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>
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
The temperature suppression was hardcoded to only match opus-4-7. Newer
Anthropic models (opus-4-8, fable) also reject the parameter with a 400.
Normalize model_name to lowercase before matching so mixed-case configs
do not fall through. Add tests for opus-4-8 and fable across adaptive,
enabled, and no-thinking paths, plus a negative test confirming ordinary
models still send temperature.
Fixes#4333
- Register SiliconFlow in transcription registry with default model
FunAudioLLM/SenseVoiceSmall and alias 'silicon'
- Reuse existing OpenAITranscriptionProvider adapter (Whisper-compatible)
- Add generic key/base resolution: fallback to registry env_key and
default_api_base when provider config is absent
- Add tests for registry entry, alias, adapter, default model, and
config resolution with env var fallback
maintainer edit: streamed timeout recovery was returning the retried response internally while the channel still treated the final outbound as already streamed. End the current stream segment before retry/fallback recovery so subsequent deltas are delivered in a new segment.
When a stream stalls mid-response, both the retry layer and
FallbackProvider blocked recovery because content had already been
emitted via on_content_delta. This left users with truncated replies
and no automatic recovery.
For error_kind="timeout" specifically:
- _run_with_retry now suppresses delta callbacks and retries the same
model instead of returning immediately
- FallbackProvider now allows failover to a different model with
delta callbacks suppressed
Non-timeout errors retain the original "skip retry/failover after
streamed content" behavior to avoid duplicate output.
- Add StepFunTranscriptionProvider class in nanobot/providers/transcription.py
- New _post_stepfun_asr_with_retry() function handling SSE stream parsing
(transcript.text.delta → transcript.text.done event sequence)
- Register 'stepfun' in transcription_registry.py with default model stepaudio-2.5-asr
- Reuse existing stepfun provider config (apiBase can point to Plan endpoint)
- Add 17 tests covering SSE parsing, retry contract, empty-text edge case, and registry integration
- Update docs/configuration.md with stepfun ASR documentation
StepFun ASR uses a dedicated SSE endpoint (/v1/audio/asr/sse) rather
than the chat-completions or Whisper multipart formats used by other
providers. Users on Step Plan can set apiBase to the Plan endpoint.
Maintainer edit: keep the GPT-5/o-series fallback on slug-boundary matching so unrelated model names are not caught by substring checks, and include o1 alongside o3/o4 because it is also an o-series chat model.
Add AssemblyAI as a third transcription provider option alongside
OpenAI and Groq. AssemblyAI offers better accuracy for certain
audio types (distant voices, noisy environments) and serves as a
reliable fallback when other providers struggle.
Changes:
- Add AssemblyAITranscriptionProvider class in providers/transcription.py
- Add 'assemblyai' option in base channel's transcribe_audio()
- Per-channel configuration via transcriptionProvider in config
Usage:
Set transcriptionProvider: 'assemblyai' and provide an AssemblyAI
API key via transcriptionApiKey in the channel config.
Add support for Xiaomi MiMo ASR as a third transcription backend alongside
Groq and OpenAI Whisper. Xiaomi ASR uses the /v1/chat/completions endpoint
with base64-encoded audio input, rather than the standard Whisper multipart
upload format.
Co-Authored-By:连 <lian@tangping.homes>
Add a `transcriptionModel` channel setting and an OpenRouter transcription
backend so voice messages can be transcribed through OpenRouter's
speech-to-text endpoint (e.g. nvidia/parakeet-tdt-0.6b-v3, openai/whisper-1),
alongside the existing Groq/OpenAI Whisper providers.
- schema: add channels.transcriptionModel (None = provider default)
- providers/transcription: extract a shared POST/retry skeleton; add a
JSON+base64 OpenRouterTranscriptionProvider; make the STT model a
constructor param on all providers instead of hardcoding it
- channels: route transcriptionProvider="openrouter" and thread the model
through the manager to each channel
- docs + tests
Only dedicated STT models work on OpenRouter's transcription endpoint;
chat LLMs (e.g. google/gemini-3.5-flash) are rejected there.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds ProviderConfig.extra_query, threaded into AsyncOpenAI(default_query)
so that Azure-style gateways requiring query params like api-version can
be configured without URL hacks.
Also updates provider_signature to track extra_query changes so per-turn
refresh rebuilds the provider when the value changes.
Addresses the extra_query portion of #4204. The max_completion_tokens
model-awareness enhancement is intentionally left separate.
maintainer edit: add SDK-object and tool-call history regressions so the empty-string reasoning_content fix is covered across both parse branches and the sanitized request path.
Custom providers (e.g. DeepSeek) may return reasoning_content as an
empty string "" to explicitly indicate no reasoning occurred. The
previous truthiness checks (, ) treated "" as falsy
and converted it to None, which caused the field to be dropped from
the message history entirely. Providers that require reasoning_content
on all assistant messages then rejected subsequent requests.
Replace truthiness checks with identity checks () so that
empty-string reasoning_content is preserved as-is. The streaming path
is unchanged since an empty join genuinely means no chunks received.
Fixes#4105
Maintainer edit: preserve provider-specific size hints for custom image generation endpoints while keeping the default 1K mapping compatible. Clarify the custom provider contract in docs and cover response_format/size overrides in tests.
Maintainer edit: require providers.custom.apiBase before making custom image requests and allow unauthenticated local endpoints by omitting Authorization when no apiKey is configured.
Every other streaming provider (anthropic, bedrock, openai_compat,
litellm) reads NANOBOT_STREAM_IDLE_TIMEOUT_S with a 90s default. The
Codex provider hardcoded 60s in _request_codex, so it could not be
tuned the same way and aborted streams sooner than its peers.
Read the same env var with the same default and pass it as the httpx
client timeout. The variable name and int parsing match anthropic /
openai_compat / bedrock verbatim.
#4009 normalized the error response when the timeout fires; this PR
fixes the timeout knob itself.
Ensure converted Responses API input items use unique replay ids when restoring assistant messages and function calls. This prevents Codex from rejecting resumed conversations with duplicate rs_* item ids while preserving call_id-based tool result linkage.
michaelxerandXubin Ren
yu-xin-candXubin Ren
qcypggsandXubin Ren
axelray-devandXubin Ren
moranandXubin Ren
chengyongruandXubin Ren
aiguozhi123456andXubin Ren
04cbandXubin Ren
chengyongruandGitHub
comadrejaandXubin Ren
chengyongruandXubin Ren
Kunal KarmakarandXubin Ren
hamb1yandXubin Ren
yeounhyeokandXubin Ren
EunHyunsuandXubin Ren
outlook84andXubin Ren
Yuxin LouandXubin Ren
Jiajun XieandXubin Ren
ZegWeandXubin Ren