fix(openai): keep Gemini 3 thought signatures server-side, keyed by tool call id

[ItszFinn]

Problem

Gemini 3 rejects a functionCall that comes back in history without the thoughtSignature it was issued with:

HTTP 400 INVALID_ARGUMENT: Function call is missing a thought signature.

The proxy already hands the signature to the client as extra_content.google.thought_signature, but ordinary OpenAI clients (OpenAI SDKs, litellm, most agent frameworks) don't echo unknown fields back. So the second request of every tool round trip loses it.

#26 fixed the 400 by sending skip_thought_signature_validator when the signature is missing. That works, but the model gets a placeholder instead of the signature it actually produced.

Change

• Store: internal/api/thoughtsig.go adds a small in-memory store keyed by tool call id. It's a FIFO capped at 1024 entries.
• Fill: both response paths (streaming and non-streaming) put the signature of every emitted tool call into the store.
• Restore: chatMessageContent puts the stored signature back when an assistant tool call comes back without extra_content.
• Precedence:
  • A client that does echo extra_content still wins.
  • A call this process never emitted (after a restart, or from another instance) still falls back to the fix(aistudio): 兼容 JSON Schema const 字段并在历史函数调用缺少签名时填充 bypass 标记 #26 placeholder.

Tests

• internal/api/thoughtsig_test.go checks two things: a stored signature is restored when the client omits it and a client-supplied one takes precedence, and the store stays bounded.
• go vet ./internal/api/ and go test ./internal/api/ ./internal/aistudio/ pass.
• The same patch ran in my deployment (litellm in front) from 2026-09-12 on the pre-fix(aistudio): 兼容 JSON Schema const 字段并在历史函数调用缺少签名时填充 bypass 标记 #26 base. Without it, every multi-turn tool round trip failed with the error above.
• I rebased it onto current main for this PR. I haven't re-run it live on the rebased base yet because my AI Studio sessions have expired.

Disclosure

This change was written with the help of an AI assistant (Claude). I reviewed it and ran the tests above.

🤖 Generated with Claude Code

[ItszFinn]

Live-verified on the rebased branch (current main + this patch), with litellm in front, which does not echo extra_content: our conformance suite passes 7/7 for gemini-3.8-flash (chat, system, stream, tool calls non-stream and SSE, tool-result round trip, parallel tool calls).

(Written with the help of an AI assistant.)

[Mag1cFall]

Thanks! We reproduced it on main, tested the round trip after merging (streaming and non-streaming), and merged it. We also fixed a few more things in a follow-up commit, mainly keying the store by call ID, function name and arguments, since upstream call IDs are short enough to collide: aec8a47

Thanks again for the contribution~