fix: resolve 400 error due to missing thought_signature in Gemini 3 Pro tool calls

[kfxmvp]

TLDR

This PR fixes a critical bug where using tools (function calls) with Gemini 3 Pro Preview resulted in an API Error 400. The error was caused by the missing thought_signature, which is required by the Gemini API for tool verification in this model.

Dive Deeper

• Issue: As reported in [Bug] API Error 400 with Gemini 3 Pro: Missing thought_signature for function call #1187, Gemini 3 Pro Preview failed to execute tool calls, returning Unable to submit request because function call ... is missing a thought_signature.
• Fix: Implemented logic in OpenAIContentConverter to extract the thoughtSignature from the streaming response chunks and attach it to the corresponding functionCall object. This ensures the signature is passed back to the API in subsequent requests, satisfying the verification requirement.
• Cache Mechanism: The implementation uses index as the key to cache thoughtSignature. This approach ensures robustness because id might not be present in every chunk during streaming or in certain model responses, whereas index is a reliable identifier for correlating chunks within the stream.

Reviewer Test Plan

1. Setup: Configure the CLI to use gemini-3-pro-preview.
2. Reproduction: Execute a prompt that requires a tool call (e.g., read_file README.md).
3. Verification: Confirm that the tool executes successfully and the previous API Error 400 is resolved.
4. Regression Test: Ensure that other models (e.g., gemini-1.5-pro, gpt-4o) still function correctly with tool calls.

Testing Matrix

	🍏	🪟	🐧
npm run	✅	❓	❓
npx	❓	❓	❓
Docker	❓	❓	❓
Podman	-	-	-
Seatbelt	❓	-	-

Linked issues / bugs

Resolves #1187

[pomelo-nwu]

@kfxmvp Appreciate it! We’ll need a bit to check it out.