[agent] fix: normalize inbound Chat images before route selection and preserve an explicit reasoning disable

[lidge-jun]

Summary

The native Chat fast path recognized only image_url content parts, while the translated path also understood Pi/MCP {type:"image", data, mimeType} parts (Aside read_file tool results) and Anthropic-shaped {type:"image", source} parts. Two user-visible failures followed from that one gap:

• A routed model an operator declared text-only kept an image-bearing body, because isNativeChatRouteEligible could not see the image and so never diverted the request to the Responses pipeline that describes or strips it.
• The native path is a whitelist passthrough, so the foreign part was forwarded verbatim to an OpenAI-compatible upstream that does not accept that shape.

Recognition now lives once in src/chat/image-parts.ts, and normalizeChatImageParts runs in handleChatCompletionsWithBudget immediately after routing-body validation and before routeModel, so the diversion decision and the forwarded wire observe the same parts.

Identity is preserved deliberately: a body with no foreign image part is returned by reference and stays byte-identical, as is one whose images are already image_url. Only the messages array, the messages holding a rewritten part, and their content arrays are rebuilt — the native path is a whitelist passthrough, so an incidental deep clone would itself be a behavior change. A remote source.type:"url" reference is recognized and rewritten, never fetched; this PR adds no outbound request.

Separately, the Chat inbound effort allowlist dropped none. That is the runtime's disable sentinel, not an unknown value: src/reasoning-effort.ts accepts it and maps it to omitting the wire parameter, and src/clients/config-export.ts maps Pi's off thinking level onto it. Dropping it let a provider default re-enable thinking the caller had explicitly turned off — not neutral for the Anthropic families that think by default and require an explicit thinking:{type:"disabled"} to stop (src/adapters/anthropic.ts:960-966).

Addresses audit findings F1 and F7 from the 2026-09-14 provider/PI compatibility audit.

Stack (merge bottom-up)

#	PR	Layer	Review focus
4	pending	modality fidelity and explicit refusal	F8, F5, F9, Kiro
3	pending	Google structured output, Anthropic parallel-tool disable	F3, F4
2	pending	Chat→Responses control fidelity	F2, F6
1	this PR ← you are here	inbound normalization + reasoning disable	F1, F7

Base is dev. Layer 2 genuinely depends on this one — both change src/server/chat-completions.ts and src/chat/inbound.ts. Review this PR's diff only.

The plan unit is devlog/_plan/260914_provider_parity_stack/; 003_blocker_corrections.md is authoritative over the per-layer docs and records the corrections independent review required.

Verification

Local verification NOT RUN BY USER INSTRUCTION. The repository owner directed that no local product check execute on this machine for this work. No bun test, bun run test, typecheck, build, lint, structure:check, privacy:scan or prepush script was run by the authoring session, and none is claimed as passing, provisional, or assumed. This PR is therefore opened as a draft.

Because red-first execution is impossible under that restriction, the regressions below were written to assert the desired behavior and reviewed statically rather than driven red first.

What real evidence exists:

• Hosted GitHub Actions at this exact head (c7e863b9e5fff52c224574981d7b6860a34ce0f2) is the gate for this layer. Results are not pre-judged here.
• Coordinator baseline at df7dc1be53, before this unit's changes — bun run typecheck exit 0, bun run structure:check exit 0, bun run privacy:scan exit 0. That is a baseline of unmodified source and is not coverage of anything this PR adds.

Regression coverage added (not executed locally):

• tests/server/chat-native-image-normalization.test.ts — recognizer accepts both OpenAI spellings, Pi data/mimeType, and both Anthropic source forms; returns null for a part with no usable reference. Normalization rewrites a Pi part, preserves a detail hint, handles image-only and tool-message content, returns the identical object reference when there is no image and when images are already image_url, and leaves other body fields untouched. Text-only diversion is asserted for Pi, Anthropic base64, Anthropic remote-url and tool-carried images; a text-only body still takes the native path and a vision-capable route still keeps an image body on it.
• tests/server/chat-inbound-reasoning-none.test.ts — flat and nested none spellings survive into reasoning.effort, the produced body still validates against responsesRequestSchema, every other ladder value is unchanged, and an unknown effort is still ignored rather than forwarded.

Both files are registered in scripts/test-layout/layout.json and tests/fixtures/test-layout-expected.json (2-line insertions each, no reordering).

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed. structure/data-planes/inbound-compat.md gains the two sections that own this behavior, per structure/INDEX.md ownership for src/chat/ and src/server/.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults. No credential, auth or network path changes. A remote image reference is never fetched and never logged; no image bytes reach any log or error message.

Current stack synchronization

The manual stack was synchronized bottom-up with dev@246b5cab432b03cbec1766c2faffac13d6e39321.
The repository's existing 2.55.0 version change came from that parent; no artificial feature-branch version bump or release-test suppression was used.

Order: #4534 → #4535 → #4536 → #4539 → #4562.
Current head: c7e863b9e5fff52c224574981d7b6860a34ce0f2. Current base: dev.
Each parent is an ancestor of its child. All five branches were pushed using git push --no-verify.

Fresh hosted CI is required at these new heads. Earlier green jobs or the historical 2.54.0 release-line failure are not represented as new-head results. The PR remains draft; no merge or release was performed. No product validation ran on the connected Mac.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

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.

[lidge-jun]

리뷰 · 우선순위 74 / 80

이 PR은 2026-09-14 프로바이더/PI 호환 감사의 1층(F1·F7) 입니다. 지금 dev HEAD는 df7dc1be5(#4520 desktop-restart 검증 docs, package.json 2.54.0)이고, Chat 네이티브 빠른 경로는 아직 image_url만 이미지로 봅니다. 번역(Responses) 경로는 이미 Pi/MCP {type:"image", data, mimeType}와 Anthropic {type:"image", source}까지 읽습니다. 그 차이 때문에 두 가지가 동시에 깨집니다.

첫째, 운영자가 텍스트만 받는 모델로 라우팅해도, 본문에 Pi/Anthropic 모양 이미지가 있으면 isNativeChatRouteEligible가 이미지를 못 보고 네이티브 경로로 보냅니다. 둘째, 네이티브 경로는 화이트리스트 패스스루라서 그 외국 파트를 OpenAI 호환 업스트림에 그대로 실어 보냅니다. 업스트림은 그 모양을 모릅니다. 이 PR은 인식을 src/chat/image-parts.ts 한곳으로 모으고, normalizeChatImageParts를 routeModel 앞에서 돌립니다. 그래서 “어느 파이프라인으로 갈지”와 “와이어에 실릴 파트”가 같은 본문을 봅니다. 외국 이미지가 없으면 같은 객체 참조를 돌려 바이트 동일성을 지키고, 원격 source.type:"url"은 인식·재작성만 하고 절대 fetch하지 않습니다.

같은 PR의 F7은 Chat inbound effort 허용 목록에 none이 빠진 구멍입니다. src/reasoning-effort.ts는 none을 “와이어에서 reasoning을 빼라”는 센티널로 이미 알고 있고, Pi export는 off를 none으로 넣습니다. 그런데 OUTPUT_CONFIG_EFFORTS가 none을 버리면, Anthropic처럼 기본으로 생각하는 계열은 thinking:{type:"disabled"}를 못 받고 다시 생각하기 시작합니다. “끈다”가 “안 보냄”이 되어 버리는 셈입니다.

스택으로 보면 이 PR이 바닥입니다. base는 dev, 브랜치 agent/provider-parity-01-ingress, 위층 #4535가 같은 chat-completions.ts/inbound.ts를 건드립니다. devlog/_plan/260914_provider_parity_stack/ 계획서가 함께 들어와 줄 수가 많아 보이지만, 감사 근거·차단 수정·4층 순서를 고정하는 문서라 범위 밖 잡청소는 아닙니다. types.ts/config.ts 분리 캠페인과 충돌하지 않고, #4501/#4511(네이티브 describer의 modelCapabilities)이나 #4513(Devin 이미지 와이어)과도 겹치지 않습니다. 로컬 suite는 의도적으로 안 돌렸고 draft입니다. tip CI가 게이트입니다. 리뷰 시점에 test 2/4가 한 번 빨간 기록이 있어, 초록 확인 전까지 merge 점수는 76이 아니라 74입니다.

라인 - src/chat/image-parts.ts normalizeChatImageParts - 이미 image_url인 파트는 손대지 않고, 외국 모양만 새 image_url로 바꿉니다. 메시지가 바뀐 배열만 새로 만들고 나머지 필드는 그대로입니다. 네이티브 패스스루 계약과 맞습니다.
라인 - src/server/chat-completions.ts handleChatCompletionsWithBudget - assertChatCompletionsRoutingBody 직후·routeModel 전에 정규화를 넣었습니다. 순서가 핵심이라, 이 줄을 뒤로 밀면 F1이 다시 생깁니다.
라인 - src/server/chat-native.ts - 로컬 chatBodyCarriesImage를 지우고 image-parts를 씁니다. 인식 폭이 번역 경로와 같아집니다.
라인 - src/chat/inbound.ts OUTPUT_CONFIG_EFFORTS - none을 넣습니다. Anthropic 기본-on 모델에서 “끔”이 다시 살아나지 않게 하는 한 줄입니다.
경로/심볼 - tests/server/chat-native-image-normalization.test.ts / chat-inbound-reasoning-none.test.ts - Pi·Anthropic base64·원격 URL·tool 메시지·참조 동일성·텍스트온리 diversion을 잠급니다. 로컬 미실행이므로 tip CI가 진짜 증거입니다.
경로/심볼 - devlog/_plan/260914_provider_parity_stack/003_blocker_corrections.md - F2를 ingress strip이 아니라 final-target sanitization으로 고친 기록이 있습니다. 위층 #4535를 읽을 때 이 문서가 우선입니다.
경로/심볼 - tip CI test 2/4 - 리뷰 시점에 fail 기록이 있습니다. flake인지 제품 회귀인지 머지 전에 확인하세요.

메인테이너의 판단이 필요한 지점

• draft를 tip Cross-platform CI 초록 뒤에 ready+merge할지, 실패 shard를 먼저 재돌릴지
• 계획서 9파일을 layer 1과 같이 dev에 넣을지, 계획만 따로 docs PR로 뺄지(지금은 한 PR에 묶여 있음)
• layer 2([agent] fix: scope the Responses control strip to the ChatGPT backend and carry Chat reasoning and penalties #4535)를 이 PR 머지 직후 이어서 올릴지, CI 안정화 하루를 둘지
• 원격 URL 이미지를 네이티브 경로에서 image_url.url로만 넘기는 현재 계약(fetch 없음)을 유지할지

너의 추천
내용 방향은 맞고 F1/F7 둘 다 현재 dev에 남아 있습니다. tip CI가 초록이면 draft를 풀고 dev에 merge하세요. 그 전에는 #4535를 단독으로 dev에 붙이지 마세요(base가 이 브랜치입니다). test 2/4 실패가 재현되면 그 shard 로그를 먼저 보고, 계획서 분량은 분리하지 않아도 됩니다.

이 댓글은 grok-bot이 작성했습니다

[lidge-jun]

Maintainer integration per MAINTAINERS.md: landing this maintainer-authored PR on dev without a second maintainer approval.\n\nExact-head verification on c7e863b: 28 checks pass, 2 skipped (macos control, windows matrix placeholder), 0 failed/cancelled. Run set: 34790779603 plus metadata workflows (enforce-target, label, resolve-pr, hygiene) — all green on this head.\n\nLocal suite/typecheck/build: NOT RUN (hosted exact-head CI is the evidence).