* fix: exit MCP anyio contexts in the lifecycle task on disable
#9070 moved MCP shutdown cleanup into the per-server lifecycle task so
that the anyio cancel scopes entered in connect_to_server() are exited
from the task that entered them. Running the cleanup through
asyncio.shield() defeats that: shield wraps the coroutine in a new
task, so disabling a server still fails with "Attempted to exit cancel
scope in a different task than it was entered in" and leaves the scope
state corrupted -- the failure mode behind the CPU spin reported in
#9068.
Await _terminate_mcp_client() directly instead. The graceful disable
path has no cancellation in flight, and on forced shutdown the
lifecycle task still runs its finally block in-task, so the shield only
served to break task affinity.
Closes#9068
* fix: absorb late cancellations during MCP shutdown cleanup
A cancellation delivered while the lifecycle task is already running its
finally-block cleanup would abort _terminate_mcp_client() halfway,
stranding the runtime entry and the MCP transport. Retry the cleanup in
the same task and uncancel() the absorbed request instead, so a forced
shutdown can no longer skip it.
* fix: guard Task.uncancel() for Python 3.10 runtimes
Task.uncancel() only exists on Python 3.11+. On 3.10 absorbing the
CancelledError is sufficient, so skip the call when unavailable.
* fix: adapt MiMo STT to V2.5 models and reject non-WAV audio
The MiMo-V2 series went offline on 2026-06-30, so the default STT model
mimo-v2-omni fails for every default configuration. Switch the default
to mimo-v2.5-asr, the dedicated speech recognition model whose official
docs use exactly the bare input_audio payload this provider sends.
For non-ASR multimodal models such as mimo-v2.5, the audio understanding
docs require a text instruction alongside the audio, so restore the
system/user transcription prompts for that model family only.
Also validate that the resolved audio payload really is RIFF/WAVE before
calling the API: when a platform voice file (e.g. Tencent SILK from QQ)
slips through the WAV conversion chain unchanged, fail locally with an
actionable error instead of the opaque HTTP 400 from the API.
Fixes#9113
* fix: accept unpadded MiMo wav headers
---------
Co-authored-by: tangtaizong666 <212687958+tangtaizong666@users.noreply.github.com>
* fix: preserve assistant messages with reasoning_content in sanitize pass
When _sanitize_assistant_messages encounters an assistant message with empty content and no tool_calls but with reasoning_content, keep it with content set to empty string instead of dropping it. Reasoning models (DeepSeek V4, MiMo, etc.) require this history for subsequent turn validation.
* Update astrbot/core/provider/sources/openai_source.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* fix: default to empty TokenUsage when completion.usage is None
When completion.usage is None (e.g. certain proxy/streaming edge cases), llm_response.usage stayed unset (None). Plugins accessing .input_tokens on it would crash with AttributeError.
Always assign llm_response.usage — extract from completion.usage if present, otherwise fall back to a zeroed TokenUsage().
Closes#8605
---------
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: renchonghan <renchonghan@users.noreply.github.com>
* sanitize orphaned tool_result blocks in Anthropic provider
Sanitize tool_result blocks and merge consecutive messages with the same role to comply with Anthropic API requirements.
* Fix content handling in message merging logic
* fix: sanitize anthropic assistant messages
* fix: validate anthropic tool result ordering
---------
Co-authored-by: Soulter <905617992@qq.com>
* fix: tool definition does not pass back to gemini properly, causing repeated tool calls.
fixes: #8789fixes: #8773fixes: #7111fixes: #6402fixes: #7684
* fix: remove unnecessary logging and improve log messages in Gemini source
* fix: improve type checking for tool calls in Google Gemini provider
* fix: unify media reference handling
* fix: accept bare base64 record media refs
* chore: update agents.md
* fix: unify file URI handling across media components and utilities
* fix: unify media reference type handling with MediaRefStr alias
* Potential fix for pull request finding 'CodeQL / Incomplete URL substring sanitization'
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
* Update astrbot/core/platform/sources/discord/discord_platform_adapter.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* fix: unify media handling and improve base64 decoding across components
* fix: simplify client_kwargs type definition and enhance media message handling in platform adapter
* fix: unify media utility documentation and enhance function descriptions
* perf: drop "pilk" requirement, improve audio outbound for tencent-related IM apps which using silk
* fix: unify Tencent Silk audio handling and enhance media resolver functionality
---
- Centralize media reference materialization and base64 resolution for local paths, http(s), base64://, data URIs, and legacy bare base64 payloads.
- Normalize incoming Record audio to wav and Image media to temporary jpg during preprocess, with event-scoped cleanup.
- Reuse the shared media resolver across OpenAI, Gemini, Anthropic, MiMo, DeerFlow, STT, and platform media paths while sanitizing logs and cleaning temporary conversion outputs.
- Ensure generated TTS audio is tracked for cleanup after the event finishes.
fix#8676fix#8543fix#7588fix#7580fix#8030fix#8034fix#7461fix#7565fix#6509fix#7144fix#7795
---------
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* fix(context): restore turn cap, serialize content parts and tool calls for llm compress, fix AftCompact debug log
Three context-compaction regression fixes after #8226:
1. Restore max_context_length -> enforce_max_turns propagation so
normal turn-based truncation works again.
2. Serialize ContentPart and ToolCall objects into plain dicts in
_message_to_dict so llm_compress no longer fails with JSON
serialization errors.
3. Print _provider_messages (compacted) instead of run_context.messages
(unchanged) in AftCompact debug log; truncate long role lists to
first4,...,last4 to avoid log spam.
Assertions in tests are also hardened to avoid coupling to exact prompt
wording.
* fix(tool_loop_agent_runner): simplify context handling by removing redundant provider messages
* fix(tool_loop_agent_runner): rename context manager variables for clarity
* fix: update context compression to use recent token ratio instead of fixed count
* fix: enhance LLMSummaryCompressor to sanitize contexts and improve message handling
* ruff format
---------
Co-authored-by: Soulter <905617992@qq.com>
* fix(provider): wrap batch embedding texts in Content to avoid collapse on gemini-embedding-2
* fix(gemini_embedding): format list comprehension for better readability
---------
Co-authored-by: Rat0323 <Rat0323@users.noreply.github.com>
Co-authored-by: Soulter <905617992@qq.com>
Set MiniMax-M3 as the default fallback model for the Token Plan provider.
The model list itself is already fetched dynamically from the MiniMax API,
so all available models including M3 are auto-discovered. This change just
updates the hardcoded fallback used when no model is configured to the
current flagship.
MiniMax-M2.7 remains fully usable; users can still configure it explicitly.
* fix: handle delta=None chunks in streaming to prevent SDK to_dict() error
When certain OpenAI-compatible providers (Gemini, DeepSeek, some proxies)
return chunks with choice.delta=None (e.g. ContentBlockDeltaEvent),
ChatCompletionStreamState._convert_initial_chunk_into_snapshot internally
calls choice.delta.to_dict() at line 747, causing:
'NoneType' object has no attribute 'to_dict'
Fix:
1. Skip handle_chunk when delta is None (delta=None chunks have no
content contribution anyway)
2. Wrap get_final_completion in try/except to gracefully fall back to
empty ChatCompletion if SDK state is corrupted
Refs: openai-python#5069, openai-python#5047
* fix: resolve bugs found by Sourcery and gemini-code-assist review
- Remove orphan logger.error that caused NameError on every chunk
- Replace broken empty ChatCompletion fallback with clean return;
streamed content already yielded, no data loss
Co-authored-by: sourcery-ai[bot] <sourcery-ai[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <gemini-code-assist[bot]@users.noreply.github.com>
* fix: properly replace get_final_completion fallback in _query_stream
Previous fix_pr_v3 wrongly injected code into terminate() instead.
Now correctly:
1. Replace empty ChatCompletion fallback with clean return in _query_stream
2. Revert terminate() to original (await self.client.close() only)
* fix: revert corrupted terminate() to original
Previous fix_pr_v3 injected wrong-indentation code into terminate().
---------
Co-authored-by: sourcery-ai[bot] <sourcery-ai[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <gemini-code-assist[bot]@users.noreply.github.com>
