[oyi77]

Summary

Re-submits the thoughtSignature bypass fix as a focused 6-file PR per @diegosouzapw's review on #3514. The previous PR was 130+ files (with ~42 i18n CHANGELOG dumps and unrelated incidental edits from a stale rebase). This PR is a clean 6-file change off current release/v3.8.20.

What's included (6 files, +87 / -103)

Source (3 files)

• open-sse/translator/request/openai-to-gemini.ts — scoped bypass injection: only openaiToGeminiCLIRequest (which serves CLI + Antigravity) sets supportsSignatureBypass: true; standard Gemini (openaiToGeminiRequest) leaves it undefined. Also removed the redundant toolCallIds.push (already at loop end).
• open-sse/executors/antigravity.ts — minor 4-line alignment
• open-sse/utils/error.ts — Gemini error parsing fix

Tests (3 files)

• tests/unit/t43-gemini-tool-call-no-thought-signature.test.ts — verify standard Gemini does NOT inject fake signature
• tests/unit/t44-antigravity-thought-signature-preserved.test.ts — verify Antigravity/CLI path preserves the bypass
• tests/unit/translator-openai-to-gemini.test.ts — update legacy evasion tests to reflect native bypass behavior

What's NOT included (removed from this PR per review)

• ❌ No ~42 docs/i18n/*/CHANGELOG.md dump
• ❌ No unrelated incidental edits from stale rebase (chatCore, requestLogger, streamHandler, plugin lifecycle hooks, package.json×3, etc.)
• ❌ No claude-to-gemini.ts (the original +1/-1 netted to 0)
• ❌ No open-sse/package.json bump (was carrying unrelated test-script noise)

Verification

• npm run check:any-budget:t11 → PASS
• prettier --check → all 6 files pass
• All 3 new tests pass (regression guards in place)

Original problem

Standard Gemini API (FORMATS.GEMINI) does not support thoughtSignature and rejects requests with it as 400. The bypass sentinel skip_thought_signature_validator was being unconditionally injected by openaiToGeminiBase, which is shared between standard Gemini, Gemini CLI, and Antigravity.

Fix: add supportsSignatureBypass option to GeminiToolNameOptions and gate it at the call site.

Refs the original review from @gemini-code-assist on PR #3414 and PR #3514.

[gemini-code-assist]

Code Review

This pull request refactors the handling of signature-less tool calls for Gemini and Antigravity models, replacing the legacy text-context evasion logic with a native bypass sentinel (skip_thought_signature_validator) to prevent 400 errors. It also updates upstream error parsing to handle array responses and adapts the test suite to verify native function calls and responses. The review feedback points out that the refactoring leaves behind several unused variables and redundant checks (such as stringifySignaturelessToolCalls and contextualizeSignaturelessToolResponses) that should be cleaned up, and identifies a test name that has become misleading and needs renaming.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

With the removal of the signatureless tool call evasion/text-context logic, the variables stringifySignaturelessToolCalls and contextualizeSignaturelessToolResponses (defined earlier in this function) are now unused/dead code. Additionally, hasSignaturelessTextResponses is redundant because toolCallIds now unconditionally contains all tool call IDs, making hasActualResponses equivalent to just checking toolCallIds.some((fid) => toolResponses[fid]). These unused variables and redundant checks should be cleaned up to maintain code readability and avoid dead code.

[gemini-code-assist]

Since the legacy XML escaping assertions have been removed and replaced with a check for native functionResponse emission, the test name at line 817 ('OpenAI -> Antigravity escapes signature-less tool response context content') is now misleading. It should be renamed to accurately describe the new test behavior, e.g., 'OpenAI -> Antigravity emits native functionResponse instead of escaped XML context for signature-less tool responses'.

[diegosouzapw]

Thanks @oyi77 — this is exactly the clean, scoped resubmission I was hoping for, and it addresses the #3414 review precisely:

• ✅ Bypass is now correctly gated to Antigravity/CLI via supportsSignatureBypass (only set in openaiToGeminiCLIRequest), so standard Gemini is untouched.
• ✅ T43 correctly asserts undefined again on the standard path (no more inverted contract).
• ✅ Client-provided signatures take precedence over the sentinel.
• ✅ The error.ts array-unwrap fix (originally from @Six7Day's fix(translator): implement thought_signature bypass and fix Gemini er… #3414) is preserved.

I pushed one follow-up so it's fully safe: the removal of the text/context signature-less representation was un-gated, which would have changed the standard Gemini direct path (FORMATS.GEMINI, mode "text" — it rejects both thoughtSignature and signature-less native tool parts, the reason #3358 added the text representation). I gated that removal behind !supportsSignatureBypass so the Antigravity/CLI path keeps your new native+sentinel behavior while standard Gemini keeps its current behavior, re-wired the now-no-longer-orphaned buildInert*/buildHistorical* helpers, and added a regression test for the text-mode path. Full translator suite + T43 + T44 + new test = 41/41 green, typecheck/lint clean.

Credit to @Six7Day for the original fix (#3414) — your commits are preserved in this branch's history. Merging now. 🙌