[ZhuLinsen]

Change Summary

Gemini 3 tool loops stop after the first tool result with 400 Function call is missing a thought_signature. On Gemini's OpenAI-compatible endpoint the signature comes back on each tool call as extra_content.google.thought_signature. Google requires it to be sent back, and ms-agent dropped it.

This PR keeps extra_content on ToolCall:

• Both OpenAI-compatible clients, transport/openai_compat.py and the legacy openai_llm.py, read it in the streaming and non-streaming parsers.
• Stream merging keeps it whichever chunk carries it.
• Message.to_dict_clean puts it back on the wire tool call.

The field is written only when the provider returned it, so requests to other OpenAI-compatible providers are unchanged. Session logs already keep it, because ToolCall is a plain dict.

Known limits, left out to keep this focused:

• If a session with Gemini tool calls is replayed against a different provider, the old calls still carry extra_content.
• The stream mergers only extend the last tool call. A signature-only delta for an earlier call that arrives after a later call has started would still be misplaced. This limit already applies to arguments.
• ResponseAdapter.to_message does not carry the field. It has no caller in the repo, and the agent loop does not use it.

The reasoning_content point in the issue comments is a separate change and is not included here.

Related issue number

fix #1040

Checklist

• The pull request title is a good summary of the changes - it will be used in the changelog
• Unit tests for the changes exist
• Run pre-commit install and pre-commit run --all-files before git commit, and passed lint check.
• Documentation reflects the changes where applicable

Tests

New tests/llm/test_tool_call_extra_content.py uses real openai ChatCompletion and ChatCompletionChunk objects with extra_content. It covers both clients for:

• a non-streaming round trip;
• streaming, with the signature arriving on different chunks of two parallel calls;
• providers that send no field, which must get no field back;
• the assistant turn sent in the next request.

Without the fix, 3 of the 4 tests fail with KeyError: 'extra_content'. The test that checks no field is added passes either way.

pytest tests/llm/test_tool_call_extra_content.py tests/llm/test_toolcall_empty_content.py -q
11 passed

tests/llm/test_message_text.py, test_continuation_guard.py and test_thinking_fallback.py also pass. pre-commit run --files was run on the changed files. yapf also wanted to reformat three unrelated existing lines in the touched files, and I left those out of the diff. I tested against the parsing code only, not against the live Gemini API.