fix(responses): strip Codex-private item metadata at the noncanonical boundary

[lidge-jun]

Summary

Carries #3066 (author @yanzhibo-bytedance) rebased onto current dev. Both commits are the author's and needed no changes.

Codex ≥0.151 attaches internal_chat_message_metadata_passthrough to Responses input[] items. OpenCodex occupies the built-in openai slot, so that private field reaches whatever destination the request is routed to, and a strict upstream rejects it as an unknown input[*] parameter. The adapter now removes it at the noncanonical boundary, copy-on-write so _rawBody stays caller-owned.

Supersedes #3038, which found the same defect. See below.

Why this one and not #3038

They differ in layer, and the layer is the whole decision.

#3038 strips inside core.ts and compact.ts after readJsonRequestBody, unconditionally — including on the canonical ChatGPT forward path, where the field is not foreign at all. isCanonicalOpenAiForwardProvider exists precisely to mark that destination (src/providers/openai-tiers.ts:34-37).

Its tests are also vacuous: tests/openai-internal-request-metadata.test.ts imports stripOpenAiInternalRequestMetadata and calls it directly. It never drives handleResponses, compact, or an adapter buildRequest, so deleting both production call sites leaves that file green.

#3066 strips inside createResponsesPassthroughAdapter().buildRequest, inside the existing if (!isCanonicalOpenAiForwardProvider(provider)) block, and its tests call that production buildRequest.

Verification

bun test tests/openai-responses-passthrough.test.ts  -> 117 pass / 0 fail / 372 expect()
bun x tsc --noEmit                                   -> exit 0

Two mutations, each restored:

mutation	result
remove the strip call site	116 pass / 1 fail — noncanonical Responses destinations strip Codex-private item metadata
move the call outside the noncanonical guard	116 pass / 1 fail — canonical ChatGPT forward preserves Codex-private item metadata

Two mutations, two different failures: stripping and ChatGPT preservation are guarded independently, which is exactly the property #3038's test layer cannot assert.

Known residual

Official openai-apikey native compact bypasses this adapter, so a strict-validator failure on that specific destination is not covered here. That destination is OpenAI-operated and outside the reported joy-openai / gpt-5.6-sol failure; it belongs in its own change if it ever reproduces.

Checklist

• Focused tests for the changed subsystem pass
• bun x tsc --noEmit clean
• Regression tests present and mutation-verified
• No docs-site change needed (wire-level request hygiene)

Triaged in the 2026-08-31 non-priority-70 bug round.

Summary by CodeRabbit

• Bug Fixes

  • Prevented private internal metadata from being sent to routed Responses API destinations.
  • Preserved required metadata when communicating with the canonical ChatGPT backend.
  • Ensured request data remains unchanged for callers while being safely filtered for forwarding.

• Tests

  • Added coverage confirming metadata filtering and preservation across supported destinations.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-31T18:59:11.993846Z	39712e4	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

✅ Deterministic PR hygiene checks passed.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 0786393c-d9ff-44ae-b2c5-4dff2a9a73ef

📥 Commits

Reviewing files that changed from the base of the PR and between 6123be3 and 39712e4.

📒 Files selected for processing (2)

• src/adapters/openai-responses.ts
• tests/openai-responses-passthrough.test.ts

Included review availability: Your plan provides up to 10 included reviews per hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

Changes

Responses metadata filtering

Layer / File(s)	Summary
Conditional metadata stripping and provider coverage src/adapters/openai-responses.ts, tests/openai-responses-passthrough.test.ts	The request builder removes internal_chat_message_metadata_passthrough from message and function-call items for non-canonical providers. The copy-on-write helper preserves the original body when no changes are needed. Tests verify input immutability, non-canonical stripping, and canonical preservation.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to 39712

The change removes private item metadata only for noncanonical Responses destinations while preserving canonical forwarding and caller-owned requests. No actionable merge-blocking risk remains after normal checks and review.

Suggested reviewers: ingwannu, olddonkey

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 33.33% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 3 functions across 2 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: removing Codex-private Responses item metadata at the noncanonical forwarding boundary.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

• Fix all pre-merge checks with AI

✨ Finishing Touches 💡 1

📝 Generate docstrings 💡

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/3066-strip-codex-private-metadata

---

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.}

[lidge-jun]

리뷰 · 우선순위 64 / 80

설명

이 PR은 기여자 @yanzhibo-bytedance 의 #3066을 지금 dev HEAD 6123be31f (#3103 forked-rollout restore) 위에 다시 올린 랜딩입니다. Codex 0.151부터 Responses input[] 항목에 internal_chat_message_metadata_passthrough 라는 ChatGPT 전용 필드가 붙습니다. OpenCodex는 히스토리 연속을 위해 루프백에서 공급자 이름을 openai 로 남겨 두기 때문에, 그 필드가 공개 Responses 게이트웨이까지 같이 갑니다. 그쪽은 input[*] 의 알 수 없는 파라미터라며 요청 전체를 거절합니다.

지금 체크아웃의 src/adapters/openai-responses.ts 를 보면, 비정규 목적지에만 쓰는 정리는 이미 isCanonicalOpenAiForwardProvider 가드로 나뉘어 있습니다. 이 함수는 src/providers/openai-tiers.ts 34–37행에서 adapter가 openai-responses이고 authMode가 forward이며 baseUrl이 Codex forward URL일 때만 참입니다. 2141행 근처에서 비정규일 때만 promoteClientLoadedTools 를 돌리고, 그 다음 블록에서 커스텀 툴·툴서치·네임스페이스 툴을 낮춥니다. 이 PR은 그 첫 비정규 블록 맨 앞에 stripInternalChatMessageMetadataPassthrough 를 넣습니다. 본문 객체를 직접 고치지 않고 input 항목만 복사한 뒤 필드를 지워서, 호출자가 넘긴 _rawBody 는 그대로 남습니다.

같은 결함을 먼저 잡은 #3038은 core.ts / compact.ts 에서 readJsonRequestBody 직후 무조건 지웁니다. 그러면 ChatGPT 정규 전달 경로의 필드까지 빠집니다. 그 경로는 이 필드가 외국 필드가 아닙니다. #3038 테스트는 헬퍼를 직접 호출해서 handleResponses 나 어댑터 buildRequest 를 한 번도 안 탑니다. 이 PR 테스트는 tests/openai-responses-passthrough.test.ts 에서 키 인증·forward 비정규 목적지 둘 다 필드가 빠지는지, 그리고 정규 ChatGPT forward에서는 남는지, 원본 rawBody 는 안 바뀌는지를 봅니다. types.ts/config.ts 분할 캠페인과는 무관합니다. #3066은 이 랜딩이 머지된 뒤에 leftover로 닫으면 됩니다.

round-2 prio≥70 남은 축은 #3029 / #3008 / #3019 입니다. 이 PR은 그 열차 밖이지만, 공개 게이트웨이로 라우트하는 사용자에게는 요청이 통째로 거절되는 체감이 큽니다. 변경은 파일 2개·함수 하나·테스트 두 개입니다. 다만 같은 시각에 열린 #3109도 src/adapters/openai-responses.ts 를 만집니다. 둘 다 머지하려면 한쪽에 리베이스가 필요합니다.

라인 src/adapters/openai-responses.ts strip 함수 - 최상위 input 배열의 직접 항목만 지웁니다. 중첩 메시지 content 안의 같은 필드가 있으면 그대로 남습니다. 지금 Codex가 붙이는 위치와 맞는지 한 번만 확인하면 됩니다.
경로 src/adapters/openai-responses.ts 비정규 가드 블록 - 같은 파일의 isCanonicalOpenAiForwardProvider 가드가 여러 개로 쪼개져 있습니다. 이 PR은 첫 블록에만 넣어서 promote보다 먼저 돕니다. 순서는 맞지만, 나중에 누군가 가드를 합치면 빠지기 쉬운 자리입니다.
경로 #3109 - 같은 어댑터 파일의 compact SSE 수집을 고칩니다. 이 PR과 한 커밋에 안 넣어도 되지만, 머지 순서를 정하지 않으면 한쪽이 CONFLICTING이 됩니다.
경로 #3038 / #3066 - 레이어 결정은 이 PR이 맞습니다. #3038을 리베이스하지 말고, 이 랜딩이 들어간 뒤 leftover로 닫는 쪽이 분할 캠페인 원칙과도 맞습니다.

메인테이너의 판단이 필요한 지점

• #3109와 누가 먼저 dev 에 들어갈지 (같은 openai-responses.ts)
• leftover fix(responses): strip Codex-private item metadata #3066 / 대체된 #3038을 이 랜딩 머지 직후 바로 닫을지
• 중첩 input 항목까지 재귀로 지울지, 최상위만으로 충분한지

너의 추천
머지 쪽으로 진행하는 것을 추천합니다. 비정규 경계에서만 지우고 정규 ChatGPT 경로는 보존하며, 테스트가 어댑터 buildRequest 를 실제로 탑니다. #3038은 닫고 이 브랜치를 쓰면 됩니다. #3109와 파일이 겹치니 둘 중 하나를 먼저 랜딩한 뒤 나머지를 고치면 됩니다. Protect dev 리뷰 후 넣으면 됩니다.

이 댓글은 grok-bot이 작성했습니다

[lidge-jun]

Administrator bypass record\n\nOwner authorization was explicit in Codex session 01a05a34-1e3a-73f2-8607-15e517cbec11 on 2026-09-01 KST. Exact head: 39712e4. Current exact-head check rollup: 0 failing, 0 pending. Independent A-gate review found no unresolved code blocker. Using the maintain/admin pull_request bypass documented in MAINTAINERS.md because the author cannot approve their own PR. This bypass does not waive CI or security review.