feat(google): preserve thought signatures for Gemini 3+ function calling

[tomc98]

Summary

Gemini 3+ models require thought signatures to be preserved and returned when using function calling with thinking mode enabled. Without this fix, the API returns a 400 error:

Function call is missing a thought_signature in functionCall parts. 
This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance.

Changes

• Add thoughtSignature optional field to FunctionCall class
• Capture thoughtSignature from Gemini API response in Google plugin's #parsePart method
• Include thoughtSignature when converting chat context back to Google format
• Update isEquivalent to compare thoughtSignature for function calls

Models affected

This fix enables compatibility with:

• gemini-3-flash-preview
• gemini-3-flash-lite-preview
• gemini-3-pro-preview

References

• Thought Signatures Documentation
• Similar issue in langchain-google

Test plan

• Build passes
• Existing tests pass (failures are due to missing API keys, not code changes)
• Lint passes

[changeset-bot]

🦋 Changeset detected

Latest commit: 3043ebf

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@livekit/agents-plugin-google	Minor

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[tomc98]

I also made an issue here #920

[tomc98]

Wait, I've got a few issues, do not merge

[tomc98]

Ok ready for review :D

[tomc98]

@toubatbrian Could you please review, not having this is blocking me

[tomc98]

Also do I need a changeset?

[toubatbrian]

LG to me!

[toubatbrian]

Hey @tomc98 thanks for the PR! I'm currently OOO and review might be slower. Please add a patch changeset on google plugin, thanks!

Also, could you run the agent with current changes using gemini model and see if the voice interaction works as usual? Just to make sure no regression is introduced.

[tomc98]

Tested for realtime regressions using examples/src/gemini_realtime_agent.ts - voice interaction works as expected with the current changes.

Note: I had to update the model to models/gemini-2.5-flash-native-audio-preview-12-2025 as the default model (gemini-2.0-flash-live-001) was not working. The default model in the realtime plugin should probably be updated in a separate PR.

[tomc98]

@toubatbrian It's ready, finished everything you needed 👍