[WRONG BRANCH] fix(google): prevent backward attachment of thought signatures in streaming parser

[luvs01]

Motivation

• Prevent a streaming-parser ordering bug where a later thought_signature in the same SSE frame could be attached backwards to an earlier unsigned functionCall, corrupting replay metadata and causing subsequent replay/serialization errors.

Description

• In src/adapters/google.ts the observer result is now captured to observedStreamThoughtSig and only published to pendingStreamThoughtSig after parts are emitted in source order, avoiding backward pairing.
• The change preserves cross-chunk carry and the existing observeAntigravityReplay behavior while ensuring per-frame fallback metadata cannot be applied to earlier parts.
• Added a regression test in tests/google-signature-history-roundtrip.test.ts that emits an SSE frame with an earlier unsigned call, a later thought part with thought_signature, and a following call, asserting only the following call receives the signature.

Testing

• git diff --check completed successfully.
• bun test tests/google-signature-history-roundtrip.test.ts tests/google-antigravity-replay.test.ts passed (focused Google adapter tests).
• bun run typecheck completed successfully.
• bun run privacy:scan completed successfully.
• A repository-wide bun run test was run and exposed unrelated pre-existing management-auth expectations failing (HTTP 403 vs expected statuses) that are not caused by this focused adapter change.

---

Codex Task

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 68bf6594-7c9a-46da-8fb7-aab8a524fd70

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

✅ Deterministic PR hygiene checks passed.

[github-actions]

⏳ DRAFT

• wrong target branch (main); retarget to dev.

What to do

• Retarget this PR to dev — all contributions go to dev.

Its title has been prefixed with [WRONG BRANCH].
This pull request was already a draft. Its draft status will be preserved after every issue above is resolved.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bcf3b528f8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve signature carry when replay observation is inactive

For streaming AI Studio requests, the replay observer is skipped by the googleMode guard, so observedStreamThoughtSig retains the value from before this frame. After the loop records a standalone thought signature in pendingStreamThoughtSig, this assignment immediately resets it; if the corresponding unsigned function call arrives in a later SSE frame, its tool_call_start lacks the required Google signature metadata and the subsequent tool-result turn can be rejected by Gemini. Keep the source-order value computed by the loop authoritative (and invoke the observer only for its replay-cache side effect), including for modes such as Claude-on-Antigravity where the observer returns without scanning.

AGENTS.md reference: src/AGENTS.md:L19-L19

Useful? React with 👍 / 👎.