Add _QWEN_THINKING_MODELS to _MODEL_THINKING_STYLES with enable_thinking style. Prevents Qwen 3.5/3.6/3.7 models from exposing raw reasoning content in chat responses. Closes#4934
Symptom
-------
LLM requests intermittently fail with:
'utf-8' codec can't encode characters in position N-N+1: surrogates not allowed
when messages contain emoji-heavy content (e.g. HTML with mixed emoji + JSON round-trips).
This blocks the affected session until the session file is quarantined.
Root cause
----------
Surrogate sanitization was only applied at the CLI entry point
(nanobot/cli/commands.py: _sanitize_surrogates). Requests entering
the LLM provider layer through other channels (Feishu, cron, webui,
tool results, memory injection) had no defensive cleaning, so any
message that happened to carry unpaired UTF-16 surrogates (from an
upstream JSON round-trip with ensure_ascii=True on ill-formed input,
memory rehydration, or third-party content) would blow up at
json.dumps -> HTTP encode time inside the provider client.
Fix
---
1. Extract sanitize_surrogates() and sanitize_surrogates_deep() into
nanobot/utils/helpers.py as the single source of truth. Both use
utf-16-le round-tripping with errors='surrogatepass' / 'replace',
so paired surrogates reconstruct back into their real code point
and lone surrogates collapse to U+FFFD.
2. Make nanobot/cli/commands.py:_sanitize_surrogates a thin wrapper
that re-exports the shared helper (backward compatible).
3. Add defense-in-depth at the LLM provider boundary in
nanobot/providers/base.py:_sanitize_empty_content by running
sanitize_surrogates_deep over each message and its content blocks
right before requests are serialized to JSON.
Non-goals
---------
- truncate_text() is intentionally left untouched. Python str slicing
cannot split a single code point into surrogate halves, so it is
not the source of lone surrogates.
- session/manager storage layer is untouched. Archived sessions
reproduced the failure only through the request path, not through
storage.
Verification
------------
- New regression suite tests/providers/test_sanitize_surrogates.py
covers: paired surrogate reconstruction, lone surrogate replacement,
identity return on clean input (zero allocation), deep recursion on
dict/list/tuple, provider _sanitize_empty_content integration, and
full utf-8 encodability of the sanitized request body.
- 14/14 new tests pass; full existing test module also green.
- Replayed 58 archived real session messages plus adversarial
lone-surrogate injection through the provider path with no encode
errors after the fix.
Impact
------
- No behaviour change for clean inputs (sanitize_surrogates_deep is
an identity return when no surrogate is present).
- Fails-safe: unpaired surrogates degrade to U+FFFD instead of
aborting the entire request.
Replace the legacy long-goal skill contract with command-scoped goal tools and runtime guidance. Keep goal state durable across continuations while restricting create and replace mutations to explicit user /goal turns.
_get_copilot_access_token had a check-then-act race: concurrent chat()
calls after token expiry both fetched new tokens and clobbered each other.
Add asyncio.Lock with double-checked locking so only one fetch happens
per expiry window.
Closes#4677
Combine malformed tool-call handling with placeholder filtering and a
no-tools fallback so a relay that returns tool_use blocks with null
id/name/input can no longer crash a turn or permanently wedge a session.
Adapted to the ContextGovernor architecture (context governance now lives
in nanobot/agent/context_governance.py, not runner.py):
- ToolCallRequest.has_valid_name(): single source of truth for "usable
name" (non-empty string).
- tool_hints.format_tool_hints(): skip tool calls with a non-string/empty
name instead of raising AttributeError on the whole turn.
- ContextGovernor.strip_placeholder_assistant_messages() and
strip_malformed_tool_calls() (plus the _tool_call_name_is_valid helper):
history-cleaning staticmethods invoked at the START of
prepare_for_model() — strip_placeholder, then strip_malformed, then the
existing drop_orphan/backfill chain. Both only repair the model-facing
copy and leave persisted history untouched (return a copy, or the same
list when nothing changes). Also wired into runner's minimal-repair path.
- AgentRunner._drop_malformed_tool_calls(): returns
(dropped, all_dropped, original_finish_reason); clears finish_reason to
"stop" when all calls are dropped.
- AgentRunner._malformed_tool_call_retry_messages() + _request_model
malformed_retry flag: when an all-dropped tool_calls response comes back,
retry once with a corrective note; if the retry STILL comes back
all-dropped, fall back to _request_no_tools for graceful text degradation.
Tests for the history-cleaning methods live with ContextGovernor in
tests/agent/test_runner_governance.py; response-layer and tool-hint tests
stay on AgentRunner / tool_hints.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
_assistant_blocks appends dict items from content lists directly
without checking for the required 'type' field. A block like
{'text': 'hi'} reaches the Anthropic payload without a 'type',
causing a 400 rejection.
Add the same missing-type check that _convert_user_content already has,
so bare dicts in assistant content lists are coerced to text blocks
instead of triggering API validation errors.
Co-authored-by: nanobot-issues <issues@nanobot.dev>
Duplicate tool call ID normalization exists in the streaming parser path
but is not shared with the non-stream parser. Non-stream parsing appends
raw provider IDs into ToolCallRequest objects without deduplication.
Some OpenAI-compatible providers reuse the same tool_call_id for parallel
tool calls in non-streaming responses. Without dedup, runner executes
both tools with the same ID, producing duplicate tool results with the
same tool_call_id, which can fail strict provider validation.
Add the same _seen_tc_ids dedup pattern used in _parse_chunks to the
_parse method so both paths handle duplicate IDs consistently.
ProviderConfig.thinking_style defaults to None (Optional field), but
create_dynamic_spec expects a string. Coalesce None to "" at all call
sites (factory.py, settings_api.py) and fix the test assertion to
expect None from the config default.
Add a dedicated provider entry for the Kimi Coding Plan endpoint
(api.kimi.com/coding) using the Anthropic Messages API transport.
- Register kimi_coding with backend=anthropic
- Use KIMI_CODING_API_KEY env key to avoid clashing with MOONSHOT_API_KEY
- Default api_base set to https://api.kimi.com/coding/v1 so that
AnthropicProvider._normalize_base_url() + the SDK produce the correct
/coding/v1/messages request path
- Keywords include kimi-coding, kimi_coding and kimi-for-coding
ClosesHKUDS/nanobot#4463
Maintainer edit: document OpenCode Zen and Go configuration, keep their registry entries with gateway providers, and add focused provider registration tests.
maintainer edit: move duplicate tool_use history repair out of AgentRunner and into Anthropic message conversion, reusing the OpenAI-compatible queue-mapping approach locally without broadening the shared runner path.
maintainer edit: remap duplicate tool_use/tool_call ids instead of dropping later calls, so Anthropic-compatible providers that reuse ids for distinct parallel tool calls keep all requested work while still sending unique ids.
Anthropic rejects any request where two tool_use blocks share an id
("messages.N.content.M: tool_use ids must be unique"). A mis-assembled
stream could surface the same tool_use block twice in one assistant turn;
the runner persisted it verbatim, so the malformed message was re-sent on
every subsequent turn and permanently bricked the session — the agent
silently stopped replying.
Fix at two layers:
- AnthropicProvider._parse_response: drop duplicate tool_use ids (keep
first) as the response enters nanobot, so corruption is never persisted.
- AgentRunner._dedup_tool_calls: a new context-governance pass that dedupes
assistant tool_calls and tool results by id before each send, healing any
history that was already corrupted.
Add regression tests covering both the dedup and the no-op fast path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The image-strip fallback (triggered when a model errors on image input)
replaced image_url blocks with [image: <path>] or [image omitted]. Both
read like a live, available image to the LLM, causing it to:
1. hallucinate about image contents it never received
2. attempt read_file on the leaked server path
3. expose internal file paths to the model
Replace with an explicit '[Image not delivered to model — do not describe
or reference it]' placeholder that tells the LLM the image was stripped.
Fixes#4345
When the primary model (e.g. DeepSeek during peak hours) returns an empty
choices response with HTTP 200, the error carries no status code or
structured error metadata. The existing _FALLBACK_ERROR_TOKENS had no
matching token, so _should_fallback() returned False and fallback models
were never tried.
Changes:
- Add 'empty' token to _FALLBACK_ERROR_TOKENS so 'Error: API returned
empty choices.' text matches the fallback path
- Set error_kind='empty' in openai_compat_provider when returning
the empty-choices error, making the classification explicit
- Add test coverage for both text-only and error_kind matching paths
Fixes: glebov reported primary never falls back when DeepSeek returns
empty responses
Mistral's API constrains reasoning_effort to "high"/"none", rejects the
kwarg entirely for Magistral (reasoning is implicit), returns assistant
content as a mixed array of {type:"thinking",...}/{type:"text",...}
blocks, and 400s on the reasoning_content key in history.
- Remap user-supplied reasoning_effort (low/medium/minimal) onto Mistral's
two-tier vocabulary; strip the kwarg for Magistral models
- Lift thinking blocks into reasoning_content for both batch and streaming
responses; pass only text through on_content_delta callbacks
- Drop reasoning_content from outbound history when the spec asks for it
- Expose per-preset reasoning_effort_values so the UI can render the
provider-specific option set
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
The Anthropic Messages API rejects tool IDs that don't match
`^[a-zA-Z0-9_-]+$` with a 400 error. Tool IDs originating from other
providers or restored multi-turn sessions can contain invalid
characters (pipes, dots). Add a deterministic _sanitize_tool_id() and
apply it to both the tool_use id and the matching tool_result
tool_use_id so the pair stays consistent.
chengyongru reviewed #4367 and identified that the cloud branch
created a bare httpx.AsyncClient that lacked the SDK's default settings
(follow_redirects, connection pool limits). Since the SDK's
DefaultAsyncHttpxClient already has trust_env=True and proper defaults,
the simplest fix is to let http_client stay None for cloud endpoints.
Also updated the test to match the new behavior (http_client is None).
When the host has HTTP_PROXY / HTTPS_PROXY / ALL_PROXY set, httpx routes
all traffic through the proxy — including requests to localhost or LAN
addresses that the proxy typically cannot reach. This breaks local model
servers (Ollama, llama.cpp, vLLM) silently.
- Local endpoints: pass transport=httpx.AsyncHTTPTransport(proxy=None)
so proxy env vars are ignored for local traffic.
- Cloud endpoints: pass trust_env=True so corporate/VPN proxies work
without explicit configuration.
Fixes#4366
chengyongru
seteiro
yrk
Kris Lu
chengyongru
bingqilinweimaotai
chengyongru
Xubin Ren
hamb1y
axelray-dev
04cb
wangjunweiandClaude Opus 4.8
NanoBot
zpljd258
Teddy YanandClaude Opus 4.8
sbyinin
michaelxer
nanobot-contributor
Javad ArjmandiandClaude Sonnet 4.6
comadreja