Skip to content

[openai-responses] Strip unsupported prompt_cache_options parameter on canonical ChatGPT Codex forward #2765

Closed
Closed
[openai-responses] Strip unsupported prompt_cache_options parameter on canonical ChatGPT Codex forward#2765
Labels
landed-via-maintainerOriginal PR closed after landing via a maintainer merge trainproviderProvider adapters, OpenAI-compat presets, upstream API quirksprovider-compatibilityProvider compatibility reports

Description

@Liang-Psych
Liang-Psych
opened on Aug 27, 2026

Client or integration

VS Code / GitHub Copilot (via /v1/responses BYOK endpoint)

Provider or upstream service

ChatGPT Codex forward / openai-responses

OpenCodex version

2.34.0

Endpoint or capability

/v1/responses, OpenAI Responses adapter

Current behaviour

Following the fixes in #2668 and #2648, VS Code / GitHub Copilot requests to gpt-5.6-luna (and other native forward models) now still fail when the client includes top-level "prompt_cache_options" (e.g. {"prompt_cache_options": { ... }}) in the /v1/responses request.

OpenCodex forwards this payload directly to the ChatGPT Codex backend, which rejects it with HTTP 502:

{
  "status": 502,
  "errorCode": "upstream_server_error",
  "upstreamError": "Unsupported parameter: prompt_cache_options"
}

As a result, VS Code Copilot fails on turn start with generic code: 0 / "The model provider reported a failed response without any error details".

Expected behaviour

The openai-responses forward adapter should automatically strip the unsupported prompt_cache_options parameter when dispatching to the canonical ChatGPT Codex forward backend (isCanonicalOpenAiForwardProvider / authMode === "forward"), similar to how truncation is stripped in #2668 / #2672.

Additionally, src/responses/schema.ts (responsesRequestSchema) should accept prompt_cache_options: z.unknown().optional() so valid Responses API requests containing prompt cache options pass schema validation.

Minimal redacted request or reproduction

  1. Send a request to POST http://127.0.0.1:7861/v1/responses:
{
  "model": "gpt-5.6-luna",
  "store": false,
  "stream": true,
  "truncation": "disabled",
  "prompt_cache_options": {
    "ttl": 3600
  },
  "input": [
    {
      "type": "message",
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "hi"
        }
      ]
    }
  ]
}
  1. Request fails with HTTP 502 (Unsupported parameter: prompt_cache_options).

Actual response or error

{
  "status": 502,
  "errorCode": "upstream_server_error",
  "upstreamError": "Unsupported parameter: prompt_cache_options"
}

Upstream documentation

https://platform.openai.com/docs/api-reference/responses

Suggested mapping or implementation notes

  1. In src/responses/schema.ts:
    Add prompt_cache_options: z.unknown().optional() to responsesRequestSchema.

  2. In src/adapters/openai-responses.ts:
    Strip prompt_cache_options in normalizeCanonicalForwardPromptEnvelope or stripUnsupportedForwardParams when dispatching to the canonical ChatGPT Codex forward destination.

Checks

  • I searched existing provider and compatibility issues.
  • The request and response were redacted.
  • The expected behaviour is based on an upstream specification or a concrete client requirement.

Activity

  1. github-actions
    added
    provider-compatibilityProvider compatibility reports
    providerProvider adapters, OpenAI-compat presets, upstream API quirks
    on Aug 27, 2026

  2. Ingwannu commented on Aug 27, 2026

    @Ingwannu
    Ingwannu
    on Aug 27, 2026
    Owner

    I confirmed the current dev@8b1b65b8d path that makes this report possible. parseRequest() deliberately retains the caller's original object in _rawBody, and the canonical forward adapter currently has an explicit regression test requiring caller-sent prompt_cache_options to survive. The reported upstream rejection therefore contradicts a real existing compatibility assumption; it is not a Zod-validation failure.

    I will keep the fix narrower than the suggested implementation:

    • strip prompt_cache_options only for isCanonicalOpenAiForwardProvider(...);
    • preserve it for OpenAI API-key and noncanonical/custom Responses providers;
    • update the existing pass-through regression and the canonical-forward compatibility evidence/documentation;
    • do not add the field to responsesRequestSchema, because passthrough already uses _rawBody and adding it would not affect the wire result.

    I am treating the redacted 502 as valid destination evidence and will open a separate PR with focused negative-path coverage. This issue will remain open until that PR has exact-head CI, non-author review, and merge.

  3. Ingwannu commented on Aug 27, 2026

    @Ingwannu
    Ingwannu
    on Aug 27, 2026
    Owner

    The bounded implementation is now in draft PR #2767: #2767

    The pre-fix canonical adapter test and versioned production fixture both proved that prompt_cache_options reached the wire. After the destination-scoped strip, focused isolated validation is 114/114 green, including gpt-5.6-luna, another canonical model, and preservation controls for OpenAI API-key plus noncanonical forward providers.

    As noted earlier, no responsesRequestSchema field was added because _rawBody already preserves passthrough input and the schema change would not alter dispatch. The issue remains open and the PR remains Draft until the independent current-dev gate regression in #2766 is repaired, exact-head CI is green, and a non-author maintainer reviews/merges it.

  4. lidge-jun commented on Aug 27, 2026

    @lidge-jun
    lidge-jun
    on Aug 27, 2026
    Owner

    리뷰 · 우선순위 77 / 80

    이 이슈는 VS Code / GitHub Copilot이 /v1/responses로 gpt-5.6-luna 같은 네이티브 forward 모델에 요청할 때, 공개 Responses 모양의 prompt_cache_options를 넣으면 ChatGPT Codex 백엔드가 Unsupported parameter: prompt_cache_options로 거절하고 OpenCodex가 HTTP 502(upstream_server_error)로 감싸서 클라이언트가 빈 실패를 보게 되는 버그입니다. 지금 dev HEAD 8b1b65b8d를 보면, #2668/#2672에서 truncation은 normalizeCanonicalForwardPromptEnvelope(약 1202행)에서 canonical forward 경계에서만 빼고 있습니다. 같은 파일의 stripDeprecatedPromptCacheRetention(약 333행)은 prompt_cache_retention만 제거하고, canonical 분기(약 2002–2006행)에는 prompt_cache_options strip이 없습니다. 그래서 이슈 재현 JSON은 현재 코드로도 그대로 502가 납니다.

    한 가지 더 중요한 점입니다. 주석(약 321행)은 GPT-5.6이 레거시 retention을 prompt_cache_options.ttl로 대체했다고 말하지만, 이슈가 보여 주듯이 canonical ChatGPT Codex 백엔드는 그 대체 필드 자체도 거절합니다. 그리고 tests/openai-responses-passthrough.test.ts의 keeps caller-sent prompt_cache_options while dropping the retired retention 테스트는 지금 canonical forward에서 options를 보존하도록 고정해 두었습니다. 즉 코드·테스트·주석이 모두 “options는 살아 남아야 한다” 쪽에 맞춰져 있어서, 이슈는 그 가정이 틀렸다는 증거입니다.

    스키마 쪽 제안도 짚겠습니다. src/responses/schema.ts의 responsesRequestSchema(약 143행)에는 prompt_cache_key와 truncation은 있지만 prompt_cache_options는 없습니다. 다만 parseRequest는 원본 body를 _rawBody에 그대로 넣으므로(parser.ts 약 810행), Zod에 필드를 추가하지 않아도 호출자 값은 이미 패스스루됩니다. 그래서 핵심 수정은 스키마 추가가 아니라, canonical forward destination에서만 필드를 제거하는 쪽입니다. API-key·커스텀 forward까지 전역 strip하면 공개 API 계약을 깨뜨릴 수 있습니다.

    라인 openai-responses.ts stripDeprecatedPromptCacheRetention 주석 L321 - GPT-5.6이 retention을 prompt_cache_options로 대체했다고 적혀 있지만, canonical 백엔드는 options도 거절합니다. 주석이 실제 wire와 어긋납니다.

    라인 openai-responses.ts L2002-2006 - isCanonicalOpenAiForwardProvider 안에서 retention strip과 normalizeCanonicalForwardPromptEnvelope만 호출합니다. prompt_cache_options 제거가 없습니다.

    라인 schema.ts responsesRequestSchema - prompt_cache_options 키가 없습니다. 다만 _rawBody 패스스루 때문에 Zod 추가만으로는 wire가 바뀌지 않습니다. 이슈의 “스키마에 추가”는 문서/검증 편의일 뿐, 실질 fix 지점은 어댑터 destination boundary입니다.

    라인 tests/openai-responses-passthrough.test.ts keeps caller-sent prompt_cache_options - 현재는 canonical에서 options 보존을 기대합니다. 수정 PR은 이 기대를 뒤집고, 커스텀 forward·API-key 보존 네거티브도 같이 고정해야 합니다.

    메인테이너의 판단이 필요한 지점

    • Zod에 prompt_cache_options를 문서용으로 넣을지, _rawBody만으로 충분한지
    • strip 범위를 gpt-5.6 family만으로 좁힐지, canonical forward 전체 모델에 적용할지(이슈·재현은 luna 포함)
    • 초안 PR #2767을 이 이슈의 정식 수정으로 받을지

    너의 추천
    이슈는 유효하고 우선순위가 높습니다. schema.ts 추가는 필수가 아닙니다. #2767처럼 isCanonicalOpenAiForwardProvider 경계에서만 prompt_cache_options를 제거하고, 커스텀/API-key는 보존하는 작은 PR을 받으세요. 기존 “options 보존” 테스트를 깨뜨린 뒤 canonical 제거·비canonical 보존으로 다시 고정하면 됩니다. types.ts/config.ts 분할 캠페인과는 무관합니다.

    이 댓글은 grok-bot이 작성했습니다

  5. lidge-jun commented on Aug 27, 2026

    @lidge-jun
    lidge-jun
    on Aug 27, 2026
    Owner

    Landed via #2767 at 50e9556

  6. lidge-jun
    added
    landed-via-maintainerOriginal PR closed after landing via a maintainer merge train
    on Aug 27, 2026

  7. lidge-jun commented on Aug 27, 2026

    @lidge-jun
    lidge-jun
    on Aug 27, 2026
    Owner

    Already on dev via #2767.

  8. lidge-jun
    closed this as completedon Aug 27, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    landed-via-maintainerOriginal PR closed after landing via a maintainer merge trainproviderProvider adapters, OpenAI-compat presets, upstream API quirksprovider-compatibilityProvider compatibility reports

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions