[BYOK Agent Mode] Gemini 3.1 Pro returns 400 "Function call is missing a thought_signature" — OpenAI-compatible client strips required field

[john-raymon]

[BYOK Agent Mode] Gemini 3.1 Pro returns 400 "Function call is missing a thought_signature" — OpenAI-compatible client strips required field

Environment

• Version of VS Code: VS Code Insiders (latest)
• Version of copilot-chat extension: 0.38.2026022002
• Operating System: macOS (or Windows/Linux)
• LLM Model: models/gemini-3.1-pro-preview-customtools via https://generativelanguage.googleapis.com/v1beta/openai/

Steps to reproduce

1. Set up GitHub Copilot BYOK in VS Code Insiders (since BYOK is currently only available there) using the Google Gemini OpenAI-compatible endpoint (https://generativelanguage.googleapis.com/v1beta/openai/).
2. Select the models/gemini-3.1-pro-preview-customtools model.
3. Try to use Agent mode (tool/function calling) in Copilot Chat.
4. It breaks on the second turn of the conversation.

Actual behavior

You immediately get hit with this error:

400 INVALID_ARGUMENT: Function call is missing a thought_signature in functionCall parts.

Root cause analysis

Gemini 3.1 does a "thinking" step before it generates tool calls. Google cryptographically signs this reasoning and attaches the signature to any tool call the model returns, stuffing it into a non-standard field:
choices[0].message.tool_calls[N].extra_content.google.thought_signature

Since VS Code's Copilot extension is a standard OpenAI client, it just drops this non-standard extra_content field entirely. On the next turn, when VS Code sends the tool result back to Google, the conversation history is missing the signature. Google sees the missing signature and rejects the request with a 400.

Expected behavior

Either VS Code should pass through unknown fields on assistant messages without stripping them, OR Google shouldn't strictly require the signature from clients hitting their OpenAI-compatible endpoint.

Workaround

I threw together a quick local Node proxy that intercepts the requests and injects Google's documented bypass sentinel (skip_thought_signature_validator) into the tool calls before forwarding them to Google.

You can run it via npx:

npx gemini-thought-signature-proxy

Then just update your chatLanguageModels.json to point to the proxy:

"url": "http://localhost:3000/v1beta/openai/"

You can find the npm package for the proxy here: gemini-thought-signature-proxy on npm

Additional note

While debugging this, I noticed VS Code constructs the final URL by appending v1/chat/completions to the base URL in chatLanguageModels.json. So with a base of http://localhost:3000/v1beta/openai/, VS Code sends requests to /v1beta/openai/v1/chat/completions. The actual Google endpoint is /v1beta/openai/chat/completions (no extra /v1). Just a secondary routing quirk worth mentioning.