fix(responses): give every reasoning input item the summary the upstream requires

[lidge-jun]

Summary

A reasoning item that reaches a Responses upstream without a summary field is refused before
inference with Missing required parameter: 'input[N].summary', and the whole turn fails. The field
is optional in responsesRequestSchema, so such an item passed every local gate and broke only on
the wire — reported from an Aside/Pi client talking to the proxy over /v1/chat/completions.

Two layers produced or passed on that item, and both are fixed:

• src/chat/inbound.ts minted the reasoning item for a replayed assistant turn (Audit F6) with
  content only. Every other item this proxy mints — src/bridge.ts, src/claude/inbound.ts —
  already carries summary; this ingress was the one that did not. It now carries the replayed
  thinking as a summary_text part, matching the Claude ingress, which also matters because
  sanitizeReasoningInputContent blanks content for every destination except a
  preserveResponsesReasoningContent provider, so summary is the only channel that survives to a
  native backend.
• sanitizeReasoningInputContent now supplies an empty summary for any reasoning input item that
  arrives without the key, so a third-party client or another translated ingress cannot reintroduce
  the same 400. This is not gated on the destination: an item missing the key is not a shape a
  canonical backend issued — Codex and this proxy always emit summary — so nothing a native
  backend sent is reshaped. An item that already has a summary is returned untouched.

Verification

• Hosted CI at the exact head of this branch is the evidence for this change. The local suite,
  typecheck, and build were NOT RUN: the maintainer explicitly prohibited local suite runs for
  this task.
• Static review of the blast radius instead: sanitizeReasoningInputContent has two direct test
  owners (tests/providers/deepseek-reasoning-replay.test.ts,
  tests/adapters/anthropic/anthropic-thinking-signature.test.ts), and the only whole-item equality
  assertion over a summary-less item is the first test in the DeepSeek file, whose expectation is
  updated here with the reason inline. Every other test that builds a reasoning item either supplies
  summary or does not compare the sanitized item by equality.
• New regression coverage: tests/providers/deepseek-reasoning-replay.test.ts (summary supplied
  when absent, existing summary untouched, repair applies on the content-preserving path) and
  tests/responses/chat-inbound-reasoning-replay.test.ts (the minted item carries the summary, and
  the replayed thinking survives sanitization).

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

Summary by CodeRabbit

• Bug Fixes

  • Fixed Responses API requests containing reasoning items without a summary so they no longer fail validation.
  • Chat Completions reasoning translations now preserve both summary and reasoning text.
  • Improved reasoning replay to retain plaintext, required metadata, and correct message ordering.

• Tests

  • Added regression coverage for missing, existing, and preserved reasoning summaries.
  • Added coverage for synthesized reasoning replay behavior and reliable recovery of prior conversation state.

[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 2026-09-15T01:43:16.814738Z	304ef17	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: Advanced

Run ID: 9570112d-bf31-4852-8c26-98daea4a2e20

📥 Commits

Reviewing files that changed from the base of the PR and between 304ef17 and ffcc189.

📒 Files selected for processing (1)

• tests/codex-integration/issue-702-expired-replay-state.test.ts

Included review availability: Your plan provides up to 10 included reviews per hour; 7 remain after this review.

---

📝 Walkthrough

Walkthrough

The change adds reasoning summaries during Chat Completions translation and repairs missing summaries before sending Responses input. Tests cover synthesized summaries, empty-summary injection, existing-summary preservation, encrypted content, raw-content preservation, and replayed requests.

Changes

Reasoning replay

Layer / File(s)	Summary
Synthesize reasoning summaries src/chat/inbound.ts, tests/responses/chat-inbound-reasoning-replay.test.ts	Translated assistant reasoning items now include plaintext in summary as well as content. Tests verify the summary and its preservation when sanitization clears content.
Repair missing Responses summaries src/adapters/openai-responses/reasoning.ts, tests/providers/deepseek-reasoning-replay.test.ts, tests/codex-integration/issue-702-expired-replay-state.test.ts	sanitizeReasoningInputContent detects a missing summary key and adds summary: []. Existing summaries remain unchanged. Tests cover encrypted content, raw-content preservation, and forwarded replay input.

Priority: ➖ Normal

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

Change: Bug fix · Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to ffcc1

The PR supplies the required reasoning summaries for translated and replayed inputs without an identified merge-blocking regression.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 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: adding the required summary to reasoning input items sent to Responses upstreams.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 2 functions across 5 files.
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.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/reasoning-input-summary

---

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.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 304ef17a2d

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]

Update the owned reasoning contract for the new summary shape

This changes translated Chat replay from a content-only reasoning item to one containing a non-empty summary, and the sanitizer now makes that summary the only text sent to canonical Responses targets, but none of the owning structure documents were updated. In particular, structure/providers/chat-compat.md:247-253 still says raw reasoning remains in content and explicitly records that routing raw chain-of-thought through the summary channel was reverted. Update structure/data-planes/inbound-compat.md and the reasoning-channel invariant to describe this new exception, or keep the raw text out of summary, so the repository's architecture contract no longer contradicts the wire behavior.

AGENTS.md reference: AGENTS.md:L33-L41

Useful? React with 👍 / 👎.

[lidge-jun]

리뷰 · 우선순위 75 / 80

이 PR은 Chat Completions → Responses 변환과 reasoning 입력 살균에서 summary 키가 빠져 업스트림이 Missing required parameter: 'input[N].summary' 로 거절하던 구멍을 막는다. 지금 dev(HEAD 9b711073a, 패키지 2.56.0, 방금 합쳐진 #4671 openai-responses 파사드 이후)의 src/chat/inbound.ts는 assistant 의 reasoning_content/reasoning_details 를 재생할 때 { type: "reasoning", content: [{ type: "reasoning_text", text }] } 만 넣고 summary 를 붙이지 않는다. 로컬 responsesRequestSchema 는 그 필드를 optional 로 두어서 게이트를 통과하고, 실제 Responses 업스트림에서만 400이 난다. Aside/Pi 클라이언트가 /v1/chat/completions 로 프록시에 붙을 때 보고된 증상과 맞다.

고침은 두 층이다. 첫째, chat 인그레스가 Claude 인그레스(src/claude/inbound.ts)처럼 재생된 thinking 을 summary: [{ type: "summary_text", text }] 와 content 양쪽에 담는다. 이유 주석이 중요하다. sanitizeReasoningInputContent 는 preserveResponsesReasoningContent 가 아닌 목적지에서 content 를 비우기 때문에, summary 가 네이티브 백엔드까지 살아남는 유일한 채널이다. 둘째, 살균 함수 자체가 summary 키가 없는 reasoning 입력 항목에 빈 배열을 넣는다. 목적지와 무관하다. Codex/이 프록시가 내보내는 항목은 이미 summary 를 가지고 있으므로, 키가 없는 항목은 번역된 인그레스나 제3자 클라이언트에서 온 것이고, 빈 배열을 넣는 것은 “빠진 키만 채우는” 형태 교정이다. 이미 summary 가 있으면 그대로 둔다.

파일 위치도 현재 dev 와 맞다. #4671 이후 살균기는 모놀리스가 아니라 src/adapters/openai-responses/reasoning.ts 리프에 있고, 파사드 src/adapters/openai-responses.ts 가 sanitizeReasoningInputContent 를 재export 한다. 테스트 import 가 그 파사드를 쓰는 것도 맞다. 회귀는 DeepSeek 살균 스위트(없을 때 채움 / 있을 때 유지 / content 보존 경로에서도 수리)와 chat-inbound 재생 스위트(요약 보유 + 살균 후 content 비고 summary 유지)를 추가했다.

같은 구멍에 대한 열린 PR #4629(Ingwannu, #4628)와 선택이 다르다. #4629 는 wire 를 맞추려고 summary: [] 만 넣고 raw thinking 을 summary_text 로 올리지 않는다(privacy / “raw 를 summary 로 재분류하지 않는다”). 이 PR(#4673) 은 summary 에 텍스트를 넣어 살균 후에도 재생 thinking 이 남게 한다. 빈 summary 만 넣으면 살균 뒤 content·summary 가 둘 다 비어 네이티브 경로에서 재생 내용이 사라질 수 있다. 그래서 “400만 막는가” vs “재생 내용까지 살리는가”가 메인테이너 판단 축이다. types.ts/config.ts 스플릿과 무관하고, close-don't-rebase 대상이 아니다. draft 가 아니고 MERGEABLE, hygiene 계열은 SUCCESS 로 보인다.

라인 / 경로 수준:

라인 ~316 (src/chat/inbound.ts) - 기존은 content 만. 이번엔 summary_text + content. Claude inbound 와 같은 “thinking → summary” 매핑이다. #4629 의 빈 배열 선택과 의도적으로 갈라선다
라인 ~62–70 (src/adapters/openai-responses/reasoning.ts) - missingSummary 가 early-return 조건을 깨고 next.summary = [] 를 넣는다. content blank / status strip 과 독립적으로 동작한다
경로 tests/providers/deepseek-reasoning-replay.test.ts - 첫 equality 기대값에 summary: [] 를 넣었고, 세 개 회귀 테스트가 살균 계약을 핀다
경로 tests/responses/chat-inbound-reasoning-replay.test.ts - 파사드에서 sanitize 를 import 해 “살균 후 summary 생존”을 증명한다. #4671 이후 경로가 맞다
심볼 assistantReasoningText / F6 adjacency - reasoning 이 assistant message 바로 앞에 오는 순서는 그대로다

메인테이너의 판단이 필요한 지점

• fix(chat): include summary on replayed reasoning items #4629(빈 summary, privacy) vs fix(responses): give every reasoning input item the summary the upstream requires #4673(summary_text 에 thinking, 재생 생존) 중 어느 계약을 dev 표준으로 잠글지
• fix(responses): give every reasoning input item the summary the upstream requires #4673 을 택하면 fix(chat): include summary on replayed reasoning items #4629/[Bug]: Chat Completions translator omits reasoning.summary; gpt-6-astra 400s after tool turns #4628 을 superseded 로 닫을지, 아니면 privacy 문구만 fix(chat): include summary on replayed reasoning items #4629 문서에서 가져와 주석에 남길지
• 로컬 reasoningItemSchema.summary 를 required(빈 배열 허용)로 올려 omit 을 게이트에서 막을지(후속 가능)
• Aside/Pi 가 summary 채널에 thinking 이 보이는 것을 제품으로 받아들일지(Claude 경로와 동일 패턴)

너의 추천
CI 초록 확인 후 #4673 을 우선 착륙 후보로 둔다. 살균기까지 고쳐서 제3자 omit 재발을 막고, content blanking 이후에도 재생 텍스트가 남는 쪽이 #4629 의 “빈 summary 만”보다 제품 회귀(#4628 계열 도구 연속 턴)에 더 가깝다. 머지하면 #4628 을 닫고 #4629 는 Landed via #4673 / superseded 로 정리하는 것을 권한다. privacy 우려가 크면 머지 코멘트에 “Chat 재생 thinking 은 Claude inbound 와 같이 summary_text 로 실린다”를 한 줄 고정한다. types/config 스플릿 때문에 닫을 PR이 아니다.

이 댓글은 grok-bot이 작성했습니다