When requests arrive via OpenAI-compatible endpoints and are converted to
Gemini format, functionCall parts were missing thoughtSignature. Gemini 3+
rejects such requests with 400 when thinking mode is enabled.
Changes:
- Extract thoughtSignature from tool_calls provider_specific_fields
- Inject into corresponding functionCall parts
- Fall back to dummy sentinel when thinking is enabled but no signature available
- Split signed/unsigned parts into separate model messages (Gemini requirement)
All existing tests pass (59/59).
weijiafu14
8bfdbda251