Port two improvements inspired by Kilo-Org/kilocode analysis:
1. Error classifier: add context overflow patterns for vLLM, Ollama,
and llama.cpp/llama-server. These local inference servers return
different error formats than cloud providers (e.g., 'exceeds the
max_model_len', 'context length exceeded', 'slot context'). Without
these patterns, context overflow errors from local servers are
misclassified as format errors, causing infinite retries instead
of triggering compression.
2. MCP initial connection retry: previously, if the very first
connection attempt to an MCP server failed (e.g., transient DNS
blip at startup), the server was permanently marked as failed with
no retry. Post-connect reconnection had 5 retries with exponential
backoff, but initial connection had zero. Now initial connections
retry up to 3 times with backoff before giving up, matching the
resilience of post-connect reconnection.
(Inspired by Kilo Code's MCP server disappearing fix in v1.3.3)
Tests: 6 new error classifier tests, 4 new MCP retry tests, 1
updated existing test. All 276 affected tests pass.
Adds Arcee AI as a standard direct provider (ARCEEAI_API_KEY) with
Trinity models: trinity-large-thinking, trinity-large-preview, trinity-mini.
Standard OpenAI-compatible provider checklist: auth.py, config.py,
models.py, main.py, providers.py, doctor.py, model_normalize.py,
model_metadata.py, setup.py, trajectory_compressor.py.
Based on PR #9274 by arthurbr11, simplified to a standard direct
provider without dual-endpoint OpenRouter routing.
- Use isinstance() with try/except import for CopilotACPClient check
in _to_async_client instead of fragile __class__.__name__ string check
- Restore accurate comment: GPT-5.x models *require* (not 'often require')
the Responses API on OpenAI/OpenRouter; ACP is the exception, not a
softening of the requirement
- Add inline comment explaining the ACP exclusion rationale
Cherry-picked from PR #7637 by hcshen0111.
Adds kimi-coding-cn provider with dedicated KIMI_CN_API_KEY env var
and api.moonshot.cn/v1 endpoint for China-region Moonshot users.
The v11→v12 migration converts custom_providers (list) into providers
(dict), then deletes the list. But all runtime resolvers read from
custom_providers — after migration, named custom endpoints silently stop
resolving and fallback chains fail with AuthError.
Add get_compatible_custom_providers() that reads from both config schemas
(legacy custom_providers list + v12+ providers dict), normalizes entries,
deduplicates, and returns a unified list. Update ALL consumers:
- hermes_cli/runtime_provider.py: _get_named_custom_provider() + key_env
- hermes_cli/auth_commands.py: credential pool provider names
- hermes_cli/main.py: model picker + _model_flow_named_custom()
- agent/auxiliary_client.py: key_env + custom_entry model fallback
- agent/credential_pool.py: _iter_custom_providers()
- cli.py + gateway/run.py: /model switch custom_providers passthrough
- run_agent.py + gateway/run.py: per-model context_length lookup
Also: use config.pop() instead of del for safer migration, fix stale
_config_version assertions in tests, add pool mock to codex test.
Co-authored-by: 墨綠BG <s5460703@gmail.com>
Closes#8776, salvaged from PR #8814
resolve_vision_provider_client() computed resolved_api_mode from config
but never passed it to downstream resolve_provider_client() or
_get_cached_client() calls, causing custom providers with
api_mode: anthropic_messages to crash when used for vision tasks.
Also remove the for_vision special case in _normalize_aux_provider()
that incorrectly discarded named custom provider identifiers.
Fixes#8857
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove the backward-compat code paths that read compression provider/model
settings from legacy config keys and env vars, which caused silent failures
when auto-detection resolved to incompatible backends.
What changed:
- Remove compression.summary_model, summary_provider, summary_base_url from
DEFAULT_CONFIG and cli.py defaults
- Remove backward-compat block in _resolve_task_provider_model() that read
from the legacy compression section
- Remove _get_auxiliary_provider() and _get_auxiliary_env_override() helper
functions (AUXILIARY_*/CONTEXT_* env var readers)
- Remove env var fallback chain for per-task overrides
- Update hermes config show to read from auxiliary.compression
- Add config migration (v16→17) that moves non-empty legacy values to
auxiliary.compression and strips the old keys
- Update example config and openclaw migration script
- Remove/update tests for deleted code paths
Compression model/provider is now configured exclusively via:
auxiliary.compression.provider / auxiliary.compression.model
Closes#8923
_query_local_context_length was checking model_info.context_length
(the GGUF training max) before num_ctx (the Modelfile runtime override),
inverse to query_ollama_num_ctx. The two helpers therefore disagreed on
the same model:
hermes-brain:qwen3-14b-ctx32k # Modelfile: num_ctx 32768
underlying qwen3:14b GGUF # qwen3.context_length: 40960
query_ollama_num_ctx correctly returned 32768 (the value Ollama will
actually allocate KV cache for). _query_local_context_length returned
40960, which let ContextCompressor grow conversations past 32768 before
triggering compression — at which point Ollama silently truncated the
prefix, corrupting context.
Swap the order so num_ctx is checked first, matching query_ollama_num_ctx.
Adds a parametrized test that seeds both values and asserts num_ctx wins.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
auxiliary_client.py had its own regex mirroring _strip_think_blocks
but was missing the <thought> variant. Also adds test coverage for
<thought> paired and orphaned tags.
The previous wording ('If one clearly matches') set too high a threshold,
and 'If none match, proceed normally' was an easy escape hatch for lazy
models. Now:
- Lowered threshold: 'matches or is even partially relevant'
- Added MUST directive and 'err on the side of loading' guidance
- Replaced permissive closer with 'only proceed without if genuinely none
are relevant'
This should reduce cases where the agent skips loading relevant skills
unless explicitly forced.
When running inside WSL (Windows Subsystem for Linux), inject a hint into
the system prompt explaining that the Windows host filesystem is mounted
at /mnt/c/, /mnt/d/, etc. This lets the agent naturally translate Windows
paths (Desktop, Documents) to their /mnt/ equivalents without the user
needing to configure anything.
Uses the existing is_wsl() detection from hermes_constants (cached,
checks /proc/version for 'microsoft'). Adds build_environment_hints()
in prompt_builder.py — extensible for Termux, Docker, etc. later.
Closes the UX gap where WSL users had to manually explain path
translation to the agent every session.
OpenAI OAuth refresh tokens are single-use and rotate on every refresh.
When Hermes refreshes a Codex token, it consumed the old refresh_token
but never wrote the new pair back to ~/.codex/auth.json. This caused
Codex CLI and VS Code to fail with 'refresh_token_reused' on their
next refresh attempt.
This mirrors the existing Anthropic write-back pattern where refreshed
tokens are written to ~/.claude/.credentials.json via
_write_claude_code_credentials().
Changes:
- Add _write_codex_cli_tokens() in hermes_cli/auth.py (parallel to
_write_claude_code_credentials in anthropic_adapter.py)
- Call it from _refresh_codex_auth_tokens() (non-pool refresh path)
- Call it from credential_pool._refresh_entry() (pool happy path + retry)
- Add tests for the new write-back behavior
- Update existing test docstring to clarify _save_codex_tokens vs
_write_codex_cli_tokens separation
Fixes refresh token conflict reported by @ec12edfae2cb221
The previous wording ('If one clearly matches') set too high a threshold,
and 'If none match, proceed normally' was an easy escape hatch for lazy
models. Now:
- Lowered threshold: 'matches or is even partially relevant'
- Added MUST directive and 'err on the side of loading' guidance
- Replaced permissive closer with 'only proceed without if genuinely none
are relevant'
This should reduce cases where the agent skips loading relevant skills
unless explicitly forced.
- Add openai/openai-codex -> openai mapping to PROVIDER_TO_MODELS_DEV
so context-length lookups use models.dev data instead of 128k fallback.
Fixes#8161.
- Set api_mode from custom_providers entry when switching via hermes model,
and clear stale api_mode when the entry has none. Also extract api_mode
in _named_custom_provider_map(). Fixes#8181.
- Convert OpenAI image_url content blocks to Anthropic image blocks when
the endpoint is Anthropic-compatible (MiniMax, MiniMax-CN, or any URL
containing /anthropic). Fixes#8147.
Users whose credentials exist only in external files — OpenAI Codex
OAuth tokens in ~/.codex/auth.json or Anthropic Claude Code credentials
in ~/.claude/.credentials.json — would not see those providers in the
/model picker, even though hermes auth and hermes model detected them.
Root cause: list_authenticated_providers() only checked the raw Hermes
auth store and env vars. External credential file fallbacks (Codex CLI
import, Claude Code file discovery) were never triggered.
Fix (three parts):
1. _seed_from_singletons() in credential_pool.py: openai-codex now
imports from ~/.codex/auth.json when the Hermes auth store is empty,
mirroring resolve_codex_runtime_credentials().
2. list_authenticated_providers() in model_switch.py: auth store + pool
checks now run for ALL providers (not just OAuth auth_type), catching
providers like anthropic that support both API key and OAuth.
3. list_authenticated_providers(): direct check for anthropic external
credential files (Claude Code, Hermes PKCE). The credential pool
intentionally gates anthropic behind is_provider_explicitly_configured()
to prevent auxiliary tasks from silently consuming tokens. The /model
picker bypasses this gate since it is discovery-oriented.
After compression, models (especially Kimi 2.5) would sometimes respond
to questions from the summary instead of the latest user message. This
happened ~30% of the time on Telegram.
Root cause: the summary's 'Next Steps' section read as active instructions,
and the SUMMARY_PREFIX didn't explicitly tell the model to ignore questions
in the summary. When the summary merged into the first tail message, there
was no clear separator between historical context and the actual user message.
Changes inspired by competitor analysis (Claude Code, OpenCode, Codex):
1. SUMMARY_PREFIX rewritten with explicit 'Do NOT answer questions from
this summary — respond ONLY to the latest user message AFTER it'
2. Summarizer preamble (shared by both prompts) adds:
- 'Do NOT respond to any questions' (from OpenCode's approach)
- 'Different assistant' framing (from Codex) to create psychological
distance between summary content and active conversation
3. New summary sections:
- '## Resolved Questions' — tracks already-answered questions with
their answers, preventing re-answering (from Claude Code's
'Pending user asks' pattern)
- '## Pending User Asks' — explicitly marks unanswered questions
- '## Remaining Work' replaces '## Next Steps' — passive framing
avoids reading as active instructions
4. merge-summary-into-tail path now inserts a clear separator:
'--- END OF CONTEXT SUMMARY — respond to the message below ---'
5. Iterative update prompt now instructs: 'Move answered questions to
Resolved Questions' to maintain the resolved/pending distinction
across multiple compactions.
Adds an optional focus topic to /compress: `/compress database schema`
guides the summariser to preserve information related to the focus topic
(60-70% of summary budget) while compressing everything else more aggressively.
Inspired by Claude Code's /compact <focus>.
Changes:
- context_compressor.py: focus_topic parameter on _generate_summary() and
compress(); appends FOCUS TOPIC guidance block to the LLM prompt
- run_agent.py: focus_topic parameter on _compress_context(), passed through
to the compressor
- cli.py: _manual_compress() extracts focus topic from command string,
preserves existing manual_compression_feedback integration (no regression)
- gateway/run.py: _handle_compress_command() extracts focus from event args
and passes through — full gateway parity
- commands.py: args_hint="[focus topic]" on /compress CommandDef
Salvaged from PR #7459 (CLI /compress focus only — /context command deferred).
15 new tests across CLI, compressor, and gateway.
Switch estimate_tokens_rough(), estimate_messages_tokens_rough(), and
estimate_request_tokens_rough() from floor division (len // 4) to
ceiling division ((len + 3) // 4). Short texts (1-3 chars) previously
estimated as 0 tokens, causing the compressor and pre-flight checks to
systematically undercount when many short tool results are present.
Also replaced the inline duplicate formula in run_conversation()
(total_chars // 4) with a call to the shared
estimate_messages_tokens_rough() function.
Updated 4 tests that hardcoded floor-division expected values.
Related: issue #6217, PR #6629
Three root causes of the 'agent stops mid-task' gateway bug:
1. Compression threshold floor (64K tokens minimum)
- The 50% threshold on a 100K-context model fired at 50K tokens,
causing premature compression that made models lose track of
multi-step plans. Now threshold_tokens = max(50% * context, 64K).
- Models with <64K context are rejected at startup with a clear error.
2. Budget warning removal — grace call instead
- Removed the 70%/90% iteration budget warnings entirely. These
injected '[BUDGET WARNING: Provide your final response NOW]' into
tool results, causing models to abandon complex tasks prematurely.
- Now: no warnings during normal execution. When the budget is
actually exhausted (90/90), inject a user message asking the model
to summarise, allow one grace API call, and only then fall back
to _handle_max_iterations.
3. Activity touches during long terminal execution
- _wait_for_process polls every 0.2s but never reported activity.
The gateway's inactivity timeout (default 1800s) would fire during
long-running commands that appeared 'idle.'
- Now: thread-local activity callback fires every 10s during the
poll loop, keeping the gateway's activity tracker alive.
- Agent wires _touch_activity into the callback before each tool call.
Also: docs update noting 64K minimum context requirement.
Closes#7915 (root cause was agent-loop termination, not Weixin delivery limits).
* fix(tools): neutralize shell injection in _write_to_sandbox via path quoting
_write_to_sandbox interpolated storage_dir and remote_path directly into
a shell command passed to env.execute(). Paths containing shell
metacharacters (spaces, semicolons, $(), backticks) could trigger
arbitrary command execution inside the sandbox.
Fix: wrap both paths with shlex.quote(). Clean paths (alphanumeric +
slashes/hyphens/dots) are left unmodified by shlex.quote, so existing
behavior is unchanged. Paths with unsafe characters get single-quoted.
Tests added for spaces, $(command) substitution, and semicolon injection.
* fix: is_local_endpoint misses Docker/Podman DNS names
host.docker.internal, host.containers.internal, gateway.docker.internal,
and host.lima.internal are well-known DNS names that container runtimes
use to resolve the host machine. Users running Ollama on the host with
the agent in Docker/Podman hit the default 120s stream timeout instead
of the bumped 1800s because these hostnames weren't recognized as local.
Add _CONTAINER_LOCAL_SUFFIXES tuple and suffix check in
is_local_endpoint(). Tests cover all three runtime families plus a
negative case for domains that merely contain the suffix as a substring.
The auxiliary client previously checked env vars (AUXILIARY_{TASK}_PROVIDER,
AUXILIARY_{TASK}_MODEL, etc.) before config.yaml's auxiliary.{task}.* section.
This violated the project's '.env is for secrets only' policy — these are
behavioral settings, not API keys.
Flipped the resolution order in _resolve_task_provider_model():
1. Explicit args (always win)
2. config.yaml auxiliary.{task}.* (PRIMARY)
3. Env var overrides (backward-compat fallback only)
4. 'auto' (full auto-detection chain)
Env var reading code is kept for backward compatibility but config.yaml
now takes precedence. Updated module docstring and function docstring.
Also removed AUXILIARY_VISION_MODEL from _EXTRA_ENV_KEYS in config.py.
Cherry-picked from PR #7702 by kshitijk4poor.
Adds Xiaomi MiMo as a direct provider (XIAOMI_API_KEY) with models:
- mimo-v2-pro (1M context), mimo-v2-omni (256K, multimodal), mimo-v2-flash (256K, cheapest)
Standard OpenAI-compatible provider checklist: auth.py, config.py, models.py,
main.py, providers.py, doctor.py, model_normalize.py, model_metadata.py,
models_dev.py, auxiliary_client.py, .env.example, cli-config.yaml.example.
Follow-up: vision tasks use mimo-v2-omni (multimodal) instead of the user's
main model. Non-vision aux uses the user's selected model. Added
_PROVIDER_VISION_MODELS dict for provider-specific vision model overrides.
On failure, falls back to aggregators (gemini flash) via existing fallback chain.
Corrects pre-existing context lengths: mimo-v2-pro 1048576→1000000,
mimo-v2-omni 1048576→256000, adds mimo-v2-flash 256000.
36 tests covering registry, aliases, auto-detect, credentials, models.dev,
normalization, URL mapping, providers module, doctor, aux client, vision
model override, and agent init.
Cherry-picked from PR #7749 by kshitijk4poor with modifications:
- Raise hard image limit from 5 MB to 20 MB (matches most restrictive provider)
- Send images at full resolution first; only auto-resize to 5 MB on API failure
- Add _is_image_size_error() helper to detect size-related API rejections
- Auto-resize uses Pillow (soft dep) with progressive downscale + JPEG quality reduction
- Fix get_model_capabilities() to check modalities.input for vision support
- Increase default vision timeout from 30s to 120s (matches hardcoded fallback intent)
- Applied retry-with-resize to both vision_analyze_tool and browser_vision
Closes#7740
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>
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>
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
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>
_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.
GPT-5+ models (except gpt-5-mini) are only accessible via the Responses
API on Copilot. When these models were configured as the compression
summary_model (or any auxiliary task), the plain OpenAI client sent them
to /chat/completions which returned a 400 error:
model "gpt-5.4-mini" is not accessible via the /chat/completions endpoint
resolve_provider_client() now checks _should_use_copilot_responses_api()
for the copilot provider and wraps the client in CodexAuxiliaryClient
when needed, routing calls through responses.stream() transparently.
Adds tests for both the wrapping (gpt-5.4-mini) and non-wrapping
(gpt-4.1-mini) paths.
Follow-up fixes for the context engine plugin slot (PR #5700):
- Enhance ContextEngine ABC: add threshold_percent, protect_first_n,
protect_last_n as class attributes; complete update_model() default
with threshold recalculation; clarify on_session_end() lifecycle docs
- Add ContextCompressor.update_model() override for model/provider/
base_url/api_key updates
- Replace all direct compressor internal access in run_agent.py with
ABC interface: switch_model(), fallback restore, context probing
all use update_model() now; _context_probed guarded with getattr/
hasattr for plugin engine compatibility
- Create plugins/context_engine/ directory with discovery module
(mirrors plugins/memory/ pattern) — discover_context_engines(),
load_context_engine()
- Add context.engine config key to DEFAULT_CONFIG (default: compressor)
- Config-driven engine selection in run_agent.__init__: checks config,
then plugins/context_engine/<name>/, then general plugin system,
falls back to built-in ContextCompressor
- Wire on_session_end() in shutdown_memory_provider() at real session
boundaries (CLI exit, /reset, gateway expiry)
- PluginContext.register_context_engine() lets plugins replace the
built-in ContextCompressor with a custom ContextEngine implementation
- PluginManager stores the registered engine; only one allowed
- run_agent.py checks for a plugin engine at init before falling back
to the default ContextCompressor
- reset_session_state() now calls engine.on_session_reset() instead of
poking internal attributes directly
- ContextCompressor.on_session_reset() handles its own internals
(_context_probed, _previous_summary, etc.)
- 19 new tests covering ABC contract, defaults, plugin slot registration,
rejection of duplicates/non-engines, and compressor reset behavior
- All 34 existing compressor tests pass unchanged
Introduces agent/context_engine.py — an abstract base class that defines
the pluggable context engine interface. ContextCompressor now inherits
from ContextEngine as the default implementation.
No behavior change. All 34 existing compressor tests pass.
This is the foundation for a context engine plugin slot, enabling
third-party engines like LCM (Lossless Context Management) to replace
the built-in compressor via the plugin system.
When two gateway messages arrived concurrently, _set_session_env wrote
HERMES_SESSION_PLATFORM/CHAT_ID/CHAT_NAME/THREAD_ID into the process-global
os.environ. Because asyncio tasks share the same process, Message B would
overwrite Message A's values mid-flight, causing background-task notifications
and tool calls to route to the wrong thread/chat.
Replace os.environ with Python's contextvars.ContextVar. Each asyncio task
(and any run_in_executor thread it spawns) gets its own copy, so concurrent
messages never interfere.
Changes:
- New gateway/session_context.py with ContextVar definitions, set/clear/get
helpers, and os.environ fallback for CLI/cron/test backward compatibility
- gateway/run.py: _set_session_env returns reset tokens, _clear_session_env
accepts them for proper cleanup in finally blocks
- All tool consumers updated: cronjob_tools, send_message_tool, skills_tool,
terminal_tool (both notify_on_complete AND check_interval blocks), tts_tool,
agent/skill_utils, agent/prompt_builder
- Tests updated for new contextvar-based API
Fixes#7358
Co-authored-by: teknium1 <127238744+teknium1@users.noreply.github.com>
Adds xAI as a first-class provider: ProviderConfig in auth.py,
HermesOverlay in providers.py, 11 curated Grok models, URL mapping
in model_metadata.py, aliases (x-ai, x.ai), and env var tests.
Uses standard OpenAI-compatible chat completions.
Closes#7050
- Remove sys.path.insert hack (leftover from standalone dev)
- Add token lock (acquire_scoped_lock/release_scoped_lock) in
connect()/disconnect() to prevent duplicate pollers across profiles
- Fix get_connected_platforms: WEIXIN check must precede generic
token/api_key check (requires both token AND account_id)
- Add WEIXIN_HOME_CHANNEL_NAME to _EXTRA_ENV_KEYS
- Add gateway setup wizard with QR login flow
- Add platform status check for partially configured state
- Add weixin.md docs page with full adapter documentation
- Update environment-variables.md reference with all 11 env vars
- Update sidebars.ts to include weixin docs page
- Wire all gateway integration points onto current main
Salvaged from PR #6747 by Zihan Huang.
Port from anomalyco/opencode#21355: Alibaba's DashScope API returns a
unique throttling message ('Request rate increased too quickly...') that
doesn't match standard rate-limit patterns ('rate limit', 'too many
requests'). This caused Alibaba errors to fall through to the 'unknown'
category rather than being properly classified as rate_limit with
appropriate backoff/rotation.
Add 'rate increased too quickly' to _RATE_LIMIT_PATTERNS and test with
the exact error message observed from the Alibaba provider.
_resolve_api_key_provider() now checks is_provider_explicitly_configured
before calling _try_anthropic(). Previously, any auxiliary fallback
(e.g. when kimi-coding key was invalid) would silently discover and use
Claude Code OAuth tokens — consuming the user's Claude Max subscription
without their knowledge.
This is the auxiliary-client counterpart of the setup-wizard gate in
PR #4210.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously, removing a claude_code credential from the anthropic pool
only printed a note — the next load_pool() re-seeded it from
~/.claude/.credentials.json. Now writes a 'suppressed_sources' flag
to auth.json that _seed_from_singletons checks before seeding.
Follows the pattern of env: source removal (clears .env var) and
device_code removal (clears auth store state).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
_seed_from_singletons('anthropic') now checks
is_provider_explicitly_configured('anthropic') before reading
~/.claude/.credentials.json. Without this, the auxiliary client
fallback chain silently discovers and uses Claude Code tokens when
the user's primary provider key is invalid — consuming their Claude
Max subscription quota without consent.
Follows the same gating pattern as PR #4210 (setup wizard gate)
but applied to the credential pool seeding path.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Automated dead code audit using vulture + coverage.py + ast-grep intersection,
confirmed by Opus deep verification pass. Every symbol verified to have zero
production callers (test imports excluded from reachability analysis).
Removes ~1,534 lines of dead production code across 46 files and ~1,382 lines
of stale test code. 3 entire files deleted (agent/builtin_memory_provider.py,
hermes_cli/checklist.py, tests/hermes_cli/test_setup_model_selection.py).
Co-authored-by: alt-glitch <balyan.sid@gmail.com>
prompt_builder.py: The `hidden_div` detection pattern uses `.*` which does not
match newlines in Python regex (re.DOTALL is not passed). An attacker can bypass
detection by splitting the style attribute across lines:
`<div style="color:red;\ndisplay: none">injected content</div>`
Replace `.*` with `[\s\S]*?` to match across line boundaries.
credential_files.py: `_load_config_files()` catches all exceptions at DEBUG level
(line 171), making YAML parse failures invisible in production logs. Users whose
credential files silently fail to mount into sandboxes have no diagnostic clue.
Promote to WARNING to match the severity pattern used by the path validation
warnings at lines 150 and 158 in the same function.
webhook.py: `_reload_dynamic_routes()` logs JSON parse failures at WARNING (line
265) but the impact — stale/corrupted dynamic routes persisting silently — warrants
ERROR level to ensure operator visibility in alerting pipelines.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
error_classifier.py: Message-only auth errors ("invalid api key", "unauthorized",
etc.) were classified as retryable=True (line 707), inconsistent with the HTTP 401
path (line 432) which correctly uses retryable=False + should_fallback=True. The
mismatch causes 3 wasted retries with the same broken credential before fallback,
while 401 errors immediately attempt fallback. Align the message-based path to
match: retryable=False, should_fallback=True.
web_tools.py: The _PREFIX_RE secret-detection check in web_extract_tool() runs
against the raw URL string (line 1196). URL-encoded secrets like %73k-1234... (
sk-1234...) bypass the filter because the regex expects literal ASCII. Add
urllib.parse.unquote() before the check so percent-encoded variants are also caught.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
xAI /v1/models does not return context_length metadata, so Hermes
probes down to the 128k default whenever a user configures a custom
provider pointing at https://api.x.ai/v1. This forces every xAI user
to manually override model.context_length in config.yaml (2M for
Grok 4.20 / 4.1-fast / 4-fast) or lose most of the usable context
window.
Add DEFAULT_CONTEXT_LENGTHS entries for the Grok family so the
fallback lookup returns the correct value via substring matching.
Values sourced from models.dev (2026-04) and cross-checked against
the xAI /v1/models listing:
- grok-4.20-* 2,000,000 (reasoning, non-reasoning, multi-agent)
- grok-4-1-fast-* 2,000,000
- grok-4-fast-* 2,000,000
- grok-4 / grok-4-0709 256,000
- grok-code-fast-1 256,000
- grok-3* 131,072
- grok-2 / latest 131,072
- grok-2-vision* 8,192
- grok (catch-all) 131,072
Keys are ordered longest-first so that specific variants match before
the catch-all, consistent with the existing Claude/Gemma/MiniMax entries.
Add TestDefaultContextLengths.test_grok_models_context_lengths and
test_grok_substring_matching to pin the values and verify the full
lookup path. All 77 tests in test_model_metadata.py pass.
Auth errors matched by message pattern were incorrectly marked retryable=True, causing futile retry loops. Aligns with _classify_by_status() which already sets retryable=False for 401/403. Fixes#7026. Contributed by @kuishou68.
The hardcoded User-Agent 'KimiCLI/1.3' is outdated — Kimi CLI is now at
v1.30.0. The stale version string causes intermittent 403 errors from
Kimi's coding endpoint ('only available for Coding Agents').
Update all 8 occurrences across run_agent.py, auxiliary_client.py, and
doctor.py to 'KimiCLI/1.30.0' to match the current official Kimi CLI.
Extends the /fast command to support Anthropic's Fast Mode beta in addition
to OpenAI Priority Processing. When enabled on Claude Opus 4.6, adds
speed:"fast" and the fast-mode-2026-02-01 beta header to API requests for
~2.5x faster output token throughput.
Changes:
- hermes_cli/models.py: Add _ANTHROPIC_FAST_MODE_MODELS registry,
model_supports_fast_mode() now recognizes Claude Opus 4.6,
resolve_fast_mode_overrides() returns {speed: fast} for Anthropic
vs {service_tier: priority} for OpenAI
- agent/anthropic_adapter.py: Add _FAST_MODE_BETA constant,
build_anthropic_kwargs() accepts fast_mode=True which injects
speed:fast + beta header via extra_headers (skipped for third-party
Anthropic-compatible endpoints like MiniMax)
- run_agent.py: Pass fast_mode to build_anthropic_kwargs in the
anthropic_messages path of _build_api_kwargs()
- cli.py: Update _handle_fast_command with provider-aware messaging
(shows 'Anthropic Fast Mode' vs 'Priority Processing')
- hermes_cli/commands.py: Update /fast description to mention both
providers
- tests: 13 new tests covering Anthropic model detection, override
resolution, CLI availability, routing, adapter kwargs, and
third-party endpoint safety
When the model mentions <think> as literal text in its response (e.g.
"(/think not producing <think> tags)"), the streaming display treated it
as a reasoning block opener and suppressed everything after it. The
response box would close with truncated content and no error — the API
response was complete but the display ate it.
Root cause: _stream_delta() matched <think> anywhere in the text stream
regardless of position. Real reasoning blocks always start at the
beginning of a line; mentions in prose appear mid-sentence.
Fix: track line position across streaming deltas with a
_stream_last_was_newline flag. Only enter reasoning suppression when
the tag appears at a block boundary (start of stream, after a newline,
or after only whitespace on the current line). Add a _flush_stream()
safety net that recovers buffered content if no closing tag is found
by end-of-stream.
Also fixes three related issues discovered during investigation:
- anthropic_adapter: _get_anthropic_max_output() now normalizes dots to
hyphens so 'claude-opus-4.6' matches the 'claude-opus-4-6' table key
(was returning 32K instead of 128K)
- run_agent: send explicit max_tokens for Claude models on Nous Portal,
same as OpenRouter — both proxy to Anthropic's API which requires it.
Without it the backend defaults to a low limit that truncates responses.
- run_agent: reset truncated_tool_call_retries after successful tool
execution so a single truncation doesn't poison the entire conversation.
The Codex retry block and valid-token short-circuit in _refresh_entry()
both return early, bypassing the auth.json sync at the end of the method.
This adds _sync_device_code_entry_to_auth_store() calls on both paths
so refreshed/synced tokens are written back to auth.json regardless of
which code path succeeds.
MiniMax's Anthropic-compatible endpoints reject requests that include
the fine-grained-tool-streaming beta header — every tool-use message
triggers a connection error (~18s timeout). Regular chat works fine.
Add _common_betas_for_base_url() that filters out the tool-streaming
beta for Bearer-auth (MiniMax) endpoints while keeping all other betas.
All four client-construction branches now use the filtered list.
Based on #6528 by @HiddenPuppy.
Original cherry-picked from PR #6688 by kshitijk4poor.
Fixes#6510, fixes#6555.
_classify_by_message had no handling for _USAGE_LIMIT_PATTERNS, so
messages like 'usage limit exceeded, try again in 5 minutes' arriving
without an HTTP status code fell through to FailoverReason.unknown
instead of rate_limit.
Apply the same billing/rate-limit disambiguation that _classify_402
already uses: USAGE_LIMIT_PATTERNS + transient signal → rate_limit,
USAGE_LIMIT_PATTERNS alone → billing.
Add 4 tests covering the no-status-code usage-limit path.
When _generate_summary() failed (no provider, timeout, model error),
the compressor silently dropped all middle turns with just a debug
log. The agent would then see head + tail with no explanation of the
gap, causing total context amnesia (generic greetings instead of
continuing the conversation).
Now generates a static fallback marker that tells the model context
was lost and to continue from the recent tail messages. The fallback
flows through the same role-alternation logic as a real summary so
message structure stays valid.
Step 1 of _resolve_auto() explicitly excluded 'custom' providers,
forcing custom endpoint users through the fragile fallback chain
instead of using their known-working main model credentials.
This caused silent compression failures for users on local OpenAI-
compatible endpoints — the summary generation would fail, middle
turns would be silently dropped, and the agent would lose all
conversation context.
Remove 'custom' from the exclusion list so custom endpoint users
get the same main-model-first treatment as DeepSeek, Anthropic,
Gemini, and other direct providers.
When the API returns "max_tokens too large given prompt" (input tokens
are within the context window, but input + requested output > window),
the old code incorrectly routed through the same handler as "prompt too
long" errors, calling get_next_probe_tier() and permanently halving
context_length. This made things worse: the window was fine, only the
requested output size needed trimming for that one call.
Two distinct error classes now handled separately:
Prompt too long — input itself exceeds context window.
Fix: compress history + halve context_length (existing behaviour,
unchanged).
Output cap too large — input OK, but input + max_tokens > window.
Fix: parse available_tokens from the error message, set a one-shot
_ephemeral_max_output_tokens override for the retry, and leave
context_length completely untouched.
Changes:
- agent/model_metadata.py: add parse_available_output_tokens_from_error()
that detects Anthropic's "available_tokens: N" error format and returns
the available output budget, or None for all other error types.
- run_agent.py: call the new parser first in the is_context_length_error
block; if it fires, set _ephemeral_max_output_tokens (with a 64-token
safety margin) and break to retry without touching context_length.
_build_api_kwargs consumes the ephemeral value exactly once then clears
it so subsequent calls use self.max_tokens normally.
- agent/anthropic_adapter.py: expand build_anthropic_kwargs docstring to
clearly document the max_tokens (output cap) vs context_length (total
window) distinction, which is a persistent source of confusion due to
the OpenAI-inherited "max_tokens" name.
- cli-config.yaml.example: add inline comments explaining both keys side
by side where users are most likely to look.
- website/docs/integrations/providers.md: add a callout box at the top
of "Context Length Detection" and clarify the troubleshooting entry.
- tests/test_ctx_halving_fix.py: 24 tests across four classes covering
the parser, build_anthropic_kwargs clamping, ephemeral one-shot
consumption, and the invariant that context_length is never mutated
on output-cap errors.
The error classifier's generic-400 heuristic only extracted err_body_msg from
the nested body structure (body['error']['message']), missing the flat body
format used by OpenAI's Responses API (body['message']). This caused
descriptive 400 errors like 'Invalid input[index].name: string does not match
pattern' to appear generic when the session was large, misclassifying them as
context overflow and triggering an infinite compression loop.
Added flat-body fallback in _classify_400() consistent with the parent
classify_api_error() function's existing handling at line 297-298.
When is explicitly set to ,
the custom-endpoint path in creates a plain
client without provider-specific headers. This means sync vision calls (e.g.
) use the generic User-Agent and get rejected by
Kimi's coding endpoint with a 403:
'Kimi For Coding is currently only available for Coding Agents such as Kimi CLI...'
The async converter already injects , and the
auto-detected API-key provider path also injects it, but the explicit custom
endpoint shortcut was missing it entirely.
This patch adds the same injection to the custom endpoint
branch, and updates all existing Kimi header sites to for
consistency.
Fixes <issue number to be filled in>
The credential pool seeder (_seed_from_env) hardcoded the base URL
for API-key providers without running provider-specific auto-detection.
For kimi-coding, this caused sk-kimi- prefixed keys to be seeded with
the legacy api.moonshot.ai/v1 endpoint instead of api.kimi.com/coding/v1,
resulting in HTTP 401 on the first request.
Import and call _resolve_kimi_base_url for kimi-coding so the pool
uses the correct endpoint based on the key prefix, matching the
runtime credential resolver behavior.
Also fix a comment: sk-kimi- keys are issued by kimi.com/code,
not platform.kimi.ai.
Fixes#5561
Two bugs in the model fallback system:
1. Nous login leaves stale model in config (provider=nous, model=opus
from previous OpenRouter setup). Fixed by deferring the config.yaml
provider write until AFTER model selection completes, and passing the
selected model atomically via _update_config_for_provider's
default_model parameter. Previously, _update_config_for_provider was
called before model selection — if selection failed (free tier, no
models, exception), config stayed as nous+opus permanently.
2. Codex/stale providers in auxiliary fallback can't connect but block
the auto-detection chain. Added _is_connection_error() detection
(APIConnectionError, APITimeoutError, DNS failures, connection
refused) alongside the existing _is_payment_error() check in
call_llm(). When a provider endpoint is unreachable, the system now
falls back to the next available provider instead of crashing.
Parse x-ratelimit-* headers from inference API responses (Nous Portal,
OpenRouter, OpenAI-compatible) and display them in the /usage command.
- New agent/rate_limit_tracker.py: parse 12 rate limit headers (RPM/RPH/
TPM/TPH limits, remaining, reset timers), format as progress bars (CLI)
or compact one-liner (gateway)
- Hook into streaming path in run_agent.py: stream.response.headers is
available on the OpenAI SDK Stream object before chunks are consumed
- CLI /usage: appends rate limit section with progress bars + warnings
when any bucket exceeds 80%
- Gateway /usage: appends compact rate limit summary
- 24 unit tests covering parsing, formatting, edge cases
Headers captured per response:
x-ratelimit-{limit,remaining,reset}-{requests,tokens}{,-1h}
Example CLI display:
Nous Rate Limits (captured just now):
Requests/min [░░░░░░░░░░░░░░░░░░░░] 0.1% 1/800 used (799 left, resets in 59s)
Tokens/hr [░░░░░░░░░░░░░░░░░░░░] 0.0% 49/336.0M (336.0M left, resets in 52m)
Wrap is_dir() in _is_valid_subdir() and is_file() in
_load_hints_for_directory() with OSError handlers so that
inaccessible directories (e.g. /root from a non-root Daytona
host user) are silently skipped instead of crashing the agent.
The existing PermissionError PRs for prompt_builder.py (#6247,
#6321, #6355) do not cover subdirectory_hints.py, which was
identified as a separate crash path in the #6214 comments.
Ref: #6214
The 24-hour default cooldown for 402-exhausted credentials was far too
aggressive — if a user tops up credits or the 402 was caused by an
oversized max_tokens request rather than true billing exhaustion, they
shouldn't have to wait a full day. Reduce to 1 hour (matching the
existing 429 TTL).
Inspired by PR #6493 (michalkomar).
Two issues resolved:
1. Add opencode.ai to _URL_TO_PROVIDER mapping so base_url routes through
models.dev lookup (which has mimo-v2-pro at 1M context) instead of
falling back to probing /models (404) and defaulting to 128K.
2. Fix _format_context_length to round cleanly: 1048576 → '1M' instead
of '1.048576M'. Applies same rounding logic to K values.
Tail protection was effectively message-count based despite having a
token budget, because protect_last_n=20 acted as a hard floor. A single
50K-token tool output would cause all 20 recent messages to be
preserved regardless of budget, leaving little room for summarization.
Changes:
- _find_tail_cut_by_tokens: min_tail reduced from protect_last_n (20)
to 3; token budget is now the primary criterion
- Soft ceiling at 1.5x budget to avoid cutting mid-oversized-message
- _prune_old_tool_results: accepts optional protect_tail_tokens so
pruning also respects the token budget instead of a fixed count
- compress() minimum message check relaxed from protect_first_n +
protect_last_n + 1 to protect_first_n + 3 + 1
- Tool group alignment (no splitting tool_call/result) preserved
Three targeted improvements to the compression system:
1. Replace hardcoded truncation limits with named class constants
(_CONTENT_MAX=6000, _CONTENT_HEAD=4000, _CONTENT_TAIL=1500,
_TOOL_ARGS_MAX=1500, _TOOL_ARGS_HEAD=1200). Previous limits
(3000/500) heavily truncated the summarizer's input — a 200-line
edit got cut to 3000 chars before the summarizer ever saw it.
2. Add '## Tools & Patterns' section to both compression prompt
templates (first-pass and iterative). Preserves working tool
invocations, preferred flags, and tool-specific discoveries
across compaction boundaries.
3. Warn users on 2nd+ compression: 'Session compressed N times —
accuracy may degrade. Consider /new to start fresh.'
Ref #499
Two linked fixes for MiniMax Anthropic-compatible fallback:
1. Normalize httpx.URL to str before calling .rstrip() in auth/provider
detection helpers. Some client objects expose base_url as httpx.URL,
not str — crashed with AttributeError in _requires_bearer_auth() and
_is_third_party_anthropic_endpoint(). Also fixes _try_activate_fallback()
to use the already-stringified fb_base_url instead of raw httpx.URL.
2. Strip Anthropic-proprietary thinking block signatures when targeting
third-party Anthropic-compatible endpoints (MiniMax, Azure AI Foundry,
self-hosted proxies). These endpoints cannot validate Anthropic's
signatures and reject them with HTTP 400 'Invalid signature in
thinking block'. Now threads base_url through convert_messages_to_anthropic()
→ build_anthropic_kwargs() so signature management is endpoint-aware.
Based on PR #4945 by kshitijk4poor (rstrip fix).
Fixes#4944.
Fixes 9 test failures on current main, incorporating ideas from PR stack
#6219-#6222 by xinbenlv with corrections:
- model_metadata: sync HF context length key casing
(minimaxai/minimax-m2.5 → MiniMaxAI/MiniMax-M2.5)
- cli.py: route quick command error output through self.console
instead of creating a new ChatConsole() instance
- docker.py: explicit docker_forward_env entries now bypass the
Hermes secret blocklist (intentional opt-in wins over generic filter)
- auxiliary_client: revert _read_main_provider() to simple
provider.strip().lower() — the _normalize_aux_provider() call
introduced in 5c03f2e7 stripped the custom: prefix, breaking
named custom provider resolution
- auxiliary_client: flip vision auto-detection order to
active provider → OpenRouter → Nous → stop (was OR → Nous → active)
- test: update vision priority test to match new order
Based on PR #6219-#6222 by xinbenlv.
- Add HERMES_QWEN_BASE_URL to OPTIONAL_ENV_VARS in config.py (was missing
despite being referenced in code)
- Remove redundant qwen-oauth entry from _API_KEY_PROVIDER_AUX_MODELS
(non-aggregator providers use their main model for aux tasks automatically)
Based on #6079 by @tunamitom with critical fixes and comprehensive tests.
Changes from #6079:
- Fix: sanitization overwrite bug — Qwen message prep now runs AFTER codex
field sanitization, not before (was silently discarding Qwen transforms)
- Fix: missing try/except AuthError in runtime_provider.py — stale Qwen
credentials now fall through to next provider on auto-detect
- Fix: 'qwen' alias conflict — bare 'qwen' stays mapped to 'alibaba'
(DashScope); use 'qwen-portal' or 'qwen-cli' for the OAuth provider
- Fix: hardcoded ['coder-model'] replaced with live API fetch + curated
fallback list (qwen3-coder-plus, qwen3-coder)
- Fix: extract _is_qwen_portal() helper + _qwen_portal_headers() to replace
5 inline 'portal.qwen.ai' string checks and share headers between init
and credential swap
- Fix: add Qwen branch to _apply_client_headers_for_base_url for mid-session
credential swaps
- Fix: remove suspicious TypeError catch blocks around _prompt_provider_choice
- Fix: handle bare string items in content lists (were silently dropped)
- Fix: remove redundant dict() copies after deepcopy in message prep
- Revert: unrelated ai-gateway test mock removal and model_switch.py comment deletion
New tests (30 test functions):
- _qwen_cli_auth_path, _read_qwen_cli_tokens (success + 3 error paths)
- _save_qwen_cli_tokens (roundtrip, parent creation, permissions)
- _qwen_access_token_is_expiring (5 edge cases: fresh, expired, within skew,
None, non-numeric)
- _refresh_qwen_cli_tokens (success, preserve old refresh, 4 error paths,
default expires_in, disk persistence)
- resolve_qwen_runtime_credentials (fresh, auto-refresh, force-refresh,
missing token, env override)
- get_qwen_auth_status (logged in, not logged in)
- Runtime provider resolution (direct, pool entry, alias)
- _build_api_kwargs (metadata, vl_high_resolution_images, message formatting,
max_tokens suppression)
Hermes Agent identified and patched its own prompting blind spots through
automated self-evaluation — running 64+ tool-use benchmarks across GPT-5.4
and Codex-5.3, diagnosing 5 failure modes, writing targeted prompt patches,
and verifying the fix in a closed loop.
Failure modes discovered and fixed:
- Mental arithmetic (wrong answers: 39,152,053 vs correct 39,151,253)
- User profile hallucination ('Windows 11' when running on Linux)
- Time guessing without verification
- Clarification-seeking instead of acting ('open where?' for port checks)
- Hash computation from memory (SHA-256, encodings)
- Confusing system RAM with agent's own persistent memory store
Two new XML sections added to OPENAI_MODEL_EXECUTION_GUIDANCE:
- <mandatory_tool_use>: explicit categories that must always use tools
- <act_dont_ask>: default to action on obvious interpretations
Results:
gpt-5.4: 68.8% → 100% tool compliance (+31.2pp)
gpt-5.3-codex: 62.5% → 100% tool compliance (+37.5pp)
Regression: 0/8 conversational prompts over-tooled
Anthropic signs thinking blocks against the full turn content. Any
upstream mutation (context compression, session truncation, orphan
stripping, message merging) invalidates the signature, causing HTTP 400
'Invalid signature in thinking block' — especially in long-lived
gateway sessions.
Strategy (following clawdbot/OpenClaw pattern):
1. Strip thinking/redacted_thinking from all assistant messages EXCEPT
the last one — preserves reasoning continuity on the current
tool-use chain while avoiding stale signature errors on older turns.
2. Downgrade unsigned thinking blocks to plain text — Anthropic can't
validate them, but the reasoning content is preserved.
3. Strip cache_control from thinking/redacted_thinking blocks to
prevent cache markers from interfering with signature validation.
4. Drop thinking blocks from the second message when merging
consecutive assistant messages (role alternation enforcement).
5. Error recovery: on HTTP 400 mentioning 'signature' and 'thinking',
strip all reasoning_details from the conversation and retry once.
This is the safety net for edge cases the proactive stripping
misses.
Addresses the issue reported in PR #6086 by @mingginwan while
preserving reasoning continuity (their PR stripped ALL thinking
blocks unconditionally).
Files changed:
- agent/anthropic_adapter.py: thinking block management in
convert_messages_to_anthropic (strip old turns, downgrade unsigned,
strip cache_control, merge-time strip)
- run_agent.py: one-shot signature error recovery in retry loop
- tests/test_anthropic_adapter.py: 10 new tests covering all cases
Simplify the vision auto-detection chain from 5 backends (openrouter,
nous, codex, anthropic, custom) down to 3:
1. OpenRouter (known vision-capable default model)
2. Nous Portal (known vision-capable default model)
3. Active provider + model (whatever the user is running)
4. Stop
This is simpler and more predictable. The active provider step uses
resolve_provider_client() which handles all provider types including
named custom providers (from #5978).
Removed the complex preferred-provider promotion logic and API-level
fallback — the chain is short enough that it doesn't need them.
Based on PR #5376 by Mibay. Closes#5366.
Salvaged fixes from community PRs:
- fix(model_switch): _read_auth_store → _load_auth_store + fix auth store
key lookup (was checking top-level dict instead of store['providers']).
OAuth providers now correctly detected in /model picker.
Cherry-picked from PR #5911 by Xule Lin (linxule).
- fix(ollama): pass num_ctx to override 2048 default context window.
Ollama defaults to 2048 context regardless of model capabilities. Now
auto-detects from /api/show metadata and injects num_ctx into every
request. Config override via model.ollama_num_ctx. Fixes#2708.
Cherry-picked from PR #5929 by kshitij (kshitijk4poor).
- fix(aux): normalize provider aliases for vision/auxiliary routing.
Adds _normalize_aux_provider() with 17 aliases (google→gemini,
claude→anthropic, glm→zai, etc). Fixes vision routing failure when
provider is set to 'google' instead of 'gemini'.
Cherry-picked from PR #5793 by e11i (Elizabeth1979).
- fix(aux): rewrite MiniMax /anthropic base URLs to /v1 for OpenAI SDK.
MiniMax's inference_base_url ends in /anthropic (Anthropic Messages API),
but auxiliary client uses OpenAI SDK which appends /chat/completions →
404 at /anthropic/chat/completions. Generic _to_openai_base_url() helper
rewrites terminal /anthropic to /v1 for OpenAI-compatible endpoint.
Inspired by PR #5786 by Lempkey.
Added debug logging to silent exception blocks across all fixes.
Co-authored-by: Hermes Agent <hermes@nousresearch.com>
Free-tier Nous Portal users were getting mimo-v2-omni (a multimodal
model) for all auxiliary tasks including compression, session search,
and web extraction. Now routes non-vision tasks to mimo-v2-pro (a
text model) which is better suited for those workloads.
- Added _NOUS_FREE_TIER_AUX_MODEL constant for text auxiliary tasks
- _try_nous() accepts vision=False param to select the right model
- Vision path (_resolve_strict_vision_backend) passes vision=True
- All other callers default to vision=False → mimo-v2-pro
* fix(telegram): replace substring caption check with exact line-by-line match
Captions in photo bursts and media group albums were silently dropped when
a shorter caption happened to be a substring of an existing one (e.g.
"Meeting" lost inside "Meeting agenda"). Extract a shared _merge_caption
static helper that splits on "\n\n" and uses exact match with whitespace
normalisation, then use it in both _enqueue_photo_event and
_queue_media_group_event.
Adds 13 unit tests covering the fixed bug scenarios.
Cherry-picked from PR #2671 by Dilee.
* fix: extend caption substring fix to all platforms
Move _merge_caption helper from TelegramAdapter to BasePlatformAdapter
so all adapters inherit it. Fix the same substring-containment bug in:
- gateway/platforms/base.py (photo burst merging)
- gateway/run.py (priority photo follow-up merging)
- gateway/platforms/feishu.py (media batch merging)
The original fix only covered telegram.py. The same bug existed in base.py
and run.py (pure substring check) and feishu.py (list membership without
whitespace normalization).
* fix(auxiliary): resolve named custom providers and 'main' alias in auxiliary routing
Two bugs caused auxiliary tasks (vision, compression, etc.) to fail when
using named custom providers defined in config.yaml:
1. 'provider: main' was hardcoded to 'custom', which only checks legacy
OPENAI_BASE_URL env vars. Now reads _read_main_provider() to resolve
to the actual provider (e.g., 'custom:beans', 'openrouter', 'deepseek').
2. Named custom provider names (e.g., 'beans') fell through to
PROVIDER_REGISTRY which doesn't know about config.yaml entries.
Now checks _get_named_custom_provider() before the registry fallback.
Fixes both resolve_provider_client() and _normalize_vision_provider()
so the fix covers all auxiliary tasks (vision, compression, web_extract,
session_search, etc.).
Adds 13 unit tests. Reported by Laura via Discord.
---------
Co-authored-by: Dilee <uzmpsk.dilekakbas@gmail.com>
