargs( + model=agent.model, + messages=_msgs_for_codex, + tools=tools_for_api, + reasoning_config=agent.reasoning_config, + session_id=getattr(agent, "session_id", None), + max_tokens=agent.max_tokens, + request_overrides=agent.request_overrides, + is_github_responses=is_github_responses, + is_codex_backend=is_codex_backend, + is_xai_responses=is_xai_responses, + github_reasoning_extra=agent._github_models_reasoning_extra_body() if is_github_responses else None, + ) + + # ── chat_completions (default) ───────────────────────────────────── + _ct = agent._get_transport() + + # Provider detection flags + _is_qwen = agent._is_qwen_portal() + _is_or = agent._is_openrouter_url() + _is_gh = ( + base_url_host_matches(agent._base_url_lower, "models.github.ai") + or base_url_host_matches(agent._base_url_lower, "api.githubcopilot.com") + ) + _is_nous = "nousresearch" in agent._base_url_lower + _is_nvidia = "integrate.api.nvidia.com" in agent._base_url_lower + _is_kimi = ( + base_url_host_matches(agent.base_url, "api.kimi.com") + or base_url_host_matches(agent.base_url, "moonshot.ai") + or base_url_host_matches(agent.base_url, "moonshot.cn") + ) + _is_tokenhub = base_url_host_matches(agent._base_url_lower, "tokenhub.tencentmaas.com") + _is_lmstudio = (agent.provider or "").strip().lower() == "lmstudio" + + # Temperature: _fixed_temperature_for_model may return OMIT_TEMPERATURE + # sentinel (temperature omitted entirely), a numeric override, or None. + try: + from agent.auxiliary_client import _fixed_temperature_for_model, OMIT_TEMPERATURE + _ft = _fixed_temperature_for_model(agent.model, agent.base_url) + _omit_temp = _ft is OMIT_TEMPERATURE + _fixed_temp = _ft if not _omit_temp else None + except Exception: + _omit_temp = False + _fixed_temp = None + + # Provider preferences (OpenRouter-style) + _prefs: Dict[str, Any] = {} + if agent.providers_allowed: + _prefs["only"] = agent.providers_allowed + if agent.providers_ignored: + _prefs["ignore"] = agent.providers_ignored + if agent.providers_order: + _prefs["order"] = agent.providers_order + if agent.provider_sort: + _prefs["sort"] = agent.provider_sort + if agent.provider_require_parameters: + _prefs["require_parameters"] = True + if agent.provider_data_collection: + _prefs["data_collection"] = agent.provider_data_collection + + # Claude max-output override on aggregators + _ant_max = None + if (_is_or or _is_nous) and "claude" in (agent.model or "").lower(): + try: + from agent.anthropic_adapter import _get_anthropic_max_output + _ant_max = _get_anthropic_max_output(agent.model) + except Exception: + pass + + # Qwen session metadata + _qwen_meta = None + if _is_qwen: + _qwen_meta = { + "sessionId": agent.session_id or "hermes", + "promptId": str(uuid.uuid4()), + } + + # ── Provider profile path (registered providers) ─────────────────── + # Profiles handle per-provider quirks via hooks. When a profile is + # found, delegate fully; otherwise fall through to the legacy flag path. + try: + from providers import get_provider_profile + _profile = get_provider_profile(agent.provider) + except Exception: + _profile = None + + if _profile: + _ephemeral_out = getattr(agent, "_ephemeral_max_output_tokens", None) + if _ephemeral_out is not None: + agent._ephemeral_max_output_tokens = None + + # Strip image parts for non-vision models that have provider profiles + # (e.g. DeepSeek, Kimi). The legacy path below already does this, but + # registered providers with profiles were bypassing the strip. + api_messages = agent._prepare_messages_for_non_vision_model(api_messages) + + return _ct.build_kwargs( + model=agent.model, + messages=api_messages, + tools=tools_for_api, + base_url=agent.base_url, + timeout=agent._resolved_api_call_timeout(), + max_tokens=agent.max_tokens, + ephemeral_max_output_tokens=_ephemeral_out, + max_tokens_param_fn=agent._max_tokens_param, + reasoning_config=agent.reasoning_config, + request_overrides=agent.request_overrides, + session_id=getattr(agent, "session_id", None), + provider_profile=_profile, + ollama_num_ctx=agent._ollama_num_ctx, + # Context forwarded to profile hooks: + provider_preferences=_prefs or None, + openrouter_min_coding_score=agent.openrouter_min_coding_score, + anthropic_max_output=_ant_max, + supports_reasoning=agent._supports_reasoning_extra_body(), + qwen_session_metadata=_qwen_meta, + ) + + # ── Legacy flag path ──────────────────────────────────────────── + # Reached only when get_provider_profile() returns None — i.e. a + # completely unknown provider not in providers/ registry. + _ephemeral_out = getattr(agent, "_ephemeral_max_output_tokens", None) + if _ephemeral_out is not None: + agent._ephemeral_max_output_tokens = None + + # Strip image parts for non-vision models (no-op when vision-capable). + _msgs_for_chat = agent._prepare_messages_for_non_vision_model(api_messages) + + return _ct.build_kwargs( + model=agent.model, + messages=_msgs_for_chat, + tools=tools_for_api, + base_url=agent.base_url, + timeout=agent._resolved_api_call_timeout(), + max_tokens=agent.max_tokens, + ephemeral_max_output_tokens=_ephemeral_out, + max_tokens_param_fn=agent._max_tokens_param, + reasoning_config=agent.reasoning_config, + request_overrides=agent.request_overrides, + session_id=getattr(agent, "session_id", None), + model_lower=(agent.model or "").lower(), + is_openrouter=_is_or, + is_nous=_is_nous, + is_qwen_portal=_is_qwen, + is_github_models=_is_gh, + is_nvidia_nim=_is_nvidia, + is_kimi=_is_kimi, + is_tokenhub=_is_tokenhub, + is_lmstudio=_is_lmstudio, + is_custom_provider=agent.provider == "custom", + ollama_num_ctx=agent._ollama_num_ctx, + provider_preferences=_prefs or None, + openrouter_min_coding_score=agent.openrouter_min_coding_score, + qwen_prepare_fn=agent._qwen_prepare_chat_messages if _is_qwen else None, + qwen_prepare_inplace_fn=agent._qwen_prepare_chat_messages_inplace if _is_qwen else None, + qwen_session_metadata=_qwen_meta, + fixed_temperature=_fixed_temp, + omit_temperature=_omit_temp, + supports_reasoning=agent._supports_reasoning_extra_body(), + github_reasoning_extra=agent._github_models_reasoning_extra_body() if _is_gh else None, + lmstudio_reasoning_options=agent._lmstudio_reasoning_options_cached() if _is_lmstudio else None, + anthropic_max_output=_ant_max, + provider_name=agent.provider, + ) + + + +def build_assistant_message(agent, assistant_message, finish_reason: str) -> dict: + """Build a normalized assistant message dict from an API response message. + + Handles reasoning extraction, reasoning_details, and optional tool_calls + so both the tool-call path and the final-response path share one builder. + """ + assistant_tool_calls = getattr(assistant_message, "tool_calls", None) + reasoning_text = agent._extract_reasoning(assistant_message) + _from_structured = bool(reasoning_text) + + # Fallback: extract inline blocks from content when no structured + # reasoning fields are present (some models/providers embed thinking + # directly in the content rather than returning separate API fields). + if not reasoning_text: + content = assistant_message.content or "" + think_blocks = re.findall(r'(.*?)', content, flags=re.DOTALL) + if think_blocks: + combined = "\n\n".join(b.strip() for b in think_blocks if b.strip()) + reasoning_text = combined or None + + if reasoning_text and agent.verbose_logging: + logging.debug(f"Captured reasoning ({len(reasoning_text)} chars): {reasoning_text}") + + if reasoning_text and agent.reasoning_callback: + # Skip callback when streaming is active — reasoning was already + # displayed during the stream via one of two paths: + # (a) _fire_reasoning_delta (structured reasoning_content deltas) + # (b) _stream_delta tag extraction (/) + # When streaming is NOT active, always fire so non-streaming modes + # (gateway, batch, quiet) still get reasoning. + # Any reasoning that wasn't shown during streaming is caught by the + # CLI post-response display fallback (cli.py _reasoning_shown_this_turn). + if not agent.stream_delta_callback and not agent._stream_callback: + try: + agent.reasoning_callback(reasoning_text) + except Exception: + pass + + # Sanitize surrogates from API response — some models (e.g. Kimi/GLM via Ollama) + # can return invalid surrogate code points that crash json.dumps() on persist. + _raw_content = assistant_message.content or "" + _san_content = _sanitize_surrogates(_raw_content) + if reasoning_text: + reasoning_text = _sanitize_surrogates(reasoning_text) + + # Strip inline reasoning tags (… etc.) from the stored + # assistant content. Reasoning was already captured into + # ``reasoning_text`` above (either from structured fields or the + # inline-block fallback), so the raw tags in content are redundant. + # Leaving them in place caused reasoning to leak to messaging + # platforms (#8878, #9568), inflate context on subsequent turns + # (#9306 observed 16% content-size reduction on a real MiniMax + # session), and pollute generated session titles. One strip at the + # storage boundary cleans content for every downstream consumer: + # API replay, session transcript, gateway delivery, CLI display, + # compression, title generation. + if isinstance(_san_content, str) and _san_content: + _san_content = agent._strip_think_blocks(_san_content).strip() + + msg = { + "role": "assistant", + "content": _san_content, + "reasoning": reasoning_text, + "finish_reason": finish_reason, + } + + raw_reasoning_content = getattr(assistant_message, "reasoning_content", None) + if raw_reasoning_content is None and hasattr(assistant_message, "model_extra"): + model_extra = getattr(assistant_message, "model_extra", None) or {} + if isinstance(model_extra, dict) and "reasoning_content" in model_extra: + raw_reasoning_content = model_extra["reasoning_content"] + if raw_reasoning_content is not None: + msg["reasoning_content"] = _sanitize_surrogates(raw_reasoning_content) + elif assistant_tool_calls and agent._needs_thinking_reasoning_pad(): + # DeepSeek v4 thinking mode and Kimi / Moonshot thinking mode + # both require reasoning_content on every assistant tool-call + # message. Without it, replaying the persisted message causes + # HTTP 400 ("The reasoning_content in the thinking mode must + # be passed back to the API"). Include streamed reasoning + # text when captured; otherwise pad with a single space — + # DeepSeek V4 Pro tightened validation and rejects empty + # string ("The reasoning content in the thinking mode must + # be passed back to the API"). A space satisfies non-empty + # checks everywhere without leaking fabricated reasoning. + # Refs #15250, #17400, #17341. + msg["reasoning_content"] = reasoning_text or " " + + # Additive fallback (refs #16844, #16884). Streaming-only providers + # (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) + # accumulate reasoning through ``delta.reasoning_content`` chunks + # but never land it on the message object as a top-level attribute, + # so neither branch above fires and the chain-of-thought is stored + # only under the internal ``reasoning`` key. When the user later + # replays that history through a DeepSeek-v4 / Kimi thinking model, + # the missing ``reasoning_content`` causes HTTP 400 ("The + # reasoning_content in the thinking mode must be passed back to the + # API."). + # + # Promote the already-sanitized streamed ``reasoning_text`` to + # ``reasoning_content`` at write time, but ONLY when no prior branch + # already set it AND we actually captured reasoning text. This + # preserves every existing behavior: + # - SDK-exposed ``reasoning_content`` (OpenAI/Moonshot/DeepSeek SDK) + # still wins. + # - DeepSeek tool-call ""-pad (#15250) still fires. + # - Non-thinking turns with no reasoning leave the field absent, + # so ``_copy_reasoning_content_for_api``'s cross-provider leak + # guard (#15748) and ``reasoning``→``reasoning_content`` + # promotion tiers still apply at replay time. + if "reasoning_content" not in msg and reasoning_text: + msg["reasoning_content"] = reasoning_text + + if hasattr(assistant_message, 'reasoning_details') and assistant_message.reasoning_details: + # Pass reasoning_details back unmodified so providers (OpenRouter, + # Anthropic, OpenAI) can maintain reasoning continuity across turns. + # Each provider may include opaque fields (signature, encrypted_content) + # that must be preserved exactly. + raw_details = assistant_message.reasoning_details + preserved = [] + for d in raw_details: + if isinstance(d, dict): + preserved.append(d) + elif hasattr(d, "__dict__"): + preserved.append(d.__dict__) + elif hasattr(d, "model_dump"): + preserved.append(d.model_dump()) + if preserved: + msg["reasoning_details"] = preserved + + # Codex Responses API: preserve encrypted reasoning items for + # multi-turn continuity. These get replayed as input on the next turn. + codex_items = getattr(assistant_message, "codex_reasoning_items", None) + if codex_items: + msg["codex_reasoning_items"] = codex_items + + # Codex Responses API: preserve exact assistant message items (with + # id/phase) so follow-up turns can replay structured items instead of + # flattening to plain text. This is required for prefix cache hits. + codex_message_items = getattr(assistant_message, "codex_message_items", None) + if codex_message_items: + msg["codex_message_items"] = codex_message_items + + if assistant_tool_calls: + tool_calls = [] + for tool_call in assistant_tool_calls: + raw_id = getattr(tool_call, "id", None) + call_id = getattr(tool_call, "call_id", None) + if not isinstance(call_id, str) or not call_id.strip(): + embedded_call_id, _ = agent._split_responses_tool_id(raw_id) + call_id = embedded_call_id + if not isinstance(call_id, str) or not call_id.strip(): + if isinstance(raw_id, str) and raw_id.strip(): + call_id = raw_id.strip() + else: + _fn = getattr(tool_call, "function", None) + _fn_name = getattr(_fn, "name", "") if _fn else "" + _fn_args = getattr(_fn, "arguments", "{}") if _fn else "{}" + call_id = agent._deterministic_call_id(_fn_name, _fn_args, len(tool_calls)) + call_id = call_id.strip() + + response_item_id = getattr(tool_call, "response_item_id", None) + if not isinstance(response_item_id, str) or not response_item_id.strip(): + _, embedded_response_item_id = agent._split_responses_tool_id(raw_id) + response_item_id = embedded_response_item_id + + response_item_id = agent._derive_responses_function_call_id( + call_id, + response_item_id if isinstance(response_item_id, str) else None, + ) + + tc_dict = { + "id": call_id, + "call_id": call_id, + "response_item_id": response_item_id, + "type": tool_call.type, + "function": { + "name": tool_call.function.name, + "arguments": tool_call.function.arguments + }, + } + # Preserve extra_content (e.g. Gemini thought_signature) so it + # is sent back on subsequent API calls. Without this, Gemini 3 + # thinking models reject the request with a 400 error. + extra = getattr(tool_call, "extra_content", None) + if extra is not None: + if hasattr(extra, "model_dump"): + extra = extra.model_dump() + tc_dict["extra_content"] = extra + tool_calls.append(tc_dict) + msg["tool_calls"] = tool_calls + + return msg + + + +def try_activate_fallback(agent, reason: "FailoverReason | None" = None) -> bool: + """Switch to the next fallback model/provider in the chain. + + Called when the current model is failing after retries. Swaps the + OpenAI client, model slug, and provider in-place so the retry loop + can continue with the new backend. Advances through the chain on + each call; returns False when exhausted. + + Uses the centralized provider router (resolve_provider_client) for + auth resolution and client construction — no duplicated provider→key + mappings. + """ + if reason in {FailoverReason.rate_limit, FailoverReason.billing}: + # Only start cooldown when leaving the primary provider. If we're + # already on a fallback and chain-switching, the primary wasn't the + # source of the 429 so the cooldown should not be reset/extended. + fallback_already_active = bool(getattr(agent, "_fallback_activated", False)) + current_provider = (getattr(agent, "provider", "") or "").strip().lower() + primary_provider = ((agent._primary_runtime or {}).get("provider") or "").strip().lower() + if (not fallback_already_active) or (primary_provider and current_provider == primary_provider): + agent._rate_limited_until = time.monotonic() + 60 + if agent._fallback_index >= len(agent._fallback_chain): + return False + + fb = agent._fallback_chain[agent._fallback_index] + agent._fallback_index += 1 + fb_provider = (fb.get("provider") or "").strip().lower() + fb_model = (fb.get("model") or "").strip() + if not fb_provider or not fb_model: + return agent._try_activate_fallback() # skip invalid, try next + + # Skip entries that resolve to the current (provider, model) — falling + # back to the same backend that just failed loops the failure. Compare + # base_url too so two distinct custom_providers entries pointing at the + # same shim/proxy URL also dedup. See issue #22548. + current_provider = (getattr(agent, "provider", "") or "").strip().lower() + current_model = (getattr(agent, "model", "") or "").strip() + current_base_url = str(getattr(agent, "base_url", "") or "").rstrip("/").lower() + fb_base_url_for_dedup = (fb.get("base_url") or "").strip().rstrip("/").lower() + if fb_provider == current_provider and fb_model == current_model: + logging.warning( + "Fallback skip: chain entry %s/%s matches current provider/model", + fb_provider, fb_model, + ) + return agent._try_activate_fallback() + if ( + fb_base_url_for_dedup + and current_base_url + and fb_base_url_for_dedup == current_base_url + and fb_model == current_model + ): + logging.warning( + "Fallback skip: chain entry base_url %s matches current backend", + fb_base_url_for_dedup, + ) + return agent._try_activate_fallback() + + # Use centralized router for client construction. + # raw_codex=True because the main agent needs direct responses.stream() + # access for Codex providers. + try: + from agent.auxiliary_client import resolve_provider_client + # Pass base_url and api_key from fallback config so custom + # endpoints (e.g. Ollama Cloud) resolve correctly instead of + # falling through to OpenRouter defaults. + fb_base_url_hint = (fb.get("base_url") or "").strip() or None + fb_api_key_hint = (fb.get("api_key") or "").strip() or None + if not fb_api_key_hint: + # key_env and api_key_env are both documented aliases (see + # _normalize_custom_provider_entry in hermes_cli/config.py). + fb_key_env = (fb.get("key_env") or fb.get("api_key_env") or "").strip() + if fb_key_env: + fb_api_key_hint = os.getenv(fb_key_env, "").strip() or None + # For Ollama Cloud endpoints, pull OLLAMA_API_KEY from env + # when no explicit key is in the fallback config. Host match + # (not substring) — see GHSA-76xc-57q6-vm5m. + if fb_base_url_hint and base_url_host_matches(fb_base_url_hint, "ollama.com") and not fb_api_key_hint: + fb_api_key_hint = os.getenv("OLLAMA_API_KEY") or None + fb_client, _resolved_fb_model = resolve_provider_client( + fb_provider, model=fb_model, raw_codex=True, + explicit_base_url=fb_base_url_hint, + explicit_api_key=fb_api_key_hint) + if fb_client is None: + logging.warning( + "Fallback to %s failed: provider not configured", + fb_provider) + return agent._try_activate_fallback() # try next in chain + try: + from hermes_cli.model_normalize import normalize_model_for_provider + + fb_model = normalize_model_for_provider(fb_model, fb_provider) + except Exception: + pass + + # Determine api_mode from provider / base URL / model + fb_api_mode = "chat_completions" + fb_base_url = str(fb_client.base_url) + _fb_is_azure = agent._is_azure_openai_url(fb_base_url) + if fb_provider == "openai-codex": + fb_api_mode = "codex_responses" + elif fb_provider == "anthropic" or fb_base_url.rstrip("/").lower().endswith("/anthropic"): + fb_api_mode = "anthropic_messages" + elif _fb_is_azure: + # Azure OpenAI serves gpt-5.x on /chat/completions — does NOT + # support the Responses API. Stay on chat_completions. + fb_api_mode = "chat_completions" + elif agent._is_direct_openai_url(fb_base_url): + fb_api_mode = "codex_responses" + elif agent._provider_model_requires_responses_api( + fb_model, + provider=fb_provider, + ): + # GPT-5.x models usually need Responses API, but keep + # provider-specific exceptions like Copilot gpt-5-mini on + # chat completions. + fb_api_mode = "codex_responses" + elif fb_provider == "bedrock" or ( + base_url_hostname(fb_base_url).startswith("bedrock-runtime.") + and base_url_host_matches(fb_base_url, "amazonaws.com") + ): + fb_api_mode = "bedrock_converse" + + old_model = agent.model + + # Clear the per-config context_length override so the fallback + # model's actual context window is resolved instead of inheriting + # the stale value from the previous model. See #22387. + agent._config_context_length = None + agent.model = fb_model + agent.provider = fb_provider + agent.base_url = fb_base_url + agent.api_mode = fb_api_mode + if hasattr(agent, "_transport_cache"): + agent._transport_cache.clear() + agent._fallback_activated = True + + # Honor per-provider / per-model request_timeout_seconds for the + # fallback target (same knob the primary client uses). None = use + # SDK default. + _fb_timeout = get_provider_request_timeout(fb_provider, fb_model) + + if fb_api_mode == "anthropic_messages": + # Build native Anthropic client instead of using OpenAI client + from agent.anthropic_adapter import build_anthropic_client, resolve_anthropic_token, _is_oauth_token + effective_key = (fb_client.api_key or resolve_anthropic_token() or "") if fb_provider == "anthropic" else (fb_client.api_key or "") + agent.api_key = effective_key + agent._anthropic_api_key = effective_key + agent._anthropic_base_url = fb_base_url + agent._anthropic_client = build_anthropic_client( + effective_key, agent._anthropic_base_url, timeout=_fb_timeout, + ) + agent._is_anthropic_oauth = _is_oauth_token(effective_key) if fb_provider == "anthropic" else False + agent.client = None + agent._client_kwargs = {} + else: + # Swap OpenAI client and config in-place + agent.api_key = fb_client.api_key + agent.client = fb_client + # Preserve provider-specific headers that + # resolve_provider_client() may have baked into + # fb_client via the default_headers kwarg. The OpenAI + # SDK stores these in _custom_headers. Without this, + # subsequent request-client rebuilds (via + # _create_request_openai_client) drop the headers, + # causing 403s from providers like Kimi Coding that + # require a User-Agent sentinel. + fb_headers = getattr(fb_client, "_custom_headers", None) + if not fb_headers: + fb_headers = getattr(fb_client, "default_headers", None) + agent._client_kwargs = { + "api_key": fb_client.api_key, + "base_url": fb_base_url, + **({"default_headers": dict(fb_headers)} if fb_headers else {}), + } + if _fb_timeout is not None: + agent._client_kwargs["timeout"] = _fb_timeout + # Rebuild the shared OpenAI client so the configured + # timeout takes effect on the very next fallback request, + # not only after a later credential-rotation rebuild. + agent._replace_primary_openai_client(reason="fallback_timeout_apply") + + # Re-evaluate prompt caching for the new provider/model + agent._use_prompt_caching, agent._use_native_cache_layout = ( + agent._anthropic_prompt_cache_policy( + provider=fb_provider, + base_url=fb_base_url, + api_mode=fb_api_mode, + model=fb_model, + ) + ) + + # LM Studio: preload before probing the fallback's context length. + agent._ensure_lmstudio_runtime_loaded() + + # Update context compressor limits for the fallback model. + # Without this, compression decisions use the primary model's + # context window (e.g. 200K) instead of the fallback's (e.g. 32K), + # causing oversized sessions to overflow the fallback. + # Also pass _config_context_length so the explicit config override + # (model.context_length in config.yaml) is respected — without this, + # the fallback activation drops to 128K even when config says 204800. + if hasattr(agent, 'context_compressor') and agent.context_compressor: + from agent.model_metadata import get_model_context_length + fb_context_length = get_model_context_length( + agent.model, base_url=agent.base_url, + api_key=agent.api_key, provider=agent.provider, + config_context_length=getattr(agent, "_config_context_length", None), + custom_providers=getattr(agent, "_custom_providers", None), + ) + agent.context_compressor.update_model( + model=agent.model, + context_length=fb_context_length, + base_url=agent.base_url, + api_key=getattr(agent, "api_key", ""), + provider=agent.provider, + ) + + agent._emit_status( + f"🔄 Primary model failed — switching to fallback: " + f"{fb_model} via {fb_provider}" + ) + logging.info( + "Fallback activated: %s → %s (%s)", + old_model, fb_model, fb_provider, + ) + return True + except Exception as e: + logging.error("Failed to activate fallback %s: %s", fb_model, e) + return agent._try_activate_fallback() # try next in chain + + + +def handle_max_iterations(agent, messages: list, api_call_count: int) -> str: + """Request a summary when max iterations are reached. Returns the final response text.""" + print(f"⚠️ Reached maximum iterations ({agent.max_iterations}). Requesting summary...") + + summary_request = ( + "You've reached the maximum number of tool-calling iterations allowed. " + "Please provide a final response summarizing what you've found and accomplished so far, " + "without calling any more tools." + ) + messages.append({"role": "user", "content": summary_request}) + + try: + # Build API messages, stripping internal-only fields + # (finish_reason, reasoning) that strict APIs like Mistral reject with 422 + _needs_sanitize = agent._should_sanitize_tool_calls() + api_messages = [] + for msg in messages: + api_msg = msg.copy() + agent._copy_reasoning_content_for_api(msg, api_msg) + for internal_field in ("reasoning", "finish_reason", "_thinking_prefill"): + api_msg.pop(internal_field, None) + if _needs_sanitize: + agent._sanitize_tool_calls_for_strict_api(api_msg) + api_messages.append(api_msg) + + effective_system = agent._cached_system_prompt or "" + if agent.ephemeral_system_prompt: + effective_system = (effective_system + "\n\n" + agent.ephemeral_system_prompt).strip() + if effective_system: + api_messages = [{"role": "system", "content": effective_system}] + api_messages + if agent.prefill_messages: + sys_offset = 1 if effective_system else 0 + for idx, pfm in enumerate(agent.prefill_messages): + api_messages.insert(sys_offset + idx, pfm.copy()) + + # Same safety net as the main loop: repair tool-call/result + # pairing before asking for a final summary. Compression and + # session resume can leave a tool result whose parent assistant + # tool_call was summarized away; Responses API rejects that as + # "No tool call found for function call output". + api_messages = agent._sanitize_api_messages(api_messages) + + # Same safety net as the main loop: drop thinking-only assistant + # turns so Anthropic-family providers don't 400 the summary call. + api_messages = agent._drop_thinking_only_and_merge_users(api_messages) + + summary_extra_body = {} + try: + from agent.auxiliary_client import _fixed_temperature_for_model, OMIT_TEMPERATURE as _OMIT_TEMP + except Exception: + _fixed_temperature_for_model = None + _OMIT_TEMP = None + _raw_summary_temp = ( + _fixed_temperature_for_model(agent.model, agent.base_url) + if _fixed_temperature_for_model is not None + else None + ) + _omit_summary_temperature = _raw_summary_temp is _OMIT_TEMP + _summary_temperature = None if _omit_summary_temperature else _raw_summary_temp + _is_nous = "nousresearch" in agent._base_url_lower + # LM Studio uses top-level `reasoning_effort` (not extra_body.reasoning). + # Mirror ChatCompletionsTransport.build_kwargs() so the summary path + # — which calls chat.completions.create() directly without going + # through the transport — sends the same shape the transport does. + _is_lmstudio_summary = ( + (agent.provider or "").strip().lower() == "lmstudio" + and agent._supports_reasoning_extra_body() + ) + _lm_reasoning_effort: str | None = ( + agent._resolve_lmstudio_summary_reasoning_effort() + if _is_lmstudio_summary else None + ) + if not _is_lmstudio_summary and agent._supports_reasoning_extra_body(): + if agent.reasoning_config is not None: + summary_extra_body["reasoning"] = agent.reasoning_config + else: + summary_extra_body["reasoning"] = { + "enabled": True, + "effort": "medium" + } + if _is_nous: + from agent.portal_tags import nous_portal_tags as _portal_tags + summary_extra_body["tags"] = _portal_tags() + + if agent.api_mode == "codex_responses": + codex_kwargs = agent._build_api_kwargs(api_messages) + codex_kwargs.pop("tools", None) + summary_response = agent._run_codex_stream(codex_kwargs) + _ct_sum = agent._get_transport() + _cnr_sum = _ct_sum.normalize_response(summary_response) + final_response = (_cnr_sum.content or "").strip() + else: + summary_kwargs = { + "model": agent.model, + "messages": api_messages, + } + if _summary_temperature is not None: + summary_kwargs["temperature"] = _summary_temperature + if agent.max_tokens is not None: + summary_kwargs.update(agent._max_tokens_param(agent.max_tokens)) + if _lm_reasoning_effort is not None: + summary_kwargs["reasoning_effort"] = _lm_reasoning_effort + + # Include provider routing preferences + provider_preferences = {} + if agent.providers_allowed: + provider_preferences["only"] = agent.providers_allowed + if agent.providers_ignored: + provider_preferences["ignore"] = agent.providers_ignored + if agent.providers_order: + provider_preferences["order"] = agent.providers_order + if agent.provider_sort: + provider_preferences["sort"] = agent.provider_sort + if provider_preferences and ( + (agent.provider or "").strip().lower() == "openrouter" + or agent._is_openrouter_url() + ): + summary_extra_body["provider"] = provider_preferences + + # Pareto Code router plugin — model-gated. Same shape as + # the main-loop emission so summary calls on + # openrouter/pareto-code respect the user's coding-score floor. + if ( + agent.model == "openrouter/pareto-code" + and ( + (agent.provider or "").strip().lower() == "openrouter" + or agent._is_openrouter_url() + ) + and agent.openrouter_min_coding_score is not None + and agent.openrouter_min_coding_score != "" + ): + try: + _ps = float(agent.openrouter_min_coding_score) + except (TypeError, ValueError): + _ps = None + if _ps is not None and 0.0 <= _ps <= 1.0: + summary_extra_body["plugins"] = [ + {"id": "pareto-router", "min_coding_score": _ps} + ] + + if summary_extra_body: + summary_kwargs["extra_body"] = summary_extra_body + + if agent.api_mode == "anthropic_messages": + _tsum = agent._get_transport() + _ant_kw = _tsum.build_kwargs(model=agent.model, messages=api_messages, tools=None, + max_tokens=agent.max_tokens, reasoning_config=agent.reasoning_config, + is_oauth=agent._is_anthropic_oauth, + preserve_dots=agent._anthropic_preserve_dots()) + summary_response = agent._anthropic_messages_create(_ant_kw) + _summary_result = _tsum.normalize_response(summary_response, strip_tool_prefix=agent._is_anthropic_oauth) + final_response = (_summary_result.content or "").strip() + else: + summary_response = agent._ensure_primary_openai_client(reason="iteration_limit_summary").chat.completions.create(**summary_kwargs) + _summary_result = agent._get_transport().normalize_response(summary_response) + final_response = (_summary_result.content or "").strip() + + if final_response: + if "" in final_response: + final_response = re.sub(r'.*?\s*', '', final_response, flags=re.DOTALL).strip() + if final_response: + messages.append({"role": "assistant", "content": final_response}) + else: + final_response = "I reached the iteration limit and couldn't generate a summary." + else: + # Retry summary generation + if agent.api_mode == "codex_responses": + codex_kwargs = agent._build_api_kwargs(api_messages) + codex_kwargs.pop("tools", None) + retry_response = agent._run_codex_stream(codex_kwargs) + _ct_retry = agent._get_transport() + _cnr_retry = _ct_retry.normalize_response(retry_response) + final_response = (_cnr_retry.content or "").strip() + elif agent.api_mode == "anthropic_messages": + _tretry = agent._get_transport() + _ant_kw2 = _tretry.build_kwargs(model=agent.model, messages=api_messages, tools=None, + is_oauth=agent._is_anthropic_oauth, + max_tokens=agent.max_tokens, reasoning_config=agent.reasoning_config, + preserve_dots=agent._anthropic_preserve_dots()) + retry_response = agent._anthropic_messages_create(_ant_kw2) + _retry_result = _tretry.normalize_response(retry_response, strip_tool_prefix=agent._is_anthropic_oauth) + final_response = (_retry_result.content or "").strip() + else: + summary_kwargs = { + "model": agent.model, + "messages": api_messages, + } + if _summary_temperature is not None: + summary_kwargs["temperature"] = _summary_temperature + if agent.max_tokens is not None: + summary_kwargs.update(agent._max_tokens_param(agent.max_tokens)) + if _lm_reasoning_effort is not None: + summary_kwargs["reasoning_effort"] = _lm_reasoning_effort + if summary_extra_body: + summary_kwargs["extra_body"] = summary_extra_body + + summary_response = agent._ensure_primary_openai_client(reason="iteration_limit_summary_retry").chat.completions.create(**summary_kwargs) + _retry_result = agent._get_transport().normalize_response(summary_response) + final_response = (_retry_result.content or "").strip() + + if final_response: + if "" in final_response: + final_response = re.sub(r'.*?\s*', '', final_response, flags=re.DOTALL).strip() + if final_response: + messages.append({"role": "assistant", "content": final_response}) + else: + final_response = "I reached the iteration limit and couldn't generate a summary." + else: + final_response = "I reached the iteration limit and couldn't generate a summary." + + except Exception as e: + logging.warning(f"Failed to get summary response: {e}") + final_response = f"I reached the maximum iterations ({agent.max_iterations}) but couldn't summarize. Error: {str(e)}" + + return final_response + + + +def cleanup_task_resources(agent, task_id: str) -> None: + """Clean up VM and browser resources for a given task. + + Skips ``cleanup_vm`` when the active terminal environment is marked + persistent (``persistent_filesystem=True``) so that long-lived sandbox + containers survive between turns. The idle reaper in + ``terminal_tool._cleanup_inactive_envs`` still tears them down once + ``terminal.lifetime_seconds`` is exceeded. Non-persistent backends are + torn down per-turn as before to prevent resource leakage (the original + intent of this hook for the Morph backend, see commit fbd3a2fd). + """ + try: + if is_persistent_env(task_id): + if agent.verbose_logging: + logging.debug( + f"Skipping per-turn cleanup_vm for persistent env {task_id}; " + f"idle reaper will handle it." + ) + else: + _ra().cleanup_vm(task_id) + except Exception as e: + if agent.verbose_logging: + logging.warning(f"Failed to cleanup VM for task {task_id}: {e}") + try: + _ra().cleanup_browser(task_id) + except Exception as e: + if agent.verbose_logging: + logging.warning(f"Failed to cleanup browser for task {task_id}: {e}") + + + + +def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=None): + """Streaming variant of _interruptible_api_call for real-time token delivery. + + Handles all three api_modes: + - chat_completions: stream=True on OpenAI-compatible endpoints + - anthropic_messages: client.messages.stream() via Anthropic SDK + - codex_responses: delegates to _run_codex_stream (already streaming) + + Fires stream_delta_callback and _stream_callback for each text token. + Tool-call turns suppress the callback — only text-only final responses + stream to the consumer. Returns a SimpleNamespace that mimics the + non-streaming response shape so the rest of the agent loop is unchanged. + + Falls back to _interruptible_api_call on provider errors indicating + streaming is not supported. + """ + if agent._interrupt_requested: + raise InterruptedError("Agent interrupted before streaming API call") + + if agent.api_mode == "codex_responses": + # Codex streams internally via _run_codex_stream. The main dispatch + # in _interruptible_api_call already calls it; we just need to + # ensure on_first_delta reaches it. Store it on the instance + # temporarily so _run_codex_stream can pick it up. + agent._codex_on_first_delta = on_first_delta + try: + return agent._interruptible_api_call(api_kwargs) + finally: + agent._codex_on_first_delta = None + + # Bedrock Converse uses boto3's converse_stream() with real-time delta + # callbacks — same UX as Anthropic and chat_completions streaming. + if agent.api_mode == "bedrock_converse": + result = {"response": None, "error": None} + first_delta_fired = {"done": False} + deltas_were_sent = {"yes": False} + + def _fire_first(): + if not first_delta_fired["done"] and on_first_delta: + first_delta_fired["done"] = True + try: + on_first_delta() + except Exception: + pass + + def _bedrock_call(): + try: + from agent.bedrock_adapter import ( + _get_bedrock_runtime_client, + invalidate_runtime_client, + is_stale_connection_error, + stream_converse_with_callbacks, + ) + region = api_kwargs.pop("__bedrock_region__", "us-east-1") + api_kwargs.pop("__bedrock_converse__", None) + client = _get_bedrock_runtime_client(region) + try: + raw_response = client.converse_stream(**api_kwargs) + except Exception as _bedrock_exc: + # Evict the cached client on stale-connection failures + # so the outer retry loop builds a fresh client/pool. + if is_stale_connection_error(_bedrock_exc): + invalidate_runtime_client(region) + raise + + def _on_text(text): + _fire_first() + agent._fire_stream_delta(text) + deltas_were_sent["yes"] = True + + def _on_tool(name): + _fire_first() + agent._fire_tool_gen_started(name) + + def _on_reasoning(text): + _fire_first() + agent._fire_reasoning_delta(text) + + result["response"] = stream_converse_with_callbacks( + raw_response, + on_text_delta=_on_text if agent._has_stream_consumers() else None, + on_tool_start=_on_tool, + on_reasoning_delta=_on_reasoning if agent.reasoning_callback or agent.stream_delta_callback else None, + on_interrupt_check=lambda: agent._interrupt_requested, + ) + except Exception as e: + result["error"] = e + + t = threading.Thread(target=_bedrock_call, daemon=True) + t.start() + while t.is_alive(): + t.join(timeout=0.3) + if agent._interrupt_requested: + raise InterruptedError("Agent interrupted during Bedrock API call") + if result["error"] is not None: + raise result["error"] + return result["response"] + + result = {"response": None, "error": None, "partial_tool_names": []} + request_client_holder = {"client": None, "diag": None} + first_delta_fired = {"done": False} + deltas_were_sent = {"yes": False} # Track if any deltas were fired (for fallback) + # Wall-clock timestamp of the last real streaming chunk. The outer + # poll loop uses this to detect stale connections that keep receiving + # SSE keep-alive pings but no actual data. + last_chunk_time = {"t": time.time()} + + def _fire_first_delta(): + if not first_delta_fired["done"] and on_first_delta: + first_delta_fired["done"] = True + try: + on_first_delta() + except Exception: + pass + + def _call_chat_completions(): + """Stream a chat completions response.""" + import httpx as _httpx + # Per-provider / per-model request_timeout_seconds (from config.yaml) + # wins over the HERMES_API_TIMEOUT env default if the user set it. + _provider_timeout_cfg = get_provider_request_timeout(agent.provider, agent.model) + _base_timeout = ( + _provider_timeout_cfg + if _provider_timeout_cfg is not None + else float(os.getenv("HERMES_API_TIMEOUT", 1800.0)) + ) + # Read timeout: config wins here too. Otherwise use + # HERMES_STREAM_READ_TIMEOUT (default 120s) for cloud providers. + if _provider_timeout_cfg is not None: + _stream_read_timeout = _provider_timeout_cfg + else: + _stream_read_timeout = float(os.getenv("HERMES_STREAM_READ_TIMEOUT", 120.0)) + # Local providers (Ollama, llama.cpp, vLLM) can take minutes for + # prefill on large contexts before producing the first token. + # Auto-increase the httpx read timeout unless the user explicitly + # overrode HERMES_STREAM_READ_TIMEOUT. + if _stream_read_timeout == 120.0 and agent.base_url and is_local_endpoint(agent.base_url): + _stream_read_timeout = _base_timeout + logger.debug( + "Local provider detected (%s) — stream read timeout raised to %.0fs", + agent.base_url, _stream_read_timeout, + ) + # Cap connect/pool at 60s even when provider timeout is higher. + # connect/pool cover TCP handshake, not model inference. + _conn_cap = min(_base_timeout, 60.0) if _provider_timeout_cfg is not None else 30.0 + stream_kwargs = { + **api_kwargs, + "stream": True, + "stream_options": {"include_usage": True}, + "timeout": _httpx.Timeout( + connect=_conn_cap, + read=_stream_read_timeout, + write=_base_timeout, + pool=_conn_cap, + ), + } + request_client_holder["client"] = agent._create_request_openai_client( + reason="chat_completion_stream_request", + api_kwargs=stream_kwargs, + ) + # Reset stale-stream timer so the detector measures from this + # attempt's start, not a previous attempt's last chunk. + last_chunk_time["t"] = time.time() + agent._touch_activity("waiting for provider response (streaming)") + # Initialize per-attempt stream diagnostics so the retry block can + # reach for them after the stream dies. Lives on + # ``request_client_holder["diag"]`` for closure access. + _diag = agent._stream_diag_init() + request_client_holder["diag"] = _diag + stream = request_client_holder["client"].chat.completions.create(**stream_kwargs) + + # Capture rate limit headers from the initial HTTP response. + # The OpenAI SDK Stream object exposes the underlying httpx + # response via .response before any chunks are consumed. + agent._capture_rate_limits(getattr(stream, "response", None)) + # Snapshot diagnostic headers (cf-ray, x-openrouter-provider, etc.) + # so they survive even when the stream dies before any chunk + # arrives. Best-effort; never raises. + agent._stream_diag_capture_response(_diag, getattr(stream, "response", None)) + + # Log OpenRouter response cache status when present. + agent._check_openrouter_cache_status(getattr(stream, "response", None)) + + content_parts: list = [] + tool_calls_acc: dict = {} + tool_gen_notified: set = set() + # Ollama-compatible endpoints reuse index 0 for every tool call + # in a parallel batch, distinguishing them only by id. Track + # the last seen id per raw index so we can detect a new tool + # call starting at the same index and redirect it to a fresh slot. + _last_id_at_idx: dict = {} # raw_index -> last seen non-empty id + _active_slot_by_idx: dict = {} # raw_index -> current slot in tool_calls_acc + finish_reason = None + model_name = None + role = "assistant" + reasoning_parts: list = [] + usage_obj = None + for chunk in stream: + last_chunk_time["t"] = time.time() + agent._touch_activity("receiving stream response") + + # Update per-attempt diagnostic counters. Best-effort — + # failures are swallowed so the streaming hot path is never + # interrupted by diagnostic accounting. + try: + _diag["chunks"] = int(_diag.get("chunks", 0)) + 1 + if _diag.get("first_chunk_at") is None: + _diag["first_chunk_at"] = last_chunk_time["t"] + # Approximate byte size from the chunk's repr — exact wire + # bytes aren't exposed by the SDK, but len(repr(chunk)) is + # a stable proxy for "how much content arrived" that + # survives stub provider differences. + try: + _diag["bytes"] = int(_diag.get("bytes", 0)) + len(repr(chunk)) + except Exception: + pass + except Exception: + pass + + if agent._interrupt_requested: + break + + if not chunk.choices: + if hasattr(chunk, "model") and chunk.model: + model_name = chunk.model + # Usage comes in the final chunk with empty choices + if hasattr(chunk, "usage") and chunk.usage: + usage_obj = chunk.usage + continue + + delta = chunk.choices[0].delta + if hasattr(chunk, "model") and chunk.model: + model_name = chunk.model + + # Accumulate reasoning content + reasoning_text = getattr(delta, "reasoning_content", None) or getattr(delta, "reasoning", None) + if reasoning_text: + reasoning_parts.append(reasoning_text) + _fire_first_delta() + agent._fire_reasoning_delta(reasoning_text) + + # Accumulate text content — fire callback only when no tool calls + if delta and delta.content: + content_parts.append(delta.content) + if not tool_calls_acc: + _fire_first_delta() + agent._fire_stream_delta(delta.content) + deltas_were_sent["yes"] = True + # Tool calls suppress regular content streaming (avoids + # displaying chatty "I'll use the tool..." text alongside + # tool calls). But reasoning tags embedded in suppressed + # content should still reach the display — otherwise the + # reasoning box only appears as a post-response fallback, + # rendering it confusingly after the already-streamed + # response. Route suppressed content through the stream + # delta callback so its tag extraction can fire the + # reasoning display. Non-reasoning text is harmlessly + # suppressed by the CLI's _stream_delta when the stream + # box is already closed (tool boundary flush). + elif agent.stream_delta_callback: + try: + agent.stream_delta_callback(delta.content) + agent._record_streamed_assistant_text(delta.content) + except Exception: + pass + + # Accumulate tool call deltas — notify display on first name + if delta and delta.tool_calls: + for tc_delta in delta.tool_calls: + raw_idx = tc_delta.index if tc_delta.index is not None else 0 + delta_id = tc_delta.id or "" + + # Ollama fix: detect a new tool call reusing the same + # raw index (different id) and redirect to a fresh slot. + if raw_idx not in _active_slot_by_idx: + _active_slot_by_idx[raw_idx] = raw_idx + if ( + delta_id + and raw_idx in _last_id_at_idx + and delta_id != _last_id_at_idx[raw_idx] + ): + new_slot = max(tool_calls_acc, default=-1) + 1 + _active_slot_by_idx[raw_idx] = new_slot + if delta_id: + _last_id_at_idx[raw_idx] = delta_id + idx = _active_slot_by_idx[raw_idx] + + if idx not in tool_calls_acc: + tool_calls_acc[idx] = { + "id": tc_delta.id or "", + "type": "function", + "function": {"name": "", "arguments": ""}, + "extra_content": None, + } + entry = tool_calls_acc[idx] + if tc_delta.id: + entry["id"] = tc_delta.id + if tc_delta.function: + if tc_delta.function.name: + # Use assignment, not +=. Function names are + # atomic identifiers delivered complete in the + # first chunk (OpenAI spec). Some providers + # (MiniMax M2.7 via NVIDIA NIM) resend the full + # name in every chunk; concatenation would + # produce "read_fileread_file". Assignment + # (matching the OpenAI Node SDK / LiteLLM / + # Vercel AI patterns) is immune to this. + entry["function"]["name"] = tc_delta.function.name + if tc_delta.function.arguments: + entry["function"]["arguments"] += tc_delta.function.arguments + extra = getattr(tc_delta, "extra_content", None) + if extra is None and hasattr(tc_delta, "model_extra"): + extra = (tc_delta.model_extra or {}).get("extra_content") + if extra is not None: + if hasattr(extra, "model_dump"): + extra = extra.model_dump() + entry["extra_content"] = extra + # Fire once per tool when the full name is available + name = entry["function"]["name"] + if name and idx not in tool_gen_notified: + tool_gen_notified.add(idx) + _fire_first_delta() + agent._fire_tool_gen_started(name) + # Record the partial tool-call name so the outer + # stub-builder can surface a user-visible warning + # if streaming dies before this tool's arguments + # are fully delivered. Without this, a stall + # during tool-call JSON generation lets the stub + # at line ~6107 return `tool_calls=None`, silently + # discarding the attempted action. + result["partial_tool_names"].append(name) + + if chunk.choices[0].finish_reason: + finish_reason = chunk.choices[0].finish_reason + + # Usage in the final chunk + if hasattr(chunk, "usage") and chunk.usage: + usage_obj = chunk.usage + + # Build mock response matching non-streaming shape + full_content = "".join(content_parts) or None + mock_tool_calls = None + has_truncated_tool_args = False + if tool_calls_acc: + mock_tool_calls = [] + for idx in sorted(tool_calls_acc): + tc = tool_calls_acc[idx] + arguments = tc["function"]["arguments"] + tool_name = tc["function"]["name"] or "?" + if arguments and arguments.strip(): + try: + json.loads(arguments) + except json.JSONDecodeError: + # Attempt repair before flagging as truncated. + # Models like GLM-5.1 via Ollama produce trailing + # commas, unclosed brackets, Python None, etc. + # Without repair, these hit the truncation handler + # and kill the session. _repair_tool_call_arguments + # returns "{}" for unrepairable args, which is far + # better than a crashed session. + repaired = _repair_tool_call_arguments(arguments, tool_name) + if repaired != "{}": + # Successfully repaired — use the fixed args + arguments = repaired + else: + # Unrepairable — flag for truncation handling + has_truncated_tool_args = True + mock_tool_calls.append(SimpleNamespace( + id=tc["id"], + type=tc["type"], + extra_content=tc.get("extra_content"), + function=SimpleNamespace( + name=tc["function"]["name"], + arguments=arguments, + ), + )) + + effective_finish_reason = finish_reason or "stop" + if has_truncated_tool_args: + effective_finish_reason = "length" + + full_reasoning = "".join(reasoning_parts) or None + mock_message = SimpleNamespace( + role=role, + content=full_content, + tool_calls=mock_tool_calls, + reasoning_content=full_reasoning, + ) + mock_choice = SimpleNamespace( + index=0, + message=mock_message, + finish_reason=effective_finish_reason, + ) + return SimpleNamespace( + id="stream-" + str(uuid.uuid4()), + model=model_name, + choices=[mock_choice], + usage=usage_obj, + ) + + def _call_anthropic(): + """Stream an Anthropic Messages API response. + + Fires delta callbacks for real-time token delivery, but returns + the native Anthropic Message object from get_final_message() so + the rest of the agent loop (validation, tool extraction, etc.) + works unchanged. + """ + has_tool_use = False + + # Reset stale-stream timer for this attempt + last_chunk_time["t"] = time.time() + # Per-attempt diagnostic dict for the retry block to consume. + _diag = agent._stream_diag_init() + request_client_holder["diag"] = _diag + # Use the Anthropic SDK's streaming context manager + with agent._anthropic_client.messages.stream(**api_kwargs) as stream: + # The Anthropic SDK exposes the raw httpx response on + # ``stream.response``. Snapshot diagnostic headers + # immediately so they survive a stream that dies before the + # first event. + try: + agent._stream_diag_capture_response( + _diag, getattr(stream, "response", None) + ) + except Exception: + pass + for event in stream: + # Update stale-stream timer on every event so the + # outer poll loop knows data is flowing. Without + # this, the detector kills healthy long-running + # Opus streams after 180 s even when events are + # actively arriving (the chat_completions path + # already does this at the top of its chunk loop). + last_chunk_time["t"] = time.time() + agent._touch_activity("receiving stream response") + + # Update per-attempt diagnostic counters (best-effort). + try: + _diag["chunks"] = int(_diag.get("chunks", 0)) + 1 + if _diag.get("first_chunk_at") is None: + _diag["first_chunk_at"] = last_chunk_time["t"] + try: + _diag["bytes"] = int(_diag.get("bytes", 0)) + len(repr(event)) + except Exception: + pass + except Exception: + pass + + if agent._interrupt_requested: + break + + event_type = getattr(event, "type", None) + + if event_type == "content_block_start": + block = getattr(event, "content_block", None) + if block and getattr(block, "type", None) == "tool_use": + has_tool_use = True + tool_name = getattr(block, "name", None) + if tool_name: + _fire_first_delta() + agent._fire_tool_gen_started(tool_name) + + elif event_type == "content_block_delta": + delta = getattr(event, "delta", None) + if delta: + delta_type = getattr(delta, "type", None) + if delta_type == "text_delta": + text = getattr(delta, "text", "") + if text and not has_tool_use: + _fire_first_delta() + agent._fire_stream_delta(text) + deltas_were_sent["yes"] = True + elif delta_type == "thinking_delta": + thinking_text = getattr(delta, "thinking", "") + if thinking_text: + _fire_first_delta() + agent._fire_reasoning_delta(thinking_text) + + # Return the native Anthropic Message for downstream processing + return stream.get_final_message() + + def _call(): + import httpx as _httpx + + _max_stream_retries = int(os.getenv("HERMES_STREAM_RETRIES", 2)) + + try: + for _stream_attempt in range(_max_stream_retries + 1): + # Check for interrupt before each retry attempt. Without + # this, /stop closes the HTTP connection (outer poll loop), + # but the retry loop opens a FRESH connection — negating the + # interrupt entirely. On slow providers (ollama-cloud) each + # retry can block for the full stream-read timeout (120s+), + # causing multi-minute delays between /stop and response. + if agent._interrupt_requested: + raise InterruptedError("Agent interrupted before stream retry") + try: + if agent.api_mode == "anthropic_messages": + agent._try_refresh_anthropic_client_credentials() + result["response"] = _call_anthropic() + else: + result["response"] = _call_chat_completions() + return # success + except Exception as e: + _is_timeout = isinstance( + e, (_httpx.ReadTimeout, _httpx.ConnectTimeout, _httpx.PoolTimeout) + ) + _is_conn_err = isinstance( + e, (_httpx.ConnectError, _httpx.RemoteProtocolError, ConnectionError) + ) + _is_stream_parse_err = agent._is_provider_stream_parse_error(e) + + # If the stream died AFTER some tokens were delivered: + # normally we don't retry (the user already saw text, + # retrying would duplicate it). BUT: if a tool call + # was in-flight when the stream died, silently aborting + # discards the tool call entirely. In that case we + # prefer to retry — the user sees a brief + # "reconnecting" marker + duplicated preamble text, + # which is strictly better than a failed action with + # a "retry manually" message. Limit this to transient + # connection errors (Clawdbot-style narrow gate): no + # tool has executed yet within this API call, so + # silent retry is safe wrt side-effects. + if deltas_were_sent["yes"]: + _partial_tool_in_flight = bool( + result.get("partial_tool_names") + ) + _is_sse_conn_err_preview = False + if not _is_timeout and not _is_conn_err: + from openai import APIError as _APIError + if isinstance(e, _APIError) and not getattr(e, "status_code", None): + _err_lower_preview = str(e).lower() + _SSE_PREVIEW_PHRASES = ( + "connection lost", + "connection reset", + "connection closed", + "connection terminated", + "network error", + "network connection", + "terminated", + "peer closed", + "broken pipe", + "upstream connect error", + ) + _is_sse_conn_err_preview = any( + phrase in _err_lower_preview + for phrase in _SSE_PREVIEW_PHRASES + ) + _is_transient = ( + _is_timeout + or _is_conn_err + or _is_sse_conn_err_preview + or _is_stream_parse_err + ) + _can_silent_retry = ( + _partial_tool_in_flight + and _is_transient + and _stream_attempt < _max_stream_retries + ) + if not _can_silent_retry: + # Either no tool call was in-flight (so the + # turn was a pure text response — current + # stub-with-recovered-text behaviour is + # correct), or retries are exhausted, or the + # error isn't transient. Fall through to the + # stub path. + logger.warning( + "Streaming failed after partial delivery, not retrying: %s", e + ) + result["error"] = e + return + # Tool call was in-flight AND error is transient: + # retry silently. Clear per-attempt state so the + # next stream starts clean. Fire a "reconnecting" + # marker so the user sees why the preamble is + # about to be re-streamed. Structured WARNING is + # emitted by ``_emit_stream_drop`` below; no + # additional INFO line needed. + try: + agent._fire_stream_delta( + "\n\n⚠ Connection dropped mid tool-call; " + "reconnecting…\n\n" + ) + except Exception: + pass + # Reset the streamed-text buffer so the retry's + # fresh preamble doesn't get double-recorded in + # _current_streamed_assistant_text (which would + # pollute the interim-visible-text comparison). + try: + agent._reset_stream_delivery_tracking() + except Exception: + pass + # Reset in-memory accumulators so the next + # attempt's chunks don't concat onto the dead + # stream's partial JSON. + result["partial_tool_names"] = [] + deltas_were_sent["yes"] = False + first_delta_fired["done"] = False + agent._emit_stream_drop( + error=e, + attempt=_stream_attempt + 2, + max_attempts=_max_stream_retries + 1, + mid_tool_call=True, + diag=request_client_holder.get("diag"), + ) + stale = request_client_holder.get("client") + if stale is not None: + agent._close_request_openai_client( + stale, reason="stream_mid_tool_retry_cleanup" + ) + request_client_holder["client"] = None + try: + agent._replace_primary_openai_client( + reason="stream_mid_tool_retry_pool_cleanup" + ) + except Exception: + pass + continue + + # SSE error events from proxies (e.g. OpenRouter sends + # {"error":{"message":"Network connection lost."}}) are + # raised as APIError by the OpenAI SDK. These are + # semantically identical to httpx connection drops — + # the upstream stream died — and should be retried with + # a fresh connection. Distinguish from HTTP errors: + # APIError from SSE has no status_code, while + # APIStatusError (4xx/5xx) always has one. + _is_sse_conn_err = False + if not _is_timeout and not _is_conn_err: + from openai import APIError as _APIError + if isinstance(e, _APIError) and not getattr(e, "status_code", None): + _err_lower_sse = str(e).lower() + _SSE_CONN_PHRASES = ( + "connection lost", + "connection reset", + "connection closed", + "connection terminated", + "network error", + "network connection", + "terminated", + "peer closed", + "broken pipe", + "upstream connect error", + ) + _is_sse_conn_err = any( + phrase in _err_lower_sse + for phrase in _SSE_CONN_PHRASES + ) + + if _is_timeout or _is_conn_err or _is_sse_conn_err or _is_stream_parse_err: + # Transient network / timeout error. Retry the + # streaming request with a fresh connection first. + if _stream_attempt < _max_stream_retries: + agent._emit_stream_drop( + error=e, + attempt=_stream_attempt + 2, + max_attempts=_max_stream_retries + 1, + mid_tool_call=False, + diag=request_client_holder.get("diag"), + ) + # Close the stale request client before retry + stale = request_client_holder.get("client") + if stale is not None: + agent._close_request_openai_client( + stale, reason="stream_retry_cleanup" + ) + request_client_holder["client"] = None + # Also rebuild the primary client to purge + # any dead connections from the pool. + try: + agent._replace_primary_openai_client( + reason="stream_retry_pool_cleanup" + ) + except Exception: + pass + continue + # Retries exhausted. Log the final failure with + # full diagnostic detail (chain, headers, + # bytes/elapsed) via the same helper used for + # mid-flight retries — subagent lines get the + # ``[subagent-N]`` log_prefix so the parent can + # attribute them. + agent._log_stream_retry( + kind="exhausted", + error=e, + attempt=_max_stream_retries + 1, + max_attempts=_max_stream_retries + 1, + mid_tool_call=False, + diag=request_client_holder.get("diag"), + ) + agent._emit_status( + "❌ Provider returned malformed streaming data after " + f"{_max_stream_retries + 1} attempts. " + "The provider may be experiencing issues — " + "try again in a moment." + if _is_stream_parse_err else + "❌ Connection to provider failed after " + f"{_max_stream_retries + 1} attempts. " + "The provider may be experiencing issues — " + "try again in a moment." + ) + else: + _err_lower = str(e).lower() + _is_stream_unsupported = ( + "stream" in _err_lower + and "not supported" in _err_lower + ) + if _is_stream_unsupported: + agent._disable_streaming = True + agent._safe_print( + "\n⚠ Streaming is not supported for this " + "model/provider. Switching to non-streaming.\n" + " To avoid this delay, set display.streaming: false " + "in config.yaml\n" + ) + logger.info( + "Streaming failed before delivery: %s", + e, + ) + + # Propagate the error to the main retry loop instead of + # falling back to non-streaming inline. The main loop has + # richer recovery: credential rotation, provider fallback, + # backoff, and — for "stream not supported" — will switch + # to non-streaming on the next attempt via _disable_streaming. + result["error"] = e + return + except InterruptedError as e: + # The interrupt may be noticed inside the worker thread before + # the polling loop sees it. Surface it through the normal result + # channel so callers never miss a fast pre-retry interrupt. + result["error"] = e + return + finally: + request_client = request_client_holder.get("client") + if request_client is not None: + agent._close_request_openai_client(request_client, reason="stream_request_complete") + + # Provider-configured stale timeout takes priority over env default. + _cfg_stale = get_provider_stale_timeout(agent.provider, agent.model) + if _cfg_stale is not None: + _stream_stale_timeout_base = _cfg_stale + else: + _stream_stale_timeout_base = float(os.getenv("HERMES_STREAM_STALE_TIMEOUT", 180.0)) + # Local providers (Ollama, oMLX, llama-cpp) can take 300+ seconds + # for prefill on large contexts. Disable the stale detector unless + # the user explicitly set HERMES_STREAM_STALE_TIMEOUT. + if _stream_stale_timeout_base == 180.0 and agent.base_url and is_local_endpoint(agent.base_url): + _stream_stale_timeout = float("inf") + logger.debug("Local provider detected (%s) — stale stream timeout disabled", agent.base_url) + else: + # Scale the stale timeout for large contexts: slow models (like Opus) + # can legitimately think for minutes before producing the first token + # when the context is large. Without this, the stale detector kills + # healthy connections during the model's thinking phase, producing + # spurious RemoteProtocolError ("peer closed connection"). + _est_tokens = sum(len(str(v)) for v in api_kwargs.get("messages", [])) // 4 + if _est_tokens > 100_000: + _stream_stale_timeout = max(_stream_stale_timeout_base, 300.0) + elif _est_tokens > 50_000: + _stream_stale_timeout = max(_stream_stale_timeout_base, 240.0) + else: + _stream_stale_timeout = _stream_stale_timeout_base + + t = threading.Thread(target=_call, daemon=True) + t.start() + _last_heartbeat = time.time() + _HEARTBEAT_INTERVAL = 30.0 # seconds between gateway activity touches + while t.is_alive(): + t.join(timeout=0.3) + + # Periodic heartbeat: touch the agent's activity tracker so the + # gateway's inactivity monitor knows we're alive while waiting + # for stream chunks. Without this, long thinking pauses (e.g. + # reasoning models) or slow prefill on local providers (Ollama) + # trigger false inactivity timeouts. The _call thread touches + # activity on each chunk, but the gap between API call start + # and first chunk can exceed the gateway timeout — especially + # when the stale-stream timeout is disabled (local providers). + _hb_now = time.time() + if _hb_now - _last_heartbeat >= _HEARTBEAT_INTERVAL: + _last_heartbeat = _hb_now + _waiting_secs = int(_hb_now - last_chunk_time["t"]) + agent._touch_activity( + f"waiting for stream response ({_waiting_secs}s, no chunks yet)" + ) + + # Detect stale streams: connections kept alive by SSE pings + # but delivering no real chunks. Kill the client so the + # inner retry loop can start a fresh connection. + _stale_elapsed = time.time() - last_chunk_time["t"] + if _stale_elapsed > _stream_stale_timeout: + _est_ctx = sum(len(str(v)) for v in api_kwargs.get("messages", [])) // 4 + logger.warning( + "Stream stale for %.0fs (threshold %.0fs) — no chunks received. " + "model=%s context=~%s tokens. Killing connection.", + _stale_elapsed, _stream_stale_timeout, + api_kwargs.get("model", "unknown"), f"{_est_ctx:,}", + ) + agent._emit_status( + f"⚠️ No response from provider for {int(_stale_elapsed)}s " + f"(model: {api_kwargs.get('model', 'unknown')}, " + f"context: ~{_est_ctx:,} tokens). " + f"Reconnecting..." + ) + try: + rc = request_client_holder.get("client") + if rc is not None: + agent._close_request_openai_client(rc, reason="stale_stream_kill") + except Exception: + pass + # Rebuild the primary client too — its connection pool + # may hold dead sockets from the same provider outage. + try: + agent._replace_primary_openai_client(reason="stale_stream_pool_cleanup") + except Exception: + pass + # Reset the timer so we don't kill repeatedly while + # the inner thread processes the closure. + last_chunk_time["t"] = time.time() + agent._touch_activity( + f"stale stream detected after {int(_stale_elapsed)}s, reconnecting" + ) + + if agent._interrupt_requested: + try: + if agent.api_mode == "anthropic_messages": + agent._anthropic_client.close() + agent._rebuild_anthropic_client() + else: + request_client = request_client_holder.get("client") + if request_client is not None: + agent._close_request_openai_client(request_client, reason="stream_interrupt_abort") + except Exception: + pass + raise InterruptedError("Agent interrupted during streaming API call") + if result["error"] is not None: + if deltas_were_sent["yes"]: + # Streaming failed AFTER some tokens were already delivered to + # the platform. Re-raising would let the outer retry loop make + # a new API call, creating a duplicate message. Return a + # partial "stop" response instead so the outer loop treats this + # turn as complete (no retry, no fallback). + # Recover whatever content was already streamed to the user. + # _current_streamed_assistant_text accumulates text fired + # through _fire_stream_delta, so it has exactly what the + # user saw before the connection died. + _partial_text = ( + getattr(agent, "_current_streamed_assistant_text", "") or "" + ).strip() or None + + # If the stream died while the model was emitting a tool call, + # the stub below will silently set `tool_calls=None` and the + # agent loop will treat the turn as complete — the attempted + # action is lost with no user-facing signal. Append a + # human-visible warning to the stub content so (a) the user + # knows something failed, and (b) the next turn's model sees + # in conversation history what was attempted and can retry. + _partial_names = list(result.get("partial_tool_names") or []) + if _partial_names: + _name_str = ", ".join(_partial_names[:3]) + if len(_partial_names) > 3: + _name_str += f", +{len(_partial_names) - 3} more" + _warn = ( + f"\n\n⚠ Stream stalled mid tool-call " + f"({_name_str}); the action was not executed. " + f"Ask me to retry if you want to continue." + ) + _partial_text = (_partial_text or "") + _warn + # Also fire as a streaming delta so the user sees it now + # instead of only in the persisted transcript. + try: + agent._fire_stream_delta(_warn) + except Exception: + pass + logger.warning( + "Partial stream dropped tool call(s) %s after %s chars " + "of text; surfaced warning to user: %s", + _partial_names, len(_partial_text or ""), result["error"], + ) + else: + logger.warning( + "Partial stream delivered before error; returning stub " + "response with %s chars of recovered content to prevent " + "duplicate messages: %s", + len(_partial_text or ""), + result["error"], + ) + _stub_msg = SimpleNamespace( + role="assistant", content=_partial_text, tool_calls=None, + reasoning_content=None, + ) + return SimpleNamespace( + id="partial-stream-stub", + model=getattr(agent, "model", "unknown"), + choices=[SimpleNamespace( + index=0, message=_stub_msg, finish_reason="stop", + )], + usage=None, + ) + raise result["error"] + return result["response"] + +# ── Provider fallback ────────────────────────────────────────────────── + + + +__all__ = [ + "interruptible_api_call", + "build_api_kwargs", + "build_assistant_message", + "try_activate_fallback", + "handle_max_iterations", + "cleanup_task_resources", + "interruptible_streaming_api_call", +] diff --git a/agent/codex_runtime.py b/agent/codex_runtime.py new file mode 100644 index 00000000000..02b788f5777 --- /dev/null +++ b/agent/codex_runtime.py @@ -0,0 +1,448 @@ +"""Codex API runtime — App Server and Responses-API streaming paths. + +Extracted from :class:`AIAgent` to keep the agent loop file focused. +Each function takes the parent ``AIAgent`` as its first argument +(``agent``). AIAgent keeps thin forwarder methods for backward +compatibility. + +* ``run_codex_app_server_turn`` — drives one turn through the + ``codex_app_server`` subprocess client (used when a Codex CLI install + is the active provider). +* ``run_codex_stream`` — streams a Codex Responses API call (the + ``codex_responses`` api_mode). +* ``run_codex_create_stream_fallback`` — recovery path when the + Responses ``stream=True`` initial create fails. +""" + +from __future__ import annotations + +import json +import logging +import os +from types import SimpleNamespace +from typing import Any, Dict, List + +logger = logging.getLogger(__name__) + + +def run_codex_app_server_turn( + agent, + *, + user_message: str, + original_user_message: Any, + messages: List[Dict[str, Any]], + effective_task_id: str, + should_review_memory: bool = False, +) -> Dict[str, Any]: + """Codex app-server runtime path. Hands the entire turn to a `codex + app-server` subprocess and projects its events back into Hermes' + messages list so memory/skill review keep working. + + Called from run_conversation() when agent.api_mode == "codex_app_server". + Returns the same dict shape as the chat_completions path. + """ + from agent.transports.codex_app_server_session import CodexAppServerSession + + # Lazy session: one CodexAppServerSession per AIAgent instance. + # Spawned on first turn, reused across turns, closed at AIAgent + # shutdown (see _cleanup hook). + if not hasattr(agent, "_codex_session") or agent._codex_session is None: + cwd = getattr(agent, "session_cwd", None) or os.getcwd() + # Approval callback: defer to Hermes' standard prompt flow if a + # CLI thread has installed one. Gateway / cron contexts get the + # codex-side fail-closed default. + try: + from tools.terminal_tool import _get_approval_callback + approval_callback = _get_approval_callback() + except Exception: + approval_callback = None + agent._codex_session = CodexAppServerSession( + cwd=cwd, + approval_callback=approval_callback, + ) + + # NOTE: the user message is ALREADY appended to messages by the + # standard run_conversation() flow (line ~11823) before the early + # return reaches us. Do NOT append again — that would duplicate. + + try: + turn = agent._codex_session.run_turn(user_input=user_message) + except Exception as exc: + logger.exception("codex app-server turn failed") + # Crash → unconditionally drop the session so the next turn + # respawns from scratch instead of reusing a dead client. + try: + agent._codex_session.close() + except Exception: + pass + agent._codex_session = None + return { + "final_response": ( + f"Codex app-server turn failed: {exc}. " + f"Fall back to default runtime with `/codex-runtime auto`." + ), + "messages": messages, + "api_calls": 0, + "completed": False, + "partial": True, + "error": str(exc), + } + + # If the turn signalled the underlying client is wedged (deadline + # blown, post-tool watchdog tripped, OAuth refresh died, subprocess + # exited), retire the session so the next turn respawns codex + # rather than riding the broken process. Mirrors openclaw beta.8's + # "retire timed-out app-server clients" fix. + if getattr(turn, "should_retire", False): + logger.warning( + "codex app-server session retired (turn error: %s)", + turn.error, + ) + try: + agent._codex_session.close() + except Exception: + pass + agent._codex_session = None + + # Splice projected messages into the conversation. The projector emits + # standard {role, content, tool_calls, tool_call_id} entries, which + # is exactly what curator.py / sessions DB expect. + if turn.projected_messages: + messages.extend(turn.projected_messages) + + # Counter ticks for the agent-improvement loop. + # _turns_since_memory and _user_turn_count are ALREADY incremented + # in the run_conversation() pre-loop block (lines ~11793-11817) so we + # do NOT touch them here — that would double-count. + # Only _iters_since_skill needs explicit increment, since the + # chat_completions loop bumps it per tool iteration (line ~12110) + # and that loop is bypassed on this path. + agent._iters_since_skill = ( + getattr(agent, "_iters_since_skill", 0) + turn.tool_iterations + ) + + # Now check the skill nudge AFTER iters were incremented — same + # pattern the chat_completions path uses (line ~15432). + should_review_skills = False + if ( + agent._skill_nudge_interval > 0 + and agent._iters_since_skill >= agent._skill_nudge_interval + and "skill_manage" in agent.valid_tool_names + ): + should_review_skills = True + agent._iters_since_skill = 0 + + # External memory provider sync (mirrors line ~15439). Skipped on + # interrupt/error to avoid feeding partial transcripts to memory. + if not turn.interrupted and turn.error is None: + try: + agent._sync_external_memory_for_turn( + original_user_message=original_user_message, + final_response=turn.final_text, + interrupted=False, + ) + except Exception: + logger.debug("external memory sync raised", exc_info=True) + + # Background review fork — same cadence + signature as the default + # path (line ~15449). Only fires when a trigger actually tripped AND + # we have a real final response. + if ( + turn.final_text + and not turn.interrupted + and (should_review_memory or should_review_skills) + ): + try: + agent._spawn_background_review( + messages_snapshot=list(messages), + review_memory=should_review_memory, + review_skills=should_review_skills, + ) + except Exception: + logger.debug("background review spawn raised", exc_info=True) + + return { + "final_response": turn.final_text, + "messages": messages, + "api_calls": 1, # one app-server "turn" maps to one logical API call + "completed": not turn.interrupted and turn.error is None, + "partial": turn.interrupted or turn.error is not None, + "error": turn.error, + "codex_thread_id": turn.thread_id, + "codex_turn_id": turn.turn_id, + } + + + + +def run_codex_stream(agent, api_kwargs: dict, client: Any = None, on_first_delta: callable = None): + """Execute one streaming Responses API request and return the final response.""" + import httpx as _httpx + + active_client = client or agent._ensure_primary_openai_client(reason="codex_stream_direct") + max_stream_retries = 1 + has_tool_calls = False + first_delta_fired = False + # Accumulate streamed text so we can recover if get_final_response() + # returns empty output (e.g. chatgpt.com backend-api sends + # response.incomplete instead of response.completed). + agent._codex_streamed_text_parts: list = [] + for attempt in range(max_stream_retries + 1): + if agent._interrupt_requested: + raise InterruptedError("Agent interrupted before Codex stream retry") + collected_output_items: list = [] + try: + with active_client.responses.stream(**api_kwargs) as stream: + for event in stream: + agent._touch_activity("receiving stream response") + if agent._interrupt_requested: + break + event_type = getattr(event, "type", "") + # Fire callbacks on text content deltas (suppress during tool calls) + if "output_text.delta" in event_type or event_type == "response.output_text.delta": + delta_text = getattr(event, "delta", "") + if delta_text: + agent._codex_streamed_text_parts.append(delta_text) + if delta_text and not has_tool_calls: + if not first_delta_fired: + first_delta_fired = True + if on_first_delta: + try: + on_first_delta() + except Exception: + pass + agent._fire_stream_delta(delta_text) + # Track tool calls to suppress text streaming + elif "function_call" in event_type: + has_tool_calls = True + # Fire reasoning callbacks + elif "reasoning" in event_type and "delta" in event_type: + reasoning_text = getattr(event, "delta", "") + if reasoning_text: + agent._fire_reasoning_delta(reasoning_text) + # Collect completed output items — some backends + # (chatgpt.com/backend-api/codex) stream valid items + # via response.output_item.done but the SDK's + # get_final_response() returns an empty output list. + elif event_type == "response.output_item.done": + done_item = getattr(event, "item", None) + if done_item is not None: + collected_output_items.append(done_item) + # Log non-completed terminal events for diagnostics + elif event_type in {"response.incomplete", "response.failed"}: + resp_obj = getattr(event, "response", None) + status = getattr(resp_obj, "status", None) if resp_obj else None + incomplete_details = getattr(resp_obj, "incomplete_details", None) if resp_obj else None + logger.warning( + "Codex Responses stream received terminal event %s " + "(status=%s, incomplete_details=%s, streamed_chars=%d). %s", + event_type, status, incomplete_details, + sum(len(p) for p in agent._codex_streamed_text_parts), + agent._client_log_context(), + ) + final_response = stream.get_final_response() + # PATCH: ChatGPT Codex backend streams valid output items + # but get_final_response() can return an empty output list. + # Backfill from collected items or synthesize from deltas. + _out = getattr(final_response, "output", None) + if isinstance(_out, list) and not _out: + if collected_output_items: + final_response.output = list(collected_output_items) + logger.debug( + "Codex stream: backfilled %d output items from stream events", + len(collected_output_items), + ) + elif agent._codex_streamed_text_parts and not has_tool_calls: + assembled = "".join(agent._codex_streamed_text_parts) + final_response.output = [SimpleNamespace( + type="message", + role="assistant", + status="completed", + content=[SimpleNamespace(type="output_text", text=assembled)], + )] + logger.debug( + "Codex stream: synthesized output from %d text deltas (%d chars)", + len(agent._codex_streamed_text_parts), len(assembled), + ) + return final_response + except (_httpx.RemoteProtocolError, _httpx.ReadTimeout, _httpx.ConnectError, ConnectionError) as exc: + if attempt < max_stream_retries: + logger.debug( + "Codex Responses stream transport failed (attempt %s/%s); retrying. %s error=%s", + attempt + 1, + max_stream_retries + 1, + agent._client_log_context(), + exc, + ) + continue + logger.debug( + "Codex Responses stream transport failed; falling back to create(stream=True). %s error=%s", + agent._client_log_context(), + exc, + ) + return agent._run_codex_create_stream_fallback(api_kwargs, client=active_client) + except RuntimeError as exc: + err_text = str(exc) + missing_completed = "response.completed" in err_text + # The OpenAI SDK's Responses streaming state machine raises + # ``RuntimeError("Expected to have received `response.created` + # before ``")`` when the first SSE event from the + # server is anything other than ``response.created`` — and it + # discards the event's payload before we can read it. Three + # real-world backends emit a different first frame: + # + # * xAI on grok-4.x OAuth — sends ``error`` (issues + # reported around the May 2026 SuperGrok rollout when + # multi-turn conversations replay encrypted reasoning + # content the OAuth tier rejects) + # * codex-lb relays — send ``codex.rate_limits`` (#14634) + # * custom Responses relays — send ``response.in_progress`` + # (#8133) + # + # In all three cases the underlying byte stream is still + # readable: a non-stream ``responses.create(stream=True)`` + # fallback succeeds and surfaces the real provider error as + # a normal exception with body+status_code attached, which + # ``_summarize_api_error`` can then translate into a useful + # user-facing line. Treat ``response.created`` prelude + # errors the same way we already treat ``response.completed`` + # postlude errors. + prelude_error = ( + "Expected to have received `response.created`" in err_text + or "Expected to have received \"response.created\"" in err_text + ) + if (missing_completed or prelude_error) and attempt < max_stream_retries: + logger.debug( + "Responses stream %s (attempt %s/%s); retrying. %s", + "prelude rejected" if prelude_error else "closed before completion", + attempt + 1, + max_stream_retries + 1, + agent._client_log_context(), + ) + continue + if missing_completed or prelude_error: + logger.debug( + "Responses stream %s; falling back to create(stream=True). %s err=%s", + "rejected before response.created" if prelude_error else "did not emit response.completed", + agent._client_log_context(), + err_text, + ) + return agent._run_codex_create_stream_fallback(api_kwargs, client=active_client) + raise + + + +def run_codex_create_stream_fallback(agent, api_kwargs: dict, client: Any = None): + """Fallback path for stream completion edge cases on Codex-style Responses backends.""" + active_client = client or agent._ensure_primary_openai_client(reason="codex_create_stream_fallback") + fallback_kwargs = dict(api_kwargs) + fallback_kwargs["stream"] = True + fallback_kwargs = agent._get_transport().preflight_kwargs(fallback_kwargs, allow_stream=True) + stream_or_response = active_client.responses.create(**fallback_kwargs) + + # Compatibility shim for mocks or providers that still return a concrete response. + if hasattr(stream_or_response, "output"): + return stream_or_response + if not hasattr(stream_or_response, "__iter__"): + return stream_or_response + + terminal_response = None + collected_output_items: list = [] + collected_text_deltas: list = [] + try: + for event in stream_or_response: + agent._touch_activity("receiving stream response") + event_type = getattr(event, "type", None) + if not event_type and isinstance(event, dict): + event_type = event.get("type") + + # ``error`` SSE frames carry the provider's real failure + # reason (subscription / quota / model-not-available / + # rejected-reasoning-replay) but never appear in the + # ``{completed, incomplete, failed}`` terminal set, so the + # raw loop below would silently consume them and end with + # "did not emit a terminal response". xAI in particular + # emits ``type=error`` as the FIRST frame for OAuth + # accounts whose Grok subscription is missing/exhausted — + # the SDK's stream helper raises ``RuntimeError(Expected + # to have received response.created before error)`` which + # the caller catches and routes here, expecting this + # fallback to surface the message. Synthesize an + # APIError-shaped exception so ``_summarize_api_error`` + # and the credential-pool entitlement detector see the + # real text instead of a generic RuntimeError. + if event_type == "error": + err_message = getattr(event, "message", None) + if not err_message and isinstance(event, dict): + err_message = event.get("message") + err_code = getattr(event, "code", None) + if not err_code and isinstance(event, dict): + err_code = event.get("code") + err_param = getattr(event, "param", None) + if not err_param and isinstance(event, dict): + err_param = event.get("param") + err_message = (err_message or "stream emitted error event").strip() + from run_agent import _StreamErrorEvent + raise _StreamErrorEvent(err_message, code=err_code, param=err_param) + + # Collect output items and text deltas for backfill + if event_type == "response.output_item.done": + done_item = getattr(event, "item", None) + if done_item is None and isinstance(event, dict): + done_item = event.get("item") + if done_item is not None: + collected_output_items.append(done_item) + elif event_type in {"response.output_text.delta",}: + delta = getattr(event, "delta", "") + if not delta and isinstance(event, dict): + delta = event.get("delta", "") + if delta: + collected_text_deltas.append(delta) + + if event_type not in {"response.completed", "response.incomplete", "response.failed"}: + continue + + terminal_response = getattr(event, "response", None) + if terminal_response is None and isinstance(event, dict): + terminal_response = event.get("response") + if terminal_response is not None: + # Backfill empty output from collected stream events + _out = getattr(terminal_response, "output", None) + if isinstance(_out, list) and not _out: + if collected_output_items: + terminal_response.output = list(collected_output_items) + logger.debug( + "Codex fallback stream: backfilled %d output items", + len(collected_output_items), + ) + elif collected_text_deltas: + assembled = "".join(collected_text_deltas) + terminal_response.output = [SimpleNamespace( + type="message", role="assistant", + status="completed", + content=[SimpleNamespace(type="output_text", text=assembled)], + )] + logger.debug( + "Codex fallback stream: synthesized from %d deltas (%d chars)", + len(collected_text_deltas), len(assembled), + ) + return terminal_response + finally: + close_fn = getattr(stream_or_response, "close", None) + if callable(close_fn): + try: + close_fn() + except Exception: + pass + + if terminal_response is not None: + return terminal_response + raise RuntimeError("Responses create(stream=True) fallback did not emit a terminal response.") + + + +__all__ = [ + "run_codex_app_server_turn", + "run_codex_stream", + "run_codex_create_stream_fallback", +] diff --git a/agent/conversation_compression.py b/agent/conversation_compression.py new file mode 100644 index 00000000000..bc70623997d --- /dev/null +++ b/agent/conversation_compression.py @@ -0,0 +1,556 @@ +"""Context compression — extract the AIAgent methods that drive summarisation. + +Three concerns live here: + +* :func:`check_compression_model_feasibility` — startup probe of the + configured auxiliary compression model. Warns when the aux context + window can't fit the main model's compression threshold; auto-lowers + the session threshold when possible; hard-rejects auxes below + ``MINIMUM_CONTEXT_LENGTH``. + +* :func:`replay_compression_warning` — re-emit a stored warning through + the gateway ``status_callback`` once it's wired up (the callback is + set after :class:`AIAgent` construction). + +* :func:`compress_context` — the actual compression call. Runs the + configured compressor, splits the SQLite session, rotates the + session_id, notifies plugin context engines / memory providers, and + returns the compressed message list and freshly-built system prompt. + +* :func:`try_shrink_image_parts_in_messages` — image-too-large recovery + helper that re-encodes ``data:image/...;base64,...`` parts at a smaller + size so retries can fit under provider ceilings (Anthropic's 5 MB). + +``run_agent`` keeps thin wrappers for each so existing call sites +(``self._compress_context(...)``) keep working. Tests that exercise +these paths see no behavioural change. +""" + +from __future__ import annotations + +import logging +import os +import tempfile +import uuid +from datetime import datetime +from pathlib import Path +from typing import Any, List, Optional, Tuple + +from agent.model_metadata import estimate_request_tokens_rough + +logger = logging.getLogger(__name__) + + +def check_compression_model_feasibility(agent: Any) -> None: + """Warn at session start if the auxiliary compression model's context + window is smaller than the main model's compression threshold. + + When the auxiliary model cannot fit the content that needs summarising, + compression will either fail outright (the LLM call errors) or produce + a severely truncated summary. + + Called during ``AIAgent.__init__`` so CLI users see the warning + immediately (via ``_vprint``). The gateway sets ``status_callback`` + *after* construction, so :func:`replay_compression_warning` re-sends + the stored warning through the callback on the first + ``run_conversation()`` call. + """ + if not agent.compression_enabled: + return + try: + from agent.auxiliary_client import ( + _resolve_task_provider_model, + get_text_auxiliary_client, + ) + from agent.model_metadata import ( + MINIMUM_CONTEXT_LENGTH, + get_model_context_length, + ) + + client, aux_model = get_text_auxiliary_client( + "compression", + main_runtime=agent._current_main_runtime(), + ) + # Best-effort aux provider label for the warning message. The + # configured provider may be "auto", in which case we fall back + # to the client's base_url hostname so the user can still tell + # where the compression model is actually being called. + try: + _aux_cfg_provider, _, _, _, _ = _resolve_task_provider_model("compression") + except Exception: + _aux_cfg_provider = "" + if client is None or not aux_model: + if _aux_cfg_provider and _aux_cfg_provider != "auto": + msg = ( + "⚠ Configured auxiliary compression provider " + f"'{_aux_cfg_provider}' is unavailable — context " + "compression will drop middle turns without a summary. " + "Check auxiliary.compression in config.yaml and " + "reauthenticate that provider." + ) + else: + msg = ( + "⚠ No auxiliary LLM provider configured — context " + "compression will drop middle turns without a summary. " + "Run `hermes setup` or set OPENROUTER_API_KEY." + ) + agent._compression_warning = msg + agent._emit_status(msg) + logger.warning( + "No auxiliary LLM provider for compression — " + "summaries will be unavailable." + ) + return + + aux_base_url = str(getattr(client, "base_url", "")) + aux_api_key = str(getattr(client, "api_key", "")) + + aux_context = get_model_context_length( + aux_model, + base_url=aux_base_url, + api_key=aux_api_key, + config_context_length=getattr(agent, "_aux_compression_context_length_config", None), + # Each model must be resolved with its own provider so that + # provider-specific paths (e.g. Bedrock static table, OpenRouter API) + # are invoked for the correct client, not inherited from the main model. + provider=(_aux_cfg_provider if _aux_cfg_provider and _aux_cfg_provider != "auto" else getattr(agent, "provider", "")), + custom_providers=agent._custom_providers, + ) + + # Hard floor: the auxiliary compression model must have at least + # MINIMUM_CONTEXT_LENGTH (64K) tokens of context. The main model + # is already required to meet this floor (checked earlier in + # __init__), so the compression model must too — otherwise it + # cannot summarise a full threshold-sized window of main-model + # content. Mirrors the main-model rejection pattern. + if aux_context and aux_context < MINIMUM_CONTEXT_LENGTH: + raise ValueError( + f"Auxiliary compression model {aux_model} has a context " + f"window of {aux_context:,} tokens, which is below the " + f"minimum {MINIMUM_CONTEXT_LENGTH:,} required by Hermes " + f"Agent. Choose a compression model with at least " + f"{MINIMUM_CONTEXT_LENGTH // 1000}K context (set " + f"auxiliary.compression.model in config.yaml), or set " + f"auxiliary.compression.context_length to override the " + f"detected value if it is wrong." + ) + + threshold = agent.context_compressor.threshold_tokens + if aux_context < threshold: + # Auto-correct: lower the live session threshold so + # compression actually works this session. The hard floor + # above guarantees aux_context >= MINIMUM_CONTEXT_LENGTH, + # so the new threshold is always >= 64K. + # + # The compression summariser sends a single user-role + # prompt (no system prompt, no tools) to the aux model, so + # new_threshold == aux_context is safe: the request is + # the raw messages plus a small summarisation instruction. + old_threshold = threshold + new_threshold = aux_context + agent.context_compressor.threshold_tokens = new_threshold + # Keep threshold_percent in sync so future main-model + # context_length changes (update_model) re-derive from a + # sensible number rather than the original too-high value. + main_ctx = agent.context_compressor.context_length + if main_ctx: + agent.context_compressor.threshold_percent = ( + new_threshold / main_ctx + ) + safe_pct = int((aux_context / main_ctx) * 100) if main_ctx else 50 + # Build human-readable "model (provider)" labels for both + # the main model and the compression model so users can + # tell at a glance which provider each side is actually + # using. When the configured provider is empty or "auto", + # fall back to the client's base_url hostname. + _main_model = getattr(agent, "model", "") or "?" + _main_provider = getattr(agent, "provider", "") or "" + _aux_provider_label = ( + _aux_cfg_provider + if _aux_cfg_provider and _aux_cfg_provider != "auto" + else "" + ) + if not _aux_provider_label: + try: + from urllib.parse import urlparse + _aux_provider_label = ( + urlparse(aux_base_url).hostname or aux_base_url + ) + except Exception: + _aux_provider_label = aux_base_url or "auto" + _main_label = ( + f"{_main_model} ({_main_provider})" + if _main_provider + else _main_model + ) + _aux_label = f"{aux_model} ({_aux_provider_label})" + msg = ( + f"⚠ Compression model {_aux_label} context is " + f"{aux_context:,} tokens, but the main model " + f"{_main_label}'s compression threshold was " + f"{old_threshold:,} tokens. " + f"Auto-lowered this session's threshold to " + f"{new_threshold:,} tokens so compression can run.\n" + f" To make this permanent, edit config.yaml — either:\n" + f" 1. Use a larger compression model:\n" + f" auxiliary:\n" + f" compression:\n" + f" model: \n" + f" 2. Lower the compression threshold:\n" + f" compression:\n" + f" threshold: 0.{safe_pct:02d}" + ) + agent._compression_warning = msg + agent._emit_status(msg) + logger.warning( + "Auxiliary compression model %s has %d token context, " + "below the main model's compression threshold of %d " + "tokens — auto-lowered session threshold to %d to " + "keep compression working.", + aux_model, + aux_context, + old_threshold, + new_threshold, + ) + except ValueError: + # Hard rejections (aux below minimum context) must propagate + # so the session refuses to start. + raise + except Exception as exc: + logger.debug( + "Compression feasibility check failed (non-fatal): %s", exc + ) + + +def replay_compression_warning(agent: Any) -> None: + """Re-send the compression warning through ``status_callback``. + + During ``__init__`` the gateway's ``status_callback`` is not yet + wired, so ``_emit_status`` only reaches ``_vprint`` (CLI). This + method is called once at the start of the first + ``run_conversation()`` — by then the gateway has set the callback, + so every platform (Telegram, Discord, Slack, etc.) receives the + warning. + """ + msg = getattr(agent, "_compression_warning", None) + if msg and agent.status_callback: + try: + agent.status_callback("lifecycle", msg) + except Exception: + pass + + +def compress_context( + agent: Any, + messages: list, + system_message: str, + *, + approx_tokens: Optional[int] = None, + task_id: str = "default", + focus_topic: Optional[str] = None, +) -> Tuple[list, str]: + """Compress conversation context and split the session in SQLite. + + Args: + agent: The owning :class:`AIAgent`. + messages: Current message history (will be summarised). + system_message: Current system prompt; rebuilt after compression. + approx_tokens: Pre-compression token estimate, logged for ops. + task_id: Tool task scope (used for clearing file-read dedup state). + focus_topic: Optional focus string for guided compression — the + summariser will prioritise preserving information related to + this topic. Inspired by Claude Code's ``/compact ``. + + Returns: + ``(compressed_messages, new_system_prompt)`` tuple. + """ + _pre_msg_count = len(messages) + logger.info( + "context compression started: session=%s messages=%d tokens=~%s model=%s focus=%r", + agent.session_id or "none", _pre_msg_count, + f"{approx_tokens:,}" if approx_tokens else "unknown", agent.model, + focus_topic, + ) + agent._emit_status( + "🗜️ Compacting context — summarizing earlier conversation so I can continue..." + ) + + # Notify external memory provider before compression discards context + if agent._memory_manager: + try: + agent._memory_manager.on_pre_compress(messages) + except Exception: + pass + + try: + compressed = agent.context_compressor.compress(messages, current_tokens=approx_tokens, focus_topic=focus_topic) + except TypeError: + # Plugin context engine with strict signature that doesn't accept + # focus_topic — fall back to calling without it. + compressed = agent.context_compressor.compress(messages, current_tokens=approx_tokens) + + summary_error = getattr(agent.context_compressor, "_last_summary_error", None) + if summary_error: + if getattr(agent, "_last_compression_summary_warning", None) != summary_error: + agent._last_compression_summary_warning = summary_error + agent._emit_warning( + f"⚠ Compression summary failed: {summary_error}. " + "Inserted a fallback context marker." + ) + else: + # No hard failure — but did the configured aux model error out + # and get recovered by retrying on main? Surface that so users + # know their auxiliary.compression.model setting is broken even + # though compression succeeded. + _aux_fail_model = getattr(agent.context_compressor, "_last_aux_model_failure_model", None) + _aux_fail_err = getattr(agent.context_compressor, "_last_aux_model_failure_error", None) + if _aux_fail_model: + # Dedup on (model, error) so we don't spam on every compaction + _aux_key = (_aux_fail_model, _aux_fail_err) + if getattr(agent, "_last_aux_fallback_warning_key", None) != _aux_key: + agent._last_aux_fallback_warning_key = _aux_key + agent._emit_warning( + f"ℹ Configured compression model '{_aux_fail_model}' failed " + f"({_aux_fail_err or 'unknown error'}). Recovered using main model — " + "check auxiliary.compression.model in config.yaml." + ) + + todo_snapshot = agent._todo_store.format_for_injection() + if todo_snapshot: + compressed.append({"role": "user", "content": todo_snapshot}) + + agent._invalidate_system_prompt() + new_system_prompt = agent._build_system_prompt(system_message) + agent._cached_system_prompt = new_system_prompt + + if agent._session_db: + try: + # Propagate title to the new session with auto-numbering + old_title = agent._session_db.get_session_title(agent.session_id) + # Trigger memory extraction on the old session before it rotates. + agent.commit_memory_session(messages) + agent._session_db.end_session(agent.session_id, "compression") + old_session_id = agent.session_id + agent.session_id = f"{datetime.now().strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:6]}" + os.environ["HERMES_SESSION_ID"] = agent.session_id + try: + from gateway.session_context import _SESSION_ID + _SESSION_ID.set(agent.session_id) + except Exception: + pass + # Update session_log_file to point to the new session's JSON file + agent.session_log_file = agent.logs_dir / f"session_{agent.session_id}.json" + agent._session_db_created = False + agent._session_db.create_session( + session_id=agent.session_id, + source=agent.platform or os.environ.get("HERMES_SESSION_SOURCE", "cli"), + model=agent.model, + model_config=agent._session_init_model_config, + parent_session_id=old_session_id, + ) + agent._session_db_created = True + # Auto-number the title for the continuation session + if old_title: + try: + new_title = agent._session_db.get_next_title_in_lineage(old_title) + agent._session_db.set_session_title(agent.session_id, new_title) + except (ValueError, Exception) as e: + logger.debug("Could not propagate title on compression: %s", e) + agent._session_db.update_system_prompt(agent.session_id, new_system_prompt) + # Reset flush cursor — new session starts with no messages written + agent._last_flushed_db_idx = 0 + except Exception as e: + logger.warning("Session DB compression split failed — new session will NOT be indexed: %s", e) + + # Notify the context engine that the session_id rotated because of + # compression (not a fresh /new). Plugin engines (e.g. hermes-lcm) use + # boundary_reason="compression" to preserve DAG lineage across the + # rollover instead of re-initializing fresh per-session state. + # See hermes-lcm#68. Built-in ContextCompressor ignores kwargs. + try: + _old_sid = locals().get("old_session_id") + if _old_sid and hasattr(agent.context_compressor, "on_session_start"): + agent.context_compressor.on_session_start( + agent.session_id or "", + boundary_reason="compression", + old_session_id=_old_sid, + ) + except Exception as _ce_err: + logger.debug("context engine on_session_start (compression): %s", _ce_err) + + # Notify memory providers of the compression-driven session_id rotation + # so provider-cached per-session state (Hindsight's _document_id, + # accumulated turn buffers, counters) refreshes. reset=False because + # the logical conversation continues; only the id and DB row rolled + # over. See #6672. + try: + _old_sid = locals().get("old_session_id") + if _old_sid and agent._memory_manager: + agent._memory_manager.on_session_switch( + agent.session_id or "", + parent_session_id=_old_sid, + reset=False, + reason="compression", + ) + except Exception as _me_err: + logger.debug("memory manager on_session_switch (compression): %s", _me_err) + + # Warn on repeated compressions (quality degrades with each pass) + _cc = agent.context_compressor.compression_count + if _cc >= 2: + agent._vprint( + f"{agent.log_prefix}⚠️ Session compressed {_cc} times — " + f"accuracy may degrade. Consider /new to start fresh.", + force=True, + ) + + # Update token estimate after compaction so pressure calculations + # use the post-compression count, not the stale pre-compression one. + # Use estimate_request_tokens_rough() so tool schemas are included — + # with 50+ tools enabled, schemas alone can add 20-30K tokens, and + # omitting them delays the next compression cycle far past the + # configured threshold (issue #14695). + _compressed_est = estimate_request_tokens_rough( + compressed, + system_prompt=new_system_prompt or "", + tools=agent.tools or None, + ) + agent.context_compressor.last_prompt_tokens = _compressed_est + agent.context_compressor.last_completion_tokens = 0 + + # Clear the file-read dedup cache. After compression the original + # read content is summarised away — if the model re-reads the same + # file it needs the full content, not a "file unchanged" stub. + try: + from tools.file_tools import reset_file_dedup + reset_file_dedup(task_id) + except Exception: + pass + + logger.info( + "context compression done: session=%s messages=%d->%d tokens=~%s", + agent.session_id or "none", _pre_msg_count, len(compressed), + f"{_compressed_est:,}", + ) + return compressed, new_system_prompt + + +def try_shrink_image_parts_in_messages(api_messages: list) -> bool: + """Re-encode all native image parts at a smaller size to recover from + image-too-large errors (Anthropic 5 MB, unknown other providers). + + Mutates ``api_messages`` in place. Returns True if any image part was + actually replaced, False if there were no image parts to shrink or + Pillow couldn't help (caller should surface the original error). + + Strategy: look for ``image_url`` / ``input_image`` parts carrying a + ``data:image/...;base64,...`` payload. For each one whose encoded + size exceeds 4 MB (a safe target that slides under Anthropic's 5 MB + ceiling with header overhead), write the base64 to a tempfile, call + ``vision_tools._resize_image_for_vision`` to produce a smaller data + URL, and substitute it in place. + + Non-data-URL images (http/https URLs) are not touched — the provider + fetches those itself and the size limit is different. + """ + if not api_messages: + return False + + try: + from tools.vision_tools import _resize_image_for_vision + except Exception as exc: + logger.warning("image-shrink recovery: vision_tools unavailable — %s", exc) + return False + + # 4 MB target leaves comfortable headroom under Anthropic's 5 MB. + # Non-Anthropic providers we haven't observed rejecting are fine with + # much larger; shrinking to 4 MB here loses quality but only fires + # after a confirmed provider rejection, so the alternative is failure. + target_bytes = 4 * 1024 * 1024 + changed_count = 0 + + def _shrink_data_url(url: str) -> Optional[str]: + """Return a smaller data URL, or None if shrink can't help.""" + if not isinstance(url, str) or not url.startswith("data:"): + return None + if len(url) <= target_bytes: + # This specific image wasn't the oversized one. + return None + try: + header, _, data = url.partition(",") + mime = "image/jpeg" + if header.startswith("data:"): + mime_part = header[len("data:"):].split(";", 1)[0].strip() + if mime_part.startswith("image/"): + mime = mime_part + import base64 as _b64 + raw = _b64.b64decode(data) + suffix = { + "image/png": ".png", "image/gif": ".gif", "image/webp": ".webp", + "image/jpeg": ".jpg", "image/jpg": ".jpg", "image/bmp": ".bmp", + }.get(mime, ".jpg") + tmp = tempfile.NamedTemporaryFile( + prefix="hermes_shrink_", suffix=suffix, delete=False, + ) + try: + tmp.write(raw) + tmp.close() + resized = _resize_image_for_vision( + Path(tmp.name), + mime_type=mime, + max_base64_bytes=target_bytes, + ) + finally: + try: + Path(tmp.name).unlink(missing_ok=True) + except Exception: + pass + if not resized or len(resized) >= len(url): + # Shrink didn't help (or made it bigger — corrupt input?). + return None + return resized + except Exception as exc: + logger.warning("image-shrink recovery: re-encode failed — %s", exc) + return None + + for msg in api_messages: + if not isinstance(msg, dict): + continue + content = msg.get("content") + if not isinstance(content, list): + continue + for part in content: + if not isinstance(part, dict): + continue + ptype = part.get("type") + if ptype not in {"image_url", "input_image"}: + continue + image_value = part.get("image_url") + # OpenAI chat.completions: {"image_url": {"url": "data:..."}} + # OpenAI Responses: {"image_url": "data:..."} + if isinstance(image_value, dict): + url = image_value.get("url", "") + resized = _shrink_data_url(url) + if resized: + image_value["url"] = resized + changed_count += 1 + elif isinstance(image_value, str): + resized = _shrink_data_url(image_value) + if resized: + part["image_url"] = resized + changed_count += 1 + + if changed_count: + logger.info( + "image-shrink recovery: re-encoded %d image part(s) to fit under %.0f MB", + changed_count, target_bytes / (1024 * 1024), + ) + return changed_count > 0 + + +__all__ = [ + "check_compression_model_feasibility", + "replay_compression_warning", + "compress_context", + "try_shrink_image_parts_in_messages", +] diff --git a/agent/conversation_loop.py b/agent/conversation_loop.py new file mode 100644 index 00000000000..8096b754298 --- /dev/null +++ b/agent/conversation_loop.py @@ -0,0 +1,4018 @@ +"""The agent conversation loop — extracted from ``run_agent.AIAgent``. + +This is the biggest single chunk pulled out of ``run_agent.py``: the +roughly 3,900-line :func:`run_conversation` body that drives one user +turn through the agent (model call, tool dispatch, retries, fallbacks, +compression, post-turn hooks, background memory/skill review nudges). + +The function takes the parent ``AIAgent`` instance as its first +argument (``agent``) and accesses its state via attribute lookup. +``_ra().AIAgent.run_conversation`` is now a thin forwarder. + +Symbols that production code or tests patch on ``run_agent`` directly +(``handle_function_call``, ``_set_interrupt``, ``OpenAI``, ...) are +resolved through :func:`_ra` so those patches keep working. +""" + +from __future__ import annotations + +import json +import logging +import os +import random +import re +import ssl +import threading +import time +import uuid +from typing import Any, Dict, List, Optional + +from agent.anthropic_adapter import _is_oauth_token +from agent.auxiliary_client import set_runtime_main +from agent.codex_responses_adapter import _summarize_user_message_for_log +from agent.display import KawaiiSpinner +from agent.error_classifier import FailoverReason, classify_api_error +from agent.iteration_budget import IterationBudget +from agent.memory_manager import build_memory_context_block +from agent.message_sanitization import ( + _repair_tool_call_arguments, + _sanitize_messages_non_ascii, + _sanitize_messages_surrogates, + _sanitize_structure_non_ascii, + _sanitize_structure_surrogates, + _sanitize_surrogates, + _sanitize_tools_non_ascii, + _strip_images_from_messages, + _strip_non_ascii, +) +from agent.model_metadata import ( + estimate_messages_tokens_rough, + estimate_request_tokens_rough, + get_next_probe_tier, + parse_available_output_tokens_from_error, + parse_context_limit_from_error, + save_context_length, +) +from agent.nous_rate_guard import ( + clear_nous_rate_limit, + is_genuine_nous_rate_limit, + nous_rate_limit_remaining, + record_nous_rate_limit, +) +from agent.process_bootstrap import _install_safe_stdio +from agent.prompt_caching import apply_anthropic_cache_control +from agent.retry_utils import jittered_backoff +from agent.trajectory import has_incomplete_scratchpad +from agent.usage_pricing import estimate_usage_cost, normalize_usage +from hermes_constants import display_hermes_home as _dhh_fn +from hermes_logging import set_session_context +from tools.schema_sanitizer import strip_pattern_and_format +from tools.skill_provenance import set_current_write_origin +from utils import base_url_host_matches, env_var_enabled + +logger = logging.getLogger(__name__) + + +def _ra(): + """Lazy reference to ``run_agent`` so callers can patch + ``run_agent.handle_function_call`` / ``run_agent._set_interrupt`` / + ``run_agent.OpenAI`` and have those patches reach this code path. + """ + import run_agent + return run_agent + + +def run_conversation( + agent, + user_message: str, + system_message: str = None, + conversation_history: List[Dict[str, Any]] = None, + task_id: str = None, + stream_callback: Optional[callable] = None, + persist_user_message: Optional[str] = None, +) -> Dict[str, Any]: + """ + Run a complete conversation with tool calling until completion. + + Args: + user_message (str): The user's message/question + system_message (str): Custom system message (optional, overrides ephemeral_system_prompt if provided) + conversation_history (List[Dict]): Previous conversation messages (optional) + task_id (str): Unique identifier for this task to isolate VMs between concurrent tasks (optional, auto-generated if not provided) + stream_callback: Optional callback invoked with each text delta during streaming. + Used by the TTS pipeline to start audio generation before the full response. + When None (default), API calls use the standard non-streaming path. + persist_user_message: Optional clean user message to store in + transcripts/history when user_message contains API-only + synthetic prefixes. + or queuing follow-up prefetch work. + + Returns: + Dict: Complete conversation result with final response and message history + """ + # Guard stdio against OSError from broken pipes (systemd/headless/daemon). + # Installed once, transparent when streams are healthy, prevents crash on write. + _install_safe_stdio() + + agent._ensure_db_session() + + # Tell auxiliary_client what the live main provider/model are for + # this turn. Used by tools whose behaviour depends on the active + # main model (e.g. vision_analyze's native fast path) so they see + # the CLI/gateway override instead of the stale config.yaml + # default. Idempotent — fine to call every turn. + try: + from agent.auxiliary_client import set_runtime_main + set_runtime_main( + getattr(agent, "provider", "") or "", + getattr(agent, "model", "") or "", + ) + except Exception: + pass + + # Tag all log records on this thread with the session ID so + # ``hermes logs --session `` can filter a single conversation. + from hermes_logging import set_session_context + set_session_context(agent.session_id) + + # Bind the skill write-origin ContextVar for this thread so tool + # handlers (e.g. skill_manage create) can tell whether they are + # running inside the background agent-improvement review fork vs. + # a foreground user-directed turn. Set at the top of each call; + # the review fork runs on its own thread with a fresh context, + # so the foreground value here does not leak into it. + from tools.skill_provenance import set_current_write_origin + set_current_write_origin(getattr(agent, "_memory_write_origin", "assistant_tool")) + + # If the previous turn activated fallback, restore the primary + # runtime so this turn gets a fresh attempt with the preferred model. + # No-op when _fallback_activated is False (gateway, first turn, etc.). + agent._restore_primary_runtime() + + # Sanitize surrogate characters from user input. Clipboard paste from + # rich-text editors (Google Docs, Word, etc.) can inject lone surrogates + # that are invalid UTF-8 and crash JSON serialization in the OpenAI SDK. + if isinstance(user_message, str): + user_message = _sanitize_surrogates(user_message) + if isinstance(persist_user_message, str): + persist_user_message = _sanitize_surrogates(persist_user_message) + + # Store stream callback for _interruptible_api_call to pick up + agent._stream_callback = stream_callback + agent._persist_user_message_idx = None + agent._persist_user_message_override = persist_user_message + # Generate unique task_id if not provided to isolate VMs between concurrent tasks + effective_task_id = task_id or str(uuid.uuid4()) + # Expose the active task_id so tools running mid-turn (e.g. delegate_task + # in delegate_tool.py) can identify this agent for the cross-agent file + # state registry. Set BEFORE any tool dispatch so snapshots taken at + # child-launch time see the parent's real id, not None. + agent._current_task_id = effective_task_id + + # Reset retry counters and iteration budget at the start of each turn + # so subagent usage from a previous turn doesn't eat into the next one. + agent._invalid_tool_retries = 0 + agent._invalid_json_retries = 0 + agent._empty_content_retries = 0 + agent._incomplete_scratchpad_retries = 0 + agent._codex_incomplete_retries = 0 + agent._thinking_prefill_retries = 0 + agent._post_tool_empty_retried = False + agent._last_content_with_tools = None + agent._last_content_tools_all_housekeeping = False + agent._mute_post_response = False + agent._unicode_sanitization_passes = 0 + agent._tool_guardrails.reset_for_turn() + agent._tool_guardrail_halt_decision = None + # True until the server rejects an image_url content part with an error + # like "Only 'text' content type is supported." Set to False on first + # rejection and kept False for the rest of the session so we never re-send + # images to a text-only endpoint. Scoped per `_run()` call, not per instance. + agent._vision_supported = True + + # Pre-turn connection health check: detect and clean up dead TCP + # connections left over from provider outages or dropped streams. + # This prevents the next API call from hanging on a zombie socket. + if agent.api_mode != "anthropic_messages": + try: + if agent._cleanup_dead_connections(): + agent._emit_status( + "🔌 Detected stale connections from a previous provider " + "issue — cleaned up automatically. Proceeding with fresh " + "connection." + ) + except Exception: + pass + # Replay compression warning through status_callback for gateway + # platforms (the callback was not wired during __init__). + if agent._compression_warning: + agent._replay_compression_warning() + agent._compression_warning = None # send once + + # NOTE: _turns_since_memory and _iters_since_skill are NOT reset here. + # They are initialized in __init__ and must persist across run_conversation + # calls so that nudge logic accumulates correctly in CLI mode. + agent.iteration_budget = IterationBudget(agent.max_iterations) + + # Log conversation turn start for debugging/observability + _preview_text = _summarize_user_message_for_log(user_message) + _msg_preview = (_preview_text[:80] + "...") if len(_preview_text) > 80 else _preview_text + _msg_preview = _msg_preview.replace("\n", " ") + logger.info( + "conversation turn: session=%s model=%s provider=%s platform=%s history=%d msg=%r", + agent.session_id or "none", agent.model, agent.provider or "unknown", + agent.platform or "unknown", len(conversation_history or []), + _msg_preview, + ) + + # Initialize conversation (copy to avoid mutating the caller's list) + messages = list(conversation_history) if conversation_history else [] + + # Hydrate todo store from conversation history (gateway creates a fresh + # AIAgent per message, so the in-memory store is empty -- we need to + # recover the todo state from the most recent todo tool response in history) + if conversation_history and not agent._todo_store.has_items(): + agent._hydrate_todo_store(conversation_history) + + # Hydrate per-session nudge counters from persisted history. + # Gateway creates a fresh AIAgent per inbound message (cache miss / + # 1h idle eviction / config-signature mismatch / process restart), so + # _turns_since_memory and _user_turn_count start at 0 every turn and + # the memory.nudge_interval trigger may never be reached. Reconstruct + # an effective count from prior user turns in conversation_history. + # Idempotent: a cached agent that already accumulated counters keeps + # them; only a freshly-built agent with empty in-memory state hydrates. + # See issue #22357. + if conversation_history and agent._user_turn_count == 0: + prior_user_turns = sum( + 1 for m in conversation_history if m.get("role") == "user" + ) + if prior_user_turns > 0: + agent._user_turn_count = prior_user_turns + if agent._memory_nudge_interval > 0 and agent._turns_since_memory == 0: + # % preserves original 1-in-N cadence rather than firing a + # review immediately on resume (which would surprise users + # whose session happened to land just past a multiple of N). + agent._turns_since_memory = prior_user_turns % agent._memory_nudge_interval + + + # Prefill messages (few-shot priming) are injected at API-call time only, + # never stored in the messages list. This keeps them ephemeral: they won't + # be saved to session DB, session logs, or batch trajectories, but they're + # automatically re-applied on every API call (including session continuations). + + # Track user turns for memory flush and periodic nudge logic + agent._user_turn_count += 1 + + # Reset the streaming context scrubber at the top of each turn so a + # hung span from a prior interrupted stream can't taint this turn's + # output. + scrubber = getattr(agent, "_stream_context_scrubber", None) + if scrubber is not None: + scrubber.reset() + # Reset the think scrubber for the same reason — an interrupted + # prior stream may have left us inside an unterminated block. + think_scrubber = getattr(agent, "_stream_think_scrubber", None) + if think_scrubber is not None: + think_scrubber.reset() + + # Preserve the original user message (no nudge injection). + original_user_message = persist_user_message if persist_user_message is not None else user_message + + # Track memory nudge trigger (turn-based, checked here). + # Skill trigger is checked AFTER the agent loop completes, based on + # how many tool iterations THIS turn used. + _should_review_memory = False + if (agent._memory_nudge_interval > 0 + and "memory" in agent.valid_tool_names + and agent._memory_store): + agent._turns_since_memory += 1 + if agent._turns_since_memory >= agent._memory_nudge_interval: + _should_review_memory = True + agent._turns_since_memory = 0 + + # Add user message + user_msg = {"role": "user", "content": user_message} + messages.append(user_msg) + current_turn_user_idx = len(messages) - 1 + agent._persist_user_message_idx = current_turn_user_idx + + if not agent.quiet_mode: + _print_preview = _summarize_user_message_for_log(user_message) + agent._safe_print(f"💬 Starting conversation: '{_print_preview[:60]}{'...' if len(_print_preview) > 60 else ''}'") + + # ── System prompt (cached per session for prefix caching) ── + # Built once on first call, reused for all subsequent calls. + # Only rebuilt after context compression events (which invalidate + # the cache and reload memory from disk). + # + # For continuing sessions (gateway creates a fresh AIAgent per + # message), we load the stored system prompt from the session DB + # instead of rebuilding. Rebuilding would pick up memory changes + # from disk that the model already knows about (it wrote them!), + # producing a different system prompt and breaking the Anthropic + # prefix cache. + if agent._cached_system_prompt is None: + stored_prompt = None + if conversation_history and agent._session_db: + try: + session_row = agent._session_db.get_session(agent.session_id) + if session_row: + stored_prompt = session_row.get("system_prompt") or None + except Exception: + pass # Fall through to build fresh + + if stored_prompt: + # Continuing session — reuse the exact system prompt from + # the previous turn so the Anthropic cache prefix matches. + agent._cached_system_prompt = stored_prompt + else: + # First turn of a new session — build from scratch. + agent._cached_system_prompt = agent._build_system_prompt(system_message) + # Plugin hook: on_session_start + # Fired once when a brand-new session is created (not on + # continuation). Plugins can use this to initialise + # session-scoped state (e.g. warm a memory cache). + try: + from hermes_cli.plugins import invoke_hook as _invoke_hook + _invoke_hook( + "on_session_start", + session_id=agent.session_id, + model=agent.model, + platform=getattr(agent, "platform", None) or "", + ) + except Exception as exc: + logger.warning("on_session_start hook failed: %s", exc) + + # Store the system prompt snapshot in SQLite + if agent._session_db: + try: + agent._session_db.update_system_prompt(agent.session_id, agent._cached_system_prompt) + except Exception as e: + logger.debug("Session DB update_system_prompt failed: %s", e) + + active_system_prompt = agent._cached_system_prompt + + # ── Preflight context compression ── + # Before entering the main loop, check if the loaded conversation + # history already exceeds the model's context threshold. This handles + # cases where a user switches to a model with a smaller context window + # while having a large existing session — compress proactively rather + # than waiting for an API error (which might be caught as a non-retryable + # 4xx and abort the request entirely). + if ( + agent.compression_enabled + and len(messages) > agent.context_compressor.protect_first_n + + agent.context_compressor.protect_last_n + 1 + ): + # Include tool schema tokens — with many tools these can add + # 20-30K+ tokens that the old sys+msg estimate missed entirely. + _preflight_tokens = estimate_request_tokens_rough( + messages, + system_prompt=active_system_prompt or "", + tools=agent.tools or None, + ) + + if _preflight_tokens >= agent.context_compressor.threshold_tokens: + logger.info( + "Preflight compression: ~%s tokens >= %s threshold (model %s, ctx %s)", + f"{_preflight_tokens:,}", + f"{agent.context_compressor.threshold_tokens:,}", + agent.model, + f"{agent.context_compressor.context_length:,}", + ) + agent._emit_status( + f"📦 Preflight compression: ~{_preflight_tokens:,} tokens " + f">= {agent.context_compressor.threshold_tokens:,} threshold. " + "This may take a moment." + ) + # May need multiple passes for very large sessions with small + # context windows (each pass summarises the middle N turns). + for _pass in range(3): + _orig_len = len(messages) + messages, active_system_prompt = agent._compress_context( + messages, system_message, approx_tokens=_preflight_tokens, + task_id=effective_task_id, + ) + if len(messages) >= _orig_len: + break # Cannot compress further + # Compression created a new session — clear the history + # reference so _flush_messages_to_session_db writes ALL + # compressed messages to the new session's SQLite, not + # skipping them because conversation_history is still the + # pre-compression length. + conversation_history = None + # Fix: reset retry counters after compression so the model + # gets a fresh budget on the compressed context. Without + # this, pre-compression retries carry over and the model + # hits "(empty)" immediately after compression-induced + # context loss. + agent._empty_content_retries = 0 + agent._thinking_prefill_retries = 0 + agent._last_content_with_tools = None + agent._last_content_tools_all_housekeeping = False + agent._mute_post_response = False + # Re-estimate after compression + _preflight_tokens = estimate_request_tokens_rough( + messages, + system_prompt=active_system_prompt or "", + tools=agent.tools or None, + ) + if _preflight_tokens < agent.context_compressor.threshold_tokens: + break # Under threshold + + # Plugin hook: pre_llm_call + # Fired once per turn before the tool-calling loop. Plugins can + # return a dict with a ``context`` key (or a plain string) whose + # value is appended to the current turn's user message. + # + # Context is ALWAYS injected into the user message, never the + # system prompt. This preserves the prompt cache prefix — the + # system prompt stays identical across turns so cached tokens + # are reused. The system prompt is Hermes's territory; plugins + # contribute context alongside the user's input. + # + # All injected context is ephemeral (not persisted to session DB). + _plugin_user_context = "" + try: + from hermes_cli.plugins import invoke_hook as _invoke_hook + _pre_results = _invoke_hook( + "pre_llm_call", + session_id=agent.session_id, + user_message=original_user_message, + conversation_history=list(messages), + is_first_turn=(not bool(conversation_history)), + model=agent.model, + platform=getattr(agent, "platform", None) or "", + sender_id=getattr(agent, "_user_id", None) or "", + ) + _ctx_parts: list[str] = [] + for r in _pre_results: + if isinstance(r, dict) and r.get("context"): + _ctx_parts.append(str(r["context"])) + elif isinstance(r, str) and r.strip(): + _ctx_parts.append(r) + if _ctx_parts: + _plugin_user_context = "\n\n".join(_ctx_parts) + except Exception as exc: + logger.warning("pre_llm_call hook failed: %s", exc) + + # Main conversation loop + api_call_count = 0 + final_response = None + interrupted = False + codex_ack_continuations = 0 + length_continue_retries = 0 + truncated_tool_call_retries = 0 + truncated_response_parts: List[str] = [] + compression_attempts = 0 + _turn_exit_reason = "unknown" # Diagnostic: why the loop ended + + # Per-turn file-mutation verifier state. Keyed by resolved path; + # each failed ``write_file`` / ``patch`` call records the error + # preview. Later successful writes to the same path remove the + # entry (the model recovered). At end-of-turn, any entries still + # present are surfaced in an advisory footer so the model cannot + # over-claim success while the file is actually unchanged on disk. + agent._turn_failed_file_mutations: Dict[str, Dict[str, Any]] = {} + + # Record the execution thread so interrupt()/clear_interrupt() can + # scope the tool-level interrupt signal to THIS agent's thread only. + # Must be set before any thread-scoped interrupt syncing. + agent._execution_thread_id = threading.current_thread().ident + + # Always clear stale per-thread state from a previous turn. If an + # interrupt arrived before startup finished, preserve it and bind it + # to this execution thread now instead of dropping it on the floor. + _ra()._set_interrupt(False, agent._execution_thread_id) + if agent._interrupt_requested: + _ra()._set_interrupt(True, agent._execution_thread_id) + agent._interrupt_thread_signal_pending = False + else: + agent._interrupt_message = None + agent._interrupt_thread_signal_pending = False + + # Notify memory providers of the new turn so cadence tracking works. + # Must happen BEFORE prefetch_all() so providers know which turn it is + # and can gate context/dialectic refresh via contextCadence/dialecticCadence. + if agent._memory_manager: + try: + _turn_msg = original_user_message if isinstance(original_user_message, str) else "" + agent._memory_manager.on_turn_start(agent._user_turn_count, _turn_msg) + except Exception: + pass + + # External memory provider: prefetch once before the tool loop. + # Reuse the cached result on every iteration to avoid re-calling + # prefetch_all() on each tool call (10 tool calls = 10x latency + cost). + # Use original_user_message (clean input) — user_message may contain + # injected skill content that bloats / breaks provider queries. + _ext_prefetch_cache = "" + if agent._memory_manager: + try: + _query = original_user_message if isinstance(original_user_message, str) else "" + _ext_prefetch_cache = agent._memory_manager.prefetch_all(_query) or "" + except Exception: + pass + + # Optional opt-in runtime: if api_mode == codex_app_server, hand the + # turn to the codex app-server subprocess (terminal/file ops/patching + # all run inside Codex). Default Hermes path is bypassed entirely. + # See agent/transports/codex_app_server_session.py for the adapter + # and references/codex-app-server-runtime.md for the rationale. + if agent.api_mode == "codex_app_server": + return agent._run_codex_app_server_turn( + user_message=user_message, + original_user_message=original_user_message, + messages=messages, + effective_task_id=effective_task_id, + should_review_memory=_should_review_memory, + ) + + while (api_call_count < agent.max_iterations and agent.iteration_budget.remaining > 0) or agent._budget_grace_call: + # Reset per-turn checkpoint dedup so each iteration can take one snapshot + agent._checkpoint_mgr.new_turn() + + # Check for interrupt request (e.g., user sent new message) + if agent._interrupt_requested: + interrupted = True + _turn_exit_reason = "interrupted_by_user" + if not agent.quiet_mode: + agent._safe_print("\n⚡ Breaking out of tool loop due to interrupt...") + break + + api_call_count += 1 + agent._api_call_count = api_call_count + agent._touch_activity(f"starting API call #{api_call_count}") + + # Grace call: the budget is exhausted but we gave the model one + # more chance. Consume the grace flag so the loop exits after + # this iteration regardless of outcome. + if agent._budget_grace_call: + agent._budget_grace_call = False + elif not agent.iteration_budget.consume(): + _turn_exit_reason = "budget_exhausted" + if not agent.quiet_mode: + agent._safe_print(f"\n⚠️ Iteration budget exhausted ({agent.iteration_budget.used}/{agent.iteration_budget.max_total} iterations used)") + break + + # Fire step_callback for gateway hooks (agent:step event) + if agent.step_callback is not None: + try: + prev_tools = [] + for _idx, _m in enumerate(reversed(messages)): + if _m.get("role") == "assistant" and _m.get("tool_calls"): + _fwd_start = len(messages) - _idx + _results_by_id = {} + for _tm in messages[_fwd_start:]: + if _tm.get("role") != "tool": + break + _tcid = _tm.get("tool_call_id") + if _tcid: + _results_by_id[_tcid] = _tm.get("content", "") + prev_tools = [ + { + "name": tc["function"]["name"], + "result": _results_by_id.get(tc.get("id")), + "arguments": tc["function"].get("arguments"), + } + for tc in _m["tool_calls"] + if isinstance(tc, dict) + ] + break + agent.step_callback(api_call_count, prev_tools) + except Exception as _step_err: + logger.debug("step_callback error (iteration %s): %s", api_call_count, _step_err) + + # Track tool-calling iterations for skill nudge. + # Counter resets whenever skill_manage is actually used. + if (agent._skill_nudge_interval > 0 + and "skill_manage" in agent.valid_tool_names): + agent._iters_since_skill += 1 + + # ── Pre-API-call /steer drain ────────────────────────────────── + # If a /steer arrived during the previous API call (while the model + # was thinking), drain it now — before we build api_messages — so + # the model sees the steer text on THIS iteration. Without this, + # steers sent during an API call only land after the NEXT tool batch, + # which may never come if the model returns a final response. + # + # We scan backwards for the last tool-role message in the messages + # list. If found, the steer is appended there. If not (first + # iteration, no tools yet), the steer stays pending for the next + # tool batch — injecting into a user message would break role + # alternation, and there's no tool output to piggyback on. + _pre_api_steer = agent._drain_pending_steer() + if _pre_api_steer: + _injected = False + for _si in range(len(messages) - 1, -1, -1): + _sm = messages[_si] + if isinstance(_sm, dict) and _sm.get("role") == "tool": + marker = f"\n\nUser guidance: {_pre_api_steer}" + existing = _sm.get("content", "") + if isinstance(existing, str): + _sm["content"] = existing + marker + else: + # Multimodal content blocks — append text block + try: + blocks = list(existing) if existing else [] + blocks.append({"type": "text", "text": marker}) + _sm["content"] = blocks + except Exception: + pass + _injected = True + logger.debug( + "Pre-API-call steer drain: injected into tool msg at index %d", + _si, + ) + break + if not _injected: + # No tool message to inject into — put it back so + # the post-tool-execution drain picks it up later. + _lock = getattr(agent, "_pending_steer_lock", None) + if _lock is not None: + with _lock: + if agent._pending_steer: + agent._pending_steer = agent._pending_steer + "\n" + _pre_api_steer + else: + agent._pending_steer = _pre_api_steer + else: + existing = getattr(agent, "_pending_steer", None) + agent._pending_steer = (existing + "\n" + _pre_api_steer) if existing else _pre_api_steer + + # Prepare messages for API call + # If we have an ephemeral system prompt, prepend it to the messages + # Note: Reasoning is embedded in content via tags for trajectory storage. + # However, providers like Moonshot AI require a separate 'reasoning_content' field + # on assistant messages with tool_calls. We handle both cases here. + request_logger = getattr(agent, "logger", None) or logging.getLogger(__name__) + repaired_tool_calls = agent._sanitize_tool_call_arguments( + messages, + logger=request_logger, + session_id=agent.session_id, + ) + if repaired_tool_calls > 0: + request_logger.info( + "Sanitized %s corrupted tool_call arguments before request (session=%s)", + repaired_tool_calls, + agent.session_id or "-", + ) + + # Defensive: repair malformed role-alternation before API call. + # Catches cases where the history got wedged into a + # ``tool → user`` or ``user → user`` tail (e.g. after empty- + # response scaffolding was stripped and a new user message + # landed after an orphan tool result). Most providers return + # empty content on malformed sequences, which would otherwise + # retrigger the empty-retry loop indefinitely. + repaired_seq = agent._repair_message_sequence(messages) + if repaired_seq > 0: + request_logger.info( + "Repaired %s message-alternation violations before request (session=%s)", + repaired_seq, + agent.session_id or "-", + ) + + api_messages = [] + for idx, msg in enumerate(messages): + api_msg = msg.copy() + + # Inject ephemeral context into the current turn's user message. + # Sources: memory manager prefetch + plugin pre_llm_call hooks + # with target="user_message" (the default). Both are + # API-call-time only — the original message in `messages` is + # never mutated, so nothing leaks into session persistence. + if idx == current_turn_user_idx and msg.get("role") == "user": + _injections = [] + if _ext_prefetch_cache: + _fenced = build_memory_context_block(_ext_prefetch_cache) + if _fenced: + _injections.append(_fenced) + if _plugin_user_context: + _injections.append(_plugin_user_context) + if _injections: + _base = api_msg.get("content", "") + if isinstance(_base, str): + api_msg["content"] = _base + "\n\n" + "\n\n".join(_injections) + + # For ALL assistant messages, pass reasoning back to the API + # This ensures multi-turn reasoning context is preserved + agent._copy_reasoning_content_for_api(msg, api_msg) + + # Remove 'reasoning' field - it's for trajectory storage only + # We've copied it to 'reasoning_content' for the API above + if "reasoning" in api_msg: + api_msg.pop("reasoning") + # Remove finish_reason - not accepted by strict APIs (e.g. Mistral) + if "finish_reason" in api_msg: + api_msg.pop("finish_reason") + # Strip internal thinking-prefill marker + api_msg.pop("_thinking_prefill", None) + # Strip Codex Responses API fields (call_id, response_item_id) for + # strict providers like Mistral, Fireworks, etc. that reject unknown fields. + # Uses new dicts so the internal messages list retains the fields + # for Codex Responses compatibility. + if agent._should_sanitize_tool_calls(): + agent._sanitize_tool_calls_for_strict_api(api_msg) + # Keep 'reasoning_details' - OpenRouter uses this for multi-turn reasoning context + # The signature field helps maintain reasoning continuity + api_messages.append(api_msg) + + # Build the final system message: cached prompt + ephemeral system prompt. + # Ephemeral additions are API-call-time only (not persisted to session DB). + # External recall context is injected into the user message, not the system + # prompt, so the stable cache prefix remains unchanged. + # + # NOTE: Plugin context from pre_llm_call hooks is injected into the + # user message (see injection block above), NOT the system prompt. + # This is intentional — system prompt modifications break the prompt + # cache prefix. The system prompt is reserved for Hermes internals. + # + # Hermes invariant: the system prompt is built ONCE per session + # (cached on ``_cached_system_prompt``) and replayed verbatim on + # every turn. We send it as a single content string so the + # bytes are byte-stable across turns and upstream prompt caches + # stay warm. + effective_system = active_system_prompt or "" + if agent.ephemeral_system_prompt: + effective_system = (effective_system + "\n\n" + agent.ephemeral_system_prompt).strip() + if effective_system: + api_messages = [{"role": "system", "content": effective_system}] + api_messages + + # Inject ephemeral prefill messages right after the system prompt + # but before conversation history. Same API-call-time-only pattern. + if agent.prefill_messages: + sys_offset = 1 if (api_messages and api_messages[0].get("role") == "system") else 0 + for idx, pfm in enumerate(agent.prefill_messages): + api_messages.insert(sys_offset + idx, pfm.copy()) + + # Apply Anthropic prompt caching for Claude models on native + # Anthropic, OpenRouter, and third-party Anthropic-compatible + # gateways. Auto-detected: if ``_use_prompt_caching`` is set, + # inject cache_control breakpoints (system + last 3 messages) + # to reduce input token costs by ~75% on multi-turn + # conversations. + if agent._use_prompt_caching: + api_messages = apply_anthropic_cache_control( + api_messages, + cache_ttl=agent._cache_ttl, + native_anthropic=agent._use_native_cache_layout, + ) + + # Safety net: strip orphaned tool results / add stubs for missing + # results before sending to the API. Runs unconditionally — not + # gated on context_compressor — so orphans from session loading or + # manual message manipulation are always caught. + api_messages = agent._sanitize_api_messages(api_messages) + + # Drop thinking-only assistant turns (reasoning but no visible + # output and no tool_calls) and merge any adjacent user messages + # left behind. Prevents Anthropic 400s ("The final block in an + # assistant message cannot be `thinking`.") and equivalent errors + # from third-party Anthropic-compatible gateways that can't replay + # a thinking-only turn. Runs on the per-call copy only — the + # stored conversation history keeps the reasoning block for the + # UI transcript and session persistence. + api_messages = agent._drop_thinking_only_and_merge_users(api_messages) + + # Normalize message whitespace and tool-call JSON for consistent + # prefix matching. Ensures bit-perfect prefixes across turns, + # which enables KV cache reuse on local inference servers + # (llama.cpp, vLLM, Ollama) and improves cache hit rates for + # cloud providers. Operates on api_messages (the API copy) so + # the original conversation history in `messages` is untouched. + for am in api_messages: + if isinstance(am.get("content"), str): + am["content"] = am["content"].strip() + for am in api_messages: + tcs = am.get("tool_calls") + if not tcs: + continue + new_tcs = [] + for tc in tcs: + if isinstance(tc, dict) and "function" in tc: + try: + args_obj = json.loads(tc["function"]["arguments"]) + tc = {**tc, "function": { + **tc["function"], + "arguments": json.dumps( + args_obj, separators=(",", ":"), + sort_keys=True, + ), + }} + except Exception: + tc["function"]["arguments"] = _repair_tool_call_arguments( + tc["function"]["arguments"], + tc["function"].get("name", "?"), + ) + new_tcs.append(tc) + am["tool_calls"] = new_tcs + + # Proactively strip any surrogate characters before the API call. + # Models served via Ollama (Kimi K2.5, GLM-5, Qwen) can return + # lone surrogates (U+D800-U+DFFF) that crash json.dumps() inside + # the OpenAI SDK. Sanitizing here prevents the 3-retry cycle. + _sanitize_messages_surrogates(api_messages) + + # Calculate approximate request size for logging + total_chars = sum(len(str(msg)) for msg in api_messages) + approx_tokens = estimate_messages_tokens_rough(api_messages) + + # Thinking spinner for quiet mode (animated during API call) + thinking_spinner = None + + if not agent.quiet_mode: + agent._vprint(f"\n{agent.log_prefix}🔄 Making API call #{api_call_count}/{agent.max_iterations}...") + agent._vprint(f"{agent.log_prefix} 📊 Request size: {len(api_messages)} messages, ~{approx_tokens:,} tokens (~{total_chars:,} chars)") + agent._vprint(f"{agent.log_prefix} 🔧 Available tools: {len(agent.tools) if agent.tools else 0}") + else: + # Animated thinking spinner in quiet mode + face = random.choice(KawaiiSpinner.get_thinking_faces()) + verb = random.choice(KawaiiSpinner.get_thinking_verbs()) + if agent.thinking_callback: + # CLI TUI mode: use prompt_toolkit widget instead of raw spinner + # (works in both streaming and non-streaming modes) + agent.thinking_callback(f"{face} {verb}...") + elif not agent._has_stream_consumers() and agent._should_start_quiet_spinner(): + # Raw KawaiiSpinner only when no streaming consumers and the + # spinner output has a safe sink. + spinner_type = random.choice(['brain', 'sparkle', 'pulse', 'moon', 'star']) + thinking_spinner = KawaiiSpinner(f"{face} {verb}...", spinner_type=spinner_type, print_fn=agent._print_fn) + thinking_spinner.start() + + # Log request details if verbose + if agent.verbose_logging: + logging.debug(f"API Request - Model: {agent.model}, Messages: {len(messages)}, Tools: {len(agent.tools) if agent.tools else 0}") + logging.debug(f"Last message role: {messages[-1]['role'] if messages else 'none'}") + logging.debug(f"Total message size: ~{approx_tokens:,} tokens") + + api_start_time = time.time() + retry_count = 0 + max_retries = agent._api_max_retries + primary_recovery_attempted = False + max_compression_attempts = 3 + codex_auth_retry_attempted=False + anthropic_auth_retry_attempted=False + nous_auth_retry_attempted=False + copilot_auth_retry_attempted=False + thinking_sig_retry_attempted = False + image_shrink_retry_attempted = False + oauth_1m_beta_retry_attempted = False + llama_cpp_grammar_retry_attempted = False + has_retried_429 = False + restart_with_compressed_messages = False + restart_with_length_continuation = False + + finish_reason = "stop" + response = None # Guard against UnboundLocalError if all retries fail + api_kwargs = None # Guard against UnboundLocalError in except handler + + while retry_count < max_retries: + # ── Nous Portal rate limit guard ────────────────────── + # If another session already recorded that Nous is rate- + # limited, skip the API call entirely. Each attempt + # (including SDK-level retries) counts against RPH and + # deepens the rate limit hole. + if agent.provider == "nous": + try: + from agent.nous_rate_guard import ( + nous_rate_limit_remaining, + format_remaining as _fmt_nous_remaining, + ) + _nous_remaining = nous_rate_limit_remaining() + if _nous_remaining is not None and _nous_remaining > 0: + _nous_msg = ( + f"Nous Portal rate limit active — " + f"resets in {_fmt_nous_remaining(_nous_remaining)}." + ) + agent._vprint( + f"{agent.log_prefix}⏳ {_nous_msg} Trying fallback...", + force=True, + ) + agent._emit_status(f"⏳ {_nous_msg}") + if agent._try_activate_fallback(): + retry_count = 0 + compression_attempts = 0 + primary_recovery_attempted = False + continue + # No fallback available — return with clear message + agent._persist_session(messages, conversation_history) + return { + "final_response": ( + f"⏳ {_nous_msg}\n\n" + "No fallback provider available. " + "Try again after the reset, or add a " + "fallback provider in config.yaml." + ), + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "failed": True, + "error": _nous_msg, + } + except ImportError: + pass + except Exception: + pass # Never let rate guard break the agent loop + + try: + agent._reset_stream_delivery_tracking() + api_kwargs = agent._build_api_kwargs(api_messages) + if agent._force_ascii_payload: + _sanitize_structure_non_ascii(api_kwargs) + if agent.api_mode == "codex_responses": + api_kwargs = agent._get_transport().preflight_kwargs(api_kwargs, allow_stream=False) + + try: + from hermes_cli.plugins import invoke_hook as _invoke_hook + request_messages = api_kwargs.get("messages") + if not isinstance(request_messages, list): + request_messages = api_kwargs.get("input") + if not isinstance(request_messages, list): + request_messages = api_messages + # Shallow-copy the outer list so plugins that retain the + # reference for async snapshotting don't observe later + # mutations of api_messages. The inner dicts are not + # mutated by the agent loop, so a shallow copy is + # sufficient; a deepcopy would walk every tool result + # and base64 image on every API call. + _invoke_hook( + "pre_api_request", + task_id=effective_task_id, + session_id=agent.session_id or "", + user_message=original_user_message, + conversation_history=list(messages), + platform=agent.platform or "", + model=agent.model, + provider=agent.provider, + base_url=agent.base_url, + api_mode=agent.api_mode, + api_call_count=api_call_count, + request_messages=list(request_messages) if isinstance(request_messages, list) else [], + message_count=len(api_messages), + tool_count=len(agent.tools or []), + approx_input_tokens=approx_tokens, + request_char_count=total_chars, + max_tokens=agent.max_tokens, + ) + except Exception: + pass + + if env_var_enabled("HERMES_DUMP_REQUESTS"): + agent._dump_api_request_debug(api_kwargs, reason="preflight") + + # Always prefer the streaming path — even without stream + # consumers. Streaming gives us fine-grained health + # checking (90s stale-stream detection, 60s read timeout) + # that the non-streaming path lacks. Without this, + # subagents and other quiet-mode callers can hang + # indefinitely when the provider keeps the connection + # alive with SSE pings but never delivers a response. + # The streaming path is a no-op for callbacks when no + # consumers are registered, and falls back to non- + # streaming automatically if the provider doesn't + # support it. + def _stop_spinner(): + nonlocal thinking_spinner + if thinking_spinner: + thinking_spinner.stop("") + thinking_spinner = None + if agent.thinking_callback: + agent.thinking_callback("") + + _use_streaming = True + # Provider signaled "stream not supported" on a previous + # attempt — switch to non-streaming for the rest of this + # session instead of re-failing every retry. + if getattr(agent, "_disable_streaming", False): + _use_streaming = False + # CopilotACPClient communicates via subprocess stdio and + # returns a plain SimpleNamespace — not an iterable + # stream. Mirror the ACP exclusion used for Responses + # API upgrade (lines ~1083-1085). + elif ( + agent.provider == "copilot-acp" + or str(agent.base_url or "").lower().startswith("acp://copilot") + or str(agent.base_url or "").lower().startswith("acp+tcp://") + ): + _use_streaming = False + elif not agent._has_stream_consumers(): + # No display/TTS consumer. Still prefer streaming for + # health checking, but skip for Mock clients in tests + # (mocks return SimpleNamespace, not stream iterators). + from unittest.mock import Mock + if isinstance(getattr(agent, "client", None), Mock): + _use_streaming = False + + if _use_streaming: + response = agent._interruptible_streaming_api_call( + api_kwargs, on_first_delta=_stop_spinner + ) + else: + response = agent._interruptible_api_call(api_kwargs) + + api_duration = time.time() - api_start_time + + # Stop thinking spinner silently -- the response box or tool + # execution messages that follow are more informative. + if thinking_spinner: + thinking_spinner.stop("") + thinking_spinner = None + if agent.thinking_callback: + agent.thinking_callback("") + + if not agent.quiet_mode: + agent._vprint(f"{agent.log_prefix}⏱️ API call completed in {api_duration:.2f}s") + + if agent.verbose_logging: + # Log response with provider info if available + resp_model = getattr(response, 'model', 'N/A') if response else 'N/A' + logging.debug(f"API Response received - Model: {resp_model}, Usage: {response.usage if hasattr(response, 'usage') else 'N/A'}") + + # Validate response shape before proceeding + response_invalid = False + error_details = [] + if agent.api_mode == "codex_responses": + _ct_v = agent._get_transport() + if not _ct_v.validate_response(response): + if response is None: + response_invalid = True + error_details.append("response is None") + else: + # Provider returned a terminal failure (e.g. quota exhaustion). + # Treat as invalid so the fallback chain is triggered instead of + # letting the error bubble up outside the retry/fallback loop. + _codex_resp_status = str(getattr(response, "status", "") or "").strip().lower() + if _codex_resp_status in {"failed", "cancelled"}: + _codex_error_obj = getattr(response, "error", None) + _codex_error_msg = ( + _codex_error_obj.get("message") if isinstance(_codex_error_obj, dict) + else str(_codex_error_obj) if _codex_error_obj + else f"Responses API returned status '{_codex_resp_status}'" + ) + logging.warning( + "Codex response status='%s' (error=%s). Routing to fallback. %s", + _codex_resp_status, _codex_error_msg, + agent._client_log_context(), + ) + response_invalid = True + error_details.append(f"response.status={_codex_resp_status}: {_codex_error_msg}") + else: + # output_text fallback: stream backfill may have failed + # but normalize can still recover from output_text + _out_text = getattr(response, "output_text", None) + _out_text_stripped = _out_text.strip() if isinstance(_out_text, str) else "" + if _out_text_stripped: + logger.debug( + "Codex response.output is empty but output_text is present " + "(%d chars); deferring to normalization.", + len(_out_text_stripped), + ) + else: + _resp_status = getattr(response, "status", None) + _resp_incomplete = getattr(response, "incomplete_details", None) + logger.warning( + "Codex response.output is empty after stream backfill " + "(status=%s, incomplete_details=%s, model=%s). %s", + _resp_status, _resp_incomplete, + getattr(response, "model", None), + f"api_mode={agent.api_mode} provider={agent.provider}", + ) + response_invalid = True + error_details.append("response.output is empty") + elif agent.api_mode == "anthropic_messages": + _tv = agent._get_transport() + if not _tv.validate_response(response): + response_invalid = True + if response is None: + error_details.append("response is None") + else: + error_details.append("response.content invalid (not a non-empty list)") + elif agent.api_mode == "bedrock_converse": + _btv = agent._get_transport() + if not _btv.validate_response(response): + response_invalid = True + if response is None: + error_details.append("response is None") + else: + error_details.append("Bedrock response invalid (no output or choices)") + else: + _ctv = agent._get_transport() + if not _ctv.validate_response(response): + response_invalid = True + if response is None: + error_details.append("response is None") + elif not hasattr(response, 'choices'): + error_details.append("response has no 'choices' attribute") + elif response.choices is None: + error_details.append("response.choices is None") + else: + error_details.append("response.choices is empty") + + if response_invalid: + # Stop spinner before printing error messages + if thinking_spinner: + thinking_spinner.stop("(´;ω;`) oops, retrying...") + thinking_spinner = None + if agent.thinking_callback: + agent.thinking_callback("") + + # Invalid response — could be rate limiting, provider timeout, + # upstream server error, or malformed response. + retry_count += 1 + + # Eager fallback: empty/malformed responses are a common + # rate-limit symptom. Switch to fallback immediately + # rather than retrying with extended backoff. + if agent._fallback_index < len(agent._fallback_chain): + agent._emit_status("⚠️ Empty/malformed response — switching to fallback...") + if agent._try_activate_fallback(): + retry_count = 0 + compression_attempts = 0 + primary_recovery_attempted = False + continue + + # Check for error field in response (some providers include this) + error_msg = "Unknown" + provider_name = "Unknown" + if response and hasattr(response, 'error') and response.error: + error_msg = str(response.error) + # Try to extract provider from error metadata + if hasattr(response.error, 'metadata') and response.error.metadata: + provider_name = response.error.metadata.get('provider_name', 'Unknown') + elif response and hasattr(response, 'message') and response.message: + error_msg = str(response.message) + + # Try to get provider from model field (OpenRouter often returns actual model used) + if provider_name == "Unknown" and response and hasattr(response, 'model') and response.model: + provider_name = f"model={response.model}" + + # Check for x-openrouter-provider or similar metadata + if provider_name == "Unknown" and response: + # Log all response attributes for debugging + resp_attrs = {k: str(v)[:100] for k, v in vars(response).items() if not k.startswith('_')} + if agent.verbose_logging: + logging.debug(f"Response attributes for invalid response: {resp_attrs}") + + # Extract error code from response for contextual diagnostics + _resp_error_code = None + if response and hasattr(response, 'error') and response.error: + _code_raw = getattr(response.error, 'code', None) + if _code_raw is None and isinstance(response.error, dict): + _code_raw = response.error.get('code') + if _code_raw is not None: + try: + _resp_error_code = int(_code_raw) + except (TypeError, ValueError): + pass + + # Build a human-readable failure hint from the error code + # and response time, instead of always assuming rate limiting. + if _resp_error_code == 524: + _failure_hint = f"upstream provider timed out (Cloudflare 524, {api_duration:.0f}s)" + elif _resp_error_code == 504: + _failure_hint = f"upstream gateway timeout (504, {api_duration:.0f}s)" + elif _resp_error_code == 429: + _failure_hint = f"rate limited by upstream provider (429)" + elif _resp_error_code in {500, 502}: + _failure_hint = f"upstream server error ({_resp_error_code}, {api_duration:.0f}s)" + elif _resp_error_code in {503, 529}: + _failure_hint = f"upstream provider overloaded ({_resp_error_code})" + elif _resp_error_code is not None: + _failure_hint = f"upstream error (code {_resp_error_code}, {api_duration:.0f}s)" + elif api_duration < 10: + _failure_hint = f"fast response ({api_duration:.1f}s) — likely rate limited" + elif api_duration > 60: + _failure_hint = f"slow response ({api_duration:.0f}s) — likely upstream timeout" + else: + _failure_hint = f"response time {api_duration:.1f}s" + + agent._vprint(f"{agent.log_prefix}⚠️ Invalid API response (attempt {retry_count}/{max_retries}): {', '.join(error_details)}", force=True) + agent._vprint(f"{agent.log_prefix} 🏢 Provider: {provider_name}", force=True) + cleaned_provider_error = agent._clean_error_message(error_msg) + agent._vprint(f"{agent.log_prefix} 📝 Provider message: {cleaned_provider_error}", force=True) + agent._vprint(f"{agent.log_prefix} ⏱️ {_failure_hint}", force=True) + + if retry_count >= max_retries: + # Try fallback before giving up + agent._emit_status(f"⚠️ Max retries ({max_retries}) for invalid responses — trying fallback...") + if agent._try_activate_fallback(): + retry_count = 0 + compression_attempts = 0 + primary_recovery_attempted = False + continue + agent._emit_status(f"❌ Max retries ({max_retries}) exceeded for invalid responses. Giving up.") + logging.error(f"{agent.log_prefix}Invalid API response after {max_retries} retries.") + agent._persist_session(messages, conversation_history) + return { + "messages": messages, + "completed": False, + "api_calls": api_call_count, + "error": f"Invalid API response after {max_retries} retries: {_failure_hint}", + "failed": True # Mark as failure for filtering + } + + # Backoff before retry — jittered exponential: 5s base, 120s cap + wait_time = jittered_backoff(retry_count, base_delay=5.0, max_delay=120.0) + agent._vprint(f"{agent.log_prefix}⏳ Retrying in {wait_time:.1f}s ({_failure_hint})...", force=True) + logging.warning(f"Invalid API response (retry {retry_count}/{max_retries}): {', '.join(error_details)} | Provider: {provider_name}") + + # Sleep in small increments to stay responsive to interrupts + sleep_end = time.time() + wait_time + _backoff_touch_counter = 0 + while time.time() < sleep_end: + if agent._interrupt_requested: + agent._vprint(f"{agent.log_prefix}⚡ Interrupt detected during retry wait, aborting.", force=True) + agent._persist_session(messages, conversation_history) + agent.clear_interrupt() + return { + "final_response": f"Operation interrupted during retry ({_failure_hint}, attempt {retry_count}/{max_retries}).", + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "interrupted": True, + } + time.sleep(0.2) + # Touch activity every ~30s so the gateway's inactivity + # monitor knows we're alive during backoff waits. + _backoff_touch_counter += 1 + if _backoff_touch_counter % 150 == 0: # 150 × 0.2s = 30s + agent._touch_activity( + f"retry backoff ({retry_count}/{max_retries}), " + f"{int(sleep_end - time.time())}s remaining" + ) + continue # Retry the API call + + # Check finish_reason before proceeding + if agent.api_mode == "codex_responses": + status = getattr(response, "status", None) + incomplete_details = getattr(response, "incomplete_details", None) + incomplete_reason = None + if isinstance(incomplete_details, dict): + incomplete_reason = incomplete_details.get("reason") + else: + incomplete_reason = getattr(incomplete_details, "reason", None) + if status == "incomplete" and incomplete_reason in {"max_output_tokens", "length"}: + finish_reason = "length" + else: + finish_reason = "stop" + elif agent.api_mode == "anthropic_messages": + _tfr = agent._get_transport() + finish_reason = _tfr.map_finish_reason(response.stop_reason) + elif agent.api_mode == "bedrock_converse": + # Bedrock response already normalized at dispatch — use transport + _bt_fr = agent._get_transport() + _bedrock_result = _bt_fr.normalize_response(response) + finish_reason = _bedrock_result.finish_reason + else: + _cc_fr = agent._get_transport() + _finish_result = _cc_fr.normalize_response(response) + finish_reason = _finish_result.finish_reason + assistant_message = _finish_result + if agent._should_treat_stop_as_truncated( + finish_reason, + assistant_message, + messages, + ): + agent._vprint( + f"{agent.log_prefix}⚠️ Treating suspicious Ollama/GLM stop response as truncated", + force=True, + ) + finish_reason = "length" + + if finish_reason == "length": + agent._vprint(f"{agent.log_prefix}⚠️ Response truncated (finish_reason='length') - model hit max output tokens", force=True) + + # Normalize the truncated response to a single OpenAI-style + # message shape so text-continuation and tool-call retry + # work uniformly across chat_completions, bedrock_converse, + # and anthropic_messages. For Anthropic we use the same + # adapter the agent loop already relies on so the rebuilt + # interim assistant message is byte-identical to what + # would have been appended in the non-truncated path. + _trunc_msg = None + _trunc_transport = agent._get_transport() + if agent.api_mode == "anthropic_messages": + _trunc_result = _trunc_transport.normalize_response( + response, strip_tool_prefix=agent._is_anthropic_oauth + ) + else: + _trunc_result = _trunc_transport.normalize_response(response) + _trunc_msg = _trunc_result + + _trunc_content = getattr(_trunc_msg, "content", None) if _trunc_msg else None + _trunc_has_tool_calls = bool(getattr(_trunc_msg, "tool_calls", None)) if _trunc_msg else False + + # ── Detect thinking-budget exhaustion ────────────── + # When the model spends ALL output tokens on reasoning + # and has none left for the response, continuation + # retries are pointless. Detect this early and give a + # targeted error instead of wasting 3 API calls. + # A response is "thinking exhausted" only when the model + # actually produced reasoning blocks but no visible text after + # them. Models that do not use tags (e.g. GLM-4.7 on + # NVIDIA Build, minimax) may return content=None or an empty + # string for unrelated reasons — treat those as normal + # truncations that deserve continuation retries, not as + # thinking-budget exhaustion. + _has_think_tags = bool( + _trunc_content and re.search( + r'<(?:think|thinking|reasoning|REASONING_SCRATCHPAD)[^>]*>', + _trunc_content, + re.IGNORECASE, + ) + ) + _thinking_exhausted = ( + not _trunc_has_tool_calls + and _has_think_tags + and ( + (_trunc_content is not None and not agent._has_content_after_think_block(_trunc_content)) + or _trunc_content is None + ) + ) + + if _thinking_exhausted: + _exhaust_error = ( + "Model used all output tokens on reasoning with none left " + "for the response. Try lowering reasoning effort or " + "increasing max_tokens." + ) + agent._vprint( + f"{agent.log_prefix}💭 Reasoning exhausted the output token budget — " + f"no visible response was produced.", + force=True, + ) + # Return a user-friendly message as the response so + # CLI (response box) and gateway (chat message) both + # display it naturally instead of a suppressed error. + _exhaust_response = ( + "⚠️ **Thinking Budget Exhausted**\n\n" + "The model used all its output tokens on reasoning " + "and had none left for the actual response.\n\n" + "To fix this:\n" + "→ Lower reasoning effort: `/thinkon low` or `/thinkon minimal`\n" + "→ Or switch to a larger/non-reasoning model with `/model`" + ) + agent._cleanup_task_resources(effective_task_id) + agent._persist_session(messages, conversation_history) + return { + "final_response": _exhaust_response, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "partial": True, + "error": _exhaust_error, + } + + if agent.api_mode in {"chat_completions", "bedrock_converse", "anthropic_messages"}: + assistant_message = _trunc_msg + if assistant_message is not None and not _trunc_has_tool_calls: + length_continue_retries += 1 + interim_msg = agent._build_assistant_message(assistant_message, finish_reason) + messages.append(interim_msg) + if assistant_message.content: + truncated_response_parts.append(assistant_message.content) + + if length_continue_retries < 3: + agent._vprint( + f"{agent.log_prefix}↻ Requesting continuation " + f"({length_continue_retries}/3)..." + ) + continue_msg = { + "role": "user", + "content": ( + "[System: Your previous response was truncated by the output " + "length limit. Continue exactly where you left off. Do not " + "restart or repeat prior text. Finish the answer directly.]" + ), + } + messages.append(continue_msg) + agent._session_messages = messages + agent._save_session_log(messages) + restart_with_length_continuation = True + break + + partial_response = agent._strip_think_blocks("".join(truncated_response_parts)).strip() + agent._cleanup_task_resources(effective_task_id) + agent._persist_session(messages, conversation_history) + return { + "final_response": partial_response or None, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "partial": True, + "error": "Response remained truncated after 3 continuation attempts", + } + + if agent.api_mode in {"chat_completions", "bedrock_converse", "anthropic_messages"}: + assistant_message = _trunc_msg + if assistant_message is not None and _trunc_has_tool_calls: + if truncated_tool_call_retries < 1: + truncated_tool_call_retries += 1 + agent._vprint( + f"{agent.log_prefix}⚠️ Truncated tool call detected — retrying API call...", + force=True, + ) + # Don't append the broken response to messages; + # just re-run the same API call from the current + # message state, giving the model another chance. + continue + agent._vprint( + f"{agent.log_prefix}⚠️ Truncated tool call response detected again — refusing to execute incomplete tool arguments.", + force=True, + ) + agent._cleanup_task_resources(effective_task_id) + agent._persist_session(messages, conversation_history) + return { + "final_response": None, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "partial": True, + "error": "Response truncated due to output length limit", + } + + # If we have prior messages, roll back to last complete state + if len(messages) > 1: + agent._vprint(f"{agent.log_prefix} ⏪ Rolling back to last complete assistant turn") + rolled_back_messages = agent._get_messages_up_to_last_assistant(messages) + + agent._cleanup_task_resources(effective_task_id) + agent._persist_session(messages, conversation_history) + + return { + "final_response": None, + "messages": rolled_back_messages, + "api_calls": api_call_count, + "completed": False, + "partial": True, + "error": "Response truncated due to output length limit" + } + else: + # First message was truncated - mark as failed + agent._vprint(f"{agent.log_prefix}❌ First response truncated - cannot recover", force=True) + agent._persist_session(messages, conversation_history) + return { + "final_response": None, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "failed": True, + "error": "First response truncated due to output length limit" + } + + # Track actual token usage from response for context management + if hasattr(response, 'usage') and response.usage: + canonical_usage = normalize_usage( + response.usage, + provider=agent.provider, + api_mode=agent.api_mode, + ) + prompt_tokens = canonical_usage.prompt_tokens + completion_tokens = canonical_usage.output_tokens + total_tokens = canonical_usage.total_tokens + usage_dict = { + "prompt_tokens": prompt_tokens, + "completion_tokens": completion_tokens, + "total_tokens": total_tokens, + } + agent.context_compressor.update_from_response(usage_dict) + + # Cache discovered context length after successful call. + # Only persist limits confirmed by the provider (parsed + # from the error message), not guessed probe tiers. + if getattr(agent.context_compressor, "_context_probed", False): + ctx = agent.context_compressor.context_length + if getattr(agent.context_compressor, "_context_probe_persistable", False): + save_context_length(agent.model, agent.base_url, ctx) + agent._safe_print(f"{agent.log_prefix}💾 Cached context length: {ctx:,} tokens for {agent.model}") + agent.context_compressor._context_probed = False + agent.context_compressor._context_probe_persistable = False + + agent.session_prompt_tokens += prompt_tokens + agent.session_completion_tokens += completion_tokens + agent.session_total_tokens += total_tokens + agent.session_api_calls += 1 + agent.session_input_tokens += canonical_usage.input_tokens + agent.session_output_tokens += canonical_usage.output_tokens + agent.session_cache_read_tokens += canonical_usage.cache_read_tokens + agent.session_cache_write_tokens += canonical_usage.cache_write_tokens + agent.session_reasoning_tokens += canonical_usage.reasoning_tokens + + # Log API call details for debugging/observability + _cache_pct = "" + if canonical_usage.cache_read_tokens and prompt_tokens: + _cache_pct = f" cache={canonical_usage.cache_read_tokens}/{prompt_tokens} ({100*canonical_usage.cache_read_tokens/prompt_tokens:.0f}%)" + logger.info( + "API call #%d: model=%s provider=%s in=%d out=%d total=%d latency=%.1fs%s", + agent.session_api_calls, agent.model, agent.provider or "unknown", + prompt_tokens, completion_tokens, total_tokens, + api_duration, _cache_pct, + ) + + cost_result = estimate_usage_cost( + agent.model, + canonical_usage, + provider=agent.provider, + base_url=agent.base_url, + api_key=getattr(agent, "api_key", ""), + ) + if cost_result.amount_usd is not None: + agent.session_estimated_cost_usd += float(cost_result.amount_usd) + agent.session_cost_status = cost_result.status + agent.session_cost_source = cost_result.source + + # Persist token counts to session DB for /insights. + # Do this for every platform with a session_id so non-CLI + # sessions (gateway, cron, delegated runs) cannot lose + # token/accounting data if a higher-level persistence path + # is skipped or fails. Gateway/session-store writes use + # absolute totals, so they safely overwrite these per-call + # deltas instead of double-counting them. + if agent._session_db and agent.session_id: + try: + # Ensure the session row exists before attempting UPDATE. + # Under concurrent load (cron/kanban), the initial + # _ensure_db_session() may have failed due to SQLite + # locking. Retry here so per-call token deltas are + # not silently lost (UPDATE on a non-existent row + # affects 0 rows without error). + if not agent._session_db_created: + agent._ensure_db_session() + agent._session_db.update_token_counts( + agent.session_id, + input_tokens=canonical_usage.input_tokens, + output_tokens=canonical_usage.output_tokens, + cache_read_tokens=canonical_usage.cache_read_tokens, + cache_write_tokens=canonical_usage.cache_write_tokens, + reasoning_tokens=canonical_usage.reasoning_tokens, + estimated_cost_usd=float(cost_result.amount_usd) + if cost_result.amount_usd is not None else None, + cost_status=cost_result.status, + cost_source=cost_result.source, + billing_provider=agent.provider, + billing_base_url=agent.base_url, + billing_mode="subscription_included" + if cost_result.status == "included" else None, + model=agent.model, + api_call_count=1, + ) + except Exception as e: + # Log token persistence failures so they're + # visible in agent.log — silent loss here is + # the root cause of undercounted analytics. + logger.debug( + "Token persistence failed (session=%s, tokens=%d): %s", + agent.session_id, total_tokens, e, + ) + + if agent.verbose_logging: + logging.debug(f"Token usage: prompt={usage_dict['prompt_tokens']:,}, completion={usage_dict['completion_tokens']:,}, total={usage_dict['total_tokens']:,}") + + # Surface cache hit stats for any provider that reports + # them — not just those where we inject cache_control + # markers. OpenAI/Kimi/DeepSeek/Qwen all do automatic + # server-side prefix caching and return + # ``prompt_tokens_details.cached_tokens``; users + # previously could not see their cache % because this + # line was gated on ``_use_prompt_caching``, which is + # only True for Anthropic-style marker injection. + # ``canonical_usage`` is already normalised from all + # three API shapes (Anthropic / Codex / OpenAI-chat) + # so we can rely on its values directly. + cached = canonical_usage.cache_read_tokens + written = canonical_usage.cache_write_tokens + prompt = usage_dict["prompt_tokens"] + if (cached or written) and not agent.quiet_mode: + hit_pct = (cached / prompt * 100) if prompt > 0 else 0 + agent._vprint( + f"{agent.log_prefix} 💾 Cache: " + f"{cached:,}/{prompt:,} tokens " + f"({hit_pct:.0f}% hit, {written:,} written)" + ) + + has_retried_429 = False # Reset on success + # Clear Nous rate limit state on successful request — + # proves the limit has reset and other sessions can + # resume hitting Nous. + if agent.provider == "nous": + try: + from agent.nous_rate_guard import clear_nous_rate_limit + clear_nous_rate_limit() + except Exception: + pass + agent._touch_activity(f"API call #{api_call_count} completed") + break # Success, exit retry loop + + except InterruptedError: + if thinking_spinner: + thinking_spinner.stop("") + thinking_spinner = None + if agent.thinking_callback: + agent.thinking_callback("") + api_elapsed = time.time() - api_start_time + agent._vprint(f"{agent.log_prefix}⚡ Interrupted during API call.", force=True) + agent._persist_session(messages, conversation_history) + interrupted = True + final_response = f"Operation interrupted: waiting for model response ({api_elapsed:.1f}s elapsed)." + break + + except Exception as api_error: + # Stop spinner before printing error messages + if thinking_spinner: + thinking_spinner.stop("(╥_╥) error, retrying...") + thinking_spinner = None + if agent.thinking_callback: + agent.thinking_callback("") + + # ----------------------------------------------------------- + # UnicodeEncodeError recovery. Two common causes: + # 1. Lone surrogates (U+D800..U+DFFF) from clipboard paste + # (Google Docs, rich-text editors) — sanitize and retry. + # 2. ASCII codec on systems with LANG=C or non-UTF-8 locale + # (e.g. Chromebooks) — any non-ASCII character fails. + # Detect via the error message mentioning 'ascii' codec. + # We sanitize messages in-place and may retry twice: + # first to strip surrogates, then once more for pure + # ASCII-only locale sanitization if needed. + # ----------------------------------------------------------- + if isinstance(api_error, UnicodeEncodeError) and getattr(agent, '_unicode_sanitization_passes', 0) < 2: + _err_str = str(api_error).lower() + _is_ascii_codec = "'ascii'" in _err_str or "ascii" in _err_str + # Detect surrogate errors — utf-8 codec refusing to + # encode U+D800..U+DFFF. The error text is: + # "'utf-8' codec can't encode characters in position + # N-M: surrogates not allowed" + _is_surrogate_error = ( + "surrogate" in _err_str + or ("'utf-8'" in _err_str and not _is_ascii_codec) + ) + # Sanitize surrogates from both the canonical `messages` + # list AND `api_messages` (the API-copy, which may carry + # `reasoning_content`/`reasoning_details` transformed + # from `reasoning` — fields the canonical list doesn't + # have directly). Also clean `api_kwargs` if built and + # `prefill_messages` if present. Mirrors the ASCII + # codec recovery below. + _surrogates_found = _sanitize_messages_surrogates(messages) + if isinstance(api_messages, list): + if _sanitize_messages_surrogates(api_messages): + _surrogates_found = True + if isinstance(api_kwargs, dict): + if _sanitize_structure_surrogates(api_kwargs): + _surrogates_found = True + if isinstance(getattr(agent, "prefill_messages", None), list): + if _sanitize_messages_surrogates(agent.prefill_messages): + _surrogates_found = True + # Gate the retry on the error type, not on whether we + # found anything — _force_ascii_payload / the extended + # surrogate walker above cover all known paths, but a + # new transformed field could still slip through. If + # the error was a surrogate encode failure, always let + # the retry run; the proactive sanitizer at line ~8781 + # runs again on the next iteration. Bounded by + # _unicode_sanitization_passes < 2 (outer guard). + if _surrogates_found or _is_surrogate_error: + agent._unicode_sanitization_passes += 1 + if _surrogates_found: + agent._vprint( + f"{agent.log_prefix}⚠️ Stripped invalid surrogate characters from messages. Retrying...", + force=True, + ) + else: + agent._vprint( + f"{agent.log_prefix}⚠️ Surrogate encoding error — retrying after full-payload sanitization...", + force=True, + ) + continue + if _is_ascii_codec: + agent._force_ascii_payload = True + # ASCII codec: the system encoding can't handle + # non-ASCII characters at all. Sanitize all + # non-ASCII content from messages/tool schemas and retry. + # Sanitize both the canonical `messages` list and + # `api_messages` (the API-copy built before the retry + # loop, which may contain extra fields like + # reasoning_content that are not in `messages`). + _messages_sanitized = _sanitize_messages_non_ascii(messages) + if isinstance(api_messages, list): + _sanitize_messages_non_ascii(api_messages) + # Also sanitize the last api_kwargs if already built, + # so a leftover non-ASCII value in a transformed field + # (e.g. extra_body, reasoning_content) doesn't survive + # into the next attempt via _build_api_kwargs cache paths. + if isinstance(api_kwargs, dict): + _sanitize_structure_non_ascii(api_kwargs) + _prefill_sanitized = False + if isinstance(getattr(agent, "prefill_messages", None), list): + _prefill_sanitized = _sanitize_messages_non_ascii(agent.prefill_messages) + + _tools_sanitized = False + if isinstance(getattr(agent, "tools", None), list): + _tools_sanitized = _sanitize_tools_non_ascii(agent.tools) + + _system_sanitized = False + if isinstance(active_system_prompt, str): + _sanitized_system = _strip_non_ascii(active_system_prompt) + if _sanitized_system != active_system_prompt: + active_system_prompt = _sanitized_system + agent._cached_system_prompt = _sanitized_system + _system_sanitized = True + if isinstance(getattr(agent, "ephemeral_system_prompt", None), str): + _sanitized_ephemeral = _strip_non_ascii(agent.ephemeral_system_prompt) + if _sanitized_ephemeral != agent.ephemeral_system_prompt: + agent.ephemeral_system_prompt = _sanitized_ephemeral + _system_sanitized = True + + _headers_sanitized = False + _default_headers = ( + agent._client_kwargs.get("default_headers") + if isinstance(getattr(agent, "_client_kwargs", None), dict) + else None + ) + if isinstance(_default_headers, dict): + _headers_sanitized = _sanitize_structure_non_ascii(_default_headers) + + # Sanitize the API key — non-ASCII characters in + # credentials (e.g. ʋ instead of v from a bad + # copy-paste) cause httpx to fail when encoding + # the Authorization header as ASCII. This is the + # most common cause of persistent UnicodeEncodeError + # that survives message/tool sanitization (#6843). + _credential_sanitized = False + _raw_key = getattr(agent, "api_key", None) or "" + if _raw_key: + _clean_key = _strip_non_ascii(_raw_key) + if _clean_key != _raw_key: + agent.api_key = _clean_key + if isinstance(getattr(agent, "_client_kwargs", None), dict): + agent._client_kwargs["api_key"] = _clean_key + # Also update the live client — it holds its + # own copy of api_key which auth_headers reads + # dynamically on every request. + if getattr(agent, "client", None) is not None and hasattr(agent.client, "api_key"): + agent.client.api_key = _clean_key + _credential_sanitized = True + agent._vprint( + f"{agent.log_prefix}⚠️ API key contained non-ASCII characters " + f"(bad copy-paste?) — stripped them. If auth fails, " + f"re-copy the key from your provider's dashboard.", + force=True, + ) + + # Always retry on ASCII codec detection — + # _force_ascii_payload guarantees the full + # api_kwargs payload is sanitized on the + # next iteration (line ~8475). Even when + # per-component checks above find nothing + # (e.g. non-ASCII only in api_messages' + # reasoning_content), the flag catches it. + # Bounded by _unicode_sanitization_passes < 2. + agent._unicode_sanitization_passes += 1 + _any_sanitized = ( + _messages_sanitized + or _prefill_sanitized + or _tools_sanitized + or _system_sanitized + or _headers_sanitized + or _credential_sanitized + ) + if _any_sanitized: + agent._vprint( + f"{agent.log_prefix}⚠️ System encoding is ASCII — stripped non-ASCII characters from request payload. Retrying...", + force=True, + ) + else: + agent._vprint( + f"{agent.log_prefix}⚠️ System encoding is ASCII — enabling full-payload sanitization for retry...", + force=True, + ) + continue + + # ── Image-rejection recovery ────────────────────────────── + # Some providers (mlx-lm, text-only endpoints, text-only + # fallbacks on multimodal models) reject any message that + # contains image_url content with a 4xx error like + # "Only 'text' content type is supported." On first hit, + # strip all images from the message list, mark the session + # as vision-unsupported, and retry with text only. + # + # Detection is best-effort English phrase matching — a + # locale-translated or heavily-reworded upstream error + # will bypass this guard and fall through to the normal + # error handler. Expand the phrase list when new + # provider wordings are observed in the wild. + _err_body = "" + try: + _err_body = str(getattr(api_error, "body", None) or + getattr(api_error, "message", None) or + str(api_error)) + except Exception: + pass + _err_status = getattr(api_error, "status_code", None) + _IMAGE_REJECTION_PHRASES = ( + "only 'text' content type is supported", + "only text content type is supported", + "image_url is not supported", + "image content is not supported", + "multimodal is not supported", + "multimodal content is not supported", + "multimodal input is not supported", + "vision is not supported", + "vision input is not supported", + "does not support images", + "does not support image input", + "does not support multimodal", + "does not support vision", + "model does not support image", + # ChatGPT-account Codex backend + # (https://chatgpt.com/backend-api/codex) rejects + # data:image/...base64 URLs in input_image fields + # with HTTP 400 "Invalid 'input[N].content[K].image_url'. + # Expected a valid URL, but got a value with an + # invalid format." The OpenAI Responses API on the + # public endpoint accepts data URLs, but the + # ChatGPT-account variant does not. Without this + # phrase the agent cascaded into compression / + # context-too-large recovery instead of just + # stripping the images. Match is narrow on + # purpose — keyed on the field-path apostrophe so + # we don't false-trip on other URL validation + # errors. (issue #23570) + "image_url'. expected", + # DeepSeek's OpenAI-compatible API reports text-only + # request-body variants as: + # "unknown variant `image_url`, expected `text`". + "unknown variant `image_url`, expected `text`", + "unknown variant image_url, expected text", + ) + _err_lower = _err_body.lower() + _looks_like_image_rejection = any( + p in _err_lower for p in _IMAGE_REJECTION_PHRASES + ) + # 4xx-only gate: never interpret 5xx/timeout as "server + # said no to images" — those are transient and must + # route to the normal retry path. + _status_ok = _err_status is None or (400 <= int(_err_status) < 500) + if ( + getattr(agent, "_vision_supported", True) + and _looks_like_image_rejection + and _status_ok + ): + agent._vision_supported = False + _imgs_removed = _strip_images_from_messages(messages) + if isinstance(api_messages, list): + _strip_images_from_messages(api_messages) + agent._vprint( + f"{agent.log_prefix}⚠️ Server rejected image content — " + f"switching to text-only mode for this session" + + (". Stripped images from history and retrying." if _imgs_removed else "."), + force=True, + ) + continue + + status_code = getattr(api_error, "status_code", None) + error_context = agent._extract_api_error_context(api_error) + + # ── Classify the error for structured recovery decisions ── + _compressor = getattr(agent, "context_compressor", None) + _ctx_len = getattr(_compressor, "context_length", 200000) if _compressor else 200000 + classified = classify_api_error( + api_error, + provider=getattr(agent, "provider", "") or "", + model=getattr(agent, "model", "") or "", + approx_tokens=approx_tokens, + context_length=_ctx_len, + num_messages=len(api_messages) if api_messages else 0, + ) + logger.debug( + "Error classified: reason=%s status=%s retryable=%s compress=%s rotate=%s fallback=%s", + classified.reason.value, classified.status_code, + classified.retryable, classified.should_compress, + classified.should_rotate_credential, classified.should_fallback, + ) + + recovered_with_pool, has_retried_429 = agent._recover_with_credential_pool( + status_code=status_code, + has_retried_429=has_retried_429, + classified_reason=classified.reason, + error_context=error_context, + ) + if recovered_with_pool: + continue + + # Image-too-large recovery: shrink oversized native image + # parts in-place and retry once. Triggered by Anthropic's + # per-image 5 MB ceiling (400 with "image exceeds 5 MB + # maximum") or any other provider that complains about + # image size. If shrink fails or a second attempt still + # fails, fall through to normal error handling. + if ( + classified.reason == FailoverReason.image_too_large + and not image_shrink_retry_attempted + ): + image_shrink_retry_attempted = True + if agent._try_shrink_image_parts_in_messages(api_messages): + agent._vprint( + f"{agent.log_prefix}📐 Image(s) exceeded provider size limit — " + f"shrank and retrying...", + force=True, + ) + continue + else: + logger.info( + "image-shrink recovery: no data-URL image parts found " + "or shrink didn't reduce size; surfacing original error." + ) + + # Anthropic OAuth subscription rejected the 1M-context beta + # header ("long context beta is not yet available for this + # subscription"). Disable the beta for the rest of this + # session, rebuild the client, and retry once. 1M-capable + # subscriptions never hit this branch — they accept the + # beta and keep full 1M context. See PR #17680 for the + # original report (we chose reactive recovery over the + # proposed unconditional omit so capable subscriptions + # don't silently lose the capability). + if ( + classified.reason == FailoverReason.oauth_long_context_beta_forbidden + and agent.api_mode == "anthropic_messages" + and agent._is_anthropic_oauth + and not oauth_1m_beta_retry_attempted + ): + oauth_1m_beta_retry_attempted = True + if not getattr(agent, "_oauth_1m_beta_disabled", False): + agent._oauth_1m_beta_disabled = True + try: + agent._anthropic_client.close() + except Exception: + pass + agent._rebuild_anthropic_client() + agent._vprint( + f"{agent.log_prefix}🔕 OAuth subscription doesn't support " + f"the 1M-context beta — disabled for this session and retrying...", + force=True, + ) + continue + + if ( + agent.api_mode == "codex_responses" + and agent.provider in {"openai-codex", "xai-oauth"} + and status_code == 401 + and not codex_auth_retry_attempted + ): + codex_auth_retry_attempted = True + if agent._try_refresh_codex_client_credentials(force=True): + _label = "xAI OAuth" if agent.provider == "xai-oauth" else "Codex" + agent._vprint(f"{agent.log_prefix}🔐 {_label} auth refreshed after 401. Retrying request...") + continue + if ( + agent.api_mode == "chat_completions" + and agent.provider == "nous" + and status_code == 401 + and not nous_auth_retry_attempted + ): + nous_auth_retry_attempted = True + if agent._try_refresh_nous_client_credentials(force=True): + print(f"{agent.log_prefix}🔐 Nous agent key refreshed after 401. Retrying request...") + continue + # Credential refresh didn't help — show diagnostic info. + # Most common causes: Portal OAuth expired/revoked, + # account out of credits, or agent key blocked. + from hermes_constants import display_hermes_home as _dhh_fn + _dhh = _dhh_fn() + _body_text = "" + try: + _body = getattr(api_error, "body", None) or getattr(api_error, "response", None) + if _body is not None: + _body_text = str(_body)[:200] + except Exception: + pass + print(f"{agent.log_prefix}🔐 Nous 401 — Portal authentication failed.") + if _body_text: + print(f"{agent.log_prefix} Response: {_body_text}") + print(f"{agent.log_prefix} Most likely: Portal OAuth expired, account out of credits, or agent key revoked.") + print(f"{agent.log_prefix} Troubleshooting:") + print(f"{agent.log_prefix} • Re-authenticate: hermes login --provider nous") + print(f"{agent.log_prefix} • Check credits / billing: https://portal.nousresearch.com") + print(f"{agent.log_prefix} • Verify stored credentials: {_dhh}/auth.json") + print(f"{agent.log_prefix} • Switch providers temporarily: /model --provider openrouter") + if ( + agent.provider == "copilot" + and status_code == 401 + and not copilot_auth_retry_attempted + ): + copilot_auth_retry_attempted = True + if agent._try_refresh_copilot_client_credentials(): + agent._vprint(f"{agent.log_prefix}🔐 Copilot credentials refreshed after 401. Retrying request...") + continue + if ( + agent.api_mode == "anthropic_messages" + and status_code == 401 + and hasattr(agent, '_anthropic_api_key') + and not anthropic_auth_retry_attempted + ): + anthropic_auth_retry_attempted = True + from agent.anthropic_adapter import _is_oauth_token + if agent._try_refresh_anthropic_client_credentials(): + print(f"{agent.log_prefix}🔐 Anthropic credentials refreshed after 401. Retrying request...") + continue + # Credential refresh didn't help — show diagnostic info + key = agent._anthropic_api_key + auth_method = "Bearer (OAuth/setup-token)" if _is_oauth_token(key) else "x-api-key (API key)" + print(f"{agent.log_prefix}🔐 Anthropic 401 — authentication failed.") + print(f"{agent.log_prefix} Auth method: {auth_method}") + print(f"{agent.log_prefix} Token prefix: {key[:12]}..." if key and len(key) > 12 else f"{agent.log_prefix} Token: (empty or short)") + print(f"{agent.log_prefix} Troubleshooting:") + from hermes_constants import display_hermes_home as _dhh_fn + _dhh = _dhh_fn() + print(f"{agent.log_prefix} • Check ANTHROPIC_TOKEN in {_dhh}/.env for Hermes-managed OAuth/setup tokens") + print(f"{agent.log_prefix} • Check ANTHROPIC_API_KEY in {_dhh}/.env for API keys or legacy token values") + print(f"{agent.log_prefix} • For API keys: verify at https://platform.claude.com/settings/keys") + print(f"{agent.log_prefix} • For Claude Code: run 'claude /login' to refresh, then retry") + print(f"{agent.log_prefix} • Legacy cleanup: hermes config set ANTHROPIC_TOKEN \"\"") + print(f"{agent.log_prefix} • Clear stale keys: hermes config set ANTHROPIC_API_KEY \"\"") + + # ── Thinking block signature recovery ───────────────── + # Anthropic signs thinking blocks against the full turn + # content. Any upstream mutation (context compression, + # session truncation, message merging) invalidates the + # signature → HTTP 400. Recovery: strip reasoning_details + # from all messages so the next retry sends no thinking + # blocks at all. One-shot — don't retry infinitely. + if ( + classified.reason == FailoverReason.thinking_signature + and not thinking_sig_retry_attempted + ): + thinking_sig_retry_attempted = True + for _m in messages: + if isinstance(_m, dict): + _m.pop("reasoning_details", None) + agent._vprint( + f"{agent.log_prefix}⚠️ Thinking block signature invalid — " + f"stripped all thinking blocks, retrying...", + force=True, + ) + logging.warning( + "%sThinking block signature recovery: stripped " + "reasoning_details from %d messages", + agent.log_prefix, len(messages), + ) + continue + + # ── llama.cpp grammar-parse recovery ────────────────── + # llama.cpp's ``json-schema-to-grammar`` converter rejects + # regex escape classes (``\d``, ``\w``, ``\s``) and most + # ``format`` values in tool schemas. MCP servers emit + # these routinely for date/phone/email params. Recovery: + # strip ``pattern``/``format`` from ``agent.tools`` and + # retry once. We keep the keywords by default so cloud + # providers get the full prompting hints; this branch + # fires only for users on llama.cpp's OAI server. + if ( + classified.reason == FailoverReason.llama_cpp_grammar_pattern + and not llama_cpp_grammar_retry_attempted + ): + llama_cpp_grammar_retry_attempted = True + try: + from tools.schema_sanitizer import strip_pattern_and_format + _, _stripped = strip_pattern_and_format(agent.tools) + except Exception as _strip_exc: # pragma: no cover — defensive + logging.warning( + "%sllama.cpp grammar recovery: strip helper failed: %s", + agent.log_prefix, _strip_exc, + ) + _stripped = 0 + if _stripped: + agent._vprint( + f"{agent.log_prefix}⚠️ llama.cpp rejected tool schema grammar — " + f"stripped {_stripped} pattern/format keyword(s), retrying...", + force=True, + ) + logging.warning( + "%sllama.cpp grammar recovery: stripped %d " + "pattern/format keyword(s) from tool schemas", + agent.log_prefix, _stripped, + ) + continue + # No keywords found to strip — fall through to normal + # retry path rather than loop forever on the same error. + logging.warning( + "%sllama.cpp grammar error but no pattern/format " + "keywords to strip — falling through to normal retry", + agent.log_prefix, + ) + + retry_count += 1 + elapsed_time = time.time() - api_start_time + agent._touch_activity( + f"API error recovery (attempt {retry_count}/{max_retries})" + ) + + error_type = type(api_error).__name__ + error_msg = str(api_error).lower() + _error_summary = agent._summarize_api_error(api_error) + logger.warning( + "API call failed (attempt %s/%s) error_type=%s %s summary=%s", + retry_count, + max_retries, + error_type, + agent._client_log_context(), + _error_summary, + ) + + _provider = getattr(agent, "provider", "unknown") + _base = getattr(agent, "base_url", "unknown") + _model = getattr(agent, "model", "unknown") + _status_code_str = f" [HTTP {status_code}]" if status_code else "" + agent._vprint(f"{agent.log_prefix}⚠️ API call failed (attempt {retry_count}/{max_retries}): {error_type}{_status_code_str}", force=True) + agent._vprint(f"{agent.log_prefix} 🔌 Provider: {_provider} Model: {_model}", force=True) + agent._vprint(f"{agent.log_prefix} 🌐 Endpoint: {_base}", force=True) + agent._vprint(f"{agent.log_prefix} 📝 Error: {_error_summary}", force=True) + if status_code and status_code < 500: + _err_body = getattr(api_error, "body", None) + _err_body_str = str(_err_body)[:300] if _err_body else None + if _err_body_str: + agent._vprint(f"{agent.log_prefix} 📋 Details: {_err_body_str}", force=True) + agent._vprint(f"{agent.log_prefix} ⏱️ Elapsed: {elapsed_time:.2f}s Context: {len(api_messages)} msgs, ~{approx_tokens:,} tokens") + + # Actionable hint for OpenRouter "no tool endpoints" error. + # This fires regardless of whether fallback succeeds — the + # user needs to know WHY their model failed so they can fix + # their provider routing, not just silently fall back. + if ( + agent._is_openrouter_url() + and "support tool use" in error_msg + ): + agent._vprint( + f"{agent.log_prefix} 💡 No OpenRouter providers for {_model} support tool calling with your current settings.", + force=True, + ) + if agent.providers_allowed: + agent._vprint( + f"{agent.log_prefix} Your provider_routing.only restriction is filtering out tool-capable providers.", + force=True, + ) + agent._vprint( + f"{agent.log_prefix} Try removing the restriction or adding providers that support tools for this model.", + force=True, + ) + agent._vprint( + f"{agent.log_prefix} Check which providers support tools: https://openrouter.ai/models/{_model}", + force=True, + ) + + # Check for interrupt before deciding to retry + if agent._interrupt_requested: + agent._vprint(f"{agent.log_prefix}⚡ Interrupt detected during error handling, aborting retries.", force=True) + agent._persist_session(messages, conversation_history) + agent.clear_interrupt() + return { + "final_response": f"Operation interrupted: handling API error ({error_type}: {agent._clean_error_message(str(api_error))}).", + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "interrupted": True, + } + + # Check for 413 payload-too-large BEFORE generic 4xx handler. + # A 413 is a payload-size error — the correct response is to + # compress history and retry, not abort immediately. + status_code = getattr(api_error, "status_code", None) + + # ── Anthropic Sonnet long-context tier gate ─────────── + # Anthropic returns HTTP 429 "Extra usage is required for + # long context requests" when a Claude Max (or similar) + # subscription doesn't include the 1M-context tier. This + # is NOT a transient rate limit — retrying or switching + # credentials won't help. Reduce context to 200k (the + # standard tier) and compress. + if classified.reason == FailoverReason.long_context_tier: + _reduced_ctx = 200000 + compressor = agent.context_compressor + old_ctx = compressor.context_length + if old_ctx > _reduced_ctx: + compressor.update_model( + model=agent.model, + context_length=_reduced_ctx, + base_url=agent.base_url, + api_key=getattr(agent, "api_key", ""), + provider=agent.provider, + ) + # Context probing flags — only set on built-in + # compressor (plugin engines manage their own). + if hasattr(compressor, "_context_probed"): + compressor._context_probed = True + # Don't persist — this is a subscription-tier + # limitation, not a model capability. If the + # user later enables extra usage the 1M limit + # should come back automatically. + compressor._context_probe_persistable = False + agent._vprint( + f"{agent.log_prefix}⚠️ Anthropic long-context tier " + f"requires extra usage — reducing context: " + f"{old_ctx:,} → {_reduced_ctx:,} tokens", + force=True, + ) + + compression_attempts += 1 + if compression_attempts <= max_compression_attempts: + original_len = len(messages) + messages, active_system_prompt = agent._compress_context( + messages, system_message, + approx_tokens=approx_tokens, + task_id=effective_task_id, + ) + # Compression created a new session — clear history + # so _flush_messages_to_session_db writes compressed + # messages to the new session, not skipping them. + conversation_history = None + if len(messages) < original_len or old_ctx > _reduced_ctx: + agent._emit_status( + f"🗜️ Context reduced to {_reduced_ctx:,} tokens " + f"(was {old_ctx:,}), retrying..." + ) + time.sleep(2) + restart_with_compressed_messages = True + break + # Fall through to normal error handling if compression + # is exhausted or didn't help. + + # Eager fallback for rate-limit errors (429 or quota exhaustion). + # When a fallback model is configured, switch immediately instead + # of burning through retries with exponential backoff -- the + # primary provider won't recover within the retry window. + is_rate_limited = classified.reason in { + FailoverReason.rate_limit, + FailoverReason.billing, + } + if is_rate_limited and agent._fallback_index < len(agent._fallback_chain): + # Don't eagerly fallback if credential pool rotation may + # still recover. See _pool_may_recover_from_rate_limit + # for the single-credential-pool and CloudCode-quota + # exceptions. Fixes #11314 and #13636. + pool_may_recover = _pool_may_recover_from_rate_limit( + agent._credential_pool, + provider=agent.provider, + base_url=getattr(agent, "base_url", None), + ) + if not pool_may_recover: + agent._emit_status("⚠️ Rate limited — switching to fallback provider...") + if agent._try_activate_fallback(reason=classified.reason): + retry_count = 0 + compression_attempts = 0 + primary_recovery_attempted = False + continue + + # ── Nous Portal: record rate limit & skip retries ───── + # When Nous returns a 429 that is a genuine account- + # level rate limit, record the reset time to a shared + # file so ALL sessions (cron, gateway, auxiliary) know + # not to pile on, then skip further retries -- each + # one burns another RPH request and deepens the hole. + # The retry loop's top-of-iteration guard will catch + # this on the next pass and try fallback or bail. + # + # IMPORTANT: Nous Portal multiplexes multiple upstream + # providers (DeepSeek, Kimi, MiMo, Hermes). A 429 can + # also mean an UPSTREAM provider is out of capacity + # for one specific model -- transient, clears in + # seconds, nothing to do with the caller's quota. + # Tripping the cross-session breaker on that would + # block every Nous model for minutes. We use + # ``is_genuine_nous_rate_limit`` to tell the two + # apart via the 429's own x-ratelimit-* headers and + # the last-known-good state captured on the previous + # successful response. + if ( + is_rate_limited + and agent.provider == "nous" + and classified.reason == FailoverReason.rate_limit + and not recovered_with_pool + ): + _genuine_nous_rate_limit = False + try: + from agent.nous_rate_guard import ( + is_genuine_nous_rate_limit, + record_nous_rate_limit, + ) + _err_resp = getattr(api_error, "response", None) + _err_hdrs = ( + getattr(_err_resp, "headers", None) + if _err_resp else None + ) + _genuine_nous_rate_limit = is_genuine_nous_rate_limit( + headers=_err_hdrs, + last_known_state=agent._rate_limit_state, + ) + if _genuine_nous_rate_limit: + record_nous_rate_limit( + headers=_err_hdrs, + error_context=error_context, + ) + else: + logging.info( + "Nous 429 looks like upstream capacity " + "(no exhausted bucket in headers or " + "last-known state) -- not tripping " + "cross-session breaker." + ) + except Exception: + pass + if _genuine_nous_rate_limit: + # Skip straight to max_retries -- the + # top-of-loop guard will handle fallback or + # bail cleanly. + retry_count = max_retries + continue + # Upstream capacity 429: fall through to normal + # retry logic. A different model (or the same + # model a moment later) will typically succeed. + + is_payload_too_large = ( + classified.reason == FailoverReason.payload_too_large + ) + + # Actionable hint for GitHub Models (Azure) 413 errors. + # The free tier enforces a hard 8K token cap per request, + # which Hermes' system prompt + tool schemas alone exceed. + # Compression can't help — the floor is the system prompt + # itself, not the conversation — so surface a clear "not + # compatible" message instead of looping into three futile + # compression attempts. + if ( + status_code == 413 + and isinstance(agent.base_url, str) + and "models.inference.ai.azure.com" in agent.base_url + ): + agent._vprint( + f"{agent.log_prefix} 💡 GitHub Models free tier (models.inference.ai.azure.com) caps every", + force=True, + ) + agent._vprint( + f"{agent.log_prefix} request at ~8K tokens. Hermes' system prompt + tool schemas baseline", + force=True, + ) + agent._vprint( + f"{agent.log_prefix} exceeds that floor, so this endpoint cannot run an agentic loop.", + force=True, + ) + agent._vprint( + f"{agent.log_prefix} Use the `copilot` provider with a Copilot subscription token (`hermes", + force=True, + ) + agent._vprint( + f"{agent.log_prefix} setup` → GitHub Copilot), or pick any other provider.", + force=True, + ) + + if is_payload_too_large: + compression_attempts += 1 + if compression_attempts > max_compression_attempts: + agent._vprint(f"{agent.log_prefix}❌ Max compression attempts ({max_compression_attempts}) reached for payload-too-large error.", force=True) + agent._vprint(f"{agent.log_prefix} 💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True) + logging.error(f"{agent.log_prefix}413 compression failed after {max_compression_attempts} attempts.") + agent._persist_session(messages, conversation_history) + return { + "messages": messages, + "completed": False, + "api_calls": api_call_count, + "error": f"Request payload too large: max compression attempts ({max_compression_attempts}) reached.", + "partial": True, + "failed": True, + "compression_exhausted": True, + } + agent._emit_status(f"⚠️ Request payload too large (413) — compression attempt {compression_attempts}/{max_compression_attempts}...") + + original_len = len(messages) + messages, active_system_prompt = agent._compress_context( + messages, system_message, approx_tokens=approx_tokens, + task_id=effective_task_id, + ) + # Compression created a new session — clear history + # so _flush_messages_to_session_db writes compressed + # messages to the new session, not skipping them. + conversation_history = None + + if len(messages) < original_len: + agent._emit_status(f"🗜️ Compressed {original_len} → {len(messages)} messages, retrying...") + time.sleep(2) # Brief pause between compression retries + restart_with_compressed_messages = True + break + else: + agent._vprint(f"{agent.log_prefix}❌ Payload too large and cannot compress further.", force=True) + agent._vprint(f"{agent.log_prefix} 💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True) + logging.error(f"{agent.log_prefix}413 payload too large. Cannot compress further.") + agent._persist_session(messages, conversation_history) + return { + "messages": messages, + "completed": False, + "api_calls": api_call_count, + "error": "Request payload too large (413). Cannot compress further.", + "partial": True, + "failed": True, + "compression_exhausted": True, + } + + # Check for context-length errors BEFORE generic 4xx handler. + # The classifier detects context overflow from: explicit error + # messages, generic 400 + large session heuristic (#1630), and + # server disconnect + large session pattern (#2153). + is_context_length_error = ( + classified.reason == FailoverReason.context_overflow + ) + + if is_context_length_error: + compressor = agent.context_compressor + old_ctx = compressor.context_length + + # ── Distinguish two very different errors ─────────── + # 1. "Prompt too long": the INPUT exceeds the context window. + # Fix: reduce context_length + compress history. + # 2. "max_tokens too large": input is fine, but + # input_tokens + requested max_tokens > context_window. + # Fix: reduce max_tokens (the OUTPUT cap) for this call. + # Do NOT shrink context_length — the window is unchanged. + # + # Note: max_tokens = output token cap (one response). + # context_length = total window (input + output combined). + available_out = parse_available_output_tokens_from_error(error_msg) + if available_out is not None: + # Error is purely about the output cap being too large. + # Cap output to the available space and retry without + # touching context_length or triggering compression. + safe_out = max(1, available_out - 64) # small safety margin + agent._ephemeral_max_output_tokens = safe_out + agent._vprint( + f"{agent.log_prefix}⚠️ Output cap too large for current prompt — " + f"retrying with max_tokens={safe_out:,} " + f"(available_tokens={available_out:,}; context_length unchanged at {old_ctx:,})", + force=True, + ) + # Still count against compression_attempts so we don't + # loop forever if the error keeps recurring. + compression_attempts += 1 + if compression_attempts > max_compression_attempts: + agent._vprint(f"{agent.log_prefix}❌ Max compression attempts ({max_compression_attempts}) reached.", force=True) + agent._vprint(f"{agent.log_prefix} 💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True) + logging.error(f"{agent.log_prefix}Context compression failed after {max_compression_attempts} attempts.") + agent._persist_session(messages, conversation_history) + return { + "messages": messages, + "completed": False, + "api_calls": api_call_count, + "error": f"Context length exceeded: max compression attempts ({max_compression_attempts}) reached.", + "partial": True, + "failed": True, + "compression_exhausted": True, + } + restart_with_compressed_messages = True + break + + # Error is about the INPUT being too large — reduce context_length. + # Try to parse the actual limit from the error message + parsed_limit = parse_context_limit_from_error(error_msg) + _provider_lower = (getattr(agent, "provider", "") or "").lower() + _base_lower = (getattr(agent, "base_url", "") or "").rstrip("/").lower() + is_minimax_provider = ( + _provider_lower in {"minimax", "minimax-cn"} + or _base_lower.startswith(( + "https://api.minimax.io/anthropic", + "https://api.minimaxi.com/anthropic", + )) + ) + minimax_delta_only_overflow = ( + is_minimax_provider + and parsed_limit is None + and "context window exceeds limit (" in error_msg + ) + if parsed_limit and parsed_limit < old_ctx: + new_ctx = parsed_limit + agent._vprint(f"{agent.log_prefix}Context limit detected from API: {new_ctx:,} tokens (was {old_ctx:,})", force=True) + elif minimax_delta_only_overflow: + new_ctx = old_ctx + agent._vprint( + f"{agent.log_prefix}Provider reported overflow amount only; " + f"keeping context_length at {old_ctx:,} tokens and compressing.", + force=True, + ) + else: + # Step down to the next probe tier + new_ctx = get_next_probe_tier(old_ctx) + + if new_ctx and new_ctx < old_ctx: + compressor.update_model( + model=agent.model, + context_length=new_ctx, + base_url=agent.base_url, + api_key=getattr(agent, "api_key", ""), + provider=agent.provider, + ) + # Context probing flags — only set on built-in + # compressor (plugin engines manage their own). + if hasattr(compressor, "_context_probed"): + compressor._context_probed = True + # Only persist limits parsed from the provider's + # error message (a real number). Guessed fallback + # tiers from get_next_probe_tier() should stay + # in-memory only — persisting them pollutes the + # cache with wrong values. + compressor._context_probe_persistable = bool( + parsed_limit and parsed_limit == new_ctx + ) + agent._vprint(f"{agent.log_prefix}⚠️ Context length exceeded — stepping down: {old_ctx:,} → {new_ctx:,} tokens", force=True) + else: + agent._vprint(f"{agent.log_prefix}⚠️ Context length exceeded at minimum tier — attempting compression...", force=True) + + compression_attempts += 1 + if compression_attempts > max_compression_attempts: + agent._vprint(f"{agent.log_prefix}❌ Max compression attempts ({max_compression_attempts}) reached.", force=True) + agent._vprint(f"{agent.log_prefix} 💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True) + logging.error(f"{agent.log_prefix}Context compression failed after {max_compression_attempts} attempts.") + agent._persist_session(messages, conversation_history) + return { + "messages": messages, + "completed": False, + "api_calls": api_call_count, + "error": f"Context length exceeded: max compression attempts ({max_compression_attempts}) reached.", + "partial": True, + "failed": True, + "compression_exhausted": True, + } + agent._emit_status(f"🗜️ Context too large (~{approx_tokens:,} tokens) — compressing ({compression_attempts}/{max_compression_attempts})...") + + original_len = len(messages) + messages, active_system_prompt = agent._compress_context( + messages, system_message, approx_tokens=approx_tokens, + task_id=effective_task_id, + ) + # Compression created a new session — clear history + # so _flush_messages_to_session_db writes compressed + # messages to the new session, not skipping them. + conversation_history = None + + if len(messages) < original_len or new_ctx and new_ctx < old_ctx: + if len(messages) < original_len: + agent._emit_status(f"🗜️ Compressed {original_len} → {len(messages)} messages, retrying...") + time.sleep(2) # Brief pause between compression retries + restart_with_compressed_messages = True + break + else: + # Can't compress further and already at minimum tier + agent._vprint(f"{agent.log_prefix}❌ Context length exceeded and cannot compress further.", force=True) + agent._vprint(f"{agent.log_prefix} 💡 The conversation has accumulated too much content. Try /new to start fresh, or /compress to manually trigger compression.", force=True) + logging.error(f"{agent.log_prefix}Context length exceeded: {approx_tokens:,} tokens. Cannot compress further.") + agent._persist_session(messages, conversation_history) + return { + "messages": messages, + "completed": False, + "api_calls": api_call_count, + "error": f"Context length exceeded ({approx_tokens:,} tokens). Cannot compress further.", + "partial": True, + "failed": True, + "compression_exhausted": True, + } + + # Check for non-retryable client errors. The classifier + # already accounts for 413, 429, 529 (transient), context + # overflow, and generic-400 heuristics. Local validation + # errors (ValueError, TypeError) are programming bugs. + # Exclude UnicodeEncodeError — it's a ValueError subclass + # but is handled separately by the surrogate sanitization + # path above. Exclude json.JSONDecodeError — also a + # ValueError subclass, but it indicates a transient + # provider/network failure (malformed response body, + # truncated stream, routing layer corruption), not a + # local programming bug, and should be retried (#14782). + is_local_validation_error = ( + isinstance(api_error, (ValueError, TypeError)) + and not isinstance( + api_error, (UnicodeEncodeError, json.JSONDecodeError) + ) + # ssl.SSLError (and its subclass SSLCertVerificationError) + # inherits from OSError *and* ValueError via Python MRO, + # so the isinstance(ValueError) check above would + # misclassify a TLS transport failure as a local + # programming bug and abort without retrying. Exclude + # ssl.SSLError explicitly so the error classifier's + # retryable=True mapping takes effect instead. + and not isinstance(api_error, ssl.SSLError) + ) + is_client_error = ( + is_local_validation_error + or ( + not classified.retryable + and not classified.should_compress + and classified.reason not in { + FailoverReason.rate_limit, + FailoverReason.billing, + FailoverReason.overloaded, + FailoverReason.context_overflow, + FailoverReason.payload_too_large, + FailoverReason.long_context_tier, + FailoverReason.thinking_signature, + } + ) + ) and not is_context_length_error + + if is_client_error: + # Try fallback before aborting — a different provider + # may not have the same issue (rate limit, auth, etc.) + agent._emit_status(f"⚠️ Non-retryable error (HTTP {status_code}) — trying fallback...") + if agent._try_activate_fallback(): + retry_count = 0 + compression_attempts = 0 + primary_recovery_attempted = False + continue + if api_kwargs is not None: + agent._dump_api_request_debug( + api_kwargs, reason="non_retryable_client_error", error=api_error, + ) + agent._emit_status( + f"❌ Non-retryable error (HTTP {status_code}): " + f"{agent._summarize_api_error(api_error)}" + ) + agent._vprint(f"{agent.log_prefix}❌ Non-retryable client error (HTTP {status_code}). Aborting.", force=True) + agent._vprint(f"{agent.log_prefix} 🔌 Provider: {_provider} Model: {_model}", force=True) + agent._vprint(f"{agent.log_prefix} 🌐 Endpoint: {_base}", force=True) + # Actionable guidance for common auth errors + if classified.is_auth or classified.reason == FailoverReason.billing: + if _provider in {"openai-codex", "xai-oauth"} and status_code == 401: + if _provider == "openai-codex": + agent._vprint(f"{agent.log_prefix} 💡 Codex OAuth token was rejected (HTTP 401). Your token may have been", force=True) + agent._vprint(f"{agent.log_prefix} refreshed by another client (Codex CLI, VS Code). To fix:", force=True) + agent._vprint(f"{agent.log_prefix} 1. Run `codex` in your terminal to generate fresh tokens.", force=True) + agent._vprint(f"{agent.log_prefix} 2. Then run `hermes auth` to re-authenticate.", force=True) + else: + agent._vprint(f"{agent.log_prefix} 💡 xAI OAuth token was rejected (HTTP 401). To fix:", force=True) + agent._vprint(f"{agent.log_prefix} re-authenticate with xAI Grok OAuth (SuperGrok Subscription) from `hermes model`.", force=True) + else: + agent._vprint(f"{agent.log_prefix} 💡 Your API key was rejected by the provider. Check:", force=True) + agent._vprint(f"{agent.log_prefix} • Is the key valid? Run: hermes setup", force=True) + agent._vprint(f"{agent.log_prefix} • Does your account have access to {_model}?", force=True) + if base_url_host_matches(str(_base), "openrouter.ai"): + agent._vprint(f"{agent.log_prefix} • Check credits: https://openrouter.ai/settings/credits", force=True) + else: + agent._vprint(f"{agent.log_prefix} 💡 This type of error won't be fixed by retrying.", force=True) + logging.error(f"{agent.log_prefix}Non-retryable client error: {api_error}") + # Skip session persistence when the error is likely + # context-overflow related (status 400 + large session). + # Persisting the failed user message would make the + # session even larger, causing the same failure on the + # next attempt. (#1630) + if status_code == 400 and (approx_tokens > 50000 or len(api_messages) > 80): + agent._vprint( + f"{agent.log_prefix}⚠️ Skipping session persistence " + f"for large failed session to prevent growth loop.", + force=True, + ) + else: + agent._persist_session(messages, conversation_history) + return { + "final_response": None, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "failed": True, + "error": str(api_error), + } + + if retry_count >= max_retries: + # Before falling back, try rebuilding the primary + # client once for transient transport errors (stale + # connection pool, TCP reset). Only attempted once + # per API call block. + if not primary_recovery_attempted and agent._try_recover_primary_transport( + api_error, retry_count=retry_count, max_retries=max_retries, + ): + primary_recovery_attempted = True + retry_count = 0 + continue + # Try fallback before giving up entirely + agent._emit_status(f"⚠️ Max retries ({max_retries}) exhausted — trying fallback...") + if agent._try_activate_fallback(): + retry_count = 0 + compression_attempts = 0 + primary_recovery_attempted = False + continue + _final_summary = agent._summarize_api_error(api_error) + if is_rate_limited: + agent._emit_status(f"❌ Rate limited after {max_retries} retries — {_final_summary}") + else: + agent._emit_status(f"❌ API failed after {max_retries} retries — {_final_summary}") + agent._vprint(f"{agent.log_prefix} 💀 Final error: {_final_summary}", force=True) + + # Detect SSE stream-drop pattern (e.g. "Network + # connection lost") and surface actionable guidance. + # This typically happens when the model generates a + # very large tool call (write_file with huge content) + # and the proxy/CDN drops the stream mid-response. + _is_stream_drop = ( + not getattr(api_error, "status_code", None) + and any(p in error_msg for p in ( + "connection lost", "connection reset", + "connection closed", "network connection", + "network error", "terminated", + )) + ) + if _is_stream_drop: + agent._vprint( + f"{agent.log_prefix} 💡 The provider's stream " + f"connection keeps dropping. This often happens " + f"when the model tries to write a very large " + f"file in a single tool call.", + force=True, + ) + agent._vprint( + f"{agent.log_prefix} Try asking the model " + f"to use execute_code with Python's open() for " + f"large files, or to write the file in smaller " + f"sections.", + force=True, + ) + + logging.error( + "%sAPI call failed after %s retries. %s | provider=%s model=%s msgs=%s tokens=~%s", + agent.log_prefix, max_retries, _final_summary, + _provider, _model, len(api_messages), f"{approx_tokens:,}", + ) + if api_kwargs is not None: + agent._dump_api_request_debug( + api_kwargs, reason="max_retries_exhausted", error=api_error, + ) + agent._persist_session(messages, conversation_history) + _final_response = f"API call failed after {max_retries} retries: {_final_summary}" + if _is_stream_drop: + _final_response += ( + "\n\nThe provider's stream connection keeps " + "dropping — this often happens when generating " + "very large tool call responses (e.g. write_file " + "with long content). Try asking me to use " + "execute_code with Python's open() for large " + "files, or to write in smaller sections." + ) + return { + "final_response": _final_response, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "failed": True, + "error": _final_summary, + } + + # For rate limits, respect the Retry-After header if present + _retry_after = None + if is_rate_limited: + _resp_headers = getattr(getattr(api_error, "response", None), "headers", None) + if _resp_headers and hasattr(_resp_headers, "get"): + _ra_raw = _resp_headers.get("retry-after") or _resp_headers.get("Retry-After") + if _ra_raw: + try: + _retry_after = min(float(_ra_raw), 120) # Cap at 2 minutes + except (TypeError, ValueError): + pass + wait_time = _retry_after if _retry_after else jittered_backoff(retry_count, base_delay=2.0, max_delay=60.0) + if is_rate_limited: + agent._emit_status(f"⏱️ Rate limited. Waiting {wait_time:.1f}s (attempt {retry_count + 1}/{max_retries})...") + else: + agent._emit_status(f"⏳ Retrying in {wait_time:.1f}s (attempt {retry_count}/{max_retries})...") + logger.warning( + "Retrying API call in %ss (attempt %s/%s) %s error=%s", + wait_time, + retry_count, + max_retries, + agent._client_log_context(), + api_error, + ) + # Sleep in small increments so we can respond to interrupts quickly + # instead of blocking the entire wait_time in one sleep() call + sleep_end = time.time() + wait_time + _backoff_touch_counter = 0 + while time.time() < sleep_end: + if agent._interrupt_requested: + agent._vprint(f"{agent.log_prefix}⚡ Interrupt detected during retry wait, aborting.", force=True) + agent._persist_session(messages, conversation_history) + agent.clear_interrupt() + return { + "final_response": f"Operation interrupted: retrying API call after error (retry {retry_count}/{max_retries}).", + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "interrupted": True, + } + time.sleep(0.2) # Check interrupt every 200ms + # Touch activity every ~30s so the gateway's inactivity + # monitor knows we're alive during backoff waits. + _backoff_touch_counter += 1 + if _backoff_touch_counter % 150 == 0: # 150 × 0.2s = 30s + agent._touch_activity( + f"error retry backoff ({retry_count}/{max_retries}), " + f"{int(sleep_end - time.time())}s remaining" + ) + + # If the API call was interrupted, skip response processing + if interrupted: + _turn_exit_reason = "interrupted_during_api_call" + break + + if restart_with_compressed_messages: + api_call_count -= 1 + agent.iteration_budget.refund() + # Count compression restarts toward the retry limit to prevent + # infinite loops when compression reduces messages but not enough + # to fit the context window. + retry_count += 1 + restart_with_compressed_messages = False + continue + + if restart_with_length_continuation: + # Progressively boost the output token budget on each retry. + # Retry 1 → 2× base, retry 2 → 3× base, capped at 32 768. + # Applies to all providers via _ephemeral_max_output_tokens. + _boost_base = agent.max_tokens if agent.max_tokens else 4096 + _boost = _boost_base * (length_continue_retries + 1) + agent._ephemeral_max_output_tokens = min(_boost, 32768) + continue + + # Guard: if all retries exhausted without a successful response + # (e.g. repeated context-length errors that exhausted retry_count), + # the `response` variable is still None. Break out cleanly. + if response is None: + _turn_exit_reason = "all_retries_exhausted_no_response" + print(f"{agent.log_prefix}❌ All API retries exhausted with no successful response.") + agent._persist_session(messages, conversation_history) + break + + try: + _transport = agent._get_transport() + _normalize_kwargs = {} + if agent.api_mode == "anthropic_messages": + _normalize_kwargs["strip_tool_prefix"] = agent._is_anthropic_oauth + normalized = _transport.normalize_response(response, **_normalize_kwargs) + assistant_message = normalized + finish_reason = normalized.finish_reason + + # Normalize content to string — some OpenAI-compatible servers + # (llama-server, etc.) return content as a dict or list instead + # of a plain string, which crashes downstream .strip() calls. + if assistant_message.content is not None and not isinstance(assistant_message.content, str): + raw = assistant_message.content + if isinstance(raw, dict): + assistant_message.content = raw.get("text", "") or raw.get("content", "") or json.dumps(raw) + elif isinstance(raw, list): + # Multimodal content list — extract text parts + parts = [] + for part in raw: + if isinstance(part, str): + parts.append(part) + elif isinstance(part, dict) and part.get("type") == "text": + parts.append(part.get("text", "")) + elif isinstance(part, dict) and "text" in part: + parts.append(str(part["text"])) + assistant_message.content = "\n".join(parts) + else: + assistant_message.content = str(raw) + + try: + from hermes_cli.plugins import invoke_hook as _invoke_hook + _assistant_tool_calls = getattr(assistant_message, "tool_calls", None) or [] + _assistant_text = assistant_message.content or "" + _invoke_hook( + "post_api_request", + task_id=effective_task_id, + session_id=agent.session_id or "", + platform=agent.platform or "", + model=agent.model, + provider=agent.provider, + base_url=agent.base_url, + api_mode=agent.api_mode, + api_call_count=api_call_count, + api_duration=api_duration, + finish_reason=finish_reason, + message_count=len(api_messages), + response_model=getattr(response, "model", None), + response=response, + usage=agent._usage_summary_for_api_request_hook(response), + assistant_message=assistant_message, + assistant_content_chars=len(_assistant_text), + assistant_tool_call_count=len(_assistant_tool_calls), + ) + except Exception: + pass + + # Handle assistant response + if assistant_message.content and not agent.quiet_mode: + if agent.verbose_logging: + agent._vprint(f"{agent.log_prefix}🤖 Assistant: {assistant_message.content}") + else: + agent._vprint(f"{agent.log_prefix}🤖 Assistant: {assistant_message.content[:100]}{'...' if len(assistant_message.content) > 100 else ''}") + + # Notify progress callback of model's thinking (used by subagent + # delegation to relay the child's reasoning to the parent display). + if (assistant_message.content and agent.tool_progress_callback): + _think_text = assistant_message.content.strip() + # Strip reasoning XML tags that shouldn't leak to parent display + _think_text = re.sub( + r'', '', _think_text + ).strip() + # For subagents: relay first line to parent display (existing behaviour). + # For all agents with a structured callback: emit reasoning.available event. + first_line = _think_text.split('\n')[0][:80] if _think_text else "" + if first_line and getattr(agent, '_delegate_depth', 0) > 0: + try: + agent.tool_progress_callback("_thinking", first_line) + except Exception: + pass + elif _think_text: + try: + agent.tool_progress_callback("reasoning.available", "_thinking", _think_text[:500], None) + except Exception: + pass + + # Check for incomplete (opened but never closed) + # This means the model ran out of output tokens mid-reasoning — retry up to 2 times + if has_incomplete_scratchpad(assistant_message.content or ""): + agent._incomplete_scratchpad_retries += 1 + + agent._vprint(f"{agent.log_prefix}⚠️ Incomplete detected (opened but never closed)") + + if agent._incomplete_scratchpad_retries <= 2: + agent._vprint(f"{agent.log_prefix}🔄 Retrying API call ({agent._incomplete_scratchpad_retries}/2)...") + # Don't add the broken message, just retry + continue + else: + # Max retries - discard this turn and save as partial + agent._vprint(f"{agent.log_prefix}❌ Max retries (2) for incomplete scratchpad. Saving as partial.", force=True) + agent._incomplete_scratchpad_retries = 0 + + rolled_back_messages = agent._get_messages_up_to_last_assistant(messages) + agent._cleanup_task_resources(effective_task_id) + agent._persist_session(messages, conversation_history) + + return { + "final_response": None, + "messages": rolled_back_messages, + "api_calls": api_call_count, + "completed": False, + "partial": True, + "error": "Incomplete REASONING_SCRATCHPAD after 2 retries" + } + + # Reset incomplete scratchpad counter on clean response + agent._incomplete_scratchpad_retries = 0 + + if agent.api_mode == "codex_responses" and finish_reason == "incomplete": + agent._codex_incomplete_retries += 1 + + interim_msg = agent._build_assistant_message(assistant_message, finish_reason) + interim_has_content = bool((interim_msg.get("content") or "").strip()) + interim_has_reasoning = bool(interim_msg.get("reasoning", "").strip()) if isinstance(interim_msg.get("reasoning"), str) else False + interim_has_codex_reasoning = bool(interim_msg.get("codex_reasoning_items")) + interim_has_codex_message_items = bool(interim_msg.get("codex_message_items")) + + if ( + interim_has_content + or interim_has_reasoning + or interim_has_codex_reasoning + or interim_has_codex_message_items + ): + last_msg = messages[-1] if messages else None + # Duplicate detection: two consecutive incomplete assistant + # messages with identical content AND reasoning are collapsed. + # For provider-state-only changes (encrypted reasoning + # items or replayable message ids/phases/statuses differ + # while visible content/reasoning are unchanged), compare + # those opaque payloads too so we don't silently drop the + # newer continuation state. + last_codex_items = last_msg.get("codex_reasoning_items") if isinstance(last_msg, dict) else None + interim_codex_items = interim_msg.get("codex_reasoning_items") + last_codex_message_items = last_msg.get("codex_message_items") if isinstance(last_msg, dict) else None + interim_codex_message_items = interim_msg.get("codex_message_items") + duplicate_interim = ( + isinstance(last_msg, dict) + and last_msg.get("role") == "assistant" + and last_msg.get("finish_reason") == "incomplete" + and (last_msg.get("content") or "") == (interim_msg.get("content") or "") + and (last_msg.get("reasoning") or "") == (interim_msg.get("reasoning") or "") + and last_codex_items == interim_codex_items + and last_codex_message_items == interim_codex_message_items + ) + if not duplicate_interim: + messages.append(interim_msg) + agent._emit_interim_assistant_message(interim_msg) + + if agent._codex_incomplete_retries < 3: + if not agent.quiet_mode: + agent._vprint(f"{agent.log_prefix}↻ Codex response incomplete; continuing turn ({agent._codex_incomplete_retries}/3)") + agent._session_messages = messages + agent._save_session_log(messages) + continue + + agent._codex_incomplete_retries = 0 + agent._persist_session(messages, conversation_history) + return { + "final_response": None, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "partial": True, + "error": "Codex response remained incomplete after 3 continuation attempts", + } + elif hasattr(agent, "_codex_incomplete_retries"): + agent._codex_incomplete_retries = 0 + + # Check for tool calls + if assistant_message.tool_calls: + if not agent.quiet_mode: + agent._vprint(f"{agent.log_prefix}🔧 Processing {len(assistant_message.tool_calls)} tool call(s)...") + + if agent.verbose_logging: + for tc in assistant_message.tool_calls: + logging.debug(f"Tool call: {tc.function.name} with args: {tc.function.arguments[:200]}...") + + # Validate tool call names - detect model hallucinations + # Repair mismatched tool names before validating + for tc in assistant_message.tool_calls: + if tc.function.name not in agent.valid_tool_names: + repaired = agent._repair_tool_call(tc.function.name) + if repaired: + print(f"{agent.log_prefix}🔧 Auto-repaired tool name: '{tc.function.name}' -> '{repaired}'") + tc.function.name = repaired + invalid_tool_calls = [ + tc.function.name for tc in assistant_message.tool_calls + if tc.function.name not in agent.valid_tool_names + ] + if invalid_tool_calls: + # Track retries for invalid tool calls + agent._invalid_tool_retries += 1 + + # Return helpful error to model — model can agent-correct next turn + available = ", ".join(sorted(agent.valid_tool_names)) + invalid_name = invalid_tool_calls[0] + invalid_preview = invalid_name[:80] + "..." if len(invalid_name) > 80 else invalid_name + agent._vprint(f"{agent.log_prefix}⚠️ Unknown tool '{invalid_preview}' — sending error to model for agent-correction ({agent._invalid_tool_retries}/3)") + + if agent._invalid_tool_retries >= 3: + agent._vprint(f"{agent.log_prefix}❌ Max retries (3) for invalid tool calls exceeded. Stopping as partial.", force=True) + agent._invalid_tool_retries = 0 + agent._persist_session(messages, conversation_history) + return { + "final_response": None, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "partial": True, + "error": f"Model generated invalid tool call: {invalid_preview}" + } + + assistant_msg = agent._build_assistant_message(assistant_message, finish_reason) + messages.append(assistant_msg) + for tc in assistant_message.tool_calls: + if tc.function.name not in agent.valid_tool_names: + content = f"Tool '{tc.function.name}' does not exist. Available tools: {available}" + else: + content = "Skipped: another tool call in this turn used an invalid name. Please retry this tool call." + messages.append({ + "role": "tool", + "name": tc.function.name, + "tool_call_id": tc.id, + "content": content, + }) + continue + # Reset retry counter on successful tool call validation + agent._invalid_tool_retries = 0 + + # Validate tool call arguments are valid JSON + # Handle empty strings as empty objects (common model quirk) + invalid_json_args = [] + for tc in assistant_message.tool_calls: + args = tc.function.arguments + if isinstance(args, (dict, list)): + tc.function.arguments = json.dumps(args) + continue + if args is not None and not isinstance(args, str): + tc.function.arguments = str(args) + args = tc.function.arguments + # Treat empty/whitespace strings as empty object + if not args or not args.strip(): + tc.function.arguments = "{}" + continue + try: + json.loads(args) + except json.JSONDecodeError as e: + invalid_json_args.append((tc.function.name, str(e))) + + if invalid_json_args: + # Check if the invalid JSON is due to truncation rather + # than a model formatting mistake. Routers sometimes + # rewrite finish_reason from "length" to "tool_calls", + # hiding the truncation from the length handler above. + # Detect truncation: args that don't end with } or ] + # (after stripping whitespace) are cut off mid-stream. + _truncated = any( + not (tc.function.arguments or "").rstrip().endswith(("}", "]")) + for tc in assistant_message.tool_calls + if tc.function.name in {n for n, _ in invalid_json_args} + ) + if _truncated: + agent._vprint( + f"{agent.log_prefix}⚠️ Truncated tool call arguments detected " + f"(finish_reason={finish_reason!r}) — refusing to execute.", + force=True, + ) + agent._invalid_json_retries = 0 + agent._cleanup_task_resources(effective_task_id) + agent._persist_session(messages, conversation_history) + return { + "final_response": None, + "messages": messages, + "api_calls": api_call_count, + "completed": False, + "partial": True, + "error": "Response truncated due to output length limit", + } + + # Track retries for invalid JSON arguments + agent._invalid_json_retries += 1 + + tool_name, error_msg = invalid_json_args[0] + agent._vprint(f"{agent.log_prefix}⚠️ Invalid JSON in tool call arguments for '{tool_name}': {error_msg}") + + if agent._invalid_json_retries < 3: + agent._vprint(f"{agent.log_prefix}🔄 Retrying API call ({agent._invalid_json_retries}/3)...") + # Don't add anything to messages, just retry the API call + continue + else: + # Instead of returning partial, inject tool error results so the model can recover. + # Using tool results (not user messages) preserves role alternation. + agent._vprint(f"{agent.log_prefix}⚠️ Injecting recovery tool results for invalid JSON...") + agent._invalid_json_retries = 0 # Reset for next attempt + + # Append the assistant message with its (broken) tool_calls + recovery_assistant = agent._build_assistant_message(assistant_message, finish_reason) + messages.append(recovery_assistant) + + # Respond with tool error results for each tool call + invalid_names = {name for name, _ in invalid_json_args} + for tc in assistant_message.tool_calls: + if tc.function.name in invalid_names: + err = next(e for n, e in invalid_json_args if n == tc.function.name) + tool_result = ( + f"Error: Invalid JSON arguments. {err}. " + f"For tools with no required parameters, use an empty object: {{}}. " + f"Please retry with valid JSON." + ) + else: + tool_result = "Skipped: other tool call in this response had invalid JSON." + messages.append({ + "role": "tool", + "name": tc.function.name, + "tool_call_id": tc.id, + "content": tool_result, + }) + continue + + # Reset retry counter on successful JSON validation + agent._invalid_json_retries = 0 + + # ── Post-call guardrails ────────────────────────── + assistant_message.tool_calls = agent._cap_delegate_task_calls( + assistant_message.tool_calls + ) + assistant_message.tool_calls = agent._deduplicate_tool_calls( + assistant_message.tool_calls + ) + + assistant_msg = agent._build_assistant_message(assistant_message, finish_reason) + + # If this turn has both content AND tool_calls, capture the content + # as a fallback final response. Common pattern: model delivers its + # answer and calls memory/skill tools as a side-effect in the same + # turn. If the follow-up turn after tools is empty, we use this. + turn_content = assistant_message.content or "" + if turn_content and agent._has_content_after_think_block(turn_content): + agent._last_content_with_tools = turn_content + # Only mute subsequent output when EVERY tool call in + # this turn is post-response housekeeping (memory, todo, + # skill_manage, etc.). If any substantive tool is present + # (search_files, read_file, write_file, terminal, ...), + # keep output visible so the user sees progress. + _HOUSEKEEPING_TOOLS = frozenset({ + "memory", "todo", "skill_manage", "session_search", + }) + _all_housekeeping = all( + tc.function.name in _HOUSEKEEPING_TOOLS + for tc in assistant_message.tool_calls + ) + agent._last_content_tools_all_housekeeping = _all_housekeeping + if _all_housekeeping and agent._has_stream_consumers(): + agent._mute_post_response = True + elif agent._should_emit_quiet_tool_messages(): + clean = agent._strip_think_blocks(turn_content).strip() + if clean: + agent._vprint(f" ┊ 💬 {clean}") + + # Pop thinking-only prefill message(s) before appending + # (tool-call path — same rationale as the final-response path). + _had_prefill = False + while ( + messages + and isinstance(messages[-1], dict) + and messages[-1].get("_thinking_prefill") + ): + messages.pop() + _had_prefill = True + + # Reset prefill counter when tool calls follow a prefill + # recovery. Without this, the counter accumulates across + # the whole conversation — a model that intermittently + # empties (empty → prefill → tools → empty → prefill → + # tools) burns both prefill attempts and the third empty + # gets zero recovery. Resetting here treats each tool- + # call success as a fresh start. + if _had_prefill: + agent._thinking_prefill_retries = 0 + agent._empty_content_retries = 0 + # Successful tool execution — reset the post-tool nudge + # flag so it can fire again if the model goes empty on + # a LATER tool round. + agent._post_tool_empty_retried = False + + messages.append(assistant_msg) + agent._emit_interim_assistant_message(assistant_msg) + + # Close any open streaming display (response box, reasoning + # box) before tool execution begins. Intermediate turns may + # have streamed early content that opened the response box; + # flushing here prevents it from wrapping tool feed lines. + # Only signal the display callback — TTS (_stream_callback) + # should NOT receive None (it uses None as end-of-stream). + if agent.stream_delta_callback: + try: + agent.stream_delta_callback(None) + except Exception: + pass + + agent._execute_tool_calls(assistant_message, messages, effective_task_id, api_call_count) + + if agent._tool_guardrail_halt_decision is not None: + decision = agent._tool_guardrail_halt_decision + _turn_exit_reason = "guardrail_halt" + final_response = agent._toolguard_controlled_halt_response(decision) + agent._emit_status( + f"⚠️ Tool guardrail halted {decision.tool_name}: {decision.code}" + ) + messages.append({"role": "assistant", "content": final_response}) + break + + # Reset per-turn retry counters after successful tool + # execution so a single truncation doesn't poison the + # entire conversation. + truncated_tool_call_retries = 0 + + # Signal that a paragraph break is needed before the next + # streamed text. We don't emit it immediately because + # multiple consecutive tool iterations would stack up + # redundant blank lines. Instead, _fire_stream_delta() + # will prepend a single "\n\n" the next time real text + # arrives. + agent._stream_needs_break = True + + # Refund the iteration if the ONLY tool(s) called were + # execute_code (programmatic tool calling). These are + # cheap RPC-style calls that shouldn't eat the budget. + _tc_names = {tc.function.name for tc in assistant_message.tool_calls} + if _tc_names == {"execute_code"}: + agent.iteration_budget.refund() + + # Use real token counts from the API response to decide + # compression. prompt_tokens + completion_tokens is the + # actual context size the provider reported plus the + # assistant turn — a tight lower bound for the next prompt. + # Tool results appended above aren't counted yet, but the + # threshold (default 50%) leaves ample headroom; if tool + # results push past it, the next API call will report the + # real total and trigger compression then. + # + # If last_prompt_tokens is 0 (stale after API disconnect + # or provider returned no usage data), fall back to rough + # estimate to avoid missing compression. Without this, + # a session can grow unbounded after disconnects because + # should_compress(0) never fires. (#2153) + _compressor = agent.context_compressor + if _compressor.last_prompt_tokens > 0: + # Only use prompt_tokens — completion/reasoning + # tokens don't consume context window space. + # Thinking models (GLM-5.1, QwQ, DeepSeek R1) + # inflate completion_tokens with reasoning, + # causing premature compression. (#12026) + _real_tokens = _compressor.last_prompt_tokens + else: + # Include tool schemas — with 50+ tools enabled + # these add 20-30K tokens the messages-only + # estimate misses, which can skip compression + # past the configured threshold (#14695). + _real_tokens = estimate_request_tokens_rough( + messages, tools=agent.tools or None + ) + + if agent.compression_enabled and _compressor.should_compress(_real_tokens): + agent._safe_print(" ⟳ compacting context…") + messages, active_system_prompt = agent._compress_context( + messages, system_message, + approx_tokens=agent.context_compressor.last_prompt_tokens, + task_id=effective_task_id, + ) + # Compression created a new session — clear history so + # _flush_messages_to_session_db writes compressed messages + # to the new session (see preflight compression comment). + conversation_history = None + + # Save session log incrementally (so progress is visible even if interrupted) + agent._session_messages = messages + agent._save_session_log(messages) + + # Continue loop for next response + continue + + else: + # No tool calls - this is the final response + final_response = assistant_message.content or "" + + # Fix: unmute output when entering the no-tool-call branch + # so the user can see empty-response warnings and recovery + # status messages. _mute_post_response was set during a + # prior housekeeping tool turn and should not silence the + # final response path. + agent._mute_post_response = False + + # Check if response only has think block with no actual content after it + if not agent._has_content_after_think_block(final_response): + # ── Partial stream recovery ───────────────────── + # If content was already streamed to the user before + # the connection died, use it as the final response + # instead of falling through to prior-turn fallback + # or wasting API calls on retries. + _partial_streamed = ( + getattr(agent, "_current_streamed_assistant_text", "") or "" + ) + if agent._has_content_after_think_block(_partial_streamed): + _turn_exit_reason = "partial_stream_recovery" + _recovered = agent._strip_think_blocks(_partial_streamed).strip() + logger.info( + "Partial stream content delivered (%d chars) " + "— using as final response", + len(_recovered), + ) + agent._emit_status( + "↻ Stream interrupted — using delivered content " + "as final response" + ) + final_response = _recovered + agent._response_was_previewed = True + break + + # If the previous turn already delivered real content alongside + # HOUSEKEEPING tool calls (e.g. "You're welcome!" + memory save), + # the model has nothing more to say. Use the earlier content + # immediately instead of wasting API calls on retries. + # NOTE: Only use this shortcut when ALL tools in that turn were + # housekeeping (memory, todo, etc.). When substantive tools + # were called (terminal, search_files, etc.), the content was + # likely mid-task narration ("I'll scan the directory...") and + # the empty follow-up means the model choked — let the + # post-tool nudge below handle that instead of exiting early. + fallback = getattr(agent, '_last_content_with_tools', None) + if fallback and getattr(agent, '_last_content_tools_all_housekeeping', False): + _turn_exit_reason = "fallback_prior_turn_content" + logger.info("Empty follow-up after tool calls — using prior turn content as final response") + agent._emit_status("↻ Empty response after tool calls — using earlier content as final answer") + agent._last_content_with_tools = None + agent._last_content_tools_all_housekeeping = False + agent._empty_content_retries = 0 + # Do NOT modify the assistant message content — the + # old code injected "Calling the X tools..." which + # poisoned the conversation history. Just use the + # fallback text as the final response and break. + final_response = agent._strip_think_blocks(fallback).strip() + agent._response_was_previewed = True + break + + # ── Post-tool-call empty response nudge ─────────── + # The model returned empty after executing tool calls. + # This covers two cases: + # (a) No prior-turn content at all — model went silent + # (b) Prior turn had content + SUBSTANTIVE tools (the + # fallback above was skipped because the content + # was mid-task narration, not a final answer) + # Instead of giving up, nudge the model to continue by + # appending a user-level hint. This is the #9400 case: + # weaker models (mimo-v2-pro, GLM-5, etc.) sometimes + # return empty after tool results instead of continuing + # to the next step. One retry with a nudge usually + # fixes it. + _prior_was_tool = any( + m.get("role") == "tool" + for m in messages[-5:] # check recent messages + ) + # Detect Qwen3/Ollama-style in-content thinking blocks. + # Ollama puts in the content field (not in + # reasoning_content), so _has_structured below would + # miss it. We check here so thinking-only responses + # after tool calls route to prefill instead of nudge. + _has_inline_thinking = bool( + re.search( + r'||', + final_response or "", + re.IGNORECASE, + ) + ) + if ( + _prior_was_tool + and not getattr(agent, "_post_tool_empty_retried", False) + and not _has_inline_thinking # thinking model still working — let prefill handle + ): + agent._post_tool_empty_retried = True + # Clear stale narration so it doesn't resurface + # on a later empty response after the nudge. + agent._last_content_with_tools = None + agent._last_content_tools_all_housekeeping = False + logger.info( + "Empty response after tool calls — nudging model " + "to continue processing" + ) + agent._emit_status( + "⚠️ Model returned empty after tool calls — " + "nudging to continue" + ) + # Append the empty assistant message first so the + # message sequence stays valid: + # tool(result) → assistant("(empty)") → user(nudge) + # Without this, we'd have tool → user which most + # APIs reject as an invalid sequence. + _nudge_msg = agent._build_assistant_message(assistant_message, finish_reason) + _nudge_msg["content"] = "(empty)" + _nudge_msg["_empty_recovery_synthetic"] = True + messages.append(_nudge_msg) + messages.append({ + "role": "user", + "content": ( + "You just executed tool calls but returned an " + "empty response. Please process the tool " + "results above and continue with the task." + ), + "_empty_recovery_synthetic": True, + }) + continue + + # ── Thinking-only prefill continuation ────────── + # The model produced structured reasoning (via API + # fields) but no visible text content. Rather than + # giving up, append the assistant message as-is and + # continue — the model will see its own reasoning + # on the next turn and produce the text portion. + # Inspired by clawdbot's "incomplete-text" recovery. + # Also covers Qwen3/Ollama in-content blocks + # (detected above as _has_inline_thinking). + _has_structured = bool( + getattr(assistant_message, "reasoning", None) + or getattr(assistant_message, "reasoning_content", None) + or getattr(assistant_message, "reasoning_details", None) + or _has_inline_thinking + ) + if _has_structured and agent._thinking_prefill_retries < 2: + agent._thinking_prefill_retries += 1 + logger.info( + "Thinking-only response (no visible content) — " + "prefilling to continue (%d/2)", + agent._thinking_prefill_retries, + ) + agent._emit_status( + f"↻ Thinking-only response — prefilling to continue " + f"({agent._thinking_prefill_retries}/2)" + ) + interim_msg = agent._build_assistant_message( + assistant_message, "incomplete" + ) + interim_msg["_thinking_prefill"] = True + messages.append(interim_msg) + agent._session_messages = messages + agent._save_session_log(messages) + continue + + # ── Empty response retry ────────────────────── + # Model returned nothing usable. Retry up to 3 + # times before attempting fallback. This covers + # both truly empty responses (no content, no + # reasoning) AND reasoning-only responses after + # prefill exhaustion — models like mimo-v2-pro + # always populate reasoning fields via OpenRouter, + # so the old `not _has_structured` guard blocked + # retries for every reasoning model after prefill. + _truly_empty = not agent._strip_think_blocks( + final_response + ).strip() + _prefill_exhausted = ( + _has_structured + and agent._thinking_prefill_retries >= 2 + ) + if _truly_empty and (not _has_structured or _prefill_exhausted) and agent._empty_content_retries < 3: + agent._empty_content_retries += 1 + logger.warning( + "Empty response (no content or reasoning) — " + "retry %d/3 (model=%s)", + agent._empty_content_retries, agent.model, + ) + agent._emit_status( + f"⚠️ Empty response from model — retrying " + f"({agent._empty_content_retries}/3)" + ) + continue + + # ── Exhausted retries — try fallback provider ── + # Before giving up with "(empty)", attempt to + # switch to the next provider in the fallback + # chain. This covers the case where a model + # (e.g. GLM-4.5-Air) consistently returns empty + # due to context degradation or provider issues. + if _truly_empty and agent._fallback_chain: + logger.warning( + "Empty response after %d retries — " + "attempting fallback (model=%s, provider=%s)", + agent._empty_content_retries, agent.model, + agent.provider, + ) + agent._emit_status( + "⚠️ Model returning empty responses — " + "switching to fallback provider..." + ) + if agent._try_activate_fallback(): + agent._empty_content_retries = 0 + agent._emit_status( + f"↻ Switched to fallback: {agent.model} " + f"({agent.provider})" + ) + logger.info( + "Fallback activated after empty responses: " + "now using %s on %s", + agent.model, agent.provider, + ) + continue + + # Exhausted retries and fallback chain (or no + # fallback configured). Fall through to the + # "(empty)" terminal. + _turn_exit_reason = "empty_response_exhausted" + reasoning_text = agent._extract_reasoning(assistant_message) + agent._drop_trailing_empty_response_scaffolding(messages) + assistant_msg = agent._build_assistant_message(assistant_message, finish_reason) + assistant_msg["content"] = "(empty)" + # This is a user-facing failure sentinel for the gateway, + # not real assistant content. Persisting it makes later + # "continue" turns replay assistant("(empty)") as if it + # were a meaningful model response, which can keep long + # tool-heavy sessions stuck in empty-response loops. + assistant_msg["_empty_terminal_sentinel"] = True + messages.append(assistant_msg) + + if reasoning_text: + reasoning_preview = reasoning_text[:500] + "..." if len(reasoning_text) > 500 else reasoning_text + logger.warning( + "Reasoning-only response (no visible content) " + "after exhausting retries and fallback. " + "Reasoning: %s", reasoning_preview, + ) + agent._emit_status( + "⚠️ Model produced reasoning but no visible " + "response after all retries. Returning empty." + ) + else: + logger.warning( + "Empty response (no content or reasoning) " + "after %d retries. No fallback available. " + "model=%s provider=%s", + agent._empty_content_retries, agent.model, + agent.provider, + ) + agent._emit_status( + "❌ Model returned no content after all retries" + + (" and fallback attempts." if agent._fallback_chain else + ". No fallback providers configured.") + ) + + final_response = "(empty)" + break + + # Reset retry counter/signature on successful content + agent._empty_content_retries = 0 + agent._thinking_prefill_retries = 0 + + if ( + agent.api_mode == "codex_responses" + and agent.valid_tool_names + and codex_ack_continuations < 2 + and agent._looks_like_codex_intermediate_ack( + user_message=user_message, + assistant_content=final_response, + messages=messages, + ) + ): + codex_ack_continuations += 1 + interim_msg = agent._build_assistant_message(assistant_message, "incomplete") + messages.append(interim_msg) + agent._emit_interim_assistant_message(interim_msg) + + continue_msg = { + "role": "user", + "content": ( + "[System: Continue now. Execute the required tool calls and only " + "send your final answer after completing the task.]" + ), + } + messages.append(continue_msg) + agent._session_messages = messages + agent._save_session_log(messages) + continue + + codex_ack_continuations = 0 + + if truncated_response_parts: + final_response = "".join(truncated_response_parts) + final_response + truncated_response_parts = [] + length_continue_retries = 0 + + final_response = agent._strip_think_blocks(final_response).strip() + + final_msg = agent._build_assistant_message(assistant_message, finish_reason) + + # Pop thinking-only prefill and empty-response retry + # scaffolding before appending the final response. These + # internal turns are only for the next API retry and should + # not become durable transcript context. + while ( + messages + and isinstance(messages[-1], dict) + and ( + messages[-1].get("_thinking_prefill") + or messages[-1].get("_empty_recovery_synthetic") + or messages[-1].get("_empty_terminal_sentinel") + ) + ): + messages.pop() + + messages.append(final_msg) + + _turn_exit_reason = f"text_response(finish_reason={finish_reason})" + if not agent.quiet_mode: + agent._safe_print(f"🎉 Conversation completed after {api_call_count} OpenAI-compatible API call(s)") + break + + except Exception as e: + error_msg = f"Error during OpenAI-compatible API call #{api_call_count}: {str(e)}" + try: + print(f"❌ {error_msg}") + except (OSError, ValueError): + logger.error(error_msg) + + logger.debug("Outer loop error in API call #%d", api_call_count, exc_info=True) + + # If an assistant message with tool_calls was already appended, + # the API expects a role="tool" result for every tool_call_id. + # Fill in error results for any that weren't answered yet. + for idx in range(len(messages) - 1, -1, -1): + msg = messages[idx] + if not isinstance(msg, dict): + break + if msg.get("role") == "tool": + continue + if msg.get("role") == "assistant" and msg.get("tool_calls"): + answered_ids = { + m["tool_call_id"] + for m in messages[idx + 1:] + if isinstance(m, dict) and m.get("role") == "tool" + } + for tc in msg["tool_calls"]: + if not tc or not isinstance(tc, dict): continue + if tc["id"] not in answered_ids: + err_msg = { + "role": "tool", + "name": _ra().AIAgent._get_tool_call_name_static(tc), + "tool_call_id": tc["id"], + "content": f"Error executing tool: {error_msg}", + } + messages.append(err_msg) + break + + # Non-tool errors don't need a synthetic message injected. + # The error is already printed to the user (line above), and + # the retry loop continues. Injecting a fake user/assistant + # message pollutes history, burns tokens, and risks violating + # role-alternation invariants. + + # If we're near the limit, break to avoid infinite loops + if api_call_count >= agent.max_iterations - 1: + _turn_exit_reason = f"error_near_max_iterations({error_msg[:80]})" + final_response = f"I apologize, but I encountered repeated errors: {error_msg}" + # Append as assistant so the history stays valid for + # session resume (avoids consecutive user messages). + messages.append({"role": "assistant", "content": final_response}) + break + + if final_response is None and ( + api_call_count >= agent.max_iterations + or agent.iteration_budget.remaining <= 0 + ): + # Budget exhausted — ask the model for a summary via one extra + # API call with tools stripped. _handle_max_iterations injects a + # user message and makes a single toolless request. + _turn_exit_reason = f"max_iterations_reached({api_call_count}/{agent.max_iterations})" + agent._emit_status( + f"⚠️ Iteration budget exhausted ({api_call_count}/{agent.max_iterations}) " + "— asking model to summarise" + ) + if not agent.quiet_mode: + agent._safe_print( + f"\n⚠️ Iteration budget exhausted ({api_call_count}/{agent.max_iterations}) " + "— requesting summary..." + ) + final_response = agent._handle_max_iterations(messages, api_call_count) + + # If running as a kanban worker, block the task so the dispatcher + # knows the worker could not complete (rather than treating it as a + # protocol violation). The agent loop strips tools before calling + # _handle_max_iterations, so the model cannot call kanban_block + # itself — we must do it on its behalf. + _kanban_task = os.environ.get("HERMES_KANBAN_TASK") + if _kanban_task: + try: + _ra().handle_function_call( + "kanban_block", + { + "task_id": _kanban_task, + "reason": ( + f"Iteration budget exhausted " + f"({api_call_count}/{agent.max_iterations}) — " + "task could not complete within the allowed " + "iterations" + ), + }, + task_id=effective_task_id, + ) + logger.info( + "kanban_block called for task %s after iteration " + "exhaustion (%d/%d)", + _kanban_task, api_call_count, agent.max_iterations, + ) + except Exception: + logger.warning( + "Failed to call kanban_block after iteration " + "exhaustion for task %s", + _kanban_task, + exc_info=True, + ) + + # Determine if conversation completed successfully + completed = final_response is not None and api_call_count < agent.max_iterations + + # Save trajectory if enabled. ``user_message`` may be a multimodal + # list of parts; the trajectory format wants a plain string. + agent._save_trajectory(messages, _summarize_user_message_for_log(user_message), completed) + + # Clean up VM and browser for this task after conversation completes + agent._cleanup_task_resources(effective_task_id) + + # Persist session to both JSON log and SQLite only after private retry + # scaffolding has been removed. Otherwise a later user "continue" turn + # can replay assistant("(empty)") / recovery nudges and fall into the + # same empty-response loop again. + agent._drop_trailing_empty_response_scaffolding(messages) + agent._persist_session(messages, conversation_history) + + # ── Turn-exit diagnostic log ───────────────────────────────────── + # Always logged at INFO so agent.log captures WHY every turn ended. + # When the last message is a tool result (agent was mid-work), log + # at WARNING — this is the "just stops" scenario users report. + _last_msg_role = messages[-1].get("role") if messages else None + _last_tool_name = None + if _last_msg_role == "tool": + # Walk back to find the assistant message with the tool call + for _m in reversed(messages): + if _m.get("role") == "assistant" and _m.get("tool_calls"): + _tcs = _m["tool_calls"] + if _tcs and isinstance(_tcs[0], dict): + _last_tool_name = _tcs[-1].get("function", {}).get("name") + break + + _turn_tool_count = sum( + 1 for m in messages + if isinstance(m, dict) and m.get("role") == "assistant" and m.get("tool_calls") + ) + _resp_len = len(final_response) if final_response else 0 + _budget_used = agent.iteration_budget.used if agent.iteration_budget else 0 + _budget_max = agent.iteration_budget.max_total if agent.iteration_budget else 0 + + _diag_msg = ( + "Turn ended: reason=%s model=%s api_calls=%d/%d budget=%d/%d " + "tool_turns=%d last_msg_role=%s response_len=%d session=%s" + ) + _diag_args = ( + _turn_exit_reason, agent.model, api_call_count, agent.max_iterations, + _budget_used, _budget_max, + _turn_tool_count, _last_msg_role, _resp_len, + agent.session_id or "none", + ) + + if _last_msg_role == "tool" and not interrupted: + # Agent was mid-work — this is the "just stops" case. + logger.warning( + "Turn ended with pending tool result (agent may appear stuck). " + + _diag_msg + " last_tool=%s", + *_diag_args, _last_tool_name, + ) + else: + logger.info(_diag_msg, *_diag_args) + + # File-mutation verifier footer. + # If one or more ``write_file`` / ``patch`` calls failed during this + # turn and were never superseded by a successful write to the same + # path, append an advisory footer to the assistant response. This + # catches the specific case — reported by Ben Eng (#15524-adjacent) + # — where a model issues a batch of parallel patches, half of them + # fail with "Could not find old_string", and the model summarises + # the turn claiming every file was edited. The user then has to + # manually run ``git status`` to catch the lie. With this footer + # the truth is surfaced on every turn, so over-claiming is + # structurally impossible past the model. + # + # Gate: only applied when a real text response exists for this + # turn and the user didn't interrupt. Empty/interrupted turns + # already have other surface text that shouldn't be augmented. + if final_response and not interrupted: + try: + _failed = getattr(agent, "_turn_failed_file_mutations", None) or {} + if _failed and agent._file_mutation_verifier_enabled(): + footer = agent._format_file_mutation_failure_footer(_failed) + if footer: + final_response = final_response.rstrip() + "\n\n" + footer + except Exception as _ver_err: + logger.debug("file-mutation verifier footer failed: %s", _ver_err) + + # Plugin hook: transform_llm_output + # Fired once per turn after the tool-calling loop completes. + # Plugins can transform the LLM's output text before it's returned. + # First hook to return a string wins; None/empty return leaves text unchanged. + if final_response and not interrupted: + try: + from hermes_cli.plugins import invoke_hook as _invoke_hook + _transform_results = _invoke_hook( + "transform_llm_output", + response_text=final_response, + session_id=agent.session_id or "", + model=agent.model, + platform=getattr(agent, "platform", None) or "", + ) + for _hook_result in _transform_results: + if isinstance(_hook_result, str) and _hook_result: + final_response = _hook_result + break # First non-empty string wins + except Exception as exc: + logger.warning("transform_llm_output hook failed: %s", exc) + + # Plugin hook: post_llm_call + # Fired once per turn after the tool-calling loop completes. + # Plugins can use this to persist conversation data (e.g. sync + # to an external memory system). + if final_response and not interrupted: + try: + from hermes_cli.plugins import invoke_hook as _invoke_hook + _invoke_hook( + "post_llm_call", + session_id=agent.session_id, + user_message=original_user_message, + assistant_response=final_response, + conversation_history=list(messages), + model=agent.model, + platform=getattr(agent, "platform", None) or "", + ) + except Exception as exc: + logger.warning("post_llm_call hook failed: %s", exc) + + # Extract reasoning from the CURRENT turn only. Walk backwards + # but stop at the user message that started this turn — anything + # earlier is from a prior turn and must not leak into the reasoning + # box (confusing stale display; #17055). Within the current turn + # we still want the *most recent* non-empty reasoning: many + # providers (Claude thinking, DeepSeek v4, Codex Responses) emit + # reasoning on the tool-call step and leave the final-answer step + # with reasoning=None, so picking only the last assistant would + # silently drop legitimate same-turn reasoning. + last_reasoning = None + for msg in reversed(messages): + if msg.get("role") == "user": + break # turn boundary — don't cross into prior turns + if msg.get("role") == "assistant" and msg.get("reasoning"): + last_reasoning = msg["reasoning"] + break + + # Build result with interrupt info if applicable + result = { + "final_response": final_response, + "last_reasoning": last_reasoning, + "messages": messages, + "api_calls": api_call_count, + "completed": completed, + "turn_exit_reason": _turn_exit_reason, + "partial": False, # True only when stopped due to invalid tool calls + "interrupted": interrupted, + "response_previewed": getattr(agent, "_response_was_previewed", False), + "model": agent.model, + "provider": agent.provider, + "base_url": agent.base_url, + "input_tokens": agent.session_input_tokens, + "output_tokens": agent.session_output_tokens, + "cache_read_tokens": agent.session_cache_read_tokens, + "cache_write_tokens": agent.session_cache_write_tokens, + "reasoning_tokens": agent.session_reasoning_tokens, + "prompt_tokens": agent.session_prompt_tokens, + "completion_tokens": agent.session_completion_tokens, + "total_tokens": agent.session_total_tokens, + "last_prompt_tokens": getattr(agent.context_compressor, "last_prompt_tokens", 0) or 0, + "estimated_cost_usd": agent.session_estimated_cost_usd, + "cost_status": agent.session_cost_status, + "cost_source": agent.session_cost_source, + } + if agent._tool_guardrail_halt_decision is not None: + result["guardrail"] = agent._tool_guardrail_halt_decision.to_metadata() + # If a /steer landed after the final assistant turn (no more tool + # batches to drain into), hand it back to the caller so it can be + # delivered as the next user turn instead of being silently lost. + _leftover_steer = agent._drain_pending_steer() + if _leftover_steer: + result["pending_steer"] = _leftover_steer + agent._response_was_previewed = False + + # Include interrupt message if one triggered the interrupt + if interrupted and agent._interrupt_message: + result["interrupt_message"] = agent._interrupt_message + + # Clear interrupt state after handling + agent.clear_interrupt() + + # Clear stream callback so it doesn't leak into future calls + agent._stream_callback = None + + # Check skill trigger NOW — based on how many tool iterations THIS turn used. + _should_review_skills = False + if (agent._skill_nudge_interval > 0 + and agent._iters_since_skill >= agent._skill_nudge_interval + and "skill_manage" in agent.valid_tool_names): + _should_review_skills = True + agent._iters_since_skill = 0 + + # External memory provider: sync the completed turn + queue next prefetch. + agent._sync_external_memory_for_turn( + original_user_message=original_user_message, + final_response=final_response, + interrupted=interrupted, + ) + + # Background memory/skill review — runs AFTER the response is delivered + # so it never competes with the user's task for model attention. + if final_response and not interrupted and (_should_review_memory or _should_review_skills): + try: + agent._spawn_background_review( + messages_snapshot=list(messages), + review_memory=_should_review_memory, + review_skills=_should_review_skills, + ) + except Exception: + pass # Background review is best-effort + + # Note: Memory provider on_session_end() + shutdown_all() are NOT + # called here — run_conversation() is called once per user message in + # multi-turn sessions. Shutting down after every turn would kill the + # provider before the second message. Actual session-end cleanup is + # handled by the CLI (atexit / /reset) and gateway (session expiry / + # _reset_session). + + # Plugin hook: on_session_end + # Fired at the very end of every run_conversation call. + # Plugins can use this for cleanup, flushing buffers, etc. + try: + from hermes_cli.plugins import invoke_hook as _invoke_hook + _invoke_hook( + "on_session_end", + session_id=agent.session_id, + completed=completed, + interrupted=interrupted, + model=agent.model, + platform=getattr(agent, "platform", None) or "", + ) + except Exception as exc: + logger.warning("on_session_end hook failed: %s", exc) + + return result + + + +__all__ = ["run_conversation"] diff --git a/agent/credential_pool.py b/agent/credential_pool.py index 7f27873a7fb..98dbaf30839 100644 --- a/agent/credential_pool.py +++ b/agent/credential_pool.py @@ -166,6 +166,8 @@ class PooledCredential: @property def runtime_api_key(self) -> str: if self.provider == "nous": + # Nous stores the runtime inference credential in agent_key for + # compatibility. It may be a NAS invoke JWT or legacy opaque key. return str(self.agent_key or self.access_token or "") return str(self.access_token or "") @@ -621,18 +623,35 @@ class CredentialPool: return entry store_refresh = state.get("refresh_token", "") store_access = state.get("access_token", "") - if store_refresh and store_refresh != entry.refresh_token: + comparable_updates = { + "access_token": store_access, + "refresh_token": store_refresh, + "expires_at": state.get("expires_at"), + "agent_key": state.get("agent_key"), + "agent_key_expires_at": state.get("agent_key_expires_at"), + "inference_base_url": state.get("inference_base_url"), + } + should_sync = any( + value not in (None, "") and getattr(entry, key, None) != value + for key, value in comparable_updates.items() + ) + if should_sync: logger.debug( - "Pool entry %s: syncing tokens from auth.json (Nous refresh token changed)", + "Pool entry %s: syncing Nous state from auth.json", entry.id, ) field_updates: Dict[str, Any] = { - "access_token": store_access, - "refresh_token": store_refresh, "last_status": None, "last_status_at": None, "last_error_code": None, + "last_error_reason": None, + "last_error_message": None, + "last_error_reset_at": None, } + if store_access: + field_updates["access_token"] = store_access + if store_refresh: + field_updates["refresh_token"] = store_refresh if state.get("expires_at"): field_updates["expires_at"] = state["expires_at"] if state.get("agent_key"): @@ -811,36 +830,15 @@ class CredentialPool: synced = self._sync_nous_entry_from_auth_store(entry) if synced is not entry: entry = synced - nous_state = { - "access_token": entry.access_token, - "refresh_token": entry.refresh_token, - "client_id": entry.client_id, - "portal_base_url": entry.portal_base_url, - "inference_base_url": entry.inference_base_url, - "token_type": entry.token_type, - "scope": entry.scope, - "obtained_at": entry.obtained_at, - "expires_at": entry.expires_at, - "agent_key": entry.agent_key, - "agent_key_expires_at": entry.agent_key_expires_at, - "tls": entry.tls, - } - refreshed = auth_mod.refresh_nous_oauth_from_state( - nous_state, + auth_mod.resolve_nous_runtime_credentials( min_key_ttl_seconds=DEFAULT_AGENT_KEY_MIN_TTL_SECONDS, - force_refresh=force, - force_mint=force, + inference_auth_mode=( + auth_mod.NOUS_INFERENCE_AUTH_MODE_LEGACY + if force + else auth_mod.NOUS_INFERENCE_AUTH_MODE_AUTO + ), ) - # Apply returned fields: dataclass fields via replace, extras via dict update - field_updates = {} - extra_updates = dict(entry.extra) - _field_names = {f.name for f in fields(entry)} - for k, v in refreshed.items(): - if k in _field_names: - field_updates[k] = v - elif k in _EXTRA_KEYS: - extra_updates[k] = v - updated = replace(entry, extra=extra_updates, **field_updates) + updated = self._sync_nous_entry_from_auth_store(entry) else: return entry except Exception as exc: @@ -929,6 +927,49 @@ class CredentialPool: self._persist() self._sync_device_code_entry_to_auth_store(updated) return updated + if auth_mod._is_terminal_nous_refresh_error(exc): + logger.debug("Nous refresh token is terminally invalid; clearing local token state") + try: + with _auth_store_lock(): + auth_store = _load_auth_store() + state = _load_provider_state(auth_store, "nous") or { + "client_id": entry.client_id, + "portal_base_url": entry.portal_base_url, + "inference_base_url": entry.inference_base_url, + "token_type": entry.token_type, + "scope": entry.scope, + "tls": entry.tls, + } + store_refresh = str(state.get("refresh_token") or "").strip() + entry_refresh = str(entry.refresh_token or "").strip() + if not store_refresh or store_refresh == entry_refresh: + auth_mod._quarantine_nous_oauth_state( + state, + exc, + reason="credential_pool_refresh_failure", + ) + auth_mod._quarantine_nous_pool_entries( + auth_store, + exc, + reason="credential_pool_refresh_failure", + ) + _save_provider_state(auth_store, "nous", state) + _save_auth_store(auth_store) + except Exception as clear_exc: + logger.debug("Failed to clear terminal Nous OAuth state: %s", clear_exc) + + singleton_sources = { + auth_mod.NOUS_DEVICE_CODE_SOURCE, + f"manual:{auth_mod.NOUS_DEVICE_CODE_SOURCE}", + } + self._entries = [ + item for item in self._entries + if item.source not in singleton_sources + ] + if self._current_id == entry.id: + self._current_id = None + self._persist() + return None self._mark_exhausted(entry, None) return None @@ -1365,7 +1406,22 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup elif provider == "nous": state = _load_provider_state(auth_store, "nous") - if state and not _is_suppressed(provider, "device_code"): + has_runtime_material = bool( + isinstance(state, dict) + and ( + str(state.get("access_token") or "").strip() + or str(state.get("agent_key") or "").strip() + ) + ) + if state and not has_runtime_material: + retained = [ + entry for entry in entries + if entry.source not in {"device_code", "manual:device_code"} + ] + if len(retained) != len(entries): + entries[:] = retained + changed = True + if state and has_runtime_material and not _is_suppressed(provider, "device_code"): active_sources.add("device_code") # Prefer a user-supplied label embedded in the singleton state # (set by persist_nous_credentials(label=...) when the user ran diff --git a/agent/iteration_budget.py b/agent/iteration_budget.py new file mode 100644 index 00000000000..213b97c0226 --- /dev/null +++ b/agent/iteration_budget.py @@ -0,0 +1,62 @@ +"""Per-agent iteration budget — thread-safe consume/refund counter. + +Extracted from ``run_agent.py``. Each ``AIAgent`` instance (parent or +subagent) holds an :class:`IterationBudget`; the parent's cap comes from +``max_iterations`` (default 90), each subagent's cap comes from +``delegation.max_iterations`` (default 50). + +``run_agent`` re-exports ``IterationBudget`` so existing +``from run_agent import IterationBudget`` imports keep working unchanged. +""" + +from __future__ import annotations + +import threading + + +class IterationBudget: + """Thread-safe iteration counter for an agent. + + Each agent (parent or subagent) gets its own ``IterationBudget``. + The parent's budget is capped at ``max_iterations`` (default 90). + Each subagent gets an independent budget capped at + ``delegation.max_iterations`` (default 50) — this means total + iterations across parent + subagents can exceed the parent's cap. + Users control the per-subagent limit via ``delegation.max_iterations`` + in config.yaml. + + ``execute_code`` (programmatic tool calling) iterations are refunded via + :meth:`refund` so they don't eat into the budget. + """ + + def __init__(self, max_total: int): + self.max_total = max_total + self._used = 0 + self._lock = threading.Lock() + + def consume(self) -> bool: + """Try to consume one iteration. Returns True if allowed.""" + with self._lock: + if self._used >= self.max_total: + return False + self._used += 1 + return True + + def refund(self) -> None: + """Give back one iteration (e.g. for execute_code turns).""" + with self._lock: + if self._used > 0: + self._used -= 1 + + @property + def used(self) -> int: + with self._lock: + return self._used + + @property + def remaining(self) -> int: + with self._lock: + return max(0, self.max_total - self._used) + + +__all__ = ["IterationBudget"] diff --git a/agent/lsp/client.py b/agent/lsp/client.py index 8f380fc7a60..06a92ae351b 100644 --- a/agent/lsp/client.py +++ b/agent/lsp/client.py @@ -232,7 +232,7 @@ class LSPClient: the process is killed and the client is left in state ``"error"`` — re-call ``start()`` to retry. """ - if self._state in ("running", "starting"): + if self._state in {"running", "starting"}: return self._state = "starting" try: diff --git a/agent/lsp/install.py b/agent/lsp/install.py index 0aaa22be744..d4a80ec195e 100644 --- a/agent/lsp/install.py +++ b/agent/lsp/install.py @@ -151,7 +151,7 @@ def try_install(pkg: str, strategy: str = "auto") -> Optional[str]: same path (or ``None``) without reinstalling. Concurrent calls are serialized. """ - if strategy not in ("auto",): + if strategy not in {"auto",}: # Only ``auto`` triggers an actual install. In manual/off, # we still check whether the binary already exists. recipe = INSTALL_RECIPES.get(pkg, {}) diff --git a/agent/lsp/manager.py b/agent/lsp/manager.py index 7f5feaa170f..4f16188de0b 100644 --- a/agent/lsp/manager.py +++ b/agent/lsp/manager.py @@ -162,7 +162,7 @@ class LSPService: idle_timeout: float = DEFAULT_IDLE_TIMEOUT, ) -> None: self._enabled = enabled - self._wait_mode = wait_mode if wait_mode in ("document", "full") else "document" + self._wait_mode = wait_mode if wait_mode in {"document", "full"} else "document" self._wait_timeout = wait_timeout self._install_strategy = install_strategy self._binary_overrides = binary_overrides or {} diff --git a/agent/lsp/reporter.py b/agent/lsp/reporter.py index fedad0d19b3..0eba96ba1ff 100644 --- a/agent/lsp/reporter.py +++ b/agent/lsp/reporter.py @@ -28,7 +28,7 @@ def format_diagnostic(d: Dict[str, Any]) -> str: col = int(start.get("character", 0)) + 1 msg = str(d.get("message") or "").rstrip() code = d.get("code") - code_part = f" [{code}]" if code not in (None, "") else "" + code_part = f" [{code}]" if code not in {None, ""} else "" source = d.get("source") source_part = f" ({source})" if source else "" return f"{sev} [{line}:{col}] {msg}{code_part}{source_part}" diff --git a/agent/lsp/servers.py b/agent/lsp/servers.py index 00ad4c40005..144b5cb2c11 100644 --- a/agent/lsp/servers.py +++ b/agent/lsp/servers.py @@ -237,7 +237,7 @@ def _spawn_pyright(root: str, ctx: ServerContext) -> Optional[SpawnSpec]: return None # If we got the cli ``pyright``, the langserver is its sibling. base = os.path.basename(bin_path) - if base in ("pyright", "pyright.exe"): + if base in {"pyright", "pyright.exe"}: sibling = os.path.join(os.path.dirname(bin_path), "pyright-langserver") if os.path.exists(sibling): bin_path = sibling diff --git a/agent/message_sanitization.py b/agent/message_sanitization.py new file mode 100644 index 00000000000..ff53d247a84 --- /dev/null +++ b/agent/message_sanitization.py @@ -0,0 +1,444 @@ +"""Message and tool-payload sanitization helpers. + +Pure functions extracted from ``run_agent.py`` so the AIAgent module can +stay focused on the conversation loop. These walk OpenAI-format message +lists and structured payloads, repairing or stripping problematic +characters that would otherwise crash ``json.dumps`` inside the OpenAI +SDK or be rejected by upstream APIs. + +All helpers are stateless and side-effect-free except for in-place +mutation of their input (where documented). Backward-compatible +re-exports from ``run_agent`` remain in place so existing imports +``from run_agent import _sanitize_surrogates`` keep working. +""" + +from __future__ import annotations + +import json +import logging +import re +from typing import Any + +logger = logging.getLogger(__name__) + +# Lone surrogate code points are invalid in UTF-8 and crash json.dumps +# inside the OpenAI SDK. Used by every surrogate-sanitization helper +# below as well as by run_agent and the CLI for paste-from-clipboard +# scrubbing. +_SURROGATE_RE = re.compile(r'[\ud800-\udfff]') + + +def _sanitize_surrogates(text: str) -> str: + """Replace lone surrogate code points with U+FFFD (replacement character). + + Surrogates are invalid in UTF-8 and will crash ``json.dumps()`` inside the + OpenAI SDK. This is a fast no-op when the text contains no surrogates. + """ + if _SURROGATE_RE.search(text): + return _SURROGATE_RE.sub('\ufffd', text) + return text + + +def _sanitize_structure_surrogates(payload: Any) -> bool: + """Replace surrogate code points in nested dict/list payloads in-place. + + Mirror of ``_sanitize_structure_non_ascii`` but for surrogate recovery. + Used to scrub nested structured fields (e.g. ``reasoning_details`` — an + array of dicts with ``summary``/``text`` strings) that flat per-field + checks don't reach. Returns True if any surrogates were replaced. + """ + found = False + + def _walk(node): + nonlocal found + if isinstance(node, dict): + for key, value in node.items(): + if isinstance(value, str): + if _SURROGATE_RE.search(value): + node[key] = _SURROGATE_RE.sub('\ufffd', value) + found = True + elif isinstance(value, (dict, list)): + _walk(value) + elif isinstance(node, list): + for idx, value in enumerate(node): + if isinstance(value, str): + if _SURROGATE_RE.search(value): + node[idx] = _SURROGATE_RE.sub('\ufffd', value) + found = True + elif isinstance(value, (dict, list)): + _walk(value) + + _walk(payload) + return found + + +def _sanitize_messages_surrogates(messages: list) -> bool: + """Sanitize surrogate characters from all string content in a messages list. + + Walks message dicts in-place. Returns True if any surrogates were found + and replaced, False otherwise. Covers content/text, name, tool call + metadata/arguments, AND any additional string or nested structured fields + (``reasoning``, ``reasoning_content``, ``reasoning_details``, etc.) so + retries don't fail on a non-content field. Byte-level reasoning models + (xiaomi/mimo, kimi, glm) can emit lone surrogates in reasoning output + that flow through to ``api_messages["reasoning_content"]`` on the next + turn and crash json.dumps inside the OpenAI SDK. + """ + found = False + for msg in messages: + if not isinstance(msg, dict): + continue + content = msg.get("content") + if isinstance(content, str) and _SURROGATE_RE.search(content): + msg["content"] = _SURROGATE_RE.sub('\ufffd', content) + found = True + elif isinstance(content, list): + for part in content: + if isinstance(part, dict): + text = part.get("text") + if isinstance(text, str) and _SURROGATE_RE.search(text): + part["text"] = _SURROGATE_RE.sub('\ufffd', text) + found = True + name = msg.get("name") + if isinstance(name, str) and _SURROGATE_RE.search(name): + msg["name"] = _SURROGATE_RE.sub('\ufffd', name) + found = True + tool_calls = msg.get("tool_calls") + if isinstance(tool_calls, list): + for tc in tool_calls: + if not isinstance(tc, dict): + continue + tc_id = tc.get("id") + if isinstance(tc_id, str) and _SURROGATE_RE.search(tc_id): + tc["id"] = _SURROGATE_RE.sub('\ufffd', tc_id) + found = True + fn = tc.get("function") + if isinstance(fn, dict): + fn_name = fn.get("name") + if isinstance(fn_name, str) and _SURROGATE_RE.search(fn_name): + fn["name"] = _SURROGATE_RE.sub('\ufffd', fn_name) + found = True + fn_args = fn.get("arguments") + if isinstance(fn_args, str) and _SURROGATE_RE.search(fn_args): + fn["arguments"] = _SURROGATE_RE.sub('\ufffd', fn_args) + found = True + # Walk any additional string / nested fields (reasoning, + # reasoning_content, reasoning_details, etc.) — surrogates from + # byte-level reasoning models (xiaomi/mimo, kimi, glm) can lurk + # in these fields and aren't covered by the per-field checks above. + # Matches _sanitize_messages_non_ascii's coverage (PR #10537). + for key, value in msg.items(): + if key in {"content", "name", "tool_calls", "role"}: + continue + if isinstance(value, str): + if _SURROGATE_RE.search(value): + msg[key] = _SURROGATE_RE.sub('\ufffd', value) + found = True + elif isinstance(value, (dict, list)): + if _sanitize_structure_surrogates(value): + found = True + return found + + +def _escape_invalid_chars_in_json_strings(raw: str) -> str: + """Escape unescaped control chars inside JSON string values. + + Walks the raw JSON character-by-character, tracking whether we are + inside a double-quoted string. Inside strings, replaces literal + control characters (0x00-0x1F) that aren't already part of an escape + sequence with their ``\\uXXXX`` equivalents. Pass-through for everything + else. + + Ported from #12093 — complements the other repair passes in + ``_repair_tool_call_arguments`` when ``json.loads(strict=False)`` is + not enough (e.g. llama.cpp backends that emit literal apostrophes or + tabs alongside other malformations). + """ + out: list[str] = [] + in_string = False + i = 0 + n = len(raw) + while i < n: + ch = raw[i] + if in_string: + if ch == "\\" and i + 1 < n: + # Already-escaped char — pass through as-is + out.append(ch) + out.append(raw[i + 1]) + i += 2 + continue + if ch == '"': + in_string = False + out.append(ch) + elif ord(ch) < 0x20: + out.append(f"\\u{ord(ch):04x}") + else: + out.append(ch) + else: + if ch == '"': + in_string = True + out.append(ch) + i += 1 + return "".join(out) + + +def _repair_tool_call_arguments(raw_args: str, tool_name: str = "?") -> str: + """Attempt to repair malformed tool_call argument JSON. + + Models like GLM-5.1 via Ollama can produce truncated JSON, trailing + commas, Python ``None``, etc. The API proxy rejects these with HTTP 400 + "invalid tool call arguments". This function applies common repairs; + if all fail it returns ``"{}"`` so the request succeeds (better than + crashing the session). All repairs are logged at WARNING level. + """ + raw_stripped = raw_args.strip() if isinstance(raw_args, str) else "" + + # Fast-path: empty / whitespace-only -> empty object + if not raw_stripped: + logger.warning("Sanitized empty tool_call arguments for %s", tool_name) + return "{}" + + # Python-literal None -> normalise to {} + if raw_stripped == "None": + logger.warning("Sanitized Python-None tool_call arguments for %s", tool_name) + return "{}" + + # Repair pass 0: llama.cpp backends sometimes emit literal control + # characters (tabs, newlines) inside JSON string values. json.loads + # with strict=False accepts these and lets us re-serialise the + # result into wire-valid JSON without any string surgery. This is + # the most common local-model repair case (#12068). + try: + parsed = json.loads(raw_stripped, strict=False) + reserialised = json.dumps(parsed, separators=(",", ":")) + if reserialised != raw_stripped: + logger.warning( + "Repaired unescaped control chars in tool_call arguments for %s", + tool_name, + ) + return reserialised + except (json.JSONDecodeError, TypeError, ValueError): + pass + + # Attempt common JSON repairs + fixed = raw_stripped + # 1. Strip trailing commas before } or ] + fixed = re.sub(r',\s*([}\]])', r'\1', fixed) + # 2. Close unclosed structures + open_curly = fixed.count('{') - fixed.count('}') + open_bracket = fixed.count('[') - fixed.count(']') + if open_curly > 0: + fixed += '}' * open_curly + if open_bracket > 0: + fixed += ']' * open_bracket + # 3. Remove excess closing braces/brackets (bounded to 50 iterations) + for _ in range(50): + try: + json.loads(fixed) + break + except json.JSONDecodeError: + if fixed.endswith('}') and fixed.count('}') > fixed.count('{'): + fixed = fixed[:-1] + elif fixed.endswith(']') and fixed.count(']') > fixed.count('['): + fixed = fixed[:-1] + else: + break + + try: + json.loads(fixed) + logger.warning( + "Repaired malformed tool_call arguments for %s: %s → %s", + tool_name, raw_stripped[:80], fixed[:80], + ) + return fixed + except json.JSONDecodeError: + pass + + # Repair pass 4: escape unescaped control chars inside JSON strings, + # then retry. Catches cases where strict=False alone fails because + # other malformations are present too. + try: + escaped = _escape_invalid_chars_in_json_strings(fixed) + if escaped != fixed: + json.loads(escaped) + logger.warning( + "Repaired control-char-laced tool_call arguments for %s: %s → %s", + tool_name, raw_stripped[:80], escaped[:80], + ) + return escaped + except (json.JSONDecodeError, TypeError, ValueError): + pass + + # Last resort: replace with empty object so the API request doesn't + # crash the entire session. + logger.warning( + "Unrepairable tool_call arguments for %s — " + "replaced with empty object (was: %s)", + tool_name, raw_stripped[:80], + ) + return "{}" + + +def _strip_non_ascii(text: str) -> str: + """Remove non-ASCII characters, replacing with closest ASCII equivalent or removing. + + Used as a last resort when the system encoding is ASCII and can't handle + any non-ASCII characters (e.g. LANG=C on Chromebooks). + """ + return text.encode('ascii', errors='ignore').decode('ascii') + + +def _sanitize_messages_non_ascii(messages: list) -> bool: + """Strip non-ASCII characters from all string content in a messages list. + + This is a last-resort recovery for systems with ASCII-only encoding + (LANG=C, Chromebooks, minimal containers). Returns True if any + non-ASCII content was found and sanitized. + """ + found = False + for msg in messages: + if not isinstance(msg, dict): + continue + # Sanitize content (string) + content = msg.get("content") + if isinstance(content, str): + sanitized = _strip_non_ascii(content) + if sanitized != content: + msg["content"] = sanitized + found = True + elif isinstance(content, list): + for part in content: + if isinstance(part, dict): + text = part.get("text") + if isinstance(text, str): + sanitized = _strip_non_ascii(text) + if sanitized != text: + part["text"] = sanitized + found = True + # Sanitize name field (can contain non-ASCII in tool results) + name = msg.get("name") + if isinstance(name, str): + sanitized = _strip_non_ascii(name) + if sanitized != name: + msg["name"] = sanitized + found = True + # Sanitize tool_calls + tool_calls = msg.get("tool_calls") + if isinstance(tool_calls, list): + for tc in tool_calls: + if isinstance(tc, dict): + fn = tc.get("function", {}) + if isinstance(fn, dict): + fn_args = fn.get("arguments") + if isinstance(fn_args, str): + sanitized = _strip_non_ascii(fn_args) + if sanitized != fn_args: + fn["arguments"] = sanitized + found = True + # Sanitize any additional top-level string fields (e.g. reasoning_content) + for key, value in msg.items(): + if key in {"content", "name", "tool_calls", "role"}: + continue + if isinstance(value, str): + sanitized = _strip_non_ascii(value) + if sanitized != value: + msg[key] = sanitized + found = True + return found + + +def _sanitize_tools_non_ascii(tools: list) -> bool: + """Strip non-ASCII characters from tool payloads in-place.""" + return _sanitize_structure_non_ascii(tools) + + +def _strip_images_from_messages(messages: list) -> bool: + """Remove image_url content parts from all messages in-place. + + Called when a server signals it does not support images (e.g. + "Only 'text' content type is supported."). Mutates messages so the + next API call sends text only. + + Preserves message alternation invariants: + * ``tool``-role messages whose content was entirely images are replaced + with a plaintext placeholder, NOT deleted — deleting them would leave + the paired ``tool_call_id`` on the prior assistant message unmatched, + which providers reject with HTTP 400. + * Non-tool messages whose content becomes empty are dropped. In + practice this only hits synthetic image-only user messages appended + for attachment delivery; real user turns always include text. + + Returns True if any image parts were removed. + """ + found = False + to_delete = [] + for i, msg in enumerate(messages): + if not isinstance(msg, dict): + continue + content = msg.get("content") + if not isinstance(content, list): + continue + new_parts = [] + for part in content: + if isinstance(part, dict) and part.get("type") in {"image_url", "image", "input_image"}: + found = True + else: + new_parts.append(part) + if len(new_parts) < len(content): + if new_parts: + msg["content"] = new_parts + elif msg.get("role") == "tool": + # Preserve tool_call_id linkage — providers require every + # assistant tool_call to have a matching tool response. + msg["content"] = "[image content removed — server does not support images]" + else: + # Synthetic image-only user/assistant message with no text; + # safe to drop. + to_delete.append(i) + for i in reversed(to_delete): + del messages[i] + return found + + +def _sanitize_structure_non_ascii(payload: Any) -> bool: + """Strip non-ASCII characters from nested dict/list payloads in-place.""" + found = False + + def _walk(node): + nonlocal found + if isinstance(node, dict): + for key, value in node.items(): + if isinstance(value, str): + sanitized = _strip_non_ascii(value) + if sanitized != value: + node[key] = sanitized + found = True + elif isinstance(value, (dict, list)): + _walk(value) + elif isinstance(node, list): + for idx, value in enumerate(node): + if isinstance(value, str): + sanitized = _strip_non_ascii(value) + if sanitized != value: + node[idx] = sanitized + found = True + elif isinstance(value, (dict, list)): + _walk(value) + + _walk(payload) + return found + + +__all__ = [ + "_SURROGATE_RE", + "_sanitize_surrogates", + "_sanitize_structure_surrogates", + "_sanitize_messages_surrogates", + "_escape_invalid_chars_in_json_strings", + "_repair_tool_call_arguments", + "_strip_non_ascii", + "_sanitize_messages_non_ascii", + "_sanitize_tools_non_ascii", + "_strip_images_from_messages", + "_sanitize_structure_non_ascii", +] diff --git a/agent/model_metadata.py b/agent/model_metadata.py index 26a844ccb92..b8ec0d6509e 100644 --- a/agent/model_metadata.py +++ b/agent/model_metadata.py @@ -194,6 +194,7 @@ DEFAULT_CONTEXT_LENGTHS = { "llama": 131072, # Qwen — specific model families before the catch-all. # Official docs: https://help.aliyun.com/zh/model-studio/developer-reference/ + "qwen3.6-plus": 1048576, # 1M context (DashScope/Alibaba & OpenRouter) "qwen3-coder-plus": 1000000, # 1M context "qwen3-coder": 262144, # 256K context "qwen": 131072, diff --git a/agent/process_bootstrap.py b/agent/process_bootstrap.py new file mode 100644 index 00000000000..fdd9053f5d8 --- /dev/null +++ b/agent/process_bootstrap.py @@ -0,0 +1,167 @@ +"""Process-level bootstrap helpers for ``run_agent``. + +Three concerns, all tied to ``AIAgent`` boot-time / runtime IO setup: + +1. **Lazy OpenAI SDK import** — ``_load_openai_cls`` + ``_OpenAIProxy`` + defer the 240ms-ish ``from openai import OpenAI`` cost until first use, + while preserving ``isinstance(client, OpenAI)`` checks and + ``patch("run_agent.OpenAI", ...)`` test patterns. + +2. **Crash-resistant stdio** — ``_SafeWriter`` wraps stdout/stderr so + ``OSError: Input/output error`` from broken pipes (systemd, Docker, + thread teardown races) cannot crash the agent. ``_install_safe_stdio`` + applies the wrapper. + +3. **HTTP proxy resolution** — ``_get_proxy_from_env`` reads + ``HTTPS_PROXY`` / ``HTTP_PROXY`` / ``ALL_PROXY``; + ``_get_proxy_for_base_url`` respects ``NO_PROXY`` for the given base URL. + +``run_agent`` re-exports every name so existing +``from run_agent import _get_proxy_from_env`` imports keep working +unchanged. +""" + +from __future__ import annotations + +import os +import sys +import urllib.request +from typing import Optional + +from utils import base_url_hostname, normalize_proxy_url + + +# Cached at module level so we only pay the OpenAI SDK import cost once +# per process (after the first lazy load). +_OPENAI_CLS_CACHE = None + + +def _load_openai_cls() -> type: + """Import and cache ``openai.OpenAI``.""" + global _OPENAI_CLS_CACHE + if _OPENAI_CLS_CACHE is None: + from openai import OpenAI as _cls + _OPENAI_CLS_CACHE = _cls + return _OPENAI_CLS_CACHE + + +class _OpenAIProxy: + """Module-level proxy that looks like ``openai.OpenAI`` but imports lazily.""" + + __slots__ = () + + def __call__(self, *args, **kwargs): + return _load_openai_cls()(*args, **kwargs) + + def __instancecheck__(self, obj): + return isinstance(obj, _load_openai_cls()) + + def __repr__(self): + return "" + + +class _SafeWriter: + """Transparent stdio wrapper that catches OSError/ValueError from broken pipes. + + When hermes-agent runs as a systemd service, Docker container, or headless + daemon, the stdout/stderr pipe can become unavailable (idle timeout, buffer + exhaustion, socket reset). Any print() call then raises + ``OSError: [Errno 5] Input/output error``, which can crash agent setup or + run_conversation() — especially via double-fault when an except handler + also tries to print. + + Additionally, when subagents run in ThreadPoolExecutor threads, the shared + stdout handle can close between thread teardown and cleanup, raising + ``ValueError: I/O operation on closed file`` instead of OSError. + + This wrapper delegates all writes to the underlying stream and silently + catches both OSError and ValueError. It is transparent when the wrapped + stream is healthy. + """ + + __slots__ = ("_inner",) + + def __init__(self, inner): + object.__setattr__(self, "_inner", inner) + + def write(self, data): + try: + return self._inner.write(data) + except (OSError, ValueError): + return len(data) if isinstance(data, str) else 0 + + def flush(self): + try: + self._inner.flush() + except (OSError, ValueError): + pass + + def fileno(self): + return self._inner.fileno() + + def isatty(self): + try: + return self._inner.isatty() + except (OSError, ValueError): + return False + + def __getattr__(self, name): + return getattr(self._inner, name) + + +def _get_proxy_from_env() -> Optional[str]: + """Read proxy URL from environment variables. + + Checks HTTPS_PROXY, HTTP_PROXY, ALL_PROXY (and lowercase variants) in order. + Returns the first valid proxy URL found, or None if no proxy is configured. + """ + for key in ("HTTPS_PROXY", "HTTP_PROXY", "ALL_PROXY", + "https_proxy", "http_proxy", "all_proxy"): + value = os.environ.get(key, "").strip() + if value: + return normalize_proxy_url(value) + return None + + +def _get_proxy_for_base_url(base_url: Optional[str]) -> Optional[str]: + """Return an env-configured proxy unless NO_PROXY excludes this base URL.""" + proxy = _get_proxy_from_env() + if not proxy or not base_url: + return proxy + + host = base_url_hostname(base_url) + if not host: + return proxy + + try: + if urllib.request.proxy_bypass_environment(host): + return None + except Exception: + pass + + return proxy + + +def _install_safe_stdio() -> None: + """Wrap stdout/stderr so best-effort console output cannot crash the agent.""" + for stream_name in ("stdout", "stderr"): + stream = getattr(sys, stream_name, None) + if stream is not None and not isinstance(stream, _SafeWriter): + setattr(sys, stream_name, _SafeWriter(stream)) + + +# Module-level proxy instance — drops in for ``openai.OpenAI``. Imported as +# ``from agent.process_bootstrap import OpenAI`` (or re-exported via +# ``run_agent`` for legacy tests). +OpenAI = _OpenAIProxy() + + +__all__ = [ + "OpenAI", + "_OpenAIProxy", + "_load_openai_cls", + "_SafeWriter", + "_install_safe_stdio", + "_get_proxy_from_env", + "_get_proxy_for_base_url", +] diff --git a/agent/shell_hooks.py b/agent/shell_hooks.py index bad5388f88b..79d494d7dcb 100644 --- a/agent/shell_hooks.py +++ b/agent/shell_hooks.py @@ -83,6 +83,7 @@ logger = logging.getLogger(__name__) DEFAULT_TIMEOUT_SECONDS = 60 MAX_TIMEOUT_SECONDS = 300 ALLOWLIST_FILENAME = "shell-hooks-allowlist.json" +_DEFAULT_BLOCK_MESSAGE = "Blocked by shell hook." # (event, matcher, command) triples that have been wired to the plugin # manager in the current process. Matcher is part of the key because @@ -481,6 +482,17 @@ def _serialize_payload(event: str, kwargs: Dict[str, Any]) -> str: return json.dumps(payload, ensure_ascii=False, default=str) +def _block_message(primary: Any, secondary: Any) -> str: + """Return a validated string block message, falling back to the default. + + Accepts two candidate fields (primary wins over secondary) so callers + can express field-priority differences between the two hook wire formats + without duplicating the type-check logic. + """ + raw = primary or secondary + return raw if isinstance(raw, str) and raw else _DEFAULT_BLOCK_MESSAGE + + def _parse_response(event: str, stdout: str) -> Optional[Dict[str, Any]]: """Translate stdout JSON into a Hermes wire-shape dict. @@ -515,13 +527,9 @@ def _parse_response(event: str, stdout: str) -> Optional[Dict[str, Any]]: if event == "pre_tool_call": if data.get("action") == "block": - message = data.get("message") or data.get("reason") or "" - if isinstance(message, str) and message: - return {"action": "block", "message": message} + return {"action": "block", "message": _block_message(data.get("message"), data.get("reason"))} if data.get("decision") == "block": - message = data.get("reason") or data.get("message") or "" - if isinstance(message, str) and message: - return {"action": "block", "message": message} + return {"action": "block", "message": _block_message(data.get("reason"), data.get("message"))} return None context = data.get("context") diff --git a/agent/stream_diag.py b/agent/stream_diag.py new file mode 100644 index 00000000000..c4d8c54f470 --- /dev/null +++ b/agent/stream_diag.py @@ -0,0 +1,280 @@ +"""Stream diagnostics — per-attempt counters, exception chains, retry logging. + +When a streaming chat-completions request dies mid-response, we want to +know why: which Cloudflare edge served the request, which OpenRouter +downstream provider answered, how many bytes/chunks we got before the +drop, the HTTP status, the underlying httpx error class. These helpers +collect that info and emit it both to ``agent.log`` (full detail) and to +the user-facing status line (compact). + +All helpers are extracted from :class:`AIAgent` for cleanliness. +``run_agent`` keeps thin forwarder methods so existing call sites and +tests that patch ``run_agent.`` keep working. +""" + +from __future__ import annotations + +import logging +import time +from typing import Any, Dict, List, Optional + +logger = logging.getLogger(__name__) + + +# Per-attempt stream diagnostic headers. Lowercased; httpx returns +# CIMultiDict so case-insensitive lookups already work, but we read .get() +# on the dict from agent.log for free-form post-hoc analysis. +STREAM_DIAG_HEADERS = ( + "cf-ray", + "cf-cache-status", + "x-openrouter-provider", + "x-openrouter-model", + "x-openrouter-id", + "x-request-id", + "x-vercel-id", + "via", + "server", + "x-forwarded-for", +) + + +def stream_diag_init() -> Dict[str, Any]: + """Return a fresh per-attempt diagnostic dict. + + Mutated in-place by the streaming functions and read from the retry + block when a stream dies. Lives on ``request_client_holder`` so it + survives across the closure boundary. + """ + return { + "started_at": time.time(), + "first_chunk_at": None, + "chunks": 0, + "bytes": 0, + "headers": {}, + "http_status": None, + } + + +def stream_diag_capture_response(agent: Any, diag: Dict[str, Any], http_response: Any) -> None: + """Snapshot interesting headers + HTTP status from the live stream. + + Called once at stream open (before iterating chunks) so the metadata + survives even if the stream dies before any chunk arrives. Failures + are swallowed — diag is best-effort. + """ + if http_response is None or not isinstance(diag, dict): + return + try: + diag["http_status"] = getattr(http_response, "status_code", None) + except Exception: + pass + try: + headers = getattr(http_response, "headers", None) or {} + captured: Dict[str, str] = {} + # Allow per-agent override of the headers list (back-compat). + target_headers = getattr(agent, "_STREAM_DIAG_HEADERS", STREAM_DIAG_HEADERS) + for name in target_headers: + try: + val = headers.get(name) + if val: + # Truncate single-value to keep log lines bounded. + captured[name] = str(val)[:120] + except Exception: + continue + diag["headers"] = captured + except Exception: + pass + + +def flatten_exception_chain(error: BaseException) -> str: + """Return a compact ``Outer(msg) <- Inner(msg) <- ...`` rendering. + + OpenAI SDK wraps httpx errors as ``APIConnectionError`` / + ``APIError`` and only the wrapper's class is visible at the catch + site — but the underlying ``RemoteProtocolError`` / + ``ConnectError`` / ``ReadError`` is what tells us WHY the stream + died. Walks ``__cause__`` then ``__context__`` (deduped, max 4 + deep) to surface the chain in one line. + """ + seen: List[BaseException] = [] + link: Optional[BaseException] = error + while link is not None and len(seen) < 4: + if link in seen: + break + seen.append(link) + nxt = getattr(link, "__cause__", None) or getattr( + link, "__context__", None + ) + if nxt is None or nxt is link: + break + link = nxt + parts: List[str] = [] + for e in seen: + msg = str(e).strip().replace("\n", " ") + if len(msg) > 140: + msg = msg[:140] + "…" + parts.append(f"{type(e).__name__}({msg})" if msg else type(e).__name__) + return " <- ".join(parts) if parts else type(error).__name__ + + +def log_stream_retry( + agent: Any, + *, + kind: str, + error: BaseException, + attempt: int, + max_attempts: int, + mid_tool_call: bool, + diag: Optional[Dict[str, Any]] = None, +) -> None: + """Record a transient stream-drop and retry to ``agent.log``. + + Always logs a structured WARNING so users have a breadcrumb regardless + of UI verbosity. Subagents in particular benefit because their + retries no longer spam the parent's terminal — but the file log keeps + full detail (provider, error class, attempt, base_url, subagent_id). + + When *diag* is provided (the per-attempt stream-diagnostic dict from + :func:`stream_diag_init`), the WARNING also captures upstream headers + (cf-ray, x-openrouter-provider, x-openrouter-id), HTTP status, bytes + streamed before the drop, and elapsed time on the dying attempt. + These are the breadcrumbs needed to answer "is one CF edge / one + downstream provider responsible, or is it random across runs?" + """ + try: + try: + _summary = agent._summarize_api_error(error) + except Exception: + _summary = str(error) + if _summary and len(_summary) > 240: + _summary = _summary[:240] + "…" + + # Inner-cause chain (httpx errors hide under openai.APIError). + try: + _chain = flatten_exception_chain(error) + except Exception: + _chain = type(error).__name__ + + # Per-attempt counters and upstream headers. + _now = time.time() + _bytes = 0 + _chunks = 0 + _elapsed = 0.0 + _ttfb = None + _headers_repr = "-" + _http_status = "-" + if isinstance(diag, dict): + try: + _bytes = int(diag.get("bytes") or 0) + _chunks = int(diag.get("chunks") or 0) + _started = float(diag.get("started_at") or _now) + _elapsed = max(0.0, _now - _started) + _first = diag.get("first_chunk_at") + if _first is not None: + _ttfb = max(0.0, float(_first) - _started) + headers = diag.get("headers") or {} + if isinstance(headers, dict) and headers: + _headers_repr = " ".join( + f"{k}={v}" for k, v in headers.items() + ) + if diag.get("http_status") is not None: + _http_status = str(diag.get("http_status")) + except Exception: + pass + + logger.warning( + "Stream %s on attempt %s/%s — retrying. " + "subagent_id=%s depth=%s provider=%s base_url=%s " + "error_type=%s error=%s " + "chain=%s " + "http_status=%s bytes=%d chunks=%d elapsed=%.2fs ttfb=%s " + "upstream=[%s]", + kind, + attempt, + max_attempts, + getattr(agent, "_subagent_id", None) or "-", + getattr(agent, "_delegate_depth", 0), + agent.provider or "-", + agent.base_url or "-", + type(error).__name__, + _summary, + _chain, + _http_status, + _bytes, + _chunks, + _elapsed, + f"{_ttfb:.2f}s" if _ttfb is not None else "-", + _headers_repr, + extra={"mid_tool_call": mid_tool_call}, + ) + except Exception: + logger.debug("stream-retry log emit failed", exc_info=True) + + +def emit_stream_drop( + agent: Any, + *, + error: BaseException, + attempt: int, + max_attempts: int, + mid_tool_call: bool, + diag: Optional[Dict[str, Any]] = None, +) -> None: + """Emit a single user-visible line for a stream drop+retry. + + Both top-level agents and subagents announce drops in the UI — the + parent prefixes subagent lines with ``[subagent-N]`` via ``log_prefix`` + so they're easy to attribute. All cases also write a structured + WARNING to ``agent.log`` via :func:`log_stream_retry` with the full + diagnostic detail (subagent_id, provider, base_url, error_type, + cf-ray, x-openrouter-provider, bytes/chunks, elapsed) for post-hoc + analysis. + + The user-visible status line is intentionally compact: provider, + error class, attempt N/M, plus ``after Xs`` when the stream dropped + mid-flight. Full diagnostic detail goes to ``agent.log`` only — + ``hermes logs --level WARNING | grep "Stream drop"`` to inspect. + """ + kind = "drop mid tool-call" if mid_tool_call else "drop" + log_stream_retry( + agent, + kind=kind, + error=error, + attempt=attempt, + max_attempts=max_attempts, + mid_tool_call=mid_tool_call, + diag=diag, + ) + provider = agent.provider or "provider" + # Compose a brief "after Xs" suffix when we have timing data — helps + # the user distinguish "couldn't connect" (0s) from "died after 30s + # of streaming" (likely upstream idle-kill or proxy timeout). + _suffix = "" + if isinstance(diag, dict): + try: + started = diag.get("started_at") + if started is not None: + _suffix = f" after {max(0.0, time.time() - float(started)):.1f}s" + except Exception: + pass + try: + agent._emit_status( + f"⚠️ {provider} stream {kind} ({type(error).__name__}){_suffix} " + f"— reconnecting, retry {attempt}/{max_attempts}" + ) + agent._touch_activity( + f"stream retry {attempt}/{max_attempts} " + f"after {type(error).__name__}" + ) + except Exception: + pass + + +__all__ = [ + "STREAM_DIAG_HEADERS", + "stream_diag_init", + "stream_diag_capture_response", + "flatten_exception_chain", + "log_stream_retry", + "emit_stream_drop", +] diff --git a/agent/system_prompt.py b/agent/system_prompt.py new file mode 100644 index 00000000000..52a574101f5 --- /dev/null +++ b/agent/system_prompt.py @@ -0,0 +1,333 @@ +"""System-prompt assembly for :class:`AIAgent`. + +The agent's system prompt is built once per session and reused across all +turns — only context compression triggers a rebuild. This keeps the +upstream prefix cache warm. See ``hermes-agent-dev``'s +``references/system-prompt-invariant.md`` for the invariants and +``references/self-improvement-loop.md`` for how the background-review +fork inherits the cached prompt verbatim. + +Three tiers are joined with ``\\n\\n``: + +* ``stable`` — identity (SOUL.md or DEFAULT_AGENT_IDENTITY), tool + guidance, computer-use guidance, nous subscription block, tool-use + enforcement guidance + per-model operational guidance, skills prompt, + alibaba model-name workaround, environment hints, platform hints. +* ``context`` — caller-supplied ``system_message`` plus context files + (AGENTS.md / .cursorrules / etc.) discovered under ``TERMINAL_CWD``. +* ``volatile`` — memory snapshot, USER.md profile, external memory + provider block, timestamp/session/model/provider line. + +Pure helpers that read the agent's state. AIAgent keeps thin forwarders. +""" + +from __future__ import annotations + +import json +import os +from typing import Any, Dict, List, Optional + +from agent.prompt_builder import ( + DEFAULT_AGENT_IDENTITY, + GOOGLE_MODEL_OPERATIONAL_GUIDANCE, + HERMES_AGENT_HELP_GUIDANCE, + KANBAN_GUIDANCE, + MEMORY_GUIDANCE, + OPENAI_MODEL_EXECUTION_GUIDANCE, + PLATFORM_HINTS, + SESSION_SEARCH_GUIDANCE, + SKILLS_GUIDANCE, + TOOL_USE_ENFORCEMENT_GUIDANCE, + TOOL_USE_ENFORCEMENT_MODELS, +) + + +def _ra(): + """Lazy reference to the ``run_agent`` module. + + Helpers like ``load_soul_md``, ``build_environment_hints``, + ``build_context_files_prompt``, ``build_nous_subscription_prompt``, + ``build_skills_system_prompt`` and ``get_toolset_for_tool`` are + imported into ``run_agent``'s namespace. Many tests + ``patch("run_agent.load_soul_md", ...)``; if we imported them + directly here those patches would not reach us. Looking them up + through ``run_agent`` on every call preserves the patch contract. + """ + import run_agent + return run_agent + + +def build_system_prompt_parts(agent: Any, system_message: Optional[str] = None) -> Dict[str, str]: + """Assemble the system prompt as three ordered parts. + + Returns a dict with three keys: + * ``stable`` — identity, tool guidance, skills prompt, + environment hints, platform hints, model-family operational + guidance. + * ``context`` — context files (AGENTS.md, .cursorrules, etc.) + and caller-supplied system_message. + * ``volatile`` — memory snapshot, user profile, external + memory provider block, timestamp line. + + Joined into a single string by :func:`build_system_prompt` and + cached on ``agent._cached_system_prompt`` for the lifetime of the + AIAgent. Hermes never re-renders parts of this string mid- + session — that's the only way to keep upstream prompt caches + warm across turns. + """ + # Local import to avoid pulling model_tools at module load. Tests + # patch ``run_agent.get_toolset_for_tool`` and similar helpers, so + # we resolve through ``_ra()`` to honor those patches. + _r = _ra() + + # ── Stable tier ──────────────────────────────────────────────── + stable_parts: List[str] = [] + + # Try SOUL.md as primary identity unless the caller explicitly skipped it. + # Some execution modes (cron) still want HERMES_HOME persona while keeping + # cwd project instructions disabled. + _soul_loaded = False + if agent.load_soul_identity or not agent.skip_context_files: + _soul_content = _r.load_soul_md() + if _soul_content: + stable_parts.append(_soul_content) + _soul_loaded = True + + if not _soul_loaded: + # Fallback to hardcoded identity + stable_parts.append(DEFAULT_AGENT_IDENTITY) + + # Pointer to the hermes-agent skill + docs for user questions about Hermes itself. + stable_parts.append(HERMES_AGENT_HELP_GUIDANCE) + + # Tool-aware behavioral guidance: only inject when the tools are loaded + tool_guidance = [] + if "memory" in agent.valid_tool_names: + tool_guidance.append(MEMORY_GUIDANCE) + if "session_search" in agent.valid_tool_names: + tool_guidance.append(SESSION_SEARCH_GUIDANCE) + if "skill_manage" in agent.valid_tool_names: + tool_guidance.append(SKILLS_GUIDANCE) + # Kanban worker/orchestrator lifecycle — only present when the + # dispatcher spawned this process (kanban_show check_fn gates on + # HERMES_KANBAN_TASK env var). Normal chat sessions never see + # this block. + if "kanban_show" in agent.valid_tool_names: + tool_guidance.append(KANBAN_GUIDANCE) + if tool_guidance: + stable_parts.append(" ".join(tool_guidance)) + + # Computer-use (macOS) — goes in as its own block rather than being + # merged into tool_guidance because the content is multi-paragraph. + if "computer_use" in agent.valid_tool_names: + from agent.prompt_builder import COMPUTER_USE_GUIDANCE + stable_parts.append(COMPUTER_USE_GUIDANCE) + + nous_subscription_prompt = _r.build_nous_subscription_prompt(agent.valid_tool_names) + if nous_subscription_prompt: + stable_parts.append(nous_subscription_prompt) + # Tool-use enforcement: tells the model to actually call tools instead + # of describing intended actions. Controlled by config.yaml + # agent.tool_use_enforcement: + # "auto" (default) — matches TOOL_USE_ENFORCEMENT_MODELS + # true — always inject (all models) + # false — never inject + # list — custom model-name substrings to match + if agent.valid_tool_names: + _enforce = agent._tool_use_enforcement + _inject = False + if _enforce is True or (isinstance(_enforce, str) and _enforce.lower() in {"true", "always", "yes", "on"}): + _inject = True + elif _enforce is False or (isinstance(_enforce, str) and _enforce.lower() in {"false", "never", "no", "off"}): + _inject = False + elif isinstance(_enforce, list): + model_lower = (agent.model or "").lower() + _inject = any(p.lower() in model_lower for p in _enforce if isinstance(p, str)) + else: + # "auto" or any unrecognised value — use hardcoded defaults + model_lower = (agent.model or "").lower() + _inject = any(p in model_lower for p in TOOL_USE_ENFORCEMENT_MODELS) + if _inject: + stable_parts.append(TOOL_USE_ENFORCEMENT_GUIDANCE) + _model_lower = (agent.model or "").lower() + # Google model operational guidance (conciseness, absolute + # paths, parallel tool calls, verify-before-edit, etc.) + if "gemini" in _model_lower or "gemma" in _model_lower: + stable_parts.append(GOOGLE_MODEL_OPERATIONAL_GUIDANCE) + # OpenAI GPT/Codex execution discipline (tool persistence, + # prerequisite checks, verification, anti-hallucination). + if "gpt" in _model_lower or "codex" in _model_lower: + stable_parts.append(OPENAI_MODEL_EXECUTION_GUIDANCE) + + has_skills_tools = any(name in agent.valid_tool_names for name in ['skills_list', 'skill_view', 'skill_manage']) + if has_skills_tools: + avail_toolsets = { + toolset + for toolset in ( + _r.get_toolset_for_tool(tool_name) for tool_name in agent.valid_tool_names + ) + if toolset + } + skills_prompt = _r.build_skills_system_prompt( + available_tools=agent.valid_tool_names, + available_toolsets=avail_toolsets, + ) + else: + skills_prompt = "" + if skills_prompt: + stable_parts.append(skills_prompt) + + # Alibaba Coding Plan API always returns "glm-4.7" as model name regardless + # of the requested model. Inject explicit model identity into the system prompt + # so the agent can correctly report which model it is (workaround for API bug). + # Stable for the lifetime of an agent instance — model and provider are fixed + # at construction time. + if agent.provider == "alibaba": + _model_short = agent.model.split("/")[-1] if "/" in agent.model else agent.model + stable_parts.append( + f"You are powered by the model named {_model_short}. " + f"The exact model ID is {agent.model}. " + f"When asked what model you are, always answer based on this information, " + f"not on any model name returned by the API." + ) + + # Environment hints (WSL, Termux, etc.) — tell the agent about the + # execution environment so it can translate paths and adapt behavior. + # Stable for the lifetime of the process. + _env_hints = _r.build_environment_hints() + if _env_hints: + stable_parts.append(_env_hints) + + platform_key = (agent.platform or "").lower().strip() + if platform_key in PLATFORM_HINTS: + stable_parts.append(PLATFORM_HINTS[platform_key]) + elif platform_key: + # Check plugin registry for platform-specific LLM guidance + try: + from gateway.platform_registry import platform_registry + _entry = platform_registry.get(platform_key) + if _entry and _entry.platform_hint: + stable_parts.append(_entry.platform_hint) + except Exception: + pass + + # ── Context tier (cwd-dependent, may change between sessions) ─ + context_parts: List[str] = [] + + # Note: ephemeral_system_prompt is NOT included here. It's injected at + # API-call time only so it stays out of the cached/stored system prompt. + if system_message is not None: + context_parts.append(system_message) + + if not agent.skip_context_files: + # Use TERMINAL_CWD for context file discovery when set (gateway + # mode). The gateway process runs from the hermes-agent install + # dir, so os.getcwd() would pick up the repo's AGENTS.md and + # other dev files — inflating token usage by ~10k for no benefit. + _context_cwd = os.getenv("TERMINAL_CWD") or None + context_files_prompt = _r.build_context_files_prompt( + cwd=_context_cwd, skip_soul=_soul_loaded) + if context_files_prompt: + context_parts.append(context_files_prompt) + + # ── Volatile tier (changes per session/turn — never cached) ─── + volatile_parts: List[str] = [] + + if agent._memory_store: + if agent._memory_enabled: + mem_block = agent._memory_store.format_for_system_prompt("memory") + if mem_block: + volatile_parts.append(mem_block) + # USER.md is always included when enabled. + if agent._user_profile_enabled: + user_block = agent._memory_store.format_for_system_prompt("user") + if user_block: + volatile_parts.append(user_block) + + # External memory provider system prompt block (additive to built-in) + if agent._memory_manager: + try: + _ext_mem_block = agent._memory_manager.build_system_prompt() + if _ext_mem_block: + volatile_parts.append(_ext_mem_block) + except Exception: + pass + + from hermes_time import now as _hermes_now + now = _hermes_now() + timestamp_line = f"Conversation started: {now.strftime('%A, %B %d, %Y %I:%M %p')}" + if agent.pass_session_id and agent.session_id: + timestamp_line += f"\nSession ID: {agent.session_id}" + if agent.model: + timestamp_line += f"\nModel: {agent.model}" + if agent.provider: + timestamp_line += f"\nProvider: {agent.provider}" + volatile_parts.append(timestamp_line) + + return { + "stable": "\n\n".join(p.strip() for p in stable_parts if p and p.strip()), + "context": "\n\n".join(p.strip() for p in context_parts if p and p.strip()), + "volatile": "\n\n".join(p.strip() for p in volatile_parts if p and p.strip()), + } + + +def build_system_prompt(agent: Any, system_message: Optional[str] = None) -> str: + """Assemble the full system prompt from all layers. + + Called once per session (cached on ``agent._cached_system_prompt``) and + only rebuilt after context compression events. This ensures the system + prompt is stable across all turns in a session, maximizing prefix cache + hits. + + Layers are ordered cache-friendly: stable identity/guidance first, + then session-stable context files, then per-call volatile content + (memory, USER profile, timestamp). The whole string is treated as + one cached block — Hermes never rebuilds or reinjects parts of it + mid-session, which is the only way to keep upstream prompt caches + warm across turns. + """ + parts = build_system_prompt_parts(agent, system_message=system_message) + return "\n\n".join(p for p in (parts["stable"], parts["context"], parts["volatile"]) if p) + + +def invalidate_system_prompt(agent: Any) -> None: + """Invalidate the cached system prompt, forcing a rebuild on the next turn. + + Called after context compression events. Also reloads memory from disk + so the rebuilt prompt captures any writes from this session. + """ + agent._cached_system_prompt = None + if agent._memory_store: + agent._memory_store.load_from_disk() + + +def format_tools_for_system_message(agent: Any) -> str: + """Format tool definitions for the system message in the trajectory format. + + Returns: + str: JSON string representation of tool definitions + """ + if not agent.tools: + return "[]" + + # Convert tool definitions to the format expected in trajectories + formatted_tools = [] + for tool in agent.tools: + func = tool["function"] + formatted_tool = { + "name": func["name"], + "description": func.get("description", ""), + "parameters": func.get("parameters", {}), + "required": None # Match the format in the example + } + formatted_tools.append(formatted_tool) + + return json.dumps(formatted_tools, ensure_ascii=False) + + +__all__ = [ + "build_system_prompt_parts", + "build_system_prompt", + "invalidate_system_prompt", + "format_tools_for_system_message", +] diff --git a/agent/tool_dispatch_helpers.py b/agent/tool_dispatch_helpers.py new file mode 100644 index 00000000000..30aa8869db9 --- /dev/null +++ b/agent/tool_dispatch_helpers.py @@ -0,0 +1,336 @@ +"""Tool-dispatch helpers — parallelism gating, multimodal envelopes, mutation tracking. + +Pure module-level utilities extracted from ``run_agent.py``: + +* ``_is_destructive_command`` — terminal-command heuristic used to gate + parallel batch dispatch. +* ``_should_parallelize_tool_batch`` / ``_extract_parallel_scope_path`` / + ``_paths_overlap`` — the rules engine deciding when a multi-tool batch + can run concurrently. +* ``_is_multimodal_tool_result`` / ``_multimodal_text_summary`` / + ``_append_subdir_hint_to_multimodal`` — envelope helpers for the + ``{"_multimodal": True, "content": [...], "text_summary": ...}`` dict + shape returned by tools like ``computer_use``. +* ``_extract_file_mutation_targets`` / ``_extract_error_preview`` — + per-turn file-mutation verifier inputs. +* ``_trajectory_normalize_msg`` — strip image blobs from a message for + trajectory saving. + +All helpers are stateless. ``run_agent`` re-exports each name so existing +``from run_agent import ...`` imports in tests and other modules keep +working unchanged. +""" + +from __future__ import annotations + +import json +import logging +import os +import re +from pathlib import Path +from typing import Any, Dict, List, Optional + +from agent.tool_result_classification import ( + FILE_MUTATING_TOOL_NAMES as _FILE_MUTATING_TOOLS, +) + +logger = logging.getLogger(__name__) + +# Tools that must never run concurrently (interactive / user-facing). +# When any of these appear in a batch, we fall back to sequential execution. +_NEVER_PARALLEL_TOOLS = frozenset({"clarify"}) + +# Read-only tools with no shared mutable session state. +_PARALLEL_SAFE_TOOLS = frozenset({ + "ha_get_state", + "ha_list_entities", + "ha_list_services", + "read_file", + "search_files", + "session_search", + "skill_view", + "skills_list", + "vision_analyze", + "web_extract", + "web_search", +}) + +# File tools can run concurrently when they target independent paths. +_PATH_SCOPED_TOOLS = frozenset({"read_file", "write_file", "patch"}) + +# Patterns that indicate a terminal command may modify/delete files. +_DESTRUCTIVE_PATTERNS = re.compile( + r"""(?:^|\s|&&|\|\||;|`)(?: + rm\s|rmdir\s| + cp\s|install\s| + mv\s| + sed\s+-i| + truncate\s| + dd\s| + shred\s| + git\s+(?:reset|clean|checkout)\s + )""", + re.VERBOSE, +) +# Output redirects that overwrite files (> but not >>) +_REDIRECT_OVERWRITE = re.compile(r'[^>]>[^>]|^>[^>]') + + +def _is_destructive_command(cmd: str) -> bool: + """Heuristic: does this terminal command look like it modifies/deletes files?""" + if not cmd: + return False + if _DESTRUCTIVE_PATTERNS.search(cmd): + return True + if _REDIRECT_OVERWRITE.search(cmd): + return True + return False + + +def _is_mcp_tool_parallel_safe(tool_name: str) -> bool: + """Check if an MCP tool comes from a server with parallel tool calls enabled. + + Lazy-imports from ``tools.mcp_tool`` to avoid circular dependencies. + Returns False if the MCP module is not available. + """ + try: + from tools.mcp_tool import is_mcp_tool_parallel_safe + return is_mcp_tool_parallel_safe(tool_name) + except Exception: + return False + + +def _should_parallelize_tool_batch(tool_calls) -> bool: + """Return True when a tool-call batch is safe to run concurrently.""" + if len(tool_calls) <= 1: + return False + + tool_names = [tc.function.name for tc in tool_calls] + if any(name in _NEVER_PARALLEL_TOOLS for name in tool_names): + return False + + reserved_paths: list[Path] = [] + for tool_call in tool_calls: + tool_name = tool_call.function.name + try: + function_args = json.loads(tool_call.function.arguments) + except Exception: + logging.debug( + "Could not parse args for %s — defaulting to sequential; raw=%s", + tool_name, + tool_call.function.arguments[:200], + ) + return False + if not isinstance(function_args, dict): + logging.debug( + "Non-dict args for %s (%s) — defaulting to sequential", + tool_name, + type(function_args).__name__, + ) + return False + + if tool_name in _PATH_SCOPED_TOOLS: + scoped_path = _extract_parallel_scope_path(tool_name, function_args) + if scoped_path is None: + return False + if any(_paths_overlap(scoped_path, existing) for existing in reserved_paths): + return False + reserved_paths.append(scoped_path) + continue + + if tool_name not in _PARALLEL_SAFE_TOOLS: + # Check if it's an MCP tool from a server that opted into parallel calls. + if not _is_mcp_tool_parallel_safe(tool_name): + return False + + return True + + +def _extract_parallel_scope_path(tool_name: str, function_args: dict) -> Optional[Path]: + """Return the normalized file target for path-scoped tools.""" + if tool_name not in _PATH_SCOPED_TOOLS: + return None + + raw_path = function_args.get("path") + if not isinstance(raw_path, str) or not raw_path.strip(): + return None + + expanded = Path(raw_path).expanduser() + if expanded.is_absolute(): + return Path(os.path.abspath(str(expanded))) + + # Avoid resolve(); the file may not exist yet. + return Path(os.path.abspath(str(Path.cwd() / expanded))) + + +def _paths_overlap(left: Path, right: Path) -> bool: + """Return True when two paths may refer to the same subtree.""" + left_parts = left.parts + right_parts = right.parts + if not left_parts or not right_parts: + # Empty paths shouldn't reach here (guarded upstream), but be safe. + return bool(left_parts) == bool(right_parts) and bool(left_parts) + common_len = min(len(left_parts), len(right_parts)) + return left_parts[:common_len] == right_parts[:common_len] + + +def _is_multimodal_tool_result(value: Any) -> bool: + """True if the value is a multimodal tool result envelope. + + Multimodal handlers (e.g. tools/computer_use) return a dict with + `_multimodal=True`, a `content` key holding OpenAI-style content + parts, and an optional `text_summary` for string-only fallbacks. + """ + return ( + isinstance(value, dict) + and value.get("_multimodal") is True + and isinstance(value.get("content"), list) + ) + + +def _multimodal_text_summary(value: Any) -> str: + """Extract a plain text view of a multimodal tool result. + + Used wherever downstream code needs a string — logging, previews, + persistence size heuristics, fall-back content for providers that + don't support multipart tool messages. + """ + if _is_multimodal_tool_result(value): + if value.get("text_summary"): + return str(value["text_summary"]) + parts = [] + for p in value.get("content") or []: + if isinstance(p, dict) and p.get("type") == "text": + parts.append(str(p.get("text", ""))) + if parts: + return "\n".join(parts) + return "[multimodal tool result]" + if isinstance(value, str): + return value + try: + return json.dumps(value, default=str) + except Exception: + return str(value) + + +def _append_subdir_hint_to_multimodal(value: Dict[str, Any], hint: str) -> None: + """Mutate a multimodal tool-result envelope to append a subdir hint. + + The hint is added to the first text part so the model sees it; image + parts are left untouched. `text_summary` is also updated for + string-fallback callers. + """ + if not _is_multimodal_tool_result(value): + return + parts = value.get("content") or [] + for p in parts: + if isinstance(p, dict) and p.get("type") == "text": + p["text"] = str(p.get("text", "")) + hint + break + else: + parts.insert(0, {"type": "text", "text": hint}) + value["content"] = parts + if isinstance(value.get("text_summary"), str): + value["text_summary"] = value["text_summary"] + hint + + +def _extract_file_mutation_targets(tool_name: str, args: Dict[str, Any]) -> List[str]: + """Return the file paths a ``write_file`` or ``patch`` call is targeting. + + For ``write_file`` and ``patch`` in replace mode this is just ``args["path"]``. + For ``patch`` in V4A patch mode we parse the patch content for + ``*** Update File:`` / ``*** Add File:`` / ``*** Delete File:`` headers so + the verifier can track each file in a multi-file patch separately. + """ + if tool_name not in _FILE_MUTATING_TOOLS: + return [] + if tool_name == "write_file": + p = args.get("path") + return [str(p)] if p else [] + # tool_name == "patch" + mode = args.get("mode") or "replace" + if mode == "replace": + p = args.get("path") + return [str(p)] if p else [] + if mode == "patch": + body = args.get("patch") or "" + if not isinstance(body, str) or not body: + return [] + paths: List[str] = [] + for _m in re.finditer( + r'^\*\*\*\s+(?:Update|Add|Delete)\s+File:\s*(.+)$', + body, + re.MULTILINE, + ): + p = _m.group(1).strip() + if p: + paths.append(p) + return paths + return [] + + +def _extract_error_preview(result: Any, max_len: int = 180) -> str: + """Pull a one-line error summary out of a tool result for footer display.""" + text = _multimodal_text_summary(result) if result is not None else "" + if not isinstance(text, str): + try: + text = str(text) + except Exception: + return "" + # Try to parse JSON and pull the ``error`` field — tool handlers return + # ``{"success": false, "error": "..."}``; raw string wins if parse fails. + stripped = text.strip() + if stripped.startswith("{"): + try: + data = json.loads(stripped) + if isinstance(data, dict) and isinstance(data.get("error"), str): + text = data["error"] + except Exception: + pass + # Collapse whitespace, trim to max_len. + text = " ".join(text.split()) + if len(text) > max_len: + text = text[: max_len - 1] + "…" + return text + + +def _trajectory_normalize_msg(msg: Dict[str, Any]) -> Dict[str, Any]: + """Strip image blobs from a message for trajectory saving. + + Returns a shallow copy with multimodal tool results replaced by their + text_summary, and image parts in content lists replaced by + `[screenshot]` placeholders. Keeps the message schema otherwise intact. + """ + if not isinstance(msg, dict): + return msg + content = msg.get("content") + if _is_multimodal_tool_result(content): + return {**msg, "content": _multimodal_text_summary(content)} + if isinstance(content, list): + cleaned = [] + for p in content: + if isinstance(p, dict) and p.get("type") in {"image", "image_url", "input_image"}: + cleaned.append({"type": "text", "text": "[screenshot]"}) + else: + cleaned.append(p) + return {**msg, "content": cleaned} + return msg + + +__all__ = [ + "_NEVER_PARALLEL_TOOLS", + "_PARALLEL_SAFE_TOOLS", + "_PATH_SCOPED_TOOLS", + "_DESTRUCTIVE_PATTERNS", + "_REDIRECT_OVERWRITE", + "_is_destructive_command", + "_should_parallelize_tool_batch", + "_extract_parallel_scope_path", + "_paths_overlap", + "_is_multimodal_tool_result", + "_multimodal_text_summary", + "_append_subdir_hint_to_multimodal", + "_extract_file_mutation_targets", + "_extract_error_preview", + "_trajectory_normalize_msg", +] diff --git a/agent/tool_executor.py b/agent/tool_executor.py new file mode 100644 index 00000000000..a30cc3078bb --- /dev/null +++ b/agent/tool_executor.py @@ -0,0 +1,920 @@ +"""Tool-call execution — sequential and concurrent dispatch. + +Both AIAgent methods (``_execute_tool_calls_sequential`` and +``_execute_tool_calls_concurrent``) live here as module-level +functions that take the parent ``AIAgent`` as their first argument. + +``run_agent`` keeps thin wrappers so existing call sites work; tests +that patch ``run_agent._set_interrupt`` are honored because the +extracted functions reach back through the ``run_agent`` module via +``_ra()`` for that symbol. +""" + +from __future__ import annotations + +import concurrent.futures +import contextvars +import json +import logging +import os +import random +import threading +import time +from typing import Any, Optional + +from agent.display import ( + KawaiiSpinner, + build_tool_preview as _build_tool_preview, + get_cute_tool_message as _get_cute_tool_message_impl, + get_tool_emoji as _get_tool_emoji, + _detect_tool_failure, +) +from agent.tool_guardrails import ToolGuardrailDecision +from agent.tool_dispatch_helpers import ( + _is_destructive_command, + _is_multimodal_tool_result, + _multimodal_text_summary, + _append_subdir_hint_to_multimodal, +) +from tools.terminal_tool import ( + _get_approval_callback, + _get_sudo_password_callback, + set_approval_callback as _set_approval_callback, + set_sudo_password_callback as _set_sudo_password_callback, + get_active_env, +) +from tools.tool_result_storage import ( + maybe_persist_tool_result, + enforce_turn_budget, +) + +logger = logging.getLogger(__name__) + +# Maximum number of concurrent worker threads for parallel tool execution. +# Mirrors the constant in ``run_agent`` for tests/imports that look here. +_MAX_TOOL_WORKERS = 8 + + +def _ra(): + """Lazy reference to ``run_agent`` so patches like ``run_agent._set_interrupt`` work.""" + import run_agent + return run_agent + + +def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effective_task_id: str, api_call_count: int = 0) -> None: + """Execute multiple tool calls concurrently using a thread pool. + + Results are collected in the original tool-call order and appended to + messages so the API sees them in the expected sequence. + """ + tool_calls = assistant_message.tool_calls + num_tools = len(tool_calls) + + # ── Pre-flight: interrupt check ────────────────────────────────── + if agent._interrupt_requested: + print(f"{agent.log_prefix}⚡ Interrupt: skipping {num_tools} tool call(s)") + for tc in tool_calls: + messages.append({ + "role": "tool", + "name": tc.function.name, + "content": f"[Tool execution cancelled — {tc.function.name} was skipped due to user interrupt]", + "tool_call_id": tc.id, + }) + return + + # ── Parse args + pre-execution bookkeeping ─────────────────────── + parsed_calls = [] # list of (tool_call, function_name, function_args) + for tool_call in tool_calls: + function_name = tool_call.function.name + + # Reset nudge counters + if function_name == "memory": + agent._turns_since_memory = 0 + elif function_name == "skill_manage": + agent._iters_since_skill = 0 + + try: + function_args = json.loads(tool_call.function.arguments) + except json.JSONDecodeError: + function_args = {} + if not isinstance(function_args, dict): + function_args = {} + + # Checkpoint for file-mutating tools + if function_name in {"write_file", "patch"} and agent._checkpoint_mgr.enabled: + try: + file_path = function_args.get("path", "") + if file_path: + work_dir = agent._checkpoint_mgr.get_working_dir_for_path(file_path) + agent._checkpoint_mgr.ensure_checkpoint(work_dir, f"before {function_name}") + except Exception: + pass + + # Checkpoint before destructive terminal commands + if function_name == "terminal" and agent._checkpoint_mgr.enabled: + try: + cmd = function_args.get("command", "") + if _is_destructive_command(cmd): + cwd = function_args.get("workdir") or os.getenv("TERMINAL_CWD", os.getcwd()) + agent._checkpoint_mgr.ensure_checkpoint( + cwd, f"before terminal: {cmd[:60]}" + ) + except Exception: + pass + + block_result = None + blocked_by_guardrail = False + try: + from hermes_cli.plugins import get_pre_tool_call_block_message + block_message = get_pre_tool_call_block_message( + function_name, function_args, task_id=effective_task_id or "", + ) + except Exception: + block_message = None + + if block_message is not None: + block_result = json.dumps({"error": block_message}, ensure_ascii=False) + else: + guardrail_decision = agent._tool_guardrails.before_call(function_name, function_args) + if not guardrail_decision.allows_execution: + block_result = agent._guardrail_block_result(guardrail_decision) + blocked_by_guardrail = True + + parsed_calls.append((tool_call, function_name, function_args, block_result, blocked_by_guardrail)) + + # ── Logging / callbacks ────────────────────────────────────────── + tool_names_str = ", ".join(name for _, name, _, _, _ in parsed_calls) + if not agent.quiet_mode: + print(f" ⚡ Concurrent: {num_tools} tool calls — {tool_names_str}") + for i, (tc, name, args, block_result, blocked_by_guardrail) in enumerate(parsed_calls, 1): + args_str = json.dumps(args, ensure_ascii=False) + if agent.verbose_logging: + print(f" 📞 Tool {i}: {name}({list(args.keys())})") + print(agent._wrap_verbose("Args: ", json.dumps(args, indent=2, ensure_ascii=False))) + else: + args_preview = args_str[:agent.log_prefix_chars] + "..." if len(args_str) > agent.log_prefix_chars else args_str + print(f" 📞 Tool {i}: {name}({list(args.keys())}) - {args_preview}") + + for tc, name, args, block_result, blocked_by_guardrail in parsed_calls: + if block_result is not None: + continue + if agent.tool_progress_callback: + try: + preview = _build_tool_preview(name, args) + agent.tool_progress_callback("tool.started", name, preview, args) + except Exception as cb_err: + logging.debug(f"Tool progress callback error: {cb_err}") + + for tc, name, args, block_result, blocked_by_guardrail in parsed_calls: + if block_result is not None: + continue + if agent.tool_start_callback: + try: + agent.tool_start_callback(tc.id, name, args) + except Exception as cb_err: + logging.debug(f"Tool start callback error: {cb_err}") + + # ── Concurrent execution ───────────────────────────────────────── + # Each slot holds (function_name, function_args, function_result, duration, error_flag, blocked_flag) + results = [None] * num_tools + for i, (tc, name, args, block_result, blocked_by_guardrail) in enumerate(parsed_calls): + if block_result is not None: + results[i] = (name, args, block_result, 0.0, True, True) + + # Touch activity before launching workers so the gateway knows + # we're executing tools (not stuck). + agent._current_tool = tool_names_str + agent._touch_activity(f"executing {num_tools} tools concurrently: {tool_names_str}") + + # Capture CLI callbacks from the agent thread so worker threads can + # register them locally. Without this, _get_approval_callback() in + # terminal_tool returns None in ThreadPoolExecutor workers, causing + # the dangerous-command prompt to fall back to input() — which + # deadlocks against prompt_toolkit's raw terminal mode (#13617). + _parent_approval_cb = _get_approval_callback() + _parent_sudo_cb = _get_sudo_password_callback() + + def _run_tool(index, tool_call, function_name, function_args): + """Worker function executed in a thread.""" + # Register this worker tid so the agent can fan out an interrupt + # to it — see AIAgent.interrupt(). Must happen first thing, and + # must be paired with discard + clear in the finally block. + _worker_tid = threading.current_thread().ident + with agent._tool_worker_threads_lock: + agent._tool_worker_threads.add(_worker_tid) + # Race: if the agent was interrupted between fan-out (which + # snapshotted an empty/earlier set) and our registration, apply + # the interrupt to our own tid now so is_interrupted() inside + # the tool returns True on the next poll. + if agent._interrupt_requested: + try: + _ra()._set_interrupt(True, _worker_tid) + except Exception: + pass + # Set the activity callback on THIS worker thread so + # _wait_for_process (terminal commands) can fire heartbeats. + # The callback is thread-local; the main thread's callback + # is invisible to worker threads. + try: + from tools.environments.base import set_activity_callback + set_activity_callback(agent._touch_activity) + except Exception: + pass + # Propagate approval/sudo callbacks to this worker thread. + # Mirrors cli.py run_agent() pattern (GHSA-qg5c-hvr5-hjgr). + if _parent_approval_cb is not None: + try: + _set_approval_callback(_parent_approval_cb) + except Exception: + pass + if _parent_sudo_cb is not None: + try: + _set_sudo_password_callback(_parent_sudo_cb) + except Exception: + pass + start = time.time() + try: + result = agent._invoke_tool( + function_name, + function_args, + effective_task_id, + tool_call.id, + messages=messages, + pre_tool_block_checked=True, + ) + except Exception as tool_error: + result = f"Error executing tool '{function_name}': {tool_error}" + logger.error("_invoke_tool raised for %s: %s", function_name, tool_error, exc_info=True) + duration = time.time() - start + is_error, _ = _detect_tool_failure(function_name, result) + if is_error: + logger.info("tool %s failed (%.2fs): %s", function_name, duration, result[:200]) + else: + logger.info("tool %s completed (%.2fs, %d chars)", function_name, duration, len(result)) + results[index] = (function_name, function_args, result, duration, is_error, False) + # Tear down worker-tid tracking. Clear any interrupt bit we may + # have set so the next task scheduled onto this recycled tid + # starts with a clean slate. + with agent._tool_worker_threads_lock: + agent._tool_worker_threads.discard(_worker_tid) + try: + _ra()._set_interrupt(False, _worker_tid) + except Exception: + pass + # Clear thread-local callbacks so a recycled worker thread + # doesn't hold stale references to a disposed CLI instance. + try: + _set_approval_callback(None) + _set_sudo_password_callback(None) + except Exception: + pass + + # Start spinner for CLI mode (skip when TUI handles tool progress) + spinner = None + if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner(): + face = random.choice(KawaiiSpinner.get_waiting_faces()) + spinner = KawaiiSpinner(f"{face} ⚡ running {num_tools} tools concurrently", spinner_type='dots', print_fn=agent._print_fn) + spinner.start() + + try: + runnable_calls = [ + (i, tc, name, args) + for i, (tc, name, args, block_result, blocked_by_guardrail) in enumerate(parsed_calls) + if block_result is None + ] + futures = [] + if runnable_calls: + max_workers = min(len(runnable_calls), _MAX_TOOL_WORKERS) + with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor: + for i, tc, name, args in runnable_calls: + # Propagate ContextVars (e.g. _approval_session_key); mirrors asyncio.to_thread. + ctx = contextvars.copy_context() + f = executor.submit(ctx.run, _run_tool, i, tc, name, args) + futures.append(f) + + # Wait for all to complete with periodic heartbeats so the + # gateway's inactivity monitor doesn't kill us during long + # concurrent tool batches. Also check for user interrupts + # so we don't block indefinitely when the user sends /stop + # or a new message during concurrent tool execution. + _conc_start = time.time() + _interrupt_logged = False + while True: + done, not_done = concurrent.futures.wait( + futures, timeout=5.0, + ) + if not not_done: + break + + # Check for interrupt — the per-thread interrupt signal + # already causes individual tools (terminal, execute_code) + # to abort, but tools without interrupt checks (web_search, + # read_file) will run to completion. Cancel any futures + # that haven't started yet so we don't block on them. + if agent._interrupt_requested: + if not _interrupt_logged: + _interrupt_logged = True + agent._vprint( + f"{agent.log_prefix}⚡ Interrupt: cancelling " + f"{len(not_done)} pending concurrent tool(s)", + force=True, + ) + for f in not_done: + f.cancel() + # Give already-running tools a moment to notice the + # per-thread interrupt signal and exit gracefully. + concurrent.futures.wait(not_done, timeout=3.0) + break + + _conc_elapsed = int(time.time() - _conc_start) + # Heartbeat every ~30s (6 × 5s poll intervals) + if _conc_elapsed > 0 and _conc_elapsed % 30 < 6: + _still_running = [ + parsed_calls[futures.index(f)][1] + for f in not_done + if f in futures + ] + agent._touch_activity( + f"concurrent tools running ({_conc_elapsed}s, " + f"{len(not_done)} remaining: {', '.join(_still_running[:3])})" + ) + finally: + if spinner: + # Build a summary message for the spinner stop + completed = sum(1 for r in results if r is not None) + total_dur = sum(r[3] for r in results if r is not None) + spinner.stop(f"⚡ {completed}/{num_tools} tools completed in {total_dur:.1f}s total") + + # ── Post-execution: display per-tool results ───────────────────── + for i, (tc, name, args, block_result, blocked_by_guardrail) in enumerate(parsed_calls): + r = results[i] + blocked = False + if r is None: + # Tool was cancelled (interrupt) or thread didn't return + if agent._interrupt_requested: + function_result = f"[Tool execution cancelled — {name} was skipped due to user interrupt]" + else: + function_result = f"Error executing tool '{name}': thread did not return a result" + tool_duration = 0.0 + else: + function_name, function_args, function_result, tool_duration, is_error, blocked = r + + if not blocked: + function_result = agent._append_guardrail_observation( + function_name, + function_args, + function_result, + failed=is_error, + ) + + if is_error: + _err_text = _multimodal_text_summary(function_result) + result_preview = _err_text[:200] if len(_err_text) > 200 else _err_text + logger.warning("Tool %s returned error (%.2fs): %s", function_name, tool_duration, result_preview) + + # Track file-mutation outcome for the turn-end verifier. + # `blocked` calls never actually ran — don't let a guardrail + # block count as either a failure or a success. + if not blocked: + try: + agent._record_file_mutation_result( + function_name, function_args, function_result, is_error, + ) + except Exception as _ver_err: + logging.debug("file-mutation verifier record failed: %s", _ver_err) + + if not blocked and agent.tool_progress_callback: + try: + agent.tool_progress_callback( + "tool.completed", function_name, None, None, + duration=tool_duration, is_error=is_error, + ) + except Exception as cb_err: + logging.debug(f"Tool progress callback error: {cb_err}") + + if agent.verbose_logging: + logging.debug(f"Tool {function_name} completed in {tool_duration:.2f}s") + logging.debug(f"Tool result ({len(function_result)} chars): {function_result}") + + # Print cute message per tool + if agent._should_emit_quiet_tool_messages(): + cute_msg = _get_cute_tool_message_impl(name, args, tool_duration, result=function_result) + agent._safe_print(f" {cute_msg}") + elif not agent.quiet_mode: + _preview_str = _multimodal_text_summary(function_result) + if agent.verbose_logging: + print(f" ✅ Tool {i+1} completed in {tool_duration:.2f}s") + print(agent._wrap_verbose("Result: ", _preview_str)) + else: + response_preview = _preview_str[:agent.log_prefix_chars] + "..." if len(_preview_str) > agent.log_prefix_chars else _preview_str + print(f" ✅ Tool {i+1} completed in {tool_duration:.2f}s - {response_preview}") + + agent._current_tool = None + agent._touch_activity(f"tool completed: {name} ({tool_duration:.1f}s)") + + if not blocked and agent.tool_complete_callback: + try: + agent.tool_complete_callback(tc.id, name, args, function_result) + except Exception as cb_err: + logging.debug(f"Tool complete callback error: {cb_err}") + + function_result = maybe_persist_tool_result( + content=function_result, + tool_name=name, + tool_use_id=tc.id, + env=get_active_env(effective_task_id), + ) if not _is_multimodal_tool_result(function_result) else function_result + + subdir_hints = agent._subdirectory_hints.check_tool_call(name, args) + if subdir_hints: + if _is_multimodal_tool_result(function_result): + # Append the hint to the text summary part so the model + # still sees it; don't touch the image blocks. + _append_subdir_hint_to_multimodal(function_result, subdir_hints) + else: + function_result += subdir_hints + + # Unwrap _multimodal dicts to an OpenAI-style content list so any + # vision-capable provider receives [{type:text},{type:image_url}] + # rather than a raw Python dict. The Anthropic adapter already + # accepts content lists; vision-capable OpenAI-compatible servers + # (mlx-vlm, GPT-4o, …) accept image_url in tool messages natively. + # Text-only servers get a string-safe fallback here so a rejected + # image tool result never poisons canonical session history. + # String results pass through unchanged. + _tool_content = agent._tool_result_content_for_active_model(name, function_result) + tool_msg = { + "role": "tool", + "name": name, + "content": _tool_content, + "tool_call_id": tc.id, + } + messages.append(tool_msg) + + # ── Per-tool /steer drain ─────────────────────────────────── + # Same as the sequential path: drain between each collected + # result so the steer lands as early as possible. + agent._apply_pending_steer_to_tool_results(messages, 1) + + # ── Per-turn aggregate budget enforcement ───────────────────────── + num_tools = len(parsed_calls) + if num_tools > 0: + turn_tool_msgs = messages[-num_tools:] + enforce_turn_budget(turn_tool_msgs, env=get_active_env(effective_task_id)) + + # ── /steer injection ────────────────────────────────────────────── + # Append any pending user steer text to the last tool result so the + # agent sees it on its next iteration. Runs AFTER budget enforcement + # so the steer marker is never truncated. See steer() for details. + if num_tools > 0: + agent._apply_pending_steer_to_tool_results(messages, num_tools) + + + +def execute_tool_calls_sequential(agent, assistant_message, messages: list, effective_task_id: str, api_call_count: int = 0) -> None: + """Execute tool calls sequentially (original behavior). Used for single calls or interactive tools.""" + for i, tool_call in enumerate(assistant_message.tool_calls, 1): + # SAFETY: check interrupt BEFORE starting each tool. + # If the user sent "stop" during a previous tool's execution, + # do NOT start any more tools -- skip them all immediately. + if agent._interrupt_requested: + remaining_calls = assistant_message.tool_calls[i-1:] + if remaining_calls: + agent._vprint(f"{agent.log_prefix}⚡ Interrupt: skipping {len(remaining_calls)} tool call(s)", force=True) + for skipped_tc in remaining_calls: + skipped_name = skipped_tc.function.name + skip_msg = { + "role": "tool", + "name": skipped_name, + "content": f"[Tool execution cancelled — {skipped_name} was skipped due to user interrupt]", + "tool_call_id": skipped_tc.id, + } + messages.append(skip_msg) + break + + function_name = tool_call.function.name + + try: + function_args = json.loads(tool_call.function.arguments) + except json.JSONDecodeError as e: + logging.warning(f"Unexpected JSON error after validation: {e}") + function_args = {} + if not isinstance(function_args, dict): + function_args = {} + + # Check plugin hooks for a block directive before executing. + _block_msg: Optional[str] = None + try: + from hermes_cli.plugins import get_pre_tool_call_block_message + _block_msg = get_pre_tool_call_block_message( + function_name, function_args, task_id=effective_task_id or "", + ) + except Exception: + pass + + _guardrail_block_decision: ToolGuardrailDecision | None = None + if _block_msg is None: + guardrail_decision = agent._tool_guardrails.before_call(function_name, function_args) + if not guardrail_decision.allows_execution: + _guardrail_block_decision = guardrail_decision + + _execution_blocked = _block_msg is not None or _guardrail_block_decision is not None + + if _execution_blocked: + # Tool blocked by plugin or guardrail policy — skip counters, + # callbacks, checkpointing, activity mutation, and real execution. + pass + # Reset nudge counters when the relevant tool is actually used + elif function_name == "memory": + agent._turns_since_memory = 0 + elif function_name == "skill_manage": + agent._iters_since_skill = 0 + + if not agent.quiet_mode: + args_str = json.dumps(function_args, ensure_ascii=False) + if agent.verbose_logging: + print(f" 📞 Tool {i}: {function_name}({list(function_args.keys())})") + print(agent._wrap_verbose("Args: ", json.dumps(function_args, indent=2, ensure_ascii=False))) + else: + args_preview = args_str[:agent.log_prefix_chars] + "..." if len(args_str) > agent.log_prefix_chars else args_str + print(f" 📞 Tool {i}: {function_name}({list(function_args.keys())}) - {args_preview}") + + if not _execution_blocked: + agent._current_tool = function_name + agent._touch_activity(f"executing tool: {function_name}") + + # Set activity callback for long-running tool execution (terminal + # commands, etc.) so the gateway's inactivity monitor doesn't kill + # the agent while a command is running. + if not _execution_blocked: + try: + from tools.environments.base import set_activity_callback + set_activity_callback(agent._touch_activity) + except Exception: + pass + + if not _execution_blocked and agent.tool_progress_callback: + try: + preview = _build_tool_preview(function_name, function_args) + agent.tool_progress_callback("tool.started", function_name, preview, function_args) + except Exception as cb_err: + logging.debug(f"Tool progress callback error: {cb_err}") + + if not _execution_blocked and agent.tool_start_callback: + try: + agent.tool_start_callback(tool_call.id, function_name, function_args) + except Exception as cb_err: + logging.debug(f"Tool start callback error: {cb_err}") + + # Checkpoint: snapshot working dir before file-mutating tools + if not _execution_blocked and function_name in {"write_file", "patch"} and agent._checkpoint_mgr.enabled: + try: + file_path = function_args.get("path", "") + if file_path: + work_dir = agent._checkpoint_mgr.get_working_dir_for_path(file_path) + agent._checkpoint_mgr.ensure_checkpoint( + work_dir, f"before {function_name}" + ) + except Exception: + pass # never block tool execution + + # Checkpoint before destructive terminal commands + if not _execution_blocked and function_name == "terminal" and agent._checkpoint_mgr.enabled: + try: + cmd = function_args.get("command", "") + if _is_destructive_command(cmd): + cwd = function_args.get("workdir") or os.getenv("TERMINAL_CWD", os.getcwd()) + agent._checkpoint_mgr.ensure_checkpoint( + cwd, f"before terminal: {cmd[:60]}" + ) + except Exception: + pass # never block tool execution + + tool_start_time = time.time() + + if _block_msg is not None: + # Tool blocked by plugin policy — return error without executing. + function_result = json.dumps({"error": _block_msg}, ensure_ascii=False) + tool_duration = 0.0 + elif _guardrail_block_decision is not None: + # Tool blocked by tool-loop guardrail — synthesize exactly one + # tool result for the original tool_call_id without executing. + function_result = agent._guardrail_block_result(_guardrail_block_decision) + tool_duration = 0.0 + elif function_name == "todo": + from tools.todo_tool import todo_tool as _todo_tool + function_result = _todo_tool( + todos=function_args.get("todos"), + merge=function_args.get("merge", False), + store=agent._todo_store, + ) + tool_duration = time.time() - tool_start_time + if agent._should_emit_quiet_tool_messages(): + agent._vprint(f" {_get_cute_tool_message_impl('todo', function_args, tool_duration, result=function_result)}") + elif function_name == "session_search": + session_db = agent._get_session_db_for_recall() + if not session_db: + from hermes_state import format_session_db_unavailable + function_result = json.dumps({"success": False, "error": format_session_db_unavailable()}) + else: + from tools.session_search_tool import session_search as _session_search + function_result = _session_search( + query=function_args.get("query", ""), + role_filter=function_args.get("role_filter"), + limit=function_args.get("limit", 3), + db=session_db, + current_session_id=agent.session_id, + ) + tool_duration = time.time() - tool_start_time + if agent._should_emit_quiet_tool_messages(): + agent._vprint(f" {_get_cute_tool_message_impl('session_search', function_args, tool_duration, result=function_result)}") + elif function_name == "memory": + target = function_args.get("target", "memory") + from tools.memory_tool import memory_tool as _memory_tool + function_result = _memory_tool( + action=function_args.get("action"), + target=target, + content=function_args.get("content"), + old_text=function_args.get("old_text"), + store=agent._memory_store, + ) + # Bridge: notify external memory provider of built-in memory writes + if agent._memory_manager and function_args.get("action") in {"add", "replace"}: + try: + agent._memory_manager.on_memory_write( + function_args.get("action", ""), + target, + function_args.get("content", ""), + metadata=agent._build_memory_write_metadata( + task_id=effective_task_id, + tool_call_id=getattr(tool_call, "id", None), + ), + ) + except Exception: + pass + tool_duration = time.time() - tool_start_time + if agent._should_emit_quiet_tool_messages(): + agent._vprint(f" {_get_cute_tool_message_impl('memory', function_args, tool_duration, result=function_result)}") + elif function_name == "clarify": + from tools.clarify_tool import clarify_tool as _clarify_tool + function_result = _clarify_tool( + question=function_args.get("question", ""), + choices=function_args.get("choices"), + callback=agent.clarify_callback, + ) + tool_duration = time.time() - tool_start_time + if agent._should_emit_quiet_tool_messages(): + agent._vprint(f" {_get_cute_tool_message_impl('clarify', function_args, tool_duration, result=function_result)}") + elif function_name == "delegate_task": + tasks_arg = function_args.get("tasks") + if tasks_arg and isinstance(tasks_arg, list): + spinner_label = f"🔀 delegating {len(tasks_arg)} tasks" + else: + goal_preview = (function_args.get("goal") or "")[:30] + spinner_label = f"🔀 {goal_preview}" if goal_preview else "🔀 delegating" + spinner = None + if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner(): + face = random.choice(KawaiiSpinner.get_waiting_faces()) + spinner = KawaiiSpinner(f"{face} {spinner_label}", spinner_type='dots', print_fn=agent._print_fn) + spinner.start() + agent._delegate_spinner = spinner + _delegate_result = None + try: + function_result = agent._dispatch_delegate_task(function_args) + _delegate_result = function_result + finally: + agent._delegate_spinner = None + tool_duration = time.time() - tool_start_time + cute_msg = _get_cute_tool_message_impl('delegate_task', function_args, tool_duration, result=_delegate_result) + if spinner: + spinner.stop(cute_msg) + elif agent._should_emit_quiet_tool_messages(): + agent._vprint(f" {cute_msg}") + elif agent._context_engine_tool_names and function_name in agent._context_engine_tool_names: + # Context engine tools (lcm_grep, lcm_describe, lcm_expand, etc.) + spinner = None + if agent._should_emit_quiet_tool_messages(): + face = random.choice(KawaiiSpinner.get_waiting_faces()) + emoji = _get_tool_emoji(function_name) + preview = _build_tool_preview(function_name, function_args) or function_name + spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn) + spinner.start() + _ce_result = None + try: + function_result = agent.context_compressor.handle_tool_call(function_name, function_args, messages=messages) + _ce_result = function_result + except Exception as tool_error: + function_result = json.dumps({"error": f"Context engine tool '{function_name}' failed: {tool_error}"}) + logger.error("context_engine.handle_tool_call raised for %s: %s", function_name, tool_error, exc_info=True) + finally: + tool_duration = time.time() - tool_start_time + cute_msg = _get_cute_tool_message_impl(function_name, function_args, tool_duration, result=_ce_result) + if spinner: + spinner.stop(cute_msg) + elif agent._should_emit_quiet_tool_messages(): + agent._vprint(f" {cute_msg}") + elif agent._memory_manager and agent._memory_manager.has_tool(function_name): + # Memory provider tools (hindsight_retain, honcho_search, etc.) + # These are not in the tool registry — route through MemoryManager. + spinner = None + if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner(): + face = random.choice(KawaiiSpinner.get_waiting_faces()) + emoji = _get_tool_emoji(function_name) + preview = _build_tool_preview(function_name, function_args) or function_name + spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn) + spinner.start() + _mem_result = None + try: + function_result = agent._memory_manager.handle_tool_call(function_name, function_args) + _mem_result = function_result + except Exception as tool_error: + function_result = json.dumps({"error": f"Memory tool '{function_name}' failed: {tool_error}"}) + logger.error("memory_manager.handle_tool_call raised for %s: %s", function_name, tool_error, exc_info=True) + finally: + tool_duration = time.time() - tool_start_time + cute_msg = _get_cute_tool_message_impl(function_name, function_args, tool_duration, result=_mem_result) + if spinner: + spinner.stop(cute_msg) + elif agent._should_emit_quiet_tool_messages(): + agent._vprint(f" {cute_msg}") + elif agent.quiet_mode: + spinner = None + if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner(): + face = random.choice(KawaiiSpinner.get_waiting_faces()) + emoji = _get_tool_emoji(function_name) + preview = _build_tool_preview(function_name, function_args) or function_name + spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn) + spinner.start() + _spinner_result = None + try: + function_result = _ra().handle_function_call( + function_name, function_args, effective_task_id, + tool_call_id=tool_call.id, + session_id=agent.session_id or "", + enabled_tools=list(agent.valid_tool_names) if agent.valid_tool_names else None, + skip_pre_tool_call_hook=True, + ) + _spinner_result = function_result + except Exception as tool_error: + function_result = f"Error executing tool '{function_name}': {tool_error}" + logger.error("handle_function_call raised for %s: %s", function_name, tool_error, exc_info=True) + finally: + tool_duration = time.time() - tool_start_time + cute_msg = _get_cute_tool_message_impl(function_name, function_args, tool_duration, result=_spinner_result) + if spinner: + spinner.stop(cute_msg) + elif agent._should_emit_quiet_tool_messages(): + agent._vprint(f" {cute_msg}") + else: + try: + function_result = _ra().handle_function_call( + function_name, function_args, effective_task_id, + tool_call_id=tool_call.id, + session_id=agent.session_id or "", + enabled_tools=list(agent.valid_tool_names) if agent.valid_tool_names else None, + skip_pre_tool_call_hook=True, + ) + except Exception as tool_error: + function_result = f"Error executing tool '{function_name}': {tool_error}" + logger.error("handle_function_call raised for %s: %s", function_name, tool_error, exc_info=True) + tool_duration = time.time() - tool_start_time + + if isinstance(function_result, str): + result_preview = function_result if agent.verbose_logging else ( + function_result[:200] if len(function_result) > 200 else function_result + ) + _result_len = len(function_result) + else: + # Multimodal dict result (_multimodal=True) — not sliceable as string + result_preview = function_result + _result_len = len(str(function_result)) + + # Log tool errors to the persistent error log so [error] tags + # in the UI always have a corresponding detailed entry on disk. + _is_error_result, _ = _detect_tool_failure(function_name, function_result) + if not _execution_blocked: + function_result = agent._append_guardrail_observation( + function_name, + function_args, + function_result, + failed=_is_error_result, + ) + result_preview = function_result if agent.verbose_logging else ( + function_result[:200] if len(function_result) > 200 else function_result + ) + if _is_error_result: + logger.warning("Tool %s returned error (%.2fs): %s", function_name, tool_duration, result_preview) + else: + logger.info("tool %s completed (%.2fs, %d chars)", function_name, tool_duration, _result_len) + + # Track file-mutation outcome for the turn-end verifier. See + # the concurrent path for the rationale; both paths must feed + # the same state so the footer reflects every tool call in the + # turn, not just the parallel ones. + if not _execution_blocked: + try: + agent._record_file_mutation_result( + function_name, function_args, function_result, _is_error_result, + ) + except Exception as _ver_err: + logging.debug("file-mutation verifier record failed: %s", _ver_err) + + if not _execution_blocked and agent.tool_progress_callback: + try: + agent.tool_progress_callback( + "tool.completed", function_name, None, None, + duration=tool_duration, is_error=_is_error_result, + ) + except Exception as cb_err: + logging.debug(f"Tool progress callback error: {cb_err}") + + agent._current_tool = None + agent._touch_activity(f"tool completed: {function_name} ({tool_duration:.1f}s)") + + if agent.verbose_logging: + logging.debug(f"Tool {function_name} completed in {tool_duration:.2f}s") + _log_result = _multimodal_text_summary(function_result) + logging.debug(f"Tool result ({len(_log_result)} chars): {_log_result}") + + if not _execution_blocked and agent.tool_complete_callback: + try: + agent.tool_complete_callback(tool_call.id, function_name, function_args, function_result) + except Exception as cb_err: + logging.debug(f"Tool complete callback error: {cb_err}") + + function_result = maybe_persist_tool_result( + content=function_result, + tool_name=function_name, + tool_use_id=tool_call.id, + env=get_active_env(effective_task_id), + ) if not _is_multimodal_tool_result(function_result) else function_result + + # Discover subdirectory context files from tool arguments + subdir_hints = agent._subdirectory_hints.check_tool_call(function_name, function_args) + if subdir_hints: + if _is_multimodal_tool_result(function_result): + _append_subdir_hint_to_multimodal(function_result, subdir_hints) + else: + function_result += subdir_hints + + # Unwrap _multimodal dicts to an OpenAI-style content list + # (see parallel path for rationale). String results pass through. + _tool_content = agent._tool_result_content_for_active_model(function_name, function_result) + tool_msg = { + "role": "tool", + "name": function_name, + "content": _tool_content, + "tool_call_id": tool_call.id + } + messages.append(tool_msg) + + # ── Per-tool /steer drain ─────────────────────────────────── + # Drain pending steer BETWEEN individual tool calls so the + # injection lands as soon as a tool finishes — not after the + # entire batch. The model sees it on the next API iteration. + agent._apply_pending_steer_to_tool_results(messages, 1) + + if not agent.quiet_mode: + if agent.verbose_logging: + print(f" ✅ Tool {i} completed in {tool_duration:.2f}s") + print(agent._wrap_verbose("Result: ", function_result)) + else: + _fr_str = function_result if isinstance(function_result, str) else str(function_result) + response_preview = _fr_str[:agent.log_prefix_chars] + "..." if len(_fr_str) > agent.log_prefix_chars else _fr_str + print(f" ✅ Tool {i} completed in {tool_duration:.2f}s - {response_preview}") + + if agent._interrupt_requested and i < len(assistant_message.tool_calls): + remaining = len(assistant_message.tool_calls) - i + agent._vprint(f"{agent.log_prefix}⚡ Interrupt: skipping {remaining} remaining tool call(s)", force=True) + for skipped_tc in assistant_message.tool_calls[i:]: + skipped_name = skipped_tc.function.name + skip_msg = { + "role": "tool", + "name": skipped_name, + "content": f"[Tool execution skipped — {skipped_name} was not started. User sent a new message]", + "tool_call_id": skipped_tc.id + } + messages.append(skip_msg) + break + + if agent.tool_delay > 0 and i < len(assistant_message.tool_calls): + time.sleep(agent.tool_delay) + + # ── Per-turn aggregate budget enforcement ───────────────────────── + num_tools_seq = len(assistant_message.tool_calls) + if num_tools_seq > 0: + enforce_turn_budget(messages[-num_tools_seq:], env=get_active_env(effective_task_id)) + + # ── /steer injection ────────────────────────────────────────────── + # See _execute_tool_calls_parallel for the rationale. Same hook, + # applied to sequential execution as well. + if num_tools_seq > 0: + agent._apply_pending_steer_to_tool_results(messages, num_tools_seq) + + + + +__all__ = [ + "execute_tool_calls_concurrent", + "execute_tool_calls_sequential", +] diff --git a/agent/transports/codex_app_server.py b/agent/transports/codex_app_server.py index b1aeaa00786..7128de9c4fa 100644 --- a/agent/transports/codex_app_server.py +++ b/agent/transports/codex_app_server.py @@ -74,12 +74,43 @@ class CodexAppServerClient: env: Optional[dict[str, str]] = None, ) -> None: self._codex_bin = codex_bin - cmd = [codex_bin, "app-server"] + list(extra_args or []) spawn_env = os.environ.copy() if env: spawn_env.update(env) if codex_home: spawn_env["CODEX_HOME"] = codex_home + + app_server_args = list(extra_args or []) + # Kanban workers must be able to write their handoff/status back to + # the board DB, which lives outside the per-task workspace. Keep the + # Codex sandbox on, but add the Kanban root as the only extra writable + # root. Without this, codex-runtime workers finish their actual work + # but crash/block when kanban_complete/kanban_block writes SQLite. + if spawn_env.get("HERMES_KANBAN_TASK"): + kanban_db = spawn_env.get("HERMES_KANBAN_DB") + kanban_root = ( + os.path.dirname(kanban_db) + if kanban_db + else spawn_env.get( + "HERMES_KANBAN_ROOT", + os.path.join( + spawn_env.get("HERMES_HOME", os.path.expanduser("~/.hermes")), + "kanban", + ), + ) + ) + app_server_args.extend( + [ + "-c", + 'sandbox_mode="workspace-write"', + "-c", + f'sandbox_workspace_write.writable_roots=["{kanban_root}"]', + "-c", + "sandbox_workspace_write.network_access=false", + ] + ) + + cmd = [codex_bin, "app-server"] + app_server_args # Codex emits tracing to stderr; default WARN keeps it quiet for users. spawn_env.setdefault("RUST_LOG", "warn") diff --git a/agent/transports/codex_app_server_session.py b/agent/transports/codex_app_server_session.py index f0cd0a196c4..d9ee92dfbf5 100644 --- a/agent/transports/codex_app_server_session.py +++ b/agent/transports/codex_app_server_session.py @@ -404,7 +404,7 @@ class CodexAppServerSession: return result result.turn_id = (ts.get("turn") or {}).get("id") - deadline = time.time() + turn_timeout + deadline = time.monotonic() + turn_timeout turn_complete = False # Post-tool watchdog state. last_tool_completion_at is set whenever # a tool-shaped item completes; if no further notification arrives @@ -412,7 +412,7 @@ class CodexAppServerSession: # fast-fail and retire the session. last_tool_completion_at: Optional[float] = None - while time.time() < deadline and not turn_complete: + while time.monotonic() < deadline and not turn_complete: if self._interrupt_event.is_set(): self._issue_interrupt(result.turn_id) result.interrupted = True @@ -440,7 +440,7 @@ class CodexAppServerSession: # up on this turn instead of waiting for the outer deadline. if ( last_tool_completion_at is not None - and (time.time() - last_tool_completion_at) + and (time.monotonic() - last_tool_completion_at) > post_tool_quiet_timeout ): self._issue_interrupt(result.turn_id) @@ -471,7 +471,7 @@ class CodexAppServerSession: result.projected_messages.extend(proj.messages) if proj.is_tool_iteration: result.tool_iterations += 1 - last_tool_completion_at = time.time() + last_tool_completion_at = time.monotonic() if proj.final_text is not None: result.final_text = proj.final_text if _has_turn_aborted_marker(proj.final_text): @@ -514,7 +514,7 @@ class CodexAppServerSession: result.tool_iterations += 1 # Arm/refresh the post-tool quiet watchdog whenever a # tool-shaped item completes. - last_tool_completion_at = time.time() + last_tool_completion_at = time.monotonic() else: # Any non-tool projected activity (assistant message, # status update, etc.) means codex is still producing @@ -541,7 +541,7 @@ class CodexAppServerSession: turn_status = ( (note.get("params") or {}).get("turn") or {} ).get("status") - if turn_status and turn_status not in ("completed", "interrupted"): + if turn_status and turn_status not in {"completed", "interrupted"}: err_obj = ( (note.get("params") or {}).get("turn") or {} ).get("error") @@ -775,9 +775,9 @@ def _approval_choice_to_codex_decision(choice: str) -> str: (verified against codex-rs/app-server-protocol/src/protocol/v2/item.rs on codex 0.130.0). """ - if choice in ("once",): + if choice in {"once",}: return "accept" - if choice in ("session", "always"): + if choice in {"session", "always"}: return "acceptForSession" return "decline" diff --git a/apps/dashboard/src/components/ChatSidebar.tsx b/apps/dashboard/src/components/ChatSidebar.tsx index 87af497237a..c78b2bf5fd9 100644 --- a/apps/dashboard/src/components/ChatSidebar.tsx +++ b/apps/dashboard/src/components/ChatSidebar.tsx @@ -30,6 +30,7 @@ import { Card } from "@/components/ui/card"; import { ModelPickerDialog } from "@/components/ModelPickerDialog"; import { ToolCall, type ToolEntry } from "@/components/ToolCall"; import { GatewayClient, type ConnectionState } from "@/lib/gatewayClient"; +import { HERMES_BASE_PATH } from "@/lib/api"; import { cn } from "@/lib/utils"; import { AlertCircle, ChevronDown, RefreshCw } from "lucide-react"; @@ -160,7 +161,7 @@ export function ChatSidebar({ channel, className }: ChatSidebarProps) { const proto = window.location.protocol === "https:" ? "wss:" : "ws:"; const qs = new URLSearchParams({ token, channel }); const ws = new WebSocket( - `${proto}//${window.location.host}/api/events?${qs.toString()}`, + `${proto}//${window.location.host}${HERMES_BASE_PATH}/api/events?${qs.toString()}`, ); // `unmounting` suppresses the banner during cleanup — `ws.close()` diff --git a/apps/dashboard/src/lib/gatewayClient.ts b/apps/dashboard/src/lib/gatewayClient.ts index e3a1486e084..3c8cdd76035 100644 --- a/apps/dashboard/src/lib/gatewayClient.ts +++ b/apps/dashboard/src/lib/gatewayClient.ts @@ -5,6 +5,8 @@ import { type GatewayEventName, } from "@hermes/shared"; +import { HERMES_BASE_PATH } from "@/lib/api"; + export type { ConnectionState, GatewayEvent, GatewayEventName }; /** @@ -24,7 +26,7 @@ export class GatewayClient extends JsonRpcGatewayClient { const scheme = location.protocol === "https:" ? "wss:" : "ws:"; await super.connect( - `${scheme}//${location.host}/api/ws?token=${encodeURIComponent(resolved)}`, + `${scheme}//${location.host}${HERMES_BASE_PATH}/api/ws?token=${encodeURIComponent(resolved)}`, ); } } diff --git a/apps/dashboard/src/pages/ChatPage.tsx b/apps/dashboard/src/pages/ChatPage.tsx index 0d092c72c04..3e3c2e3268b 100644 --- a/apps/dashboard/src/pages/ChatPage.tsx +++ b/apps/dashboard/src/pages/ChatPage.tsx @@ -24,6 +24,7 @@ import { Terminal } from "@xterm/xterm"; import "@xterm/xterm/css/xterm.css"; import { Button } from "@nous-research/ui/ui/components/button"; import { Typography } from "@/components/NouiTypography"; +import { HERMES_BASE_PATH } from "@/lib/api"; import { cn } from "@/lib/utils"; import { Copy, PanelRight, X } from "lucide-react"; import { useCallback, useEffect, useMemo, useRef, useState } from "react"; @@ -44,7 +45,7 @@ function buildWsUrl( const proto = window.location.protocol === "https:" ? "wss:" : "ws:"; const qs = new URLSearchParams({ token, channel }); if (resume) qs.set("resume", resume); - return `${proto}//${window.location.host}/api/pty?${qs.toString()}`; + return `${proto}//${window.location.host}${HERMES_BASE_PATH}/api/pty?${qs.toString()}`; } // Channel id ties this chat tab's PTY child (publisher) to its sidebar @@ -286,6 +287,17 @@ export default function ChatPage({ isActive = true }: { isActive?: boolean }) { fontWeight: "400", fontWeightBold: "700", macOptionIsMeta: true, + // Hold Option (Alt on Linux/Windows) to force native text selection + // even when the inner Hermes TUI has enabled xterm mouse-events + // mode (CSI ?1000h family). Without this, click-and-drag in the + // chat canvas selects nothing and Cmd+C falls back to copying the + // entire visible buffer, which is rarely what the user wants. + // See #25720. + macOptionClickForcesSelection: true, + // Right-click selects the word under the pointer. xterm.js default + // is false; enabling it gives users a single-action selection + // path on top of the modifier-based bypass above. + rightClickSelectsWord: true, // Single-scroll-system experiment: // let the inner Hermes TUI own transcript history/scroll behavior. // The outer browser xterm should act as a display/input bridge only. diff --git a/cli.py b/cli.py index 00b3af44df6..f42ba973835 100644 --- a/cli.py +++ b/cli.py @@ -1396,7 +1396,7 @@ def _detect_light_mode() -> bool: last = cfgbg.split(";")[-1] if ";" in cfgbg else cfgbg if last.isdigit(): bg = int(last) - if bg in (7, 15): + if bg in {7, 15}: result = True _LIGHT_MODE_CACHE = result return result @@ -2412,6 +2412,7 @@ def _looks_like_slash_command(text: str) -> bool: from agent.skill_commands import ( scan_skill_commands, + get_skill_commands, build_skill_invocation_message, build_preloaded_skills_prompt, ) @@ -2824,6 +2825,11 @@ class HermesCLI: # turn (which would make Ctrl+C feel like it did nothing). self._last_turn_interrupted = False self._should_exit = False + # /exit --delete: when True, the current session's SQLite history and + # on-disk transcripts are deleted during shutdown. Set by + # process_command() when the user runs /exit --delete or /quit --delete. + # Ported from google-gemini/gemini-cli#19332. + self._delete_session_on_exit = False self._last_ctrl_c_time = 0 self._clarify_state = None self._clarify_freetext = False @@ -7653,6 +7659,16 @@ class HermesCLI: canonical = _cmd_def.name if _cmd_def else _base_word if canonical in {"quit", "exit"}: + # Parse --delete flag: /exit --delete also removes the current + # session's transcripts + SQLite history. Ported from + # google-gemini/gemini-cli#19332. + _rest = cmd_original.split(None, 1) + _args = (_rest[1] if len(_rest) > 1 else "").strip().lower() + if _args in {"--delete", "-d"}: + self._delete_session_on_exit = True + elif _args: + _cprint(f" {_DIM}✗ Unknown argument: {_escape(_args)}. Use /exit --delete to also remove session history.{_RST}") + return True return False elif canonical == "help": self.show_help() @@ -9598,12 +9614,18 @@ class HermesCLI: prompt caching intact. """ try: - from agent.skill_commands import reload_skills + from agent.skill_commands import reload_skills, get_skill_commands if not self._command_running: print("🔄 Reloading skills...") result = reload_skills() + + # Sync cli.py's module-level _skill_commands so all consumers + # (help display, command dispatch, Tab-completion lambda) see the + # updated dict without needing to restart the session. + global _skill_commands + _skill_commands = get_skill_commands() added = result.get("added", []) # [{"name", "description"}, ...] removed = result.get("removed", []) # [{"name", "description"}, ...] total = result.get("total", 0) @@ -12609,7 +12631,7 @@ class HermesCLI: _completer = SlashCommandCompleter( - skill_commands_provider=lambda: _skill_commands, + skill_commands_provider=lambda: get_skill_commands(), command_filter=cli_ref._command_available, ) input_area = TextArea( @@ -13777,7 +13799,7 @@ class HermesCLI: if _errno == errno.EIO: pass # suppress broken-stdout I/O errors on interrupt (#13710) elif ( - _errno in (errno.EINVAL, errno.EBADF) + _errno in {errno.EINVAL, errno.EBADF} or "is not registered" in _msg or "Bad file descriptor" in _msg or "Invalid argument" in _msg @@ -13824,6 +13846,19 @@ class HermesCLI: self._session_db.end_session(self.agent.session_id, "cli_close") except (Exception, KeyboardInterrupt) as e: logger.debug("Could not close session in DB: %s", e) + # /exit --delete: also remove the current session's transcripts + # and SQLite history. Ported from google-gemini/gemini-cli#19332. + if getattr(self, '_delete_session_on_exit', False): + try: + from hermes_constants import get_hermes_home as _ghh + _sessions_dir = _ghh() / "sessions" + _sid = self.agent.session_id + if self._session_db.delete_session(_sid, sessions_dir=_sessions_dir): + _cprint(f" {_DIM}✓ Session {_escape(_sid)} deleted{_RST}") + else: + _cprint(f" {_DIM}✗ Session {_escape(_sid)} not found for deletion{_RST}") + except (Exception, KeyboardInterrupt) as e: + logger.debug("Could not delete session on exit: %s", e) # Plugin hook: on_session_end — safety net for interrupted exits. # run_conversation() already fires this per-turn on normal completion, # so only fire here if the agent was mid-turn (_agent_running) when diff --git a/cron/scheduler.py b/cron/scheduler.py index d470e8c2c74..322fa64906f 100644 --- a/cron/scheduler.py +++ b/cron/scheduler.py @@ -1802,7 +1802,12 @@ def tick(verbose: bool = True, adapters=None, loop=None) -> int: for job in parallel_jobs: _ctx = contextvars.copy_context() _futures.append(_tick_pool.submit(_ctx.run, _process_job, job)) - _results.extend(f.result() for f in _futures) + for f in concurrent.futures.as_completed(_futures, timeout=600): + try: + _results.append(f.result()) + except Exception as exc: + logger.error("Parallel cron job future failed: %s", exc) + _results.append(False) # Best-effort sweep of MCP stdio subprocesses that survived their # session teardown during this tick. Runs AFTER every job has diff --git a/gateway/platforms/api_server.py b/gateway/platforms/api_server.py index 809d6cd8a03..0668896e170 100644 --- a/gateway/platforms/api_server.py +++ b/gateway/platforms/api_server.py @@ -71,6 +71,35 @@ def _coerce_port(value: Any, default: int = DEFAULT_PORT) -> int: return default +_TRUE_REQUEST_BOOL_STRINGS = frozenset({"1", "true", "yes", "on"}) +_FALSE_REQUEST_BOOL_STRINGS = frozenset({"0", "false", "no", "off"}) + + +def _coerce_request_bool(value: Any, default: bool = False) -> bool: + """Normalize boolean-like API payload values. + + External clients should send real JSON booleans, but some OpenAI-compatible + frontends and middleware serialize flags like ``stream`` as strings. Using + Python truthiness on those values misroutes requests because ``"false"`` is + still truthy. Treat only explicit bool-ish scalars as booleans; everything + else falls back to the caller's default. + """ + if isinstance(value, bool): + return value + if value is None: + return default + if isinstance(value, str): + normalized = value.strip().lower() + if normalized in _TRUE_REQUEST_BOOL_STRINGS: + return True + if normalized in _FALSE_REQUEST_BOOL_STRINGS: + return False + return default + if isinstance(value, (int, float)): + return bool(value) + return default + + def _normalize_chat_content( content: Any, *, _max_depth: int = 10, _depth: int = 0, ) -> str: @@ -481,7 +510,12 @@ else: body_limit_middleware = None # type: ignore[assignment] _SECURITY_HEADERS = { + "Content-Security-Policy": "default-src 'none'; frame-ancestors 'none'", + "Permissions-Policy": "camera=(), microphone=(), geolocation=()", + "Strict-Transport-Security": "max-age=31536000; includeSubDomains", "X-Content-Type-Options": "nosniff", + "X-Frame-Options": "DENY", + "X-XSS-Protection": "0", "Referrer-Policy": "no-referrer", } @@ -1005,7 +1039,7 @@ class APIServerAdapter(BasePlatformAdapter): status=400, ) - stream = body.get("stream", False) + stream = _coerce_request_bool(body.get("stream"), default=False) # Extract system message (becomes ephemeral system prompt layered ON TOP of core) system_prompt = None @@ -2082,7 +2116,7 @@ class APIServerAdapter(BasePlatformAdapter): instructions = body.get("instructions") previous_response_id = body.get("previous_response_id") conversation = body.get("conversation") - store = body.get("store", True) + store = _coerce_request_bool(body.get("store"), default=True) # conversation and previous_response_id are mutually exclusive if conversation and previous_response_id: @@ -2165,7 +2199,7 @@ class APIServerAdapter(BasePlatformAdapter): # groups the entire conversation under one session entry. session_id = stored_session_id or str(uuid.uuid4()) - stream = bool(body.get("stream", False)) + stream = _coerce_request_bool(body.get("stream"), default=False) if stream: # Streaming branch — emit OpenAI Responses SSE events as the # agent runs so frontends can render text deltas and tool @@ -3228,7 +3262,10 @@ class APIServerAdapter(BasePlatformAdapter): status=409, ) - resolve_all = bool(body.get("all") or body.get("resolve_all")) + resolve_all = ( + _coerce_request_bool(body.get("all"), default=False) + or _coerce_request_bool(body.get("resolve_all"), default=False) + ) try: from tools.approval import resolve_gateway_approval diff --git a/gateway/platforms/base.py b/gateway/platforms/base.py index 7b3147e21f4..96b56d29cc7 100644 --- a/gateway/platforms/base.py +++ b/gateway/platforms/base.py @@ -2014,6 +2014,13 @@ class BasePlatformAdapter(ABC): text = f"{caption}\n{text}" return await self.send(chat_id=chat_id, content=text, reply_to=reply_to, metadata=metadata) + def prepare_tts_text(self, text: str) -> str: + """Prepare text for TTS. Override to filter tool output, code, etc. + + Default strips markdown formatting and truncates to 4000 chars. + """ + return re.sub(r'[*_`#\[\]()]', '', text)[:4000].strip() + async def play_tts( self, chat_id: str, @@ -3144,7 +3151,7 @@ class BasePlatformAdapter(ABC): from tools.tts_tool import text_to_speech_tool, check_tts_requirements if check_tts_requirements(): import json as _json - speech_text = re.sub(r'[*_`#\[\]()]', '', text_content)[:4000].strip() + speech_text = self.prepare_tts_text(text_content) if not speech_text: raise ValueError("Empty text after markdown cleanup") tts_result_str = await asyncio.to_thread( diff --git a/gateway/platforms/discord.py b/gateway/platforms/discord.py index 9b8285e2a36..f79678bc61a 100644 --- a/gateway/platforms/discord.py +++ b/gateway/platforms/discord.py @@ -3639,18 +3639,18 @@ class DiscordAdapter(BasePlatformAdapter): configured = self.config.extra.get("thread_require_mention") if configured is not None: if isinstance(configured, str): - return configured.lower() not in ("false", "0", "no", "off") + return configured.lower() not in {"false", "0", "no", "off"} return bool(configured) - return os.getenv("DISCORD_THREAD_REQUIRE_MENTION", "false").lower() in ("true", "1", "yes", "on") + return os.getenv("DISCORD_THREAD_REQUIRE_MENTION", "false").lower() in {"true", "1", "yes", "on"} def _discord_history_backfill(self) -> bool: """Return whether history backfill is enabled for shared sessions.""" configured = self.config.extra.get("history_backfill") if configured is not None: if isinstance(configured, str): - return configured.lower() not in ("false", "0", "no", "off") + return configured.lower() not in {"false", "0", "no", "off"} return bool(configured) - return os.getenv("DISCORD_HISTORY_BACKFILL", "true").lower() in ("true", "1", "yes") + return os.getenv("DISCORD_HISTORY_BACKFILL", "true").lower() in {"true", "1", "yes"} def _discord_history_backfill_limit(self) -> int: """Return the max number of messages to scan backwards for context. @@ -3737,7 +3737,7 @@ class DiscordAdapter(BasePlatformAdapter): break # Skip system messages (pins, joins, thread renames, etc.) - if msg.type not in (discord.MessageType.default, discord.MessageType.reply): + if msg.type not in {discord.MessageType.default, discord.MessageType.reply}: continue # Respect DISCORD_ALLOW_BOTS for other bots. diff --git a/gateway/platforms/helpers.py b/gateway/platforms/helpers.py index 1c4f451585a..a3704bf50cf 100644 --- a/gateway/platforms/helpers.py +++ b/gateway/platforms/helpers.py @@ -168,8 +168,8 @@ class TextBatchAggregator: # Pre-compiled regexes for performance _RE_BOLD = re.compile(r"\*\*(.+?)\*\*", re.DOTALL) _RE_ITALIC_STAR = re.compile(r"\*(.+?)\*", re.DOTALL) -_RE_BOLD_UNDER = re.compile(r"__(.+?)__", re.DOTALL) -_RE_ITALIC_UNDER = re.compile(r"_(.+?)_", re.DOTALL) +_RE_BOLD_UNDER = re.compile(r"\b__(?![\s_])(.+?)(?30s past startup (initial-sync replay window closed) + # - the skew is *consistent* across consecutive drops, which is + # the signature of a constant clock offset rather than a + # variable-age room history. Backfill from a freshly invited + # room can deliver events spanning hours/days — those skews + # will be all over the place and reset the counter. + if not self._clock_skew_warned and ( + time.time() - self._startup_ts > 30 + ): + skew = self._startup_ts - event_ts + # Sanity bound: malformed events with negative or absurd + # timestamps shouldn't count. + if 5 < skew < 86400: + if self._late_grace_drops == 0: + self._late_grace_skew = skew + self._late_grace_drops = 1 + elif abs(skew - self._late_grace_skew) < 60: + # Consistent offset → likely real clock skew. + self._late_grace_drops += 1 + else: + # Varied skew → likely backfill, restart sampling. + self._late_grace_skew = skew + self._late_grace_drops = 1 + if self._late_grace_drops >= 3: + logger.warning( + "Matrix: dropped %d consecutive live events as " + "'too old' more than 30s after startup (skew " + "≈ %.0fs). The host system clock is likely set " + "ahead of real time, which causes the startup " + "grace filter to silently discard every incoming " + "message. Run `timedatectl set-ntp true` (or " + "sync NTP) and restart the bot.", + self._late_grace_drops, + skew, + ) + self._clock_skew_warned = True return # Extract content from the event. diff --git a/gateway/platforms/slack.py b/gateway/platforms/slack.py index 2116b569f96..5accfdb4108 100644 --- a/gateway/platforms/slack.py +++ b/gateway/platforms/slack.py @@ -482,7 +482,7 @@ class SlackAdapter(BasePlatformAdapter): "text": text, } try: - async with aiohttp.ClientSession() as session: + async with aiohttp.ClientSession(trust_env=True) as session: async with session.post( ctx["response_url"], json=payload, diff --git a/gateway/platforms/sms.py b/gateway/platforms/sms.py index 2cf7db69b74..9d9957d5ea1 100644 --- a/gateway/platforms/sms.py +++ b/gateway/platforms/sms.py @@ -128,6 +128,7 @@ class SmsAdapter(BasePlatformAdapter): await site.start() self._http_session = aiohttp.ClientSession( timeout=aiohttp.ClientTimeout(total=30), + trust_env=True, ) self._running = True @@ -169,6 +170,7 @@ class SmsAdapter(BasePlatformAdapter): session = self._http_session or aiohttp.ClientSession( timeout=aiohttp.ClientTimeout(total=30), + trust_env=True, ) try: for chunk in chunks: diff --git a/gateway/platforms/telegram.py b/gateway/platforms/telegram.py index 4c56937e5cb..77af24765d9 100644 --- a/gateway/platforms/telegram.py +++ b/gateway/platforms/telegram.py @@ -1663,7 +1663,17 @@ class TelegramAdapter(BasePlatformAdapter): continue raise message_ids.append(str(msg.message_id)) - + + # Re-trigger typing indicator after sending a message. + # Telegram clears the typing state when a new message is delivered, + # so without this the "...typing" bubble disappears mid-response + # (especially noticeable when the agent sends intermediate progress + # messages like "Checking:" before running tools). + try: + await self.send_typing(chat_id, metadata=metadata) + except Exception: + pass # Typing failures are non-fatal + return SendResult( success=True, message_id=message_ids[0] if message_ids else None, diff --git a/gateway/run.py b/gateway/run.py index f9a282a413f..25156932bfe 100644 --- a/gateway/run.py +++ b/gateway/run.py @@ -4763,11 +4763,106 @@ class GatewayRunner: pass return False + # Auto-decompose: turn fresh triage tasks into ready workgraphs + # before the dispatcher fans out workers. Gated by + # ``kanban.auto_decompose`` (default True). Capped by + # ``kanban.auto_decompose_per_tick`` (default 3) so a bulk-load + # of triage tasks doesn't burst-spend the aux LLM in one tick; + # remainder defers to subsequent ticks. + auto_decompose_enabled = bool(kanban_cfg.get("auto_decompose", True)) + try: + auto_decompose_per_tick = int( + kanban_cfg.get("auto_decompose_per_tick", 3) or 3 + ) + except (TypeError, ValueError): + auto_decompose_per_tick = 3 + if auto_decompose_per_tick < 1: + auto_decompose_per_tick = 1 + + def _auto_decompose_tick() -> int: + """Run the auto-decomposer for up to N triage tasks across all + boards. Returns the number of triage tasks that were + successfully decomposed or specified this tick. + """ + try: + from hermes_cli import kanban_decompose as _decomp + except Exception as exc: # pragma: no cover + logger.warning( + "kanban auto-decompose: import failed (%s); skipping", exc, + ) + return 0 + try: + boards = _kb.list_boards(include_archived=False) + except Exception: + boards = [_kb.read_board_metadata(_kb.DEFAULT_BOARD)] + attempted = 0 + successes = 0 + for b in boards: + slug = b.get("slug") or _kb.DEFAULT_BOARD + if attempted >= auto_decompose_per_tick: + break + # Pin this board for the duration of the call — same + # pattern as the dashboard specify endpoint. The + # decomposer module connects with no board kwarg and + # relies on the env var. + prev_env = os.environ.get("HERMES_KANBAN_BOARD") + try: + os.environ["HERMES_KANBAN_BOARD"] = slug + try: + triage_ids = _decomp.list_triage_ids() + except Exception as exc: + logger.debug( + "kanban auto-decompose: list_triage_ids failed on board %s (%s)", + slug, exc, + ) + triage_ids = [] + for tid in triage_ids: + if attempted >= auto_decompose_per_tick: + break + attempted += 1 + try: + outcome = _decomp.decompose_task( + tid, author="auto-decomposer", + ) + except Exception: + logger.exception( + "kanban auto-decompose: decompose_task crashed on %s", + tid, + ) + continue + if outcome.ok: + successes += 1 + if outcome.fanout and outcome.child_ids: + logger.info( + "kanban auto-decompose [%s]: %s → %d children", + slug, tid, len(outcome.child_ids), + ) + else: + logger.info( + "kanban auto-decompose [%s]: %s → single task (no fanout)", + slug, tid, + ) + else: + # Common no-op reasons (no aux client configured) shouldn't + # spam logs every tick. Log at debug. + logger.debug( + "kanban auto-decompose [%s]: %s skipped: %s", + slug, tid, outcome.reason, + ) + finally: + if prev_env is None: + os.environ.pop("HERMES_KANBAN_BOARD", None) + else: + os.environ["HERMES_KANBAN_BOARD"] = prev_env + return successes + logger.info( "kanban dispatcher: embedded in gateway (interval=%.1fs)", interval ) while self._running: try: + if auto_decompose_enabled: + await asyncio.to_thread(_auto_decompose_tick) results = await asyncio.to_thread(_tick_once) any_spawned = False for slug, res in (results or []): @@ -8845,7 +8940,7 @@ class GatewayRunner: lines.append("Failed/paused: (none)") return "\n".join(lines) - if action in ("pause", "resume"): + if action in {"pause", "resume"}: if not target: return f"Usage: /platform {action} " platform = _resolve_platform(target) @@ -8953,13 +9048,15 @@ class GatewayRunner: logger.debug("Failed to write restart dedup marker: %s", e) active_agents = self._running_agent_count() - # When running under a service manager (systemd/launchd), use the - # service restart path: exit with code 75 so the service manager - # restarts us. The detached subprocess approach (setsid + bash) - # doesn't work under systemd because KillMode=mixed kills all - # processes in the cgroup, including the detached helper. + # When running under a service manager (systemd/launchd) or inside a + # Docker/Podman container, use the service restart path: exit with + # code 75 so the service manager / container restart policy restarts + # us. The detached subprocess approach (setsid + bash) doesn't work + # under systemd (KillMode=mixed kills the cgroup) or Docker (tini + # exits when the gateway dies, taking the detached helper with it). _under_service = bool(os.environ.get("INVOCATION_ID")) # systemd sets this - if _under_service: + _in_container = os.path.exists("/.dockerenv") or os.path.exists("/run/.containerenv") + if _under_service or _in_container: self.request_restart(detached=False, via_service=True) else: self.request_restart(detached=True, via_service=False) @@ -12528,6 +12625,12 @@ class GatewayRunner: and getattr(source, "chat_type", None) == "dm" ): metadata["telegram_dm_topic_reply_fallback"] = True + # Telegram DM topic lanes need direct_messages_topic_id in metadata + # so synthetic/queued messages (goal continuations, status notices) + # route to the correct topic even when reply anchor is unavailable. + tid = str(thread_id) + if tid and tid not in {"", "1"}: + metadata["direct_messages_topic_id"] = tid anchor = reply_to_message_id or getattr(source, "message_id", None) if anchor is not None: metadata["telegram_reply_to_message_id"] = str(anchor) @@ -12813,7 +12916,11 @@ class GatewayRunner: update_cmd = ( f"PYTHONUNBUFFERED=1 {hermes_cmd_str} update --gateway" f" > {shlex.quote(str(output_path))} 2>&1; " - f"status=$?; printf '%s' \"$status\" > {shlex.quote(str(exit_code_path))}" + # Avoid `status=$?`: `status` is a read-only special parameter + # in zsh, and this command string is copied/reused in macOS/zsh + # operator wrappers. Keep the template zsh-safe even though this + # specific subprocess currently runs under bash. + f"rc=$?; printf '%s' \"$rc\" > {shlex.quote(str(exit_code_path))}" ) setsid_bin = shutil.which("setsid") if setsid_bin: diff --git a/hermes_cli/auth.py b/hermes_cli/auth.py index 6cabb61570d..cb97a4c2300 100644 --- a/hermes_cli/auth.py +++ b/hermes_cli/auth.py @@ -11,6 +11,12 @@ Architecture: - resolve_provider() picks the active provider via priority chain - resolve_*_runtime_credentials() handles token refresh and key minting - logout_command() is the CLI entry point for clearing auth + +Nous authentication paths: +- Invoke JWT (preferred): use a scoped access_token directly for inference. +- Legacy session key (fallback): mint an opaque 24h key when JWT auth is + unavailable, or when HERMES_AGENT_USE_LEGACY_SESSION_KEYS is set for + debugging or rollback. """ from __future__ import annotations @@ -35,7 +41,7 @@ from dataclasses import dataclass, field from datetime import datetime, timezone from http.server import BaseHTTPRequestHandler, HTTPServer from pathlib import Path -from typing import Any, Dict, List, Optional, Tuple +from typing import Any, Callable, Dict, List, Optional, Tuple from urllib.parse import parse_qs, urlencode, urlparse import httpx @@ -67,9 +73,25 @@ AUTH_LOCK_TIMEOUT_SECONDS = 15.0 DEFAULT_NOUS_PORTAL_URL = "https://portal.nousresearch.com" DEFAULT_NOUS_INFERENCE_URL = "https://inference-api.nousresearch.com/v1" DEFAULT_NOUS_CLIENT_ID = "hermes-cli" -DEFAULT_NOUS_SCOPE = "inference:mint_agent_key" +NOUS_LEGACY_AGENT_KEY_SCOPE = "inference:mint_agent_key" +NOUS_INFERENCE_INVOKE_SCOPE = "inference:invoke" +DEFAULT_NOUS_SCOPE = f"{NOUS_INFERENCE_INVOKE_SCOPE} {NOUS_LEGACY_AGENT_KEY_SCOPE}" +NOUS_LEGACY_SESSION_KEYS_ENV = "HERMES_AGENT_USE_LEGACY_SESSION_KEYS" +NOUS_DEVICE_CODE_SOURCE = "device_code" +NOUS_INFERENCE_AUTH_MODE_AUTO = "auto" +NOUS_INFERENCE_AUTH_MODE_FRESH = "fresh" +NOUS_INFERENCE_AUTH_MODE_LEGACY = "legacy" +NOUS_INFERENCE_AUTH_MODES = frozenset({ + NOUS_INFERENCE_AUTH_MODE_AUTO, + NOUS_INFERENCE_AUTH_MODE_FRESH, + NOUS_INFERENCE_AUTH_MODE_LEGACY, +}) +NOUS_AUTH_PATH_INVOKE_JWT = "invoke_jwt" +NOUS_AUTH_PATH_LEGACY_SESSION_KEY_CACHE = "legacy_session_key_cache" +NOUS_AUTH_PATH_LEGACY_SESSION_KEY_MINT = "legacy_session_key_mint" DEFAULT_AGENT_KEY_MIN_TTL_SECONDS = 30 * 60 # 30 minutes ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120 # refresh 2 min before expiry +NOUS_INVOKE_JWT_MIN_TTL_SECONDS = ACCESS_TOKEN_REFRESH_SKEW_SECONDS DEVICE_AUTH_POLL_INTERVAL_CAP_SECONDS = 1 # poll at most every 1s DEFAULT_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex" DEFAULT_XAI_OAUTH_BASE_URL = "https://api.x.ai/v1" @@ -1549,6 +1571,255 @@ def _decode_jwt_claims(token: Any) -> Dict[str, Any]: return claims if isinstance(claims, dict) else {} +def _scope_values(raw_scope: Any) -> set[str]: + # OAuth token responses normally return a space-separated string. Keep + # collection support for JWT ``scp`` claims and older stored test fixtures. + scopes: set[str] = set() + if isinstance(raw_scope, str): + for part in raw_scope.replace(",", " ").split(): + cleaned = part.strip() + if cleaned: + scopes.add(cleaned) + elif isinstance(raw_scope, (list, tuple, set, frozenset)): + for item in raw_scope: + if isinstance(item, str): + scopes.update(_scope_values(item)) + return scopes + + +def _nous_legacy_session_keys_forced() -> bool: + return is_truthy_value(os.getenv(NOUS_LEGACY_SESSION_KEYS_ENV), default=False) + + +def _nous_scope_has_invoke(raw_scope: Any) -> bool: + return NOUS_INFERENCE_INVOKE_SCOPE in _scope_values(raw_scope) + + +def _normalize_nous_inference_auth_mode(inference_auth_mode: Optional[str]) -> str: + mode = str(inference_auth_mode or NOUS_INFERENCE_AUTH_MODE_AUTO).strip().lower() + if mode not in NOUS_INFERENCE_AUTH_MODES: + allowed = ", ".join(sorted(NOUS_INFERENCE_AUTH_MODES)) + raise ValueError( + "Invalid Nous inference auth mode " + f"{inference_auth_mode!r}; expected one of: {allowed}" + ) + return mode + + +def _nous_invoke_jwt_status( + token: Any, + *, + scope: Any = None, + expires_at: Any = None, + min_ttl_seconds: int = NOUS_INVOKE_JWT_MIN_TTL_SECONDS, +) -> Optional[str]: + """Return None when the token can be used for inference, else a reason.""" + claims = _decode_jwt_claims(token) + if not claims: + return "access_token_not_jwt" + scopes = ( + _scope_values(scope) + | _scope_values(claims.get("scope")) + | _scope_values(claims.get("scp")) + ) + if NOUS_INFERENCE_INVOKE_SCOPE not in scopes: + return "missing_inference_invoke_scope" + exp = claims.get("exp") + skew = max(0, int(min_ttl_seconds)) + if isinstance(exp, (int, float)): + if float(exp) <= (time.time() + skew): + return "invoke_jwt_expiring" + return None + if _is_expiring(expires_at, skew): + return "invoke_jwt_expiry_unknown_or_expiring" + return None + + +def _nous_invoke_jwt_is_usable( + token: Any, + *, + scope: Any = None, + expires_at: Any = None, + min_ttl_seconds: int = NOUS_INVOKE_JWT_MIN_TTL_SECONDS, +) -> bool: + return ( + _nous_invoke_jwt_status( + token, + scope=scope, + expires_at=expires_at, + min_ttl_seconds=min_ttl_seconds, + ) + is None + ) + + +def _nous_legacy_session_key_reason( + token: Any, + *, + scope: Any = None, + expires_at: Any = None, + inference_auth_mode: str = NOUS_INFERENCE_AUTH_MODE_AUTO, +) -> str: + if inference_auth_mode == NOUS_INFERENCE_AUTH_MODE_LEGACY: + return "forced_legacy_session_key" + if _nous_legacy_session_keys_forced(): + return "forced_legacy_session_keys" + return ( + _nous_invoke_jwt_status(token, scope=scope, expires_at=expires_at) + or "invoke_jwt_unavailable" + ) + + +def _choose_nous_inference_auth_path( + state: Dict[str, Any], + *, + access_token: Any = None, + min_key_ttl_seconds: int = DEFAULT_AGENT_KEY_MIN_TTL_SECONDS, + inference_auth_mode: str = NOUS_INFERENCE_AUTH_MODE_AUTO, +) -> Tuple[str, Optional[str]]: + inference_auth_mode = _normalize_nous_inference_auth_mode(inference_auth_mode) + token = state.get("access_token") if access_token is None else access_token + if ( + not _nous_legacy_session_keys_forced() + and inference_auth_mode != NOUS_INFERENCE_AUTH_MODE_LEGACY + and _nous_invoke_jwt_is_usable( + token, + scope=state.get("scope"), + expires_at=state.get("expires_at"), + ) + ): + return NOUS_AUTH_PATH_INVOKE_JWT, None + if ( + inference_auth_mode == NOUS_INFERENCE_AUTH_MODE_AUTO + and _agent_key_is_usable( + state, + max(60, int(min_key_ttl_seconds)), + ) + ): + return NOUS_AUTH_PATH_LEGACY_SESSION_KEY_CACHE, None + return ( + NOUS_AUTH_PATH_LEGACY_SESSION_KEY_MINT, + _nous_legacy_session_key_reason( + token, + scope=state.get("scope"), + expires_at=state.get("expires_at"), + inference_auth_mode=inference_auth_mode, + ), + ) + + +def _log_nous_invoke_jwt_selected( + *, + access_token: Any, + sequence_id: Optional[str] = None, +) -> None: + logger.info("Nous inference auth: using NAS invoke JWT") + _oauth_trace( + "nous_invoke_jwt_selected", + sequence_id=sequence_id, + access_token_fp=_token_fingerprint(access_token), + ) + + +def _log_nous_legacy_session_key_selected( + reason: str, + *, + access_token: Any, + sequence_id: Optional[str] = None, +) -> None: + logger.info( + "Nous inference auth: using legacy session key path (%s)", + reason, + ) + _oauth_trace( + "nous_legacy_session_key_selected", + sequence_id=sequence_id, + reason=reason, + access_token_fp=_token_fingerprint(access_token), + ) + + +def _nous_jwt_expires_at(token: Any, fallback_expires_at: Any = None) -> Optional[str]: + claims = _decode_jwt_claims(token) + exp = claims.get("exp") + if isinstance(exp, (int, float)): + try: + return datetime.fromtimestamp(float(exp), tz=timezone.utc).isoformat() + except Exception: + pass + return fallback_expires_at if isinstance(fallback_expires_at, str) else None + + +def _set_nous_agent_key_from_invoke_jwt( + state: Dict[str, Any], + *, + obtained_at: Optional[str] = None, +) -> None: + access_token = state.get("access_token") + if not isinstance(access_token, str) or not access_token.strip(): + return + now = datetime.now(timezone.utc) + existing_obtained_at = state.get("agent_key_obtained_at") + if obtained_at: + effective_obtained_at = obtained_at + elif ( + state.get("agent_key") == access_token + and isinstance(existing_obtained_at, str) + and existing_obtained_at.strip() + ): + effective_obtained_at = existing_obtained_at + else: + effective_obtained_at = now.isoformat() + expires_at = _nous_jwt_expires_at(access_token, state.get("expires_at")) + expires_epoch = _parse_iso_timestamp(expires_at) + expires_in = ( + max(0, int(expires_epoch - time.time())) + if expires_epoch is not None + else _coerce_ttl_seconds(state.get("expires_in")) + ) + if expires_at: + state["expires_at"] = expires_at + state["expires_in"] = expires_in + state["agent_key"] = access_token + state["agent_key_id"] = None + state["agent_key_expires_at"] = expires_at + state["agent_key_expires_in"] = expires_in + state["agent_key_reused"] = False + state["agent_key_obtained_at"] = effective_obtained_at + + +def _select_nous_invoke_jwt( + state: Dict[str, Any], + *, + access_token: Any = None, + sequence_id: Optional[str] = None, +) -> None: + if isinstance(access_token, str) and access_token.strip(): + state["access_token"] = access_token + _set_nous_agent_key_from_invoke_jwt(state) + _log_nous_invoke_jwt_selected( + access_token=state.get("access_token"), + sequence_id=sequence_id, + ) + + +_NOUS_EFFECTIVE_STATE_IGNORED_KEYS = frozenset({ + # These are derived from expires_at/JWT exp and naturally tick down between + # reads. Persisting only these changes makes auth.json noisy and defeats + # the mtime-keyed auth-status cache. + "expires_in", + "agent_key_expires_in", +}) + + +def _nous_effective_provider_state(state: Dict[str, Any]) -> Dict[str, Any]: + return { + key: value + for key, value in state.items() + if key not in _NOUS_EFFECTIVE_STATE_IGNORED_KEYS + } + + def _codex_access_token_is_expiring(access_token: Any, skew_seconds: int) -> bool: claims = _decode_jwt_claims(access_token) exp = claims.get("exp") @@ -2610,7 +2881,7 @@ def _print_loopback_ssh_hint(redirect_uri: str, *, docs_url: str | None = None) return host = parsed.hostname or "" port = parsed.port - if host not in ("127.0.0.1", "::1", "localhost") or not port: + if host not in {"127.0.0.1", "::1", "localhost"} or not port: return print() print("Remote session detected. Your browser will redirect to") @@ -3333,6 +3604,85 @@ def _request_device_code( return data +def _is_nous_invoke_scope_refusal(exc: Exception) -> bool: + if not isinstance(exc, httpx.HTTPStatusError): + return False + response = exc.response + if response.status_code not in {400, 401, 403}: + return False + try: + payload = response.json() + except Exception: + payload = {} + text = " ".join( + str(value) + for value in ( + payload.get("error") if isinstance(payload, dict) else None, + payload.get("error_description") if isinstance(payload, dict) else None, + response.text, + ) + if value + ).lower() + if not text: + return False + return ( + "invalid_scope" in text + or "unsupported_scope" in text + or "scope" in text and NOUS_INFERENCE_INVOKE_SCOPE in text + ) + + +def _nous_device_scope_with_env_override( + requested_scope: Optional[str], + *, + default_scope: str = DEFAULT_NOUS_SCOPE, +) -> Tuple[str, bool]: + explicit_scope = requested_scope is not None + scope = requested_scope or default_scope + if _nous_legacy_session_keys_forced(): + scope = NOUS_LEGACY_AGENT_KEY_SCOPE + return scope, explicit_scope + + +def _request_nous_device_code_with_scope_fallback( + *, + client: httpx.Client, + portal_base_url: str, + client_id: str, + scope: str, + allow_legacy_fallback: bool, +) -> Tuple[Dict[str, Any], str]: + try: + return ( + _request_device_code( + client=client, + portal_base_url=portal_base_url, + client_id=client_id, + scope=scope, + ), + scope, + ) + except Exception as exc: + if ( + allow_legacy_fallback + and _nous_scope_has_invoke(scope) + and _is_nous_invoke_scope_refusal(exc) + ): + logger.info("Nous inference auth: NAS refused invoke scope, retrying legacy scope") + _oauth_trace("nous_device_code_invoke_scope_refused") + retry_scope = NOUS_LEGACY_AGENT_KEY_SCOPE + return ( + _request_device_code( + client=client, + portal_base_url=portal_base_url, + client_id=client_id, + scope=retry_scope, + ), + retry_scope, + ) + raise + + def _poll_for_token( client: httpx.Client, portal_base_url: str, @@ -3524,8 +3874,9 @@ def _write_shared_nous_state(state: Dict[str, Any]) -> None: is a convenience layer; the per-profile auth.json remains the source of truth. - We deliberately omit the short-lived ``agent_key`` (24h TTL, profile- - specific) — only the long-lived OAuth tokens are cross-profile useful. + We deliberately omit the runtime ``agent_key`` compatibility field + (either an invoke JWT or legacy opaque session key) — only OAuth tokens + are cross-profile useful. """ refresh_token = state.get("refresh_token") access_token = state.get("access_token") @@ -3616,6 +3967,96 @@ def _read_shared_nous_state() -> Optional[Dict[str, Any]]: return payload +def _clear_shared_nous_state(reason: str) -> None: + """Remove the shared Nous OAuth store after a terminal token failure.""" + try: + with _nous_shared_store_lock(): + path = _nous_shared_store_path() + try: + path.unlink() + except FileNotFoundError: + pass + _oauth_trace("nous_shared_store_cleared", reason=reason) + except Exception as exc: + logger.debug("Failed to clear shared Nous auth store: %s", exc) + + +def _is_terminal_nous_refresh_error(exc: Exception) -> bool: + """True when retrying the same Nous refresh token cannot succeed.""" + return ( + isinstance(exc, AuthError) + and exc.provider == "nous" + and exc.code in {"invalid_grant", "invalid_token", "refresh_token_reused"} + and bool(exc.relogin_required) + ) + + +def _quarantine_nous_oauth_state( + state: Dict[str, Any], + error: AuthError, + *, + reason: str, +) -> None: + """Keep routing metadata but remove dead OAuth material so it is not replayed.""" + for key in ( + "access_token", + "refresh_token", + "expires_at", + "expires_in", + "obtained_at", + "agent_key", + "agent_key_id", + "agent_key_expires_at", + "agent_key_expires_in", + "agent_key_reused", + "agent_key_obtained_at", + ): + state.pop(key, None) + state["last_auth_error"] = { + "provider": "nous", + "code": error.code, + "message": str(error), + "reason": reason, + "relogin_required": True, + "at": datetime.now(timezone.utc).isoformat(), + } + _clear_shared_nous_state(reason) + invalidate_nous_auth_status_cache() + + +def _quarantine_nous_pool_entries( + auth_store: Dict[str, Any], + error: AuthError, + *, + reason: str, +) -> bool: + """Remove singleton-seeded Nous pool entries that contain dead OAuth state.""" + pool = auth_store.get("credential_pool") + if not isinstance(pool, dict): + return False + entries = pool.get("nous") + if not isinstance(entries, list): + return False + + retained = [] + removed = False + singleton_sources = {NOUS_DEVICE_CODE_SOURCE, f"manual:{NOUS_DEVICE_CODE_SOURCE}"} + for entry in entries: + if isinstance(entry, dict) and entry.get("source") in singleton_sources: + removed = True + continue + retained.append(entry) + + if removed: + pool["nous"] = retained + _oauth_trace( + "nous_pool_device_code_quarantined", + reason=reason, + error_code=error.code, + ) + return removed + + def _try_import_shared_nous_state( *, timeout_seconds: float = 15.0, @@ -3641,7 +4082,7 @@ def _try_import_shared_nous_state( # Build a full state dict so refresh_nous_oauth_from_state has every # field it needs. force_refresh=True gets us a fresh access_token - # for this profile; force_mint=True gets us a fresh agent_key. + # for this profile; fresh auth mode avoids stale cached legacy keys. state: Dict[str, Any] = { "access_token": shared.get("access_token"), "refresh_token": shared.get("refresh_token"), @@ -3657,12 +4098,16 @@ def _try_import_shared_nous_state( "tls": {"insecure": False, "ca_bundle": None}, } + def _persist_shared_refresh(updated_state: Dict[str, Any], _reason: str) -> None: + _write_shared_nous_state(updated_state) + refreshed = refresh_nous_oauth_from_state( state, min_key_ttl_seconds=min_key_ttl_seconds, timeout_seconds=timeout_seconds, force_refresh=True, - force_mint=True, + inference_auth_mode=NOUS_INFERENCE_AUTH_MODE_FRESH, + on_state_update=_persist_shared_refresh, ) _write_shared_nous_state(refreshed) except AuthError as exc: @@ -3671,6 +4116,8 @@ def _try_import_shared_nous_state( error_type=type(exc).__name__, error_code=getattr(exc, "code", None), ) + if _is_terminal_nous_refresh_error(exc): + _clear_shared_nous_state("shared_import_terminal_refresh_failure") logger.debug("Shared Nous import failed: %s", exc) return None except Exception as exc: @@ -3715,7 +4162,7 @@ def _refresh_access_token( code = str(error_payload.get("error", "invalid_grant")) description = str(error_payload.get("error_description") or "Refresh token exchange failed") - relogin = code in {"invalid_grant", "invalid_token"} + relogin = code in {"invalid_grant", "invalid_token", "refresh_token_reused"} # Detect the OAuth 2.1 "refresh token reuse" signal from the Nous portal # server and surface an actionable message. This fires when an external @@ -3725,7 +4172,7 @@ def _refresh_access_token( # retires the original RT, Hermes's next refresh uses it, and the whole # session chain gets revoked as a token-theft signal (#15099). lowered = description.lower() - if "reuse" in lowered or "reuse detected" in lowered: + if code == "refresh_token_reused" or "reuse" in lowered or "reuse detected" in lowered: description = ( "Nous Portal detected refresh-token reuse and revoked this session.\n" "This usually means an external process (monitoring script, " @@ -3737,6 +4184,7 @@ def _refresh_access_token( "instead.\n" "Re-authenticate with: hermes auth add nous" ) + relogin = True raise AuthError(description, provider="nous", code=code, relogin_required=relogin) @@ -3835,6 +4283,14 @@ def _agent_key_is_usable(state: Dict[str, Any], min_ttl_seconds: int) -> bool: key = state.get("agent_key") if not isinstance(key, str) or not key.strip(): return False + if _decode_jwt_claims(key): + if _nous_legacy_session_keys_forced(): + return False + return _nous_invoke_jwt_is_usable( + key, + scope=state.get("scope"), + expires_at=state.get("agent_key_expires_at"), + ) return not _is_expiring(state.get("agent_key_expires_at"), min_ttl_seconds) @@ -3896,12 +4352,28 @@ def resolve_nous_access_token( headers={"Accept": "application/json"}, verify=verify, ) as client: - refreshed = _refresh_access_token( - client=client, - portal_base_url=portal_base_url, - client_id=client_id, - refresh_token=refresh_token, - ) + try: + refreshed = _refresh_access_token( + client=client, + portal_base_url=portal_base_url, + client_id=client_id, + refresh_token=refresh_token, + ) + except AuthError as exc: + if _is_terminal_nous_refresh_error(exc): + _quarantine_nous_oauth_state( + state, + exc, + reason="managed_access_token_refresh_failure", + ) + _quarantine_nous_pool_entries( + auth_store, + exc, + reason="managed_access_token_refresh_failure", + ) + _save_provider_state(auth_store, "nous", state) + _save_auth_store(auth_store) + raise now = datetime.now(timezone.utc) access_ttl = _coerce_ttl_seconds(refreshed.get("expires_in")) @@ -3945,9 +4417,16 @@ def refresh_nous_oauth_pure( insecure: Optional[bool] = None, ca_bundle: Optional[str] = None, force_refresh: bool = False, - force_mint: bool = False, + inference_auth_mode: str = NOUS_INFERENCE_AUTH_MODE_AUTO, + on_state_update: Optional[Callable[[Dict[str, Any], str], None]] = None, ) -> Dict[str, Any]: - """Refresh Nous OAuth state without mutating auth.json.""" + """Refresh Nous OAuth state without mutating auth.json directly. + + ``on_state_update`` is called after a successful access-token refresh and + before any subsequent agent-key mint. Callers that own persistent state can + use it to save the newly rotated refresh token before later work can fail. + """ + inference_auth_mode = _normalize_nous_inference_auth_mode(inference_auth_mode) state: Dict[str, Any] = { "access_token": access_token, "refresh_token": refresh_token, @@ -3969,7 +4448,23 @@ def refresh_nous_oauth_pure( timeout = httpx.Timeout(timeout_seconds if timeout_seconds else 15.0) with httpx.Client(timeout=timeout, headers={"Accept": "application/json"}, verify=verify) as client: - if force_refresh or _is_expiring(state.get("expires_at"), ACCESS_TOKEN_REFRESH_SKEW_SECONDS): + min_agent_key_ttl = max(60, int(min_key_ttl_seconds)) + legacy_session_keys = _nous_legacy_session_keys_forced() + current_invoke_jwt_usable = ( + not legacy_session_keys + and _nous_invoke_jwt_is_usable( + state.get("access_token"), + scope=state.get("scope"), + expires_at=state.get("expires_at"), + ) + ) + if ( + force_refresh + or ( + _is_expiring(state.get("expires_at"), ACCESS_TOKEN_REFRESH_SKEW_SECONDS) + and not current_invoke_jwt_usable + ) + ): refreshed = _refresh_access_token( client=client, portal_base_url=state["portal_base_url"], @@ -3990,8 +4485,21 @@ def refresh_nous_oauth_pure( state["expires_at"] = datetime.fromtimestamp( now.timestamp() + access_ttl, tz=timezone.utc ).isoformat() + if on_state_update is not None: + on_state_update(dict(state), "post_refresh_access_token") - if force_mint or not _agent_key_is_usable(state, max(60, int(min_key_ttl_seconds))): + selected_auth_path, fallback_reason = _choose_nous_inference_auth_path( + state, + min_key_ttl_seconds=min_agent_key_ttl, + inference_auth_mode=inference_auth_mode, + ) + if selected_auth_path == NOUS_AUTH_PATH_INVOKE_JWT: + _select_nous_invoke_jwt(state) + elif selected_auth_path == NOUS_AUTH_PATH_LEGACY_SESSION_KEY_MINT: + _log_nous_legacy_session_key_selected( + fallback_reason or "legacy_session_key_required", + access_token=state.get("access_token"), + ) mint_payload = _mint_agent_key( client=client, portal_base_url=state["portal_base_url"], @@ -4018,7 +4526,8 @@ def refresh_nous_oauth_from_state( min_key_ttl_seconds: int = DEFAULT_AGENT_KEY_MIN_TTL_SECONDS, timeout_seconds: float = 15.0, force_refresh: bool = False, - force_mint: bool = False, + inference_auth_mode: str = NOUS_INFERENCE_AUTH_MODE_AUTO, + on_state_update: Optional[Callable[[Dict[str, Any], str], None]] = None, ) -> Dict[str, Any]: """Refresh Nous OAuth from a state dict. Thin wrapper around refresh_nous_oauth_pure.""" tls = state.get("tls") or {} @@ -4039,13 +4548,11 @@ def refresh_nous_oauth_from_state( insecure=tls.get("insecure"), ca_bundle=tls.get("ca_bundle"), force_refresh=force_refresh, - force_mint=force_mint, + inference_auth_mode=inference_auth_mode, + on_state_update=on_state_update, ) -NOUS_DEVICE_CODE_SOURCE = "device_code" - - def persist_nous_credentials( creds: Dict[str, Any], *, @@ -4105,13 +4612,23 @@ def persist_nous_credentials( ) +def _sync_nous_pool_from_auth_store() -> None: + """Best-effort pool reseed after providers.nous changes; never fail login.""" + try: + from agent.credential_pool import load_pool + + load_pool("nous") + except Exception as exc: + logger.debug("Failed to sync Nous credential pool from auth store: %s", exc) + + def resolve_nous_runtime_credentials( *, min_key_ttl_seconds: int = DEFAULT_AGENT_KEY_MIN_TTL_SECONDS, timeout_seconds: float = 15.0, insecure: Optional[bool] = None, ca_bundle: Optional[str] = None, - force_mint: bool = False, + inference_auth_mode: str = NOUS_INFERENCE_AUTH_MODE_AUTO, ) -> Dict[str, Any]: """ Resolve Nous inference credentials for runtime use. @@ -4121,8 +4638,9 @@ def resolve_nous_runtime_credentials( Concurrent processes coordinate through the auth store file lock. Returns dict with: provider, base_url, api_key, key_id, expires_at, - expires_in, source ("cache" or "portal"). + expires_in, source ("invoke_jwt", "cache", or "portal"), and auth_path. """ + inference_auth_mode = _normalize_nous_inference_auth_mode(inference_auth_mode) min_key_ttl_seconds = max(60, int(min_key_ttl_seconds)) sequence_id = uuid.uuid4().hex[:12] @@ -4134,6 +4652,9 @@ def resolve_nous_runtime_credentials( raise AuthError("Hermes is not logged into Nous Portal.", provider="nous", relogin_required=True) + persisted_state = dict(state) + state_persisted = False + portal_base_url = ( _optional_base_url(state.get("portal_base_url")) or os.getenv("HERMES_PORTAL_BASE_URL") @@ -4148,6 +4669,19 @@ def resolve_nous_runtime_credentials( client_id = str(state.get("client_id") or DEFAULT_NOUS_CLIENT_ID) def _persist_state(reason: str) -> None: + nonlocal persisted_state, state_persisted + # Skip writes where only derived TTL countdowns changed; this keeps + # the mtime-keyed Nous auth-status cache warm during read paths. + if ( + _nous_effective_provider_state(state) + == _nous_effective_provider_state(persisted_state) + ): + _oauth_trace( + "nous_state_persist_skipped", + sequence_id=sequence_id, + reason=reason, + ) + return try: _save_provider_state(auth_store, "nous", state) _save_auth_store(auth_store) @@ -4166,6 +4700,8 @@ def resolve_nous_runtime_credentials( refresh_token_fp=_token_fingerprint(state.get("refresh_token")), access_token_fp=_token_fingerprint(state.get("access_token")), ) + persisted_state = dict(state) + state_persisted = True # Mirror post-refresh state to the shared store so sibling # profiles don't hold stale refresh_tokens after rotation. # Best-effort — any failure is logged and swallowed inside @@ -4177,7 +4713,7 @@ def resolve_nous_runtime_credentials( _oauth_trace( "nous_runtime_credentials_start", sequence_id=sequence_id, - force_mint=bool(force_mint), + inference_auth_mode=inference_auth_mode, min_key_ttl_seconds=min_key_ttl_seconds, refresh_token_fp=_token_fingerprint(state.get("refresh_token")), ) @@ -4190,15 +4726,35 @@ def resolve_nous_runtime_credentials( raise AuthError("No access token found for Nous Portal login.", provider="nous", relogin_required=True) - # Step 1: refresh access token if expiring - if _is_expiring(state.get("expires_at"), ACCESS_TOKEN_REFRESH_SKEW_SECONDS): + # Step 1: refresh access token if expiring. If the access token + # is already a valid invoke JWT, trust its own exp claim even when + # older auth.json metadata has a stale/missing expires_at. + current_invoke_jwt_usable = ( + not _nous_legacy_session_keys_forced() + and _nous_invoke_jwt_is_usable( + access_token, + scope=state.get("scope"), + expires_at=state.get("expires_at"), + ) + ) + if ( + _is_expiring(state.get("expires_at"), ACCESS_TOKEN_REFRESH_SKEW_SECONDS) + and not current_invoke_jwt_usable + ): with _nous_shared_store_lock(timeout_seconds=max(timeout_seconds + 5.0, AUTH_LOCK_TIMEOUT_SECONDS)): if _merge_shared_nous_oauth_state(state): access_token = state.get("access_token") refresh_token = state.get("refresh_token") _persist_state("post_shared_merge_access_expiring") - if _is_expiring(state.get("expires_at"), ACCESS_TOKEN_REFRESH_SKEW_SECONDS): + if ( + _is_expiring(state.get("expires_at"), ACCESS_TOKEN_REFRESH_SKEW_SECONDS) + and not _nous_invoke_jwt_is_usable( + access_token, + scope=state.get("scope"), + expires_at=state.get("expires_at"), + ) + ): if not isinstance(refresh_token, str) or not refresh_token: raise AuthError("Session expired and no refresh token is available.", provider="nous", relogin_required=True) @@ -4209,10 +4765,25 @@ def resolve_nous_runtime_credentials( reason="access_expiring", refresh_token_fp=_token_fingerprint(refresh_token), ) - refreshed = _refresh_access_token( - client=client, portal_base_url=portal_base_url, - client_id=client_id, refresh_token=refresh_token, - ) + try: + refreshed = _refresh_access_token( + client=client, portal_base_url=portal_base_url, + client_id=client_id, refresh_token=refresh_token, + ) + except AuthError as exc: + if _is_terminal_nous_refresh_error(exc): + _quarantine_nous_oauth_state( + state, + exc, + reason="runtime_access_refresh_failure", + ) + _quarantine_nous_pool_entries( + auth_store, + exc, + reason="runtime_access_refresh_failure", + ) + _persist_state("terminal_runtime_access_refresh_failure") + raise now = datetime.now(timezone.utc) access_ttl = _coerce_ttl_seconds(refreshed.get("expires_in")) previous_refresh_token = refresh_token @@ -4240,14 +4811,34 @@ def resolve_nous_runtime_credentials( # Persist immediately so downstream mint failures cannot drop rotated refresh tokens. _persist_state("post_refresh_access_expiring") - # Step 2: mint agent key if missing/expiring + # Step 2: resolve the compatibility ``agent_key`` field. Preferred + # path stores the NAS invoke JWT there; legacy path mints/reuses + # the opaque session key. used_cached_key = False mint_payload: Optional[Dict[str, Any]] = None + selected_auth_path, fallback_reason = _choose_nous_inference_auth_path( + state, + access_token=access_token, + min_key_ttl_seconds=min_key_ttl_seconds, + inference_auth_mode=inference_auth_mode, + ) - if not force_mint and _agent_key_is_usable(state, min_key_ttl_seconds): + if selected_auth_path == NOUS_AUTH_PATH_INVOKE_JWT: + _select_nous_invoke_jwt( + state, + access_token=access_token, + sequence_id=sequence_id, + ) + elif selected_auth_path == NOUS_AUTH_PATH_LEGACY_SESSION_KEY_CACHE: used_cached_key = True + logger.info("Nous inference auth: using cached agent_key") _oauth_trace("agent_key_reuse", sequence_id=sequence_id) else: + _log_nous_legacy_session_key_selected( + fallback_reason or "legacy_session_key_required", + access_token=access_token, + sequence_id=sequence_id, + ) try: _oauth_trace( "mint_start", @@ -4283,10 +4874,25 @@ def resolve_nous_runtime_credentials( reason="mint_retry_after_invalid_token", refresh_token_fp=_token_fingerprint(latest_refresh_token), ) - refreshed = _refresh_access_token( - client=client, portal_base_url=portal_base_url, - client_id=client_id, refresh_token=latest_refresh_token, - ) + try: + refreshed = _refresh_access_token( + client=client, portal_base_url=portal_base_url, + client_id=client_id, refresh_token=latest_refresh_token, + ) + except AuthError as exc: + if _is_terminal_nous_refresh_error(exc): + _quarantine_nous_oauth_state( + state, + exc, + reason="runtime_mint_retry_refresh_failure", + ) + _quarantine_nous_pool_entries( + auth_store, + exc, + reason="runtime_mint_retry_refresh_failure", + ) + _persist_state("terminal_runtime_mint_retry_refresh_failure") + raise now = datetime.now(timezone.utc) access_ttl = _coerce_ttl_seconds(refreshed.get("expires_in")) state["access_token"] = refreshed["access_token"] @@ -4313,10 +4919,30 @@ def resolve_nous_runtime_credentials( # Persist retry refresh immediately for crash safety and cross-process visibility. _persist_state("post_refresh_mint_retry") - mint_payload = _mint_agent_key( - client=client, portal_base_url=portal_base_url, - access_token=access_token, min_ttl_seconds=min_key_ttl_seconds, + retry_inference_auth_mode = ( + NOUS_INFERENCE_AUTH_MODE_LEGACY + if inference_auth_mode == NOUS_INFERENCE_AUTH_MODE_LEGACY + else NOUS_INFERENCE_AUTH_MODE_FRESH ) + retry_auth_path, _ = _choose_nous_inference_auth_path( + state, + access_token=access_token, + min_key_ttl_seconds=min_key_ttl_seconds, + inference_auth_mode=retry_inference_auth_mode, + ) + if retry_auth_path == NOUS_AUTH_PATH_INVOKE_JWT: + mint_payload = None + selected_auth_path = NOUS_AUTH_PATH_INVOKE_JWT + _select_nous_invoke_jwt( + state, + access_token=access_token, + sequence_id=sequence_id, + ) + else: + mint_payload = _mint_agent_key( + client=client, portal_base_url=portal_base_url, + access_token=access_token, min_ttl_seconds=min_key_ttl_seconds, + ) else: raise @@ -4348,6 +4974,9 @@ def resolve_nous_runtime_credentials( _persist_state("resolve_nous_runtime_credentials_final") + if state_persisted: + _sync_nous_pool_from_auth_store() + api_key = state.get("agent_key") if not isinstance(api_key, str) or not api_key: raise AuthError("Failed to resolve a Nous inference API key", @@ -4368,7 +4997,12 @@ def resolve_nous_runtime_credentials( "key_id": state.get("agent_key_id"), "expires_at": expires_at, "expires_in": expires_in, - "source": "cache" if used_cached_key else "portal", + "source": ( + NOUS_AUTH_PATH_INVOKE_JWT + if selected_auth_path == NOUS_AUTH_PATH_INVOKE_JWT + else ("cache" if used_cached_key else "portal") + ), + "auth_path": selected_auth_path, } @@ -5246,7 +5880,7 @@ def _login_xai_oauth( reuse = input("Use existing credentials? [Y/n]: ").strip().lower() except (EOFError, KeyboardInterrupt): reuse = "y" - if reuse in ("", "y", "yes"): + if reuse in {"", "y", "yes"}: config_path = _update_config_for_provider( "xai-oauth", existing.get("base_url", DEFAULT_XAI_OAUTH_BASE_URL), @@ -5312,6 +5946,107 @@ def _xai_oauth_build_authorize_url( return f"{authorization_endpoint}?{urlencode(authorize_params)}" +def _xai_oauth_exchange_code_for_tokens( + *, + token_endpoint: str, + code: str, + redirect_uri: str, + code_verifier: str, + code_challenge: str, + timeout_seconds: float = 20.0, +) -> Dict[str, Any]: + """POST the authorization code to xAI's token endpoint and return + the parsed JSON payload. + + Sends ``code_verifier`` as required by RFC 7636 §4.5. Also echoes + ``code_challenge`` + ``code_challenge_method`` in the request body + as a defense-in-depth measure for OAuth servers (xAI's among them, + per #26990) that re-validate the challenge at the token step + instead of relying solely on server-side session state captured + during the authorize step. Echoing the challenge is harmless for + strict RFC-compliant servers — RFC 7636 doesn't forbid additional + parameters at the token endpoint — and decisively fixes the + ``code_challenge is required`` failure mode users hit on the + loopback flow. + + Raises :class:`AuthError` on any non-2xx response or transport + failure; the error message embeds the HTTP status code and the + full response body so users can disambiguate cause at a glance. + """ + # Paranoia: if upstream call sites ever drop ``code_verifier`` we + # want to surface a precise, local error rather than send a + # missing-PKCE request to xAI and receive their generic "code + # challenge required" message back. + if not code_verifier: + raise AuthError( + "xAI token exchange refused locally: PKCE code_verifier is empty. " + "This is a bug in Hermes — please report at " + "https://github.com/NousResearch/hermes-agent/issues/26990.", + provider="xai-oauth", + code="xai_pkce_verifier_missing", + ) + + data = { + "grant_type": "authorization_code", + "code": code, + "redirect_uri": redirect_uri, + "client_id": XAI_OAUTH_CLIENT_ID, + "code_verifier": code_verifier, + } + # Defense-in-depth: include the original ``code_challenge`` and + # ``code_challenge_method``. Some OAuth servers (including xAI's + # auth.x.ai implementation, per the symptom reported in #26990) + # validate these at the token endpoint instead of relying purely on + # state captured during the authorize step — without them, xAI + # rejects the exchange with ``code_challenge is required`` even + # though we sent a valid ``code_verifier``. + if code_challenge: + data["code_challenge"] = code_challenge + data["code_challenge_method"] = "S256" + + try: + response = httpx.post( + token_endpoint, + headers={ + "Content-Type": "application/x-www-form-urlencoded", + "Accept": "application/json", + }, + data=data, + timeout=max(20.0, timeout_seconds), + ) + except Exception as exc: + raise AuthError( + f"xAI token exchange failed: {exc}", + provider="xai-oauth", + code="xai_token_exchange_failed", + ) from exc + + if response.status_code != 200: + body = response.text.strip() + raise AuthError( + f"xAI token exchange failed (HTTP {response.status_code})." + + (f" Response: {body}" if body else ""), + provider="xai-oauth", + code="xai_token_exchange_failed", + ) + + try: + payload = response.json() + except Exception as exc: + raise AuthError( + f"xAI token exchange returned invalid JSON: {exc}", + provider="xai-oauth", + code="xai_token_exchange_invalid", + ) from exc + if not isinstance(payload, dict): + raise AuthError( + "xAI token exchange response was not a JSON object.", + provider="xai-oauth", + code="xai_token_exchange_invalid", + ) + return payload + + def _xai_oauth_loopback_login( *, timeout_seconds: float = 20.0, @@ -5392,47 +6127,14 @@ def _xai_oauth_loopback_login( code="xai_code_missing", ) - try: - response = httpx.post( - token_endpoint, - headers={"Content-Type": "application/x-www-form-urlencoded", "Accept": "application/json"}, - data={ - "grant_type": "authorization_code", - "code": code, - "redirect_uri": redirect_uri, - "client_id": XAI_OAUTH_CLIENT_ID, - "code_verifier": code_verifier, - }, - timeout=max(20.0, timeout_seconds), - ) - except Exception as exc: - raise AuthError( - f"xAI token exchange failed: {exc}", - provider="xai-oauth", - code="xai_token_exchange_failed", - ) from exc - if response.status_code != 200: - detail = response.text.strip() - raise AuthError( - "xAI token exchange failed." - + (f" Response: {detail}" if detail else ""), - provider="xai-oauth", - code="xai_token_exchange_failed", - ) - try: - payload = response.json() - except Exception as exc: - raise AuthError( - f"xAI token exchange returned invalid JSON: {exc}", - provider="xai-oauth", - code="xai_token_exchange_invalid", - ) from exc - if not isinstance(payload, dict): - raise AuthError( - "xAI token exchange response was not a JSON object.", - provider="xai-oauth", - code="xai_token_exchange_invalid", - ) + payload = _xai_oauth_exchange_code_for_tokens( + token_endpoint=token_endpoint, + code=code, + redirect_uri=redirect_uri, + code_verifier=code_verifier, + code_challenge=code_challenge, + timeout_seconds=timeout_seconds, + ) access_token = str(payload.get("access_token", "") or "").strip() refresh_token = str(payload.get("refresh_token", "") or "").strip() if not access_token: @@ -5979,7 +6681,10 @@ def _nous_device_code_login( or pconfig.inference_base_url ).rstrip("/") client_id = client_id or pconfig.client_id - scope = scope or pconfig.scope + scope, explicit_scope = _nous_device_scope_with_env_override( + scope, + default_scope=pconfig.scope, + ) timeout = httpx.Timeout(timeout_seconds) verify: bool | str = False if insecure else (ca_bundle if ca_bundle else True) @@ -5994,11 +6699,12 @@ def _nous_device_code_login( print(f"TLS verification: custom CA bundle ({ca_bundle})") with httpx.Client(timeout=timeout, headers={"Accept": "application/json"}, verify=verify) as client: - device_data = _request_device_code( + device_data, scope = _request_nous_device_code_with_scope_fallback( client=client, portal_base_url=portal_base_url, client_id=client_id, scope=scope, + allow_legacy_fallback=not explicit_scope, ) verification_url = str(device_data["verification_uri_complete"]) @@ -6068,7 +6774,7 @@ def _nous_device_code_login( min_key_ttl_seconds=min_key_ttl_seconds, timeout_seconds=timeout_seconds, force_refresh=False, - force_mint=True, + inference_auth_mode=NOUS_INFERENCE_AUTH_MODE_FRESH, ) except AuthError as exc: if exc.code == "subscription_required": @@ -6129,7 +6835,7 @@ def _login_nous(args, pconfig: ProviderConfig) -> None: portal_base_url=getattr(args, "portal_url", None), inference_base_url=getattr(args, "inference_url", None), client_id=getattr(args, "client_id", None) or pconfig.client_id, - scope=getattr(args, "scope", None) or pconfig.scope, + scope=getattr(args, "scope", None), open_browser=not getattr(args, "no_browser", False), timeout_seconds=timeout_seconds, insecure=insecure, @@ -6156,6 +6862,7 @@ def _login_nous(args, pconfig: ProviderConfig) -> None: # these credentials. Best-effort: any I/O failure is logged and # swallowed inside the helper. _write_shared_nous_state(auth_state) + _sync_nous_pool_from_auth_store() print() print("Login successful!") diff --git a/hermes_cli/codex_runtime_switch.py b/hermes_cli/codex_runtime_switch.py index b3adda12b54..98b40b1e8f2 100644 --- a/hermes_cli/codex_runtime_switch.py +++ b/hermes_cli/codex_runtime_switch.py @@ -48,9 +48,9 @@ def parse_args(arg_string: str) -> tuple[Optional[str], list[str]]: if not raw: return None, [] # Accept human-friendly synonyms - if raw in ("on", "codex", "enable"): + if raw in {"on", "codex", "enable"}: return "codex_app_server", [] - if raw in ("off", "default", "disable", "hermes"): + if raw in {"off", "default", "disable", "hermes"}: return "auto", [] if raw in VALID_RUNTIMES: return raw, [] diff --git a/hermes_cli/commands.py b/hermes_cli/commands.py index 83d86c4a3a9..a62077af100 100644 --- a/hermes_cli/commands.py +++ b/hermes_cli/commands.py @@ -123,7 +123,8 @@ COMMAND_REGISTRY: list[CommandDef] = [ CommandDef("model", "Switch model for this session", "Configuration", aliases=("provider",), args_hint="[model] [--provider name] [--global]"), CommandDef("codex-runtime", "Toggle codex app-server runtime for OpenAI/Codex models", - "Configuration", args_hint="[auto|codex_app_server]"), + "Configuration", aliases=("codex_runtime",), + args_hint="[auto|codex_app_server]"), CommandDef("gquota", "Show Google Gemini Code Assist quota usage", "Info", cli_only=True), diff --git a/hermes_cli/config.py b/hermes_cli/config.py index 22656b5c81e..b7649126b58 100644 --- a/hermes_cli/config.py +++ b/hermes_cli/config.py @@ -926,6 +926,31 @@ DEFAULT_CONFIG = { "timeout": 120, "extra_body": {}, }, + # Kanban decomposer — decomposes a triage task into a graph of + # child tasks routed to specialist profiles by description. + # Invoked by ``hermes kanban decompose`` and the kanban + # auto-decompose dispatcher tick. Returns a JSON task graph; + # uses more tokens than the specifier so allow more headroom. + "kanban_decomposer": { + "provider": "auto", + "model": "", + "base_url": "", + "api_key": "", + "timeout": 180, + "extra_body": {}, + }, + # Profile describer — auto-generates a 1-2 sentence description + # of what a profile is good at. Invoked by + # ``hermes profile describe --auto`` and the dashboard's + # auto-generate button. Short, cheap call. + "profile_describer": { + "provider": "auto", + "model": "", + "base_url": "", + "api_key": "", + "timeout": 60, + "extra_body": {}, + }, # Curator — skill-usage review fork. Timeout is generous because the # review pass can take several minutes on reasoning models (umbrella # building over hundreds of candidate skills). "auto" = use main chat @@ -1473,6 +1498,25 @@ DEFAULT_CONFIG = { # same task/profile (spawn_failed, timed_out, or crashed). Reassignment # resets the streak for the new profile. "failure_limit": 2, + # Profile that decomposes tasks in the Triage column. When unset, + # falls back to the default profile (the one `hermes` launches with + # no -p flag). Set this to a dedicated 'orchestrator' profile if you + # want decomposition to use a different model/skills from your main + # working profile. + "orchestrator_profile": "", + # Where a child task lands if the orchestrator can't match an + # assignee to any installed profile. When unset, falls back to the + # default profile. A task never ends up with assignee=None. + "default_assignee": "", + # When true, the kanban dispatcher auto-runs the decomposer on + # tasks that land in Triage (every dispatcher tick). When false, + # decomposition is manual via `hermes kanban decompose ` or + # the dashboard's Decompose button. + "auto_decompose": True, + # Max triage tasks to decompose per dispatcher tick. Prevents a + # large bulk-load of triage tasks from spending a burst of aux + # LLM calls in one tick. Excess tasks defer to the next tick. + "auto_decompose_per_tick": 3, }, # execute_code settings — controls the tool used for programmatic tool calls. @@ -2913,6 +2957,7 @@ def _normalize_custom_provider_entry( "api_mode", "transport", "model", "default_model", "models", "context_length", "rate_limit_delay", "request_timeout_seconds", "stale_timeout_seconds", + "discover_models", } for camel, snake in _CAMEL_ALIASES.items(): if camel in entry and snake not in entry: @@ -3003,6 +3048,10 @@ def _normalize_custom_provider_entry( if isinstance(rate_limit_delay, (int, float)) and rate_limit_delay >= 0: normalized["rate_limit_delay"] = rate_limit_delay + discover_models = entry.get("discover_models") + if isinstance(discover_models, bool): + normalized["discover_models"] = discover_models + return normalized diff --git a/hermes_cli/dep_ensure.py b/hermes_cli/dep_ensure.py index 3312726c36d..1067b428f7b 100644 --- a/hermes_cli/dep_ensure.py +++ b/hermes_cli/dep_ensure.py @@ -91,7 +91,7 @@ def ensure_dependency(dep: str, interactive: bool = True) -> bool: reply = input(f"{desc} is not installed. Install now? [Y/n] ").strip().lower() except (EOFError, KeyboardInterrupt): return False - if reply not in ("", "y", "yes"): + if reply not in {"", "y", "yes"}: return False result = subprocess.run( diff --git a/hermes_cli/doctor.py b/hermes_cli/doctor.py index 9d3b6e3c01a..87043bc2611 100644 --- a/hermes_cli/doctor.py +++ b/hermes_cli/doctor.py @@ -160,19 +160,25 @@ def _has_healthy_oauth_fallback_for_apikey_provider(provider_label: str) -> bool still show a failed API-key connectivity row, but it should not promote that direct-key problem into the final blocking summary. """ - try: - from hermes_cli.auth import ( - get_gemini_oauth_auth_status, - get_minimax_oauth_auth_status, - ) - except Exception: - return False - normalized = (provider_label or "").strip().lower() if normalized in {"google / gemini", "gemini"}: - return bool((get_gemini_oauth_auth_status() or {}).get("logged_in")) + try: + from hermes_cli.auth import get_gemini_oauth_auth_status + return bool((get_gemini_oauth_auth_status() or {}).get("logged_in")) + except Exception: + return False if normalized == "minimax": - return bool((get_minimax_oauth_auth_status() or {}).get("logged_in")) + try: + from hermes_cli.auth import get_minimax_oauth_auth_status + return bool((get_minimax_oauth_auth_status() or {}).get("logged_in")) + except Exception: + return False + if normalized == "xai": + try: + from hermes_cli.auth import get_xai_oauth_auth_status + return bool((get_xai_oauth_auth_status() or {}).get("logged_in")) + except Exception: + return False return False @@ -645,31 +651,41 @@ def run_doctor(args): # Check credentials for the configured provider. # Limit to API-key providers in PROVIDER_REGISTRY — other provider - # types (OAuth, SDK, openrouter/anthropic/custom/auto) have their - # own env-var checks elsewhere in doctor, and get_auth_status() - # returns a bare {logged_in: False} for anything it doesn't - # explicitly dispatch, which would produce false positives. - if runtime_provider and runtime_provider not in {"auto", "custom", "openrouter"}: + # types (OAuth, SDK, anthropic/custom/auto) have their own env-var + # checks elsewhere in doctor, and get_auth_status() returns a bare + # {logged_in: False} for anything it doesn't explicitly dispatch, + # which would produce false positives. + if runtime_provider and runtime_provider not in ("auto", "custom"): try: - from hermes_cli.auth import PROVIDER_REGISTRY, get_auth_status - pconfig = PROVIDER_REGISTRY.get(runtime_provider) - if pconfig and getattr(pconfig, "auth_type", "") == "api_key": - status = get_auth_status(runtime_provider) or {} + if runtime_provider == "openrouter": + from hermes_cli.config import get_env_value + configured = bool( - status.get("configured") - or status.get("logged_in") - or status.get("api_key") + str(get_env_value("OPENROUTER_API_KEY") or "").strip() + or str(get_env_value("OPENAI_API_KEY") or "").strip() ) - if not configured: - check_fail( - f"model.provider '{runtime_provider}' is set but no API key is configured", - "(check ~/.hermes/.env or run 'hermes setup')", - ) - issues.append( - f"No credentials found for provider '{runtime_provider}'. " - f"Run 'hermes setup' or set the provider's API key in {_DHH}/.env, " - f"or switch providers with 'hermes config set model.provider '" + else: + from hermes_cli.auth import PROVIDER_REGISTRY, get_auth_status + + pconfig = PROVIDER_REGISTRY.get(runtime_provider) + configured = True + if pconfig and getattr(pconfig, "auth_type", "") == "api_key": + status = get_auth_status(runtime_provider) or {} + configured = bool( + status.get("configured") + or status.get("logged_in") + or status.get("api_key") ) + if not configured: + check_fail( + f"model.provider '{runtime_provider}' is set but no API key is configured", + "(check ~/.hermes/.env or run 'hermes setup')", + ) + issues.append( + f"No credentials found for provider '{runtime_provider}'. " + f"Run 'hermes setup' or set the provider's API key in {_DHH}/.env, " + f"or switch providers with 'hermes config set model.provider '" + ) except Exception: pass @@ -817,6 +833,20 @@ def run_doctor(args): except Exception as e: check_warn("Auth provider status", f"(could not check: {e})") + # xAI OAuth — separate try/except so an import failure here cannot + # disrupt the already-printed Nous/Codex/Gemini/MiniMax rows above. + try: + from hermes_cli.auth import get_xai_oauth_auth_status + xai_oauth_status = get_xai_oauth_auth_status() or {} + if xai_oauth_status.get("logged_in"): + check_ok("xAI OAuth", "(logged in)") + else: + check_warn("xAI OAuth", "(not logged in)") + if xai_oauth_status.get("error"): + check_info(xai_oauth_status["error"]) + except Exception: + pass + if _safe_which("codex"): check_ok("codex CLI") else: @@ -1073,10 +1103,20 @@ def run_doctor(args): if terminal_env == "ssh": ssh_host = os.getenv("TERMINAL_SSH_HOST") if ssh_host: + ssh_user = os.getenv("TERMINAL_SSH_USER") + ssh_port = os.getenv("TERMINAL_SSH_PORT") + ssh_key = os.getenv("TERMINAL_SSH_KEY") + target = f"{ssh_user}@{ssh_host}" if ssh_user else ssh_host + cmd = ["ssh", "-o", "ConnectTimeout=5", "-o", "BatchMode=yes"] + if ssh_port: + cmd += ["-p", ssh_port] + if ssh_key: + cmd += ["-i", os.path.expanduser(ssh_key)] + cmd += [target, "echo ok"] # Try to connect try: result = subprocess.run( - ["ssh", "-o", "ConnectTimeout=5", "-o", "BatchMode=yes", ssh_host, "echo ok"], + cmd, capture_output=True, text=True, timeout=15 @@ -1474,6 +1514,15 @@ def run_doctor(args): } if base_url_host_matches(base, "api.kimi.com"): headers["User-Agent"] = "claude-code/0.1.0" + # Google's Generative Language API (generativelanguage.googleapis.com) + # rejects ``Authorization: Bearer `` with 401 + # ``ACCESS_TOKEN_TYPE_UNSUPPORTED`` — that header is reserved for + # OAuth 2 access tokens, not plain API keys. Plain keys use + # ``x-goog-api-key`` (or ``?key=``). Without this, a perfectly valid + # GOOGLE_API_KEY/GEMINI_API_KEY always shows red in ``hermes doctor``. + if url and base_url_host_matches(url, "generativelanguage.googleapis.com"): + headers.pop("Authorization", None) + headers["x-goog-api-key"] = key r = httpx.get(url, headers=headers, timeout=10) if ( pname == "Alibaba/DashScope" diff --git a/hermes_cli/gateway.py b/hermes_cli/gateway.py index c5303e32799..ef57d5ce9fe 100644 --- a/hermes_cli/gateway.py +++ b/hermes_cli/gateway.py @@ -2110,24 +2110,30 @@ def _build_service_path_dirs(project_root: Path | None = None) -> list[str]: if project_root is None: project_root = PROJECT_ROOT + def _is_dir(path: Path) -> bool: + try: + return path.is_dir() + except OSError: + return False + candidates = [] venv_bin = project_root / "venv" / "bin" - if venv_bin.is_dir(): + if _is_dir(venv_bin): candidates.append(str(venv_bin)) elif sys.prefix != sys.base_prefix: candidates.append(str(Path(sys.prefix) / "bin")) node_bin = project_root / "node_modules" / ".bin" - if node_bin.is_dir(): + if _is_dir(node_bin): candidates.append(str(node_bin)) hermes_home = get_hermes_home() hermes_node = hermes_home / "node" / "bin" - if hermes_node.is_dir(): + if _is_dir(hermes_node): candidates.append(str(hermes_node)) hermes_nm = hermes_home / "node_modules" / ".bin" - if hermes_nm.is_dir(): + if _is_dir(hermes_nm): candidates.append(str(hermes_nm)) return candidates diff --git a/hermes_cli/goals.py b/hermes_cli/goals.py index 62ee00547c1..d6a139419a7 100644 --- a/hermes_cli/goals.py +++ b/hermes_cli/goals.py @@ -34,6 +34,7 @@ import logging import re import time from dataclasses import dataclass, field, asdict +from datetime import datetime, timezone from typing import Any, Dict, List, Optional, Tuple logger = logging.getLogger(__name__) @@ -110,6 +111,7 @@ JUDGE_SYSTEM_PROMPT = ( JUDGE_USER_PROMPT_TEMPLATE = ( "Goal:\n{goal}\n\n" "Agent's most recent response:\n{response}\n\n" + "Current time: {current_time}\n\n" "Is the goal satisfied?" ) @@ -120,6 +122,7 @@ JUDGE_USER_PROMPT_WITH_SUBGOALS_TEMPLATE = ( "Additional criteria the user added mid-loop (all must also be " "satisfied for the goal to be DONE):\n{subgoals_block}\n\n" "Agent's most recent response:\n{response}\n\n" + "Current time: {current_time}\n\n" "Decision: For each numbered criterion above, find concrete " "evidence in the agent's response that the criterion is " "satisfied. Do not accept generic phrases like 'all requirements " @@ -415,6 +418,7 @@ def judge_goal( # Build the prompt — pick the with-subgoals variant when applicable. clean_subgoals = [s.strip() for s in (subgoals or []) if s and s.strip()] + current_time = datetime.now(tz=timezone.utc).astimezone().strftime("%Y-%m-%d %H:%M:%S %Z") if clean_subgoals: subgoals_block = "\n".join( f"- {i}. {text}" for i, text in enumerate(clean_subgoals, start=1) @@ -423,11 +427,13 @@ def judge_goal( goal=_truncate(goal, 2000), subgoals_block=_truncate(subgoals_block, 2000), response=_truncate(last_response, _JUDGE_RESPONSE_SNIPPET_CHARS), + current_time=current_time, ) else: prompt = JUDGE_USER_PROMPT_TEMPLATE.format( goal=_truncate(goal, 2000), response=_truncate(last_response, _JUDGE_RESPONSE_SNIPPET_CHARS), + current_time=current_time, ) try: diff --git a/hermes_cli/kanban.py b/hermes_cli/kanban.py index b4024e2e70e..55b1d4125a2 100644 --- a/hermes_cli/kanban.py +++ b/hermes_cli/kanban.py @@ -610,6 +610,43 @@ def build_parser(parent_subparsers: argparse._SubParsersAction) -> argparse.Argu help="Emit one JSON object per task on stdout", ) + # --- decompose --- (triage → fan-out via auxiliary LLM + orchestrator) + p_decompose = sub.add_parser( + "decompose", + help="Decompose a triage-column task into a graph of child tasks " + "routed to specialist profiles by description. Falls back to " + "specify-style single-task promotion when the task doesn't " + "benefit from fan-out. Uses auxiliary.kanban_decomposer.", + ) + p_decompose.add_argument( + "task_id", + nargs="?", + default=None, + help="Task id to decompose (required unless --all is given)", + ) + p_decompose.add_argument( + "--all", + dest="all_triage", + action="store_true", + help="Decompose every task currently in the triage column", + ) + p_decompose.add_argument( + "--tenant", + default=None, + help="When used with --all, restrict the sweep to this tenant", + ) + p_decompose.add_argument( + "--author", + default=None, + help="Author name recorded on the audit comment " + "(default: $HERMES_PROFILE or 'decomposer')", + ) + p_decompose.add_argument( + "--json", + action="store_true", + help="Emit one JSON object per task on stdout", + ) + # --- gc --- p_gc = sub.add_parser( "gc", help="Garbage-collect archived-task workspaces, old events, and old logs", @@ -740,6 +777,7 @@ def kanban_command(args: argparse.Namespace) -> int: "notify-unsubscribe": _cmd_notify_unsubscribe, "context": _cmd_context, "specify": _cmd_specify, + "decompose": _cmd_decompose, "gc": _cmd_gc, } handler = handlers.get(action) @@ -2115,6 +2153,87 @@ def _cmd_specify(args: argparse.Namespace) -> int: return 0 if (ok_count > 0 or not ids) else 1 +def _cmd_decompose(args: argparse.Namespace) -> int: + """Fan a triage task (or all of them) out into a graph of child + tasks via the auxiliary LLM, routed to specialist profiles by + description. Thin wrapper over ``kanban_decompose``.""" + from hermes_cli import kanban_decompose as decomp + + all_flag = bool(getattr(args, "all_triage", False)) + tenant = getattr(args, "tenant", None) + author = getattr(args, "author", None) or _profile_author() + want_json = bool(getattr(args, "json", False)) + + if args.task_id and all_flag: + print( + "kanban: pass either a task id OR --all, not both", + file=sys.stderr, + ) + return 2 + + if all_flag: + ids = decomp.list_triage_ids(tenant=tenant) + if not ids: + msg = ( + "No triage tasks" + + (f" for tenant {tenant!r}" if tenant else "") + + "." + ) + if want_json: + print(json.dumps({"decomposed": 0, "total": 0})) + else: + print(msg) + return 0 + elif args.task_id: + ids = [args.task_id] + else: + print( + "kanban: decompose requires a task id or --all", + file=sys.stderr, + ) + return 2 + + ok_count = 0 + for tid in ids: + outcome = decomp.decompose_task(tid, author=author) + if outcome.ok: + ok_count += 1 + if want_json: + print(json.dumps({ + "task_id": outcome.task_id, + "ok": outcome.ok, + "reason": outcome.reason, + "fanout": outcome.fanout, + "child_ids": outcome.child_ids, + "new_title": outcome.new_title, + })) + elif outcome.ok: + if outcome.fanout and outcome.child_ids: + child_summary = ", ".join(outcome.child_ids) + print( + f"Decomposed {outcome.task_id} → {len(outcome.child_ids)} " + f"children ({child_summary}); root promoted to todo" + ) + else: + title_suffix = ( + f" — retitled: {outcome.new_title!r}" + if outcome.new_title + else "" + ) + print( + f"Specified {outcome.task_id} → todo " + f"(no fanout){title_suffix}" + ) + else: + print( + f"kanban: decompose {outcome.task_id}: {outcome.reason}", + file=sys.stderr, + ) + if not all_flag: + return 0 if ok_count == 1 else 1 + return 0 if (ok_count > 0 or not ids) else 1 + + def _cmd_gc(args: argparse.Namespace) -> int: """Remove scratch workspaces of archived tasks, prune old events, and delete old worker logs.""" diff --git a/hermes_cli/kanban_db.py b/hermes_cli/kanban_db.py index 0db694ff5b1..4bd4827e386 100644 --- a/hermes_cli/kanban_db.py +++ b/hermes_cli/kanban_db.py @@ -93,6 +93,7 @@ from toolsets import get_toolset_names VALID_STATUSES = {"triage", "todo", "ready", "running", "blocked", "done", "archived"} VALID_WORKSPACE_KINDS = {"scratch", "worktree", "dir"} KNOWN_TOOLSET_NAMES = frozenset(name.casefold() for name in get_toolset_names()) +_IS_WINDOWS = sys.platform == "win32" # A running task's claim is valid for 15 minutes; after that the next # dispatcher tick reclaims it. Workers that outlive this window should call @@ -2776,6 +2777,180 @@ def specify_triage_task( return True +def decompose_triage_task( + conn: sqlite3.Connection, + task_id: str, + *, + root_assignee: Optional[str], + children: list[dict], + author: Optional[str] = None, +) -> Optional[list[str]]: + """Fan a triage task out into child tasks and promote the root to ``todo``. + + The root task stays alive and becomes the parent of every child — + when all children reach ``done``, the root promotes to ``ready`` and + its assignee (typically the orchestrator profile) wakes back up to + judge completion or spawn more work. + + ``children`` is a list of dicts, each shaped like:: + + { + "title": "...", + "body": "...", # optional + "assignee": "profile-name", # optional, None -> default fallback + "parents": [0, 2], # indices into this same children list + } + + Returns the list of created child task ids (in input order) on + success. Returns ``None`` when: + - The root task does not exist + - The root task is not in ``triage`` + - A cycle would result (caller built a bad graph) + + Validation of titles/assignees happens inside the same write_txn as + the inserts so a malformed entry aborts the whole decomposition + cleanly (no orphan children). + """ + if not children: + return None + if root_assignee is not None: + root_assignee = _canonical_assignee(root_assignee) + + # Pre-validate the children list shape outside the txn. Cheap checks + # that don't need DB access. Bad input aborts before we touch the DB. + for idx, child in enumerate(children): + if not isinstance(child, dict): + raise ValueError(f"child[{idx}] is not a dict") + title = child.get("title") + if not isinstance(title, str) or not title.strip(): + raise ValueError(f"child[{idx}].title is required") + parents_idx = child.get("parents") or [] + if not isinstance(parents_idx, list): + raise ValueError(f"child[{idx}].parents must be a list") + for p in parents_idx: + if not isinstance(p, int) or p < 0 or p >= len(children): + raise ValueError( + f"child[{idx}].parents[{p}] is not a valid index into children" + ) + if p == idx: + raise ValueError(f"child[{idx}] cannot list itself as a parent") + + # We do the full decomposition in a SINGLE write_txn so it's + # atomic: either every child is created AND the root flips to + # ``todo``, or nothing changes. We deliberately do NOT call any + # kb helper that opens its own write_txn (create_task, link_tasks, + # add_comment) from inside this block — see architecture.md + # write_txn pitfalls. Instead we inline the INSERTs and + # _append_event calls. + now = int(time.time()) + child_ids: list[str] = [] + with write_txn(conn): + root_row = conn.execute( + "SELECT id, status, tenant FROM tasks WHERE id = ?", (task_id,) + ).fetchone() + if root_row is None: + return None + if root_row["status"] != "triage": + return None + tenant = root_row["tenant"] + + # Create children. Status is 'todo' regardless of parents — we + # link them under the root AFTER creation so the dispatcher + # sees a coherent state, and recompute_ready() at the end + # promotes parent-free children to 'ready'. + for idx, child in enumerate(children): + new_id = _new_task_id() + title = child["title"].strip() + body = child.get("body") + assignee = _canonical_assignee(child.get("assignee")) + conn.execute( + "INSERT INTO tasks " + "(id, title, body, assignee, status, workspace_kind, " + " tenant, created_at, created_by) " + "VALUES (?, ?, ?, ?, 'todo', 'scratch', ?, ?, ?)", + ( + new_id, + title, + body if isinstance(body, str) else None, + assignee, + tenant, + now, + (author or "decomposer"), + ), + ) + _append_event( + conn, new_id, "created", + {"by": author or "decomposer", "from_decompose_of": task_id}, + ) + child_ids.append(new_id) + + # Link children to their sibling parents (within the decomposed graph). + for idx, child in enumerate(children): + for p_idx in child.get("parents") or []: + parent_id = child_ids[p_idx] + child_id = child_ids[idx] + conn.execute( + "INSERT OR IGNORE INTO task_links (parent_id, child_id) " + "VALUES (?, ?)", + (parent_id, child_id), + ) + _append_event( + conn, child_id, "linked", + {"parent": parent_id, "child": child_id}, + ) + + # Link the ROOT task as a child of every leaf child — i.e. the + # root waits for the whole graph. Simpler than computing leaves: + # link root under every child. Cycle-free because the root is + # only ever a child here, never a parent of children. + for cid in child_ids: + conn.execute( + "INSERT OR IGNORE INTO task_links (parent_id, child_id) " + "VALUES (?, ?)", + (cid, task_id), + ) + + # Flip the root: triage -> todo, set assignee to the orchestrator. + sets = ["status = 'todo'"] + params: list[Any] = [] + if root_assignee is not None: + sets.append("assignee = ?") + params.append(root_assignee) + params.append(task_id) + conn.execute( + f"UPDATE tasks SET {', '.join(sets)} WHERE id = ?", + tuple(params), + ) + + # Audit comment + event on the root so the timeline shows the fan-out. + if author and author.strip(): + conn.execute( + "INSERT INTO task_comments (task_id, author, body, created_at) " + "VALUES (?, ?, ?, ?)", + ( + task_id, + author.strip(), + "Decomposed into " + + ", ".join(child_ids) + + ". Root will wake when all children complete.", + now, + ), + ) + _append_event( + conn, task_id, "decomposed", + { + "child_ids": child_ids, + "root_assignee": root_assignee, + }, + ) + + # Outside the write_txn: promote parent-free children to 'ready' + # so the dispatcher picks them up on its next tick. Same pattern + # specify_triage_task uses. + recompute_ready(conn) + return child_ids + + def archive_task(conn: sqlite3.Connection, task_id: str) -> bool: with write_txn(conn): cur = conn.execute( @@ -4024,6 +4199,7 @@ def _default_spawn( stderr=subprocess.STDOUT, env=env, start_new_session=True, + creationflags=subprocess.CREATE_NO_WINDOW if _IS_WINDOWS else 0, ) except FileNotFoundError: log_f.close() diff --git a/hermes_cli/kanban_decompose.py b/hermes_cli/kanban_decompose.py new file mode 100644 index 00000000000..2ebe3f04c6e --- /dev/null +++ b/hermes_cli/kanban_decompose.py @@ -0,0 +1,440 @@ +"""Kanban decomposer — fan a triage task out into a graph of child tasks. + +Invoked by ``hermes kanban decompose [task_id | --all]`` and the +auto-decompose path in the gateway dispatcher loop. Reads the user's +profile roster (with descriptions) and asks the auxiliary LLM to +return a task graph in JSON. Then atomically creates the children, +links them under the root, and flips the root ``triage -> todo``. + +The root task stays alive and becomes the parent of every leaf child, +so when the whole graph completes the root wakes back up — its +assignee (the orchestrator profile) gets a chance to judge completion +and add more tasks if the work isn't done yet. + +Design notes +------------ + +* Mirrors the shape of ``hermes_cli/kanban_specify.py``: lazy aux + client import inside the function, lenient response parse, never + raises on expected failure modes. + +* The system prompt sees the *configured* profile roster — names plus + descriptions plus the default fallback. Profiles without a + description are still listed (with a note) so the orchestrator can + match on name as a fallback, but the user has an obvious incentive + to describe them. + +* ``fanout=false`` collapses to the same effect as ``kanban specify``: + we tighten the body and flip ``triage -> todo`` as a single task, + no children created. This makes ``decompose`` a strict superset of + ``specify`` from the user's perspective. + +* If the LLM picks an assignee that doesn't exist as a profile, we + rewrite it to the configured ``default_assignee`` (or the default + profile if unset). A child task NEVER ends up with ``assignee=None``. +""" + +from __future__ import annotations + +import json +import logging +import os +import re +from dataclasses import dataclass +from typing import Optional + +from hermes_cli import kanban_db as kb +from hermes_cli import profiles as profiles_mod + +logger = logging.getLogger(__name__) + + +_SYSTEM_PROMPT = """You are the Kanban decomposer for the Hermes Agent board. + +A user dropped a rough idea into the Triage column. Your job is to break it +into a small graph of concrete child tasks and route each one to the best- +matching profile from the available roster. + +You will be given: + - The original task title and body + - The list of available profiles (each with name + description) + - The fallback "default_assignee" used when no profile fits + +Output a single JSON object with this exact shape: + + { + "fanout": true, + "rationale": "", + "tasks": [ + { + "title": "", + "body": "", + "assignee": "", + "parents": [, ...] + }, + ... + ] + } + +Rules: + - "parents" is a list of INDICES (0-based) into this same "tasks" list, + expressing actual data dependencies. Tasks with no parents run in + PARALLEL. Tasks with parents wait until every parent completes. + - Prefer parallelism. If two tasks can be done independently, give + them no parents so the dispatcher fans them out at once. + - Use 2-6 tasks for normal work. Don't create 20 tiny tasks. Don't + cram everything into 1 task. + - Pick assignees from the roster by matching the task to the profile's + DESCRIPTION (not just the name). When nothing matches well, use null + and the system will route to the default_assignee. + - Each child task body is what a fresh worker will read with no other + context — be specific about goal, approach, and acceptance criteria. + +When the task is genuinely a single unit of work (no useful decomposition), +return: + + { + "fanout": false, + "rationale": "", + "title": "", + "body": "" + } + +In that case the task stays as one work item, just with a tightened spec. + +No preamble, no closing remarks, no code fences. Output only the JSON object. +""" + + +_USER_TEMPLATE = """Task id: {task_id} +Title: {title} +Body: +{body} + +Available profiles (assignees you may pick from): +{roster} + +Default assignee (used when no profile fits a task): {default_assignee} +""" + + +_FENCE_RE = re.compile(r"^```(?:json)?\s*|\s*```$", re.MULTILINE) + + +@dataclass +class DecomposeOutcome: + """Result of decomposing a single triage task.""" + + task_id: str + ok: bool + reason: str = "" + fanout: bool = False + child_ids: list[str] | None = None + new_title: Optional[str] = None + + +def _truncate(text: str, limit: int) -> str: + if len(text) <= limit: + return text + return text[: limit - 1] + "…" + + +def _extract_json_blob(raw: str) -> Optional[dict]: + if not raw: + return None + stripped = _FENCE_RE.sub("", raw.strip()) + first = stripped.find("{") + last = stripped.rfind("}") + if first == -1 or last == -1 or last <= first: + return None + candidate = stripped[first : last + 1] + try: + val = json.loads(candidate) + except (ValueError, json.JSONDecodeError): + return None + if not isinstance(val, dict): + return None + return val + + +def _profile_author() -> str: + """Mirror of ``hermes_cli.kanban._profile_author``.""" + return ( + os.environ.get("HERMES_PROFILE") + or os.environ.get("USER") + or "decomposer" + ) + + +def _load_config() -> dict: + try: + from hermes_cli.config import load_config + return load_config() or {} + except Exception: + return {} + + +def _resolve_orchestrator_profile(cfg: dict) -> str: + """Resolve which profile owns decomposition. + + Falls back to the active default profile when ``kanban.orchestrator_profile`` + is unset, so a task is never stranded for lack of an orchestrator. + """ + kanban_cfg = cfg.get("kanban", {}) if isinstance(cfg, dict) else {} + explicit = (kanban_cfg.get("orchestrator_profile") or "").strip() + if explicit: + try: + if profiles_mod.profile_exists(explicit): + return explicit + except Exception: + pass + # Fall back to the active default profile. + try: + return profiles_mod.get_active_profile_name() or "default" + except Exception: + return "default" + + +def _resolve_default_assignee(cfg: dict) -> str: + """Resolve which profile catches child tasks the orchestrator can't route.""" + kanban_cfg = cfg.get("kanban", {}) if isinstance(cfg, dict) else {} + explicit = (kanban_cfg.get("default_assignee") or "").strip() + if explicit: + try: + if profiles_mod.profile_exists(explicit): + return explicit + except Exception: + pass + try: + return profiles_mod.get_active_profile_name() or "default" + except Exception: + return "default" + + +def _build_roster() -> tuple[list[dict], set[str]]: + """Return (roster_for_prompt, valid_assignee_names). + + Each roster entry is ``{name, description, has_description}``. The + valid-set is used after the LLM responds to rewrite invalid + assignees to the default fallback. + """ + roster: list[dict] = [] + valid: set[str] = set() + try: + all_profiles = profiles_mod.list_profiles() + except Exception as exc: + logger.warning("decompose: failed to list profiles: %s", exc) + return roster, valid + for p in all_profiles: + desc = (p.description or "").strip() + roster.append({ + "name": p.name, + "description": desc or f"(no description; profile named {p.name!r})", + "has_description": bool(desc), + }) + valid.add(p.name) + return roster, valid + + +def _format_roster(roster: list[dict]) -> str: + if not roster: + return " (no profiles installed — decomposer cannot route work)" + lines = [] + for entry in roster: + tag = "" if entry["has_description"] else " ⚠ undescribed" + lines.append(f" - {entry['name']}{tag}: {entry['description']}") + return "\n".join(lines) + + +def decompose_task( + task_id: str, + *, + author: Optional[str] = None, + timeout: Optional[int] = None, +) -> DecomposeOutcome: + """Decompose a triage task into a graph of child tasks. + + Returns an outcome describing what happened. Never raises for + expected failure modes (task not in triage, no aux client + configured, API error, malformed response, decomposer returned + fanout=true with empty task list) — those surface via ``ok=False``. + """ + with kb.connect() as conn: + task = kb.get_task(conn, task_id) + if task is None: + return DecomposeOutcome(task_id, False, "unknown task id") + if task.status != "triage": + return DecomposeOutcome( + task_id, False, f"task is not in triage (status={task.status!r})" + ) + + cfg = _load_config() + orchestrator = _resolve_orchestrator_profile(cfg) + default_assignee = _resolve_default_assignee(cfg) + roster, valid_names = _build_roster() + + try: + from agent.auxiliary_client import ( # type: ignore + get_auxiliary_extra_body, + get_text_auxiliary_client, + ) + except Exception as exc: + logger.debug("decompose: auxiliary client import failed: %s", exc) + return DecomposeOutcome(task_id, False, "auxiliary client unavailable") + + try: + client, model = get_text_auxiliary_client("kanban_decomposer") + except Exception as exc: + logger.debug("decompose: get_text_auxiliary_client failed: %s", exc) + return DecomposeOutcome(task_id, False, "auxiliary client unavailable") + + if client is None or not model: + return DecomposeOutcome(task_id, False, "no auxiliary client configured") + + user_msg = _USER_TEMPLATE.format( + task_id=task.id, + title=_truncate(task.title or "", 400), + body=_truncate(task.body or "(no body)", 4000), + roster=_format_roster(roster), + default_assignee=default_assignee, + ) + + try: + resp = client.chat.completions.create( + model=model, + messages=[ + {"role": "system", "content": _SYSTEM_PROMPT}, + {"role": "user", "content": user_msg}, + ], + temperature=0.3, + max_tokens=4000, + timeout=timeout or 180, + extra_body=get_auxiliary_extra_body() or None, + ) + except Exception as exc: + logger.info( + "decompose: API call failed for %s (%s)", task_id, exc, + ) + return DecomposeOutcome(task_id, False, f"LLM error: {type(exc).__name__}") + + try: + raw = resp.choices[0].message.content or "" + except Exception: + raw = "" + + parsed = _extract_json_blob(raw) + if parsed is None: + return DecomposeOutcome(task_id, False, "LLM returned malformed JSON") + + fanout = bool(parsed.get("fanout")) + audit_author = author or _profile_author() + + if not fanout: + # Fall back to single-task spec promotion (same effect as specify). + new_title = parsed.get("title") + new_body = parsed.get("body") + title_val = new_title.strip() if isinstance(new_title, str) and new_title.strip() else None + body_val = new_body if isinstance(new_body, str) and new_body.strip() else None + if title_val is None and body_val is None: + return DecomposeOutcome( + task_id, False, "decomposer returned fanout=false with no title/body", + ) + with kb.connect() as conn: + ok = kb.specify_triage_task( + conn, + task_id, + title=title_val, + body=body_val, + author=audit_author, + ) + if not ok: + return DecomposeOutcome( + task_id, False, "task moved out of triage before promotion", + ) + return DecomposeOutcome( + task_id, True, "single task (no fanout)", + fanout=False, new_title=title_val, + ) + + raw_tasks = parsed.get("tasks") or [] + if not isinstance(raw_tasks, list) or not raw_tasks: + return DecomposeOutcome( + task_id, False, "decomposer returned fanout=true with empty tasks list", + ) + + # Rewrite invalid assignees to the default fallback. Never leave a + # task with assignee=None — the user explicitly does not want that. + children: list[dict] = [] + for idx, entry in enumerate(raw_tasks): + if not isinstance(entry, dict): + return DecomposeOutcome( + task_id, False, f"tasks[{idx}] is not an object", + ) + title = entry.get("title") + if not isinstance(title, str) or not title.strip(): + return DecomposeOutcome( + task_id, False, f"tasks[{idx}].title is missing or empty", + ) + body = entry.get("body") + if not isinstance(body, str): + body = "" + assignee = entry.get("assignee") + if not isinstance(assignee, str) or not assignee.strip(): + chosen = default_assignee + elif assignee not in valid_names: + logger.info( + "decompose: task %s child %d picked unknown assignee %r — " + "routing to default_assignee %r", + task_id, idx, assignee, default_assignee, + ) + chosen = default_assignee + else: + chosen = assignee + parents = entry.get("parents") or [] + if not isinstance(parents, list): + parents = [] + # Clean parent indices: drop non-int and out-of-range. + clean_parents = [p for p in parents if isinstance(p, int) and 0 <= p < len(raw_tasks) and p != idx] + children.append({ + "title": title.strip()[:200], + "body": body.strip(), + "assignee": chosen, + "parents": clean_parents, + }) + + try: + with kb.connect() as conn: + child_ids = kb.decompose_triage_task( + conn, + task_id, + root_assignee=orchestrator, + children=children, + author=audit_author, + ) + except ValueError as exc: + return DecomposeOutcome(task_id, False, f"DB rejected graph: {exc}") + except Exception as exc: + logger.exception("decompose: DB error on task %s", task_id) + return DecomposeOutcome(task_id, False, f"DB error: {type(exc).__name__}") + + if child_ids is None: + return DecomposeOutcome( + task_id, False, "task moved out of triage before decomposition", + ) + + return DecomposeOutcome( + task_id, True, f"decomposed into {len(child_ids)} children", + fanout=True, child_ids=child_ids, + ) + + +def list_triage_ids(*, tenant: Optional[str] = None) -> list[str]: + """Return task ids currently in the triage column.""" + with kb.connect() as conn: + rows = kb.list_tasks( + conn, + status="triage", + tenant=tenant, + limit=1000, + ) + return [row.id for row in rows] diff --git a/hermes_cli/main.py b/hermes_cli/main.py index 4503d31da2a..bdb43cc504c 100644 --- a/hermes_cli/main.py +++ b/hermes_cli/main.py @@ -9082,6 +9082,7 @@ def cmd_profile(args): clone_config=clone, no_alias=no_alias, no_skills=no_skills, + description=getattr(args, "description", None), ) print(f"\nProfile '{name}' created at {profile_dir}") @@ -9181,6 +9182,107 @@ def cmd_profile(args): print(f"Error: {e}") sys.exit(1) + elif action == "describe": + # Read or write a profile's description. The description is + # consumed by the kanban decomposer to route tasks based on + # role instead of name alone. + from hermes_cli import profiles as _profiles_mod + + all_flag = bool(getattr(args, "all_missing", False)) + auto_flag = bool(getattr(args, "auto", False)) + overwrite_flag = bool(getattr(args, "overwrite", False)) + text_value = getattr(args, "text", None) + name = getattr(args, "profile_name", None) + + if all_flag and not auto_flag: + print("profile describe: --all requires --auto", file=sys.stderr) + sys.exit(2) + if all_flag and (text_value or name): + print( + "profile describe: --all is mutually exclusive with a profile name / --text", + file=sys.stderr, + ) + sys.exit(2) + if not all_flag and not name: + print("profile describe: profile name is required (or --all --auto)", file=sys.stderr) + sys.exit(2) + if text_value and auto_flag: + print( + "profile describe: --text is mutually exclusive with --auto", + file=sys.stderr, + ) + sys.exit(2) + + # Show current description if no operation requested. + if name and not text_value and not auto_flag: + try: + if _profiles_mod.normalize_profile_name(name) == "default": + from hermes_constants import get_hermes_home as _hh + profile_dir = Path(_hh()) + else: + profile_dir = _profiles_mod.get_profile_dir(name) + except Exception as exc: + print(f"Error: {exc}", file=sys.stderr) + sys.exit(1) + if not profile_dir.is_dir(): + print(f"Error: profile '{name}' not found", file=sys.stderr) + sys.exit(1) + meta = _profiles_mod.read_profile_meta(profile_dir) + desc = meta.get("description") or "" + if not desc: + print(f"(no description set for '{name}')") + else: + tag = "[auto] " if meta.get("description_auto") else "" + print(f"{tag}{desc}") + sys.exit(0) + + # --text path: just write the user-authored description. + if text_value: + try: + if _profiles_mod.normalize_profile_name(name) == "default": + from hermes_constants import get_hermes_home as _hh + profile_dir = Path(_hh()) + else: + profile_dir = _profiles_mod.get_profile_dir(name) + _profiles_mod.write_profile_meta( + profile_dir, + description=text_value, + description_auto=False, + ) + print(f"Description updated for '{name}'.") + except Exception as exc: + print(f"Error: {exc}", file=sys.stderr) + sys.exit(1) + sys.exit(0) + + # --auto path: invoke the LLM describer. + from hermes_cli import profile_describer as _pd + + if all_flag: + targets = _pd.list_describable_profiles(missing_only=True) + if not targets: + print("All profiles already have descriptions.") + sys.exit(0) + else: + targets = [name] + + ok_count = 0 + fail_count = 0 + for tgt in targets: + outcome = _pd.describe_profile(tgt, overwrite=overwrite_flag) + if outcome.ok: + ok_count += 1 + print(f"Described '{outcome.profile_name}': {outcome.description}") + else: + fail_count += 1 + print( + f"profile describe {outcome.profile_name}: {outcome.reason}", + file=sys.stderr, + ) + if not all_flag: + sys.exit(0 if ok_count == 1 else 1) + sys.exit(0 if ok_count > 0 else 1) + elif action == "show": name = args.profile_name from hermes_cli.profiles import ( @@ -9684,8 +9786,8 @@ _BUILTIN_SUBCOMMANDS = frozenset( "config", "cron", "curator", "dashboard", "debug", "doctor", "dump", "fallback", "gateway", "hooks", "import", "insights", "kanban", "login", "logout", "logs", "lsp", "mcp", "memory", - "model", "pairing", "plugins", "postinstall", "profile", "proxy", "send", - "sessions", "setup", + "model", "pairing", "plugins", "postinstall", "profile", "proxy", + "send", "sessions", "setup", "skills", "slack", "status", "tools", "uninstall", "update", "version", "webhook", "whatsapp", "chat", # Help-ish invocations — plugin commands not being listed in @@ -12076,6 +12178,13 @@ Examples: action="store_true", help="Create an empty profile with no bundled skills (opts out of `hermes update` skill sync)", ) + profile_create.add_argument( + "--description", + default=None, + help="One- or two-sentence description of what this profile is good at. " + "Used by the kanban decomposer to route tasks based on role instead " + "of profile name alone. Skip and add later via `hermes profile describe`.", + ) profile_delete = profile_subparsers.add_parser("delete", help="Delete a profile") profile_delete.add_argument("profile_name", help="Profile to delete") @@ -12083,6 +12192,40 @@ Examples: "-y", "--yes", action="store_true", help="Skip confirmation prompt" ) + profile_describe = profile_subparsers.add_parser( + "describe", + help="Read or set a profile's description (used by the kanban orchestrator)", + ) + profile_describe.add_argument( + "profile_name", + nargs="?", + default=None, + help="Profile to describe (omit + use --all --auto to sweep)", + ) + profile_describe.add_argument( + "--text", + default=None, + help="Set description to this exact text (overwrites any existing description)", + ) + profile_describe.add_argument( + "--auto", + action="store_true", + help="Auto-generate description via the auxiliary LLM " + "(uses auxiliary.profile_describer)", + ) + profile_describe.add_argument( + "--overwrite", + action="store_true", + help="With --auto, replace user-authored descriptions too (default: only " + "fill in missing or previously-auto descriptions)", + ) + profile_describe.add_argument( + "--all", + dest="all_missing", + action="store_true", + help="With --auto, run on every profile missing a description", + ) + profile_show = profile_subparsers.add_parser("show", help="Show profile details") profile_show.add_argument("profile_name", help="Profile to show") diff --git a/hermes_cli/model_switch.py b/hermes_cli/model_switch.py index fec1f33d092..727905270e1 100644 --- a/hermes_cli/model_switch.py +++ b/hermes_cli/model_switch.py @@ -1688,7 +1688,26 @@ def list_authenticated_providers( continue # Live model discovery from custom provider endpoints (matches # Section 3 behavior for user ``providers:`` entries). - if api_url and api_key: + # Also probes when no api_key is set (e.g. local llama.cpp / + # Ollama servers) — the /models endpoint often works without + # auth. The CLI's _model_flow_named_custom always probes, so + # the Telegram/Discord picker should do the same for parity. + # Live-discovery policy: + # - With an api_key, the user has explicitly opted into the + # endpoint and live /models is the source of truth — replace + # the (possibly partial) ``models:`` subset configured for + # context-length overrides with the full live catalog. + # This is the Bifrost / aggregator-gateway case. + # - Without an api_key but with an explicit ``models:`` list + # (or top-level ``model:``), the user is narrowing a public + # endpoint to a specific subset (e.g. ollama.com /v1/models + # returns 35 models but the user only wants 4). Preserve the + # explicit list and skip live discovery. + # - Without an api_key AND no explicit models, fall through to + # live discovery so bare-endpoint custom providers (local + # llama.cpp / Ollama servers) still appear populated. + should_probe = bool(api_url) and (bool(api_key) or not grp["models"]) + if should_probe: try: from hermes_cli.models import fetch_api_models diff --git a/hermes_cli/plugins.py b/hermes_cli/plugins.py index d0bbee6ce63..6150bf016d1 100644 --- a/hermes_cli/plugins.py +++ b/hermes_cli/plugins.py @@ -608,6 +608,38 @@ class PluginContext: self.manifest.name, provider.name, ) + # -- browser provider registration --------------------------------------- + + def register_browser_provider(self, provider) -> None: + """Register a cloud browser backend. + + ``provider`` must be an instance of + :class:`agent.browser_provider.BrowserProvider`. The + ``provider.name`` attribute is what ``browser.cloud_provider`` in + ``config.yaml`` matches against when routing cloud-mode + ``browser_*`` tool calls. + + Mirrors :meth:`register_web_search_provider` exactly — same + registration shape, same gating, same logging. The browser + subsystem's dispatcher (:func:`tools.browser_tool._get_cloud_provider`) + consults the registry built up by these calls. + """ + from agent.browser_provider import BrowserProvider + from agent.browser_registry import register_provider as _register_browser_provider + + if not isinstance(provider, BrowserProvider): + logger.warning( + "Plugin '%s' tried to register a browser provider that does " + "not inherit from BrowserProvider. Ignoring.", + self.manifest.name, + ) + return + _register_browser_provider(provider) + logger.info( + "Plugin '%s' registered browser provider: %s", + self.manifest.name, provider.name, + ) + # -- platform adapter registration --------------------------------------- def register_platform( diff --git a/hermes_cli/profile_describer.py b/hermes_cli/profile_describer.py new file mode 100644 index 00000000000..55d646d92cd --- /dev/null +++ b/hermes_cli/profile_describer.py @@ -0,0 +1,299 @@ +"""Profile describer — auto-generate ``description`` for a profile. + +Used by ``hermes profile describe --auto`` and the dashboard's +"auto-generate description" button. Reads the profile's installed +skills, model+provider, name, and optionally a small slice of memory, +then asks the auxiliary LLM to produce a 1-2 sentence description of +what the profile is good at. + +Result is written to ``/profile.yaml`` with +``description_auto: true`` so the dashboard can surface a "review" +badge. User can edit afterward to confirm. + +Design notes +------------ +- Mirrors the shape of ``hermes_cli/kanban_specify.py``: lazy aux + client import inside the function, lenient response parse, never + raises on expected failure modes. +- Reads at most ``MAX_SKILLS_FOR_PROMPT`` skill names to keep the + prompt bounded. No skill body — names + categories are enough + signal and avoid blowing context on profiles with 100+ skills. +- Memory is intentionally NOT read here. Memories are personal and + the orchestrator routes work to a *role* not a *biography*. If we + find later that memory adds signal we can wire it; for now, + skills + name + model is plenty. +""" + +from __future__ import annotations + +import json +import logging +import os +import re +from dataclasses import dataclass +from pathlib import Path +from typing import Optional + +from hermes_cli import profiles as profiles_mod + +logger = logging.getLogger(__name__) + +# Cap on how many skill names we feed the LLM. Profiles with 200+ +# skills (uncommon but possible) would blow context otherwise. The cap +# is per-category — see _collect_skills. +MAX_SKILLS_FOR_PROMPT = 60 + + +_SYSTEM_PROMPT = """You are a profile-describer for the Hermes Agent kanban board. + +A user runs multiple "profiles" — distinct agent identities, each with their +own skills, model, and configuration. The kanban board's orchestrator routes +work to whichever profile best fits each task. To do that well, every +profile needs a short, concrete description of what it's good at. + +You are given a profile's: + - Name + - Model / provider + - List of installed skill names (a strong signal of role / domain) + +Produce a single JSON object with exactly one key: + + { + "description": "<1-2 sentence description, plain prose, no preamble>" + } + +Rules: + - The description is what an orchestrator will read to decide whether to + route a task here. Lead with the profile's strongest capability. + - Stay concrete. Bad: "an AI agent that helps users." + Good: "Reads and modifies Python codebases — runs tests, + refactors functions, opens GitHub PRs." + - 1-2 sentences, <= 280 characters total. + - Never invent capabilities the skills don't suggest. + - Never write "Hermes Agent profile" or other meta-narration. + - No code fences, no preamble, no closing remarks. Output only JSON. +""" + + +_USER_TEMPLATE = """Profile name: {name} +Default model: {model} +Provider: {provider} +Installed skill count: {skill_count} +Notable skills (up to {skill_cap}): +{skill_list} +""" + + +_FENCE_RE = re.compile(r"^```(?:json)?\s*|\s*```$", re.MULTILINE) + + +@dataclass +class DescribeOutcome: + """Result of describing a single profile.""" + + profile_name: str + ok: bool + reason: str = "" + description: Optional[str] = None + + +def _collect_skills(profile_dir: Path) -> list[str]: + """Return a stable, capped list of skill names for the prompt. + + Format: ``category/skill_name`` where category is the immediate + subdir under ``skills/`` (e.g. ``devops``, ``research``). Skills + that live directly under ``skills/`` show as bare ``skill_name``. + """ + skills_dir = profile_dir / "skills" + if not skills_dir.is_dir(): + return [] + names: list[str] = [] + for md in skills_dir.rglob("SKILL.md"): + path_str = str(md) + if "/.hub/" in path_str or "/.git/" in path_str: + continue + try: + rel = md.relative_to(skills_dir) + except ValueError: + continue + parts = rel.parts[:-1] # drop SKILL.md filename + if not parts: + continue + # parts[-1] is the skill dir name; parts[:-1] is the category path + if len(parts) == 1: + names.append(parts[0]) + else: + names.append(f"{parts[0]}/{parts[-1]}") + names.sort() + # Keep within prompt budget. Skills earlier in alphabet aren't more + # important — we'll let the LLM see a sample. Pick evenly-spaced + # entries instead of just the head so a profile with skills A..Z + # doesn't get described as "starts with A". + if len(names) <= MAX_SKILLS_FOR_PROMPT: + return names + step = len(names) / MAX_SKILLS_FOR_PROMPT + sampled = [names[int(i * step)] for i in range(MAX_SKILLS_FOR_PROMPT)] + return sampled + + +def _extract_json_blob(raw: str) -> Optional[dict]: + if not raw: + return None + stripped = _FENCE_RE.sub("", raw.strip()) + first = stripped.find("{") + last = stripped.rfind("}") + if first == -1 or last == -1 or last <= first: + return None + candidate = stripped[first : last + 1] + try: + val = json.loads(candidate) + except (ValueError, json.JSONDecodeError): + return None + if not isinstance(val, dict): + return None + return val + + +def describe_profile( + profile_name: str, + *, + overwrite: bool = False, + timeout: Optional[int] = None, +) -> DescribeOutcome: + """Auto-generate a description for one profile. + + Returns an outcome describing what happened. Never raises for + expected failure modes (profile missing, no aux client configured, + API error, malformed response) — those surface via ``ok=False`` so + a sweep can continue past individual failures. + + ``overwrite`` controls whether an existing user-authored description + is replaced. By default we refuse to overwrite a description with + ``description_auto: false`` to protect curated text. Auto-generated + descriptions (``description_auto: true``) are always replaceable. + """ + canon = profiles_mod.normalize_profile_name(profile_name) + if not profiles_mod.profile_exists(canon): + # Special case: "default" exists as a virtual profile name + # mapped to the default home dir. profile_exists() handles it. + return DescribeOutcome(canon, False, "profile not found") + + try: + if canon == "default": + from hermes_constants import get_hermes_home # type: ignore + profile_dir = Path(get_hermes_home()) + else: + profile_dir = profiles_mod.get_profile_dir(canon) + except Exception as exc: + return DescribeOutcome(canon, False, f"cannot resolve profile dir: {exc}") + + # Honor curated descriptions unless --overwrite. + existing = profiles_mod.read_profile_meta(profile_dir) + if existing.get("description") and not existing.get("description_auto") and not overwrite: + return DescribeOutcome( + canon, + False, + "profile already has a user-authored description " + "(use --overwrite to replace)", + ) + + skill_names = _collect_skills(profile_dir) + skill_list = "\n".join(f" - {n}" for n in skill_names) or " (no skills installed)" + skill_count = sum( + 1 for _ in (profile_dir / "skills").rglob("SKILL.md") + if "/.hub/" not in str(_) and "/.git/" not in str(_) + ) if (profile_dir / "skills").is_dir() else 0 + + # Read model + provider from the profile's config. + try: + model, provider = profiles_mod._read_config_model(profile_dir) + except Exception: + model, provider = None, None + + try: + from agent.auxiliary_client import ( # type: ignore + get_auxiliary_extra_body, + get_text_auxiliary_client, + ) + except Exception as exc: + logger.debug("describe: auxiliary client import failed: %s", exc) + return DescribeOutcome(canon, False, "auxiliary client unavailable") + + try: + client, aux_model = get_text_auxiliary_client("profile_describer") + except Exception as exc: + logger.debug("describe: get_text_auxiliary_client failed: %s", exc) + return DescribeOutcome(canon, False, "auxiliary client unavailable") + + if client is None or not aux_model: + return DescribeOutcome(canon, False, "no auxiliary client configured") + + user_msg = _USER_TEMPLATE.format( + name=canon, + model=(model or "(unset)"), + provider=(provider or "(unset)"), + skill_count=skill_count, + skill_cap=MAX_SKILLS_FOR_PROMPT, + skill_list=skill_list, + ) + + try: + resp = client.chat.completions.create( + model=aux_model, + messages=[ + {"role": "system", "content": _SYSTEM_PROMPT}, + {"role": "user", "content": user_msg}, + ], + temperature=0.3, + max_tokens=400, + timeout=timeout or 60, + extra_body=get_auxiliary_extra_body() or None, + ) + except Exception as exc: + logger.info("describe: API call failed for %s (%s)", canon, exc) + return DescribeOutcome(canon, False, f"LLM error: {type(exc).__name__}") + + try: + raw = resp.choices[0].message.content or "" + except Exception: + raw = "" + + parsed = _extract_json_blob(raw) + if parsed is None: + # Fall back: take the raw text trimmed to one paragraph. + text = raw.strip().split("\n\n", 1)[0] + if not text: + return DescribeOutcome(canon, False, "LLM returned an empty response") + description = text[:280] + else: + val = parsed.get("description") + if not isinstance(val, str) or not val.strip(): + return DescribeOutcome( + canon, False, "LLM response missing 'description' field" + ) + description = val.strip()[:280] + + try: + profiles_mod.write_profile_meta( + profile_dir, + description=description, + description_auto=True, + ) + except Exception as exc: + return DescribeOutcome(canon, False, f"failed to write profile.yaml: {exc}") + + return DescribeOutcome(canon, True, "described", description=description) + + +def list_describable_profiles(*, missing_only: bool = True) -> list[str]: + """Return profile names that can be described. + + ``missing_only=True`` (default) returns only profiles without a + description. ``missing_only=False`` returns every profile. + """ + out: list[str] = [] + for p in profiles_mod.list_profiles(): + if missing_only and (p.description or "").strip() and not p.description_auto: + continue + out.append(p.name) + return out diff --git a/hermes_cli/profiles.py b/hermes_cli/profiles.py index de555caf9be..d35669c6243 100644 --- a/hermes_cli/profiles.py +++ b/hermes_cli/profiles.py @@ -412,6 +412,17 @@ class ProfileInfo: distribution_name: Optional[str] = None distribution_version: Optional[str] = None distribution_source: Optional[str] = None + # Free-form description (1-2 sentences) of what this profile is good + # at. Persisted in ``/profile.yaml``. Empty when the + # user has not described the profile (legacy profiles, fresh + # installs). Surfaced to the kanban decomposer so it can route work + # to the right profile based on role rather than name alone. + description: str = "" + # When True, ``description`` was auto-generated by the LLM + # describer and has not been confirmed by the user. The dashboard + # surfaces a "review" badge in this case so the user can edit or + # accept. + description_auto: bool = False def _read_distribution_meta(profile_dir: Path) -> tuple: @@ -479,6 +490,82 @@ def _count_skills(profile_dir: Path) -> int: return count +# --------------------------------------------------------------------------- +# profile.yaml — per-profile metadata (description, role, etc.) +# --------------------------------------------------------------------------- +# +# We keep this file deliberately tiny and separate from the profile's +# ``config.yaml``. ``config.yaml`` is the user-facing Hermes config +# (~5000 lines of defaults); ``profile.yaml`` is metadata ABOUT the +# profile itself (its role, who described it). Mixing them makes both +# harder to read. +# +# Missing file -> empty defaults; never an error. The kanban decomposer +# tolerates empty descriptions and just falls back to the profile name. + + +def _profile_yaml_path(profile_dir: Path) -> Path: + return profile_dir / "profile.yaml" + + +def read_profile_meta(profile_dir: Path) -> dict: + """Read ``/profile.yaml`` and return a dict. + + Returns ``{"description": "", "description_auto": False}`` when the + file is missing or unreadable. Never raises — a corrupt + profile.yaml on an unrelated profile must not break + ``hermes profile list``. + """ + path = _profile_yaml_path(profile_dir) + if not path.is_file(): + return {"description": "", "description_auto": False} + try: + import yaml + with open(path, "r", encoding="utf-8") as f: + data = yaml.safe_load(f) or {} + except Exception: + return {"description": "", "description_auto": False} + if not isinstance(data, dict): + return {"description": "", "description_auto": False} + return { + "description": str(data.get("description") or "").strip(), + "description_auto": bool(data.get("description_auto", False)), + } + + +def write_profile_meta( + profile_dir: Path, + *, + description: Optional[str] = None, + description_auto: Optional[bool] = None, +) -> None: + """Update ``/profile.yaml`` in place. + + Only the explicitly passed fields are overwritten; unspecified + fields preserve existing values. Creates the file if missing. + Profile directory itself must exist. + """ + if not profile_dir.is_dir(): + raise FileNotFoundError(f"profile directory does not exist: {profile_dir}") + import yaml + path = _profile_yaml_path(profile_dir) + existing: dict = {} + if path.is_file(): + try: + with open(path, "r", encoding="utf-8") as f: + loaded = yaml.safe_load(f) or {} + if isinstance(loaded, dict): + existing = loaded + except Exception: + existing = {} + if description is not None: + existing["description"] = description.strip() + if description_auto is not None: + existing["description_auto"] = bool(description_auto) + with open(path, "w", encoding="utf-8") as f: + yaml.safe_dump(existing, f, sort_keys=False, default_flow_style=False) + + # --------------------------------------------------------------------------- # CRUD operations # --------------------------------------------------------------------------- @@ -493,6 +580,7 @@ def list_profiles() -> List[ProfileInfo]: if default_home.is_dir(): model, provider = _read_config_model(default_home) dist_name, dist_version, dist_source = _read_distribution_meta(default_home) + meta = read_profile_meta(default_home) profiles.append(ProfileInfo( name="default", path=default_home, @@ -505,6 +593,8 @@ def list_profiles() -> List[ProfileInfo]: distribution_name=dist_name, distribution_version=dist_version, distribution_source=dist_source, + description=meta.get("description", ""), + description_auto=meta.get("description_auto", False), )) # Named profiles @@ -519,6 +609,7 @@ def list_profiles() -> List[ProfileInfo]: model, provider = _read_config_model(entry) alias_path = wrapper_dir / name dist_name, dist_version, dist_source = _read_distribution_meta(entry) + meta = read_profile_meta(entry) profiles.append(ProfileInfo( name=name, path=entry, @@ -532,6 +623,8 @@ def list_profiles() -> List[ProfileInfo]: distribution_name=dist_name, distribution_version=dist_version, distribution_source=dist_source, + description=meta.get("description", ""), + description_auto=meta.get("description_auto", False), )) return profiles @@ -544,6 +637,7 @@ def create_profile( clone_config: bool = False, no_alias: bool = False, no_skills: bool = False, + description: Optional[str] = None, ) -> Path: """Create a new profile directory. @@ -667,6 +761,19 @@ def create_profile( except OSError: pass # best-effort — the feature still works via the empty skills/ dir + # Persist description if the caller provided one. Done last so a + # partial-create failure doesn't strand a description file in an + # incomplete profile. + if description and description.strip(): + try: + write_profile_meta( + profile_dir, + description=description.strip(), + description_auto=False, + ) + except Exception: + pass # non-fatal — user can describe later with `hermes profile describe` + return profile_dir diff --git a/hermes_cli/proxy/adapters/base.py b/hermes_cli/proxy/adapters/base.py index 5ac8a5dcedd..db778e18fa9 100644 --- a/hermes_cli/proxy/adapters/base.py +++ b/hermes_cli/proxy/adapters/base.py @@ -81,6 +81,21 @@ class UpstreamAdapter(ABC): refresh fails. The proxy will return 401 to the client. """ + def get_retry_credential( + self, + *, + failed_credential: UpstreamCredential, + status_code: int, + ) -> Optional[UpstreamCredential]: + """Return an alternate credential after an upstream auth failure. + + The default is no retry. Providers can override this for one-shot + fallback paths, such as switching from a preferred token type to a + legacy bearer after the upstream rejects the first request. + """ + _ = failed_credential, status_code + return None + def describe(self) -> str: """One-line status summary for ``proxy status``.""" try: diff --git a/hermes_cli/proxy/adapters/nous_portal.py b/hermes_cli/proxy/adapters/nous_portal.py index b72cbd305b3..9fb07a9c053 100644 --- a/hermes_cli/proxy/adapters/nous_portal.py +++ b/hermes_cli/proxy/adapters/nous_portal.py @@ -1,12 +1,13 @@ """Nous Portal upstream adapter. -Reads the user's Nous OAuth state from ``~/.hermes/auth.json``, refreshes -the access token and mints a fresh agent key when needed, and exposes the -upstream base URL plus minted bearer for the proxy server to forward to. +Reads the user's Nous OAuth state from ``~/.hermes/auth.json`` through the +shared runtime resolver, refreshes the access token and resolves the +``agent_key`` compatibility credential when needed, then exposes the upstream +base URL plus bearer for the proxy server to forward to. -The minted ``agent_key`` (not the OAuth ``access_token``) is what -``inference-api.nousresearch.com`` accepts as a bearer. The refresh helper -already handles both — see :func:`hermes_cli.auth.refresh_nous_oauth_from_state`. +The ``agent_key`` field may hold either a NAS invoke JWT or the legacy +opaque session key. The refresh helper handles both — see +:func:`hermes_cli.auth.resolve_nous_runtime_credentials`. """ from __future__ import annotations @@ -16,11 +17,18 @@ import threading from typing import Any, Dict, FrozenSet, Optional from hermes_cli.auth import ( + AuthError, DEFAULT_NOUS_INFERENCE_URL, + NOUS_INFERENCE_AUTH_MODE_AUTO, + NOUS_INFERENCE_AUTH_MODE_LEGACY, _load_auth_store, + _auth_store_lock, + _is_terminal_nous_refresh_error, + _quarantine_nous_oauth_state, + _quarantine_nous_pool_entries, _save_auth_store, _write_shared_nous_state, - refresh_nous_oauth_from_state, + resolve_nous_runtime_credentials, ) from hermes_cli.proxy.adapters.base import UpstreamAdapter, UpstreamCredential @@ -43,9 +51,8 @@ class NousPortalAdapter(UpstreamAdapter): """Proxy upstream for the Nous Portal inference API.""" def __init__(self) -> None: - # Lock guards _load → refresh → _save against parallel proxy requests - # racing to refresh expired tokens. Refresh itself is HTTP, so we - # hold the lock across the network call (brief; OAuth refresh is fast). + # Serialize proxy requests in this process; cross-process token refresh + # and persistence are handled by resolve_nous_runtime_credentials(). self._lock = threading.Lock() @property @@ -72,6 +79,26 @@ class NousPortalAdapter(UpstreamAdapter): ) def get_credential(self) -> UpstreamCredential: + return self._get_credential( + inference_auth_mode=NOUS_INFERENCE_AUTH_MODE_AUTO, + ) + + def get_retry_credential( + self, + *, + failed_credential: UpstreamCredential, + status_code: int, + ) -> Optional[UpstreamCredential]: + if status_code != 401: + return None + if failed_credential.bearer.count(".") != 2: + return None + logger.info("proxy: Nous upstream rejected bearer; retrying with legacy session key") + return self._get_credential( + inference_auth_mode=NOUS_INFERENCE_AUTH_MODE_LEGACY, + ) + + def _get_credential(self, *, inference_auth_mode: str) -> UpstreamCredential: with self._lock: state = self._read_state() if state is None: @@ -80,28 +107,43 @@ class NousPortalAdapter(UpstreamAdapter): ) try: - refreshed = refresh_nous_oauth_from_state(state) + refreshed = resolve_nous_runtime_credentials( + inference_auth_mode=inference_auth_mode, + ) + except AuthError as exc: + if _is_terminal_nous_refresh_error(exc): + _quarantine_nous_oauth_state( + state, + exc, + reason="proxy_refresh_failure", + ) + self._save_state( + state, + quarantine_error=exc, + quarantine_reason="proxy_refresh_failure", + ) + raise RuntimeError( + f"Failed to refresh Nous Portal credentials: {exc}" + ) from exc except Exception as exc: raise RuntimeError( f"Failed to refresh Nous Portal credentials: {exc}" ) from exc - self._save_state(refreshed) - - agent_key = refreshed.get("agent_key") + agent_key = refreshed.get("api_key") if not agent_key: raise RuntimeError( "Nous Portal refresh did not return a usable agent_key. " "Try `hermes login nous` to re-authenticate." ) - base_url = refreshed.get("inference_base_url") or DEFAULT_NOUS_INFERENCE_URL + base_url = refreshed.get("base_url") or DEFAULT_NOUS_INFERENCE_URL base_url = base_url.rstrip("/") return UpstreamCredential( bearer=agent_key, base_url=base_url, - expires_at=refreshed.get("agent_key_expires_at"), + expires_at=refreshed.get("expires_at"), ) # ------------------------------------------------------------------ @@ -111,7 +153,8 @@ class NousPortalAdapter(UpstreamAdapter): def _read_state(self) -> Optional[Dict[str, Any]]: try: - store = _load_auth_store() + with _auth_store_lock(): + store = _load_auth_store() except Exception as exc: logger.warning("proxy: failed to load auth store: %s", exc) return None @@ -121,17 +164,28 @@ class NousPortalAdapter(UpstreamAdapter): return None return dict(state) # copy so the refresh helper can mutate freely - def _save_state(self, state: Dict[str, Any]) -> None: + def _save_state( + self, + state: Dict[str, Any], + *, + quarantine_error: Optional[AuthError] = None, + quarantine_reason: Optional[str] = None, + ) -> None: try: - store = _load_auth_store() - providers = store.setdefault("providers", {}) - providers["nous"] = state - _save_auth_store(store) + with _auth_store_lock(): + store = _load_auth_store() + if quarantine_error is not None and quarantine_reason: + _quarantine_nous_pool_entries( + store, + quarantine_error, + reason=quarantine_reason, + ) + providers = store.setdefault("providers", {}) + providers["nous"] = state + _save_auth_store(store) _write_shared_nous_state(state) except Exception as exc: - # Best effort — we still return the fresh credential. The next - # request just won't see cached state, which means another refresh. - logger.warning("proxy: failed to persist refreshed Nous state: %s", exc) + logger.warning("proxy: failed to persist Nous quarantine state: %s", exc) __all__ = ["NousPortalAdapter"] diff --git a/hermes_cli/proxy/cli.py b/hermes_cli/proxy/cli.py index 83c2d34035b..c35b14f7835 100644 --- a/hermes_cli/proxy/cli.py +++ b/hermes_cli/proxy/cli.py @@ -114,7 +114,7 @@ def cmd_proxy(args: Any) -> int: return cmd_proxy_start(args) if sub == "status": return cmd_proxy_status(args) - if sub in ("providers", "list"): + if sub in {"providers", "list"}: return cmd_proxy_list_providers(args) # No subcommand → print short help. print( diff --git a/hermes_cli/proxy/server.py b/hermes_cli/proxy/server.py index 48de784afe4..a72f75d67ee 100644 --- a/hermes_cli/proxy/server.py +++ b/hermes_cli/proxy/server.py @@ -26,7 +26,7 @@ except ImportError: web = None # type: ignore[assignment] AIOHTTP_AVAILABLE = False -from hermes_cli.proxy.adapters.base import UpstreamAdapter +from hermes_cli.proxy.adapters.base import UpstreamAdapter, UpstreamCredential logger = logging.getLogger(__name__) @@ -76,7 +76,7 @@ def _filter_response_headers(headers) -> dict: if key.lower() in _HOP_BY_HOP_HEADERS: continue # aiohttp recomputes Content-Encoding/Content-Length on stream — let it. - if key.lower() in ("content-encoding", "content-length"): + if key.lower() in {"content-encoding", "content-length"}: continue out[key] = value return out @@ -136,50 +136,93 @@ def create_app(adapter: UpstreamAdapter) -> "web.Application": logger.warning("proxy: credential resolution failed: %s", exc) return _json_error(401, str(exc), code="upstream_auth_failed") - upstream_url = f"{cred.base_url.rstrip('/')}{rel_path}" - # Preserve query string verbatim. - if request.query_string: - upstream_url = f"{upstream_url}?{request.query_string}" - # Forward body verbatim. Read into memory once — request bodies for # chat/completions/embeddings are small (<1MB typically). If we ever # need to forward large multipart uploads we'll switch to streaming # the request body too. body = await request.read() - fwd_headers = _filter_request_headers(request.headers) - fwd_headers["Authorization"] = f"{cred.token_type} {cred.bearer}" - - logger.debug( - "proxy: forwarding %s %s -> %s (body=%d bytes)", - request.method, rel_path, upstream_url, len(body), - ) - - # Use a per-request session so connection state doesn't leak between - # clients. Could be optimized to a shared session later. timeout = aiohttp.ClientTimeout(total=None, sock_connect=15, sock_read=300) - try: - session = aiohttp.ClientSession(timeout=timeout) - except Exception as exc: # pragma: no cover - aiohttp setup issue - return _json_error(500, f"proxy session init failed: {exc}") - try: - upstream_resp = await session.request( - request.method, - upstream_url, - data=body if body else None, - headers=fwd_headers, - allow_redirects=False, + async def _send_upstream(active_cred: UpstreamCredential): + upstream_url = f"{active_cred.base_url.rstrip('/')}{rel_path}" + # Preserve query string verbatim. + if request.query_string: + upstream_url = f"{upstream_url}?{request.query_string}" + + fwd_headers = _filter_request_headers(request.headers) + fwd_headers["Authorization"] = f"{active_cred.token_type} {active_cred.bearer}" + + logger.debug( + "proxy: forwarding %s %s -> %s (body=%d bytes)", + request.method, rel_path, upstream_url, len(body), ) - except aiohttp.ClientError as exc: - await session.close() - logger.warning("proxy: upstream connection failed: %s", exc) - return _json_error(502, f"upstream connection failed: {exc}", - code="upstream_unreachable") - except asyncio.TimeoutError: - await session.close() - return _json_error(504, "upstream request timed out", - code="upstream_timeout") + + try: + session = aiohttp.ClientSession(timeout=timeout) + except Exception as exc: # pragma: no cover - aiohttp setup issue + raise RuntimeError(f"proxy session init failed: {exc}") from exc + + try: + upstream_resp = await session.request( + request.method, + upstream_url, + data=body if body else None, + headers=fwd_headers, + allow_redirects=False, + ) + except Exception: + await session.close() + raise + return session, upstream_resp + + async def _open_upstream(active_cred: UpstreamCredential): + try: + return await _send_upstream(active_cred) + except RuntimeError as exc: + return _json_error(500, str(exc)), None + except aiohttp.ClientError as exc: + logger.warning("proxy: upstream connection failed: %s", exc) + return ( + _json_error( + 502, + f"upstream connection failed: {exc}", + code="upstream_unreachable", + ), + None, + ) + except asyncio.TimeoutError: + return ( + _json_error( + 504, + "upstream request timed out", + code="upstream_timeout", + ), + None, + ) + + session_or_response, upstream_resp = await _open_upstream(cred) + if upstream_resp is None: + return session_or_response + session = session_or_response + + if upstream_resp.status == 401: + try: + retry_cred = adapter.get_retry_credential( + failed_credential=cred, + status_code=upstream_resp.status, + ) + except Exception as exc: + logger.warning("proxy: retry credential resolution failed: %s", exc) + retry_cred = None + + if retry_cred is not None: + upstream_resp.release() + await session.close() + session_or_response, upstream_resp = await _open_upstream(retry_cred) + if upstream_resp is None: + return session_or_response + session = session_or_response # Stream response back. Headers first, then chunked body. resp = web.StreamResponse( diff --git a/hermes_cli/runtime_provider.py b/hermes_cli/runtime_provider.py index c0baf14db92..de32131d861 100644 --- a/hermes_cli/runtime_provider.py +++ b/hermes_cli/runtime_provider.py @@ -209,7 +209,7 @@ def _maybe_apply_codex_app_server_runtime( Returns the (possibly-rewritten) api_mode.""" if not model_cfg: return api_mode - if provider not in ("openai", "openai-codex"): + if provider not in {"openai", "openai-codex"}: return api_mode runtime = str(model_cfg.get("openai_runtime") or "").strip().lower() if runtime == "codex_app_server": @@ -875,10 +875,9 @@ def _resolve_explicit_runtime( explicit_base_url or str(state.get("inference_base_url") or auth_mod.DEFAULT_NOUS_INFERENCE_URL).strip().rstrip("/") ) - # Only use agent_key for inference — access_token is an OAuth token for the - # portal API (minting keys, refreshing tokens), not for the inference API. - # Falling back to access_token sends an OAuth bearer token to the inference - # endpoint, which returns 404 because it is not a valid inference credential. + # Only use the agent_key compatibility field for inference. It may be + # either a NAS invoke JWT or a legacy opaque session key; raw OAuth + # access_token fallback is handled by resolve_nous_runtime_credentials(). api_key = explicit_api_key or str(state.get("agent_key") or "").strip() expires_at = state.get("agent_key_expires_at") or state.get("expires_at") if not api_key: @@ -1069,17 +1068,19 @@ def resolve_runtime_provider( getattr(entry, "runtime_api_key", None) or getattr(entry, "access_token", "") ) - # For Nous, the pool entry's runtime_api_key is the agent_key — a - # short-lived inference credential (~30 min TTL). The pool doesn't + # For Nous, the pool entry's runtime_api_key is the agent_key + # compatibility field: either an invoke JWT or legacy opaque key. + # The pool doesn't # refresh it during selection (that would trigger network calls in # non-runtime contexts like `hermes auth list`). If the key is # expired, clear pool_api_key so we fall through to - # resolve_nous_runtime_credentials() which handles refresh + mint. + # resolve_nous_runtime_credentials() which handles refresh + fallback. if provider == "nous" and entry is not None and pool_api_key: min_ttl = max(60, int(os.getenv("HERMES_NOUS_MIN_KEY_TTL_SECONDS", "1800"))) nous_state = { "agent_key": getattr(entry, "agent_key", None), "agent_key_expires_at": getattr(entry, "agent_key_expires_at", None), + "scope": getattr(entry, "scope", None), } if not _agent_key_is_usable(nous_state, min_ttl): logger.debug("Nous pool entry agent_key expired/missing, falling through to runtime resolution") diff --git a/hermes_cli/session_recap.py b/hermes_cli/session_recap.py index d67f737d799..111da117485 100644 --- a/hermes_cli/session_recap.py +++ b/hermes_cli/session_recap.py @@ -171,7 +171,7 @@ def _recent_window( cut = 0 for i in range(len(messages) - 1, -1, -1): msg = messages[i] - if isinstance(msg, Mapping) and msg.get("role") in ("user", "assistant"): + if isinstance(msg, Mapping) and msg.get("role") in {"user", "assistant"}: count += 1 if count >= window: cut = i diff --git a/hermes_cli/status.py b/hermes_cli/status.py index f2164ac8a4d..5629da03fe3 100644 --- a/hermes_cli/status.py +++ b/hermes_cli/status.py @@ -259,6 +259,27 @@ def show_status(args): if minimax_status.get("error") and not minimax_logged_in: print(f" Error: {minimax_status.get('error')}") + # xAI OAuth — separate try/except so an import failure here cannot + # disrupt the already-printed Nous/Codex/Qwen/MiniMax rows above. + try: + from hermes_cli.auth import get_xai_oauth_auth_status + xai_oauth_status = get_xai_oauth_auth_status() or {} + except Exception: + xai_oauth_status = {} + + xai_oauth_logged_in = bool(xai_oauth_status.get("logged_in")) + print( + f" {'xAI OAuth':<12} {check_mark(xai_oauth_logged_in)} " + f"{'logged in' if xai_oauth_logged_in else 'not logged in (run: hermes auth add xai-oauth)'}" + ) + xai_auth_file = xai_oauth_status.get("auth_store") + if xai_auth_file: + print(f" Auth file: {xai_auth_file}") + if xai_oauth_status.get("last_refresh"): + print(f" Refreshed: {_format_iso_timestamp(xai_oauth_status.get('last_refresh'))}") + if xai_oauth_status.get("error") and not xai_oauth_logged_in: + print(f" Error: {xai_oauth_status.get('error')}") + # ========================================================================= # Nous Subscription Features # ========================================================================= diff --git a/hermes_cli/tools_config.py b/hermes_cli/tools_config.py index 074bd04aa64..07974800089 100644 --- a/hermes_cli/tools_config.py +++ b/hermes_cli/tools_config.py @@ -88,12 +88,40 @@ CONFIGURABLE_TOOLSETS = [ # who want it opt in via `hermes tools` → Video Generation, which walks # them through provider + model selection. # -# X search is off by default — gated on xAI credentials (SuperGrok OAuth -# or XAI_API_KEY). Users opt in via `hermes tools` → X (Twitter) Search, -# which walks them through credential setup. The tool's check_fn means -# the schema won't appear to the model even if enabled without credentials. +# X search is off by default for users without xAI credentials, but +# auto-enables when SuperGrok OAuth tokens are stored OR XAI_API_KEY is +# set — mirroring the HASS_TOKEN → homeassistant auto-enable below. The +# `hermes tools` → X (Twitter) Search setup walks users through credential +# setup. The tool's check_fn means the schema still won't appear to the +# model if the credential later goes missing or expires. _DEFAULT_OFF_TOOLSETS = {"moa", "homeassistant", "spotify", "discord", "discord_admin", "video", "video_gen", "x_search"} + +def _xai_credentials_present() -> bool: + """Cheap, side-effect-free check for usable xAI credentials. + + Used to auto-enable the ``x_search`` toolset when the user has either + completed xAI Grok OAuth (SuperGrok subscription) or set + ``XAI_API_KEY``. Does NOT hit the network — only inspects the local + auth store and environment. The tool's runtime ``check_fn`` still + gates schema registration if creds later expire or get revoked. + """ + try: + from hermes_cli.auth import _read_xai_oauth_tokens + + _read_xai_oauth_tokens() + return True + except Exception: + pass + try: + from tools.xai_http import get_env_value as _xai_get_env_value + + if str(_xai_get_env_value("XAI_API_KEY") or "").strip(): + return True + except Exception: + pass + return bool(str(os.environ.get("XAI_API_KEY") or "").strip()) + # Platform-scoped toolsets: only appear in the `hermes tools` checklist for # these platforms, and only resolve/save for these platforms. A toolset # absent from this map is available on every platform (current behaviour). @@ -350,6 +378,17 @@ TOOL_CATEGORIES = { "browser": { "name": "Browser Automation", "icon": "🌐", + # Per-provider rows for Browserbase, Browser Use, and Firecrawl are + # injected at runtime from plugins.browser..provider via + # _plugin_browser_providers() in _visible_providers(). Only + # non-provider UX setup-flow rows remain here: + # - "Nous Subscription (Browser Use cloud)" — managed Browser Use + # billed via Nous subscription (requires_nous_auth + + # override_env_vars). Uses the browser-use plugin as the + # underlying backend but has a distinct setup UX. + # - "Local Browser" — non-cloud option, no CloudBrowserProvider. + # - "Camofox" — anti-detection local Firefox; short-circuits the + # cloud-provider dispatch path via _is_camofox_mode(). "providers": [ { "name": "Nous Subscription (Browser Use cloud)", @@ -370,37 +409,6 @@ TOOL_CATEGORIES = { "browser_provider": "local", "post_setup": "agent_browser", }, - { - "name": "Browserbase", - "badge": "paid", - "tag": "Cloud browser with stealth and proxies", - "env_vars": [ - {"key": "BROWSERBASE_API_KEY", "prompt": "Browserbase API key", "url": "https://browserbase.com"}, - {"key": "BROWSERBASE_PROJECT_ID", "prompt": "Browserbase project ID"}, - ], - "browser_provider": "browserbase", - "post_setup": "agent_browser", - }, - { - "name": "Browser Use", - "badge": "paid", - "tag": "Cloud browser with remote execution", - "env_vars": [ - {"key": "BROWSER_USE_API_KEY", "prompt": "Browser Use API key", "url": "https://browser-use.com"}, - ], - "browser_provider": "browser-use", - "post_setup": "agent_browser", - }, - { - "name": "Firecrawl", - "badge": "paid", - "tag": "Cloud browser with remote execution", - "env_vars": [ - {"key": "FIRECRAWL_API_KEY", "prompt": "Firecrawl API key", "url": "https://firecrawl.dev"}, - ], - "browser_provider": "firecrawl", - "post_setup": "agent_browser", - }, { "name": "Camofox", "badge": "free · local", @@ -1170,6 +1178,23 @@ def _get_platform_tools( if ts_tools and ts_tools.issubset(all_tool_names): enabled_toolsets.add(ts_key) + # Auto-enable ``x_search`` when xAI credentials are configured. + # Unlike ``homeassistant`` (whose ``ha_*`` tools live inside the + # platform composite and thus pass the subset check above), + # ``x_search`` is its own one-tool toolset that the composite does + # NOT include, so the subset loop never picks it up. Inject it + # directly here, mirroring the HASS_TOKEN → ``homeassistant`` rule + # below: once you have working creds, you don't have to also click + # through ``hermes tools`` to flip the toolset on. Only fires when + # the user has not yet saved an explicit toolset list — once they + # do, the saved list is authoritative. + x_search_auto_enabled = ( + _toolset_allowed_for_platform("x_search", platform) + and _xai_credentials_present() + ) + if x_search_auto_enabled: + enabled_toolsets.add("x_search") + default_off = set(_DEFAULT_OFF_TOOLSETS) # Legacy safety: if the platform's own name matches a default-off # toolset (e.g. `homeassistant` platform + `homeassistant` toolset), @@ -1187,6 +1212,11 @@ def _get_platform_tools( # regressed after #14798 made cron honor per-platform tool config. if "homeassistant" in default_off and os.getenv("HASS_TOKEN"): default_off.remove("homeassistant") + # Symmetric carve-out for x_search auto-enable (see the inject + # block above). Without this, the default_off subtraction would + # strip the entry we just added. + if x_search_auto_enabled and "x_search" in default_off: + default_off.remove("x_search") enabled_toolsets -= default_off # Recover non-configurable platform toolsets (e.g. discord, feishu_doc, @@ -1653,6 +1683,61 @@ def _plugin_web_search_providers() -> list[dict]: return rows +# Mirror of _plugin_web_search_providers for cloud browser backends. After +# PR #25214, Browserbase / Browser Use / Firecrawl live as plugins under +# plugins/browser//; this helper is the sole source of provider rows +# for those three in the "Browser Automation" picker. The hardcoded +# ``TOOL_CATEGORIES["browser"]`` entries that drove the category before +# were deleted in the same PR; only non-provider UX setup-flow rows remain +# ("Nous Subscription", "Local Browser", "Camofox") — see the comment block +# in ``TOOL_CATEGORIES["browser"]`` for why each one stays hardcoded. +def _plugin_browser_providers() -> list[dict]: + """Build picker-row dicts from plugin-registered cloud browser providers. + + Each returned dict mirrors the legacy ``TOOL_CATEGORIES["browser"]`` + schema (``name`` / ``badge`` / ``tag`` / ``env_vars`` / + ``browser_provider`` / ``post_setup``) so the picker behaves identically + whether a provider was hardcoded or plugin-registered. + + Populates ``browser_provider`` (the legacy config key written to + ``browser.cloud_provider``) and a ``browser_plugin_name`` marker so + setup / write paths can route through the registry when they want to. + """ + try: + from agent.browser_registry import list_providers as _list_browser_providers + from hermes_cli.plugins import _ensure_plugins_discovered + + _ensure_plugins_discovered() + providers = _list_browser_providers() + except Exception: + return [] + + rows: list[dict] = [] + for provider in providers: + name = getattr(provider, "name", None) + if not name: + continue + try: + schema = provider.get_setup_schema() + except Exception: + continue + if not isinstance(schema, dict): + continue + row = { + "name": schema.get("name", provider.display_name), + "badge": schema.get("badge", ""), + "tag": schema.get("tag", ""), + "env_vars": schema.get("env_vars", []), + "browser_provider": name, + "browser_plugin_name": name, + } + # Pass-through optional fields the schema can opt into. + if schema.get("post_setup"): + row["post_setup"] = schema["post_setup"] + rows.append(row) + return rows + + def _visible_providers(cat: dict, config: dict) -> list[dict]: """Return provider entries visible for the current auth/config state.""" features = get_nous_subscription_features(config) @@ -1682,6 +1767,14 @@ def _visible_providers(cat: dict, config: dict) -> list[dict]: if cat.get("name") == "Web Search & Extract": visible.extend(_plugin_web_search_providers()) + # Inject plugin-registered cloud browser backends. After PR #25214, + # Browserbase / Browser Use / Firecrawl are the plugin-supplied rows; + # the hardcoded "Nous Subscription" / "Local Browser" / "Camofox" rows + # stay because they're non-provider UX setup flows (subscription auth, + # local fallback, and the REST-API anti-detection backend respectively). + if cat.get("name") == "Browser Automation": + visible.extend(_plugin_browser_providers()) + return visible @@ -2590,6 +2683,9 @@ def _reconfigure_provider(provider: dict, config: dict): else: _print_info(" Kept current") + if provider.get("post_setup"): + _run_post_setup(provider["post_setup"]) + # Imagegen backends prompt for model selection on reconfig too. plugin_name = provider.get("image_gen_plugin_name") if plugin_name: diff --git a/hermes_cli/web_server.py b/hermes_cli/web_server.py index 0f3f1e8f4cd..e6a5b0110dc 100644 --- a/hermes_cli/web_server.py +++ b/hermes_cli/web_server.py @@ -2609,7 +2609,11 @@ async def _start_device_code_flow(provider_id: str) -> Dict[str, Any]: so the UI can render the verification page link + user code. """ if provider_id == "nous": - from hermes_cli.auth import _request_device_code, PROVIDER_REGISTRY + from hermes_cli.auth import ( + _nous_device_scope_with_env_override, + _request_nous_device_code_with_scope_fallback, + PROVIDER_REGISTRY, + ) import httpx pconfig = PROVIDER_REGISTRY["nous"] portal_base_url = ( @@ -2618,22 +2622,34 @@ async def _start_device_code_flow(provider_id: str) -> Dict[str, Any]: or pconfig.portal_base_url ).rstrip("/") client_id = pconfig.client_id - scope = pconfig.scope + scope, explicit_scope = _nous_device_scope_with_env_override( + None, + default_scope=pconfig.scope, + ) + def _do_nous_device_request(): - with httpx.Client(timeout=httpx.Timeout(15.0), headers={"Accept": "application/json"}) as client: - return _request_device_code( + with httpx.Client( + timeout=httpx.Timeout(15.0), + headers={"Accept": "application/json"}, + ) as client: + return _request_nous_device_code_with_scope_fallback( client=client, portal_base_url=portal_base_url, client_id=client_id, scope=scope, + allow_legacy_fallback=not explicit_scope, ) - device_data = await asyncio.get_running_loop().run_in_executor(None, _do_nous_device_request) + + device_data, effective_scope = await asyncio.get_running_loop().run_in_executor( + None, _do_nous_device_request + ) sid, sess = _new_oauth_session("nous", "device_code") sess["device_code"] = str(device_data["device_code"]) sess["interval"] = int(device_data["interval"]) sess["expires_at"] = time.time() + int(device_data["expires_in"]) sess["portal_base_url"] = portal_base_url sess["client_id"] = client_id + sess["scope"] = effective_scope threading.Thread( target=_nous_poller, args=(sid,), daemon=True, name=f"oauth-poll-{sid[:6]}" ).start() @@ -2762,7 +2778,11 @@ async def _start_device_code_flow(provider_id: str) -> Dict[str, Any]: def _nous_poller(session_id: str) -> None: """Background poller that drives a Nous device-code flow to completion.""" - from hermes_cli.auth import _poll_for_token, refresh_nous_oauth_from_state + from hermes_cli.auth import ( + NOUS_INFERENCE_AUTH_MODE_FRESH, + _poll_for_token, + refresh_nous_oauth_from_state, + ) from datetime import datetime, timezone import httpx with _oauth_sessions_lock: @@ -2773,6 +2793,7 @@ def _nous_poller(session_id: str) -> None: client_id = sess["client_id"] device_code = sess["device_code"] interval = sess["interval"] + scope = sess.get("scope") expires_in = max(60, int(sess["expires_at"] - time.time())) try: with httpx.Client(timeout=httpx.Timeout(15.0), headers={"Accept": "application/json"}) as client: @@ -2791,7 +2812,7 @@ def _nous_poller(session_id: str) -> None: "portal_base_url": portal_base_url, "inference_base_url": token_data.get("inference_base_url"), "client_id": client_id, - "scope": token_data.get("scope"), + "scope": token_data.get("scope") or scope, "token_type": token_data.get("token_type", "Bearer"), "access_token": token_data["access_token"], "refresh_token": token_data.get("refresh_token"), @@ -2803,8 +2824,11 @@ def _nous_poller(session_id: str) -> None: "expires_in": token_ttl, } full_state = refresh_nous_oauth_from_state( - auth_state, min_key_ttl_seconds=300, timeout_seconds=15.0, - force_refresh=False, force_mint=True, + auth_state, + min_key_ttl_seconds=300, + timeout_seconds=15.0, + force_refresh=False, + inference_auth_mode=NOUS_INFERENCE_AUTH_MODE_FRESH, ) from hermes_cli.auth import persist_nous_credentials persist_nous_credentials(full_state) @@ -5381,4 +5405,7 @@ def start_server( open_browser, ) print(f" Hermes Web UI → http://{host}:{port}") - uvicorn.run(app, host=host, port=port, log_level="warning") + # proxy_headers=False so _ws_client_is_allowed sees the real connection peer + # rather than X-Forwarded-For's rewritten value (which would defeat the + # loopback gate when behind a reverse proxy). + uvicorn.run(app, host=host, port=port, log_level="warning", proxy_headers=False) diff --git a/optional-skills/creative/meme-generation/scripts/generate_meme.py b/optional-skills/creative/meme-generation/scripts/generate_meme.py index 288c3838367..807fee71165 100644 --- a/optional-skills/creative/meme-generation/scripts/generate_meme.py +++ b/optional-skills/creative/meme-generation/scripts/generate_meme.py @@ -358,7 +358,7 @@ def generate_meme(template_id: str, texts: list[str], output_path: str) -> str: img = _overlay_on_image(img, texts, fields) output = Path(output_path) - if output.suffix.lower() in (".jpg", ".jpeg"): + if output.suffix.lower() in {".jpg", ".jpeg"}: img = img.convert("RGB") img.save(str(output), quality=95) return str(output) @@ -378,7 +378,7 @@ def generate_from_image( result = _overlay_on_image(img, texts, fields) output = Path(output_path) - if output.suffix.lower() in (".jpg", ".jpeg"): + if output.suffix.lower() in {".jpg", ".jpeg"}: result = result.convert("RGB") result.save(str(output), quality=95) return str(output) diff --git a/optional-skills/devops/watchers/scripts/watch_rss.py b/optional-skills/devops/watchers/scripts/watch_rss.py index cc729f91b13..6e09630404f 100755 --- a/optional-skills/devops/watchers/scripts/watch_rss.py +++ b/optional-skills/devops/watchers/scripts/watch_rss.py @@ -43,7 +43,7 @@ def _parse_feed(xml_bytes: bytes): entries = [] for item in root.iter(): tag = _strip_ns(item.tag) - if tag not in ("item", "entry"): + if tag not in {"item", "entry"}: continue # ElementTree Elements without children are *falsy* — use `is not None`. children = {_strip_ns(c.tag): c for c in item} diff --git a/optional-skills/finance/stocks/scripts/stocks_client.py b/optional-skills/finance/stocks/scripts/stocks_client.py index 7b98fd9dc66..c0bf97dce4a 100755 --- a/optional-skills/finance/stocks/scripts/stocks_client.py +++ b/optional-skills/finance/stocks/scripts/stocks_client.py @@ -125,7 +125,7 @@ def fetch_url(url: str, headers: dict | None = None, retries: int = MAX_RETRIES) return json.loads(raw.decode("utf-8", errors="replace")) except urllib.error.HTTPError as e: last_err = e - if e.code in (404, 400): + if e.code in {404, 400}: break # no point retrying wait = BACKOFF_BASE ** attempt time.sleep(wait) diff --git a/optional-skills/health/fitness-nutrition/scripts/body_calc.py b/optional-skills/health/fitness-nutrition/scripts/body_calc.py index 2d07129cecc..2ce65fd336e 100644 --- a/optional-skills/health/fitness-nutrition/scripts/body_calc.py +++ b/optional-skills/health/fitness-nutrition/scripts/body_calc.py @@ -95,11 +95,11 @@ def one_rep_max(weight, reps): def macros(tdee_kcal, goal): goal = goal.lower() - if goal in ("cut", "lose", "deficit"): + if goal in {"cut", "lose", "deficit"}: cals = tdee_kcal - 500 p, f, c = 0.40, 0.30, 0.30 label = "Fat Loss (-500 kcal)" - elif goal in ("bulk", "gain", "surplus"): + elif goal in {"bulk", "gain", "surplus"}: cals = tdee_kcal + 400 p, f, c = 0.30, 0.25, 0.45 label = "Lean Bulk (+400 kcal)" @@ -184,7 +184,7 @@ def main(): int(sys.argv[4]), sys.argv[5], int(sys.argv[6]), ) - elif cmd in ("1rm", "orm"): + elif cmd in {"1rm", "orm"}: one_rep_max(float(sys.argv[2]), int(sys.argv[3])) elif cmd == "macros": diff --git a/optional-skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py b/optional-skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py index 6ebb1d75400..d9d53a97a24 100644 --- a/optional-skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py +++ b/optional-skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py @@ -610,7 +610,7 @@ def _is_secret_key(key: str) -> bool: normalized = _normalize_secret_key(key) if normalized == "token" or normalized.endswith("token"): return True - if normalized in ("auth", "authorization"): + if normalized in {"auth", "authorization"}: return True return any(marker in normalized for marker in _SECRET_KEY_MARKERS) @@ -831,7 +831,7 @@ class Migrator: # Flip the config-block flag when a conflict/error occurs on a # config.yaml write. Later config-mutating options will skip rather # than attempting a partial write. - if status in (STATUS_CONFLICT, STATUS_ERROR) and destination is not None: + if status in {STATUS_CONFLICT, STATUS_ERROR} and destination is not None: dest_str = str(destination) if dest_str.endswith("config.yaml") or dest_str.endswith("config.yml"): self._config_apply_blocked = True @@ -1526,7 +1526,7 @@ class Migrator: api_key = resolve_secret_input(raw_key, openclaw_env) if not api_key: # Warn if a SecretRef with file/exec source was silently unresolvable - if isinstance(raw_key, dict) and raw_key.get("source") in ("file", "exec"): + if isinstance(raw_key, dict) and raw_key.get("source") in {"file", "exec"}: self.record( "provider-keys", self.source_root / "openclaw.json", @@ -1736,7 +1736,7 @@ class Migrator: tts_data: Dict[str, Any] = {} provider = tts.get("provider") - if isinstance(provider, str) and provider in ("elevenlabs", "openai", "edge", "microsoft"): + if isinstance(provider, str) and provider in {"elevenlabs", "openai", "edge", "microsoft"}: # OpenClaw renamed "edge" to "microsoft"; Hermes still uses "edge" tts_data["provider"] = "edge" if provider == "microsoft" else provider @@ -2304,11 +2304,11 @@ class Migrator: if defaults.get("thinkingDefault"): # Map OpenClaw thinking -> Hermes reasoning_effort thinking = defaults["thinkingDefault"] - if thinking in ("always", "high", "xhigh"): + if thinking in {"always", "high", "xhigh"}: agent_cfg["reasoning_effort"] = "high" - elif thinking in ("auto", "medium", "adaptive"): + elif thinking in {"auto", "medium", "adaptive"}: agent_cfg["reasoning_effort"] = "medium" - elif thinking in ("off", "low", "none", "minimal"): + elif thinking in {"off", "low", "none", "minimal"}: agent_cfg["reasoning_effort"] = "low" changes = True @@ -2626,8 +2626,8 @@ class Migrator: if not isinstance(ch_cfg, dict): continue complex_keys = {k: v for k, v in ch_cfg.items() - if k not in ("botToken", "appToken", "allowFrom", "enabled") - and v and k not in ("requireMention", "autoThread")} + if k not in {"botToken", "appToken", "allowFrom", "enabled"} + and v and k not in {"requireMention", "autoThread"}} if complex_keys: complex_archive[ch_name] = complex_keys @@ -2671,7 +2671,7 @@ class Migrator: # Archive remaining browser settings advanced = {k: v for k, v in browser.items() - if k not in ("cdpUrl", "headless") and v} + if k not in {"cdpUrl", "headless"} and v} if advanced and self.archive_dir: if self.execute: self.archive_dir.mkdir(parents=True, exist_ok=True) diff --git a/optional-skills/productivity/telephony/scripts/telephony.py b/optional-skills/productivity/telephony/scripts/telephony.py index c9233647f3f..188b6be2ad9 100644 --- a/optional-skills/productivity/telephony/scripts/telephony.py +++ b/optional-skills/productivity/telephony/scripts/telephony.py @@ -109,7 +109,7 @@ def _config_lookup(*paths: tuple[str, ...], default: str = "") -> str: node = None break node = node.get(key) - if node not in (None, "") and not isinstance(node, dict): + if node not in {None, ""} and not isinstance(node, dict): return str(node) return default diff --git a/optional-skills/research/darwinian-evolver/scripts/show_snapshot.py b/optional-skills/research/darwinian-evolver/scripts/show_snapshot.py index 10e3a03dca9..5dd559570dd 100644 --- a/optional-skills/research/darwinian-evolver/scripts/show_snapshot.py +++ b/optional-skills/research/darwinian-evolver/scripts/show_snapshot.py @@ -51,7 +51,7 @@ def main() -> int: field = args.field if field is None: for k, v in vars(org).items(): - if isinstance(v, str) and not k.startswith("_") and k not in ("id",): + if isinstance(v, str) and not k.startswith("_") and k not in {"id",}: field = k break val = getattr(org, field, None) if field else None diff --git a/optional-skills/research/domain-intel/scripts/domain_intel.py b/optional-skills/research/domain-intel/scripts/domain_intel.py index 1a69f6528f2..c25e9286d40 100644 --- a/optional-skills/research/domain-intel/scripts/domain_intel.py +++ b/optional-skills/research/domain-intel/scripts/domain_intel.py @@ -185,7 +185,7 @@ def whois_lookup(domain): for key, pat in patterns.items(): matches = re.findall(pat, raw, re.IGNORECASE) if matches: - if key in ("name_servers", "status"): + if key in {"name_servers", "status"}: result[key] = list(dict.fromkeys(m.strip().lower() for m in matches)) else: result[key] = matches[0].strip() diff --git a/optional-skills/research/osint-investigation/scripts/_http.py b/optional-skills/research/osint-investigation/scripts/_http.py index 5da62310b9f..0936548a92a 100644 --- a/optional-skills/research/osint-investigation/scripts/_http.py +++ b/optional-skills/research/osint-investigation/scripts/_http.py @@ -60,7 +60,7 @@ def get( f"HTTP 429 rate-limited by {urllib.parse.urlsplit(url).netloc}. " f"Slow down or supply a real API key. Body: {body[:300]}" ) from e - if e.code in (500, 502, 503, 504) and attempt < max_retries: + if e.code in {500, 502, 503, 504} and attempt < max_retries: retry_after = e.headers.get("Retry-After") if e.headers else None wait = float(retry_after) if (retry_after and retry_after.isdigit()) else backoff ** (attempt + 1) time.sleep(wait) diff --git a/optional-skills/research/osint-investigation/scripts/fetch_icij_offshore.py b/optional-skills/research/osint-investigation/scripts/fetch_icij_offshore.py index 8d050b62bf1..3108681e20c 100644 --- a/optional-skills/research/osint-investigation/scripts/fetch_icij_offshore.py +++ b/optional-skills/research/osint-investigation/scripts/fetch_icij_offshore.py @@ -122,7 +122,7 @@ def fetch( with zipfile.ZipFile(zip_path) as zf: for node_type, csv_substring in targets: - relevant_needles = [n for (k, n) in needles if k in (node_type, "Entity", "Officer")] or [] + relevant_needles = [n for (k, n) in needles if k in {node_type, "Entity", "Officer"}] or [] # Only scan a CSV if we have a needle that could plausibly match it, # or if we have ONLY a jurisdiction filter. applicable_needles = [n for (k, n) in needles if k == node_type] diff --git a/plugins/browser/browser_use/__init__.py b/plugins/browser/browser_use/__init__.py new file mode 100644 index 00000000000..b07db13913a --- /dev/null +++ b/plugins/browser/browser_use/__init__.py @@ -0,0 +1,14 @@ +"""Browser Use cloud browser plugin — bundled, auto-loaded. + +Mirrors the ``plugins/web//`` layout: ``provider.py`` holds the +provider class; ``__init__.py::register`` instantiates and registers it. +""" + +from __future__ import annotations + +from plugins.browser.browser_use.provider import BrowserUseBrowserProvider + + +def register(ctx) -> None: + """Register the Browser Use provider with the plugin context.""" + ctx.register_browser_provider(BrowserUseBrowserProvider()) diff --git a/plugins/browser/browser_use/plugin.yaml b/plugins/browser/browser_use/plugin.yaml new file mode 100644 index 00000000000..ff926a50ea7 --- /dev/null +++ b/plugins/browser/browser_use/plugin.yaml @@ -0,0 +1,7 @@ +name: browser-browser-use +version: 1.0.0 +description: "Browser Use (https://browser-use.com) cloud browser backend. Supports both direct BROWSER_USE_API_KEY and the managed Nous tool gateway. Also powers the 'Nous Subscription' UX flow that bills usage to a Nous subscription." +author: NousResearch +kind: backend +provides_browser_providers: + - browser-use diff --git a/tools/browser_providers/browser_use.py b/plugins/browser/browser_use/provider.py similarity index 63% rename from tools/browser_providers/browser_use.py rename to plugins/browser/browser_use/provider.py index a1f4f425ba0..3d371bdd88a 100644 --- a/tools/browser_providers/browser_use.py +++ b/plugins/browser/browser_use/provider.py @@ -1,4 +1,32 @@ -"""Browser Use cloud browser provider.""" +"""Browser Use cloud browser provider — plugin form. + +Subclasses :class:`agent.browser_provider.BrowserProvider` (the plugin-facing +ABC introduced in PR #25214). The legacy in-tree module +``tools.browser_providers.browser_use`` was removed in the same PR; this file +is now the canonical implementation. + +Browser Use is the only browser backend with dual auth: a direct +``BROWSER_USE_API_KEY`` for self-billed users, or the managed Nous tool +gateway (which Hermes uses to bill Browser Use sessions to a Nous +subscription). The dispatch order — direct API key first, managed gateway +second — preserves the pre-migration behaviour in +``tools.browser_providers.browser_use.BrowserUseProvider._get_config_or_none``. + +Config keys this provider responds to:: + + browser: + cloud_provider: "browser-use" # explicit selection + tool_gateway: + browser: "gateway" # optional: prefer managed gateway + # even when BROWSER_USE_API_KEY is set + +Auth env vars (one of):: + + BROWSER_USE_API_KEY=... # https://browser-use.com + # OR a managed Nous gateway entry (configured via 'hermes setup') +""" + +from __future__ import annotations import logging import os @@ -8,11 +36,14 @@ from typing import Any, Dict, Optional import requests -from tools.browser_providers.base import CloudBrowserProvider -from tools.managed_tool_gateway import resolve_managed_tool_gateway -from tools.tool_backend_helpers import managed_nous_tools_enabled, prefers_gateway +from agent.browser_provider import BrowserProvider logger = logging.getLogger(__name__) + +# Idempotency tracking for managed-mode session creation. The managed Nous +# gateway returns 409 "already in progress" on retried POSTs; we forward the +# original idempotency key so the gateway can deduplicate. Cleared on +# success or terminal failure. _pending_create_keys: Dict[str, str] = {} _pending_create_keys_lock = threading.Lock() @@ -38,6 +69,16 @@ def _clear_pending_create_key(task_id: str) -> None: def _should_preserve_pending_create_key(response: requests.Response) -> bool: + """Decide whether to keep the idempotency key after a failed create. + + Preserve the key when the failure looks retryable (5xx) OR when the + gateway reports the original request is still in flight (409 "already + in progress") — in either case, retrying with the same key lets the + gateway deduplicate. + + Drop the key on any other 4xx (auth failure, bad request, etc.) — those + won't succeed by being retried. + """ if response.status_code >= 500: return True @@ -60,13 +101,24 @@ def _should_preserve_pending_create_key(response: requests.Response) -> bool: return "already in progress" in message -class BrowserUseProvider(CloudBrowserProvider): - """Browser Use (https://browser-use.com) cloud browser backend.""" +class BrowserUseBrowserProvider(BrowserProvider): + """Browser Use (https://browser-use.com) cloud browser backend. - def provider_name(self) -> str: + Dual auth: prefers a direct BROWSER_USE_API_KEY when set, falling back + to the managed Nous tool gateway when ``tool_gateway.browser`` config + routes through it. Setting ``tool_gateway.browser: gateway`` flips the + order so managed billing wins even when BROWSER_USE_API_KEY is present. + """ + + @property + def name(self) -> str: + return "browser-use" + + @property + def display_name(self) -> str: return "Browser Use" - def is_configured(self) -> bool: + def is_available(self) -> bool: return self._get_config_or_none() is not None # ------------------------------------------------------------------ @@ -74,6 +126,14 @@ class BrowserUseProvider(CloudBrowserProvider): # ------------------------------------------------------------------ def _get_config_or_none(self) -> Optional[Dict[str, Any]]: + # Import here to avoid a hard dependency at module-import time — + # managed_tool_gateway pulls in the Nous auth stack which can be + # heavy and is not needed for direct-API-key users. + from tools.managed_tool_gateway import resolve_managed_tool_gateway + from tools.tool_backend_helpers import prefers_gateway + + # Direct API key wins unless the user has explicitly opted into the + # managed Nous gateway via ``tool_gateway.browser: gateway``. api_key = os.environ.get("BROWSER_USE_API_KEY") if api_key and not prefers_gateway("browser"): return { @@ -93,6 +153,8 @@ class BrowserUseProvider(CloudBrowserProvider): } def _get_config(self) -> Dict[str, Any]: + from tools.tool_backend_helpers import managed_nous_tools_enabled + config = self._get_config_or_none() if config is None: message = ( @@ -111,11 +173,10 @@ class BrowserUseProvider(CloudBrowserProvider): # ------------------------------------------------------------------ def _headers(self, config: Dict[str, Any]) -> Dict[str, str]: - headers = { + return { "Content-Type": "application/json", "X-Browser-Use-API-Key": config["api_key"], } - return headers def create_session(self, task_id: str) -> Dict[str, object]: config = self._get_config() @@ -166,7 +227,9 @@ class BrowserUseProvider(CloudBrowserProvider): if managed_mode: _clear_pending_create_key(task_id) session_name = f"hermes_{task_id}_{uuid.uuid4().hex[:8]}" - external_call_id = response.headers.get("x-external-call-id") if managed_mode else None + external_call_id = ( + response.headers.get("x-external-call-id") if managed_mode else None + ) logger.info("Created Browser Use session %s", session_name) @@ -184,7 +247,9 @@ class BrowserUseProvider(CloudBrowserProvider): try: config = self._get_config() except ValueError: - logger.warning("Cannot close Browser Use session %s — missing credentials", session_id) + logger.warning( + "Cannot close Browser Use session %s — missing credentials", session_id + ) return False try: @@ -212,7 +277,10 @@ class BrowserUseProvider(CloudBrowserProvider): def emergency_cleanup(self, session_id: str) -> None: config = self._get_config_or_none() if config is None: - logger.warning("Cannot emergency-cleanup Browser Use session %s — missing credentials", session_id) + logger.warning( + "Cannot emergency-cleanup Browser Use session %s — missing credentials", + session_id, + ) return try: requests.patch( @@ -222,4 +290,21 @@ class BrowserUseProvider(CloudBrowserProvider): timeout=5, ) except Exception as e: - logger.debug("Emergency cleanup failed for Browser Use session %s: %s", session_id, e) + logger.debug( + "Emergency cleanup failed for Browser Use session %s: %s", session_id, e + ) + + def get_setup_schema(self) -> Dict[str, Any]: + return { + "name": "Browser Use", + "badge": "paid", + "tag": "Cloud browser with remote execution", + "env_vars": [ + { + "key": "BROWSER_USE_API_KEY", + "prompt": "Browser Use API key", + "url": "https://browser-use.com", + }, + ], + "post_setup": "agent_browser", + } diff --git a/plugins/browser/browserbase/__init__.py b/plugins/browser/browserbase/__init__.py new file mode 100644 index 00000000000..1e0269e2733 --- /dev/null +++ b/plugins/browser/browserbase/__init__.py @@ -0,0 +1,15 @@ +"""Browserbase cloud browser plugin — bundled, auto-loaded. + +Mirrors the ``plugins/web//`` and ``plugins/image_gen/openai/`` +layout: ``provider.py`` holds the provider class; ``__init__.py::register`` +instantiates and registers it via the plugin context. +""" + +from __future__ import annotations + +from plugins.browser.browserbase.provider import BrowserbaseBrowserProvider + + +def register(ctx) -> None: + """Register the Browserbase provider with the plugin context.""" + ctx.register_browser_provider(BrowserbaseBrowserProvider()) diff --git a/plugins/browser/browserbase/plugin.yaml b/plugins/browser/browserbase/plugin.yaml new file mode 100644 index 00000000000..5d976328a23 --- /dev/null +++ b/plugins/browser/browserbase/plugin.yaml @@ -0,0 +1,7 @@ +name: browser-browserbase +version: 1.0.0 +description: "Browserbase (https://browserbase.com) cloud browser backend. Requires BROWSERBASE_API_KEY + BROWSERBASE_PROJECT_ID. Supports stealth, proxies, and keep-alive sessions; auto-falls-back when paid features are unavailable." +author: NousResearch +kind: backend +provides_browser_providers: + - browserbase diff --git a/tools/browser_providers/browserbase.py b/plugins/browser/browserbase/provider.py similarity index 67% rename from tools/browser_providers/browserbase.py rename to plugins/browser/browserbase/provider.py index 4807345214b..2b05d01d03b 100644 --- a/tools/browser_providers/browserbase.py +++ b/plugins/browser/browserbase/provider.py @@ -1,4 +1,35 @@ -"""Browserbase cloud browser provider (direct credentials only).""" +"""Browserbase cloud browser provider — plugin form. + +Subclasses :class:`agent.browser_provider.BrowserProvider` (the plugin-facing +ABC introduced in PR #25214). The legacy in-tree module +``tools.browser_providers.browserbase`` was removed in the same PR; this file +is now the canonical implementation. + +Browserbase requires direct ``BROWSERBASE_API_KEY`` and ``BROWSERBASE_PROJECT_ID`` +credentials. Managed Nous gateway support has been removed — the Nous +subscription now routes through Browser Use instead (see +``plugins/browser/browser_use/``). + +Config keys this provider responds to:: + + browser: + cloud_provider: "browserbase" + +Auth env vars:: + + BROWSERBASE_API_KEY=... # https://browserbase.com + BROWSERBASE_PROJECT_ID=... + +Optional feature knobs:: + + BROWSERBASE_BASE_URL=... # default https://api.browserbase.com + BROWSERBASE_PROXIES=true # default true + BROWSERBASE_ADVANCED_STEALTH=false + BROWSERBASE_KEEP_ALIVE=true # default true + BROWSERBASE_SESSION_TIMEOUT=... (ms, integer) +""" + +from __future__ import annotations import logging import os @@ -7,27 +38,31 @@ from typing import Any, Dict, Optional import requests -from tools.browser_providers.base import CloudBrowserProvider +from agent.browser_provider import BrowserProvider logger = logging.getLogger(__name__) -class BrowserbaseProvider(CloudBrowserProvider): +class BrowserbaseBrowserProvider(BrowserProvider): """Browserbase (https://browserbase.com) cloud browser backend. - This provider requires direct BROWSERBASE_API_KEY and BROWSERBASE_PROJECT_ID - credentials. Managed Nous gateway support has been removed — the Nous - subscription now routes through Browser Use instead. + Direct credentials only — managed-Nous-gateway support lives on the + Browser Use provider now. """ - def provider_name(self) -> str: + @property + def name(self) -> str: + return "browserbase" + + @property + def display_name(self) -> str: return "Browserbase" - def is_configured(self) -> bool: + def is_available(self) -> bool: return self._get_config_or_none() is not None # ------------------------------------------------------------------ - # Session lifecycle + # Config resolution # ------------------------------------------------------------------ def _get_config_or_none(self) -> Optional[Dict[str, Any]]: @@ -37,7 +72,9 @@ class BrowserbaseProvider(CloudBrowserProvider): return { "api_key": api_key, "project_id": project_id, - "base_url": os.environ.get("BROWSERBASE_BASE_URL", "https://api.browserbase.com").rstrip("/"), + "base_url": os.environ.get( + "BROWSERBASE_BASE_URL", "https://api.browserbase.com" + ).rstrip("/"), } return None @@ -50,13 +87,21 @@ class BrowserbaseProvider(CloudBrowserProvider): ) return config + # ------------------------------------------------------------------ + # Session lifecycle + # ------------------------------------------------------------------ + def create_session(self, task_id: str) -> Dict[str, object]: config = self._get_config() # Optional env-var knobs enable_proxies = os.environ.get("BROWSERBASE_PROXIES", "true").lower() != "false" - enable_advanced_stealth = os.environ.get("BROWSERBASE_ADVANCED_STEALTH", "false").lower() == "true" - enable_keep_alive = os.environ.get("BROWSERBASE_KEEP_ALIVE", "true").lower() != "false" + enable_advanced_stealth = ( + os.environ.get("BROWSERBASE_ADVANCED_STEALTH", "false").lower() == "true" + ) + enable_keep_alive = ( + os.environ.get("BROWSERBASE_KEEP_ALIVE", "true").lower() != "false" + ) custom_timeout_ms = os.environ.get("BROWSERBASE_SESSION_TIMEOUT") features_enabled = { @@ -78,7 +123,9 @@ class BrowserbaseProvider(CloudBrowserProvider): if timeout_val > 0: session_config["timeout"] = timeout_val except ValueError: - logger.warning("Invalid BROWSERBASE_SESSION_TIMEOUT value: %s", custom_timeout_ms) + logger.warning( + "Invalid BROWSERBASE_SESSION_TIMEOUT value: %s", custom_timeout_ms + ) if enable_proxies: session_config["proxies"] = True @@ -156,7 +203,9 @@ class BrowserbaseProvider(CloudBrowserProvider): features_enabled["custom_timeout"] = True feature_str = ", ".join(k for k, v in features_enabled.items() if v) - logger.info("Created Browserbase session %s with features: %s", session_name, feature_str) + logger.info( + "Created Browserbase session %s with features: %s", session_name, feature_str + ) return { "session_name": session_name, @@ -169,7 +218,9 @@ class BrowserbaseProvider(CloudBrowserProvider): try: config = self._get_config() except ValueError: - logger.warning("Cannot close Browserbase session %s — missing credentials", session_id) + logger.warning( + "Cannot close Browserbase session %s — missing credentials", session_id + ) return False try: @@ -203,7 +254,10 @@ class BrowserbaseProvider(CloudBrowserProvider): def emergency_cleanup(self, session_id: str) -> None: config = self._get_config_or_none() if config is None: - logger.warning("Cannot emergency-cleanup Browserbase session %s — missing credentials", session_id) + logger.warning( + "Cannot emergency-cleanup Browserbase session %s — missing credentials", + session_id, + ) return try: requests.post( @@ -219,4 +273,25 @@ class BrowserbaseProvider(CloudBrowserProvider): timeout=5, ) except Exception as e: - logger.debug("Emergency cleanup failed for Browserbase session %s: %s", session_id, e) + logger.debug( + "Emergency cleanup failed for Browserbase session %s: %s", session_id, e + ) + + def get_setup_schema(self) -> Dict[str, Any]: + return { + "name": "Browserbase", + "badge": "paid", + "tag": "Cloud browser with stealth and proxies", + "env_vars": [ + { + "key": "BROWSERBASE_API_KEY", + "prompt": "Browserbase API key", + "url": "https://browserbase.com", + }, + { + "key": "BROWSERBASE_PROJECT_ID", + "prompt": "Browserbase project ID", + }, + ], + "post_setup": "agent_browser", + } diff --git a/plugins/browser/firecrawl/__init__.py b/plugins/browser/firecrawl/__init__.py new file mode 100644 index 00000000000..b045b636302 --- /dev/null +++ b/plugins/browser/firecrawl/__init__.py @@ -0,0 +1,16 @@ +"""Firecrawl cloud browser plugin — bundled, auto-loaded. + +Distinct from ``plugins/web/firecrawl/`` (the web search/extract/crawl +plugin); both share the FIRECRAWL_API_KEY but speak to different endpoints +(``/v2/browser`` here vs ``/v2/search`` / ``/v2/scrape`` / ``/v2/crawl`` +over there). +""" + +from __future__ import annotations + +from plugins.browser.firecrawl.provider import FirecrawlBrowserProvider + + +def register(ctx) -> None: + """Register the Firecrawl cloud-browser provider with the plugin context.""" + ctx.register_browser_provider(FirecrawlBrowserProvider()) diff --git a/plugins/browser/firecrawl/plugin.yaml b/plugins/browser/firecrawl/plugin.yaml new file mode 100644 index 00000000000..22da6a7f4b5 --- /dev/null +++ b/plugins/browser/firecrawl/plugin.yaml @@ -0,0 +1,7 @@ +name: browser-firecrawl +version: 1.0.0 +description: "Firecrawl (https://firecrawl.dev) cloud browser backend. Requires FIRECRAWL_API_KEY. Distinct from the firecrawl WEB search/extract plugin — the two share an API key but operate on different endpoints." +author: NousResearch +kind: backend +provides_browser_providers: + - firecrawl diff --git a/tools/browser_providers/firecrawl.py b/plugins/browser/firecrawl/provider.py similarity index 57% rename from tools/browser_providers/firecrawl.py rename to plugins/browser/firecrawl/provider.py index 4a8ae82a2d2..2c605134a01 100644 --- a/tools/browser_providers/firecrawl.py +++ b/plugins/browser/firecrawl/provider.py @@ -1,26 +1,61 @@ -"""Firecrawl cloud browser provider.""" +"""Firecrawl cloud browser provider — plugin form. + +Subclasses :class:`agent.browser_provider.BrowserProvider` (the plugin-facing +ABC introduced in PR #25214). The legacy in-tree module +``tools.browser_providers.firecrawl`` was removed in the same PR; this file +is now the canonical implementation. + +This is the cloud-browser path — distinct from the firecrawl WEB plugin at +``plugins/web/firecrawl/`` which handles search/extract/crawl on +``/v2/search`` / ``/v2/scrape`` / ``/v2/crawl``. The two plugins share the +``FIRECRAWL_API_KEY`` env var but talk to different endpoints (this one +hits ``/v2/browser``). + +Config keys this provider responds to:: + + browser: + cloud_provider: "firecrawl" # explicit selection only — not in the + # legacy auto-detect walk + +Auth env vars:: + + FIRECRAWL_API_KEY=... # https://firecrawl.dev + FIRECRAWL_API_URL=... # optional override (default https://api.firecrawl.dev) + FIRECRAWL_BROWSER_TTL=... # optional, default 300 seconds +""" + +from __future__ import annotations import logging import os import uuid -from typing import Dict +from typing import Any, Dict import requests -from tools.browser_providers.base import CloudBrowserProvider +from agent.browser_provider import BrowserProvider logger = logging.getLogger(__name__) _BASE_URL = "https://api.firecrawl.dev" -class FirecrawlProvider(CloudBrowserProvider): - """Firecrawl (https://firecrawl.dev) cloud browser backend.""" +class FirecrawlBrowserProvider(BrowserProvider): + """Firecrawl (https://firecrawl.dev) cloud browser backend. - def provider_name(self) -> str: + Cloud-browser path only — search/extract/crawl live in the separate + ``plugins/web/firecrawl/`` plugin. + """ + + @property + def name(self) -> str: + return "firecrawl" + + @property + def display_name(self) -> str: return "Firecrawl" - def is_configured(self) -> bool: + def is_available(self) -> bool: return bool(os.environ.get("FIRECRAWL_API_KEY")) # ------------------------------------------------------------------ @@ -100,13 +135,34 @@ class FirecrawlProvider(CloudBrowserProvider): return False def emergency_cleanup(self, session_id: str) -> None: + if not self.is_available(): + logger.warning( + "Cannot emergency-cleanup Firecrawl session %s — missing credentials", + session_id, + ) + return try: requests.delete( f"{self._api_url()}/v2/browser/{session_id}", headers=self._headers(), timeout=5, ) - except ValueError: - logger.warning("Cannot emergency-cleanup Firecrawl session %s — missing credentials", session_id) except Exception as e: - logger.debug("Emergency cleanup failed for Firecrawl session %s: %s", session_id, e) + logger.debug( + "Emergency cleanup failed for Firecrawl session %s: %s", session_id, e + ) + + def get_setup_schema(self) -> Dict[str, Any]: + return { + "name": "Firecrawl", + "badge": "paid", + "tag": "Cloud browser with remote execution", + "env_vars": [ + { + "key": "FIRECRAWL_API_KEY", + "prompt": "Firecrawl API key", + "url": "https://firecrawl.dev", + }, + ], + "post_setup": "agent_browser", + } diff --git a/plugins/disk-cleanup/__init__.py b/plugins/disk-cleanup/__init__.py index 0a4b6c7ae16..71d44b1c891 100644 --- a/plugins/disk-cleanup/__init__.py +++ b/plugins/disk-cleanup/__init__.py @@ -222,7 +222,7 @@ def _fmt_summary(summary: Dict[str, Any]) -> str: def _handle_slash(raw_args: str) -> Optional[str]: argv = raw_args.strip().split() - if not argv or argv[0] in ("help", "-h", "--help"): + if not argv or argv[0] in {"help", "-h", "--help"}: return _HELP_TEXT sub = argv[0] diff --git a/plugins/google_meet/__init__.py b/plugins/google_meet/__init__.py index feca75667b5..df401e1a680 100644 --- a/plugins/google_meet/__init__.py +++ b/plugins/google_meet/__init__.py @@ -72,7 +72,7 @@ def register(ctx) -> None: # tested path there and guest-join Chromium is flakier. Refuse to register # rather than half-working. system = platform.system().lower() - if system not in ("linux", "darwin"): + if system not in {"linux", "darwin"}: logger.info( "google_meet plugin: platform=%s not supported (linux/macos only)", system, diff --git a/plugins/google_meet/cli.py b/plugins/google_meet/cli.py index b7d8097fc76..0e9b08881b3 100644 --- a/plugins/google_meet/cli.py +++ b/plugins/google_meet/cli.py @@ -159,7 +159,7 @@ def _cmd_setup() -> int: print("---------------------") system = _p.system() - system_ok = system in ("Linux", "Darwin") + system_ok = system in {"Linux", "Darwin"} print(f" platform : {system} [{'ok' if system_ok else 'unsupported'}]") try: @@ -231,7 +231,7 @@ def _cmd_install(*, realtime: bool, assume_yes: bool) -> int: import subprocess as _sp system = _p.system() - if system not in ("Linux", "Darwin"): + if system not in {"Linux", "Darwin"}: print(f"google_meet install: {system} is not supported (linux/macos only)") return 1 @@ -242,7 +242,7 @@ def _cmd_install(*, realtime: bool, assume_yes: bool) -> int: ans = input(f"{prompt} [y/N] ").strip().lower() except EOFError: return False - return ans in ("y", "yes") + return ans in {"y", "yes"} print("google_meet install") print("-------------------") diff --git a/plugins/google_meet/meet_bot.py b/plugins/google_meet/meet_bot.py index eb9318ae4a5..9040d9a789a 100644 --- a/plugins/google_meet/meet_bot.py +++ b/plugins/google_meet/meet_bot.py @@ -447,7 +447,7 @@ def _mac_audio_device_index(device_name: str) -> str: def run_bot() -> int: # noqa: C901 — orchestration, explicit branches url = os.environ.get("HERMES_MEET_URL", "").strip() out_dir_env = os.environ.get("HERMES_MEET_OUT_DIR", "").strip() - headed = os.environ.get("HERMES_MEET_HEADED", "").lower() in ("1", "true", "yes") + headed = os.environ.get("HERMES_MEET_HEADED", "").lower() in {"1", "true", "yes"} auth_state = os.environ.get("HERMES_MEET_AUTH_STATE", "").strip() guest_name = os.environ.get("HERMES_MEET_GUEST_NAME", "Hermes Agent") duration_s = _parse_duration(os.environ.get("HERMES_MEET_DURATION", "")) @@ -808,7 +808,7 @@ def _looks_like_human_speaker(speaker: str, bot_guest_name: str) -> bool: if not speaker or not speaker.strip(): return False spk = speaker.strip().lower() - if spk in ("unknown", "you", bot_guest_name.strip().lower()): + if spk in {"unknown", "you", bot_guest_name.strip().lower()}: return False return True diff --git a/plugins/google_meet/node/cli.py b/plugins/google_meet/node/cli.py index 4e10161e0cc..255b851ba6a 100644 --- a/plugins/google_meet/node/cli.py +++ b/plugins/google_meet/node/cli.py @@ -103,7 +103,7 @@ def node_command(args: argparse.Namespace) -> int: print(f"removed {args.name!r}" if ok else f"no such node: {args.name!r}") return 0 if ok else 1 - if cmd in ("status", "ping"): + if cmd in {"status", "ping"}: entry = reg.get(args.name) if entry is None: print(f"no such node: {args.name!r}", file=sys.stderr) diff --git a/plugins/google_meet/realtime/openai_client.py b/plugins/google_meet/realtime/openai_client.py index e9738d106ae..24527603e52 100644 --- a/plugins/google_meet/realtime/openai_client.py +++ b/plugins/google_meet/realtime/openai_client.py @@ -183,7 +183,7 @@ class RealtimeSession: rid = (frame.get("response") or {}).get("id") if rid: self._last_response_id = rid - elif ftype in ("response.done", "response.completed", "response.cancelled"): + elif ftype in {"response.done", "response.completed", "response.cancelled"}: break elif ftype == "error": err = frame.get("error") or frame diff --git a/plugins/google_meet/tools.py b/plugins/google_meet/tools.py index 9af804288c7..034116b88af 100644 --- a/plugins/google_meet/tools.py +++ b/plugins/google_meet/tools.py @@ -36,7 +36,7 @@ def check_meet_requirements() -> bool: handlers relax the requirement when a node is addressed. """ import platform as _p - if _p.system().lower() not in ("linux", "darwin"): + if _p.system().lower() not in {"linux", "darwin"}: return False try: import playwright # noqa: F401 @@ -238,7 +238,7 @@ def handle_meet_join(args: Dict[str, Any], **_kw) -> str: if not url: return _err("url is required") mode = (args.get("mode") or "transcribe").strip().lower() - if mode not in ("transcribe", "realtime"): + if mode not in {"transcribe", "realtime"}: return _err(f"mode must be 'transcribe' or 'realtime' (got {mode!r})") node = args.get("node") diff --git a/plugins/kanban/dashboard/dist/index.js b/plugins/kanban/dashboard/dist/index.js index 6f05df72bf6..3f6def61cef 100644 --- a/plugins/kanban/dashboard/dist/index.js +++ b/plugins/kanban/dashboard/dist/index.js @@ -908,6 +908,7 @@ return createNewBoard(payload).then(function () { setShowNewBoard(false); }); }, }) : null, + h(OrchestrationPanel, null), h(AttentionStrip, { boardData, onOpen: setSelectedTaskId, @@ -1386,6 +1387,288 @@ }, "?"); } + // --------------------------------------------------------------------- + // OrchestrationPanel — collapsible settings panel for the kanban + // orchestrator (orchestrator profile picker, default assignee picker, + // auto-decompose toggle, plus per-profile description editing with + // auto-generate). Backed by /orchestration + /profiles endpoints. + // --------------------------------------------------------------------- + + function OrchestrationPanel() { + const [expanded, setExpanded] = useState(false); + const [settings, setSettings] = useState(null); + const [profiles, setProfiles] = useState([]); + const [busy, setBusy] = useState({}); + const [msg, setMsg] = useState(null); + + const loadAll = useCallback(function () { + Promise.all([ + SDK.fetchJSON(`${API}/orchestration`), + SDK.fetchJSON(`${API}/profiles`), + ]).then(function (results) { + setSettings(results[0] || null); + setProfiles((results[1] && results[1].profiles) || []); + setMsg(null); + }).catch(function (err) { + setMsg({ ok: false, text: "Failed to load: " + (err.message || String(err)) }); + }); + }, []); + + useEffect(function () { + // Load on mount so the collapsed pill shows the real mode without + // requiring the user to expand the panel first. + if (settings === null) loadAll(); + }, [settings, loadAll]); + + const saveSettings = function (patch) { + setMsg(null); + return SDK.fetchJSON(`${API}/orchestration`, { + method: "PUT", + headers: { "Content-Type": "application/json" }, + body: JSON.stringify(patch), + }).then(function (res) { + setSettings(res); + setMsg({ ok: true, text: "Settings saved." }); + return res; + }).catch(function (err) { + setMsg({ ok: false, text: "Save failed: " + (err.message || String(err)) }); + }); + }; + + const saveProfileDescription = function (name, description) { + setBusy(function (b) { return Object.assign({}, b, { [name]: "save" }); }); + return SDK.fetchJSON(`${API}/profiles/${encodeURIComponent(name)}`, { + method: "PATCH", + headers: { "Content-Type": "application/json" }, + body: JSON.stringify({ description: description }), + }).then(function () { + loadAll(); + setMsg({ ok: true, text: `Description saved for ${name}.` }); + }).catch(function (err) { + setMsg({ ok: false, text: "Save failed: " + (err.message || String(err)) }); + }).then(function () { + setBusy(function (b) { + const next = Object.assign({}, b); delete next[name]; return next; + }); + }); + }; + + const autoGenerateDescription = function (name, overwrite) { + setBusy(function (b) { return Object.assign({}, b, { [name]: "auto" }); }); + return SDK.fetchJSON(`${API}/profiles/${encodeURIComponent(name)}/describe-auto`, { + method: "POST", + headers: { "Content-Type": "application/json" }, + body: JSON.stringify({ overwrite: !!overwrite }), + }).then(function (res) { + if (res && res.ok) { + loadAll(); + setMsg({ ok: true, text: `Auto-generated description for ${name}.` }); + } else { + setMsg({ + ok: false, + text: "Auto-generate failed: " + ((res && res.reason) || "unknown error"), + }); + } + }).catch(function (err) { + setMsg({ ok: false, text: "Auto-generate failed: " + (err.message || String(err)) }); + }).then(function () { + setBusy(function (b) { + const next = Object.assign({}, b); delete next[name]; return next; + }); + }); + }; + + const headerLabel = expanded + ? "▾ Orchestration settings" + : "▸ Orchestration settings"; + + // Mode pill — always visible (collapsed or expanded). One click flips + // between Auto and Manual. Auto = dispatcher decomposes new triage tasks + // every tick. Manual = pre-PR behavior, the user clicks ⚗ Decompose on + // each triage card (or runs `hermes kanban decompose `) and tasks + // stay in triage until then. + const autoOn = !!(settings && settings.auto_decompose); + const modePillTitle = settings === null + ? "Loading mode…" + : (autoOn + ? "Orchestration: Auto — the dispatcher decomposes new triage tasks automatically every tick. Click to switch to Manual (pre-PR behavior)." + : "Orchestration: Manual — triage tasks stay in triage until you click ⚗ Decompose on each card. Click to switch to Auto."); + const modePill = h("button", { + type: "button", + onClick: function () { + if (settings === null) return; // not loaded yet + saveSettings({ auto_decompose: !autoOn }); + }, + disabled: settings === null, + title: modePillTitle, + className: "inline-flex items-center gap-1 rounded-full border px-2 py-0.5 " + + "text-xs font-medium " + + (autoOn + ? "border-emerald-500/40 bg-emerald-500/10 text-emerald-700 dark:text-emerald-300" + : "border-muted-foreground/30 bg-muted/30 text-muted-foreground"), + }, + "Orchestration: ", + h("span", { className: "ml-1 font-semibold" }, + settings === null ? "…" : (autoOn ? "Auto" : "Manual")) + ); + + if (!expanded) { + return h("div", { className: "flex items-center gap-3 text-xs" }, + modePill, + h("button", { + type: "button", + onClick: function () { setExpanded(true); }, + className: "underline text-muted-foreground hover:text-foreground", + title: "Configure the kanban orchestrator (profile picker, default assignee, auto-decompose, profile descriptions)", + }, headerLabel), + ); + } + + const profileOptions = profiles.map(function (p) { + const tag = p.is_default ? " (default)" : ""; + return h(SelectOption, { key: p.name, value: p.name }, p.name + tag); + }); + + return h(Card, { className: "p-3" }, + h(CardContent, { className: "p-2 flex flex-col gap-3" }, + h("div", { className: "flex items-center justify-between" }, + h("button", { + type: "button", + onClick: function () { setExpanded(false); }, + className: "text-sm font-medium underline-offset-2 hover:underline", + }, headerLabel), + modePill, + h(Button, { onClick: loadAll, size: "sm" }, "Reload"), + ), + msg ? h("div", { + className: msg.ok ? "hermes-kanban-msg-ok" : "hermes-kanban-msg-err", + }, msg.text) : null, + + settings ? h("div", { className: "grid gap-3 sm:grid-cols-3" }, + h("div", { className: "flex flex-col gap-1" }, + h(Label, { className: "text-xs text-muted-foreground" }, + "Orchestrator profile"), + h(Select, { + value: settings.orchestrator_profile || "", + className: "h-8", + onChange: function (e) { + const v = (e && e.target ? e.target.value : e) || ""; + saveSettings({ orchestrator_profile: v }); + }, + }, + h(SelectOption, { value: "" }, + "(default: " + (settings.active_profile || "default") + ")"), + profileOptions, + ), + h("div", { className: "text-[10px] text-muted-foreground" }, + "Resolved: " + (settings.resolved_orchestrator_profile || "default")), + ), + h("div", { className: "flex flex-col gap-1" }, + h(Label, { className: "text-xs text-muted-foreground" }, + "Default assignee"), + h(Select, { + value: settings.default_assignee || "", + className: "h-8", + onChange: function (e) { + const v = (e && e.target ? e.target.value : e) || ""; + saveSettings({ default_assignee: v }); + }, + }, + h(SelectOption, { value: "" }, + "(default: " + (settings.active_profile || "default") + ")"), + profileOptions, + ), + h("div", { className: "text-[10px] text-muted-foreground" }, + "Resolved: " + (settings.resolved_default_assignee || "default")), + ), + h("div", { className: "flex flex-col gap-1" }, + h(Label, { className: "text-xs text-muted-foreground" }, + "Orchestration mode"), + h("label", { className: "flex items-center gap-2 text-xs h-8" }, + h("input", { + type: "checkbox", + checked: !!settings.auto_decompose, + onChange: function (e) { + saveSettings({ auto_decompose: !!e.target.checked }); + }, + }), + settings.auto_decompose ? "Auto (default)" : "Manual", + ), + h("div", { className: "text-[10px] text-muted-foreground" }, + "When on, the dispatcher decomposes new triage tasks automatically."), + ), + ) : h("div", { className: "text-xs text-muted-foreground" }, + "Loading…"), + + h("div", { className: "border-t pt-3" }, + h(Label, { className: "text-xs text-muted-foreground" }, + "Profile descriptions"), + h("div", { className: "text-[10px] text-muted-foreground pb-2" }, + "Descriptions guide the orchestrator's routing. Click ⚗ to auto-generate, or edit and save."), + profiles.length === 0 + ? h("div", { className: "text-xs text-muted-foreground" }, "No profiles installed.") + : h("div", { className: "flex flex-col gap-2" }, + profiles.map(function (p) { + return h(ProfileDescriptionRow, { + key: p.name, + profile: p, + busy: busy[p.name] || null, + onSave: saveProfileDescription, + onAuto: autoGenerateDescription, + }); + }), + ), + ), + ), + ); + } + + function ProfileDescriptionRow(props) { + const p = props.profile; + const [draft, setDraft] = useState(p.description || ""); + const busy = props.busy; + // Re-sync the local draft if the server-side description changes (e.g. + // after auto-generate). Cheap because re-runs only happen on prop change. + useEffect(function () { + setDraft(p.description || ""); + }, [p.description]); + + const tag = p.description_auto && p.description ? " [auto, review]" : ""; + return h("div", { className: "flex flex-col gap-1 border-l-2 pl-2", + style: { borderColor: p.description ? "#888" : "#cc6" } }, + h("div", { className: "flex items-center gap-2 text-xs" }, + h("span", { className: "font-medium" }, p.name), + p.is_default ? h("span", { className: "text-[10px] text-muted-foreground" }, "(default)") : null, + p.description_auto && p.description + ? h("span", { className: "text-[10px] text-yellow-600" }, "auto — review") + : null, + !p.description + ? h("span", { className: "text-[10px] text-yellow-600" }, "⚠ no description") + : null, + ), + h("div", { className: "flex items-center gap-2" }, + h(Input, { + value: draft, + onChange: function (e) { setDraft(e.target.value); }, + placeholder: "What is this profile good at?", + className: "h-7 text-xs flex-1", + }), + h(Button, { + onClick: function () { props.onSave(p.name, draft); }, + size: "sm", + disabled: !!busy || draft === (p.description || ""), + title: "Save the description above as user-authored", + }, busy === "save" ? "Saving…" : "Save"), + h(Button, { + onClick: function () { props.onAuto(p.name, true); }, + size: "sm", + disabled: !!busy, + title: "Auto-generate a description from this profile's skills and model", + }, busy === "auto" ? "Generating…" : "⚗ Auto"), + ), + ); + } + function BoardSwitcher(props) { const { t } = useI18n(); const list = props.boardList || []; @@ -2395,6 +2678,25 @@ }); }; + // POST /tasks/:id/decompose — fan a triage task out into a graph + // of child tasks routed to specialist profiles by description. + // Refreshes both the drawer (so the user sees the root flip to + // todo) and the board (so the new children appear in the columns). + const doDecompose = function () { + return SDK.fetchJSON( + withBoard(`${API}/tasks/${encodeURIComponent(props.taskId)}/decompose`, boardSlug), + { + method: "POST", + headers: { "Content-Type": "application/json" }, + body: JSON.stringify({}), + } + ).then(function (res) { + load(); + props.onRefresh(); + return res; + }); + }; + const addLink = function (parentId) { return SDK.fetchJSON(withBoard(`${API}/links`, boardSlug), { method: "POST", @@ -2486,6 +2788,7 @@ boardSlug: boardSlug, onPatch: doPatch, onSpecify: doSpecify, + onDecompose: doDecompose, onAddParent: addLink, onRemoveParent: removeLink, onAddChild: addChild, @@ -2559,6 +2862,7 @@ task: t, onPatch: props.onPatch, onSpecify: props.onSpecify, + onDecompose: props.onDecompose, }), h(DiagnosticsSection, { task: t, @@ -3023,6 +3327,8 @@ const task = props.task; const [specifyBusy, setSpecifyBusy] = useState(false); const [specifyMsg, setSpecifyMsg] = useState(null); + const [decomposeBusy, setDecomposeBusy] = useState(false); + const [decomposeMsg, setDecomposeMsg] = useState(null); const b = function (label, patch, enabled, confirmMsg) { return h(Button, { onClick: function () { if (enabled !== false) props.onPatch(patch, { confirm: confirmMsg }); }, @@ -3067,9 +3373,57 @@ }, specifyBusy ? "Specifying…" : "✨ Specify") : null; + // "Decompose" is the orchestrator-driven fan-out. Like Specify, only + // makes sense on triage-column tasks — elsewhere the backend short- + // circuits with ok:false. When the orchestrator returns fanout:false + // we render the same single-task message as Specify; when it fans + // out we report the child count for quick at-a-glance verification. + const decomposeButton = (task.status === "triage" && props.onDecompose) + ? h(Button, { + onClick: function () { + if (decomposeBusy) return; + setDecomposeBusy(true); + setDecomposeMsg(null); + props.onDecompose().then(function (res) { + if (res && res.ok) { + if (res.fanout && res.child_ids && res.child_ids.length) { + setDecomposeMsg({ + ok: true, + text: `Decomposed into ${res.child_ids.length} children: ${res.child_ids.join(", ")}`, + }); + } else { + const suffix = res.new_title + ? ` — retitled: ${res.new_title}` + : ""; + setDecomposeMsg({ + ok: true, + text: `Single task (no fanout)${suffix}`, + }); + } + } else { + setDecomposeMsg({ + ok: false, + text: "Decompose failed: " + ((res && res.reason) || "unknown error"), + }); + } + }).catch(function (err) { + setDecomposeMsg({ + ok: false, + text: "Decompose failed: " + (err.message || String(err)), + }); + }).then(function () { + setDecomposeBusy(false); + }); + }, + disabled: decomposeBusy, + size: "sm", + }, decomposeBusy ? "Decomposing…" : "⚗ Decompose") + : null; + return h("div", null, h("div", { className: "hermes-kanban-actions" }, specifyButton, + decomposeButton, b("→ triage", { status: "triage" }, task.status !== "triage"), b("→ ready", { status: "ready" }, task.status !== "ready"), // No direct → running button: /tasks/:id PATCH rejects status=running @@ -3091,6 +3445,11 @@ ? "hermes-kanban-msg-ok" : "hermes-kanban-msg-err", }, specifyMsg.text) : null, + decomposeMsg ? h("div", { + className: decomposeMsg.ok + ? "hermes-kanban-msg-ok" + : "hermes-kanban-msg-err", + }, decomposeMsg.text) : null, ); } diff --git a/plugins/kanban/dashboard/plugin_api.py b/plugins/kanban/dashboard/plugin_api.py index 7b0cb1d791a..16e60663854 100644 --- a/plugins/kanban/dashboard/plugin_api.py +++ b/plugins/kanban/dashboard/plugin_api.py @@ -628,7 +628,7 @@ def update_task(task_id: str, payload: UpdateTaskBody, board: Optional[str] = Qu status_code=400, detail="Cannot set status to 'running' directly; use the dispatcher/claim path", ) - elif s in ("todo", "triage"): + elif s in {"todo", "triage"}: ok = _set_status_direct(conn, task_id, s) else: raise HTTPException(status_code=400, detail=f"unknown status: {s}") @@ -742,7 +742,7 @@ def _set_status_direct( (task_id, run_id, json.dumps({"status": new_status}), int(time.time())), ) # If we re-opened something, children may have gone stale. - if new_status in ("done", "ready"): + if new_status in {"done", "ready"}: kanban_db.recompute_ready(conn) return True @@ -868,7 +868,7 @@ def bulk_update(payload: BulkTaskBody, board: Optional[str] = Query(None)): ok = kanban_db.unblock_task(conn, tid) else: ok = _set_status_direct(conn, tid, "ready") - elif s in ("todo", "running", "triage"): + elif s in {"todo", "running", "triage"}: ok = _set_status_direct(conn, tid, s) else: entry.update(ok=False, error=f"unknown status {s!r}") @@ -1535,6 +1535,279 @@ def switch_board(slug: str): _EVENT_POLL_SECONDS = 0.3 +# --------------------------------------------------------------------------- +# Profile metadata & description editing (consumed by the kanban orchestrator) +# --------------------------------------------------------------------------- + +class DescribeBody(BaseModel): + description: Optional[str] = None # explicit user-authored text + + +class DescribeAutoBody(BaseModel): + overwrite: bool = False + + +@router.get("/profiles") +def list_profile_roster(): + """Return every installed profile with its description. + + Consumed by the dashboard's settings panel (orchestrator picker) + and the profile-description editing UI. Profiles without a + description still appear here — they're routable on name alone, + just less precisely. + """ + try: + from hermes_cli import profiles as profiles_mod + profiles = profiles_mod.list_profiles() + except Exception as exc: + raise HTTPException(status_code=500, detail=f"failed to list profiles: {exc}") + return { + "profiles": [ + { + "name": p.name, + "is_default": bool(p.is_default), + "model": p.model or "", + "provider": p.provider or "", + "description": p.description or "", + "description_auto": bool(p.description_auto), + "skill_count": int(p.skill_count or 0), + } + for p in profiles + ], + } + + +@router.patch("/profiles/{profile_name}") +def update_profile_description(profile_name: str, payload: DescribeBody): + """Set or clear the description of a profile. + + Empty string clears the description; non-empty stores it as a + user-authored description (``description_auto: false``) so the + auto-describer won't overwrite it on a sweep without + ``--overwrite``. + """ + try: + from hermes_cli import profiles as profiles_mod + canon = profiles_mod.normalize_profile_name(profile_name) + if canon == "default": + from hermes_constants import get_hermes_home # type: ignore + from pathlib import Path as _Path + profile_dir = _Path(get_hermes_home()) + else: + profile_dir = profiles_mod.get_profile_dir(canon) + if not profile_dir.is_dir(): + raise HTTPException(status_code=404, detail=f"profile '{profile_name}' not found") + text = (payload.description or "").strip() + profiles_mod.write_profile_meta( + profile_dir, + description=text, + description_auto=False, + ) + except HTTPException: + raise + except Exception as exc: + raise HTTPException(status_code=500, detail=f"failed to update profile: {exc}") + return {"ok": True, "profile": canon, "description": text} + + +@router.post("/profiles/{profile_name}/describe-auto") +def auto_describe_profile(profile_name: str, payload: DescribeAutoBody): + """Generate a description for the named profile via the auxiliary + LLM (``auxiliary.profile_describer``). Persists with + ``description_auto: true`` so the dashboard can surface a "review" + badge. + + Maps 1:1 to ``hermes profile describe --auto``. Non-OK + outcomes are NOT HTTP errors — the UI renders the reason inline + (e.g. "no auxiliary client configured") so the operator can fix + config and retry without a page reload. + """ + try: + from hermes_cli import profile_describer # noqa: WPS433 (intentional) + outcome = profile_describer.describe_profile( + profile_name, + overwrite=bool(payload.overwrite), + ) + except Exception as exc: + raise HTTPException(status_code=500, detail=f"describer crashed: {exc}") + return { + "ok": bool(outcome.ok), + "profile": outcome.profile_name, + "reason": outcome.reason, + "description": outcome.description, + } + + +# --------------------------------------------------------------------------- +# Decompose endpoint (orchestrator-driven fan-out) +# --------------------------------------------------------------------------- + +class DecomposeBody(BaseModel): + author: Optional[str] = None + + +@router.post("/tasks/{task_id}/decompose") +def decompose_task_endpoint( + task_id: str, + payload: DecomposeBody, + board: Optional[str] = Query(None), +): + """Fan a triage-column task out into a graph of child tasks via the + auxiliary LLM, routed to specialist profiles by description. Maps + 1:1 to ``hermes kanban decompose ``. + + Returns the outcome shape used by the CLI: ``{ok, task_id, reason, + fanout, child_ids, new_title}``. A non-OK outcome is NOT an HTTP + error — the UI renders the reason inline. + + Runs in FastAPI's threadpool (sync ``def``) because the LLM call + can take minutes on reasoning models. + """ + board = _resolve_board(board) + prev_env = os.environ.get("HERMES_KANBAN_BOARD") + try: + os.environ["HERMES_KANBAN_BOARD"] = board or kanban_db.DEFAULT_BOARD + from hermes_cli import kanban_decompose # noqa: WPS433 (intentional) + outcome = kanban_decompose.decompose_task( + task_id, + author=(payload.author or None), + ) + finally: + if prev_env is None: + os.environ.pop("HERMES_KANBAN_BOARD", None) + else: + os.environ["HERMES_KANBAN_BOARD"] = prev_env + + return { + "ok": bool(outcome.ok), + "task_id": outcome.task_id, + "reason": outcome.reason, + "fanout": bool(outcome.fanout), + "child_ids": outcome.child_ids or [], + "new_title": outcome.new_title, + } + + +# --------------------------------------------------------------------------- +# Orchestration settings (kanban.orchestrator_profile / default_assignee / +# auto_decompose) — surfaced to the dashboard's settings panel +# --------------------------------------------------------------------------- + +class OrchestrationSettingsBody(BaseModel): + orchestrator_profile: Optional[str] = None + default_assignee: Optional[str] = None + auto_decompose: Optional[bool] = None + + +@router.get("/orchestration") +def get_orchestration_settings(): + """Return the current kanban orchestration knobs from config.yaml + plus the resolved effective values (filling in fallbacks).""" + try: + from hermes_cli.config import load_config + cfg = load_config() or {} + except Exception: + cfg = {} + kanban_cfg = (cfg.get("kanban") or {}) if isinstance(cfg, dict) else {} + explicit_orch = (kanban_cfg.get("orchestrator_profile") or "").strip() + explicit_default = (kanban_cfg.get("default_assignee") or "").strip() + auto_decompose = bool(kanban_cfg.get("auto_decompose", True)) + + # Resolve fallbacks the same way the decomposer does. + resolved_orch = explicit_orch + resolved_default = explicit_default + try: + from hermes_cli import profiles as profiles_mod + active_default = profiles_mod.get_active_profile_name() or "default" + if not resolved_orch or not profiles_mod.profile_exists(resolved_orch): + resolved_orch = active_default + if not resolved_default or not profiles_mod.profile_exists(resolved_default): + resolved_default = active_default + except Exception: + active_default = "default" + if not resolved_orch: + resolved_orch = active_default + if not resolved_default: + resolved_default = active_default + + return { + "orchestrator_profile": explicit_orch, + "default_assignee": explicit_default, + "auto_decompose": auto_decompose, + "resolved_orchestrator_profile": resolved_orch, + "resolved_default_assignee": resolved_default, + "active_profile": active_default, + } + + +@router.put("/orchestration") +def set_orchestration_settings(payload: OrchestrationSettingsBody): + """Update the kanban orchestration knobs in ~/.hermes/config.yaml. + + Each field is optional — only fields explicitly passed are + written. ``orchestrator_profile`` / ``default_assignee`` accept + empty strings to clear the override and fall back to the default + profile. + """ + try: + from hermes_cli.config import load_config, save_config + cfg = load_config() or {} + except Exception as exc: + raise HTTPException(status_code=500, detail=f"failed to load config: {exc}") + + kanban_section = cfg.setdefault("kanban", {}) + if not isinstance(kanban_section, dict): + kanban_section = {} + cfg["kanban"] = kanban_section + + # Validate any non-empty profile names exist before saving. + try: + from hermes_cli import profiles as profiles_mod + except Exception: + profiles_mod = None # type: ignore + + if payload.orchestrator_profile is not None: + name = (payload.orchestrator_profile or "").strip() + if name and profiles_mod is not None: + try: + if not profiles_mod.profile_exists(name): + raise HTTPException( + status_code=400, + detail=f"profile '{name}' does not exist", + ) + except HTTPException: + raise + except Exception: + pass # fail open if the lookup itself errors + kanban_section["orchestrator_profile"] = name + + if payload.default_assignee is not None: + name = (payload.default_assignee or "").strip() + if name and profiles_mod is not None: + try: + if not profiles_mod.profile_exists(name): + raise HTTPException( + status_code=400, + detail=f"profile '{name}' does not exist", + ) + except HTTPException: + raise + except Exception: + pass + kanban_section["default_assignee"] = name + + if payload.auto_decompose is not None: + kanban_section["auto_decompose"] = bool(payload.auto_decompose) + + try: + save_config(cfg) + except Exception as exc: + raise HTTPException(status_code=500, detail=f"failed to save config: {exc}") + + # Echo back the resolved state (callers usually re-render from it). + return get_orchestration_settings() + + @router.websocket("/events") async def stream_events(ws: WebSocket): # Enforce the dashboard session token as a query param — browsers can't diff --git a/plugins/memory/byterover/__init__.py b/plugins/memory/byterover/__init__.py index 1870e9ab865..eafd9b2cfe5 100644 --- a/plugins/memory/byterover/__init__.py +++ b/plugins/memory/byterover/__init__.py @@ -263,7 +263,7 @@ class ByteRoverMemoryProvider(MemoryProvider): def on_memory_write(self, action: str, target: str, content: str) -> None: """Mirror built-in memory writes to ByteRover.""" - if action not in ("add", "replace") or not content: + if action not in {"add", "replace"} or not content: return def _write(): @@ -289,7 +289,7 @@ class ByteRoverMemoryProvider(MemoryProvider): for msg in messages[-10:]: # last 10 messages role = msg.get("role", "") content = msg.get("content", "") - if isinstance(content, str) and content.strip() and role in ("user", "assistant"): + if isinstance(content, str) and content.strip() and role in {"user", "assistant"}: parts.append(f"{role}: {content[:500]}") if not parts: diff --git a/plugins/memory/hindsight/__init__.py b/plugins/memory/hindsight/__init__.py index 52b1ac247f1..40772f79d8a 100644 --- a/plugins/memory/hindsight/__init__.py +++ b/plugins/memory/hindsight/__init__.py @@ -416,7 +416,7 @@ def _build_embedded_profile_env(config: dict[str, Any], *, llm_api_key: str | No current_base_url = config.get("llm_base_url") or os.environ.get("HINDSIGHT_API_LLM_BASE_URL", "") # The embedded daemon expects OpenAI wire format for these providers. - daemon_provider = "openai" if current_provider in ("openai_compatible", "openrouter") else current_provider + daemon_provider = "openai" if current_provider in {"openai_compatible", "openrouter"} else current_provider env_values = { "HINDSIGHT_API_LLM_PROVIDER": str(daemon_provider), @@ -596,7 +596,7 @@ class HindsightMemoryProvider(MemoryProvider): try: cfg = _load_config() mode = cfg.get("mode", "cloud") - if mode in ("local", "local_embedded"): + if mode in {"local", "local_embedded"}: available, _ = _check_local_runtime() return available if mode == "local_external": @@ -888,7 +888,7 @@ class HindsightMemoryProvider(MemoryProvider): from hindsight import HindsightEmbedded HindsightEmbedded.__del__ = lambda self: None llm_provider = self._config.get("llm_provider", "") - if llm_provider in ("openai_compatible", "openrouter"): + if llm_provider in {"openai_compatible", "openrouter"}: llm_provider = "openai" logger.debug("Creating HindsightEmbedded client (profile=%s, provider=%s)", self._config.get("profile", "hermes"), llm_provider) @@ -1132,7 +1132,7 @@ class HindsightMemoryProvider(MemoryProvider): self._mode = "disabled" return self._api_key = self._config.get("apiKey") or self._config.get("api_key") or os.environ.get("HINDSIGHT_API_KEY", "") - default_url = _DEFAULT_LOCAL_URL if self._mode in ("local_embedded", "local_external") else _DEFAULT_API_URL + default_url = _DEFAULT_LOCAL_URL if self._mode in {"local_embedded", "local_external"} else _DEFAULT_API_URL self._api_url = self._config.get("api_url") or os.environ.get("HINDSIGHT_API_URL", default_url) self._llm_base_url = self._config.get("llm_base_url", "") @@ -1152,10 +1152,10 @@ class HindsightMemoryProvider(MemoryProvider): self._budget = budget if budget in _VALID_BUDGETS else "mid" memory_mode = self._config.get("memory_mode", "hybrid") - self._memory_mode = memory_mode if memory_mode in ("context", "tools", "hybrid") else "hybrid" + self._memory_mode = memory_mode if memory_mode in {"context", "tools", "hybrid"} else "hybrid" prefetch_method = self._config.get("recall_prefetch_method") or self._config.get("prefetch_method", "recall") - self._prefetch_method = prefetch_method if prefetch_method in ("recall", "reflect") else "recall" + self._prefetch_method = prefetch_method if prefetch_method in {"recall", "reflect"} else "recall" # Bank options self._bank_mission = self._config.get("bank_mission", "") diff --git a/plugins/memory/honcho/__init__.py b/plugins/memory/honcho/__init__.py index d97f459acef..efbba937a4d 100644 --- a/plugins/memory/honcho/__init__.py +++ b/plugins/memory/honcho/__init__.py @@ -283,7 +283,7 @@ class HonchoMemoryProvider(MemoryProvider): # ----- Port #4053: cron guard ----- agent_context = kwargs.get("agent_context", "") platform = kwargs.get("platform", "cli") - if agent_context in ("cron", "flush") or platform == "cron": + if agent_context in {"cron", "flush"} or platform == "cron": logger.debug("Honcho skipped: cron/flush context (agent_context=%s, platform=%s)", agent_context, platform) self._cron_skipped = True @@ -404,7 +404,7 @@ class HonchoMemoryProvider(MemoryProvider): # pop_context_result() in prefetch(). Dialectic prewarm runs the # full configured depth and writes into _prefetch_result so turn 1 # consumes the result directly. - if self._recall_mode in ("context", "hybrid"): + if self._recall_mode in {"context", "hybrid"}: try: self._manager.prefetch_context(self._session_key) except Exception as e: diff --git a/plugins/memory/honcho/cli.py b/plugins/memory/honcho/cli.py index 402389ab962..28f213a1a66 100644 --- a/plugins/memory/honcho/cli.py +++ b/plugins/memory/honcho/cli.py @@ -233,7 +233,7 @@ _profile_override: str | None = None def _host_key() -> str: """Return the active Honcho host key, derived from the current Hermes profile.""" if _profile_override: - if _profile_override in ("default", "custom"): + if _profile_override in {"default", "custom"}: return HOST return f"{HOST}.{_profile_override}" return resolve_active_host() @@ -295,13 +295,13 @@ def _resolve_api_key(cfg: dict) -> str: parsed = urlparse(base_url) except (TypeError, ValueError): parsed = None - if parsed and parsed.scheme in ("http", "https") and parsed.netloc: + if parsed and parsed.scheme in {"http", "https"} and parsed.netloc: return "local" # Schemeless but looks like a host (contains '.' or ':' and isn't # a boolean literal): let it through so legacy configs don't # regress into "no API key configured" when they previously worked. lowered = base_url.lower() - if lowered not in ("true", "false", "none", "null") and any( + if lowered not in {"true", "false", "none", "null"} and any( c in base_url for c in ".:" ) and not base_url.isdigit(): return "local" @@ -334,7 +334,7 @@ def _ensure_sdk_installed() -> bool: print(" honcho-ai is not installed.") answer = _prompt("Install it now? (honcho-ai>=2.0.1)", default="y") - if answer.lower() not in ("y", "yes"): + if answer.lower() not in {"y", "yes"}: print(" Skipping install. Run: pip install 'honcho-ai>=2.0.1'\n") return False @@ -382,7 +382,7 @@ def cmd_setup(args) -> None: for h in ("localhost", "127.0.0.1", "::1") ) else "cloud" deploy = _prompt("Cloud or local?", default=current_deploy) - is_local = deploy.lower() in ("local", "l") + is_local = deploy.lower() in {"local", "l"} # Clean up legacy snake_case key cfg.pop("base_url", None) @@ -441,7 +441,7 @@ def cmd_setup(args) -> None: print(" directional -- all observations on, each AI peer builds its own view (default)") print(" unified -- shared pool, user observes self, AI observes others only") new_obs = _prompt("Observation mode", default=current_obs) - if new_obs in ("unified", "directional"): + if new_obs in {"unified", "directional"}: hermes_host["observationMode"] = new_obs else: hermes_host["observationMode"] = "directional" @@ -457,17 +457,17 @@ def cmd_setup(args) -> None: try: hermes_host["writeFrequency"] = int(new_wf) except (ValueError, TypeError): - hermes_host["writeFrequency"] = new_wf if new_wf in ("async", "turn", "session") else "async" + hermes_host["writeFrequency"] = new_wf if new_wf in {"async", "turn", "session"} else "async" # --- 6. Recall mode --- _raw_recall = hermes_host.get("recallMode") or cfg.get("recallMode", "hybrid") - current_recall = "hybrid" if _raw_recall not in ("hybrid", "context", "tools") else _raw_recall + current_recall = "hybrid" if _raw_recall not in {"hybrid", "context", "tools"} else _raw_recall print("\n Recall mode:") print(" hybrid -- auto-injected context + Honcho tools available (default)") print(" context -- auto-injected context only, Honcho tools hidden") print(" tools -- Honcho tools only, no auto-injected context") new_recall = _prompt("Recall mode", default=current_recall) - if new_recall in ("hybrid", "context", "tools"): + if new_recall in {"hybrid", "context", "tools"}: hermes_host["recallMode"] = new_recall # --- 7. Context token budget --- @@ -477,7 +477,7 @@ def cmd_setup(args) -> None: print(" uncapped -- no limit (default)") print(" N -- token limit per turn (e.g. 1200)") new_ctx_tokens = _prompt("Context tokens", default=current_display) - if new_ctx_tokens.strip().lower() in ("none", "uncapped", "no limit"): + if new_ctx_tokens.strip().lower() in {"none", "uncapped", "no limit"}: hermes_host.pop("contextTokens", None) elif new_ctx_tokens.strip() == "": pass # keep current @@ -517,7 +517,7 @@ def cmd_setup(args) -> None: print(" high -- complex behavioral patterns") print(" max -- thorough audit-level analysis") new_reasoning = _prompt("Reasoning level", default=current_reasoning) - if new_reasoning in ("minimal", "low", "medium", "high", "max"): + if new_reasoning in {"minimal", "low", "medium", "high", "max"}: hermes_host["dialecticReasoningLevel"] = new_reasoning else: hermes_host["dialecticReasoningLevel"] = "low" @@ -530,7 +530,7 @@ def cmd_setup(args) -> None: print(" per-repo -- one session per git repository") print(" global -- single session across all directories") new_strat = _prompt("Session strategy", default=current_strat) - if new_strat in ("per-session", "per-repo", "per-directory", "global"): + if new_strat in {"per-session", "per-repo", "per-directory", "global"}: hermes_host["sessionStrategy"] = new_strat hermes_host["enabled"] = True @@ -1130,7 +1130,7 @@ def cmd_migrate(args) -> None: print(" Paste the key when prompted.") print() answer = _prompt(" Run 'hermes honcho setup' now?", default="y") - if answer.lower() in ("y", "yes"): + if answer.lower() in {"y", "yes"}: cmd_setup(args) cfg = _read_config() has_key = bool(cfg.get("apiKey", "")) @@ -1176,7 +1176,7 @@ def cmd_migrate(args) -> None: print(" hermes honcho migrate — this step handles it interactively") if has_key: answer = _prompt(" Upload user memory files to Honcho now?", default="y") - if answer.lower() in ("y", "yes"): + if answer.lower() in {"y", "yes"}: try: from plugins.memory.honcho.client import ( HonchoClientConfig, @@ -1226,7 +1226,7 @@ def cmd_migrate(args) -> None: print() if has_key: answer = _prompt(" Seed AI identity from all detected files now?", default="y") - if answer.lower() in ("y", "yes"): + if answer.lower() in {"y", "yes"}: try: from plugins.memory.honcho.client import ( HonchoClientConfig, diff --git a/plugins/memory/honcho/client.py b/plugins/memory/honcho/client.py index de34642911e..eb268216c9b 100644 --- a/plugins/memory/honcho/client.py +++ b/plugins/memory/honcho/client.py @@ -47,7 +47,7 @@ def resolve_active_host() -> str: try: from hermes_cli.profiles import get_active_profile_name profile = get_active_profile_name() - if profile and profile not in ("default", "custom"): + if profile and profile not in {"default", "custom"}: return f"{HOST}.{profile}" except Exception: pass @@ -653,7 +653,7 @@ class HonchoClientConfig: return base # per-directory: one Honcho session per working directory (default) - if self.session_strategy in ("per-directory", "per-session"): + if self.session_strategy in {"per-directory", "per-session"}: base = Path(cwd).name if self.session_peer_prefix and self.peer_name: return f"{self.peer_name}-{base}" diff --git a/plugins/memory/openviking/__init__.py b/plugins/memory/openviking/__init__.py index ecb02b3de7e..ff01bbf402e 100644 --- a/plugins/memory/openviking/__init__.py +++ b/plugins/memory/openviking/__init__.py @@ -357,7 +357,7 @@ def _is_windows_absolute_path(value: str) -> bool: len(value) >= 3 and value[0].isalpha() and value[1] == ":" - and value[2] in ("/", "\\") + and value[2] in {"/", "\\"} ) @@ -381,7 +381,7 @@ def _is_local_path_reference(value: str) -> bool: def _path_from_file_uri(uri: str) -> Path | str: parsed = urlparse(uri) - if parsed.netloc not in ("", "localhost"): + if parsed.netloc not in {"", "localhost"}: return f"Unsupported non-local file URI: {uri}" return Path(url2pathname(parsed.path)).expanduser() @@ -755,7 +755,7 @@ class OpenVikingMemoryProvider(MemoryProvider): level = args.get("level", "overview") - summary_level = level in ("abstract", "overview") + summary_level = level in {"abstract", "overview"} # OpenViking expects directory URIs for pseudo summary files # (e.g. viking://user/hermes/.overview.md). resolved_uri = self._normalize_summary_uri(uri) if summary_level else uri @@ -832,7 +832,7 @@ class OpenVikingMemoryProvider(MemoryProvider): result = self._unwrap_result(resp) # Format list/tree results for readability - if action in ("list", "tree"): + if action in {"list", "tree"}: raw_entries = result if isinstance(result, dict): raw_entries = result.get("entries") or result.get("items") or result.get("children") or [] @@ -887,7 +887,7 @@ class OpenVikingMemoryProvider(MemoryProvider): payload: Dict[str, Any] = {} for key in ("reason", "to", "parent", "instruction", "wait", "timeout"): - if key in args and args[key] not in (None, ""): + if key in args and args[key] not in {None, ""}: payload[key] = args[key] parsed_url = urlparse(url) diff --git a/plugins/memory/supermemory/__init__.py b/plugins/memory/supermemory/__init__.py index f0cbfd60276..35b5b6fd649 100644 --- a/plugins/memory/supermemory/__init__.py +++ b/plugins/memory/supermemory/__init__.py @@ -88,9 +88,9 @@ def _as_bool(value: Any, default: bool) -> bool: return value if isinstance(value, str): lowered = value.strip().lower() - if lowered in ("true", "1", "yes", "y", "on"): + if lowered in {"true", "1", "yes", "y", "on"}: return True - if lowered in ("false", "0", "no", "n", "off"): + if lowered in {"false", "0", "no", "n", "off"}: return False return default @@ -508,7 +508,7 @@ class SupermemoryMemoryProvider(MemoryProvider): self._allowed_containers = [self._container_tag] + list(self._custom_containers) agent_context = kwargs.get("agent_context", "") - self._write_enabled = agent_context not in ("cron", "flush", "subagent") + self._write_enabled = agent_context not in {"cron", "flush", "subagent"} self._active = bool(self._api_key) self._client = None if self._active: @@ -598,7 +598,7 @@ class SupermemoryMemoryProvider(MemoryProvider): cleaned = [] for message in messages or []: role = message.get("role") - if role not in ("user", "assistant"): + if role not in {"user", "assistant"}: continue content = _clean_text_for_capture(str(message.get("content", ""))) if content: diff --git a/plugins/model-providers/deepseek/__init__.py b/plugins/model-providers/deepseek/__init__.py index 525766f87eb..34a8017b76e 100644 --- a/plugins/model-providers/deepseek/__init__.py +++ b/plugins/model-providers/deepseek/__init__.py @@ -74,9 +74,9 @@ class DeepSeekProfile(ProviderProfile): # its server default (currently high). if isinstance(reasoning_config, dict): effort = (reasoning_config.get("effort") or "").strip().lower() - if effort in ("xhigh", "max"): + if effort in {"xhigh", "max"}: top_level["reasoning_effort"] = "max" - elif effort in ("low", "medium", "high"): + elif effort in {"low", "medium", "high"}: top_level["reasoning_effort"] = effort return extra_body, top_level diff --git a/plugins/model-providers/kimi-coding/__init__.py b/plugins/model-providers/kimi-coding/__init__.py index b5cf53a8010..ed96ec514ef 100644 --- a/plugins/model-providers/kimi-coding/__init__.py +++ b/plugins/model-providers/kimi-coding/__init__.py @@ -37,7 +37,7 @@ class KimiProfile(ProviderProfile): # Enabled extra_body["thinking"] = {"type": "enabled"} effort = (reasoning_config.get("effort") or "").strip().lower() - if effort in ("low", "medium", "high"): + if effort in {"low", "medium", "high"}: top_level["reasoning_effort"] = effort else: top_level["reasoning_effort"] = "medium" diff --git a/plugins/platforms/google_chat/adapter.py b/plugins/platforms/google_chat/adapter.py index d8777bf7101..0fdf1ea9d86 100644 --- a/plugins/platforms/google_chat/adapter.py +++ b/plugins/platforms/google_chat/adapter.py @@ -1539,7 +1539,7 @@ class GoogleChatAdapter(BasePlatformAdapter): if sender_email and space_name: self._last_sender_by_chat[space_name] = sender_email.strip().lower() - chat_type = "dm" if space_type in ("DIRECT_MESSAGE", "DM") else "group" + chat_type = "dm" if space_type in {"DIRECT_MESSAGE", "DM"} else "group" text = msg.get("argumentText") or msg.get("text") or "" text = text.strip() @@ -1935,7 +1935,7 @@ class GoogleChatAdapter(BasePlatformAdapter): return True except HttpError as exc: status = getattr(getattr(exc, "resp", None), "status", None) - if status in (403, 404): + if status in {403, 404}: return False logger.debug( "[GoogleChat] delete_message failed: %s", @@ -1958,7 +1958,7 @@ class GoogleChatAdapter(BasePlatformAdapter): update_mask = ",".join(update_mask_fields) or "text" # Patch body cannot carry thread (immutable). - patch_body = {k: v for k, v in body.items() if k not in ("thread",)} + patch_body = {k: v for k, v in body.items() if k not in {"thread",}} def _do_patch() -> Dict[str, Any]: return ( @@ -2791,7 +2791,7 @@ class GoogleChatAdapter(BasePlatformAdapter): upload_resp = await asyncio.to_thread(_upload) except HttpError as exc: status = getattr(getattr(exc, "resp", None), "status", None) - if status in (401, 403): + if status in {401, 403}: logger.warning( "[GoogleChat] media.upload auth failure for identity=%s " "(token revoked or scope missing) — falling back to " @@ -2927,7 +2927,7 @@ class GoogleChatAdapter(BasePlatformAdapter): display = info.get("displayName") or chat_id return { "name": display, - "type": "dm" if space_type in ("DIRECT_MESSAGE", "DM") else "group", + "type": "dm" if space_type in {"DIRECT_MESSAGE", "DM"} else "group", "chat_id": chat_id, } @@ -3246,7 +3246,7 @@ async def _standalone_send( return {"error": "Google Chat standalone send: aiohttp not installed"} try: - async with _aiohttp.ClientSession(timeout=_aiohttp.ClientTimeout(total=30.0)) as session: + async with _aiohttp.ClientSession(timeout=_aiohttp.ClientTimeout(total=30.0), trust_env=True) as session: async with session.post( url, json=body, diff --git a/plugins/platforms/irc/adapter.py b/plugins/platforms/irc/adapter.py index ff10475d4e1..3358fa5b188 100644 --- a/plugins/platforms/irc/adapter.py +++ b/plugins/platforms/irc/adapter.py @@ -112,7 +112,7 @@ class IRCAdapter(BasePlatformAdapter): self.nickname = os.getenv("IRC_NICKNAME") or extra.get("nickname", "hermes-bot") self.channel = os.getenv("IRC_CHANNEL") or extra.get("channel", "") self.use_tls = ( - os.getenv("IRC_USE_TLS", "").lower() in ("1", "true", "yes") + os.getenv("IRC_USE_TLS", "").lower() in {"1", "true", "yes"} if os.getenv("IRC_USE_TLS") else extra.get("use_tls", True) ) @@ -680,7 +680,7 @@ def _env_enablement() -> dict | None: seed["nickname"] = nickname use_tls = os.getenv("IRC_USE_TLS", "").strip().lower() if use_tls: - seed["use_tls"] = use_tls in ("1", "true", "yes") + seed["use_tls"] = use_tls in {"1", "true", "yes"} # Passwords live in PlatformConfig.extra as well for back-compat with # existing config.yaml users; env-reads at construct time still win. if os.getenv("IRC_SERVER_PASSWORD"): @@ -756,7 +756,7 @@ async def _standalone_send( nickname = os.getenv("IRC_NICKNAME") or extra.get("nickname", "hermes-bot") use_tls_env = os.getenv("IRC_USE_TLS") if use_tls_env is not None: - use_tls = use_tls_env.lower() in ("1", "true", "yes") + use_tls = use_tls_env.lower() in {"1", "true", "yes"} else: use_tls = bool(extra.get("use_tls", True)) @@ -821,7 +821,7 @@ async def _standalone_send( await _raw(f"PONG :{payload}") elif cmd == "001": registered = True - elif cmd in ("432", "433"): + elif cmd in {"432", "433"}: nick_attempts += 1 if nick_attempts > max_nick_attempts: return {"error": "IRC standalone send: too many nick collisions"} @@ -829,7 +829,7 @@ async def _standalone_send( # mutated value, so the suffix stays bounded. standalone_nick = f"{nick_base}-cron-{nick_attempts}"[:30] await _raw(f"NICK {standalone_nick}") - elif cmd in ("464", "465"): + elif cmd in {"464", "465"}: return {"error": f"IRC standalone send: server rejected client ({cmd})"} if nickserv_password: @@ -860,9 +860,9 @@ async def _standalone_send( if jcmd == "PING": payload = jmsg["params"][0] if jmsg["params"] else "" await _raw(f"PONG :{payload}") - elif jcmd in ("366", "JOIN"): + elif jcmd in {"366", "JOIN"}: joined = True - elif jcmd in ("403", "405", "471", "473", "474", "475"): + elif jcmd in {"403", "405", "471", "473", "474", "475"}: return {"error": f"IRC standalone send: JOIN {target} rejected ({jcmd})"} # Bytes-aware per-line splitting so multi-line plain text never diff --git a/plugins/platforms/line/adapter.py b/plugins/platforms/line/adapter.py index db5d3564d32..49931aa57ab 100644 --- a/plugins/platforms/line/adapter.py +++ b/plugins/platforms/line/adapter.py @@ -325,7 +325,7 @@ class RequestCache: def mark_delivered(self, request_id: str) -> None: entry = self._entries.get(request_id) - if entry is None or entry.state not in (State.READY, State.ERROR): + if entry is None or entry.state not in {State.READY, State.ERROR}: return entry.state = State.DELIVERED entry.updated_at = time.time() @@ -447,7 +447,7 @@ class _LineClient: async def reply(self, reply_token: str, messages: List[Dict[str, Any]]) -> None: import aiohttp timeout = aiohttp.ClientTimeout(total=self._timeout) - async with aiohttp.ClientSession(timeout=timeout) as session: + async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session: async with session.post( LINE_REPLY_URL, headers=self._headers, @@ -460,7 +460,7 @@ class _LineClient: async def push(self, chat_id: str, messages: List[Dict[str, Any]]) -> None: import aiohttp timeout = aiohttp.ClientTimeout(total=self._timeout) - async with aiohttp.ClientSession(timeout=timeout) as session: + async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session: async with session.post( LINE_PUSH_URL, headers=self._headers, @@ -479,7 +479,7 @@ class _LineClient: clamped = max(5, min(60, (seconds // 5) * 5 or 5)) try: timeout = aiohttp.ClientTimeout(total=5.0) - async with aiohttp.ClientSession(timeout=timeout) as session: + async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session: await session.post( LINE_LOADING_URL, headers=self._headers, @@ -493,7 +493,7 @@ class _LineClient: import aiohttp url = LINE_CONTENT_URL_FMT.format(message_id=message_id) timeout = aiohttp.ClientTimeout(total=30.0) - async with aiohttp.ClientSession(timeout=timeout) as session: + async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session: async with session.get(url, headers={"Authorization": f"Bearer {self._token}"}) as resp: if resp.status >= 400: raise RuntimeError(f"LINE content {resp.status}") @@ -504,7 +504,7 @@ class _LineClient: import aiohttp timeout = aiohttp.ClientTimeout(total=10.0) try: - async with aiohttp.ClientSession(timeout=timeout) as session: + async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session: async with session.get(LINE_BOT_INFO_URL, headers=self._headers) as resp: if resp.status >= 400: return None @@ -614,7 +614,7 @@ def _truthy_env(name: str, default: bool = False) -> bool: v = os.getenv(name) if v is None: return default - return v.strip().lower() in ("1", "true", "yes", "on") + return v.strip().lower() in {"1", "true", "yes", "on"} # --------------------------------------------------------------------------- @@ -910,7 +910,7 @@ class LineAdapter(BasePlatformAdapter): await self._handle_message_event(event) elif event_type == "postback": await self._handle_postback_event(event) - elif event_type in ("follow", "unfollow", "join", "leave"): + elif event_type in {"follow", "unfollow", "join", "leave"}: logger.info("LINE: lifecycle event %s from %s", event_type, source) else: logger.debug("LINE: ignoring event type %r", event_type) @@ -939,7 +939,7 @@ class LineAdapter(BasePlatformAdapter): if msg_type == "text": text = msg.get("text", "") or "" - elif msg_type in ("image", "audio", "video", "file"): + elif msg_type in {"image", "audio", "video", "file"}: local_path = await self._download_media(message_id, msg_type) if local_path: media_urls.append(local_path) diff --git a/plugins/platforms/simplex/adapter.py b/plugins/platforms/simplex/adapter.py index b568f29bbb5..264deb89608 100644 --- a/plugins/platforms/simplex/adapter.py +++ b/plugins/platforms/simplex/adapter.py @@ -101,11 +101,11 @@ def _guess_extension(data: bytes) -> str: def _is_image_ext(ext: str) -> bool: - return ext.lower() in (".jpg", ".jpeg", ".png", ".gif", ".webp") + return ext.lower() in {".jpg", ".jpeg", ".png", ".gif", ".webp"} def _is_audio_ext(ext: str) -> bool: - return ext.lower() in (".mp3", ".wav", ".ogg", ".m4a", ".aac") + return ext.lower() in {".mp3", ".wav", ".ogg", ".m4a", ".aac"} # --------------------------------------------------------------------------- @@ -326,12 +326,12 @@ class SimplexAdapter(BasePlatformAdapter): # Filter out messages sent by us (direction == "snd") meta = chat_item.get("meta") or {} direction = (meta.get("itemStatus") or {}).get("type", "") - if direction in ("sndSent", "sndSentDirect", "sndSentViaProxy", "sndNew"): + if direction in {"sndSent", "sndSentDirect", "sndSentViaProxy", "sndNew"}: return # Determine chat type and IDs chat_type_raw = chat_info.get("type", "") - is_group = chat_type_raw in ("group", "groupInfo") + is_group = chat_type_raw in {"group", "groupInfo"} if is_group: group_info = chat_info.get("groupInfo") or chat_info.get("group") or {} @@ -374,7 +374,7 @@ class SimplexAdapter(BasePlatformAdapter): media_urls: List[str] = [] media_types: List[str] = [] file_info = chat_item.get("file") or {} - if file_info and file_info.get("fileStatus") not in ("cancelled", "error"): + if file_info and file_info.get("fileStatus") not in {"cancelled", "error"}: file_id = file_info.get("fileId") file_name = file_info.get("fileName", "file") if file_id: diff --git a/plugins/platforms/teams/adapter.py b/plugins/platforms/teams/adapter.py index c71baeb9d93..975ef5b4093 100644 --- a/plugins/platforms/teams/adapter.py +++ b/plugins/platforms/teams/adapter.py @@ -566,7 +566,7 @@ async def _standalone_send( # Per-request timeouts so a slow STS endpoint cannot starve the # subsequent activity POST of its budget. per_request_timeout = _aiohttp.ClientTimeout(total=15.0) - async with _aiohttp.ClientSession() as session: + async with _aiohttp.ClientSession(trust_env=True) as session: async with session.post( token_url, data={ @@ -841,7 +841,7 @@ class TeamsAdapter(BasePlatformAdapter): # bot silently treated every clicker as authorized — meaning any # Teams user who could message the bot could approve dangerous commands. allowed_csv = os.getenv("TEAMS_ALLOWED_USERS", "").strip() - allow_all = os.getenv("TEAMS_ALLOW_ALL_USERS", "").strip().lower() in ("1", "true", "yes") + allow_all = os.getenv("TEAMS_ALLOW_ALL_USERS", "").strip().lower() in {"1", "true", "yes"} if not allow_all: if not allowed_csv: diff --git a/plugins/teams_pipeline/cli.py b/plugins/teams_pipeline/cli.py index 0e1114e3e74..7afaa3888a0 100644 --- a/plugins/teams_pipeline/cli.py +++ b/plugins/teams_pipeline/cli.py @@ -99,15 +99,15 @@ def teams_pipeline_command(args: argparse.Namespace) -> int: return 2 try: - if action in ("list", "ls"): + if action in {"list", "ls"}: _cmd_list(args) elif action == "show": _cmd_show(args) - elif action in ("run", "replay"): + elif action in {"run", "replay"}: _cmd_run(args) - elif action in ("fetch", "test"): + elif action in {"fetch", "test"}: _cmd_fetch(args) - elif action in ("subscriptions", "subs"): + elif action in {"subscriptions", "subs"}: _cmd_subscriptions(args) elif action == "subscribe": _cmd_subscribe(args) @@ -117,7 +117,7 @@ def teams_pipeline_command(args: argparse.Namespace) -> int: _cmd_delete_subscription(args) elif action == "maintain-subscriptions": _cmd_maintain_subscriptions(args) - elif action in ("token-health", "token"): + elif action in {"token-health", "token"}: _cmd_token_health(args) elif action == "validate": _cmd_validate(args) diff --git a/plugins/teams_pipeline/meetings.py b/plugins/teams_pipeline/meetings.py index 6d2648abd52..ed024bc7e31 100644 --- a/plugins/teams_pipeline/meetings.py +++ b/plugins/teams_pipeline/meetings.py @@ -33,7 +33,7 @@ def _meeting_path(meeting_ref: TeamsMeetingRef | str) -> str: def _wrap_graph_error(exc: MicrosoftGraphAPIError, *, missing_message: str) -> TeamsMeetingError: - if exc.status_code in (401, 403): + if exc.status_code in {401, 403}: return TeamsMeetingPermissionError(str(exc)) if exc.status_code == 404: return TeamsMeetingNotFoundError(missing_message) @@ -286,7 +286,7 @@ async def fetch_call_record_artifact( try: payload = await client.get_json(f"/communications/callRecords/{quote(call_record_id, safe='')}") except MicrosoftGraphAPIError as exc: - if exc.status_code in (401, 403) and allow_permission_errors: + if exc.status_code in {401, 403} and allow_permission_errors: return None if exc.status_code == 404: return None diff --git a/plugins/teams_pipeline/models.py b/plugins/teams_pipeline/models.py index 8d85092be96..b1ae5196f51 100644 --- a/plugins/teams_pipeline/models.py +++ b/plugins/teams_pipeline/models.py @@ -145,7 +145,7 @@ class MeetingArtifact: metadata: dict[str, Any] = field(default_factory=dict) def __post_init__(self) -> None: - if self.artifact_type not in ("transcript", "recording", "call_record"): + if self.artifact_type not in {"transcript", "recording", "call_record"}: raise ValueError( "MeetingArtifact.artifact_type must be transcript, recording, or call_record." ) diff --git a/plugins/teams_pipeline/runtime.py b/plugins/teams_pipeline/runtime.py index e8d3ada710c..f51be5e19e3 100644 --- a/plugins/teams_pipeline/runtime.py +++ b/plugins/teams_pipeline/runtime.py @@ -62,7 +62,7 @@ def build_pipeline_runtime_config(gateway_config: Any) -> dict[str, Any]: "chat_id", ): value = teams_extra.get(key) - if value not in (None, ""): + if value not in {None, ""}: teams_delivery[key] = value if teams_delivery: diff --git a/run_agent.py b/run_agent.py index 329163f12b3..28dfa905b92 100644 --- a/run_agent.py +++ b/run_agent.py @@ -70,38 +70,20 @@ from pathlib import Path from hermes_constants import get_hermes_home - -_OPENAI_CLS_CACHE: Optional[type] = None +# OpenAI lazy proxy + safe stdio + proxy URL helpers — see agent/process_bootstrap.py. +# `OpenAI` is re-exported here so `patch("run_agent.OpenAI", ...)` in tests works. +from agent.process_bootstrap import ( + OpenAI, + _OpenAIProxy, + _load_openai_cls, + _SafeWriter, + _install_safe_stdio, + _get_proxy_from_env, + _get_proxy_for_base_url, +) +from agent.iteration_budget import IterationBudget -def _load_openai_cls() -> type: - """Import and cache ``openai.OpenAI``.""" - global _OPENAI_CLS_CACHE - if _OPENAI_CLS_CACHE is None: - from openai import OpenAI as _cls - _OPENAI_CLS_CACHE = _cls - return _OPENAI_CLS_CACHE - - -class _OpenAIProxy: - """Module-level proxy that looks like ``openai.OpenAI`` but imports lazily.""" - - __slots__ = () - - def __call__(self, *args, **kwargs): - return _load_openai_cls()(*args, **kwargs) - - def __instancecheck__(self, obj): - return isinstance(obj, _load_openai_cls()) - - def __repr__(self): - return "" - - -OpenAI = _OpenAIProxy() - -# Load .env from ~/.hermes/.env first, then project root as dev fallback. -# User-managed env files should override stale shell exports on restart. from hermes_cli.env_loader import load_hermes_dotenv from hermes_cli.timeouts import ( get_provider_request_timeout, @@ -189,173 +171,41 @@ from agent.trajectory import ( convert_scratchpad_to_think, has_incomplete_scratchpad, save_trajectory as _save_trajectory_to_file, ) +from agent.message_sanitization import ( + _SURROGATE_RE, + _sanitize_surrogates, + _sanitize_structure_surrogates, + _sanitize_messages_surrogates, + _escape_invalid_chars_in_json_strings, + _repair_tool_call_arguments, + _strip_non_ascii, + _sanitize_messages_non_ascii, + _sanitize_tools_non_ascii, + _strip_images_from_messages, + _sanitize_structure_non_ascii, +) +from agent.tool_dispatch_helpers import ( + _NEVER_PARALLEL_TOOLS, + _PARALLEL_SAFE_TOOLS, + _PATH_SCOPED_TOOLS, + _DESTRUCTIVE_PATTERNS, + _REDIRECT_OVERWRITE, + _is_destructive_command, + _should_parallelize_tool_batch, + _extract_parallel_scope_path, + _paths_overlap, + _is_multimodal_tool_result, + _multimodal_text_summary, + _append_subdir_hint_to_multimodal, + _extract_file_mutation_targets, + _extract_error_preview, + _trajectory_normalize_msg, +) from utils import atomic_json_write, base_url_host_matches, base_url_hostname, env_var_enabled, normalize_proxy_url from hermes_cli.config import cfg_get -class _SafeWriter: - """Transparent stdio wrapper that catches OSError/ValueError from broken pipes. - - When hermes-agent runs as a systemd service, Docker container, or headless - daemon, the stdout/stderr pipe can become unavailable (idle timeout, buffer - exhaustion, socket reset). Any print() call then raises - ``OSError: [Errno 5] Input/output error``, which can crash agent setup or - run_conversation() — especially via double-fault when an except handler - also tries to print. - - Additionally, when subagents run in ThreadPoolExecutor threads, the shared - stdout handle can close between thread teardown and cleanup, raising - ``ValueError: I/O operation on closed file`` instead of OSError. - - This wrapper delegates all writes to the underlying stream and silently - catches both OSError and ValueError. It is transparent when the wrapped - stream is healthy. - """ - - __slots__ = ("_inner",) - - def __init__(self, inner): - object.__setattr__(self, "_inner", inner) - - def write(self, data): - try: - return self._inner.write(data) - except (OSError, ValueError): - return len(data) if isinstance(data, str) else 0 - - def flush(self): - try: - self._inner.flush() - except (OSError, ValueError): - pass - - def fileno(self): - return self._inner.fileno() - - def isatty(self): - try: - return self._inner.isatty() - except (OSError, ValueError): - return False - - def __getattr__(self, name): - return getattr(self._inner, name) - - -def _get_proxy_from_env() -> Optional[str]: - """Read proxy URL from environment variables. - - Checks HTTPS_PROXY, HTTP_PROXY, ALL_PROXY (and lowercase variants) in order. - Returns the first valid proxy URL found, or None if no proxy is configured. - """ - for key in ("HTTPS_PROXY", "HTTP_PROXY", "ALL_PROXY", - "https_proxy", "http_proxy", "all_proxy"): - value = os.environ.get(key, "").strip() - if value: - return normalize_proxy_url(value) - return None - - -def _get_proxy_for_base_url(base_url: Optional[str]) -> Optional[str]: - """Return an env-configured proxy unless NO_PROXY excludes this base URL.""" - proxy = _get_proxy_from_env() - if not proxy or not base_url: - return proxy - - host = base_url_hostname(base_url) - if not host: - return proxy - - try: - if urllib.request.proxy_bypass_environment(host): - return None - except Exception: - pass - - return proxy - - -def _install_safe_stdio() -> None: - """Wrap stdout/stderr so best-effort console output cannot crash the agent.""" - for stream_name in ("stdout", "stderr"): - stream = getattr(sys, stream_name, None) - if stream is not None and not isinstance(stream, _SafeWriter): - setattr(sys, stream_name, _SafeWriter(stream)) - - -class IterationBudget: - """Thread-safe iteration counter for an agent. - - Each agent (parent or subagent) gets its own ``IterationBudget``. - The parent's budget is capped at ``max_iterations`` (default 90). - Each subagent gets an independent budget capped at - ``delegation.max_iterations`` (default 50) — this means total - iterations across parent + subagents can exceed the parent's cap. - Users control the per-subagent limit via ``delegation.max_iterations`` - in config.yaml. - - ``execute_code`` (programmatic tool calling) iterations are refunded via - :meth:`refund` so they don't eat into the budget. - """ - - def __init__(self, max_total: int): - self.max_total = max_total - self._used = 0 - self._lock = threading.Lock() - - def consume(self) -> bool: - """Try to consume one iteration. Returns True if allowed.""" - with self._lock: - if self._used >= self.max_total: - return False - self._used += 1 - return True - - def refund(self) -> None: - """Give back one iteration (e.g. for execute_code turns).""" - with self._lock: - if self._used > 0: - self._used -= 1 - - @property - def used(self) -> int: - with self._lock: - return self._used - - @property - def remaining(self) -> int: - with self._lock: - return max(0, self.max_total - self._used) - - -# Tools that must never run concurrently (interactive / user-facing). -# When any of these appear in a batch, we fall back to sequential execution. -_NEVER_PARALLEL_TOOLS = frozenset({"clarify"}) - -# Read-only tools with no shared mutable session state. -_PARALLEL_SAFE_TOOLS = frozenset({ - "ha_get_state", - "ha_list_entities", - "ha_list_services", - "read_file", - "search_files", - "session_search", - "skill_view", - "skills_list", - "vision_analyze", - "web_extract", - "web_search", -}) - -# File tools can run concurrently when they target independent paths. -_PATH_SCOPED_TOOLS = frozenset({"read_file", "write_file", "patch"}) - -# Tools that mutate files on disk. Used by the per-turn verifier that -# surfaces silently-failed file edits so the model can't over-claim success. -# Imported above as `_FILE_MUTATING_TOOLS` from `agent.tool_result_classification`. - -# Maximum number of concurrent worker threads for parallel tool execution. _MAX_TOOL_WORKERS = 8 # Guard so the OpenRouter metadata pre-warm thread is only spawned once per @@ -364,682 +214,6 @@ _MAX_TOOL_WORKERS = 8 # exhaust the system thread limit (RuntimeError: can't start new thread). _openrouter_prewarm_done = threading.Event() -# Patterns that indicate a terminal command may modify/delete files. -_DESTRUCTIVE_PATTERNS = re.compile( - r"""(?:^|\s|&&|\|\||;|`)(?: - rm\s|rmdir\s| - cp\s|install\s| - mv\s| - sed\s+-i| - truncate\s| - dd\s| - shred\s| - git\s+(?:reset|clean|checkout)\s - )""", - re.VERBOSE, -) -# Output redirects that overwrite files (> but not >>) -_REDIRECT_OVERWRITE = re.compile(r'[^>]>[^>]|^>[^>]') - - -def _is_destructive_command(cmd: str) -> bool: - """Heuristic: does this terminal command look like it modifies/deletes files?""" - if not cmd: - return False - if _DESTRUCTIVE_PATTERNS.search(cmd): - return True - if _REDIRECT_OVERWRITE.search(cmd): - return True - return False - - -def _is_mcp_tool_parallel_safe(tool_name: str) -> bool: - """Check if an MCP tool comes from a server with parallel tool calls enabled. - - Lazy-imports from ``tools.mcp_tool`` to avoid circular dependencies. - Returns False if the MCP module is not available. - """ - try: - from tools.mcp_tool import is_mcp_tool_parallel_safe - return is_mcp_tool_parallel_safe(tool_name) - except Exception: - return False - - -def _should_parallelize_tool_batch(tool_calls) -> bool: - """Return True when a tool-call batch is safe to run concurrently.""" - if len(tool_calls) <= 1: - return False - - tool_names = [tc.function.name for tc in tool_calls] - if any(name in _NEVER_PARALLEL_TOOLS for name in tool_names): - return False - - reserved_paths: list[Path] = [] - for tool_call in tool_calls: - tool_name = tool_call.function.name - try: - function_args = json.loads(tool_call.function.arguments) - except Exception: - logging.debug( - "Could not parse args for %s — defaulting to sequential; raw=%s", - tool_name, - tool_call.function.arguments[:200], - ) - return False - if not isinstance(function_args, dict): - logging.debug( - "Non-dict args for %s (%s) — defaulting to sequential", - tool_name, - type(function_args).__name__, - ) - return False - - if tool_name in _PATH_SCOPED_TOOLS: - scoped_path = _extract_parallel_scope_path(tool_name, function_args) - if scoped_path is None: - return False - if any(_paths_overlap(scoped_path, existing) for existing in reserved_paths): - return False - reserved_paths.append(scoped_path) - continue - - if tool_name not in _PARALLEL_SAFE_TOOLS: - # Check if it's an MCP tool from a server that opted into parallel calls. - if not _is_mcp_tool_parallel_safe(tool_name): - return False - - return True - - -def _extract_parallel_scope_path(tool_name: str, function_args: dict) -> Path | None: - """Return the normalized file target for path-scoped tools.""" - if tool_name not in _PATH_SCOPED_TOOLS: - return None - - raw_path = function_args.get("path") - if not isinstance(raw_path, str) or not raw_path.strip(): - return None - - expanded = Path(raw_path).expanduser() - if expanded.is_absolute(): - return Path(os.path.abspath(str(expanded))) - - # Avoid resolve(); the file may not exist yet. - return Path(os.path.abspath(str(Path.cwd() / expanded))) - - -def _paths_overlap(left: Path, right: Path) -> bool: - """Return True when two paths may refer to the same subtree.""" - left_parts = left.parts - right_parts = right.parts - if not left_parts or not right_parts: - # Empty paths shouldn't reach here (guarded upstream), but be safe. - return bool(left_parts) == bool(right_parts) and bool(left_parts) - common_len = min(len(left_parts), len(right_parts)) - return left_parts[:common_len] == right_parts[:common_len] - - - -_SURROGATE_RE = re.compile(r'[\ud800-\udfff]') - - - - -def _is_multimodal_tool_result(value: Any) -> bool: - """True if the value is a multimodal tool result envelope. - - Multimodal handlers (e.g. tools/computer_use) return a dict with - `_multimodal=True`, a `content` key holding OpenAI-style content - parts, and an optional `text_summary` for string-only fallbacks. - """ - return ( - isinstance(value, dict) - and value.get("_multimodal") is True - and isinstance(value.get("content"), list) - ) - - -def _multimodal_text_summary(value: Any) -> str: - """Extract a plain text view of a multimodal tool result. - - Used wherever downstream code needs a string — logging, previews, - persistence size heuristics, fall-back content for providers that - don't support multipart tool messages. - """ - if _is_multimodal_tool_result(value): - if value.get("text_summary"): - return str(value["text_summary"]) - parts = [] - for p in value.get("content") or []: - if isinstance(p, dict) and p.get("type") == "text": - parts.append(str(p.get("text", ""))) - if parts: - return "\n".join(parts) - return "[multimodal tool result]" - if isinstance(value, str): - return value - try: - import json as _json - return _json.dumps(value, default=str) - except Exception: - return str(value) - - -def _append_subdir_hint_to_multimodal(value: Dict[str, Any], hint: str) -> None: - """Mutate a multimodal tool-result envelope to append a subdir hint. - - The hint is added to the first text part so the model sees it; image - parts are left untouched. `text_summary` is also updated for - string-fallback callers. - """ - if not _is_multimodal_tool_result(value): - return - parts = value.get("content") or [] - for p in parts: - if isinstance(p, dict) and p.get("type") == "text": - p["text"] = str(p.get("text", "")) + hint - break - else: - parts.insert(0, {"type": "text", "text": hint}) - value["content"] = parts - if isinstance(value.get("text_summary"), str): - value["text_summary"] = value["text_summary"] + hint - - -def _extract_file_mutation_targets(tool_name: str, args: Dict[str, Any]) -> List[str]: - """Return the file paths a ``write_file`` or ``patch`` call is targeting. - - For ``write_file`` and ``patch`` in replace mode this is just ``args["path"]``. - For ``patch`` in V4A patch mode we parse the patch content for - ``*** Update File:`` / ``*** Add File:`` / ``*** Delete File:`` headers so - the verifier can track each file in a multi-file patch separately. - """ - if tool_name not in _FILE_MUTATING_TOOLS: - return [] - if tool_name == "write_file": - p = args.get("path") - return [str(p)] if p else [] - # tool_name == "patch" - mode = args.get("mode") or "replace" - if mode == "replace": - p = args.get("path") - return [str(p)] if p else [] - if mode == "patch": - body = args.get("patch") or "" - if not isinstance(body, str) or not body: - return [] - import re as _re - paths: List[str] = [] - for _m in _re.finditer( - r'^\*\*\*\s+(?:Update|Add|Delete)\s+File:\s*(.+)$', - body, - _re.MULTILINE, - ): - p = _m.group(1).strip() - if p: - paths.append(p) - return paths - return [] - - -def _extract_error_preview(result: Any, max_len: int = 180) -> str: - """Pull a one-line error summary out of a tool result for footer display.""" - text = _multimodal_text_summary(result) if result is not None else "" - if not isinstance(text, str): - try: - text = str(text) - except Exception: - return "" - # Try to parse JSON and pull the ``error`` field — tool handlers return - # ``{"success": false, "error": "..."}``; raw string wins if parse fails. - stripped = text.strip() - if stripped.startswith("{"): - try: - import json as _json - data = _json.loads(stripped) - if isinstance(data, dict) and isinstance(data.get("error"), str): - text = data["error"] - except Exception: - pass - # Collapse whitespace, trim to max_len. - text = " ".join(text.split()) - if len(text) > max_len: - text = text[: max_len - 1] + "…" - return text - - -def _trajectory_normalize_msg(msg: Dict[str, Any]) -> Dict[str, Any]: - """Strip image blobs from a message for trajectory saving. - - Returns a shallow copy with multimodal tool results replaced by their - text_summary, and image parts in content lists replaced by - `[screenshot]` placeholders. Keeps the message schema otherwise intact. - """ - if not isinstance(msg, dict): - return msg - content = msg.get("content") - if _is_multimodal_tool_result(content): - return {**msg, "content": _multimodal_text_summary(content)} - if isinstance(content, list): - cleaned = [] - for p in content: - if isinstance(p, dict) and p.get("type") in {"image", "image_url", "input_image"}: - cleaned.append({"type": "text", "text": "[screenshot]"}) - else: - cleaned.append(p) - return {**msg, "content": cleaned} - return msg - - -def _sanitize_surrogates(text: str) -> str: - """Replace lone surrogate code points with U+FFFD (replacement character). - - Surrogates are invalid in UTF-8 and will crash ``json.dumps()`` inside the - OpenAI SDK. This is a fast no-op when the text contains no surrogates. - """ - if _SURROGATE_RE.search(text): - return _SURROGATE_RE.sub('\ufffd', text) - return text - - -# _summarize_user_message_for_log is imported from agent.codex_responses_adapter -# (see import block above). Remains importable from run_agent for backward compat. - - -def _sanitize_structure_surrogates(payload: Any) -> bool: - """Replace surrogate code points in nested dict/list payloads in-place. - - Mirror of ``_sanitize_structure_non_ascii`` but for surrogate recovery. - Used to scrub nested structured fields (e.g. ``reasoning_details`` — an - array of dicts with ``summary``/``text`` strings) that flat per-field - checks don't reach. Returns True if any surrogates were replaced. - """ - found = False - - def _walk(node): - nonlocal found - if isinstance(node, dict): - for key, value in node.items(): - if isinstance(value, str): - if _SURROGATE_RE.search(value): - node[key] = _SURROGATE_RE.sub('\ufffd', value) - found = True - elif isinstance(value, (dict, list)): - _walk(value) - elif isinstance(node, list): - for idx, value in enumerate(node): - if isinstance(value, str): - if _SURROGATE_RE.search(value): - node[idx] = _SURROGATE_RE.sub('\ufffd', value) - found = True - elif isinstance(value, (dict, list)): - _walk(value) - - _walk(payload) - return found - - -def _sanitize_messages_surrogates(messages: list) -> bool: - """Sanitize surrogate characters from all string content in a messages list. - - Walks message dicts in-place. Returns True if any surrogates were found - and replaced, False otherwise. Covers content/text, name, tool call - metadata/arguments, AND any additional string or nested structured fields - (``reasoning``, ``reasoning_content``, ``reasoning_details``, etc.) so - retries don't fail on a non-content field. Byte-level reasoning models - (xiaomi/mimo, kimi, glm) can emit lone surrogates in reasoning output - that flow through to ``api_messages["reasoning_content"]`` on the next - turn and crash json.dumps inside the OpenAI SDK. - """ - found = False - for msg in messages: - if not isinstance(msg, dict): - continue - content = msg.get("content") - if isinstance(content, str) and _SURROGATE_RE.search(content): - msg["content"] = _SURROGATE_RE.sub('\ufffd', content) - found = True - elif isinstance(content, list): - for part in content: - if isinstance(part, dict): - text = part.get("text") - if isinstance(text, str) and _SURROGATE_RE.search(text): - part["text"] = _SURROGATE_RE.sub('\ufffd', text) - found = True - name = msg.get("name") - if isinstance(name, str) and _SURROGATE_RE.search(name): - msg["name"] = _SURROGATE_RE.sub('\ufffd', name) - found = True - tool_calls = msg.get("tool_calls") - if isinstance(tool_calls, list): - for tc in tool_calls: - if not isinstance(tc, dict): - continue - tc_id = tc.get("id") - if isinstance(tc_id, str) and _SURROGATE_RE.search(tc_id): - tc["id"] = _SURROGATE_RE.sub('\ufffd', tc_id) - found = True - fn = tc.get("function") - if isinstance(fn, dict): - fn_name = fn.get("name") - if isinstance(fn_name, str) and _SURROGATE_RE.search(fn_name): - fn["name"] = _SURROGATE_RE.sub('\ufffd', fn_name) - found = True - fn_args = fn.get("arguments") - if isinstance(fn_args, str) and _SURROGATE_RE.search(fn_args): - fn["arguments"] = _SURROGATE_RE.sub('\ufffd', fn_args) - found = True - # Walk any additional string / nested fields (reasoning, - # reasoning_content, reasoning_details, etc.) — surrogates from - # byte-level reasoning models (xiaomi/mimo, kimi, glm) can lurk - # in these fields and aren't covered by the per-field checks above. - # Matches _sanitize_messages_non_ascii's coverage (PR #10537). - for key, value in msg.items(): - if key in {"content", "name", "tool_calls", "role"}: - continue - if isinstance(value, str): - if _SURROGATE_RE.search(value): - msg[key] = _SURROGATE_RE.sub('\ufffd', value) - found = True - elif isinstance(value, (dict, list)): - if _sanitize_structure_surrogates(value): - found = True - return found - - -def _escape_invalid_chars_in_json_strings(raw: str) -> str: - """Escape unescaped control chars inside JSON string values. - - Walks the raw JSON character-by-character, tracking whether we are - inside a double-quoted string. Inside strings, replaces literal - control characters (0x00-0x1F) that aren't already part of an escape - sequence with their ``\\uXXXX`` equivalents. Pass-through for everything - else. - - Ported from #12093 — complements the other repair passes in - ``_repair_tool_call_arguments`` when ``json.loads(strict=False)`` is - not enough (e.g. llama.cpp backends that emit literal apostrophes or - tabs alongside other malformations). - """ - out: list[str] = [] - in_string = False - i = 0 - n = len(raw) - while i < n: - ch = raw[i] - if in_string: - if ch == "\\" and i + 1 < n: - # Already-escaped char — pass through as-is - out.append(ch) - out.append(raw[i + 1]) - i += 2 - continue - if ch == '"': - in_string = False - out.append(ch) - elif ord(ch) < 0x20: - out.append(f"\\u{ord(ch):04x}") - else: - out.append(ch) - else: - if ch == '"': - in_string = True - out.append(ch) - i += 1 - return "".join(out) - - -def _repair_tool_call_arguments(raw_args: str, tool_name: str = "?") -> str: - """Attempt to repair malformed tool_call argument JSON. - - Models like GLM-5.1 via Ollama can produce truncated JSON, trailing - commas, Python ``None``, etc. The API proxy rejects these with HTTP 400 - "invalid tool call arguments". This function applies common repairs; - if all fail it returns ``"{}"`` so the request succeeds (better than - crashing the session). All repairs are logged at WARNING level. - """ - raw_stripped = raw_args.strip() if isinstance(raw_args, str) else "" - - # Fast-path: empty / whitespace-only -> empty object - if not raw_stripped: - logger.warning("Sanitized empty tool_call arguments for %s", tool_name) - return "{}" - - # Python-literal None -> normalise to {} - if raw_stripped == "None": - logger.warning("Sanitized Python-None tool_call arguments for %s", tool_name) - return "{}" - - # Repair pass 0: llama.cpp backends sometimes emit literal control - # characters (tabs, newlines) inside JSON string values. json.loads - # with strict=False accepts these and lets us re-serialise the - # result into wire-valid JSON without any string surgery. This is - # the most common local-model repair case (#12068). - try: - parsed = json.loads(raw_stripped, strict=False) - reserialised = json.dumps(parsed, separators=(",", ":")) - if reserialised != raw_stripped: - logger.warning( - "Repaired unescaped control chars in tool_call arguments for %s", - tool_name, - ) - return reserialised - except (json.JSONDecodeError, TypeError, ValueError): - pass - - # Attempt common JSON repairs - fixed = raw_stripped - # 1. Strip trailing commas before } or ] - fixed = re.sub(r',\s*([}\]])', r'\1', fixed) - # 2. Close unclosed structures - open_curly = fixed.count('{') - fixed.count('}') - open_bracket = fixed.count('[') - fixed.count(']') - if open_curly > 0: - fixed += '}' * open_curly - if open_bracket > 0: - fixed += ']' * open_bracket - # 3. Remove excess closing braces/brackets (bounded to 50 iterations) - for _ in range(50): - try: - json.loads(fixed) - break - except json.JSONDecodeError: - if fixed.endswith('}') and fixed.count('}') > fixed.count('{'): - fixed = fixed[:-1] - elif fixed.endswith(']') and fixed.count(']') > fixed.count('['): - fixed = fixed[:-1] - else: - break - - try: - json.loads(fixed) - logger.warning( - "Repaired malformed tool_call arguments for %s: %s → %s", - tool_name, raw_stripped[:80], fixed[:80], - ) - return fixed - except json.JSONDecodeError: - pass - - # Repair pass 4: escape unescaped control chars inside JSON strings, - # then retry. Catches cases where strict=False alone fails because - # other malformations are present too. - try: - escaped = _escape_invalid_chars_in_json_strings(fixed) - if escaped != fixed: - json.loads(escaped) - logger.warning( - "Repaired control-char-laced tool_call arguments for %s: %s → %s", - tool_name, raw_stripped[:80], escaped[:80], - ) - return escaped - except (json.JSONDecodeError, TypeError, ValueError): - pass - - # Last resort: replace with empty object so the API request doesn't - # crash the entire session. - logger.warning( - "Unrepairable tool_call arguments for %s — " - "replaced with empty object (was: %s)", - tool_name, raw_stripped[:80], - ) - return "{}" - - -def _strip_non_ascii(text: str) -> str: - """Remove non-ASCII characters, replacing with closest ASCII equivalent or removing. - - Used as a last resort when the system encoding is ASCII and can't handle - any non-ASCII characters (e.g. LANG=C on Chromebooks). - """ - return text.encode('ascii', errors='ignore').decode('ascii') - - -def _sanitize_messages_non_ascii(messages: list) -> bool: - """Strip non-ASCII characters from all string content in a messages list. - - This is a last-resort recovery for systems with ASCII-only encoding - (LANG=C, Chromebooks, minimal containers). Returns True if any - non-ASCII content was found and sanitized. - """ - found = False - for msg in messages: - if not isinstance(msg, dict): - continue - # Sanitize content (string) - content = msg.get("content") - if isinstance(content, str): - sanitized = _strip_non_ascii(content) - if sanitized != content: - msg["content"] = sanitized - found = True - elif isinstance(content, list): - for part in content: - if isinstance(part, dict): - text = part.get("text") - if isinstance(text, str): - sanitized = _strip_non_ascii(text) - if sanitized != text: - part["text"] = sanitized - found = True - # Sanitize name field (can contain non-ASCII in tool results) - name = msg.get("name") - if isinstance(name, str): - sanitized = _strip_non_ascii(name) - if sanitized != name: - msg["name"] = sanitized - found = True - # Sanitize tool_calls - tool_calls = msg.get("tool_calls") - if isinstance(tool_calls, list): - for tc in tool_calls: - if isinstance(tc, dict): - fn = tc.get("function", {}) - if isinstance(fn, dict): - fn_args = fn.get("arguments") - if isinstance(fn_args, str): - sanitized = _strip_non_ascii(fn_args) - if sanitized != fn_args: - fn["arguments"] = sanitized - found = True - # Sanitize any additional top-level string fields (e.g. reasoning_content) - for key, value in msg.items(): - if key in {"content", "name", "tool_calls", "role"}: - continue - if isinstance(value, str): - sanitized = _strip_non_ascii(value) - if sanitized != value: - msg[key] = sanitized - found = True - return found - - -def _sanitize_tools_non_ascii(tools: list) -> bool: - """Strip non-ASCII characters from tool payloads in-place.""" - return _sanitize_structure_non_ascii(tools) - - -def _strip_images_from_messages(messages: list) -> bool: - """Remove image_url content parts from all messages in-place. - - Called when a server signals it does not support images (e.g. - "Only 'text' content type is supported."). Mutates messages so the - next API call sends text only. - - Preserves message alternation invariants: - * ``tool``-role messages whose content was entirely images are replaced - with a plaintext placeholder, NOT deleted — deleting them would leave - the paired ``tool_call_id`` on the prior assistant message unmatched, - which providers reject with HTTP 400. - * Non-tool messages whose content becomes empty are dropped. In - practice this only hits synthetic image-only user messages appended - for attachment delivery; real user turns always include text. - - Returns True if any image parts were removed. - """ - found = False - to_delete = [] - for i, msg in enumerate(messages): - if not isinstance(msg, dict): - continue - content = msg.get("content") - if not isinstance(content, list): - continue - new_parts = [] - for part in content: - if isinstance(part, dict) and part.get("type") in {"image_url", "image", "input_image"}: - found = True - else: - new_parts.append(part) - if len(new_parts) < len(content): - if new_parts: - msg["content"] = new_parts - elif msg.get("role") == "tool": - # Preserve tool_call_id linkage — providers require every - # assistant tool_call to have a matching tool response. - msg["content"] = "[image content removed — server does not support images]" - else: - # Synthetic image-only user/assistant message with no text; - # safe to drop. - to_delete.append(i) - for i in reversed(to_delete): - del messages[i] - return found - - -def _sanitize_structure_non_ascii(payload: Any) -> bool: - """Strip non-ASCII characters from nested dict/list payloads in-place.""" - found = False - - def _walk(node): - nonlocal found - if isinstance(node, dict): - for key, value in node.items(): - if isinstance(value, str): - sanitized = _strip_non_ascii(value) - if sanitized != value: - node[key] = sanitized - found = True - elif isinstance(value, (dict, list)): - _walk(value) - elif isinstance(node, list): - for idx, value in enumerate(node): - if isinstance(value, str): - sanitized = _strip_non_ascii(value) - if sanitized != value: - node[idx] = sanitized - found = True - elif isinstance(value, (dict, list)): - _walk(value) - - _walk(payload) - return found - - - - - # ========================================================================= # Large tool result handler — save oversized output to temp file # ========================================================================= @@ -1200,1331 +374,75 @@ class AIAgent: checkpoint_max_file_size_mb: int = 10, pass_session_id: bool = False, ): - """ - Initialize the AI Agent. - - Args: - base_url (str): Base URL for the model API (optional) - api_key (str): API key for authentication (optional, uses env var if not provided) - provider (str): Provider identifier (optional; used for telemetry/routing hints) - api_mode (str): API mode override: "chat_completions" or "codex_responses" - model (str): Model name to use (default: "anthropic/claude-opus-4.6") - max_iterations (int): Maximum number of tool calling iterations (default: 90) - tool_delay (float): Delay between tool calls in seconds (default: 1.0) - enabled_toolsets (List[str]): Only enable tools from these toolsets (optional) - disabled_toolsets (List[str]): Disable tools from these toolsets (optional) - save_trajectories (bool): Whether to save conversation trajectories to JSONL files (default: False) - verbose_logging (bool): Enable verbose logging for debugging (default: False) - quiet_mode (bool): Suppress progress output for clean CLI experience (default: False) - ephemeral_system_prompt (str): System prompt used during agent execution but NOT saved to trajectories (optional) - log_prefix_chars (int): Number of characters to show in log previews for tool calls/responses (default: 100) - log_prefix (str): Prefix to add to all log messages for identification in parallel processing (default: "") - providers_allowed (List[str]): OpenRouter providers to allow (optional) - providers_ignored (List[str]): OpenRouter providers to ignore (optional) - providers_order (List[str]): OpenRouter providers to try in order (optional) - provider_sort (str): Sort providers by price/throughput/latency (optional) - openrouter_min_coding_score (float): Coding-score floor (0.0-1.0) for the - openrouter/pareto-code router. Only applied when model == "openrouter/pareto-code". - None or empty = let OpenRouter pick the strongest available coder. - session_id (str): Pre-generated session ID for logging (optional, auto-generated if not provided) - tool_progress_callback (callable): Callback function(tool_name, args_preview) for progress notifications - clarify_callback (callable): Callback function(question, choices) -> str for interactive user questions. - Provided by the platform layer (CLI or gateway). If None, the clarify tool returns an error. - max_tokens (int): Maximum tokens for model responses (optional, uses model default if not set) - reasoning_config (Dict): OpenRouter reasoning configuration override (e.g. {"effort": "none"} to disable thinking). - If None, defaults to {"enabled": True, "effort": "medium"} for OpenRouter. Set to disable/customize reasoning. - prefill_messages (List[Dict]): Messages to prepend to conversation history as prefilled context. - Useful for injecting a few-shot example or priming the model's response style. - Example: [{"role": "user", "content": "Hi!"}, {"role": "assistant", "content": "Hello!"}] - NOTE: Anthropic Sonnet 4.6+ and Opus 4.6+ reject a conversation that ends on an - assistant-role message (400 error). For those models use structured outputs or - output_config.format instead of a trailing-assistant prefill. - platform (str): The interface platform the user is on (e.g. "cli", "telegram", "discord", "whatsapp"). - Used to inject platform-specific formatting hints into the system prompt. - skip_context_files (bool): If True, skip auto-injection of SOUL.md, AGENTS.md, and .cursorrules - into the system prompt. Use this for batch processing and data generation to avoid - polluting trajectories with user-specific persona or project instructions. - load_soul_identity (bool): If True, still use ~/.hermes/SOUL.md as the primary - identity even when skip_context_files=True. Project context files from the cwd - remain skipped. - """ - _install_safe_stdio() - - self.model = model - self.max_iterations = max_iterations - # Shared iteration budget — parent creates, children inherit. - # Consumed by every LLM turn across parent + all subagents. - self.iteration_budget = iteration_budget or IterationBudget(max_iterations) - self.tool_delay = tool_delay - self.save_trajectories = save_trajectories - self.verbose_logging = verbose_logging - self.quiet_mode = quiet_mode - self.ephemeral_system_prompt = ephemeral_system_prompt - self.platform = platform # "cli", "telegram", "discord", "whatsapp", etc. - self._user_id = user_id # Platform user identifier (gateway sessions) - self._user_name = user_name - self._chat_id = chat_id - self._chat_name = chat_name - self._chat_type = chat_type - self._thread_id = thread_id - self._gateway_session_key = gateway_session_key # Stable per-chat key (e.g. agent:main:telegram:dm:123) - # Pluggable print function — CLI replaces this with _cprint so that - # raw ANSI status lines are routed through prompt_toolkit's renderer - # instead of going directly to stdout where patch_stdout's StdoutProxy - # would mangle the escape sequences. None = use builtins.print. - self._print_fn = None - self.background_review_callback = None # Optional sync callback for gateway delivery - self.skip_context_files = skip_context_files - self.load_soul_identity = load_soul_identity - self.pass_session_id = pass_session_id - self._credential_pool = credential_pool - self.log_prefix_chars = log_prefix_chars - self.log_prefix = f"{log_prefix} " if log_prefix else "" - # Store effective base URL for feature detection (prompt caching, reasoning, etc.) - self.base_url = base_url or "" - provider_name = provider.strip().lower() if isinstance(provider, str) and provider.strip() else None - self.provider = provider_name or "" - self.acp_command = acp_command or command - self.acp_args = list(acp_args or args or []) - if api_mode in {"chat_completions", "codex_responses", "anthropic_messages", "bedrock_converse", "codex_app_server"}: - self.api_mode = api_mode - elif self.provider == "openai-codex": - self.api_mode = "codex_responses" - elif self.provider in {"xai", "xai-oauth"}: - self.api_mode = "codex_responses" - elif (provider_name is None) and ( - self._base_url_hostname == "chatgpt.com" - and "/backend-api/codex" in self._base_url_lower - ): - self.api_mode = "codex_responses" - self.provider = "openai-codex" - elif (provider_name is None) and self._base_url_hostname == "api.x.ai": - self.api_mode = "codex_responses" - self.provider = "xai" - elif self.provider == "anthropic" or (provider_name is None and self._base_url_hostname == "api.anthropic.com"): - self.api_mode = "anthropic_messages" - self.provider = "anthropic" - elif self._base_url_lower.rstrip("/").endswith("/anthropic"): - # Third-party Anthropic-compatible endpoints (e.g. MiniMax, DashScope) - # use a URL convention ending in /anthropic. Auto-detect these so the - # Anthropic Messages API adapter is used instead of chat completions. - self.api_mode = "anthropic_messages" - elif self.provider == "bedrock" or ( - self._base_url_hostname.startswith("bedrock-runtime.") - and base_url_host_matches(self._base_url_lower, "amazonaws.com") - ): - # AWS Bedrock — auto-detect from provider name or base URL - # (bedrock-runtime..amazonaws.com). - self.api_mode = "bedrock_converse" - else: - self.api_mode = "chat_completions" - - # Eagerly warm the transport cache so import errors surface at init, - # not mid-conversation. Also validates the api_mode is registered. - try: - self._get_transport() - except Exception: - pass # Non-fatal — transport may not exist for all modes yet - - try: - from hermes_cli.model_normalize import ( - _AGGREGATOR_PROVIDERS, - normalize_model_for_provider, - ) - - if self.provider not in _AGGREGATOR_PROVIDERS: - self.model = normalize_model_for_provider(self.model, self.provider) - except Exception: - pass - - # GPT-5.x models usually require the Responses API path, but some - # providers have exceptions (for example Copilot's gpt-5-mini still - # uses chat completions). Also auto-upgrade for direct OpenAI URLs - # (api.openai.com) since all newer tool-calling models prefer - # Responses there. ACP runtimes are excluded: CopilotACPClient - # handles its own routing and does not implement the Responses API - # surface. - # When api_mode was explicitly provided, respect it — the user - # knows what their endpoint supports (#10473). - # Exception: Azure OpenAI serves gpt-5.x on /chat/completions and - # does NOT support the Responses API — skip the upgrade for Azure - # (openai.azure.com), even though it looks OpenAI-compatible. - if ( - api_mode is None - and self.api_mode == "chat_completions" - and self.provider != "copilot-acp" - and not str(self.base_url or "").lower().startswith("acp://copilot") - and not str(self.base_url or "").lower().startswith("acp+tcp://") - and not self._is_azure_openai_url() - and ( - self._is_direct_openai_url() - or self._provider_model_requires_responses_api( - self.model, - provider=self.provider, - ) - ) - ): - self.api_mode = "codex_responses" - # Invalidate the eager-warmed transport cache — api_mode changed - # from chat_completions to codex_responses after the warm at __init__. - if hasattr(self, "_transport_cache"): - self._transport_cache.clear() - - # Pre-warm OpenRouter model metadata cache in a background thread. - # fetch_model_metadata() is cached for 1 hour; this avoids a blocking - # HTTP request on the first API response when pricing is estimated. - # Use a process-level Event so this thread is only spawned once — a new - # AIAgent is created for every gateway request, so without the guard - # each message leaks one OS thread and the process eventually exhausts - # the system thread limit (RuntimeError: can't start new thread). - if (self.provider == "openrouter" or self._is_openrouter_url()) and \ - not _openrouter_prewarm_done.is_set(): - _openrouter_prewarm_done.set() - threading.Thread( - target=fetch_model_metadata, - daemon=True, - name="openrouter-prewarm", - ).start() - - self.tool_progress_callback = tool_progress_callback - self.tool_start_callback = tool_start_callback - self.tool_complete_callback = tool_complete_callback - self.suppress_status_output = False - self.thinking_callback = thinking_callback - self.reasoning_callback = reasoning_callback - self.clarify_callback = clarify_callback - self.step_callback = step_callback - self.stream_delta_callback = stream_delta_callback - self.interim_assistant_callback = interim_assistant_callback - self.status_callback = status_callback - self.tool_gen_callback = tool_gen_callback - - - # Tool execution state — allows _vprint during tool execution - # even when stream consumers are registered (no tokens streaming then) - self._executing_tools = False - self._tool_guardrails = ToolCallGuardrailController() - self._tool_guardrail_halt_decision: ToolGuardrailDecision | None = None - - # Interrupt mechanism for breaking out of tool loops - self._interrupt_requested = False - self._interrupt_message = None # Optional message that triggered interrupt - self._execution_thread_id: int | None = None # Set at run_conversation() start - self._interrupt_thread_signal_pending = False - self._client_lock = threading.RLock() - - # /steer mechanism — inject a user note into the next tool result - # without interrupting the agent. Unlike interrupt(), steer() does - # NOT set _interrupt_requested; it waits for the current tool batch - # to finish naturally, then the drain hook appends the text to the - # last tool result's content so the model sees it on its next - # iteration. Message-role alternation is preserved (we modify an - # existing tool message rather than inserting a new user turn). - self._pending_steer: Optional[str] = None - self._pending_steer_lock = threading.Lock() - - # Concurrent-tool worker thread tracking. `_execute_tool_calls_concurrent` - # runs each tool on its own ThreadPoolExecutor worker — those worker - # threads have tids distinct from `_execution_thread_id`, so - # `_set_interrupt(True, _execution_thread_id)` alone does NOT cause - # `is_interrupted()` inside the worker to return True. Track the - # workers here so `interrupt()` / `clear_interrupt()` can fan out to - # their tids explicitly. - self._tool_worker_threads: set[int] = set() - self._tool_worker_threads_lock = threading.Lock() - - # Subagent delegation state - self._delegate_depth = 0 # 0 = top-level agent, incremented for children - self._active_children = [] # Running child AIAgents (for interrupt propagation) - self._active_children_lock = threading.Lock() - - # Store OpenRouter provider preferences - self.providers_allowed = providers_allowed - self.providers_ignored = providers_ignored - self.providers_order = providers_order - self.provider_sort = provider_sort - self.provider_require_parameters = provider_require_parameters - self.provider_data_collection = provider_data_collection - self.openrouter_min_coding_score = openrouter_min_coding_score - - # Store toolset filtering options - self.enabled_toolsets = enabled_toolsets - self.disabled_toolsets = disabled_toolsets - - # Model response configuration - self.max_tokens = max_tokens # None = use model default - self.reasoning_config = reasoning_config # None = use default (medium for OpenRouter) - self.service_tier = service_tier - self.request_overrides = dict(request_overrides or {}) - self.prefill_messages = prefill_messages or [] # Prefilled conversation turns - self._force_ascii_payload = False - - # Anthropic prompt caching: auto-enabled for Claude models on native - # Anthropic, OpenRouter, and third-party gateways that speak the - # Anthropic protocol (``api_mode == 'anthropic_messages'``). Reduces - # input costs by ~75% on multi-turn conversations. Uses system_and_3 - # strategy (4 breakpoints). See ``_anthropic_prompt_cache_policy`` - # for the layout-vs-transport decision. - self._use_prompt_caching, self._use_native_cache_layout = ( - self._anthropic_prompt_cache_policy() - ) - # Anthropic supports "5m" (default) and "1h" cache TTL tiers. Read from - # config.yaml under prompt_caching.cache_ttl; unknown values keep "5m". - # 1h tier costs 2x on write vs 1.25x for 5m, but amortizes across long - # sessions with >5-minute pauses between turns (#14971). - self._cache_ttl = "5m" - try: - from hermes_cli.config import load_config as _load_pc_cfg - - _pc_cfg = _load_pc_cfg().get("prompt_caching", {}) or {} - _ttl = _pc_cfg.get("cache_ttl", "5m") - if _ttl in {"5m", "1h"}: - self._cache_ttl = _ttl - except Exception: - pass - - # Iteration budget: the LLM is only notified when it actually exhausts - # the iteration budget (api_call_count >= max_iterations). At that - # point we inject ONE message, allow one final API call, and if the - # model doesn't produce a text response, force a user-message asking - # it to summarise. No intermediate pressure warnings — they caused - # models to "give up" prematurely on complex tasks (#7915). - self._budget_exhausted_injected = False - self._budget_grace_call = False - - # Activity tracking — updated on each API call, tool execution, and - # stream chunk. Used by the gateway timeout handler to report what the - # agent was doing when it was killed, and by the "still working" - # notifications to show progress. - self._last_activity_ts: float = time.time() - self._last_activity_desc: str = "initializing" - self._current_tool: str | None = None - self._api_call_count: int = 0 - - # Rate limit tracking — updated from x-ratelimit-* response headers - # after each API call. Accessed by /usage slash command. - self._rate_limit_state: Optional["RateLimitState"] = None - - # OpenRouter response cache hit counter — incremented when - # X-OpenRouter-Cache-Status: HIT is seen in streaming response headers. - self._or_cache_hits: int = 0 - - # Centralized logging — agent.log (INFO+) and errors.log (WARNING+) - # both live under ~/.hermes/logs/. Idempotent, so gateway mode - # (which creates a new AIAgent per message) won't duplicate handlers. - from hermes_logging import setup_logging, setup_verbose_logging - setup_logging(hermes_home=_hermes_home) - - if self.verbose_logging: - setup_verbose_logging() - logger.info("Verbose logging enabled (third-party library logs suppressed)") - elif self.quiet_mode: - # In quiet mode (CLI default), keep console output clean — - # but DO NOT raise per-logger levels. Doing so prevents the - # root logger's file handlers (agent.log, errors.log) from - # ever seeing the records, because Python checks - # logger.isEnabledFor() before handler propagation. We rely - # on the fact that hermes_logging.setup_logging() does not - # install a console StreamHandler in quiet mode — so INFO - # records flow to the file handlers but never reach a - # console. Any future noise reduction belongs at the - # handler level inside hermes_logging.py, not here. - pass - - # Internal stream callback (set during streaming TTS). - # Initialized here so _vprint can reference it before run_conversation. - self._stream_callback = None - # Deferred paragraph break flag — set after tool iterations so a - # single "\n\n" is prepended to the next real text delta. - self._stream_needs_break = False - # Stateful scrubber for spans split across stream - # deltas (#5719). sanitize_context() alone can't survive chunk - # boundaries because the block regex needs both tags in one string. - self._stream_context_scrubber = StreamingContextScrubber() - # Stateful scrubber for reasoning/thinking tags in streamed deltas - # (#17924). Replaces the per-delta _strip_think_blocks regex that - # destroyed downstream state (e.g. MiniMax-M2.7 streaming - # '' as delta1 and 'Let me check' as delta2 — the regex - # erased delta1, so downstream state machines never learned a - # block was open and leaked delta2 as content). - self._stream_think_scrubber = StreamingThinkScrubber() - # Visible assistant text already delivered through live token callbacks - # during the current model response. Used to avoid re-sending the same - # commentary when the provider later returns it as a completed interim - # assistant message. - self._current_streamed_assistant_text = "" - - # Optional current-turn user-message override used when the API-facing - # user message intentionally differs from the persisted transcript - # (e.g. CLI voice mode adds a temporary prefix for the live call only). - self._persist_user_message_idx = None - self._persist_user_message_override = None - - # Cache anthropic image-to-text fallbacks per image payload/URL so a - # single tool loop does not repeatedly re-run auxiliary vision on the - # same image history. - self._anthropic_image_fallback_cache: Dict[str, str] = {} - - # Initialize LLM client via centralized provider router. - # The router handles auth resolution, base URL, headers, and - # Codex/Anthropic wrapping for all known providers. - # raw_codex=True because the main agent needs direct responses.stream() - # access for Codex Responses API streaming. - self._anthropic_client = None - self._is_anthropic_oauth = False - - # Resolve per-provider / per-model request timeout once up front so - # every client construction path below (Anthropic native, OpenAI-wire, - # router-based implicit auth) can apply it consistently. Bedrock - # Claude uses its own timeout path and is not covered here. - _provider_timeout = get_provider_request_timeout(self.provider, self.model) - - if self.api_mode == "anthropic_messages": - from agent.anthropic_adapter import build_anthropic_client, resolve_anthropic_token - # Bedrock + Claude → use AnthropicBedrock SDK for full feature parity - # (prompt caching, thinking budgets, adaptive thinking). - _is_bedrock_anthropic = self.provider == "bedrock" - if _is_bedrock_anthropic: - from agent.anthropic_adapter import build_anthropic_bedrock_client - _region_match = re.search(r"bedrock-runtime\.([a-z0-9-]+)\.", base_url or "") - _br_region = _region_match.group(1) if _region_match else "us-east-1" - self._bedrock_region = _br_region - self._anthropic_client = build_anthropic_bedrock_client(_br_region) - self._anthropic_api_key = "aws-sdk" - self._anthropic_base_url = base_url - self._is_anthropic_oauth = False - self.api_key = "aws-sdk" - self.client = None - self._client_kwargs = {} - if not self.quiet_mode: - print(f"🤖 AI Agent initialized with model: {self.model} (AWS Bedrock + AnthropicBedrock SDK, {_br_region})") - else: - # Only fall back to ANTHROPIC_TOKEN when the provider is actually Anthropic. - # Other anthropic_messages providers (MiniMax, Alibaba, etc.) must use their own API key. - # Falling back would send Anthropic credentials to third-party endpoints (Fixes #1739, #minimax-401). - _is_native_anthropic = self.provider == "anthropic" - effective_key = (api_key or resolve_anthropic_token() or "") if _is_native_anthropic else (api_key or "") - self.api_key = effective_key - self._anthropic_api_key = effective_key - self._anthropic_base_url = base_url - # Only mark the session as OAuth-authenticated when the token - # genuinely belongs to native Anthropic. Third-party providers - # (MiniMax, Kimi, GLM, LiteLLM proxies) that accept the - # Anthropic protocol must never trip OAuth code paths — doing - # so injects Claude-Code identity headers and system prompts - # that cause 401/403 on their endpoints. Guards #1739 and - # the third-party identity-injection bug. - from agent.anthropic_adapter import _is_oauth_token as _is_oat - self._is_anthropic_oauth = _is_oat(effective_key) if _is_native_anthropic else False - self._anthropic_client = build_anthropic_client(effective_key, base_url, timeout=_provider_timeout) - # No OpenAI client needed for Anthropic mode - self.client = None - self._client_kwargs = {} - if not self.quiet_mode: - print(f"🤖 AI Agent initialized with model: {self.model} (Anthropic native)") - if effective_key and len(effective_key) > 12: - print(f"🔑 Using token: {effective_key[:8]}...{effective_key[-4:]}") - elif self.api_mode == "bedrock_converse": - # AWS Bedrock — uses boto3 directly, no OpenAI client needed. - # Region is extracted from the base_url or defaults to us-east-1. - _region_match = re.search(r"bedrock-runtime\.([a-z0-9-]+)\.", base_url or "") - self._bedrock_region = _region_match.group(1) if _region_match else "us-east-1" - # Guardrail config — read from config.yaml at init time. - self._bedrock_guardrail_config = None - try: - from hermes_cli.config import load_config as _load_br_cfg - _gr = _load_br_cfg().get("bedrock", {}).get("guardrail", {}) - if _gr.get("guardrail_identifier") and _gr.get("guardrail_version"): - self._bedrock_guardrail_config = { - "guardrailIdentifier": _gr["guardrail_identifier"], - "guardrailVersion": _gr["guardrail_version"], - } - if _gr.get("stream_processing_mode"): - self._bedrock_guardrail_config["streamProcessingMode"] = _gr["stream_processing_mode"] - if _gr.get("trace"): - self._bedrock_guardrail_config["trace"] = _gr["trace"] - except Exception: - pass - self.client = None - self._client_kwargs = {} - if not self.quiet_mode: - _gr_label = " + Guardrails" if self._bedrock_guardrail_config else "" - print(f"🤖 AI Agent initialized with model: {self.model} (AWS Bedrock, {self._bedrock_region}{_gr_label})") - else: - if api_key and base_url: - # Explicit credentials from CLI/gateway — construct directly. - # The runtime provider resolver already handled auth for us. - # Extract query params (e.g. Azure api-version) from base_url - # and pass via default_query to prevent loss during SDK URL - # joining (httpx drops query string when joining paths). - _parsed_url = urlparse(base_url) - if _parsed_url.query: - _clean_url = urlunparse(_parsed_url._replace(query="")) - _query_params = { - k: v[0] for k, v in parse_qs(_parsed_url.query).items() - } - client_kwargs = { - "api_key": api_key, - "base_url": _clean_url, - "default_query": _query_params, - } - else: - client_kwargs = {"api_key": api_key, "base_url": base_url} - if _provider_timeout is not None: - client_kwargs["timeout"] = _provider_timeout - if self.provider == "copilot-acp": - client_kwargs["command"] = self.acp_command - client_kwargs["args"] = self.acp_args - effective_base = base_url - if base_url_host_matches(effective_base, "openrouter.ai"): - from agent.auxiliary_client import build_or_headers - client_kwargs["default_headers"] = build_or_headers() - elif base_url_host_matches(effective_base, "integrate.api.nvidia.com"): - from agent.auxiliary_client import build_nvidia_nim_headers - client_kwargs["default_headers"] = build_nvidia_nim_headers(effective_base) - elif base_url_host_matches(effective_base, "api.routermint.com"): - client_kwargs["default_headers"] = _routermint_headers() - elif base_url_host_matches(effective_base, "api.githubcopilot.com"): - from hermes_cli.models import copilot_default_headers - - client_kwargs["default_headers"] = copilot_default_headers() - elif base_url_host_matches(effective_base, "api.kimi.com"): - client_kwargs["default_headers"] = { - "User-Agent": "claude-code/0.1.0", - } - elif base_url_host_matches(effective_base, "portal.qwen.ai"): - client_kwargs["default_headers"] = _qwen_portal_headers() - elif base_url_host_matches(effective_base, "chatgpt.com"): - from agent.auxiliary_client import _codex_cloudflare_headers - client_kwargs["default_headers"] = _codex_cloudflare_headers(api_key) - elif "default_headers" not in client_kwargs: - # Fall back to profile.default_headers for providers that - # declare custom headers (e.g. Vercel AI Gateway attribution, - # Kimi User-Agent on non-kimi.com endpoints). - try: - from providers import get_provider_profile as _gpf - _ph = _gpf(self.provider) - if _ph and _ph.default_headers: - client_kwargs["default_headers"] = dict(_ph.default_headers) - except Exception: - pass - else: - # No explicit creds — use the centralized provider router - from agent.auxiliary_client import resolve_provider_client - _routed_client, _ = resolve_provider_client( - self.provider or "auto", model=self.model, raw_codex=True) - if _routed_client is not None: - client_kwargs = { - "api_key": _routed_client.api_key, - "base_url": str(_routed_client.base_url), - } - if _provider_timeout is not None: - client_kwargs["timeout"] = _provider_timeout - # Preserve provider-specific headers the router set. The - # OpenAI SDK stores caller-provided default_headers in - # _custom_headers; older/mocked clients may expose - # _default_headers instead. - _routed_headers = getattr(_routed_client, "_custom_headers", None) - if not _routed_headers: - _routed_headers = getattr(_routed_client, "_default_headers", None) - if _routed_headers: - client_kwargs["default_headers"] = dict(_routed_headers) - else: - # When the user explicitly chose a non-OpenRouter provider - # but no credentials were found, fail fast with a clear - # message instead of silently routing through OpenRouter. - _explicit = (self.provider or "").strip().lower() - if _explicit and _explicit not in {"auto", "openrouter", "custom"}: - # Look up the actual env var name from the provider - # config — some providers use non-standard names - # (e.g. alibaba → DASHSCOPE_API_KEY, not ALIBABA_API_KEY). - _env_hint = f"{_explicit.upper()}_API_KEY" - try: - from hermes_cli.auth import PROVIDER_REGISTRY - _pcfg = PROVIDER_REGISTRY.get(_explicit) - if _pcfg and _pcfg.api_key_env_vars: - _env_hint = _pcfg.api_key_env_vars[0] - except Exception: - pass - # --- Init-time fallback (#17929) --- - _fb_entries = [] - if isinstance(fallback_model, list): - _fb_entries = [ - f for f in fallback_model - if isinstance(f, dict) and f.get("provider") and f.get("model") - ] - elif isinstance(fallback_model, dict) and fallback_model.get("provider") and fallback_model.get("model"): - _fb_entries = [fallback_model] - _fb_resolved = False - for _fb in _fb_entries: - _fb_explicit_key = (_fb.get("api_key") or "").strip() or None - if not _fb_explicit_key: - _fb_key_env = (_fb.get("key_env") or _fb.get("api_key_env") or "").strip() - if _fb_key_env: - _fb_explicit_key = os.getenv(_fb_key_env, "").strip() or None - _fb_client, _fb_model = resolve_provider_client( - _fb["provider"], model=_fb["model"], raw_codex=True, - explicit_base_url=_fb.get("base_url"), - explicit_api_key=_fb_explicit_key, - ) - if _fb_client is not None: - self.provider = _fb["provider"] - self.model = _fb_model or _fb["model"] - self._fallback_activated = True - client_kwargs = { - "api_key": _fb_client.api_key, - "base_url": str(_fb_client.base_url), - } - if _provider_timeout is not None: - client_kwargs["timeout"] = _provider_timeout - _fb_headers = getattr(_fb_client, "_custom_headers", None) - if not _fb_headers: - _fb_headers = getattr(_fb_client, "_default_headers", None) - if _fb_headers: - client_kwargs["default_headers"] = dict(_fb_headers) - _fb_resolved = True - break - if not _fb_resolved: - raise RuntimeError( - f"Provider '{_explicit}' is set in config.yaml but no API key " - f"was found. Set the {_env_hint} environment " - f"variable, or switch to a different provider with `hermes model`." - ) - if not getattr(self, "_fallback_activated", False): - # No provider configured — reject with a clear message. - raise RuntimeError( - "No LLM provider configured. Run `hermes model` to " - "select a provider, or run `hermes setup` for first-time " - "configuration." - ) - - self._client_kwargs = client_kwargs # stored for rebuilding after interrupt - - # Enable fine-grained tool streaming for Claude on OpenRouter. - # Without this, Anthropic buffers the entire tool call and goes - # silent for minutes while thinking — OpenRouter's upstream proxy - # times out during the silence. The beta header makes Anthropic - # stream tool call arguments token-by-token, keeping the - # connection alive. - _effective_base = str(client_kwargs.get("base_url", "")).lower() - if base_url_host_matches(_effective_base, "openrouter.ai") and "claude" in (self.model or "").lower(): - headers = client_kwargs.get("default_headers") or {} - existing_beta = headers.get("x-anthropic-beta", "") - _FINE_GRAINED = "fine-grained-tool-streaming-2025-05-14" - if _FINE_GRAINED not in existing_beta: - if existing_beta: - headers["x-anthropic-beta"] = f"{existing_beta},{_FINE_GRAINED}" - else: - headers["x-anthropic-beta"] = _FINE_GRAINED - client_kwargs["default_headers"] = headers - - self.api_key = client_kwargs.get("api_key", "") - self.base_url = client_kwargs.get("base_url", self.base_url) - try: - self.client = self._create_openai_client(client_kwargs, reason="agent_init", shared=True) - if not self.quiet_mode: - print(f"🤖 AI Agent initialized with model: {self.model}") - if base_url: - print(f"🔗 Using custom base URL: {base_url}") - # Always show API key info (masked) for debugging auth issues - key_used = client_kwargs.get("api_key", "none") - if key_used and key_used != "dummy-key" and len(key_used) > 12: - print(f"🔑 Using API key: {key_used[:8]}...{key_used[-4:]}") - else: - print(f"⚠️ Warning: API key appears invalid or missing (got: '{key_used[:20] if key_used else 'none'}...')") - except Exception as e: - raise RuntimeError(f"Failed to initialize OpenAI client: {e}") - - # Provider fallback chain — ordered list of backup providers tried - # when the primary is exhausted (rate-limit, overload, connection - # failure). Supports both legacy single-dict ``fallback_model`` and - # new list ``fallback_providers`` format. - if isinstance(fallback_model, list): - self._fallback_chain = [ - f for f in fallback_model - if isinstance(f, dict) and f.get("provider") and f.get("model") - ] - elif isinstance(fallback_model, dict) and fallback_model.get("provider") and fallback_model.get("model"): - self._fallback_chain = [fallback_model] - else: - self._fallback_chain = [] - self._fallback_index = 0 - self._fallback_activated = getattr(self, "_fallback_activated", False) - # Legacy attribute kept for backward compat (tests, external callers) - self._fallback_model = self._fallback_chain[0] if self._fallback_chain else None - if self._fallback_chain and not self.quiet_mode: - if len(self._fallback_chain) == 1: - fb = self._fallback_chain[0] - print(f"🔄 Fallback model: {fb['model']} ({fb['provider']})") - else: - print(f"🔄 Fallback chain ({len(self._fallback_chain)} providers): " + - " → ".join(f"{f['model']} ({f['provider']})" for f in self._fallback_chain)) - - # Get available tools with filtering - self.tools = get_tool_definitions( + """Forwarder — see ``agent.agent_init.init_agent``.""" + from agent.agent_init import init_agent + init_agent( + self, + base_url=base_url, + api_key=api_key, + provider=provider, + api_mode=api_mode, + acp_command=acp_command, + acp_args=acp_args, + command=command, + args=args, + model=model, + max_iterations=max_iterations, + tool_delay=tool_delay, enabled_toolsets=enabled_toolsets, disabled_toolsets=disabled_toolsets, - quiet_mode=self.quiet_mode, + save_trajectories=save_trajectories, + verbose_logging=verbose_logging, + quiet_mode=quiet_mode, + ephemeral_system_prompt=ephemeral_system_prompt, + log_prefix_chars=log_prefix_chars, + log_prefix=log_prefix, + providers_allowed=providers_allowed, + providers_ignored=providers_ignored, + providers_order=providers_order, + provider_sort=provider_sort, + provider_require_parameters=provider_require_parameters, + provider_data_collection=provider_data_collection, + openrouter_min_coding_score=openrouter_min_coding_score, + session_id=session_id, + tool_progress_callback=tool_progress_callback, + tool_start_callback=tool_start_callback, + tool_complete_callback=tool_complete_callback, + thinking_callback=thinking_callback, + reasoning_callback=reasoning_callback, + clarify_callback=clarify_callback, + step_callback=step_callback, + stream_delta_callback=stream_delta_callback, + interim_assistant_callback=interim_assistant_callback, + tool_gen_callback=tool_gen_callback, + status_callback=status_callback, + max_tokens=max_tokens, + reasoning_config=reasoning_config, + service_tier=service_tier, + request_overrides=request_overrides, + prefill_messages=prefill_messages, + platform=platform, + user_id=user_id, + user_name=user_name, + chat_id=chat_id, + chat_name=chat_name, + chat_type=chat_type, + thread_id=thread_id, + gateway_session_key=gateway_session_key, + skip_context_files=skip_context_files, + load_soul_identity=load_soul_identity, + skip_memory=skip_memory, + session_db=session_db, + parent_session_id=parent_session_id, + iteration_budget=iteration_budget, + fallback_model=fallback_model, + credential_pool=credential_pool, + checkpoints_enabled=checkpoints_enabled, + checkpoint_max_snapshots=checkpoint_max_snapshots, + checkpoint_max_total_size_mb=checkpoint_max_total_size_mb, + checkpoint_max_file_size_mb=checkpoint_max_file_size_mb, + pass_session_id=pass_session_id, ) - - # Show tool configuration and store valid tool names for validation - self.valid_tool_names = set() - if self.tools: - self.valid_tool_names = {tool["function"]["name"] for tool in self.tools} - tool_names = sorted(self.valid_tool_names) - if not self.quiet_mode: - print(f"🛠️ Loaded {len(self.tools)} tools: {', '.join(tool_names)}") - - # Show filtering info if applied - if enabled_toolsets: - print(f" ✅ Enabled toolsets: {', '.join(enabled_toolsets)}") - if disabled_toolsets: - print(f" ❌ Disabled toolsets: {', '.join(disabled_toolsets)}") - elif not self.quiet_mode: - print("🛠️ No tools loaded (all tools filtered out or unavailable)") - - # Check tool requirements - if self.tools and not self.quiet_mode: - requirements = check_toolset_requirements() - missing_reqs = [name for name, available in requirements.items() if not available] - if missing_reqs: - print(f"⚠️ Some tools may not work due to missing requirements: {missing_reqs}") - - # Show trajectory saving status - if self.save_trajectories and not self.quiet_mode: - print("📝 Trajectory saving enabled") - - # Show ephemeral system prompt status - if self.ephemeral_system_prompt and not self.quiet_mode: - prompt_preview = self.ephemeral_system_prompt[:60] + "..." if len(self.ephemeral_system_prompt) > 60 else self.ephemeral_system_prompt - print(f"🔒 Ephemeral system prompt: '{prompt_preview}' (not saved to trajectories)") - - # Show prompt caching status - if self._use_prompt_caching and not self.quiet_mode: - if self._use_native_cache_layout and self.provider == "anthropic": - source = "native Anthropic" - elif self._use_native_cache_layout: - source = "Anthropic-compatible endpoint" - else: - source = "Claude via OpenRouter" - print(f"💾 Prompt caching: ENABLED ({source}, {self._cache_ttl} TTL)") - - # Session logging setup - auto-save conversation trajectories for debugging - self.session_start = datetime.now() - if session_id: - # Use provided session ID (e.g., from CLI) - self.session_id = session_id - else: - # Generate a new session ID - timestamp_str = self.session_start.strftime("%Y%m%d_%H%M%S") - short_uuid = uuid.uuid4().hex[:6] - self.session_id = f"{timestamp_str}_{short_uuid}" - - # Expose session ID to tools (terminal, execute_code) so agents can - # reference their own session for --resume commands, cross-session - # coordination, and logging. Uses the ContextVar system from - # session_context.py for concurrency safety (gateway runs multiple - # sessions in one process). Also writes os.environ as fallback for - # CLI mode where ContextVars aren't used. - os.environ["HERMES_SESSION_ID"] = self.session_id - try: - from gateway.session_context import _SESSION_ID - _SESSION_ID.set(self.session_id) - except Exception: - pass # CLI/test mode — ContextVar not needed - - # Session logs go into ~/.hermes/sessions/ alongside gateway sessions - hermes_home = get_hermes_home() - self.logs_dir = hermes_home / "sessions" - self.logs_dir.mkdir(parents=True, exist_ok=True) - self.session_log_file = self.logs_dir / f"session_{self.session_id}.json" - - # Track conversation messages for session logging - self._session_messages: List[Dict[str, Any]] = [] - self._memory_write_origin = "assistant_tool" - self._memory_write_context = "foreground" - - # Cached system prompt -- built once per session, only rebuilt on compression - self._cached_system_prompt: Optional[str] = None - - # Filesystem checkpoint manager (transparent — not a tool) - from tools.checkpoint_manager import CheckpointManager - self._checkpoint_mgr = CheckpointManager( - enabled=checkpoints_enabled, - max_snapshots=checkpoint_max_snapshots, - max_total_size_mb=checkpoint_max_total_size_mb, - max_file_size_mb=checkpoint_max_file_size_mb, - ) - - # SQLite session store (optional -- provided by CLI or gateway) - self._session_db = session_db - self._parent_session_id = parent_session_id - self._last_flushed_db_idx = 0 # tracks DB-write cursor to prevent duplicate writes - self._session_db_created = False # DB row deferred to run_conversation() - self._session_init_model_config = { - "max_iterations": self.max_iterations, - "reasoning_config": reasoning_config, - "max_tokens": max_tokens, - } - - # In-memory todo list for task planning (one per agent/session) - from tools.todo_tool import TodoStore - self._todo_store = TodoStore() - - # Load config once for memory, skills, and compression sections - try: - from hermes_cli.config import load_config as _load_agent_config - _agent_cfg = _load_agent_config() - except Exception: - _agent_cfg = {} - try: - self._tool_guardrails = ToolCallGuardrailController( - ToolCallGuardrailConfig.from_mapping( - _agent_cfg.get("tool_loop_guardrails", {}) - ) - ) - except Exception as _tlg_err: - logger.warning("Tool loop guardrail config ignored: %s", _tlg_err) - # Cache only the derived auxiliary compression context override that is - # needed later by the startup feasibility check. Avoid exposing a - # broad pseudo-public config object on the agent instance. - self._aux_compression_context_length_config = None - - # Persistent memory (MEMORY.md + USER.md) -- loaded from disk - self._memory_store = None - self._memory_enabled = False - self._user_profile_enabled = False - self._memory_nudge_interval = 10 - self._turns_since_memory = 0 - self._iters_since_skill = 0 - if not skip_memory: - try: - mem_config = _agent_cfg.get("memory", {}) - self._memory_enabled = mem_config.get("memory_enabled", False) - self._user_profile_enabled = mem_config.get("user_profile_enabled", False) - self._memory_nudge_interval = int(mem_config.get("nudge_interval", 10)) - if self._memory_enabled or self._user_profile_enabled: - from tools.memory_tool import MemoryStore - self._memory_store = MemoryStore( - memory_char_limit=mem_config.get("memory_char_limit", 2200), - user_char_limit=mem_config.get("user_char_limit", 1375), - ) - self._memory_store.load_from_disk() - except Exception: - pass # Memory is optional -- don't break agent init - - - - # Memory provider plugin (external — one at a time, alongside built-in) - # Reads memory.provider from config to select which plugin to activate. - self._memory_manager = None - if not skip_memory: - try: - _mem_provider_name = mem_config.get("provider", "") if mem_config else "" - - if _mem_provider_name and _mem_provider_name.strip(): - from agent.memory_manager import MemoryManager as _MemoryManager - from plugins.memory import load_memory_provider as _load_mem - self._memory_manager = _MemoryManager() - _mp = _load_mem(_mem_provider_name) - if _mp and _mp.is_available(): - self._memory_manager.add_provider(_mp) - if self._memory_manager.providers: - _init_kwargs = { - "session_id": self.session_id, - "platform": platform or "cli", - "hermes_home": str(get_hermes_home()), - "agent_context": "primary", - } - # Thread session title for memory provider scoping - # (e.g. honcho uses this to derive chat-scoped session keys) - if self._session_db: - try: - _st = self._session_db.get_session_title(self.session_id) - if _st: - _init_kwargs["session_title"] = _st - except Exception: - pass - # Thread gateway user identity for per-user memory scoping - if self._user_id: - _init_kwargs["user_id"] = self._user_id - if self._user_name: - _init_kwargs["user_name"] = self._user_name - if self._chat_id: - _init_kwargs["chat_id"] = self._chat_id - if self._chat_name: - _init_kwargs["chat_name"] = self._chat_name - if self._chat_type: - _init_kwargs["chat_type"] = self._chat_type - if self._thread_id: - _init_kwargs["thread_id"] = self._thread_id - # Thread gateway session key for stable per-chat Honcho session isolation - if self._gateway_session_key: - _init_kwargs["gateway_session_key"] = self._gateway_session_key - # Profile identity for per-profile provider scoping - try: - from hermes_cli.profiles import get_active_profile_name - _profile = get_active_profile_name() - _init_kwargs["agent_identity"] = _profile - _init_kwargs["agent_workspace"] = "hermes" - except Exception: - pass - self._memory_manager.initialize_all(**_init_kwargs) - logger.info("Memory provider '%s' activated", _mem_provider_name) - else: - logger.debug("Memory provider '%s' not found or not available", _mem_provider_name) - self._memory_manager = None - except Exception as _mpe: - logger.warning("Memory provider plugin init failed: %s", _mpe) - self._memory_manager = None - - # Inject memory provider tool schemas into the tool surface. - # Skip tools whose names already exist (plugins may register the - # same tools via ctx.register_tool(), which lands in self.tools - # through get_tool_definitions()). Duplicate function names cause - # 400 errors on providers that enforce unique names (e.g. Xiaomi - # MiMo via Nous Portal). - if self._memory_manager and self.tools is not None: - _existing_tool_names = { - t.get("function", {}).get("name") - for t in self.tools - if isinstance(t, dict) - } - for _schema in self._memory_manager.get_all_tool_schemas(): - _tname = _schema.get("name", "") - if _tname and _tname in _existing_tool_names: - continue # already registered via plugin path - _wrapped = {"type": "function", "function": _schema} - self.tools.append(_wrapped) - if _tname: - self.valid_tool_names.add(_tname) - _existing_tool_names.add(_tname) - - # Skills config: nudge interval for skill creation reminders - self._skill_nudge_interval = 10 - try: - skills_config = _agent_cfg.get("skills", {}) - self._skill_nudge_interval = int(skills_config.get("creation_nudge_interval", 10)) - except Exception: - pass - - # Tool-use enforcement config: "auto" (default — matches hardcoded - # model list), true (always), false (never), or list of substrings. - _agent_section = _agent_cfg.get("agent", {}) - if not isinstance(_agent_section, dict): - _agent_section = {} - self._tool_use_enforcement = _agent_section.get("tool_use_enforcement", "auto") - - # App-level API retry count (wraps each model API call). Default 3, - # overridable via agent.api_max_retries in config.yaml. See #11616. - try: - _raw_api_retries = _agent_section.get("api_max_retries", 3) - _api_retries = int(_raw_api_retries) - _api_retries = max(_api_retries, 1) # 1 = no retry (single attempt) - except (TypeError, ValueError): - _api_retries = 3 - self._api_max_retries = _api_retries - - # Initialize context compressor for automatic context management - # Compresses conversation when approaching model's context limit - # Configuration via config.yaml (compression section) - _compression_cfg = _agent_cfg.get("compression", {}) - if not isinstance(_compression_cfg, dict): - _compression_cfg = {} - compression_threshold = float(_compression_cfg.get("threshold", 0.50)) - try: - from agent.auxiliary_client import _compression_threshold_for_model as _cthresh_fn - _model_cthresh = _cthresh_fn(self.model) - if _model_cthresh is not None: - compression_threshold = _model_cthresh - except Exception: - pass - compression_enabled = str(_compression_cfg.get("enabled", True)).lower() in {"true", "1", "yes"} - compression_target_ratio = float(_compression_cfg.get("target_ratio", 0.20)) - compression_protect_last = int(_compression_cfg.get("protect_last_n", 20)) - # protect_first_n is the number of non-system messages to protect at - # the head, in addition to the system prompt (which is always - # implicitly protected by the compressor). Floor at 0 — a value of - # 0 means "preserve only the system prompt + summary + tail", which - # is a legitimate (and common) configuration for long-running - # rolling-compaction sessions. - compression_protect_first = max( - 0, int(_compression_cfg.get("protect_first_n", 3)) - ) - - # Read optional explicit context_length override for the auxiliary - # compression model. Custom endpoints often cannot report this via - # /models, so the startup feasibility check needs the config hint. - try: - _aux_cfg = cfg_get(_agent_cfg, "auxiliary", "compression", default={}) - except Exception: - _aux_cfg = {} - if isinstance(_aux_cfg, dict): - _aux_context_config = _aux_cfg.get("context_length") - else: - _aux_context_config = None - if _aux_context_config is not None: - try: - _aux_context_config = int(_aux_context_config) - except (TypeError, ValueError): - _aux_context_config = None - self._aux_compression_context_length_config = _aux_context_config - - # Read explicit model output-token override from config when the - # caller did not pass one directly. - _model_cfg = _agent_cfg.get("model", {}) - if self.max_tokens is None and isinstance(_model_cfg, dict): - _config_max_tokens = _model_cfg.get("max_tokens") - if _config_max_tokens is not None: - try: - if isinstance(_config_max_tokens, bool): - raise ValueError - _parsed_max_tokens = int(_config_max_tokens) - if _parsed_max_tokens <= 0: - raise ValueError - self.max_tokens = _parsed_max_tokens - except (TypeError, ValueError): - logger.warning( - "Invalid model.max_tokens in config.yaml: %r — " - "must be a positive integer (e.g. 4096). " - "Falling back to provider default.", - _config_max_tokens, - ) - print( - f"\n⚠ Invalid model.max_tokens in config.yaml: {_config_max_tokens!r}\n" - f" Must be a positive integer (e.g. 4096).\n" - f" Falling back to provider default.\n", - file=sys.stderr, - ) - self._session_init_model_config["max_tokens"] = self.max_tokens - - # Read explicit context_length override from model config - if isinstance(_model_cfg, dict): - _config_context_length = _model_cfg.get("context_length") - else: - _config_context_length = None - if _config_context_length is not None: - try: - _config_context_length = int(_config_context_length) - except (TypeError, ValueError): - logger.warning( - "Invalid model.context_length in config.yaml: %r — " - "must be a plain integer (e.g. 256000, not '256K'). " - "Falling back to auto-detection.", - _config_context_length, - ) - print( - f"\n⚠ Invalid model.context_length in config.yaml: {_config_context_length!r}\n" - f" Must be a plain integer (e.g. 256000, not '256K').\n" - f" Falling back to auto-detected context window.\n", - file=sys.stderr, - ) - _config_context_length = None - - # Resolve custom_providers list once for reuse below (startup - # context-length override and plugin context-engine init). - try: - from hermes_cli.config import get_compatible_custom_providers - _custom_providers = get_compatible_custom_providers(_agent_cfg) - except Exception: - _custom_providers = _agent_cfg.get("custom_providers") - if not isinstance(_custom_providers, list): - _custom_providers = [] - - # Store for reuse by _check_compression_model_feasibility (auxiliary - # compression model context-length detection needs the same list). - self._custom_providers = _custom_providers - - # Check custom_providers per-model context_length - if _config_context_length is None and _custom_providers: - try: - from hermes_cli.config import get_custom_provider_context_length - _cp_ctx_resolved = get_custom_provider_context_length( - model=self.model, - base_url=self.base_url, - custom_providers=_custom_providers, - ) - if _cp_ctx_resolved: - _config_context_length = int(_cp_ctx_resolved) - except Exception: - _cp_ctx_resolved = None - - # Surface a clear warning if the user set a context_length but it - # wasn't a valid positive int — the helper silently skips those. - if _config_context_length is None: - _target = self.base_url.rstrip("/") if self.base_url else "" - for _cp_entry in _custom_providers: - if not isinstance(_cp_entry, dict): - continue - _cp_url = (_cp_entry.get("base_url") or "").rstrip("/") - if _target and _cp_url == _target: - _cp_models = _cp_entry.get("models", {}) - if isinstance(_cp_models, dict): - _cp_model_cfg = _cp_models.get(self.model, {}) - if isinstance(_cp_model_cfg, dict): - _cp_ctx = _cp_model_cfg.get("context_length") - if _cp_ctx is not None: - try: - _parsed = int(_cp_ctx) - if _parsed <= 0: - raise ValueError - except (TypeError, ValueError): - logger.warning( - "Invalid context_length for model %r in " - "custom_providers: %r — must be a positive " - "integer (e.g. 256000, not '256K'). " - "Falling back to auto-detection.", - self.model, _cp_ctx, - ) - print( - f"\n⚠ Invalid context_length for model {self.model!r} in custom_providers: {_cp_ctx!r}\n" - f" Must be a positive integer (e.g. 256000, not '256K').\n" - f" Falling back to auto-detected context window.\n", - file=sys.stderr, - ) - break - - # Persist for reuse on switch_model / fallback activation. Must come - # AFTER the custom_providers branch so per-model overrides aren't lost. - self._config_context_length = _config_context_length - - self._ensure_lmstudio_runtime_loaded(_config_context_length) - - - - # Select context engine: config-driven (like memory providers). - # 1. Check config.yaml context.engine setting - # 2. Check plugins/context_engine// directory (repo-shipped) - # 3. Check general plugin system (user-installed plugins) - # 4. Fall back to built-in ContextCompressor - _selected_engine = None - _engine_name = "compressor" # default - try: - _ctx_cfg = _agent_cfg.get("context", {}) if isinstance(_agent_cfg, dict) else {} - _engine_name = _ctx_cfg.get("engine", "compressor") or "compressor" - except Exception: - pass - - if _engine_name != "compressor": - # Try loading from plugins/context_engine// - try: - from plugins.context_engine import load_context_engine - _selected_engine = load_context_engine(_engine_name) - except Exception as _ce_load_err: - logger.debug("Context engine load from plugins/context_engine/: %s", _ce_load_err) - - # Try general plugin system as fallback - if _selected_engine is None: - try: - from hermes_cli.plugins import get_plugin_context_engine - _candidate = get_plugin_context_engine() - if _candidate and _candidate.name == _engine_name: - _selected_engine = _candidate - except Exception: - pass - - if _selected_engine is None: - logger.warning( - "Context engine '%s' not found — falling back to built-in compressor", - _engine_name, - ) - # else: config says "compressor" — use built-in, don't auto-activate plugins - - if _selected_engine is not None: - self.context_compressor = _selected_engine - # Resolve context_length for plugin engines — mirrors switch_model() path - from agent.model_metadata import get_model_context_length - _plugin_ctx_len = get_model_context_length( - self.model, - base_url=self.base_url, - api_key=getattr(self, "api_key", ""), - config_context_length=_config_context_length, - provider=self.provider, - custom_providers=_custom_providers, - ) - self.context_compressor.update_model( - model=self.model, - context_length=_plugin_ctx_len, - base_url=self.base_url, - api_key=getattr(self, "api_key", ""), - provider=self.provider, - ) - if not self.quiet_mode: - logger.info("Using context engine: %s", _selected_engine.name) - else: - self.context_compressor = ContextCompressor( - model=self.model, - threshold_percent=compression_threshold, - protect_first_n=compression_protect_first, - protect_last_n=compression_protect_last, - summary_target_ratio=compression_target_ratio, - summary_model_override=None, - quiet_mode=self.quiet_mode, - base_url=self.base_url, - api_key=getattr(self, "api_key", ""), - config_context_length=_config_context_length, - provider=self.provider, - api_mode=self.api_mode, - ) - self.compression_enabled = compression_enabled - - # Reject models whose context window is below the minimum required - # for reliable tool-calling workflows (64K tokens). - from agent.model_metadata import MINIMUM_CONTEXT_LENGTH - _ctx = getattr(self.context_compressor, "context_length", 0) - if _ctx and _ctx < MINIMUM_CONTEXT_LENGTH: - raise ValueError( - f"Model {self.model} has a context window of {_ctx:,} tokens, " - f"which is below the minimum {MINIMUM_CONTEXT_LENGTH:,} required " - f"by Hermes Agent. Choose a model with at least " - f"{MINIMUM_CONTEXT_LENGTH // 1000}K context, or set " - f"model.context_length in config.yaml to override." - ) - - # Inject context engine tool schemas (e.g. lcm_grep, lcm_describe, lcm_expand). - # Skip names that are already present — the get_tool_definitions() - # quiet_mode cache returned a shared list pre-#17335, so a stray - # mutation here would poison subsequent agent inits in the same - # Gateway process and trip provider-side 'duplicate tool name' - # errors. Even with the cache fix, dedup is the right defense - # against plugin paths that may register the same schemas via - # ctx.register_tool(). Mirrors the memory tools dedup above. - self._context_engine_tool_names: set = set() - if hasattr(self, "context_compressor") and self.context_compressor and self.tools is not None: - _existing_tool_names = { - t.get("function", {}).get("name") - for t in self.tools - if isinstance(t, dict) - } - for _schema in self.context_compressor.get_tool_schemas(): - _tname = _schema.get("name", "") - if _tname and _tname in _existing_tool_names: - continue # already registered via plugin/cache path - _wrapped = {"type": "function", "function": _schema} - self.tools.append(_wrapped) - if _tname: - self.valid_tool_names.add(_tname) - self._context_engine_tool_names.add(_tname) - _existing_tool_names.add(_tname) - - # Notify context engine of session start - if hasattr(self, "context_compressor") and self.context_compressor: - try: - self.context_compressor.on_session_start( - self.session_id, - hermes_home=str(get_hermes_home()), - platform=self.platform or "cli", - model=self.model, - context_length=getattr(self.context_compressor, "context_length", 0), - ) - except Exception as _ce_err: - logger.debug("Context engine on_session_start: %s", _ce_err) - - self._subdirectory_hints = SubdirectoryHintTracker( - working_dir=os.getenv("TERMINAL_CWD") or None, - ) - self._user_turn_count = 0 - - # Cumulative token usage for the session - self.session_prompt_tokens = 0 - self.session_completion_tokens = 0 - self.session_total_tokens = 0 - self.session_api_calls = 0 - self.session_input_tokens = 0 - self.session_output_tokens = 0 - self.session_cache_read_tokens = 0 - self.session_cache_write_tokens = 0 - self.session_reasoning_tokens = 0 - self.session_estimated_cost_usd = 0.0 - self.session_cost_status = "unknown" - self.session_cost_source = "none" - - # ── Ollama num_ctx injection ── - # Ollama defaults to 2048 context regardless of the model's capabilities. - # When running against an Ollama server, detect the model's max context - # and pass num_ctx on every chat request so the full window is used. - # User override: set model.ollama_num_ctx in config.yaml to cap VRAM use. - # If model.context_length is set, it caps num_ctx so the user's VRAM - # budget is respected even when GGUF metadata advertises a larger window. - self._ollama_num_ctx: int | None = None - _ollama_num_ctx_override = None - if isinstance(_model_cfg, dict): - _ollama_num_ctx_override = _model_cfg.get("ollama_num_ctx") - if _ollama_num_ctx_override is not None: - try: - self._ollama_num_ctx = int(_ollama_num_ctx_override) - except (TypeError, ValueError): - logger.debug("Invalid ollama_num_ctx config value: %r", _ollama_num_ctx_override) - if self._ollama_num_ctx is None and self.base_url and is_local_endpoint(self.base_url): - try: - _detected = query_ollama_num_ctx(self.model, self.base_url, api_key=self.api_key or "") - if _detected and _detected > 0: - self._ollama_num_ctx = _detected - except Exception as exc: - logger.debug("Ollama num_ctx detection failed: %s", exc) - # Cap auto-detected ollama_num_ctx to the user's explicit context_length. - # Without this, GGUF metadata can advertise 256K+ which Ollama honours - # by allocating that much VRAM — blowing up small GPUs even though the - # user explicitly set a smaller context_length in config.yaml. - if ( - self._ollama_num_ctx - and _config_context_length - and _ollama_num_ctx_override is None # don't override explicit ollama_num_ctx - and self._ollama_num_ctx > _config_context_length - ): - logger.info( - "Ollama num_ctx capped: %d -> %d (model.context_length override)", - self._ollama_num_ctx, _config_context_length, - ) - self._ollama_num_ctx = _config_context_length - if self._ollama_num_ctx and not self.quiet_mode: - logger.info( - "Ollama num_ctx: will request %d tokens (model max from /api/show)", - self._ollama_num_ctx, - ) - - if not self.quiet_mode: - if compression_enabled: - print(f"📊 Context limit: {self.context_compressor.context_length:,} tokens (compress at {int(compression_threshold*100)}% = {self.context_compressor.threshold_tokens:,})") - else: - print(f"📊 Context limit: {self.context_compressor.context_length:,} tokens (auto-compression disabled)") - - # Check immediately so CLI users see the warning at startup. - # Gateway status_callback is not yet wired, so any warning is stored - # in _compression_warning and replayed in the first run_conversation(). - self._compression_warning = None - self._check_compression_model_feasibility() - - # Snapshot primary runtime for per-turn restoration. When fallback - # activates during a turn, the next turn restores these values so the - # preferred model gets a fresh attempt each time. Uses a single dict - # so new state fields are easy to add without N individual attributes. - _cc = self.context_compressor - self._primary_runtime = { - "model": self.model, - "provider": self.provider, - "base_url": self.base_url, - "api_mode": self.api_mode, - "api_key": getattr(self, "api_key", ""), - "client_kwargs": dict(self._client_kwargs), - "use_prompt_caching": self._use_prompt_caching, - "use_native_cache_layout": self._use_native_cache_layout, - # Context engine state that _try_activate_fallback() overwrites. - # Use getattr for model/base_url/api_key/provider since plugin - # engines may not have these (they're ContextCompressor-specific). - "compressor_model": getattr(_cc, "model", self.model), - "compressor_base_url": getattr(_cc, "base_url", self.base_url), - "compressor_api_key": getattr(_cc, "api_key", ""), - "compressor_provider": getattr(_cc, "provider", self.provider), - "compressor_context_length": _cc.context_length, - "compressor_threshold_tokens": _cc.threshold_tokens, - } - if self.api_mode == "anthropic_messages": - self._primary_runtime.update({ - "anthropic_api_key": self._anthropic_api_key, - "anthropic_base_url": self._anthropic_base_url, - "is_anthropic_oauth": self._is_anthropic_oauth, - }) def _get_session_db_for_recall(self): """Return a SessionDB for recall, lazily creating it if an entrypoint forgot. @@ -2640,198 +558,9 @@ class AIAgent: logger.debug("LM Studio preload skipped: %s", err) def switch_model(self, new_model, new_provider, api_key='', base_url='', api_mode=''): - """Switch the model/provider in-place for a live agent. - - Called by the /model command handlers (CLI and gateway) after - ``model_switch.switch_model()`` has resolved credentials and - validated the model. This method performs the actual runtime - swap: rebuilding clients, updating caching flags, and refreshing - the context compressor. - - The implementation mirrors ``_try_activate_fallback()`` for the - client-swap logic but also updates ``_primary_runtime`` so the - change persists across turns (unlike fallback which is - turn-scoped). - """ - from hermes_cli.providers import determine_api_mode - - # ── Determine api_mode if not provided ── - if not api_mode: - api_mode = determine_api_mode(new_provider, base_url) - - # Defense-in-depth: ensure OpenCode base_url doesn't carry a trailing - # /v1 into the anthropic_messages client, which would cause the SDK to - # hit /v1/v1/messages. `model_switch.switch_model()` already strips - # this, but we guard here so any direct callers (future code paths, - # tests) can't reintroduce the double-/v1 404 bug. - if ( - api_mode == "anthropic_messages" - and new_provider in {"opencode-zen", "opencode-go"} - and isinstance(base_url, str) - and base_url - ): - base_url = re.sub(r"/v1/?$", "", base_url) - - old_model = self.model - old_provider = self.provider - - # Clear the per-config context_length override so the new model's - # actual context window is resolved via get_model_context_length() - # instead of inheriting the stale value from the previous model. - self._config_context_length = None - - # ── Swap core runtime fields ── - self.model = new_model - self.provider = new_provider - # Use new base_url when provided; only fall back to current when the - # new provider genuinely has no endpoint (e.g. native SDK providers). - # Without this guard the old provider's URL (e.g. Ollama's localhost - # address) would persist silently after switching to a cloud provider - # that returns an empty base_url string. - if base_url: - self.base_url = base_url - self.api_mode = api_mode - # Invalidate transport cache — new api_mode may need a different transport - if hasattr(self, "_transport_cache"): - self._transport_cache.clear() - if api_key: - self.api_key = api_key - - # ── Build new client ── - if api_mode == "anthropic_messages": - from agent.anthropic_adapter import ( - build_anthropic_client, - resolve_anthropic_token, - _is_oauth_token, - ) - # Only fall back to ANTHROPIC_TOKEN when the provider is actually Anthropic. - # Other anthropic_messages providers (MiniMax, Alibaba, etc.) must use their own - # API key — falling back would send Anthropic credentials to third-party endpoints. - _is_native_anthropic = new_provider == "anthropic" - effective_key = (api_key or self.api_key or resolve_anthropic_token() or "") if _is_native_anthropic else (api_key or self.api_key or "") - self.api_key = effective_key - self._anthropic_api_key = effective_key - self._anthropic_base_url = base_url or getattr(self, "_anthropic_base_url", None) - self._anthropic_client = build_anthropic_client( - effective_key, self._anthropic_base_url, - timeout=get_provider_request_timeout(self.provider, self.model), - ) - self._is_anthropic_oauth = _is_oauth_token(effective_key) if _is_native_anthropic else False - self.client = None - self._client_kwargs = {} - else: - effective_key = api_key or self.api_key - effective_base = base_url or self.base_url - self._client_kwargs = { - "api_key": effective_key, - "base_url": effective_base, - } - _sm_timeout = get_provider_request_timeout(self.provider, self.model) - if _sm_timeout is not None: - self._client_kwargs["timeout"] = _sm_timeout - self.client = self._create_openai_client( - dict(self._client_kwargs), - reason="switch_model", - shared=True, - ) - - # ── Re-evaluate prompt caching ── - self._use_prompt_caching, self._use_native_cache_layout = ( - self._anthropic_prompt_cache_policy( - provider=new_provider, - base_url=self.base_url, - api_mode=api_mode, - model=new_model, - ) - ) - - # ── LM Studio: preload before probing context length ── - self._ensure_lmstudio_runtime_loaded() - - # ── Update context compressor ── - if hasattr(self, "context_compressor") and self.context_compressor: - from agent.model_metadata import get_model_context_length - # Re-read custom_providers from live config so per-model - # context_length overrides are honored when switching to a - # custom provider mid-session (closes #15779). - _sm_custom_providers = None - try: - from hermes_cli.config import load_config, get_compatible_custom_providers - _sm_cfg = load_config() - _sm_custom_providers = get_compatible_custom_providers(_sm_cfg) - except Exception: - _sm_custom_providers = None - new_context_length = get_model_context_length( - self.model, - base_url=self.base_url, - api_key=self.api_key, - provider=self.provider, - config_context_length=getattr(self, "_config_context_length", None), - custom_providers=_sm_custom_providers, - ) - self.context_compressor.update_model( - model=self.model, - context_length=new_context_length, - base_url=self.base_url, - api_key=getattr(self, "api_key", ""), - provider=self.provider, - api_mode=self.api_mode, - ) - - # ── Invalidate cached system prompt so it rebuilds next turn ── - self._cached_system_prompt = None - - # ── Update _primary_runtime so the change persists across turns ── - _cc = self.context_compressor if hasattr(self, "context_compressor") and self.context_compressor else None - self._primary_runtime = { - "model": self.model, - "provider": self.provider, - "base_url": self.base_url, - "api_mode": self.api_mode, - "api_key": getattr(self, "api_key", ""), - "client_kwargs": dict(self._client_kwargs), - "use_prompt_caching": self._use_prompt_caching, - "use_native_cache_layout": self._use_native_cache_layout, - "compressor_model": getattr(_cc, "model", self.model) if _cc else self.model, - "compressor_base_url": getattr(_cc, "base_url", self.base_url) if _cc else self.base_url, - "compressor_api_key": getattr(_cc, "api_key", "") if _cc else "", - "compressor_provider": getattr(_cc, "provider", self.provider) if _cc else self.provider, - "compressor_context_length": _cc.context_length if _cc else 0, - "compressor_threshold_tokens": _cc.threshold_tokens if _cc else 0, - } - if api_mode == "anthropic_messages": - self._primary_runtime.update({ - "anthropic_api_key": self._anthropic_api_key, - "anthropic_base_url": self._anthropic_base_url, - "is_anthropic_oauth": self._is_anthropic_oauth, - }) - - # ── Reset fallback state ── - self._fallback_activated = False - self._fallback_index = 0 - - # When the user deliberately swaps primary providers (e.g. openrouter - # → anthropic), drop any fallback entries that target the OLD primary - # or the NEW one. The chain was seeded from config at agent init for - # the original provider — without pruning, a failed turn on the new - # primary silently re-activates the provider the user just rejected, - # which is exactly what was reported during TUI v2 blitz testing - # ("switched to anthropic, tui keeps trying openrouter"). - old_norm = (old_provider or "").strip().lower() - new_norm = (new_provider or "").strip().lower() - fallback_chain = list(getattr(self, "_fallback_chain", []) or []) - if old_norm and new_norm and old_norm != new_norm: - fallback_chain = [ - entry for entry in fallback_chain - if (entry.get("provider") or "").strip().lower() not in {old_norm, new_norm} - ] - self._fallback_chain = fallback_chain - self._fallback_model = fallback_chain[0] if fallback_chain else None - - logging.info( - "Model switched in-place: %s (%s) -> %s (%s)", - old_model, old_provider, new_model, new_provider, - ) + """Forwarder — see ``agent.agent_runtime_helpers.switch_model``.""" + from agent.agent_runtime_helpers import switch_model + return switch_model(self, new_model, new_provider, api_key, base_url, api_mode) def _safe_print(self, *args, **kwargs): """Print that silently handles broken pipes / closed stdout. @@ -2948,99 +677,28 @@ class AIAgent: except Exception: logger.debug("status_callback error in _emit_warning", exc_info=True) - # Headers we capture from the dying stream's HTTP response so post-mortem - # diagnosis can answer "which CF edge / which OpenRouter downstream - # provider / which request id". Lowercased; httpx returns CIMultiDict. - _STREAM_DIAG_HEADERS = ( - "cf-ray", - "cf-cache-status", - "x-openrouter-provider", - "x-openrouter-model", - "x-openrouter-id", - "x-request-id", - "x-vercel-id", - "via", - "server", - "x-forwarded-for", - ) + # Stream-diagnostic class header preserved for backward compat — + # actual list lives in ``agent.stream_diag.STREAM_DIAG_HEADERS``. + from agent.stream_diag import STREAM_DIAG_HEADERS as _STREAM_DIAG_HEADERS # noqa: E402 @staticmethod def _stream_diag_init() -> Dict[str, Any]: - """Return a fresh per-attempt diagnostic dict. - - Mutated in-place by the streaming functions and read from the retry - block when a stream dies. Lives on ``request_client_holder`` so it - survives across the closure boundary. - """ - return { - "started_at": time.time(), - "first_chunk_at": None, - "chunks": 0, - "bytes": 0, - "headers": {}, - "http_status": None, - } + """Forwarder — see ``agent.stream_diag.stream_diag_init``.""" + from agent.stream_diag import stream_diag_init + return stream_diag_init() def _stream_diag_capture_response( self, diag: Dict[str, Any], http_response: Any ) -> None: - """Snapshot interesting headers + HTTP status from the live stream. - - Called once at stream open (before iterating chunks) so the metadata - survives even if the stream dies before any chunk arrives. Failures - are swallowed — diag is best-effort. - """ - if http_response is None or not isinstance(diag, dict): - return - try: - diag["http_status"] = getattr(http_response, "status_code", None) - except Exception: - pass - try: - headers = getattr(http_response, "headers", None) or {} - captured: Dict[str, str] = {} - for name in self._STREAM_DIAG_HEADERS: - try: - val = headers.get(name) - if val: - # Truncate single-value to keep log lines bounded. - captured[name] = str(val)[:120] - except Exception: - continue - diag["headers"] = captured - except Exception: - pass + """Forwarder — see ``agent.stream_diag.stream_diag_capture_response``.""" + from agent.stream_diag import stream_diag_capture_response + stream_diag_capture_response(self, diag, http_response) @staticmethod def _flatten_exception_chain(error: BaseException) -> str: - """Return a compact ``Outer(msg) <- Inner(msg) <- ...`` rendering. - - OpenAI SDK wraps httpx errors as ``APIConnectionError`` / - ``APIError`` and only the wrapper's class is visible at the catch - site — but the underlying ``RemoteProtocolError`` / - ``ConnectError`` / ``ReadError`` is what tells us WHY the stream - died. Walks ``__cause__`` then ``__context__`` (deduped, max 4 - deep) to surface the chain in one line. - """ - seen: List[BaseException] = [] - link: Optional[BaseException] = error - while link is not None and len(seen) < 4: - if link in seen: - break - seen.append(link) - nxt = getattr(link, "__cause__", None) or getattr( - link, "__context__", None - ) - if nxt is None or nxt is link: - break - link = nxt - parts: List[str] = [] - for e in seen: - msg = str(e).strip().replace("\n", " ") - if len(msg) > 140: - msg = msg[:140] + "…" - parts.append(f"{type(e).__name__}({msg})" if msg else type(e).__name__) - return " <- ".join(parts) if parts else type(error).__name__ + """Forwarder — see ``agent.stream_diag.flatten_exception_chain``.""" + from agent.stream_diag import flatten_exception_chain + return flatten_exception_chain(error) def _is_provider_stream_parse_error(self, error: BaseException) -> bool: """Return True for malformed provider streaming data from SDK parsers. @@ -3070,88 +728,12 @@ class AIAgent: mid_tool_call: bool, diag: Optional[Dict[str, Any]] = None, ) -> None: - """Record a transient stream-drop and retry to ``agent.log``. - - Always logs a structured WARNING so users have a breadcrumb regardless - of UI verbosity. Subagents in particular benefit because their - retries no longer spam the parent's terminal — but the file log keeps - full detail (provider, error class, attempt, base_url, subagent_id). - - When *diag* is provided (the per-attempt stream-diagnostic dict from - ``_stream_diag_init``), the WARNING also captures upstream headers - (cf-ray, x-openrouter-provider, x-openrouter-id), HTTP status, bytes - streamed before the drop, and elapsed time on the dying attempt. - These are the breadcrumbs needed to answer "is one CF edge / one - downstream provider responsible, or is it random across runs?" - """ - try: - try: - _summary = self._summarize_api_error(error) - except Exception: - _summary = str(error) - if _summary and len(_summary) > 240: - _summary = _summary[:240] + "…" - - # Inner-cause chain (httpx errors hide under openai.APIError). - try: - _chain = self._flatten_exception_chain(error) - except Exception: - _chain = type(error).__name__ - - # Per-attempt counters and upstream headers. - _now = time.time() - _bytes = 0 - _chunks = 0 - _elapsed = 0.0 - _ttfb = None - _headers_repr = "-" - _http_status = "-" - if isinstance(diag, dict): - try: - _bytes = int(diag.get("bytes") or 0) - _chunks = int(diag.get("chunks") or 0) - _started = float(diag.get("started_at") or _now) - _elapsed = max(0.0, _now - _started) - _first = diag.get("first_chunk_at") - if _first is not None: - _ttfb = max(0.0, float(_first) - _started) - headers = diag.get("headers") or {} - if isinstance(headers, dict) and headers: - _headers_repr = " ".join( - f"{k}={v}" for k, v in headers.items() - ) - if diag.get("http_status") is not None: - _http_status = str(diag.get("http_status")) - except Exception: - pass - - logger.warning( - "Stream %s on attempt %s/%s — retrying. " - "subagent_id=%s depth=%s provider=%s base_url=%s " - "error_type=%s error=%s " - "chain=%s " - "http_status=%s bytes=%d chunks=%d elapsed=%.2fs ttfb=%s " - "upstream=[%s]", - kind, - attempt, - max_attempts, - getattr(self, "_subagent_id", None) or "-", - getattr(self, "_delegate_depth", 0), - self.provider or "-", - self.base_url or "-", - type(error).__name__, - _summary, - _chain, - _http_status, - _bytes, - _chunks, - _elapsed, - f"{_ttfb:.2f}s" if _ttfb is not None else "-", - _headers_repr, - extra={"mid_tool_call": mid_tool_call}, - ) - except Exception: - logger.debug("stream-retry log emit failed", exc_info=True) + """Forwarder — see ``agent.stream_diag.log_stream_retry``.""" + from agent.stream_diag import log_stream_retry + log_stream_retry( + self, kind=kind, error=error, attempt=attempt, + max_attempts=max_attempts, mid_tool_call=mid_tool_call, diag=diag, + ) def _emit_stream_drop( self, @@ -3162,53 +744,12 @@ class AIAgent: mid_tool_call: bool, diag: Optional[Dict[str, Any]] = None, ) -> None: - """Emit a single user-visible line for a stream drop+retry. - - Both top-level agents and subagents announce drops in the UI — the - parent prefixes subagent lines with ``[subagent-N]`` via ``log_prefix`` - so they're easy to attribute. All cases also write a structured - WARNING to ``agent.log`` via :meth:`_log_stream_retry` with the full - diagnostic detail (subagent_id, provider, base_url, error_type, - cf-ray, x-openrouter-provider, bytes/chunks, elapsed) for post-hoc - analysis. - - The user-visible status line is intentionally compact: provider, - error class, attempt N/M, plus ``after Xs`` when the stream dropped - mid-flight. Full diagnostic detail goes to ``agent.log`` only — - ``hermes logs --level WARNING | grep "Stream drop"`` to inspect. - """ - kind = "drop mid tool-call" if mid_tool_call else "drop" - self._log_stream_retry( - kind=kind, - error=error, - attempt=attempt, - max_attempts=max_attempts, - mid_tool_call=mid_tool_call, - diag=diag, + """Forwarder — see ``agent.stream_diag.emit_stream_drop``.""" + from agent.stream_diag import emit_stream_drop + emit_stream_drop( + self, error=error, attempt=attempt, max_attempts=max_attempts, + mid_tool_call=mid_tool_call, diag=diag, ) - provider = self.provider or "provider" - # Compose a brief "after Xs" suffix when we have timing data — helps - # the user distinguish "couldn't connect" (0s) from "died after 30s - # of streaming" (likely upstream idle-kill or proxy timeout). - _suffix = "" - if isinstance(diag, dict): - try: - started = diag.get("started_at") - if started is not None: - _suffix = f" after {max(0.0, time.time() - float(started)):.1f}s" - except Exception: - pass - try: - self._emit_status( - f"⚠️ {provider} stream {kind} ({type(error).__name__}){_suffix} " - f"— reconnecting, retry {attempt}/{max_attempts}" - ) - self._touch_activity( - f"stream retry {attempt}/{max_attempts} " - f"after {type(error).__name__}" - ) - except Exception: - pass def _emit_auxiliary_failure(self, task: str, exc: BaseException) -> None: """Surface a compact warning for failed auxiliary work.""" @@ -3232,201 +773,14 @@ class AIAgent: } def _check_compression_model_feasibility(self) -> None: - """Warn at session start if the auxiliary compression model's context - window is smaller than the main model's compression threshold. - - When the auxiliary model cannot fit the content that needs summarising, - compression will either fail outright (the LLM call errors) or produce - a severely truncated summary. - - Called during ``__init__`` so CLI users see the warning immediately - (via ``_vprint``). The gateway sets ``status_callback`` *after* - construction, so ``_replay_compression_warning()`` re-sends the - stored warning through the callback on the first - ``run_conversation()`` call. - """ - if not self.compression_enabled: - return - try: - from agent.auxiliary_client import ( - _resolve_task_provider_model, - get_text_auxiliary_client, - ) - from agent.model_metadata import ( - MINIMUM_CONTEXT_LENGTH, - get_model_context_length, - ) - - client, aux_model = get_text_auxiliary_client( - "compression", - main_runtime=self._current_main_runtime(), - ) - # Best-effort aux provider label for the warning message. The - # configured provider may be "auto", in which case we fall back - # to the client's base_url hostname so the user can still tell - # where the compression model is actually being called. - try: - _aux_cfg_provider, _, _, _, _ = _resolve_task_provider_model("compression") - except Exception: - _aux_cfg_provider = "" - if client is None or not aux_model: - if _aux_cfg_provider and _aux_cfg_provider != "auto": - msg = ( - "⚠ Configured auxiliary compression provider " - f"'{_aux_cfg_provider}' is unavailable — context " - "compression will drop middle turns without a summary. " - "Check auxiliary.compression in config.yaml and " - "reauthenticate that provider." - ) - else: - msg = ( - "⚠ No auxiliary LLM provider configured — context " - "compression will drop middle turns without a summary. " - "Run `hermes setup` or set OPENROUTER_API_KEY." - ) - self._compression_warning = msg - self._emit_status(msg) - logger.warning( - "No auxiliary LLM provider for compression — " - "summaries will be unavailable." - ) - return - - aux_base_url = str(getattr(client, "base_url", "")) - aux_api_key = str(getattr(client, "api_key", "")) - - aux_context = get_model_context_length( - aux_model, - base_url=aux_base_url, - api_key=aux_api_key, - config_context_length=getattr(self, "_aux_compression_context_length_config", None), - # Each model must be resolved with its own provider so that - # provider-specific paths (e.g. Bedrock static table, OpenRouter API) - # are invoked for the correct client, not inherited from the main model. - provider=(_aux_cfg_provider if _aux_cfg_provider and _aux_cfg_provider != "auto" else getattr(self, "provider", "")), - custom_providers=self._custom_providers, - ) - - # Hard floor: the auxiliary compression model must have at least - # MINIMUM_CONTEXT_LENGTH (64K) tokens of context. The main model - # is already required to meet this floor (checked earlier in - # __init__), so the compression model must too — otherwise it - # cannot summarise a full threshold-sized window of main-model - # content. Mirrors the main-model rejection pattern. - if aux_context and aux_context < MINIMUM_CONTEXT_LENGTH: - raise ValueError( - f"Auxiliary compression model {aux_model} has a context " - f"window of {aux_context:,} tokens, which is below the " - f"minimum {MINIMUM_CONTEXT_LENGTH:,} required by Hermes " - f"Agent. Choose a compression model with at least " - f"{MINIMUM_CONTEXT_LENGTH // 1000}K context (set " - f"auxiliary.compression.model in config.yaml), or set " - f"auxiliary.compression.context_length to override the " - f"detected value if it is wrong." - ) - - threshold = self.context_compressor.threshold_tokens - if aux_context < threshold: - # Auto-correct: lower the live session threshold so - # compression actually works this session. The hard floor - # above guarantees aux_context >= MINIMUM_CONTEXT_LENGTH, - # so the new threshold is always >= 64K. - # - # The compression summariser sends a single user-role - # prompt (no system prompt, no tools) to the aux model, so - # new_threshold == aux_context is safe: the request is - # the raw messages plus a small summarisation instruction. - old_threshold = threshold - new_threshold = aux_context - self.context_compressor.threshold_tokens = new_threshold - # Keep threshold_percent in sync so future main-model - # context_length changes (update_model) re-derive from a - # sensible number rather than the original too-high value. - main_ctx = self.context_compressor.context_length - if main_ctx: - self.context_compressor.threshold_percent = ( - new_threshold / main_ctx - ) - safe_pct = int((aux_context / main_ctx) * 100) if main_ctx else 50 - # Build human-readable "model (provider)" labels for both - # the main model and the compression model so users can - # tell at a glance which provider each side is actually - # using. When the configured provider is empty or "auto", - # fall back to the client's base_url hostname. - _main_model = getattr(self, "model", "") or "?" - _main_provider = getattr(self, "provider", "") or "" - _aux_provider_label = ( - _aux_cfg_provider - if _aux_cfg_provider and _aux_cfg_provider != "auto" - else "" - ) - if not _aux_provider_label: - try: - from urllib.parse import urlparse - _aux_provider_label = ( - urlparse(aux_base_url).hostname or aux_base_url - ) - except Exception: - _aux_provider_label = aux_base_url or "auto" - _main_label = ( - f"{_main_model} ({_main_provider})" - if _main_provider - else _main_model - ) - _aux_label = f"{aux_model} ({_aux_provider_label})" - msg = ( - f"⚠ Compression model {_aux_label} context is " - f"{aux_context:,} tokens, but the main model " - f"{_main_label}'s compression threshold was " - f"{old_threshold:,} tokens. " - f"Auto-lowered this session's threshold to " - f"{new_threshold:,} tokens so compression can run.\n" - f" To make this permanent, edit config.yaml — either:\n" - f" 1. Use a larger compression model:\n" - f" auxiliary:\n" - f" compression:\n" - f" model: \n" - f" 2. Lower the compression threshold:\n" - f" compression:\n" - f" threshold: 0.{safe_pct:02d}" - ) - self._compression_warning = msg - self._emit_status(msg) - logger.warning( - "Auxiliary compression model %s has %d token context, " - "below the main model's compression threshold of %d " - "tokens — auto-lowered session threshold to %d to " - "keep compression working.", - aux_model, - aux_context, - old_threshold, - new_threshold, - ) - except ValueError: - # Hard rejections (aux below minimum context) must propagate - # so the session refuses to start. - raise - except Exception as exc: - logger.debug( - "Compression feasibility check failed (non-fatal): %s", exc - ) + """Forwarder — see ``agent.conversation_compression.check_compression_model_feasibility``.""" + from agent.conversation_compression import check_compression_model_feasibility + check_compression_model_feasibility(self) def _replay_compression_warning(self) -> None: - """Re-send the compression warning through ``status_callback``. - - During ``__init__`` the gateway's ``status_callback`` is not yet - wired, so ``_emit_status`` only reaches ``_vprint`` (CLI). This - method is called once at the start of the first - ``run_conversation()`` — by then the gateway has set the callback, - so every platform (Telegram, Discord, Slack, etc.) receives the - warning. - """ - msg = getattr(self, "_compression_warning", None) - if msg and self.status_callback: - try: - self.status_callback("lifecycle", msg) - except Exception: - pass + """Forwarder — see ``agent.conversation_compression.replay_compression_warning``.""" + from agent.conversation_compression import replay_compression_warning + replay_compression_warning(self) def _is_direct_openai_url(self, base_url: str = None) -> bool: """Return True when a base URL targets OpenAI's native API.""" @@ -3534,101 +888,9 @@ class AIAgent: api_mode: Optional[str] = None, model: Optional[str] = None, ) -> tuple[bool, bool]: - """Decide whether to apply Anthropic prompt caching and which layout to use. - - Returns ``(should_cache, use_native_layout)``: - * ``should_cache`` — inject ``cache_control`` breakpoints for this - request (applies to OpenRouter Claude, native Anthropic, and - third-party gateways that speak the native Anthropic protocol). - * ``use_native_layout`` — place markers on the *inner* content - blocks (native Anthropic accepts and requires this layout); - when False markers go on the message envelope (OpenRouter and - OpenAI-wire proxies expect the looser layout). - - Third-party providers using the native Anthropic transport - (``api_mode == 'anthropic_messages'`` + Claude-named model) get - caching with the native layout so they benefit from the same - cost reduction as direct Anthropic callers, provided their - gateway implements the Anthropic cache_control contract - (MiniMax, Zhipu GLM, LiteLLM's Anthropic proxy mode all do). - - Qwen / Alibaba-family models on OpenCode, OpenCode Go, and direct - Alibaba (DashScope) also honour Anthropic-style ``cache_control`` - markers on OpenAI-wire chat completions. Upstream pi-mono #3392 / - pi #3393 documented this for opencode-go Qwen. Without markers - these providers serve zero cache hits, re-billing the full prompt - on every turn. - """ - eff_provider = (provider if provider is not None else self.provider) or "" - eff_base_url = base_url if base_url is not None else (self.base_url or "") - eff_api_mode = api_mode if api_mode is not None else (self.api_mode or "") - eff_model = (model if model is not None else self.model) or "" - - model_lower = eff_model.lower() - provider_lower = eff_provider.lower() - is_claude = "claude" in model_lower - is_openrouter = base_url_host_matches(eff_base_url, "openrouter.ai") - # Nous Portal proxies to OpenRouter behind the scenes — identical - # OpenAI-wire envelope cache_control semantics. Treat it as an - # OpenRouter-equivalent endpoint for caching layout purposes. - is_nous_portal = "nousresearch" in eff_base_url.lower() - is_anthropic_wire = eff_api_mode == "anthropic_messages" - is_native_anthropic = ( - is_anthropic_wire - and (eff_provider == "anthropic" or base_url_hostname(eff_base_url) == "api.anthropic.com") - ) - - if is_native_anthropic: - return True, True - if (is_openrouter or is_nous_portal) and is_claude: - return True, False - # Nous Portal Qwen (e.g. qwen3.6-plus) takes the same envelope-layout - # cache_control path as Portal Claude. Portal proxies to OpenRouter - # and the upstream Qwen route accepts cache_control markers; without - # this branch the alibaba-family check below only matches - # provider=opencode/alibaba and Portal traffic falls through to - # (False, False), serving 0% cache hits and re-billing the full - # prompt on every turn. - if is_nous_portal and "qwen" in model_lower: - return True, False - if is_anthropic_wire and is_claude: - # Third-party Anthropic-compatible gateway. - return True, True - - # MiniMax on its Anthropic-compatible endpoint serves its own - # model family (MiniMax-M2.7, M2.5, M2.1, M2) with documented - # cache_control support (0.1× read pricing, 5-minute TTL). The - # blanket is_claude gate above excludes these — opt them in - # explicitly via provider id or host match so users on - # provider=minimax / minimax-cn (or custom endpoints pointing at - # api.minimax.io/anthropic / api.minimaxi.com/anthropic) get the - # same cost reduction as Claude traffic. - # Docs: https://platform.minimax.io/docs/api-reference/anthropic-api-compatible-cache - if is_anthropic_wire: - is_minimax_provider = provider_lower in {"minimax", "minimax-cn"} - is_minimax_host = ( - base_url_host_matches(eff_base_url, "api.minimax.io") - or base_url_host_matches(eff_base_url, "api.minimaxi.com") - ) - if is_minimax_provider or is_minimax_host: - return True, True - - # Qwen/Alibaba on OpenCode (Zen/Go) and native DashScope: OpenAI-wire - # transport that accepts Anthropic-style cache_control markers and - # rewards them with real cache hits. Without this branch - # qwen3.6-plus on opencode-go reports 0% cached tokens and burns - # through the subscription on every turn. - model_is_qwen = "qwen" in model_lower - provider_is_alibaba_family = provider_lower in { - "opencode", "opencode-zen", "opencode-go", "alibaba", - } - if provider_is_alibaba_family and model_is_qwen: - # Envelope layout (native_anthropic=False): markers on inner - # content parts, not top-level tool messages. Matches - # pi-mono's "alibaba" cacheControlFormat. - return True, False - - return False, False + """Forwarder — see ``agent.agent_runtime_helpers.anthropic_prompt_cache_policy``.""" + from agent.agent_runtime_helpers import anthropic_prompt_cache_policy + return anthropic_prompt_cache_policy(self, provider=provider, base_url=base_url, api_mode=api_mode, model=model) @staticmethod def _model_requires_responses_api(model: str) -> bool: @@ -3704,98 +966,9 @@ class AIAgent: return bool(cleaned.strip()) def _strip_think_blocks(self, content: str) -> str: - """Remove reasoning/thinking blocks from content, returning only visible text. - - Handles four cases: - 1. Closed tag pairs (``…``) — the common path when - the provider emits complete reasoning blocks. - 2. Unterminated open tag at a block boundary (start of text or - after a newline) — e.g. MiniMax M2.7 / NIM endpoints where the - closing tag is dropped. Everything from the open tag to end - of string is stripped. The block-boundary check mirrors - ``gateway/stream_consumer.py``'s filter so models that mention - ```` in prose aren't over-stripped. - 3. Stray orphan open/close tags that slip through. - 4. Tag variants: ````, ````, ````, - ````, ```` (Gemma 4), all - case-insensitive. - - Additionally strips standalone tool-call XML blocks that some open - models (notably Gemma variants on OpenRouter) emit inside assistant - content instead of via the structured ``tool_calls`` field: - * ``…`` - * ``…`` - * ``…`` - * ``…`` - * ``…`` - * ``…`` (Gemma style) - Ported from openclaw/openclaw#67318. The ```` variant is - boundary-gated (only strips when the tag sits at start-of-line or - after punctuation and carries a ``name="..."`` attribute) so prose - mentions like "Use in JavaScript" are preserved. - """ - if not content: - return "" - # 1. Closed tag pairs — case-insensitive for all variants so - # mixed-case tags (, ) don't slip through to - # the unterminated-tag pass and take trailing content with them. - content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) - content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) - content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) - content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) - content = re.sub(r'.*?', '', content, flags=re.DOTALL | re.IGNORECASE) - # 1b. Tool-call XML blocks (openclaw/openclaw#67318). Handle the - # generic tag names first — they have no attribute gating since - # a literal in prose is already vanishingly rare. - for _tc_name in ("tool_call", "tool_calls", "tool_result", - "function_call", "function_calls"): - content = re.sub( - rf'<{_tc_name}\b[^>]*>.*?', - '', - content, - flags=re.DOTALL | re.IGNORECASE, - ) - # 1c. ... — Gemma-style standalone - # tool call. Only strip when the tag sits at a block boundary - # (start of text, after a newline, or after sentence-ending - # punctuation) AND carries a name="..." attribute. This keeps - # prose mentions like "Use to declare" safe. - content = re.sub( - r'(?:(?<=^)|(?<=[\n\r.!?:]))[ \t]*' - r']*\bname\s*=[^>]*>' - r'(?:(?:(?!).)*)', - '', - content, - flags=re.DOTALL | re.IGNORECASE, - ) - # 2. Unterminated reasoning block — open tag at a block boundary - # (start of text, or after a newline) with no matching close. - # Strip from the tag to end of string. Fixes #8878 / #9568 - # (MiniMax M2.7 leaking raw reasoning into assistant content). - content = re.sub( - r'(?:^|\n)[ \t]*<(?:think|thinking|reasoning|thought|REASONING_SCRATCHPAD)\b[^>]*>.*$', - '', - content, - flags=re.DOTALL | re.IGNORECASE, - ) - # 3. Stray orphan open/close tags that slipped through. - content = re.sub( - r'\s*', - '', - content, - flags=re.IGNORECASE, - ) - # 3b. Stray tool-call closers. (We do NOT strip bare or - # unterminated because a truncated tail - # during streaming may still be valuable to the user; matches - # OpenClaw's intentional asymmetry.) - content = re.sub( - r'\s*', - '', - content, - flags=re.IGNORECASE, - ) - return content + """Forwarder — see ``agent.agent_runtime_helpers.strip_think_blocks``.""" + from agent.agent_runtime_helpers import strip_think_blocks + return strip_think_blocks(self, content) @staticmethod def _has_natural_response_ending(content: str) -> bool: @@ -3856,366 +1029,27 @@ class AIAgent: assistant_content: str, messages: List[Dict[str, Any]], ) -> bool: - """Detect a planning/ack message that should continue instead of ending the turn.""" - if any(isinstance(msg, dict) and msg.get("role") == "tool" for msg in messages): - return False - - assistant_text = self._strip_think_blocks(assistant_content or "").strip().lower() - if not assistant_text: - return False - if len(assistant_text) > 1200: - return False - - has_future_ack = bool( - re.search(r"\b(i['’]ll|i will|let me|i can do that|i can help with that)\b", assistant_text) - ) - if not has_future_ack: - return False - - action_markers = ( - "look into", - "look at", - "inspect", - "scan", - "check", - "analyz", - "review", - "explore", - "read", - "open", - "run", - "test", - "fix", - "debug", - "search", - "find", - "walkthrough", - "report back", - "summarize", - ) - workspace_markers = ( - "directory", - "current directory", - "current dir", - "cwd", - "repo", - "repository", - "codebase", - "project", - "folder", - "filesystem", - "file tree", - "files", - "path", - ) - - user_text = (user_message or "").strip().lower() - user_targets_workspace = ( - any(marker in user_text for marker in workspace_markers) - or "~/" in user_text - or "/" in user_text - ) - assistant_mentions_action = any(marker in assistant_text for marker in action_markers) - assistant_targets_workspace = any( - marker in assistant_text for marker in workspace_markers - ) - return (user_targets_workspace or assistant_targets_workspace) and assistant_mentions_action - + """Forwarder — see ``agent.agent_runtime_helpers.looks_like_codex_intermediate_ack``.""" + from agent.agent_runtime_helpers import looks_like_codex_intermediate_ack + return looks_like_codex_intermediate_ack(self, user_message, assistant_content, messages) def _extract_reasoning(self, assistant_message) -> Optional[str]: - """ - Extract reasoning/thinking content from an assistant message. - - OpenRouter and various providers can return reasoning in multiple formats: - 1. message.reasoning - Direct reasoning field (DeepSeek, Qwen, etc.) - 2. message.reasoning_content - Alternative field (Moonshot AI, Novita, etc.) - 3. message.reasoning_details - Array of {type, summary, ...} objects (OpenRouter unified) - - Args: - assistant_message: The assistant message object from the API response - - Returns: - Combined reasoning text, or None if no reasoning found - """ - reasoning_parts = [] - - # Check direct reasoning field - if hasattr(assistant_message, 'reasoning') and assistant_message.reasoning: - reasoning_parts.append(assistant_message.reasoning) - - # Check reasoning_content field (alternative name used by some providers) - if hasattr(assistant_message, 'reasoning_content') and assistant_message.reasoning_content: - # Don't duplicate if same as reasoning - if assistant_message.reasoning_content not in reasoning_parts: - reasoning_parts.append(assistant_message.reasoning_content) - - # Check reasoning_details array (OpenRouter unified format) - # Format: [{"type": "reasoning.summary", "summary": "...", ...}, ...] - if hasattr(assistant_message, 'reasoning_details') and assistant_message.reasoning_details: - for detail in assistant_message.reasoning_details: - if isinstance(detail, dict): - # Extract summary from reasoning detail object - summary = ( - detail.get('summary') - or detail.get('thinking') - or detail.get('content') - or detail.get('text') - ) - if summary and summary not in reasoning_parts: - reasoning_parts.append(summary) - - # Some providers embed reasoning directly inside assistant content - # instead of returning structured reasoning fields. Only fall back - # to inline extraction when no structured reasoning was found. - content = getattr(assistant_message, "content", None) - if not reasoning_parts and isinstance(content, list): - # DeepSeek V4 Pro (and compatible providers) return content as a - # list of typed blocks, e.g.: - # [{"type": "thinking", "thinking": "..."}, {"type": "output", ...}] - # Without this branch the thinking text is silently dropped and the - # next turn fails with HTTP 400 ("thinking must be passed back"). - # Refs #21944. - for block in content: - if isinstance(block, dict) and block.get("type") == "thinking": - thinking_text = block.get("thinking") or block.get("text") or "" - thinking_text = thinking_text.strip() - if thinking_text and thinking_text not in reasoning_parts: - reasoning_parts.append(thinking_text) - if not reasoning_parts and isinstance(content, str) and content: - inline_patterns = ( - r"(.*?)", - r"(.*?)", - r"(.*?)", - r"(.*?)", - r"(.*?)", - ) - for pattern in inline_patterns: - flags = re.DOTALL | re.IGNORECASE - for block in re.findall(pattern, content, flags=flags): - cleaned = block.strip() - if cleaned and cleaned not in reasoning_parts: - reasoning_parts.append(cleaned) - - # Combine all reasoning parts - if reasoning_parts: - return "\n\n".join(reasoning_parts) - - return None + """Forwarder — see ``agent.agent_runtime_helpers.extract_reasoning``.""" + from agent.agent_runtime_helpers import extract_reasoning + return extract_reasoning(self, assistant_message) def _cleanup_task_resources(self, task_id: str) -> None: - """Clean up VM and browser resources for a given task. - - Skips ``cleanup_vm`` when the active terminal environment is marked - persistent (``persistent_filesystem=True``) so that long-lived sandbox - containers survive between turns. The idle reaper in - ``terminal_tool._cleanup_inactive_envs`` still tears them down once - ``terminal.lifetime_seconds`` is exceeded. Non-persistent backends are - torn down per-turn as before to prevent resource leakage (the original - intent of this hook for the Morph backend, see commit fbd3a2fd). - """ - try: - if is_persistent_env(task_id): - if self.verbose_logging: - logging.debug( - f"Skipping per-turn cleanup_vm for persistent env {task_id}; " - f"idle reaper will handle it." - ) - else: - cleanup_vm(task_id) - except Exception as e: - if self.verbose_logging: - logging.warning(f"Failed to cleanup VM for task {task_id}: {e}") - try: - cleanup_browser(task_id) - except Exception as e: - if self.verbose_logging: - logging.warning(f"Failed to cleanup browser for task {task_id}: {e}") + """Forwarder — see ``agent.chat_completion_helpers.cleanup_task_resources``.""" + from agent.chat_completion_helpers import cleanup_task_resources + return cleanup_task_resources(self, task_id) # ------------------------------------------------------------------ - # Background memory/skill review + # Background memory/skill review — prompts live in agent.background_review # ------------------------------------------------------------------ - - _MEMORY_REVIEW_PROMPT = ( - "Review the conversation above and consider saving to memory if appropriate.\n\n" - "Focus on:\n" - "1. Has the user revealed things about themselves — their persona, desires, " - "preferences, or personal details worth remembering?\n" - "2. Has the user expressed expectations about how you should behave, their work " - "style, or ways they want you to operate?\n\n" - "If something stands out, save it using the memory tool. " - "If nothing is worth saving, just say 'Nothing to save.' and stop." - ) - - _SKILL_REVIEW_PROMPT = ( - "Review the conversation above and update the skill library. Be " - "ACTIVE — most sessions produce at least one skill update, even if " - "small. A pass that does nothing is a missed learning opportunity, " - "not a neutral outcome.\n\n" - "Target shape of the library: CLASS-LEVEL skills, each with a rich " - "SKILL.md and a `references/` directory for session-specific detail. " - "Not a long flat list of narrow one-session-one-skill entries. This " - "shapes HOW you update, not WHETHER you update.\n\n" - "Signals to look for (any one of these warrants action):\n" - " • User corrected your style, tone, format, legibility, or " - "verbosity. Frustration signals like 'stop doing X', 'this is too " - "verbose', 'don't format like this', 'why are you explaining', " - "'just give me the answer', 'you always do Y and I hate it', or an " - "explicit 'remember this' are FIRST-CLASS skill signals, not just " - "memory signals. Update the relevant skill(s) to embed the " - "preference so the next session starts already knowing.\n" - " • User corrected your workflow, approach, or sequence of steps. " - "Encode the correction as a pitfall or explicit step in the skill " - "that governs that class of task.\n" - " • Non-trivial technique, fix, workaround, debugging path, or " - "tool-usage pattern emerged that a future session would benefit " - "from. Capture it.\n" - " • A skill that got loaded or consulted this session turned out " - "to be wrong, missing a step, or outdated. Patch it NOW.\n\n" - "Preference order — prefer the earliest action that fits, but do " - "pick one when a signal above fired:\n" - " 1. UPDATE A CURRENTLY-LOADED SKILL. Look back through the " - "conversation for skills the user loaded via /skill-name or you " - "read via skill_view. If any of them covers the territory of the " - "new learning, PATCH that one first. It is the skill that was in " - "play, so it's the right one to extend.\n" - " 2. UPDATE AN EXISTING UMBRELLA (via skills_list + skill_view). " - "If no loaded skill fits but an existing class-level skill does, " - "patch it. Add a subsection, a pitfall, or broaden a trigger.\n" - " 3. ADD A SUPPORT FILE under an existing umbrella. Skills can be " - "packaged with three kinds of support files — use the right " - "directory per kind:\n" - " • `references/.md` — session-specific detail (error " - "transcripts, reproduction recipes, provider quirks) AND " - "condensed knowledge banks: quoted research, API docs, external " - "authoritative excerpts, or domain notes you found while working " - "on the problem. Write it concise and for the value of the task, " - "not as a full mirror of upstream docs.\n" - " • `templates/.` — starter files meant to be " - "copied and modified (boilerplate configs, scaffolding, a " - "known-good example the agent can `reproduce with modifications`).\n" - " • `scripts/.` — statically re-runnable actions " - "the skill can invoke directly (verification scripts, fixture " - "generators, deterministic probes, anything the agent should run " - "rather than hand-type each time).\n" - " Add support files via skill_manage action=write_file with " - "file_path starting 'references/', 'templates/', or 'scripts/'. " - "The umbrella's SKILL.md should gain a one-line pointer to any " - "new support file so future agents know it exists.\n" - " 4. CREATE A NEW CLASS-LEVEL UMBRELLA SKILL when no existing " - "skill covers the class. The name MUST be at the class level. " - "The name MUST NOT be a specific PR number, error string, feature " - "codename, library-alone name, or 'fix-X / debug-Y / audit-Z-today' " - "session artifact. If the proposed name only makes sense for " - "today's task, it's wrong — fall back to (1), (2), or (3).\n\n" - "User-preference embedding (important): when the user expressed a " - "style/format/workflow preference, the update belongs in the " - "SKILL.md body, not just in memory. Memory captures 'who the user " - "is and what the current situation and state of your operations " - "are'; skills capture 'how to do this class of task for this " - "user'. When they complain about how you handled a task, the " - "skill that governs that task needs to carry the lesson.\n\n" - "If you notice two existing skills that overlap, note it in your " - "reply — the background curator handles consolidation at scale.\n\n" - "Do NOT capture (these become persistent self-imposed constraints " - "that bite you later when the environment changes):\n" - " • Environment-dependent failures: missing binaries, fresh-install " - "errors, post-migration path mismatches, 'command not found', " - "unconfigured credentials, uninstalled packages. The user can fix " - "these — they are not durable rules.\n" - " • Negative claims about tools or features ('browser tools do not " - "work', 'X tool is broken', 'cannot use Y from execute_code'). These " - "harden into refusals the agent cites against itself for months " - "after the actual problem was fixed.\n" - " • Session-specific transient errors that resolved before the " - "conversation ended. If retrying worked, the lesson is the retry " - "pattern, not the original failure.\n" - " • One-off task narratives. A user asking 'summarize today's " - "market' or 'analyze this PR' is not a class of work that warrants " - "a skill.\n\n" - "If a tool failed because of setup state, capture the FIX (install " - "command, config step, env var to set) under an existing setup or " - "troubleshooting skill — never 'this tool does not work' as a " - "standalone constraint.\n\n" - "'Nothing to save.' is a real option but should NOT be the " - "default. If the session ran smoothly with no corrections and " - "produced no new technique, just say 'Nothing to save.' and stop. " - "Otherwise, act." - ) - - _COMBINED_REVIEW_PROMPT = ( - "Review the conversation above and update two things:\n\n" - "**Memory**: who the user is. Did the user reveal persona, " - "desires, preferences, personal details, or expectations about " - "how you should behave? Save facts about the user and durable " - "preferences with the memory tool.\n\n" - "**Skills**: how to do this class of task. Be ACTIVE — most " - "sessions produce at least one skill update. A pass that does " - "nothing is a missed learning opportunity, not a neutral outcome.\n\n" - "Target shape of the skill library: CLASS-LEVEL skills with a rich " - "SKILL.md and a `references/` directory for session-specific detail. " - "Not a long flat list of narrow one-session-one-skill entries.\n\n" - "Signals that warrant a skill update (any one is enough):\n" - " • User corrected your style, tone, format, legibility, " - "verbosity, or approach. Frustration is a FIRST-CLASS skill " - "signal, not just a memory signal. 'stop doing X', 'don't format " - "like this', 'I hate when you Y' — embed the lesson in the skill " - "that governs that task so the next session starts fixed.\n" - " • Non-trivial technique, fix, workaround, or debugging path " - "emerged.\n" - " • A skill that was loaded or consulted turned out wrong, " - "missing, or outdated — patch it now.\n\n" - "Preference order for skills — pick the earliest that fits:\n" - " 1. UPDATE A CURRENTLY-LOADED SKILL. Check what skills were " - "loaded via /skill-name or skill_view in the conversation. If one " - "of them covers the learning, PATCH it first. It was in play; " - "it's the right place.\n" - " 2. UPDATE AN EXISTING UMBRELLA (skills_list + skill_view to " - "find the right one). Patch it.\n" - " 3. ADD A SUPPORT FILE under an existing umbrella via " - "skill_manage action=write_file. Three kinds: " - "`references/.md` for session-specific detail OR condensed " - "knowledge banks (quoted research, API docs excerpts, domain " - "notes) written concise and task-focused; `templates/.` " - "for starter files meant to be copied and modified; " - "`scripts/.` for statically re-runnable actions " - "(verification, fixture generators, probes). Add a one-line " - "pointer in SKILL.md so future agents find them.\n" - " 4. CREATE A NEW CLASS-LEVEL UMBRELLA when nothing exists. " - "Name at the class level — NOT a PR number, error string, " - "codename, library-alone name, or 'fix-X / debug-Y' session " - "artifact. If the name only fits today's task, fall back to (1), " - "(2), or (3).\n\n" - "User-preference embedding: when the user complains about how " - "you handled a task, update the skill that governs that task — " - "memory alone isn't enough. Memory says 'who the user is and " - "what the current situation and state of your operations are'; " - "skills say 'how to do this class of task for this user'. Both " - "should carry user-preference lessons when relevant.\n\n" - "If you notice overlapping existing skills, mention it — the " - "background curator handles consolidation.\n\n" - "Do NOT capture as skills (these become persistent self-imposed " - "constraints that bite you later when the environment changes):\n" - " • Environment-dependent failures: missing binaries, fresh-install " - "errors, post-migration path mismatches, 'command not found', " - "unconfigured credentials, uninstalled packages. The user can fix " - "these — they are not durable rules.\n" - " • Negative claims about tools or features ('browser tools do not " - "work', 'X tool is broken', 'cannot use Y from execute_code'). These " - "harden into refusals the agent cites against itself for months " - "after the actual problem was fixed.\n" - " • Session-specific transient errors that resolved before the " - "conversation ended. If retrying worked, the lesson is the retry " - "pattern, not the original failure.\n" - " • One-off task narratives. A user asking 'summarize today's " - "market' or 'analyze this PR' is not a class of work that warrants " - "a skill.\n\n" - "If a tool failed because of setup state, capture the FIX (install " - "command, config step, env var to set) under an existing setup or " - "troubleshooting skill — never 'this tool does not work' as a " - "standalone constraint.\n\n" - "Act on whichever of the two dimensions has real signal. If " - "genuinely nothing stands out on either, say 'Nothing to save.' " - "and stop — but don't reach for that conclusion as a default." + from agent.background_review import ( + _MEMORY_REVIEW_PROMPT, + _SKILL_REVIEW_PROMPT, + _COMBINED_REVIEW_PROMPT, ) @staticmethod @@ -4223,63 +1057,9 @@ class AIAgent: review_messages: List[Dict], prior_snapshot: List[Dict], ) -> List[str]: - """Build the human-facing action summary for a background review pass. - - Walks the review agent's session messages and collects "successful tool - action" descriptions to surface to the user (e.g. "Memory updated"). - Tool messages already present in ``prior_snapshot`` are skipped so we - don't re-surface stale results from the prior conversation that the - review agent inherited via ``conversation_history`` (issue #14944). - - Matching is by ``tool_call_id`` when available, with a content-equality - fallback for tool messages that lack one. - """ - existing_tool_call_ids = set() - existing_tool_contents = set() - for prior in prior_snapshot or []: - if not isinstance(prior, dict) or prior.get("role") != "tool": - continue - tcid = prior.get("tool_call_id") - if tcid: - existing_tool_call_ids.add(tcid) - else: - content = prior.get("content") - if isinstance(content, str): - existing_tool_contents.add(content) - - actions: List[str] = [] - for msg in review_messages or []: - if not isinstance(msg, dict) or msg.get("role") != "tool": - continue - tcid = msg.get("tool_call_id") - if tcid and tcid in existing_tool_call_ids: - continue - if not tcid: - content_str = msg.get("content") - if isinstance(content_str, str) and content_str in existing_tool_contents: - continue - try: - data = json.loads(msg.get("content", "{}")) - except (json.JSONDecodeError, TypeError): - continue - if not isinstance(data, dict) or not data.get("success"): - continue - message = data.get("message", "") - target = data.get("target", "") - if "created" in message.lower(): - actions.append(message) - elif "updated" in message.lower(): - actions.append(message) - elif "added" in message.lower() or (target and "add" in message.lower()): - label = "Memory" if target == "memory" else "User profile" if target == "user" else target - actions.append(f"{label} updated") - elif "Entry added" in message: - label = "Memory" if target == "memory" else "User profile" if target == "user" else target - actions.append(f"{label} updated") - elif "removed" in message.lower() or "replaced" in message.lower(): - label = "Memory" if target == "memory" else "User profile" if target == "user" else target - actions.append(f"{label} updated") - return actions + """Forwarder — see ``agent.background_review.summarize_background_review_actions``.""" + from agent.background_review import summarize_background_review_actions + return summarize_background_review_actions(review_messages, prior_snapshot) def _spawn_background_review( self, @@ -4287,235 +1067,22 @@ class AIAgent: review_memory: bool = False, review_skills: bool = False, ) -> None: - """Spawn a background thread to review the conversation for memory/skill saves. + """Spawn the background memory/skill review thread. - Creates a full AIAgent fork with the same model, tools, and context as the - main session. The review prompt is appended as the next user turn in the - forked conversation. Writes directly to the shared memory/skill stores. - Never modifies the main conversation history or produces user-visible output. + Thin wrapper — the heavy lifting lives in + ``agent.background_review.spawn_background_review_thread`` which + returns the thread target. ``threading.Thread`` is constructed + here so existing tests that patch ``run_agent.threading.Thread`` + keep working. """ - import threading - - # Pick the right prompt based on which triggers fired - if review_memory and review_skills: - prompt = self._COMBINED_REVIEW_PROMPT - elif review_memory: - prompt = self._MEMORY_REVIEW_PROMPT - else: - prompt = self._SKILL_REVIEW_PROMPT - - def _run_review(): - import contextlib - # Install a non-interactive approval callback on this worker - # thread so any dangerous-command guard the review agent trips - # resolves to "deny" instead of falling back to input() -- which - # deadlocks against the parent's prompt_toolkit TUI (#15216). - # Same pattern as _subagent_auto_deny in tools/delegate_tool.py. - def _bg_review_auto_deny(command, description, **kwargs): - logger.warning( - "Background review auto-denied dangerous command: %s (%s)", - command, description, - ) - return "deny" - try: - _set_approval_callback(_bg_review_auto_deny) - except Exception: - pass - review_agent = None - review_messages = [] - try: - with open(os.devnull, "w", encoding="utf-8") as _devnull, \ - contextlib.redirect_stdout(_devnull), \ - contextlib.redirect_stderr(_devnull): - # Inherit the parent agent's live runtime (provider, model, - # base_url, api_key, api_mode) so the fork uses the exact - # same credentials the main turn is using. Without this, - # AIAgent.__init__ re-runs auto-resolution from env vars, - # which fails for OAuth-only providers, session-scoped - # creds, or credential-pool setups where the resolver can't - # reconstruct auth from scratch -- producing the spurious - # "No LLM provider configured" warning at end of turn. - _parent_runtime = self._current_main_runtime() - _parent_api_mode = _parent_runtime.get("api_mode") or None - # The review fork needs to call agent-loop tools (memory, - # skill_manage). Those tools require Hermes' own dispatch, - # which the codex_app_server runtime bypasses entirely - # (it runs the turn inside codex's subprocess). So when - # the parent is on codex_app_server, downgrade the review - # fork to codex_responses — same auth/credentials, but - # talks to the OpenAI Responses API directly so Hermes - # owns the loop and the agent-loop tools dispatch. - if _parent_api_mode == "codex_app_server": - _parent_api_mode = "codex_responses" - # skip_memory=True keeps the review fork from - # touching external memory plugins (honcho, mem0, - # supermemory, etc.). Without it, the fork's - # __init__ rebuilds its own _memory_manager from - # config, scoped to the parent's session_id, and - # run_conversation() then leaks the harness prompt - # into the user's real memory namespace via three - # ingestion sites: on_turn_start (cadence + turn - # message), prefetch_all (recall query), and - # sync_all (harness prompt + review output recorded - # as a (user, assistant) turn pair). Built-in - # MEMORY.md / USER.md state is re-bound from the - # parent below so memory(action="add") writes from - # the review still land on disk; the review just - # has zero side effects on external providers. - review_agent = AIAgent( - model=self.model, - max_iterations=16, - quiet_mode=True, - platform=self.platform, - provider=self.provider, - api_mode=_parent_api_mode, - base_url=_parent_runtime.get("base_url") or None, - api_key=_parent_runtime.get("api_key") or None, - credential_pool=getattr(self, "_credential_pool", None), - parent_session_id=self.session_id, - skip_memory=True, - ) - review_agent._memory_write_origin = "background_review" - review_agent._memory_write_context = "background_review" - review_agent._memory_store = self._memory_store - review_agent._memory_enabled = self._memory_enabled - review_agent._user_profile_enabled = self._user_profile_enabled - review_agent._memory_nudge_interval = 0 - review_agent._skill_nudge_interval = 0 - # Suppress all status/warning emits from the fork so the - # user only sees the final successful-action summary. - # Without this, mid-review "Iteration budget exhausted", - # rate-limit retries, compression warnings, and other - # lifecycle messages bubble up through _emit_status -> - # _vprint and leak past the stdout redirect (they go via - # _print_fn/status_callback, which bypass sys.stdout). - review_agent.suppress_status_output = True - # Inherit the parent's cached system prompt verbatim so - # the review fork's outbound HTTP request hits the same - # Anthropic/OpenRouter prefix cache the parent warmed. - # Without this, the fork rebuilds the system prompt from - # scratch (fresh _hermes_now() timestamp, fresh - # session_id, narrower toolset → different skills_prompt) - # and the byte-exact prefix-cache key misses. See - # issue #25322 and PR #17276 for the full analysis + - # measured impact (~26% end-to-end cost reduction on - # Sonnet 4.5). - review_agent._cached_system_prompt = self._cached_system_prompt - # Defensive: pin session_start + session_id to the - # parent's so any code path that re-renders parts of - # the system prompt (compression, plugin hooks) still - # produces byte-identical output. The cached-prompt - # assignment above already short-circuits the normal - # rebuild path, but these pins guarantee parity even - # if a future code path bypasses the cache. - review_agent.session_start = self.session_start - review_agent.session_id = self.session_id - - from model_tools import get_tool_definitions - from hermes_cli.plugins import ( - set_thread_tool_whitelist, - clear_thread_tool_whitelist, - ) - - review_whitelist = { - t["function"]["name"] - for t in get_tool_definitions( - enabled_toolsets=["memory", "skills"], - quiet_mode=True, - ) - } - set_thread_tool_whitelist( - review_whitelist, - deny_msg_fmt=( - "Background review denied non-whitelisted tool: " - "{tool_name}. Only memory/skill tools are allowed." - ), - ) - try: - review_agent.run_conversation( - user_message=( - prompt - + "\n\nYou can only call memory and skill " - "management tools. Other tools will be denied " - "at runtime — do not attempt them." - ), - conversation_history=messages_snapshot, - ) - finally: - clear_thread_tool_whitelist() - - # Tear down memory providers while stdout is still - # redirected so background thread teardown (Honcho flush, - # Hindsight sync, etc.) stays silent. The finally block - # below is a safety net for the exception path. - try: - review_agent.shutdown_memory_provider() - except Exception: - pass - try: - review_agent.close() - except Exception: - pass - review_messages = list(getattr(review_agent, "_session_messages", [])) - review_agent = None - - # Scan the review agent's messages for successful tool actions - # and surface a compact summary to the user. Tool messages - # already present in messages_snapshot must be skipped, since - # the review agent inherits that history and would otherwise - # re-surface stale "created"/"updated" messages from the prior - # conversation as if they just happened (issue #14944). - actions = self._summarize_background_review_actions( - review_messages, - messages_snapshot, - ) - - if actions: - summary = " · ".join(dict.fromkeys(actions)) - self._safe_print( - f" 💾 Self-improvement review: {summary}" - ) - _bg_cb = self.background_review_callback - if _bg_cb: - try: - _bg_cb( - f"💾 Self-improvement review: {summary}" - ) - except Exception: - pass - - except Exception as e: - logger.warning("Background memory/skill review failed: %s", e) - self._emit_auxiliary_failure("background review", e) - finally: - # Safety-net cleanup for the exception path. Normal - # completion already shut down inside redirect_stdout above. - # Re-open devnull here so any teardown output (Honcho flush, - # Hindsight sync, background thread joins) stays silent even - # on the exception path where redirect_stdout already exited. - if review_agent is not None: - try: - with open(os.devnull, "w", encoding="utf-8") as _fn, \ - contextlib.redirect_stdout(_fn), \ - contextlib.redirect_stderr(_fn): - try: - review_agent.shutdown_memory_provider() - except Exception: - pass - try: - review_agent.close() - except Exception: - pass - except Exception: - pass - # Clear the approval callback on this bg-review thread so a - # recycled thread-id doesn't inherit a stale reference. - try: - _set_approval_callback(None) - except Exception: - pass - - t = threading.Thread(target=_run_review, daemon=True, name="bg-review") + from agent.background_review import spawn_background_review_thread + target, _prompt = spawn_background_review_thread( + self, + messages_snapshot, + review_memory=review_memory, + review_skills=review_skills, + ) + t = threading.Thread(target=target, daemon=True, name="bg-review") t.start() def _build_memory_write_metadata( @@ -4526,23 +1093,15 @@ class AIAgent: task_id: Optional[str] = None, tool_call_id: Optional[str] = None, ) -> Dict[str, Any]: - """Build provenance metadata for external memory-provider mirrors.""" - metadata: Dict[str, Any] = { - "write_origin": write_origin or getattr(self, "_memory_write_origin", "assistant_tool"), - "execution_context": ( - execution_context - or getattr(self, "_memory_write_context", "foreground") - ), - "session_id": self.session_id or "", - "parent_session_id": self._parent_session_id or "", - "platform": self.platform or os.environ.get("HERMES_SESSION_SOURCE", "cli"), - "tool_name": "memory", - } - if task_id: - metadata["task_id"] = task_id - if tool_call_id: - metadata["tool_call_id"] = tool_call_id - return {k: v for k, v in metadata.items() if v not in {None, ""}} + """Forwarder — see ``agent.background_review.build_memory_write_metadata``.""" + from agent.background_review import build_memory_write_metadata + return build_memory_write_metadata( + self, + write_origin=write_origin, + execution_context=execution_context, + task_id=task_id, + tool_call_id=tool_call_id, + ) def _apply_persist_user_message_override(self, messages: List[Dict]) -> None: """Rewrite the current-turn user message before persistence/return. @@ -4627,104 +1186,9 @@ class AIAgent: messages.pop() def _repair_message_sequence(self, messages: List[Dict]) -> int: - """Collapse malformed role-alternation left in the live history. - - Providers (OpenAI, OpenRouter, Anthropic) expect strict alternation: - after the system message, user/tool alternates with assistant, with - no two consecutive user messages and no tool-result that doesn't - follow an assistant-with-tool_calls. Violations cause silent empty - responses on most providers, which triggers the empty-retry loop. - - This runs right before the API call as a defensive belt — by the - time it fires, the scaffolding strip should already have prevented - most shapes, but external callers (gateway multi-queue replay, - session resume, cron, explicit conversation_history passed in by - host code) can feed in already-broken histories. - - Repairs applied: - 1. Stray ``tool`` messages whose ``tool_call_id`` doesn't match - any preceding assistant tool_call — dropped. - 2. Consecutive ``user`` messages — merged with newline separator - so no user input is lost. - - Deliberately does NOT rewind orphan ``assistant(tool_calls)+tool`` - pairs that precede a user message — that pattern IS valid when the - previous turn completed normally and the user jumped in to redirect - before the model got a continuation turn (the ongoing dialog - pattern). The empty-response scaffolding stripper handles the - genuinely-broken variant via its flag-gated rewind. - - Returns the number of repairs made (for logging/telemetry). - """ - if not messages: - return 0 - - repairs = 0 - - # Pass 1: drop stray tool messages that don't follow a known - # assistant tool_call_id. Uses a rolling set of known ids refreshed - # on each assistant message. - known_tool_ids: set = set() - filtered: List[Dict] = [] - for msg in messages: - if not isinstance(msg, dict): - filtered.append(msg) - continue - role = msg.get("role") - if role == "assistant": - known_tool_ids = set() - for tc in (msg.get("tool_calls") or []): - tc_id = tc.get("id") if isinstance(tc, dict) else None - if tc_id: - known_tool_ids.add(tc_id) - filtered.append(msg) - elif role == "tool": - tc_id = msg.get("tool_call_id") - if tc_id and tc_id in known_tool_ids: - filtered.append(msg) - else: - repairs += 1 - else: - if role == "user": - # A user turn closes the tool-result run; subsequent - # tool messages without a fresh assistant tool_call - # are orphans. - known_tool_ids = set() - filtered.append(msg) - - # Pass 2: merge consecutive user messages. Preserves all user input - # so nothing the user typed is lost. - merged: List[Dict] = [] - for msg in filtered: - if ( - merged - and isinstance(msg, dict) - and msg.get("role") == "user" - and isinstance(merged[-1], dict) - and merged[-1].get("role") == "user" - ): - prev = merged[-1] - prev_content = prev.get("content", "") - new_content = msg.get("content", "") - # Only merge plain-text content; leave multimodal (list) - # content alone — collapsing image/audio blocks risks - # mangling the attachment structure. - if isinstance(prev_content, str) and isinstance(new_content, str): - prev["content"] = ( - (prev_content + "\n\n" + new_content) - if prev_content and new_content - else (prev_content or new_content) - ) - repairs += 1 - continue - merged.append(msg) - - if repairs > 0: - # Rewrite in place so downstream paths (persistence, return - # value, session DB flush) see the repaired sequence. - messages[:] = merged - - return repairs + """Forwarder — see ``agent.agent_runtime_helpers.repair_message_sequence``.""" + from agent.agent_runtime_helpers import repair_message_sequence + return repair_message_sequence(self, messages) def _flush_messages_to_session_db(self, messages: List[Dict], conversation_history: List[Dict] = None): """Persist any un-flushed messages to the SQLite session store. @@ -4817,197 +1281,14 @@ class AIAgent: return messages[:last_assistant_idx] def _format_tools_for_system_message(self) -> str: - """ - Format tool definitions for the system message in the trajectory format. - - Returns: - str: JSON string representation of tool definitions - """ - if not self.tools: - return "[]" - - # Convert tool definitions to the format expected in trajectories - formatted_tools = [] - for tool in self.tools: - func = tool["function"] - formatted_tool = { - "name": func["name"], - "description": func.get("description", ""), - "parameters": func.get("parameters", {}), - "required": None # Match the format in the example - } - formatted_tools.append(formatted_tool) - - return json.dumps(formatted_tools, ensure_ascii=False) + """Forwarder — see ``agent.system_prompt.format_tools_for_system_message``.""" + from agent.system_prompt import format_tools_for_system_message + return format_tools_for_system_message(self) def _convert_to_trajectory_format(self, messages: List[Dict[str, Any]], user_query: str, completed: bool) -> List[Dict[str, Any]]: - """ - Convert internal message format to trajectory format for saving. - - Args: - messages (List[Dict]): Internal message history - user_query (str): Original user query - completed (bool): Whether the conversation completed successfully - - Returns: - List[Dict]: Messages in trajectory format - """ - # Normalize multimodal tool results — trajectories are text-only, so - # replace image-bearing tool messages with their text_summary to avoid - # embedding ~1MB base64 blobs into every saved trajectory. - messages = [_trajectory_normalize_msg(m) for m in messages] - trajectory = [] - - # Add system message with tool definitions - system_msg = ( - "You are a function calling AI model. You are provided with function signatures within XML tags. " - "You may call one or more functions to assist with the user query. If available tools are not relevant in assisting " - "with user query, just respond in natural conversational language. Don't make assumptions about what values to plug " - "into functions. After calling & executing the functions, you will be provided with function results within " - " XML tags. Here are the available tools:\n" - f"\n{self._format_tools_for_system_message()}\n\n" - "For each function call return a JSON object, with the following pydantic model json schema for each:\n" - "{'title': 'FunctionCall', 'type': 'object', 'properties': {'name': {'title': 'Name', 'type': 'string'}, " - "'arguments': {'title': 'Arguments', 'type': 'object'}}, 'required': ['name', 'arguments']}\n" - "Each function call should be enclosed within XML tags.\n" - "Example:\n\n{'name': ,'arguments': }\n" - ) - - trajectory.append({ - "from": "system", - "value": system_msg - }) - - # Add the actual user prompt (from the dataset) as the first human message - trajectory.append({ - "from": "human", - "value": user_query - }) - - # Skip the first message (the user query) since we already added it above. - # Prefill messages are injected at API-call time only (not in the messages - # list), so no offset adjustment is needed here. - i = 1 - - while i < len(messages): - msg = messages[i] - - if msg["role"] == "assistant": - # Check if this message has tool calls - if "tool_calls" in msg and msg["tool_calls"]: - # Format assistant message with tool calls - # Add tags around reasoning for trajectory storage - content = "" - - # Prepend reasoning in tags if available (native thinking tokens) - if msg.get("reasoning") and msg["reasoning"].strip(): - content = f"\n{msg['reasoning']}\n\n" - - if msg.get("content") and msg["content"].strip(): - # Convert any tags to tags - # (used when native thinking is disabled and model reasons via XML) - content += convert_scratchpad_to_think(msg["content"]) + "\n" - - # Add tool calls wrapped in XML tags - for tool_call in msg["tool_calls"]: - if not tool_call or not isinstance(tool_call, dict): continue - # Parse arguments - should always succeed since we validate during conversation - # but keep try-except as safety net - try: - arguments = json.loads(tool_call["function"]["arguments"]) if isinstance(tool_call["function"]["arguments"], str) else tool_call["function"]["arguments"] - except json.JSONDecodeError: - # This shouldn't happen since we validate and retry during conversation, - # but if it does, log warning and use empty dict - logging.warning(f"Unexpected invalid JSON in trajectory conversion: {tool_call['function']['arguments'][:100]}") - arguments = {} - - tool_call_json = { - "name": tool_call["function"]["name"], - "arguments": arguments - } - content += f"\n{json.dumps(tool_call_json, ensure_ascii=False)}\n\n" - - # Ensure every gpt turn has a block (empty if no reasoning) - # so the format is consistent for training data - if "" not in content: - content = "\n\n" + content - - trajectory.append({ - "from": "gpt", - "value": content.rstrip() - }) - - # Collect all subsequent tool responses - tool_responses = [] - j = i + 1 - while j < len(messages) and messages[j]["role"] == "tool": - tool_msg = messages[j] - # Format tool response with XML tags - tool_response = "\n" - - # Try to parse tool content as JSON if it looks like JSON - tool_content = tool_msg["content"] - try: - if tool_content.strip().startswith(("{", "[")): - tool_content = json.loads(tool_content) - except (json.JSONDecodeError, AttributeError): - pass # Keep as string if not valid JSON - - tool_index = len(tool_responses) - tool_name = ( - msg["tool_calls"][tool_index]["function"]["name"] - if tool_index < len(msg["tool_calls"]) - else "unknown" - ) - tool_response += json.dumps({ - "tool_call_id": tool_msg.get("tool_call_id", ""), - "name": tool_name, - "content": tool_content - }, ensure_ascii=False) - tool_response += "\n" - tool_responses.append(tool_response) - j += 1 - - # Add all tool responses as a single message - if tool_responses: - trajectory.append({ - "from": "tool", - "value": "\n".join(tool_responses) - }) - i = j - 1 # Skip the tool messages we just processed - - else: - # Regular assistant message without tool calls - # Add tags around reasoning for trajectory storage - content = "" - - # Prepend reasoning in tags if available (native thinking tokens) - if msg.get("reasoning") and msg["reasoning"].strip(): - content = f"\n{msg['reasoning']}\n\n" - - # Convert any tags to tags - # (used when native thinking is disabled and model reasons via XML) - raw_content = msg["content"] or "" - content += convert_scratchpad_to_think(raw_content) - - # Ensure every gpt turn has a block (empty if no reasoning) - if "" not in content: - content = "\n\n" + content - - trajectory.append({ - "from": "gpt", - "value": content.strip() - }) - - elif msg["role"] == "user": - trajectory.append({ - "from": "human", - "value": msg["content"] - }) - - i += 1 - - return trajectory + """Forwarder — see ``agent.agent_runtime_helpers.convert_to_trajectory_format``.""" + from agent.agent_runtime_helpers import convert_to_trajectory_format + return convert_to_trajectory_format(self, messages, user_query, completed) def _save_trajectory(self, messages: List[Dict[str, Any]], user_query: str, completed: bool): """ @@ -5045,7 +1326,7 @@ class AIAgent: the existing 1M-context-beta branch handles them; revisit if other subscription tiers start producing the same loop signature). """ - if status_code not in (401, 403, None): + if status_code not in {401, 403, None}: return False if not isinstance(error_context, dict): return False @@ -5200,68 +1481,9 @@ class AIAgent: @staticmethod def _extract_api_error_context(error: Exception) -> Dict[str, Any]: - """Extract structured rate-limit details from provider errors.""" - context: Dict[str, Any] = {} - - body = getattr(error, "body", None) - payload = None - if isinstance(body, dict): - payload = body.get("error") if isinstance(body.get("error"), dict) else body - if isinstance(payload, dict): - reason = payload.get("code") or payload.get("error") - if isinstance(reason, str) and reason.strip(): - context["reason"] = reason.strip() - message = payload.get("message") or payload.get("error_description") - if isinstance(message, str) and message.strip(): - context["message"] = message.strip() - for key in ("resets_at", "reset_at"): - value = payload.get(key) - if value not in {None, ""}: - context["reset_at"] = value - break - retry_after = payload.get("retry_after") - if retry_after not in {None, ""} and "reset_at" not in context: - try: - context["reset_at"] = time.time() + float(retry_after) - except (TypeError, ValueError): - pass - - response = getattr(error, "response", None) - headers = getattr(response, "headers", None) - if headers: - retry_after = headers.get("retry-after") or headers.get("Retry-After") - if retry_after and "reset_at" not in context: - try: - context["reset_at"] = time.time() + float(retry_after) - except (TypeError, ValueError): - pass - ratelimit_reset = headers.get("x-ratelimit-reset") - if ratelimit_reset and "reset_at" not in context: - context["reset_at"] = ratelimit_reset - - if "message" not in context: - raw_message = str(error).strip() - if raw_message: - context["message"] = raw_message[:500] - - if "reset_at" not in context: - message = context.get("message") or "" - if isinstance(message, str): - delay_match = re.search(r"quotaResetDelay[:\s\"]+(\\d+(?:\\.\\d+)?)(ms|s)", message, re.IGNORECASE) - if delay_match: - value = float(delay_match.group(1)) - seconds = value / 1000.0 if delay_match.group(2).lower() == "ms" else value - context["reset_at"] = time.time() + seconds - else: - sec_match = re.search( - r"retry\s+(?:after\s+)?(\d+(?:\.\d+)?)\s*(?:sec|secs|seconds|s\b)", - message, - re.IGNORECASE, - ) - if sec_match: - context["reset_at"] = time.time() + float(sec_match.group(1)) - - return context + """Forwarder — see ``agent.agent_runtime_helpers.extract_api_error_context``.""" + from agent.agent_runtime_helpers import extract_api_error_context + return extract_api_error_context(error) def _usage_summary_for_api_request_hook(self, response: Any) -> Optional[Dict[str, Any]]: """Token buckets for ``post_api_request`` plugins (no raw ``response`` object).""" @@ -5286,80 +1508,9 @@ class AIAgent: reason: str, error: Optional[Exception] = None, ) -> Optional[Path]: - """ - Dump a debug-friendly HTTP request record for the active inference API. - - Captures the request body from api_kwargs (excluding transport-only keys - like timeout). Intended for debugging provider-side 4xx failures where - retries are not useful. - """ - try: - body = copy.deepcopy(api_kwargs) - body.pop("timeout", None) - body = {k: v for k, v in body.items() if v is not None} - - api_key = None - try: - api_key = getattr(self.client, "api_key", None) - except Exception as e: - logger.debug("Could not extract API key for debug dump: %s", e) - - dump_payload: Dict[str, Any] = { - "timestamp": datetime.now().isoformat(), - "session_id": self.session_id, - "reason": reason, - "request": { - "method": "POST", - "url": f"{self.base_url.rstrip('/')}{'/responses' if self.api_mode == 'codex_responses' else '/chat/completions'}", - "headers": { - "Authorization": f"Bearer {self._mask_api_key_for_logs(api_key)}", - "Content-Type": "application/json", - }, - "body": body, - }, - } - - if error is not None: - error_info: Dict[str, Any] = { - "type": type(error).__name__, - "message": str(error), - } - for attr_name in ("status_code", "request_id", "code", "param", "type"): - attr_value = getattr(error, attr_name, None) - if attr_value is not None: - error_info[attr_name] = attr_value - - body_attr = getattr(error, "body", None) - if body_attr is not None: - error_info["body"] = body_attr - - response_obj = getattr(error, "response", None) - if response_obj is not None: - try: - error_info["response_status"] = getattr(response_obj, "status_code", None) - error_info["response_text"] = response_obj.text - except Exception as e: - logger.debug("Could not extract error response details: %s", e) - - dump_payload["error"] = error_info - - timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f") - dump_file = self.logs_dir / f"request_dump_{self.session_id}_{timestamp}.json" - dump_file.write_text( - json.dumps(dump_payload, ensure_ascii=False, indent=2, default=str), - encoding="utf-8", - ) - - self._vprint(f"{self.log_prefix}🧾 Request debug dump written to: {dump_file}") - - if env_var_enabled("HERMES_DUMP_REQUEST_STDOUT"): - print(json.dumps(dump_payload, ensure_ascii=False, indent=2, default=str)) - - return dump_file - except Exception as dump_error: - if self.verbose_logging: - logging.warning(f"Failed to dump API request debug payload: {dump_error}") - return None + """Forwarder — see ``agent.agent_runtime_helpers.dump_api_request_debug``.""" + from agent.agent_runtime_helpers import dump_api_request_debug + return dump_api_request_debug(self, api_kwargs, reason=reason, error=error) @staticmethod def _clean_session_content(content: str) -> str: @@ -5641,7 +1792,7 @@ class AIAgent: import os as _os env = _os.environ.get("HERMES_FILE_MUTATION_VERIFIER") if env is not None: - return env.strip().lower() not in ("0", "false", "no", "off") + return env.strip().lower() not in {"0", "false", "no", "off"} # Read from the persisted config.yaml so gateway and CLI share # the same setting. Import lazily to avoid a startup-time cycle. try: @@ -5689,67 +1840,9 @@ class AIAgent: return "\n".join(lines) def _apply_pending_steer_to_tool_results(self, messages: list, num_tool_msgs: int) -> None: - """Append any pending /steer text to the last tool result in this turn. - - Called at the end of a tool-call batch, before the next API call. - The steer is appended to the last ``role:"tool"`` message's content - with a clear marker so the model understands it came from the user - and NOT from the tool itself. Role alternation is preserved — - nothing new is inserted, we only modify existing content. - - Args: - messages: The running messages list. - num_tool_msgs: Number of tool results appended in this batch; - used to locate the tail slice safely. - """ - if num_tool_msgs <= 0 or not messages: - return - steer_text = self._drain_pending_steer() - if not steer_text: - return - # Find the last tool-role message in the recent tail. Skipping - # non-tool messages defends against future code appending - # something else at the boundary. - target_idx = None - for j in range(len(messages) - 1, max(len(messages) - num_tool_msgs - 1, -1), -1): - msg = messages[j] - if isinstance(msg, dict) and msg.get("role") == "tool": - target_idx = j - break - if target_idx is None: - # No tool result in this batch (e.g. all skipped by interrupt); - # put the steer back so the caller's fallback path can deliver - # it as a normal next-turn user message. - _lock = getattr(self, "_pending_steer_lock", None) - if _lock is not None: - with _lock: - if self._pending_steer: - self._pending_steer = self._pending_steer + "\n" + steer_text - else: - self._pending_steer = steer_text - else: - existing = getattr(self, "_pending_steer", None) - self._pending_steer = (existing + "\n" + steer_text) if existing else steer_text - return - marker = f"\n\nUser guidance: {steer_text}" - existing_content = messages[target_idx].get("content", "") - if not isinstance(existing_content, str): - # Anthropic multimodal content blocks — preserve them and append - # a text block at the end. - try: - blocks = list(existing_content) if existing_content else [] - blocks.append({"type": "text", "text": marker.lstrip()}) - messages[target_idx]["content"] = blocks - except Exception: - # Fall back to string replacement if content shape is unexpected. - messages[target_idx]["content"] = f"{existing_content}{marker}" - else: - messages[target_idx]["content"] = existing_content + marker - logger.info( - "Delivered /steer to agent after tool batch (%d chars): %s", - len(steer_text), - steer_text[:120] + ("..." if len(steer_text) > 120 else ""), - ) + """Forwarder — see ``agent.agent_runtime_helpers.apply_pending_steer_to_tool_results``.""" + from agent.agent_runtime_helpers import apply_pending_steer_to_tool_results + return apply_pending_steer_to_tool_results(self, messages, num_tool_msgs) def _touch_activity(self, desc: str) -> None: """Update the last-activity timestamp and description (thread-safe).""" @@ -6070,235 +2163,14 @@ class AIAgent: def _build_system_prompt_parts(self, system_message: str = None) -> Dict[str, str]: - """Assemble the system prompt as three ordered parts. - - Returns a dict with three keys: - * ``stable`` — identity, tool guidance, skills prompt, - environment hints, platform hints, model-family operational - guidance. - * ``context`` — context files (AGENTS.md, .cursorrules, etc.) - and caller-supplied system_message. - * ``volatile`` — memory snapshot, user profile, external - memory provider block, timestamp line. - - Joined into a single string by ``_build_system_prompt`` and - cached on ``_cached_system_prompt`` for the lifetime of the - AIAgent. Hermes never re-renders parts of this string mid- - session — that's the only way to keep upstream prompt caches - warm across turns. - """ - # ── Stable tier ──────────────────────────────────────────────── - stable_parts: List[str] = [] - - # Try SOUL.md as primary identity unless the caller explicitly skipped it. - # Some execution modes (cron) still want HERMES_HOME persona while keeping - # cwd project instructions disabled. - _soul_loaded = False - if self.load_soul_identity or not self.skip_context_files: - _soul_content = load_soul_md() - if _soul_content: - stable_parts.append(_soul_content) - _soul_loaded = True - - if not _soul_loaded: - # Fallback to hardcoded identity - stable_parts.append(DEFAULT_AGENT_IDENTITY) - - # Pointer to the hermes-agent skill + docs for user questions about Hermes itself. - stable_parts.append(HERMES_AGENT_HELP_GUIDANCE) - - # Tool-aware behavioral guidance: only inject when the tools are loaded - tool_guidance = [] - if "memory" in self.valid_tool_names: - tool_guidance.append(MEMORY_GUIDANCE) - if "session_search" in self.valid_tool_names: - tool_guidance.append(SESSION_SEARCH_GUIDANCE) - if "skill_manage" in self.valid_tool_names: - tool_guidance.append(SKILLS_GUIDANCE) - # Kanban worker/orchestrator lifecycle — only present when the - # dispatcher spawned this process (kanban_show check_fn gates on - # HERMES_KANBAN_TASK env var). Normal chat sessions never see - # this block. - if "kanban_show" in self.valid_tool_names: - tool_guidance.append(KANBAN_GUIDANCE) - if tool_guidance: - stable_parts.append(" ".join(tool_guidance)) - - # Computer-use (macOS) — goes in as its own block rather than being - # merged into tool_guidance because the content is multi-paragraph. - if "computer_use" in self.valid_tool_names: - from agent.prompt_builder import COMPUTER_USE_GUIDANCE - stable_parts.append(COMPUTER_USE_GUIDANCE) - - nous_subscription_prompt = build_nous_subscription_prompt(self.valid_tool_names) - if nous_subscription_prompt: - stable_parts.append(nous_subscription_prompt) - # Tool-use enforcement: tells the model to actually call tools instead - # of describing intended actions. Controlled by config.yaml - # agent.tool_use_enforcement: - # "auto" (default) — matches TOOL_USE_ENFORCEMENT_MODELS - # true — always inject (all models) - # false — never inject - # list — custom model-name substrings to match - if self.valid_tool_names: - _enforce = self._tool_use_enforcement - _inject = False - if _enforce is True or (isinstance(_enforce, str) and _enforce.lower() in {"true", "always", "yes", "on"}): - _inject = True - elif _enforce is False or (isinstance(_enforce, str) and _enforce.lower() in {"false", "never", "no", "off"}): - _inject = False - elif isinstance(_enforce, list): - model_lower = (self.model or "").lower() - _inject = any(p.lower() in model_lower for p in _enforce if isinstance(p, str)) - else: - # "auto" or any unrecognised value — use hardcoded defaults - model_lower = (self.model or "").lower() - _inject = any(p in model_lower for p in TOOL_USE_ENFORCEMENT_MODELS) - if _inject: - stable_parts.append(TOOL_USE_ENFORCEMENT_GUIDANCE) - _model_lower = (self.model or "").lower() - # Google model operational guidance (conciseness, absolute - # paths, parallel tool calls, verify-before-edit, etc.) - if "gemini" in _model_lower or "gemma" in _model_lower: - stable_parts.append(GOOGLE_MODEL_OPERATIONAL_GUIDANCE) - # OpenAI GPT/Codex execution discipline (tool persistence, - # prerequisite checks, verification, anti-hallucination). - if "gpt" in _model_lower or "codex" in _model_lower: - stable_parts.append(OPENAI_MODEL_EXECUTION_GUIDANCE) - - has_skills_tools = any(name in self.valid_tool_names for name in ['skills_list', 'skill_view', 'skill_manage']) - if has_skills_tools: - avail_toolsets = { - toolset - for toolset in ( - get_toolset_for_tool(tool_name) for tool_name in self.valid_tool_names - ) - if toolset - } - skills_prompt = build_skills_system_prompt( - available_tools=self.valid_tool_names, - available_toolsets=avail_toolsets, - ) - else: - skills_prompt = "" - if skills_prompt: - stable_parts.append(skills_prompt) - - # Alibaba Coding Plan API always returns "glm-4.7" as model name regardless - # of the requested model. Inject explicit model identity into the system prompt - # so the agent can correctly report which model it is (workaround for API bug). - # Stable for the lifetime of an agent instance — model and provider are fixed - # at construction time. - if self.provider == "alibaba": - _model_short = self.model.split("/")[-1] if "/" in self.model else self.model - stable_parts.append( - f"You are powered by the model named {_model_short}. " - f"The exact model ID is {self.model}. " - f"When asked what model you are, always answer based on this information, " - f"not on any model name returned by the API." - ) - - # Environment hints (WSL, Termux, etc.) — tell the agent about the - # execution environment so it can translate paths and adapt behavior. - # Stable for the lifetime of the process. - _env_hints = build_environment_hints() - if _env_hints: - stable_parts.append(_env_hints) - - platform_key = (self.platform or "").lower().strip() - if platform_key in PLATFORM_HINTS: - stable_parts.append(PLATFORM_HINTS[platform_key]) - elif platform_key: - # Check plugin registry for platform-specific LLM guidance - try: - from gateway.platform_registry import platform_registry - _entry = platform_registry.get(platform_key) - if _entry and _entry.platform_hint: - stable_parts.append(_entry.platform_hint) - except Exception: - pass - - # ── Context tier (cwd-dependent, may change between sessions) ─ - context_parts: List[str] = [] - - # Note: ephemeral_system_prompt is NOT included here. It's injected at - # API-call time only so it stays out of the cached/stored system prompt. - if system_message is not None: - context_parts.append(system_message) - - if not self.skip_context_files: - # Use TERMINAL_CWD for context file discovery when set (gateway - # mode). The gateway process runs from the hermes-agent install - # dir, so os.getcwd() would pick up the repo's AGENTS.md and - # other dev files — inflating token usage by ~10k for no benefit. - _context_cwd = os.getenv("TERMINAL_CWD") or None - context_files_prompt = build_context_files_prompt( - cwd=_context_cwd, skip_soul=_soul_loaded) - if context_files_prompt: - context_parts.append(context_files_prompt) - - # ── Volatile tier (changes per session/turn — never cached) ─── - volatile_parts: List[str] = [] - - if self._memory_store: - if self._memory_enabled: - mem_block = self._memory_store.format_for_system_prompt("memory") - if mem_block: - volatile_parts.append(mem_block) - # USER.md is always included when enabled. - if self._user_profile_enabled: - user_block = self._memory_store.format_for_system_prompt("user") - if user_block: - volatile_parts.append(user_block) - - # External memory provider system prompt block (additive to built-in) - if self._memory_manager: - try: - _ext_mem_block = self._memory_manager.build_system_prompt() - if _ext_mem_block: - volatile_parts.append(_ext_mem_block) - except Exception: - pass - - from hermes_time import now as _hermes_now - now = _hermes_now() - timestamp_line = f"Conversation started: {now.strftime('%A, %B %d, %Y %I:%M %p')}" - if self.pass_session_id and self.session_id: - timestamp_line += f"\nSession ID: {self.session_id}" - if self.model: - timestamp_line += f"\nModel: {self.model}" - if self.provider: - timestamp_line += f"\nProvider: {self.provider}" - volatile_parts.append(timestamp_line) - - return { - "stable": "\n\n".join(p.strip() for p in stable_parts if p and p.strip()), - "context": "\n\n".join(p.strip() for p in context_parts if p and p.strip()), - "volatile": "\n\n".join(p.strip() for p in volatile_parts if p and p.strip()), - } + """Forwarder — see ``agent.system_prompt.build_system_prompt_parts``.""" + from agent.system_prompt import build_system_prompt_parts + return build_system_prompt_parts(self, system_message=system_message) def _build_system_prompt(self, system_message: str = None) -> str: - """ - Assemble the full system prompt from all layers. - - Called once per session (cached on self._cached_system_prompt) and only - rebuilt after context compression events. This ensures the system prompt - is stable across all turns in a session, maximizing prefix cache hits. - - Layers are ordered cache-friendly: stable identity/guidance first, - then session-stable context files, then per-call volatile content - (memory, USER profile, timestamp). The whole string is treated as - one cached block — Hermes never rebuilds or reinjects parts of it - mid-session, which is the only way to keep upstream prompt caches - warm across turns. - """ - parts = self._build_system_prompt_parts(system_message=system_message) - joined = "\n\n".join(p for p in (parts["stable"], parts["context"], parts["volatile"]) if p) - return joined - - # ========================================================================= - # Pre/post-call guardrails (inspired by PR #1321 — @alireza78a) - # ========================================================================= + """Forwarder — see ``agent.system_prompt.build_system_prompt``.""" + from agent.system_prompt import build_system_prompt + return build_system_prompt(self, system_message=system_message) @staticmethod def _get_tool_call_id_static(tc) -> str: @@ -6328,74 +2200,9 @@ class AIAgent: @staticmethod def _sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]: - """Fix orphaned tool_call / tool_result pairs before every LLM call. - - Runs unconditionally — not gated on whether the context compressor - is present — so orphans from session loading or manual message - manipulation are always caught. - """ - # --- Role allowlist: drop messages with roles the API won't accept --- - filtered = [] - for msg in messages: - role = msg.get("role") - if role not in AIAgent._VALID_API_ROLES: - logger.debug( - "Pre-call sanitizer: dropping message with invalid role %r", - role, - ) - continue - filtered.append(msg) - messages = filtered - - surviving_call_ids: set = set() - for msg in messages: - if msg.get("role") == "assistant": - for tc in msg.get("tool_calls") or []: - cid = AIAgent._get_tool_call_id_static(tc) - if cid: - surviving_call_ids.add(cid) - - result_call_ids: set = set() - for msg in messages: - if msg.get("role") == "tool": - cid = msg.get("tool_call_id") - if cid: - result_call_ids.add(cid) - - # 1. Drop tool results with no matching assistant call - orphaned_results = result_call_ids - surviving_call_ids - if orphaned_results: - messages = [ - m for m in messages - if not (m.get("role") == "tool" and m.get("tool_call_id") in orphaned_results) - ] - logger.debug( - "Pre-call sanitizer: removed %d orphaned tool result(s)", - len(orphaned_results), - ) - - # 2. Inject stub results for calls whose result was dropped - missing_results = surviving_call_ids - result_call_ids - if missing_results: - patched: List[Dict[str, Any]] = [] - for msg in messages: - patched.append(msg) - if msg.get("role") == "assistant": - for tc in msg.get("tool_calls") or []: - cid = AIAgent._get_tool_call_id_static(tc) - if cid in missing_results: - patched.append({ - "role": "tool", - "name": AIAgent._get_tool_call_name_static(tc), - "content": "[Result unavailable — see context summary above]", - "tool_call_id": cid, - }) - messages = patched - logger.debug( - "Pre-call sanitizer: added %d stub tool result(s)", - len(missing_results), - ) - return messages + """Forwarder — see ``agent.agent_runtime_helpers.sanitize_api_messages``.""" + from agent.agent_runtime_helpers import sanitize_api_messages + return sanitize_api_messages(messages) @staticmethod def _is_thinking_only_assistant(msg: Dict[str, Any]) -> bool: @@ -6455,86 +2262,9 @@ class AIAgent: def _drop_thinking_only_and_merge_users( messages: List[Dict[str, Any]], ) -> List[Dict[str, Any]]: - """Drop thinking-only assistant turns; merge any adjacent user messages left behind. - - Runs on the per-call ``api_messages`` copy only. The stored - conversation history (``self.messages``) is never mutated, so the - user still sees the thinking block in the CLI/gateway transcript and - session persistence keeps the full trace. Only the wire copy sent to - the provider is cleaned. - - Why drop-and-merge rather than inject stub text: - - Fabricating ``"."`` / ``"(continued)"`` text lies in the history - and makes future turns see model output the model didn't emit. - - Dropping the turn preserves honesty; merging adjacent user messages - preserves the provider's role-alternation invariant. - - This is the pattern used by Claude Code's ``normalizeMessagesForAPI`` - (filterOrphanedThinkingOnlyMessages + mergeAdjacentUserMessages). - """ - if not messages: - return messages - - # Pass 1: drop thinking-only assistant turns. - kept = [m for m in messages if not AIAgent._is_thinking_only_assistant(m)] - dropped = len(messages) - len(kept) - if dropped == 0: - return messages - - # Pass 2: merge any newly-adjacent user messages. - merged: List[Dict[str, Any]] = [] - merges = 0 - for m in kept: - prev = merged[-1] if merged else None - if ( - prev is not None - and prev.get("role") == "user" - and m.get("role") == "user" - ): - prev_content = prev.get("content", "") - cur_content = m.get("content", "") - # Work on a copy of ``prev`` so the caller's input dicts are - # never mutated. ``_sanitize_api_messages`` upstream already - # hands us per-call copies, but staying pure here means we - # can be called safely from anywhere (tests, other loops). - prev_copy = dict(prev) - # Only string-content merge is meaningful for role-alternation - # purposes. If either side is a list (multimodal), append as a - # separate block rather than collapsing. - if isinstance(prev_content, str) and isinstance(cur_content, str): - sep = "\n\n" if prev_content and cur_content else "" - prev_copy["content"] = prev_content + sep + cur_content - elif isinstance(prev_content, list) and isinstance(cur_content, list): - prev_copy["content"] = list(prev_content) + list(cur_content) - elif isinstance(prev_content, list) and isinstance(cur_content, str): - if cur_content: - prev_copy["content"] = list(prev_content) + [ - {"type": "text", "text": cur_content} - ] - else: - prev_copy["content"] = list(prev_content) - elif isinstance(prev_content, str) and isinstance(cur_content, list): - new_blocks: List[Dict[str, Any]] = [] - if prev_content: - new_blocks.append({"type": "text", "text": prev_content}) - new_blocks.extend(cur_content) - prev_copy["content"] = new_blocks - else: - # Unknown content shape — fall back to appending separately - # (violates alternation, but safer than raising in a hot path). - merged.append(m) - continue - merged[-1] = prev_copy - merges += 1 - else: - merged.append(m) - - logger.debug( - "Pre-call sanitizer: dropped %d thinking-only assistant turn(s), " - "merged %d adjacent user message(s)", - dropped, - merges, - ) - return merged + """Forwarder — see ``agent.agent_runtime_helpers.drop_thinking_only_and_merge_users``.""" + from agent.agent_runtime_helpers import drop_thinking_only_and_merge_users + return drop_thinking_only_and_merge_users(messages) @staticmethod def _cap_delegate_task_calls(tool_calls: list) -> list: @@ -6586,87 +2316,14 @@ class AIAgent: return unique if len(unique) < len(tool_calls) else tool_calls def _repair_tool_call(self, tool_name: str) -> str | None: - """Attempt to repair a mismatched tool name before aborting. - - Models sometimes emit variants of a tool name that differ only - in casing, separators, or class-like suffixes. Normalize - aggressively before falling back to fuzzy match: - - 1. Lowercase direct match. - 2. Lowercase + hyphens/spaces -> underscores. - 3. CamelCase -> snake_case (TodoTool -> todo_tool). - 4. Strip trailing ``_tool`` / ``-tool`` / ``tool`` suffix that - Claude-style models sometimes tack on (TodoTool_tool -> - TodoTool -> Todo -> todo). Applied twice so double-tacked - suffixes like ``TodoTool_tool`` reduce all the way. - 5. Fuzzy match (difflib, cutoff=0.7). - - See #14784 for the original reports (TodoTool_tool, Patch_tool, - BrowserClick_tool were all returning "Unknown tool" before). - - Returns the repaired name if found in valid_tool_names, else None. - """ - import re - from difflib import get_close_matches - - if not tool_name: - return None - - def _norm(s: str) -> str: - return s.lower().replace("-", "_").replace(" ", "_") - - def _camel_snake(s: str) -> str: - return re.sub(r"(? str | None: - lc = s.lower() - for suffix in ("_tool", "-tool", "tool"): - if lc.endswith(suffix): - return s[: -len(suffix)].rstrip("_-") - return None - - # Cheap fast-paths first — these cover the common case. - lowered = tool_name.lower() - if lowered in self.valid_tool_names: - return lowered - normalized = _norm(tool_name) - if normalized in self.valid_tool_names: - return normalized - - # Build the full candidate set for class-like emissions. - cands: set[str] = {tool_name, lowered, normalized, _camel_snake(tool_name)} - # Strip trailing tool-suffix up to twice — TodoTool_tool needs it. - for _ in range(2): - extra: set[str] = set() - for c in cands: - stripped = _strip_tool_suffix(c) - if stripped: - extra.add(stripped) - extra.add(_norm(stripped)) - extra.add(_camel_snake(stripped)) - cands |= extra - - for c in cands: - if c and c in self.valid_tool_names: - return c - - # Fuzzy match as last resort. - matches = get_close_matches(lowered, self.valid_tool_names, n=1, cutoff=0.7) - if matches: - return matches[0] - - return None + """Forwarder — see ``agent.agent_runtime_helpers.repair_tool_call``.""" + from agent.agent_runtime_helpers import repair_tool_call + return repair_tool_call(self, tool_name) def _invalidate_system_prompt(self): - """ - Invalidate the cached system prompt, forcing a rebuild on the next turn. - - Called after context compression events. Also reloads memory from disk - so the rebuilt prompt captures any writes from this session. - """ - self._cached_system_prompt = None - if self._memory_store: - self._memory_store.load_from_disk() + """Forwarder — see ``agent.system_prompt.invalidate_system_prompt``.""" + from agent.system_prompt import invalidate_system_prompt + invalidate_system_prompt(self) @staticmethod def _deterministic_call_id(fn_name: str, arguments: str, index: int = 0) -> str: @@ -6767,156 +2424,15 @@ class AIAgent: return None def _create_openai_client(self, client_kwargs: dict, *, reason: str, shared: bool) -> Any: - from agent.auxiliary_client import _validate_base_url, _validate_proxy_env_urls - # Treat client_kwargs as read-only. Callers pass self._client_kwargs (or shallow - # copies of it) in; any in-place mutation leaks back into the stored dict and is - # reused on subsequent requests. #10933 hit this by injecting an httpx.Client - # transport that was torn down after the first request, so the next request - # wrapped a closed transport and raised "Cannot send a request, as the client - # has been closed" on every retry. The revert resolved that specific path; this - # copy locks the contract so future transport/keepalive work can't reintroduce - # the same class of bug. - client_kwargs = dict(client_kwargs) - _validate_proxy_env_urls() - _validate_base_url(client_kwargs.get("base_url")) - if self.provider == "copilot-acp" or str(client_kwargs.get("base_url", "")).startswith("acp://copilot"): - from agent.copilot_acp_client import CopilotACPClient - - client = CopilotACPClient(**client_kwargs) - logger.info( - "Copilot ACP client created (%s, shared=%s) %s", - reason, - shared, - self._client_log_context(), - ) - return client - if self.provider == "google-gemini-cli" or str(client_kwargs.get("base_url", "")).startswith("cloudcode-pa://"): - from agent.gemini_cloudcode_adapter import GeminiCloudCodeClient - - # Strip OpenAI-specific kwargs the Gemini client doesn't accept - safe_kwargs = { - k: v for k, v in client_kwargs.items() - if k in {"api_key", "base_url", "default_headers", "project_id", "timeout"} - } - client = GeminiCloudCodeClient(**safe_kwargs) - logger.info( - "Gemini Cloud Code Assist client created (%s, shared=%s) %s", - reason, - shared, - self._client_log_context(), - ) - return client - if self.provider == "gemini": - from agent.gemini_native_adapter import GeminiNativeClient, is_native_gemini_base_url - - base_url = str(client_kwargs.get("base_url", "") or "") - if is_native_gemini_base_url(base_url): - safe_kwargs = { - k: v for k, v in client_kwargs.items() - if k in {"api_key", "base_url", "default_headers", "timeout", "http_client"} - } - if "http_client" not in safe_kwargs: - keepalive_http = self._build_keepalive_http_client(base_url) - if keepalive_http is not None: - safe_kwargs["http_client"] = keepalive_http - client = GeminiNativeClient(**safe_kwargs) - logger.info( - "Gemini native client created (%s, shared=%s) %s", - reason, - shared, - self._client_log_context(), - ) - return client - # Inject TCP keepalives so the kernel detects dead provider connections - # instead of letting them sit silently in CLOSE-WAIT (#10324). Without - # this, a peer that drops mid-stream leaves the socket in a state where - # epoll_wait never fires, ``httpx`` read timeout may not trigger, and - # the agent hangs until manually killed. Probes after 30s idle, retry - # every 10s, give up after 3 → dead peer detected within ~60s. - # - # Safety against #10933: the ``client_kwargs = dict(client_kwargs)`` - # above means this injection only lands in the local per-call copy, - # never back into ``self._client_kwargs``. Each ``_create_openai_client`` - # invocation therefore gets its OWN fresh ``httpx.Client`` whose - # lifetime is tied to the OpenAI client it is passed to. When the - # OpenAI client is closed (rebuild, teardown, credential rotation), - # the paired ``httpx.Client`` closes with it, and the next call - # constructs a fresh one — no stale closed transport can be reused. - # Tests in ``tests/run_agent/test_create_openai_client_reuse.py`` and - # ``tests/run_agent/test_sequential_chats_live.py`` pin this invariant. - if "http_client" not in client_kwargs: - keepalive_http = self._build_keepalive_http_client(client_kwargs.get("base_url", "")) - if keepalive_http is not None: - client_kwargs["http_client"] = keepalive_http - # Uses the module-level `OpenAI` name, resolved lazily on first - # access via __getattr__ below. Tests patch via `run_agent.OpenAI`. - client = OpenAI(**client_kwargs) - logger.info( - "OpenAI client created (%s, shared=%s) %s", - reason, - shared, - self._client_log_context(), - ) - return client + """Forwarder — see ``agent.agent_runtime_helpers.create_openai_client``.""" + from agent.agent_runtime_helpers import create_openai_client + return create_openai_client(self, client_kwargs, reason=reason, shared=shared) @staticmethod def _force_close_tcp_sockets(client: Any) -> int: - """Force-close underlying TCP sockets to prevent CLOSE-WAIT accumulation. - - When a provider drops a connection mid-stream, httpx's ``client.close()`` - performs a graceful shutdown which leaves sockets in CLOSE-WAIT until the - OS times them out (often minutes). This method walks the httpx transport - pool and issues ``socket.shutdown(SHUT_RDWR)`` + ``socket.close()`` to - force an immediate TCP RST, freeing the file descriptors. - - Returns the number of sockets force-closed. - """ - import socket as _socket - - closed = 0 - try: - http_client = getattr(client, "_client", None) - if http_client is None: - return 0 - transport = getattr(http_client, "_transport", None) - if transport is None: - return 0 - pool = getattr(transport, "_pool", None) - if pool is None: - return 0 - # httpx uses httpcore connection pools; connections live in - # _connections (list) or _pool (list) depending on version. - connections = ( - getattr(pool, "_connections", None) - or getattr(pool, "_pool", None) - or [] - ) - for conn in list(connections): - stream = ( - getattr(conn, "_network_stream", None) - or getattr(conn, "_stream", None) - ) - if stream is None: - continue - sock = getattr(stream, "_sock", None) - if sock is None: - sock = getattr(stream, "stream", None) - if sock is not None: - sock = getattr(sock, "_sock", None) - if sock is None: - continue - try: - sock.shutdown(_socket.SHUT_RDWR) - except OSError: - pass - try: - sock.close() - except OSError: - pass - closed += 1 - except Exception as exc: - logger.debug("Force-close TCP sockets sweep error: %s", exc) - return closed + """Forwarder — see ``agent.agent_runtime_helpers.force_close_tcp_sockets``.""" + from agent.agent_runtime_helpers import force_close_tcp_sockets + return force_close_tcp_sockets(client) def _close_openai_client(self, client: Any, *, reason: str, shared: bool) -> None: if client is None: @@ -6976,74 +2492,9 @@ class AIAgent: return self.client def _cleanup_dead_connections(self) -> bool: - """Detect and clean up dead TCP connections on the primary client. - - Inspects the httpx connection pool for sockets in unhealthy states - (CLOSE-WAIT, errors). If any are found, force-closes all sockets - and rebuilds the primary client from scratch. - - Returns True if dead connections were found and cleaned up. - """ - client = getattr(self, "client", None) - if client is None: - return False - try: - http_client = getattr(client, "_client", None) - if http_client is None: - return False - transport = getattr(http_client, "_transport", None) - if transport is None: - return False - pool = getattr(transport, "_pool", None) - if pool is None: - return False - connections = ( - getattr(pool, "_connections", None) - or getattr(pool, "_pool", None) - or [] - ) - dead_count = 0 - for conn in list(connections): - # Check for connections that are idle but have closed sockets - stream = ( - getattr(conn, "_network_stream", None) - or getattr(conn, "_stream", None) - ) - if stream is None: - continue - sock = getattr(stream, "_sock", None) - if sock is None: - sock = getattr(stream, "stream", None) - if sock is not None: - sock = getattr(sock, "_sock", None) - if sock is None: - continue - # Probe socket health with a non-blocking recv peek - import socket as _socket - try: - sock.setblocking(False) - data = sock.recv(1, _socket.MSG_PEEK | _socket.MSG_DONTWAIT) - if data == b"": - dead_count += 1 - except BlockingIOError: - pass # No data available — socket is healthy - except OSError: - dead_count += 1 - finally: - try: - sock.setblocking(True) - except OSError: - pass - if dead_count > 0: - logger.warning( - "Found %d dead connection(s) in client pool — rebuilding client", - dead_count, - ) - self._replace_primary_openai_client(reason="dead_connection_cleanup") - return True - except Exception as exc: - logger.debug("Dead connection check error: %s", exc) - return False + """Forwarder — see ``agent.agent_runtime_helpers.cleanup_dead_connections``.""" + from agent.agent_runtime_helpers import cleanup_dead_connections + return cleanup_dead_connections(self) @staticmethod def _api_kwargs_have_image_parts(api_kwargs: dict) -> bool: @@ -7107,237 +2558,14 @@ class AIAgent: self._close_openai_client(client, reason=reason, shared=False) def _run_codex_stream(self, api_kwargs: dict, client: Any = None, on_first_delta: callable = None): - """Execute one streaming Responses API request and return the final response.""" - import httpx as _httpx - - active_client = client or self._ensure_primary_openai_client(reason="codex_stream_direct") - max_stream_retries = 1 - has_tool_calls = False - first_delta_fired = False - # Accumulate streamed text so we can recover if get_final_response() - # returns empty output (e.g. chatgpt.com backend-api sends - # response.incomplete instead of response.completed). - self._codex_streamed_text_parts: list = [] - for attempt in range(max_stream_retries + 1): - if self._interrupt_requested: - raise InterruptedError("Agent interrupted before Codex stream retry") - collected_output_items: list = [] - try: - with active_client.responses.stream(**api_kwargs) as stream: - for event in stream: - self._touch_activity("receiving stream response") - if self._interrupt_requested: - break - event_type = getattr(event, "type", "") - # Fire callbacks on text content deltas (suppress during tool calls) - if "output_text.delta" in event_type or event_type == "response.output_text.delta": - delta_text = getattr(event, "delta", "") - if delta_text: - self._codex_streamed_text_parts.append(delta_text) - if delta_text and not has_tool_calls: - if not first_delta_fired: - first_delta_fired = True - if on_first_delta: - try: - on_first_delta() - except Exception: - pass - self._fire_stream_delta(delta_text) - # Track tool calls to suppress text streaming - elif "function_call" in event_type: - has_tool_calls = True - # Fire reasoning callbacks - elif "reasoning" in event_type and "delta" in event_type: - reasoning_text = getattr(event, "delta", "") - if reasoning_text: - self._fire_reasoning_delta(reasoning_text) - # Collect completed output items — some backends - # (chatgpt.com/backend-api/codex) stream valid items - # via response.output_item.done but the SDK's - # get_final_response() returns an empty output list. - elif event_type == "response.output_item.done": - done_item = getattr(event, "item", None) - if done_item is not None: - collected_output_items.append(done_item) - # Log non-completed terminal events for diagnostics - elif event_type in {"response.incomplete", "response.failed"}: - resp_obj = getattr(event, "response", None) - status = getattr(resp_obj, "status", None) if resp_obj else None - incomplete_details = getattr(resp_obj, "incomplete_details", None) if resp_obj else None - logger.warning( - "Codex Responses stream received terminal event %s " - "(status=%s, incomplete_details=%s, streamed_chars=%d). %s", - event_type, status, incomplete_details, - sum(len(p) for p in self._codex_streamed_text_parts), - self._client_log_context(), - ) - final_response = stream.get_final_response() - # PATCH: ChatGPT Codex backend streams valid output items - # but get_final_response() can return an empty output list. - # Backfill from collected items or synthesize from deltas. - _out = getattr(final_response, "output", None) - if isinstance(_out, list) and not _out: - if collected_output_items: - final_response.output = list(collected_output_items) - logger.debug( - "Codex stream: backfilled %d output items from stream events", - len(collected_output_items), - ) - elif self._codex_streamed_text_parts and not has_tool_calls: - assembled = "".join(self._codex_streamed_text_parts) - final_response.output = [SimpleNamespace( - type="message", - role="assistant", - status="completed", - content=[SimpleNamespace(type="output_text", text=assembled)], - )] - logger.debug( - "Codex stream: synthesized output from %d text deltas (%d chars)", - len(self._codex_streamed_text_parts), len(assembled), - ) - return final_response - except (_httpx.RemoteProtocolError, _httpx.ReadTimeout, _httpx.ConnectError, ConnectionError) as exc: - if attempt < max_stream_retries: - logger.debug( - "Codex Responses stream transport failed (attempt %s/%s); retrying. %s error=%s", - attempt + 1, - max_stream_retries + 1, - self._client_log_context(), - exc, - ) - continue - logger.debug( - "Codex Responses stream transport failed; falling back to create(stream=True). %s error=%s", - self._client_log_context(), - exc, - ) - return self._run_codex_create_stream_fallback(api_kwargs, client=active_client) - except RuntimeError as exc: - err_text = str(exc) - missing_completed = "response.completed" in err_text - # The OpenAI SDK's Responses streaming state machine raises - # ``RuntimeError("Expected to have received `response.created` - # before ``")`` when the first SSE event from the - # server is anything other than ``response.created`` — and it - # discards the event's payload before we can read it. Three - # real-world backends emit a different first frame: - # - # * xAI on grok-4.x OAuth — sends ``error`` (issues - # reported around the May 2026 SuperGrok rollout when - # multi-turn conversations replay encrypted reasoning - # content the OAuth tier rejects) - # * codex-lb relays — send ``codex.rate_limits`` (#14634) - # * custom Responses relays — send ``response.in_progress`` - # (#8133) - # - # In all three cases the underlying byte stream is still - # readable: a non-stream ``responses.create(stream=True)`` - # fallback succeeds and surfaces the real provider error as - # a normal exception with body+status_code attached, which - # ``_summarize_api_error`` can then translate into a useful - # user-facing line. Treat ``response.created`` prelude - # errors the same way we already treat ``response.completed`` - # postlude errors. - prelude_error = ( - "Expected to have received `response.created`" in err_text - or "Expected to have received \"response.created\"" in err_text - ) - if (missing_completed or prelude_error) and attempt < max_stream_retries: - logger.debug( - "Responses stream %s (attempt %s/%s); retrying. %s", - "prelude rejected" if prelude_error else "closed before completion", - attempt + 1, - max_stream_retries + 1, - self._client_log_context(), - ) - continue - if missing_completed or prelude_error: - logger.debug( - "Responses stream %s; falling back to create(stream=True). %s err=%s", - "rejected before response.created" if prelude_error else "did not emit response.completed", - self._client_log_context(), - err_text, - ) - return self._run_codex_create_stream_fallback(api_kwargs, client=active_client) - raise + """Forwarder — see ``agent.codex_runtime.run_codex_stream``.""" + from agent.codex_runtime import run_codex_stream + return run_codex_stream(self, api_kwargs, client, on_first_delta) def _run_codex_create_stream_fallback(self, api_kwargs: dict, client: Any = None): - """Fallback path for stream completion edge cases on Codex-style Responses backends.""" - active_client = client or self._ensure_primary_openai_client(reason="codex_create_stream_fallback") - fallback_kwargs = dict(api_kwargs) - fallback_kwargs["stream"] = True - fallback_kwargs = self._get_transport().preflight_kwargs(fallback_kwargs, allow_stream=True) - stream_or_response = active_client.responses.create(**fallback_kwargs) - - # Compatibility shim for mocks or providers that still return a concrete response. - if hasattr(stream_or_response, "output"): - return stream_or_response - if not hasattr(stream_or_response, "__iter__"): - return stream_or_response - - terminal_response = None - collected_output_items: list = [] - collected_text_deltas: list = [] - try: - for event in stream_or_response: - self._touch_activity("receiving stream response") - event_type = getattr(event, "type", None) - if not event_type and isinstance(event, dict): - event_type = event.get("type") - - # Collect output items and text deltas for backfill - if event_type == "response.output_item.done": - done_item = getattr(event, "item", None) - if done_item is None and isinstance(event, dict): - done_item = event.get("item") - if done_item is not None: - collected_output_items.append(done_item) - elif event_type in {"response.output_text.delta",}: - delta = getattr(event, "delta", "") - if not delta and isinstance(event, dict): - delta = event.get("delta", "") - if delta: - collected_text_deltas.append(delta) - - if event_type not in {"response.completed", "response.incomplete", "response.failed"}: - continue - - terminal_response = getattr(event, "response", None) - if terminal_response is None and isinstance(event, dict): - terminal_response = event.get("response") - if terminal_response is not None: - # Backfill empty output from collected stream events - _out = getattr(terminal_response, "output", None) - if isinstance(_out, list) and not _out: - if collected_output_items: - terminal_response.output = list(collected_output_items) - logger.debug( - "Codex fallback stream: backfilled %d output items", - len(collected_output_items), - ) - elif collected_text_deltas: - assembled = "".join(collected_text_deltas) - terminal_response.output = [SimpleNamespace( - type="message", role="assistant", - status="completed", - content=[SimpleNamespace(type="output_text", text=assembled)], - )] - logger.debug( - "Codex fallback stream: synthesized from %d deltas (%d chars)", - len(collected_text_deltas), len(assembled), - ) - return terminal_response - finally: - close_fn = getattr(stream_or_response, "close", None) - if callable(close_fn): - try: - close_fn() - except Exception: - pass - - if terminal_response is not None: - return terminal_response - raise RuntimeError("Responses create(stream=True) fallback did not emit a terminal response.") + """Forwarder — see ``agent.codex_runtime.run_codex_create_stream_fallback``.""" + from agent.codex_runtime import run_codex_create_stream_fallback + return run_codex_create_stream_fallback(self, api_kwargs, client) def _try_refresh_codex_client_credentials(self, *, force: bool = True) -> bool: if self.api_mode != "codex_responses" or self.provider not in {"openai-codex", "xai-oauth"}: @@ -7418,12 +2646,20 @@ class AIAgent: return False try: - from hermes_cli.auth import resolve_nous_runtime_credentials + from hermes_cli.auth import ( + NOUS_INFERENCE_AUTH_MODE_AUTO, + NOUS_INFERENCE_AUTH_MODE_LEGACY, + resolve_nous_runtime_credentials, + ) creds = resolve_nous_runtime_credentials( min_key_ttl_seconds=max(60, int(os.getenv("HERMES_NOUS_MIN_KEY_TTL_SECONDS", "1800"))), timeout_seconds=float(os.getenv("HERMES_NOUS_TIMEOUT_SECONDS", "15")), - force_mint=force, + inference_auth_mode=( + NOUS_INFERENCE_AUTH_MODE_LEGACY + if force + else NOUS_INFERENCE_AUTH_MODE_AUTO + ), ) except Exception as exc: logger.debug("Nous credential refresh failed: %s", exc) @@ -7615,99 +2851,9 @@ class AIAgent: classified_reason: Optional[FailoverReason] = None, error_context: Optional[Dict[str, Any]] = None, ) -> tuple[bool, bool]: - """Attempt credential recovery via pool rotation. - - Returns (recovered, has_retried_429). - On rate limits: first occurrence retries same credential (sets flag True). - second consecutive failure rotates to next credential. - On billing exhaustion: immediately rotates. - On auth failures: attempts token refresh before rotating. - - `classified_reason` lets the recovery path honor the structured error - classifier instead of relying only on raw HTTP codes. This matters for - providers that surface billing/rate-limit/auth conditions under a - different status code, such as Anthropic returning HTTP 400 for - "out of extra usage". - """ - pool = self._credential_pool - if pool is None: - return False, has_retried_429 - - effective_reason = classified_reason - if effective_reason is None: - if status_code == 402: - effective_reason = FailoverReason.billing - elif status_code == 429: - effective_reason = FailoverReason.rate_limit - elif status_code in {401, 403}: - effective_reason = FailoverReason.auth - - if effective_reason == FailoverReason.billing: - rotate_status = status_code if status_code is not None else 402 - next_entry = pool.mark_exhausted_and_rotate(status_code=rotate_status, error_context=error_context) - if next_entry is not None: - logger.info( - "Credential %s (billing) — rotated to pool entry %s", - rotate_status, - getattr(next_entry, "id", "?"), - ) - self._swap_credential(next_entry) - return True, False - return False, has_retried_429 - - if effective_reason == FailoverReason.rate_limit: - if not has_retried_429: - return False, True - rotate_status = status_code if status_code is not None else 429 - next_entry = pool.mark_exhausted_and_rotate(status_code=rotate_status, error_context=error_context) - if next_entry is not None: - logger.info( - "Credential %s (rate limit) — rotated to pool entry %s", - rotate_status, - getattr(next_entry, "id", "?"), - ) - self._swap_credential(next_entry) - return True, False - return False, True - - if effective_reason == FailoverReason.auth: - # Subscription/entitlement 403s look like auth failures on the - # wire but refresh cannot fix them — the OAuth token is - # already valid; the account simply lacks the entitlement - # (e.g. xAI OAuth without SuperGrok/X Premium for grok-4.3). - # Without this guard, ``try_refresh_current()`` keeps minting - # fresh tokens against the same unsubscribed account and the - # main agent loop spins re-issuing the same 403 until the - # user Ctrl+C's. Surface the error instead so the friendly - # entitlement hint from ``_summarize_api_error`` can land. - if self._is_entitlement_failure(error_context, status_code): - logger.info( - "Credential %s — entitlement-shaped 403 from %s; " - "skipping pool refresh (account lacks subscription, " - "not a transient auth failure).", - status_code if status_code is not None else "auth", - self.provider or "provider", - ) - return False, has_retried_429 - refreshed = pool.try_refresh_current() - if refreshed is not None: - logger.info(f"Credential auth failure — refreshed pool entry {getattr(refreshed, 'id', '?')}") - self._swap_credential(refreshed) - return True, has_retried_429 - # Refresh failed — rotate to next credential instead of giving up. - # The failed entry is already marked exhausted by try_refresh_current(). - rotate_status = status_code if status_code is not None else 401 - next_entry = pool.mark_exhausted_and_rotate(status_code=rotate_status, error_context=error_context) - if next_entry is not None: - logger.info( - "Credential %s (auth refresh failed) — rotated to pool entry %s", - rotate_status, - getattr(next_entry, "id", "?"), - ) - self._swap_credential(next_entry) - return True, False - - return False, has_retried_429 + """Forwarder — see ``agent.agent_runtime_helpers.recover_with_credential_pool``.""" + from agent.agent_runtime_helpers import recover_with_credential_pool + return recover_with_credential_pool(self, status_code=status_code, has_retried_429=has_retried_429, classified_reason=classified_reason, error_context=error_context) def _credential_pool_may_recover_rate_limit(self) -> bool: """Whether a rate-limit retry should wait for same-provider credentials.""" @@ -7756,156 +2902,9 @@ class AIAgent: ) def _interruptible_api_call(self, api_kwargs: dict): - """ - Run the API call in a background thread so the main conversation loop - can detect interrupts without waiting for the full HTTP round-trip. - - Each worker thread gets its own OpenAI client instance. Interrupts only - close that worker-local client, so retries and other requests never - inherit a closed transport. - - Includes a stale-call detector: if no response arrives within the - configured timeout, the connection is killed and an error raised so - the main retry loop can try again with backoff / credential rotation / - provider fallback. - """ - result = {"response": None, "error": None} - request_client_holder = {"client": None} - - def _call(): - try: - if self.api_mode == "codex_responses": - request_client_holder["client"] = self._create_request_openai_client( - reason="codex_stream_request", - api_kwargs=api_kwargs, - ) - result["response"] = self._run_codex_stream( - api_kwargs, - client=request_client_holder["client"], - on_first_delta=getattr(self, "_codex_on_first_delta", None), - ) - elif self.api_mode == "anthropic_messages": - result["response"] = self._anthropic_messages_create(api_kwargs) - elif self.api_mode == "bedrock_converse": - # Bedrock uses boto3 directly — no OpenAI client needed. - # normalize_converse_response produces an OpenAI-compatible - # SimpleNamespace so the rest of the agent loop can treat - # bedrock responses like chat_completions responses. - from agent.bedrock_adapter import ( - _get_bedrock_runtime_client, - invalidate_runtime_client, - is_stale_connection_error, - normalize_converse_response, - ) - region = api_kwargs.pop("__bedrock_region__", "us-east-1") - api_kwargs.pop("__bedrock_converse__", None) - client = _get_bedrock_runtime_client(region) - try: - raw_response = client.converse(**api_kwargs) - except Exception as _bedrock_exc: - # Evict the cached client on stale-connection failures - # so the outer retry loop builds a fresh client/pool. - if is_stale_connection_error(_bedrock_exc): - invalidate_runtime_client(region) - raise - result["response"] = normalize_converse_response(raw_response) - else: - request_client_holder["client"] = self._create_request_openai_client( - reason="chat_completion_request", - api_kwargs=api_kwargs, - ) - result["response"] = request_client_holder["client"].chat.completions.create(**api_kwargs) - except Exception as e: - result["error"] = e - finally: - request_client = request_client_holder.get("client") - if request_client is not None: - self._close_request_openai_client(request_client, reason="request_complete") - - # ── Stale-call timeout (mirrors streaming stale detector) ──────── - # Non-streaming calls return nothing until the full response is - # ready. Without this, a hung provider can block for the full - # httpx timeout (default 1800s) with zero feedback. The stale - # detector kills the connection early so the main retry loop can - # apply richer recovery (credential rotation, provider fallback). - _stale_timeout = self._compute_non_stream_stale_timeout( - api_kwargs.get("messages", []) - ) - - _call_start = time.time() - self._touch_activity("waiting for non-streaming API response") - - t = threading.Thread(target=_call, daemon=True) - t.start() - _poll_count = 0 - while t.is_alive(): - t.join(timeout=0.3) - _poll_count += 1 - - # Touch activity every ~30s so the gateway's inactivity - # monitor knows we're alive while waiting for the response. - if _poll_count % 100 == 0: # 100 × 0.3s = 30s - _elapsed = time.time() - _call_start - self._touch_activity( - f"waiting for non-streaming response ({int(_elapsed)}s elapsed)" - ) - - # Stale-call detector: kill the connection if no response - # arrives within the configured timeout. - _elapsed = time.time() - _call_start - if _elapsed > _stale_timeout: - _est_ctx = sum(len(str(v)) for v in api_kwargs.get("messages", [])) // 4 - logger.warning( - "Non-streaming API call stale for %.0fs (threshold %.0fs). " - "model=%s context=~%s tokens. Killing connection.", - _elapsed, _stale_timeout, - api_kwargs.get("model", "unknown"), f"{_est_ctx:,}", - ) - self._emit_status( - f"⚠️ No response from provider for {int(_elapsed)}s " - f"(non-streaming, model: {api_kwargs.get('model', 'unknown')}). " - f"Aborting call." - ) - try: - if self.api_mode == "anthropic_messages": - self._anthropic_client.close() - self._rebuild_anthropic_client() - else: - rc = request_client_holder.get("client") - if rc is not None: - self._close_request_openai_client(rc, reason="stale_call_kill") - except Exception: - pass - self._touch_activity( - f"stale non-streaming call killed after {int(_elapsed)}s" - ) - # Wait briefly for the thread to notice the closed connection. - t.join(timeout=2.0) - if result["error"] is None and result["response"] is None: - result["error"] = TimeoutError( - f"Non-streaming API call timed out after {int(_elapsed)}s " - f"with no response (threshold: {int(_stale_timeout)}s)" - ) - break - - if self._interrupt_requested: - # Force-close the in-flight worker-local HTTP connection to stop - # token generation without poisoning the shared client used to - # seed future retries. - try: - if self.api_mode == "anthropic_messages": - self._anthropic_client.close() - self._rebuild_anthropic_client() - else: - request_client = request_client_holder.get("client") - if request_client is not None: - self._close_request_openai_client(request_client, reason="interrupt_abort") - except Exception: - pass - raise InterruptedError("Agent interrupted during API call") - if result["error"] is not None: - raise result["error"] - return result["response"] + """Forwarder — see ``agent.chat_completion_helpers.interruptible_api_call``.""" + from agent.chat_completion_helpers import interruptible_api_call + return interruptible_api_call(self, api_kwargs) # ── Unified streaming API call ───────────────────────────────────────── @@ -8076,1306 +3075,28 @@ class AIAgent: def _interruptible_streaming_api_call( self, api_kwargs: dict, *, on_first_delta: callable = None ): - """Streaming variant of _interruptible_api_call for real-time token delivery. - - Handles all three api_modes: - - chat_completions: stream=True on OpenAI-compatible endpoints - - anthropic_messages: client.messages.stream() via Anthropic SDK - - codex_responses: delegates to _run_codex_stream (already streaming) - - Fires stream_delta_callback and _stream_callback for each text token. - Tool-call turns suppress the callback — only text-only final responses - stream to the consumer. Returns a SimpleNamespace that mimics the - non-streaming response shape so the rest of the agent loop is unchanged. - - Falls back to _interruptible_api_call on provider errors indicating - streaming is not supported. - """ - if self._interrupt_requested: - raise InterruptedError("Agent interrupted before streaming API call") - - if self.api_mode == "codex_responses": - # Codex streams internally via _run_codex_stream. The main dispatch - # in _interruptible_api_call already calls it; we just need to - # ensure on_first_delta reaches it. Store it on the instance - # temporarily so _run_codex_stream can pick it up. - self._codex_on_first_delta = on_first_delta - try: - return self._interruptible_api_call(api_kwargs) - finally: - self._codex_on_first_delta = None - - # Bedrock Converse uses boto3's converse_stream() with real-time delta - # callbacks — same UX as Anthropic and chat_completions streaming. - if self.api_mode == "bedrock_converse": - result = {"response": None, "error": None} - first_delta_fired = {"done": False} - deltas_were_sent = {"yes": False} - - def _fire_first(): - if not first_delta_fired["done"] and on_first_delta: - first_delta_fired["done"] = True - try: - on_first_delta() - except Exception: - pass - - def _bedrock_call(): - try: - from agent.bedrock_adapter import ( - _get_bedrock_runtime_client, - invalidate_runtime_client, - is_stale_connection_error, - stream_converse_with_callbacks, - ) - region = api_kwargs.pop("__bedrock_region__", "us-east-1") - api_kwargs.pop("__bedrock_converse__", None) - client = _get_bedrock_runtime_client(region) - try: - raw_response = client.converse_stream(**api_kwargs) - except Exception as _bedrock_exc: - # Evict the cached client on stale-connection failures - # so the outer retry loop builds a fresh client/pool. - if is_stale_connection_error(_bedrock_exc): - invalidate_runtime_client(region) - raise - - def _on_text(text): - _fire_first() - self._fire_stream_delta(text) - deltas_were_sent["yes"] = True - - def _on_tool(name): - _fire_first() - self._fire_tool_gen_started(name) - - def _on_reasoning(text): - _fire_first() - self._fire_reasoning_delta(text) - - result["response"] = stream_converse_with_callbacks( - raw_response, - on_text_delta=_on_text if self._has_stream_consumers() else None, - on_tool_start=_on_tool, - on_reasoning_delta=_on_reasoning if self.reasoning_callback or self.stream_delta_callback else None, - on_interrupt_check=lambda: self._interrupt_requested, - ) - except Exception as e: - result["error"] = e - - t = threading.Thread(target=_bedrock_call, daemon=True) - t.start() - while t.is_alive(): - t.join(timeout=0.3) - if self._interrupt_requested: - raise InterruptedError("Agent interrupted during Bedrock API call") - if result["error"] is not None: - raise result["error"] - return result["response"] - - result = {"response": None, "error": None, "partial_tool_names": []} - request_client_holder = {"client": None, "diag": None} - first_delta_fired = {"done": False} - deltas_were_sent = {"yes": False} # Track if any deltas were fired (for fallback) - # Wall-clock timestamp of the last real streaming chunk. The outer - # poll loop uses this to detect stale connections that keep receiving - # SSE keep-alive pings but no actual data. - last_chunk_time = {"t": time.time()} - - def _fire_first_delta(): - if not first_delta_fired["done"] and on_first_delta: - first_delta_fired["done"] = True - try: - on_first_delta() - except Exception: - pass - - def _call_chat_completions(): - """Stream a chat completions response.""" - import httpx as _httpx - # Per-provider / per-model request_timeout_seconds (from config.yaml) - # wins over the HERMES_API_TIMEOUT env default if the user set it. - _provider_timeout_cfg = get_provider_request_timeout(self.provider, self.model) - _base_timeout = ( - _provider_timeout_cfg - if _provider_timeout_cfg is not None - else float(os.getenv("HERMES_API_TIMEOUT", 1800.0)) - ) - # Read timeout: config wins here too. Otherwise use - # HERMES_STREAM_READ_TIMEOUT (default 120s) for cloud providers. - if _provider_timeout_cfg is not None: - _stream_read_timeout = _provider_timeout_cfg - else: - _stream_read_timeout = float(os.getenv("HERMES_STREAM_READ_TIMEOUT", 120.0)) - # Local providers (Ollama, llama.cpp, vLLM) can take minutes for - # prefill on large contexts before producing the first token. - # Auto-increase the httpx read timeout unless the user explicitly - # overrode HERMES_STREAM_READ_TIMEOUT. - if _stream_read_timeout == 120.0 and self.base_url and is_local_endpoint(self.base_url): - _stream_read_timeout = _base_timeout - logger.debug( - "Local provider detected (%s) — stream read timeout raised to %.0fs", - self.base_url, _stream_read_timeout, - ) - stream_kwargs = { - **api_kwargs, - "stream": True, - "stream_options": {"include_usage": True}, - "timeout": _httpx.Timeout( - connect=30.0, - read=_stream_read_timeout, - write=_base_timeout, - pool=30.0, - ), - } - request_client_holder["client"] = self._create_request_openai_client( - reason="chat_completion_stream_request", - api_kwargs=stream_kwargs, - ) - # Reset stale-stream timer so the detector measures from this - # attempt's start, not a previous attempt's last chunk. - last_chunk_time["t"] = time.time() - self._touch_activity("waiting for provider response (streaming)") - # Initialize per-attempt stream diagnostics so the retry block can - # reach for them after the stream dies. Lives on - # ``request_client_holder["diag"]`` for closure access. - _diag = self._stream_diag_init() - request_client_holder["diag"] = _diag - stream = request_client_holder["client"].chat.completions.create(**stream_kwargs) - - # Capture rate limit headers from the initial HTTP response. - # The OpenAI SDK Stream object exposes the underlying httpx - # response via .response before any chunks are consumed. - self._capture_rate_limits(getattr(stream, "response", None)) - # Snapshot diagnostic headers (cf-ray, x-openrouter-provider, etc.) - # so they survive even when the stream dies before any chunk - # arrives. Best-effort; never raises. - self._stream_diag_capture_response(_diag, getattr(stream, "response", None)) - - # Log OpenRouter response cache status when present. - self._check_openrouter_cache_status(getattr(stream, "response", None)) - - content_parts: list = [] - tool_calls_acc: dict = {} - tool_gen_notified: set = set() - # Ollama-compatible endpoints reuse index 0 for every tool call - # in a parallel batch, distinguishing them only by id. Track - # the last seen id per raw index so we can detect a new tool - # call starting at the same index and redirect it to a fresh slot. - _last_id_at_idx: dict = {} # raw_index -> last seen non-empty id - _active_slot_by_idx: dict = {} # raw_index -> current slot in tool_calls_acc - finish_reason = None - model_name = None - role = "assistant" - reasoning_parts: list = [] - usage_obj = None - for chunk in stream: - last_chunk_time["t"] = time.time() - self._touch_activity("receiving stream response") - - # Update per-attempt diagnostic counters. Best-effort — - # failures are swallowed so the streaming hot path is never - # interrupted by diagnostic accounting. - try: - _diag["chunks"] = int(_diag.get("chunks", 0)) + 1 - if _diag.get("first_chunk_at") is None: - _diag["first_chunk_at"] = last_chunk_time["t"] - # Approximate byte size from the chunk's repr — exact wire - # bytes aren't exposed by the SDK, but len(repr(chunk)) is - # a stable proxy for "how much content arrived" that - # survives stub provider differences. - try: - _diag["bytes"] = int(_diag.get("bytes", 0)) + len(repr(chunk)) - except Exception: - pass - except Exception: - pass - - if self._interrupt_requested: - break - - if not chunk.choices: - if hasattr(chunk, "model") and chunk.model: - model_name = chunk.model - # Usage comes in the final chunk with empty choices - if hasattr(chunk, "usage") and chunk.usage: - usage_obj = chunk.usage - continue - - delta = chunk.choices[0].delta - if hasattr(chunk, "model") and chunk.model: - model_name = chunk.model - - # Accumulate reasoning content - reasoning_text = getattr(delta, "reasoning_content", None) or getattr(delta, "reasoning", None) - if reasoning_text: - reasoning_parts.append(reasoning_text) - _fire_first_delta() - self._fire_reasoning_delta(reasoning_text) - - # Accumulate text content — fire callback only when no tool calls - if delta and delta.content: - content_parts.append(delta.content) - if not tool_calls_acc: - _fire_first_delta() - self._fire_stream_delta(delta.content) - deltas_were_sent["yes"] = True - # Tool calls suppress regular content streaming (avoids - # displaying chatty "I'll use the tool..." text alongside - # tool calls). But reasoning tags embedded in suppressed - # content should still reach the display — otherwise the - # reasoning box only appears as a post-response fallback, - # rendering it confusingly after the already-streamed - # response. Route suppressed content through the stream - # delta callback so its tag extraction can fire the - # reasoning display. Non-reasoning text is harmlessly - # suppressed by the CLI's _stream_delta when the stream - # box is already closed (tool boundary flush). - elif self.stream_delta_callback: - try: - self.stream_delta_callback(delta.content) - self._record_streamed_assistant_text(delta.content) - except Exception: - pass - - # Accumulate tool call deltas — notify display on first name - if delta and delta.tool_calls: - for tc_delta in delta.tool_calls: - raw_idx = tc_delta.index if tc_delta.index is not None else 0 - delta_id = tc_delta.id or "" - - # Ollama fix: detect a new tool call reusing the same - # raw index (different id) and redirect to a fresh slot. - if raw_idx not in _active_slot_by_idx: - _active_slot_by_idx[raw_idx] = raw_idx - if ( - delta_id - and raw_idx in _last_id_at_idx - and delta_id != _last_id_at_idx[raw_idx] - ): - new_slot = max(tool_calls_acc, default=-1) + 1 - _active_slot_by_idx[raw_idx] = new_slot - if delta_id: - _last_id_at_idx[raw_idx] = delta_id - idx = _active_slot_by_idx[raw_idx] - - if idx not in tool_calls_acc: - tool_calls_acc[idx] = { - "id": tc_delta.id or "", - "type": "function", - "function": {"name": "", "arguments": ""}, - "extra_content": None, - } - entry = tool_calls_acc[idx] - if tc_delta.id: - entry["id"] = tc_delta.id - if tc_delta.function: - if tc_delta.function.name: - # Use assignment, not +=. Function names are - # atomic identifiers delivered complete in the - # first chunk (OpenAI spec). Some providers - # (MiniMax M2.7 via NVIDIA NIM) resend the full - # name in every chunk; concatenation would - # produce "read_fileread_file". Assignment - # (matching the OpenAI Node SDK / LiteLLM / - # Vercel AI patterns) is immune to this. - entry["function"]["name"] = tc_delta.function.name - if tc_delta.function.arguments: - entry["function"]["arguments"] += tc_delta.function.arguments - extra = getattr(tc_delta, "extra_content", None) - if extra is None and hasattr(tc_delta, "model_extra"): - extra = (tc_delta.model_extra or {}).get("extra_content") - if extra is not None: - if hasattr(extra, "model_dump"): - extra = extra.model_dump() - entry["extra_content"] = extra - # Fire once per tool when the full name is available - name = entry["function"]["name"] - if name and idx not in tool_gen_notified: - tool_gen_notified.add(idx) - _fire_first_delta() - self._fire_tool_gen_started(name) - # Record the partial tool-call name so the outer - # stub-builder can surface a user-visible warning - # if streaming dies before this tool's arguments - # are fully delivered. Without this, a stall - # during tool-call JSON generation lets the stub - # at line ~6107 return `tool_calls=None`, silently - # discarding the attempted action. - result["partial_tool_names"].append(name) - - if chunk.choices[0].finish_reason: - finish_reason = chunk.choices[0].finish_reason - - # Usage in the final chunk - if hasattr(chunk, "usage") and chunk.usage: - usage_obj = chunk.usage - - # Build mock response matching non-streaming shape - full_content = "".join(content_parts) or None - mock_tool_calls = None - has_truncated_tool_args = False - if tool_calls_acc: - mock_tool_calls = [] - for idx in sorted(tool_calls_acc): - tc = tool_calls_acc[idx] - arguments = tc["function"]["arguments"] - tool_name = tc["function"]["name"] or "?" - if arguments and arguments.strip(): - try: - json.loads(arguments) - except json.JSONDecodeError: - # Attempt repair before flagging as truncated. - # Models like GLM-5.1 via Ollama produce trailing - # commas, unclosed brackets, Python None, etc. - # Without repair, these hit the truncation handler - # and kill the session. _repair_tool_call_arguments - # returns "{}" for unrepairable args, which is far - # better than a crashed session. - repaired = _repair_tool_call_arguments(arguments, tool_name) - if repaired != "{}": - # Successfully repaired — use the fixed args - arguments = repaired - else: - # Unrepairable — flag for truncation handling - has_truncated_tool_args = True - mock_tool_calls.append(SimpleNamespace( - id=tc["id"], - type=tc["type"], - extra_content=tc.get("extra_content"), - function=SimpleNamespace( - name=tc["function"]["name"], - arguments=arguments, - ), - )) - - effective_finish_reason = finish_reason or "stop" - if has_truncated_tool_args: - effective_finish_reason = "length" - - full_reasoning = "".join(reasoning_parts) or None - mock_message = SimpleNamespace( - role=role, - content=full_content, - tool_calls=mock_tool_calls, - reasoning_content=full_reasoning, - ) - mock_choice = SimpleNamespace( - index=0, - message=mock_message, - finish_reason=effective_finish_reason, - ) - return SimpleNamespace( - id="stream-" + str(uuid.uuid4()), - model=model_name, - choices=[mock_choice], - usage=usage_obj, - ) - - def _call_anthropic(): - """Stream an Anthropic Messages API response. - - Fires delta callbacks for real-time token delivery, but returns - the native Anthropic Message object from get_final_message() so - the rest of the agent loop (validation, tool extraction, etc.) - works unchanged. - """ - has_tool_use = False - - # Reset stale-stream timer for this attempt - last_chunk_time["t"] = time.time() - # Per-attempt diagnostic dict for the retry block to consume. - _diag = self._stream_diag_init() - request_client_holder["diag"] = _diag - # Use the Anthropic SDK's streaming context manager - with self._anthropic_client.messages.stream(**api_kwargs) as stream: - # The Anthropic SDK exposes the raw httpx response on - # ``stream.response``. Snapshot diagnostic headers - # immediately so they survive a stream that dies before the - # first event. - try: - self._stream_diag_capture_response( - _diag, getattr(stream, "response", None) - ) - except Exception: - pass - for event in stream: - # Update stale-stream timer on every event so the - # outer poll loop knows data is flowing. Without - # this, the detector kills healthy long-running - # Opus streams after 180 s even when events are - # actively arriving (the chat_completions path - # already does this at the top of its chunk loop). - last_chunk_time["t"] = time.time() - self._touch_activity("receiving stream response") - - # Update per-attempt diagnostic counters (best-effort). - try: - _diag["chunks"] = int(_diag.get("chunks", 0)) + 1 - if _diag.get("first_chunk_at") is None: - _diag["first_chunk_at"] = last_chunk_time["t"] - try: - _diag["bytes"] = int(_diag.get("bytes", 0)) + len(repr(event)) - except Exception: - pass - except Exception: - pass - - if self._interrupt_requested: - break - - event_type = getattr(event, "type", None) - - if event_type == "content_block_start": - block = getattr(event, "content_block", None) - if block and getattr(block, "type", None) == "tool_use": - has_tool_use = True - tool_name = getattr(block, "name", None) - if tool_name: - _fire_first_delta() - self._fire_tool_gen_started(tool_name) - - elif event_type == "content_block_delta": - delta = getattr(event, "delta", None) - if delta: - delta_type = getattr(delta, "type", None) - if delta_type == "text_delta": - text = getattr(delta, "text", "") - if text and not has_tool_use: - _fire_first_delta() - self._fire_stream_delta(text) - deltas_were_sent["yes"] = True - elif delta_type == "thinking_delta": - thinking_text = getattr(delta, "thinking", "") - if thinking_text: - _fire_first_delta() - self._fire_reasoning_delta(thinking_text) - - # Return the native Anthropic Message for downstream processing - return stream.get_final_message() - - def _call(): - import httpx as _httpx - - _max_stream_retries = int(os.getenv("HERMES_STREAM_RETRIES", 2)) - - try: - for _stream_attempt in range(_max_stream_retries + 1): - # Check for interrupt before each retry attempt. Without - # this, /stop closes the HTTP connection (outer poll loop), - # but the retry loop opens a FRESH connection — negating the - # interrupt entirely. On slow providers (ollama-cloud) each - # retry can block for the full stream-read timeout (120s+), - # causing multi-minute delays between /stop and response. - if self._interrupt_requested: - raise InterruptedError("Agent interrupted before stream retry") - try: - if self.api_mode == "anthropic_messages": - self._try_refresh_anthropic_client_credentials() - result["response"] = _call_anthropic() - else: - result["response"] = _call_chat_completions() - return # success - except Exception as e: - _is_timeout = isinstance( - e, (_httpx.ReadTimeout, _httpx.ConnectTimeout, _httpx.PoolTimeout) - ) - _is_conn_err = isinstance( - e, (_httpx.ConnectError, _httpx.RemoteProtocolError, ConnectionError) - ) - _is_stream_parse_err = self._is_provider_stream_parse_error(e) - - # If the stream died AFTER some tokens were delivered: - # normally we don't retry (the user already saw text, - # retrying would duplicate it). BUT: if a tool call - # was in-flight when the stream died, silently aborting - # discards the tool call entirely. In that case we - # prefer to retry — the user sees a brief - # "reconnecting" marker + duplicated preamble text, - # which is strictly better than a failed action with - # a "retry manually" message. Limit this to transient - # connection errors (Clawdbot-style narrow gate): no - # tool has executed yet within this API call, so - # silent retry is safe wrt side-effects. - if deltas_were_sent["yes"]: - _partial_tool_in_flight = bool( - result.get("partial_tool_names") - ) - _is_sse_conn_err_preview = False - if not _is_timeout and not _is_conn_err: - from openai import APIError as _APIError - if isinstance(e, _APIError) and not getattr(e, "status_code", None): - _err_lower_preview = str(e).lower() - _SSE_PREVIEW_PHRASES = ( - "connection lost", - "connection reset", - "connection closed", - "connection terminated", - "network error", - "network connection", - "terminated", - "peer closed", - "broken pipe", - "upstream connect error", - ) - _is_sse_conn_err_preview = any( - phrase in _err_lower_preview - for phrase in _SSE_PREVIEW_PHRASES - ) - _is_transient = ( - _is_timeout - or _is_conn_err - or _is_sse_conn_err_preview - or _is_stream_parse_err - ) - _can_silent_retry = ( - _partial_tool_in_flight - and _is_transient - and _stream_attempt < _max_stream_retries - ) - if not _can_silent_retry: - # Either no tool call was in-flight (so the - # turn was a pure text response — current - # stub-with-recovered-text behaviour is - # correct), or retries are exhausted, or the - # error isn't transient. Fall through to the - # stub path. - logger.warning( - "Streaming failed after partial delivery, not retrying: %s", e - ) - result["error"] = e - return - # Tool call was in-flight AND error is transient: - # retry silently. Clear per-attempt state so the - # next stream starts clean. Fire a "reconnecting" - # marker so the user sees why the preamble is - # about to be re-streamed. Structured WARNING is - # emitted by ``_emit_stream_drop`` below; no - # additional INFO line needed. - try: - self._fire_stream_delta( - "\n\n⚠ Connection dropped mid tool-call; " - "reconnecting…\n\n" - ) - except Exception: - pass - # Reset the streamed-text buffer so the retry's - # fresh preamble doesn't get double-recorded in - # _current_streamed_assistant_text (which would - # pollute the interim-visible-text comparison). - try: - self._reset_stream_delivery_tracking() - except Exception: - pass - # Reset in-memory accumulators so the next - # attempt's chunks don't concat onto the dead - # stream's partial JSON. - result["partial_tool_names"] = [] - deltas_were_sent["yes"] = False - first_delta_fired["done"] = False - self._emit_stream_drop( - error=e, - attempt=_stream_attempt + 2, - max_attempts=_max_stream_retries + 1, - mid_tool_call=True, - diag=request_client_holder.get("diag"), - ) - stale = request_client_holder.get("client") - if stale is not None: - self._close_request_openai_client( - stale, reason="stream_mid_tool_retry_cleanup" - ) - request_client_holder["client"] = None - try: - self._replace_primary_openai_client( - reason="stream_mid_tool_retry_pool_cleanup" - ) - except Exception: - pass - continue - - # SSE error events from proxies (e.g. OpenRouter sends - # {"error":{"message":"Network connection lost."}}) are - # raised as APIError by the OpenAI SDK. These are - # semantically identical to httpx connection drops — - # the upstream stream died — and should be retried with - # a fresh connection. Distinguish from HTTP errors: - # APIError from SSE has no status_code, while - # APIStatusError (4xx/5xx) always has one. - _is_sse_conn_err = False - if not _is_timeout and not _is_conn_err: - from openai import APIError as _APIError - if isinstance(e, _APIError) and not getattr(e, "status_code", None): - _err_lower_sse = str(e).lower() - _SSE_CONN_PHRASES = ( - "connection lost", - "connection reset", - "connection closed", - "connection terminated", - "network error", - "network connection", - "terminated", - "peer closed", - "broken pipe", - "upstream connect error", - ) - _is_sse_conn_err = any( - phrase in _err_lower_sse - for phrase in _SSE_CONN_PHRASES - ) - - if _is_timeout or _is_conn_err or _is_sse_conn_err or _is_stream_parse_err: - # Transient network / timeout error. Retry the - # streaming request with a fresh connection first. - if _stream_attempt < _max_stream_retries: - self._emit_stream_drop( - error=e, - attempt=_stream_attempt + 2, - max_attempts=_max_stream_retries + 1, - mid_tool_call=False, - diag=request_client_holder.get("diag"), - ) - # Close the stale request client before retry - stale = request_client_holder.get("client") - if stale is not None: - self._close_request_openai_client( - stale, reason="stream_retry_cleanup" - ) - request_client_holder["client"] = None - # Also rebuild the primary client to purge - # any dead connections from the pool. - try: - self._replace_primary_openai_client( - reason="stream_retry_pool_cleanup" - ) - except Exception: - pass - continue - # Retries exhausted. Log the final failure with - # full diagnostic detail (chain, headers, - # bytes/elapsed) via the same helper used for - # mid-flight retries — subagent lines get the - # ``[subagent-N]`` log_prefix so the parent can - # attribute them. - self._log_stream_retry( - kind="exhausted", - error=e, - attempt=_max_stream_retries + 1, - max_attempts=_max_stream_retries + 1, - mid_tool_call=False, - diag=request_client_holder.get("diag"), - ) - if _is_stream_parse_err: - self._emit_status( - "❌ Provider returned malformed streaming data after " - f"{_max_stream_retries + 1} attempts. " - "The provider may be experiencing issues — " - "try again in a moment." - ) - else: - self._emit_status( - "❌ Connection to provider failed after " - f"{_max_stream_retries + 1} attempts. " - "The provider may be experiencing issues — " - "try again in a moment." - ) - else: - _err_lower = str(e).lower() - _is_stream_unsupported = ( - "stream" in _err_lower - and "not supported" in _err_lower - ) - if _is_stream_unsupported: - self._disable_streaming = True - self._safe_print( - "\n⚠ Streaming is not supported for this " - "model/provider. Switching to non-streaming.\n" - " To avoid this delay, set display.streaming: false " - "in config.yaml\n" - ) - logger.info( - "Streaming failed before delivery: %s", - e, - ) - - # Propagate the error to the main retry loop instead of - # falling back to non-streaming inline. The main loop has - # richer recovery: credential rotation, provider fallback, - # backoff, and — for "stream not supported" — will switch - # to non-streaming on the next attempt via _disable_streaming. - result["error"] = e - return - except InterruptedError as e: - # The interrupt may be noticed inside the worker thread before - # the polling loop sees it. Surface it through the normal result - # channel so callers never miss a fast pre-retry interrupt. - result["error"] = e - return - finally: - request_client = request_client_holder.get("client") - if request_client is not None: - self._close_request_openai_client(request_client, reason="stream_request_complete") - - _stream_stale_timeout_base = float(os.getenv("HERMES_STREAM_STALE_TIMEOUT", 180.0)) - # Local providers (Ollama, oMLX, llama-cpp) can take 300+ seconds - # for prefill on large contexts. Disable the stale detector unless - # the user explicitly set HERMES_STREAM_STALE_TIMEOUT. - if _stream_stale_timeout_base == 180.0 and self.base_url and is_local_endpoint(self.base_url): - _stream_stale_timeout = float("inf") - logger.debug("Local provider detected (%s) — stale stream timeout disabled", self.base_url) - else: - # Scale the stale timeout for large contexts: slow models (like Opus) - # can legitimately think for minutes before producing the first token - # when the context is large. Without this, the stale detector kills - # healthy connections during the model's thinking phase, producing - # spurious RemoteProtocolError ("peer closed connection"). - _est_tokens = sum(len(str(v)) for v in api_kwargs.get("messages", [])) // 4 - if _est_tokens > 100_000: - _stream_stale_timeout = max(_stream_stale_timeout_base, 300.0) - elif _est_tokens > 50_000: - _stream_stale_timeout = max(_stream_stale_timeout_base, 240.0) - else: - _stream_stale_timeout = _stream_stale_timeout_base - - t = threading.Thread(target=_call, daemon=True) - t.start() - _last_heartbeat = time.time() - _HEARTBEAT_INTERVAL = 30.0 # seconds between gateway activity touches - while t.is_alive(): - t.join(timeout=0.3) - - # Periodic heartbeat: touch the agent's activity tracker so the - # gateway's inactivity monitor knows we're alive while waiting - # for stream chunks. Without this, long thinking pauses (e.g. - # reasoning models) or slow prefill on local providers (Ollama) - # trigger false inactivity timeouts. The _call thread touches - # activity on each chunk, but the gap between API call start - # and first chunk can exceed the gateway timeout — especially - # when the stale-stream timeout is disabled (local providers). - _hb_now = time.time() - if _hb_now - _last_heartbeat >= _HEARTBEAT_INTERVAL: - _last_heartbeat = _hb_now - _waiting_secs = int(_hb_now - last_chunk_time["t"]) - self._touch_activity( - f"waiting for stream response ({_waiting_secs}s, no chunks yet)" - ) - - # Detect stale streams: connections kept alive by SSE pings - # but delivering no real chunks. Kill the client so the - # inner retry loop can start a fresh connection. - _stale_elapsed = time.time() - last_chunk_time["t"] - if _stale_elapsed > _stream_stale_timeout: - _est_ctx = sum(len(str(v)) for v in api_kwargs.get("messages", [])) // 4 - logger.warning( - "Stream stale for %.0fs (threshold %.0fs) — no chunks received. " - "model=%s context=~%s tokens. Killing connection.", - _stale_elapsed, _stream_stale_timeout, - api_kwargs.get("model", "unknown"), f"{_est_ctx:,}", - ) - self._emit_status( - f"⚠️ No response from provider for {int(_stale_elapsed)}s " - f"(model: {api_kwargs.get('model', 'unknown')}, " - f"context: ~{_est_ctx:,} tokens). " - f"Reconnecting..." - ) - try: - rc = request_client_holder.get("client") - if rc is not None: - self._close_request_openai_client(rc, reason="stale_stream_kill") - except Exception: - pass - # Rebuild the primary client too — its connection pool - # may hold dead sockets from the same provider outage. - try: - self._replace_primary_openai_client(reason="stale_stream_pool_cleanup") - except Exception: - pass - # Reset the timer so we don't kill repeatedly while - # the inner thread processes the closure. - last_chunk_time["t"] = time.time() - self._touch_activity( - f"stale stream detected after {int(_stale_elapsed)}s, reconnecting" - ) - - if self._interrupt_requested: - try: - if self.api_mode == "anthropic_messages": - self._anthropic_client.close() - self._rebuild_anthropic_client() - else: - request_client = request_client_holder.get("client") - if request_client is not None: - self._close_request_openai_client(request_client, reason="stream_interrupt_abort") - except Exception: - pass - raise InterruptedError("Agent interrupted during streaming API call") - if result["error"] is not None: - if deltas_were_sent["yes"]: - # Streaming failed AFTER some tokens were already delivered to - # the platform. Re-raising would let the outer retry loop make - # a new API call, creating a duplicate message. Return a - # partial "stop" response instead so the outer loop treats this - # turn as complete (no retry, no fallback). - # Recover whatever content was already streamed to the user. - # _current_streamed_assistant_text accumulates text fired - # through _fire_stream_delta, so it has exactly what the - # user saw before the connection died. - _partial_text = ( - getattr(self, "_current_streamed_assistant_text", "") or "" - ).strip() or None - - # If the stream died while the model was emitting a tool call, - # the stub below will silently set `tool_calls=None` and the - # agent loop will treat the turn as complete — the attempted - # action is lost with no user-facing signal. Append a - # human-visible warning to the stub content so (a) the user - # knows something failed, and (b) the next turn's model sees - # in conversation history what was attempted and can retry. - _partial_names = list(result.get("partial_tool_names") or []) - if _partial_names: - _name_str = ", ".join(_partial_names[:3]) - if len(_partial_names) > 3: - _name_str += f", +{len(_partial_names) - 3} more" - _warn = ( - f"\n\n⚠ Stream stalled mid tool-call " - f"({_name_str}); the action was not executed. " - f"Ask me to retry if you want to continue." - ) - _partial_text = (_partial_text or "") + _warn - # Also fire as a streaming delta so the user sees it now - # instead of only in the persisted transcript. - try: - self._fire_stream_delta(_warn) - except Exception: - pass - logger.warning( - "Partial stream dropped tool call(s) %s after %s chars " - "of text; surfaced warning to user: %s", - _partial_names, len(_partial_text or ""), result["error"], - ) - else: - logger.warning( - "Partial stream delivered before error; returning stub " - "response with %s chars of recovered content to prevent " - "duplicate messages: %s", - len(_partial_text or ""), - result["error"], - ) - _stub_msg = SimpleNamespace( - role="assistant", content=_partial_text, tool_calls=None, - reasoning_content=None, - ) - return SimpleNamespace( - id="partial-stream-stub", - model=getattr(self, "model", "unknown"), - choices=[SimpleNamespace( - index=0, message=_stub_msg, finish_reason="stop", - )], - usage=None, - ) - raise result["error"] - return result["response"] - - # ── Provider fallback ────────────────────────────────────────────────── + """Forwarder — see ``agent.chat_completion_helpers.interruptible_streaming_api_call``.""" + from agent.chat_completion_helpers import interruptible_streaming_api_call + return interruptible_streaming_api_call(self, api_kwargs, on_first_delta=on_first_delta) def _try_activate_fallback(self, reason: "FailoverReason | None" = None) -> bool: - """Switch to the next fallback model/provider in the chain. - - Called when the current model is failing after retries. Swaps the - OpenAI client, model slug, and provider in-place so the retry loop - can continue with the new backend. Advances through the chain on - each call; returns False when exhausted. - - Uses the centralized provider router (resolve_provider_client) for - auth resolution and client construction — no duplicated provider→key - mappings. - """ - if reason in {FailoverReason.rate_limit, FailoverReason.billing}: - # Only start cooldown when leaving the primary provider. If we're - # already on a fallback and chain-switching, the primary wasn't the - # source of the 429 so the cooldown should not be reset/extended. - fallback_already_active = bool(getattr(self, "_fallback_activated", False)) - current_provider = (getattr(self, "provider", "") or "").strip().lower() - primary_provider = ((self._primary_runtime or {}).get("provider") or "").strip().lower() - if (not fallback_already_active) or (primary_provider and current_provider == primary_provider): - self._rate_limited_until = time.monotonic() + 60 - if self._fallback_index >= len(self._fallback_chain): - return False - - fb = self._fallback_chain[self._fallback_index] - self._fallback_index += 1 - fb_provider = (fb.get("provider") or "").strip().lower() - fb_model = (fb.get("model") or "").strip() - if not fb_provider or not fb_model: - return self._try_activate_fallback() # skip invalid, try next - - # Skip entries that resolve to the current (provider, model) — falling - # back to the same backend that just failed loops the failure. Compare - # base_url too so two distinct custom_providers entries pointing at the - # same shim/proxy URL also dedup. See issue #22548. - current_provider = (getattr(self, "provider", "") or "").strip().lower() - current_model = (getattr(self, "model", "") or "").strip() - current_base_url = str(getattr(self, "base_url", "") or "").rstrip("/").lower() - fb_base_url_for_dedup = (fb.get("base_url") or "").strip().rstrip("/").lower() - if fb_provider == current_provider and fb_model == current_model: - logging.warning( - "Fallback skip: chain entry %s/%s matches current provider/model", - fb_provider, fb_model, - ) - return self._try_activate_fallback() - if ( - fb_base_url_for_dedup - and current_base_url - and fb_base_url_for_dedup == current_base_url - and fb_model == current_model - ): - logging.warning( - "Fallback skip: chain entry base_url %s matches current backend", - fb_base_url_for_dedup, - ) - return self._try_activate_fallback() - - # Use centralized router for client construction. - # raw_codex=True because the main agent needs direct responses.stream() - # access for Codex providers. - try: - from agent.auxiliary_client import resolve_provider_client - # Pass base_url and api_key from fallback config so custom - # endpoints (e.g. Ollama Cloud) resolve correctly instead of - # falling through to OpenRouter defaults. - fb_base_url_hint = (fb.get("base_url") or "").strip() or None - fb_api_key_hint = (fb.get("api_key") or "").strip() or None - if not fb_api_key_hint: - # key_env and api_key_env are both documented aliases (see - # _normalize_custom_provider_entry in hermes_cli/config.py). - fb_key_env = (fb.get("key_env") or fb.get("api_key_env") or "").strip() - if fb_key_env: - fb_api_key_hint = os.getenv(fb_key_env, "").strip() or None - # For Ollama Cloud endpoints, pull OLLAMA_API_KEY from env - # when no explicit key is in the fallback config. Host match - # (not substring) — see GHSA-76xc-57q6-vm5m. - if fb_base_url_hint and base_url_host_matches(fb_base_url_hint, "ollama.com") and not fb_api_key_hint: - fb_api_key_hint = os.getenv("OLLAMA_API_KEY") or None - fb_client, _resolved_fb_model = resolve_provider_client( - fb_provider, model=fb_model, raw_codex=True, - explicit_base_url=fb_base_url_hint, - explicit_api_key=fb_api_key_hint) - if fb_client is None: - logging.warning( - "Fallback to %s failed: provider not configured", - fb_provider) - return self._try_activate_fallback() # try next in chain - try: - from hermes_cli.model_normalize import normalize_model_for_provider - - fb_model = normalize_model_for_provider(fb_model, fb_provider) - except Exception: - pass - - # Determine api_mode from provider / base URL / model - fb_api_mode = "chat_completions" - fb_base_url = str(fb_client.base_url) - _fb_is_azure = self._is_azure_openai_url(fb_base_url) - if fb_provider == "openai-codex": - fb_api_mode = "codex_responses" - elif fb_provider == "anthropic" or fb_base_url.rstrip("/").lower().endswith("/anthropic"): - fb_api_mode = "anthropic_messages" - elif _fb_is_azure: - # Azure OpenAI serves gpt-5.x on /chat/completions — does NOT - # support the Responses API. Stay on chat_completions. - fb_api_mode = "chat_completions" - elif self._is_direct_openai_url(fb_base_url): - fb_api_mode = "codex_responses" - elif self._provider_model_requires_responses_api( - fb_model, - provider=fb_provider, - ): - # GPT-5.x models usually need Responses API, but keep - # provider-specific exceptions like Copilot gpt-5-mini on - # chat completions. - fb_api_mode = "codex_responses" - elif fb_provider == "bedrock" or ( - base_url_hostname(fb_base_url).startswith("bedrock-runtime.") - and base_url_host_matches(fb_base_url, "amazonaws.com") - ): - fb_api_mode = "bedrock_converse" - - old_model = self.model - - # Clear the per-config context_length override so the fallback - # model's actual context window is resolved instead of inheriting - # the stale value from the previous model. See #22387. - self._config_context_length = None - self.model = fb_model - self.provider = fb_provider - self.base_url = fb_base_url - self.api_mode = fb_api_mode - if hasattr(self, "_transport_cache"): - self._transport_cache.clear() - self._fallback_activated = True - - # Honor per-provider / per-model request_timeout_seconds for the - # fallback target (same knob the primary client uses). None = use - # SDK default. - _fb_timeout = get_provider_request_timeout(fb_provider, fb_model) - - if fb_api_mode == "anthropic_messages": - # Build native Anthropic client instead of using OpenAI client - from agent.anthropic_adapter import build_anthropic_client, resolve_anthropic_token, _is_oauth_token - effective_key = (fb_client.api_key or resolve_anthropic_token() or "") if fb_provider == "anthropic" else (fb_client.api_key or "") - self.api_key = effective_key - self._anthropic_api_key = effective_key - self._anthropic_base_url = fb_base_url - self._anthropic_client = build_anthropic_client( - effective_key, self._anthropic_base_url, timeout=_fb_timeout, - ) - self._is_anthropic_oauth = _is_oauth_token(effective_key) if fb_provider == "anthropic" else False - self.client = None - self._client_kwargs = {} - else: - # Swap OpenAI client and config in-place - self.api_key = fb_client.api_key - self.client = fb_client - # Preserve provider-specific headers that - # resolve_provider_client() may have baked into - # fb_client via the default_headers kwarg. The OpenAI - # SDK stores these in _custom_headers. Without this, - # subsequent request-client rebuilds (via - # _create_request_openai_client) drop the headers, - # causing 403s from providers like Kimi Coding that - # require a User-Agent sentinel. - fb_headers = getattr(fb_client, "_custom_headers", None) - if not fb_headers: - fb_headers = getattr(fb_client, "default_headers", None) - self._client_kwargs = { - "api_key": fb_client.api_key, - "base_url": fb_base_url, - **({"default_headers": dict(fb_headers)} if fb_headers else {}), - } - if _fb_timeout is not None: - self._client_kwargs["timeout"] = _fb_timeout - # Rebuild the shared OpenAI client so the configured - # timeout takes effect on the very next fallback request, - # not only after a later credential-rotation rebuild. - self._replace_primary_openai_client(reason="fallback_timeout_apply") - - # Re-evaluate prompt caching for the new provider/model - self._use_prompt_caching, self._use_native_cache_layout = ( - self._anthropic_prompt_cache_policy( - provider=fb_provider, - base_url=fb_base_url, - api_mode=fb_api_mode, - model=fb_model, - ) - ) - - # LM Studio: preload before probing the fallback's context length. - self._ensure_lmstudio_runtime_loaded() - - # Update context compressor limits for the fallback model. - # Without this, compression decisions use the primary model's - # context window (e.g. 200K) instead of the fallback's (e.g. 32K), - # causing oversized sessions to overflow the fallback. - # Also pass _config_context_length so the explicit config override - # (model.context_length in config.yaml) is respected — without this, - # the fallback activation drops to 128K even when config says 204800. - if hasattr(self, 'context_compressor') and self.context_compressor: - from agent.model_metadata import get_model_context_length - fb_context_length = get_model_context_length( - self.model, base_url=self.base_url, - api_key=self.api_key, provider=self.provider, - config_context_length=getattr(self, "_config_context_length", None), - custom_providers=self._custom_providers, - ) - self.context_compressor.update_model( - model=self.model, - context_length=fb_context_length, - base_url=self.base_url, - api_key=getattr(self, "api_key", ""), - provider=self.provider, - ) - - self._emit_status( - f"🔄 Primary model failed — switching to fallback: " - f"{fb_model} via {fb_provider}" - ) - logging.info( - "Fallback activated: %s → %s (%s)", - old_model, fb_model, fb_provider, - ) - return True - except Exception as e: - logging.error("Failed to activate fallback %s: %s", fb_model, e) - return self._try_activate_fallback() # try next in chain + """Forwarder — see ``agent.chat_completion_helpers.try_activate_fallback``.""" + from agent.chat_completion_helpers import try_activate_fallback + return try_activate_fallback(self, reason) # ── Per-turn primary restoration ───────────────────────────────────── def _restore_primary_runtime(self) -> bool: - """Restore the primary runtime at the start of a new turn. - - In long-lived CLI sessions a single AIAgent instance spans multiple - turns. Without restoration, one transient failure pins the session - to the fallback provider for every subsequent turn. Calling this at - the top of ``run_conversation()`` makes fallback turn-scoped. - - The gateway caches agents across messages (``_agent_cache`` in - ``gateway/run.py``), so this restoration IS needed there too. - """ - if not self._fallback_activated: - return False - - if getattr(self, "_rate_limited_until", 0) > time.monotonic(): - return False # primary still in rate-limit cooldown, stay on fallback - - rt = self._primary_runtime - try: - # ── Core runtime state ── - self.model = rt["model"] - self.provider = rt["provider"] - self.base_url = rt["base_url"] # setter updates _base_url_lower - self.api_mode = rt["api_mode"] - if hasattr(self, "_transport_cache"): - self._transport_cache.clear() - self.api_key = rt["api_key"] - self._client_kwargs = dict(rt["client_kwargs"]) - self._use_prompt_caching = rt["use_prompt_caching"] - # Default to native layout when the restored snapshot predates the - # native-vs-proxy split (older sessions saved before this PR). - self._use_native_cache_layout = rt.get( - "use_native_cache_layout", - self.api_mode == "anthropic_messages" and self.provider == "anthropic", - ) - - # ── Rebuild client for the primary provider ── - if self.api_mode == "anthropic_messages": - from agent.anthropic_adapter import build_anthropic_client - self._anthropic_api_key = rt["anthropic_api_key"] - self._anthropic_base_url = rt["anthropic_base_url"] - self._anthropic_client = build_anthropic_client( - rt["anthropic_api_key"], rt["anthropic_base_url"], - timeout=get_provider_request_timeout(self.provider, self.model), - ) - self._is_anthropic_oauth = rt["is_anthropic_oauth"] - self.client = None - else: - self.client = self._create_openai_client( - dict(rt["client_kwargs"]), - reason="restore_primary", - shared=True, - ) - - # ── Restore context engine state ── - cc = self.context_compressor - cc.update_model( - model=rt["compressor_model"], - context_length=rt["compressor_context_length"], - base_url=rt["compressor_base_url"], - api_key=rt["compressor_api_key"], - provider=rt["compressor_provider"], - ) - - # ── Reset fallback chain for the new turn ── - self._fallback_activated = False - self._fallback_index = 0 - - logging.info( - "Primary runtime restored for new turn: %s (%s)", - self.model, self.provider, - ) - return True - except Exception as e: - logging.warning("Failed to restore primary runtime: %s", e) - return False - - # Which error types indicate a transient transport failure worth - # one more attempt with a rebuilt client / connection pool. - _TRANSIENT_TRANSPORT_ERRORS = frozenset({ - "ReadTimeout", "ConnectTimeout", "PoolTimeout", - "ConnectError", "RemoteProtocolError", - "APIConnectionError", "APITimeoutError", - }) + """Forwarder — see ``agent.agent_runtime_helpers.restore_primary_runtime``.""" + from agent.agent_runtime_helpers import restore_primary_runtime + return restore_primary_runtime(self) def _try_recover_primary_transport( self, api_error: Exception, *, retry_count: int, max_retries: int, ) -> bool: - """Attempt one extra primary-provider recovery cycle for transient transport failures. - - After ``max_retries`` exhaust, rebuild the primary client (clearing - stale connection pools) and give it one more attempt before falling - back. This is most useful for direct endpoints (custom, Z.AI, - Anthropic, OpenAI, local models) where a TCP-level hiccup does not - mean the provider is down. - - Skipped for proxy/aggregator providers (OpenRouter, Nous) which - already manage connection pools and retries server-side — if our - retries through them are exhausted, one more rebuilt client won't help. - """ - if self._fallback_activated: - return False - - # Only for transient transport errors - error_type = type(api_error).__name__ - if error_type not in self._TRANSIENT_TRANSPORT_ERRORS: - return False - - # Skip for aggregator providers — they manage their own retry infra - if self._is_openrouter_url(): - return False - provider_lower = (self.provider or "").strip().lower() - if provider_lower in {"nous", "nous-research"}: - return False - - try: - # Close existing client to release stale connections - if getattr(self, "client", None) is not None: - try: - self._close_openai_client( - self.client, reason="primary_recovery", shared=True, - ) - except Exception: - pass - - # Rebuild from primary snapshot - rt = self._primary_runtime - self._client_kwargs = dict(rt["client_kwargs"]) - self.model = rt["model"] - self.provider = rt["provider"] - self.base_url = rt["base_url"] - self.api_mode = rt["api_mode"] - if hasattr(self, "_transport_cache"): - self._transport_cache.clear() - self.api_key = rt["api_key"] - - if self.api_mode == "anthropic_messages": - from agent.anthropic_adapter import build_anthropic_client - self._anthropic_api_key = rt["anthropic_api_key"] - self._anthropic_base_url = rt["anthropic_base_url"] - self._anthropic_client = build_anthropic_client( - rt["anthropic_api_key"], rt["anthropic_base_url"], - timeout=get_provider_request_timeout(self.provider, self.model), - ) - self._is_anthropic_oauth = rt["is_anthropic_oauth"] - self.client = None - else: - self.client = self._create_openai_client( - dict(rt["client_kwargs"]), - reason="primary_recovery", - shared=True, - ) - - wait_time = min(3 + retry_count, 8) - self._vprint( - f"{self.log_prefix}🔁 Transient {error_type} on {self.provider} — " - f"rebuilt client, waiting {wait_time}s before one last primary attempt.", - force=True, - ) - time.sleep(wait_time) - return True - except Exception as e: - logging.warning("Primary transport recovery failed: %s", e) - return False - - # ── End provider fallback ────────────────────────────────────────────── + """Forwarder — see ``agent.agent_runtime_helpers.try_recover_primary_transport``.""" + from agent.agent_runtime_helpers import try_recover_primary_transport + return try_recover_primary_transport(self, api_error, retry_count=retry_count, max_retries=max_retries) @staticmethod def _content_has_image_parts(content: Any) -> bool: @@ -9650,116 +3371,9 @@ class AIAgent: return summary def _try_shrink_image_parts_in_messages(self, api_messages: list) -> bool: - """Re-encode all native image parts at a smaller size to recover from - image-too-large errors (Anthropic 5 MB, unknown other providers). - - Mutates ``api_messages`` in place. Returns True if any image part was - actually replaced, False if there were no image parts to shrink or - Pillow couldn't help (caller should surface the original error). - - Strategy: look for ``image_url`` / ``input_image`` parts carrying a - ``data:image/...;base64,...`` payload. For each one whose encoded - size exceeds 4 MB (a safe target that slides under Anthropic's 5 MB - ceiling with header overhead), write the base64 to a tempfile, call - ``vision_tools._resize_image_for_vision`` to produce a smaller data - URL, and substitute it in place. - - Non-data-URL images (http/https URLs) are not touched — the provider - fetches those itself and the size limit is different. - """ - if not api_messages: - return False - - try: - from tools.vision_tools import _resize_image_for_vision - except Exception as exc: - logger.warning("image-shrink recovery: vision_tools unavailable — %s", exc) - return False - - # 4 MB target leaves comfortable headroom under Anthropic's 5 MB. - # Non-Anthropic providers we haven't observed rejecting are fine with - # much larger; shrinking to 4 MB here loses quality but only fires - # after a confirmed provider rejection, so the alternative is failure. - target_bytes = 4 * 1024 * 1024 - changed_count = 0 - - def _shrink_data_url(url: str) -> Optional[str]: - """Return a smaller data URL, or None if shrink can't help.""" - if not isinstance(url, str) or not url.startswith("data:"): - return None - if len(url) <= target_bytes: - # This specific image wasn't the oversized one. - return None - try: - header, _, data = url.partition(",") - mime = "image/jpeg" - if header.startswith("data:"): - mime_part = header[len("data:"):].split(";", 1)[0].strip() - if mime_part.startswith("image/"): - mime = mime_part - import base64 as _b64 - raw = _b64.b64decode(data) - suffix = { - "image/png": ".png", "image/gif": ".gif", "image/webp": ".webp", - "image/jpeg": ".jpg", "image/jpg": ".jpg", "image/bmp": ".bmp", - }.get(mime, ".jpg") - tmp = tempfile.NamedTemporaryFile( - prefix="hermes_shrink_", suffix=suffix, delete=False, - ) - try: - tmp.write(raw) - tmp.close() - resized = _resize_image_for_vision( - Path(tmp.name), - mime_type=mime, - max_base64_bytes=target_bytes, - ) - finally: - try: - Path(tmp.name).unlink(missing_ok=True) - except Exception: - pass - if not resized or len(resized) >= len(url): - # Shrink didn't help (or made it bigger — corrupt input?). - return None - return resized - except Exception as exc: - logger.warning("image-shrink recovery: re-encode failed — %s", exc) - return None - - for msg in api_messages: - if not isinstance(msg, dict): - continue - content = msg.get("content") - if not isinstance(content, list): - continue - for part in content: - if not isinstance(part, dict): - continue - ptype = part.get("type") - if ptype not in {"image_url", "input_image"}: - continue - image_value = part.get("image_url") - # OpenAI chat.completions: {"image_url": {"url": "data:..."}} - # OpenAI Responses: {"image_url": "data:..."} - if isinstance(image_value, dict): - url = image_value.get("url", "") - resized = _shrink_data_url(url) - if resized: - image_value["url"] = resized - changed_count += 1 - elif isinstance(image_value, str): - resized = _shrink_data_url(image_value) - if resized: - part["image_url"] = resized - changed_count += 1 - - if changed_count: - logger.info( - "image-shrink recovery: re-encoded %d image part(s) to fit under %.0f MB", - changed_count, target_bytes / (1024 * 1024), - ) - return changed_count > 0 + """Forwarder — see ``agent.conversation_compression.try_shrink_image_parts_in_messages``.""" + from agent.conversation_compression import try_shrink_image_parts_in_messages + return try_shrink_image_parts_in_messages(api_messages) def _anthropic_preserve_dots(self) -> bool: """True when using an anthropic-compatible endpoint that preserves dots in model names. @@ -9861,220 +3475,9 @@ class AIAgent: break def _build_api_kwargs(self, api_messages: list) -> dict: - """Build the keyword arguments dict for the active API mode.""" - tools_for_api = self.tools - - if self.api_mode == "anthropic_messages": - _transport = self._get_transport() - anthropic_messages = self._prepare_anthropic_messages_for_api(api_messages) - ctx_len = getattr(self, "context_compressor", None) - ctx_len = ctx_len.context_length if ctx_len else None - ephemeral_out = getattr(self, "_ephemeral_max_output_tokens", None) - if ephemeral_out is not None: - self._ephemeral_max_output_tokens = None # consume immediately - return _transport.build_kwargs( - model=self.model, - messages=anthropic_messages, - tools=tools_for_api, - max_tokens=ephemeral_out if ephemeral_out is not None else self.max_tokens, - reasoning_config=self.reasoning_config, - is_oauth=self._is_anthropic_oauth, - preserve_dots=self._anthropic_preserve_dots(), - context_length=ctx_len, - base_url=getattr(self, "_anthropic_base_url", None), - fast_mode=(self.request_overrides or {}).get("speed") == "fast", - drop_context_1m_beta=bool(getattr(self, "_oauth_1m_beta_disabled", False)), - ) - - # AWS Bedrock native Converse API — bypasses the OpenAI client entirely. - # The adapter handles message/tool conversion and boto3 calls directly. - if self.api_mode == "bedrock_converse": - _bt = self._get_transport() - region = getattr(self, "_bedrock_region", None) or "us-east-1" - guardrail = getattr(self, "_bedrock_guardrail_config", None) - return _bt.build_kwargs( - model=self.model, - messages=api_messages, - tools=tools_for_api, - max_tokens=self.max_tokens or 4096, - region=region, - guardrail_config=guardrail, - ) - - if self.api_mode == "codex_responses": - _ct = self._get_transport() - is_github_responses = ( - base_url_host_matches(self.base_url, "models.github.ai") - or base_url_host_matches(self.base_url, "api.githubcopilot.com") - ) - is_codex_backend = ( - self.provider == "openai-codex" - or ( - self._base_url_hostname == "chatgpt.com" - and "/backend-api/codex" in self._base_url_lower - ) - ) - is_xai_responses = self.provider in {"xai", "xai-oauth"} or self._base_url_hostname == "api.x.ai" - _msgs_for_codex = self._prepare_messages_for_non_vision_model(api_messages) - return _ct.build_kwargs( - model=self.model, - messages=_msgs_for_codex, - tools=tools_for_api, - reasoning_config=self.reasoning_config, - session_id=getattr(self, "session_id", None), - max_tokens=self.max_tokens, - request_overrides=self.request_overrides, - is_github_responses=is_github_responses, - is_codex_backend=is_codex_backend, - is_xai_responses=is_xai_responses, - github_reasoning_extra=self._github_models_reasoning_extra_body() if is_github_responses else None, - ) - - # ── chat_completions (default) ───────────────────────────────────── - _ct = self._get_transport() - - # Provider detection flags - _is_qwen = self._is_qwen_portal() - _is_or = self._is_openrouter_url() - _is_gh = ( - base_url_host_matches(self._base_url_lower, "models.github.ai") - or base_url_host_matches(self._base_url_lower, "api.githubcopilot.com") - ) - _is_nous = "nousresearch" in self._base_url_lower - _is_nvidia = "integrate.api.nvidia.com" in self._base_url_lower - _is_kimi = ( - base_url_host_matches(self.base_url, "api.kimi.com") - or base_url_host_matches(self.base_url, "moonshot.ai") - or base_url_host_matches(self.base_url, "moonshot.cn") - ) - _is_tokenhub = base_url_host_matches(self._base_url_lower, "tokenhub.tencentmaas.com") - _is_lmstudio = (self.provider or "").strip().lower() == "lmstudio" - - # Temperature: _fixed_temperature_for_model may return OMIT_TEMPERATURE - # sentinel (temperature omitted entirely), a numeric override, or None. - try: - from agent.auxiliary_client import _fixed_temperature_for_model, OMIT_TEMPERATURE - _ft = _fixed_temperature_for_model(self.model, self.base_url) - _omit_temp = _ft is OMIT_TEMPERATURE - _fixed_temp = _ft if not _omit_temp else None - except Exception: - _omit_temp = False - _fixed_temp = None - - # Provider preferences (OpenRouter-style) - _prefs: Dict[str, Any] = {} - if self.providers_allowed: - _prefs["only"] = self.providers_allowed - if self.providers_ignored: - _prefs["ignore"] = self.providers_ignored - if self.providers_order: - _prefs["order"] = self.providers_order - if self.provider_sort: - _prefs["sort"] = self.provider_sort - if self.provider_require_parameters: - _prefs["require_parameters"] = True - if self.provider_data_collection: - _prefs["data_collection"] = self.provider_data_collection - - # Claude max-output override on aggregators - _ant_max = None - if (_is_or or _is_nous) and "claude" in (self.model or "").lower(): - try: - from agent.anthropic_adapter import _get_anthropic_max_output - _ant_max = _get_anthropic_max_output(self.model) - except Exception: - pass - - # Qwen session metadata - _qwen_meta = None - if _is_qwen: - _qwen_meta = { - "sessionId": self.session_id or "hermes", - "promptId": str(uuid.uuid4()), - } - - # ── Provider profile path (registered providers) ─────────────────── - # Profiles handle per-provider quirks via hooks. When a profile is - # found, delegate fully; otherwise fall through to the legacy flag path. - try: - from providers import get_provider_profile - _profile = get_provider_profile(self.provider) - except Exception: - _profile = None - - if _profile: - _ephemeral_out = getattr(self, "_ephemeral_max_output_tokens", None) - if _ephemeral_out is not None: - self._ephemeral_max_output_tokens = None - - return _ct.build_kwargs( - model=self.model, - messages=api_messages, - tools=tools_for_api, - base_url=self.base_url, - timeout=self._resolved_api_call_timeout(), - max_tokens=self.max_tokens, - ephemeral_max_output_tokens=_ephemeral_out, - max_tokens_param_fn=self._max_tokens_param, - reasoning_config=self.reasoning_config, - request_overrides=self.request_overrides, - session_id=getattr(self, "session_id", None), - provider_profile=_profile, - ollama_num_ctx=self._ollama_num_ctx, - # Context forwarded to profile hooks: - provider_preferences=_prefs or None, - openrouter_min_coding_score=self.openrouter_min_coding_score, - anthropic_max_output=_ant_max, - supports_reasoning=self._supports_reasoning_extra_body(), - qwen_session_metadata=_qwen_meta, - ) - - # ── Legacy flag path ──────────────────────────────────────────── - # Reached only when get_provider_profile() returns None — i.e. a - # completely unknown provider not in providers/ registry. - _ephemeral_out = getattr(self, "_ephemeral_max_output_tokens", None) - if _ephemeral_out is not None: - self._ephemeral_max_output_tokens = None - - # Strip image parts for non-vision models (no-op when vision-capable). - _msgs_for_chat = self._prepare_messages_for_non_vision_model(api_messages) - - return _ct.build_kwargs( - model=self.model, - messages=_msgs_for_chat, - tools=tools_for_api, - base_url=self.base_url, - timeout=self._resolved_api_call_timeout(), - max_tokens=self.max_tokens, - ephemeral_max_output_tokens=_ephemeral_out, - max_tokens_param_fn=self._max_tokens_param, - reasoning_config=self.reasoning_config, - request_overrides=self.request_overrides, - session_id=getattr(self, "session_id", None), - model_lower=(self.model or "").lower(), - is_openrouter=_is_or, - is_nous=_is_nous, - is_qwen_portal=_is_qwen, - is_github_models=_is_gh, - is_nvidia_nim=_is_nvidia, - is_kimi=_is_kimi, - is_tokenhub=_is_tokenhub, - is_lmstudio=_is_lmstudio, - is_custom_provider=self.provider == "custom", - ollama_num_ctx=self._ollama_num_ctx, - provider_preferences=_prefs or None, - openrouter_min_coding_score=self.openrouter_min_coding_score, - qwen_prepare_fn=self._qwen_prepare_chat_messages if _is_qwen else None, - qwen_prepare_inplace_fn=self._qwen_prepare_chat_messages_inplace if _is_qwen else None, - qwen_session_metadata=_qwen_meta, - fixed_temperature=_fixed_temp, - omit_temperature=_omit_temp, - supports_reasoning=self._supports_reasoning_extra_body(), - github_reasoning_extra=self._github_models_reasoning_extra_body() if _is_gh else None, - lmstudio_reasoning_options=self._lmstudio_reasoning_options_cached() if _is_lmstudio else None, - anthropic_max_output=_ant_max, - provider_name=self.provider, - ) + """Forwarder — see ``agent.chat_completion_helpers.build_api_kwargs``.""" + from agent.chat_completion_helpers import build_api_kwargs + return build_api_kwargs(self, api_messages) def _supports_reasoning_extra_body(self) -> bool: """Return True when reasoning extra_body is safe to send for this route/model. @@ -10200,197 +3603,9 @@ class AIAgent: return {"effort": requested_effort} def _build_assistant_message(self, assistant_message, finish_reason: str) -> dict: - """Build a normalized assistant message dict from an API response message. - - Handles reasoning extraction, reasoning_details, and optional tool_calls - so both the tool-call path and the final-response path share one builder. - """ - assistant_tool_calls = getattr(assistant_message, "tool_calls", None) - reasoning_text = self._extract_reasoning(assistant_message) - _from_structured = bool(reasoning_text) - - # Fallback: extract inline blocks from content when no structured - # reasoning fields are present (some models/providers embed thinking - # directly in the content rather than returning separate API fields). - if not reasoning_text: - content = assistant_message.content or "" - think_blocks = re.findall(r'(.*?)', content, flags=re.DOTALL) - if think_blocks: - combined = "\n\n".join(b.strip() for b in think_blocks if b.strip()) - reasoning_text = combined or None - - if reasoning_text and self.verbose_logging: - logging.debug(f"Captured reasoning ({len(reasoning_text)} chars): {reasoning_text}") - - if reasoning_text and self.reasoning_callback: - # Skip callback when streaming is active — reasoning was already - # displayed during the stream via one of two paths: - # (a) _fire_reasoning_delta (structured reasoning_content deltas) - # (b) _stream_delta tag extraction (/) - # When streaming is NOT active, always fire so non-streaming modes - # (gateway, batch, quiet) still get reasoning. - # Any reasoning that wasn't shown during streaming is caught by the - # CLI post-response display fallback (cli.py _reasoning_shown_this_turn). - if not self.stream_delta_callback and not self._stream_callback: - try: - self.reasoning_callback(reasoning_text) - except Exception: - pass - - # Sanitize surrogates from API response — some models (e.g. Kimi/GLM via Ollama) - # can return invalid surrogate code points that crash json.dumps() on persist. - _raw_content = assistant_message.content or "" - _san_content = _sanitize_surrogates(_raw_content) - if reasoning_text: - reasoning_text = _sanitize_surrogates(reasoning_text) - - # Strip inline reasoning tags (… etc.) from the stored - # assistant content. Reasoning was already captured into - # ``reasoning_text`` above (either from structured fields or the - # inline-block fallback), so the raw tags in content are redundant. - # Leaving them in place caused reasoning to leak to messaging - # platforms (#8878, #9568), inflate context on subsequent turns - # (#9306 observed 16% content-size reduction on a real MiniMax - # session), and pollute generated session titles. One strip at the - # storage boundary cleans content for every downstream consumer: - # API replay, session transcript, gateway delivery, CLI display, - # compression, title generation. - if isinstance(_san_content, str) and _san_content: - _san_content = self._strip_think_blocks(_san_content).strip() - - msg = { - "role": "assistant", - "content": _san_content, - "reasoning": reasoning_text, - "finish_reason": finish_reason, - } - - raw_reasoning_content = getattr(assistant_message, "reasoning_content", None) - if raw_reasoning_content is None and hasattr(assistant_message, "model_extra"): - model_extra = getattr(assistant_message, "model_extra", None) or {} - if isinstance(model_extra, dict) and "reasoning_content" in model_extra: - raw_reasoning_content = model_extra["reasoning_content"] - if raw_reasoning_content is not None: - msg["reasoning_content"] = _sanitize_surrogates(raw_reasoning_content) - elif assistant_tool_calls and self._needs_thinking_reasoning_pad(): - # DeepSeek v4 thinking mode and Kimi / Moonshot thinking mode - # both require reasoning_content on every assistant tool-call - # message. Without it, replaying the persisted message causes - # HTTP 400 ("The reasoning_content in the thinking mode must - # be passed back to the API"). Include streamed reasoning - # text when captured; otherwise pad with a single space — - # DeepSeek V4 Pro tightened validation and rejects empty - # string ("The reasoning content in the thinking mode must - # be passed back to the API"). A space satisfies non-empty - # checks everywhere without leaking fabricated reasoning. - # Refs #15250, #17400, #17341. - msg["reasoning_content"] = reasoning_text or " " - - # Additive fallback (refs #16844, #16884). Streaming-only providers - # (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) - # accumulate reasoning through ``delta.reasoning_content`` chunks - # but never land it on the message object as a top-level attribute, - # so neither branch above fires and the chain-of-thought is stored - # only under the internal ``reasoning`` key. When the user later - # replays that history through a DeepSeek-v4 / Kimi thinking model, - # the missing ``reasoning_content`` causes HTTP 400 ("The - # reasoning_content in the thinking mode must be passed back to the - # API."). - # - # Promote the already-sanitized streamed ``reasoning_text`` to - # ``reasoning_content`` at write time, but ONLY when no prior branch - # already set it AND we actually captured reasoning text. This - # preserves every existing behavior: - # - SDK-exposed ``reasoning_content`` (OpenAI/Moonshot/DeepSeek SDK) - # still wins. - # - DeepSeek tool-call ""-pad (#15250) still fires. - # - Non-thinking turns with no reasoning leave the field absent, - # so ``_copy_reasoning_content_for_api``'s cross-provider leak - # guard (#15748) and ``reasoning``→``reasoning_content`` - # promotion tiers still apply at replay time. - if "reasoning_content" not in msg and reasoning_text: - msg["reasoning_content"] = reasoning_text - - if hasattr(assistant_message, 'reasoning_details') and assistant_message.reasoning_details: - # Pass reasoning_details back unmodified so providers (OpenRouter, - # Anthropic, OpenAI) can maintain reasoning continuity across turns. - # Each provider may include opaque fields (signature, encrypted_content) - # that must be preserved exactly. - raw_details = assistant_message.reasoning_details - preserved = [] - for d in raw_details: - if isinstance(d, dict): - preserved.append(d) - elif hasattr(d, "__dict__"): - preserved.append(d.__dict__) - elif hasattr(d, "model_dump"): - preserved.append(d.model_dump()) - if preserved: - msg["reasoning_details"] = preserved - - # Codex Responses API: preserve encrypted reasoning items for - # multi-turn continuity. These get replayed as input on the next turn. - codex_items = getattr(assistant_message, "codex_reasoning_items", None) - if codex_items: - msg["codex_reasoning_items"] = codex_items - - # Codex Responses API: preserve exact assistant message items (with - # id/phase) so follow-up turns can replay structured items instead of - # flattening to plain text. This is required for prefix cache hits. - codex_message_items = getattr(assistant_message, "codex_message_items", None) - if codex_message_items: - msg["codex_message_items"] = codex_message_items - - if assistant_tool_calls: - tool_calls = [] - for tool_call in assistant_tool_calls: - raw_id = getattr(tool_call, "id", None) - call_id = getattr(tool_call, "call_id", None) - if not isinstance(call_id, str) or not call_id.strip(): - embedded_call_id, _ = self._split_responses_tool_id(raw_id) - call_id = embedded_call_id - if not isinstance(call_id, str) or not call_id.strip(): - if isinstance(raw_id, str) and raw_id.strip(): - call_id = raw_id.strip() - else: - _fn = getattr(tool_call, "function", None) - _fn_name = getattr(_fn, "name", "") if _fn else "" - _fn_args = getattr(_fn, "arguments", "{}") if _fn else "{}" - call_id = self._deterministic_call_id(_fn_name, _fn_args, len(tool_calls)) - call_id = call_id.strip() - - response_item_id = getattr(tool_call, "response_item_id", None) - if not isinstance(response_item_id, str) or not response_item_id.strip(): - _, embedded_response_item_id = self._split_responses_tool_id(raw_id) - response_item_id = embedded_response_item_id - - response_item_id = self._derive_responses_function_call_id( - call_id, - response_item_id if isinstance(response_item_id, str) else None, - ) - - tc_dict = { - "id": call_id, - "call_id": call_id, - "response_item_id": response_item_id, - "type": tool_call.type, - "function": { - "name": tool_call.function.name, - "arguments": tool_call.function.arguments - }, - } - # Preserve extra_content (e.g. Gemini thought_signature) so it - # is sent back on subsequent API calls. Without this, Gemini 3 - # thinking models reject the request with a 400 error. - extra = getattr(tool_call, "extra_content", None) - if extra is not None: - if hasattr(extra, "model_dump"): - extra = extra.model_dump() - tc_dict["extra_content"] = extra - tool_calls.append(tc_dict) - msg["tool_calls"] = tool_calls - - return msg + """Forwarder — see ``agent.chat_completion_helpers.build_assistant_message``.""" + from agent.chat_completion_helpers import build_assistant_message + return build_assistant_message(self, assistant_message, finish_reason) def _needs_thinking_reasoning_pad(self) -> bool: """Return True when the active provider enforces reasoning_content echo-back. @@ -10412,15 +3627,17 @@ class AIAgent: ``reasoning_content`` on every assistant tool-call message; omitting it causes the next replay to fail with HTTP 400. - Also detects Kimi models served through third-party providers (e.g. - ollama-cloud) by matching ``kimi`` in the model name. + Detection is host-driven, not model-name-driven: aggregators like + OpenRouter that re-export Kimi/Moonshot models speak their own + protocol and reject ``reasoning_content`` echoes. We only enable the + kimi-reasoning replay when the request actually targets a + kimi/moonshot endpoint or the dedicated kimi-coding provider. """ return ( self.provider in {"kimi-coding", "kimi-coding-cn"} or base_url_host_matches(self.base_url, "api.kimi.com") or base_url_host_matches(self.base_url, "moonshot.ai") or base_url_host_matches(self.base_url, "moonshot.cn") - or "kimi" in (self.model or "").lower() ) def _needs_deepseek_tool_reasoning(self) -> bool: @@ -10455,74 +3672,9 @@ class AIAgent: ) def _copy_reasoning_content_for_api(self, source_msg: dict, api_msg: dict) -> None: - """Copy provider-facing reasoning fields onto an API replay message.""" - if source_msg.get("role") != "assistant": - return - - # 1. Explicit reasoning_content already set — preserve it verbatim - # (includes DeepSeek/Kimi's own space-placeholder written at creation - # time, and any valid reasoning content from the same provider). - # - # Exception: sessions persisted BEFORE #17341 have empty-string - # placeholders pinned at creation time. DeepSeek V4 Pro rejects - # those with HTTP 400. When the active provider enforces the - # thinking-mode echo, upgrade "" → " " on replay so stale history - # doesn't 400 the user on the next turn. - existing = source_msg.get("reasoning_content") - if isinstance(existing, str): - if existing == "" and self._needs_thinking_reasoning_pad(): - api_msg["reasoning_content"] = " " - else: - api_msg["reasoning_content"] = existing - return - - needs_thinking_pad = self._needs_thinking_reasoning_pad() - - # 2. Cross-provider poisoned history (#15748): on DeepSeek/Kimi, - # if the source turn has tool_calls AND a 'reasoning' field but no - # 'reasoning_content' key, the 'reasoning' text was written by a - # prior provider (e.g. MiniMax) — DeepSeek's own _build_assistant_message - # pins reasoning_content at creation time for tool-call turns, so the - # shape (reasoning set, reasoning_content absent, tool_calls present) - # is unreachable from same-provider DeepSeek history after this fix. - # Inject a single space to satisfy the API without leaking another - # provider's chain of thought to DeepSeek/Kimi. Space (not "") - # because DeepSeek V4 Pro rejects empty-string reasoning_content - # in thinking mode (refs #17341). - normalized_reasoning = source_msg.get("reasoning") - if ( - needs_thinking_pad - and source_msg.get("tool_calls") - and isinstance(normalized_reasoning, str) - and normalized_reasoning - ): - api_msg["reasoning_content"] = " " - return - - # 3. Healthy session: promote 'reasoning' field to 'reasoning_content' - # for providers that use the internal 'reasoning' key. - # This must happen before the unconditional empty-string fallback so - # genuine reasoning content is not overwritten (#15812 regression in - # PR #15478). - if isinstance(normalized_reasoning, str) and normalized_reasoning: - api_msg["reasoning_content"] = normalized_reasoning - return - - # 4. DeepSeek / Kimi thinking mode: all assistant messages need - # reasoning_content. Inject a single space to satisfy the provider's - # requirement when no explicit reasoning content is present. Covers - # both tool-call turns (already-poisoned history with no reasoning - # at all) and plain text turns. Space (not "") because DeepSeek V4 - # Pro tightened validation and rejects empty string with HTTP 400 - # ("The reasoning content in the thinking mode must be passed back - # to the API"). Refs #17341. - if needs_thinking_pad: - api_msg["reasoning_content"] = " " - return - - # 5. reasoning_content was present but not a string (e.g. None after - # context compaction). Don't pass null to the API. - api_msg.pop("reasoning_content", None) + """Forwarder — see ``agent.agent_runtime_helpers.copy_reasoning_content_for_api``.""" + from agent.agent_runtime_helpers import copy_reasoning_content_for_api + return copy_reasoning_content_for_api(self, source_msg, api_msg) @staticmethod def _sanitize_tool_calls_for_strict_api(api_msg: dict) -> dict: @@ -10559,108 +3711,9 @@ class AIAgent: logger=None, session_id: str = None, ) -> int: - """Repair corrupted assistant tool-call argument JSON in-place.""" - log = logger or logging.getLogger(__name__) - if not isinstance(messages, list): - return 0 - - repaired = 0 - marker = AIAgent._TOOL_CALL_ARGUMENTS_CORRUPTION_MARKER - - def _prepend_marker(tool_msg: dict) -> None: - existing = tool_msg.get("content") - if isinstance(existing, str): - if not existing: - tool_msg["content"] = marker - elif not existing.startswith(marker): - tool_msg["content"] = f"{marker}\n{existing}" - return - if existing is None: - tool_msg["content"] = marker - return - try: - existing_text = json.dumps(existing) - except TypeError: - existing_text = str(existing) - tool_msg["content"] = f"{marker}\n{existing_text}" - - message_index = 0 - while message_index < len(messages): - msg = messages[message_index] - if not isinstance(msg, dict) or msg.get("role") != "assistant": - message_index += 1 - continue - - tool_calls = msg.get("tool_calls") - if not isinstance(tool_calls, list) or not tool_calls: - message_index += 1 - continue - - insert_at = message_index + 1 - for tool_call in tool_calls: - if not isinstance(tool_call, dict): - continue - function = tool_call.get("function") - if not isinstance(function, dict): - continue - - arguments = function.get("arguments") - if arguments is None or arguments == "": - function["arguments"] = "{}" - continue - if isinstance(arguments, str) and not arguments.strip(): - function["arguments"] = "{}" - continue - if not isinstance(arguments, str): - continue - - try: - json.loads(arguments) - except json.JSONDecodeError: - tool_call_id = tool_call.get("id") - function_name = function.get("name", "?") - preview = arguments[:80] - log.warning( - "Corrupted tool_call arguments repaired before request " - "(session=%s, message_index=%s, tool_call_id=%s, function=%s, preview=%r)", - session_id or "-", - message_index, - tool_call_id or "-", - function_name, - preview, - ) - function["arguments"] = "{}" - - existing_tool_msg = None - scan_index = message_index + 1 - while scan_index < len(messages): - candidate = messages[scan_index] - if not isinstance(candidate, dict) or candidate.get("role") != "tool": - break - if candidate.get("tool_call_id") == tool_call_id: - existing_tool_msg = candidate - break - scan_index += 1 - - if existing_tool_msg is None: - messages.insert( - insert_at, - { - "role": "tool", - "name": function_name if function_name != "?" else "", - "tool_call_id": tool_call_id, - "content": marker, - }, - ) - insert_at += 1 - else: - _prepend_marker(existing_tool_msg) - - repaired += 1 - - message_index += 1 - - return repaired + """Forwarder — see ``agent.agent_runtime_helpers.sanitize_tool_call_arguments``.""" + from agent.agent_runtime_helpers import sanitize_tool_call_arguments + return sanitize_tool_call_arguments(messages, logger=logger, session_id=session_id) def _should_sanitize_tool_calls(self) -> bool: """Determine if tool_calls need sanitization for strict APIs. @@ -10676,185 +3729,12 @@ class AIAgent: return self.api_mode != "codex_responses" def _compress_context(self, messages: list, system_message: str, *, approx_tokens: int = None, task_id: str = "default", focus_topic: str = None) -> tuple: - """Compress conversation context and split the session in SQLite. - - Args: - focus_topic: Optional focus string for guided compression — the - summariser will prioritise preserving information related to - this topic. Inspired by Claude Code's ``/compact ``. - - Returns: - (compressed_messages, new_system_prompt) tuple - """ - _pre_msg_count = len(messages) - logger.info( - "context compression started: session=%s messages=%d tokens=~%s model=%s focus=%r", - self.session_id or "none", _pre_msg_count, - f"{approx_tokens:,}" if approx_tokens else "unknown", self.model, - focus_topic, + """Forwarder — see ``agent.conversation_compression.compress_context``.""" + from agent.conversation_compression import compress_context + return compress_context( + self, messages, system_message, + approx_tokens=approx_tokens, task_id=task_id, focus_topic=focus_topic, ) - self._emit_status( - "🗜️ Compacting context — summarizing earlier conversation so I can continue..." - ) - - # Notify external memory provider before compression discards context - if self._memory_manager: - try: - self._memory_manager.on_pre_compress(messages) - except Exception: - pass - - try: - compressed = self.context_compressor.compress(messages, current_tokens=approx_tokens, focus_topic=focus_topic) - except TypeError: - # Plugin context engine with strict signature that doesn't accept - # focus_topic — fall back to calling without it. - compressed = self.context_compressor.compress(messages, current_tokens=approx_tokens) - - summary_error = getattr(self.context_compressor, "_last_summary_error", None) - if summary_error: - if getattr(self, "_last_compression_summary_warning", None) != summary_error: - self._last_compression_summary_warning = summary_error - self._emit_warning( - f"⚠ Compression summary failed: {summary_error}. " - "Inserted a fallback context marker." - ) - else: - # No hard failure — but did the configured aux model error out - # and get recovered by retrying on main? Surface that so users - # know their auxiliary.compression.model setting is broken even - # though compression succeeded. - _aux_fail_model = getattr(self.context_compressor, "_last_aux_model_failure_model", None) - _aux_fail_err = getattr(self.context_compressor, "_last_aux_model_failure_error", None) - if _aux_fail_model: - # Dedup on (model, error) so we don't spam on every compaction - _aux_key = (_aux_fail_model, _aux_fail_err) - if getattr(self, "_last_aux_fallback_warning_key", None) != _aux_key: - self._last_aux_fallback_warning_key = _aux_key - self._emit_warning( - f"ℹ Configured compression model '{_aux_fail_model}' failed " - f"({_aux_fail_err or 'unknown error'}). Recovered using main model — " - "check auxiliary.compression.model in config.yaml." - ) - - todo_snapshot = self._todo_store.format_for_injection() - if todo_snapshot: - compressed.append({"role": "user", "content": todo_snapshot}) - - self._invalidate_system_prompt() - new_system_prompt = self._build_system_prompt(system_message) - self._cached_system_prompt = new_system_prompt - - if self._session_db: - try: - # Propagate title to the new session with auto-numbering - old_title = self._session_db.get_session_title(self.session_id) - # Trigger memory extraction on the old session before it rotates. - self.commit_memory_session(messages) - self._session_db.end_session(self.session_id, "compression") - old_session_id = self.session_id - self.session_id = f"{datetime.now().strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:6]}" - os.environ["HERMES_SESSION_ID"] = self.session_id - try: - from gateway.session_context import _SESSION_ID - _SESSION_ID.set(self.session_id) - except Exception: - pass - # Update session_log_file to point to the new session's JSON file - self.session_log_file = self.logs_dir / f"session_{self.session_id}.json" - self._session_db_created = False - self._session_db.create_session( - session_id=self.session_id, - source=self.platform or os.environ.get("HERMES_SESSION_SOURCE", "cli"), - model=self.model, - model_config=self._session_init_model_config, - parent_session_id=old_session_id, - ) - self._session_db_created = True - # Auto-number the title for the continuation session - if old_title: - try: - new_title = self._session_db.get_next_title_in_lineage(old_title) - self._session_db.set_session_title(self.session_id, new_title) - except (ValueError, Exception) as e: - logger.debug("Could not propagate title on compression: %s", e) - self._session_db.update_system_prompt(self.session_id, new_system_prompt) - # Reset flush cursor — new session starts with no messages written - self._last_flushed_db_idx = 0 - except Exception as e: - logger.warning("Session DB compression split failed — new session will NOT be indexed: %s", e) - - # Notify the context engine that the session_id rotated because of - # compression (not a fresh /new). Plugin engines (e.g. hermes-lcm) use - # boundary_reason="compression" to preserve DAG lineage across the - # rollover instead of re-initializing fresh per-session state. - # See hermes-lcm#68. Built-in ContextCompressor ignores kwargs. - try: - _old_sid = locals().get("old_session_id") - if _old_sid and hasattr(self.context_compressor, "on_session_start"): - self.context_compressor.on_session_start( - self.session_id or "", - boundary_reason="compression", - old_session_id=_old_sid, - ) - except Exception as _ce_err: - logger.debug("context engine on_session_start (compression): %s", _ce_err) - - # Notify memory providers of the compression-driven session_id rotation - # so provider-cached per-session state (Hindsight's _document_id, - # accumulated turn buffers, counters) refreshes. reset=False because - # the logical conversation continues; only the id and DB row rolled - # over. See #6672. - try: - _old_sid = locals().get("old_session_id") - if _old_sid and self._memory_manager: - self._memory_manager.on_session_switch( - self.session_id or "", - parent_session_id=_old_sid, - reset=False, - reason="compression", - ) - except Exception as _me_err: - logger.debug("memory manager on_session_switch (compression): %s", _me_err) - - # Warn on repeated compressions (quality degrades with each pass) - _cc = self.context_compressor.compression_count - if _cc >= 2: - self._vprint( - f"{self.log_prefix}⚠️ Session compressed {_cc} times — " - f"accuracy may degrade. Consider /new to start fresh.", - force=True, - ) - - # Update token estimate after compaction so pressure calculations - # use the post-compression count, not the stale pre-compression one. - # Use estimate_request_tokens_rough() so tool schemas are included — - # with 50+ tools enabled, schemas alone can add 20-30K tokens, and - # omitting them delays the next compression cycle far past the - # configured threshold (issue #14695). - _compressed_est = estimate_request_tokens_rough( - compressed, - system_prompt=new_system_prompt or "", - tools=self.tools or None, - ) - self.context_compressor.last_prompt_tokens = _compressed_est - self.context_compressor.last_completion_tokens = 0 - - # Clear the file-read dedup cache. After compression the original - # read content is summarised away — if the model re-reads the same - # file it needs the full content, not a "file unchanged" stub. - try: - from tools.file_tools import reset_file_dedup - reset_file_dedup(task_id) - except Exception: - pass - - logger.info( - "context compression done: session=%s messages=%d->%d tokens=~%s", - self.session_id or "none", _pre_msg_count, len(compressed), - f"{_compressed_est:,}", - ) - return compressed, new_system_prompt def _set_tool_guardrail_halt(self, decision: ToolGuardrailDecision) -> None: """Record the first guardrail decision that should stop this turn.""" @@ -10939,89 +3819,9 @@ class AIAgent: def _invoke_tool(self, function_name: str, function_args: dict, effective_task_id: str, tool_call_id: Optional[str] = None, messages: list = None, pre_tool_block_checked: bool = False) -> str: - """Invoke a single tool and return the result string. No display logic. - - Handles both agent-level tools (todo, memory, etc.) and registry-dispatched - tools. Used by the concurrent execution path; the sequential path retains - its own inline invocation for backward-compatible display handling. - """ - # Check plugin hooks for a block directive before executing anything. - block_message: Optional[str] = None - if not pre_tool_block_checked: - try: - from hermes_cli.plugins import get_pre_tool_call_block_message - block_message = get_pre_tool_call_block_message( - function_name, function_args, task_id=effective_task_id or "", - ) - except Exception: - pass - if block_message is not None: - return json.dumps({"error": block_message}, ensure_ascii=False) - - if function_name == "todo": - from tools.todo_tool import todo_tool as _todo_tool - return _todo_tool( - todos=function_args.get("todos"), - merge=function_args.get("merge", False), - store=self._todo_store, - ) - elif function_name == "session_search": - session_db = self._get_session_db_for_recall() - if not session_db: - from hermes_state import format_session_db_unavailable - return json.dumps({"success": False, "error": format_session_db_unavailable()}) - from tools.session_search_tool import session_search as _session_search - return _session_search( - query=function_args.get("query", ""), - role_filter=function_args.get("role_filter"), - limit=function_args.get("limit", 3), - db=session_db, - current_session_id=self.session_id, - ) - elif function_name == "memory": - target = function_args.get("target", "memory") - from tools.memory_tool import memory_tool as _memory_tool - result = _memory_tool( - action=function_args.get("action"), - target=target, - content=function_args.get("content"), - old_text=function_args.get("old_text"), - store=self._memory_store, - ) - # Bridge: notify external memory provider of built-in memory writes - if self._memory_manager and function_args.get("action") in {"add", "replace"}: - try: - self._memory_manager.on_memory_write( - function_args.get("action", ""), - target, - function_args.get("content", ""), - metadata=self._build_memory_write_metadata( - task_id=effective_task_id, - tool_call_id=tool_call_id, - ), - ) - except Exception: - pass - return result - elif self._memory_manager and self._memory_manager.has_tool(function_name): - return self._memory_manager.handle_tool_call(function_name, function_args) - elif function_name == "clarify": - from tools.clarify_tool import clarify_tool as _clarify_tool - return _clarify_tool( - question=function_args.get("question", ""), - choices=function_args.get("choices"), - callback=self.clarify_callback, - ) - elif function_name == "delegate_task": - return self._dispatch_delegate_task(function_args) - else: - return handle_function_call( - function_name, function_args, effective_task_id, - tool_call_id=tool_call_id, - session_id=self.session_id or "", - enabled_tools=list(self.valid_tool_names) if self.valid_tool_names else None, - skip_pre_tool_call_hook=True, - ) + """Forwarder — see ``agent.agent_runtime_helpers.invoke_tool``.""" + from agent.agent_runtime_helpers import invoke_tool + return invoke_tool(self, function_name, function_args, effective_task_id, tool_call_id, messages, pre_tool_block_checked) @staticmethod def _wrap_verbose(label: str, text: str, indent: str = " ") -> str: @@ -11049,1069 +3849,19 @@ class AIAgent: return f"{indent}{label}{body}" def _execute_tool_calls_concurrent(self, assistant_message, messages: list, effective_task_id: str, api_call_count: int = 0) -> None: - """Execute multiple tool calls concurrently using a thread pool. - - Results are collected in the original tool-call order and appended to - messages so the API sees them in the expected sequence. - """ - tool_calls = assistant_message.tool_calls - num_tools = len(tool_calls) - - # ── Pre-flight: interrupt check ────────────────────────────────── - if self._interrupt_requested: - print(f"{self.log_prefix}⚡ Interrupt: skipping {num_tools} tool call(s)") - for tc in tool_calls: - messages.append({ - "role": "tool", - "name": tc.function.name, - "content": f"[Tool execution cancelled — {tc.function.name} was skipped due to user interrupt]", - "tool_call_id": tc.id, - }) - return - - # ── Parse args + pre-execution bookkeeping ─────────────────────── - parsed_calls = [] # list of (tool_call, function_name, function_args) - for tool_call in tool_calls: - function_name = tool_call.function.name - - # Reset nudge counters - if function_name == "memory": - self._turns_since_memory = 0 - elif function_name == "skill_manage": - self._iters_since_skill = 0 - - try: - function_args = json.loads(tool_call.function.arguments) - except json.JSONDecodeError: - function_args = {} - if not isinstance(function_args, dict): - function_args = {} - - # Checkpoint for file-mutating tools - if function_name in {"write_file", "patch"} and self._checkpoint_mgr.enabled: - try: - file_path = function_args.get("path", "") - if file_path: - work_dir = self._checkpoint_mgr.get_working_dir_for_path(file_path) - self._checkpoint_mgr.ensure_checkpoint(work_dir, f"before {function_name}") - except Exception: - pass - - # Checkpoint before destructive terminal commands - if function_name == "terminal" and self._checkpoint_mgr.enabled: - try: - cmd = function_args.get("command", "") - if _is_destructive_command(cmd): - cwd = function_args.get("workdir") or os.getenv("TERMINAL_CWD", os.getcwd()) - self._checkpoint_mgr.ensure_checkpoint( - cwd, f"before terminal: {cmd[:60]}" - ) - except Exception: - pass - - block_result = None - blocked_by_guardrail = False - try: - from hermes_cli.plugins import get_pre_tool_call_block_message - block_message = get_pre_tool_call_block_message( - function_name, function_args, task_id=effective_task_id or "", - ) - except Exception: - block_message = None - - if block_message is not None: - block_result = json.dumps({"error": block_message}, ensure_ascii=False) - else: - guardrail_decision = self._tool_guardrails.before_call(function_name, function_args) - if not guardrail_decision.allows_execution: - block_result = self._guardrail_block_result(guardrail_decision) - blocked_by_guardrail = True - - parsed_calls.append((tool_call, function_name, function_args, block_result, blocked_by_guardrail)) - - # ── Logging / callbacks ────────────────────────────────────────── - tool_names_str = ", ".join(name for _, name, _, _, _ in parsed_calls) - if not self.quiet_mode: - print(f" ⚡ Concurrent: {num_tools} tool calls — {tool_names_str}") - for i, (tc, name, args, block_result, blocked_by_guardrail) in enumerate(parsed_calls, 1): - args_str = json.dumps(args, ensure_ascii=False) - if self.verbose_logging: - print(f" 📞 Tool {i}: {name}({list(args.keys())})") - print(self._wrap_verbose("Args: ", json.dumps(args, indent=2, ensure_ascii=False))) - else: - args_preview = args_str[:self.log_prefix_chars] + "..." if len(args_str) > self.log_prefix_chars else args_str - print(f" 📞 Tool {i}: {name}({list(args.keys())}) - {args_preview}") - - for tc, name, args, block_result, blocked_by_guardrail in parsed_calls: - if block_result is not None: - continue - if self.tool_progress_callback: - try: - preview = _build_tool_preview(name, args) - self.tool_progress_callback("tool.started", name, preview, args) - except Exception as cb_err: - logging.debug(f"Tool progress callback error: {cb_err}") - - for tc, name, args, block_result, blocked_by_guardrail in parsed_calls: - if block_result is not None: - continue - if self.tool_start_callback: - try: - self.tool_start_callback(tc.id, name, args) - except Exception as cb_err: - logging.debug(f"Tool start callback error: {cb_err}") - - # ── Concurrent execution ───────────────────────────────────────── - # Each slot holds (function_name, function_args, function_result, duration, error_flag, blocked_flag) - results = [None] * num_tools - for i, (tc, name, args, block_result, blocked_by_guardrail) in enumerate(parsed_calls): - if block_result is not None: - results[i] = (name, args, block_result, 0.0, True, True) - - # Touch activity before launching workers so the gateway knows - # we're executing tools (not stuck). - self._current_tool = tool_names_str - self._touch_activity(f"executing {num_tools} tools concurrently: {tool_names_str}") - - # Capture CLI callbacks from the agent thread so worker threads can - # register them locally. Without this, _get_approval_callback() in - # terminal_tool returns None in ThreadPoolExecutor workers, causing - # the dangerous-command prompt to fall back to input() — which - # deadlocks against prompt_toolkit's raw terminal mode (#13617). - _parent_approval_cb = _get_approval_callback() - _parent_sudo_cb = _get_sudo_password_callback() - - def _run_tool(index, tool_call, function_name, function_args): - """Worker function executed in a thread.""" - # Register this worker tid so the agent can fan out an interrupt - # to it — see AIAgent.interrupt(). Must happen first thing, and - # must be paired with discard + clear in the finally block. - _worker_tid = threading.current_thread().ident - with self._tool_worker_threads_lock: - self._tool_worker_threads.add(_worker_tid) - # Race: if the agent was interrupted between fan-out (which - # snapshotted an empty/earlier set) and our registration, apply - # the interrupt to our own tid now so is_interrupted() inside - # the tool returns True on the next poll. - if self._interrupt_requested: - try: - _set_interrupt(True, _worker_tid) - except Exception: - pass - # Set the activity callback on THIS worker thread so - # _wait_for_process (terminal commands) can fire heartbeats. - # The callback is thread-local; the main thread's callback - # is invisible to worker threads. - try: - from tools.environments.base import set_activity_callback - set_activity_callback(self._touch_activity) - except Exception: - pass - # Propagate approval/sudo callbacks to this worker thread. - # Mirrors cli.py run_agent() pattern (GHSA-qg5c-hvr5-hjgr). - if _parent_approval_cb is not None: - try: - _set_approval_callback(_parent_approval_cb) - except Exception: - pass - if _parent_sudo_cb is not None: - try: - _set_sudo_password_callback(_parent_sudo_cb) - except Exception: - pass - start = time.time() - try: - result = self._invoke_tool( - function_name, - function_args, - effective_task_id, - tool_call.id, - messages=messages, - pre_tool_block_checked=True, - ) - except Exception as tool_error: - result = f"Error executing tool '{function_name}': {tool_error}" - logger.error("_invoke_tool raised for %s: %s", function_name, tool_error, exc_info=True) - duration = time.time() - start - is_error, _ = _detect_tool_failure(function_name, result) - if is_error: - logger.info("tool %s failed (%.2fs): %s", function_name, duration, result[:200]) - else: - logger.info("tool %s completed (%.2fs, %d chars)", function_name, duration, len(result)) - results[index] = (function_name, function_args, result, duration, is_error, False) - # Tear down worker-tid tracking. Clear any interrupt bit we may - # have set so the next task scheduled onto this recycled tid - # starts with a clean slate. - with self._tool_worker_threads_lock: - self._tool_worker_threads.discard(_worker_tid) - try: - _set_interrupt(False, _worker_tid) - except Exception: - pass - # Clear thread-local callbacks so a recycled worker thread - # doesn't hold stale references to a disposed CLI instance. - try: - _set_approval_callback(None) - _set_sudo_password_callback(None) - except Exception: - pass - - # Start spinner for CLI mode (skip when TUI handles tool progress) - spinner = None - if self._should_emit_quiet_tool_messages() and self._should_start_quiet_spinner(): - face = random.choice(KawaiiSpinner.get_waiting_faces()) - spinner = KawaiiSpinner(f"{face} ⚡ running {num_tools} tools concurrently", spinner_type='dots', print_fn=self._print_fn) - spinner.start() - - try: - runnable_calls = [ - (i, tc, name, args) - for i, (tc, name, args, block_result, blocked_by_guardrail) in enumerate(parsed_calls) - if block_result is None - ] - futures = [] - if runnable_calls: - max_workers = min(len(runnable_calls), _MAX_TOOL_WORKERS) - with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor: - for i, tc, name, args in runnable_calls: - # Propagate ContextVars (e.g. _approval_session_key); mirrors asyncio.to_thread. - ctx = contextvars.copy_context() - f = executor.submit(ctx.run, _run_tool, i, tc, name, args) - futures.append(f) - - # Wait for all to complete with periodic heartbeats so the - # gateway's inactivity monitor doesn't kill us during long - # concurrent tool batches. Also check for user interrupts - # so we don't block indefinitely when the user sends /stop - # or a new message during concurrent tool execution. - _conc_start = time.time() - _interrupt_logged = False - while True: - done, not_done = concurrent.futures.wait( - futures, timeout=5.0, - ) - if not not_done: - break - - # Check for interrupt — the per-thread interrupt signal - # already causes individual tools (terminal, execute_code) - # to abort, but tools without interrupt checks (web_search, - # read_file) will run to completion. Cancel any futures - # that haven't started yet so we don't block on them. - if self._interrupt_requested: - if not _interrupt_logged: - _interrupt_logged = True - self._vprint( - f"{self.log_prefix}⚡ Interrupt: cancelling " - f"{len(not_done)} pending concurrent tool(s)", - force=True, - ) - for f in not_done: - f.cancel() - # Give already-running tools a moment to notice the - # per-thread interrupt signal and exit gracefully. - concurrent.futures.wait(not_done, timeout=3.0) - break - - _conc_elapsed = int(time.time() - _conc_start) - # Heartbeat every ~30s (6 × 5s poll intervals) - if _conc_elapsed > 0 and _conc_elapsed % 30 < 6: - _still_running = [ - parsed_calls[futures.index(f)][1] - for f in not_done - if f in futures - ] - self._touch_activity( - f"concurrent tools running ({_conc_elapsed}s, " - f"{len(not_done)} remaining: {', '.join(_still_running[:3])})" - ) - finally: - if spinner: - # Build a summary message for the spinner stop - completed = sum(1 for r in results if r is not None) - total_dur = sum(r[3] for r in results if r is not None) - spinner.stop(f"⚡ {completed}/{num_tools} tools completed in {total_dur:.1f}s total") - - # ── Post-execution: display per-tool results ───────────────────── - for i, (tc, name, args, block_result, blocked_by_guardrail) in enumerate(parsed_calls): - r = results[i] - blocked = False - if r is None: - # Tool was cancelled (interrupt) or thread didn't return - if self._interrupt_requested: - function_result = f"[Tool execution cancelled — {name} was skipped due to user interrupt]" - else: - function_result = f"Error executing tool '{name}': thread did not return a result" - tool_duration = 0.0 - else: - function_name, function_args, function_result, tool_duration, is_error, blocked = r - - if not blocked: - function_result = self._append_guardrail_observation( - function_name, - function_args, - function_result, - failed=is_error, - ) - - if is_error: - _err_text = _multimodal_text_summary(function_result) - result_preview = _err_text[:200] if len(_err_text) > 200 else _err_text - logger.warning("Tool %s returned error (%.2fs): %s", function_name, tool_duration, result_preview) - - # Track file-mutation outcome for the turn-end verifier. - # `blocked` calls never actually ran — don't let a guardrail - # block count as either a failure or a success. - if not blocked: - try: - self._record_file_mutation_result( - function_name, function_args, function_result, is_error, - ) - except Exception as _ver_err: - logging.debug("file-mutation verifier record failed: %s", _ver_err) - - if not blocked and self.tool_progress_callback: - try: - self.tool_progress_callback( - "tool.completed", function_name, None, None, - duration=tool_duration, is_error=is_error, - ) - except Exception as cb_err: - logging.debug(f"Tool progress callback error: {cb_err}") - - if self.verbose_logging: - logging.debug(f"Tool {function_name} completed in {tool_duration:.2f}s") - logging.debug(f"Tool result ({len(function_result)} chars): {function_result}") - - # Print cute message per tool - if self._should_emit_quiet_tool_messages(): - cute_msg = _get_cute_tool_message_impl(name, args, tool_duration, result=function_result) - self._safe_print(f" {cute_msg}") - elif not self.quiet_mode: - _preview_str = _multimodal_text_summary(function_result) - if self.verbose_logging: - print(f" ✅ Tool {i+1} completed in {tool_duration:.2f}s") - print(self._wrap_verbose("Result: ", _preview_str)) - else: - response_preview = _preview_str[:self.log_prefix_chars] + "..." if len(_preview_str) > self.log_prefix_chars else _preview_str - print(f" ✅ Tool {i+1} completed in {tool_duration:.2f}s - {response_preview}") - - self._current_tool = None - self._touch_activity(f"tool completed: {name} ({tool_duration:.1f}s)") - - if not blocked and self.tool_complete_callback: - try: - self.tool_complete_callback(tc.id, name, args, function_result) - except Exception as cb_err: - logging.debug(f"Tool complete callback error: {cb_err}") - - function_result = maybe_persist_tool_result( - content=function_result, - tool_name=name, - tool_use_id=tc.id, - env=get_active_env(effective_task_id), - ) if not _is_multimodal_tool_result(function_result) else function_result - - subdir_hints = self._subdirectory_hints.check_tool_call(name, args) - if subdir_hints: - if _is_multimodal_tool_result(function_result): - # Append the hint to the text summary part so the model - # still sees it; don't touch the image blocks. - _append_subdir_hint_to_multimodal(function_result, subdir_hints) - else: - function_result += subdir_hints - - # Unwrap _multimodal dicts to an OpenAI-style content list so any - # vision-capable provider receives [{type:text},{type:image_url}] - # rather than a raw Python dict. The Anthropic adapter already - # accepts content lists; vision-capable OpenAI-compatible servers - # (mlx-vlm, GPT-4o, …) accept image_url in tool messages natively. - # Text-only servers get a string-safe fallback here so a rejected - # image tool result never poisons canonical session history. - # String results pass through unchanged. - _tool_content = self._tool_result_content_for_active_model(name, function_result) - tool_msg = { - "role": "tool", - "name": name, - "content": _tool_content, - "tool_call_id": tc.id, - } - messages.append(tool_msg) - - # ── Per-tool /steer drain ─────────────────────────────────── - # Same as the sequential path: drain between each collected - # result so the steer lands as early as possible. - self._apply_pending_steer_to_tool_results(messages, 1) - - # ── Per-turn aggregate budget enforcement ───────────────────────── - num_tools = len(parsed_calls) - if num_tools > 0: - turn_tool_msgs = messages[-num_tools:] - enforce_turn_budget(turn_tool_msgs, env=get_active_env(effective_task_id)) - - # ── /steer injection ────────────────────────────────────────────── - # Append any pending user steer text to the last tool result so the - # agent sees it on its next iteration. Runs AFTER budget enforcement - # so the steer marker is never truncated. See steer() for details. - if num_tools > 0: - self._apply_pending_steer_to_tool_results(messages, num_tools) + """Forwarder — see ``agent.tool_executor.execute_tool_calls_concurrent``.""" + from agent.tool_executor import execute_tool_calls_concurrent + return execute_tool_calls_concurrent(self, assistant_message, messages, effective_task_id, api_call_count) def _execute_tool_calls_sequential(self, assistant_message, messages: list, effective_task_id: str, api_call_count: int = 0) -> None: - """Execute tool calls sequentially (original behavior). Used for single calls or interactive tools.""" - for i, tool_call in enumerate(assistant_message.tool_calls, 1): - # SAFETY: check interrupt BEFORE starting each tool. - # If the user sent "stop" during a previous tool's execution, - # do NOT start any more tools -- skip them all immediately. - if self._interrupt_requested: - remaining_calls = assistant_message.tool_calls[i-1:] - if remaining_calls: - self._vprint(f"{self.log_prefix}⚡ Interrupt: skipping {len(remaining_calls)} tool call(s)", force=True) - for skipped_tc in remaining_calls: - skipped_name = skipped_tc.function.name - skip_msg = { - "role": "tool", - "name": skipped_name, - "content": f"[Tool execution cancelled — {skipped_name} was skipped due to user interrupt]", - "tool_call_id": skipped_tc.id, - } - messages.append(skip_msg) - break - - function_name = tool_call.function.name - - try: - function_args = json.loads(tool_call.function.arguments) - except json.JSONDecodeError as e: - logging.warning(f"Unexpected JSON error after validation: {e}") - function_args = {} - if not isinstance(function_args, dict): - function_args = {} - - # Check plugin hooks for a block directive before executing. - _block_msg: Optional[str] = None - try: - from hermes_cli.plugins import get_pre_tool_call_block_message - _block_msg = get_pre_tool_call_block_message( - function_name, function_args, task_id=effective_task_id or "", - ) - except Exception: - pass - - _guardrail_block_decision: ToolGuardrailDecision | None = None - if _block_msg is None: - guardrail_decision = self._tool_guardrails.before_call(function_name, function_args) - if not guardrail_decision.allows_execution: - _guardrail_block_decision = guardrail_decision - - _execution_blocked = _block_msg is not None or _guardrail_block_decision is not None - - if _execution_blocked: - # Tool blocked by plugin or guardrail policy — skip counters, - # callbacks, checkpointing, activity mutation, and real execution. - pass - # Reset nudge counters when the relevant tool is actually used - elif function_name == "memory": - self._turns_since_memory = 0 - elif function_name == "skill_manage": - self._iters_since_skill = 0 - - if not self.quiet_mode: - args_str = json.dumps(function_args, ensure_ascii=False) - if self.verbose_logging: - print(f" 📞 Tool {i}: {function_name}({list(function_args.keys())})") - print(self._wrap_verbose("Args: ", json.dumps(function_args, indent=2, ensure_ascii=False))) - else: - args_preview = args_str[:self.log_prefix_chars] + "..." if len(args_str) > self.log_prefix_chars else args_str - print(f" 📞 Tool {i}: {function_name}({list(function_args.keys())}) - {args_preview}") - - if not _execution_blocked: - self._current_tool = function_name - self._touch_activity(f"executing tool: {function_name}") - - # Set activity callback for long-running tool execution (terminal - # commands, etc.) so the gateway's inactivity monitor doesn't kill - # the agent while a command is running. - if not _execution_blocked: - try: - from tools.environments.base import set_activity_callback - set_activity_callback(self._touch_activity) - except Exception: - pass - - if not _execution_blocked and self.tool_progress_callback: - try: - preview = _build_tool_preview(function_name, function_args) - self.tool_progress_callback("tool.started", function_name, preview, function_args) - except Exception as cb_err: - logging.debug(f"Tool progress callback error: {cb_err}") - - if not _execution_blocked and self.tool_start_callback: - try: - self.tool_start_callback(tool_call.id, function_name, function_args) - except Exception as cb_err: - logging.debug(f"Tool start callback error: {cb_err}") - - # Checkpoint: snapshot working dir before file-mutating tools - if not _execution_blocked and function_name in {"write_file", "patch"} and self._checkpoint_mgr.enabled: - try: - file_path = function_args.get("path", "") - if file_path: - work_dir = self._checkpoint_mgr.get_working_dir_for_path(file_path) - self._checkpoint_mgr.ensure_checkpoint( - work_dir, f"before {function_name}" - ) - except Exception: - pass # never block tool execution - - # Checkpoint before destructive terminal commands - if not _execution_blocked and function_name == "terminal" and self._checkpoint_mgr.enabled: - try: - cmd = function_args.get("command", "") - if _is_destructive_command(cmd): - cwd = function_args.get("workdir") or os.getenv("TERMINAL_CWD", os.getcwd()) - self._checkpoint_mgr.ensure_checkpoint( - cwd, f"before terminal: {cmd[:60]}" - ) - except Exception: - pass # never block tool execution - - tool_start_time = time.time() - - if _block_msg is not None: - # Tool blocked by plugin policy — return error without executing. - function_result = json.dumps({"error": _block_msg}, ensure_ascii=False) - tool_duration = 0.0 - elif _guardrail_block_decision is not None: - # Tool blocked by tool-loop guardrail — synthesize exactly one - # tool result for the original tool_call_id without executing. - function_result = self._guardrail_block_result(_guardrail_block_decision) - tool_duration = 0.0 - elif function_name == "todo": - from tools.todo_tool import todo_tool as _todo_tool - function_result = _todo_tool( - todos=function_args.get("todos"), - merge=function_args.get("merge", False), - store=self._todo_store, - ) - tool_duration = time.time() - tool_start_time - if self._should_emit_quiet_tool_messages(): - self._vprint(f" {_get_cute_tool_message_impl('todo', function_args, tool_duration, result=function_result)}") - elif function_name == "session_search": - session_db = self._get_session_db_for_recall() - if not session_db: - from hermes_state import format_session_db_unavailable - function_result = json.dumps({"success": False, "error": format_session_db_unavailable()}) - else: - from tools.session_search_tool import session_search as _session_search - function_result = _session_search( - query=function_args.get("query", ""), - role_filter=function_args.get("role_filter"), - limit=function_args.get("limit", 3), - db=session_db, - current_session_id=self.session_id, - ) - tool_duration = time.time() - tool_start_time - if self._should_emit_quiet_tool_messages(): - self._vprint(f" {_get_cute_tool_message_impl('session_search', function_args, tool_duration, result=function_result)}") - elif function_name == "memory": - target = function_args.get("target", "memory") - from tools.memory_tool import memory_tool as _memory_tool - function_result = _memory_tool( - action=function_args.get("action"), - target=target, - content=function_args.get("content"), - old_text=function_args.get("old_text"), - store=self._memory_store, - ) - # Bridge: notify external memory provider of built-in memory writes - if self._memory_manager and function_args.get("action") in {"add", "replace"}: - try: - self._memory_manager.on_memory_write( - function_args.get("action", ""), - target, - function_args.get("content", ""), - metadata=self._build_memory_write_metadata( - task_id=effective_task_id, - tool_call_id=getattr(tool_call, "id", None), - ), - ) - except Exception: - pass - tool_duration = time.time() - tool_start_time - if self._should_emit_quiet_tool_messages(): - self._vprint(f" {_get_cute_tool_message_impl('memory', function_args, tool_duration, result=function_result)}") - elif function_name == "clarify": - from tools.clarify_tool import clarify_tool as _clarify_tool - function_result = _clarify_tool( - question=function_args.get("question", ""), - choices=function_args.get("choices"), - callback=self.clarify_callback, - ) - tool_duration = time.time() - tool_start_time - if self._should_emit_quiet_tool_messages(): - self._vprint(f" {_get_cute_tool_message_impl('clarify', function_args, tool_duration, result=function_result)}") - elif function_name == "delegate_task": - tasks_arg = function_args.get("tasks") - if tasks_arg and isinstance(tasks_arg, list): - spinner_label = f"🔀 delegating {len(tasks_arg)} tasks" - else: - goal_preview = (function_args.get("goal") or "")[:30] - spinner_label = f"🔀 {goal_preview}" if goal_preview else "🔀 delegating" - spinner = None - if self._should_emit_quiet_tool_messages() and self._should_start_quiet_spinner(): - face = random.choice(KawaiiSpinner.get_waiting_faces()) - spinner = KawaiiSpinner(f"{face} {spinner_label}", spinner_type='dots', print_fn=self._print_fn) - spinner.start() - self._delegate_spinner = spinner - _delegate_result = None - try: - function_result = self._dispatch_delegate_task(function_args) - _delegate_result = function_result - finally: - self._delegate_spinner = None - tool_duration = time.time() - tool_start_time - cute_msg = _get_cute_tool_message_impl('delegate_task', function_args, tool_duration, result=_delegate_result) - if spinner: - spinner.stop(cute_msg) - elif self._should_emit_quiet_tool_messages(): - self._vprint(f" {cute_msg}") - elif self._context_engine_tool_names and function_name in self._context_engine_tool_names: - # Context engine tools (lcm_grep, lcm_describe, lcm_expand, etc.) - spinner = None - if self._should_emit_quiet_tool_messages(): - face = random.choice(KawaiiSpinner.get_waiting_faces()) - emoji = _get_tool_emoji(function_name) - preview = _build_tool_preview(function_name, function_args) or function_name - spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=self._print_fn) - spinner.start() - _ce_result = None - try: - function_result = self.context_compressor.handle_tool_call(function_name, function_args, messages=messages) - _ce_result = function_result - except Exception as tool_error: - function_result = json.dumps({"error": f"Context engine tool '{function_name}' failed: {tool_error}"}) - logger.error("context_engine.handle_tool_call raised for %s: %s", function_name, tool_error, exc_info=True) - finally: - tool_duration = time.time() - tool_start_time - cute_msg = _get_cute_tool_message_impl(function_name, function_args, tool_duration, result=_ce_result) - if spinner: - spinner.stop(cute_msg) - elif self._should_emit_quiet_tool_messages(): - self._vprint(f" {cute_msg}") - elif self._memory_manager and self._memory_manager.has_tool(function_name): - # Memory provider tools (hindsight_retain, honcho_search, etc.) - # These are not in the tool registry — route through MemoryManager. - spinner = None - if self._should_emit_quiet_tool_messages() and self._should_start_quiet_spinner(): - face = random.choice(KawaiiSpinner.get_waiting_faces()) - emoji = _get_tool_emoji(function_name) - preview = _build_tool_preview(function_name, function_args) or function_name - spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=self._print_fn) - spinner.start() - _mem_result = None - try: - function_result = self._memory_manager.handle_tool_call(function_name, function_args) - _mem_result = function_result - except Exception as tool_error: - function_result = json.dumps({"error": f"Memory tool '{function_name}' failed: {tool_error}"}) - logger.error("memory_manager.handle_tool_call raised for %s: %s", function_name, tool_error, exc_info=True) - finally: - tool_duration = time.time() - tool_start_time - cute_msg = _get_cute_tool_message_impl(function_name, function_args, tool_duration, result=_mem_result) - if spinner: - spinner.stop(cute_msg) - elif self._should_emit_quiet_tool_messages(): - self._vprint(f" {cute_msg}") - elif self.quiet_mode: - spinner = None - if self._should_emit_quiet_tool_messages() and self._should_start_quiet_spinner(): - face = random.choice(KawaiiSpinner.get_waiting_faces()) - emoji = _get_tool_emoji(function_name) - preview = _build_tool_preview(function_name, function_args) or function_name - spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=self._print_fn) - spinner.start() - _spinner_result = None - try: - function_result = handle_function_call( - function_name, function_args, effective_task_id, - tool_call_id=tool_call.id, - session_id=self.session_id or "", - enabled_tools=list(self.valid_tool_names) if self.valid_tool_names else None, - skip_pre_tool_call_hook=True, - ) - _spinner_result = function_result - except Exception as tool_error: - function_result = f"Error executing tool '{function_name}': {tool_error}" - logger.error("handle_function_call raised for %s: %s", function_name, tool_error, exc_info=True) - finally: - tool_duration = time.time() - tool_start_time - cute_msg = _get_cute_tool_message_impl(function_name, function_args, tool_duration, result=_spinner_result) - if spinner: - spinner.stop(cute_msg) - elif self._should_emit_quiet_tool_messages(): - self._vprint(f" {cute_msg}") - else: - try: - function_result = handle_function_call( - function_name, function_args, effective_task_id, - tool_call_id=tool_call.id, - session_id=self.session_id or "", - enabled_tools=list(self.valid_tool_names) if self.valid_tool_names else None, - skip_pre_tool_call_hook=True, - ) - except Exception as tool_error: - function_result = f"Error executing tool '{function_name}': {tool_error}" - logger.error("handle_function_call raised for %s: %s", function_name, tool_error, exc_info=True) - tool_duration = time.time() - tool_start_time - - if isinstance(function_result, str): - result_preview = function_result if self.verbose_logging else ( - function_result[:200] if len(function_result) > 200 else function_result - ) - _result_len = len(function_result) - else: - # Multimodal dict result (_multimodal=True) — not sliceable as string - result_preview = function_result - _result_len = len(str(function_result)) - - # Log tool errors to the persistent error log so [error] tags - # in the UI always have a corresponding detailed entry on disk. - _is_error_result, _ = _detect_tool_failure(function_name, function_result) - if not _execution_blocked: - function_result = self._append_guardrail_observation( - function_name, - function_args, - function_result, - failed=_is_error_result, - ) - result_preview = function_result if self.verbose_logging else ( - function_result[:200] if len(function_result) > 200 else function_result - ) - if _is_error_result: - logger.warning("Tool %s returned error (%.2fs): %s", function_name, tool_duration, result_preview) - else: - logger.info("tool %s completed (%.2fs, %d chars)", function_name, tool_duration, _result_len) - - # Track file-mutation outcome for the turn-end verifier. See - # the concurrent path for the rationale; both paths must feed - # the same state so the footer reflects every tool call in the - # turn, not just the parallel ones. - if not _execution_blocked: - try: - self._record_file_mutation_result( - function_name, function_args, function_result, _is_error_result, - ) - except Exception as _ver_err: - logging.debug("file-mutation verifier record failed: %s", _ver_err) - - if not _execution_blocked and self.tool_progress_callback: - try: - self.tool_progress_callback( - "tool.completed", function_name, None, None, - duration=tool_duration, is_error=_is_error_result, - ) - except Exception as cb_err: - logging.debug(f"Tool progress callback error: {cb_err}") - - self._current_tool = None - self._touch_activity(f"tool completed: {function_name} ({tool_duration:.1f}s)") - - if self.verbose_logging: - logging.debug(f"Tool {function_name} completed in {tool_duration:.2f}s") - _log_result = _multimodal_text_summary(function_result) - logging.debug(f"Tool result ({len(_log_result)} chars): {_log_result}") - - if not _execution_blocked and self.tool_complete_callback: - try: - self.tool_complete_callback(tool_call.id, function_name, function_args, function_result) - except Exception as cb_err: - logging.debug(f"Tool complete callback error: {cb_err}") - - function_result = maybe_persist_tool_result( - content=function_result, - tool_name=function_name, - tool_use_id=tool_call.id, - env=get_active_env(effective_task_id), - ) if not _is_multimodal_tool_result(function_result) else function_result - - # Discover subdirectory context files from tool arguments - subdir_hints = self._subdirectory_hints.check_tool_call(function_name, function_args) - if subdir_hints: - if _is_multimodal_tool_result(function_result): - _append_subdir_hint_to_multimodal(function_result, subdir_hints) - else: - function_result += subdir_hints - - # Unwrap _multimodal dicts to an OpenAI-style content list - # (see parallel path for rationale). String results pass through. - _tool_content = self._tool_result_content_for_active_model(function_name, function_result) - tool_msg = { - "role": "tool", - "name": function_name, - "content": _tool_content, - "tool_call_id": tool_call.id - } - messages.append(tool_msg) - - # ── Per-tool /steer drain ─────────────────────────────────── - # Drain pending steer BETWEEN individual tool calls so the - # injection lands as soon as a tool finishes — not after the - # entire batch. The model sees it on the next API iteration. - self._apply_pending_steer_to_tool_results(messages, 1) - - if not self.quiet_mode: - if self.verbose_logging: - print(f" ✅ Tool {i} completed in {tool_duration:.2f}s") - print(self._wrap_verbose("Result: ", function_result)) - else: - _fr_str = function_result if isinstance(function_result, str) else str(function_result) - response_preview = _fr_str[:self.log_prefix_chars] + "..." if len(_fr_str) > self.log_prefix_chars else _fr_str - print(f" ✅ Tool {i} completed in {tool_duration:.2f}s - {response_preview}") - - if self._interrupt_requested and i < len(assistant_message.tool_calls): - remaining = len(assistant_message.tool_calls) - i - self._vprint(f"{self.log_prefix}⚡ Interrupt: skipping {remaining} remaining tool call(s)", force=True) - for skipped_tc in assistant_message.tool_calls[i:]: - skipped_name = skipped_tc.function.name - skip_msg = { - "role": "tool", - "name": skipped_name, - "content": f"[Tool execution skipped — {skipped_name} was not started. User sent a new message]", - "tool_call_id": skipped_tc.id - } - messages.append(skip_msg) - break - - if self.tool_delay > 0 and i < len(assistant_message.tool_calls): - time.sleep(self.tool_delay) - - # ── Per-turn aggregate budget enforcement ───────────────────────── - num_tools_seq = len(assistant_message.tool_calls) - if num_tools_seq > 0: - enforce_turn_budget(messages[-num_tools_seq:], env=get_active_env(effective_task_id)) - - # ── /steer injection ────────────────────────────────────────────── - # See _execute_tool_calls_parallel for the rationale. Same hook, - # applied to sequential execution as well. - if num_tools_seq > 0: - self._apply_pending_steer_to_tool_results(messages, num_tools_seq) - + """Forwarder — see ``agent.tool_executor.execute_tool_calls_sequential``.""" + from agent.tool_executor import execute_tool_calls_sequential + return execute_tool_calls_sequential(self, assistant_message, messages, effective_task_id, api_call_count) def _handle_max_iterations(self, messages: list, api_call_count: int) -> str: - """Request a summary when max iterations are reached. Returns the final response text.""" - print(f"⚠️ Reached maximum iterations ({self.max_iterations}). Requesting summary...") - - summary_request = ( - "You've reached the maximum number of tool-calling iterations allowed. " - "Please provide a final response summarizing what you've found and accomplished so far, " - "without calling any more tools." - ) - messages.append({"role": "user", "content": summary_request}) - - try: - # Build API messages, stripping internal-only fields - # (finish_reason, reasoning) that strict APIs like Mistral reject with 422 - _needs_sanitize = self._should_sanitize_tool_calls() - api_messages = [] - for msg in messages: - api_msg = msg.copy() - self._copy_reasoning_content_for_api(msg, api_msg) - for internal_field in ("reasoning", "finish_reason", "_thinking_prefill"): - api_msg.pop(internal_field, None) - if _needs_sanitize: - self._sanitize_tool_calls_for_strict_api(api_msg) - api_messages.append(api_msg) - - effective_system = self._cached_system_prompt or "" - if self.ephemeral_system_prompt: - effective_system = (effective_system + "\n\n" + self.ephemeral_system_prompt).strip() - if effective_system: - api_messages = [{"role": "system", "content": effective_system}] + api_messages - if self.prefill_messages: - sys_offset = 1 if effective_system else 0 - for idx, pfm in enumerate(self.prefill_messages): - api_messages.insert(sys_offset + idx, pfm.copy()) - - # Same safety net as the main loop: repair tool-call/result - # pairing before asking for a final summary. Compression and - # session resume can leave a tool result whose parent assistant - # tool_call was summarized away; Responses API rejects that as - # "No tool call found for function call output". - api_messages = self._sanitize_api_messages(api_messages) - - # Same safety net as the main loop: drop thinking-only assistant - # turns so Anthropic-family providers don't 400 the summary call. - api_messages = self._drop_thinking_only_and_merge_users(api_messages) - - summary_extra_body = {} - try: - from agent.auxiliary_client import _fixed_temperature_for_model, OMIT_TEMPERATURE as _OMIT_TEMP - except Exception: - _fixed_temperature_for_model = None - _OMIT_TEMP = None - _raw_summary_temp = ( - _fixed_temperature_for_model(self.model, self.base_url) - if _fixed_temperature_for_model is not None - else None - ) - _omit_summary_temperature = _raw_summary_temp is _OMIT_TEMP - _summary_temperature = None if _omit_summary_temperature else _raw_summary_temp - _is_nous = "nousresearch" in self._base_url_lower - # LM Studio uses top-level `reasoning_effort` (not extra_body.reasoning). - # Mirror ChatCompletionsTransport.build_kwargs() so the summary path - # — which calls chat.completions.create() directly without going - # through the transport — sends the same shape the transport does. - _is_lmstudio_summary = ( - (self.provider or "").strip().lower() == "lmstudio" - and self._supports_reasoning_extra_body() - ) - _lm_reasoning_effort: str | None = ( - self._resolve_lmstudio_summary_reasoning_effort() - if _is_lmstudio_summary else None - ) - if not _is_lmstudio_summary and self._supports_reasoning_extra_body(): - if self.reasoning_config is not None: - summary_extra_body["reasoning"] = self.reasoning_config - else: - summary_extra_body["reasoning"] = { - "enabled": True, - "effort": "medium" - } - if _is_nous: - from agent.portal_tags import nous_portal_tags as _portal_tags - summary_extra_body["tags"] = _portal_tags() - - if self.api_mode == "codex_responses": - codex_kwargs = self._build_api_kwargs(api_messages) - codex_kwargs.pop("tools", None) - summary_response = self._run_codex_stream(codex_kwargs) - _ct_sum = self._get_transport() - _cnr_sum = _ct_sum.normalize_response(summary_response) - final_response = (_cnr_sum.content or "").strip() - else: - summary_kwargs = { - "model": self.model, - "messages": api_messages, - } - if _summary_temperature is not None: - summary_kwargs["temperature"] = _summary_temperature - if self.max_tokens is not None: - summary_kwargs.update(self._max_tokens_param(self.max_tokens)) - if _lm_reasoning_effort is not None: - summary_kwargs["reasoning_effort"] = _lm_reasoning_effort - - # Include provider routing preferences - provider_preferences = {} - if self.providers_allowed: - provider_preferences["only"] = self.providers_allowed - if self.providers_ignored: - provider_preferences["ignore"] = self.providers_ignored - if self.providers_order: - provider_preferences["order"] = self.providers_order - if self.provider_sort: - provider_preferences["sort"] = self.provider_sort - if provider_preferences and ( - (self.provider or "").strip().lower() == "openrouter" - or self._is_openrouter_url() - ): - summary_extra_body["provider"] = provider_preferences - - # Pareto Code router plugin — model-gated. Same shape as - # the main-loop emission so summary calls on - # openrouter/pareto-code respect the user's coding-score floor. - if ( - self.model == "openrouter/pareto-code" - and ( - (self.provider or "").strip().lower() == "openrouter" - or self._is_openrouter_url() - ) - and self.openrouter_min_coding_score is not None - and self.openrouter_min_coding_score != "" - ): - try: - _ps = float(self.openrouter_min_coding_score) - except (TypeError, ValueError): - _ps = None - if _ps is not None and 0.0 <= _ps <= 1.0: - summary_extra_body["plugins"] = [ - {"id": "pareto-router", "min_coding_score": _ps} - ] - - if summary_extra_body: - summary_kwargs["extra_body"] = summary_extra_body - - if self.api_mode == "anthropic_messages": - _tsum = self._get_transport() - _ant_kw = _tsum.build_kwargs(model=self.model, messages=api_messages, tools=None, - max_tokens=self.max_tokens, reasoning_config=self.reasoning_config, - is_oauth=self._is_anthropic_oauth, - preserve_dots=self._anthropic_preserve_dots()) - summary_response = self._anthropic_messages_create(_ant_kw) - _summary_result = _tsum.normalize_response(summary_response, strip_tool_prefix=self._is_anthropic_oauth) - final_response = (_summary_result.content or "").strip() - else: - summary_response = self._ensure_primary_openai_client(reason="iteration_limit_summary").chat.completions.create(**summary_kwargs) - _summary_result = self._get_transport().normalize_response(summary_response) - final_response = (_summary_result.content or "").strip() - - if final_response: - if "" in final_response: - final_response = re.sub(r'.*?\s*', '', final_response, flags=re.DOTALL).strip() - if final_response: - messages.append({"role": "assistant", "content": final_response}) - else: - final_response = "I reached the iteration limit and couldn't generate a summary." - else: - # Retry summary generation - if self.api_mode == "codex_responses": - codex_kwargs = self._build_api_kwargs(api_messages) - codex_kwargs.pop("tools", None) - retry_response = self._run_codex_stream(codex_kwargs) - _ct_retry = self._get_transport() - _cnr_retry = _ct_retry.normalize_response(retry_response) - final_response = (_cnr_retry.content or "").strip() - elif self.api_mode == "anthropic_messages": - _tretry = self._get_transport() - _ant_kw2 = _tretry.build_kwargs(model=self.model, messages=api_messages, tools=None, - is_oauth=self._is_anthropic_oauth, - max_tokens=self.max_tokens, reasoning_config=self.reasoning_config, - preserve_dots=self._anthropic_preserve_dots()) - retry_response = self._anthropic_messages_create(_ant_kw2) - _retry_result = _tretry.normalize_response(retry_response, strip_tool_prefix=self._is_anthropic_oauth) - final_response = (_retry_result.content or "").strip() - else: - summary_kwargs = { - "model": self.model, - "messages": api_messages, - } - if _summary_temperature is not None: - summary_kwargs["temperature"] = _summary_temperature - if self.max_tokens is not None: - summary_kwargs.update(self._max_tokens_param(self.max_tokens)) - if _lm_reasoning_effort is not None: - summary_kwargs["reasoning_effort"] = _lm_reasoning_effort - if summary_extra_body: - summary_kwargs["extra_body"] = summary_extra_body - - summary_response = self._ensure_primary_openai_client(reason="iteration_limit_summary_retry").chat.completions.create(**summary_kwargs) - _retry_result = self._get_transport().normalize_response(summary_response) - final_response = (_retry_result.content or "").strip() - - if final_response: - if "" in final_response: - final_response = re.sub(r'.*?\s*', '', final_response, flags=re.DOTALL).strip() - if final_response: - messages.append({"role": "assistant", "content": final_response}) - else: - final_response = "I reached the iteration limit and couldn't generate a summary." - else: - final_response = "I reached the iteration limit and couldn't generate a summary." - - except Exception as e: - logging.warning(f"Failed to get summary response: {e}") - final_response = f"I reached the maximum iterations ({self.max_iterations}) but couldn't summarize. Error: {str(e)}" - - return final_response + """Forwarder — see ``agent.chat_completion_helpers.handle_max_iterations``.""" + from agent.chat_completion_helpers import handle_max_iterations + return handle_max_iterations(self, messages, api_call_count) def run_conversation( self, @@ -12122,3932 +3872,9 @@ class AIAgent: stream_callback: Optional[callable] = None, persist_user_message: Optional[str] = None, ) -> Dict[str, Any]: - """ - Run a complete conversation with tool calling until completion. - - Args: - user_message (str): The user's message/question - system_message (str): Custom system message (optional, overrides ephemeral_system_prompt if provided) - conversation_history (List[Dict]): Previous conversation messages (optional) - task_id (str): Unique identifier for this task to isolate VMs between concurrent tasks (optional, auto-generated if not provided) - stream_callback: Optional callback invoked with each text delta during streaming. - Used by the TTS pipeline to start audio generation before the full response. - When None (default), API calls use the standard non-streaming path. - persist_user_message: Optional clean user message to store in - transcripts/history when user_message contains API-only - synthetic prefixes. - or queuing follow-up prefetch work. - - Returns: - Dict: Complete conversation result with final response and message history - """ - # Guard stdio against OSError from broken pipes (systemd/headless/daemon). - # Installed once, transparent when streams are healthy, prevents crash on write. - _install_safe_stdio() - - self._ensure_db_session() - - # Tell auxiliary_client what the live main provider/model are for - # this turn. Used by tools whose behaviour depends on the active - # main model (e.g. vision_analyze's native fast path) so they see - # the CLI/gateway override instead of the stale config.yaml - # default. Idempotent — fine to call every turn. - try: - from agent.auxiliary_client import set_runtime_main - set_runtime_main( - getattr(self, "provider", "") or "", - getattr(self, "model", "") or "", - ) - except Exception: - pass - - # Tag all log records on this thread with the session ID so - # ``hermes logs --session `` can filter a single conversation. - from hermes_logging import set_session_context - set_session_context(self.session_id) - - # Bind the skill write-origin ContextVar for this thread so tool - # handlers (e.g. skill_manage create) can tell whether they are - # running inside the background self-improvement review fork vs. - # a foreground user-directed turn. Set at the top of each call; - # the review fork runs on its own thread with a fresh context, - # so the foreground value here does not leak into it. - from tools.skill_provenance import set_current_write_origin - set_current_write_origin(getattr(self, "_memory_write_origin", "assistant_tool")) - - # If the previous turn activated fallback, restore the primary - # runtime so this turn gets a fresh attempt with the preferred model. - # No-op when _fallback_activated is False (gateway, first turn, etc.). - self._restore_primary_runtime() - - # Sanitize surrogate characters from user input. Clipboard paste from - # rich-text editors (Google Docs, Word, etc.) can inject lone surrogates - # that are invalid UTF-8 and crash JSON serialization in the OpenAI SDK. - if isinstance(user_message, str): - user_message = _sanitize_surrogates(user_message) - if isinstance(persist_user_message, str): - persist_user_message = _sanitize_surrogates(persist_user_message) - - # Store stream callback for _interruptible_api_call to pick up - self._stream_callback = stream_callback - self._persist_user_message_idx = None - self._persist_user_message_override = persist_user_message - # Generate unique task_id if not provided to isolate VMs between concurrent tasks - effective_task_id = task_id or str(uuid.uuid4()) - # Expose the active task_id so tools running mid-turn (e.g. delegate_task - # in delegate_tool.py) can identify this agent for the cross-agent file - # state registry. Set BEFORE any tool dispatch so snapshots taken at - # child-launch time see the parent's real id, not None. - self._current_task_id = effective_task_id - - # Reset retry counters and iteration budget at the start of each turn - # so subagent usage from a previous turn doesn't eat into the next one. - self._invalid_tool_retries = 0 - self._invalid_json_retries = 0 - self._empty_content_retries = 0 - self._incomplete_scratchpad_retries = 0 - self._codex_incomplete_retries = 0 - self._thinking_prefill_retries = 0 - self._post_tool_empty_retried = False - self._last_content_with_tools = None - self._last_content_tools_all_housekeeping = False - self._mute_post_response = False - self._unicode_sanitization_passes = 0 - self._tool_guardrails.reset_for_turn() - self._tool_guardrail_halt_decision = None - # True until the server rejects an image_url content part with an error - # like "Only 'text' content type is supported." Set to False on first - # rejection and kept False for the rest of the session so we never re-send - # images to a text-only endpoint. Scoped per `_run()` call, not per instance. - self._vision_supported = True - - # Pre-turn connection health check: detect and clean up dead TCP - # connections left over from provider outages or dropped streams. - # This prevents the next API call from hanging on a zombie socket. - if self.api_mode != "anthropic_messages": - try: - if self._cleanup_dead_connections(): - self._emit_status( - "🔌 Detected stale connections from a previous provider " - "issue — cleaned up automatically. Proceeding with fresh " - "connection." - ) - except Exception: - pass - # Replay compression warning through status_callback for gateway - # platforms (the callback was not wired during __init__). - if self._compression_warning: - self._replay_compression_warning() - self._compression_warning = None # send once - - # NOTE: _turns_since_memory and _iters_since_skill are NOT reset here. - # They are initialized in __init__ and must persist across run_conversation - # calls so that nudge logic accumulates correctly in CLI mode. - self.iteration_budget = IterationBudget(self.max_iterations) - - # Log conversation turn start for debugging/observability - _preview_text = _summarize_user_message_for_log(user_message) - _msg_preview = (_preview_text[:80] + "...") if len(_preview_text) > 80 else _preview_text - _msg_preview = _msg_preview.replace("\n", " ") - logger.info( - "conversation turn: session=%s model=%s provider=%s platform=%s history=%d msg=%r", - self.session_id or "none", self.model, self.provider or "unknown", - self.platform or "unknown", len(conversation_history or []), - _msg_preview, - ) - - # Initialize conversation (copy to avoid mutating the caller's list) - messages = list(conversation_history) if conversation_history else [] - - # Hydrate todo store from conversation history (gateway creates a fresh - # AIAgent per message, so the in-memory store is empty -- we need to - # recover the todo state from the most recent todo tool response in history) - if conversation_history and not self._todo_store.has_items(): - self._hydrate_todo_store(conversation_history) - - # Hydrate per-session nudge counters from persisted history. - # Gateway creates a fresh AIAgent per inbound message (cache miss / - # 1h idle eviction / config-signature mismatch / process restart), so - # _turns_since_memory and _user_turn_count start at 0 every turn and - # the memory.nudge_interval trigger may never be reached. Reconstruct - # an effective count from prior user turns in conversation_history. - # Idempotent: a cached agent that already accumulated counters keeps - # them; only a freshly-built agent with empty in-memory state hydrates. - # See issue #22357. - if conversation_history and self._user_turn_count == 0: - prior_user_turns = sum( - 1 for m in conversation_history if m.get("role") == "user" - ) - if prior_user_turns > 0: - self._user_turn_count = prior_user_turns - if self._memory_nudge_interval > 0 and self._turns_since_memory == 0: - # % preserves original 1-in-N cadence rather than firing a - # review immediately on resume (which would surprise users - # whose session happened to land just past a multiple of N). - self._turns_since_memory = prior_user_turns % self._memory_nudge_interval - - - # Prefill messages (few-shot priming) are injected at API-call time only, - # never stored in the messages list. This keeps them ephemeral: they won't - # be saved to session DB, session logs, or batch trajectories, but they're - # automatically re-applied on every API call (including session continuations). - - # Track user turns for memory flush and periodic nudge logic - self._user_turn_count += 1 - - # Reset the streaming context scrubber at the top of each turn so a - # hung span from a prior interrupted stream can't taint this turn's - # output. - scrubber = getattr(self, "_stream_context_scrubber", None) - if scrubber is not None: - scrubber.reset() - # Reset the think scrubber for the same reason — an interrupted - # prior stream may have left us inside an unterminated block. - think_scrubber = getattr(self, "_stream_think_scrubber", None) - if think_scrubber is not None: - think_scrubber.reset() - - # Preserve the original user message (no nudge injection). - original_user_message = persist_user_message if persist_user_message is not None else user_message - - # Track memory nudge trigger (turn-based, checked here). - # Skill trigger is checked AFTER the agent loop completes, based on - # how many tool iterations THIS turn used. - _should_review_memory = False - if (self._memory_nudge_interval > 0 - and "memory" in self.valid_tool_names - and self._memory_store): - self._turns_since_memory += 1 - if self._turns_since_memory >= self._memory_nudge_interval: - _should_review_memory = True - self._turns_since_memory = 0 - - # Add user message - user_msg = {"role": "user", "content": user_message} - messages.append(user_msg) - current_turn_user_idx = len(messages) - 1 - self._persist_user_message_idx = current_turn_user_idx - - if not self.quiet_mode: - _print_preview = _summarize_user_message_for_log(user_message) - self._safe_print(f"💬 Starting conversation: '{_print_preview[:60]}{'...' if len(_print_preview) > 60 else ''}'") - - # ── System prompt (cached per session for prefix caching) ── - # Built once on first call, reused for all subsequent calls. - # Only rebuilt after context compression events (which invalidate - # the cache and reload memory from disk). - # - # For continuing sessions (gateway creates a fresh AIAgent per - # message), we load the stored system prompt from the session DB - # instead of rebuilding. Rebuilding would pick up memory changes - # from disk that the model already knows about (it wrote them!), - # producing a different system prompt and breaking the Anthropic - # prefix cache. - if self._cached_system_prompt is None: - stored_prompt = None - if conversation_history and self._session_db: - try: - session_row = self._session_db.get_session(self.session_id) - if session_row: - stored_prompt = session_row.get("system_prompt") or None - except Exception: - pass # Fall through to build fresh - - if stored_prompt: - # Continuing session — reuse the exact system prompt from - # the previous turn so the Anthropic cache prefix matches. - self._cached_system_prompt = stored_prompt - else: - # First turn of a new session — build from scratch. - self._cached_system_prompt = self._build_system_prompt(system_message) - # Plugin hook: on_session_start - # Fired once when a brand-new session is created (not on - # continuation). Plugins can use this to initialise - # session-scoped state (e.g. warm a memory cache). - try: - from hermes_cli.plugins import invoke_hook as _invoke_hook - _invoke_hook( - "on_session_start", - session_id=self.session_id, - model=self.model, - platform=getattr(self, "platform", None) or "", - ) - except Exception as exc: - logger.warning("on_session_start hook failed: %s", exc) - - # Store the system prompt snapshot in SQLite - if self._session_db: - try: - self._session_db.update_system_prompt(self.session_id, self._cached_system_prompt) - except Exception as e: - logger.debug("Session DB update_system_prompt failed: %s", e) - - active_system_prompt = self._cached_system_prompt - - # ── Preflight context compression ── - # Before entering the main loop, check if the loaded conversation - # history already exceeds the model's context threshold. This handles - # cases where a user switches to a model with a smaller context window - # while having a large existing session — compress proactively rather - # than waiting for an API error (which might be caught as a non-retryable - # 4xx and abort the request entirely). - if ( - self.compression_enabled - and len(messages) > self.context_compressor.protect_first_n - + self.context_compressor.protect_last_n + 1 - ): - # Include tool schema tokens — with many tools these can add - # 20-30K+ tokens that the old sys+msg estimate missed entirely. - _preflight_tokens = estimate_request_tokens_rough( - messages, - system_prompt=active_system_prompt or "", - tools=self.tools or None, - ) - - if _preflight_tokens >= self.context_compressor.threshold_tokens: - logger.info( - "Preflight compression: ~%s tokens >= %s threshold (model %s, ctx %s)", - f"{_preflight_tokens:,}", - f"{self.context_compressor.threshold_tokens:,}", - self.model, - f"{self.context_compressor.context_length:,}", - ) - self._emit_status( - f"📦 Preflight compression: ~{_preflight_tokens:,} tokens " - f">= {self.context_compressor.threshold_tokens:,} threshold. " - "This may take a moment." - ) - # May need multiple passes for very large sessions with small - # context windows (each pass summarises the middle N turns). - for _pass in range(3): - _orig_len = len(messages) - messages, active_system_prompt = self._compress_context( - messages, system_message, approx_tokens=_preflight_tokens, - task_id=effective_task_id, - ) - if len(messages) >= _orig_len: - break # Cannot compress further - # Compression created a new session — clear the history - # reference so _flush_messages_to_session_db writes ALL - # compressed messages to the new session's SQLite, not - # skipping them because conversation_history is still the - # pre-compression length. - conversation_history = None - # Fix: reset retry counters after compression so the model - # gets a fresh budget on the compressed context. Without - # this, pre-compression retries carry over and the model - # hits "(empty)" immediately after compression-induced - # context loss. - self._empty_content_retries = 0 - self._thinking_prefill_retries = 0 - self._last_content_with_tools = None - self._last_content_tools_all_housekeeping = False - self._mute_post_response = False - # Re-estimate after compression - _preflight_tokens = estimate_request_tokens_rough( - messages, - system_prompt=active_system_prompt or "", - tools=self.tools or None, - ) - if _preflight_tokens < self.context_compressor.threshold_tokens: - break # Under threshold - - # Plugin hook: pre_llm_call - # Fired once per turn before the tool-calling loop. Plugins can - # return a dict with a ``context`` key (or a plain string) whose - # value is appended to the current turn's user message. - # - # Context is ALWAYS injected into the user message, never the - # system prompt. This preserves the prompt cache prefix — the - # system prompt stays identical across turns so cached tokens - # are reused. The system prompt is Hermes's territory; plugins - # contribute context alongside the user's input. - # - # All injected context is ephemeral (not persisted to session DB). - _plugin_user_context = "" - try: - from hermes_cli.plugins import invoke_hook as _invoke_hook - _pre_results = _invoke_hook( - "pre_llm_call", - session_id=self.session_id, - user_message=original_user_message, - conversation_history=list(messages), - is_first_turn=(not bool(conversation_history)), - model=self.model, - platform=getattr(self, "platform", None) or "", - sender_id=getattr(self, "_user_id", None) or "", - ) - _ctx_parts: list[str] = [] - for r in _pre_results: - if isinstance(r, dict) and r.get("context"): - _ctx_parts.append(str(r["context"])) - elif isinstance(r, str) and r.strip(): - _ctx_parts.append(r) - if _ctx_parts: - _plugin_user_context = "\n\n".join(_ctx_parts) - except Exception as exc: - logger.warning("pre_llm_call hook failed: %s", exc) - - # Main conversation loop - api_call_count = 0 - final_response = None - interrupted = False - codex_ack_continuations = 0 - length_continue_retries = 0 - truncated_tool_call_retries = 0 - truncated_response_parts: List[str] = [] - compression_attempts = 0 - _turn_exit_reason = "unknown" # Diagnostic: why the loop ended - - # Per-turn file-mutation verifier state. Keyed by resolved path; - # each failed ``write_file`` / ``patch`` call records the error - # preview. Later successful writes to the same path remove the - # entry (the model recovered). At end-of-turn, any entries still - # present are surfaced in an advisory footer so the model cannot - # over-claim success while the file is actually unchanged on disk. - self._turn_failed_file_mutations: Dict[str, Dict[str, Any]] = {} - - # Record the execution thread so interrupt()/clear_interrupt() can - # scope the tool-level interrupt signal to THIS agent's thread only. - # Must be set before any thread-scoped interrupt syncing. - self._execution_thread_id = threading.current_thread().ident - - # Always clear stale per-thread state from a previous turn. If an - # interrupt arrived before startup finished, preserve it and bind it - # to this execution thread now instead of dropping it on the floor. - _set_interrupt(False, self._execution_thread_id) - if self._interrupt_requested: - _set_interrupt(True, self._execution_thread_id) - self._interrupt_thread_signal_pending = False - else: - self._interrupt_message = None - self._interrupt_thread_signal_pending = False - - # Notify memory providers of the new turn so cadence tracking works. - # Must happen BEFORE prefetch_all() so providers know which turn it is - # and can gate context/dialectic refresh via contextCadence/dialecticCadence. - if self._memory_manager: - try: - _turn_msg = original_user_message if isinstance(original_user_message, str) else "" - self._memory_manager.on_turn_start(self._user_turn_count, _turn_msg) - except Exception: - pass - - # External memory provider: prefetch once before the tool loop. - # Reuse the cached result on every iteration to avoid re-calling - # prefetch_all() on each tool call (10 tool calls = 10x latency + cost). - # Use original_user_message (clean input) — user_message may contain - # injected skill content that bloats / breaks provider queries. - _ext_prefetch_cache = "" - if self._memory_manager: - try: - _query = original_user_message if isinstance(original_user_message, str) else "" - _ext_prefetch_cache = self._memory_manager.prefetch_all(_query) or "" - except Exception: - pass - - # Optional opt-in runtime: if api_mode == codex_app_server, hand the - # turn to the codex app-server subprocess (terminal/file ops/patching - # all run inside Codex). Default Hermes path is bypassed entirely. - # See agent/transports/codex_app_server_session.py for the adapter - # and references/codex-app-server-runtime.md for the rationale. - if self.api_mode == "codex_app_server": - return self._run_codex_app_server_turn( - user_message=user_message, - original_user_message=original_user_message, - messages=messages, - effective_task_id=effective_task_id, - should_review_memory=_should_review_memory, - ) - - while (api_call_count < self.max_iterations and self.iteration_budget.remaining > 0) or self._budget_grace_call: - # Reset per-turn checkpoint dedup so each iteration can take one snapshot - self._checkpoint_mgr.new_turn() - - # Check for interrupt request (e.g., user sent new message) - if self._interrupt_requested: - interrupted = True - _turn_exit_reason = "interrupted_by_user" - if not self.quiet_mode: - self._safe_print("\n⚡ Breaking out of tool loop due to interrupt...") - break - - api_call_count += 1 - self._api_call_count = api_call_count - self._touch_activity(f"starting API call #{api_call_count}") - - # Grace call: the budget is exhausted but we gave the model one - # more chance. Consume the grace flag so the loop exits after - # this iteration regardless of outcome. - if self._budget_grace_call: - self._budget_grace_call = False - elif not self.iteration_budget.consume(): - _turn_exit_reason = "budget_exhausted" - if not self.quiet_mode: - self._safe_print(f"\n⚠️ Iteration budget exhausted ({self.iteration_budget.used}/{self.iteration_budget.max_total} iterations used)") - break - - # Fire step_callback for gateway hooks (agent:step event) - if self.step_callback is not None: - try: - prev_tools = [] - for _idx, _m in enumerate(reversed(messages)): - if _m.get("role") == "assistant" and _m.get("tool_calls"): - _fwd_start = len(messages) - _idx - _results_by_id = {} - for _tm in messages[_fwd_start:]: - if _tm.get("role") != "tool": - break - _tcid = _tm.get("tool_call_id") - if _tcid: - _results_by_id[_tcid] = _tm.get("content", "") - prev_tools = [ - { - "name": tc["function"]["name"], - "result": _results_by_id.get(tc.get("id")), - "arguments": tc["function"].get("arguments"), - } - for tc in _m["tool_calls"] - if isinstance(tc, dict) - ] - break - self.step_callback(api_call_count, prev_tools) - except Exception as _step_err: - logger.debug("step_callback error (iteration %s): %s", api_call_count, _step_err) - - # Track tool-calling iterations for skill nudge. - # Counter resets whenever skill_manage is actually used. - if (self._skill_nudge_interval > 0 - and "skill_manage" in self.valid_tool_names): - self._iters_since_skill += 1 - - # ── Pre-API-call /steer drain ────────────────────────────────── - # If a /steer arrived during the previous API call (while the model - # was thinking), drain it now — before we build api_messages — so - # the model sees the steer text on THIS iteration. Without this, - # steers sent during an API call only land after the NEXT tool batch, - # which may never come if the model returns a final response. - # - # We scan backwards for the last tool-role message in the messages - # list. If found, the steer is appended there. If not (first - # iteration, no tools yet), the steer stays pending for the next - # tool batch — injecting into a user message would break role - # alternation, and there's no tool output to piggyback on. - _pre_api_steer = self._drain_pending_steer() - if _pre_api_steer: - _injected = False - for _si in range(len(messages) - 1, -1, -1): - _sm = messages[_si] - if isinstance(_sm, dict) and _sm.get("role") == "tool": - marker = f"\n\nUser guidance: {_pre_api_steer}" - existing = _sm.get("content", "") - if isinstance(existing, str): - _sm["content"] = existing + marker - else: - # Multimodal content blocks — append text block - try: - blocks = list(existing) if existing else [] - blocks.append({"type": "text", "text": marker}) - _sm["content"] = blocks - except Exception: - pass - _injected = True - logger.debug( - "Pre-API-call steer drain: injected into tool msg at index %d", - _si, - ) - break - if not _injected: - # No tool message to inject into — put it back so - # the post-tool-execution drain picks it up later. - _lock = getattr(self, "_pending_steer_lock", None) - if _lock is not None: - with _lock: - if self._pending_steer: - self._pending_steer = self._pending_steer + "\n" + _pre_api_steer - else: - self._pending_steer = _pre_api_steer - else: - existing = getattr(self, "_pending_steer", None) - self._pending_steer = (existing + "\n" + _pre_api_steer) if existing else _pre_api_steer - - # Prepare messages for API call - # If we have an ephemeral system prompt, prepend it to the messages - # Note: Reasoning is embedded in content via tags for trajectory storage. - # However, providers like Moonshot AI require a separate 'reasoning_content' field - # on assistant messages with tool_calls. We handle both cases here. - request_logger = getattr(self, "logger", None) or logging.getLogger(__name__) - repaired_tool_calls = self._sanitize_tool_call_arguments( - messages, - logger=request_logger, - session_id=self.session_id, - ) - if repaired_tool_calls > 0: - request_logger.info( - "Sanitized %s corrupted tool_call arguments before request (session=%s)", - repaired_tool_calls, - self.session_id or "-", - ) - - # Defensive: repair malformed role-alternation before API call. - # Catches cases where the history got wedged into a - # ``tool → user`` or ``user → user`` tail (e.g. after empty- - # response scaffolding was stripped and a new user message - # landed after an orphan tool result). Most providers return - # empty content on malformed sequences, which would otherwise - # retrigger the empty-retry loop indefinitely. - repaired_seq = self._repair_message_sequence(messages) - if repaired_seq > 0: - request_logger.info( - "Repaired %s message-alternation violations before request (session=%s)", - repaired_seq, - self.session_id or "-", - ) - - api_messages = [] - for idx, msg in enumerate(messages): - api_msg = msg.copy() - - # Inject ephemeral context into the current turn's user message. - # Sources: memory manager prefetch + plugin pre_llm_call hooks - # with target="user_message" (the default). Both are - # API-call-time only — the original message in `messages` is - # never mutated, so nothing leaks into session persistence. - if idx == current_turn_user_idx and msg.get("role") == "user": - _injections = [] - if _ext_prefetch_cache: - _fenced = build_memory_context_block(_ext_prefetch_cache) - if _fenced: - _injections.append(_fenced) - if _plugin_user_context: - _injections.append(_plugin_user_context) - if _injections: - _base = api_msg.get("content", "") - if isinstance(_base, str): - api_msg["content"] = _base + "\n\n" + "\n\n".join(_injections) - - # For ALL assistant messages, pass reasoning back to the API - # This ensures multi-turn reasoning context is preserved - self._copy_reasoning_content_for_api(msg, api_msg) - - # Remove 'reasoning' field - it's for trajectory storage only - # We've copied it to 'reasoning_content' for the API above - if "reasoning" in api_msg: - api_msg.pop("reasoning") - # Remove finish_reason - not accepted by strict APIs (e.g. Mistral) - if "finish_reason" in api_msg: - api_msg.pop("finish_reason") - # Strip internal thinking-prefill marker - api_msg.pop("_thinking_prefill", None) - # Strip Codex Responses API fields (call_id, response_item_id) for - # strict providers like Mistral, Fireworks, etc. that reject unknown fields. - # Uses new dicts so the internal messages list retains the fields - # for Codex Responses compatibility. - if self._should_sanitize_tool_calls(): - self._sanitize_tool_calls_for_strict_api(api_msg) - # Keep 'reasoning_details' - OpenRouter uses this for multi-turn reasoning context - # The signature field helps maintain reasoning continuity - api_messages.append(api_msg) - - # Build the final system message: cached prompt + ephemeral system prompt. - # Ephemeral additions are API-call-time only (not persisted to session DB). - # External recall context is injected into the user message, not the system - # prompt, so the stable cache prefix remains unchanged. - # - # NOTE: Plugin context from pre_llm_call hooks is injected into the - # user message (see injection block above), NOT the system prompt. - # This is intentional — system prompt modifications break the prompt - # cache prefix. The system prompt is reserved for Hermes internals. - # - # Hermes invariant: the system prompt is built ONCE per session - # (cached on ``_cached_system_prompt``) and replayed verbatim on - # every turn. We send it as a single content string so the - # bytes are byte-stable across turns and upstream prompt caches - # stay warm. - effective_system = active_system_prompt or "" - if self.ephemeral_system_prompt: - effective_system = (effective_system + "\n\n" + self.ephemeral_system_prompt).strip() - if effective_system: - api_messages = [{"role": "system", "content": effective_system}] + api_messages - - # Inject ephemeral prefill messages right after the system prompt - # but before conversation history. Same API-call-time-only pattern. - if self.prefill_messages: - sys_offset = 1 if (api_messages and api_messages[0].get("role") == "system") else 0 - for idx, pfm in enumerate(self.prefill_messages): - api_messages.insert(sys_offset + idx, pfm.copy()) - - # Apply Anthropic prompt caching for Claude models on native - # Anthropic, OpenRouter, and third-party Anthropic-compatible - # gateways. Auto-detected: if ``_use_prompt_caching`` is set, - # inject cache_control breakpoints (system + last 3 messages) - # to reduce input token costs by ~75% on multi-turn - # conversations. - if self._use_prompt_caching: - api_messages = apply_anthropic_cache_control( - api_messages, - cache_ttl=self._cache_ttl, - native_anthropic=self._use_native_cache_layout, - ) - - # Safety net: strip orphaned tool results / add stubs for missing - # results before sending to the API. Runs unconditionally — not - # gated on context_compressor — so orphans from session loading or - # manual message manipulation are always caught. - api_messages = self._sanitize_api_messages(api_messages) - - # Drop thinking-only assistant turns (reasoning but no visible - # output and no tool_calls) and merge any adjacent user messages - # left behind. Prevents Anthropic 400s ("The final block in an - # assistant message cannot be `thinking`.") and equivalent errors - # from third-party Anthropic-compatible gateways that can't replay - # a thinking-only turn. Runs on the per-call copy only — the - # stored conversation history keeps the reasoning block for the - # UI transcript and session persistence. - api_messages = self._drop_thinking_only_and_merge_users(api_messages) - - # Normalize message whitespace and tool-call JSON for consistent - # prefix matching. Ensures bit-perfect prefixes across turns, - # which enables KV cache reuse on local inference servers - # (llama.cpp, vLLM, Ollama) and improves cache hit rates for - # cloud providers. Operates on api_messages (the API copy) so - # the original conversation history in `messages` is untouched. - for am in api_messages: - if isinstance(am.get("content"), str): - am["content"] = am["content"].strip() - for am in api_messages: - tcs = am.get("tool_calls") - if not tcs: - continue - new_tcs = [] - for tc in tcs: - if isinstance(tc, dict) and "function" in tc: - try: - args_obj = json.loads(tc["function"]["arguments"]) - tc = {**tc, "function": { - **tc["function"], - "arguments": json.dumps( - args_obj, separators=(",", ":"), - sort_keys=True, - ), - }} - except Exception: - tc["function"]["arguments"] = _repair_tool_call_arguments( - tc["function"]["arguments"], - tc["function"].get("name", "?"), - ) - new_tcs.append(tc) - am["tool_calls"] = new_tcs - - # Proactively strip any surrogate characters before the API call. - # Models served via Ollama (Kimi K2.5, GLM-5, Qwen) can return - # lone surrogates (U+D800-U+DFFF) that crash json.dumps() inside - # the OpenAI SDK. Sanitizing here prevents the 3-retry cycle. - _sanitize_messages_surrogates(api_messages) - - # Calculate approximate request size for logging - total_chars = sum(len(str(msg)) for msg in api_messages) - approx_tokens = estimate_messages_tokens_rough(api_messages) - - # Thinking spinner for quiet mode (animated during API call) - thinking_spinner = None - - if not self.quiet_mode: - self._vprint(f"\n{self.log_prefix}🔄 Making API call #{api_call_count}/{self.max_iterations}...") - self._vprint(f"{self.log_prefix} 📊 Request size: {len(api_messages)} messages, ~{approx_tokens:,} tokens (~{total_chars:,} chars)") - self._vprint(f"{self.log_prefix} 🔧 Available tools: {len(self.tools) if self.tools else 0}") - else: - # Animated thinking spinner in quiet mode - face = random.choice(KawaiiSpinner.get_thinking_faces()) - verb = random.choice(KawaiiSpinner.get_thinking_verbs()) - if self.thinking_callback: - # CLI TUI mode: use prompt_toolkit widget instead of raw spinner - # (works in both streaming and non-streaming modes) - self.thinking_callback(f"{face} {verb}...") - elif not self._has_stream_consumers() and self._should_start_quiet_spinner(): - # Raw KawaiiSpinner only when no streaming consumers and the - # spinner output has a safe sink. - spinner_type = random.choice(['brain', 'sparkle', 'pulse', 'moon', 'star']) - thinking_spinner = KawaiiSpinner(f"{face} {verb}...", spinner_type=spinner_type, print_fn=self._print_fn) - thinking_spinner.start() - - # Log request details if verbose - if self.verbose_logging: - logging.debug(f"API Request - Model: {self.model}, Messages: {len(messages)}, Tools: {len(self.tools) if self.tools else 0}") - logging.debug(f"Last message role: {messages[-1]['role'] if messages else 'none'}") - logging.debug(f"Total message size: ~{approx_tokens:,} tokens") - - api_start_time = time.time() - retry_count = 0 - max_retries = self._api_max_retries - primary_recovery_attempted = False - max_compression_attempts = 3 - codex_auth_retry_attempted=False - anthropic_auth_retry_attempted=False - nous_auth_retry_attempted=False - copilot_auth_retry_attempted=False - thinking_sig_retry_attempted = False - image_shrink_retry_attempted = False - oauth_1m_beta_retry_attempted = False - llama_cpp_grammar_retry_attempted = False - has_retried_429 = False - restart_with_compressed_messages = False - restart_with_length_continuation = False - - finish_reason = "stop" - response = None # Guard against UnboundLocalError if all retries fail - api_kwargs = None # Guard against UnboundLocalError in except handler - - while retry_count < max_retries: - # ── Nous Portal rate limit guard ────────────────────── - # If another session already recorded that Nous is rate- - # limited, skip the API call entirely. Each attempt - # (including SDK-level retries) counts against RPH and - # deepens the rate limit hole. - if self.provider == "nous": - try: - from agent.nous_rate_guard import ( - nous_rate_limit_remaining, - format_remaining as _fmt_nous_remaining, - ) - _nous_remaining = nous_rate_limit_remaining() - if _nous_remaining is not None and _nous_remaining > 0: - _nous_msg = ( - f"Nous Portal rate limit active — " - f"resets in {_fmt_nous_remaining(_nous_remaining)}." - ) - self._vprint( - f"{self.log_prefix}⏳ {_nous_msg} Trying fallback...", - force=True, - ) - self._emit_status(f"⏳ {_nous_msg}") - if self._try_activate_fallback(): - retry_count = 0 - compression_attempts = 0 - primary_recovery_attempted = False - continue - # No fallback available — return with clear message - self._persist_session(messages, conversation_history) - return { - "final_response": ( - f"⏳ {_nous_msg}\n\n" - "No fallback provider available. " - "Try again after the reset, or add a " - "fallback provider in config.yaml." - ), - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "failed": True, - "error": _nous_msg, - } - except ImportError: - pass - except Exception: - pass # Never let rate guard break the agent loop - - try: - self._reset_stream_delivery_tracking() - api_kwargs = self._build_api_kwargs(api_messages) - if self._force_ascii_payload: - _sanitize_structure_non_ascii(api_kwargs) - if self.api_mode == "codex_responses": - api_kwargs = self._get_transport().preflight_kwargs(api_kwargs, allow_stream=False) - - try: - from hermes_cli.plugins import invoke_hook as _invoke_hook - request_messages = api_kwargs.get("messages") - if not isinstance(request_messages, list): - request_messages = api_kwargs.get("input") - if not isinstance(request_messages, list): - request_messages = api_messages - # Shallow-copy the outer list so plugins that retain the - # reference for async snapshotting don't observe later - # mutations of api_messages. The inner dicts are not - # mutated by the agent loop, so a shallow copy is - # sufficient; a deepcopy would walk every tool result - # and base64 image on every API call. - _invoke_hook( - "pre_api_request", - task_id=effective_task_id, - session_id=self.session_id or "", - user_message=original_user_message, - conversation_history=list(messages), - platform=self.platform or "", - model=self.model, - provider=self.provider, - base_url=self.base_url, - api_mode=self.api_mode, - api_call_count=api_call_count, - request_messages=list(request_messages) if isinstance(request_messages, list) else [], - message_count=len(api_messages), - tool_count=len(self.tools or []), - approx_input_tokens=approx_tokens, - request_char_count=total_chars, - max_tokens=self.max_tokens, - ) - except Exception: - pass - - if env_var_enabled("HERMES_DUMP_REQUESTS"): - self._dump_api_request_debug(api_kwargs, reason="preflight") - - # Always prefer the streaming path — even without stream - # consumers. Streaming gives us fine-grained health - # checking (90s stale-stream detection, 60s read timeout) - # that the non-streaming path lacks. Without this, - # subagents and other quiet-mode callers can hang - # indefinitely when the provider keeps the connection - # alive with SSE pings but never delivers a response. - # The streaming path is a no-op for callbacks when no - # consumers are registered, and falls back to non- - # streaming automatically if the provider doesn't - # support it. - def _stop_spinner(): - nonlocal thinking_spinner - if thinking_spinner: - thinking_spinner.stop("") - thinking_spinner = None - if self.thinking_callback: - self.thinking_callback("") - - _use_streaming = True - # Provider signaled "stream not supported" on a previous - # attempt — switch to non-streaming for the rest of this - # session instead of re-failing every retry. - if getattr(self, "_disable_streaming", False): - _use_streaming = False - # CopilotACPClient communicates via subprocess stdio and - # returns a plain SimpleNamespace — not an iterable - # stream. Mirror the ACP exclusion used for Responses - # API upgrade (lines ~1083-1085). - elif ( - self.provider == "copilot-acp" - or str(self.base_url or "").lower().startswith("acp://copilot") - or str(self.base_url or "").lower().startswith("acp+tcp://") - ): - _use_streaming = False - elif not self._has_stream_consumers(): - # No display/TTS consumer. Still prefer streaming for - # health checking, but skip for Mock clients in tests - # (mocks return SimpleNamespace, not stream iterators). - from unittest.mock import Mock - if isinstance(getattr(self, "client", None), Mock): - _use_streaming = False - - if _use_streaming: - response = self._interruptible_streaming_api_call( - api_kwargs, on_first_delta=_stop_spinner - ) - else: - response = self._interruptible_api_call(api_kwargs) - - api_duration = time.time() - api_start_time - - # Stop thinking spinner silently -- the response box or tool - # execution messages that follow are more informative. - if thinking_spinner: - thinking_spinner.stop("") - thinking_spinner = None - if self.thinking_callback: - self.thinking_callback("") - - if not self.quiet_mode: - self._vprint(f"{self.log_prefix}⏱️ API call completed in {api_duration:.2f}s") - - if self.verbose_logging: - # Log response with provider info if available - resp_model = getattr(response, 'model', 'N/A') if response else 'N/A' - logging.debug(f"API Response received - Model: {resp_model}, Usage: {response.usage if hasattr(response, 'usage') else 'N/A'}") - - # Validate response shape before proceeding - response_invalid = False - error_details = [] - if self.api_mode == "codex_responses": - _ct_v = self._get_transport() - if not _ct_v.validate_response(response): - if response is None: - response_invalid = True - error_details.append("response is None") - else: - # Provider returned a terminal failure (e.g. quota exhaustion). - # Treat as invalid so the fallback chain is triggered instead of - # letting the error bubble up outside the retry/fallback loop. - _codex_resp_status = str(getattr(response, "status", "") or "").strip().lower() - if _codex_resp_status in {"failed", "cancelled"}: - _codex_error_obj = getattr(response, "error", None) - _codex_error_msg = ( - _codex_error_obj.get("message") if isinstance(_codex_error_obj, dict) - else str(_codex_error_obj) if _codex_error_obj - else f"Responses API returned status '{_codex_resp_status}'" - ) - logging.warning( - "Codex response status='%s' (error=%s). Routing to fallback. %s", - _codex_resp_status, _codex_error_msg, - self._client_log_context(), - ) - response_invalid = True - error_details.append(f"response.status={_codex_resp_status}: {_codex_error_msg}") - else: - # output_text fallback: stream backfill may have failed - # but normalize can still recover from output_text - _out_text = getattr(response, "output_text", None) - _out_text_stripped = _out_text.strip() if isinstance(_out_text, str) else "" - if _out_text_stripped: - logger.debug( - "Codex response.output is empty but output_text is present " - "(%d chars); deferring to normalization.", - len(_out_text_stripped), - ) - else: - _resp_status = getattr(response, "status", None) - _resp_incomplete = getattr(response, "incomplete_details", None) - logger.warning( - "Codex response.output is empty after stream backfill " - "(status=%s, incomplete_details=%s, model=%s). %s", - _resp_status, _resp_incomplete, - getattr(response, "model", None), - f"api_mode={self.api_mode} provider={self.provider}", - ) - response_invalid = True - error_details.append("response.output is empty") - elif self.api_mode == "anthropic_messages": - _tv = self._get_transport() - if not _tv.validate_response(response): - response_invalid = True - if response is None: - error_details.append("response is None") - else: - error_details.append("response.content invalid (not a non-empty list)") - elif self.api_mode == "bedrock_converse": - _btv = self._get_transport() - if not _btv.validate_response(response): - response_invalid = True - if response is None: - error_details.append("response is None") - else: - error_details.append("Bedrock response invalid (no output or choices)") - else: - _ctv = self._get_transport() - if not _ctv.validate_response(response): - response_invalid = True - if response is None: - error_details.append("response is None") - elif not hasattr(response, 'choices'): - error_details.append("response has no 'choices' attribute") - elif response.choices is None: - error_details.append("response.choices is None") - else: - error_details.append("response.choices is empty") - - if response_invalid: - # Stop spinner before printing error messages - if thinking_spinner: - thinking_spinner.stop("(´;ω;`) oops, retrying...") - thinking_spinner = None - if self.thinking_callback: - self.thinking_callback("") - - # Invalid response — could be rate limiting, provider timeout, - # upstream server error, or malformed response. - retry_count += 1 - - # Eager fallback: empty/malformed responses are a common - # rate-limit symptom. Switch to fallback immediately - # rather than retrying with extended backoff. - if self._fallback_index < len(self._fallback_chain): - self._emit_status("⚠️ Empty/malformed response — switching to fallback...") - if self._try_activate_fallback(): - retry_count = 0 - compression_attempts = 0 - primary_recovery_attempted = False - continue - - # Check for error field in response (some providers include this) - error_msg = "Unknown" - provider_name = "Unknown" - if response and hasattr(response, 'error') and response.error: - error_msg = str(response.error) - # Try to extract provider from error metadata - if hasattr(response.error, 'metadata') and response.error.metadata: - provider_name = response.error.metadata.get('provider_name', 'Unknown') - elif response and hasattr(response, 'message') and response.message: - error_msg = str(response.message) - - # Try to get provider from model field (OpenRouter often returns actual model used) - if provider_name == "Unknown" and response and hasattr(response, 'model') and response.model: - provider_name = f"model={response.model}" - - # Check for x-openrouter-provider or similar metadata - if provider_name == "Unknown" and response: - # Log all response attributes for debugging - resp_attrs = {k: str(v)[:100] for k, v in vars(response).items() if not k.startswith('_')} - if self.verbose_logging: - logging.debug(f"Response attributes for invalid response: {resp_attrs}") - - # Extract error code from response for contextual diagnostics - _resp_error_code = None - if response and hasattr(response, 'error') and response.error: - _code_raw = getattr(response.error, 'code', None) - if _code_raw is None and isinstance(response.error, dict): - _code_raw = response.error.get('code') - if _code_raw is not None: - try: - _resp_error_code = int(_code_raw) - except (TypeError, ValueError): - pass - - # Build a human-readable failure hint from the error code - # and response time, instead of always assuming rate limiting. - if _resp_error_code == 524: - _failure_hint = f"upstream provider timed out (Cloudflare 524, {api_duration:.0f}s)" - elif _resp_error_code == 504: - _failure_hint = f"upstream gateway timeout (504, {api_duration:.0f}s)" - elif _resp_error_code == 429: - _failure_hint = f"rate limited by upstream provider (429)" - elif _resp_error_code in {500, 502}: - _failure_hint = f"upstream server error ({_resp_error_code}, {api_duration:.0f}s)" - elif _resp_error_code in {503, 529}: - _failure_hint = f"upstream provider overloaded ({_resp_error_code})" - elif _resp_error_code is not None: - _failure_hint = f"upstream error (code {_resp_error_code}, {api_duration:.0f}s)" - elif api_duration < 10: - _failure_hint = f"fast response ({api_duration:.1f}s) — likely rate limited" - elif api_duration > 60: - _failure_hint = f"slow response ({api_duration:.0f}s) — likely upstream timeout" - else: - _failure_hint = f"response time {api_duration:.1f}s" - - self._vprint(f"{self.log_prefix}⚠️ Invalid API response (attempt {retry_count}/{max_retries}): {', '.join(error_details)}", force=True) - self._vprint(f"{self.log_prefix} 🏢 Provider: {provider_name}", force=True) - cleaned_provider_error = self._clean_error_message(error_msg) - self._vprint(f"{self.log_prefix} 📝 Provider message: {cleaned_provider_error}", force=True) - self._vprint(f"{self.log_prefix} ⏱️ {_failure_hint}", force=True) - - if retry_count >= max_retries: - # Try fallback before giving up - self._emit_status(f"⚠️ Max retries ({max_retries}) for invalid responses — trying fallback...") - if self._try_activate_fallback(): - retry_count = 0 - compression_attempts = 0 - primary_recovery_attempted = False - continue - self._emit_status(f"❌ Max retries ({max_retries}) exceeded for invalid responses. Giving up.") - logging.error(f"{self.log_prefix}Invalid API response after {max_retries} retries.") - self._persist_session(messages, conversation_history) - return { - "messages": messages, - "completed": False, - "api_calls": api_call_count, - "error": f"Invalid API response after {max_retries} retries: {_failure_hint}", - "failed": True # Mark as failure for filtering - } - - # Backoff before retry — jittered exponential: 5s base, 120s cap - wait_time = jittered_backoff(retry_count, base_delay=5.0, max_delay=120.0) - self._vprint(f"{self.log_prefix}⏳ Retrying in {wait_time:.1f}s ({_failure_hint})...", force=True) - logging.warning(f"Invalid API response (retry {retry_count}/{max_retries}): {', '.join(error_details)} | Provider: {provider_name}") - - # Sleep in small increments to stay responsive to interrupts - sleep_end = time.time() + wait_time - _backoff_touch_counter = 0 - while time.time() < sleep_end: - if self._interrupt_requested: - self._vprint(f"{self.log_prefix}⚡ Interrupt detected during retry wait, aborting.", force=True) - self._persist_session(messages, conversation_history) - self.clear_interrupt() - return { - "final_response": f"Operation interrupted during retry ({_failure_hint}, attempt {retry_count}/{max_retries}).", - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "interrupted": True, - } - time.sleep(0.2) - # Touch activity every ~30s so the gateway's inactivity - # monitor knows we're alive during backoff waits. - _backoff_touch_counter += 1 - if _backoff_touch_counter % 150 == 0: # 150 × 0.2s = 30s - self._touch_activity( - f"retry backoff ({retry_count}/{max_retries}), " - f"{int(sleep_end - time.time())}s remaining" - ) - continue # Retry the API call - - # Check finish_reason before proceeding - if self.api_mode == "codex_responses": - status = getattr(response, "status", None) - incomplete_details = getattr(response, "incomplete_details", None) - incomplete_reason = None - if isinstance(incomplete_details, dict): - incomplete_reason = incomplete_details.get("reason") - else: - incomplete_reason = getattr(incomplete_details, "reason", None) - if status == "incomplete" and incomplete_reason in {"max_output_tokens", "length"}: - finish_reason = "length" - else: - finish_reason = "stop" - elif self.api_mode == "anthropic_messages": - _tfr = self._get_transport() - finish_reason = _tfr.map_finish_reason(response.stop_reason) - elif self.api_mode == "bedrock_converse": - # Bedrock response already normalized at dispatch — use transport - _bt_fr = self._get_transport() - _bedrock_result = _bt_fr.normalize_response(response) - finish_reason = _bedrock_result.finish_reason - else: - _cc_fr = self._get_transport() - _finish_result = _cc_fr.normalize_response(response) - finish_reason = _finish_result.finish_reason - assistant_message = _finish_result - if self._should_treat_stop_as_truncated( - finish_reason, - assistant_message, - messages, - ): - self._vprint( - f"{self.log_prefix}⚠️ Treating suspicious Ollama/GLM stop response as truncated", - force=True, - ) - finish_reason = "length" - - if finish_reason == "length": - self._vprint(f"{self.log_prefix}⚠️ Response truncated (finish_reason='length') - model hit max output tokens", force=True) - - # Normalize the truncated response to a single OpenAI-style - # message shape so text-continuation and tool-call retry - # work uniformly across chat_completions, bedrock_converse, - # and anthropic_messages. For Anthropic we use the same - # adapter the agent loop already relies on so the rebuilt - # interim assistant message is byte-identical to what - # would have been appended in the non-truncated path. - _trunc_msg = None - _trunc_transport = self._get_transport() - if self.api_mode == "anthropic_messages": - _trunc_result = _trunc_transport.normalize_response( - response, strip_tool_prefix=self._is_anthropic_oauth - ) - else: - _trunc_result = _trunc_transport.normalize_response(response) - _trunc_msg = _trunc_result - - _trunc_content = getattr(_trunc_msg, "content", None) if _trunc_msg else None - _trunc_has_tool_calls = bool(getattr(_trunc_msg, "tool_calls", None)) if _trunc_msg else False - - # ── Detect thinking-budget exhaustion ────────────── - # When the model spends ALL output tokens on reasoning - # and has none left for the response, continuation - # retries are pointless. Detect this early and give a - # targeted error instead of wasting 3 API calls. - # A response is "thinking exhausted" only when the model - # actually produced reasoning blocks but no visible text after - # them. Models that do not use tags (e.g. GLM-4.7 on - # NVIDIA Build, minimax) may return content=None or an empty - # string for unrelated reasons — treat those as normal - # truncations that deserve continuation retries, not as - # thinking-budget exhaustion. - _has_think_tags = bool( - _trunc_content and re.search( - r'<(?:think|thinking|reasoning|REASONING_SCRATCHPAD)[^>]*>', - _trunc_content, - re.IGNORECASE, - ) - ) - _thinking_exhausted = ( - not _trunc_has_tool_calls - and _has_think_tags - and ( - (_trunc_content is not None and not self._has_content_after_think_block(_trunc_content)) - or _trunc_content is None - ) - ) - - if _thinking_exhausted: - _exhaust_error = ( - "Model used all output tokens on reasoning with none left " - "for the response. Try lowering reasoning effort or " - "increasing max_tokens." - ) - self._vprint( - f"{self.log_prefix}💭 Reasoning exhausted the output token budget — " - f"no visible response was produced.", - force=True, - ) - # Return a user-friendly message as the response so - # CLI (response box) and gateway (chat message) both - # display it naturally instead of a suppressed error. - _exhaust_response = ( - "⚠️ **Thinking Budget Exhausted**\n\n" - "The model used all its output tokens on reasoning " - "and had none left for the actual response.\n\n" - "To fix this:\n" - "→ Lower reasoning effort: `/thinkon low` or `/thinkon minimal`\n" - "→ Or switch to a larger/non-reasoning model with `/model`" - ) - self._cleanup_task_resources(effective_task_id) - self._persist_session(messages, conversation_history) - return { - "final_response": _exhaust_response, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "partial": True, - "error": _exhaust_error, - } - - if self.api_mode in {"chat_completions", "bedrock_converse", "anthropic_messages"}: - assistant_message = _trunc_msg - if assistant_message is not None and not _trunc_has_tool_calls: - length_continue_retries += 1 - interim_msg = self._build_assistant_message(assistant_message, finish_reason) - messages.append(interim_msg) - if assistant_message.content: - truncated_response_parts.append(assistant_message.content) - - if length_continue_retries < 3: - self._vprint( - f"{self.log_prefix}↻ Requesting continuation " - f"({length_continue_retries}/3)..." - ) - continue_msg = { - "role": "user", - "content": ( - "[System: Your previous response was truncated by the output " - "length limit. Continue exactly where you left off. Do not " - "restart or repeat prior text. Finish the answer directly.]" - ), - } - messages.append(continue_msg) - self._session_messages = messages - self._save_session_log(messages) - restart_with_length_continuation = True - break - - partial_response = self._strip_think_blocks("".join(truncated_response_parts)).strip() - self._cleanup_task_resources(effective_task_id) - self._persist_session(messages, conversation_history) - return { - "final_response": partial_response or None, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "partial": True, - "error": "Response remained truncated after 3 continuation attempts", - } - - if self.api_mode in {"chat_completions", "bedrock_converse", "anthropic_messages"}: - assistant_message = _trunc_msg - if assistant_message is not None and _trunc_has_tool_calls: - if truncated_tool_call_retries < 1: - truncated_tool_call_retries += 1 - self._vprint( - f"{self.log_prefix}⚠️ Truncated tool call detected — retrying API call...", - force=True, - ) - # Don't append the broken response to messages; - # just re-run the same API call from the current - # message state, giving the model another chance. - continue - self._vprint( - f"{self.log_prefix}⚠️ Truncated tool call response detected again — refusing to execute incomplete tool arguments.", - force=True, - ) - self._cleanup_task_resources(effective_task_id) - self._persist_session(messages, conversation_history) - return { - "final_response": None, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "partial": True, - "error": "Response truncated due to output length limit", - } - - # If we have prior messages, roll back to last complete state - if len(messages) > 1: - self._vprint(f"{self.log_prefix} ⏪ Rolling back to last complete assistant turn") - rolled_back_messages = self._get_messages_up_to_last_assistant(messages) - - self._cleanup_task_resources(effective_task_id) - self._persist_session(messages, conversation_history) - - return { - "final_response": None, - "messages": rolled_back_messages, - "api_calls": api_call_count, - "completed": False, - "partial": True, - "error": "Response truncated due to output length limit" - } - else: - # First message was truncated - mark as failed - self._vprint(f"{self.log_prefix}❌ First response truncated - cannot recover", force=True) - self._persist_session(messages, conversation_history) - return { - "final_response": None, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "failed": True, - "error": "First response truncated due to output length limit" - } - - # Track actual token usage from response for context management - if hasattr(response, 'usage') and response.usage: - canonical_usage = normalize_usage( - response.usage, - provider=self.provider, - api_mode=self.api_mode, - ) - prompt_tokens = canonical_usage.prompt_tokens - completion_tokens = canonical_usage.output_tokens - total_tokens = canonical_usage.total_tokens - usage_dict = { - "prompt_tokens": prompt_tokens, - "completion_tokens": completion_tokens, - "total_tokens": total_tokens, - } - self.context_compressor.update_from_response(usage_dict) - - # Cache discovered context length after successful call. - # Only persist limits confirmed by the provider (parsed - # from the error message), not guessed probe tiers. - if getattr(self.context_compressor, "_context_probed", False): - ctx = self.context_compressor.context_length - if getattr(self.context_compressor, "_context_probe_persistable", False): - save_context_length(self.model, self.base_url, ctx) - self._safe_print(f"{self.log_prefix}💾 Cached context length: {ctx:,} tokens for {self.model}") - self.context_compressor._context_probed = False - self.context_compressor._context_probe_persistable = False - - self.session_prompt_tokens += prompt_tokens - self.session_completion_tokens += completion_tokens - self.session_total_tokens += total_tokens - self.session_api_calls += 1 - self.session_input_tokens += canonical_usage.input_tokens - self.session_output_tokens += canonical_usage.output_tokens - self.session_cache_read_tokens += canonical_usage.cache_read_tokens - self.session_cache_write_tokens += canonical_usage.cache_write_tokens - self.session_reasoning_tokens += canonical_usage.reasoning_tokens - - # Log API call details for debugging/observability - _cache_pct = "" - if canonical_usage.cache_read_tokens and prompt_tokens: - _cache_pct = f" cache={canonical_usage.cache_read_tokens}/{prompt_tokens} ({100*canonical_usage.cache_read_tokens/prompt_tokens:.0f}%)" - logger.info( - "API call #%d: model=%s provider=%s in=%d out=%d total=%d latency=%.1fs%s", - self.session_api_calls, self.model, self.provider or "unknown", - prompt_tokens, completion_tokens, total_tokens, - api_duration, _cache_pct, - ) - - cost_result = estimate_usage_cost( - self.model, - canonical_usage, - provider=self.provider, - base_url=self.base_url, - api_key=getattr(self, "api_key", ""), - ) - if cost_result.amount_usd is not None: - self.session_estimated_cost_usd += float(cost_result.amount_usd) - self.session_cost_status = cost_result.status - self.session_cost_source = cost_result.source - - # Persist token counts to session DB for /insights. - # Do this for every platform with a session_id so non-CLI - # sessions (gateway, cron, delegated runs) cannot lose - # token/accounting data if a higher-level persistence path - # is skipped or fails. Gateway/session-store writes use - # absolute totals, so they safely overwrite these per-call - # deltas instead of double-counting them. - if self._session_db and self.session_id: - try: - # Ensure the session row exists before attempting UPDATE. - # Under concurrent load (cron/kanban), the initial - # _ensure_db_session() may have failed due to SQLite - # locking. Retry here so per-call token deltas are - # not silently lost (UPDATE on a non-existent row - # affects 0 rows without error). - if not self._session_db_created: - self._ensure_db_session() - self._session_db.update_token_counts( - self.session_id, - input_tokens=canonical_usage.input_tokens, - output_tokens=canonical_usage.output_tokens, - cache_read_tokens=canonical_usage.cache_read_tokens, - cache_write_tokens=canonical_usage.cache_write_tokens, - reasoning_tokens=canonical_usage.reasoning_tokens, - estimated_cost_usd=float(cost_result.amount_usd) - if cost_result.amount_usd is not None else None, - cost_status=cost_result.status, - cost_source=cost_result.source, - billing_provider=self.provider, - billing_base_url=self.base_url, - billing_mode="subscription_included" - if cost_result.status == "included" else None, - model=self.model, - api_call_count=1, - ) - except Exception as e: - # Log token persistence failures so they're - # visible in agent.log — silent loss here is - # the root cause of undercounted analytics. - logger.debug( - "Token persistence failed (session=%s, tokens=%d): %s", - self.session_id, total_tokens, e, - ) - - if self.verbose_logging: - logging.debug(f"Token usage: prompt={usage_dict['prompt_tokens']:,}, completion={usage_dict['completion_tokens']:,}, total={usage_dict['total_tokens']:,}") - - # Surface cache hit stats for any provider that reports - # them — not just those where we inject cache_control - # markers. OpenAI/Kimi/DeepSeek/Qwen all do automatic - # server-side prefix caching and return - # ``prompt_tokens_details.cached_tokens``; users - # previously could not see their cache % because this - # line was gated on ``_use_prompt_caching``, which is - # only True for Anthropic-style marker injection. - # ``canonical_usage`` is already normalised from all - # three API shapes (Anthropic / Codex / OpenAI-chat) - # so we can rely on its values directly. - cached = canonical_usage.cache_read_tokens - written = canonical_usage.cache_write_tokens - prompt = usage_dict["prompt_tokens"] - if (cached or written) and not self.quiet_mode: - hit_pct = (cached / prompt * 100) if prompt > 0 else 0 - self._vprint( - f"{self.log_prefix} 💾 Cache: " - f"{cached:,}/{prompt:,} tokens " - f"({hit_pct:.0f}% hit, {written:,} written)" - ) - - has_retried_429 = False # Reset on success - # Clear Nous rate limit state on successful request — - # proves the limit has reset and other sessions can - # resume hitting Nous. - if self.provider == "nous": - try: - from agent.nous_rate_guard import clear_nous_rate_limit - clear_nous_rate_limit() - except Exception: - pass - self._touch_activity(f"API call #{api_call_count} completed") - break # Success, exit retry loop - - except InterruptedError: - if thinking_spinner: - thinking_spinner.stop("") - thinking_spinner = None - if self.thinking_callback: - self.thinking_callback("") - api_elapsed = time.time() - api_start_time - self._vprint(f"{self.log_prefix}⚡ Interrupted during API call.", force=True) - self._persist_session(messages, conversation_history) - interrupted = True - final_response = f"Operation interrupted: waiting for model response ({api_elapsed:.1f}s elapsed)." - break - - except Exception as api_error: - # Stop spinner before printing error messages - if thinking_spinner: - thinking_spinner.stop("(╥_╥) error, retrying...") - thinking_spinner = None - if self.thinking_callback: - self.thinking_callback("") - - # ----------------------------------------------------------- - # UnicodeEncodeError recovery. Two common causes: - # 1. Lone surrogates (U+D800..U+DFFF) from clipboard paste - # (Google Docs, rich-text editors) — sanitize and retry. - # 2. ASCII codec on systems with LANG=C or non-UTF-8 locale - # (e.g. Chromebooks) — any non-ASCII character fails. - # Detect via the error message mentioning 'ascii' codec. - # We sanitize messages in-place and may retry twice: - # first to strip surrogates, then once more for pure - # ASCII-only locale sanitization if needed. - # ----------------------------------------------------------- - if isinstance(api_error, UnicodeEncodeError) and getattr(self, '_unicode_sanitization_passes', 0) < 2: - _err_str = str(api_error).lower() - _is_ascii_codec = "'ascii'" in _err_str or "ascii" in _err_str - # Detect surrogate errors — utf-8 codec refusing to - # encode U+D800..U+DFFF. The error text is: - # "'utf-8' codec can't encode characters in position - # N-M: surrogates not allowed" - _is_surrogate_error = ( - "surrogate" in _err_str - or ("'utf-8'" in _err_str and not _is_ascii_codec) - ) - # Sanitize surrogates from both the canonical `messages` - # list AND `api_messages` (the API-copy, which may carry - # `reasoning_content`/`reasoning_details` transformed - # from `reasoning` — fields the canonical list doesn't - # have directly). Also clean `api_kwargs` if built and - # `prefill_messages` if present. Mirrors the ASCII - # codec recovery below. - _surrogates_found = _sanitize_messages_surrogates(messages) - if isinstance(api_messages, list): - if _sanitize_messages_surrogates(api_messages): - _surrogates_found = True - if isinstance(api_kwargs, dict): - if _sanitize_structure_surrogates(api_kwargs): - _surrogates_found = True - if isinstance(getattr(self, "prefill_messages", None), list): - if _sanitize_messages_surrogates(self.prefill_messages): - _surrogates_found = True - # Gate the retry on the error type, not on whether we - # found anything — _force_ascii_payload / the extended - # surrogate walker above cover all known paths, but a - # new transformed field could still slip through. If - # the error was a surrogate encode failure, always let - # the retry run; the proactive sanitizer at line ~8781 - # runs again on the next iteration. Bounded by - # _unicode_sanitization_passes < 2 (outer guard). - if _surrogates_found or _is_surrogate_error: - self._unicode_sanitization_passes += 1 - if _surrogates_found: - self._vprint( - f"{self.log_prefix}⚠️ Stripped invalid surrogate characters from messages. Retrying...", - force=True, - ) - else: - self._vprint( - f"{self.log_prefix}⚠️ Surrogate encoding error — retrying after full-payload sanitization...", - force=True, - ) - continue - if _is_ascii_codec: - self._force_ascii_payload = True - # ASCII codec: the system encoding can't handle - # non-ASCII characters at all. Sanitize all - # non-ASCII content from messages/tool schemas and retry. - # Sanitize both the canonical `messages` list and - # `api_messages` (the API-copy built before the retry - # loop, which may contain extra fields like - # reasoning_content that are not in `messages`). - _messages_sanitized = _sanitize_messages_non_ascii(messages) - if isinstance(api_messages, list): - _sanitize_messages_non_ascii(api_messages) - # Also sanitize the last api_kwargs if already built, - # so a leftover non-ASCII value in a transformed field - # (e.g. extra_body, reasoning_content) doesn't survive - # into the next attempt via _build_api_kwargs cache paths. - if isinstance(api_kwargs, dict): - _sanitize_structure_non_ascii(api_kwargs) - _prefill_sanitized = False - if isinstance(getattr(self, "prefill_messages", None), list): - _prefill_sanitized = _sanitize_messages_non_ascii(self.prefill_messages) - - _tools_sanitized = False - if isinstance(getattr(self, "tools", None), list): - _tools_sanitized = _sanitize_tools_non_ascii(self.tools) - - _system_sanitized = False - if isinstance(active_system_prompt, str): - _sanitized_system = _strip_non_ascii(active_system_prompt) - if _sanitized_system != active_system_prompt: - active_system_prompt = _sanitized_system - self._cached_system_prompt = _sanitized_system - _system_sanitized = True - if isinstance(getattr(self, "ephemeral_system_prompt", None), str): - _sanitized_ephemeral = _strip_non_ascii(self.ephemeral_system_prompt) - if _sanitized_ephemeral != self.ephemeral_system_prompt: - self.ephemeral_system_prompt = _sanitized_ephemeral - _system_sanitized = True - - _headers_sanitized = False - _default_headers = ( - self._client_kwargs.get("default_headers") - if isinstance(getattr(self, "_client_kwargs", None), dict) - else None - ) - if isinstance(_default_headers, dict): - _headers_sanitized = _sanitize_structure_non_ascii(_default_headers) - - # Sanitize the API key — non-ASCII characters in - # credentials (e.g. ʋ instead of v from a bad - # copy-paste) cause httpx to fail when encoding - # the Authorization header as ASCII. This is the - # most common cause of persistent UnicodeEncodeError - # that survives message/tool sanitization (#6843). - _credential_sanitized = False - _raw_key = getattr(self, "api_key", None) or "" - if _raw_key: - _clean_key = _strip_non_ascii(_raw_key) - if _clean_key != _raw_key: - self.api_key = _clean_key - if isinstance(getattr(self, "_client_kwargs", None), dict): - self._client_kwargs["api_key"] = _clean_key - # Also update the live client — it holds its - # own copy of api_key which auth_headers reads - # dynamically on every request. - if getattr(self, "client", None) is not None and hasattr(self.client, "api_key"): - self.client.api_key = _clean_key - _credential_sanitized = True - self._vprint( - f"{self.log_prefix}⚠️ API key contained non-ASCII characters " - f"(bad copy-paste?) — stripped them. If auth fails, " - f"re-copy the key from your provider's dashboard.", - force=True, - ) - - # Always retry on ASCII codec detection — - # _force_ascii_payload guarantees the full - # api_kwargs payload is sanitized on the - # next iteration (line ~8475). Even when - # per-component checks above find nothing - # (e.g. non-ASCII only in api_messages' - # reasoning_content), the flag catches it. - # Bounded by _unicode_sanitization_passes < 2. - self._unicode_sanitization_passes += 1 - _any_sanitized = ( - _messages_sanitized - or _prefill_sanitized - or _tools_sanitized - or _system_sanitized - or _headers_sanitized - or _credential_sanitized - ) - if _any_sanitized: - self._vprint( - f"{self.log_prefix}⚠️ System encoding is ASCII — stripped non-ASCII characters from request payload. Retrying...", - force=True, - ) - else: - self._vprint( - f"{self.log_prefix}⚠️ System encoding is ASCII — enabling full-payload sanitization for retry...", - force=True, - ) - continue - - # ── Image-rejection recovery ────────────────────────────── - # Some providers (mlx-lm, text-only endpoints, text-only - # fallbacks on multimodal models) reject any message that - # contains image_url content with a 4xx error like - # "Only 'text' content type is supported." On first hit, - # strip all images from the message list, mark the session - # as vision-unsupported, and retry with text only. - # - # Detection is best-effort English phrase matching — a - # locale-translated or heavily-reworded upstream error - # will bypass this guard and fall through to the normal - # error handler. Expand the phrase list when new - # provider wordings are observed in the wild. - _err_body = "" - try: - _err_body = str(getattr(api_error, "body", None) or - getattr(api_error, "message", None) or - str(api_error)) - except Exception: - pass - _err_status = getattr(api_error, "status_code", None) - _IMAGE_REJECTION_PHRASES = ( - "only 'text' content type is supported", - "only text content type is supported", - "image_url is not supported", - "image content is not supported", - "multimodal is not supported", - "multimodal content is not supported", - "multimodal input is not supported", - "vision is not supported", - "vision input is not supported", - "does not support images", - "does not support image input", - "does not support multimodal", - "does not support vision", - "model does not support image", - # ChatGPT-account Codex backend - # (https://chatgpt.com/backend-api/codex) rejects - # data:image/...base64 URLs in input_image fields - # with HTTP 400 "Invalid 'input[N].content[K].image_url'. - # Expected a valid URL, but got a value with an - # invalid format." The OpenAI Responses API on the - # public endpoint accepts data URLs, but the - # ChatGPT-account variant does not. Without this - # phrase the agent cascaded into compression / - # context-too-large recovery instead of just - # stripping the images. Match is narrow on - # purpose — keyed on the field-path apostrophe so - # we don't false-trip on other URL validation - # errors. (issue #23570) - "image_url'. expected", - # DeepSeek's OpenAI-compatible API reports text-only - # request-body variants as: - # "unknown variant `image_url`, expected `text`". - "unknown variant `image_url`, expected `text`", - "unknown variant image_url, expected text", - ) - _err_lower = _err_body.lower() - _looks_like_image_rejection = any( - p in _err_lower for p in _IMAGE_REJECTION_PHRASES - ) - # 4xx-only gate: never interpret 5xx/timeout as "server - # said no to images" — those are transient and must - # route to the normal retry path. - _status_ok = _err_status is None or (400 <= int(_err_status) < 500) - if ( - getattr(self, "_vision_supported", True) - and _looks_like_image_rejection - and _status_ok - ): - self._vision_supported = False - _imgs_removed = _strip_images_from_messages(messages) - if isinstance(api_messages, list): - _strip_images_from_messages(api_messages) - self._vprint( - f"{self.log_prefix}⚠️ Server rejected image content — " - f"switching to text-only mode for this session" - + (". Stripped images from history and retrying." if _imgs_removed else "."), - force=True, - ) - continue - - status_code = getattr(api_error, "status_code", None) - error_context = self._extract_api_error_context(api_error) - - # ── Classify the error for structured recovery decisions ── - _compressor = getattr(self, "context_compressor", None) - _ctx_len = getattr(_compressor, "context_length", 200000) if _compressor else 200000 - classified = classify_api_error( - api_error, - provider=getattr(self, "provider", "") or "", - model=getattr(self, "model", "") or "", - approx_tokens=approx_tokens, - context_length=_ctx_len, - num_messages=len(api_messages) if api_messages else 0, - ) - logger.debug( - "Error classified: reason=%s status=%s retryable=%s compress=%s rotate=%s fallback=%s", - classified.reason.value, classified.status_code, - classified.retryable, classified.should_compress, - classified.should_rotate_credential, classified.should_fallback, - ) - - recovered_with_pool, has_retried_429 = self._recover_with_credential_pool( - status_code=status_code, - has_retried_429=has_retried_429, - classified_reason=classified.reason, - error_context=error_context, - ) - if recovered_with_pool: - continue - - # Image-too-large recovery: shrink oversized native image - # parts in-place and retry once. Triggered by Anthropic's - # per-image 5 MB ceiling (400 with "image exceeds 5 MB - # maximum") or any other provider that complains about - # image size. If shrink fails or a second attempt still - # fails, fall through to normal error handling. - if ( - classified.reason == FailoverReason.image_too_large - and not image_shrink_retry_attempted - ): - image_shrink_retry_attempted = True - if self._try_shrink_image_parts_in_messages(api_messages): - self._vprint( - f"{self.log_prefix}📐 Image(s) exceeded provider size limit — " - f"shrank and retrying...", - force=True, - ) - continue - else: - logger.info( - "image-shrink recovery: no data-URL image parts found " - "or shrink didn't reduce size; surfacing original error." - ) - - # Anthropic OAuth subscription rejected the 1M-context beta - # header ("long context beta is not yet available for this - # subscription"). Disable the beta for the rest of this - # session, rebuild the client, and retry once. 1M-capable - # subscriptions never hit this branch — they accept the - # beta and keep full 1M context. See PR #17680 for the - # original report (we chose reactive recovery over the - # proposed unconditional omit so capable subscriptions - # don't silently lose the capability). - if ( - classified.reason == FailoverReason.oauth_long_context_beta_forbidden - and self.api_mode == "anthropic_messages" - and self._is_anthropic_oauth - and not oauth_1m_beta_retry_attempted - ): - oauth_1m_beta_retry_attempted = True - if not getattr(self, "_oauth_1m_beta_disabled", False): - self._oauth_1m_beta_disabled = True - try: - self._anthropic_client.close() - except Exception: - pass - self._rebuild_anthropic_client() - self._vprint( - f"{self.log_prefix}🔕 OAuth subscription doesn't support " - f"the 1M-context beta — disabled for this session and retrying...", - force=True, - ) - continue - - if ( - self.api_mode == "codex_responses" - and self.provider in {"openai-codex", "xai-oauth"} - and status_code == 401 - and not codex_auth_retry_attempted - ): - codex_auth_retry_attempted = True - if self._try_refresh_codex_client_credentials(force=True): - _label = "xAI OAuth" if self.provider == "xai-oauth" else "Codex" - self._vprint(f"{self.log_prefix}🔐 {_label} auth refreshed after 401. Retrying request...") - continue - if ( - self.api_mode == "chat_completions" - and self.provider == "nous" - and status_code == 401 - and not nous_auth_retry_attempted - ): - nous_auth_retry_attempted = True - if self._try_refresh_nous_client_credentials(force=True): - print(f"{self.log_prefix}🔐 Nous agent key refreshed after 401. Retrying request...") - continue - # Credential refresh didn't help — show diagnostic info. - # Most common causes: Portal OAuth expired/revoked, - # account out of credits, or agent key blocked. - from hermes_constants import display_hermes_home as _dhh_fn - _dhh = _dhh_fn() - _body_text = "" - try: - _body = getattr(api_error, "body", None) or getattr(api_error, "response", None) - if _body is not None: - _body_text = str(_body)[:200] - except Exception: - pass - print(f"{self.log_prefix}🔐 Nous 401 — Portal authentication failed.") - if _body_text: - print(f"{self.log_prefix} Response: {_body_text}") - print(f"{self.log_prefix} Most likely: Portal OAuth expired, account out of credits, or agent key revoked.") - print(f"{self.log_prefix} Troubleshooting:") - print(f"{self.log_prefix} • Re-authenticate: hermes login --provider nous") - print(f"{self.log_prefix} • Check credits / billing: https://portal.nousresearch.com") - print(f"{self.log_prefix} • Verify stored credentials: {_dhh}/auth.json") - print(f"{self.log_prefix} • Switch providers temporarily: /model --provider openrouter") - if ( - self.provider == "copilot" - and status_code == 401 - and not copilot_auth_retry_attempted - ): - copilot_auth_retry_attempted = True - if self._try_refresh_copilot_client_credentials(): - self._vprint(f"{self.log_prefix}🔐 Copilot credentials refreshed after 401. Retrying request...") - continue - if ( - self.api_mode == "anthropic_messages" - and status_code == 401 - and hasattr(self, '_anthropic_api_key') - and not anthropic_auth_retry_attempted - ): - anthropic_auth_retry_attempted = True - from agent.anthropic_adapter import _is_oauth_token - if self._try_refresh_anthropic_client_credentials(): - print(f"{self.log_prefix}🔐 Anthropic credentials refreshed after 401. Retrying request...") - continue - # Credential refresh didn't help — show diagnostic info - key = self._anthropic_api_key - auth_method = "Bearer (OAuth/setup-token)" if _is_oauth_token(key) else "x-api-key (API key)" - print(f"{self.log_prefix}🔐 Anthropic 401 — authentication failed.") - print(f"{self.log_prefix} Auth method: {auth_method}") - print(f"{self.log_prefix} Token prefix: {key[:12]}..." if key and len(key) > 12 else f"{self.log_prefix} Token: (empty or short)") - print(f"{self.log_prefix} Troubleshooting:") - from hermes_constants import display_hermes_home as _dhh_fn - _dhh = _dhh_fn() - print(f"{self.log_prefix} • Check ANTHROPIC_TOKEN in {_dhh}/.env for Hermes-managed OAuth/setup tokens") - print(f"{self.log_prefix} • Check ANTHROPIC_API_KEY in {_dhh}/.env for API keys or legacy token values") - print(f"{self.log_prefix} • For API keys: verify at https://platform.claude.com/settings/keys") - print(f"{self.log_prefix} • For Claude Code: run 'claude /login' to refresh, then retry") - print(f"{self.log_prefix} • Legacy cleanup: hermes config set ANTHROPIC_TOKEN \"\"") - print(f"{self.log_prefix} • Clear stale keys: hermes config set ANTHROPIC_API_KEY \"\"") - - # ── Thinking block signature recovery ───────────────── - # Anthropic signs thinking blocks against the full turn - # content. Any upstream mutation (context compression, - # session truncation, message merging) invalidates the - # signature → HTTP 400. Recovery: strip reasoning_details - # from all messages so the next retry sends no thinking - # blocks at all. One-shot — don't retry infinitely. - if ( - classified.reason == FailoverReason.thinking_signature - and not thinking_sig_retry_attempted - ): - thinking_sig_retry_attempted = True - for _m in messages: - if isinstance(_m, dict): - _m.pop("reasoning_details", None) - self._vprint( - f"{self.log_prefix}⚠️ Thinking block signature invalid — " - f"stripped all thinking blocks, retrying...", - force=True, - ) - logging.warning( - "%sThinking block signature recovery: stripped " - "reasoning_details from %d messages", - self.log_prefix, len(messages), - ) - continue - - # ── llama.cpp grammar-parse recovery ────────────────── - # llama.cpp's ``json-schema-to-grammar`` converter rejects - # regex escape classes (``\d``, ``\w``, ``\s``) and most - # ``format`` values in tool schemas. MCP servers emit - # these routinely for date/phone/email params. Recovery: - # strip ``pattern``/``format`` from ``self.tools`` and - # retry once. We keep the keywords by default so cloud - # providers get the full prompting hints; this branch - # fires only for users on llama.cpp's OAI server. - if ( - classified.reason == FailoverReason.llama_cpp_grammar_pattern - and not llama_cpp_grammar_retry_attempted - ): - llama_cpp_grammar_retry_attempted = True - try: - from tools.schema_sanitizer import strip_pattern_and_format - _, _stripped = strip_pattern_and_format(self.tools) - except Exception as _strip_exc: # pragma: no cover — defensive - logging.warning( - "%sllama.cpp grammar recovery: strip helper failed: %s", - self.log_prefix, _strip_exc, - ) - _stripped = 0 - if _stripped: - self._vprint( - f"{self.log_prefix}⚠️ llama.cpp rejected tool schema grammar — " - f"stripped {_stripped} pattern/format keyword(s), retrying...", - force=True, - ) - logging.warning( - "%sllama.cpp grammar recovery: stripped %d " - "pattern/format keyword(s) from tool schemas", - self.log_prefix, _stripped, - ) - continue - # No keywords found to strip — fall through to normal - # retry path rather than loop forever on the same error. - logging.warning( - "%sllama.cpp grammar error but no pattern/format " - "keywords to strip — falling through to normal retry", - self.log_prefix, - ) - - retry_count += 1 - elapsed_time = time.time() - api_start_time - self._touch_activity( - f"API error recovery (attempt {retry_count}/{max_retries})" - ) - - error_type = type(api_error).__name__ - error_msg = str(api_error).lower() - _error_summary = self._summarize_api_error(api_error) - logger.warning( - "API call failed (attempt %s/%s) error_type=%s %s summary=%s", - retry_count, - max_retries, - error_type, - self._client_log_context(), - _error_summary, - ) - - _provider = getattr(self, "provider", "unknown") - _base = getattr(self, "base_url", "unknown") - _model = getattr(self, "model", "unknown") - _status_code_str = f" [HTTP {status_code}]" if status_code else "" - self._vprint(f"{self.log_prefix}⚠️ API call failed (attempt {retry_count}/{max_retries}): {error_type}{_status_code_str}", force=True) - self._vprint(f"{self.log_prefix} 🔌 Provider: {_provider} Model: {_model}", force=True) - self._vprint(f"{self.log_prefix} 🌐 Endpoint: {_base}", force=True) - self._vprint(f"{self.log_prefix} 📝 Error: {_error_summary}", force=True) - if status_code and status_code < 500: - _err_body = getattr(api_error, "body", None) - _err_body_str = str(_err_body)[:300] if _err_body else None - if _err_body_str: - self._vprint(f"{self.log_prefix} 📋 Details: {_err_body_str}", force=True) - self._vprint(f"{self.log_prefix} ⏱️ Elapsed: {elapsed_time:.2f}s Context: {len(api_messages)} msgs, ~{approx_tokens:,} tokens") - - # Actionable hint for OpenRouter "no tool endpoints" error. - # This fires regardless of whether fallback succeeds — the - # user needs to know WHY their model failed so they can fix - # their provider routing, not just silently fall back. - if ( - self._is_openrouter_url() - and "support tool use" in error_msg - ): - self._vprint( - f"{self.log_prefix} 💡 No OpenRouter providers for {_model} support tool calling with your current settings.", - force=True, - ) - if self.providers_allowed: - self._vprint( - f"{self.log_prefix} Your provider_routing.only restriction is filtering out tool-capable providers.", - force=True, - ) - self._vprint( - f"{self.log_prefix} Try removing the restriction or adding providers that support tools for this model.", - force=True, - ) - self._vprint( - f"{self.log_prefix} Check which providers support tools: https://openrouter.ai/models/{_model}", - force=True, - ) - - # Check for interrupt before deciding to retry - if self._interrupt_requested: - self._vprint(f"{self.log_prefix}⚡ Interrupt detected during error handling, aborting retries.", force=True) - self._persist_session(messages, conversation_history) - self.clear_interrupt() - return { - "final_response": f"Operation interrupted: handling API error ({error_type}: {self._clean_error_message(str(api_error))}).", - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "interrupted": True, - } - - # Actionable hint for GitHub Models (Azure) 413 errors. - # The free tier enforces a hard 8K token cap per request, - # which Hermes' system prompt + tool schemas alone exceed. - # Compression can't help — the floor is the system prompt - # itself, not the conversation — so surface a clear "not - # compatible" message instead of looping into three futile - # compression attempts. - if ( - status_code == 413 - and isinstance(_base, str) - and "models.inference.ai.azure.com" in _base - ): - self._vprint( - f"{self.log_prefix} 💡 GitHub Models free tier (models.inference.ai.azure.com) caps every", - force=True, - ) - self._vprint( - f"{self.log_prefix} request at ~8K tokens. Hermes' system prompt + tool schemas baseline", - force=True, - ) - self._vprint( - f"{self.log_prefix} exceeds that floor, so this endpoint cannot run an agentic loop.", - force=True, - ) - self._vprint( - f"{self.log_prefix} Use the `copilot` provider with a Copilot subscription token (`hermes", - force=True, - ) - self._vprint( - f"{self.log_prefix} setup` → GitHub Copilot), or pick any other provider.", - force=True, - ) - - # Check for 413 payload-too-large BEFORE generic 4xx handler. - # A 413 is a payload-size error — the correct response is to - # compress history and retry, not abort immediately. - status_code = getattr(api_error, "status_code", None) - - # ── Anthropic Sonnet long-context tier gate ─────────── - # Anthropic returns HTTP 429 "Extra usage is required for - # long context requests" when a Claude Max (or similar) - # subscription doesn't include the 1M-context tier. This - # is NOT a transient rate limit — retrying or switching - # credentials won't help. Reduce context to 200k (the - # standard tier) and compress. - if classified.reason == FailoverReason.long_context_tier: - _reduced_ctx = 200000 - compressor = self.context_compressor - old_ctx = compressor.context_length - if old_ctx > _reduced_ctx: - compressor.update_model( - model=self.model, - context_length=_reduced_ctx, - base_url=self.base_url, - api_key=getattr(self, "api_key", ""), - provider=self.provider, - ) - # Context probing flags — only set on built-in - # compressor (plugin engines manage their own). - if hasattr(compressor, "_context_probed"): - compressor._context_probed = True - # Don't persist — this is a subscription-tier - # limitation, not a model capability. If the - # user later enables extra usage the 1M limit - # should come back automatically. - compressor._context_probe_persistable = False - self._vprint( - f"{self.log_prefix}⚠️ Anthropic long-context tier " - f"requires extra usage — reducing context: " - f"{old_ctx:,} → {_reduced_ctx:,} tokens", - force=True, - ) - - compression_attempts += 1 - if compression_attempts <= max_compression_attempts: - original_len = len(messages) - messages, active_system_prompt = self._compress_context( - messages, system_message, - approx_tokens=approx_tokens, - task_id=effective_task_id, - ) - # Compression created a new session — clear history - # so _flush_messages_to_session_db writes compressed - # messages to the new session, not skipping them. - conversation_history = None - if len(messages) < original_len or old_ctx > _reduced_ctx: - self._emit_status( - f"🗜️ Context reduced to {_reduced_ctx:,} tokens " - f"(was {old_ctx:,}), retrying..." - ) - time.sleep(2) - restart_with_compressed_messages = True - break - # Fall through to normal error handling if compression - # is exhausted or didn't help. - - # Eager fallback for rate-limit errors (429 or quota exhaustion). - # When a fallback model is configured, switch immediately instead - # of burning through retries with exponential backoff -- the - # primary provider won't recover within the retry window. - is_rate_limited = classified.reason in { - FailoverReason.rate_limit, - FailoverReason.billing, - } - if is_rate_limited and self._fallback_index < len(self._fallback_chain): - # Don't eagerly fallback if credential pool rotation may - # still recover. See _pool_may_recover_from_rate_limit - # for the single-credential-pool and CloudCode-quota - # exceptions. Fixes #11314 and #13636. - pool_may_recover = _pool_may_recover_from_rate_limit( - self._credential_pool, - provider=self.provider, - base_url=getattr(self, "base_url", None), - ) - if not pool_may_recover: - self._emit_status("⚠️ Rate limited — switching to fallback provider...") - if self._try_activate_fallback(reason=classified.reason): - retry_count = 0 - compression_attempts = 0 - primary_recovery_attempted = False - continue - - # ── Nous Portal: record rate limit & skip retries ───── - # When Nous returns a 429 that is a genuine account- - # level rate limit, record the reset time to a shared - # file so ALL sessions (cron, gateway, auxiliary) know - # not to pile on, then skip further retries -- each - # one burns another RPH request and deepens the hole. - # The retry loop's top-of-iteration guard will catch - # this on the next pass and try fallback or bail. - # - # IMPORTANT: Nous Portal multiplexes multiple upstream - # providers (DeepSeek, Kimi, MiMo, Hermes). A 429 can - # also mean an UPSTREAM provider is out of capacity - # for one specific model -- transient, clears in - # seconds, nothing to do with the caller's quota. - # Tripping the cross-session breaker on that would - # block every Nous model for minutes. We use - # ``is_genuine_nous_rate_limit`` to tell the two - # apart via the 429's own x-ratelimit-* headers and - # the last-known-good state captured on the previous - # successful response. - if ( - is_rate_limited - and self.provider == "nous" - and classified.reason == FailoverReason.rate_limit - and not recovered_with_pool - ): - _genuine_nous_rate_limit = False - try: - from agent.nous_rate_guard import ( - is_genuine_nous_rate_limit, - record_nous_rate_limit, - ) - _err_resp = getattr(api_error, "response", None) - _err_hdrs = ( - getattr(_err_resp, "headers", None) - if _err_resp else None - ) - _genuine_nous_rate_limit = is_genuine_nous_rate_limit( - headers=_err_hdrs, - last_known_state=self._rate_limit_state, - ) - if _genuine_nous_rate_limit: - record_nous_rate_limit( - headers=_err_hdrs, - error_context=error_context, - ) - else: - logging.info( - "Nous 429 looks like upstream capacity " - "(no exhausted bucket in headers or " - "last-known state) -- not tripping " - "cross-session breaker." - ) - except Exception: - pass - if _genuine_nous_rate_limit: - # Skip straight to max_retries -- the - # top-of-loop guard will handle fallback or - # bail cleanly. - retry_count = max_retries - continue - # Upstream capacity 429: fall through to normal - # retry logic. A different model (or the same - # model a moment later) will typically succeed. - - is_payload_too_large = ( - classified.reason == FailoverReason.payload_too_large - ) - - if is_payload_too_large: - compression_attempts += 1 - if compression_attempts > max_compression_attempts: - self._vprint(f"{self.log_prefix}❌ Max compression attempts ({max_compression_attempts}) reached for payload-too-large error.", force=True) - self._vprint(f"{self.log_prefix} 💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True) - logging.error(f"{self.log_prefix}413 compression failed after {max_compression_attempts} attempts.") - self._persist_session(messages, conversation_history) - return { - "messages": messages, - "completed": False, - "api_calls": api_call_count, - "error": f"Request payload too large: max compression attempts ({max_compression_attempts}) reached.", - "partial": True, - "failed": True, - "compression_exhausted": True, - } - self._emit_status(f"⚠️ Request payload too large (413) — compression attempt {compression_attempts}/{max_compression_attempts}...") - - original_len = len(messages) - messages, active_system_prompt = self._compress_context( - messages, system_message, approx_tokens=approx_tokens, - task_id=effective_task_id, - ) - # Compression created a new session — clear history - # so _flush_messages_to_session_db writes compressed - # messages to the new session, not skipping them. - conversation_history = None - - if len(messages) < original_len: - self._emit_status(f"🗜️ Compressed {original_len} → {len(messages)} messages, retrying...") - time.sleep(2) # Brief pause between compression retries - restart_with_compressed_messages = True - break - else: - self._vprint(f"{self.log_prefix}❌ Payload too large and cannot compress further.", force=True) - self._vprint(f"{self.log_prefix} 💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True) - logging.error(f"{self.log_prefix}413 payload too large. Cannot compress further.") - self._persist_session(messages, conversation_history) - return { - "messages": messages, - "completed": False, - "api_calls": api_call_count, - "error": "Request payload too large (413). Cannot compress further.", - "partial": True, - "failed": True, - "compression_exhausted": True, - } - - # Check for context-length errors BEFORE generic 4xx handler. - # The classifier detects context overflow from: explicit error - # messages, generic 400 + large session heuristic (#1630), and - # server disconnect + large session pattern (#2153). - is_context_length_error = ( - classified.reason == FailoverReason.context_overflow - ) - - if is_context_length_error: - compressor = self.context_compressor - old_ctx = compressor.context_length - - # ── Distinguish two very different errors ─────────── - # 1. "Prompt too long": the INPUT exceeds the context window. - # Fix: reduce context_length + compress history. - # 2. "max_tokens too large": input is fine, but - # input_tokens + requested max_tokens > context_window. - # Fix: reduce max_tokens (the OUTPUT cap) for this call. - # Do NOT shrink context_length — the window is unchanged. - # - # Note: max_tokens = output token cap (one response). - # context_length = total window (input + output combined). - available_out = parse_available_output_tokens_from_error(error_msg) - if available_out is not None: - # Error is purely about the output cap being too large. - # Cap output to the available space and retry without - # touching context_length or triggering compression. - safe_out = max(1, available_out - 64) # small safety margin - self._ephemeral_max_output_tokens = safe_out - self._vprint( - f"{self.log_prefix}⚠️ Output cap too large for current prompt — " - f"retrying with max_tokens={safe_out:,} " - f"(available_tokens={available_out:,}; context_length unchanged at {old_ctx:,})", - force=True, - ) - # Still count against compression_attempts so we don't - # loop forever if the error keeps recurring. - compression_attempts += 1 - if compression_attempts > max_compression_attempts: - self._vprint(f"{self.log_prefix}❌ Max compression attempts ({max_compression_attempts}) reached.", force=True) - self._vprint(f"{self.log_prefix} 💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True) - logging.error(f"{self.log_prefix}Context compression failed after {max_compression_attempts} attempts.") - self._persist_session(messages, conversation_history) - return { - "messages": messages, - "completed": False, - "api_calls": api_call_count, - "error": f"Context length exceeded: max compression attempts ({max_compression_attempts}) reached.", - "partial": True, - "failed": True, - "compression_exhausted": True, - } - restart_with_compressed_messages = True - break - - # Error is about the INPUT being too large — reduce context_length. - # Try to parse the actual limit from the error message - parsed_limit = parse_context_limit_from_error(error_msg) - _provider_lower = (getattr(self, "provider", "") or "").lower() - _base_lower = (getattr(self, "base_url", "") or "").rstrip("/").lower() - is_minimax_provider = ( - _provider_lower in {"minimax", "minimax-cn"} - or _base_lower.startswith(( - "https://api.minimax.io/anthropic", - "https://api.minimaxi.com/anthropic", - )) - ) - minimax_delta_only_overflow = ( - is_minimax_provider - and parsed_limit is None - and "context window exceeds limit (" in error_msg - ) - if parsed_limit and parsed_limit < old_ctx: - new_ctx = parsed_limit - self._vprint(f"{self.log_prefix}Context limit detected from API: {new_ctx:,} tokens (was {old_ctx:,})", force=True) - elif minimax_delta_only_overflow: - new_ctx = old_ctx - self._vprint( - f"{self.log_prefix}Provider reported overflow amount only; " - f"keeping context_length at {old_ctx:,} tokens and compressing.", - force=True, - ) - else: - # Step down to the next probe tier - new_ctx = get_next_probe_tier(old_ctx) - - if new_ctx and new_ctx < old_ctx: - compressor.update_model( - model=self.model, - context_length=new_ctx, - base_url=self.base_url, - api_key=getattr(self, "api_key", ""), - provider=self.provider, - ) - # Context probing flags — only set on built-in - # compressor (plugin engines manage their own). - if hasattr(compressor, "_context_probed"): - compressor._context_probed = True - # Only persist limits parsed from the provider's - # error message (a real number). Guessed fallback - # tiers from get_next_probe_tier() should stay - # in-memory only — persisting them pollutes the - # cache with wrong values. - compressor._context_probe_persistable = bool( - parsed_limit and parsed_limit == new_ctx - ) - self._vprint(f"{self.log_prefix}⚠️ Context length exceeded — stepping down: {old_ctx:,} → {new_ctx:,} tokens", force=True) - else: - self._vprint(f"{self.log_prefix}⚠️ Context length exceeded at minimum tier — attempting compression...", force=True) - - compression_attempts += 1 - if compression_attempts > max_compression_attempts: - self._vprint(f"{self.log_prefix}❌ Max compression attempts ({max_compression_attempts}) reached.", force=True) - self._vprint(f"{self.log_prefix} 💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True) - logging.error(f"{self.log_prefix}Context compression failed after {max_compression_attempts} attempts.") - self._persist_session(messages, conversation_history) - return { - "messages": messages, - "completed": False, - "api_calls": api_call_count, - "error": f"Context length exceeded: max compression attempts ({max_compression_attempts}) reached.", - "partial": True, - "failed": True, - "compression_exhausted": True, - } - self._emit_status(f"🗜️ Context too large (~{approx_tokens:,} tokens) — compressing ({compression_attempts}/{max_compression_attempts})...") - - original_len = len(messages) - messages, active_system_prompt = self._compress_context( - messages, system_message, approx_tokens=approx_tokens, - task_id=effective_task_id, - ) - # Compression created a new session — clear history - # so _flush_messages_to_session_db writes compressed - # messages to the new session, not skipping them. - conversation_history = None - - if len(messages) < original_len or new_ctx and new_ctx < old_ctx: - if len(messages) < original_len: - self._emit_status(f"🗜️ Compressed {original_len} → {len(messages)} messages, retrying...") - time.sleep(2) # Brief pause between compression retries - restart_with_compressed_messages = True - break - else: - # Can't compress further and already at minimum tier - self._vprint(f"{self.log_prefix}❌ Context length exceeded and cannot compress further.", force=True) - self._vprint(f"{self.log_prefix} 💡 The conversation has accumulated too much content. Try /new to start fresh, or /compress to manually trigger compression.", force=True) - logging.error(f"{self.log_prefix}Context length exceeded: {approx_tokens:,} tokens. Cannot compress further.") - self._persist_session(messages, conversation_history) - return { - "messages": messages, - "completed": False, - "api_calls": api_call_count, - "error": f"Context length exceeded ({approx_tokens:,} tokens). Cannot compress further.", - "partial": True, - "failed": True, - "compression_exhausted": True, - } - - # Check for non-retryable client errors. The classifier - # already accounts for 413, 429, 529 (transient), context - # overflow, and generic-400 heuristics. Local validation - # errors (ValueError, TypeError) are programming bugs. - # Exclude UnicodeEncodeError — it's a ValueError subclass - # but is handled separately by the surrogate sanitization - # path above. Exclude json.JSONDecodeError — also a - # ValueError subclass, but it indicates a transient - # provider/network failure (malformed response body, - # truncated stream, routing layer corruption), not a - # local programming bug, and should be retried (#14782). - # Exclude Anthropic stream parser ValueErrors for the - # same reason: third-party Anthropic-compatible providers - # can emit malformed event-stream frames that SDK parsers - # raise as plain ValueError. - is_local_validation_error = ( - isinstance(api_error, (ValueError, TypeError)) - and not isinstance( - api_error, (UnicodeEncodeError, json.JSONDecodeError) - ) - and not self._is_provider_stream_parse_error(api_error) - # ssl.SSLError (and its subclass SSLCertVerificationError) - # inherits from OSError *and* ValueError via Python MRO, - # so the isinstance(ValueError) check above would - # misclassify a TLS transport failure as a local - # programming bug and abort without retrying. Exclude - # ssl.SSLError explicitly so the error classifier's - # retryable=True mapping takes effect instead. - and not isinstance(api_error, ssl.SSLError) - ) - is_client_error = ( - is_local_validation_error - or ( - not classified.retryable - and not classified.should_compress - and classified.reason not in { - FailoverReason.rate_limit, - FailoverReason.billing, - FailoverReason.overloaded, - FailoverReason.context_overflow, - FailoverReason.payload_too_large, - FailoverReason.long_context_tier, - FailoverReason.thinking_signature, - } - ) - ) and not is_context_length_error - - if is_client_error: - # Try fallback before aborting — a different provider - # may not have the same issue (rate limit, auth, etc.) - self._emit_status(f"⚠️ Non-retryable error (HTTP {status_code}) — trying fallback...") - if self._try_activate_fallback(): - retry_count = 0 - compression_attempts = 0 - primary_recovery_attempted = False - continue - if api_kwargs is not None: - self._dump_api_request_debug( - api_kwargs, reason="non_retryable_client_error", error=api_error, - ) - self._emit_status( - f"❌ Non-retryable error (HTTP {status_code}): " - f"{self._summarize_api_error(api_error)}" - ) - self._vprint(f"{self.log_prefix}❌ Non-retryable client error (HTTP {status_code}). Aborting.", force=True) - self._vprint(f"{self.log_prefix} 🔌 Provider: {_provider} Model: {_model}", force=True) - self._vprint(f"{self.log_prefix} 🌐 Endpoint: {_base}", force=True) - # Actionable guidance for common auth errors - if classified.is_auth or classified.reason == FailoverReason.billing: - if _provider in {"openai-codex", "xai-oauth"} and status_code == 401: - if _provider == "openai-codex": - self._vprint(f"{self.log_prefix} 💡 Codex OAuth token was rejected (HTTP 401). Your token may have been", force=True) - self._vprint(f"{self.log_prefix} refreshed by another client (Codex CLI, VS Code). To fix:", force=True) - self._vprint(f"{self.log_prefix} 1. Run `codex` in your terminal to generate fresh tokens.", force=True) - self._vprint(f"{self.log_prefix} 2. Then run `hermes auth` to re-authenticate.", force=True) - else: - self._vprint(f"{self.log_prefix} 💡 xAI OAuth token was rejected (HTTP 401). To fix:", force=True) - self._vprint(f"{self.log_prefix} re-authenticate with xAI Grok OAuth (SuperGrok Subscription) from `hermes model`.", force=True) - else: - self._vprint(f"{self.log_prefix} 💡 Your API key was rejected by the provider. Check:", force=True) - self._vprint(f"{self.log_prefix} • Is the key valid? Run: hermes setup", force=True) - self._vprint(f"{self.log_prefix} • Does your account have access to {_model}?", force=True) - if base_url_host_matches(str(_base), "openrouter.ai"): - self._vprint(f"{self.log_prefix} • Check credits: https://openrouter.ai/settings/credits", force=True) - else: - self._vprint(f"{self.log_prefix} 💡 This type of error won't be fixed by retrying.", force=True) - logging.error(f"{self.log_prefix}Non-retryable client error: {api_error}") - # Skip session persistence when the error is likely - # context-overflow related (status 400 + large session). - # Persisting the failed user message would make the - # session even larger, causing the same failure on the - # next attempt. (#1630) - if status_code == 400 and (approx_tokens > 50000 or len(api_messages) > 80): - self._vprint( - f"{self.log_prefix}⚠️ Skipping session persistence " - f"for large failed session to prevent growth loop.", - force=True, - ) - else: - self._persist_session(messages, conversation_history) - return { - "final_response": None, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "failed": True, - "error": str(api_error), - } - - if retry_count >= max_retries: - # Before falling back, try rebuilding the primary - # client once for transient transport errors (stale - # connection pool, TCP reset). Only attempted once - # per API call block. - if not primary_recovery_attempted and self._try_recover_primary_transport( - api_error, retry_count=retry_count, max_retries=max_retries, - ): - primary_recovery_attempted = True - retry_count = 0 - continue - # Try fallback before giving up entirely - self._emit_status(f"⚠️ Max retries ({max_retries}) exhausted — trying fallback...") - if self._try_activate_fallback(): - retry_count = 0 - compression_attempts = 0 - primary_recovery_attempted = False - continue - _final_summary = self._summarize_api_error(api_error) - if is_rate_limited: - self._emit_status(f"❌ Rate limited after {max_retries} retries — {_final_summary}") - else: - self._emit_status(f"❌ API failed after {max_retries} retries — {_final_summary}") - self._vprint(f"{self.log_prefix} 💀 Final error: {_final_summary}", force=True) - - # Detect SSE stream-drop pattern (e.g. "Network - # connection lost") and surface actionable guidance. - # This typically happens when the model generates a - # very large tool call (write_file with huge content) - # and the proxy/CDN drops the stream mid-response. - _is_stream_drop = ( - not getattr(api_error, "status_code", None) - and any(p in error_msg for p in ( - "connection lost", "connection reset", - "connection closed", "network connection", - "network error", "terminated", - )) - ) - if _is_stream_drop: - self._vprint( - f"{self.log_prefix} 💡 The provider's stream " - f"connection keeps dropping. This often happens " - f"when the model tries to write a very large " - f"file in a single tool call.", - force=True, - ) - self._vprint( - f"{self.log_prefix} Try asking the model " - f"to use execute_code with Python's open() for " - f"large files, or to write the file in smaller " - f"sections.", - force=True, - ) - - logging.error( - "%sAPI call failed after %s retries. %s | provider=%s model=%s msgs=%s tokens=~%s", - self.log_prefix, max_retries, _final_summary, - _provider, _model, len(api_messages), f"{approx_tokens:,}", - ) - if api_kwargs is not None: - self._dump_api_request_debug( - api_kwargs, reason="max_retries_exhausted", error=api_error, - ) - self._persist_session(messages, conversation_history) - _final_response = f"API call failed after {max_retries} retries: {_final_summary}" - if _is_stream_drop: - _final_response += ( - "\n\nThe provider's stream connection keeps " - "dropping — this often happens when generating " - "very large tool call responses (e.g. write_file " - "with long content). Try asking me to use " - "execute_code with Python's open() for large " - "files, or to write in smaller sections." - ) - return { - "final_response": _final_response, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "failed": True, - "error": _final_summary, - } - - # For rate limits, respect the Retry-After header if present - _retry_after = None - if is_rate_limited: - _resp_headers = getattr(getattr(api_error, "response", None), "headers", None) - if _resp_headers and hasattr(_resp_headers, "get"): - _ra_raw = _resp_headers.get("retry-after") or _resp_headers.get("Retry-After") - if _ra_raw: - try: - _retry_after = min(float(_ra_raw), 120) # Cap at 2 minutes - except (TypeError, ValueError): - pass - wait_time = _retry_after if _retry_after else jittered_backoff(retry_count, base_delay=2.0, max_delay=60.0) - if is_rate_limited: - self._emit_status(f"⏱️ Rate limited. Waiting {wait_time:.1f}s (attempt {retry_count + 1}/{max_retries})...") - else: - self._emit_status(f"⏳ Retrying in {wait_time:.1f}s (attempt {retry_count}/{max_retries})...") - logger.warning( - "Retrying API call in %ss (attempt %s/%s) %s error=%s", - wait_time, - retry_count, - max_retries, - self._client_log_context(), - api_error, - ) - # Sleep in small increments so we can respond to interrupts quickly - # instead of blocking the entire wait_time in one sleep() call - sleep_end = time.time() + wait_time - _backoff_touch_counter = 0 - while time.time() < sleep_end: - if self._interrupt_requested: - self._vprint(f"{self.log_prefix}⚡ Interrupt detected during retry wait, aborting.", force=True) - self._persist_session(messages, conversation_history) - self.clear_interrupt() - return { - "final_response": f"Operation interrupted: retrying API call after error (retry {retry_count}/{max_retries}).", - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "interrupted": True, - } - time.sleep(0.2) # Check interrupt every 200ms - # Touch activity every ~30s so the gateway's inactivity - # monitor knows we're alive during backoff waits. - _backoff_touch_counter += 1 - if _backoff_touch_counter % 150 == 0: # 150 × 0.2s = 30s - self._touch_activity( - f"error retry backoff ({retry_count}/{max_retries}), " - f"{int(sleep_end - time.time())}s remaining" - ) - - # If the API call was interrupted, skip response processing - if interrupted: - _turn_exit_reason = "interrupted_during_api_call" - break - - if restart_with_compressed_messages: - api_call_count -= 1 - self.iteration_budget.refund() - # Count compression restarts toward the retry limit to prevent - # infinite loops when compression reduces messages but not enough - # to fit the context window. - retry_count += 1 - restart_with_compressed_messages = False - continue - - if restart_with_length_continuation: - # Progressively boost the output token budget on each retry. - # Retry 1 → 2× base, retry 2 → 3× base, capped at 32 768. - # Applies to all providers via _ephemeral_max_output_tokens. - _boost_base = self.max_tokens if self.max_tokens else 4096 - _boost = _boost_base * (length_continue_retries + 1) - self._ephemeral_max_output_tokens = min(_boost, 32768) - continue - - # Guard: if all retries exhausted without a successful response - # (e.g. repeated context-length errors that exhausted retry_count), - # the `response` variable is still None. Break out cleanly. - if response is None: - _turn_exit_reason = "all_retries_exhausted_no_response" - print(f"{self.log_prefix}❌ All API retries exhausted with no successful response.") - self._persist_session(messages, conversation_history) - break - - try: - _transport = self._get_transport() - _normalize_kwargs = {} - if self.api_mode == "anthropic_messages": - _normalize_kwargs["strip_tool_prefix"] = self._is_anthropic_oauth - normalized = _transport.normalize_response(response, **_normalize_kwargs) - assistant_message = normalized - finish_reason = normalized.finish_reason - - # Normalize content to string — some OpenAI-compatible servers - # (llama-server, etc.) return content as a dict or list instead - # of a plain string, which crashes downstream .strip() calls. - if assistant_message.content is not None and not isinstance(assistant_message.content, str): - raw = assistant_message.content - if isinstance(raw, dict): - assistant_message.content = raw.get("text", "") or raw.get("content", "") or json.dumps(raw) - elif isinstance(raw, list): - # Multimodal content list — extract text parts - parts = [] - for part in raw: - if isinstance(part, str): - parts.append(part) - elif isinstance(part, dict) and part.get("type") == "text": - parts.append(part.get("text", "")) - elif isinstance(part, dict) and "text" in part: - parts.append(str(part["text"])) - assistant_message.content = "\n".join(parts) - else: - assistant_message.content = str(raw) - - try: - from hermes_cli.plugins import invoke_hook as _invoke_hook - _assistant_tool_calls = getattr(assistant_message, "tool_calls", None) or [] - _assistant_text = assistant_message.content or "" - _invoke_hook( - "post_api_request", - task_id=effective_task_id, - session_id=self.session_id or "", - platform=self.platform or "", - model=self.model, - provider=self.provider, - base_url=self.base_url, - api_mode=self.api_mode, - api_call_count=api_call_count, - api_duration=api_duration, - finish_reason=finish_reason, - message_count=len(api_messages), - response_model=getattr(response, "model", None), - response=response, - usage=self._usage_summary_for_api_request_hook(response), - assistant_message=assistant_message, - assistant_content_chars=len(_assistant_text), - assistant_tool_call_count=len(_assistant_tool_calls), - ) - except Exception: - pass - - # Handle assistant response - if assistant_message.content and not self.quiet_mode: - if self.verbose_logging: - self._vprint(f"{self.log_prefix}🤖 Assistant: {assistant_message.content}") - else: - self._vprint(f"{self.log_prefix}🤖 Assistant: {assistant_message.content[:100]}{'...' if len(assistant_message.content) > 100 else ''}") - - # Notify progress callback of model's thinking (used by subagent - # delegation to relay the child's reasoning to the parent display). - if (assistant_message.content and self.tool_progress_callback): - _think_text = assistant_message.content.strip() - # Strip reasoning XML tags that shouldn't leak to parent display - _think_text = re.sub( - r'', '', _think_text - ).strip() - # For subagents: relay first line to parent display (existing behaviour). - # For all agents with a structured callback: emit reasoning.available event. - first_line = _think_text.split('\n')[0][:80] if _think_text else "" - if first_line and getattr(self, '_delegate_depth', 0) > 0: - try: - self.tool_progress_callback("_thinking", first_line) - except Exception: - pass - elif _think_text: - try: - self.tool_progress_callback("reasoning.available", "_thinking", _think_text[:500], None) - except Exception: - pass - - # Check for incomplete (opened but never closed) - # This means the model ran out of output tokens mid-reasoning — retry up to 2 times - if has_incomplete_scratchpad(assistant_message.content or ""): - self._incomplete_scratchpad_retries += 1 - - self._vprint(f"{self.log_prefix}⚠️ Incomplete detected (opened but never closed)") - - if self._incomplete_scratchpad_retries <= 2: - self._vprint(f"{self.log_prefix}🔄 Retrying API call ({self._incomplete_scratchpad_retries}/2)...") - # Don't add the broken message, just retry - continue - else: - # Max retries - discard this turn and save as partial - self._vprint(f"{self.log_prefix}❌ Max retries (2) for incomplete scratchpad. Saving as partial.", force=True) - self._incomplete_scratchpad_retries = 0 - - rolled_back_messages = self._get_messages_up_to_last_assistant(messages) - self._cleanup_task_resources(effective_task_id) - self._persist_session(messages, conversation_history) - - return { - "final_response": None, - "messages": rolled_back_messages, - "api_calls": api_call_count, - "completed": False, - "partial": True, - "error": "Incomplete REASONING_SCRATCHPAD after 2 retries" - } - - # Reset incomplete scratchpad counter on clean response - self._incomplete_scratchpad_retries = 0 - - if self.api_mode == "codex_responses" and finish_reason == "incomplete": - self._codex_incomplete_retries += 1 - - interim_msg = self._build_assistant_message(assistant_message, finish_reason) - interim_has_content = bool((interim_msg.get("content") or "").strip()) - interim_has_reasoning = bool(interim_msg.get("reasoning", "").strip()) if isinstance(interim_msg.get("reasoning"), str) else False - interim_has_codex_reasoning = bool(interim_msg.get("codex_reasoning_items")) - interim_has_codex_message_items = bool(interim_msg.get("codex_message_items")) - - if ( - interim_has_content - or interim_has_reasoning - or interim_has_codex_reasoning - or interim_has_codex_message_items - ): - last_msg = messages[-1] if messages else None - # Duplicate detection: two consecutive incomplete assistant - # messages with identical content AND reasoning are collapsed. - # For provider-state-only changes (encrypted reasoning - # items or replayable message ids/phases/statuses differ - # while visible content/reasoning are unchanged), compare - # those opaque payloads too so we don't silently drop the - # newer continuation state. - last_codex_items = last_msg.get("codex_reasoning_items") if isinstance(last_msg, dict) else None - interim_codex_items = interim_msg.get("codex_reasoning_items") - last_codex_message_items = last_msg.get("codex_message_items") if isinstance(last_msg, dict) else None - interim_codex_message_items = interim_msg.get("codex_message_items") - duplicate_interim = ( - isinstance(last_msg, dict) - and last_msg.get("role") == "assistant" - and last_msg.get("finish_reason") == "incomplete" - and (last_msg.get("content") or "") == (interim_msg.get("content") or "") - and (last_msg.get("reasoning") or "") == (interim_msg.get("reasoning") or "") - and last_codex_items == interim_codex_items - and last_codex_message_items == interim_codex_message_items - ) - if not duplicate_interim: - messages.append(interim_msg) - self._emit_interim_assistant_message(interim_msg) - - if self._codex_incomplete_retries < 3: - if not self.quiet_mode: - self._vprint(f"{self.log_prefix}↻ Codex response incomplete; continuing turn ({self._codex_incomplete_retries}/3)") - self._session_messages = messages - self._save_session_log(messages) - continue - - self._codex_incomplete_retries = 0 - self._persist_session(messages, conversation_history) - return { - "final_response": None, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "partial": True, - "error": "Codex response remained incomplete after 3 continuation attempts", - } - elif hasattr(self, "_codex_incomplete_retries"): - self._codex_incomplete_retries = 0 - - # Check for tool calls - if assistant_message.tool_calls: - if not self.quiet_mode: - self._vprint(f"{self.log_prefix}🔧 Processing {len(assistant_message.tool_calls)} tool call(s)...") - - if self.verbose_logging: - for tc in assistant_message.tool_calls: - logging.debug(f"Tool call: {tc.function.name} with args: {tc.function.arguments[:200]}...") - - # Validate tool call names - detect model hallucinations - # Repair mismatched tool names before validating - for tc in assistant_message.tool_calls: - if tc.function.name not in self.valid_tool_names: - repaired = self._repair_tool_call(tc.function.name) - if repaired: - print(f"{self.log_prefix}🔧 Auto-repaired tool name: '{tc.function.name}' -> '{repaired}'") - tc.function.name = repaired - invalid_tool_calls = [ - tc.function.name for tc in assistant_message.tool_calls - if tc.function.name not in self.valid_tool_names - ] - if invalid_tool_calls: - # Track retries for invalid tool calls - self._invalid_tool_retries += 1 - - # Return helpful error to model — model can self-correct next turn - available = ", ".join(sorted(self.valid_tool_names)) - invalid_name = invalid_tool_calls[0] - invalid_preview = invalid_name[:80] + "..." if len(invalid_name) > 80 else invalid_name - self._vprint(f"{self.log_prefix}⚠️ Unknown tool '{invalid_preview}' — sending error to model for self-correction ({self._invalid_tool_retries}/3)") - - if self._invalid_tool_retries >= 3: - self._vprint(f"{self.log_prefix}❌ Max retries (3) for invalid tool calls exceeded. Stopping as partial.", force=True) - self._invalid_tool_retries = 0 - self._persist_session(messages, conversation_history) - return { - "final_response": None, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "partial": True, - "error": f"Model generated invalid tool call: {invalid_preview}" - } - - assistant_msg = self._build_assistant_message(assistant_message, finish_reason) - messages.append(assistant_msg) - for tc in assistant_message.tool_calls: - if tc.function.name not in self.valid_tool_names: - content = f"Tool '{tc.function.name}' does not exist. Available tools: {available}" - else: - content = "Skipped: another tool call in this turn used an invalid name. Please retry this tool call." - messages.append({ - "role": "tool", - "name": tc.function.name, - "tool_call_id": tc.id, - "content": content, - }) - continue - # Reset retry counter on successful tool call validation - self._invalid_tool_retries = 0 - - # Validate tool call arguments are valid JSON - # Handle empty strings as empty objects (common model quirk) - invalid_json_args = [] - for tc in assistant_message.tool_calls: - args = tc.function.arguments - if isinstance(args, (dict, list)): - tc.function.arguments = json.dumps(args) - continue - if args is not None and not isinstance(args, str): - tc.function.arguments = str(args) - args = tc.function.arguments - # Treat empty/whitespace strings as empty object - if not args or not args.strip(): - tc.function.arguments = "{}" - continue - try: - json.loads(args) - except json.JSONDecodeError as e: - invalid_json_args.append((tc.function.name, str(e))) - - if invalid_json_args: - # Check if the invalid JSON is due to truncation rather - # than a model formatting mistake. Routers sometimes - # rewrite finish_reason from "length" to "tool_calls", - # hiding the truncation from the length handler above. - # Detect truncation: args that don't end with } or ] - # (after stripping whitespace) are cut off mid-stream. - _truncated = any( - not (tc.function.arguments or "").rstrip().endswith(("}", "]")) - for tc in assistant_message.tool_calls - if tc.function.name in {n for n, _ in invalid_json_args} - ) - if _truncated: - self._vprint( - f"{self.log_prefix}⚠️ Truncated tool call arguments detected " - f"(finish_reason={finish_reason!r}) — refusing to execute.", - force=True, - ) - self._invalid_json_retries = 0 - self._cleanup_task_resources(effective_task_id) - self._persist_session(messages, conversation_history) - return { - "final_response": None, - "messages": messages, - "api_calls": api_call_count, - "completed": False, - "partial": True, - "error": "Response truncated due to output length limit", - } - - # Track retries for invalid JSON arguments - self._invalid_json_retries += 1 - - tool_name, error_msg = invalid_json_args[0] - self._vprint(f"{self.log_prefix}⚠️ Invalid JSON in tool call arguments for '{tool_name}': {error_msg}") - - if self._invalid_json_retries < 3: - self._vprint(f"{self.log_prefix}🔄 Retrying API call ({self._invalid_json_retries}/3)...") - # Don't add anything to messages, just retry the API call - continue - else: - # Instead of returning partial, inject tool error results so the model can recover. - # Using tool results (not user messages) preserves role alternation. - self._vprint(f"{self.log_prefix}⚠️ Injecting recovery tool results for invalid JSON...") - self._invalid_json_retries = 0 # Reset for next attempt - - # Append the assistant message with its (broken) tool_calls - recovery_assistant = self._build_assistant_message(assistant_message, finish_reason) - messages.append(recovery_assistant) - - # Respond with tool error results for each tool call - invalid_names = {name for name, _ in invalid_json_args} - for tc in assistant_message.tool_calls: - if tc.function.name in invalid_names: - err = next(e for n, e in invalid_json_args if n == tc.function.name) - tool_result = ( - f"Error: Invalid JSON arguments. {err}. " - f"For tools with no required parameters, use an empty object: {{}}. " - f"Please retry with valid JSON." - ) - else: - tool_result = "Skipped: other tool call in this response had invalid JSON." - messages.append({ - "role": "tool", - "name": tc.function.name, - "tool_call_id": tc.id, - "content": tool_result, - }) - continue - - # Reset retry counter on successful JSON validation - self._invalid_json_retries = 0 - - # ── Post-call guardrails ────────────────────────── - assistant_message.tool_calls = self._cap_delegate_task_calls( - assistant_message.tool_calls - ) - assistant_message.tool_calls = self._deduplicate_tool_calls( - assistant_message.tool_calls - ) - - assistant_msg = self._build_assistant_message(assistant_message, finish_reason) - - # If this turn has both content AND tool_calls, capture the content - # as a fallback final response. Common pattern: model delivers its - # answer and calls memory/skill tools as a side-effect in the same - # turn. If the follow-up turn after tools is empty, we use this. - turn_content = assistant_message.content or "" - if turn_content and self._has_content_after_think_block(turn_content): - self._last_content_with_tools = turn_content - # Only mute subsequent output when EVERY tool call in - # this turn is post-response housekeeping (memory, todo, - # skill_manage, etc.). If any substantive tool is present - # (search_files, read_file, write_file, terminal, ...), - # keep output visible so the user sees progress. - _HOUSEKEEPING_TOOLS = frozenset({ - "memory", "todo", "skill_manage", "session_search", - }) - _all_housekeeping = all( - tc.function.name in _HOUSEKEEPING_TOOLS - for tc in assistant_message.tool_calls - ) - self._last_content_tools_all_housekeeping = _all_housekeeping - if _all_housekeeping and self._has_stream_consumers(): - self._mute_post_response = True - elif self._should_emit_quiet_tool_messages(): - clean = self._strip_think_blocks(turn_content).strip() - if clean: - self._vprint(f" ┊ 💬 {clean}") - - # Pop thinking-only prefill message(s) before appending - # (tool-call path — same rationale as the final-response path). - _had_prefill = False - while ( - messages - and isinstance(messages[-1], dict) - and messages[-1].get("_thinking_prefill") - ): - messages.pop() - _had_prefill = True - - # Reset prefill counter when tool calls follow a prefill - # recovery. Without this, the counter accumulates across - # the whole conversation — a model that intermittently - # empties (empty → prefill → tools → empty → prefill → - # tools) burns both prefill attempts and the third empty - # gets zero recovery. Resetting here treats each tool- - # call success as a fresh start. - if _had_prefill: - self._thinking_prefill_retries = 0 - self._empty_content_retries = 0 - # Successful tool execution — reset the post-tool nudge - # flag so it can fire again if the model goes empty on - # a LATER tool round. - self._post_tool_empty_retried = False - - messages.append(assistant_msg) - self._emit_interim_assistant_message(assistant_msg) - - # Close any open streaming display (response box, reasoning - # box) before tool execution begins. Intermediate turns may - # have streamed early content that opened the response box; - # flushing here prevents it from wrapping tool feed lines. - # Only signal the display callback — TTS (_stream_callback) - # should NOT receive None (it uses None as end-of-stream). - if self.stream_delta_callback: - try: - self.stream_delta_callback(None) - except Exception: - pass - - self._execute_tool_calls(assistant_message, messages, effective_task_id, api_call_count) - - if self._tool_guardrail_halt_decision is not None: - decision = self._tool_guardrail_halt_decision - _turn_exit_reason = "guardrail_halt" - final_response = self._toolguard_controlled_halt_response(decision) - self._emit_status( - f"⚠️ Tool guardrail halted {decision.tool_name}: {decision.code}" - ) - messages.append({"role": "assistant", "content": final_response}) - break - - # Reset per-turn retry counters after successful tool - # execution so a single truncation doesn't poison the - # entire conversation. - truncated_tool_call_retries = 0 - - # Signal that a paragraph break is needed before the next - # streamed text. We don't emit it immediately because - # multiple consecutive tool iterations would stack up - # redundant blank lines. Instead, _fire_stream_delta() - # will prepend a single "\n\n" the next time real text - # arrives. - self._stream_needs_break = True - - # Refund the iteration if the ONLY tool(s) called were - # execute_code (programmatic tool calling). These are - # cheap RPC-style calls that shouldn't eat the budget. - _tc_names = {tc.function.name for tc in assistant_message.tool_calls} - if _tc_names == {"execute_code"}: - self.iteration_budget.refund() - - # Use real token counts from the API response to decide - # compression. prompt_tokens + completion_tokens is the - # actual context size the provider reported plus the - # assistant turn — a tight lower bound for the next prompt. - # Tool results appended above aren't counted yet, but the - # threshold (default 50%) leaves ample headroom; if tool - # results push past it, the next API call will report the - # real total and trigger compression then. - # - # If last_prompt_tokens is 0 (stale after API disconnect - # or provider returned no usage data), fall back to rough - # estimate to avoid missing compression. Without this, - # a session can grow unbounded after disconnects because - # should_compress(0) never fires. (#2153) - _compressor = self.context_compressor - if _compressor.last_prompt_tokens > 0: - # Only use prompt_tokens — completion/reasoning - # tokens don't consume context window space. - # Thinking models (GLM-5.1, QwQ, DeepSeek R1) - # inflate completion_tokens with reasoning, - # causing premature compression. (#12026) - _real_tokens = _compressor.last_prompt_tokens - else: - # Include tool schemas — with 50+ tools enabled - # these add 20-30K tokens the messages-only - # estimate misses, which can skip compression - # past the configured threshold (#14695). - _real_tokens = estimate_request_tokens_rough( - messages, tools=self.tools or None - ) - - if self.compression_enabled and _compressor.should_compress(_real_tokens): - self._safe_print(" ⟳ compacting context…") - messages, active_system_prompt = self._compress_context( - messages, system_message, - approx_tokens=self.context_compressor.last_prompt_tokens, - task_id=effective_task_id, - ) - # Compression created a new session — clear history so - # _flush_messages_to_session_db writes compressed messages - # to the new session (see preflight compression comment). - conversation_history = None - - # Save session log incrementally (so progress is visible even if interrupted) - self._session_messages = messages - self._save_session_log(messages) - - # Continue loop for next response - continue - - else: - # No tool calls - this is the final response - final_response = assistant_message.content or "" - - # Fix: unmute output when entering the no-tool-call branch - # so the user can see empty-response warnings and recovery - # status messages. _mute_post_response was set during a - # prior housekeeping tool turn and should not silence the - # final response path. - self._mute_post_response = False - - # Check if response only has think block with no actual content after it - if not self._has_content_after_think_block(final_response): - # ── Partial stream recovery ───────────────────── - # If content was already streamed to the user before - # the connection died, use it as the final response - # instead of falling through to prior-turn fallback - # or wasting API calls on retries. - _partial_streamed = ( - getattr(self, "_current_streamed_assistant_text", "") or "" - ) - if self._has_content_after_think_block(_partial_streamed): - _turn_exit_reason = "partial_stream_recovery" - _recovered = self._strip_think_blocks(_partial_streamed).strip() - logger.info( - "Partial stream content delivered (%d chars) " - "— using as final response", - len(_recovered), - ) - self._emit_status( - "↻ Stream interrupted — using delivered content " - "as final response" - ) - final_response = _recovered - self._response_was_previewed = True - break - - # If the previous turn already delivered real content alongside - # HOUSEKEEPING tool calls (e.g. "You're welcome!" + memory save), - # the model has nothing more to say. Use the earlier content - # immediately instead of wasting API calls on retries. - # NOTE: Only use this shortcut when ALL tools in that turn were - # housekeeping (memory, todo, etc.). When substantive tools - # were called (terminal, search_files, etc.), the content was - # likely mid-task narration ("I'll scan the directory...") and - # the empty follow-up means the model choked — let the - # post-tool nudge below handle that instead of exiting early. - fallback = getattr(self, '_last_content_with_tools', None) - if fallback and getattr(self, '_last_content_tools_all_housekeeping', False): - _turn_exit_reason = "fallback_prior_turn_content" - logger.info("Empty follow-up after tool calls — using prior turn content as final response") - self._emit_status("↻ Empty response after tool calls — using earlier content as final answer") - self._last_content_with_tools = None - self._last_content_tools_all_housekeeping = False - self._empty_content_retries = 0 - # Do NOT modify the assistant message content — the - # old code injected "Calling the X tools..." which - # poisoned the conversation history. Just use the - # fallback text as the final response and break. - final_response = self._strip_think_blocks(fallback).strip() - self._response_was_previewed = True - break - - # ── Post-tool-call empty response nudge ─────────── - # The model returned empty after executing tool calls. - # This covers two cases: - # (a) No prior-turn content at all — model went silent - # (b) Prior turn had content + SUBSTANTIVE tools (the - # fallback above was skipped because the content - # was mid-task narration, not a final answer) - # Instead of giving up, nudge the model to continue by - # appending a user-level hint. This is the #9400 case: - # weaker models (mimo-v2-pro, GLM-5, etc.) sometimes - # return empty after tool results instead of continuing - # to the next step. One retry with a nudge usually - # fixes it. - _prior_was_tool = any( - m.get("role") == "tool" - for m in messages[-5:] # check recent messages - ) - # Detect Qwen3/Ollama-style in-content thinking blocks. - # Ollama puts in the content field (not in - # reasoning_content), so _has_structured below would - # miss it. We check here so thinking-only responses - # after tool calls route to prefill instead of nudge. - _has_inline_thinking = bool( - re.search( - r'||', - final_response or "", - re.IGNORECASE, - ) - ) - if ( - _prior_was_tool - and not getattr(self, "_post_tool_empty_retried", False) - and not _has_inline_thinking # thinking model still working — let prefill handle - ): - self._post_tool_empty_retried = True - # Clear stale narration so it doesn't resurface - # on a later empty response after the nudge. - self._last_content_with_tools = None - self._last_content_tools_all_housekeeping = False - logger.info( - "Empty response after tool calls — nudging model " - "to continue processing" - ) - self._emit_status( - "⚠️ Model returned empty after tool calls — " - "nudging to continue" - ) - # Append the empty assistant message first so the - # message sequence stays valid: - # tool(result) → assistant("(empty)") → user(nudge) - # Without this, we'd have tool → user which most - # APIs reject as an invalid sequence. - _nudge_msg = self._build_assistant_message(assistant_message, finish_reason) - _nudge_msg["content"] = "(empty)" - _nudge_msg["_empty_recovery_synthetic"] = True - messages.append(_nudge_msg) - messages.append({ - "role": "user", - "content": ( - "You just executed tool calls but returned an " - "empty response. Please process the tool " - "results above and continue with the task." - ), - "_empty_recovery_synthetic": True, - }) - continue - - # ── Thinking-only prefill continuation ────────── - # The model produced structured reasoning (via API - # fields) but no visible text content. Rather than - # giving up, append the assistant message as-is and - # continue — the model will see its own reasoning - # on the next turn and produce the text portion. - # Inspired by clawdbot's "incomplete-text" recovery. - # Also covers Qwen3/Ollama in-content blocks - # (detected above as _has_inline_thinking). - _has_structured = bool( - getattr(assistant_message, "reasoning", None) - or getattr(assistant_message, "reasoning_content", None) - or getattr(assistant_message, "reasoning_details", None) - or _has_inline_thinking - ) - if _has_structured and self._thinking_prefill_retries < 2: - self._thinking_prefill_retries += 1 - logger.info( - "Thinking-only response (no visible content) — " - "prefilling to continue (%d/2)", - self._thinking_prefill_retries, - ) - self._emit_status( - f"↻ Thinking-only response — prefilling to continue " - f"({self._thinking_prefill_retries}/2)" - ) - interim_msg = self._build_assistant_message( - assistant_message, "incomplete" - ) - interim_msg["_thinking_prefill"] = True - messages.append(interim_msg) - self._session_messages = messages - self._save_session_log(messages) - continue - - # ── Empty response retry ────────────────────── - # Model returned nothing usable. Retry up to 3 - # times before attempting fallback. This covers - # both truly empty responses (no content, no - # reasoning) AND reasoning-only responses after - # prefill exhaustion — models like mimo-v2-pro - # always populate reasoning fields via OpenRouter, - # so the old `not _has_structured` guard blocked - # retries for every reasoning model after prefill. - _truly_empty = not self._strip_think_blocks( - final_response - ).strip() - _prefill_exhausted = ( - _has_structured - and self._thinking_prefill_retries >= 2 - ) - if _truly_empty and (not _has_structured or _prefill_exhausted) and self._empty_content_retries < 3: - self._empty_content_retries += 1 - logger.warning( - "Empty response (no content or reasoning) — " - "retry %d/3 (model=%s)", - self._empty_content_retries, self.model, - ) - self._emit_status( - f"⚠️ Empty response from model — retrying " - f"({self._empty_content_retries}/3)" - ) - continue - - # ── Exhausted retries — try fallback provider ── - # Before giving up with "(empty)", attempt to - # switch to the next provider in the fallback - # chain. This covers the case where a model - # (e.g. GLM-4.5-Air) consistently returns empty - # due to context degradation or provider issues. - if _truly_empty and self._fallback_chain: - logger.warning( - "Empty response after %d retries — " - "attempting fallback (model=%s, provider=%s)", - self._empty_content_retries, self.model, - self.provider, - ) - self._emit_status( - "⚠️ Model returning empty responses — " - "switching to fallback provider..." - ) - if self._try_activate_fallback(): - self._empty_content_retries = 0 - self._emit_status( - f"↻ Switched to fallback: {self.model} " - f"({self.provider})" - ) - logger.info( - "Fallback activated after empty responses: " - "now using %s on %s", - self.model, self.provider, - ) - continue - - # Exhausted retries and fallback chain (or no - # fallback configured). Fall through to the - # "(empty)" terminal. - _turn_exit_reason = "empty_response_exhausted" - reasoning_text = self._extract_reasoning(assistant_message) - self._drop_trailing_empty_response_scaffolding(messages) - assistant_msg = self._build_assistant_message(assistant_message, finish_reason) - assistant_msg["content"] = "(empty)" - # This is a user-facing failure sentinel for the gateway, - # not real assistant content. Persisting it makes later - # "continue" turns replay assistant("(empty)") as if it - # were a meaningful model response, which can keep long - # tool-heavy sessions stuck in empty-response loops. - assistant_msg["_empty_terminal_sentinel"] = True - messages.append(assistant_msg) - - if reasoning_text: - reasoning_preview = reasoning_text[:500] + "..." if len(reasoning_text) > 500 else reasoning_text - logger.warning( - "Reasoning-only response (no visible content) " - "after exhausting retries and fallback. " - "Reasoning: %s", reasoning_preview, - ) - self._emit_status( - "⚠️ Model produced reasoning but no visible " - "response after all retries. Returning empty." - ) - else: - logger.warning( - "Empty response (no content or reasoning) " - "after %d retries. No fallback available. " - "model=%s provider=%s", - self._empty_content_retries, self.model, - self.provider, - ) - self._emit_status( - "❌ Model returned no content after all retries" - + (" and fallback attempts." if self._fallback_chain else - ". No fallback providers configured.") - ) - - final_response = "(empty)" - break - - # Reset retry counter/signature on successful content - self._empty_content_retries = 0 - self._thinking_prefill_retries = 0 - - if ( - self.api_mode == "codex_responses" - and self.valid_tool_names - and codex_ack_continuations < 2 - and self._looks_like_codex_intermediate_ack( - user_message=user_message, - assistant_content=final_response, - messages=messages, - ) - ): - codex_ack_continuations += 1 - interim_msg = self._build_assistant_message(assistant_message, "incomplete") - messages.append(interim_msg) - self._emit_interim_assistant_message(interim_msg) - - continue_msg = { - "role": "user", - "content": ( - "[System: Continue now. Execute the required tool calls and only " - "send your final answer after completing the task.]" - ), - } - messages.append(continue_msg) - self._session_messages = messages - self._save_session_log(messages) - continue - - codex_ack_continuations = 0 - - if truncated_response_parts: - final_response = "".join(truncated_response_parts) + final_response - truncated_response_parts = [] - length_continue_retries = 0 - - final_response = self._strip_think_blocks(final_response).strip() - - final_msg = self._build_assistant_message(assistant_message, finish_reason) - - # Pop thinking-only prefill and empty-response retry - # scaffolding before appending the final response. These - # internal turns are only for the next API retry and should - # not become durable transcript context. - while ( - messages - and isinstance(messages[-1], dict) - and ( - messages[-1].get("_thinking_prefill") - or messages[-1].get("_empty_recovery_synthetic") - or messages[-1].get("_empty_terminal_sentinel") - ) - ): - messages.pop() - - messages.append(final_msg) - - _turn_exit_reason = f"text_response(finish_reason={finish_reason})" - if not self.quiet_mode: - self._safe_print(f"🎉 Conversation completed after {api_call_count} OpenAI-compatible API call(s)") - break - - except Exception as e: - error_msg = f"Error during OpenAI-compatible API call #{api_call_count}: {str(e)}" - try: - print(f"❌ {error_msg}") - except (OSError, ValueError): - logger.error(error_msg) - - logger.debug("Outer loop error in API call #%d", api_call_count, exc_info=True) - - # If an assistant message with tool_calls was already appended, - # the API expects a role="tool" result for every tool_call_id. - # Fill in error results for any that weren't answered yet. - for idx in range(len(messages) - 1, -1, -1): - msg = messages[idx] - if not isinstance(msg, dict): - break - if msg.get("role") == "tool": - continue - if msg.get("role") == "assistant" and msg.get("tool_calls"): - answered_ids = { - m["tool_call_id"] - for m in messages[idx + 1:] - if isinstance(m, dict) and m.get("role") == "tool" - } - for tc in msg["tool_calls"]: - if not tc or not isinstance(tc, dict): continue - if tc["id"] not in answered_ids: - err_msg = { - "role": "tool", - "name": AIAgent._get_tool_call_name_static(tc), - "tool_call_id": tc["id"], - "content": f"Error executing tool: {error_msg}", - } - messages.append(err_msg) - break - - # Non-tool errors don't need a synthetic message injected. - # The error is already printed to the user (line above), and - # the retry loop continues. Injecting a fake user/assistant - # message pollutes history, burns tokens, and risks violating - # role-alternation invariants. - - # If we're near the limit, break to avoid infinite loops - if api_call_count >= self.max_iterations - 1: - _turn_exit_reason = f"error_near_max_iterations({error_msg[:80]})" - final_response = f"I apologize, but I encountered repeated errors: {error_msg}" - # Append as assistant so the history stays valid for - # session resume (avoids consecutive user messages). - messages.append({"role": "assistant", "content": final_response}) - break - - if final_response is None and ( - api_call_count >= self.max_iterations - or self.iteration_budget.remaining <= 0 - ): - # Budget exhausted — ask the model for a summary via one extra - # API call with tools stripped. _handle_max_iterations injects a - # user message and makes a single toolless request. - _turn_exit_reason = f"max_iterations_reached({api_call_count}/{self.max_iterations})" - self._emit_status( - f"⚠️ Iteration budget exhausted ({api_call_count}/{self.max_iterations}) " - "— asking model to summarise" - ) - if not self.quiet_mode: - self._safe_print( - f"\n⚠️ Iteration budget exhausted ({api_call_count}/{self.max_iterations}) " - "— requesting summary..." - ) - final_response = self._handle_max_iterations(messages, api_call_count) - - # If running as a kanban worker, block the task so the dispatcher - # knows the worker could not complete (rather than treating it as a - # protocol violation). The agent loop strips tools before calling - # _handle_max_iterations, so the model cannot call kanban_block - # itself — we must do it on its behalf. - _kanban_task = os.environ.get("HERMES_KANBAN_TASK") - if _kanban_task: - try: - handle_function_call( - "kanban_block", - { - "task_id": _kanban_task, - "reason": ( - f"Iteration budget exhausted " - f"({api_call_count}/{self.max_iterations}) — " - "task could not complete within the allowed " - "iterations" - ), - }, - task_id=effective_task_id, - ) - logger.info( - "kanban_block called for task %s after iteration " - "exhaustion (%d/%d)", - _kanban_task, api_call_count, self.max_iterations, - ) - except Exception: - logger.warning( - "Failed to call kanban_block after iteration " - "exhaustion for task %s", - _kanban_task, - exc_info=True, - ) - - # Determine if conversation completed successfully - completed = final_response is not None and api_call_count < self.max_iterations - - # Save trajectory if enabled. ``user_message`` may be a multimodal - # list of parts; the trajectory format wants a plain string. - self._save_trajectory(messages, _summarize_user_message_for_log(user_message), completed) - - # Clean up VM and browser for this task after conversation completes - self._cleanup_task_resources(effective_task_id) - - # Persist session to both JSON log and SQLite only after private retry - # scaffolding has been removed. Otherwise a later user "continue" turn - # can replay assistant("(empty)") / recovery nudges and fall into the - # same empty-response loop again. - self._drop_trailing_empty_response_scaffolding(messages) - self._persist_session(messages, conversation_history) - - # ── Turn-exit diagnostic log ───────────────────────────────────── - # Always logged at INFO so agent.log captures WHY every turn ended. - # When the last message is a tool result (agent was mid-work), log - # at WARNING — this is the "just stops" scenario users report. - _last_msg_role = messages[-1].get("role") if messages else None - _last_tool_name = None - if _last_msg_role == "tool": - # Walk back to find the assistant message with the tool call - for _m in reversed(messages): - if _m.get("role") == "assistant" and _m.get("tool_calls"): - _tcs = _m["tool_calls"] - if _tcs and isinstance(_tcs[0], dict): - _last_tool_name = _tcs[-1].get("function", {}).get("name") - break - - _turn_tool_count = sum( - 1 for m in messages - if isinstance(m, dict) and m.get("role") == "assistant" and m.get("tool_calls") - ) - _resp_len = len(final_response) if final_response else 0 - _budget_used = self.iteration_budget.used if self.iteration_budget else 0 - _budget_max = self.iteration_budget.max_total if self.iteration_budget else 0 - - _diag_msg = ( - "Turn ended: reason=%s model=%s api_calls=%d/%d budget=%d/%d " - "tool_turns=%d last_msg_role=%s response_len=%d session=%s" - ) - _diag_args = ( - _turn_exit_reason, self.model, api_call_count, self.max_iterations, - _budget_used, _budget_max, - _turn_tool_count, _last_msg_role, _resp_len, - self.session_id or "none", - ) - - if _last_msg_role == "tool" and not interrupted: - # Agent was mid-work — this is the "just stops" case. - logger.warning( - "Turn ended with pending tool result (agent may appear stuck). " - + _diag_msg + " last_tool=%s", - *_diag_args, _last_tool_name, - ) - else: - logger.info(_diag_msg, *_diag_args) - - # File-mutation verifier footer. - # If one or more ``write_file`` / ``patch`` calls failed during this - # turn and were never superseded by a successful write to the same - # path, append an advisory footer to the assistant response. This - # catches the specific case — reported by Ben Eng (#15524-adjacent) - # — where a model issues a batch of parallel patches, half of them - # fail with "Could not find old_string", and the model summarises - # the turn claiming every file was edited. The user then has to - # manually run ``git status`` to catch the lie. With this footer - # the truth is surfaced on every turn, so over-claiming is - # structurally impossible past the model. - # - # Gate: only applied when a real text response exists for this - # turn and the user didn't interrupt. Empty/interrupted turns - # already have other surface text that shouldn't be augmented. - if final_response and not interrupted: - try: - _failed = getattr(self, "_turn_failed_file_mutations", None) or {} - if _failed and self._file_mutation_verifier_enabled(): - footer = self._format_file_mutation_failure_footer(_failed) - if footer: - final_response = final_response.rstrip() + "\n\n" + footer - except Exception as _ver_err: - logger.debug("file-mutation verifier footer failed: %s", _ver_err) - - # Plugin hook: transform_llm_output - # Fired once per turn after the tool-calling loop completes. - # Plugins can transform the LLM's output text before it's returned. - # First hook to return a string wins; None/empty return leaves text unchanged. - if final_response and not interrupted: - try: - from hermes_cli.plugins import invoke_hook as _invoke_hook - _transform_results = _invoke_hook( - "transform_llm_output", - response_text=final_response, - session_id=self.session_id or "", - model=self.model, - platform=getattr(self, "platform", None) or "", - ) - for _hook_result in _transform_results: - if isinstance(_hook_result, str) and _hook_result: - final_response = _hook_result - break # First non-empty string wins - except Exception as exc: - logger.warning("transform_llm_output hook failed: %s", exc) - - # Plugin hook: post_llm_call - # Fired once per turn after the tool-calling loop completes. - # Plugins can use this to persist conversation data (e.g. sync - # to an external memory system). - if final_response and not interrupted: - try: - from hermes_cli.plugins import invoke_hook as _invoke_hook - _invoke_hook( - "post_llm_call", - session_id=self.session_id, - user_message=original_user_message, - assistant_response=final_response, - conversation_history=list(messages), - model=self.model, - platform=getattr(self, "platform", None) or "", - ) - except Exception as exc: - logger.warning("post_llm_call hook failed: %s", exc) - - # Extract reasoning from the CURRENT turn only. Walk backwards - # but stop at the user message that started this turn — anything - # earlier is from a prior turn and must not leak into the reasoning - # box (confusing stale display; #17055). Within the current turn - # we still want the *most recent* non-empty reasoning: many - # providers (Claude thinking, DeepSeek v4, Codex Responses) emit - # reasoning on the tool-call step and leave the final-answer step - # with reasoning=None, so picking only the last assistant would - # silently drop legitimate same-turn reasoning. - last_reasoning = None - for msg in reversed(messages): - if msg.get("role") == "user": - break # turn boundary — don't cross into prior turns - if msg.get("role") == "assistant" and msg.get("reasoning"): - last_reasoning = msg["reasoning"] - break - - # Build result with interrupt info if applicable - result = { - "final_response": final_response, - "last_reasoning": last_reasoning, - "messages": messages, - "api_calls": api_call_count, - "completed": completed, - "turn_exit_reason": _turn_exit_reason, - "partial": False, # True only when stopped due to invalid tool calls - "interrupted": interrupted, - "response_previewed": getattr(self, "_response_was_previewed", False), - "model": self.model, - "provider": self.provider, - "base_url": self.base_url, - "input_tokens": self.session_input_tokens, - "output_tokens": self.session_output_tokens, - "cache_read_tokens": self.session_cache_read_tokens, - "cache_write_tokens": self.session_cache_write_tokens, - "reasoning_tokens": self.session_reasoning_tokens, - "prompt_tokens": self.session_prompt_tokens, - "completion_tokens": self.session_completion_tokens, - "total_tokens": self.session_total_tokens, - "last_prompt_tokens": getattr(self.context_compressor, "last_prompt_tokens", 0) or 0, - "estimated_cost_usd": self.session_estimated_cost_usd, - "cost_status": self.session_cost_status, - "cost_source": self.session_cost_source, - } - if self._tool_guardrail_halt_decision is not None: - result["guardrail"] = self._tool_guardrail_halt_decision.to_metadata() - # If a /steer landed after the final assistant turn (no more tool - # batches to drain into), hand it back to the caller so it can be - # delivered as the next user turn instead of being silently lost. - _leftover_steer = self._drain_pending_steer() - if _leftover_steer: - result["pending_steer"] = _leftover_steer - self._response_was_previewed = False - - # Include interrupt message if one triggered the interrupt - if interrupted and self._interrupt_message: - result["interrupt_message"] = self._interrupt_message - - # Clear interrupt state after handling - self.clear_interrupt() - - # Clear stream callback so it doesn't leak into future calls - self._stream_callback = None - - # Check skill trigger NOW — based on how many tool iterations THIS turn used. - _should_review_skills = False - if (self._skill_nudge_interval > 0 - and self._iters_since_skill >= self._skill_nudge_interval - and "skill_manage" in self.valid_tool_names): - _should_review_skills = True - self._iters_since_skill = 0 - - # External memory provider: sync the completed turn + queue next prefetch. - self._sync_external_memory_for_turn( - original_user_message=original_user_message, - final_response=final_response, - interrupted=interrupted, - ) - - # Background memory/skill review — runs AFTER the response is delivered - # so it never competes with the user's task for model attention. - if final_response and not interrupted and (_should_review_memory or _should_review_skills): - try: - self._spawn_background_review( - messages_snapshot=list(messages), - review_memory=_should_review_memory, - review_skills=_should_review_skills, - ) - except Exception: - pass # Background review is best-effort - - # Note: Memory provider on_session_end() + shutdown_all() are NOT - # called here — run_conversation() is called once per user message in - # multi-turn sessions. Shutting down after every turn would kill the - # provider before the second message. Actual session-end cleanup is - # handled by the CLI (atexit / /reset) and gateway (session expiry / - # _reset_session). - - # Plugin hook: on_session_end - # Fired at the very end of every run_conversation call. - # Plugins can use this for cleanup, flushing buffers, etc. - try: - from hermes_cli.plugins import invoke_hook as _invoke_hook - _invoke_hook( - "on_session_end", - session_id=self.session_id, - completed=completed, - interrupted=interrupted, - model=self.model, - platform=getattr(self, "platform", None) or "", - ) - except Exception as exc: - logger.warning("on_session_end hook failed: %s", exc) - - return result + """Forwarder — see ``agent.conversation_loop.run_conversation``.""" + from agent.conversation_loop import run_conversation + return run_conversation(self, user_message, system_message, conversation_history, task_id, stream_callback, persist_user_message) def chat(self, message: str, stream_callback: Optional[callable] = None) -> str: """ @@ -16072,144 +3899,9 @@ class AIAgent: effective_task_id: str, should_review_memory: bool = False, ) -> Dict[str, Any]: - """Codex app-server runtime path. Hands the entire turn to a `codex - app-server` subprocess and projects its events back into Hermes' - messages list so memory/skill review keep working. - - Called from run_conversation() when self.api_mode == "codex_app_server". - Returns the same dict shape as the chat_completions path. - """ - from agent.transports.codex_app_server_session import CodexAppServerSession - - # Lazy session: one CodexAppServerSession per AIAgent instance. - # Spawned on first turn, reused across turns, closed at AIAgent - # shutdown (see _cleanup hook). - if not hasattr(self, "_codex_session") or self._codex_session is None: - cwd = getattr(self, "session_cwd", None) or os.getcwd() - # Approval callback: defer to Hermes' standard prompt flow if a - # CLI thread has installed one. Gateway / cron contexts get the - # codex-side fail-closed default. - try: - from tools.terminal_tool import _get_approval_callback - approval_callback = _get_approval_callback() - except Exception: - approval_callback = None - self._codex_session = CodexAppServerSession( - cwd=cwd, - approval_callback=approval_callback, - ) - - # NOTE: the user message is ALREADY appended to messages by the - # standard run_conversation() flow (line ~11823) before the early - # return reaches us. Do NOT append again — that would duplicate. - - try: - turn = self._codex_session.run_turn(user_input=user_message) - except Exception as exc: - logger.exception("codex app-server turn failed") - # Crash → unconditionally drop the session so the next turn - # respawns from scratch instead of reusing a dead client. - try: - self._codex_session.close() - except Exception: - pass - self._codex_session = None - return { - "final_response": ( - f"Codex app-server turn failed: {exc}. " - f"Fall back to default runtime with `/codex-runtime auto`." - ), - "messages": messages, - "api_calls": 0, - "completed": False, - "partial": True, - "error": str(exc), - } - - # If the turn signalled the underlying client is wedged (deadline - # blown, post-tool watchdog tripped, OAuth refresh died, subprocess - # exited), retire the session so the next turn respawns codex - # rather than riding the broken process. Mirrors openclaw beta.8's - # "retire timed-out app-server clients" fix. - if getattr(turn, "should_retire", False): - logger.warning( - "codex app-server session retired (turn error: %s)", - turn.error, - ) - try: - self._codex_session.close() - except Exception: - pass - self._codex_session = None - - # Splice projected messages into the conversation. The projector emits - # standard {role, content, tool_calls, tool_call_id} entries, which - # is exactly what curator.py / sessions DB expect. - if turn.projected_messages: - messages.extend(turn.projected_messages) - - # Counter ticks for the self-improvement loop. - # _turns_since_memory and _user_turn_count are ALREADY incremented - # in the run_conversation() pre-loop block (lines ~11793-11817) so we - # do NOT touch them here — that would double-count. - # Only _iters_since_skill needs explicit increment, since the - # chat_completions loop bumps it per tool iteration (line ~12110) - # and that loop is bypassed on this path. - self._iters_since_skill = ( - getattr(self, "_iters_since_skill", 0) + turn.tool_iterations - ) - - # Now check the skill nudge AFTER iters were incremented — same - # pattern the chat_completions path uses (line ~15432). - should_review_skills = False - if ( - self._skill_nudge_interval > 0 - and self._iters_since_skill >= self._skill_nudge_interval - and "skill_manage" in self.valid_tool_names - ): - should_review_skills = True - self._iters_since_skill = 0 - - # External memory provider sync (mirrors line ~15439). Skipped on - # interrupt/error to avoid feeding partial transcripts to memory. - if not turn.interrupted and turn.error is None: - try: - self._sync_external_memory_for_turn( - original_user_message=original_user_message, - final_response=turn.final_text, - interrupted=False, - ) - except Exception: - logger.debug("external memory sync raised", exc_info=True) - - # Background review fork — same cadence + signature as the default - # path (line ~15449). Only fires when a trigger actually tripped AND - # we have a real final response. - if ( - turn.final_text - and not turn.interrupted - and (should_review_memory or should_review_skills) - ): - try: - self._spawn_background_review( - messages_snapshot=list(messages), - review_memory=should_review_memory, - review_skills=should_review_skills, - ) - except Exception: - logger.debug("background review spawn raised", exc_info=True) - - return { - "final_response": turn.final_text, - "messages": messages, - "api_calls": 1, # one app-server "turn" maps to one logical API call - "completed": not turn.interrupted and turn.error is None, - "partial": turn.interrupted or turn.error is not None, - "error": turn.error, - "codex_thread_id": turn.thread_id, - "codex_turn_id": turn.turn_id, - } - + """Forwarder — see ``agent.codex_runtime.run_codex_app_server_turn``.""" + from agent.codex_runtime import run_codex_app_server_turn + return run_codex_app_server_turn(self, user_message=user_message, original_user_message=original_user_message, messages=messages, effective_task_id=effective_task_id, should_review_memory=should_review_memory) def main( query: str = None, diff --git a/scripts/check-windows-footguns.py b/scripts/check-windows-footguns.py index f424be90710..7ae7ca50c4e 100644 --- a/scripts/check-windows-footguns.py +++ b/scripts/check-windows-footguns.py @@ -551,6 +551,14 @@ def print_rules() -> None: def main(argv: list[str]) -> int: + # Windows terminals default to cp1252, which can't encode the ✓/✗ + # characters used in the output. Reconfigure streams to UTF-8 so the + # script works correctly on the very platform it is designed to help. + if hasattr(sys.stdout, "reconfigure"): + sys.stdout.reconfigure(encoding="utf-8") + if hasattr(sys.stderr, "reconfigure"): + sys.stderr.reconfigure(encoding="utf-8") + args = parse_args(argv) if args.list: diff --git a/scripts/release.py b/scripts/release.py index 6bbc2ad4ae3..d554e474fe6 100755 --- a/scripts/release.py +++ b/scripts/release.py @@ -184,6 +184,7 @@ AUTHOR_MAP = { "santoshhumagain1887@gmail.com": "npmisantosh", "39641663+luarss@users.noreply.github.com": "luarss", "16263913+zccyman@users.noreply.github.com": "zccyman", + "zccyman@users.noreply.github.com": "zccyman", # PR #26998 (auxiliary fallback chain) "ahmetosrak@Ahmet-MacBook-Air.local": "Osraka", "98612432+Osraka@users.noreply.github.com": "Osraka", "112634774+ryptotalent@users.noreply.github.com": "ryptotalent", @@ -1096,6 +1097,7 @@ AUTHOR_MAP = { "4296245+matthewlai@users.noreply.github.com": "matthewlai", "109617724+0xchainer@users.noreply.github.com": "0xchainer", # PR #27154/27138/27147 salvage "201800237+kronexoi@users.noreply.github.com": "kronexoi", # PR #27167 salvage (Teams port fallback) + "283442588+EloquentBrush0x@users.noreply.github.com": "EloquentBrush0x", # PR #26642 salvage (post_setup parity) # batch salvage (May 2026 LHF run, group 2) "shellybotmoyer@example.com": "shellybotmoyer", # PR #26661 (kanban --severity >=) "coulson@shellybotmoyer.com": "shellybotmoyer", # PR #25576 (credential_pool ISO rehydrate) @@ -1113,6 +1115,54 @@ AUTHOR_MAP = { "hermesagent26@gmail.com": "hermesagent26", # PR #26438 (kimi model-name reasoning pad) "276067471+hermesagent26@users.noreply.github.com": "hermesagent26", "71590782+kriscolab@users.noreply.github.com": "kriscolab", # PR #26926 (deepseek default_aux_model) + # batch salvage (May 2026 LHF run, group 3) + "darvsum@users.noreply.github.com": "darvsum", # PR #26766 (preserve discover_models in normalize) + "peter@Peters-Mac-mini.local": "hueilau", # PR #26498 (strip image parts for non-vision) + "33933019+hueilau@users.noreply.github.com": "hueilau", + "32297275+Timur00Kh@users.noreply.github.com": "Timur00Kh", # PR #27114 (telegram DM topic for synthetic events) + "al.bellemare@gmail.com": "Grogger", # PR #27061 (windows console flash suppress) + "clement@nousresearch.com": "lemassykoi", # PR #27042 (model-switch probe keyless providers) + "16377344+lemassykoi@users.noreply.github.com": "lemassykoi", + "draplater@icloud.com": "draplater", # PR #26707 (goal judge current time) + "6349758+draplater@users.noreply.github.com": "draplater", + "pr7426@users.noreply.github.com": "pr7426", # PR #27048 (cron parallel job loss) + "rahulnilvan43@gmail.com": "therahul-yo", # PR #26215 (mock keychain in tests) + "kingsleyemeka117@gmail.com": "flamiinngo", # PR #27205 (UnicodeEncodeError footgun checker) + # batch salvage (May 2026 LHF run, group 4) + "283442588+EloquentBrush0x@users.noreply.github.com": "EloquentBrush0x", # PR #26657 (trust_env aiohttp) + "205509009+subtract0@users.noreply.github.com": "subtract0", # PR #25658 (zsh $status -> $rc) + "patryk@jarmakowicz.me": "zwolniony", # PR #26961 (gemini x-goog-api-key) + "12735938+zwolniony@users.noreply.github.com": "zwolniony", + "ambuj@dodopayments.com": "that-ambuj", # PR #26582 (preserve underscores) + "zccyman@163.com": "zccyman", # PR #25294 (custom provider api_key_env alias) + "bitkyc08@gmail.com": "lidge-jun", # PR #26814 (api server browser security headers) + "sp_ps@Mac-mini.lan": "phoenixshen", # PR #26768 (respect user-configured vision model) + "1594534+phoenixshen@users.noreply.github.com": "phoenixshen", + "147827411+AhmetArif0@users.noreply.github.com": "AhmetArif0", # PR #26635 (line proxy env vars) + # batch salvage (May 2026 LHF run, group 5) + "hari@Hariharans-MacBook-Air-8.local": "haran2001", # PR #27070 (i18n catalog test) + "hariharan15151@gmail.com": "haran2001", # PR #27068 (qwen3.6-plus 1M context) + "56040092+haran2001@users.noreply.github.com": "haran2001", + "1472110+ms-alan@users.noreply.github.com": "ms-alan", # PR #26443 (reload-skills tab completion) + "ganlinbupt@gmail.com": "godlin-gh", # PR #26118 (ACP polished tools) + "wesley.simplicio.ext@siemens-energy.com": "wesleysimplicio", # PR #25777 (xterm.js native selection) + "6108320+wesleysimplicio@users.noreply.github.com": "wesleysimplicio", + "carryzuo00@gmail.com": "Carry00", # PR #26851 (doctor SSH env vars) + "alaamohanad169-ship-it@users.noreply.github.com": "alaamohanad169-ship-it", # PR #26036 (telegram typing after send) + "vigo@hermes": "hawknewton", # PR #26294 (bedrock boto3 lazy_deps) + "211668+hawknewton@users.noreply.github.com": "hawknewton", + "quenvix00@gmail.com": "QuenVix", # PR #26761/26772 salvage + "164776164+QuenVix@users.noreply.github.com": "QuenVix", + "262945885+Mind-Dragon@users.noreply.github.com": "Mind-Dragon", # PR #26966 salvage + "soynchuux@gmail.com": "soynchux", # PR #27060 salvage + "209694554+soynchux@users.noreply.github.com": "soynchux", + # batch salvage (May 2026 LHF run, group 6 — final) + "6666242+bird@users.noreply.github.com": "bird", # PR #25219 (gateway docker exit-75 restart) + "david@loadmagic.ai": "davidcampbelldc", # PR #26834 (web_server proxy_headers=False) + "165905879+davidcampbelldc@users.noreply.github.com": "davidcampbelldc", + "hoangv.pham0803@gmail.com": "hehehe0803", # PR #26212 salvage (codex kanban writable root) + "26063003+hehehe0803@users.noreply.github.com": "hehehe0803", + "38348871+vaddisrinivas@users.noreply.github.com": "vaddisrinivas", # PR #26394 salvage (Docker messaging extra) } diff --git a/skills/creative/comfyui/scripts/_common.py b/skills/creative/comfyui/scripts/_common.py index ef742733eb5..efe592a1b33 100644 --- a/skills/creative/comfyui/scripts/_common.py +++ b/skills/creative/comfyui/scripts/_common.py @@ -592,7 +592,7 @@ def _http_once( # Build a new request with cleaned headers clean_headers = { k: v for k, v in req2.header_items() - if k.lower() not in ("x-api-key", "authorization", "cookie") + if k.lower() not in {"x-api-key", "authorization", "cookie"} } new_req = urllib.request.Request(newurl, headers=clean_headers, method="GET") return new_req @@ -743,13 +743,13 @@ def safe_path_join(base: Path, *parts: str) -> Path: def media_type_from_filename(filename: str) -> str: ext = Path(filename).suffix.lower() - if ext in (".mp4", ".webm", ".avi", ".mov", ".mkv", ".gif", ".webp"): + if ext in {".mp4", ".webm", ".avi", ".mov", ".mkv", ".gif", ".webp"}: return "video" - if ext in (".wav", ".mp3", ".flac", ".ogg", ".m4a"): + if ext in {".wav", ".mp3", ".flac", ".ogg", ".m4a"}: return "audio" - if ext in (".glb", ".obj", ".ply", ".gltf"): + if ext in {".glb", ".obj", ".ply", ".gltf"}: return "3d" - if ext in (".json", ".txt", ".md"): + if ext in {".json", ".txt", ".md"}: return "text" return "image" diff --git a/skills/creative/comfyui/scripts/extract_schema.py b/skills/creative/comfyui/scripts/extract_schema.py index ba44cfdf6a2..0eab65b20fd 100755 --- a/skills/creative/comfyui/scripts/extract_schema.py +++ b/skills/creative/comfyui/scripts/extract_schema.py @@ -81,7 +81,7 @@ def trace_to_node(workflow: dict, link: list, *, max_hops: int = 8) -> str | Non return None cls = node.get("class_type", "") # Reroute / Primitive / passthrough wrappers - if cls in ("Reroute", "PrimitiveNode", "Note", "easy showAnything"): + if cls in {"Reroute", "PrimitiveNode", "Note", "easy showAnything"}: inputs = node.get("inputs", {}) or {} # Find first link-shaped input and follow it next_link = next((v for v in inputs.values() if is_link(v)), None) @@ -105,7 +105,7 @@ def find_negative_prompt_node(workflow: dict) -> str | None: src = trace_to_node(workflow, neg) if src and isinstance(workflow.get(src), dict): cls = workflow[src].get("class_type", "") - if cls.startswith("CLIPTextEncode") or cls in ("smZ CLIPTextEncode", "BNK_CLIPTextEncodeAdvanced"): + if cls.startswith("CLIPTextEncode") or cls in {"smZ CLIPTextEncode", "BNK_CLIPTextEncodeAdvanced"}: return src return None @@ -121,7 +121,7 @@ def find_positive_prompt_node(workflow: dict) -> str | None: src = trace_to_node(workflow, pos) if src and isinstance(workflow.get(src), dict): cls = workflow[src].get("class_type", "") - if cls.startswith("CLIPTextEncode") or cls in ("smZ CLIPTextEncode", "BNK_CLIPTextEncodeAdvanced"): + if cls.startswith("CLIPTextEncode") or cls in {"smZ CLIPTextEncode", "BNK_CLIPTextEncodeAdvanced"}: return src return None diff --git a/skills/creative/comfyui/scripts/fetch_logs.py b/skills/creative/comfyui/scripts/fetch_logs.py index c7b3b084807..e0b6e12ac75 100755 --- a/skills/creative/comfyui/scripts/fetch_logs.py +++ b/skills/creative/comfyui/scripts/fetch_logs.py @@ -151,7 +151,7 @@ def main(argv: list[str] | None = None) -> int: diag["source"] = res.get("source") diag["prompt_id"] = args.prompt_id emit_json(diag) - return 0 if diag.get("status_str") not in ("error",) else 1 + return 0 if diag.get("status_str") not in {"error",} else 1 if __name__ == "__main__": diff --git a/skills/creative/comfyui/scripts/hardware_check.py b/skills/creative/comfyui/scripts/hardware_check.py index 6a4d6c6d406..083d018acc6 100755 --- a/skills/creative/comfyui/scripts/hardware_check.py +++ b/skills/creative/comfyui/scripts/hardware_check.py @@ -203,7 +203,7 @@ def detect_apple_silicon() -> dict | None: def detect_intel_arc() -> dict | None: - if platform.system() not in ("Linux", "Windows"): + if platform.system() not in {"Linux", "Windows"}: return None if shutil.which("clinfo"): out = _run(["clinfo", "--list"]) diff --git a/skills/creative/comfyui/scripts/run_workflow.py b/skills/creative/comfyui/scripts/run_workflow.py index 444957960b6..05afb1e319f 100755 --- a/skills/creative/comfyui/scripts/run_workflow.py +++ b/skills/creative/comfyui/scripts/run_workflow.py @@ -204,7 +204,7 @@ class ComfyRunner: s = data.get("status") if s == "completed": return {"status": "success", "data": data} - if s in ("failed",): + if s in {"failed",}: return {"status": "error", "data": data} if s == "cancelled": return {"status": "cancelled", "data": data} @@ -386,7 +386,7 @@ class ComfyRunner: # local path; otherwise put the file in output_dir flat. target_parts: list[str] = [] if preserve_subfolder and subfolder: - target_parts.extend(p for p in subfolder.split("/") if p and p not in (".", "..")) + target_parts.extend(p for p in subfolder.split("/") if p and p not in {".", ".."}) target_parts.append(filename) out_path = safe_path_join(output_dir, *target_parts) @@ -467,7 +467,7 @@ def inject_params( # Auto-randomize seed when it's -1 in args, or when randomize_seed_if_unset # and user didn't pass a seed. if "seed" in params: - if "seed" in args and args["seed"] in (None, -1, "-1"): + if "seed" in args and args["seed"] in {None, -1, "-1"}: args = dict(args) args["seed"] = coerce_seed(args["seed"]) warnings.append(f"seed=-1 expanded to {args['seed']}") diff --git a/skills/creative/comfyui/scripts/ws_monitor.py b/skills/creative/comfyui/scripts/ws_monitor.py index b8689655bd0..e2b6689423a 100755 --- a/skills/creative/comfyui/scripts/ws_monitor.py +++ b/skills/creative/comfyui/scripts/ws_monitor.py @@ -170,7 +170,7 @@ def main(argv: list[str] | None = None) -> int: parsed = parse_binary_frame(msg) if parsed is None: continue - if parsed["kind"] in ("preview", "preview_with_metadata") and preview_dir: + if parsed["kind"] in {"preview", "preview_with_metadata"} and preview_dir: img_bytes = parsed.get("image_bytes", b"") if img_bytes: ext = parsed.get("ext", "png") diff --git a/skills/creative/comfyui/tests/test_cloud_integration.py b/skills/creative/comfyui/tests/test_cloud_integration.py index eb7b04ca225..0ce88efe3c2 100644 --- a/skills/creative/comfyui/tests/test_cloud_integration.py +++ b/skills/creative/comfyui/tests/test_cloud_integration.py @@ -53,7 +53,7 @@ class TestCloudEndpointsLive: url = resolve_url("https://cloud.comfy.org", "/object_info") r = http_get(url, headers={"X-API-Key": cloud_key}) # Should be either 200 (paid) or 403 (free) — not 404 / 500 - assert r.status in (200, 403) + assert r.status in {200, 403} if r.status == 403: # Body should mention the limitation assert "free tier" in r.text().lower() or "subscription" in r.text().lower() diff --git a/skills/creative/comfyui/tests/test_extract_schema.py b/skills/creative/comfyui/tests/test_extract_schema.py index 1cb965a1fa8..072a788f318 100644 --- a/skills/creative/comfyui/tests/test_extract_schema.py +++ b/skills/creative/comfyui/tests/test_extract_schema.py @@ -40,7 +40,7 @@ class TestConnectionTracing: } # Should hit max_hops without infinite loop result = trace_to_node(wf, ["1", 0], max_hops=5) - assert result in ("1", "2") # any node, just don't hang + assert result in {"1", "2"} # any node, just don't hang class TestPositiveNegativeDetection: diff --git a/skills/productivity/google-workspace/scripts/google_api.py b/skills/productivity/google-workspace/scripts/google_api.py index 7b8350ab34a..231b1b6849f 100644 --- a/skills/productivity/google-workspace/scripts/google_api.py +++ b/skills/productivity/google-workspace/scripts/google_api.py @@ -721,7 +721,7 @@ def drive_share(args): "type": args.type, "role": args.role, } - if args.type in ("user", "group"): + if args.type in {"user", "group"}: if not args.email: print("ERROR: --email is required for type=user or type=group", file=sys.stderr) sys.exit(1) diff --git a/skills/productivity/maps/scripts/maps_client.py b/skills/productivity/maps/scripts/maps_client.py index 279a41aad64..d272b4a7566 100644 --- a/skills/productivity/maps/scripts/maps_client.py +++ b/skills/productivity/maps/scripts/maps_client.py @@ -181,7 +181,7 @@ def http_get(url, params=None, retries=MAX_RETRIES, silent=False): return json.loads(raw) except urllib.error.HTTPError as exc: last_error = f"HTTP {exc.code}: {exc.reason} for {url}" - if exc.code in (429, 503, 502, 504): + if exc.code in {429, 503, 502, 504}: time.sleep(RETRY_DELAY * attempt) else: if silent: @@ -217,7 +217,7 @@ def http_get_text(url, params=None, retries=MAX_RETRIES, silent=False): return resp.read().decode("utf-8") except urllib.error.HTTPError as exc: last_error = f"HTTP {exc.code}: {exc.reason} for {url}" - if exc.code in (429, 503, 502, 504): + if exc.code in {429, 503, 502, 504}: time.sleep(RETRY_DELAY * attempt) else: if silent: @@ -256,7 +256,7 @@ def http_post(url, data_str, retries=MAX_RETRIES): return json.loads(raw) except urllib.error.HTTPError as exc: last_error = f"HTTP {exc.code}: {exc.reason}" - if exc.code in (429, 503, 502, 504): + if exc.code in {429, 503, 502, 504}: time.sleep(RETRY_DELAY * attempt) else: error_exit(last_error) @@ -459,8 +459,8 @@ def parse_overpass_elements(elements, ref_lat=None, ref_lon=None): "maps_url": f"https://www.google.com/maps/search/?api=1&query={el_lat},{el_lon}", "tags": { k: v for k, v in tags.items() - if k not in ("name", "name:en", - "addr:housenumber", "addr:street", "addr:city") + if k not in {"name", "name:en", + "addr:housenumber", "addr:street", "addr:city"} }, } diff --git a/skills/productivity/ocr-and-documents/scripts/extract_marker.py b/skills/productivity/ocr-and-documents/scripts/extract_marker.py index 4f301aac7b2..d48fd10bb02 100644 --- a/skills/productivity/ocr-and-documents/scripts/extract_marker.py +++ b/skills/productivity/ocr-and-documents/scripts/extract_marker.py @@ -63,7 +63,7 @@ def check_requirements(): if __name__ == "__main__": args = sys.argv[1:] - if not args or args[0] in ("-h", "--help"): + if not args or args[0] in {"-h", "--help"}: print(__doc__) sys.exit(0) diff --git a/skills/productivity/ocr-and-documents/scripts/extract_pymupdf.py b/skills/productivity/ocr-and-documents/scripts/extract_pymupdf.py index 22063e73489..50cb8ee86c4 100644 --- a/skills/productivity/ocr-and-documents/scripts/extract_pymupdf.py +++ b/skills/productivity/ocr-and-documents/scripts/extract_pymupdf.py @@ -68,7 +68,7 @@ def show_metadata(path): if __name__ == "__main__": args = sys.argv[1:] - if not args or args[0] in ("-h", "--help"): + if not args or args[0] in {"-h", "--help"}: print(__doc__) sys.exit(0) diff --git a/skills/research/arxiv/scripts/search_arxiv.py b/skills/research/arxiv/scripts/search_arxiv.py index 9acd8b97ec9..0bd6b2370f4 100644 --- a/skills/research/arxiv/scripts/search_arxiv.py +++ b/skills/research/arxiv/scripts/search_arxiv.py @@ -81,7 +81,7 @@ def search(query=None, author=None, category=None, ids=None, max_results=5, sort if __name__ == "__main__": args = sys.argv[1:] - if not args or args[0] in ("-h", "--help"): + if not args or args[0] in {"-h", "--help"}: print(__doc__) sys.exit(0) diff --git a/skills/research/polymarket/scripts/polymarket.py b/skills/research/polymarket/scripts/polymarket.py index 417e0b1747e..b76e7aa5f9b 100644 --- a/skills/research/polymarket/scripts/polymarket.py +++ b/skills/research/polymarket/scripts/polymarket.py @@ -233,7 +233,7 @@ def cmd_trades(limit: int = 10, market: str = None): def main(): args = sys.argv[1:] - if not args or args[0] in ("-h", "--help", "help"): + if not args or args[0] in {"-h", "--help", "help"}: print(__doc__) return diff --git a/tests/acp/test_tools.py b/tests/acp/test_tools.py index f9b0dac6d66..dc62b296c69 100644 --- a/tests/acp/test_tools.py +++ b/tests/acp/test_tools.py @@ -207,6 +207,16 @@ class TestBuildToolStart: assert result.content is None assert result.raw_input is None + def test_build_tool_start_for_browser_navigate(self): + """browser_navigate should emit a polished start event.""" + args = {"url": "https://x.com"} + result = build_tool_start("tc-browser-start", "browser_navigate", args) + assert isinstance(result, ToolCallStart) + assert result.title == "navigate: https://x.com" + assert result.kind == "fetch" + assert result.content[0].content.text == '{\n "url": "https://x.com"\n}' + assert result.raw_input is None + def test_build_tool_start_for_search(self): """search_files should include pattern in content.""" args = {"pattern": "TODO", "target": "content"} diff --git a/tests/agent/lsp/_mock_lsp_server.py b/tests/agent/lsp/_mock_lsp_server.py index 0220fec195d..619b8da233f 100644 --- a/tests/agent/lsp/_mock_lsp_server.py +++ b/tests/agent/lsp/_mock_lsp_server.py @@ -91,7 +91,7 @@ def main(): if msg.get("method") == "workspace/didChangeWatchedFiles": continue - if msg.get("method") in ("textDocument/didOpen", "textDocument/didChange"): + if msg.get("method") in {"textDocument/didOpen", "textDocument/didChange"}: params = msg.get("params") or {} td = params.get("textDocument") or {} uri = td.get("uri", "") diff --git a/tests/agent/lsp/test_install_and_lint_fixes.py b/tests/agent/lsp/test_install_and_lint_fixes.py index 9046d01295e..e9f862a6d8e 100644 --- a/tests/agent/lsp/test_install_and_lint_fixes.py +++ b/tests/agent/lsp/test_install_and_lint_fixes.py @@ -87,10 +87,10 @@ def test_install_npm_works_without_extras(tmp_path, monkeypatch): cmd = captured["cmd"] assert "pyright" in cmd # Should not blow up when extra_pkgs is omitted/None - install_targets = [c for c in cmd if not c.startswith("-") and c not in ( + install_targets = [c for c in cmd if not c.startswith("-") and c not in { "install", "--prefix", str(install_mod.hermes_lsp_bin_dir().parent), "/usr/bin/npm", - )] + }] assert install_targets == ["pyright"] diff --git a/tests/agent/test_anthropic_adapter.py b/tests/agent/test_anthropic_adapter.py index 0ba2ba29f51..c7119dfd3b0 100644 --- a/tests/agent/test_anthropic_adapter.py +++ b/tests/agent/test_anthropic_adapter.py @@ -157,6 +157,13 @@ class TestBuildAnthropicClient: class TestReadClaudeCodeCredentials: + @pytest.fixture(autouse=True) + def no_keychain(self, monkeypatch): + monkeypatch.setattr( + "agent.anthropic_adapter._read_claude_code_credentials_from_keychain", + lambda: None, + ) + def test_reads_valid_credentials(self, tmp_path, monkeypatch): cred_file = tmp_path / ".claude" / ".credentials.json" cred_file.parent.mkdir(parents=True) @@ -1651,7 +1658,7 @@ class TestThinkingBlockSignatureManagement: _, result = convert_messages_to_anthropic(messages) assistant = next(m for m in result if m["role"] == "assistant") for block in assistant["content"]: - if block.get("type") in ("thinking", "redacted_thinking"): + if block.get("type") in {"thinking", "redacted_thinking"}: assert "cache_control" not in block def test_thinking_stripped_from_merged_consecutive_assistants(self): @@ -1741,7 +1748,7 @@ class TestThinkingBlockSignatureManagement: # First two: no thinking blocks for a in assistants[:2]: assert not any( - b.get("type") in ("thinking", "redacted_thinking") + b.get("type") in {"thinking", "redacted_thinking"} for b in a["content"] if isinstance(b, dict) ) diff --git a/tests/agent/test_auxiliary_client.py b/tests/agent/test_auxiliary_client.py index 96f5802f839..2522fa16197 100644 --- a/tests/agent/test_auxiliary_client.py +++ b/tests/agent/test_auxiliary_client.py @@ -673,6 +673,8 @@ class TestGetTextAuxiliaryClient: def test_custom_endpoint_uses_codex_wrapper_when_runtime_requests_responses_api(self): with patch("agent.auxiliary_client._resolve_custom_runtime", return_value=("https://api.openai.com/v1", "sk-test", "codex_responses")), \ + patch("agent.auxiliary_client._read_nous_auth", return_value=None), \ + patch("agent.auxiliary_client._resolve_nous_runtime_api", return_value=None), \ patch("agent.auxiliary_client._read_main_model", return_value="gpt-5.3-codex"), \ patch("agent.auxiliary_client.OpenAI") as mock_openai: client, model = get_text_auxiliary_client() @@ -923,6 +925,44 @@ class TestIsPaymentError: exc = Exception("connection reset") assert _is_payment_error(exc) is False + # ── Daily / monthly quota exhaustion (#26803) ──────────────────────────── + + def test_429_quota_exceeded(self): + """Cloud provider quota exhaustion (e.g. Vertex AI) is a payment error.""" + exc = Exception("RESOURCE_EXHAUSTED: quota exceeded for project") + exc.status_code = 429 + assert _is_payment_error(exc) is True + + def test_429_too_many_tokens_per_day(self): + """Bedrock / LiteLLM daily token limit is a payment error.""" + exc = Exception("Too many tokens per day: 1000000 used, 1000000 limit") + exc.status_code = 429 + assert _is_payment_error(exc) is True + + def test_429_daily_limit_phrase(self): + """Generic 'daily limit' phrasing is a payment error.""" + exc = Exception("You have exceeded your daily limit.") + exc.status_code = 429 + assert _is_payment_error(exc) is True + + def test_429_resource_exhausted_grpc(self): + """Vertex AI gRPC RESOURCE_EXHAUSTED maps to payment error.""" + exc = Exception("resource exhausted") + exc.status_code = 429 + assert _is_payment_error(exc) is True + + def test_429_daily_quota_phrase(self): + """'daily quota' phrasing is a payment error.""" + exc = Exception("Daily quota of 500 requests reached.") + exc.status_code = 429 + assert _is_payment_error(exc) is True + + def test_429_transient_rate_limit_not_quota(self): + """Transient 429 rate limit without quota keywords is NOT a payment error.""" + exc = Exception("Rate limit exceeded. Retry after 10s.") + exc.status_code = 429 + assert _is_payment_error(exc) is False + class TestIsRateLimitError: """_is_rate_limit_error detects 429 rate-limit errors warranting fallback.""" @@ -1111,6 +1151,140 @@ class TestCallLlmPaymentFallback: # Fallback client should have been used assert fallback_client.chat.completions.create.called + +class TestAuxiliaryFallbackLayering: + """Explicit-provider users get layered fallback: configured_chain → main agent → warn.""" + + def _make_payment_err(self): + exc = Exception("Payment Required: insufficient credits") + exc.status_code = 402 + return exc + + def test_explicit_provider_uses_configured_chain_first(self, monkeypatch, caplog): + """When a user has fallback_chain configured, it's tried BEFORE the main agent model.""" + monkeypatch.setenv("OPENROUTER_API_KEY", "or-key") + + primary_client = MagicMock() + primary_client.chat.completions.create.side_effect = self._make_payment_err() + + chain_client = MagicMock() + chain_client.chat.completions.create.return_value = MagicMock(choices=[ + MagicMock(message=MagicMock(content="from configured chain")) + ]) + + main_called = MagicMock() + + with patch("agent.auxiliary_client._get_cached_client", + return_value=(primary_client, "glm-4v-flash")), \ + patch("agent.auxiliary_client._resolve_task_provider_model", + return_value=("glm", "glm-4v-flash", None, None, None)), \ + patch("agent.auxiliary_client._try_configured_fallback_chain", + return_value=(chain_client, "gpt-4o-mini", "fallback_chain[0](openai)")), \ + patch("agent.auxiliary_client._try_main_agent_model_fallback", + side_effect=main_called): + result = call_llm( + task="vision", + messages=[{"role": "user", "content": "hello"}], + ) + + assert chain_client.chat.completions.create.called + # Main agent fallback should NOT have been consulted — chain succeeded first + main_called.assert_not_called() + + def test_explicit_provider_falls_back_to_main_when_chain_exhausted(self, monkeypatch): + """If configured fallback_chain returns nothing, main agent model is tried next.""" + monkeypatch.setenv("OPENROUTER_API_KEY", "or-key") + + primary_client = MagicMock() + primary_client.chat.completions.create.side_effect = self._make_payment_err() + + main_client = MagicMock() + main_client.chat.completions.create.return_value = MagicMock(choices=[ + MagicMock(message=MagicMock(content="from main agent")) + ]) + + with patch("agent.auxiliary_client._get_cached_client", + return_value=(primary_client, "glm-4v-flash")), \ + patch("agent.auxiliary_client._resolve_task_provider_model", + return_value=("glm", "glm-4v-flash", None, None, None)), \ + patch("agent.auxiliary_client._try_configured_fallback_chain", + return_value=(None, None, "")), \ + patch("agent.auxiliary_client._try_main_agent_model_fallback", + return_value=(main_client, "claude-sonnet-4", "main-agent(openrouter)")): + result = call_llm( + task="vision", + messages=[{"role": "user", "content": "hello"}], + ) + + assert main_client.chat.completions.create.called + + def test_warning_emitted_when_all_fallbacks_exhausted(self, monkeypatch, caplog): + """When chain AND main model both fail, a user-visible warning fires before re-raise.""" + monkeypatch.setenv("OPENROUTER_API_KEY", "or-key") + + primary_client = MagicMock() + primary_client.chat.completions.create.side_effect = self._make_payment_err() + + with patch("agent.auxiliary_client._get_cached_client", + return_value=(primary_client, "glm-4v-flash")), \ + patch("agent.auxiliary_client._resolve_task_provider_model", + return_value=("glm", "glm-4v-flash", None, None, None)), \ + patch("agent.auxiliary_client._try_configured_fallback_chain", + return_value=(None, None, "")), \ + patch("agent.auxiliary_client._try_main_agent_model_fallback", + return_value=(None, None, "")), \ + caplog.at_level("WARNING", logger="agent.auxiliary_client"): + with pytest.raises(Exception, match="Payment Required"): + call_llm( + task="vision", + messages=[{"role": "user", "content": "hello"}], + ) + + assert any( + "all fallbacks exhausted" in r.message for r in caplog.records + ), f"Expected exhaustion warning, got: {[r.message for r in caplog.records]}" + + +class TestTryMainAgentModelFallback: + """_try_main_agent_model_fallback resolves the user's main provider+model as a safety net.""" + + def test_returns_none_when_main_provider_is_auto(self): + from agent.auxiliary_client import _try_main_agent_model_fallback + with patch("agent.auxiliary_client._read_main_provider", return_value="auto"), \ + patch("agent.auxiliary_client._read_main_model", return_value="some-model"): + client, model, label = _try_main_agent_model_fallback("glm", task="vision") + assert client is None and model is None and label == "" + + def test_returns_none_when_failed_provider_equals_main(self): + """If the thing that failed IS the main model, no point retrying it.""" + from agent.auxiliary_client import _try_main_agent_model_fallback + with patch("agent.auxiliary_client._read_main_provider", return_value="openrouter"), \ + patch("agent.auxiliary_client._read_main_model", return_value="anthropic/claude-sonnet-4"): + client, model, label = _try_main_agent_model_fallback("openrouter", task="vision") + assert client is None and label == "" + + def test_resolves_main_provider_client(self): + from agent.auxiliary_client import _try_main_agent_model_fallback + fake_client = MagicMock() + with patch("agent.auxiliary_client._read_main_provider", return_value="openrouter"), \ + patch("agent.auxiliary_client._read_main_model", return_value="anthropic/claude-sonnet-4"), \ + patch("agent.auxiliary_client._is_provider_unhealthy", return_value=False), \ + patch("agent.auxiliary_client.resolve_provider_client", + return_value=(fake_client, "anthropic/claude-sonnet-4")): + client, model, label = _try_main_agent_model_fallback("glm", task="vision") + assert client is fake_client + assert model == "anthropic/claude-sonnet-4" + assert label == "main-agent(openrouter)" + + def test_skips_when_main_provider_is_unhealthy(self): + from agent.auxiliary_client import _try_main_agent_model_fallback + with patch("agent.auxiliary_client._read_main_provider", return_value="openrouter"), \ + patch("agent.auxiliary_client._read_main_model", return_value="anthropic/claude-sonnet-4"), \ + patch("agent.auxiliary_client._is_provider_unhealthy", return_value=True): + client, model, label = _try_main_agent_model_fallback("glm", task="vision") + assert client is None + + # --------------------------------------------------------------------------- # Gate: _resolve_api_key_provider must skip anthropic when not configured # --------------------------------------------------------------------------- @@ -2349,10 +2523,13 @@ class TestAuxiliaryClientPoisonedCacheEviction: def test_call_llm_evicts_on_connection_error_with_explicit_provider(self): """Connection error on an explicit provider must drop the cached client. - This is the exact reporter scenario: ``auxiliary.compression.provider: - main`` (resolves to ``openai-codex``) → no fallback chain runs (not - auto), but the cached client was poisoned by a prior timeout and must - be evicted so the next call rebuilds. + Reporter scenario: ``auxiliary.compression.provider: main`` (resolves + to ``openai-codex``). After #26803, capacity errors (payment/quota/ + connection) DO trigger fallback even on explicit providers — so we + also stub ``_try_payment_fallback`` to ``(None, None, "")`` so the + connection error re-raises after eviction instead of escaping into + a real network call. The contract under test is cache eviction, + not the fallback gate. """ from agent.auxiliary_client import _client_cache, _client_cache_lock @@ -2372,6 +2549,9 @@ class TestAuxiliaryClientPoisonedCacheEviction: ), patch( "agent.auxiliary_client._get_cached_client", return_value=(poisoned, "gpt-5.5"), + ), patch( + "agent.auxiliary_client._try_payment_fallback", + return_value=(None, None, ""), ): with pytest.raises(ConnectionError): call_llm( @@ -2405,6 +2585,9 @@ class TestAuxiliaryClientPoisonedCacheEviction: ), patch( "agent.auxiliary_client._get_cached_client", return_value=(poisoned, "gpt-5.5"), + ), patch( + "agent.auxiliary_client._try_payment_fallback", + return_value=(None, None, ""), ): with pytest.raises(ConnectionError): await async_call_llm( diff --git a/tests/agent/test_auxiliary_main_first.py b/tests/agent/test_auxiliary_main_first.py index 6ac69b27b7c..d1b758c2884 100644 --- a/tests/agent/test_auxiliary_main_first.py +++ b/tests/agent/test_auxiliary_main_first.py @@ -371,7 +371,7 @@ class TestResolveVisionMainFirst: provider, client, model = resolve_vision_provider_client() assert client is fallback_client - assert provider in ("openrouter", "nous") + assert provider in {"openrouter", "nous"} def test_explicit_provider_override_still_wins(self): """Explicit config override bypasses main-first policy.""" diff --git a/tests/agent/test_context_compressor.py b/tests/agent/test_context_compressor.py index 559cf2237a2..2d1a40445d7 100644 --- a/tests/agent/test_context_compressor.py +++ b/tests/agent/test_context_compressor.py @@ -1046,7 +1046,7 @@ class TestCompressWithClient: for i in range(1, len(result)): r1 = result[i - 1].get("role") r2 = result[i].get("role") - if r1 in ("user", "assistant") and r2 in ("user", "assistant"): + if r1 in {"user", "assistant"} and r2 in {"user", "assistant"}: assert r1 != r2, f"consecutive {r1} at indices {i-1},{i}" def test_double_collision_merges_summary_into_tail(self): @@ -1087,7 +1087,7 @@ class TestCompressWithClient: for i in range(1, len(result)): r1 = result[i - 1].get("role") r2 = result[i].get("role") - if r1 in ("user", "assistant") and r2 in ("user", "assistant"): + if r1 in {"user", "assistant"} and r2 in {"user", "assistant"}: assert r1 != r2, f"consecutive {r1} at indices {i-1},{i}" # The summary text should be merged into the first tail message @@ -1164,7 +1164,7 @@ class TestCompressWithClient: for i in range(1, len(result)): r1 = result[i - 1].get("role") r2 = result[i].get("role") - if r1 in ("user", "assistant") and r2 in ("user", "assistant"): + if r1 in {"user", "assistant"} and r2 in {"user", "assistant"}: assert r1 != r2, f"consecutive {r1} at indices {i-1},{i}" # The summary should be merged into the first tail message (assistant at index 5) diff --git a/tests/agent/test_credential_pool.py b/tests/agent/test_credential_pool.py index 299567a9a6f..c288619aedf 100644 --- a/tests/agent/test_credential_pool.py +++ b/tests/agent/test_credential_pool.py @@ -2,8 +2,10 @@ from __future__ import annotations +import base64 import json import time +from datetime import datetime, timezone import pytest @@ -14,6 +16,14 @@ def _write_auth_store(tmp_path, payload: dict) -> None: (hermes_home / "auth.json").write_text(json.dumps(payload, indent=2)) +def _jwt_with_claims(claims: dict) -> str: + def _part(payload: dict) -> str: + raw = json.dumps(payload, separators=(",", ":")).encode("utf-8") + return base64.urlsafe_b64encode(raw).decode("ascii").rstrip("=") + + return f"{_part({'alg': 'none', 'typ': 'JWT'})}.{_part(claims)}.sig" + + def test_fill_first_selection_skips_recently_exhausted_entry(tmp_path, monkeypatch): monkeypatch.setenv("HERMES_HOME", str(tmp_path / "hermes")) _write_auth_store( @@ -510,6 +520,180 @@ def test_load_pool_migrates_nous_provider_state(tmp_path, monkeypatch): assert entry.agent_key == "agent-key" +def test_load_pool_mirrors_nous_invoke_jwt_agent_key_runtime_api_key(tmp_path, monkeypatch): + monkeypatch.setenv("HERMES_HOME", str(tmp_path / "hermes")) + expires_at = datetime.fromtimestamp(time.time() + 3600, tz=timezone.utc).isoformat() + token = _jwt_with_claims({ + "sub": "test-user", + "scope": ["inference:invoke", "inference:mint_agent_key"], + "exp": int(time.time() + 3600), + }) + _write_auth_store( + tmp_path, + { + "version": 1, + "active_provider": "nous", + "providers": { + "nous": { + "portal_base_url": "https://portal.example.com", + "inference_base_url": "https://inference.example.com/v1", + "client_id": "hermes-cli", + "token_type": "Bearer", + "scope": "inference:invoke inference:mint_agent_key", + "access_token": token, + "refresh_token": "refresh-token", + "expires_at": expires_at, + "agent_key": token, + "agent_key_expires_at": expires_at, + } + }, + }, + ) + + from agent.credential_pool import load_pool + + pool = load_pool("nous") + entry = pool.select() + + assert entry is not None + assert entry.source == "device_code" + assert entry.agent_key == token + assert entry.runtime_api_key == token + + auth_payload = json.loads((tmp_path / "hermes" / "auth.json").read_text()) + pool_entry = auth_payload["credential_pool"]["nous"][0] + assert pool_entry["agent_key"] == token + assert pool_entry["agent_key_expires_at"] == expires_at + + +def test_nous_pool_terminal_refresh_removes_device_code_entry(tmp_path, monkeypatch): + monkeypatch.setenv("HERMES_HOME", str(tmp_path / "hermes")) + monkeypatch.setenv("HERMES_SHARED_AUTH_DIR", str(tmp_path / "shared")) + _write_auth_store( + tmp_path, + { + "version": 1, + "active_provider": "nous", + "providers": { + "nous": { + "portal_base_url": "https://portal.example.com", + "inference_base_url": "https://inference.example.com/v1", + "client_id": "hermes-cli", + "token_type": "Bearer", + "scope": "inference:mint_agent_key", + "access_token": "access-token", + "refresh_token": "refresh-token", + "expires_at": "2026-03-24T12:00:00+00:00", + "agent_key": "agent-key", + "agent_key_expires_at": "2026-03-24T13:30:00+00:00", + } + }, + }, + ) + + from agent.credential_pool import PooledCredential, load_pool + from hermes_cli import auth as auth_mod + from hermes_cli.auth import AuthError + + refresh_calls = {"count": 0} + + def _terminal_refresh_failure(*_args, **_kwargs): + refresh_calls["count"] += 1 + raise AuthError( + "Refresh session has been revoked", + provider="nous", + code="invalid_grant", + relogin_required=True, + ) + + pool = load_pool("nous") + selected = pool.select() + assert selected is not None + assert selected.source == "device_code" + pool.add_entry(PooledCredential.from_dict("nous", { + "id": "legacy-seeded", + "source": "manual:device_code", + "auth_type": "oauth", + "access_token": "old-access-token", + "refresh_token": "old-refresh-token", + "agent_key": "old-agent-key", + })) + pool.add_entry(PooledCredential.from_dict("nous", { + "id": "manual-key", + "source": "manual", + "auth_type": "api_key", + "access_token": "manual-nous-key", + })) + + monkeypatch.setattr(auth_mod, "resolve_nous_runtime_credentials", _terminal_refresh_failure) + + assert pool.try_refresh_current() is None + + assert [entry.id for entry in pool.entries()] == ["manual-key"] + + auth_payload = json.loads((tmp_path / "hermes" / "auth.json").read_text()) + nous_state = auth_payload["providers"]["nous"] + assert not nous_state.get("refresh_token") + assert not nous_state.get("access_token") + assert not nous_state.get("agent_key") + assert nous_state["last_auth_error"]["code"] == "invalid_grant" + assert [entry["id"] for entry in auth_payload["credential_pool"]["nous"]] == ["manual-key"] + + assert pool.try_refresh_current() is None + assert refresh_calls["count"] == 1 + + +def test_load_pool_removes_nous_device_code_when_singleton_quarantined(tmp_path, monkeypatch): + monkeypatch.setenv("HERMES_HOME", str(tmp_path / "hermes")) + _write_auth_store( + tmp_path, + { + "version": 1, + "active_provider": "nous", + "providers": { + "nous": { + "portal_base_url": "https://portal.example.com", + "inference_base_url": "https://inference.example.com/v1", + "client_id": "hermes-cli", + "last_auth_error": {"code": "invalid_grant"}, + } + }, + "credential_pool": { + "nous": [ + { + "id": "seeded-current", + "source": "device_code", + "auth_type": "oauth", + "access_token": "stale-access", + "refresh_token": "stale-refresh", + "agent_key": "stale-agent", + }, + { + "id": "seeded-legacy", + "source": "manual:device_code", + "auth_type": "oauth", + "access_token": "older-stale-access", + }, + { + "id": "manual-key", + "source": "manual", + "auth_type": "api_key", + "access_token": "manual-nous-key", + }, + ] + }, + }, + ) + + from agent.credential_pool import load_pool + + pool = load_pool("nous") + + assert [entry.id for entry in pool.entries()] == ["manual-key"] + auth_payload = json.loads((tmp_path / "hermes" / "auth.json").read_text()) + assert [entry["id"] for entry in auth_payload["credential_pool"]["nous"]] == ["manual-key"] + + def test_load_pool_removes_stale_file_backed_singleton_entry(tmp_path, monkeypatch): monkeypatch.setenv("HERMES_HOME", str(tmp_path / "hermes")) monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False) diff --git a/tests/agent/test_deepseek_anthropic_thinking.py b/tests/agent/test_deepseek_anthropic_thinking.py index 4d032fa3595..67534adc3e8 100644 --- a/tests/agent/test_deepseek_anthropic_thinking.py +++ b/tests/agent/test_deepseek_anthropic_thinking.py @@ -191,7 +191,7 @@ class TestDeepSeekAnthropicPreservesThinking: if not isinstance(m.get("content"), list): continue for b in m["content"]: - if isinstance(b, dict) and b.get("type") in ("thinking", "redacted_thinking"): + if isinstance(b, dict) and b.get("type") in {"thinking", "redacted_thinking"}: assert "cache_control" not in b def test_openai_compat_deepseek_base_is_not_matched(self) -> None: diff --git a/tests/agent/test_model_metadata.py b/tests/agent/test_model_metadata.py index 7686364dcac..4f2b51293a6 100644 --- a/tests/agent/test_model_metadata.py +++ b/tests/agent/test_model_metadata.py @@ -746,6 +746,16 @@ class TestGetModelContextLength: mock_fetch.return_value = {} assert get_model_context_length("qwen3-coder") == 262144 + @patch("agent.model_metadata.fetch_model_metadata") + def test_qwen3_6_plus_context_length(self, mock_fetch): + """qwen3.6-plus has a 1M context window, not the generic 128K Qwen default.""" + mock_fetch.return_value = {} + assert get_model_context_length("qwen3.6-plus") == 1048576 + # Provider-prefixed variants must resolve to the same explicit entry + # via the longest-substring fallback (no portal/OR cache available). + assert get_model_context_length("qwen/qwen3.6-plus") == 1048576 + assert get_model_context_length("dashscope/qwen3.6-plus") == 1048576 + @patch("agent.model_metadata.fetch_model_metadata") def test_qwen_generic_context_length(self, mock_fetch): """Generic qwen models still get the 128K default.""" diff --git a/tests/agent/test_shell_hooks.py b/tests/agent/test_shell_hooks.py index 088c23eb466..743c9acb843 100644 --- a/tests/agent/test_shell_hooks.py +++ b/tests/agent/test_shell_hooks.py @@ -100,6 +100,30 @@ class TestParseResponse: ) assert r is None + def test_block_action_without_message_uses_default(self): + """Block is honored even when message/reason is absent.""" + r = shell_hooks._parse_response("pre_tool_call", '{"action": "block"}') + assert r == {"action": "block", "message": shell_hooks._DEFAULT_BLOCK_MESSAGE} + + def test_block_decision_without_reason_uses_default(self): + """Block is honored even when reason/message is absent.""" + r = shell_hooks._parse_response("pre_tool_call", '{"decision": "block"}') + assert r == {"action": "block", "message": shell_hooks._DEFAULT_BLOCK_MESSAGE} + + def test_block_action_empty_message_uses_default(self): + """Empty string message falls back to default, not empty string.""" + r = shell_hooks._parse_response( + "pre_tool_call", '{"action": "block", "message": ""}', + ) + assert r == {"action": "block", "message": shell_hooks._DEFAULT_BLOCK_MESSAGE} + + def test_block_action_non_string_message_uses_default(self): + """Non-string message (e.g. integer) falls back to default.""" + r = shell_hooks._parse_response( + "pre_tool_call", '{"action": "block", "message": 42}', + ) + assert r == {"action": "block", "message": shell_hooks._DEFAULT_BLOCK_MESSAGE} + # ── _serialize_payload ──────────────────────────────────────────────────── diff --git a/tests/agent/transports/test_codex_app_server_runtime.py b/tests/agent/transports/test_codex_app_server_runtime.py index d12ac227254..55bbc8bc6d3 100644 --- a/tests/agent/transports/test_codex_app_server_runtime.py +++ b/tests/agent/transports/test_codex_app_server_runtime.py @@ -241,3 +241,58 @@ class TestSpawnEnvIsolation: assert captured["env"].get("CODEX_HOME") == "/tmp/profile/codex" # And HOME still passes through unchanged assert captured["env"].get("HOME") == "/users/alice" + + def test_kanban_worker_adds_only_kanban_writable_root(self, monkeypatch): + """Codex-runtime Kanban workers need to write board state outside + their scratch/worktree workspace, but should not fall back to + danger-full-access. Hermes passes a narrow app-server config override + for the Kanban root only. + """ + import subprocess + from agent.transports import codex_app_server as cas + + captured = {} + + class FakePopen: + def __init__(self, cmd, *args, **kwargs): + captured["cmd"] = list(cmd) + captured["env"] = kwargs.get("env", {}).copy() + self.stdin = None + self.stdout = None + self.stderr = None + self.pid = 1 + self.returncode = None + + def poll(self): + return None + + def terminate(self): + pass + + def wait(self, timeout=None): + return 0 + + def kill(self): + pass + + monkeypatch.setattr(subprocess, "Popen", FakePopen) + monkeypatch.setenv("HOME", "/users/alice") + monkeypatch.setenv("HERMES_HOME", "/users/alice/.hermes/profiles/backend-worker") + monkeypatch.setenv("HERMES_KANBAN_TASK", "t_smoke") + monkeypatch.setenv( + "HERMES_KANBAN_DB", + "/users/alice/.hermes/kanban/boards/smoke/kanban.db", + ) + + client = cas.CodexAppServerClient(codex_bin="codex") + client._closed = True + + cmd = captured["cmd"] + assert cmd[:2] == ["codex", "app-server"] + assert 'sandbox_mode="workspace-write"' in cmd + assert ( + 'sandbox_workspace_write.writable_roots=["/users/alice/.hermes/kanban/boards/smoke"]' + in cmd + ) + assert "sandbox_workspace_write.network_access=false" in cmd + assert all("danger" not in part for part in cmd) diff --git a/tests/agent/transports/test_codex_app_server_session.py b/tests/agent/transports/test_codex_app_server_session.py index f51996dd067..b192d64e1c8 100644 --- a/tests/agent/transports/test_codex_app_server_session.py +++ b/tests/agent/transports/test_codex_app_server_session.py @@ -9,10 +9,12 @@ from __future__ import annotations import threading import time +from unittest.mock import patch from typing import Any, Optional import pytest +import agent.transports.codex_app_server_session as session_mod from agent.transports.codex_app_server_session import ( CodexAppServerSession, TurnResult, @@ -344,6 +346,23 @@ class TestRunTurn: assert r.interrupted is True assert r.error and "timed out" in r.error + def test_deadline_uses_monotonic_clock(self): + client = FakeClient() + s = make_session(client) + monotonic_values = iter([1000.0, 999.0, 999.0, 1001.0]) + with patch.object( + session_mod.time, + "monotonic", + side_effect=lambda: next(monotonic_values), + ): + r = s.run_turn( + "never finishes", + turn_timeout=0.1, + notification_poll_timeout=0.0, + ) + assert r.interrupted is True + assert r.error and "timed out" in r.error + def test_failed_turn_records_error_from_turn_completed(self): client = FakeClient() client.queue_notification( @@ -666,6 +685,35 @@ class TestSessionRetirement: # Confirm we issued turn/interrupt to free codex compute assert any(method == "turn/interrupt" for (method, _) in client.requests) + def test_post_tool_watchdog_uses_monotonic_clock(self): + client = FakeClient() + client.queue_notification( + "item/completed", + item={ + "type": "commandExecution", "id": "ex1", + "command": "echo hi", "cwd": "/tmp", + "status": "completed", "aggregatedOutput": "hi", + "exitCode": 0, "commandActions": [], + }, + threadId="t", turnId="tu1", + ) + s = make_session(client) + monotonic_values = iter([1000.0, 999.0, 999.0, 999.0, 1000.2]) + with patch.object( + session_mod.time, + "monotonic", + side_effect=lambda: next(monotonic_values), + ): + r = s.run_turn( + "tool then silence", + turn_timeout=5.0, + notification_poll_timeout=0.0, + post_tool_quiet_timeout=0.15, + ) + assert r.interrupted is True + assert r.should_retire is True + assert r.error and "silent" in r.error + def test_post_tool_watchdog_resets_on_further_activity(self): """A tool completion followed by an agent message should NOT trip the watchdog — further activity = codex still alive.""" diff --git a/tests/cli/test_cli_init.py b/tests/cli/test_cli_init.py index 8417d64e746..b05df5220c5 100644 --- a/tests/cli/test_cli_init.py +++ b/tests/cli/test_cli_init.py @@ -99,7 +99,7 @@ class TestVerboseAndToolProgress: def test_tool_progress_mode_is_string(self): cli = _make_cli() assert isinstance(cli.tool_progress_mode, str) - assert cli.tool_progress_mode in ("off", "new", "all", "verbose") + assert cli.tool_progress_mode in {"off", "new", "all", "verbose"} class TestBusyInputMode: diff --git a/tests/cli/test_reasoning_command.py b/tests/cli/test_reasoning_command.py index f5f7e35cbe7..5091256a399 100644 --- a/tests/cli/test_reasoning_command.py +++ b/tests/cli/test_reasoning_command.py @@ -70,7 +70,7 @@ class TestHandleReasoningCommand(unittest.TestCase): stub = self._make_cli(show_reasoning=False) # Simulate /reasoning show arg = "show" - if arg in ("show", "on"): + if arg in {"show", "on"}: stub.show_reasoning = True stub.agent.reasoning_callback = lambda x: None self.assertTrue(stub.show_reasoning) @@ -79,7 +79,7 @@ class TestHandleReasoningCommand(unittest.TestCase): stub = self._make_cli(show_reasoning=True) # Simulate /reasoning hide arg = "hide" - if arg in ("hide", "off"): + if arg in {"hide", "off"}: stub.show_reasoning = False stub.agent.reasoning_callback = None self.assertFalse(stub.show_reasoning) @@ -88,14 +88,14 @@ class TestHandleReasoningCommand(unittest.TestCase): def test_on_enables_display(self): stub = self._make_cli(show_reasoning=False) arg = "on" - if arg in ("show", "on"): + if arg in {"show", "on"}: stub.show_reasoning = True self.assertTrue(stub.show_reasoning) def test_off_disables_display(self): stub = self._make_cli(show_reasoning=True) arg = "off" - if arg in ("hide", "off"): + if arg in {"hide", "off"}: stub.show_reasoning = False self.assertFalse(stub.show_reasoning) diff --git a/tests/conftest.py b/tests/conftest.py index aa2b1b1fbcb..176089d5691 100644 --- a/tests/conftest.py +++ b/tests/conftest.py @@ -187,6 +187,7 @@ _HERMES_BEHAVIORAL_VARS = frozenset({ "HERMES_BACKGROUND_NOTIFICATIONS", "HERMES_EXEC_ASK", "HERMES_HOME_MODE", + "HERMES_AGENT_USE_LEGACY_SESSION_KEYS", # Kanban path/board pins must never leak from a developer shell or # dispatched worker into tests; otherwise tests can write fake tasks to # the real ~/.hermes/kanban.db instead of the per-test HERMES_HOME. diff --git a/tests/cron/test_cron_no_agent.py b/tests/cron/test_cron_no_agent.py index 117cb8c7d9a..583cd34099e 100644 --- a/tests/cron/test_cron_no_agent.py +++ b/tests/cron/test_cron_no_agent.py @@ -68,7 +68,7 @@ def test_create_job_no_agent_stores_field(hermes_env): assert job["no_agent"] is True assert job["script"] == "watchdog.sh" # Prompt can be empty/None for no_agent jobs. - assert job["prompt"] in (None, "") + assert job["prompt"] in {None, ""} def test_create_job_default_is_not_no_agent(hermes_env): @@ -148,7 +148,7 @@ def test_cronjob_tool_update_toggles_no_agent(hermes_env): off = json.loads(cronjob(action="update", job_id=job_id, no_agent=False, prompt="run")) assert off["success"] is True - assert off["job"].get("no_agent") in (False, None) + assert off["job"].get("no_agent") in {False, None} on = json.loads(cronjob(action="update", job_id=job_id, no_agent=True)) assert on["success"] is True diff --git a/tests/gateway/conftest.py b/tests/gateway/conftest.py index b6bcc28c506..965933de41b 100644 --- a/tests/gateway/conftest.py +++ b/tests/gateway/conftest.py @@ -269,7 +269,7 @@ def _scan_for_plugin_adapter_antipattern(source: str) -> list[str]: and isinstance(func.value.value, ast.Name) and func.value.value.id == "sys" and func.value.attr == "path" - and func.attr in ("insert", "append", "extend") + and func.attr in {"insert", "append", "extend"} ): target_name = f"sys.path.{func.attr}" diff --git a/tests/gateway/test_allowlist_startup_check.py b/tests/gateway/test_allowlist_startup_check.py index 96441c05213..abb2db7db12 100644 --- a/tests/gateway/test_allowlist_startup_check.py +++ b/tests/gateway/test_allowlist_startup_check.py @@ -16,8 +16,8 @@ def _would_warn(): "MATRIX_ALLOWED_USERS", "DINGTALK_ALLOWED_USERS", "FEISHU_ALLOWED_USERS", "WECOM_ALLOWED_USERS", "GATEWAY_ALLOWED_USERS") ) - _allow_all = os.getenv("GATEWAY_ALLOW_ALL_USERS", "").lower() in ("true", "1", "yes") or any( - os.getenv(v, "").lower() in ("true", "1", "yes") + _allow_all = os.getenv("GATEWAY_ALLOW_ALL_USERS", "").lower() in {"true", "1", "yes"} or any( + os.getenv(v, "").lower() in {"true", "1", "yes"} for v in ("TELEGRAM_ALLOW_ALL_USERS", "DISCORD_ALLOW_ALL_USERS", "WHATSAPP_ALLOW_ALL_USERS", "SLACK_ALLOW_ALL_USERS", "SIGNAL_ALLOW_ALL_USERS", "EMAIL_ALLOW_ALL_USERS", diff --git a/tests/gateway/test_api_server.py b/tests/gateway/test_api_server.py index 032af7109a5..aae5f550532 100644 --- a/tests/gateway/test_api_server.py +++ b/tests/gateway/test_api_server.py @@ -445,7 +445,12 @@ class TestHealthEndpoint: async with TestClient(TestServer(app)) as cli: resp = await cli.get("/health") assert resp.status == 200 + assert resp.headers.get("Content-Security-Policy") == "default-src 'none'; frame-ancestors 'none'" + assert resp.headers.get("Permissions-Policy") == "camera=(), microphone=(), geolocation=()" + assert resp.headers.get("Strict-Transport-Security") == "max-age=31536000; includeSubDomains" assert resp.headers.get("X-Content-Type-Options") == "nosniff" + assert resp.headers.get("X-Frame-Options") == "DENY" + assert resp.headers.get("X-XSS-Protection") == "0" assert resp.headers.get("Referrer-Policy") == "no-referrer" @pytest.mark.asyncio @@ -704,6 +709,37 @@ class TestChatCompletionsEndpoint: assert "[DONE]" in body assert "Hello!" in body + @pytest.mark.asyncio + async def test_stream_string_false_returns_json_completion(self, adapter): + """Quoted false must not route chat completions into SSE mode.""" + mock_result = { + "final_response": "Hello! How can I help you today?", + "messages": [], + "api_calls": 1, + } + + app = _create_app(adapter) + async with TestClient(TestServer(app)) as cli: + with patch.object(adapter, "_run_agent", new_callable=AsyncMock) as mock_run: + mock_run.return_value = ( + mock_result, + {"input_tokens": 10, "output_tokens": 5, "total_tokens": 15}, + ) + resp = await cli.post( + "/v1/chat/completions", + json={ + "model": "hermes-agent", + "messages": [{"role": "user", "content": "Hello"}], + "stream": "false", + }, + ) + + assert resp.status == 200 + assert "text/event-stream" not in resp.headers.get("Content-Type", "") + data = await resp.json() + assert data["object"] == "chat.completion" + assert data["choices"][0]["message"]["content"] == mock_result["final_response"] + @pytest.mark.asyncio async def test_stream_task_done_callback_enqueues_eos_for_chat_completions(self, adapter): """Regression guard for #24451: completion callback must signal SSE EOS.""" @@ -1655,6 +1691,31 @@ class TestResponsesEndpoint: # The response has an ID but it shouldn't be retrievable assert adapter._response_store.get(data["id"]) is None + @pytest.mark.asyncio + async def test_store_string_false_does_not_store(self, adapter): + """Quoted false must preserve ephemeral store=false semantics.""" + mock_result = {"final_response": "OK", "messages": [], "api_calls": 1} + + app = _create_app(adapter) + async with TestClient(TestServer(app)) as cli: + with patch.object(adapter, "_run_agent", new_callable=AsyncMock) as mock_run: + mock_run.return_value = ( + mock_result, + {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0}, + ) + resp = await cli.post( + "/v1/responses", + json={ + "model": "hermes-agent", + "input": "Hello", + "store": "false", + }, + ) + + assert resp.status == 200 + data = await resp.json() + assert adapter._response_store.get(data["id"]) is None + @pytest.mark.asyncio async def test_instructions_inherited_from_previous(self, adapter): """If no instructions provided, carry forward from previous response.""" @@ -1749,6 +1810,37 @@ class TestResponsesStreaming: assert "Hello" in body assert " world" in body + @pytest.mark.asyncio + async def test_stream_string_false_returns_json_response(self, adapter): + """Quoted false must not route Responses API requests into SSE mode.""" + mock_result = { + "final_response": "Paris is the capital of France.", + "messages": [], + "api_calls": 1, + } + + app = _create_app(adapter) + async with TestClient(TestServer(app)) as cli: + with patch.object(adapter, "_run_agent", new_callable=AsyncMock) as mock_run: + mock_run.return_value = ( + mock_result, + {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0}, + ) + resp = await cli.post( + "/v1/responses", + json={ + "model": "hermes-agent", + "input": "What is the capital of France?", + "stream": "false", + }, + ) + + assert resp.status == 200 + assert "text/event-stream" not in resp.headers.get("Content-Type", "") + data = await resp.json() + assert data["object"] == "response" + assert data["output"][0]["content"][0]["text"] == mock_result["final_response"] + @pytest.mark.asyncio async def test_stream_task_done_callback_enqueues_eos_for_responses(self, adapter): """Regression guard for #24451 on /v1/responses streaming path.""" diff --git a/tests/gateway/test_api_server_runs.py b/tests/gateway/test_api_server_runs.py index bdb00d74a7b..8e7169a658d 100644 --- a/tests/gateway/test_api_server_runs.py +++ b/tests/gateway/test_api_server_runs.py @@ -335,6 +335,28 @@ class TestRunEvents: "approval_not_pending", } + @pytest.mark.asyncio + async def test_approval_string_false_does_not_resolve_all(self, adapter): + """Quoted false must not fan out approval resolution across the queue.""" + app = _create_runs_app(adapter) + run_id = "run_bool_parse" + adapter._run_statuses[run_id] = {"run_id": run_id, "status": "running"} + adapter._run_approval_sessions[run_id] = "session-123" + + async with TestClient(TestServer(app)) as cli: + with patch("tools.approval.resolve_gateway_approval", return_value=1) as mock_resolve: + approval_resp = await cli.post( + f"/v1/runs/{run_id}/approval", + json={"choice": "once", "all": "false"}, + ) + + assert approval_resp.status == 200 + mock_resolve.assert_called_once_with( + "session-123", + "once", + resolve_all=False, + ) + @pytest.mark.asyncio async def test_events_not_found_returns_404(self, adapter): app = _create_runs_app(adapter) diff --git a/tests/gateway/test_background_command.py b/tests/gateway/test_background_command.py index 9c156960c70..9e0d71921cd 100644 --- a/tests/gateway/test_background_command.py +++ b/tests/gateway/test_background_command.py @@ -316,6 +316,7 @@ class TestRunBackgroundTask: assert mock_adapter.send.call_args.kwargs["metadata"] == { "thread_id": "20197", "telegram_dm_topic_reply_fallback": True, + "direct_messages_topic_id": "20197", "telegram_reply_to_message_id": "463", } diff --git a/tests/gateway/test_bluebubbles.py b/tests/gateway/test_bluebubbles.py index e3ff26cc695..6f93c1d4dba 100644 --- a/tests/gateway/test_bluebubbles.py +++ b/tests/gateway/test_bluebubbles.py @@ -101,6 +101,11 @@ class TestBlueBubblesHelpers: adapter = _make_adapter(monkeypatch) assert adapter.format_message("**Hello** `world`") == "Hello world" + def test_format_message_preserves_underscores_in_identifiers(self, monkeypatch): + adapter = _make_adapter(monkeypatch) + text = "Use /api_v2 with FEATURE_FLAG_NAME and config_file.json" + assert adapter.format_message(text) == text + def test_strip_markdown_headers(self, monkeypatch): adapter = _make_adapter(monkeypatch) assert adapter.format_message("## Heading\ntext") == "Heading\ntext" diff --git a/tests/gateway/test_config_cwd_bridge.py b/tests/gateway/test_config_cwd_bridge.py index 23666253882..f7349d073f7 100644 --- a/tests/gateway/test_config_cwd_bridge.py +++ b/tests/gateway/test_config_cwd_bridge.py @@ -44,7 +44,7 @@ def _simulate_config_bridge(cfg: dict, initial_env: dict | None = None): val = terminal_cfg[cfg_key] # Skip cwd placeholder values — don't overwrite already-resolved # TERMINAL_CWD. Mirrors the fix in gateway/run.py. - if cfg_key == "cwd" and str(val) in (".", "auto", "cwd"): + if cfg_key == "cwd" and str(val) in {".", "auto", "cwd"}: continue # Expand shell tilde so subprocess.Popen never receives a literal # "~/" which the kernel rejects. @@ -70,7 +70,7 @@ def _simulate_config_bridge(cfg: dict, initial_env: dict | None = None): # --- Replicate lines 144-147: MESSAGING_CWD fallback --- configured_cwd = env.get("TERMINAL_CWD", "") - if not configured_cwd or configured_cwd in (".", "auto", "cwd"): + if not configured_cwd or configured_cwd in {".", "auto", "cwd"}: messaging_cwd = env.get("MESSAGING_CWD") or "/root" # Path.home() for root env["TERMINAL_CWD"] = messaging_cwd diff --git a/tests/gateway/test_discord_system_messages.py b/tests/gateway/test_discord_system_messages.py index 8e2fb27e788..e58f2812745 100644 --- a/tests/gateway/test_discord_system_messages.py +++ b/tests/gateway/test_discord_system_messages.py @@ -48,7 +48,7 @@ class TestDiscordSystemMessageFilter(unittest.TestCase): return False # System message filter (the fix being tested) - if message.type not in (discord.MessageType.default, discord.MessageType.reply): + if message.type not in {discord.MessageType.default, discord.MessageType.reply}: return False return True # message accepted diff --git a/tests/gateway/test_google_chat.py b/tests/gateway/test_google_chat.py index 3f093bcea1d..9d36945a357 100644 --- a/tests/gateway/test_google_chat.py +++ b/tests/gateway/test_google_chat.py @@ -2740,7 +2740,7 @@ class _FakeAiohttpSession: def _install_fake_aiohttp(monkeypatch, session): fake_aiohttp = types.SimpleNamespace( - ClientSession=lambda timeout=None: session, + ClientSession=lambda timeout=None, **kwargs: session, ClientTimeout=lambda total=None: None, ) monkeypatch.setitem(sys.modules, "aiohttp", fake_aiohttp) diff --git a/tests/gateway/test_matrix.py b/tests/gateway/test_matrix.py index c329441531d..a0fb8f086d8 100644 --- a/tests/gateway/test_matrix.py +++ b/tests/gateway/test_matrix.py @@ -2257,6 +2257,210 @@ class TestMatrixOnRoomMessageFilter: ev = self._mk_event(sender="@alice:example.org", body="hello bot") await self.adapter._on_room_message(ev) self.adapter._handle_text_message.assert_awaited_once() + + +class TestMatrixClockSkewWarning: + """Clock-skew detector for #12614. + + Reporter's host clock was set ~2 hours ahead of real time. The grace + filter `event_ts < startup_ts - 5` then drops every live event because + server timestamps look "older than startup". When this happens well + after startup (>30s), the adapter logs a one-shot WARNING pointing the + user at NTP instead of failing silently. + """ + + def setup_method(self): + self.adapter = _make_adapter() + self.adapter._user_id = "@bot:example.org" + self.adapter._handle_text_message = AsyncMock() + self.adapter._handle_media_message = AsyncMock() + + @staticmethod + def _mk_event(sender, ts_ms, event_id=None): + ev = MagicMock() + ev.room_id = "!room:example.org" + ev.sender = sender + ev.event_id = event_id or f"$evt-{sender}-{ts_ms}" + ev.timestamp = ts_ms + ev.server_timestamp = ts_ms + ev.content = {"msgtype": "m.text", "body": "hi"} + return ev + + @pytest.mark.asyncio + async def test_late_drops_emit_one_shot_clock_skew_warning(self, caplog): + import logging + import time as _t + + # Simulate the reporter's environment: host clock is ~2 hours ahead + # of server time. Startup happened "in the future" relative to the + # real-world events we're now receiving. + now = _t.time() + self.adapter._startup_ts = now - 60 # bot started 60s ago (wall clock) + # Server events are dated 2h before startup_ts (skewed clock). + skewed_event_ts_ms = int((self.adapter._startup_ts - 7200) * 1000) + + with caplog.at_level(logging.WARNING, logger="gateway.platforms.matrix"): + for i in range(5): + ev = self._mk_event( + sender=f"@alice{i}:example.org", ts_ms=skewed_event_ts_ms + ) + await self.adapter._on_room_message(ev) + + # Handler should never be invoked — all events failed the grace check. + self.adapter._handle_text_message.assert_not_called() + # Exactly one WARNING from THIS logger should be emitted. Filter by + # logger name so unrelated stdlib/library warnings can't satisfy the + # assertion. + skew_warnings = [ + r for r in caplog.records + if r.name == "gateway.platforms.matrix" + and r.levelname == "WARNING" + and "set-ntp" in r.getMessage() + ] + assert len(skew_warnings) == 1, ( + f"expected exactly 1 clock-skew warning, got {len(skew_warnings)}" + ) + msg = skew_warnings[0].getMessage() + assert "7200" in msg, f"skew value missing from message: {msg!r}" + # Pin the counter so a regression in the gating logic (e.g. warning + # at threshold 1 or 5, or not stopping after warn) is caught. + assert self.adapter._late_grace_drops == 3 + assert self.adapter._clock_skew_warned is True + + @pytest.mark.asyncio + async def test_initial_sync_drops_do_not_warn(self, caplog): + """During the first 30s after startup, old events are normal backfill.""" + import logging + import time as _t + + now = _t.time() + # Startup was 1s ago — we're still in the initial-sync window. + self.adapter._startup_ts = now - 1 + old_ts_ms = int((self.adapter._startup_ts - 3600) * 1000) + + with caplog.at_level(logging.WARNING, logger="gateway.platforms.matrix"): + for i in range(5): + ev = self._mk_event( + sender=f"@alice{i}:example.org", ts_ms=old_ts_ms + ) + await self.adapter._on_room_message(ev) + + # Backfill drops are silent — no clock-skew warning fired. + assert self.adapter._clock_skew_warned is False + skew_warnings = [ + r for r in caplog.records + if r.name == "gateway.platforms.matrix" + and "set-ntp" in r.getMessage() + ] + assert skew_warnings == [] + + @pytest.mark.asyncio + async def test_fewer_than_three_late_drops_do_not_warn(self, caplog): + """A single delayed backfill event after 30s shouldn't trigger NTP advice.""" + import logging + import time as _t + + now = _t.time() + self.adapter._startup_ts = now - 120 # extra slack vs the 30s gate + old_ts_ms = int((self.adapter._startup_ts - 3600) * 1000) + + with caplog.at_level(logging.WARNING, logger="gateway.platforms.matrix"): + for i in range(2): # only 2 late drops — under the threshold + ev = self._mk_event( + sender=f"@alice{i}:example.org", ts_ms=old_ts_ms + ) + await self.adapter._on_room_message(ev) + + assert self.adapter._late_grace_drops == 2 + assert self.adapter._clock_skew_warned is False + + @pytest.mark.asyncio + async def test_varied_backfill_skews_do_not_warn(self, caplog): + """Backfill from a freshly-invited room delivers events of varied age. + + A genuine clock-skew bug produces drops with a *constant* offset + (every event is ~X seconds older than wall clock). Joining an old + room post-startup delivers events spanning hours-to-days; those + skews vary wildly and must NOT trigger the NTP warning. + """ + import logging + import time as _t + + now = _t.time() + self.adapter._startup_ts = now - 120 + # Each event has a different age, ranging from 1h to 30d ago. + ages_in_hours = [1, 24, 168, 720, 4] # 1h, 1d, 1w, 30d, 4h + with caplog.at_level(logging.WARNING, logger="gateway.platforms.matrix"): + for i, hrs in enumerate(ages_in_hours): + ts_ms = int((self.adapter._startup_ts - hrs * 3600) * 1000) + ev = self._mk_event( + sender=f"@alice{i}:example.org", ts_ms=ts_ms + ) + await self.adapter._on_room_message(ev) + + # The varied-skew guard should keep the counter from reaching 3. + assert self.adapter._late_grace_drops < 3 + assert self.adapter._clock_skew_warned is False + skew_warnings = [ + r for r in caplog.records + if r.name == "gateway.platforms.matrix" + and "set-ntp" in r.getMessage() + ] + assert skew_warnings == [] + + @pytest.mark.asyncio + async def test_state_reset_allows_warning_to_fire_again(self, caplog): + """After the reset block at top of connect() runs, the warning is rearmed. + + Reconnect lifecycle: the user fixes NTP, restarts the bot, and the + new connect() call resets _late_grace_drops / _clock_skew_warned at + the top. This test exercises the rearm path by: + 1. Tripping the warning once (state: warned=True). + 2. Running the same reset block connect() runs. + 3. Tripping the warning a second time — the second warning should + fire because the state was cleared. + """ + import logging + import time as _t + + now = _t.time() + self.adapter._startup_ts = now - 60 + skewed_ms = int((self.adapter._startup_ts - 7200) * 1000) + + with caplog.at_level(logging.WARNING, logger="gateway.platforms.matrix"): + for i in range(3): + ev = self._mk_event( + sender=f"@alice{i}:example.org", ts_ms=skewed_ms, + event_id=f"$first-{i}", + ) + await self.adapter._on_room_message(ev) + assert self.adapter._clock_skew_warned is True + + # Mirror the reset block in connect() (matrix.py around line 855). + self.adapter._startup_ts = _t.time() - 60 + self.adapter._late_grace_drops = 0 + self.adapter._late_grace_skew = 0.0 + self.adapter._clock_skew_warned = False + + # Same skewed-clock scenario should warn AGAIN after reset. + skewed_ms2 = int((self.adapter._startup_ts - 7200) * 1000) + for i in range(3): + ev = self._mk_event( + sender=f"@bob{i}:example.org", ts_ms=skewed_ms2, + event_id=f"$second-{i}", + ) + await self.adapter._on_room_message(ev) + + skew_warnings = [ + r for r in caplog.records + if r.name == "gateway.platforms.matrix" + and "set-ntp" in r.getMessage() + ] + assert len(skew_warnings) == 2, ( + f"expected 2 warnings (one per connect cycle), got {len(skew_warnings)}" + ) + + # --------------------------------------------------------------------------- # DM auto-thread # --------------------------------------------------------------------------- diff --git a/tests/gateway/test_platform_connected_checkers.py b/tests/gateway/test_platform_connected_checkers.py index 307c79b3086..941b8c74506 100644 --- a/tests/gateway/test_platform_connected_checkers.py +++ b/tests/gateway/test_platform_connected_checkers.py @@ -76,12 +76,12 @@ def test_checker_returns_true_when_configured(platform, checker, monkeypatch): elif platform == Platform.SMS: monkeypatch.setenv("TWILIO_ACCOUNT_SID", "ACtest") mock_config.extra = {} - elif platform in ( + elif platform in { Platform.API_SERVER, Platform.WEBHOOK, Platform.MSGRAPH_WEBHOOK, Platform.WHATSAPP, - ): + }: mock_config.extra = {} elif platform == Platform.FEISHU: mock_config.extra = {"app_id": "app"} diff --git a/tests/gateway/test_qqbot.py b/tests/gateway/test_qqbot.py index 5d5cac54bd3..4b3402387a4 100644 --- a/tests/gateway/test_qqbot.py +++ b/tests/gateway/test_qqbot.py @@ -1076,7 +1076,7 @@ class TestBuildApprovalKeyboard: parsed = parse_approval_button_data(btn.action.data) assert parsed is not None assert parsed[0] == session_key - assert parsed[1] in ("allow-once", "allow-always", "deny") + assert parsed[1] in {"allow-once", "allow-always", "deny"} class TestBuildUpdatePromptKeyboard: diff --git a/tests/gateway/test_restart_drain.py b/tests/gateway/test_restart_drain.py index 844af427308..9000e4d4820 100644 --- a/tests/gateway/test_restart_drain.py +++ b/tests/gateway/test_restart_drain.py @@ -33,7 +33,16 @@ async def test_restart_command_while_busy_requests_drain_without_interrupt(monke result = await runner._handle_message(event) - assert result == t("gateway.draining", count=1) + expected = t("gateway.draining", count=1) + assert result == expected + # Guard against the silent-degradation regression in #22266: if the i18n + # catalog cannot be resolved (e.g. xdist workers losing the locales path) + # then ``t("gateway.draining", count=1)`` returns the bare key + # ``"gateway.draining"`` instead of the formatted English string, and both + # sides of the equality above would still match. Assert on the catalog + # output explicitly so a broken locale resolution fails loudly here. + assert expected != "gateway.draining" + assert "Draining" in expected and "1" in expected running_agent.interrupt.assert_not_called() runner.request_restart.assert_called_once_with(detached=True, via_service=False) diff --git a/tests/gateway/test_restart_resume_pending.py b/tests/gateway/test_restart_resume_pending.py index 13ef2f6f99e..55d9b4a497b 100644 --- a/tests/gateway/test_restart_resume_pending.py +++ b/tests/gateway/test_restart_resume_pending.py @@ -89,7 +89,7 @@ def _build_agent_history(history: list) -> list: agent_history: list = [] for msg in history: role = msg.get("role") - if not role or role in ("session_meta", "system"): + if not role or role in {"session_meta", "system"}: continue has_tool_calls = "tool_calls" in msg has_tool_call_id = "tool_call_id" in msg diff --git a/tests/gateway/test_session_boundary_hooks.py b/tests/gateway/test_session_boundary_hooks.py index 255795492fc..30584513325 100644 --- a/tests/gateway/test_session_boundary_hooks.py +++ b/tests/gateway/test_session_boundary_hooks.py @@ -108,7 +108,7 @@ async def test_finalize_before_reset(mock_invoke_hook): await runner._handle_reset_command(_make_event("/new")) calls = [c for c in mock_invoke_hook.call_args_list - if c[0][0] in ("on_session_finalize", "on_session_reset")] + if c[0][0] in {"on_session_finalize", "on_session_reset"}] hook_names = [c[0][0] for c in calls] assert hook_names == ["on_session_finalize", "on_session_reset"] diff --git a/tests/gateway/test_session_model_override_routing.py b/tests/gateway/test_session_model_override_routing.py index 3530744e223..26acdc157aa 100644 --- a/tests/gateway/test_session_model_override_routing.py +++ b/tests/gateway/test_session_model_override_routing.py @@ -187,7 +187,7 @@ fallback_providers: monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path) def fake_resolve_runtime_provider(*, requested=None, explicit_base_url=None, explicit_api_key=None): - if requested in (None, "", "openai-codex"): + if requested in {None, "", "openai-codex"}: from hermes_cli.auth import AuthError raise AuthError("No Codex credentials stored. Run `hermes auth` to authenticate.") assert requested == "openrouter" diff --git a/tests/gateway/test_teams.py b/tests/gateway/test_teams.py index 58b8c35a5c2..6c7173fe931 100644 --- a/tests/gateway/test_teams.py +++ b/tests/gateway/test_teams.py @@ -763,7 +763,7 @@ def _install_fake_aiohttp(monkeypatch, session): """Replace ``aiohttp`` in ``sys.modules`` so ``import aiohttp as _aiohttp`` inside ``_standalone_send`` picks up our fake.""" fake_aiohttp = types.SimpleNamespace( - ClientSession=lambda timeout=None: session, + ClientSession=lambda timeout=None, **kwargs: session, ClientTimeout=lambda total=None: None, ) monkeypatch.setitem(sys.modules, "aiohttp", fake_aiohttp) diff --git a/tests/gateway/test_telegram_thread_fallback.py b/tests/gateway/test_telegram_thread_fallback.py index e31753cc2b7..fda1ebc3007 100644 --- a/tests/gateway/test_telegram_thread_fallback.py +++ b/tests/gateway/test_telegram_thread_fallback.py @@ -407,6 +407,7 @@ async def test_gateway_runner_busy_ack_replies_to_triggering_message_for_telegra assert adapter.calls[0]["metadata"] == { "thread_id": "20197", "telegram_dm_topic_reply_fallback": True, + "direct_messages_topic_id": "20197", "telegram_reply_to_message_id": "463", } diff --git a/tests/gateway/test_transcript_offset.py b/tests/gateway/test_transcript_offset.py index d8a2672f4d6..7cbb519ee3a 100644 --- a/tests/gateway/test_transcript_offset.py +++ b/tests/gateway/test_transcript_offset.py @@ -31,7 +31,7 @@ def _filter_history(history: list) -> list: role = msg.get("role") if not role: continue - if role in ("session_meta",): + if role in {"session_meta",}: continue if role == "system": continue diff --git a/tests/gateway/test_update_streaming.py b/tests/gateway/test_update_streaming.py index 932bd1b0579..eb0f0cfa890 100644 --- a/tests/gateway/test_update_streaming.py +++ b/tests/gateway/test_update_streaming.py @@ -237,6 +237,8 @@ class TestUpdateCommandGatewayFlag: cmd_string = call_args[-1] if isinstance(call_args, list) else str(call_args) assert "--gateway" in cmd_string assert "PYTHONUNBUFFERED" in cmd_string + assert "rc=$?" in cmd_string + assert "status=$?" not in cmd_string assert "stream progress" in result diff --git a/tests/gateway/test_voice_command.py b/tests/gateway/test_voice_command.py index a877730dcec..d792a48e0cf 100644 --- a/tests/gateway/test_voice_command.py +++ b/tests/gateway/test_voice_command.py @@ -461,6 +461,7 @@ class TestSendVoiceReply: assert call_kwargs["metadata"] == { "thread_id": "20197", "telegram_dm_topic_reply_fallback": True, + "direct_messages_topic_id": "20197", "telegram_reply_to_message_id": "462", } diff --git a/tests/hermes_cli/test_auth_commands.py b/tests/hermes_cli/test_auth_commands.py index 74e2a64d312..22182ba43a8 100644 --- a/tests/hermes_cli/test_auth_commands.py +++ b/tests/hermes_cli/test_auth_commands.py @@ -107,7 +107,7 @@ def test_auth_add_nous_oauth_persists_pool_entry(tmp_path, monkeypatch): "portal_base_url": "https://portal.example.com", "inference_base_url": "https://inference.example.com/v1", "client_id": "hermes-cli", - "scope": "inference:mint_agent_key", + "scope": "inference:invoke inference:mint_agent_key", "token_type": "Bearer", "access_token": token, "refresh_token": "refresh-token", @@ -228,7 +228,7 @@ def test_auth_add_nous_oauth_honors_custom_label(tmp_path, monkeypatch): "portal_base_url": "https://portal.example.com", "inference_base_url": "https://inference.example.com/v1", "client_id": "hermes-cli", - "scope": "inference:mint_agent_key", + "scope": "inference:invoke inference:mint_agent_key", "token_type": "Bearer", "access_token": token, "refresh_token": "refresh-token", diff --git a/tests/hermes_cli/test_auth_nous_provider.py b/tests/hermes_cli/test_auth_nous_provider.py index bd6098d3746..55903b11816 100644 --- a/tests/hermes_cli/test_auth_nous_provider.py +++ b/tests/hermes_cli/test_auth_nous_provider.py @@ -1,6 +1,9 @@ """Regression tests for Nous OAuth refresh + agent-key mint interactions.""" +import base64 import json +import logging +import time from datetime import datetime, timezone from pathlib import Path @@ -125,6 +128,11 @@ def _setup_nous_auth( *, access_token: str = "access-old", refresh_token: str = "refresh-old", + scope: str = "inference:mint_agent_key", + expires_at: str = "2026-02-01T00:00:00+00:00", + expires_in: int = 0, + agent_key: str | None = None, + agent_key_expires_at: str | None = None, ) -> None: hermes_home.mkdir(parents=True, exist_ok=True) auth_store = { @@ -136,15 +144,15 @@ def _setup_nous_auth( "inference_base_url": "https://inference.example.com/v1", "client_id": "hermes-cli", "token_type": "Bearer", - "scope": "inference:mint_agent_key", + "scope": scope, "access_token": access_token, "refresh_token": refresh_token, "obtained_at": "2026-02-01T00:00:00+00:00", - "expires_in": 0, - "expires_at": "2026-02-01T00:00:00+00:00", - "agent_key": None, + "expires_in": expires_in, + "expires_at": expires_at, + "agent_key": agent_key, "agent_key_id": None, - "agent_key_expires_at": None, + "agent_key_expires_at": agent_key_expires_at, "agent_key_expires_in": None, "agent_key_reused": None, "agent_key_obtained_at": None, @@ -164,6 +172,463 @@ def _mint_payload(api_key: str = "agent-key") -> dict: } +def _jwt_with_claims(claims: dict) -> str: + def _part(payload: dict) -> str: + raw = json.dumps(payload, separators=(",", ":")).encode("utf-8") + return base64.urlsafe_b64encode(raw).decode("ascii").rstrip("=") + + return f"{_part({'alg': 'none', 'typ': 'JWT'})}.{_part(claims)}.sig" + + +def _future_iso(seconds: int = 3600) -> str: + return datetime.fromtimestamp(time.time() + seconds, tz=timezone.utc).isoformat() + + +def _invoke_jwt(*, seconds: int = 3600, scope: object = "inference:invoke inference:mint_agent_key") -> str: + return _jwt_with_claims({ + "sub": "test-user", + "scope": scope, + "exp": int(time.time() + seconds), + }) + + +def test_resolve_nous_runtime_credentials_prefers_invoke_jwt_and_mirrors( + tmp_path, + monkeypatch, +): + import hermes_cli.auth as auth_mod + + hermes_home = tmp_path / "hermes" + token = _invoke_jwt(seconds=3600) + _setup_nous_auth( + hermes_home, + access_token=token, + scope=auth_mod.DEFAULT_NOUS_SCOPE, + expires_at=_future_iso(3600), + expires_in=3600, + ) + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + + def _unexpected_mint(*args, **kwargs): + raise AssertionError("legacy agent-key mint should not run for invoke JWT") + + monkeypatch.setattr(auth_mod, "_mint_agent_key", _unexpected_mint) + + creds = auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=300) + + assert creds["api_key"] == token + assert creds["source"] == auth_mod.NOUS_AUTH_PATH_INVOKE_JWT + assert creds["auth_path"] == auth_mod.NOUS_AUTH_PATH_INVOKE_JWT + + payload = json.loads((hermes_home / "auth.json").read_text()) + singleton = payload["providers"]["nous"] + assert singleton["agent_key"] == token + assert datetime.fromisoformat(singleton["agent_key_expires_at"]).timestamp() > time.time() + 300 + + pool_entries = payload["credential_pool"]["nous"] + assert len(pool_entries) == 1 + assert pool_entries[0]["agent_key"] == token + assert pool_entries[0]["source"] == auth_mod.NOUS_DEVICE_CODE_SOURCE + + +def test_resolve_nous_runtime_credentials_invoke_jwt_is_idempotent( + tmp_path, + monkeypatch, +): + import hermes_cli.auth as auth_mod + + hermes_home = tmp_path / "hermes" + hermes_home.mkdir(parents=True, exist_ok=True) + exp = int(time.time() + 3600) + expires_at = datetime.fromtimestamp(exp, tz=timezone.utc).isoformat() + token = _jwt_with_claims({ + "sub": "test-user", + "scope": auth_mod.DEFAULT_NOUS_SCOPE, + "exp": exp, + }) + original_obtained_at = "2026-04-17T22:00:10+00:00" + auth_store = { + "version": 1, + "active_provider": "nous", + "providers": { + "nous": { + "portal_base_url": "https://portal.example.com", + "inference_base_url": "https://inference.example.com/v1", + "client_id": "hermes-cli", + "token_type": "Bearer", + "scope": auth_mod.DEFAULT_NOUS_SCOPE, + "access_token": token, + "refresh_token": "refresh-token", + "obtained_at": "2026-02-01T00:00:00+00:00", + "expires_in": 123, + "expires_at": expires_at, + "agent_key": token, + "agent_key_id": None, + "agent_key_expires_at": expires_at, + "agent_key_expires_in": 123, + "agent_key_reused": False, + "agent_key_obtained_at": original_obtained_at, + "tls": {"insecure": False, "ca_bundle": None}, + }, + }, + } + auth_path = hermes_home / "auth.json" + auth_path.write_text(json.dumps(auth_store, indent=2)) + before_content = auth_path.read_text() + before_mtime = auth_path.stat().st_mtime_ns + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + + def _unexpected_mint(*args, **kwargs): + raise AssertionError("stable invoke JWT should not mint a legacy key") + + def _unexpected_shared_write(*args, **kwargs): + raise AssertionError("unchanged invoke JWT resolution should not sync shared store") + + sync_calls = [] + + monkeypatch.setattr(auth_mod, "_mint_agent_key", _unexpected_mint) + monkeypatch.setattr(auth_mod, "_write_shared_nous_state", _unexpected_shared_write) + monkeypatch.setattr( + auth_mod, + "_sync_nous_pool_from_auth_store", + lambda: sync_calls.append(True), + ) + + creds = auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=300) + + assert creds["api_key"] == token + assert creds["source"] == auth_mod.NOUS_AUTH_PATH_INVOKE_JWT + assert auth_path.read_text() == before_content + assert auth_path.stat().st_mtime_ns == before_mtime + assert sync_calls == [] + payload = json.loads(auth_path.read_text()) + assert ( + payload["providers"]["nous"]["agent_key_obtained_at"] + == original_obtained_at + ) + + +def test_resolve_nous_runtime_credentials_trusts_invoke_jwt_exp_over_stale_metadata( + tmp_path, + monkeypatch, +): + import hermes_cli.auth as auth_mod + + hermes_home = tmp_path / "hermes" + token = _invoke_jwt(seconds=3600) + _setup_nous_auth( + hermes_home, + access_token=token, + scope=auth_mod.DEFAULT_NOUS_SCOPE, + expires_at="2000-01-01T00:00:00+00:00", + expires_in=0, + agent_key=token, + agent_key_expires_at="2000-01-01T00:00:00+00:00", + ) + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + + def _unexpected_refresh(*args, **kwargs): + raise AssertionError("valid invoke JWT should not be refreshed because metadata is stale") + + def _unexpected_mint(*args, **kwargs): + raise AssertionError("valid invoke JWT should not fall back to legacy mint") + + monkeypatch.setattr(auth_mod, "_refresh_access_token", _unexpected_refresh) + monkeypatch.setattr(auth_mod, "_mint_agent_key", _unexpected_mint) + + creds = auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=300) + + assert creds["api_key"] == token + assert creds["source"] == auth_mod.NOUS_AUTH_PATH_INVOKE_JWT + payload = json.loads((hermes_home / "auth.json").read_text()) + singleton = payload["providers"]["nous"] + assert singleton["agent_key"] == token + assert datetime.fromisoformat(singleton["expires_at"]).timestamp() > time.time() + 300 + assert datetime.fromisoformat(singleton["agent_key_expires_at"]).timestamp() > time.time() + 300 + + +def test_resolve_nous_runtime_credentials_does_not_apply_legacy_ttl_to_invoke_jwt( + tmp_path, + monkeypatch, +): + import hermes_cli.auth as auth_mod + + hermes_home = tmp_path / "hermes" + token = _invoke_jwt(seconds=900) + _setup_nous_auth( + hermes_home, + access_token=token, + scope=auth_mod.DEFAULT_NOUS_SCOPE, + expires_at=_future_iso(900), + expires_in=900, + ) + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + + def _unexpected_mint(*args, **kwargs): + raise AssertionError("1800s legacy min TTL should not force opaque mint for invoke JWT") + + monkeypatch.setattr(auth_mod, "_mint_agent_key", _unexpected_mint) + + creds = auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=1800) + + assert creds["api_key"] == token + assert creds["source"] == auth_mod.NOUS_AUTH_PATH_INVOKE_JWT + payload = json.loads((hermes_home / "auth.json").read_text()) + assert payload["providers"]["nous"]["agent_key"] == token + assert payload["credential_pool"]["nous"][0]["agent_key"] == token + + +def test_legacy_auth_mode_bypasses_usable_invoke_jwt(tmp_path, monkeypatch): + import hermes_cli.auth as auth_mod + + hermes_home = tmp_path / "hermes" + token = _invoke_jwt(seconds=3600) + _setup_nous_auth( + hermes_home, + access_token=token, + scope=auth_mod.DEFAULT_NOUS_SCOPE, + expires_at=_future_iso(3600), + expires_in=3600, + ) + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + + mint_calls = [] + + def _fake_mint_agent_key(*, client, portal_base_url, access_token, min_ttl_seconds): + del client, portal_base_url, min_ttl_seconds + mint_calls.append(access_token) + return _mint_payload(api_key="legacy-after-jwt-401") + + monkeypatch.setattr(auth_mod, "_mint_agent_key", _fake_mint_agent_key) + + creds = auth_mod.resolve_nous_runtime_credentials( + min_key_ttl_seconds=300, + inference_auth_mode=auth_mod.NOUS_INFERENCE_AUTH_MODE_LEGACY, + ) + + assert mint_calls == [token] + assert creds["api_key"] == "legacy-after-jwt-401" + assert creds["auth_path"] == auth_mod.NOUS_AUTH_PATH_LEGACY_SESSION_KEY_MINT + payload = json.loads((hermes_home / "auth.json").read_text()) + assert payload["providers"]["nous"]["agent_key"] == "legacy-after-jwt-401" + + +def test_resolve_nous_runtime_credentials_falls_back_when_invoke_scope_missing( + tmp_path, + monkeypatch, +): + import hermes_cli.auth as auth_mod + + hermes_home = tmp_path / "hermes" + token = _jwt_with_claims({ + "sub": "test-user", + "scope": "inference:mint_agent_key", + "exp": int(time.time() + 3600), + }) + _setup_nous_auth( + hermes_home, + access_token=token, + scope=auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE, + expires_at=_future_iso(3600), + expires_in=3600, + ) + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + + calls = [] + + def _fake_mint_agent_key(*, client, portal_base_url, access_token, min_ttl_seconds): + del client, portal_base_url, min_ttl_seconds + calls.append(access_token) + return _mint_payload(api_key="opaque-agent-key") + + monkeypatch.setattr(auth_mod, "_mint_agent_key", _fake_mint_agent_key) + + creds = auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=300) + + assert calls == [token] + assert creds["api_key"] == "opaque-agent-key" + assert creds["source"] == "portal" + payload = json.loads((hermes_home / "auth.json").read_text()) + assert payload["providers"]["nous"]["agent_key"] == "opaque-agent-key" + assert payload["credential_pool"]["nous"][0]["agent_key"] == "opaque-agent-key" + + +def test_nous_device_code_login_retries_legacy_scope_when_invoke_refused(monkeypatch): + import hermes_cli.auth as auth_mod + + scopes = [] + + def _fake_request_device_code(*, client, portal_base_url, client_id, scope): + del client, portal_base_url, client_id + scopes.append(scope) + if len(scopes) == 1: + request = httpx.Request("POST", "https://portal.example.com/api/oauth/device/code") + response = httpx.Response( + 400, + json={ + "error": "invalid_scope", + "error_description": "unsupported inference:invoke", + }, + request=request, + ) + raise httpx.HTTPStatusError("invalid_scope", request=request, response=response) + return { + "device_code": "device", + "user_code": "user", + "verification_uri": "https://portal.example.com/device", + "verification_uri_complete": "https://portal.example.com/device?code=user", + "expires_in": 600, + "interval": 1, + } + + def _fake_poll_for_token(**kwargs): + del kwargs + return { + "access_token": "access-legacy", + "refresh_token": "refresh-legacy", + "expires_in": 900, + "scope": auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE, + } + + def _fake_refresh(state, **kwargs): + del kwargs + refreshed = dict(state) + refreshed["agent_key"] = "opaque-agent-key" + refreshed["agent_key_expires_at"] = _future_iso(1800) + return refreshed + + monkeypatch.setattr(auth_mod, "_request_device_code", _fake_request_device_code) + monkeypatch.setattr(auth_mod, "_poll_for_token", _fake_poll_for_token) + monkeypatch.setattr(auth_mod, "refresh_nous_oauth_from_state", _fake_refresh) + + result = auth_mod._nous_device_code_login( + portal_base_url="https://portal.example.com", + inference_base_url="https://inference.example.com/v1", + open_browser=False, + timeout_seconds=1, + ) + + assert scopes == [auth_mod.DEFAULT_NOUS_SCOPE, auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE] + assert result["scope"] == auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE + assert result["agent_key"] == "opaque-agent-key" + + +def test_forced_legacy_env_skips_invoke_scope_and_jwt_storage(tmp_path, monkeypatch): + import hermes_cli.auth as auth_mod + + hermes_home = tmp_path / "hermes" + token = _invoke_jwt(seconds=3600) + _setup_nous_auth( + hermes_home, + access_token=token, + scope=auth_mod.DEFAULT_NOUS_SCOPE, + expires_at=_future_iso(3600), + expires_in=3600, + ) + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + monkeypatch.setenv(auth_mod.NOUS_LEGACY_SESSION_KEYS_ENV, "true") + + mint_calls = [] + + def _fake_mint_agent_key(*, client, portal_base_url, access_token, min_ttl_seconds): + del client, portal_base_url, min_ttl_seconds + mint_calls.append(access_token) + return _mint_payload(api_key="forced-legacy-key") + + monkeypatch.setattr(auth_mod, "_mint_agent_key", _fake_mint_agent_key) + + creds = auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=300) + + assert mint_calls == [token] + assert creds["api_key"] == "forced-legacy-key" + payload = json.loads((hermes_home / "auth.json").read_text()) + assert payload["providers"]["nous"]["agent_key"] == "forced-legacy-key" + + requested_scopes = [] + + def _fake_request_device_code(*, client, portal_base_url, client_id, scope): + del client, portal_base_url, client_id + requested_scopes.append(scope) + return { + "device_code": "device", + "user_code": "user", + "verification_uri": "https://portal.example.com/device", + "verification_uri_complete": "https://portal.example.com/device?code=user", + "expires_in": 600, + "interval": 1, + } + + def _fake_poll_for_token(**kwargs): + del kwargs + return { + "access_token": "access-legacy", + "refresh_token": "refresh-legacy", + "expires_in": 900, + "scope": auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE, + } + + def _fake_refresh(state, **kwargs): + del kwargs + refreshed = dict(state) + refreshed["agent_key"] = "forced-legacy-login-key" + refreshed["agent_key_expires_at"] = _future_iso(1800) + return refreshed + + monkeypatch.setattr(auth_mod, "_request_device_code", _fake_request_device_code) + monkeypatch.setattr(auth_mod, "_poll_for_token", _fake_poll_for_token) + monkeypatch.setattr(auth_mod, "refresh_nous_oauth_from_state", _fake_refresh) + + auth_mod._nous_device_code_login( + portal_base_url="https://portal.example.com", + inference_base_url="https://inference.example.com/v1", + open_browser=False, + timeout_seconds=1, + ) + + assert requested_scopes == [auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE] + + +def test_nous_inference_auth_logs_do_not_include_secret_values( + tmp_path, + monkeypatch, + caplog, +): + import hermes_cli.auth as auth_mod + + hermes_home = tmp_path / "hermes" + token = _jwt_with_claims({ + "sub": "secret-user", + "scope": "inference:mint_agent_key", + "exp": int(time.time() + 3600), + }) + refresh_token = "refresh-secret-token" + opaque_key = "opaque-secret-agent-key" + _setup_nous_auth( + hermes_home, + access_token=token, + refresh_token=refresh_token, + scope=auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE, + expires_at=_future_iso(3600), + expires_in=3600, + ) + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + + def _fake_mint_agent_key(*, client, portal_base_url, access_token, min_ttl_seconds): + del client, portal_base_url, access_token, min_ttl_seconds + return _mint_payload(api_key=opaque_key) + + monkeypatch.setattr(auth_mod, "_mint_agent_key", _fake_mint_agent_key) + + caplog.set_level(logging.INFO, logger="hermes_cli.auth") + auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=300) + + logged = caplog.text + assert "legacy session key path" in logged + assert token not in logged + assert refresh_token not in logged + assert opaque_key not in logged + + def test_get_nous_auth_status_checks_credential_pool(tmp_path, monkeypatch): """get_nous_auth_status() should find Nous credentials in the pool even when the auth store has no Nous provider entry — this is the @@ -373,6 +838,99 @@ def test_refresh_token_persisted_when_mint_times_out(tmp_path, monkeypatch): assert state_after_failure["access_token"] == "access-1" +def test_terminal_refresh_failure_quarantines_tokens( + tmp_path, monkeypatch, shared_store_env, +): + """A revoked/invalid Nous refresh token must not be replayed forever.""" + from hermes_cli import auth as auth_mod + + hermes_home = tmp_path / "hermes" + _setup_nous_auth(hermes_home, refresh_token="refresh-old") + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + from agent.credential_pool import load_pool + + assert load_pool("nous").select() is not None + + shared_state = _full_state_fixture() + shared_state["access_token"] = "access-old" + shared_state["refresh_token"] = "refresh-old" + shared_state["expires_at"] = "2026-02-01T00:00:00+00:00" + auth_mod._write_shared_nous_state(shared_state) + + refresh_calls: list[str] = [] + + def _terminal_refresh_failure(*, client, portal_base_url, client_id, refresh_token): + refresh_calls.append(refresh_token) + raise AuthError( + "Refresh session has been revoked", + provider="nous", + code="invalid_grant", + relogin_required=True, + ) + + monkeypatch.setattr(auth_mod, "_refresh_access_token", _terminal_refresh_failure) + + with pytest.raises(AuthError, match="Refresh session has been revoked"): + auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=300) + + state_after_failure = auth_mod.get_provider_auth_state("nous") + assert state_after_failure is not None + assert not state_after_failure.get("refresh_token") + assert not state_after_failure.get("access_token") + assert not state_after_failure.get("agent_key") + assert state_after_failure["last_auth_error"]["code"] == "invalid_grant" + assert auth_mod._read_shared_nous_state() is None + payload = json.loads((hermes_home / "auth.json").read_text()) + assert payload.get("credential_pool", {}).get("nous") == [] + + with pytest.raises(AuthError, match="No access token found"): + auth_mod.resolve_nous_runtime_credentials(min_key_ttl_seconds=300) + + assert refresh_calls == ["refresh-old"] + + +def test_managed_access_token_refresh_failure_quarantines_tokens( + tmp_path, monkeypatch, shared_store_env, +): + from hermes_cli import auth as auth_mod + + hermes_home = tmp_path / "hermes" + _setup_nous_auth(hermes_home, refresh_token="refresh-old") + monkeypatch.setenv("HERMES_HOME", str(hermes_home)) + from agent.credential_pool import load_pool + + assert load_pool("nous").select() is not None + + refresh_calls: list[str] = [] + + def _terminal_refresh_failure(*, client, portal_base_url, client_id, refresh_token): + refresh_calls.append(refresh_token) + raise AuthError( + "Invalid refresh token", + provider="nous", + code="invalid_grant", + relogin_required=True, + ) + + monkeypatch.setattr(auth_mod, "_refresh_access_token", _terminal_refresh_failure) + + with pytest.raises(AuthError, match="Invalid refresh token"): + auth_mod.resolve_nous_access_token() + + state_after_failure = auth_mod.get_provider_auth_state("nous") + assert state_after_failure is not None + assert not state_after_failure.get("refresh_token") + assert not state_after_failure.get("access_token") + assert state_after_failure["last_auth_error"]["message"] == "Invalid refresh token" + payload = json.loads((hermes_home / "auth.json").read_text()) + assert payload.get("credential_pool", {}).get("nous") == [] + + with pytest.raises(AuthError, match="No access token found"): + auth_mod.resolve_nous_access_token() + + assert refresh_calls == ["refresh-old"] + + def test_mint_retry_uses_latest_rotated_refresh_token(tmp_path, monkeypatch): hermes_home = tmp_path / "hermes" _setup_nous_auth(hermes_home, refresh_token="refresh-old") @@ -555,7 +1113,7 @@ class TestLoginNousSkipKeepsCurrent: auth_path = hermes_home / "auth.json" auth_after = json.loads(auth_path.read_text()) # active_provider should NOT be set to "nous" after Skip - assert auth_after.get("active_provider") in (None, "") + assert auth_after.get("active_provider") in {None, ""} # But Nous creds are still saved assert "nous" in auth_after.get("providers", {}) @@ -640,7 +1198,11 @@ def test_persist_nous_credentials_allows_recovery_from_401(tmp_path, monkeypatch calls after a Nous 401 — before the fix it would raise AuthError because providers.nous was empty. """ - from hermes_cli.auth import persist_nous_credentials, resolve_nous_runtime_credentials + from hermes_cli.auth import ( + NOUS_INFERENCE_AUTH_MODE_FRESH, + persist_nous_credentials, + resolve_nous_runtime_credentials, + ) hermes_home = tmp_path / "hermes" hermes_home.mkdir(parents=True, exist_ok=True) @@ -668,7 +1230,10 @@ def test_persist_nous_credentials_allows_recovery_from_401(tmp_path, monkeypatch monkeypatch.setattr("hermes_cli.auth._refresh_access_token", _fake_refresh_access_token) monkeypatch.setattr("hermes_cli.auth._mint_agent_key", _fake_mint_agent_key) - creds = resolve_nous_runtime_credentials(min_key_ttl_seconds=300, force_mint=True) + creds = resolve_nous_runtime_credentials( + min_key_ttl_seconds=300, + inference_auth_mode=NOUS_INFERENCE_AUTH_MODE_FRESH, + ) assert creds["api_key"] == "new-agent-key" @@ -861,6 +1426,36 @@ def test_refresh_token_reuse_detection_surfaces_actionable_message(): assert exc_info.value.relogin_required is True +def test_refresh_token_reuse_error_code_is_terminal(): + """Nous may return refresh_token_reused as the OAuth error code itself.""" + from hermes_cli import auth as auth_mod + + class _FakeResponse: + status_code = 400 + + def json(self): + return { + "error": "refresh_token_reused", + "error_description": "Refresh token reuse detected", + } + + class _FakeClient: + def post(self, *args, **kwargs): + return _FakeResponse() + + with pytest.raises(AuthError) as exc_info: + auth_mod._refresh_access_token( + client=_FakeClient(), + portal_base_url="https://portal.nousresearch.com", + client_id="hermes-cli", + refresh_token="rt_consumed_elsewhere", + ) + + assert exc_info.value.code == "refresh_token_reused" + assert exc_info.value.relogin_required is True + assert auth_mod._is_terminal_nous_refresh_error(exc_info.value) is True + + def test_refresh_token_exchange_sends_refresh_token_header(): """Nous refresh tokens must be sent in a header so sandbox proxies can substitute placeholder credentials without parsing form bodies. @@ -1118,6 +1713,47 @@ def test_try_import_shared_returns_none_on_refresh_failure( monkeypatch.setattr(auth_mod, "refresh_nous_oauth_from_state", _boom) assert auth_mod._try_import_shared_nous_state() is None + assert auth_mod._read_shared_nous_state() is None + + +def test_try_import_shared_persists_rotated_token_when_mint_fails( + shared_store_env, monkeypatch, +): + """A forced shared import refresh rotates the single-use token before minting. + + If the later agent-key mint fails, the shared store must still keep the + rotated refresh token; otherwise the next import attempt replays the + consumed token and trips refresh-token reuse. + """ + from hermes_cli import auth as auth_mod + + shared_state = _full_state_fixture() + shared_state["refresh_token"] = "refresh-old" + shared_state["access_token"] = "access-old" + auth_mod._write_shared_nous_state(shared_state) + + def _fake_refresh_access_token(*, client, portal_base_url, client_id, refresh_token): + assert refresh_token == "refresh-old" + return { + "access_token": "access-new", + "refresh_token": "refresh-new", + "expires_in": 900, + "token_type": "Bearer", + } + + def _fake_mint_agent_key(*, client, portal_base_url, access_token, min_ttl_seconds): + assert access_token == "access-new" + raise AuthError("credits exhausted", provider="nous", code="insufficient_credits") + + monkeypatch.setattr(auth_mod, "_refresh_access_token", _fake_refresh_access_token) + monkeypatch.setattr(auth_mod, "_mint_agent_key", _fake_mint_agent_key) + + assert auth_mod._try_import_shared_nous_state() is None + + shared_after = auth_mod._read_shared_nous_state() + assert shared_after is not None + assert shared_after["refresh_token"] == "refresh-new" + assert shared_after["access_token"] == "access-new" def test_try_import_shared_rehydrates_on_success(shared_store_env, monkeypatch): @@ -1132,7 +1768,10 @@ def test_try_import_shared_rehydrates_on_success(shared_store_env, monkeypatch): def _fake_refresh(state, **kwargs): # Simulate portal returning fresh tokens + a new agent_key assert kwargs.get("force_refresh") is True - assert kwargs.get("force_mint") is True + assert ( + kwargs.get("inference_auth_mode") + == auth_mod.NOUS_INFERENCE_AUTH_MODE_FRESH + ) return { **state, "access_token": "fresh-access-tok", @@ -1260,7 +1899,7 @@ def test_runtime_refresh_uses_newer_shared_token_before_local_stale_token( creds = auth_mod.resolve_nous_runtime_credentials( min_key_ttl_seconds=300, - force_mint=True, + inference_auth_mode=auth_mod.NOUS_INFERENCE_AUTH_MODE_FRESH, ) assert creds["api_key"] == "agent-key-from-shared-token" diff --git a/tests/hermes_cli/test_cmd_update.py b/tests/hermes_cli/test_cmd_update.py index 6719a1fe532..9bafaa480b9 100644 --- a/tests/hermes_cli/test_cmd_update.py +++ b/tests/hermes_cli/test_cmd_update.py @@ -157,6 +157,24 @@ class TestCmdUpdateBranchFallback: (["/usr/bin/npm", "run", "build"], PROJECT_ROOT / "apps" / "dashboard"), ] + # Regression for #18840: repo root + ui-tui installs must stream + # output (capture_output=False) so postinstall progress is visible + # to the user. + repo_and_tui_calls = [ + call + for call in mock_run.call_args_list + if call.args + and call.args[0][0] == "/usr/bin/npm" + and call.args[0][1] == "ci" + and call.kwargs.get("cwd") in {PROJECT_ROOT, PROJECT_ROOT / "ui-tui"} + ] + assert len(repo_and_tui_calls) == 2 + for call in repo_and_tui_calls: + assert call.kwargs.get("capture_output") is False, ( + "repo-root / ui-tui npm install must stream output " + "(no capture_output) so postinstall progress is visible" + ) + def test_update_non_interactive_runs_safe_config_migrations(self, mock_args, capsys): """Dashboard/web updates apply non-interactive migrations before restart.""" with patch("shutil.which", return_value=None), patch( diff --git a/tests/hermes_cli/test_codex_runtime_switch.py b/tests/hermes_cli/test_codex_runtime_switch.py index 7bf1a59e1e7..a0b4aa5fd41 100644 --- a/tests/hermes_cli/test_codex_runtime_switch.py +++ b/tests/hermes_cli/test_codex_runtime_switch.py @@ -105,7 +105,7 @@ class TestApply: assert "Cannot enable" in r.message assert "npm i -g @openai/codex" in r.message # Config NOT mutated on failure - assert cfg.get("model", {}).get("openai_runtime") in (None, "") + assert cfg.get("model", {}).get("openai_runtime") in {None, ""} def test_enable_succeeds_when_codex_present(self): cfg = {} diff --git a/tests/hermes_cli/test_commands.py b/tests/hermes_cli/test_commands.py index d08f886fa6a..6de778347e1 100644 --- a/tests/hermes_cli/test_commands.py +++ b/tests/hermes_cli/test_commands.py @@ -107,6 +107,7 @@ class TestResolveCommand: assert resolve_command("gateway").name == "platforms" assert resolve_command("set-home").name == "sethome" assert resolve_command("reload_mcp").name == "reload-mcp" + assert resolve_command("codex_runtime").name == "codex-runtime" assert resolve_command("tasks").name == "agents" def test_topic_is_gateway_command(self): @@ -251,6 +252,12 @@ class TestTelegramBotCommands: assert "queue" in names assert "steer" in names + def test_hyphenated_codex_runtime_is_exposed_as_underscore_command(self): + """Telegram autocomplete exposes /codex-runtime as /codex_runtime.""" + names = {name for name, _ in telegram_bot_commands()} + assert "codex_runtime" in names + assert "codex-runtime" not in names + class TestSlackSubcommandMap: def test_returns_dict(self): diff --git a/tests/hermes_cli/test_doctor.py b/tests/hermes_cli/test_doctor.py index ee419656a71..be8c35239b3 100644 --- a/tests/hermes_cli/test_doctor.py +++ b/tests/hermes_cli/test_doctor.py @@ -320,6 +320,7 @@ class TestDoctorMemoryProviderSection: from hermes_cli import auth as _auth_mod monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except Exception: pass @@ -426,6 +427,7 @@ def test_run_doctor_accepts_named_provider_from_providers_section(monkeypatch, t from hermes_cli import auth as _auth_mod monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except Exception: pass @@ -463,6 +465,7 @@ def test_run_doctor_accepts_bare_custom_provider(monkeypatch, tmp_path): from hermes_cli import auth as _auth_mod monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except Exception: pass @@ -474,6 +477,48 @@ def test_run_doctor_accepts_bare_custom_provider(monkeypatch, tmp_path): assert "model.provider 'custom' is not a recognised provider" not in out +def test_run_doctor_flags_missing_credentials_for_active_openrouter_provider(monkeypatch, tmp_path): + home = tmp_path / ".hermes" + home.mkdir(parents=True, exist_ok=True) + (home / "config.yaml").write_text( + "model:\n" + " provider: openrouter\n" + " default: openai/gpt-4.1-mini\n", + encoding="utf-8", + ) + + monkeypatch.setattr(doctor_mod, "HERMES_HOME", home) + monkeypatch.setattr(doctor_mod, "PROJECT_ROOT", tmp_path / "project") + monkeypatch.setattr(doctor_mod, "_DHH", str(home)) + (tmp_path / "project").mkdir(exist_ok=True) + + fake_model_tools = types.SimpleNamespace( + check_tool_availability=lambda *a, **kw: ([], []), + TOOLSET_REQUIREMENTS={}, + ) + monkeypatch.setitem(sys.modules, "model_tools", fake_model_tools) + monkeypatch.delenv("OPENROUTER_API_KEY", raising=False) + monkeypatch.delenv("OPENAI_API_KEY", raising=False) + + try: + from hermes_cli import auth as _auth_mod + + monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_gemini_oauth_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_minimax_oauth_auth_status", lambda: {}) + except Exception: + pass + + buf = io.StringIO() + with contextlib.redirect_stdout(buf): + doctor_mod.run_doctor(Namespace(fix=False)) + + out = buf.getvalue() + assert "model.provider 'openrouter' is set but no API key is configured" in out + assert "No credentials found for provider 'openrouter'." in out + + @pytest.mark.parametrize( ("provider", "default_model"), [ @@ -510,6 +555,7 @@ def test_run_doctor_accepts_hermes_provider_ids_that_catalog_aliases( from hermes_cli import auth as _auth_mod monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except Exception: pass @@ -556,6 +602,7 @@ def test_run_doctor_accepts_kimi_coding_cn_provider(monkeypatch, tmp_path): monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_auth_status", lambda provider: {"logged_in": True}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except Exception: pass @@ -594,6 +641,7 @@ def test_run_doctor_termux_does_not_mark_browser_available_without_agent_browser from hermes_cli import auth as _auth_mod monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except Exception: pass @@ -633,6 +681,7 @@ def test_run_doctor_kimi_cn_env_is_detected_and_probe_is_null_safe(monkeypatch, from hermes_cli import auth as _auth_mod monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except Exception: pass @@ -681,6 +730,7 @@ def test_run_doctor_dashscope_retries_china_endpoint_after_intl_unauthorized(mon from hermes_cli import auth as _auth_mod monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except ImportError: pass @@ -739,6 +789,7 @@ def test_run_doctor_opencode_go_skips_invalid_models_probe(monkeypatch, tmp_path from hermes_cli import auth as _auth_mod monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {}) except ImportError: pass @@ -850,6 +901,7 @@ def _run_doctor_with_healthy_oauth_fallback( failing_host: str, gemini_oauth_status: dict, minimax_oauth_status: dict, + xai_oauth_status: dict | None = None, ) -> str: home = tmp_path / ".hermes" home.mkdir(parents=True, exist_ok=True) @@ -886,6 +938,8 @@ def _run_doctor_with_healthy_oauth_fallback( monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {}) monkeypatch.setattr(_auth_mod, "get_gemini_oauth_auth_status", lambda: gemini_oauth_status) monkeypatch.setattr(_auth_mod, "get_minimax_oauth_auth_status", lambda: minimax_oauth_status) + _xai_status = xai_oauth_status if xai_oauth_status is not None else {} + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: _xai_status) def fake_get(url, headers=None, timeout=None): status = 401 if failing_host in url else 200 @@ -902,7 +956,7 @@ def _run_doctor_with_healthy_oauth_fallback( @pytest.mark.parametrize( - ("env_key", "bad_key", "failing_host", "gemini_oauth_status", "minimax_oauth_status", "unexpected_issue"), + ("env_key", "bad_key", "failing_host", "gemini_oauth_status", "minimax_oauth_status", "xai_oauth_status", "unexpected_issue"), [ ( "GOOGLE_API_KEY", @@ -910,6 +964,7 @@ def _run_doctor_with_healthy_oauth_fallback( "googleapis.com", {"logged_in": True, "email": "user@example.com"}, {}, + None, "Check GOOGLE_API_KEY in .env", ), ( @@ -918,8 +973,18 @@ def _run_doctor_with_healthy_oauth_fallback( "minimax.io", {}, {"logged_in": True, "region": "global"}, + None, "Check MINIMAX_API_KEY in .env", ), + ( + "XAI_API_KEY", + "bad-xai-key", + "api.x.ai", + {}, + {}, + {"logged_in": True, "auth_mode": "oauth_pkce"}, + "Check XAI_API_KEY in .env", + ), ], ) def test_run_doctor_ignores_invalid_direct_keys_when_oauth_fallback_is_healthy( @@ -930,6 +995,7 @@ def test_run_doctor_ignores_invalid_direct_keys_when_oauth_fallback_is_healthy( failing_host, gemini_oauth_status, minimax_oauth_status, + xai_oauth_status, unexpected_issue, ): out = _run_doctor_with_healthy_oauth_fallback( @@ -940,7 +1006,220 @@ def test_run_doctor_ignores_invalid_direct_keys_when_oauth_fallback_is_healthy( failing_host=failing_host, gemini_oauth_status=gemini_oauth_status, minimax_oauth_status=minimax_oauth_status, + xai_oauth_status=xai_oauth_status, ) assert "invalid API key" in out assert unexpected_issue not in out + + +def test_has_healthy_oauth_fallback_returns_false_for_unknown_provider(): + from hermes_cli.doctor import _has_healthy_oauth_fallback_for_apikey_provider + assert _has_healthy_oauth_fallback_for_apikey_provider("unknown-provider") is False + + +class TestHasHealthyOauthFallbackForXai: + def test_returns_true_when_xai_oauth_healthy(self, monkeypatch): + from hermes_cli import auth as _auth_mod + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {"logged_in": True}) + from hermes_cli.doctor import _has_healthy_oauth_fallback_for_apikey_provider + assert _has_healthy_oauth_fallback_for_apikey_provider("xai") is True + + def test_returns_false_when_xai_oauth_not_logged_in(self, monkeypatch): + from hermes_cli import auth as _auth_mod + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: {"logged_in": False}) + from hermes_cli.doctor import _has_healthy_oauth_fallback_for_apikey_provider + assert _has_healthy_oauth_fallback_for_apikey_provider("xai") is False + + def test_returns_false_when_xai_oauth_returns_none(self, monkeypatch): + from hermes_cli import auth as _auth_mod + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", lambda: None) + from hermes_cli.doctor import _has_healthy_oauth_fallback_for_apikey_provider + assert _has_healthy_oauth_fallback_for_apikey_provider("xai") is False + + def test_returns_false_when_xai_import_unavailable(self, monkeypatch): + import sys + # Simulate get_xai_oauth_auth_status missing from auth module + monkeypatch.delattr("hermes_cli.auth.get_xai_oauth_auth_status", raising=False) + # Force doctor module to re-import the function + monkeypatch.delitem(sys.modules, "hermes_cli.doctor", raising=False) + from hermes_cli.doctor import _has_healthy_oauth_fallback_for_apikey_provider + assert _has_healthy_oauth_fallback_for_apikey_provider("xai") is False + + def test_xai_import_failure_does_not_affect_gemini(self, monkeypatch): + import sys + from hermes_cli import auth as _auth_mod + # xAI function missing, but Gemini is healthy + monkeypatch.delattr(_auth_mod, "get_xai_oauth_auth_status", raising=False) + monkeypatch.setattr(_auth_mod, "get_gemini_oauth_auth_status", lambda: {"logged_in": True}) + monkeypatch.delitem(sys.modules, "hermes_cli.doctor", raising=False) + from hermes_cli.doctor import _has_healthy_oauth_fallback_for_apikey_provider + assert _has_healthy_oauth_fallback_for_apikey_provider("gemini") is True + + +# --------------------------------------------------------------------------- +# ◆ Auth Providers — xAI OAuth display in run_doctor() +# --------------------------------------------------------------------------- + + +class TestDoctorXaiOAuthStatus: + """The ◆ Auth Providers section must show xAI OAuth login state. + + xAI OAuth is checked in a *separate* try/except block so that an import + failure (or runtime exception) cannot silence the Nous / Codex / Gemini / + MiniMax rows that were already printed above it. + """ + + def _run(self, monkeypatch, tmp_path, *, xai_auth_fn) -> str: + """Run doctor with a controlled xAI auth callable; return stdout.""" + home = tmp_path / ".hermes" + home.mkdir(parents=True, exist_ok=True) + (home / "config.yaml").write_text("memory: {}\n", encoding="utf-8") + project = tmp_path / "project" + project.mkdir(exist_ok=True) + + monkeypatch.setattr(doctor_mod, "HERMES_HOME", home) + monkeypatch.setattr(doctor_mod, "PROJECT_ROOT", project) + monkeypatch.setattr(doctor_mod, "_DHH", str(home)) + + fake_model_tools = types.SimpleNamespace( + check_tool_availability=lambda *a, **kw: ([], []), + TOOLSET_REQUIREMENTS={}, + ) + monkeypatch.setitem(sys.modules, "model_tools", fake_model_tools) + + from hermes_cli import auth as _auth_mod + monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_gemini_oauth_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_minimax_oauth_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_xai_oauth_auth_status", xai_auth_fn) + + buf = io.StringIO() + with contextlib.redirect_stdout(buf): + doctor_mod.run_doctor(Namespace(fix=False)) + return buf.getvalue() + + def test_logged_in_shows_ok(self, monkeypatch, tmp_path): + out = self._run( + monkeypatch, tmp_path, + xai_auth_fn=lambda: {"logged_in": True}, + ) + assert "xAI OAuth" in out + assert "(logged in)" in out + + def test_not_logged_in_shows_warn(self, monkeypatch, tmp_path): + out = self._run( + monkeypatch, tmp_path, + xai_auth_fn=lambda: {"logged_in": False}, + ) + assert "xAI OAuth" in out + assert "(not logged in)" in out + + def test_error_shown_when_not_logged_in_and_error_present(self, monkeypatch, tmp_path): + out = self._run( + monkeypatch, tmp_path, + xai_auth_fn=lambda: {"logged_in": False, "error": "refresh token expired"}, + ) + assert "xAI OAuth" in out + assert "refresh token expired" in out + + def test_no_error_line_when_error_key_absent(self, monkeypatch, tmp_path): + out = self._run( + monkeypatch, tmp_path, + xai_auth_fn=lambda: {"logged_in": False}, + ) + assert "xAI OAuth" in out + # The check_info line is only emitted when the "error" key is present. + # Pick a token that would appear in no ordinary doctor output. + assert "refresh token expired" not in out + + def test_logged_in_does_not_emit_not_logged_in_on_xai_line(self, monkeypatch, tmp_path): + out = self._run( + monkeypatch, tmp_path, + xai_auth_fn=lambda: {"logged_in": True}, + ) + assert "xAI OAuth" in out + # The xAI OAuth line itself must say "(logged in)", not "(not logged in)". + xai_line = next(l for l in out.splitlines() if "xAI OAuth" in l) + assert "(logged in)" in xai_line + assert "(not logged in)" not in xai_line + + def test_import_failure_does_not_crash_doctor(self, monkeypatch, tmp_path): + """Doctor must not crash when get_xai_oauth_auth_status cannot be imported.""" + home = tmp_path / ".hermes" + home.mkdir(parents=True, exist_ok=True) + (home / "config.yaml").write_text("memory: {}\n", encoding="utf-8") + project = tmp_path / "project" + project.mkdir(exist_ok=True) + + monkeypatch.setattr(doctor_mod, "HERMES_HOME", home) + monkeypatch.setattr(doctor_mod, "PROJECT_ROOT", project) + monkeypatch.setattr(doctor_mod, "_DHH", str(home)) + + fake_model_tools = types.SimpleNamespace( + check_tool_availability=lambda *a, **kw: ([], []), + TOOLSET_REQUIREMENTS={}, + ) + monkeypatch.setitem(sys.modules, "model_tools", fake_model_tools) + + from hermes_cli import auth as _auth_mod + monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_gemini_oauth_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_minimax_oauth_auth_status", lambda: {"logged_in": False}) + monkeypatch.delattr(_auth_mod, "get_xai_oauth_auth_status", raising=False) + + buf = io.StringIO() + with contextlib.redirect_stdout(buf): + doctor_mod.run_doctor(Namespace(fix=False)) + out = buf.getvalue() + # The ◆ Auth Providers header must still appear — other providers unaffected. + assert "Auth Providers" in out + + def test_import_failure_does_not_affect_other_providers(self, monkeypatch, tmp_path): + """Nous / Codex / Gemini / MiniMax rows must survive an xAI import failure.""" + home = tmp_path / ".hermes" + home.mkdir(parents=True, exist_ok=True) + (home / "config.yaml").write_text("memory: {}\n", encoding="utf-8") + project = tmp_path / "project" + project.mkdir(exist_ok=True) + + monkeypatch.setattr(doctor_mod, "HERMES_HOME", home) + monkeypatch.setattr(doctor_mod, "PROJECT_ROOT", project) + monkeypatch.setattr(doctor_mod, "_DHH", str(home)) + + fake_model_tools = types.SimpleNamespace( + check_tool_availability=lambda *a, **kw: ([], []), + TOOLSET_REQUIREMENTS={}, + ) + monkeypatch.setitem(sys.modules, "model_tools", fake_model_tools) + + from hermes_cli import auth as _auth_mod + monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {"logged_in": True}) + monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_gemini_oauth_auth_status", lambda: {"logged_in": False}) + monkeypatch.setattr(_auth_mod, "get_minimax_oauth_auth_status", lambda: {"logged_in": False}) + monkeypatch.delattr(_auth_mod, "get_xai_oauth_auth_status", raising=False) + + buf = io.StringIO() + with contextlib.redirect_stdout(buf): + doctor_mod.run_doctor(Namespace(fix=False)) + out = buf.getvalue() + assert "Nous Portal auth" in out + assert "logged in" in out + + def test_function_raises_does_not_crash_doctor(self, monkeypatch, tmp_path): + """A runtime exception from get_xai_oauth_auth_status must be swallowed.""" + def _raise(): + raise RuntimeError("simulated xAI status failure") + + out = self._run(monkeypatch, tmp_path, xai_auth_fn=_raise) + assert "Auth Providers" in out + + def test_function_returns_none_does_not_crash_doctor(self, monkeypatch, tmp_path): + """None return is normalised to {} via `or {}` — must not AttributeError.""" + out = self._run(monkeypatch, tmp_path, xai_auth_fn=lambda: None) + # None → {} → logged_in falsy → shows not-logged-in warn + assert "xAI OAuth" in out + assert "(not logged in)" in out diff --git a/tests/hermes_cli/test_install_cua_driver.py b/tests/hermes_cli/test_install_cua_driver.py index 42a49e22b5d..6cd50261694 100644 --- a/tests/hermes_cli/test_install_cua_driver.py +++ b/tests/hermes_cli/test_install_cua_driver.py @@ -48,7 +48,7 @@ class TestInstallCuaDriverUpgrade: with patch("platform.system", return_value="Darwin"), \ patch.object(tools_config.shutil, "which", side_effect=lambda n: "/usr/local/bin/" + n - if n in ("cua-driver", "curl") else None), \ + if n in {"cua-driver", "curl"} else None), \ patch.object(tools_config, "_run_cua_driver_installer", return_value=True) as runner, \ patch("subprocess.run"): @@ -82,7 +82,7 @@ class TestInstallCuaDriverUpgrade: with patch("platform.system", return_value="Darwin"), \ patch.object(tools_config.shutil, "which", side_effect=lambda n: "/usr/local/bin/" + n - if n in ("cua-driver", "curl") else None), \ + if n in {"cua-driver", "curl"} else None), \ patch.object(tools_config, "_run_cua_driver_installer") as runner, \ patch("subprocess.run"): assert tools_config.install_cua_driver(upgrade=False) is True diff --git a/tests/hermes_cli/test_kanban_core_functionality.py b/tests/hermes_cli/test_kanban_core_functionality.py index 17252af827a..35dc7ace951 100644 --- a/tests/hermes_cli/test_kanban_core_functionality.py +++ b/tests/hermes_cli/test_kanban_core_functionality.py @@ -1046,7 +1046,7 @@ def test_enforce_max_runtime_integrates_with_dispatch(kanban_home, monkeypatch): task = kb.get_task(conn, tid) # After timeout, task is back in 'ready' and will be re-spawned # by the same pass. That's the intended behaviour. - assert task.status in ("ready", "running") + assert task.status in {"ready", "running"} finally: conn.close() diff --git a/tests/hermes_cli/test_kanban_decompose.py b/tests/hermes_cli/test_kanban_decompose.py new file mode 100644 index 00000000000..f55e10e2f8e --- /dev/null +++ b/tests/hermes_cli/test_kanban_decompose.py @@ -0,0 +1,242 @@ +"""Tests for the decomposer module + `hermes kanban decompose` CLI surface. + +The auxiliary LLM client is mocked — no network calls. Tests exercise the +prompt plumbing, response parsing, DB writes (via the real DB helper), +and the assignee-fallback logic. +""" + +from __future__ import annotations + +import argparse +import json as jsonlib +from pathlib import Path +from unittest.mock import MagicMock, patch + +import pytest + +from hermes_cli import kanban as kanban_cli +from hermes_cli import kanban_db as kb +from hermes_cli import kanban_decompose as decomp + + +@pytest.fixture +def kanban_home(tmp_path, monkeypatch): + home = tmp_path / ".hermes" + home.mkdir() + monkeypatch.setenv("HERMES_HOME", str(home)) + monkeypatch.setattr(Path, "home", lambda: tmp_path) + kb.init_db() + return home + + +def _fake_aux_response(content: str): + resp = MagicMock() + resp.choices = [MagicMock()] + resp.choices[0].message.content = content + return resp + + +def _mock_client_returning(content: str): + client = MagicMock() + client.chat.completions.create = MagicMock(return_value=_fake_aux_response(content)) + return client + + +def _patch_aux_client(content: str, *, model: str = "test-model"): + client = _mock_client_returning(content) + return patch( + "agent.auxiliary_client.get_text_auxiliary_client", + return_value=(client, model), + ) + + +def _patch_extra_body(): + return patch( + "agent.auxiliary_client.get_auxiliary_extra_body", + return_value={}, + ) + + +def _patch_list_profiles(names: list[str]): + """Pretend the named profiles exist. The decomposer uses + profiles_mod.list_profiles() to build the roster + valid-set, and + profiles_mod.profile_exists() to resolve orchestrator/default.""" + from types import SimpleNamespace + fake_profiles = [ + SimpleNamespace( + name=n, is_default=(i == 0), description=f"desc for {n}", + description_auto=False, model="m", provider="p", skill_count=1, + ) + for i, n in enumerate(names) + ] + return [ + patch("hermes_cli.profiles.list_profiles", return_value=fake_profiles), + patch("hermes_cli.profiles.profile_exists", side_effect=lambda x: x in names), + patch("hermes_cli.profiles.get_active_profile_name", return_value=names[0] if names else "default"), + ] + + +def test_decompose_with_fanout_creates_children(kanban_home): + with kb.connect() as conn: + tid = kb.create_task(conn, title="ship a feature", triage=True) + + llm_payload = jsonlib.dumps({ + "fanout": True, + "rationale": "test split", + "tasks": [ + {"title": "research", "body": "look it up", "assignee": "researcher", "parents": []}, + {"title": "build", "body": "code it", "assignee": "engineer", "parents": [0]}, + ], + }) + + patches = _patch_list_profiles(["orchestrator", "researcher", "engineer"]) + for p in patches: + p.start() + try: + with _patch_aux_client(llm_payload), _patch_extra_body(): + outcome = decomp.decompose_task(tid, author="me") + finally: + for p in patches: + p.stop() + + assert outcome.ok, outcome.reason + assert outcome.fanout is True + assert outcome.child_ids and len(outcome.child_ids) == 2 + + with kb.connect() as conn: + root = kb.get_task(conn, tid) + c0 = kb.get_task(conn, outcome.child_ids[0]) + c1 = kb.get_task(conn, outcome.child_ids[1]) + assert root.status == "todo" + assert c0.status == "ready" + assert c1.status == "todo" + assert c0.assignee == "researcher" + assert c1.assignee == "engineer" + + +def test_decompose_fanout_false_falls_back_to_specify(kanban_home): + with kb.connect() as conn: + tid = kb.create_task(conn, title="just one thing", triage=True) + + llm_payload = jsonlib.dumps({ + "fanout": False, + "rationale": "single unit", + "title": "Tightened title", + "body": "**Goal**\nDo the thing.", + }) + + patches = _patch_list_profiles(["orchestrator"]) + for p in patches: + p.start() + try: + with _patch_aux_client(llm_payload), _patch_extra_body(): + outcome = decomp.decompose_task(tid, author="me") + finally: + for p in patches: + p.stop() + + assert outcome.ok, outcome.reason + assert outcome.fanout is False + assert outcome.new_title == "Tightened title" + with kb.connect() as conn: + task = kb.get_task(conn, tid) + # specify path with no parents -> recompute_ready flips to 'ready' + assert task.status == "ready" + assert task.title == "Tightened title" + + +def test_decompose_unknown_assignee_falls_back_to_default(kanban_home): + with kb.connect() as conn: + tid = kb.create_task(conn, title="x", triage=True) + + # Roster only has 'orchestrator' and 'fallback'; LLM picks 'made_up'. + llm_payload = jsonlib.dumps({ + "fanout": True, + "rationale": "test", + "tasks": [ + {"title": "do X", "body": "", "assignee": "made_up", "parents": []}, + ], + }) + + patches = _patch_list_profiles(["orchestrator", "fallback"]) + for p in patches: + p.start() + try: + with patch.dict( + "os.environ", {}, clear=False, + ), _patch_aux_client(llm_payload), _patch_extra_body(), \ + patch( + "hermes_cli.kanban_decompose._load_config", + return_value={ + "kanban": { + "orchestrator_profile": "orchestrator", + "default_assignee": "fallback", + } + }, + ): + outcome = decomp.decompose_task(tid, author="me") + finally: + for p in patches: + p.stop() + + assert outcome.ok, outcome.reason + assert outcome.child_ids and len(outcome.child_ids) == 1 + with kb.connect() as conn: + child = kb.get_task(conn, outcome.child_ids[0]) + # 'made_up' wasn't in roster, so assignee rewritten to 'fallback' + assert child.assignee == "fallback" + + +def test_decompose_handles_malformed_llm_json(kanban_home): + with kb.connect() as conn: + tid = kb.create_task(conn, title="x", triage=True) + + patches = _patch_list_profiles(["orchestrator"]) + for p in patches: + p.start() + try: + with _patch_aux_client("not json at all, sorry"), _patch_extra_body(): + outcome = decomp.decompose_task(tid, author="me") + finally: + for p in patches: + p.stop() + + assert outcome.ok is False + assert "malformed JSON" in outcome.reason + + +def test_decompose_returns_false_when_task_not_triage(kanban_home): + with kb.connect() as conn: + tid = kb.create_task(conn, title="x") # ready, not triage + + patches = _patch_list_profiles(["orchestrator"]) + for p in patches: + p.start() + try: + outcome = decomp.decompose_task(tid, author="me") + finally: + for p in patches: + p.stop() + assert outcome.ok is False + assert "not in triage" in outcome.reason + + +def test_decompose_no_aux_client_configured(kanban_home): + with kb.connect() as conn: + tid = kb.create_task(conn, title="x", triage=True) + + patches = _patch_list_profiles(["orchestrator"]) + for p in patches: + p.start() + try: + with patch( + "agent.auxiliary_client.get_text_auxiliary_client", + return_value=(None, ""), + ): + outcome = decomp.decompose_task(tid, author="me") + finally: + for p in patches: + p.stop() + + assert outcome.ok is False + assert "no auxiliary client" in outcome.reason diff --git a/tests/hermes_cli/test_kanban_decompose_db.py b/tests/hermes_cli/test_kanban_decompose_db.py new file mode 100644 index 00000000000..236fb1fff1b --- /dev/null +++ b/tests/hermes_cli/test_kanban_decompose_db.py @@ -0,0 +1,152 @@ +"""Tests for kb.decompose_triage_task — the DB-layer atomic fan-out +from the triage column. LLM-free by design. +""" + +from __future__ import annotations + +from pathlib import Path + +import pytest + +from hermes_cli import kanban_db as kb + + +@pytest.fixture +def kanban_home(tmp_path, monkeypatch): + home = tmp_path / ".hermes" + home.mkdir() + monkeypatch.setenv("HERMES_HOME", str(home)) + monkeypatch.setattr(Path, "home", lambda: tmp_path) + kb.init_db() + return home + + +def _create_triage(conn, title="rough idea", body=None, assignee=None, tenant=None): + return kb.create_task( + conn, + title=title, + body=body, + assignee=assignee, + tenant=tenant, + triage=True, + ) + + +def test_decompose_creates_children_and_promotes_root(kanban_home): + with kb.connect() as conn: + tid = _create_triage(conn, title="ship a feature") + assert kb.get_task(conn, tid).status == "triage" + + children = [ + {"title": "research", "body": "look at prior art", "assignee": "researcher", "parents": []}, + {"title": "build it", "body": "write code", "assignee": "engineer", "parents": [0]}, + ] + with kb.connect() as conn: + child_ids = kb.decompose_triage_task( + conn, + tid, + root_assignee="orchestrator", + children=children, + author="decomposer", + ) + assert child_ids is not None + assert len(child_ids) == 2 + + with kb.connect() as conn: + root = kb.get_task(conn, tid) + c0 = kb.get_task(conn, child_ids[0]) + c1 = kb.get_task(conn, child_ids[1]) + + # Root flipped to todo with orchestrator assignee, gated by children. + assert root.status == "todo" + assert root.assignee == "orchestrator" + # First child has no internal parents → ready on recompute_ready. + assert c0.status == "ready" + assert c0.assignee == "researcher" + # Second child has parents=[0] → stays in todo until c0 completes. + assert c1.status == "todo" + assert c1.assignee == "engineer" + + +def test_decompose_returns_none_when_task_missing(kanban_home): + with kb.connect() as conn: + result = kb.decompose_triage_task( + conn, + "nonexistent", + root_assignee="orch", + children=[{"title": "x"}], + author="me", + ) + assert result is None + + +def test_decompose_returns_none_when_task_not_in_triage(kanban_home): + with kb.connect() as conn: + tid = kb.create_task(conn, title="already a real task") # not triage + result = kb.decompose_triage_task( + conn, + tid, + root_assignee="orch", + children=[{"title": "x"}], + author="me", + ) + assert result is None + + +def test_decompose_empty_children_returns_none(kanban_home): + with kb.connect() as conn: + tid = _create_triage(conn) + result = kb.decompose_triage_task( + conn, + tid, + root_assignee="orch", + children=[], + author="me", + ) + assert result is None + + +def test_decompose_rejects_self_parent(kanban_home): + with kb.connect() as conn: + tid = _create_triage(conn) + with pytest.raises(ValueError, match="cannot list itself"): + kb.decompose_triage_task( + conn, + tid, + root_assignee="orch", + children=[{"title": "x", "parents": [0]}], + author="me", + ) + + +def test_decompose_rejects_out_of_range_parent(kanban_home): + with kb.connect() as conn: + tid = _create_triage(conn) + with pytest.raises(ValueError, match="not a valid index"): + kb.decompose_triage_task( + conn, + tid, + root_assignee="orch", + children=[{"title": "x", "parents": [5]}], + author="me", + ) + + +def test_decompose_records_audit_comment_and_event(kanban_home): + with kb.connect() as conn: + tid = _create_triage(conn) + child_ids = kb.decompose_triage_task( + conn, + tid, + root_assignee="orch", + children=[{"title": "task A", "assignee": "researcher"}], + author="alice", + ) + assert child_ids is not None + + with kb.connect() as conn: + comments = kb.list_comments(conn, tid) + events = kb.list_events(conn, tid) + + assert any("Decomposed into" in (c.body or "") for c in comments) + assert any(ev.kind == "decomposed" for ev in events) diff --git a/tests/hermes_cli/test_memory_reset.py b/tests/hermes_cli/test_memory_reset.py index 3b91326de20..48f1cfda6a7 100644 --- a/tests/hermes_cli/test_memory_reset.py +++ b/tests/hermes_cli/test_memory_reset.py @@ -43,9 +43,9 @@ def _run_memory_reset(target="all", yes=False, monkeypatch=None, confirm_input=" mem_dir = get_hermes_home() / "memories" files_to_reset = [] - if target in ("all", "memory"): + if target in {"all", "memory"}: files_to_reset.append(("MEMORY.md", "agent notes")) - if target in ("all", "user"): + if target in {"all", "user"}: files_to_reset.append(("USER.md", "user profile")) existing = [(f, desc) for f, desc in files_to_reset if (mem_dir / f).exists()] diff --git a/tests/hermes_cli/test_models.py b/tests/hermes_cli/test_models.py index 8ccf5b57f2d..78568f81f2c 100644 --- a/tests/hermes_cli/test_models.py +++ b/tests/hermes_cli/test_models.py @@ -252,7 +252,7 @@ class TestDetectProviderForModel: result = detect_provider_for_model("deepseek-chat", "openai-codex") assert result is not None # Provider is deepseek (direct) or openrouter (fallback) depending on creds - assert result[0] in ("deepseek", "openrouter") + assert result[0] in {"deepseek", "openrouter"} def test_current_provider_model_returns_none(self): """Models belonging to the current provider should not trigger a switch.""" @@ -302,7 +302,7 @@ class TestDetectProviderForModel: with patch("hermes_cli.models.fetch_openrouter_models", return_value=LIVE_OPENROUTER_MODELS): result = detect_provider_for_model("claude-opus-4-6", "openai-codex") assert result is not None - assert result[0] not in ("nous",) # nous has claude models but shouldn't be suggested + assert result[0] not in {"nous",} # nous has claude models but shouldn't be suggested class TestIsNousFreeTier: diff --git a/tests/hermes_cli/test_opencode_go_in_model_list.py b/tests/hermes_cli/test_opencode_go_in_model_list.py index 6020c817979..f784f75f31b 100644 --- a/tests/hermes_cli/test_opencode_go_in_model_list.py +++ b/tests/hermes_cli/test_opencode_go_in_model_list.py @@ -44,7 +44,7 @@ def test_opencode_go_appears_when_api_key_set(): # opencode-go can appear as "built-in" (from PROVIDER_TO_MODELS_DEV when # models.dev is reachable) or "hermes" (from HERMES_OVERLAYS fallback when # the API is unavailable, e.g. in CI). - assert opencode_go["source"] in ("built-in", "hermes") + assert opencode_go["source"] in {"built-in", "hermes"} def test_opencode_go_not_appears_when_no_creds(): diff --git a/tests/hermes_cli/test_profile_describer.py b/tests/hermes_cli/test_profile_describer.py new file mode 100644 index 00000000000..3fc5fa3a6be --- /dev/null +++ b/tests/hermes_cli/test_profile_describer.py @@ -0,0 +1,168 @@ +"""Tests for the profile.yaml metadata layer (description + description_auto) +and the profile_describer LLM module. +""" + +from __future__ import annotations + +import json as jsonlib +from pathlib import Path +from unittest.mock import MagicMock, patch + +import pytest + +from hermes_cli import profiles as profiles_mod +from hermes_cli import profile_describer as describer + + +@pytest.fixture +def profile_env(tmp_path, monkeypatch): + """Set up an isolated HERMES_HOME with a default profile dir.""" + home = tmp_path / ".hermes" + home.mkdir() + monkeypatch.setenv("HERMES_HOME", str(home)) + monkeypatch.setattr(Path, "home", lambda: tmp_path) + return home + + +def test_read_profile_meta_empty_when_missing(profile_env): + meta = profiles_mod.read_profile_meta(profile_env) + assert meta == {"description": "", "description_auto": False} + + +def test_write_and_read_profile_meta(profile_env): + profiles_mod.write_profile_meta( + profile_env, + description="a useful researcher", + description_auto=False, + ) + meta = profiles_mod.read_profile_meta(profile_env) + assert meta["description"] == "a useful researcher" + assert meta["description_auto"] is False + + +def test_write_profile_meta_preserves_other_fields(profile_env): + # First write sets description_auto=True; second write only updates + # description and leaves description_auto unchanged. + profiles_mod.write_profile_meta( + profile_env, + description="auto-gen", + description_auto=True, + ) + profiles_mod.write_profile_meta(profile_env, description="edited by hand") + meta = profiles_mod.read_profile_meta(profile_env) + assert meta["description"] == "edited by hand" + assert meta["description_auto"] is True + + +def test_write_profile_meta_rejects_missing_dir(tmp_path): + bogus = tmp_path / "does_not_exist" + with pytest.raises(FileNotFoundError): + profiles_mod.write_profile_meta(bogus, description="x") + + +def test_read_profile_meta_tolerates_corrupt_yaml(profile_env): + (profile_env / "profile.yaml").write_text("not: valid: yaml: [unclosed") + meta = profiles_mod.read_profile_meta(profile_env) + assert meta == {"description": "", "description_auto": False} + + +# --------------------------------------------------------------------------- +# profile_describer module +# --------------------------------------------------------------------------- + + +def _fake_aux_response(content: str): + resp = MagicMock() + resp.choices = [MagicMock()] + resp.choices[0].message.content = content + return resp + + +def _patch_aux_client(content: str): + client = MagicMock() + client.chat.completions.create = MagicMock(return_value=_fake_aux_response(content)) + return patch( + "agent.auxiliary_client.get_text_auxiliary_client", + return_value=(client, "test-model"), + ) + + +def test_describer_writes_description_with_auto_true(profile_env, monkeypatch): + # Pretend "myprof" is a registered profile pointing at profile_env. + monkeypatch.setattr( + profiles_mod, "profile_exists", lambda n: n == "myprof", + ) + monkeypatch.setattr( + profiles_mod, "normalize_profile_name", lambda n: n, + ) + monkeypatch.setattr( + profiles_mod, "get_profile_dir", lambda n: profile_env, + ) + + payload = jsonlib.dumps({"description": "writes Python codebases"}) + with _patch_aux_client(payload), patch( + "agent.auxiliary_client.get_auxiliary_extra_body", return_value={} + ): + outcome = describer.describe_profile("myprof") + + assert outcome.ok, outcome.reason + assert outcome.description == "writes Python codebases" + meta = profiles_mod.read_profile_meta(profile_env) + assert meta["description"] == "writes Python codebases" + assert meta["description_auto"] is True + + +def test_describer_refuses_to_overwrite_user_authored(profile_env, monkeypatch): + profiles_mod.write_profile_meta( + profile_env, description="curated", description_auto=False, + ) + monkeypatch.setattr(profiles_mod, "profile_exists", lambda n: n == "myprof") + monkeypatch.setattr(profiles_mod, "normalize_profile_name", lambda n: n) + monkeypatch.setattr(profiles_mod, "get_profile_dir", lambda n: profile_env) + + outcome = describer.describe_profile("myprof") + assert outcome.ok is False + assert "already has a user-authored description" in outcome.reason + # Description unchanged + assert profiles_mod.read_profile_meta(profile_env)["description"] == "curated" + + +def test_describer_overwrite_flag_replaces_user_authored(profile_env, monkeypatch): + profiles_mod.write_profile_meta( + profile_env, description="curated", description_auto=False, + ) + monkeypatch.setattr(profiles_mod, "profile_exists", lambda n: n == "myprof") + monkeypatch.setattr(profiles_mod, "normalize_profile_name", lambda n: n) + monkeypatch.setattr(profiles_mod, "get_profile_dir", lambda n: profile_env) + + payload = jsonlib.dumps({"description": "new auto-gen"}) + with _patch_aux_client(payload), patch( + "agent.auxiliary_client.get_auxiliary_extra_body", return_value={} + ): + outcome = describer.describe_profile("myprof", overwrite=True) + assert outcome.ok, outcome.reason + meta = profiles_mod.read_profile_meta(profile_env) + assert meta["description"] == "new auto-gen" + assert meta["description_auto"] is True + + +def test_describer_handles_malformed_llm_response(profile_env, monkeypatch): + monkeypatch.setattr(profiles_mod, "profile_exists", lambda n: n == "myprof") + monkeypatch.setattr(profiles_mod, "normalize_profile_name", lambda n: n) + monkeypatch.setattr(profiles_mod, "get_profile_dir", lambda n: profile_env) + + # Non-JSON: describer falls back to taking the first paragraph as the description. + with _patch_aux_client("Plain text description that sneaks in"), patch( + "agent.auxiliary_client.get_auxiliary_extra_body", return_value={} + ): + outcome = describer.describe_profile("myprof") + assert outcome.ok + assert "Plain text description" in (outcome.description or "") + + +def test_describer_returns_false_when_profile_missing(profile_env, monkeypatch): + monkeypatch.setattr(profiles_mod, "profile_exists", lambda n: False) + monkeypatch.setattr(profiles_mod, "normalize_profile_name", lambda n: n) + outcome = describer.describe_profile("ghost") + assert outcome.ok is False + assert "not found" in outcome.reason diff --git a/tests/hermes_cli/test_proxy.py b/tests/hermes_cli/test_proxy.py index 0c874facac7..34a10bfa5ff 100644 --- a/tests/hermes_cli/test_proxy.py +++ b/tests/hermes_cli/test_proxy.py @@ -103,7 +103,7 @@ def test_nous_adapter_authenticated_with_refresh_token_only(tmp_path, monkeypatc assert NousPortalAdapter().is_authenticated() -def test_nous_adapter_get_credential_refreshes_and_persists(tmp_path, monkeypatch): +def test_nous_adapter_get_credential_uses_runtime_resolver(tmp_path, monkeypatch): monkeypatch.setenv("HERMES_HOME", str(tmp_path)) _write_auth_store(tmp_path, { "access_token": "access-tok", @@ -114,31 +114,82 @@ def test_nous_adapter_get_credential_refreshes_and_persists(tmp_path, monkeypatc }) refreshed_state = { - "access_token": "access-tok", - "refresh_token": "refresh-tok", - "client_id": "hermes-cli", - "portal_base_url": "https://portal.nousresearch.com", - "inference_base_url": "https://inference-api.nousresearch.com/v1", - "agent_key": "minted-bearer", - "agent_key_expires_at": "2099-01-01T00:00:00Z", + "api_key": "minted-bearer", + "base_url": "https://inference-api.nousresearch.com/v1", + "expires_at": "2099-01-01T00:00:00Z", } with patch( - "hermes_cli.proxy.adapters.nous_portal.refresh_nous_oauth_from_state", + "hermes_cli.proxy.adapters.nous_portal.resolve_nous_runtime_credentials", return_value=refreshed_state, - ) as mock_refresh: + ) as mock_resolve: adapter = NousPortalAdapter() cred = adapter.get_credential() - mock_refresh.assert_called_once() + mock_resolve.assert_called_once() assert cred.bearer == "minted-bearer" assert cred.base_url == "https://inference-api.nousresearch.com/v1" assert cred.expires_at == "2099-01-01T00:00:00Z" assert cred.token_type == "Bearer" - # Verify state was persisted back - stored = json.loads((tmp_path / "auth.json").read_text()) - assert stored["providers"]["nous"]["agent_key"] == "minted-bearer" + +def test_nous_adapter_retry_credential_forces_legacy_mint(tmp_path, monkeypatch): + monkeypatch.setenv("HERMES_HOME", str(tmp_path)) + _write_auth_store(tmp_path, { + "access_token": "jwt-access", + "refresh_token": "refresh-tok", + "client_id": "hermes-cli", + "portal_base_url": "https://portal.nousresearch.com", + "inference_base_url": "https://inference-api.nousresearch.com/v1", + "agent_key": "jwt-access", + }) + + refreshed_state = { + "api_key": "legacy-bearer", + "base_url": "https://inference-api.nousresearch.com/v1", + "expires_at": "2099-01-01T00:00:00Z", + } + + with patch( + "hermes_cli.proxy.adapters.nous_portal.resolve_nous_runtime_credentials", + return_value=refreshed_state, + ) as mock_resolve: + adapter = NousPortalAdapter() + cred = adapter.get_retry_credential( + failed_credential=UpstreamCredential( + bearer="header.jwt.signature", + base_url="https://inference-api.nousresearch.com/v1", + ), + status_code=401, + ) + + assert cred is not None + assert cred.bearer == "legacy-bearer" + assert mock_resolve.call_args.kwargs["inference_auth_mode"] == "legacy" + + +def test_nous_adapter_retry_credential_skips_opaque_bearer(tmp_path, monkeypatch): + monkeypatch.setenv("HERMES_HOME", str(tmp_path)) + _write_auth_store(tmp_path, { + "access_token": "jwt-access", + "refresh_token": "refresh-tok", + "agent_key": "opaque-bearer", + }) + + with patch( + "hermes_cli.proxy.adapters.nous_portal.resolve_nous_runtime_credentials", + ) as mock_resolve: + adapter = NousPortalAdapter() + cred = adapter.get_retry_credential( + failed_credential=UpstreamCredential( + bearer="opaque-bearer", + base_url="https://inference-api.nousresearch.com/v1", + ), + status_code=401, + ) + + assert cred is None + mock_resolve.assert_not_called() def test_nous_adapter_get_credential_raises_when_not_logged_in(tmp_path, monkeypatch): @@ -156,7 +207,7 @@ def test_nous_adapter_get_credential_raises_on_refresh_failure(tmp_path, monkeyp }) with patch( - "hermes_cli.proxy.adapters.nous_portal.refresh_nous_oauth_from_state", + "hermes_cli.proxy.adapters.nous_portal.resolve_nous_runtime_credentials", side_effect=RuntimeError("Refresh session has been revoked"), ): adapter = NousPortalAdapter() @@ -164,6 +215,40 @@ def test_nous_adapter_get_credential_raises_on_refresh_failure(tmp_path, monkeyp adapter.get_credential() +def test_nous_adapter_quarantines_terminal_refresh_failure(tmp_path, monkeypatch): + from hermes_cli.auth import AuthError + from agent.credential_pool import load_pool + + monkeypatch.setenv("HERMES_HOME", str(tmp_path)) + _write_auth_store(tmp_path, { + "access_token": "access-tok", + "refresh_token": "refresh-tok", + "agent_key": "stale-agent-key", + }) + assert load_pool("nous").select() is not None + + with patch( + "hermes_cli.proxy.adapters.nous_portal.resolve_nous_runtime_credentials", + side_effect=AuthError( + "Refresh session has been revoked", + provider="nous", + code="invalid_grant", + relogin_required=True, + ), + ): + adapter = NousPortalAdapter() + with pytest.raises(RuntimeError, match="Refresh session has been revoked"): + adapter.get_credential() + + stored = json.loads((tmp_path / "auth.json").read_text()) + nous_state = stored["providers"]["nous"] + assert not nous_state.get("refresh_token") + assert not nous_state.get("access_token") + assert not nous_state.get("agent_key") + assert nous_state["last_auth_error"]["code"] == "invalid_grant" + assert stored.get("credential_pool", {}).get("nous") == [] + + def test_nous_adapter_get_credential_raises_when_no_agent_key_returned(tmp_path, monkeypatch): """If the refresh helper succeeds but produces no agent_key, we surface a clear error.""" monkeypatch.setenv("HERMES_HOME", str(tmp_path)) @@ -173,7 +258,7 @@ def test_nous_adapter_get_credential_raises_when_no_agent_key_returned(tmp_path, }) with patch( - "hermes_cli.proxy.adapters.nous_portal.refresh_nous_oauth_from_state", + "hermes_cli.proxy.adapters.nous_portal.resolve_nous_runtime_credentials", return_value={"access_token": "a", "refresh_token": "r"}, ): adapter = NousPortalAdapter() @@ -194,7 +279,7 @@ def test_nous_adapter_concurrent_refresh_serialized(tmp_path, monkeypatch): counter = [0] counter_lock = threading.Lock() - def serializing_refresh(state, **kwargs): + def serializing_refresh(**kwargs): # If another thread is already inside refresh, the lock is broken. if in_flight.is_set(): overlap_detected.set() @@ -208,10 +293,9 @@ def test_nous_adapter_concurrent_refresh_serialized(tmp_path, monkeypatch): counter[0] += 1 idx = counter[0] return { - **state, - "agent_key": f"key-{idx}", - "agent_key_expires_at": "2099-01-01T00:00:00Z", - "inference_base_url": "https://inference-api.nousresearch.com/v1", + "api_key": f"key-{idx}", + "expires_at": "2099-01-01T00:00:00Z", + "base_url": "https://inference-api.nousresearch.com/v1", } finally: in_flight.clear() @@ -227,7 +311,7 @@ def test_nous_adapter_concurrent_refresh_serialized(tmp_path, monkeypatch): errors.append(exc) with patch( - "hermes_cli.proxy.adapters.nous_portal.refresh_nous_oauth_from_state", + "hermes_cli.proxy.adapters.nous_portal.resolve_nous_runtime_credentials", side_effect=serializing_refresh, ): threads = [threading.Thread(target=worker) for _ in range(3)] @@ -260,12 +344,15 @@ class FakeAdapter(UpstreamAdapter): """A test adapter that returns a fixed credential without touching disk.""" def __init__(self, base_url: str, bearer: str = "test-bearer", - allowed=None, raise_on_credential=False): + allowed=None, raise_on_credential=False, + retry_bearer: str | None = None): self._base_url = base_url self._bearer = bearer self._allowed = frozenset(allowed or ["/chat/completions"]) self._raise = raise_on_credential + self._retry_bearer = retry_bearer self.calls = 0 + self.retry_calls = 0 @property def name(self): return "fake" @@ -287,6 +374,17 @@ class FakeAdapter(UpstreamAdapter): expires_at="2099-01-01T00:00:00Z", ) + def get_retry_credential(self, *, failed_credential, status_code): + _ = failed_credential + self.retry_calls += 1 + if status_code != 401 or not self._retry_bearer: + return None + return UpstreamCredential( + bearer=self._retry_bearer, + base_url=self._base_url, + expires_at="2099-01-01T00:00:00Z", + ) + async def _start_runner(app: "web.Application"): """Spin up an aiohttp app on an ephemeral localhost port. Returns (runner, base_url).""" @@ -327,6 +425,25 @@ def _build_fake_upstream(captured: Dict[str, Any]) -> "web.Application": return app +def _build_retrying_fake_upstream(captured: Dict[str, Any]) -> "web.Application": + async def maybe_unauthorized(request): + body = await request.read() + auth = request.headers.get("Authorization") + captured["requests"].append({ + "method": request.method, + "path": request.path, + "auth": auth, + "body": body.decode("utf-8") if body else "", + }) + if auth == "Bearer jwt-bearer": + return web.json_response({"error": "bad token"}, status=401) + return web.json_response({"ok": True}) + + app = web.Application() + app.router.add_route("*", "/v1/chat/completions", maybe_unauthorized) + return app + + def test_server_forwards_chat_completions(): async def run(): captured: Dict[str, Any] = {"requests": []} @@ -357,6 +474,41 @@ def test_server_forwards_chat_completions(): asyncio.run(run()) +def test_server_retries_once_with_adapter_retry_credential_on_401(): + async def run(): + captured: Dict[str, Any] = {"requests": []} + upstream_runner, upstream_base = await _start_runner( + _build_retrying_fake_upstream(captured) + ) + adapter = FakeAdapter( + f"{upstream_base}/v1", + bearer="jwt-bearer", + retry_bearer="legacy-bearer", + ) + proxy_runner, proxy_base = await _start_runner(create_app(adapter)) + + try: + async with aiohttp.ClientSession() as session: + async with session.post( + f"{proxy_base}/v1/chat/completions", + json={"model": "Hermes-4-70B"}, + ) as resp: + assert resp.status == 200 + data = await resp.json() + assert data["ok"] is True + + assert adapter.retry_calls == 1 + assert [req["auth"] for req in captured["requests"]] == [ + "Bearer jwt-bearer", + "Bearer legacy-bearer", + ] + finally: + await proxy_runner.cleanup() + await upstream_runner.cleanup() + + asyncio.run(run()) + + def test_server_rejects_disallowed_path(): async def run(): adapter = FakeAdapter("http://unused.example/v1", allowed=["/chat/completions"]) diff --git a/tests/hermes_cli/test_status.py b/tests/hermes_cli/test_status.py index a13e843faf8..3cee9ab10ba 100644 --- a/tests/hermes_cli/test_status.py +++ b/tests/hermes_cli/test_status.py @@ -29,6 +29,7 @@ def test_show_status_termux_gateway_section_skips_systemctl(monkeypatch, capsys, monkeypatch.setattr(status_mod, "provider_label", lambda provider: "OpenAI Codex", raising=False) monkeypatch.setattr(auth_mod, "get_nous_auth_status", lambda: {}, raising=False) monkeypatch.setattr(auth_mod, "get_codex_auth_status", lambda: {}, raising=False) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", lambda: {}, raising=False) monkeypatch.setattr(gateway_mod, "find_gateway_pids", lambda exclude_pids=None: [], raising=False) def _unexpected_systemctl(*args, **kwargs): @@ -70,6 +71,7 @@ def test_show_status_reports_nous_auth_error(monkeypatch, capsys, tmp_path): ) monkeypatch.setattr(auth_mod, "get_codex_auth_status", lambda: {}, raising=False) monkeypatch.setattr(auth_mod, "get_qwen_auth_status", lambda: {}, raising=False) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", lambda: {}, raising=False) monkeypatch.setattr(gateway_mod, "find_gateway_pids", lambda exclude_pids=None: [], raising=False) status_mod.show_status(SimpleNamespace(all=False, deep=False)) @@ -96,6 +98,7 @@ def test_show_status_reports_vercel_backend_contract(monkeypatch, capsys, tmp_pa monkeypatch.setattr(auth_mod, "get_nous_auth_status", lambda: {}, raising=False) monkeypatch.setattr(auth_mod, "get_codex_auth_status", lambda: {}, raising=False) monkeypatch.setattr(auth_mod, "get_qwen_auth_status", lambda: {}, raising=False) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", lambda: {}, raising=False) monkeypatch.setattr(gateway_mod, "find_gateway_pids", lambda exclude_pids=None: [], raising=False) status_mod.show_status(SimpleNamespace(all=False, deep=False)) @@ -109,3 +112,223 @@ def test_show_status_reports_vercel_backend_contract(monkeypatch, capsys, tmp_pa assert "oidc-token" not in output assert "snapshot filesystem" in output assert "live processes do not survive" in output + + +# --------------------------------------------------------------------------- +# Helpers shared by xAI OAuth status tests +# --------------------------------------------------------------------------- + +def _base_xai_mocks(monkeypatch, tmp_path): + """Set up the minimal environment for show_status, returning status_mod.""" + from hermes_cli import status as status_mod + import hermes_cli.auth as auth_mod + import hermes_cli.gateway as gateway_mod + + monkeypatch.setattr(status_mod, "get_env_path", lambda: tmp_path / ".env", raising=False) + monkeypatch.setattr(status_mod, "get_hermes_home", lambda: tmp_path, raising=False) + monkeypatch.setattr(status_mod, "load_config", lambda: {"model": "gpt-5.4"}, raising=False) + monkeypatch.setattr(status_mod, "resolve_requested_provider", lambda requested=None: "openai-codex", raising=False) + monkeypatch.setattr(status_mod, "resolve_provider", lambda requested=None, **kwargs: "openai-codex", raising=False) + monkeypatch.setattr(status_mod, "provider_label", lambda provider: "OpenAI Codex", raising=False) + monkeypatch.setattr(auth_mod, "get_nous_auth_status", lambda: {}, raising=False) + monkeypatch.setattr(auth_mod, "get_codex_auth_status", lambda: {}, raising=False) + monkeypatch.setattr(auth_mod, "get_qwen_auth_status", lambda: {}, raising=False) + monkeypatch.setattr(auth_mod, "get_minimax_oauth_auth_status", lambda: {}, raising=False) + monkeypatch.setattr(gateway_mod, "find_gateway_pids", lambda exclude_pids=None: [], raising=False) + return status_mod + + +class TestShowStatusXaiOAuth: + """xAI OAuth row in hermes status.""" + + # ------------------------------------------------------------------ + # Logged-in branch + # ------------------------------------------------------------------ + + def test_logged_in_shows_check_mark_and_label(self, monkeypatch, capsys, tmp_path): + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: {"logged_in": True, "auth_store": "/a/auth.json"}, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "xAI OAuth" in out + # The logged-in label must appear; the "not logged in" label must not + assert "✓" in out or "logged in" in out + assert "not logged in" not in out.split("xAI OAuth", 1)[1].split("\n")[0] + + def test_logged_in_shows_auth_store(self, monkeypatch, capsys, tmp_path): + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: {"logged_in": True, "auth_store": "/home/u/.hermes/auth.json"}, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "Auth file: /home/u/.hermes/auth.json" in out + + def test_logged_in_shows_last_refresh(self, monkeypatch, capsys, tmp_path): + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: { + "logged_in": True, + "auth_store": "/a/auth.json", + "last_refresh": "2026-05-17T10:00:00+00:00", + }, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "Refreshed:" in out + + def test_logged_in_does_not_show_error_line(self, monkeypatch, capsys, tmp_path): + """Error field must be suppressed when logged_in is True.""" + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: { + "logged_in": True, + "auth_store": "/a/auth.json", + "error": "stale-error-must-not-appear", + }, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + xai_section = out.split("xAI OAuth", 1)[1] + assert "stale-error-must-not-appear" not in xai_section + + def test_no_auth_store_line_when_field_absent(self, monkeypatch, capsys, tmp_path): + """Auth file line must not appear when auth_store is missing.""" + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: {"logged_in": True}, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + xai_section = out.split("xAI OAuth", 1)[1].split("◆", 1)[0] + assert "Auth file:" not in xai_section + + def test_no_refreshed_line_when_last_refresh_absent(self, monkeypatch, capsys, tmp_path): + """Refreshed line must not appear when last_refresh is not present.""" + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: {"logged_in": True, "auth_store": "/a/auth.json"}, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + xai_section = out.split("xAI OAuth", 1)[1].split("◆", 1)[0] + assert "Refreshed:" not in xai_section + + # ------------------------------------------------------------------ + # Not-logged-in branch + # ------------------------------------------------------------------ + + def test_not_logged_in_shows_login_command(self, monkeypatch, capsys, tmp_path): + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: {"logged_in": False, "error": "no credentials"}, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "not logged in (run: hermes auth add xai-oauth)" in out + + def test_not_logged_in_shows_error(self, monkeypatch, capsys, tmp_path): + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: {"logged_in": False, "error": "Token has expired"}, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "Error: Token has expired" in out + + def test_not_logged_in_omits_error_line_when_error_absent(self, monkeypatch, capsys, tmp_path): + """No Error: line when not logged in but error key is missing.""" + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: {"logged_in": False}, + raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + xai_section = out.split("xAI OAuth", 1)[1].split("◆", 1)[0] + assert "Error:" not in xai_section + + # ------------------------------------------------------------------ + # Resilience: import failure and runtime exception + # ------------------------------------------------------------------ + + def test_import_failure_does_not_crash_show_status(self, monkeypatch, capsys, tmp_path): + """show_status must complete even when get_xai_oauth_auth_status cannot be imported.""" + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.delattr(auth_mod, "get_xai_oauth_auth_status", raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "◆ Auth Providers" in out + + def test_import_failure_does_not_break_other_oauth_providers(self, monkeypatch, capsys, tmp_path): + """Nous/Codex/MiniMax rows must still appear when xAI import fails.""" + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_nous_auth_status", + lambda: {"logged_in": True}, raising=False) + monkeypatch.delattr(auth_mod, "get_xai_oauth_auth_status", raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "Nous Portal" in out + assert "MiniMax OAuth" in out + + def test_status_function_exception_does_not_crash(self, monkeypatch, capsys, tmp_path): + """show_status must not propagate an exception raised by get_xai_oauth_auth_status.""" + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + + def _raises(): + raise RuntimeError("backend unreachable") + + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", _raises, raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "◆ Auth Providers" in out + + def test_status_function_returns_none_does_not_crash(self, monkeypatch, capsys, tmp_path): + """get_xai_oauth_auth_status returning None must be handled gracefully.""" + import hermes_cli.auth as auth_mod + status_mod = _base_xai_mocks(monkeypatch, tmp_path) + monkeypatch.setattr(auth_mod, "get_xai_oauth_auth_status", + lambda: None, raising=False) + + status_mod.show_status(SimpleNamespace(all=False, deep=False)) + out = capsys.readouterr().out + + assert "xAI OAuth" in out + assert "not logged in (run: hermes auth add xai-oauth)" in out diff --git a/tests/hermes_cli/test_tools_config.py b/tests/hermes_cli/test_tools_config.py index 8a94ce4302f..787292d83a4 100644 --- a/tests/hermes_cli/test_tools_config.py +++ b/tests/hermes_cli/test_tools_config.py @@ -125,6 +125,62 @@ def test_get_platform_tools_homeassistant_toolset_off_for_cron_when_hass_token_m assert "homeassistant" not in cron_enabled +def test_get_platform_tools_x_search_auto_enabled_when_xai_oauth_present(monkeypatch): + """x_search toolset auto-enables across platforms when xAI Grok OAuth + tokens are present, mirroring the HASS_TOKEN → homeassistant rule. + + The user already authenticated via SuperGrok OAuth; they shouldn't have + to also click through `hermes tools` → X (Twitter) Search to flip the + toolset on. Tool's check_fn still gates schema registration if creds + later go missing. + """ + monkeypatch.delenv("XAI_API_KEY", raising=False) + monkeypatch.setattr( + "hermes_cli.tools_config._xai_credentials_present", lambda: True + ) + + for plat in ("cli", "cron", "telegram"): + enabled = _get_platform_tools({}, plat) + assert "x_search" in enabled, f"x_search missing for {plat}" + + +def test_get_platform_tools_x_search_auto_enabled_when_xai_api_key_present(monkeypatch): + """x_search toolset auto-enables when XAI_API_KEY is set, even without + OAuth tokens — the API-key path is a supported credential source.""" + monkeypatch.setenv("XAI_API_KEY", "fake-xai-key") + + cli_enabled = _get_platform_tools({}, "cli") + assert "x_search" in cli_enabled + + +def test_get_platform_tools_x_search_off_when_no_xai_credentials(monkeypatch): + """Without any xAI credentials, x_search stays off — preserves the + "don't ship the schema to users who can't use it" default.""" + monkeypatch.delenv("XAI_API_KEY", raising=False) + monkeypatch.setattr( + "hermes_cli.tools_config._xai_credentials_present", lambda: False + ) + + cli_enabled = _get_platform_tools({}, "cli") + assert "x_search" not in cli_enabled + + +def test_get_platform_tools_x_search_respects_explicit_config(monkeypatch): + """Once the user has saved an explicit toolset list via `hermes tools`, + that list is authoritative — x_search auto-enable does NOT fire even + when xAI creds exist. The saved list represents deliberate choices.""" + monkeypatch.delenv("XAI_API_KEY", raising=False) + monkeypatch.setattr( + "hermes_cli.tools_config._xai_credentials_present", lambda: True + ) + + # User explicitly opted into spotify but not x_search via `hermes tools`. + config = {"platform_toolsets": {"cli": ["hermes-cli", "spotify"]}} + enabled = _get_platform_tools(config, "cli") + assert "x_search" not in enabled + assert "spotify" in enabled + + def test_get_platform_tools_expands_composite_when_mixed_with_configurable(): """``[hermes-cli, spotify]`` (composite + configurable) must keep the full ``hermes-cli`` toolset alongside the explicit Spotify opt-in. The @@ -989,3 +1045,27 @@ def test_reconfigure_browser_provider_overwrites_stale_use_gateway(): provider = {"name": "Browserbase", "browser_provider": "browserbase", "env_vars": []} _reconfigure_provider(provider, config) assert config["browser"]["use_gateway"] is False + + +@pytest.mark.parametrize("provider_name,post_setup_key", [ + ("Camofox", "camofox"), +]) +def test_reconfigure_provider_runs_post_setup_for_env_var_providers( + monkeypatch, provider_name, post_setup_key +): + """_reconfigure_provider() must call _run_post_setup() for providers that have + both env_vars and post_setup — parity with _configure_provider() line 2286.""" + called = [] + monkeypatch.setattr("hermes_cli.tools_config._run_post_setup", lambda key: called.append(key)) + monkeypatch.setattr("hermes_cli.tools_config.get_env_value", lambda k: None) + monkeypatch.setattr("hermes_cli.tools_config._prompt", lambda *a, **kw: "") + monkeypatch.setattr("hermes_cli.tools_config.save_env_value", lambda k, v: None) + + provider = next( + p + for p in TOOL_CATEGORIES["browser"]["providers"] + if p["name"] == provider_name + ) + _reconfigure_provider(provider, {}) + + assert called == [post_setup_key] diff --git a/tests/hermes_cli/test_update_stale_dashboard.py b/tests/hermes_cli/test_update_stale_dashboard.py index 546fd489911..e79caeb9dc6 100644 --- a/tests/hermes_cli/test_update_stale_dashboard.py +++ b/tests/hermes_cli/test_update_stale_dashboard.py @@ -237,7 +237,7 @@ class TestKillStaleDashboardPosix: sent.append((pid, sig)) # Simulate stubborn process: probe (sig 0) always succeeds, # SIGTERM does nothing, SIGKILL is where it "dies". - if sig in (_signal.SIGTERM, 0, _signal.SIGKILL): + if sig in {_signal.SIGTERM, 0, _signal.SIGKILL}: return # Any other signal — also fine. diff --git a/tests/hermes_cli/test_web_oauth_dispatch.py b/tests/hermes_cli/test_web_oauth_dispatch.py index 23b72a303cf..b9ee20ccae8 100644 --- a/tests/hermes_cli/test_web_oauth_dispatch.py +++ b/tests/hermes_cli/test_web_oauth_dispatch.py @@ -19,11 +19,12 @@ The fix: These tests pin the corrected behavior. """ +import asyncio import time from datetime import datetime, timezone from unittest.mock import patch -import pytest +import httpx from fastapi.testclient import TestClient from hermes_cli.web_server import _SESSION_TOKEN, app @@ -32,6 +33,32 @@ client = TestClient(app) HEADERS = {"X-Hermes-Session-Token": _SESSION_TOKEN} +def _fake_nous_device_data(): + return { + "device_code": "device-code", + "user_code": "NOUS-1234", + "verification_uri": "https://portal.nousresearch.com/device", + "verification_uri_complete": ( + "https://portal.nousresearch.com/device?user_code=NOUS-1234" + ), + "expires_in": 600, + "interval": 5, + } + + +def _invoke_scope_refusal(): + request = httpx.Request("POST", "https://portal.nousresearch.com/oauth/device/code") + response = httpx.Response( + 400, + json={ + "error": "invalid_scope", + "error_description": "unsupported scope inference:invoke", + }, + request=request, + ) + return httpx.HTTPStatusError("invalid scope", request=request, response=response) + + def test_minimax_login_does_not_launch_anthropic_flow(): """Click 'Login' on MiniMax → MUST NOT return claude.ai auth_url.""" fake_user_code_resp = { @@ -48,6 +75,9 @@ def test_minimax_login_does_not_launch_anthropic_flow(): ), patch( "hermes_cli.auth._minimax_pkce_pair", return_value=("verifier-stub", "challenge-stub", "stub-state"), + ), patch( + "hermes_cli.web_server._minimax_poller", + return_value=None, ): resp = client.post( "/api/providers/oauth/minimax-oauth/start", @@ -69,6 +99,113 @@ def test_minimax_login_does_not_launch_anthropic_flow(): assert body["expires_in"] == 600 +def test_nous_dashboard_device_flow_honors_legacy_scope_override(monkeypatch): + from hermes_cli import auth as auth_mod + from hermes_cli import web_server as ws + + requested_scopes = [] + + def fake_request_device_code(**kwargs): + requested_scopes.append(kwargs["scope"]) + return _fake_nous_device_data() + + monkeypatch.setenv(auth_mod.NOUS_LEGACY_SESSION_KEYS_ENV, "true") + monkeypatch.setattr(auth_mod, "_request_device_code", fake_request_device_code) + monkeypatch.setattr(ws, "_nous_poller", lambda sid: None) + + result = asyncio.run(ws._start_device_code_flow("nous")) + try: + assert requested_scopes == [auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE] + assert result["flow"] == "device_code" + assert result["user_code"] == "NOUS-1234" + assert ( + ws._oauth_sessions[result["session_id"]]["scope"] + == auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE + ) + finally: + ws._oauth_sessions.pop(result["session_id"], None) + + +def test_nous_dashboard_device_flow_retries_legacy_scope_on_invoke_refusal(monkeypatch): + from hermes_cli import auth as auth_mod + from hermes_cli import web_server as ws + + requested_scopes = [] + + def fake_request_device_code(**kwargs): + requested_scopes.append(kwargs["scope"]) + if len(requested_scopes) == 1: + raise _invoke_scope_refusal() + return _fake_nous_device_data() + + monkeypatch.delenv(auth_mod.NOUS_LEGACY_SESSION_KEYS_ENV, raising=False) + monkeypatch.setattr(auth_mod, "_request_device_code", fake_request_device_code) + monkeypatch.setattr(ws, "_nous_poller", lambda sid: None) + + result = asyncio.run(ws._start_device_code_flow("nous")) + try: + assert requested_scopes == [ + auth_mod.DEFAULT_NOUS_SCOPE, + auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE, + ] + assert ( + ws._oauth_sessions[result["session_id"]]["scope"] + == auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE + ) + finally: + ws._oauth_sessions.pop(result["session_id"], None) + + +def test_nous_dashboard_poller_preserves_effective_scope_when_token_omits_scope(monkeypatch): + from hermes_cli import auth as auth_mod + from hermes_cli import web_server as ws + + session_id = "nous-effective-scope-test" + ws._oauth_sessions[session_id] = { + "session_id": session_id, + "provider": "nous", + "flow": "device_code", + "created_at": time.time(), + "status": "pending", + "error_message": None, + "portal_base_url": "https://portal.nousresearch.com", + "client_id": "hermes-cli", + "device_code": "device-code", + "interval": 5, + "expires_at": time.time() + 600, + "scope": auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE, + } + captured_state = {} + + def fake_refresh_nous_oauth_from_state(state, **kwargs): + captured_state.update(state) + return {**state, "agent_key": "legacy-agent-key"} + + monkeypatch.setattr( + auth_mod, + "_poll_for_token", + lambda **kwargs: { + "access_token": "access-token", + "refresh_token": "refresh-token", + "expires_in": 3600, + "token_type": "Bearer", + }, + ) + monkeypatch.setattr( + auth_mod, + "refresh_nous_oauth_from_state", + fake_refresh_nous_oauth_from_state, + ) + monkeypatch.setattr(auth_mod, "persist_nous_credentials", lambda state: None) + + try: + ws._nous_poller(session_id) + assert captured_state["scope"] == auth_mod.NOUS_LEGACY_AGENT_KEY_SCOPE + assert ws._oauth_sessions[session_id]["status"] == "approved" + finally: + ws._oauth_sessions.pop(session_id, None) + + def test_minimax_dashboard_poller_accepts_absolute_ms_expired_in(): """Dashboard MiniMax completion must accept unix-ms token expiry values.""" from hermes_cli import web_server as ws diff --git a/tests/hermes_cli/test_web_server.py b/tests/hermes_cli/test_web_server.py index 61de9145acc..f7f120b6356 100644 --- a/tests/hermes_cli/test_web_server.py +++ b/tests/hermes_cli/test_web_server.py @@ -449,7 +449,7 @@ class TestWebServerEndpoints: resp = self.client.get("/api/auth/session-token") # The endpoint is gone — the catch-all SPA route serves index.html # or the middleware returns 401 for unauthenticated /api/ paths. - assert resp.status_code in (200, 404) + assert resp.status_code in {200, 404} # Either way, it must NOT return the token as JSON try: data = resp.json() @@ -476,7 +476,7 @@ class TestWebServerEndpoints: # %2e%2e = .. resp = self.client.get("/%2e%2e/%2e%2e/etc/passwd") # Should return 200 with index.html (SPA fallback), not the actual file - assert resp.status_code in (200, 404) + assert resp.status_code in {200, 404} if resp.status_code == 200: # Should be the SPA fallback, not the system file assert "root:" not in resp.text @@ -484,7 +484,7 @@ class TestWebServerEndpoints: def test_path_traversal_dotdot_blocked(self): """Direct .. path traversal via encoded sequences.""" resp = self.client.get("/%2e%2e/hermes_cli/web_server.py") - assert resp.status_code in (200, 404) + assert resp.status_code in {200, 404} if resp.status_code == 200: assert "FastAPI" not in resp.text # Should not serve the actual source @@ -678,7 +678,7 @@ class TestConfigRoundTrip: if val is None: continue # not set in user config — fine expected = entry["type"] - if expected in ("string", "select") and not isinstance(val, str): + if expected in {"string", "select"} and not isinstance(val, str): mismatches.append(f"{key}: expected str, got {type(val).__name__}") elif expected == "number" and not isinstance(val, (int, float)): mismatches.append(f"{key}: expected number, got {type(val).__name__}") @@ -1175,7 +1175,7 @@ class TestNewEndpoints: """GET /api/auth/session-token no longer exists.""" resp = self.client.get("/api/auth/session-token") # Should not return a JSON token object - assert resp.status_code in (200, 404) + assert resp.status_code in {200, 404} try: data = resp.json() assert "token" not in data diff --git a/tests/hermes_cli/test_xai_oauth_pkce_token_exchange.py b/tests/hermes_cli/test_xai_oauth_pkce_token_exchange.py new file mode 100644 index 00000000000..98b81ff140e --- /dev/null +++ b/tests/hermes_cli/test_xai_oauth_pkce_token_exchange.py @@ -0,0 +1,359 @@ +"""Regression coverage for xAI OAuth PKCE token exchange (issue #26990). + +Issue [#26990] reported that ``hermes auth add xai-oauth`` succeeds at the +browser-side authorize step but fails at the token endpoint with +``code_challenge is required`` — the symptom of an OAuth server that +re-validates PKCE at the token step instead of relying purely on +state captured during the authorize redirect. + +The fix in ``hermes_cli/auth.py`` extracts the token POST into +:func:`_xai_oauth_exchange_code_for_tokens` and: + +* Sends ``code_verifier`` (RFC 7636 §4.5 requirement). +* **Also** echoes ``code_challenge`` and ``code_challenge_method`` + in the request body as defense-in-depth — strictly compliant + servers ignore extras at the token endpoint, but xAI's server + needs them. +* Refuses to fire the POST locally when ``code_verifier`` is empty + (avoids leaking the auth code to a server that can't redeem it). +* Surfaces the HTTP status code prominently in the error message so + users / maintainers can tell a 400 (bad request) from a 403 + (entitlement denied) at a glance. + +These tests pin all three behaviors so the fix can't silently regress. +""" + +from __future__ import annotations + +from typing import Any, Dict, List +from urllib.parse import parse_qs + +import httpx +import pytest + +from hermes_cli.auth import ( + AuthError, + XAI_OAUTH_CLIENT_ID, + _xai_oauth_exchange_code_for_tokens, +) + + +# --------------------------------------------------------------------------- +# httpx.post recorder +# --------------------------------------------------------------------------- + + +class _PostRecorder: + """Capture every ``httpx.post`` call without touching the network.""" + + def __init__(self, response: httpx.Response) -> None: + self.response = response + self.calls: List[Dict[str, Any]] = [] + + def __call__(self, url, *, headers=None, data=None, timeout=None, **kw): + self.calls.append( + {"url": url, "headers": headers or {}, "data": data or {}, + "timeout": timeout, "extra": kw} + ) + return self.response + + +def _ok_response(payload: dict) -> httpx.Response: + return httpx.Response(200, json=payload) + + +def _err_response(status: int, body: str) -> httpx.Response: + return httpx.Response(status, text=body) + + +@pytest.fixture +def post_recorder(monkeypatch): + """Default: 200 response with a full xAI token payload.""" + recorder = _PostRecorder( + _ok_response( + { + "access_token": "AT-fresh", + "refresh_token": "RT-fresh", + "id_token": "ID", + "expires_in": 3600, + "token_type": "Bearer", + } + ) + ) + monkeypatch.setattr("hermes_cli.auth.httpx.post", recorder) + return recorder + + +# --------------------------------------------------------------------------- +# Core contract: which fields go on the wire? +# --------------------------------------------------------------------------- + + +def test_token_exchange_includes_code_verifier(post_recorder): + """RFC 7636 §4.5 — ``code_verifier`` MUST be sent.""" + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="theVerifier_43_to_128_chars_____________________", + code_challenge="aBcDeF", + ) + sent = post_recorder.calls[-1]["data"] + assert sent["code_verifier"] == "theVerifier_43_to_128_chars_____________________" + + +def test_token_exchange_also_echoes_code_challenge_for_xai(post_recorder): + """Defense-in-depth for #26990 — xAI re-validates the challenge + at the token endpoint, not just at authorize. Without this echo + we get ``code_challenge is required`` even though we send a valid + ``code_verifier``.""" + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="aBcDeF", + ) + sent = post_recorder.calls[-1]["data"] + assert sent["code_challenge"] == "aBcDeF" + assert sent["code_challenge_method"] == "S256" + + +def test_token_exchange_uses_correct_grant_and_client(post_recorder): + """Lock the static fields too — a future refactor must not flip + these to ``client_credentials`` or drop ``client_id``.""" + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + ) + sent = post_recorder.calls[-1]["data"] + assert sent["grant_type"] == "authorization_code" + assert sent["code"] == "AUTHCODE" + assert sent["redirect_uri"] == "http://127.0.0.1:56121/callback" + assert sent["client_id"] == XAI_OAUTH_CLIENT_ID + + +def test_token_exchange_uses_form_urlencoded_content_type(post_recorder): + """xAI's token endpoint expects ``application/x-www-form-urlencoded``.""" + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + ) + headers = post_recorder.calls[-1]["headers"] + assert headers["Content-Type"] == "application/x-www-form-urlencoded" + assert headers["Accept"] == "application/json" + + +def test_token_exchange_targets_the_supplied_endpoint(post_recorder): + """Some test fixtures sniff the discovered token endpoint dynamically. + We must POST to the URL the caller passed, not a hard-coded constant.""" + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/some/other/token/path", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + ) + assert post_recorder.calls[-1]["url"] == "https://auth.x.ai/some/other/token/path" + + +def test_token_exchange_passes_timeout_through(post_recorder): + """Operators on slow networks pass a higher ``timeout_seconds``; + the helper must forward it (and bump the floor to 20s).""" + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + timeout_seconds=45.0, + ) + assert post_recorder.calls[-1]["timeout"] == 45.0 + + +def test_token_exchange_floor_timeout_is_20s(post_recorder): + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + timeout_seconds=2.0, + ) + assert post_recorder.calls[-1]["timeout"] == 20.0 + + +# --------------------------------------------------------------------------- +# Sanity guard: refuse to POST with an empty code_verifier +# --------------------------------------------------------------------------- + + +def test_empty_code_verifier_raises_without_posting(post_recorder): + """If ``code_verifier`` is somehow lost upstream, we must refuse to + send the request — leaking an authorization code to xAI without a + verifier is worse than failing locally with an actionable error.""" + with pytest.raises(AuthError) as exc_info: + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="", + code_challenge="c" * 43, + ) + assert exc_info.value.code == "xai_pkce_verifier_missing" + assert "26990" in str(exc_info.value) + # And critically: nothing was sent. + assert post_recorder.calls == [] + + +def test_missing_code_challenge_omits_echo_but_still_sends_verifier(post_recorder): + """``code_challenge`` is defensive — if a caller doesn't have it + handy, we must still send the standards-compliant request rather + than refusing. This keeps RFC-compliant servers happy.""" + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="", + ) + sent = post_recorder.calls[-1]["data"] + assert sent["code_verifier"] == "v" * 64 + assert "code_challenge" not in sent + assert "code_challenge_method" not in sent + + +# --------------------------------------------------------------------------- +# Error surfacing +# --------------------------------------------------------------------------- + + +def test_non_200_response_surfaces_status_and_body(monkeypatch): + """When xAI returns a 4xx, the operator needs both the HTTP status + code (to tell 400 from 401 from 403 at a glance) and the response + body (the actual server-side reason).""" + recorder = _PostRecorder( + _err_response(400, '{"error":"invalid_grant","error_description":"code_challenge is required"}') + ) + monkeypatch.setattr("hermes_cli.auth.httpx.post", recorder) + with pytest.raises(AuthError) as exc_info: + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + ) + msg = str(exc_info.value) + assert "HTTP 400" in msg, ( + "Status code must be in the error so callers can disambiguate " + "tier-denied (403) from bad-request (400) without inspecting " + "exc.code." + ) + assert "code_challenge is required" in msg + assert exc_info.value.code == "xai_token_exchange_failed" + + +def test_transport_error_wraps_as_auth_error(monkeypatch): + """A connection failure must come back as ``AuthError`` so the + surrounding ``format_auth_error`` UI mapping fires correctly.""" + + def _boom(*args, **kwargs): + raise httpx.ConnectError("dns failure") + + monkeypatch.setattr("hermes_cli.auth.httpx.post", _boom) + with pytest.raises(AuthError) as exc_info: + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + ) + assert exc_info.value.code == "xai_token_exchange_failed" + assert "dns failure" in str(exc_info.value) + + +def test_non_dict_payload_raises_invalid_json(monkeypatch): + """xAI returning ``[]`` or a string at 200 is a server bug — fail + with a precise error rather than crashing later in token storage.""" + recorder = _PostRecorder(_ok_response([1, 2, 3])) # type: ignore[arg-type] + monkeypatch.setattr("hermes_cli.auth.httpx.post", recorder) + with pytest.raises(AuthError) as exc_info: + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + ) + assert exc_info.value.code == "xai_token_exchange_invalid" + + +def test_success_returns_full_payload_dict(post_recorder): + """200 happy path: the parsed JSON dict comes back verbatim so the + caller can pluck ``access_token`` / ``refresh_token`` etc.""" + out = _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="v" * 64, + code_challenge="c" * 43, + ) + assert out["access_token"] == "AT-fresh" + assert out["refresh_token"] == "RT-fresh" + + +# --------------------------------------------------------------------------- +# Wire-format guard: httpx must serialise ``data`` as form-urlencoded +# --------------------------------------------------------------------------- + + +def test_wire_format_is_form_urlencoded_with_all_pkce_fields(monkeypatch): + """End-to-end check on the actual bytes httpx puts on the wire. + If anyone ever swaps ``data=`` for ``json=`` or refactors the dict, + xAI will start rejecting again — this catches it locally.""" + + captured: Dict[str, Any] = {} + + class _Transport(httpx.BaseTransport): + def handle_request(self, request): + captured["body"] = bytes(request.read()) + captured["content_type"] = request.headers.get("content-type", "") + return httpx.Response( + 200, + json={"access_token": "AT", "refresh_token": "RT", + "id_token": "", "expires_in": 60, "token_type": "Bearer"}, + ) + + real_post = httpx.post + + def _post(*args, **kwargs): + with httpx.Client(transport=_Transport()) as c: + return c.post(*args, **kwargs) + + monkeypatch.setattr("hermes_cli.auth.httpx.post", _post) + + _xai_oauth_exchange_code_for_tokens( + token_endpoint="https://auth.x.ai/oauth2/token", + code="AUTHCODE", + redirect_uri="http://127.0.0.1:56121/callback", + code_verifier="theVerifier_43+", + code_challenge="theChallenge_43+", + ) + + assert "application/x-www-form-urlencoded" in captured["content_type"] + parsed = parse_qs(captured["body"].decode()) + assert parsed["grant_type"] == ["authorization_code"] + assert parsed["code"] == ["AUTHCODE"] + assert parsed["redirect_uri"] == ["http://127.0.0.1:56121/callback"] + assert parsed["client_id"] == [XAI_OAUTH_CLIENT_ID] + assert parsed["code_verifier"] == ["theVerifier_43+"] + assert parsed["code_challenge"] == ["theChallenge_43+"] + assert parsed["code_challenge_method"] == ["S256"] diff --git a/tests/honcho_plugin/test_session.py b/tests/honcho_plugin/test_session.py index 64fcfc7ebfd..57724432348 100644 --- a/tests/honcho_plugin/test_session.py +++ b/tests/honcho_plugin/test_session.py @@ -1570,7 +1570,7 @@ class TestDialecticLifecycleSmoke: self._await_thread(provider) assert mgr.dialectic_query.call_count == 2, "turn 4 cadence fire" _, kwargs = mgr.dialectic_query.call_args - assert kwargs.get("reasoning_level") in ("medium", "high"), \ + assert kwargs.get("reasoning_level") in {"medium", "high"}, \ f"long query must bump reasoning level above 'low'; got {kwargs.get('reasoning_level')}" assert provider._last_dialectic_turn == 4, "cadence tracker advances on success" diff --git a/tests/plugins/browser/__init__.py b/tests/plugins/browser/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/plugins/browser/check_parity_vs_main.py b/tests/plugins/browser/check_parity_vs_main.py new file mode 100644 index 00000000000..b706ce3e9c0 --- /dev/null +++ b/tests/plugins/browser/check_parity_vs_main.py @@ -0,0 +1,273 @@ +"""Behavior-parity check for the browser-provider plugin migration (#25214). + +Spawns one subprocess per (version, scenario) cell — pinned to either +origin/main (legacy in-tree providers + class-instantiation lookup) or +this PR's worktree (plugin-based registry) via `sys.path[0]`. Each +subprocess clears all browser-related env vars + writes a config.yaml, +loads `tools.browser_tool._get_cloud_provider()`, and emits a reduced +"shape tuple" {is_local, provider_name, is_available} as JSON. + +The parent process diffs the shapes per scenario. A diff means the +migration introduced an observable behaviour change vs origin/main — +which would be a real regression for users on the existing config keys. + +Run from the PR worktree: + + cd ~/.hermes/hermes-agent/.worktrees/browser-providers-plugin + python tests/plugins/browser/check_parity_vs_main.py +""" +from __future__ import annotations + +import json +import subprocess +import sys +from pathlib import Path + + +REPO_ROOT = Path(__file__).resolve().parents[3] + + +# Pin one path to current main, one to the PR worktree. +# ``REPO_ROOT`` is ``.../.worktrees/browser-providers-plugin``; the main +# checkout lives two levels up at ``~/.hermes/hermes-agent``. +MAIN_DIR = REPO_ROOT.parent.parent # ~/.hermes/hermes-agent +PR_DIR = REPO_ROOT # the worktree we're in +assert (MAIN_DIR / "tools" / "browser_tool.py").exists(), ( + f"MAIN_DIR={MAIN_DIR} doesn't look like a hermes-agent checkout" +) +assert (PR_DIR / "tools" / "browser_tool.py").exists(), ( + f"PR_DIR={PR_DIR} doesn't look like a hermes-agent checkout" +) + + +# Reduced shape comparison — exact instance addresses obviously differ +# between subprocesses, so we compare the parts that matter for users. +SUBPROCESS_SCRIPT = r""" +import json, os, sys, tempfile +sys.path.insert(0, sys.argv[1]) + +# Isolated HERMES_HOME for the config write. +home = tempfile.mkdtemp() +os.environ["HERMES_HOME"] = home + +# Clear every browser-related env var so is_available() is deterministic. +for k in ( + "BROWSERBASE_API_KEY", "BROWSERBASE_PROJECT_ID", "BROWSERBASE_BASE_URL", + "BROWSER_USE_API_KEY", "BROWSER_USE_GATEWAY_URL", + "FIRECRAWL_API_KEY", "FIRECRAWL_API_URL", "FIRECRAWL_BROWSER_TTL", + "TOOL_GATEWAY_DOMAIN", "TOOL_GATEWAY_USER_TOKEN", +): + os.environ.pop(k, None) + +# Apply per-scenario env (passed as JSON via argv[2]). +scenario_env = json.loads(sys.argv[2]) +os.environ.update(scenario_env) + +# Apply per-scenario config (passed as YAML body via argv[3]). +config_yaml = sys.argv[3] +config_path = os.path.join(home, "config.yaml") +with open(config_path, "w") as f: + f.write(config_yaml) + +# Fresh import — must not have any browser modules cached. +for name in list(sys.modules): + if name.startswith("tools.") or name.startswith("agent.") or name.startswith("plugins."): + sys.modules.pop(name, None) + +from tools.browser_tool import _get_cloud_provider, _is_local_mode + +provider = _get_cloud_provider() + +# Pull the human-readable backend name via the API that exists on BOTH +# legacy (origin/main: CloudBrowserProvider.provider_name()) and the new +# ABC (BrowserProvider exposes provider_name() as a backward-compat alias +# returning display_name). Both shapes resolve to the same string — +# 'Browserbase' / 'Browser Use' / 'Firecrawl' — so we can compare safely. +provider_name = None +is_available = None +if provider is not None: + pn = getattr(provider, "provider_name", None) + if callable(pn): + provider_name = pn() + elif isinstance(pn, str): + provider_name = pn + is_conf = getattr(provider, "is_configured", None) + if callable(is_conf): + is_available = bool(is_conf()) + +shape = { + "is_local": _is_local_mode(), + "provider_name": provider_name, + "is_available": is_available, +} +print(json.dumps(shape)) +""" + + +SCENARIOS: list[tuple[str, str, dict[str, str]]] = [ + # (label, config.yaml body, extra env vars) + ("no-config-no-env", "", {}), + ("explicit-local-no-env", "browser:\n cloud_provider: local\n", {}), + ( + "explicit-browserbase-no-creds", + "browser:\n cloud_provider: browserbase\n", + {}, + ), + ( + "explicit-browserbase-with-creds", + "browser:\n cloud_provider: browserbase\n", + {"BROWSERBASE_API_KEY": "x", "BROWSERBASE_PROJECT_ID": "y"}, + ), + ( + "explicit-browser-use-no-creds", + "browser:\n cloud_provider: browser-use\n", + {}, + ), + ( + "explicit-browser-use-with-creds", + "browser:\n cloud_provider: browser-use\n", + {"BROWSER_USE_API_KEY": "k"}, + ), + ( + "explicit-firecrawl-no-creds", + "browser:\n cloud_provider: firecrawl\n", + {}, + ), + ( + "explicit-firecrawl-with-creds", + "browser:\n cloud_provider: firecrawl\n", + {"FIRECRAWL_API_KEY": "k"}, + ), + ( + "no-config-bu-creds", + "", + {"BROWSER_USE_API_KEY": "k"}, + ), + ( + "no-config-bb-creds", + "", + {"BROWSERBASE_API_KEY": "x", "BROWSERBASE_PROJECT_ID": "y"}, + ), + ( + "no-config-both-creds", + "", + { + "BROWSER_USE_API_KEY": "k", + "BROWSERBASE_API_KEY": "x", + "BROWSERBASE_PROJECT_ID": "y", + }, + ), + ( + "no-config-firecrawl-only", + "", + {"FIRECRAWL_API_KEY": "k"}, + ), + ( + "no-config-firecrawl-and-bb", + "", + { + "FIRECRAWL_API_KEY": "k", + "BROWSERBASE_API_KEY": "x", + "BROWSERBASE_PROJECT_ID": "y", + }, + ), +] + + +def _run_scenario(repo_path: Path, label: str, config_yaml: str, env: dict) -> dict: + """Run one (version, scenario) cell. Returns the shape dict.""" + venv_python = repo_path / ".venv" / "bin" / "python" + if not venv_python.exists(): + # Worktrees share the main repo's venv. + venv_python = MAIN_DIR / ".venv" / "bin" / "python" + if not venv_python.exists(): + venv_python = Path("python3") + + out = subprocess.run( + [ + str(venv_python), + "-c", + SUBPROCESS_SCRIPT, + str(repo_path), + json.dumps(env), + config_yaml, + ], + capture_output=True, + text=True, + timeout=30, + ) + if out.returncode != 0: + return { + "error": "subprocess failed", + "stdout": out.stdout, + "stderr": out.stderr[-500:], + } + try: + return json.loads(out.stdout.strip().splitlines()[-1]) + except Exception as exc: + return {"error": f"could not parse output: {exc}", "stdout": out.stdout} + + +def _reduce_for_comparison(shape: dict) -> dict: + """Reduce a shape dict to the parts that matter for user-visible parity. + + We compare ``(is_local, provider_name, is_available)`` — the trio that + decides what the dispatcher does with each tool call. ``provider_name`` + is the legacy ``provider_name()`` return value ('Browserbase' / 'Browser + Use' / 'Firecrawl'), which is identical between legacy and plugin + classes (the plugin's ``display_name`` matches the legacy + ``provider_name()`` return). + """ + return { + "is_local": shape.get("is_local"), + "provider_name": shape.get("provider_name"), + "is_available": shape.get("is_available"), + } + + +def main() -> int: + print(f"main: {MAIN_DIR}") + print(f"pr: {PR_DIR}") + print() + + failures: list[str] = [] + errors: list[str] = [] + for label, config_yaml, env in SCENARIOS: + main_shape = _run_scenario(MAIN_DIR, label, config_yaml, env) + pr_shape = _run_scenario(PR_DIR, label, config_yaml, env) + + if "error" in main_shape or "error" in pr_shape: + print(f" [ERR ] {label}: subprocess failed") + print(f" main: {main_shape}") + print(f" pr: {pr_shape}") + errors.append(label) + continue + + main_reduced = _reduce_for_comparison(main_shape) + pr_reduced = _reduce_for_comparison(pr_shape) + + if main_reduced == pr_reduced: + print(f" [OK] {label}: {main_reduced}") + else: + print(f" [FAIL] {label}") + print(f" main: {main_reduced}") + print(f" pr: {pr_reduced}") + failures.append(label) + + print() + if errors: + print(f"SUBPROCESS ERRORS in {len(errors)} scenario(s):") + for e in errors: + print(f" - {e}") + if failures: + print(f"BEHAVIOUR REGRESSION in {len(failures)} scenario(s):") + for f in failures: + print(f" - {f}") + if failures or errors: + return 1 + print(f"PARITY OK across {len(SCENARIOS)} scenarios.") + return 0 + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/tests/plugins/browser/test_browser_provider_plugins.py b/tests/plugins/browser/test_browser_provider_plugins.py new file mode 100644 index 00000000000..986a1d635bf --- /dev/null +++ b/tests/plugins/browser/test_browser_provider_plugins.py @@ -0,0 +1,379 @@ +"""Plugin-side tests for the browser provider migration (PR #25214). + +Covers: + +- All three bundled plugins (browserbase, browser-use, firecrawl) + instantiate and self-report the expected ABC defaults. +- Each plugin's ``is_available()`` correctly reflects env-var presence. +- The browser_registry resolves an active provider in the documented + scenarios: + * explicit config wins ignoring availability (so dispatcher surfaces + a typed credentials error) + * legacy preference walk: browser-use → browserbase (filtered by + availability) + * firecrawl is NOT in the legacy walk — explicit-only + * unknown name falls through to auto-detect + * ``local`` short-circuits to None + +These tests use *real* imports from the plugin modules — no mocking of +provider classes themselves — so the test catches drift in the ABC +interface, the registry, and the plugin glue layer simultaneously. +Mirrors ``tests/plugins/web/test_web_search_provider_plugins.py`` from +PR #25182. +""" +from __future__ import annotations + +import pytest + + +# --------------------------------------------------------------------------- +# Helpers +# --------------------------------------------------------------------------- + + +def _clear_browser_env(monkeypatch: pytest.MonkeyPatch) -> None: + """Strip every browser-provider env var so is_available() returns False.""" + for k in ( + "BROWSERBASE_API_KEY", + "BROWSERBASE_PROJECT_ID", + "BROWSERBASE_BASE_URL", + "BROWSER_USE_API_KEY", + "BROWSER_USE_GATEWAY_URL", + "FIRECRAWL_API_KEY", + "FIRECRAWL_API_URL", + "FIRECRAWL_BROWSER_TTL", + "TOOL_GATEWAY_DOMAIN", + "TOOL_GATEWAY_USER_TOKEN", + ): + monkeypatch.delenv(k, raising=False) + + +def _ensure_plugins_loaded() -> None: + """Idempotently load plugins so the registry is populated.""" + from hermes_cli.plugins import _ensure_plugins_discovered + + _ensure_plugins_discovered() + + +# --------------------------------------------------------------------------- +# Per-test isolation +# --------------------------------------------------------------------------- + + +@pytest.fixture(autouse=True) +def _isolate_env(monkeypatch: pytest.MonkeyPatch) -> None: + """Each test starts with a clean browser-provider env.""" + _clear_browser_env(monkeypatch) + + +# --------------------------------------------------------------------------- +# Bundled plugins register +# --------------------------------------------------------------------------- + + +class TestBundledPluginsRegister: + """All three bundled browser plugins discover and register correctly.""" + + def test_all_three_plugins_present_in_registry(self) -> None: + _ensure_plugins_loaded() + from agent.browser_registry import list_providers + + names = sorted(p.name for p in list_providers()) + assert names == ["browser-use", "browserbase", "firecrawl"] + + @pytest.mark.parametrize( + "plugin_name,expected_display", + [ + ("browserbase", "Browserbase"), + ("browser-use", "Browser Use"), + ("firecrawl", "Firecrawl"), + ], + ) + def test_each_plugin_has_name_and_display_name( + self, plugin_name: str, expected_display: str + ) -> None: + _ensure_plugins_loaded() + from agent.browser_registry import get_provider + + provider = get_provider(plugin_name) + assert provider is not None, f"plugin {plugin_name!r} not registered" + assert provider.name == plugin_name + assert provider.display_name == expected_display + + @pytest.mark.parametrize( + "plugin_name", + ["browserbase", "browser-use", "firecrawl"], + ) + def test_each_plugin_has_setup_schema(self, plugin_name: str) -> None: + """``get_setup_schema()`` returns a dict the picker can consume.""" + _ensure_plugins_loaded() + from agent.browser_registry import get_provider + + provider = get_provider(plugin_name) + assert provider is not None + schema = provider.get_setup_schema() + assert isinstance(schema, dict) + assert "name" in schema + assert "env_vars" in schema + # Every cloud-browser plugin needs the agent-browser post-setup hook + # so the picker auto-installs the CLI on selection. + assert schema.get("post_setup") == "agent_browser" + + @pytest.mark.parametrize( + "plugin_name", + ["browserbase", "browser-use", "firecrawl"], + ) + def test_each_plugin_implements_full_lifecycle(self, plugin_name: str) -> None: + """The ABC's three lifecycle methods are all overridden.""" + _ensure_plugins_loaded() + from agent.browser_provider import BrowserProvider + from agent.browser_registry import get_provider + + provider = get_provider(plugin_name) + assert provider is not None + # Each method must be a real override, not the ABC's NotImplementedError + # default — we check by comparing the function reference. + assert type(provider).create_session is not BrowserProvider.create_session + assert type(provider).close_session is not BrowserProvider.close_session + assert ( + type(provider).emergency_cleanup is not BrowserProvider.emergency_cleanup + ) + + +# --------------------------------------------------------------------------- +# is_available() behavior +# --------------------------------------------------------------------------- + + +class TestIsAvailable: + """Each plugin's ``is_available()`` reflects env-var presence accurately.""" + + def test_browserbase_requires_both_api_key_and_project_id( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + _ensure_plugins_loaded() + from agent.browser_registry import get_provider + + p = get_provider("browserbase") + assert p is not None + assert p.is_available() is False + + # API key alone is insufficient. + monkeypatch.setenv("BROWSERBASE_API_KEY", "key") + assert p.is_available() is False + + # Both env vars set → available. + monkeypatch.setenv("BROWSERBASE_PROJECT_ID", "proj") + assert p.is_available() is True + + def test_browserbase_project_id_alone_insufficient( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + _ensure_plugins_loaded() + from agent.browser_registry import get_provider + + p = get_provider("browserbase") + assert p is not None + monkeypatch.setenv("BROWSERBASE_PROJECT_ID", "proj") + assert p.is_available() is False + + def test_browser_use_satisfied_by_api_key( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + _ensure_plugins_loaded() + from agent.browser_registry import get_provider + + p = get_provider("browser-use") + assert p is not None + assert p.is_available() is False + monkeypatch.setenv("BROWSER_USE_API_KEY", "key") + assert p.is_available() is True + + def test_firecrawl_requires_api_key(self, monkeypatch: pytest.MonkeyPatch) -> None: + _ensure_plugins_loaded() + from agent.browser_registry import get_provider + + p = get_provider("firecrawl") + assert p is not None + assert p.is_available() is False + monkeypatch.setenv("FIRECRAWL_API_KEY", "key") + assert p.is_available() is True + + +# --------------------------------------------------------------------------- +# Registry resolution semantics +# --------------------------------------------------------------------------- + + +class TestRegistryResolution: + """``_resolve()`` implements the documented three-rule precedence.""" + + def test_resolve_none_with_no_creds_returns_none(self) -> None: + """No config, no env → local mode (None).""" + _ensure_plugins_loaded() + from agent.browser_registry import _resolve + + assert _resolve(None) is None + + def test_explicit_local_returns_none(self) -> None: + """``cloud_provider: local`` is a positive choice; short-circuits to None.""" + _ensure_plugins_loaded() + from agent.browser_registry import _resolve + + assert _resolve("local") is None + + def test_explicit_browserbase_returns_provider_even_when_unavailable(self) -> None: + """Rule 1: explicit-config wins even when credentials are missing. + + This is critical — the dispatcher needs to surface a typed + credentials error rather than silently switching backends. + """ + _ensure_plugins_loaded() + from agent.browser_registry import _resolve + + provider = _resolve("browserbase") + assert provider is not None + assert provider.name == "browserbase" + assert provider.is_available() is False # confirms "ignoring availability" + + def test_explicit_firecrawl_returns_provider_even_when_unavailable(self) -> None: + """Firecrawl behaves the same as browserbase under explicit config.""" + _ensure_plugins_loaded() + from agent.browser_registry import _resolve + + provider = _resolve("firecrawl") + assert provider is not None + assert provider.name == "firecrawl" + + def test_explicit_unknown_falls_back_to_auto_detect(self) -> None: + """Rule 1 miss: unknown name → fall through to legacy walk.""" + _ensure_plugins_loaded() + from agent.browser_registry import _resolve + + # With no credentials anywhere, auto-detect should also fail. + assert _resolve("not-a-real-provider") is None + + def test_legacy_walk_prefers_browser_use_over_browserbase( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + """Rule 3: walk order is browser-use → browserbase.""" + _ensure_plugins_loaded() + from agent.browser_registry import _resolve + + # Both available — browser-use should win. + monkeypatch.setenv("BROWSER_USE_API_KEY", "k1") + monkeypatch.setenv("BROWSERBASE_API_KEY", "k2") + monkeypatch.setenv("BROWSERBASE_PROJECT_ID", "p") + + provider = _resolve(None) + assert provider is not None + assert provider.name == "browser-use" + + def test_legacy_walk_falls_through_to_browserbase( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + """Rule 3: browser-use unavailable → browserbase picked.""" + _ensure_plugins_loaded() + from agent.browser_registry import _resolve + + monkeypatch.setenv("BROWSERBASE_API_KEY", "k") + monkeypatch.setenv("BROWSERBASE_PROJECT_ID", "p") + + provider = _resolve(None) + assert provider is not None + assert provider.name == "browserbase" + + def test_firecrawl_not_in_legacy_walk_even_when_only_one_available( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + """Regression: firecrawl is NEVER auto-selected even when single-eligible. + + Pre-PR-#25214, the dispatcher only auto-detected between Browser Use + and Browserbase; firecrawl was reachable solely via explicit + config. We preserve that gate because FIRECRAWL_API_KEY is shared + with the *web* firecrawl plugin — auto-routing a web-extract user + to a paid cloud browser would be a real behaviour regression. + """ + _ensure_plugins_loaded() + from agent.browser_registry import _resolve + + monkeypatch.setenv("FIRECRAWL_API_KEY", "k") + + # Only firecrawl is_available() — but it's not in the legacy walk. + assert _resolve(None) is None + + +# --------------------------------------------------------------------------- +# Legacy ABC backward-compat aliases (is_configured / provider_name) +# --------------------------------------------------------------------------- + + +class TestLegacyAbcAliases: + """is_configured() and provider_name() delegate to the new API.""" + + @pytest.mark.parametrize( + "plugin_name", + ["browserbase", "browser-use", "firecrawl"], + ) + def test_is_configured_delegates_to_is_available(self, plugin_name: str) -> None: + _ensure_plugins_loaded() + from agent.browser_registry import get_provider + + p = get_provider(plugin_name) + assert p is not None + assert p.is_configured() is p.is_available() + + @pytest.mark.parametrize( + "plugin_name,expected_label", + [ + ("browserbase", "Browserbase"), + ("browser-use", "Browser Use"), + ("firecrawl", "Firecrawl"), + ], + ) + def test_provider_name_returns_display_name( + self, plugin_name: str, expected_label: str + ) -> None: + _ensure_plugins_loaded() + from agent.browser_registry import get_provider + + p = get_provider(plugin_name) + assert p is not None + assert p.provider_name() == expected_label + + +# --------------------------------------------------------------------------- +# Picker integration +# --------------------------------------------------------------------------- + + +class TestPickerIntegration: + """`_plugin_browser_providers()` exposes all three plugins as picker rows.""" + + def test_picker_rows_match_registered_plugins(self) -> None: + _ensure_plugins_loaded() + from hermes_cli.tools_config import _plugin_browser_providers + + rows = _plugin_browser_providers() + names = sorted(r.get("browser_provider") for r in rows) + assert names == ["browser-use", "browserbase", "firecrawl"] + + def test_picker_rows_carry_post_setup_hook(self) -> None: + """Every browser plugin row has post_setup='agent_browser' so + selecting it triggers the agent-browser CLI install.""" + _ensure_plugins_loaded() + from hermes_cli.tools_config import _plugin_browser_providers + + for row in _plugin_browser_providers(): + assert row.get("post_setup") == "agent_browser", ( + f"plugin row {row['browser_provider']!r} missing post_setup hook" + ) + + def test_picker_rows_carry_browser_plugin_name_marker(self) -> None: + """`browser_plugin_name` matches `browser_provider` so downstream + code can route through the registry when it wants to.""" + _ensure_plugins_loaded() + from hermes_cli.tools_config import _plugin_browser_providers + + for row in _plugin_browser_providers(): + assert row.get("browser_plugin_name") == row.get("browser_provider") diff --git a/tests/plugins/test_achievements_plugin.py b/tests/plugins/test_achievements_plugin.py index 782aea7b397..2d908b3d46e 100644 --- a/tests/plugins/test_achievements_plugin.py +++ b/tests/plugins/test_achievements_plugin.py @@ -271,7 +271,7 @@ def test_evaluate_all_force_runs_synchronously(plugin_api): # Synchronous — snapshot is fresh on return. assert result["scan_meta"].get("sessions_total") == 25 - assert result["scan_meta"]["mode"] in ("full", "incremental") + assert result["scan_meta"]["mode"] in {"full", "incremental"} def test_start_background_scan_is_idempotent_while_running(plugin_api): diff --git a/tests/plugins/video_gen/test_xai_plugin.py b/tests/plugins/video_gen/test_xai_plugin.py index bd7a880fdee..4c365020a32 100644 --- a/tests/plugins/video_gen/test_xai_plugin.py +++ b/tests/plugins/video_gen/test_xai_plugin.py @@ -110,4 +110,4 @@ def test_xai_no_operation_kwarg(): result = XAIVideoGenProvider().generate("x", operation="generate") assert result["success"] is False # auth_required, NOT some signature error - assert result["error_type"] in ("auth_required", "api_error") + assert result["error_type"] in {"auth_required", "api_error"} diff --git a/tests/run_agent/test_anthropic_truncation_continuation.py b/tests/run_agent/test_anthropic_truncation_continuation.py index 872015bc0bc..4e87a33e9d8 100644 --- a/tests/run_agent/test_anthropic_truncation_continuation.py +++ b/tests/run_agent/test_anthropic_truncation_continuation.py @@ -106,9 +106,9 @@ class TestContinuationLogicBranching: def test_all_three_api_modes_hit_continuation_branch(self, api_mode): # The guard in run_agent.py is: # if self.api_mode in ("chat_completions", "bedrock_converse", "anthropic_messages"): - assert api_mode in ("chat_completions", "bedrock_converse", "anthropic_messages") + assert api_mode in {"chat_completions", "bedrock_converse", "anthropic_messages"} def test_codex_responses_still_excluded(self): # codex_responses has its own truncation path (not continuation-based) # and should NOT be routed through the shared block. - assert "codex_responses" not in ("chat_completions", "bedrock_converse", "anthropic_messages") + assert "codex_responses" not in {"chat_completions", "bedrock_converse", "anthropic_messages"} diff --git a/tests/run_agent/test_jsondecodeerror_retryable.py b/tests/run_agent/test_jsondecodeerror_retryable.py index 201521ddb22..0bd4fc09f9f 100644 --- a/tests/run_agent/test_jsondecodeerror_retryable.py +++ b/tests/run_agent/test_jsondecodeerror_retryable.py @@ -73,15 +73,20 @@ class TestAgentLoopSourceStillHasCarveOut: revert that happens to leave the test file intact.""" def test_run_agent_excludes_jsondecodeerror_from_local_validation(self): - import run_agent import inspect - src = inspect.getsource(run_agent) + from agent import conversation_loop + # The agent loop body lives in agent/conversation_loop.py after + # the run_agent.py refactor. Assert the carve-out is present in + # the extracted module specifically — if it ever moves back or + # disappears, this fails loudly rather than silently passing + # against a non-existent inline replica. + src = inspect.getsource(conversation_loop) # The predicate we care about must reference json.JSONDecodeError # in its exclusion tuple. We check for the specific co-occurrence # rather than the literal string so harmless reformatting doesn't # break us. assert "is_local_validation_error" in src assert "JSONDecodeError" in src, ( - "run_agent.py must carve out json.JSONDecodeError from the " - "is_local_validation_error classification — see #14782." + "agent/conversation_loop.py must carve out json.JSONDecodeError " + "from the is_local_validation_error classification — see #14782." ) diff --git a/tests/run_agent/test_memory_nudge_counter_hydration.py b/tests/run_agent/test_memory_nudge_counter_hydration.py index abf97d265a6..1b9bf56005d 100644 --- a/tests/run_agent/test_memory_nudge_counter_hydration.py +++ b/tests/run_agent/test_memory_nudge_counter_hydration.py @@ -120,10 +120,22 @@ def test_production_code_contains_hydration_block(): """Smoke test: confirm the hydration code is actually wired into run_conversation(). If someone deletes it, tests above still pass against the inline replica — this fails them awake. + + After the run_agent.py refactor the agent-loop body lives in + ``agent/conversation_loop.py`` and uses ``agent.X`` rather than + ``self.X``. Assert the block is present in the extracted module + specifically — if it ever drifts back into run_agent.py or + disappears entirely, this guard fails loudly. """ from pathlib import Path - src = Path(__file__).resolve().parents[2] / "run_agent.py" - content = src.read_text(encoding="utf-8") + repo = Path(__file__).resolve().parents[2] + cl_path = repo / "agent" / "conversation_loop.py" + src_cl = cl_path.read_text(encoding="utf-8") # Anchor on the unique comment + the modulo line. - assert "Hydrate per-session nudge counters from persisted history" in content - assert "self._turns_since_memory = prior_user_turns % self._memory_nudge_interval" in content + assert "Hydrate per-session nudge counters from persisted history" in src_cl, ( + f"Hydration comment missing from {cl_path}" + ) + assert ( + "agent._turns_since_memory = prior_user_turns % agent._memory_nudge_interval" + in src_cl + ), f"Hydration modulo assignment missing from {cl_path}" diff --git a/tests/run_agent/test_provider_parity.py b/tests/run_agent/test_provider_parity.py index c65c22004a9..cf619ea9743 100644 --- a/tests/run_agent/test_provider_parity.py +++ b/tests/run_agent/test_provider_parity.py @@ -254,8 +254,12 @@ class TestDeveloperRoleSwap: assert messages[0]["role"] == "system" def test_developer_role_via_nous_portal(self, monkeypatch): - agent = _make_agent(monkeypatch, "nous", base_url="https://inference-api.nousresearch.com/v1") - agent.model = "gpt-5" + agent = _make_agent( + monkeypatch, + "nous", + base_url="https://inference-api.nousresearch.com/v1", + model="gpt-5", + ) messages = [ {"role": "system", "content": "You are helpful."}, {"role": "user", "content": "hi"}, @@ -346,14 +350,24 @@ class TestBuildApiKwargsAIGateway: class TestBuildApiKwargsNousPortal: def test_includes_nous_product_tags(self, monkeypatch): from agent.portal_tags import nous_portal_tags - agent = _make_agent(monkeypatch, "nous", base_url="https://inference-api.nousresearch.com/v1") + agent = _make_agent( + monkeypatch, + "nous", + base_url="https://inference-api.nousresearch.com/v1", + model="gpt-5", + ) messages = [{"role": "user", "content": "hi"}] kwargs = agent._build_api_kwargs(messages) extra = kwargs.get("extra_body", {}) assert extra.get("tags") == nous_portal_tags() def test_uses_chat_completions_format(self, monkeypatch): - agent = _make_agent(monkeypatch, "nous", base_url="https://inference-api.nousresearch.com/v1") + agent = _make_agent( + monkeypatch, + "nous", + base_url="https://inference-api.nousresearch.com/v1", + model="gpt-5", + ) messages = [{"role": "user", "content": "hi"}] kwargs = agent._build_api_kwargs(messages) assert "messages" in kwargs diff --git a/tests/run_agent/test_run_agent.py b/tests/run_agent/test_run_agent.py index cd62cd41ded..fd9b01187eb 100644 --- a/tests/run_agent/test_run_agent.py +++ b/tests/run_agent/test_run_agent.py @@ -2282,9 +2282,11 @@ class TestMcpParallelToolBatch: def test_mcp_tools_parallel_when_server_opted_in(self): """MCP tools from a parallel-safe server can run concurrently.""" from run_agent import _should_parallelize_tool_batch - from tools.mcp_tool import _parallel_safe_servers, _lock + from tools.mcp_tool import _mcp_tool_server_names, _parallel_safe_servers, _lock with _lock: _parallel_safe_servers.add("github") + _mcp_tool_server_names["mcp_github_list_repos"] = "github" + _mcp_tool_server_names["mcp_github_search_code"] = "github" try: tc1 = _mock_tool_call(name="mcp_github_list_repos", arguments='{"org":"openai"}', call_id="c1") tc2 = _mock_tool_call(name="mcp_github_search_code", arguments='{"q":"test"}', call_id="c2") @@ -2292,13 +2294,16 @@ class TestMcpParallelToolBatch: finally: with _lock: _parallel_safe_servers.discard("github") + _mcp_tool_server_names.pop("mcp_github_list_repos", None) + _mcp_tool_server_names.pop("mcp_github_search_code", None) def test_mixed_mcp_and_builtin_parallel(self): """MCP parallel tools mixed with built-in parallel-safe tools.""" from run_agent import _should_parallelize_tool_batch - from tools.mcp_tool import _parallel_safe_servers, _lock + from tools.mcp_tool import _mcp_tool_server_names, _parallel_safe_servers, _lock with _lock: _parallel_safe_servers.add("docs") + _mcp_tool_server_names["mcp_docs_search"] = "docs" try: tc1 = _mock_tool_call(name="mcp_docs_search", arguments='{"query":"api"}', call_id="c1") tc2 = _mock_tool_call(name="web_search", arguments='{"query":"test"}', call_id="c2") @@ -2306,14 +2311,17 @@ class TestMcpParallelToolBatch: finally: with _lock: _parallel_safe_servers.discard("docs") + _mcp_tool_server_names.pop("mcp_docs_search", None) def test_mixed_parallel_and_serial_mcp_servers(self): """One parallel MCP server + one non-parallel MCP server = sequential.""" from run_agent import _should_parallelize_tool_batch - from tools.mcp_tool import _parallel_safe_servers, _lock + from tools.mcp_tool import _mcp_tool_server_names, _parallel_safe_servers, _lock with _lock: _parallel_safe_servers.add("docs") # "github" is NOT in _parallel_safe_servers + _mcp_tool_server_names["mcp_docs_search"] = "docs" + _mcp_tool_server_names["mcp_github_list_repos"] = "github" try: tc1 = _mock_tool_call(name="mcp_docs_search", arguments='{"query":"api"}', call_id="c1") tc2 = _mock_tool_call(name="mcp_github_list_repos", arguments='{"org":"openai"}', call_id="c2") @@ -2321,6 +2329,8 @@ class TestMcpParallelToolBatch: finally: with _lock: _parallel_safe_servers.discard("docs") + _mcp_tool_server_names.pop("mcp_docs_search", None) + _mcp_tool_server_names.pop("mcp_github_list_repos", None) class TestHandleMaxIterations: @@ -3657,7 +3667,7 @@ class TestNousCredentialRefresh: assert ok is True assert closed["value"] is True - assert captured["force_mint"] is True + assert captured["inference_auth_mode"] == "legacy" assert rebuilt["kwargs"]["api_key"] == "new-nous-key" assert ( rebuilt["kwargs"]["base_url"] == "https://inference-api.nousresearch.com/v1" @@ -4832,23 +4842,26 @@ class TestAnthropicInterruptHandler: def test_interruptible_has_anthropic_branch(self): """The interrupt handler must check api_mode == 'anthropic_messages'.""" import inspect - source = inspect.getsource(AIAgent._interruptible_api_call) + from agent.chat_completion_helpers import interruptible_api_call + source = inspect.getsource(interruptible_api_call) assert "anthropic_messages" in source, \ - "_interruptible_api_call must handle Anthropic interrupt (api_mode check)" + "interruptible_api_call must handle Anthropic interrupt (api_mode check)" def test_interruptible_rebuilds_anthropic_client(self): """After interrupting, the Anthropic client should be rebuilt.""" import inspect - source = inspect.getsource(AIAgent._interruptible_api_call) + from agent.chat_completion_helpers import interruptible_api_call + source = inspect.getsource(interruptible_api_call) assert "build_anthropic_client" in source, \ - "_interruptible_api_call must rebuild Anthropic client after interrupt" + "interruptible_api_call must rebuild Anthropic client after interrupt" def test_streaming_has_anthropic_branch(self): """_streaming_api_call must also handle Anthropic interrupt.""" import inspect - source = inspect.getsource(AIAgent._interruptible_streaming_api_call) + from agent.chat_completion_helpers import interruptible_streaming_api_call + source = inspect.getsource(interruptible_streaming_api_call) assert "anthropic_messages" in source, \ - "_streaming_api_call must handle Anthropic interrupt" + "interruptible_streaming_api_call must handle Anthropic interrupt" # --------------------------------------------------------------------------- @@ -5257,14 +5270,20 @@ class TestMemoryNudgeCounterPersistence: def test_counters_not_reset_in_preamble(self): """The run_conversation preamble must not zero the nudge counters.""" import inspect - src = inspect.getsource(AIAgent.run_conversation) + from agent.conversation_loop import run_conversation as _rc + src = inspect.getsource(_rc) # The preamble resets many fields (retry counts, budget, etc.) # before the main loop. Find that reset block and verify our # counters aren't in it. The reset block ends at iteration_budget. - preamble_end = src.index("self.iteration_budget = IterationBudget") + # The extracted body uses ``agent.X`` (not ``self.X``). Anchor + # exactly on ``agent.iteration_budget = IterationBudget`` so an + # unrelated identifier ending in ``iteration_budget`` (e.g. + # ``_iteration_budget`` or ``shared_iteration_budget``) can't + # match the boundary. + preamble_end = src.index("agent.iteration_budget = IterationBudget") preamble = src[:preamble_end] - assert "self._turns_since_memory = 0" not in preamble - assert "self._iters_since_skill = 0" not in preamble + assert "agent._turns_since_memory = 0" not in preamble + assert "agent._iters_since_skill = 0" not in preamble class TestDeadRetryCode: @@ -5272,7 +5291,8 @@ class TestDeadRetryCode: def test_no_unreachable_max_retries_after_backoff(self): import inspect - source = inspect.getsource(AIAgent.run_conversation) + from agent.conversation_loop import run_conversation as _rc + source = inspect.getsource(_rc) occurrences = source.count("if retry_count >= max_retries:") assert occurrences == 2, ( f"Expected 2 occurrences of 'if retry_count >= max_retries:' " @@ -5310,7 +5330,8 @@ class TestMemoryContextSanitization: a literal tag we don't silently delete their text. The streaming scrubber + plugin-side scrub cover real leak paths.""" import inspect - src = inspect.getsource(AIAgent.run_conversation) + from agent.conversation_loop import run_conversation as _rc + src = inspect.getsource(_rc) assert "sanitize_context(user_message)" not in src assert "sanitize_context(persist_user_message)" not in src @@ -5346,7 +5367,8 @@ class TestMemoryProviderTurnStart: def test_on_turn_start_called_before_prefetch(self): """Source-level check: on_turn_start appears before prefetch_all in run_conversation.""" import inspect - src = inspect.getsource(AIAgent.run_conversation) + from agent.conversation_loop import run_conversation as _rc + src = inspect.getsource(_rc) # Find the actual method calls, not comments idx_turn_start = src.index(".on_turn_start(") idx_prefetch = src.index(".prefetch_all(") @@ -5356,7 +5378,10 @@ class TestMemoryProviderTurnStart: ) def test_on_turn_start_uses_user_turn_count(self): - """Source-level check: on_turn_start receives self._user_turn_count.""" + """Source-level check: on_turn_start receives the user_turn_count.""" import inspect - src = inspect.getsource(AIAgent.run_conversation) - assert "on_turn_start(self._user_turn_count" in src + from agent.conversation_loop import run_conversation as _rc + src = inspect.getsource(_rc) + # The extracted body uses ``agent.X`` rather than ``self.X``; + # assert the extracted-form spelling directly. + assert "on_turn_start(agent._user_turn_count" in src diff --git a/tests/run_agent/test_tool_executor_contextvar_propagation.py b/tests/run_agent/test_tool_executor_contextvar_propagation.py index 652ecf05def..2e1d543705a 100644 --- a/tests/run_agent/test_tool_executor_contextvar_propagation.py +++ b/tests/run_agent/test_tool_executor_contextvar_propagation.py @@ -152,19 +152,28 @@ def test_run_agent_concurrent_executor_wraps_submit_with_copy_context(): import inspect import run_agent + from agent import tool_executor as tool_executor_module - src_path = inspect.getsourcefile(run_agent) - assert src_path is not None - tree = ast.parse(open(src_path, encoding="utf-8").read()) + # Source for both modules — the concurrent-executor body lives in + # ``agent/tool_executor.py`` after the run_agent.py refactor (PR + # following #16660). Search both so this guard keeps firing + # regardless of where the call site lives. + sources = [] + for mod in (run_agent, tool_executor_module): + src_path = inspect.getsourcefile(mod) + assert src_path is not None + sources.append((src_path, open(src_path, encoding="utf-8").read())) submit_calls_in_agent: list[ast.Call] = [] - for node in ast.walk(tree): - if not isinstance(node, ast.Call): - continue - func = node.func - # Match executor.submit(...) style calls. - if isinstance(func, ast.Attribute) and func.attr == "submit": - submit_calls_in_agent.append(node) + for _src_path, src_text in sources: + tree = ast.parse(src_text) + for node in ast.walk(tree): + if not isinstance(node, ast.Call): + continue + func = node.func + # Match executor.submit(...) style calls. + if isinstance(func, ast.Attribute) and func.attr == "submit": + submit_calls_in_agent.append(node) # Filter to the submit call inside the concurrent tool executor — # identifiable by passing `_run_tool` as its target. Other submit() diff --git a/tests/skills/test_openclaw_migration.py b/tests/skills/test_openclaw_migration.py index 708484027be..0b331c40238 100644 --- a/tests/skills/test_openclaw_migration.py +++ b/tests/skills/test_openclaw_migration.py @@ -846,7 +846,7 @@ def test_skill_installs_cleanly_under_skills_guard(): # the script never writes to that file # # Accept "caution" or "safe" — just not "dangerous" from a *real* threat. - assert result.verdict in ("safe", "caution", "dangerous"), f"Unexpected verdict: {result.verdict}" + assert result.verdict in {"safe", "caution", "dangerous"}, f"Unexpected verdict: {result.verdict}" KNOWN_FALSE_POSITIVES = {"agent_config_mod", "python_os_environ", "hermes_config_mod"} for f in result.findings: assert f.pattern_id in KNOWN_FALSE_POSITIVES, f"Unexpected finding: {f}" diff --git a/tests/stress/test_atypical_scenarios.py b/tests/stress/test_atypical_scenarios.py index 2010049e14f..e7e83eabccb 100644 --- a/tests/stress/test_atypical_scenarios.py +++ b/tests/stress/test_atypical_scenarios.py @@ -902,7 +902,7 @@ def _(home, kb): pass # Empty body → accept (legitimate: just title says it all) tid = kb.create_task(conn, title="empty body ok", body="", assignee="w") - assert kb.get_task(conn, tid).body in ("", None) + assert kb.get_task(conn, tid).body in {"", None} # Empty summary on complete → accept kb.claim_task(conn, tid) kb.complete_task(conn, tid, summary="") @@ -994,7 +994,7 @@ def _(home, kb): # Empty title r = client.post("/api/plugins/kanban/tasks", json={"title": ""}) - assert r.status_code in (400, 422), f"empty title should 4xx, got {r.status_code}" + assert r.status_code in {400, 422}, f"empty title should 4xx, got {r.status_code}" # Title only r = client.post("/api/plugins/kanban/tasks", json={"title": "x"}) @@ -1019,7 +1019,7 @@ def _(home, kb): r = client.post("/api/plugins/kanban/tasks", json={ "title": "fine", "nonexistent_field": "whatever", }) - assert r.status_code in (200, 422) + assert r.status_code in {200, 422} # Priority as non-int r = client.post("/api/plugins/kanban/tasks", json={"title": "prio", "priority": "high"}) @@ -1028,7 +1028,7 @@ def _(home, kb): # PATCH with empty body (no changes requested) r = client.patch(f"/api/plugins/kanban/tasks/{tid}", json={}) # Accept either success-no-op or 400 - assert r.status_code in (200, 400) + assert r.status_code in {200, 400} print(" dashboard REST handles weird inputs correctly") # ============================================================================= diff --git a/tests/test_live_system_guard_self_test.py b/tests/test_live_system_guard_self_test.py index 1856935b240..3bbe8c9f3b0 100644 --- a/tests/test_live_system_guard_self_test.py +++ b/tests/test_live_system_guard_self_test.py @@ -259,7 +259,7 @@ def test_kill_own_subtree_passes_through(): finally: p.wait(timeout=2) # SIGTERM = 15; subprocess returncode is -15 on POSIX. - assert p.returncode in (-signal.SIGTERM, 128 + int(signal.SIGTERM)) + assert p.returncode in {-signal.SIGTERM, 128 + int(signal.SIGTERM)} def test_subprocess_pkill_with_unrelated_pattern_passes_through(): diff --git a/tests/test_timezone.py b/tests/test_timezone.py index ffb831617d9..f91a27b6a75 100644 --- a/tests/test_timezone.py +++ b/tests/test_timezone.py @@ -63,7 +63,7 @@ class TestHermesTimeNow: assert result.tzinfo is not None # Offset is -5h or -4h depending on DST offset_hours = result.utcoffset().total_seconds() / 3600 - assert offset_hours in (-5, -4) + assert offset_hours in {-5, -4} def test_invalid_timezone_falls_back(self, caplog): """Invalid timezone logs warning and falls back to server-local.""" diff --git a/tests/test_tui_gateway_server.py b/tests/test_tui_gateway_server.py index cc735e2dcfc..46253c86810 100644 --- a/tests/test_tui_gateway_server.py +++ b/tests/test_tui_gateway_server.py @@ -3799,7 +3799,7 @@ def test_prompt_submit_preserves_empty_response_without_error(monkeypatch): assert payload.get("status") == "complete" # Text stays empty — we did NOT fabricate an "Error:" string text = payload.get("text", "") - assert text in ("", None), f"expected empty text, got {text!r}" + assert text in {"", None}, f"expected empty text, got {text!r}" # ── session.most_recent ────────────────────────────────────────────── diff --git a/tests/tools/test_browser_homebrew_paths.py b/tests/tools/test_browser_homebrew_paths.py index 7e4d1c70222..7edf6f6c67d 100644 --- a/tests/tools/test_browser_homebrew_paths.py +++ b/tests/tools/test_browser_homebrew_paths.py @@ -68,10 +68,10 @@ class TestDiscoverHomebrewNodeDirs: if p == "/opt/homebrew/opt": return True # node@20/bin and node@24/bin exist - if p in ( + if p in { "/opt/homebrew/opt/node@20/bin", "/opt/homebrew/opt/node@24/bin", - ): + }: return True return False @@ -171,10 +171,10 @@ class TestFindAgentBrowser: real_isdir = os.path.isdir def selective_isdir(path): - if path in ( + if path in { "/data/data/com.termux/files/usr/bin", "/data/data/com.termux/files/usr/sbin", - ): + }: return True return real_isdir(path) @@ -486,10 +486,10 @@ class TestRunBrowserCommandPathConstruction: real_isdir = os.path.isdir def selective_isdir(path): - if path in ( + if path in { "/data/data/com.termux/files/usr/bin", "/data/data/com.termux/files/usr/sbin", - ): + }: return True if path.startswith(str(tmp_path)): return True diff --git a/tests/tools/test_code_execution_modes.py b/tests/tools/test_code_execution_modes.py index 4e22fe6e7a2..e5e2d2262ff 100644 --- a/tests/tools/test_code_execution_modes.py +++ b/tests/tools/test_code_execution_modes.py @@ -125,7 +125,7 @@ class TestResolveChildPython(unittest.TestCase): def test_project_with_no_venv_falls_back(self): """Project mode without VIRTUAL_ENV or CONDA_PREFIX → sys.executable.""" env = {k: v for k, v in os.environ.items() - if k not in ("VIRTUAL_ENV", "CONDA_PREFIX")} + if k not in {"VIRTUAL_ENV", "CONDA_PREFIX"}} with patch.dict(os.environ, env, clear=True): self.assertEqual(_resolve_child_python("project"), sys.executable) diff --git a/tests/tools/test_delegate.py b/tests/tools/test_delegate.py index 684f24f5da8..72c4c67f570 100644 --- a/tests/tools/test_delegate.py +++ b/tests/tools/test_delegate.py @@ -1014,6 +1014,89 @@ class TestDelegationCredentialResolution(unittest.TestCase): self.assertIsNone(creds["model"]) self.assertIsNone(creds["provider"]) + @patch("hermes_cli.runtime_provider.resolve_runtime_provider") + def test_named_custom_provider_preserves_provider_name(self, mock_resolve): + """Named custom provider (e.g. crof.ai) resolves to 'custom' at runtime level + but the subagent must retain the original provider identity so that + resolve_provider_client routes to the correct endpoint on retry/fallback. + Regression test for #26954. + """ + mock_resolve.return_value = { + "provider": "custom", # runtime marks it as "custom" type + "model": "deepseek-v4-pro-CEER", + "base_url": "https://api.crof.ai/v1", + "api_key": "crof-key-abc", + "api_mode": "chat_completions", + } + parent = _make_mock_parent(depth=0) + cfg = {"model": "deepseek-v4-pro-CEER", "provider": "crof.ai"} + creds = _resolve_delegation_credentials(cfg, parent) + # The key assertion: subagent must keep "crof.ai", NOT "custom" + self.assertEqual(creds["provider"], "crof.ai") + self.assertEqual(creds["model"], "deepseek-v4-pro-CEER") + self.assertEqual(creds["base_url"], "https://api.crof.ai/v1") + self.assertEqual(creds["api_key"], "crof-key-abc") + # Verify resolve_runtime_provider was called with the configured name + mock_resolve.assert_called_once_with( + requested="crof.ai", target_model="deepseek-v4-pro-CEER" + ) + + @patch("hermes_cli.runtime_provider.resolve_runtime_provider") + def test_standard_provider_not_overwritten_by_configured_name(self, mock_resolve): + """Standard (non-custom) providers must still return runtime identity, + not the configured name, to preserve existing behaviour for openrouter, + nous, etc. + """ + mock_resolve.return_value = { + "provider": "openrouter", + "model": "anthropic/claude-sonnet-4", + "base_url": "https://openrouter.ai/api/v1", + "api_key": "or-key-xyz", + "api_mode": "chat_completions", + } + parent = _make_mock_parent(depth=0) + cfg = {"model": "anthropic/claude-sonnet-4", "provider": "openrouter"} + creds = _resolve_delegation_credentials(cfg, parent) + # Standard provider returns its own name, not "custom" + self.assertEqual(creds["provider"], "openrouter") + + @patch("hermes_cli.runtime_provider.resolve_runtime_provider") + def test_custom_provider_with_empty_configured_provider_falls_back_to_runtime(self, mock_resolve): + """When configured_provider is empty/None, the early return kicks in and + we return provider=None regardless of what runtime resolved. The runtime + path is only reached when configured_provider is a non-empty string. + """ + mock_resolve.return_value = { + "provider": "custom", + "model": "some-model", + "base_url": "https://fallback.example.com/v1", + "api_key": "key-fallback", + "api_mode": "chat_completions", + } + parent = _make_mock_parent(depth=0) + cfg = {"model": "some-model", "provider": ""} + creds = _resolve_delegation_credentials(cfg, parent) + # Empty provider → early return with None (child inherits parent) + self.assertIsNone(creds["provider"]) + + @patch("hermes_cli.runtime_provider.resolve_runtime_provider") + def test_runtime_missing_provider_key_returns_none(self, mock_resolve): + """When resolve_runtime_provider returns a dict without 'provider' key, + the result must be None regardless of configured_provider. + This protects against malformed runtime responses. + """ + mock_resolve.return_value = { + # deliberately missing "provider" + "model": "some-model", + "base_url": "https://example.com/v1", + "api_key": "key-123", + "api_mode": "chat_completions", + } + parent = _make_mock_parent(depth=0) + cfg = {"model": "some-model", "provider": "crof.ai"} + creds = _resolve_delegation_credentials(cfg, parent) + self.assertIsNone(creds["provider"]) + class TestDelegationProviderIntegration(unittest.TestCase): """Integration tests: delegation config → _run_single_child → AIAgent construction.""" diff --git a/tests/tools/test_discord_tool.py b/tests/tools/test_discord_tool.py index 41d2cc957be..19a31d10457 100644 --- a/tests/tools/test_discord_tool.py +++ b/tests/tools/test_discord_tool.py @@ -633,7 +633,7 @@ class TestToolsetInclusion: def test_discord_tools_not_in_other_toolsets(self): from toolsets import TOOLSETS for name, ts in TOOLSETS.items(): - if name in ("hermes-discord", "hermes-gateway", "discord", "discord_admin"): + if name in {"hermes-discord", "hermes-gateway", "discord", "discord_admin"}: continue tools = ts.get("tools", []) assert "discord" not in tools or name == "discord", ( diff --git a/tests/tools/test_dockerfile_pid1_reaping.py b/tests/tools/test_dockerfile_pid1_reaping.py index e578d8a69fd..70d95807aa7 100644 --- a/tests/tools/test_dockerfile_pid1_reaping.py +++ b/tests/tools/test_dockerfile_pid1_reaping.py @@ -121,6 +121,20 @@ def test_dockerfile_installs_tui_dependencies(dockerfile_text): ) +def test_dockerfile_preinstalls_gateway_messaging_dependencies(dockerfile_text): + sync_steps = [ + step for step in _run_steps(dockerfile_text) + if "uv sync" in step and "--no-install-project" in step + ] + + assert sync_steps, "Dockerfile must install Python dependencies with uv sync" + assert any("--extra messaging" in step for step in sync_steps), ( + "Published Docker images must preload the [messaging] extra so " + "Telegram/Discord gateway adapters do not depend on first-boot " + "lazy installation (#24698)." + ) + + def test_dockerfile_builds_tui_assets(dockerfile_text): assert any( "ui-tui" in step and "npm" in step and "run build" in step diff --git a/tests/tools/test_hidden_dir_filter.py b/tests/tools/test_hidden_dir_filter.py index d7c10846bea..c7757864f74 100644 --- a/tests/tools/test_hidden_dir_filter.py +++ b/tests/tools/test_hidden_dir_filter.py @@ -24,7 +24,7 @@ def _new_filter_matches(path: Path) -> bool: Returns True when the path SHOULD be filtered out. """ - return any(part in ('.git', '.github', '.hub') for part in path.parts) + return any(part in {'.git', '.github', '.hub'} for part in path.parts) class TestOldFilterBrokenOnWindows: diff --git a/tests/tools/test_managed_browserbase_and_modal.py b/tests/tools/test_managed_browserbase_and_modal.py index 6c963be6207..d88789706ba 100644 --- a/tests/tools/test_managed_browserbase_and_modal.py +++ b/tests/tools/test_managed_browserbase_and_modal.py @@ -10,7 +10,9 @@ from unittest.mock import patch import pytest -TOOLS_DIR = Path(__file__).resolve().parents[2] / "tools" +REPO_ROOT = Path(__file__).resolve().parents[2] +TOOLS_DIR = REPO_ROOT / "tools" +PLUGINS_DIR = REPO_ROOT / "plugins" def _load_tool_module(module_name: str, filename: str): @@ -22,6 +24,21 @@ def _load_tool_module(module_name: str, filename: str): return module +def _load_plugin_module(module_name: str, relpath: str): + """Load a plugin module by file path from ``plugins/``. + + Mirror of :func:`_load_tool_module` for the plugin tree. Used by tests + that exercise the per-vendor browser plugins' session-lifecycle + behaviour after the PR #25214 migration. + """ + spec = spec_from_file_location(module_name, PLUGINS_DIR / relpath) + assert spec and spec.loader + module = module_from_spec(spec) + sys.modules[module_name] = module + spec.loader.exec_module(module) + return module + + def _reset_modules(prefixes: tuple[str, ...]): for name in list(sys.modules): if name.startswith(prefixes): @@ -76,6 +93,48 @@ def _install_fake_tools_package(): call_llm=lambda *args, **kwargs: "", ) + # Stubs for the browser-provider plugin layer introduced in PR #25214. + # The fake `agent` package has an empty __path__ so real submodules + # aren't reachable; we install just enough stand-ins to satisfy + # ``tools.browser_tool``'s top-level imports. The actual lifecycle + # tests instantiate the real plugin classes via _load_tool_module + # below, so the stubs only need to satisfy import + isinstance. + class _StubBrowserProvider: + """Minimal BrowserProvider stub for ``from agent.browser_provider import BrowserProvider``.""" + + sys.modules["agent.browser_provider"] = types.SimpleNamespace( + BrowserProvider=_StubBrowserProvider, + ) + sys.modules["agent.browser_registry"] = types.SimpleNamespace( + get_provider=lambda name: None, + list_providers=lambda: [], + register_provider=lambda provider: None, + _resolve=lambda configured: None, + ) + + # Plugin module stubs — the real plugin classes are loaded from disk by + # the lifecycle tests below via _load_tool_module(). For the import + # phase, we just need the class names to exist on the right module path. + plugins_package = types.ModuleType("plugins") + plugins_package.__path__ = [] # type: ignore[attr-defined] + sys.modules["plugins"] = plugins_package + plugins_browser_package = types.ModuleType("plugins.browser") + plugins_browser_package.__path__ = [] # type: ignore[attr-defined] + sys.modules["plugins.browser"] = plugins_browser_package + + for _name, _classname in ( + ("browserbase", "BrowserbaseBrowserProvider"), + ("browser_use", "BrowserUseBrowserProvider"), + ("firecrawl", "FirecrawlBrowserProvider"), + ): + _vendor_pkg = types.ModuleType(f"plugins.browser.{_name}") + _vendor_pkg.__path__ = [] # type: ignore[attr-defined] + sys.modules[f"plugins.browser.{_name}"] = _vendor_pkg + _provider_stub_cls = type(_classname, (_StubBrowserProvider,), {}) + sys.modules[f"plugins.browser.{_name}.provider"] = types.SimpleNamespace( + **{_classname: _provider_stub_cls}, + ) + sys.modules["tools.managed_tool_gateway"] = _load_tool_module( "tools.managed_tool_gateway", "managed_tool_gateway.py", @@ -157,13 +216,13 @@ def test_browserbase_does_not_use_gateway_only_configuration(): }) with patch.dict(os.environ, env, clear=True): - browserbase_module = _load_tool_module( - "tools.browser_providers.browserbase", - "browser_providers/browserbase.py", + browserbase_module = _load_plugin_module( + "plugins.browser.browserbase.provider", + "browser/browserbase/provider.py", ) - provider = browserbase_module.BrowserbaseProvider() + provider = browserbase_module.BrowserbaseBrowserProvider() - assert provider.is_configured() is False + assert provider.is_available() is False def test_browser_use_managed_gateway_adds_idempotency_key_and_persists_external_call_id(): @@ -188,13 +247,13 @@ def test_browser_use_managed_gateway_adds_idempotency_key_and_persists_external_ } with patch.dict(os.environ, env, clear=True): - browser_use_module = _load_tool_module( - "tools.browser_providers.browser_use", - "browser_providers/browser_use.py", + browser_use_module = _load_plugin_module( + "plugins.browser.browser_use.provider", + "browser/browser_use/provider.py", ) with patch.object(browser_use_module.requests, "post", return_value=_Response()) as post: - provider = browser_use_module.BrowserUseProvider() + provider = browser_use_module.BrowserUseBrowserProvider() session = provider.create_session("task-browser-use-managed") sent_headers = post.call_args.kwargs["headers"] @@ -228,11 +287,11 @@ def test_browser_use_managed_gateway_reuses_pending_idempotency_key_after_timeou } with patch.dict(os.environ, env, clear=True): - browser_use_module = _load_tool_module( - "tools.browser_providers.browser_use", - "browser_providers/browser_use.py", + browser_use_module = _load_plugin_module( + "plugins.browser.browser_use.provider", + "browser/browser_use/provider.py", ) - provider = browser_use_module.BrowserUseProvider() + provider = browser_use_module.BrowserUseBrowserProvider() timeout = browser_use_module.requests.Timeout("timed out") with patch.object( @@ -290,11 +349,11 @@ def test_browser_use_managed_gateway_preserves_pending_idempotency_key_for_in_pr } with patch.dict(os.environ, env, clear=True): - browser_use_module = _load_tool_module( - "tools.browser_providers.browser_use", - "browser_providers/browser_use.py", + browser_use_module = _load_plugin_module( + "plugins.browser.browser_use.provider", + "browser/browser_use/provider.py", ) - provider = browser_use_module.BrowserUseProvider() + provider = browser_use_module.BrowserUseBrowserProvider() with patch.object( browser_use_module.requests, @@ -337,11 +396,11 @@ def test_browser_use_managed_gateway_uses_new_idempotency_key_for_a_new_session_ } with patch.dict(os.environ, env, clear=True): - browser_use_module = _load_tool_module( - "tools.browser_providers.browser_use", - "browser_providers/browser_use.py", + browser_use_module = _load_plugin_module( + "plugins.browser.browser_use.provider", + "browser/browser_use/provider.py", ) - provider = browser_use_module.BrowserUseProvider() + provider = browser_use_module.BrowserUseBrowserProvider() with patch.object(browser_use_module.requests, "post", side_effect=[_Response(), _Response()]) as post: provider.create_session("task-browser-use-new") diff --git a/tests/tools/test_managed_modal_environment.py b/tests/tools/test_managed_modal_environment.py index d36418336cc..8380e49058c 100644 --- a/tests/tools/test_managed_modal_environment.py +++ b/tests/tools/test_managed_modal_environment.py @@ -33,7 +33,7 @@ def _restore_tool_and_agent_modules(): original_modules = { name: module for name, module in sys.modules.items() - if name in ("tools", "agent", "hermes_cli") + if name in {"tools", "agent", "hermes_cli"} or name.startswith("tools.") or name.startswith("agent.") or name.startswith("hermes_cli.") diff --git a/tests/tools/test_mcp_cancelled_error_propagation.py b/tests/tools/test_mcp_cancelled_error_propagation.py index ce05d03f43a..c0e91f31531 100644 --- a/tests/tools/test_mcp_cancelled_error_propagation.py +++ b/tests/tools/test_mcp_cancelled_error_propagation.py @@ -62,7 +62,7 @@ class TestCancelledErrorPropagation: return "clean_return" outcome = asyncio.run(drive()) - assert outcome in ("cancelled_cleanly", "clean_return"), ( + assert outcome in {"cancelled_cleanly", "clean_return"}, ( f"MCPServerTask.run wedged on cancel (outcome={outcome}) — " f"#9930 regression" ) diff --git a/tests/tools/test_mcp_oauth.py b/tests/tools/test_mcp_oauth.py index 2dfebd80b9c..e12149a45d3 100644 --- a/tests/tools/test_mcp_oauth.py +++ b/tests/tools/test_mcp_oauth.py @@ -10,6 +10,8 @@ from unittest.mock import patch, MagicMock, AsyncMock import pytest +import asyncio + from tools.mcp_oauth import ( HermesTokenStorage, OAuthNonInteractiveError, @@ -20,6 +22,7 @@ from tools.mcp_oauth import ( _is_interactive, _wait_for_callback, _make_callback_handler, + _redirect_handler, ) @@ -241,6 +244,64 @@ class TestUtilities: assert _can_open_browser() is True +class TestRedirectHandlerSshHint: + """_redirect_handler must print an SSH tunnel hint on remote sessions.""" + + def _run(self, coro): + return asyncio.get_event_loop().run_until_complete(coro) + + def test_ssh_hint_shown_on_ssh_session(self, monkeypatch, capsys): + import tools.mcp_oauth as mco + monkeypatch.setattr(mco, "_oauth_port", 49200) + monkeypatch.setenv("SSH_CLIENT", "1.2.3.4 1234 22") + monkeypatch.delenv("SSH_TTY", raising=False) + monkeypatch.setattr(mco, "_can_open_browser", lambda: False) + + self._run(_redirect_handler("https://example.com/auth?foo=bar")) + + err = capsys.readouterr().err + assert "49200" in err + assert "ssh -N -L" in err + assert "Remote session detected" in err + + def test_ssh_hint_shown_via_ssh_tty(self, monkeypatch, capsys): + import tools.mcp_oauth as mco + monkeypatch.setattr(mco, "_oauth_port", 49201) + monkeypatch.delenv("SSH_CLIENT", raising=False) + monkeypatch.setenv("SSH_TTY", "/dev/pts/1") + monkeypatch.setattr(mco, "_can_open_browser", lambda: False) + + self._run(_redirect_handler("https://example.com/auth")) + + err = capsys.readouterr().err + assert "49201" in err + assert "ssh -N -L" in err + + def test_no_ssh_hint_on_local_session(self, monkeypatch, capsys): + import tools.mcp_oauth as mco + monkeypatch.setattr(mco, "_oauth_port", 49202) + monkeypatch.delenv("SSH_CLIENT", raising=False) + monkeypatch.delenv("SSH_TTY", raising=False) + monkeypatch.setattr(mco, "_can_open_browser", lambda: True) + monkeypatch.setattr("webbrowser.open", lambda url, **kw: True) + + self._run(_redirect_handler("https://example.com/auth")) + + err = capsys.readouterr().err + assert "ssh -N -L" not in err + + def test_no_ssh_hint_when_port_not_set(self, monkeypatch, capsys): + import tools.mcp_oauth as mco + monkeypatch.setattr(mco, "_oauth_port", None) + monkeypatch.setenv("SSH_CLIENT", "1.2.3.4 1234 22") + monkeypatch.setattr(mco, "_can_open_browser", lambda: False) + + self._run(_redirect_handler("https://example.com/auth")) + + err = capsys.readouterr().err + assert "ssh -N -L" not in err + + # --------------------------------------------------------------------------- # Path traversal protection # --------------------------------------------------------------------------- diff --git a/tests/tools/test_mcp_stability.py b/tests/tools/test_mcp_stability.py index 238696feba2..163a05963e0 100644 --- a/tests/tools/test_mcp_stability.py +++ b/tests/tools/test_mcp_stability.py @@ -135,7 +135,7 @@ class TestStdioPidTracking: # bpo-14484). Return True so the SIGKILL escalation fires. with patch("tools.mcp_tool.os.kill") as mock_kill, \ patch("gateway.status._pid_exists", return_value=True), \ - patch("time.sleep") as mock_sleep: + patch("tools.mcp_tool.time.sleep") as mock_sleep: _kill_orphaned_mcp_children() # SIGTERM then SIGKILL; the alive check no longer touches os.kill. @@ -163,7 +163,7 @@ class TestStdioPidTracking: monkeypatch.delattr(signal, "SIGKILL", raising=False) with patch("tools.mcp_tool.os.kill") as mock_kill, \ - patch("time.sleep") as mock_sleep: + patch("tools.mcp_tool.time.sleep") as mock_sleep: _kill_orphaned_mcp_children() # SIGTERM phase, alive check raises (process gone), no escalation diff --git a/tests/tools/test_mcp_tool.py b/tests/tools/test_mcp_tool.py index 0a094eb5467..3212a350c37 100644 --- a/tests/tools/test_mcp_tool.py +++ b/tests/tools/test_mcp_tool.py @@ -3781,16 +3781,26 @@ class TestMcpParallelToolCalls: def test_is_mcp_tool_parallel_safe_no_servers(self): """MCP tool from unknown server returns False.""" - from tools.mcp_tool import is_mcp_tool_parallel_safe, _parallel_safe_servers, _lock + from tools.mcp_tool import ( + is_mcp_tool_parallel_safe, _mcp_tool_server_names, + _parallel_safe_servers, _lock, + ) with _lock: _parallel_safe_servers.clear() + _mcp_tool_server_names.clear() assert is_mcp_tool_parallel_safe("mcp_docs_search") is False def test_is_mcp_tool_parallel_safe_with_flag(self): """MCP tool from a parallel-safe server returns True.""" - from tools.mcp_tool import is_mcp_tool_parallel_safe, _parallel_safe_servers, _lock + from tools.mcp_tool import ( + is_mcp_tool_parallel_safe, _mcp_tool_server_names, + _parallel_safe_servers, _lock, + ) with _lock: _parallel_safe_servers.add("docs") + _mcp_tool_server_names["mcp_docs_search"] = "docs" + _mcp_tool_server_names["mcp_docs_read_file"] = "docs" + _mcp_tool_server_names["mcp_github_list_repos"] = "github" try: assert is_mcp_tool_parallel_safe("mcp_docs_search") is True assert is_mcp_tool_parallel_safe("mcp_docs_read_file") is True @@ -3799,23 +3809,86 @@ class TestMcpParallelToolCalls: finally: with _lock: _parallel_safe_servers.discard("docs") + _mcp_tool_server_names.pop("mcp_docs_search", None) + _mcp_tool_server_names.pop("mcp_docs_read_file", None) + _mcp_tool_server_names.pop("mcp_github_list_repos", None) def test_is_mcp_tool_parallel_safe_server_with_underscores(self): """Server names containing underscores are correctly matched.""" - from tools.mcp_tool import is_mcp_tool_parallel_safe, _parallel_safe_servers, _lock + from tools.mcp_tool import ( + is_mcp_tool_parallel_safe, _mcp_tool_server_names, + _parallel_safe_servers, _lock, + ) with _lock: _parallel_safe_servers.add("my_server") + _mcp_tool_server_names["mcp_my_server_query"] = "my_server" try: assert is_mcp_tool_parallel_safe("mcp_my_server_query") is True finally: with _lock: _parallel_safe_servers.discard("my_server") + _mcp_tool_server_names.pop("mcp_my_server_query", None) + + def test_is_mcp_tool_parallel_safe_uses_exact_registered_server(self): + """Ambiguous MCP names must not match a shorter parallel-safe prefix.""" + from tools.mcp_tool import ( + is_mcp_tool_parallel_safe, _mcp_tool_server_names, + _parallel_safe_servers, _lock, + ) + with _lock: + _parallel_safe_servers.add("a") + _mcp_tool_server_names["mcp_a_search"] = "a" + _mcp_tool_server_names["mcp_a_b_tool"] = "a_b" + try: + assert is_mcp_tool_parallel_safe("mcp_a_search") is True + assert is_mcp_tool_parallel_safe("mcp_a_b_tool") is False + finally: + with _lock: + _parallel_safe_servers.discard("a") + _mcp_tool_server_names.pop("mcp_a_search", None) + _mcp_tool_server_names.pop("mcp_a_b_tool", None) + + def test_registered_tool_provenance_prevents_prefix_collision(self): + """Registration records exact server ownership for ambiguous names.""" + from tools.registry import registry + from tools.mcp_tool import ( + _mcp_tool_server_names, _parallel_safe_servers, + _register_server_tools, is_mcp_tool_parallel_safe, _lock, + ) + + server = _make_mock_server( + "a_b", + tools=[_make_mcp_tool("tool", "Ambiguous tool name")], + ) + registered = _register_server_tools("a_b", server, {}) + try: + assert registered == ["mcp_a_b_tool"] + with _lock: + assert _mcp_tool_server_names["mcp_a_b_tool"] == "a_b" + _parallel_safe_servers.add("a") + assert is_mcp_tool_parallel_safe("mcp_a_b_tool") is False + + with _lock: + _parallel_safe_servers.add("a_b") + assert is_mcp_tool_parallel_safe("mcp_a_b_tool") is True + finally: + for tool_name in registered: + registry.deregister(tool_name) + with _lock: + _parallel_safe_servers.discard("a") + _parallel_safe_servers.discard("a_b") + _mcp_tool_server_names.pop("mcp_a_b_tool", None) def test_is_mcp_tool_parallel_safe_no_tool_suffix(self): """Tool name that is just 'mcp_{server}' without a tool part returns False.""" - from tools.mcp_tool import is_mcp_tool_parallel_safe, _parallel_safe_servers, _lock + from tools.mcp_tool import ( + is_mcp_tool_parallel_safe, _mcp_tool_server_names, + _parallel_safe_servers, _lock, + ) with _lock: _parallel_safe_servers.add("docs") + _mcp_tool_server_names.pop("mcp_docs", None) + _mcp_tool_server_names.pop("mcp_docs_", None) try: # "mcp_docs" has no tool part after the server name assert is_mcp_tool_parallel_safe("mcp_docs") is False diff --git a/tests/tools/test_schema_sanitizer.py b/tests/tools/test_schema_sanitizer.py index 89fbcd91d2b..8c865e87b8d 100644 --- a/tests/tools/test_schema_sanitizer.py +++ b/tests/tools/test_schema_sanitizer.py @@ -304,6 +304,30 @@ def test_strip_none_returns_zero(): assert stripped == 0 + +def test_strip_responses_format_strips_format_keyword(): + """Responses-format: keyword should be stripped.""" + from tools.schema_sanitizer import strip_pattern_and_format + + tools = [ + { + "name": "get_event", + "parameters": { + "type": "object", + "properties": { + "ts": {"type": "string", "format": "date-time"}, + } + }, + "type": "function" + } + ] + + result, stripped = strip_pattern_and_format(tools) + assert stripped == 1, f"Expected 1 format stripped, got {stripped}" + assert "format" not in result[0]["parameters"]["properties"]["ts"], "format should be stripped" + assert result[0]["parameters"]["properties"]["ts"]["type"] == "string", "type should be preserved" + + def test_top_level_allof_stripped_for_codex_backend_compat(): """OpenAI Codex backend rejects top-level allOf/oneOf/anyOf/enum/not.""" tools = [_tool("memory", { @@ -360,3 +384,110 @@ def test_nested_allof_preserved(): nested = out[0]["function"]["parameters"]["properties"]["config"] assert "allOf" in nested assert nested["allOf"] == [{"required": ["mode"]}] + + +def test_strip_responses_format_tools(): + """strip_pattern_and_format should handle Responses-format tools (no function wrapper).""" + from tools.schema_sanitizer import strip_pattern_and_format + + # Responses-format: {"name": "...", "parameters": {...}, "type": "function"} + tools = [ + { + "name": "mcp_firecrawl_search", + "parameters": { + "type": "object", + "properties": { + "query": {"type": "string"}, + "includeDomains": { + "type": "array", + "items": { + "type": "string", + "pattern": "^(?=.{1,253}$)(?:[a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?\\.)+[a-z0-9][a-z0-9-]{0,61}[a-z0-9]$" + } + } + } + }, + "type": "function" + } + ] + + result, stripped = strip_pattern_and_format(tools) + assert stripped == 1, f"Expected 1 pattern stripped, got {stripped}" + + # Verify pattern keyword was removed from includeDomains + domains = result[0]["parameters"]["properties"]["includeDomains"]["items"] + assert "pattern" not in domains, f"pattern should be stripped: {domains}" + assert domains["type"] == "string", "type should be preserved" + + +def test_strip_responses_idempotent(): + """Second call on already-stripped Responses-format tools should return 0.""" + from tools.schema_sanitizer import strip_pattern_and_format + + tools = [ + { + "name": "search_files", + "parameters": { + "type": "object", + "properties": { + "pattern": {"type": "string"} # This is a property named pattern, NOT schema keyword + } + } + } + ] + + # Pass 1 - property named 'pattern' should NOT be stripped + result, first = strip_pattern_and_format(tools) + assert first == 0, f"Expected 0 stripped (property pattern preserved), got {first}" + assert "pattern" in result[0]["parameters"]["properties"], "property named pattern should survive" + + # Pass 2 - idempotent + _, second = strip_pattern_and_format(tools) + assert second == 0, f"Expected 0 on second pass, got {second}" + + +def test_strip_responses_mixed_formats(): + """Mixed list of OpenAI-format and Responses-format tools should both be sanitized.""" + from tools.schema_sanitizer import strip_pattern_and_format + + tools = [ + # OpenAI-format: {"function": {"parameters": {...}}} + { + "type": "function", + "function": { + "name": "search", + "parameters": { + "type": "object", + "properties": { + "query": {"type": "string", "pattern": "^[a-z]+$"} + } + } + } + }, + # Responses-format: {"name": "...", "parameters": {...}} + { + "name": "get_time", + "parameters": { + "type": "object", + "properties": { + "tz": {"type": "string", "format": "date-time"} + } + }, + "type": "function" + } + ] + + result, stripped = strip_pattern_and_format(tools) + assert stripped == 2, f"Expected 2 stripped (1 pattern + 1 format), got {stripped}" + + # OpenAI-format tool: pattern stripped from parameters + openai_params = result[0]["function"]["parameters"]["properties"]["query"] + assert "pattern" not in openai_params, f"pattern should be stripped: {openai_params}" + + # Responses-format tool: format stripped + resp_params = result[1]["parameters"]["properties"]["tz"] + assert "format" not in resp_params, f"format should be stripped: {resp_params}" + + # Verify structure preserved + assert result[0]["function"]["parameters"]["type"] == "object" + assert result[1]["parameters"]["type"] == "object" diff --git a/tests/tools/test_send_message_tool.py b/tests/tools/test_send_message_tool.py index fa810eb5c54..dac476749fd 100644 --- a/tests/tools/test_send_message_tool.py +++ b/tests/tools/test_send_message_tool.py @@ -182,6 +182,81 @@ class TestSendMessageTool: force_document=False, ) + def test_resolved_slack_thread_name_preserves_thread_id(self): + slack_cfg = SimpleNamespace(enabled=True, token="xoxb-test", extra={}) + config = SimpleNamespace( + platforms={Platform.SLACK: slack_cfg}, + get_home_channel=lambda _platform: None, + ) + + with patch("gateway.config.load_gateway_config", return_value=config), \ + patch("tools.interrupt.is_interrupted", return_value=False), \ + patch("gateway.channel_directory.resolve_channel_name", return_value="C123ABCDEF:171.000001"), \ + patch("model_tools._run_async", side_effect=_run_async_immediately), \ + patch("tools.send_message_tool._send_to_platform", new=AsyncMock(return_value={"success": True})) as send_mock, \ + patch("gateway.mirror.mirror_to_session", return_value=True): + result = json.loads( + send_message_tool( + { + "action": "send", + "target": "slack:ops / topic 171.000001", + "message": "hello", + } + ) + ) + + assert result["success"] is True + send_mock.assert_awaited_once_with( + Platform.SLACK, + slack_cfg, + "C123ABCDEF", + "hello", + thread_id="171.000001", + media_files=[], + force_document=False, + ) + + def test_resolved_matrix_thread_name_preserves_thread_id(self): + matrix_cfg = SimpleNamespace( + enabled=True, + token="tok", + extra={"homeserver": "https://matrix.example.com"}, + ) + config = SimpleNamespace( + platforms={Platform.MATRIX: matrix_cfg}, + get_home_channel=lambda _platform: None, + ) + + with patch("gateway.config.load_gateway_config", return_value=config), \ + patch("tools.interrupt.is_interrupted", return_value=False), \ + patch( + "gateway.channel_directory.resolve_channel_name", + return_value="!roomid:matrix.example.org:$thread123:matrix.example.org", + ), \ + patch("model_tools._run_async", side_effect=_run_async_immediately), \ + patch("tools.send_message_tool._send_to_platform", new=AsyncMock(return_value={"success": True})) as send_mock, \ + patch("gateway.mirror.mirror_to_session", return_value=True): + result = json.loads( + send_message_tool( + { + "action": "send", + "target": "matrix:Ops / topic $thread123", + "message": "hello", + } + ) + ) + + assert result["success"] is True + send_mock.assert_awaited_once_with( + Platform.MATRIX, + matrix_cfg, + "!roomid:matrix.example.org", + "hello", + thread_id="$thread123:matrix.example.org", + media_files=[], + force_document=False, + ) + def test_mirror_receives_current_session_user_id(self): config, _telegram_cfg = _make_config() @@ -503,9 +578,8 @@ class TestSendToPlatformChunking: assert all(call == [] for call in sent_calls[:-1]) assert sent_calls[-1] == media - def test_matrix_media_uses_native_adapter_helper(self): - - doc_path = Path("/tmp/test-send-message-matrix.pdf") + def test_matrix_media_uses_native_adapter_helper(self, tmp_path): + doc_path = tmp_path / "test-send-message-matrix.pdf" doc_path.write_bytes(b"%PDF-1.4 test") try: @@ -847,6 +921,16 @@ class TestParseTargetRefDiscord: class TestParseTargetRefMatrix: """_parse_target_ref correctly handles Matrix room IDs and user MXIDs.""" + def test_matrix_thread_target_is_explicit(self): + """Session-derived Matrix thread targets round-trip as room + event id.""" + chat_id, thread_id, is_explicit = _parse_target_ref( + "matrix", + "!HLOQwxYGgFPMPJUSNR:matrix.org:$thread123:matrix.org", + ) + assert chat_id == "!HLOQwxYGgFPMPJUSNR:matrix.org" + assert thread_id == "$thread123:matrix.org" + assert is_explicit is True + def test_matrix_room_id_is_explicit(self): """Matrix room IDs (!) are recognized as explicit targets.""" chat_id, thread_id, is_explicit = _parse_target_ref("matrix", "!HLOQwxYGgFPMPJUSNR:matrix.org") @@ -919,6 +1003,12 @@ class TestParseTargetRefE164: class TestParseTargetRefSlack: """_parse_target_ref recognizes Slack channel/user IDs as explicit.""" + def test_thread_target_is_explicit(self): + chat_id, thread_id, is_explicit = _parse_target_ref("slack", "C0B0QV5434G:171.000001") + assert chat_id == "C0B0QV5434G" + assert thread_id == "171.000001" + assert is_explicit is True + def test_public_channel_id_is_explicit(self): chat_id, thread_id, is_explicit = _parse_target_ref("slack", "C0B0QV5434G") assert chat_id == "C0B0QV5434G" diff --git a/tests/tools/test_singularity_preflight.py b/tests/tools/test_singularity_preflight.py index 0ba50c3e93d..fa0a0ea4d52 100644 --- a/tests/tools/test_singularity_preflight.py +++ b/tests/tools/test_singularity_preflight.py @@ -23,7 +23,7 @@ class TestFindSingularityExecutable: def test_prefers_apptainer(self): """When both are available, apptainer should be preferred.""" def which_both(name): - return f"/usr/bin/{name}" if name in ("apptainer", "singularity") else None + return f"/usr/bin/{name}" if name in {"apptainer", "singularity"} else None with patch("shutil.which", side_effect=which_both): assert _find_singularity_executable() == "apptainer" diff --git a/tests/tools/test_skill_manager_tool.py b/tests/tools/test_skill_manager_tool.py index 96c3a361f0c..33efbb98ae8 100644 --- a/tests/tools/test_skill_manager_tool.py +++ b/tests/tools/test_skill_manager_tool.py @@ -547,7 +547,7 @@ class TestSkillManageDispatcher: # No provenance marker on a foreground create — record either missing # entirely (telemetry best-effort) or present with created_by unset. rec = usage.get("test-skill") or {} - assert rec.get("created_by") in (None, "", False) + assert rec.get("created_by") in {None, "", False} def test_create_from_background_review_marks_agent_created(self, tmp_path): """Background-review fork creates ARE marked as agent-created.""" diff --git a/tests/tools/test_skills_hub.py b/tests/tools/test_skills_hub.py index b7c483d1a16..e831b50943e 100644 --- a/tests/tools/test_skills_hub.py +++ b/tests/tools/test_skills_hub.py @@ -101,7 +101,7 @@ class TestTrustLevelFor: src = self._source() result = src.trust_level_for("owner/repo") # No path part — still resolves repo correctly - assert result in ("trusted", "community") + assert result in {"trusted", "community"} # --------------------------------------------------------------------------- diff --git a/tests/tools/test_transcription_dotenv_fallback.py b/tests/tools/test_transcription_dotenv_fallback.py index 7188487a96b..a49d3fa455c 100644 --- a/tests/tools/test_transcription_dotenv_fallback.py +++ b/tests/tools/test_transcription_dotenv_fallback.py @@ -60,6 +60,33 @@ class TestProviderSelectionGate: finally: importlib.reload(tt) + def test_xai_resolver_import_after_config_env_patch_uses_restored_dotenv_loader(self): + """xAI HTTP auth must not cache a temporarily patched env helper.""" + import importlib + import hermes_cli.config as config_mod + from tools import xai_http + + with pytest.MonkeyPatch.context() as mp: + mp.setattr(config_mod, "get_env_value", lambda name, default=None: "") + xai_http = importlib.reload(xai_http) + + try: + with patch( + "hermes_cli.runtime_provider.resolve_runtime_provider", + side_effect=RuntimeError("no oauth"), + ), patch( + "hermes_cli.auth.resolve_xai_oauth_runtime_credentials", + return_value={}, + ), patch( + "hermes_cli.config.load_env", + return_value={"XAI_API_KEY": "dotenv-secret"}, + ): + creds = xai_http.resolve_xai_http_credentials() + finally: + importlib.reload(xai_http) + + assert creds["api_key"] == "dotenv-secret" + def test_explicit_groq_sees_dotenv(self): from tools import transcription_tools as tt diff --git a/tests/tools/test_voice_cli_integration.py b/tests/tools/test_voice_cli_integration.py index 93dffa649a7..a6cf5e36627 100644 --- a/tests/tools/test_voice_cli_integration.py +++ b/tests/tools/test_voice_cli_integration.py @@ -482,8 +482,11 @@ class TestVprintForceParameter: else: unforced_error_count += 1 - assert forced_error_count > 0, \ - "Expected at least one _vprint with force=True for error messages" + # Invariant: no critical-error _vprint call may silently drop under + # streaming suppression — every ❌-prefixed _vprint must pass force=True. + # The codebase may legitimately have zero such calls if errors are + # routed through print() or higher-level Rich panels; what matters is + # that none are quietly suppressed. assert unforced_error_count == 0, \ f"Found {unforced_error_count} critical error _vprint calls without force=True" diff --git a/tests/tui_gateway/test_entry_sys_path.py b/tests/tui_gateway/test_entry_sys_path.py index f8741b18e4b..e7f9e47cee0 100644 --- a/tests/tui_gateway/test_entry_sys_path.py +++ b/tests/tui_gateway/test_entry_sys_path.py @@ -25,7 +25,7 @@ def _reload_entry_with_env(env_overrides: dict) -> None: _src_root = os.environ.get("HERMES_PYTHON_SRC_ROOT", "") if _src_root and _src_root not in sys.path: sys.path.insert(0, _src_root) - sys.path = [p for p in sys.path if p not in ("", ".")] + sys.path = [p for p in sys.path if p not in {"", "."}] return sys.path[:] finally: sys.path = original_path @@ -45,7 +45,7 @@ def test_empty_string_and_dot_removed_from_sys_path(): assert "." in sys.path # Run the entry.py fixup logic directly - sys.path = [p for p in sys.path if p not in ("", ".")] + sys.path = [p for p in sys.path if p not in {"", "."}] assert "" not in sys.path assert "." not in sys.path @@ -61,7 +61,7 @@ def test_hermes_src_root_inserted_at_front(): _src_root = os.environ.get("HERMES_PYTHON_SRC_ROOT", "") if _src_root and _src_root not in sys.path: sys.path.insert(0, _src_root) - sys.path = [p for p in sys.path if p not in ("", ".")] + sys.path = [p for p in sys.path if p not in {"", "."}] assert sys.path[0] == fake_root finally: @@ -79,7 +79,7 @@ def test_src_root_not_duplicated_if_already_present(): _src_root = os.environ.get("HERMES_PYTHON_SRC_ROOT", "") if _src_root and _src_root not in sys.path: sys.path.insert(0, _src_root) - sys.path = [p for p in sys.path if p not in ("", ".")] + sys.path = [p for p in sys.path if p not in {"", "."}] assert sys.path.count(fake_root) == count_before finally: @@ -95,7 +95,7 @@ def test_no_src_root_env_does_not_crash(): _src_root = os.environ.get("HERMES_PYTHON_SRC_ROOT", "") if _src_root and _src_root not in sys.path: sys.path.insert(0, _src_root) - sys.path = [p for p in sys.path if p not in ("", ".")] + sys.path = [p for p in sys.path if p not in {"", "."}] # No exception raised finally: sys.path = original diff --git a/tools/browser_providers/__init__.py b/tools/browser_providers/__init__.py deleted file mode 100644 index 7fa59ef04ee..00000000000 --- a/tools/browser_providers/__init__.py +++ /dev/null @@ -1,10 +0,0 @@ -"""Cloud browser provider abstraction. - -Import the ABC so callers can do:: - - from tools.browser_providers import CloudBrowserProvider -""" - -from tools.browser_providers.base import CloudBrowserProvider - -__all__ = ["CloudBrowserProvider"] diff --git a/tools/browser_providers/base.py b/tools/browser_providers/base.py deleted file mode 100644 index 6b8e1ed4f6b..00000000000 --- a/tools/browser_providers/base.py +++ /dev/null @@ -1,59 +0,0 @@ -"""Abstract base class for cloud browser providers.""" - -from abc import ABC, abstractmethod -from typing import Dict - - -class CloudBrowserProvider(ABC): - """Interface for cloud browser backends (Browserbase, Steel, etc.). - - Implementations live in sibling modules and are registered in - ``browser_tool._PROVIDER_REGISTRY``. The user selects a provider via - ``hermes setup`` / ``hermes tools``; the choice is persisted as - ``config["browser"]["cloud_provider"]``. - """ - - @abstractmethod - def provider_name(self) -> str: - """Short, human-readable name shown in logs and diagnostics.""" - - @abstractmethod - def is_configured(self) -> bool: - """Return True when all required env vars / credentials are present. - - Called at tool-registration time (``check_browser_requirements``) to - gate availability. Must be cheap — no network calls. - """ - - @abstractmethod - def create_session(self, task_id: str) -> Dict[str, object]: - """Create a cloud browser session and return session metadata. - - Must return a dict with at least:: - - { - "session_name": str, # unique name for agent-browser --session - "bb_session_id": str, # provider session ID (for close/cleanup) - "cdp_url": str, # CDP websocket URL - "features": dict, # feature flags that were enabled - } - - ``bb_session_id`` is a legacy key name kept for backward compat with - the rest of browser_tool.py — it holds the provider's session ID - regardless of which provider is in use. - """ - - @abstractmethod - def close_session(self, session_id: str) -> bool: - """Release / terminate a cloud session by its provider session ID. - - Returns True on success, False on failure. Should not raise. - """ - - @abstractmethod - def emergency_cleanup(self, session_id: str) -> None: - """Best-effort session teardown during process exit. - - Called from atexit / signal handlers. Must tolerate missing - credentials, network errors, etc. — log and move on. - """ diff --git a/tools/browser_tool.py b/tools/browser_tool.py index b3eb24ee044..fb96649cb38 100644 --- a/tools/browser_tool.py +++ b/tools/browser_tool.py @@ -83,10 +83,24 @@ try: except Exception: _is_safe_url = lambda url: False # noqa: E731 — fail-closed: block all if safety module unavailable _is_always_blocked_url = lambda url: True # noqa: E731 — fail-closed on the floor too -from tools.browser_providers.base import CloudBrowserProvider -from tools.browser_providers.browserbase import BrowserbaseProvider -from tools.browser_providers.browser_use import BrowserUseProvider -from tools.browser_providers.firecrawl import FirecrawlProvider +# Browser-provider ABC + registry — PR #25214 moved the per-vendor providers +# (Browserbase / Browser Use / Firecrawl) out of ``tools/browser_providers/`` +# and into ``plugins/browser//``. The dispatcher consults the +# registry; the legacy class names are re-exported below as backward-compat +# shims for callers that import them from this module. +from agent.browser_provider import BrowserProvider as CloudBrowserProvider # noqa: F401 (legacy alias) +from agent.browser_registry import ( # noqa: F401 (test-patchable surface) + get_provider as _registry_get_browser_provider, +) +from plugins.browser.browserbase.provider import ( # noqa: F401 (legacy import surface) + BrowserbaseBrowserProvider as BrowserbaseProvider, +) +from plugins.browser.browser_use.provider import ( # noqa: F401 + BrowserUseBrowserProvider as BrowserUseProvider, +) +from plugins.browser.firecrawl.provider import ( # noqa: F401 + FirecrawlBrowserProvider as FirecrawlProvider, +) from tools.tool_backend_helpers import normalize_browser_cloud_provider # Camofox local anti-detection browser backend (optional). @@ -391,12 +405,29 @@ def _stop_cdp_supervisor(task_id: str) -> None: # ============================================================================ # Cloud Provider Registry # ============================================================================ +# +# Per-vendor browser providers (Browserbase / Browser Use / Firecrawl) live as +# plugins under ``plugins/browser//`` and self-register through +# :mod:`agent.browser_registry` at plugin-discovery time. The legacy +# class-name registry below is preserved as a backward-compat shim so test +# fixtures that ``monkeypatch.setattr(browser_tool, "_PROVIDER_REGISTRY", ...)`` +# keep working — but ``_get_cloud_provider()`` now consults +# :mod:`agent.browser_registry` for the actual lookup. +# +# When the test patches ``_PROVIDER_REGISTRY``, we honour it (so the cache +# unit tests still drive the function); otherwise the registry-backed path +# wins. This keeps the test surface stable while letting third-party +# plugins drop in under ``~/.hermes/plugins/browser//``. _PROVIDER_REGISTRY: Dict[str, type] = { "browserbase": BrowserbaseProvider, "browser-use": BrowserUseProvider, "firecrawl": FirecrawlProvider, } +# Frozen copy of the import-time _PROVIDER_REGISTRY, used by +# ``_is_legacy_provider_registry_overridden`` to detect test-time +# monkeypatching. NEVER mutate this dict. +_DEFAULT_PROVIDER_REGISTRY: Dict[str, type] = dict(_PROVIDER_REGISTRY) _cached_cloud_provider: Optional[CloudBrowserProvider] = None _cloud_provider_resolved = False @@ -411,13 +442,65 @@ _cached_browser_engine: Optional[str] = None _browser_engine_resolved = False +def _is_legacy_provider_registry_overridden() -> bool: + """Return True when a test has patched ``_PROVIDER_REGISTRY`` to a custom value. + + Detected by spotting any registered class that *isn't* the canonical + plugin-backed class for that name. Tests that + ``monkeypatch.setattr(browser_tool, "_PROVIDER_REGISTRY", ...)`` install + custom factories (`exploding_factory`, `lambda: fake_provider`, etc.); + those entries fail the canonical-class identity check below. + + Note: a future maintainer adding a 4th built-in provider only needs to + extend ``_DEFAULT_PROVIDER_REGISTRY`` below — they do NOT need to update + a hardcoded set of keys here. The detection just compares each registered + value against the corresponding canonical class. + """ + try: + for key, default_cls in _DEFAULT_PROVIDER_REGISTRY.items(): + if _PROVIDER_REGISTRY.get(key) is not default_cls: + return True + # Extra keys not in the default registry → also an override. + return len(_PROVIDER_REGISTRY) != len(_DEFAULT_PROVIDER_REGISTRY) + except Exception: + return False + + +def _ensure_browser_plugins_loaded() -> None: + """Idempotently trigger plugin discovery so the browser registry is populated. + + Normally `model_tools` is imported early in any session and that + triggers `discover_plugins()` as a side effect. But `_get_cloud_provider` + can be called from contexts that haven't gone through `model_tools` — + standalone scripts, certain unit-test paths, the parity-sweep harness. + Make discovery idempotent and side-effect-only here so users always + see registered plugins regardless of import order. Cheap: subsequent + calls early-return inside `_ensure_plugins_discovered`. + """ + try: + from hermes_cli.plugins import _ensure_plugins_discovered + + _ensure_plugins_discovered() + except Exception as exc: + logger.debug("Browser plugin discovery failed (non-fatal): %s", exc) + + def _get_cloud_provider() -> Optional[CloudBrowserProvider]: """Return the configured cloud browser provider, or None for local mode. Reads ``config["browser"]["cloud_provider"]`` once and caches the result for the process lifetime. An explicit ``local`` provider disables cloud - fallback. If unset, fall back to Browserbase when direct or managed - Browserbase credentials are available. + fallback. If unset, fall back to Browser Use (managed Nous gateway or + direct API key) and then Browserbase (direct credentials only) — the + historic auto-detect order, now expressed as the + :data:`agent.browser_registry._LEGACY_PREFERENCE` walk. + + Selection routes through :mod:`agent.browser_registry` so third-party + browser plugins (``~/.hermes/plugins/browser//``) participate + in explicit-config resolution. Test fixtures that override + ``_PROVIDER_REGISTRY`` or ``BrowserUseProvider`` / ``BrowserbaseProvider`` + on this module still drive the function — see + ``_is_legacy_provider_registry_overridden``. """ global _cached_cloud_provider, _cloud_provider_resolved if _cloud_provider_resolved: @@ -437,9 +520,33 @@ def _get_cloud_provider() -> Optional[CloudBrowserProvider]: _cached_cloud_provider = None _cloud_provider_resolved = True return None - if provider_key and provider_key in _PROVIDER_REGISTRY: + if provider_key: try: - resolved = _PROVIDER_REGISTRY[provider_key]() + if _is_legacy_provider_registry_overridden(): + # Test fixture path: honour the patched dict so the + # cache-policy unit tests keep working. + factory = _PROVIDER_REGISTRY.get(provider_key) + if factory is not None: + resolved = factory() + else: + # Ensure plugins are discovered so the registry is + # populated. Idempotent — cheap on subsequent calls. + _ensure_browser_plugins_loaded() + resolved = _registry_get_browser_provider(provider_key) + if resolved is None: + # Explicit config name unknown to the registry — + # might be a typo, an uninstalled plugin, or a + # registry-population failure. Warn the user + # (legacy code would have surfaced a typed + # credentials error via direct class instantiation; + # post-migration we surface this WARNING instead). + logger.warning( + "browser.cloud_provider=%r is not a registered " + "browser plugin; falling back to auto-detect " + "(install the corresponding plugin or fix the " + "config key spelling).", + provider_key, + ) except Exception: logger.warning( "Failed to instantiate explicit cloud_provider %r; will retry on next call", @@ -453,8 +560,15 @@ def _get_cloud_provider() -> Optional[CloudBrowserProvider]: logger.debug("Could not read cloud_provider from config: %s", e) if resolved is None: - # Prefer Browser Use (managed Nous gateway or direct API key), - # fall back to Browserbase (direct credentials only). + # Auto-detect path: Browser Use first (managed Nous gateway or + # direct API key), then Browserbase (direct credentials). Uses + # the legacy class names imported at the top of this module so + # tests that ``monkeypatch.setattr(browser_tool, "BrowserUseProvider", ...)`` + # keep driving this branch deterministically. Third-party browser + # plugins are intentionally NOT reachable from auto-detect — they + # participate only via explicit ``browser.cloud_provider: ``, + # mirroring the firecrawl gate documented on + # :data:`agent.browser_registry._LEGACY_PREFERENCE`. try: fallback_provider = BrowserUseProvider() if fallback_provider.is_configured(): diff --git a/tools/code_execution_tool.py b/tools/code_execution_tool.py index 3822ce539f2..bdbc4bfbe1b 100644 --- a/tools/code_execution_tool.py +++ b/tools/code_execution_tool.py @@ -1238,6 +1238,7 @@ def execute_code( stderr=subprocess.PIPE, stdin=subprocess.DEVNULL, preexec_fn=None if _IS_WINDOWS else os.setsid, + creationflags=subprocess.CREATE_NO_WINDOW if _IS_WINDOWS else 0, ) # --- Poll loop: watch for exit, timeout, and interrupt --- @@ -1568,6 +1569,7 @@ def _is_usable_python(python_path: str) -> bool: "import sys; sys.exit(0 if sys.version_info >= (3, 8) else 1)"], timeout=5, capture_output=True, + creationflags=subprocess.CREATE_NO_WINDOW if _IS_WINDOWS else 0, ) return result.returncode == 0 except (OSError, subprocess.TimeoutExpired, subprocess.SubprocessError): diff --git a/tools/delegate_tool.py b/tools/delegate_tool.py index e9ad32e0d3a..86dcd0715cc 100644 --- a/tools/delegate_tool.py +++ b/tools/delegate_tool.py @@ -31,6 +31,11 @@ from concurrent.futures import ( from typing import Any, Dict, List, Optional from toolsets import TOOLSETS + +# Sentinel value used by the runtime provider system for providers that are +# not natively known (named custom providers, third-party aggregators, etc.). +# Must match hermes_cli.runtime_provider.RUNTIME_PROVIDER_TYPE_CUSTOM. +_RUNTIME_PROVIDER_CUSTOM = "custom" from tools import file_state from tools.terminal_tool import set_approval_callback as _set_subagent_approval_cb from utils import base_url_hostname, is_truthy_value @@ -2442,7 +2447,7 @@ def _resolve_delegation_credentials(cfg: dict, parent_agent) -> dict: return { "model": configured_model or runtime.get("model") or None, - "provider": runtime.get("provider"), + "provider": configured_provider if runtime.get("provider") == _RUNTIME_PROVIDER_CUSTOM else runtime.get("provider"), "base_url": runtime.get("base_url"), "api_key": api_key, "api_mode": runtime.get("api_mode"), diff --git a/tools/environments/local.py b/tools/environments/local.py index 3b9d65449fa..177e5efab15 100644 --- a/tools/environments/local.py +++ b/tools/environments/local.py @@ -513,6 +513,7 @@ class LocalEnvironment(BaseEnvironment): stderr=subprocess.STDOUT, stdin=subprocess.PIPE if stdin_data is not None else subprocess.DEVNULL, preexec_fn=None if _IS_WINDOWS else os.setsid, + creationflags=subprocess.CREATE_NO_WINDOW if _IS_WINDOWS else 0, cwd=_popen_cwd, ) if not _IS_WINDOWS: diff --git a/tools/lazy_deps.py b/tools/lazy_deps.py index faaf7ec42bf..c7d7730c756 100644 --- a/tools/lazy_deps.py +++ b/tools/lazy_deps.py @@ -450,7 +450,7 @@ def ensure(feature: str, *, prompt: bool = True) -> None: ).strip().lower() except (EOFError, KeyboardInterrupt): answer = "n" - if answer and answer not in ("y", "yes"): + if answer and answer not in {"y", "yes"}: raise FeatureUnavailable( feature, missing, "user declined install at prompt" ) diff --git a/tools/mcp_oauth.py b/tools/mcp_oauth.py index d7bf135da47..8d48eedf0e8 100644 --- a/tools/mcp_oauth.py +++ b/tools/mcp_oauth.py @@ -401,6 +401,23 @@ async def _redirect_handler(authorization_url: str) -> None: ) print(msg, file=sys.stderr) + # On a remote SSH session the OAuth provider redirects to + # http://127.0.0.1:/callback, which reaches the callback server on + # the *remote* machine — not the user's local machine where the browser + # opened. Print a port-forward hint so the user knows to tunnel first. + if _oauth_port and (os.getenv("SSH_CLIENT") or os.getenv("SSH_TTY")): + print( + f" Remote session detected. The OAuth provider will redirect your browser to\n" + f" http://127.0.0.1:{_oauth_port}/callback\n" + f" which the callback listener on THIS machine is waiting on. If your browser\n" + f" is on a different machine, forward the port first in a separate terminal:\n" + f"\n" + f" ssh -N -L {_oauth_port}:127.0.0.1:{_oauth_port} @\n" + f"\n" + f" Then open the URL above. See: https://hermes-agent.nousresearch.com/docs/guides/oauth-over-ssh\n", + file=sys.stderr, + ) + if _can_open_browser(): try: opened = webbrowser.open(authorization_url) diff --git a/tools/mcp_tool.py b/tools/mcp_tool.py index b24bb9705ad..dfe0b8c4750 100644 --- a/tools/mcp_tool.py +++ b/tools/mcp_tool.py @@ -91,6 +91,7 @@ import threading import time from datetime import datetime from typing import Any, Dict, List, Optional +from urllib.parse import urlparse logger = logging.getLogger(__name__) @@ -492,6 +493,73 @@ def _cache_mcp_image_block(block) -> str: return f"MEDIA:{image_path}" +# --------------------------------------------------------------------------- +# Remote MCP URL validation +# --------------------------------------------------------------------------- + + +class InvalidMcpUrlError(ValueError): + """Raised when a remote MCP server's ``url`` cannot be parsed as http(s)://. + + Validated once at startup so we fail fast with a clear message instead of + burning through the reconnect-backoff loop on every attempt. (Ported from + anomalyco/opencode#25019.) + """ + + +def _validate_remote_mcp_url(server_name: str, url: Any) -> str: + """Return the URL as a string if it's a valid http(s) remote MCP URL. + + Raises :class:`InvalidMcpUrlError` otherwise with a message naming the + offending server, so users can spot the bad entry in their config. + + Accepts: + - ``http://host`` / ``https://host`` with optional port, path, query + - IPv4, IPv6 (bracketed), DNS hostnames + + Rejects: + - Non-string values (``None``, dicts, ints) + - Missing scheme (``example.com/mcp``) + - Non-http(s) schemes (``file://``, ``ws://``, ``stdio:`` — stdio servers + use the ``command`` key, not ``url``) + - Empty host (``http://``, ``https:///path``) + """ + if not isinstance(url, str): + raise InvalidMcpUrlError( + f"Invalid MCP URL for '{server_name}': expected a string, got " + f"{type(url).__name__}" + ) + stripped = url.strip() + if not stripped: + raise InvalidMcpUrlError( + f"Invalid MCP URL for '{server_name}': empty url" + ) + try: + parsed = urlparse(stripped) + except Exception as exc: # urlparse is very permissive — belt and braces + raise InvalidMcpUrlError( + f"Invalid MCP URL for '{server_name}': {stripped!r} ({exc})" + ) from exc + if parsed.scheme.lower() not in {"http", "https"}: + raise InvalidMcpUrlError( + f"Invalid MCP URL for '{server_name}': scheme must be http or " + f"https, got {parsed.scheme!r} ({stripped!r})" + ) + if not parsed.netloc: + raise InvalidMcpUrlError( + f"Invalid MCP URL for '{server_name}': missing host ({stripped!r})" + ) + # ``urlparse`` accepts ``http://:8080`` (empty host, explicit port). + # Reject that — we need a real host. + if not parsed.hostname: + raise InvalidMcpUrlError( + f"Invalid MCP URL for '{server_name}': missing hostname " + f"({stripped!r})" + ) + return stripped + + + def _format_connect_error(exc: BaseException) -> str: """Render nested MCP connection errors into an actionable short message.""" @@ -1094,6 +1162,7 @@ class MCPServerTask: } for tool_name in stale_tool_names: registry.deregister(tool_name) + _forget_mcp_tool_server(tool_name) # 3. Re-register with fresh tool list self._tools = new_mcp_tools @@ -1614,6 +1683,7 @@ class MCPServerTask: self._pending_refresh_tasks.clear() for tool_name in list(getattr(self, "_registered_tool_names", [])): registry.deregister(tool_name) + _forget_mcp_tool_server(tool_name) self._registered_tool_names = [] self.session = None @@ -1984,11 +2054,20 @@ def _handle_session_expired_and_retry( # ``is_mcp_tool_parallel_safe()`` for the parallel-execution check in run_agent. _parallel_safe_servers: set = set() +# Exact MCP tool-name provenance. MCP tool names are formatted as +# ``mcp_{sanitized_server}_{sanitized_tool}``, which is ambiguous when server +# names contain underscores (``mcp_a_b_tool`` could be server ``a`` + tool +# ``b_tool`` or server ``a_b`` + tool ``tool``). Keep the server component +# captured at registration time so parallel safety never relies on prefix +# guessing. +_mcp_tool_server_names: Dict[str, str] = {} + # Dedicated event loop running in a background daemon thread. _mcp_loop: Optional[asyncio.AbstractEventLoop] = None _mcp_thread: Optional[threading.Thread] = None -# Protects _mcp_loop, _mcp_thread, _servers, _parallel_safe_servers, and _stdio_pids. +# Protects _mcp_loop, _mcp_thread, _servers, _parallel_safe_servers, +# _mcp_tool_server_names, and _stdio_pids. _lock = threading.Lock() # PIDs of stdio MCP server subprocesses. Tracked so we can force-kill @@ -2871,6 +2950,19 @@ _UTILITY_CAPABILITY_ATTRS = { } +def _track_mcp_tool_server(tool_name: str, server_name: str) -> None: + """Remember the exact MCP server that registered *tool_name*.""" + safe_server_name = sanitize_mcp_name_component(server_name) + with _lock: + _mcp_tool_server_names[tool_name] = safe_server_name + + +def _forget_mcp_tool_server(tool_name: str) -> None: + """Forget MCP server provenance for a deregistered tool.""" + with _lock: + _mcp_tool_server_names.pop(tool_name, None) + + def _select_utility_schemas(server_name: str, server: MCPServerTask, config: dict) -> List[dict]: """Select utility schemas based on config and server capabilities.""" tools_filter = config.get("tools") or {} @@ -3005,6 +3097,7 @@ def _register_server_tools(name: str, server: MCPServerTask, config: dict) -> Li is_async=False, description=schema["description"], ) + _track_mcp_tool_server(tool_name_prefixed, name) registered_names.append(tool_name_prefixed) # Register MCP Resources & Prompts utility tools, filtered by config and @@ -3041,6 +3134,7 @@ def _register_server_tools(name: str, server: MCPServerTask, config: dict) -> Li is_async=False, description=schema["description"], ) + _track_mcp_tool_server(util_name, name) registered_names.append(util_name) if registered_names: @@ -3225,24 +3319,19 @@ def discover_mcp_tools() -> List[str]: def is_mcp_tool_parallel_safe(tool_name: str) -> bool: """Check if an MCP tool belongs to a server that supports parallel tool calls. - MCP tool names follow the pattern ``mcp_{server}_{tool}``. This extracts - the server component and checks it against the set of servers whose config - includes ``supports_parallel_tool_calls: true``. + MCP tool names follow the pattern ``mcp_{server}_{tool}``, but that string + shape is ambiguous when server names contain underscores. Use the exact + server provenance captured at registration time rather than prefix + matching, then check whether that server's config includes + ``supports_parallel_tool_calls: true``. Returns False for non-MCP tools or tools from servers without the flag. """ if not tool_name.startswith("mcp_"): return False - # Strip the "mcp_" prefix and extract the server name. - # Tool names are: mcp_{sanitized_server}_{sanitized_tool} - # We need to check all possible server prefixes because the server name - # itself may contain underscores after sanitization. - rest = tool_name[4:] # strip "mcp_" with _lock: - for server_name in _parallel_safe_servers: - if rest.startswith(server_name + "_") and len(rest) > len(server_name) + 1: - return True - return False + server_name = _mcp_tool_server_names.get(tool_name) + return bool(server_name and server_name in _parallel_safe_servers) def get_mcp_status() -> List[dict]: @@ -3415,7 +3504,6 @@ def _kill_orphaned_mcp_children(include_active: bool = False) -> None: sessions can still be in flight. """ import signal as _signal - import time as _time with _lock: pids: Dict[int, str] = {} @@ -3440,7 +3528,7 @@ def _kill_orphaned_mcp_children(include_active: bool = False) -> None: pass # Phase 2: Wait for graceful exit - _time.sleep(2) + time.sleep(2) # Phase 3: SIGKILL any survivors _sigkill = getattr(_signal, "SIGKILL", _signal.SIGTERM) diff --git a/tools/process_registry.py b/tools/process_registry.py index 1f31527475a..5a48f9f3906 100644 --- a/tools/process_registry.py +++ b/tools/process_registry.py @@ -557,6 +557,7 @@ class ProcessRegistry: stderr=subprocess.STDOUT, stdin=subprocess.PIPE, preexec_fn=None if _IS_WINDOWS else os.setsid, + creationflags=subprocess.CREATE_NO_WINDOW if _IS_WINDOWS else 0, ) session.process = proc diff --git a/tools/schema_sanitizer.py b/tools/schema_sanitizer.py index 87587c7fed5..0d03998d366 100644 --- a/tools/schema_sanitizer.py +++ b/tools/schema_sanitizer.py @@ -355,11 +355,23 @@ def strip_pattern_and_format(tools: list[dict]) -> tuple[list[dict], int]: _walk(item) for tool in tools: - fn = tool.get("function") if isinstance(tool, dict) else None + if not isinstance(tool, dict): + continue + + # OpenAI-format: {"function": {"parameters": {...}}} + fn = tool.get("function") if isinstance(fn, dict): params = fn.get("parameters") if isinstance(params, dict): _walk(params) + continue + + # Responses-format: {"name": "...", "parameters": {...}} + # (used by codex_responses API mode — xAI, OpenAI Codex, etc.) + params = tool.get("parameters") + if isinstance(params, dict): + _walk(params) + continue if stripped: logger.info( diff --git a/tools/send_message_tool.py b/tools/send_message_tool.py index d5b2c0c782c..bfe1a630707 100644 --- a/tools/send_message_tool.py +++ b/tools/send_message_tool.py @@ -28,6 +28,8 @@ _FEISHU_TARGET_RE = re.compile(r"^\s*((?:oc|ou|on|chat|open)_[-A-Za-z0-9]+)(?::( # conversations.open to obtain a D... ID. Without this gate, Slack IDs fall # through to channel-name resolution, which only matches by name and fails. _SLACK_TARGET_RE = re.compile(r"^\s*([CGD][A-Z0-9]{8,})\s*$") +# Session-derived Slack thread targets use ":". +_SLACK_THREAD_TARGET_RE = re.compile(r"^\s*([CGD][A-Z0-9]{8,}):([^\s:]+)\s*$") _WEIXIN_TARGET_RE = re.compile(r"^\s*((?:wxid|gh|v\d+|wm|wb)_[A-Za-z0-9_-]+|[A-Za-z0-9._-]+@chatroom|filehelper)\s*$") _YUANBAO_TARGET_RE = re.compile(r"^\s*((?:group|direct):[^:]+)\s*$") # Discord snowflake IDs are numeric, same regex pattern as Telegram topic targets. @@ -330,9 +332,17 @@ def _parse_target_ref(platform_name: str, target_ref: str): if match: return match.group(1), match.group(2), True if platform_name == "slack": + match = _SLACK_THREAD_TARGET_RE.fullmatch(target_ref) + if match: + return match.group(1), match.group(2), True match = _SLACK_TARGET_RE.fullmatch(target_ref) if match: return match.group(1), None, True + if platform_name == "matrix": + trimmed = target_ref.strip() + split_idx = trimmed.rfind(":$") + if split_idx > 0: + return trimmed[:split_idx], trimmed[split_idx + 1 :], True if platform_name == "weixin": match = _WEIXIN_TARGET_RE.fullmatch(target_ref) if match: diff --git a/tools/video_generation_tool.py b/tools/video_generation_tool.py index 63d80165dc0..472b8409255 100644 --- a/tools/video_generation_tool.py +++ b/tools/video_generation_tool.py @@ -286,9 +286,9 @@ def _coerce_bool(value: Any) -> Optional[bool]: return value if isinstance(value, str): v = value.strip().lower() - if v in ("true", "1", "yes", "on"): + if v in {"true", "1", "yes", "on"}: return True - if v in ("false", "0", "no", "off"): + if v in {"false", "0", "no", "off"}: return False return None diff --git a/tools/x_search_tool.py b/tools/x_search_tool.py index 8b242ee0ca8..1b7685a897d 100644 --- a/tools/x_search_tool.py +++ b/tools/x_search_tool.py @@ -147,7 +147,7 @@ def _extract_response_text(payload: Dict[str, Any]) -> str: continue for content in item.get("content", []) or []: ctype = content.get("type") - if ctype in ("output_text", "text"): + if ctype in {"output_text", "text"}: text = str(content.get("text") or "").strip() if text: parts.append(text) diff --git a/tools/xai_http.py b/tools/xai_http.py index 216a51ff10d..848ad8fc748 100644 --- a/tools/xai_http.py +++ b/tools/xai_http.py @@ -5,12 +5,6 @@ from __future__ import annotations import os from typing import Dict -try: - from hermes_cli.config import get_env_value as _hermes_get_env_value -except Exception: - _hermes_get_env_value = None - - def get_env_value(name: str, default=None): """Read ``name`` from ``~/.hermes/.env`` first, then ``os.environ``. @@ -18,10 +12,14 @@ def get_env_value(name: str, default=None): ``tools.xai_http.get_env_value`` to inject dotenv-only secrets into the xAI credential resolver. """ - if _hermes_get_env_value is not None: + try: + from hermes_cli.config import get_env_value as _hermes_get_env_value + value = _hermes_get_env_value(name) if value is not None: return value + except Exception: + pass return os.environ.get(name, default) diff --git a/ui-tui/packages/hermes-ink/index.d.ts b/ui-tui/packages/hermes-ink/index.d.ts index 5d5ae9387c0..66fed32ae60 100644 --- a/ui-tui/packages/hermes-ink/index.d.ts +++ b/ui-tui/packages/hermes-ink/index.d.ts @@ -34,5 +34,6 @@ export { default as measureElement } from './src/ink/measure-element.ts' export { createRoot, forceRedraw, default as render, renderSync } from './src/ink/root.ts' export type { Instance, RenderOptions, Root } from './src/ink/root.ts' export { stringWidth } from './src/ink/stringWidth.ts' +export { wrapAnsi } from './src/ink/wrapAnsi.ts' export { default as TextInput, UncontrolledTextInput } from 'ink-text-input' export type { Props as TextInputProps } from 'ink-text-input' diff --git a/ui-tui/packages/hermes-ink/src/entry-exports.ts b/ui-tui/packages/hermes-ink/src/entry-exports.ts index d173e0c9bb1..a113660385f 100644 --- a/ui-tui/packages/hermes-ink/src/entry-exports.ts +++ b/ui-tui/packages/hermes-ink/src/entry-exports.ts @@ -26,5 +26,6 @@ export { default as measureElement } from './ink/measure-element.js' export { scrollFastPathStats, type ScrollFastPathStats } from './ink/render-node-to-output.js' export { createRoot, forceRedraw, default as render, renderSync } from './ink/root.js' export { stringWidth } from './ink/stringWidth.js' +export { wrapAnsi } from './ink/wrapAnsi.js' export { isXtermJs } from './ink/terminal.js' export { default as TextInput, UncontrolledTextInput } from 'ink-text-input' diff --git a/ui-tui/src/__tests__/cursorDriftRegression.test.ts b/ui-tui/src/__tests__/cursorDriftRegression.test.ts new file mode 100644 index 00000000000..3f9082dcefc --- /dev/null +++ b/ui-tui/src/__tests__/cursorDriftRegression.test.ts @@ -0,0 +1,114 @@ +/** + * Pinned regression for the multi-line composer cursor-drift bug. + * + * Symptom: in `hermes --tui`, typing into the composer until the input + * wraps across multiple visual rows would leave several blank cells + * between the last typed character and the (hardware) cursor block. + * Worse on narrow terminals (the Cursor IDE built-in terminal in + * particular). + * + * Root cause: the composer's `cursorLayout` (used by `useDeclaredCursor` + * to place the hardware cursor) ran a hand-rolled word-wrap algorithm, + * while Ink's `` renders via `wrap-ansi`. The two + * disagreed on many real inputs — wrap-ansi would keep "branch + * investigate" on one row while cursorLayout claimed it had wrapped, + * etc. — so the declared cursor position drifted from where the text + * was actually rendered. The fix sources cursorLayout's line breaks + * directly from wrap-ansi, guaranteeing agreement. + * + * This test pins the contract: for every char that would be typed into + * the composer, the cursor position reported by cursorLayout MUST equal + * the end-of-text position that wrap-ansi would render. Any future + * regression that lets the two diverge re-introduces the drift. + */ +import { wrapAnsi } from '@hermes/ink' +import { describe, expect, it } from 'vitest' + +import { cursorLayout, inputVisualHeight } from '../lib/inputMetrics.js' + +function wrapAnsiEnd(text: string, cols: number): { line: number; column: number } { + const wrapped = wrapAnsi(text, cols, { hard: true, trim: false }) + const lines = wrapped.split('\n') + const last = lines[lines.length - 1] ?? '' + + return { line: lines.length - 1, column: last.length } +} + +const USER_REPORT_MESSAGE = + // Paraphrase of the user's actual bug report, included verbatim so the + // test is grounded in a realistic typing pattern (long single line, + // mixed-length words, punctuation, no hard newlines). + 'im in cursor terminal using hermes --tui and as i type multiline my caret at the end will often ' + + 'go.. randomly.. like multiple spaces away lol and idk why. theres no rhyme/reason really but ' + + 'there should literally never be a non-user added space at the end of my composer input right? ' + + 'i dont think it happens on new sessions but only existing ones. there have been a few prs to ' + + 'try to fix this and all not working. ok it just happened, to me, nowso attaching screenshot ' + + 'and you can see its multiline, new session. on a new bb/ branch investigate' + +describe('cursor-drift regression — composer cursorLayout matches Ink rendering', () => { + it('agrees with wrap-ansi at every typing-prefix of the user-reported message', () => { + // Walks the message char-by-char (mirroring what the TUI sees when a + // user types). At every prefix, cursorLayout must place the cursor + // exactly where wrap-ansi would render the end of the text. + // + // Pre-fix: this failed on most narrow widths because the hand-rolled + // wrap algorithm broke at slightly different points than wrap-ansi. + for (const cols of [40, 50, 55, 60, 65, 70, 80]) { + let acc = '' + + for (const ch of USER_REPORT_MESSAGE) { + acc += ch + const layout = cursorLayout(acc, acc.length, cols) + const expected = wrapAnsiEnd(acc, cols) + + expect( + layout, + `mismatch at cols=${cols}, len=${acc.length}, last-char=${JSON.stringify(ch)}, ` + + `tail=${JSON.stringify(acc.slice(-30))}` + ).toEqual(expected) + } + } + }) + + it('keeps cursor on the same row when text exactly fills the terminal width', () => { + // wrap-ansi does NOT push exact-fill text onto a phantom next line. + // The previous algorithm did — that's what produced the visible + // "cursor parked one row below the last char" symptom on narrow + // terminals at certain message lengths. + for (const cols of [8, 12, 18, 24]) { + const text = 'a'.repeat(cols) + const layout = cursorLayout(text, text.length, cols) + const inkLines = wrapAnsi(text, cols, { hard: true, trim: false }).split('\n') + + expect(layout.line).toBe(0) + expect(layout.column).toBe(cols) + expect(inkLines).toHaveLength(1) + expect(inputVisualHeight(text, cols)).toBe(1) + } + }) + + it('does not stuff a trailing whitespace word onto a phantom line', () => { + // "branch investigate" at cols=20 fits on one row in wrap-ansi. The + // bug claimed otherwise, parking the cursor at (line=1, col=?) and + // leaving the user's "branch investigate" rendered alone on row 0 + // with the cursor block several cells past it. + const text = 'branch investigate' + const cols = 20 + + expect(cursorLayout(text, text.length, cols)).toEqual({ column: text.length, line: 0 }) + expect(cursorLayout(text, text.length, cols)).toEqual(wrapAnsiEnd(text, cols)) + }) + + it('agrees with wrap-ansi for word-wrap that pushes a word onto the next line', () => { + // "hello world" at cols=8 wraps to ["hello ", "world"] in wrap-ansi. + // The cursor at end-of-text must land at line=1, col=5 — where Ink + // actually renders the last 'd'. The previous algorithm reported + // (line=2, col=0) here (phantom extra wrap), which parked the + // cursor on a row Ink never painted. + const text = 'hello world' + const cols = 8 + + expect(cursorLayout(text, text.length, cols)).toEqual({ column: 5, line: 1 }) + expect(cursorLayout(text, text.length, cols)).toEqual(wrapAnsiEnd(text, cols)) + }) +}) diff --git a/ui-tui/src/__tests__/textInputWrap.test.ts b/ui-tui/src/__tests__/textInputWrap.test.ts index c25c9629e77..22b33c9480e 100644 --- a/ui-tui/src/__tests__/textInputWrap.test.ts +++ b/ui-tui/src/__tests__/textInputWrap.test.ts @@ -1,8 +1,20 @@ +import { wrapAnsi } from '@hermes/ink' import { describe, expect, it } from 'vitest' import { offsetFromPosition } from '../components/textInput.js' import { composerPromptWidth, cursorLayout, inputVisualHeight, stableComposerColumns } from '../lib/inputMetrics.js' +// Helper: compute the "end of text" position that wrap-ansi would render +// the input to. This is what Ink's uses, so cursorLayout +// MUST agree. Disagreement is the cursor-drift bug. +function wrapAnsiEndPosition(text: string, cols: number): { line: number; column: number } { + const wrapped = wrapAnsi(text, cols, { hard: true, trim: false }) + const lines = wrapped.split('\n') + const last = lines[lines.length - 1] ?? '' + + return { line: lines.length - 1, column: last.length } +} + describe('cursorLayout — word-wrap parity with wrap-ansi', () => { it('places cursor mid-line at its column', () => { expect(cursorLayout('hello world', 6, 40)).toEqual({ column: 6, line: 0 }) @@ -12,19 +24,36 @@ describe('cursorLayout — word-wrap parity with wrap-ansi', () => { expect(cursorLayout('hi', 2, 10)).toEqual({ column: 2, line: 0 }) }) - it('wraps to next line when cursor lands exactly at the right edge', () => { - // 8 chars on an 8-col line: text fills the row exactly; the cursor's - // inverted-space cell overflows to col 0 of the next row. - expect(cursorLayout('abcdefgh', 8, 8)).toEqual({ column: 0, line: 1 }) + it('does not push exact-fill text onto a phantom next line', () => { + // Regression: the previous hand-rolled wrap algorithm forced the cursor + // onto (line+1, 0) when the text exactly filled the row. wrap-ansi keeps + // it on the same row (no soft-wrap), so the cursor must too — otherwise + // useDeclaredCursor parks the hardware cursor below the last char and + // the user sees several blank cells between text and cursor block + // (#cursor-drift-multiline). + expect(cursorLayout('abcdefgh', 8, 8)).toEqual({ column: 8, line: 0 }) + expect(cursorLayout('abcdefgh', 8, 8)).toEqual(wrapAnsiEndPosition('abcdefgh', 8)) + }) + + it('keeps short words on the current line when they fit (no phantom wrap)', () => { + // wrap-ansi: "hello wo" at cols=8 stays as one line "hello wo". + // The old cursorLayout incorrectly pushed to (1,0) because column=8 hit + // the column>=width check, but that disagreed with what Ink actually + // rendered. + expect(cursorLayout('hello wo', 8, 8)).toEqual({ column: 8, line: 0 }) + expect(cursorLayout('hello wo', 8, 8)).toEqual(wrapAnsiEndPosition('hello wo', 8)) }) it('moves words across wrap boundaries instead of splitting them', () => { - // With wordWrap:true, "hello wor" at cols=8 is "hello \nwor" rather - // than "hello wo\nr". - expect(cursorLayout('hello wo', 8, 8)).toEqual({ column: 0, line: 1 }) + // "hello wor" at cols=8: wrap-ansi breaks at the space, "hello \nwor". expect(cursorLayout('hello wor', 9, 8)).toEqual({ column: 3, line: 1 }) expect(cursorLayout('hello worl', 10, 8)).toEqual({ column: 4, line: 1 }) expect(cursorLayout('hello world', 11, 8)).toEqual({ column: 5, line: 1 }) + + // Each must match what wrap-ansi would actually render. + expect(cursorLayout('hello wor', 9, 8)).toEqual(wrapAnsiEndPosition('hello wor', 8)) + expect(cursorLayout('hello worl', 10, 8)).toEqual(wrapAnsiEndPosition('hello worl', 8)) + expect(cursorLayout('hello world', 11, 8)).toEqual(wrapAnsiEndPosition('hello world', 8)) }) it('wraps the next word instead of splitting it at the right edge', () => { @@ -42,12 +71,33 @@ describe('cursorLayout — word-wrap parity with wrap-ansi', () => { it('does not wrap when cursor is before the right edge', () => { expect(cursorLayout('abcdefg', 7, 8)).toEqual({ column: 7, line: 0 }) }) + + it('matches wrap-ansi end-position for typing-style incremental input', () => { + // Pins the actual fix: type a long message char-by-char at a narrow + // width and assert the cursor follows wrap-ansi every step of the way. + // Before the fix, ~5 boundary positions per pass disagreed and Ink + // parked the cursor several cells past the last rendered character. + const MSG = 'on a new bb branch investigate and fix the cursor drift bug here' + + for (const cols of [10, 14, 20, 30, 50, 80]) { + let acc = '' + + for (const ch of MSG) { + acc += ch + expect(cursorLayout(acc, acc.length, cols)).toEqual(wrapAnsiEndPosition(acc, cols)) + } + } + }) }) describe('input metrics helpers', () => { - it('computes visual height from the wrapped cursor line', () => { - expect(inputVisualHeight('abcdefgh', 8)).toBe(2) + it('computes visual height matching wrap-ansi line count', () => { + // Exact-fill text stays on one line in wrap-ansi (no phantom wrap), so + // visual height is 1. The previous implementation reported 2 here. + expect(inputVisualHeight('abcdefgh', 8)).toBe(1) expect(inputVisualHeight('one\ntwo', 40)).toBe(2) + // Multi-line wrap case sanity + expect(inputVisualHeight('hello world', 8)).toBe(2) }) it('counts the prompt gap as its own cell', () => { diff --git a/ui-tui/src/components/textInput.tsx b/ui-tui/src/components/textInput.tsx index ace2f479dc1..92082280a04 100644 --- a/ui-tui/src/components/textInput.tsx +++ b/ui-tui/src/components/textInput.tsx @@ -272,10 +272,22 @@ export function canFastBackspaceShape(current: string, cursor: number, columns?: } // If we know the wrap width, reject at the soft-wrap boundary: the - // caret's visual column is 0, so "\b \b" can't represent the physical - // move back to the previous visual line. - if (columns !== undefined && cursorLayout(current, cursor, columns).column === 0) { - return false + // caret's physical column would be at (or past) the terminal's right + // edge, so the terminal has already auto-wrapped to the next row. + // "\b \b" can't represent the physical move back across that wrap. + // + // We check `column === 0` for the "wrap-ansi broke onto a new line" + // case AND `column >= columns` for the "exact-fill, terminal auto-wraps" + // case. Both manifest as the same physical state (cursor parked at + // col 0 of the next row) but cursorLayout reports them differently + // because it now mirrors wrap-ansi's break points exactly (see the + // cursor-drift-multiline fix in lib/inputMetrics.ts). + if (columns !== undefined) { + const layout = cursorLayout(current, cursor, columns) + + if (layout.column === 0 || layout.column >= columns) { + return false + } } const removed = current.slice(prevPos(current, cursor), cursor) diff --git a/ui-tui/src/lib/inputMetrics.ts b/ui-tui/src/lib/inputMetrics.ts index b5645b43310..4c624da167a 100644 --- a/ui-tui/src/lib/inputMetrics.ts +++ b/ui-tui/src/lib/inputMetrics.ts @@ -1,4 +1,4 @@ -import { stringWidth } from '@hermes/ink' +import { stringWidth, wrapAnsi } from '@hermes/ink' import type { Role } from '../types.js' @@ -12,8 +12,6 @@ interface VisualLine { start: number } -const isWhitespace = (value: string) => /\s/.test(value) - const graphemes = (value: string) => [...seg().segment(value)].map(({ segment, index }) => ({ end: index + segment.length, @@ -22,76 +20,81 @@ const graphemes = (value: string) => width: Math.max(1, stringWidth(segment)) })) +// Build VisualLines from wrap-ansi's output by mapping each emitted character +// back to its original offset in `value`. wrap-ansi only INSERTS '\n' at wrap +// boundaries — it never drops, reorders, or substitutes existing characters — +// so a parallel walk uniquely identifies each line's source range. +// +// This used to be a hand-rolled word-wrap whose break points disagreed with +// wrap-ansi in subtle but visible ways: exact-fill rows pushed the cursor to +// a phantom next line, mid-word breaks landed one grapheme off, etc. The +// composer's TextInput renders text via Ink's , which +// delegates to wrap-ansi — so any drift between the two algorithms parks the +// hardware cursor several cells away from the last rendered character. +// Sourcing both from wrap-ansi guarantees agreement. function visualLines(value: string, cols: number): VisualLine[] { + if (!value.length) { + return [{ start: 0, end: 0 }] + } + const width = Math.max(1, cols) + const wrapped = wrapAnsi(value, width, { hard: true, trim: false }) const lines: VisualLine[] = [] - let sourceLineStart = 0 - for (const sourceLine of value.split('\n')) { - const parts = graphemes(sourceLine) + let originalIdx = 0 + let lineStart = 0 - if (!parts.length) { - lines.push({ start: sourceLineStart, end: sourceLineStart }) - sourceLineStart += 1 + for (let i = 0; i < wrapped.length; i += 1) { + const ch = wrapped[i]! + + if (ch === '\n') { + // wrap-ansi inserts '\n' to mark a soft-wrap boundary OR copies a + // literal '\n' from the input. Either way the next char in `wrapped` + // begins a new visual line. If the source character is a hard '\n', + // consume it (it doesn't appear in either line). Otherwise the '\n' + // is purely a wrap marker and originalIdx stays put. + lines.push({ start: lineStart, end: originalIdx }) + const isHardNewline = originalIdx < value.length && value[originalIdx] === '\n' + + if (isHardNewline) { + originalIdx += 1 + } + + lineStart = originalIdx continue } - let lineStartPart = 0 - let lineStartOffset = sourceLineStart - let column = 0 - let breakPart: null | number = null - let i = 0 - - while (i < parts.length) { - const part = parts[i]! - const partStart = sourceLineStart + part.index - - if (column + part.width > width && i > lineStartPart) { - if (breakPart !== null && breakPart > lineStartPart) { - const breakOffset = sourceLineStart + parts[breakPart - 1]!.end - lines.push({ start: lineStartOffset, end: breakOffset }) - lineStartPart = breakPart - lineStartOffset = breakOffset - } else { - lines.push({ start: lineStartOffset, end: partStart }) - lineStartPart = i - lineStartOffset = partStart - } - - column = 0 - breakPart = null - i = lineStartPart - continue - } - - column += part.width - - if (isWhitespace(part.segment)) { - breakPart = i + 1 - } - - i += 1 - - if (column >= width && i < parts.length) { - const next = parts[i]! - const nextStartsWord = !isWhitespace(next.segment) - - if (breakPart !== null && breakPart > lineStartPart && nextStartsWord) { - const breakOffset = sourceLineStart + parts[breakPart - 1]!.end - lines.push({ start: lineStartOffset, end: breakOffset }) - lineStartPart = breakPart - lineStartOffset = breakOffset - column = 0 - breakPart = null - i = lineStartPart - } - } + // Defensive sync check. wrap-ansi (with `hard: true, trim: false`, no + // styled input) is documented to only insert '\n' at break points and + // never substitute, drop, or reorder source characters — so under those + // options `wrapped[i]` should always equal `value[originalIdx]`. But + // future option changes, library upgrades, or callers that start passing + // styled input (ANSI escapes) could violate that invariant silently. If + // they do, we'd slide `originalIdx` past the end of `value` and emit + // garbage line ranges with no diagnostic. Realign by scanning forward + // for the matching character; bail out (return whatever we have) if the + // sync is unrecoverable rather than producing wrong-but-plausible output. + if (originalIdx >= value.length) { + break } - lines.push({ start: lineStartOffset, end: sourceLineStart + sourceLine.length }) - sourceLineStart += sourceLine.length + 1 + if (value[originalIdx] !== ch) { + const reSync = value.indexOf(ch, originalIdx) + + if (reSync === -1) { + break + } + + originalIdx = reSync + } + + originalIdx += 1 } + lines.push({ start: lineStart, end: originalIdx }) + + // wrap-ansi collapses an empty input into [""] which we already handled + // above; preserve the invariant that lines is never empty for any input. return lines.length ? lines : [{ start: 0, end: 0 }] } @@ -108,6 +111,12 @@ function widthBetween(value: string, start: number, end: number) { /** * Mirrors the word-wrap behavior used by the composer TextInput. * Returns the zero-based visual line and column of the cursor cell. + * + * IMPORTANT: this MUST stay in lock-step with how Ink's `` + * lays the value out (which uses `wrap-ansi`). Any divergence parks the + * hardware cursor several cells off the last rendered character — see the + * "cursor drift past blank cells" bug. `visualLines` is sourced directly + * from wrap-ansi to enforce that invariant. */ export function cursorLayout(value: string, cursor: number, cols: number) { const pos = Math.max(0, Math.min(cursor, value.length)) @@ -124,14 +133,14 @@ export function cursorLayout(value: string, cursor: number, cols: number) { } const line = lines[lineIndex]! - let column = widthBetween(value, line.start, Math.min(pos, line.end)) - - // trailing cursor-cell overflows to the next row at the wrap column - if (column >= w) { - lineIndex++ - column = 0 - } + const column = widthBetween(value, line.start, Math.min(pos, line.end)) + // NOTE: the previous implementation forced an extra line break when + // `column >= w` (the "trailing cursor-cell overflows" rule). With + // `visualLines` sourcing breaks from wrap-ansi, the line wrapping + // above already matches what Ink will actually render. Pushing the + // cursor onto a phantom next line here would re-introduce the same + // drift we're fixing, so we don't. return { column, line: lineIndex } } diff --git a/website/docs/reference/cli-commands.md b/website/docs/reference/cli-commands.md index 4cfc80191f1..37e52707cae 100644 --- a/website/docs/reference/cli-commands.md +++ b/website/docs/reference/cli-commands.md @@ -411,6 +411,7 @@ Multi-profile, multi-project collaboration board. Each install can host many boa | `dispatch` | One dispatcher pass on the active board. Flags: `--dry-run`, `--max N`, `--json`. | | `context ` | Print the full context a worker would see (title + body + parent results + comments). | | `specify ` / `specify --all` | Flesh out a triage-column task into a concrete spec (title + body with goal, approach, acceptance criteria) via the auxiliary LLM, then promote it to `todo`. Flags: `--tenant` (scope `--all` to one tenant), `--author`, `--json`. Configure the model under `auxiliary.triage_specifier` in `config.yaml`. | +| `decompose ` / `decompose --all` | Fan a triage-column task out into a graph of child tasks routed to specialist profiles by description (the orchestrator-driven path). Falls back to specify-style single-task promotion when the LLM decides the task doesn't benefit from fan-out. Same flags as `specify`. Configure the model under `auxiliary.kanban_decomposer` in `config.yaml`. Also runs automatically every dispatcher tick when `kanban.auto_decompose: true` (the default). See [Auto vs Manual orchestration](/docs/user-guide/features/kanban#auto-vs-manual-orchestration). | | `gc` | Remove scratch workspaces for archived tasks. | Examples: diff --git a/website/docs/reference/profile-commands.md b/website/docs/reference/profile-commands.md index 376394a637e..467134b6d05 100644 --- a/website/docs/reference/profile-commands.md +++ b/website/docs/reference/profile-commands.md @@ -83,6 +83,7 @@ Creates a new profile. | `--clone-all` | Copy everything (config, memories, skills, sessions, state) from the current profile. | | `--clone-from ` | Clone from a specific profile instead of the current one. Used with `--clone` or `--clone-all`. | | `--no-alias` | Skip wrapper script creation. | +| `--description ""` | One- or two-sentence description of what this profile is good at. Used by the kanban orchestrator to route tasks based on role instead of profile name alone. Skip and add later via `hermes profile describe`. Persisted in `/profile.yaml`. | Creating a profile does **not** make that profile directory the default project/workspace directory for terminal commands. If you want a profile to start in a specific project, set `terminal.cwd` in that profile's `config.yaml`. @@ -102,6 +103,40 @@ hermes profile create backup --clone-all hermes profile create work2 --clone --clone-from work ``` +## `hermes profile describe` + +```bash +hermes profile describe [] [options] +``` + +Read or set a profile's description. The description is consumed by the kanban orchestrator to route tasks based on what each profile is good at, rather than guessing from the profile name alone. Persisted in `/profile.yaml` so it survives reboots and is shared with the gateway. + +With no flags, prints the current description (or `(no description set for '')` if empty). + +| Argument / Option | Description | +|-------------------|-------------| +| `` | Profile to describe. Required unless `--all --auto` is used. | +| `--text ""` | Set the description to this exact text (user-authored). Overwrites any existing description. | +| `--auto` | Auto-generate a 1-2 sentence description via the auxiliary LLM, based on the profile's installed skills, configured model, and name. Configure the model under `auxiliary.profile_describer` in `config.yaml`. Auto-generated descriptions are marked `description_auto: true` so the dashboard can flag them for review. | +| `--overwrite` | With `--auto`, replace user-authored descriptions too (default: skip profiles whose description was set explicitly). | +| `--all` | With `--auto`, sweep every profile missing a description. | + +**Examples:** + +```bash +# Read the current description +hermes profile describe researcher + +# Set it explicitly +hermes profile describe researcher --text "Reads source code and writes findings." + +# Let the LLM generate one +hermes profile describe researcher --auto + +# Fill in descriptions for every profile that doesn't have one +hermes profile describe --all --auto +``` + ## `hermes profile delete` ```bash diff --git a/website/docs/user-guide/configuration.md b/website/docs/user-guide/configuration.md index 5ac0d8c9df2..d972b38b384 100644 --- a/website/docs/user-guide/configuration.md +++ b/website/docs/user-guide/configuration.md @@ -785,6 +785,8 @@ $ hermes model [ ] compression currently: auto / main model [ ] approval currently: auto / main model [ ] triage_specifier currently: auto / main model +[ ] kanban_decomposer currently: auto / main model +[ ] profile_describer currently: auto / main model ``` Select a task, pick a provider (OAuth flows open a browser; API-key providers prompt), pick a model. The change persists to `auxiliary..*` in `config.yaml`. Same machinery as the main-model picker — no extra syntax to learn. diff --git a/website/docs/user-guide/features/codex-app-server-runtime.md b/website/docs/user-guide/features/codex-app-server-runtime.md index a1aa6a0776e..575250d9b01 100644 --- a/website/docs/user-guide/features/codex-app-server-runtime.md +++ b/website/docs/user-guide/features/codex-app-server-runtime.md @@ -91,11 +91,11 @@ What works inside a codex-runtime worker: - The Hermes tool callback for browser_*, vision, image_gen, skills, TTS What also works because the MCP callback exposes them: -- **`kanban_complete` / `kanban_block` / `kanban_comment` / `kanban_heartbeat`** — the worker handoff tools. These read `HERMES_KANBAN_TASK` from env (set by the dispatcher), gate access correctly, and write to `~/.hermes/kanban.db`. Without these in the callback, a worker on this runtime could do its task but couldn't report back, hanging until the dispatcher's timeout. +- **`kanban_complete` / `kanban_block` / `kanban_comment` / `kanban_heartbeat`** — the worker handoff tools. These read `HERMES_KANBAN_TASK` from env (set by the dispatcher), gate access correctly, and write to the per-board SQLite DB pinned by `HERMES_KANBAN_DB`. Without these in the callback, a worker on this runtime could do its task but couldn't report back, hanging until the dispatcher's timeout. - **`kanban_show` / `kanban_list`** — read-only board queries for the worker to check its own context. - **`kanban_create` / `kanban_unblock` / `kanban_link`** — orchestrator-only operations. Available for orchestrator agents running on the codex runtime that need to dispatch new tasks. -The kanban tools are gated by `HERMES_KANBAN_TASK` env var the dispatcher sets — that var is propagated to the codex subprocess (codex inherits env) and from there to the spawned `hermes-tools` MCP server subprocess. So the tools see the right task id and gate correctly. +The kanban tools are gated by `HERMES_KANBAN_TASK` env var the dispatcher sets — that var is propagated to the codex subprocess (codex inherits env) and from there to the spawned `hermes-tools` MCP server subprocess. So the tools see the right task id and gate correctly. For Codex app-server workers, Hermes also passes narrow app-server sandbox overrides when `HERMES_KANBAN_TASK` is present: keep `workspace-write` sandboxing, add only the current board directory (derived from `HERMES_KANBAN_DB`) as an extra writable root, and keep network disabled by default. This avoids the brittle `:danger-no-sandbox` workaround while letting `kanban_complete` / `kanban_block` update the board DB. ### Cron jobs diff --git a/website/docs/user-guide/features/fallback-providers.md b/website/docs/user-guide/features/fallback-providers.md index 72528796d55..b17102cb82e 100644 --- a/website/docs/user-guide/features/fallback-providers.md +++ b/website/docs/user-guide/features/fallback-providers.md @@ -320,6 +320,55 @@ auxiliary: --- +## Auxiliary Capacity-Error Fallback + +When you set an explicit auxiliary provider (e.g. `auxiliary.vision.provider: glm`), Hermes treats that as your preferred choice — but if the provider literally cannot serve the request because of a **capacity error** (HTTP 402 payment required, HTTP 429 daily-quota exhaustion, connection failure), Hermes falls back through a layered chain instead of failing silently: + +1. **Primary aux provider** — the one you configured (tried first, always) +2. **`auxiliary..fallback_chain`** — your per-task override list, if you wrote one +3. **Main agent provider + model** — last-resort safety net (always tried, even if you didn't write a chain) +4. **Warn + re-raise** — if every layer fails, Hermes logs `Auxiliary : ... all fallbacks exhausted` at WARNING level and re-raises the original error + +Transient HTTP 429 rate limits (`Retry-After: ...`) are treated as request constraints, not capacity problems — they respect your explicit provider choice and do **not** trigger the fallback ladder. Only daily/monthly quota exhaustion, payment errors, and connection failures bypass the explicit-provider gate. + +For users on `provider: auto` (no explicit aux provider), the existing auto-detection chain runs in place of steps 2–3. Its first step is already the main agent model, so `auto` users get the same outcome with zero config. + +### Optional: per-task fallback chain + +If you want a different fallback ordering than "main agent model first", configure `fallback_chain` explicitly. Each entry needs at least `provider`; `model`, `base_url`, and `api_key` are optional. + +```yaml +auxiliary: + vision: + provider: glm + model: glm-4v-flash + fallback_chain: + - provider: openrouter + model: google/gemini-3-flash-preview + - provider: nous + model: anthropic/claude-sonnet-4 + + compression: + provider: openrouter + fallback_chain: + - provider: openai + model: gpt-4o-mini +``` + +You do **not** need to configure `fallback_chain` to get fallback — the main-agent safety net runs regardless. Use it only when you specifically want a different order than the default. + +### Provider quota errors that trigger fallback + +Hermes recognizes these as capacity-equivalent to 402 credit exhaustion (not transient rate limits): + +- Bedrock / LiteLLM: `Too many tokens per day`, `daily limit`, `tokens per day` +- Vertex AI / GCP: `quota exceeded`, `resource exhausted`, `RESOURCE_EXHAUSTED` +- Generic: `daily quota`, `quota_exceeded` + +If your provider returns a different phrase for daily-quota exhaustion and Hermes doesn't trigger fallback, that's a bug — open an issue with the exact error string. + +--- + ## Context Compression Fallback Context compression uses the `auxiliary.compression` config block to control which model and provider handles summarization: @@ -378,14 +427,16 @@ See [Scheduled Tasks (Cron)](/docs/user-guide/features/cron) for full configurat | Feature | Fallback Mechanism | Config Location | |---------|-------------------|----------------| | Main agent model | `fallback_model` in config.yaml — per-turn failover on errors (primary restored each turn) | `fallback_model:` (top-level) | -| Vision | Auto-detection chain + internal OpenRouter retry | `auxiliary.vision` | -| Web extraction | Auto-detection chain + internal OpenRouter retry | `auxiliary.web_extract` | -| Context compression | Auto-detection chain, degrades to no-summary if unavailable | `auxiliary.compression` | -| Session search | Auto-detection chain | `auxiliary.session_search` | -| Skills hub | Auto-detection chain | `auxiliary.skills_hub` | -| MCP helpers | Auto-detection chain | `auxiliary.mcp` | -| Approval classification | Auto-detection chain | `auxiliary.approval` | -| Title generation | Auto-detection chain | `auxiliary.title_generation` | -| Triage specifier | Auto-detection chain | `auxiliary.triage_specifier` | +| Auxiliary tasks (any) — auto users | Full auto-detection chain (main agent model first, then provider chain) on capacity errors | `auxiliary..provider: auto` | +| Auxiliary tasks (any) — explicit provider | `fallback_chain` (if set) → main agent model → warn + raise, on capacity errors only | `auxiliary..fallback_chain` | +| Vision | Layered (see above) + internal OpenRouter retry | `auxiliary.vision` | +| Web extraction | Layered (see above) + internal OpenRouter retry | `auxiliary.web_extract` | +| Context compression | Layered (see above); degrades to no-summary if all layers unavailable | `auxiliary.compression` | +| Session search | Layered (see above) | `auxiliary.session_search` | +| Skills hub | Layered (see above) | `auxiliary.skills_hub` | +| MCP helpers | Layered (see above) | `auxiliary.mcp` | +| Approval classification | Layered (see above) | `auxiliary.approval` | +| Title generation | Layered (see above) | `auxiliary.title_generation` | +| Triage specifier | Layered (see above) | `auxiliary.triage_specifier` | | Delegation | Provider override only (no automatic fallback) | `delegation.provider` / `delegation.model` | | Cron jobs | Per-job provider override only (no automatic fallback) | Per-job `provider` / `model` | diff --git a/website/docs/user-guide/features/kanban-tutorial.md b/website/docs/user-guide/features/kanban-tutorial.md index 5f79569c7bc..88a0f9cf5ec 100644 --- a/website/docs/user-guide/features/kanban-tutorial.md +++ b/website/docs/user-guide/features/kanban-tutorial.md @@ -22,7 +22,7 @@ Throughout the tutorial, **code blocks labelled `bash` are commands *you* run.** Six columns, left to right: -- **Triage** — raw ideas, a specifier will flesh out the spec before anyone works on them. Click the **✨ Specify** button on any triage card (or run `hermes kanban specify ` / `/kanban specify ` from a chat) to have the auxiliary LLM turn a one-liner into a full spec (goal, approach, acceptance criteria) and promote it to `todo` in one shot. Configure which model runs it under `auxiliary.triage_specifier` in `config.yaml`. +- **Triage** — raw ideas. By default the dispatcher auto-runs the **decomposer** (orchestrator-driven fan-out) on tasks here: it reads your profile roster + descriptions and produces a graph of child tasks routed to the best-fit specialists, with the original task held alive as the parent so the orchestrator wakes back up to judge completion when everything finishes. Flip the **Orchestration: Auto/Manual** pill at the top of the kanban page to switch modes. In Manual mode (or for setups without an orchestrator profile) click **⚗ Decompose** on a card, or run `hermes kanban decompose ` / `/kanban decompose `. For single tasks that don't need fan-out, **✨ Specify** does a one-shot spec rewrite (goal, approach, acceptance criteria) and promotes to `todo`. Configure the models under `auxiliary.kanban_decomposer` and `auxiliary.triage_specifier` in `config.yaml`. See [Auto vs Manual orchestration](./kanban#auto-vs-manual-orchestration) in the main Kanban guide. - **Todo** — created but waiting on dependencies, or not yet assigned. - **Ready** — assigned and waiting for the dispatcher to claim. - **In progress** — a worker is actively running the task. With "Lanes by profile" on (the default), this column sub-groups by assignee so you can see at a glance what each worker is doing. diff --git a/website/docs/user-guide/features/kanban.md b/website/docs/user-guide/features/kanban.md index 91c6dacde67..7328fc4b615 100644 --- a/website/docs/user-guide/features/kanban.md +++ b/website/docs/user-guide/features/kanban.md @@ -444,7 +444,7 @@ hermes dashboard # "Kanban" tab appears in the nav, after "Skills" ### What the plugin gives you - A **Kanban** tab showing one column per status: `triage`, `todo`, `ready`, `running`, `blocked`, `done` (plus `archived` when the toggle is on). - - `triage` is the parking column for rough ideas a specifier is expected to flesh out. Tasks created with `hermes kanban create --triage` (or via the Triage column's inline create) land here and the dispatcher leaves them alone until a human or specifier promotes them to `todo` / `ready`. Run `hermes kanban specify ` to have the auxiliary LLM expand a triage task into a concrete spec (title + body with goal, approach, acceptance criteria) and promote it to `todo` in one shot; `--all` sweeps every triage task at once. Configure which model runs the specifier under `auxiliary.triage_specifier` in `config.yaml`. + - `triage` is the parking column for rough ideas. By default (`kanban.auto_decompose: true`), the dispatcher auto-runs the **decomposer** on tasks that land here — the orchestrator profile reads the rough idea, looks at your profile roster (with descriptions), and fans the task out into a small graph of child tasks routed to the best-fit specialists. The original task stays alive as the parent of every child so the orchestrator wakes back up to judge completion when everything finishes. Flip the **Orchestration: Auto/Manual** pill at the top of the page (or set `kanban.auto_decompose: false`) to switch to manual mode, where triage tasks stay put until you click **⚗ Decompose** on a card or run `hermes kanban decompose `. For tasks that don't need fan-out (or for setups without an orchestrator profile), the **✨ Specify** button does a single-task spec rewrite (title + body with goal, approach, acceptance criteria) via the same LLM machinery. See [Auto vs Manual orchestration](#auto-vs-manual-orchestration) below. - Cards show the task id, title, priority badge, tenant tag, assigned profile, comment/link counts, a **progress pill** (`N/M` children done when the task has dependents), and "created N ago". A per-card checkbox enables multi-select. - **Per-profile lanes inside Running** — toolbar checkbox toggles sub-grouping of the Running column by assignee. - **Live updates via WebSocket** — the plugin tails the append-only `task_events` table on a short poll interval; the board reflects changes the instant any profile (CLI, gateway, or another dashboard tab) acts. Reloads are debounced so a burst of events triggers a single refetch. @@ -456,12 +456,40 @@ hermes dashboard # "Kanban" tab appears in the nav, after "Skills" - **Editable assignee / priority** — click the meta row to rewrite. - **Editable description** — markdown-rendered by default (headings, bold, italic, inline code, fenced code, `http(s)` / `mailto:` links, bullet lists), with an "edit" button that swaps in a textarea. Markdown rendering is a tiny, XSS-safe renderer — every substitution runs on HTML-escaped input, only `http(s)` / `mailto:` links pass through, and `target="_blank"` + `rel="noopener noreferrer"` are always set. - **Dependency editor** — chip list of parents and children, each with an `×` to unlink, plus dropdowns over every other task to add a new parent or child. Cycle attempts are rejected server-side with a clear message. - - **Status action row** (→ triage / → ready / → running / block / unblock / complete / archive) with confirm prompts for destructive transitions. For cards in the **Triage** column the row also exposes a **✨ Specify** button that calls the auxiliary LLM (`auxiliary.triage_specifier` in `config.yaml`) to expand the one-liner into a concrete spec (title + body with goal, approach, acceptance criteria) and promote the task to `todo`. The same behaviour is reachable from the CLI (`hermes kanban specify ` / `--all`), from any gateway platform (`/kanban specify `), and programmatically via `POST /api/plugins/kanban/tasks/:id/specify`. + - **Status action row** (→ triage / → ready / → running / block / unblock / complete / archive) with confirm prompts for destructive transitions. For cards in the **Triage** column the row also exposes two LLM-driven actions: **⚗ Decompose** fans the task out into a graph of child tasks routed to specialist profiles by description (the orchestrator-driven path), and **✨ Specify** does a single-task spec rewrite. Decompose falls back to specify-style promotion when the LLM decides the task doesn't benefit from fan-out, so it's a strict superset. Both are reachable from the CLI (`hermes kanban decompose ` / `specify ` / `--all`), from any gateway platform (`/kanban decompose `), and programmatically via `POST /api/plugins/kanban/tasks/:id/decompose` and `…/specify`. Configure the models under `auxiliary.kanban_decomposer` and `auxiliary.triage_specifier` in `config.yaml`. - Result section (also markdown-rendered), comment thread with Enter-to-submit, the last 20 events. - **Toolbar filters** — free-text search, tenant dropdown (defaults to `dashboard.kanban.default_tenant` from `config.yaml`), assignee dropdown, "show archived" toggle, "lanes by profile" toggle, and a **Nudge dispatcher** button so you don't have to wait for the next 60 s tick. Visually the target is the familiar Linear / Fusion layout: dark theme, column headers with counts, coloured status dots, pill chips for priority and tenant. The plugin reads only theme CSS vars (`--color-*`, `--radius`, `--font-mono`, ...), so it reskins automatically with whichever dashboard theme is active. +### Auto vs Manual orchestration + +The kanban board has two ways to handle a task you drop into the Triage column: + +**Auto (default)** — `kanban.auto_decompose: true`. The gateway-embedded dispatcher runs the **decomposer** on each tick, capped by `kanban.auto_decompose_per_tick` (default 3 tasks per tick) so a bulk-load of triage tasks doesn't burst-spend the auxiliary LLM. The decomposer reads the rough idea, looks at your installed profiles + their descriptions, and asks the LLM to produce a JSON task graph: which tasks to spawn, who they go to, and which depend on which. The original triage task becomes the parent of every leaf in the graph, so it stays alive until the whole graph completes — and then promotes back to `ready` so its assignee (the orchestrator profile) can judge completion and add more tasks if the work isn't done. This is the "drop a one-liner, walk away" flow. + +**Manual** — `kanban.auto_decompose: false`. Triage tasks stay in triage until you act. Click the **⚗ Decompose** button on a card, run `hermes kanban decompose ` (or `--all`), or use `/kanban decompose ` from a chat. This matches the pre-decomposer behavior of the board, useful when you want full control over what runs when. + +Flip between the two modes from the **Orchestration: Auto/Manual** pill at the top of the kanban page (emerald = Auto, muted gray = Manual), or by editing `config.yaml` directly. Both modes coexist with `hermes kanban specify` — that's still available as a single-task spec rewrite when you don't want fan-out. + +The decomposer's routing decisions depend on profile descriptions, which is a per-profile labeling primitive you set with `hermes profile create --description "..."`, `hermes profile describe --text "..."`, `hermes profile describe --auto` (LLM-generates from the profile's installed skills + model), or the dashboard's per-profile editor in the expanded **Orchestration settings** panel. Profiles without a description still appear in the roster — they're routable by name, just less precisely. The decomposer NEVER lands a child task with `assignee=None`: when the LLM picks an unknown profile, the child gets routed to `kanban.default_assignee` (or the active default profile if that's unset). + +Config knobs (all under `kanban:` in `~/.hermes/config.yaml`): + +| Key | Default | Purpose | +|---|---|---| +| `auto_decompose` | `true` | Dispatcher auto-runs the decomposer every tick. | +| `auto_decompose_per_tick` | `3` | Cap on decompositions per dispatcher tick. Excess defers to the next tick. | +| `orchestrator_profile` | `""` | Profile that owns decomposition. Empty = fall back to active default profile. | +| `default_assignee` | `""` | Where a child task lands when the LLM picks an unknown profile. Empty = fall back to active default. | + +And the two auxiliary LLM slots: + +| Key | Purpose | +|---|---| +| `auxiliary.kanban_decomposer` | Model that produces the task graph (called by Decompose). Set `provider`/`model` to override the main chat model. | +| `auxiliary.profile_describer` | Model that auto-generates profile descriptions (called by `hermes profile describe --auto`). | + ### Architecture The GUI is strictly a **read-through-the-DB + write-through-kanban_db** layer with no domain logic of its own: @@ -499,6 +527,12 @@ All routes are mounted under `/api/plugins/kanban/` and protected by the dashboa | `POST` | `/tasks/bulk` | Apply the same patch (status / archive / assignee / priority) to every id in `ids`. Per-id failures reported without aborting siblings | | `POST` | `/tasks/:id/comments` | Append a comment | | `POST` | `/tasks/:id/specify` | Run the triage specifier — auxiliary LLM fleshes out the task body and promotes it from `triage` to `todo`. Returns `{ok, task_id, reason, new_title}`; `ok=false` with a human-readable reason on "not in triage" / no aux client / LLM error is a 200, not a 4xx | +| `POST` | `/tasks/:id/decompose` | Run the kanban decomposer — auxiliary LLM produces a task graph and the helper atomically creates the children + links the root + flips `triage → todo`. Returns `{ok, task_id, reason, fanout, child_ids, new_title}`. Same 200-on-LLM-error convention as `/specify`. | +| `GET` | `/profiles` | List installed profiles with their descriptions (consumed by the dashboard's profile-description editor and the orchestrator picker). | +| `PATCH` | `/profiles/:name` | Set or clear a profile's description (user-authored — `description_auto: false`). Returns `{ok, profile, description}`. | +| `POST` | `/profiles/:name/describe-auto` | Generate a description for a profile via `auxiliary.profile_describer`. Persists with `description_auto: true` so the dashboard can surface a "review" badge. | +| `GET` | `/orchestration` | Read the kanban orchestration settings (`orchestrator_profile`, `default_assignee`, `auto_decompose`) plus the *resolved* effective values after fallbacks. | +| `PUT` | `/orchestration` | Update one or more of the three orchestration keys in `config.yaml`. Validates that non-empty profile names actually exist. | | `POST` | `/links` | Add a dependency (`parent_id` → `child_id`) | | `DELETE` | `/links?parent_id=…&child_id=…` | Remove a dependency | | `POST` | `/dispatch?max=…&dry_run=…` | Nudge the dispatcher — skip the 60 s wait | diff --git a/website/docs/user-guide/messaging/index.md b/website/docs/user-guide/messaging/index.md index acd12872812..ef02bc7fe16 100644 --- a/website/docs/user-guide/messaging/index.md +++ b/website/docs/user-guide/messaging/index.md @@ -222,9 +222,22 @@ hermes pairing revoke telegram 123456789 # Remove access Pairing codes expire after 1 hour, are rate-limited, and use cryptographic randomness. -### Slash Command Access Control +### Admins vs Regular Users -Once users are allowed in, you can split them into **admins** (full slash command access) and **regular users** (only the slash commands you explicitly enable). This applies per platform and per scope (DM vs group/channel) and works through the live command registry, so it covers built-in AND plugin-registered slash commands without per-feature wiring. +Allowlists answer "can this person reach the bot at all?" The **admin / user split** answers "now that they're in, what are they allowed to do?" + +Every allowed user falls into one of two tiers per scope (DM vs group/channel): + +- **Admin** — full access. Can run every registered slash command (built-in + plugin) and use every gated capability. +- **Regular user** — restricted access. Can chat with the agent normally, but can only run the slash commands you explicitly enable. The always-allowed floor is `/help` and `/whoami`. + +The tiers are configured per platform and per scope. DM admin status does not imply group/channel admin status — each scope has its own admin list. + +**What the tiers gate today:** slash commands. The split runs through the live command registry, so it covers built-ins and plugin-registered commands without per-feature wiring. Plain chat is not affected — non-admins can still talk to the agent. + +**What may be gated in the future:** more capability surfaces (tool access, model switching, expensive operations) will hang off the same admin / user distinction as we add them. Configuring the split now means those future restrictions land cleanly without you having to re-model who's an admin. + +#### Configuration ```yaml gateway: @@ -239,13 +252,9 @@ gateway: group_user_allowed_commands: [status] ``` -Behavior: +**Backward compat:** if `allow_admin_from` is not set for a scope, the tier split is disabled for that scope and every allowed user has full access. Existing installs keep working with no changes — opt in when you want the distinction. -- A user in `allow_admin_from` for a scope can run **every** registered slash command. -- A user in `allow_from` but not in `allow_admin_from` can only run commands in `user_allowed_commands`, plus the always-allowed floor: `/help` and `/whoami`. -- Plain chat is unaffected. Non-admins can still talk to the agent normally; they just can't trigger arbitrary commands. -- **Backward compat:** if `allow_admin_from` is not set for a scope, slash gating is disabled for that scope. Existing installs keep working with no changes. -- DM admin status does not imply group/channel admin status. Each scope has its own admin list. +#### Inspecting your access Use `/whoami` from any platform to see the active scope, your tier (admin / user / unrestricted), and which slash commands you can run. See the [Telegram](/docs/user-guide/messaging/telegram#slash-command-access-control) and [Discord](/docs/user-guide/messaging/discord#slash-command-access-control) pages for platform-specific examples. diff --git a/website/docs/user-guide/messaging/matrix.md b/website/docs/user-guide/messaging/matrix.md index 255806c01ba..b03f7a655d4 100644 --- a/website/docs/user-guide/messaging/matrix.md +++ b/website/docs/user-guide/messaging/matrix.md @@ -357,6 +357,23 @@ To find a Room ID: in Element, go to the room → **Settings** → **Advanced** **Fix**: Invite the bot to the room — it auto-joins on invite. Verify your User ID is in `MATRIX_ALLOWED_USERS` (use the full `@user:server` format). Restart the gateway. +### Bot joins rooms but silently drops every message (clock skew) + +**Cause**: The host's system clock is set ahead of real time. The Matrix adapter applies a 5-second startup-grace filter (`event_ts < startup_ts - 5`) to ignore events replayed from initial sync. When the wall clock is ahead, every incoming event looks "older than startup" and is dropped before reaching the message handler — the bot appears connected but never replies. See [#12614](https://github.com/NousResearch/hermes-agent/issues/12614). + +**Symptom**: Gateway log shows `Matrix: dropped N live events as 'too old' more than 30s after startup`. + +**Fix**: Sync the host clock with NTP and restart the bot: + +```bash +# Debian/Ubuntu +sudo timedatectl set-ntp true +timedatectl status # confirm "System clock synchronized: yes" + +# macOS +sudo sntp -sS time.apple.com +``` + ### "Failed to authenticate" / "whoami failed" on startup **Cause**: The access token or homeserver URL is incorrect. diff --git a/website/docs/user-guide/profiles.md b/website/docs/user-guide/profiles.md index 522b24cb770..73ea0a8cadd 100644 --- a/website/docs/user-guide/profiles.md +++ b/website/docs/user-guide/profiles.md @@ -32,6 +32,14 @@ hermes profile create mybot Creates a fresh profile with bundled skills seeded. Run `mybot setup` to configure API keys, model, and gateway tokens. +If you plan to use this profile as a kanban worker (or want the kanban orchestrator to route work to it), pass `--description ""` at create time so the orchestrator knows what it's good at: + +```bash +hermes profile create researcher --description "Reads source code and external docs, writes findings." +``` + +You can also set or auto-generate the description later with `hermes profile describe` — see the [Kanban guide](./features/kanban#auto-vs-manual-orchestration) for the full routing model. + ### Clone config only (`--clone`) ```bash