fix: preserve Gemini thought_signature in tool call messages

[0xbyt4]

Summary

• Gemini 3 thinking models (Flash/Pro) attach extra_content.google.thought_signature to function call responses. This signature must be echoed back on subsequent API calls or the server rejects with a 400 error.
• _build_assistant_message was only extracting id, type, and function from tool calls, dropping the extra_content field entirely. This caused all Gemini 3 tool-calling flows to fail after the first function call.
• Fix preserves extra_content (with Pydantic model_dump safety) so it round-trips through the conversation history. Non-thinking models are unaffected -- the field is only included when present.

Changes

• run_agent.py: Preserve extra_content from tool call objects in _build_assistant_message
• tests/test_run_agent.py: Add 2 tests verifying extra_content is preserved when present and omitted when absent

[teknium1]

Which provider are we talking about here?

We pushed a commit earlier to preserve all OpenRouter & Nous Portal reasoning - and one that makes it so our Codex Provider works like that too.

But openrouter doesn't work like this (neither does nous portal) - so not sure where this fix applies

[0xbyt4]

Which provider are we talking about here?

We pushed a commit earlier to preserve all OpenRouter & Nous Portal reasoning - and one that makes it so our Codex Provider works like that too.

But openrouter doesn't work like this (neither does nous portal) - so not sure where this fix applies

This fix is for direct Google AI Studio usage via the OpenAI-compatible endpoint
(OPENAI_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/), not OpenRouter or Nous Portal. Gemini 3 thinking models (Flash/Pro) attach
extra_content.google.thought_signature to function calls through this endpoint, and reject subsequent requests with a 400 error if it's not echoed back.
The fix simply preserves extra fields on tool_call objects it's a no-op for providers that don't send them (OpenRouter, Nous, OpenAI etc.).