mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-04-25 00:51:20 +00:00
Merge branch 'main' of github.com:NousResearch/hermes-agent into feat/ink-refactor
This commit is contained in:
Brooklyn Nicholson
commit
f81dba0da2
128 changed files with 8357 additions and 842 deletions
246
run_agent.py
246
run_agent.py
|
|
@ -75,7 +75,7 @@ from tools.browser_tool import cleanup_browser
|
|||
from hermes_constants import OPENROUTER_BASE_URL
|
||||
|
||||
# Agent internals extracted to agent/ package for modularity
|
||||
from agent.memory_manager import build_memory_context_block
|
||||
from agent.memory_manager import build_memory_context_block, sanitize_context
|
||||
from agent.retry_utils import jittered_backoff
|
||||
from agent.error_classifier import classify_api_error, FailoverReason
|
||||
from agent.prompt_builder import (
|
||||
|
|
@ -602,6 +602,7 @@ class AIAgent:
|
|||
prefill_messages: List[Dict[str, Any]] = None,
|
||||
platform: str = None,
|
||||
user_id: str = None,
|
||||
gateway_session_key: str = None,
|
||||
skip_context_files: bool = False,
|
||||
skip_memory: bool = False,
|
||||
session_db=None,
|
||||
|
|
@ -667,6 +668,7 @@ class AIAgent:
|
|||
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._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
|
||||
|
|
@ -689,9 +691,14 @@ class AIAgent:
|
|||
self.api_mode = api_mode
|
||||
elif self.provider == "openai-codex":
|
||||
self.api_mode = "codex_responses"
|
||||
elif self.provider == "xai":
|
||||
self.api_mode = "codex_responses"
|
||||
elif (provider_name is None) and "chatgpt.com/backend-api/codex" in self._base_url_lower:
|
||||
self.api_mode = "codex_responses"
|
||||
self.provider = "openai-codex"
|
||||
elif (provider_name is None) and "api.x.ai" in self._base_url_lower:
|
||||
self.api_mode = "codex_responses"
|
||||
self.provider = "xai"
|
||||
elif self.provider == "anthropic" or (provider_name is None and "api.anthropic.com" in self._base_url_lower):
|
||||
self.api_mode = "anthropic_messages"
|
||||
self.provider = "anthropic"
|
||||
|
|
@ -1019,16 +1026,12 @@ class AIAgent:
|
|||
f"was found. Set the {_env_hint} environment "
|
||||
f"variable, or switch to a different provider with `hermes model`."
|
||||
)
|
||||
# Final fallback: try raw OpenRouter key
|
||||
client_kwargs = {
|
||||
"api_key": os.getenv("OPENROUTER_API_KEY", ""),
|
||||
"base_url": OPENROUTER_BASE_URL,
|
||||
"default_headers": {
|
||||
"HTTP-Referer": "https://hermes-agent.nousresearch.com",
|
||||
"X-OpenRouter-Title": "Hermes Agent",
|
||||
"X-OpenRouter-Categories": "productivity,cli-agent",
|
||||
},
|
||||
}
|
||||
# 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
|
||||
|
||||
|
|
@ -1292,6 +1295,9 @@ class AIAgent:
|
|||
# Thread gateway user identity for per-user memory scoping
|
||||
if self._user_id:
|
||||
_init_kwargs["user_id"] = self._user_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
|
||||
|
|
@ -2102,6 +2108,59 @@ class AIAgent:
|
|||
content = re.sub(r'</?(?:think|thinking|reasoning|thought|REASONING_SCRATCHPAD)>\s*', '', content, flags=re.IGNORECASE)
|
||||
return content
|
||||
|
||||
@staticmethod
|
||||
def _has_natural_response_ending(content: str) -> bool:
|
||||
"""Heuristic: does visible assistant text look intentionally finished?"""
|
||||
if not content:
|
||||
return False
|
||||
stripped = content.rstrip()
|
||||
if not stripped:
|
||||
return False
|
||||
if stripped.endswith("```"):
|
||||
return True
|
||||
return stripped[-1] in '.!?:)"\']}。！？：）】」』》'
|
||||
|
||||
def _is_ollama_glm_backend(self) -> bool:
|
||||
"""Detect the narrow backend family affected by Ollama/GLM stop misreports."""
|
||||
model_lower = (self.model or "").lower()
|
||||
provider_lower = (self.provider or "").lower()
|
||||
if "glm" not in model_lower and provider_lower != "zai":
|
||||
return False
|
||||
if "ollama" in self._base_url_lower or ":11434" in self._base_url_lower:
|
||||
return True
|
||||
return bool(self.base_url and is_local_endpoint(self.base_url))
|
||||
|
||||
def _should_treat_stop_as_truncated(
|
||||
self,
|
||||
finish_reason: str,
|
||||
assistant_message,
|
||||
messages: Optional[list] = None,
|
||||
) -> bool:
|
||||
"""Detect conservative stop->length misreports for Ollama-hosted GLM models."""
|
||||
if finish_reason != "stop" or self.api_mode != "chat_completions":
|
||||
return False
|
||||
if not self._is_ollama_glm_backend():
|
||||
return False
|
||||
if not any(
|
||||
isinstance(msg, dict) and msg.get("role") == "tool"
|
||||
for msg in (messages or [])
|
||||
):
|
||||
return False
|
||||
if assistant_message is None or getattr(assistant_message, "tool_calls", None):
|
||||
return False
|
||||
|
||||
content = getattr(assistant_message, "content", None)
|
||||
if not isinstance(content, str):
|
||||
return False
|
||||
|
||||
visible_text = self._strip_think_blocks(content).strip()
|
||||
if not visible_text:
|
||||
return False
|
||||
if len(visible_text) < 20 or not re.search(r"\s", visible_text):
|
||||
return False
|
||||
|
||||
return not self._has_natural_response_ending(visible_text)
|
||||
|
||||
def _looks_like_codex_intermediate_ack(
|
||||
self,
|
||||
user_message: str,
|
||||
|
|
@ -3978,6 +4037,7 @@ class AIAgent:
|
|||
"model", "instructions", "input", "tools", "store",
|
||||
"reasoning", "include", "max_output_tokens", "temperature",
|
||||
"tool_choice", "parallel_tool_calls", "prompt_cache_key", "service_tier",
|
||||
"extra_headers",
|
||||
}
|
||||
normalized: Dict[str, Any] = {
|
||||
"model": model,
|
||||
|
|
@ -4013,6 +4073,20 @@ class AIAgent:
|
|||
if val is not None:
|
||||
normalized[passthrough_key] = val
|
||||
|
||||
extra_headers = api_kwargs.get("extra_headers")
|
||||
if extra_headers is not None:
|
||||
if not isinstance(extra_headers, dict):
|
||||
raise ValueError("Codex Responses request 'extra_headers' must be an object.")
|
||||
normalized_headers: Dict[str, str] = {}
|
||||
for key, value in extra_headers.items():
|
||||
if not isinstance(key, str) or not key.strip():
|
||||
raise ValueError("Codex Responses request 'extra_headers' keys must be non-empty strings.")
|
||||
if value is None:
|
||||
continue
|
||||
normalized_headers[key.strip()] = str(value)
|
||||
if normalized_headers:
|
||||
normalized["extra_headers"] = normalized_headers
|
||||
|
||||
if allow_stream:
|
||||
stream = api_kwargs.get("stream")
|
||||
if stream is not None and stream is not True:
|
||||
|
|
@ -6451,7 +6525,12 @@ class AIAgent:
|
|||
if not is_github_responses:
|
||||
kwargs["prompt_cache_key"] = self.session_id
|
||||
|
||||
if reasoning_enabled:
|
||||
is_xai_responses = self.provider == "xai" or "api.x.ai" in (self.base_url or "").lower()
|
||||
|
||||
if reasoning_enabled and is_xai_responses:
|
||||
# xAI reasons automatically — no effort param, just include encrypted content
|
||||
kwargs["include"] = ["reasoning.encrypted_content"]
|
||||
elif reasoning_enabled:
|
||||
if is_github_responses:
|
||||
# Copilot's Responses route advertises reasoning-effort support,
|
||||
# but not OpenAI-specific prompt cache or encrypted reasoning
|
||||
|
|
@ -6462,7 +6541,7 @@ class AIAgent:
|
|||
else:
|
||||
kwargs["reasoning"] = {"effort": reasoning_effort, "summary": "auto"}
|
||||
kwargs["include"] = ["reasoning.encrypted_content"]
|
||||
elif not is_github_responses:
|
||||
elif not is_github_responses and not is_xai_responses:
|
||||
kwargs["include"] = []
|
||||
|
||||
if self.request_overrides:
|
||||
|
|
@ -6471,6 +6550,9 @@ class AIAgent:
|
|||
if self.max_tokens is not None and not is_codex_backend:
|
||||
kwargs["max_output_tokens"] = self.max_tokens
|
||||
|
||||
if is_xai_responses and getattr(self, "session_id", None):
|
||||
kwargs["extra_headers"] = {"x-grok-conv-id": self.session_id}
|
||||
|
||||
return kwargs
|
||||
|
||||
sanitized_messages = api_messages
|
||||
|
|
@ -6635,18 +6717,24 @@ class AIAgent:
|
|||
options["num_ctx"] = self._ollama_num_ctx
|
||||
extra_body["options"] = options
|
||||
|
||||
# Ollama / custom provider: pass think=false when reasoning is disabled.
|
||||
# Ollama does not recognise the OpenRouter-style `reasoning` extra_body
|
||||
# field, so we use its native `think` parameter instead.
|
||||
# This prevents thinking-capable models (Qwen3, etc.) from generating
|
||||
# <think> blocks and producing empty-response errors when the user has
|
||||
# set reasoning_effort: none.
|
||||
if self.provider == "custom" and self.reasoning_config and isinstance(self.reasoning_config, dict):
|
||||
_effort = (self.reasoning_config.get("effort") or "").strip().lower()
|
||||
_enabled = self.reasoning_config.get("enabled", True)
|
||||
if _effort == "none" or _enabled is False:
|
||||
extra_body["think"] = False
|
||||
|
||||
if self._is_qwen_portal():
|
||||
extra_body["vl_high_resolution_images"] = True
|
||||
|
||||
if extra_body:
|
||||
api_kwargs["extra_body"] = extra_body
|
||||
|
||||
# xAI prompt caching: send x-grok-conv-id header to route requests
|
||||
# to the same server, maximizing automatic cache hits.
|
||||
# https://docs.x.ai/developers/advanced-api-usage/prompt-caching
|
||||
if "x.ai" in self._base_url_lower and hasattr(self, "session_id") and self.session_id:
|
||||
api_kwargs["extra_headers"] = {"x-grok-conv-id": self.session_id}
|
||||
|
||||
# Priority Processing / generic request overrides (e.g. service_tier).
|
||||
# Applied last so overrides win over any defaults set above.
|
||||
if self.request_overrides:
|
||||
|
|
@ -6757,9 +6845,16 @@ class AIAgent:
|
|||
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)
|
||||
|
||||
msg = {
|
||||
"role": "assistant",
|
||||
"content": assistant_message.content or "",
|
||||
"content": _san_content,
|
||||
"reasoning": reasoning_text,
|
||||
"finish_reason": finish_reason,
|
||||
}
|
||||
|
|
@ -7418,7 +7513,7 @@ class AIAgent:
|
|||
# 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.KAWAII_WAITING)
|
||||
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()
|
||||
|
||||
|
|
@ -7432,24 +7527,50 @@ class AIAgent:
|
|||
|
||||
# Wait for all to complete with periodic heartbeats so the
|
||||
# gateway's inactivity monitor doesn't kill us during long
|
||||
# concurrent tool batches.
|
||||
# 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=30.0,
|
||||
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)
|
||||
_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])})"
|
||||
)
|
||||
# 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
|
||||
|
|
@ -7461,8 +7582,11 @@ class AIAgent:
|
|||
for i, (tc, name, args) in enumerate(parsed_calls):
|
||||
r = results[i]
|
||||
if r is None:
|
||||
# Shouldn't happen, but safety fallback
|
||||
function_result = f"Error executing tool '{name}': thread did not return a result"
|
||||
# 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 = r
|
||||
|
|
@ -7714,7 +7838,7 @@ class AIAgent:
|
|||
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.KAWAII_WAITING)
|
||||
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
|
||||
|
|
@ -7741,7 +7865,7 @@ class AIAgent:
|
|||
# Context engine tools (lcm_grep, lcm_describe, lcm_expand, etc.)
|
||||
spinner = None
|
||||
if self.quiet_mode and not self.tool_progress_callback:
|
||||
face = random.choice(KawaiiSpinner.KAWAII_WAITING)
|
||||
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)
|
||||
|
|
@ -7765,7 +7889,7 @@ class AIAgent:
|
|||
# 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.KAWAII_WAITING)
|
||||
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)
|
||||
|
|
@ -7787,7 +7911,7 @@ class AIAgent:
|
|||
elif self.quiet_mode:
|
||||
spinner = None
|
||||
if self._should_emit_quiet_tool_messages() and self._should_start_quiet_spinner():
|
||||
face = random.choice(KawaiiSpinner.KAWAII_WAITING)
|
||||
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)
|
||||
|
|
@ -8150,6 +8274,16 @@ class AIAgent:
|
|||
if isinstance(persist_user_message, str):
|
||||
persist_user_message = _sanitize_surrogates(persist_user_message)
|
||||
|
||||
# Strip leaked <memory-context> blocks from user input. When Honcho's
|
||||
# saveMessages persists a turn that included injected context, the block
|
||||
# can reappear in the next turn's user message via message history.
|
||||
# Stripping here prevents stale memory tags from leaking into the
|
||||
# conversation and being visible to the user or the model as user text.
|
||||
if isinstance(user_message, str):
|
||||
user_message = sanitize_context(user_message)
|
||||
if isinstance(persist_user_message, str):
|
||||
persist_user_message = sanitize_context(persist_user_message)
|
||||
|
||||
# Store stream callback for _interruptible_api_call to pick up
|
||||
self._stream_callback = stream_callback
|
||||
self._persist_user_message_idx = None
|
||||
|
|
@ -8429,6 +8563,16 @@ class AIAgent:
|
|||
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).
|
||||
|
|
@ -8620,6 +8764,12 @@ class AIAgent:
|
|||
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)
|
||||
|
|
@ -8633,8 +8783,8 @@ class AIAgent:
|
|||
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.KAWAII_THINKING)
|
||||
verb = random.choice(KawaiiSpinner.THINKING_VERBS)
|
||||
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)
|
||||
|
|
@ -9018,6 +9168,17 @@ class AIAgent:
|
|||
finish_reason = stop_reason_map.get(response.stop_reason, "stop")
|
||||
else:
|
||||
finish_reason = response.choices[0].finish_reason
|
||||
assistant_message = response.choices[0].message
|
||||
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)
|
||||
|
|
@ -10792,8 +10953,9 @@ class AIAgent:
|
|||
# tool(result) → assistant("(empty)") → user(nudge)
|
||||
# Without this, we'd have tool → user which most
|
||||
# APIs reject as an invalid sequence.
|
||||
assistant_msg["content"] = "(empty)"
|
||||
messages.append(assistant_msg)
|
||||
_nudge_msg = self._build_assistant_message(assistant_message, finish_reason)
|
||||
_nudge_msg["content"] = "(empty)"
|
||||
messages.append(_nudge_msg)
|
||||
messages.append({
|
||||
"role": "user",
|
||||
"content": (
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue