feat(providers): translate chat completions onto Responses-only upstreams

[weselben]

TL;DR

The chatgpt provider answers every chat completions request with 501 unsupported_provider_operation, so chat-only harnesses and mixed virtual-model pools cannot use ChatGPT subscription models. Adds the inverse of the existing ResponsesViaChat adapter: chat requests translate onto the Responses API upstream, and Responses streams translate back to chat chunks.

Closes #1084. Part of #723. Follow-up for the Messages ingress: #1085.

Files to review (12, +3228 / -26):

File	Why
internal/providers/chat_via_responses.go (start here)	Entry points ChatViaResponses / StreamChatViaResponses, request mapping, and the rejection rules.
internal/providers/chat_via_responses_input.go	Messages to Responses input items. System messages become instructions. Tool calls become function_call items.
internal/providers/chat_via_responses_output.go	Responses output back to one chat choice. Mints the chatcmpl- ID. Renames usage fields.
internal/providers/chat_via_responses_stream.go	Responses SSE to chat chunks. The state machine lives here.
internal/providers/chatgpt/chatgpt.go	Wiring. Chat surface delegates to the adapter. Embeddings stay 501.
docs/providers/chatgpt.mdx, docs/providers/overview.mdx	Rejection list, drop table, capability matrix.
*_test.go (4 files)	Mapping, rejection, and streaming cases.

How

• Translation follows the mirror adapter. Same package, same naming, same error style.
• Parameters without a Responses equivalent fail with a 400 that names the field, before any upstream call: n > 1, logit_bias, stop, seed, both penalties, logprobs, top_logprobs, modalities, audio, web_search_options, deprecated functions / function_call. Explicit JSON nulls pass. prediction and n: 1 are stripped, not rejected.
• The Codex backend pins store: false and rejects previous_response_id. Minted chatcmpl- IDs are labels, not handles. No chaining exists. Chat clients resend full history as usual.

Reviewer notes

• Fast-fail 400 is not failover-eligible. Default policy retries 429 and 5xx only. A bad request fails once, in place.
• response.failed on the non-streaming path returns a 502 provider error. Without the check it would surface as a 200 with empty content.
• temperature and max_tokens map to valid Responses fields, then Codex drops them upstream. Same behavior as the native Responses path onto this backend. Documented in docs/providers/chatgpt.mdx.
• Usage accounting works on both paths. The stream converter emits the usage chunk only when the client asked or the enforcement flag is on.
• Focus area: the item tracking in chat_via_responses_stream.go — item_id to dense tool_calls[].index mapping under interleaved parallel calls.

Tests

go build ./... and go test ./... pass. New tests cover every mapped field, every rejected field (typed and via ExtraFields), null tolerance, parallel tool calls, refusal, the finish-reason table, truncation, mid-stream failure, usage gating, and ID stability. A wiring test locks the chatgpt delegation with zero network calls.

One pre-existing local failure is unrelated: internal/testconventions scans .worktrees/ directories and lints other branches' files. Fixed separately in #1087.

Follow-up

• feat(providers): Messages ingress onto Responses-only upstreams (MessagesViaResponses) #1085: Messages ingress onto Responses-only upstreams, after this merges.
• Emulation of n > 1 (fan-out) and stop (lookbehind cut) is possible later. Not in scope.

---

This PR description was generated with AI assistance.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
gomodel	🟢 Ready	View Preview	Sep 30, 2026, 6:37 PM

💡 Tip: Enable Automations to automatically generate PRs for you.

[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

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

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.}

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 96.93593% with 22 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/providers/chat_via_responses_stream.go	94.98%	15 Missing ⚠️
internal/providers/chat_via_responses.go	97.59%	4 Missing ⚠️
internal/providers/chat_via_responses_input.go	98.83%	2 Missing ⚠️
internal/providers/chat_via_responses_output.go	98.73%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!