Based on PR #7285 by @kshitijk4poor.
Two bugs affecting Qwen OAuth users:
1. Wrong context window — qwen3-coder-plus showed 128K instead of 1M.
Added specific entries before the generic qwen catch-all:
- qwen3-coder-plus: 1,000,000 (corrected from PR's 1,048,576 per
official Alibaba Cloud docs and OpenRouter)
- qwen3-coder: 262,144
2. Random stopping — max_tokens was suppressed for Qwen Portal, so the
server applied its own low default. Reasoning models exhaust that on
thinking tokens. Now: honor explicit max_tokens, default to 65536
when unset.
Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>
* feat: add watch_patterns to background processes for output monitoring
Adds a new 'watch_patterns' parameter to terminal(background=true) that
lets the agent specify strings to watch for in process output. When a
matching line appears, a notification is queued and injected as a
synthetic message — triggering a new agent turn, similar to
notify_on_complete but mid-process.
Implementation:
- ProcessSession gets watch_patterns field + rate-limit state
- _check_watch_patterns() in ProcessRegistry scans new output chunks
from all three reader threads (local, PTY, env-poller)
- Rate limited: max 8 notifications per 10s window
- Sustained overload (45s) permanently disables watching for that process
- watch_queue alongside completion_queue, same consumption pattern
- CLI drains watch_queue in both idle loop and post-turn drain
- Gateway drains after agent runs via _inject_watch_notification()
- Checkpoint persistence + crash recovery includes watch_patterns
- Blocked in execute_code sandbox (like other bg params)
- 20 new tests covering matching, rate limiting, overload kill,
checkpoint persistence, schema, and handler passthrough
Usage:
terminal(
command='npm run dev',
background=true,
watch_patterns=['ERROR', 'WARN', 'listening on port']
)
* refactor: merge watch_queue into completion_queue
Unified queue with 'type' field distinguishing 'completion',
'watch_match', and 'watch_disabled' events. Extracted
_format_process_notification() in CLI and gateway to handle
all event types in a single drain loop. Removes duplication
across both CLI drain sites and the gateway.
The _PROVIDER_MODELS['openai-codex'] list was missing gpt-5.4 and gpt-5.4-mini,
causing them to not appear in the /model picker for ChatGPT OAuth users.
codex_models.py already had these models in DEFAULT_CODEX_MODELS, but the
curated list that feeds the Telegram/Discord /model picker was never updated.
Reported by @chongdashu
The System Overview ASCII diagram had inconsistent box widths:
- Entry Points box bottom border was 73 chars instead of 71
This caused the docs-site-checks CI to fail on every docs-only PR
due to pre-existing errors in the diagram.
Fix: normalize Entry Points bottom border to 71 characters,
matching the top border width.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Cover all public functions with 50 test cases:
- managed_nous_tools_enabled() feature flag toggling
- normalize_browser_cloud_provider() coercion and defaults
- coerce_modal_mode() / normalize_modal_mode() validation
- has_direct_modal_credentials() env vars and config file detection
- resolve_modal_backend_state() full backend selection matrix
- resolve_openai_audio_api_key() priority chain and edge cases
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Follow-up fixes for cherry-pick conflicts:
- Removed test_context_keeps_pending_approval test that referenced
pop_pending() which doesn't exist on current main
- Added headers attribute to FakeResponse in vision test (needed
after #6949 added Content-Length check)
Three fixes for vision_analyze returning cryptic 400 "Invalid request data":
1. Pre-flight base64 size check — base64 inflates data ~33%, so a 3.8 MB
file exceeds the 5 MB API limit. Reject early with a clear message
instead of letting the provider return a generic 400.
2. Handle file:// URIs — strip the scheme and resolve as a local path.
Previously file:///path/to/image.png fell through to the "invalid
image source" error since it matched neither is_file() nor http(s).
3. Separate invalid_request errors from "does not support vision" errors
so the user gets actionable guidance (resize/compress/retry) instead
of a misleading "model does not support vision" message.
Closes#6677
vision_tools.py: _download_image() loads the full HTTP response body into
memory via response.content (line 190) with no Content-Length check and no
max file size limit. An attacker-hosted multi-gigabyte file causes OOM.
Add a 50 MB hard cap: check Content-Length header before download, and
verify actual body size before writing to disk.
hermes_parser.py: tc_data["name"] at line 57 raises KeyError when the LLM
outputs a tool call JSON without a "name" field. The outer except catches
it silently, causing the entire tool call to be lost with zero diagnostics.
Add "name" field validation before constructing the ChatCompletionMessage.
mistral_parser.py: tc["name"] at line 101 has the same KeyError issue in
the pre-v11 format path. The fallback decoder (line 112) already checks
"name" correctly, but the primary path does not. Add validation to match.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
process_registry.py: _reader_loop() has process.wait() after the try-except
block (line 380). If the reader thread crashes with an unexpected exception
(e.g. MemoryError, KeyboardInterrupt), control exits the except handler but
skips wait() — leaving the child as a zombie process. Move wait() and the
cleanup into a finally block so the child is always reaped.
cron/scheduler.py: _run_job_script() only redacts secrets in stdout on the
SUCCESS path (line 417-421). When a cron script fails (non-zero exit), both
stdout and stderr are returned WITHOUT redaction (lines 407-413). A script
that accidentally prints an API key to stderr during a failure would leak it
into the LLM context. Move redaction before the success/failure branch so
both paths benefit.
skill_commands.py: _build_skill_message() enumerates supporting files using
rglob("*") but only checks is_file() (line 171) without filtering symlinks.
PR #6693 added symlink protection to scan_skill_commands() but missed this
function. A malicious skill can create symlinks in references/ pointing to
arbitrary files, exposing their paths (and potentially content via skill_view)
to the LLM. Add is_symlink() check to match the guard in scan_skill_commands.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
YAML parses bare numeric keys (e.g. `12306:`) as int, causing
TypeError when sorted() is called on mixed int/str collections.
Changes:
- Normalize toolset_names entries to str in _get_platform_tools()
- Cast MCP server name to str(name) when building enabled_mcp_servers
- Add regression test
When the stream consumer has sent at least one message (already_sent=True),
the gateway skips sending the final response to avoid duplicates. But this
also suppressed error messages when the agent failed mid-loop — rate limit
exhaustion, context overflow, compression failure, etc.
The user would see the last streamed content and then nothing: no error
message, no explanation. The agent appeared to 'stop responding.'
Fix: check the 'failed' flag at both the producer (_run_agent marks
already_sent) and consumer (_handle_message_with_agent checks it) sites.
Error messages are always delivered regardless of streaming state.
async_call_llm (and call_llm) can return non-OpenAI objects from
custom providers or adapter shims, crashing downstream consumers
with misleading AttributeError ('str' has no attribute 'choices').
Add _validate_llm_response() that checks the response has the
expected .choices[0].message shape before returning. Wraps all
return paths in call_llm, async_call_llm, and fallback paths.
Fails fast with a clear RuntimeError identifying the task, response
type, and a preview of the malformed payload.
Closes#7264
`resolve_provider_client()` already drops OpenRouter-format model slugs
(containing "/") when the resolved provider is not OpenRouter (line 1097).
However, `_get_cached_client()` returns `model or cached_default` directly
on cache hits, bypassing this check entirely.
When the main provider is openai-codex, the auto-detection chain (Step 1
of `_resolve_auto`) caches a CodexAuxiliaryClient. Subsequent auxiliary
calls for different tasks (e.g. compression with `summary_model:
google/gemini-3-flash-preview`) hit the cache and pass the OpenRouter-
format model slug straight to the Codex Responses API, which does not
understand it and returns an empty `response.output`.
This causes two user-visible failures:
- "Invalid API response shape" (empty output after 3 retries)
- "Context length exceeded, cannot compress further" (compression itself
fails through the same path)
Add `_compat_model()` helper that mirrors the "/" check from
`resolve_provider_client()` and call it on the cache-hit return path.
Four fixes to auxiliary_client.py:
1. Respect explicit provider as hard constraint (#7559)
When auxiliary.{task}.provider is explicitly set (not 'auto'),
connection/payment errors no longer silently fallback to cloud
providers. Local-only users (Ollama, vLLM) will no longer get
unexpected OpenRouter billing from auxiliary tasks.
2. Eliminate model='default' sentinel (#7512)
_resolve_api_key_provider() no longer sends literal 'default' as
model name to APIs. Providers without a known aux model in
_API_KEY_PROVIDER_AUX_MODELS are skipped instead of producing
model_not_supported errors.
3. Add payment/connection fallback to async_call_llm (#7512)
async_call_llm now mirrors sync call_llm's fallback logic for
payment (402) and connection errors. Previously, async consumers
(session_search, web_tools, vision) got hard failures with no
recovery. Also fixes hardcoded 'openrouter' fallback to use the
full auto-detection chain.
4. Use accurate error reason in fallback logs (#7512)
_try_payment_fallback() now accepts a reason parameter and uses
it in log messages. Connection timeouts are no longer misleadingly
logged as 'payment error'.
Closes#7559Closes#7512
The auxiliary client always calls client.chat.completions.create(),
ignoring the api_mode config flag. This breaks codex-family models
(e.g. gpt-5.3-codex) on direct OpenAI API keys, which need the
/v1/responses endpoint.
Changes:
- Expand _resolve_task_provider_model to return api_mode (5-tuple)
- Read api_mode from auxiliary.{task}.api_mode config and env vars
(AUXILIARY_{TASK}_API_MODE)
- Pass api_mode through _get_cached_client to resolve_provider_client
- Add _needs_codex_wrap/_wrap_if_needed helpers that wrap plain OpenAI
clients in CodexAuxiliaryClient when api_mode=codex_responses or
when auto-detection finds api.openai.com + codex model pattern
- Apply wrapping at all custom endpoint, named custom provider, and
API-key provider return paths
- Update test mocks for the new 5-tuple return format
Users can now set:
auxiliary:
compression:
model: gpt-5.3-codex
base_url: https://api.openai.com/v1
api_mode: codex_responses
Closes#6800
Refactor hardcoded color constants throughout the CLI to resolve from
the active skin engine, so custom themes fully control the visual
appearance.
cli.py:
- Replace _GOLD constant with _ACCENT (_SkinAwareAnsi class) that
lazily resolves response_border from the active skin
- Rename _GOLD_DEFAULT to _ACCENT_ANSI_DEFAULT
- Make _build_compact_banner() read banner_title/accent/dim from skin
- Make session resume notifications use _accent_hex()
- Make status line use skin colors (accent_color, separator_color,
label_color instead of cryptic _dim_c/_dim_c2/_accent_c/_label_c)
- Reset _ACCENT cache on /skin switch
agent/display.py:
- Replace hardcoded diff ANSI escapes with skin-aware functions:
_diff_dim(), _diff_file(), _diff_hunk(), _diff_minus(), _diff_plus()
(renamed from SCREAMING_CASE _ANSI_* to snake_case)
- Add reset_diff_colors() for cache invalidation on skin switch
_discover_bundled_skills() used the directory name to identify skills,
but skills_tool.py and skills_hub.py use the `name:` field from SKILL.md
frontmatter. This mismatch caused 9 builtin skills whose directory name
differs from their SKILL.md name to be written to .bundled_manifest
under the wrong key, so `hermes skills list` showed them as "local"
instead of "builtin".
Read the frontmatter name field (with directory-name fallback) so the
manifest keys match what the rest of the codebase expects.
Closes#6835
Aligns MiniMax provider with official API documentation. Fixes 6 bugs:
transport mismatch (openai_chat -> anthropic_messages), credential leak
in switch_model(), prompt caching sent to non-Anthropic endpoints,
dot-to-hyphen model name corruption, trajectory compressor URL routing,
and stale doctor health check.
Also corrects context window (204,800), thinking support (manual mode),
max output (131,072), and model catalog (M2 family only on /anthropic).
Source: https://platform.minimax.io/docs/api-reference/text-anthropic-api
Co-authored-by: kshitijk4poor <kshitijk4poor@users.noreply.github.com>
Two fixes for the honcho memory plugin: (1) initOnSessionStart — opt-in eager session init in tools mode so sync_turn() works from turn 1 (default false, non-breaking). (2) peerName fix — gateway user_id no longer silently overwrites an explicitly configured peerName. 11 new tests. Contributed by @Kathie-yu.
The pre_llm_call plugin hook receives session_id, user_message,
conversation_history, is_first_turn, model, and platform — but not
the sender's user_id. This means plugins cannot perform per-user
access control (e.g. restricting knowledge base recall to authorized
users).
The gateway already passes source.user_id as user_id to AIAgent,
which stores it in self._user_id. This change forwards it as
sender_id in the pre_llm_call kwargs so plugins can use it for
ACL decisions.
For CLI sessions where no user_id exists, sender_id defaults to
empty string. Plugins can treat empty sender_id as a trusted local
call (the owner is at the terminal) or deny it depending on their
ACL policy.
_is_oauth_token() returned True for any key not starting with 'sk-ant-api',
which means MiniMax and Alibaba API keys were falsely treated as Anthropic
OAuth tokens. This triggered the Claude Code compatibility path:
- All tool names prefixed with mcp_ (e.g. mcp_terminal, mcp_web_search)
- System prompt injected with 'You are Claude Code' identity
- 'Hermes Agent' replaced with 'Claude Code' throughout
Fix: Make _is_oauth_token() positively identify Anthropic OAuth tokens by
their key format instead of using a broad catch-all:
- sk-ant-* (but not sk-ant-api-*) -> setup tokens, managed keys
- eyJ* -> JWTs from Anthropic OAuth flow
- Everything else -> False (MiniMax, Alibaba, etc.)
Reported by stefan171.
Resumed sessions showed raw JSON tool output in content boxes instead
of the compact trail lines seen during live use. The root cause was
two separate rendering paths with no shared code.
Extract buildToolTrailLine() into lib/text.ts as the single source
of truth for formatting tool trail lines. Both the live tool.complete
handler and toTranscriptMessages now call it.
Server-side, reconstruct tool name and args from the assistant
message's tool_calls field (tool_name column is unpopulated) and
pass them through _tool_ctx/build_tool_preview — the same path
the live tool.start callback uses.
Generate project ideas through creative constraints. Constraint + direction
= creativity.
Core skill (SKILL.md, 147 lines):
- 15 curated constraints across 3 categories: developers, makers, anyone
- Developer-focused prompts: 'solve your own itch', 'the CLI tool that
should exist', 'automate the annoying thing', 'nothing new except glue'
- Matching table: maps user mood/intent to appropriate constraints
- Complete worked example with 3 concrete project ideas
- Output format for consistent, actionable idea presentation
Extended library (references/full-prompt-library.md, 110 lines):
- 30+ additional constraints: communication, screens, philosophy,
transformation, identity, scale, starting points
Constraint approach inspired by wttdotm.com/prompts.html. Adapted for
software development and general-purpose ideation.
session.resume was building conversation history with only role and
content, stripping tool_call_id, tool_calls, and tool_name. The API
requires tool messages to reference their parent tool_call, so resumed
sessions with tool history would fail with HTTP 500.
Use get_messages_as_conversation() which already preserves the full
message structure including tool metadata and reasoning fields.
- Add agent.close() call to _finalize_shutdown_agents() to prevent
zombie processes (terminal sandboxes, browser daemons, httpx clients)
- Global cleanup (process_registry, environments, browsers) preserved
in _stop_impl() during conflict resolution
- Move /restart CommandDef from 'Info' to 'Session' category to match
/stop and /status
* fix: circuit breaker stops CPU-burning restart loops on persistent errors
When a gateway session hits a non-retryable error (e.g. invalid model
ID → HTTP 400), the agent fails and returns. But if the session keeps
receiving messages (or something periodically recreates agents), each
attempt spawns a new AIAgent — reinitializing MCP server connections,
burning CPU — only to hit the same 400 error again. On a 4-core server,
this pegs an entire core per stuck session and accumulates 300+ minutes
of CPU time over hours.
Fix: add a per-session consecutive failure counter in the gateway runner.
- Track consecutive non-retryable failures per session key
- After 3 consecutive failures (_MAX_CONSECUTIVE_FAILURES), block
further agent creation for that session and notify the user:
'⚠️ This session has failed N times in a row with a non-retryable
error. Use /reset to start a new session.'
- Evict the cached agent when the circuit breaker engages to prevent
stale state from accumulating
- Reset the counter on successful agent runs
- Clear the counter on /reset and /new so users can recover
- Uses getattr() pattern so bare GatewayRunner instances (common in
tests using object.__new__) don't crash
Tests:
- 8 new tests in test_circuit_breaker.py covering counter behavior,
threshold, reset, session isolation, and bare-runner safety
Addresses #7130.
* Revert "fix: circuit breaker stops CPU-burning restart loops on persistent errors"
This reverts commit d848ea7109.
* fix: don't evict cached agent on failed runs — prevents MCP restart loop
When a run fails (e.g. invalid model ID → 400) and fallback activated,
the gateway was evicting the cached agent to 'retry primary next time.'
But evicting a failed agent forces a full AIAgent recreation on the next
message — reinitializing MCP server connections, spawning stdio
processes — only to hit the same 400 again. This created a CPU-burning
loop (91%+ for hours, #7130).
The fix: add `and not _run_failed` to the fallback-eviction check.
Failed runs keep the cached agent. The next message reuses it (no MCP
reinit), hits the same error, returns it to the user quickly. The user
can /reset or /model to fix their config.
Successful fallback runs still evict as before so the next message
retries the primary model.
Addresses #7130.
- Remove unreachable `if not content_sample` branch inside the truthy
`if content_sample` block in `_is_likely_binary()` (dead code that
could never execute).
- Replace `linter_cmd.format(file=...)` with `linter_cmd.replace("{file}", ...)`
in `_check_lint()` so file paths containing curly braces (e.g.
`src/{test}.py`) no longer raise KeyError/ValueError.
- Add 16 unit tests covering both fixes and edge cases.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
