Problem (one or two sentences)
Concretely, GPT-5.6-class models from several gateways now return:
{"error":{"code":"unsupported_api_for_model","message":"model \"gpt-5.6-sol\" is not accessible via the /chat/completions endpoint"}}
The same model, sent to /v1/responses, returns a normal 200 completion.
Context (who is affected and when)
Anyone routing the OpenAI-Compatible provider at a gateway that exposes
Responses-only models. This is increasingly common:
- GitHub Copilot OAuth (via a local OpenAI-compatible proxy) serves
gpt-5.6-terra, gpt-5.6-sol, gpt-5.6-luna, gpt-5.5, gpt-5.4,
gpt-5.3-codex only through /v1/responses.
- Codex serves its GPT coding models the same way.
These models are visible in the model list and selectable in Zoo Code, but
every request errors out — which is misleading, because the model appears
supported until it's actually used.
Desired behavior (conceptual, not technical)
The OpenAI-Compatible provider should be able to reach /v1/responses for
models that are served only on that endpoint, so those models complete
normally instead of erroring out. This is already how the Opencode Go provider
handles GPT-5.6 models (PR #1443); extending the same capability to the generic
OpenAI-Compatible provider would let Responses-only GPT models be used in
Zoo Code.
Constraints / preferences (optional)
Request checklist
Zoo Code Task Links (optional)
No response
Acceptance criteria (optional)
No response
Proposed approach (optional)
No response
Trade-offs / risks (optional)
No response
Problem (one or two sentences)
Concretely, GPT-5.6-class models from several gateways now return:
The same model, sent to
/v1/responses, returns a normal 200 completion.Context (who is affected and when)
Anyone routing the OpenAI-Compatible provider at a gateway that exposes
Responses-only models. This is increasingly common:
gpt-5.6-terra,gpt-5.6-sol,gpt-5.6-luna,gpt-5.5,gpt-5.4,gpt-5.3-codexonly through/v1/responses.These models are visible in the model list and selectable in Zoo Code, but
every request errors out — which is misleading, because the model appears
supported until it's actually used.
Desired behavior (conceptual, not technical)
The OpenAI-Compatible provider should be able to reach /v1/responses for
models that are served only on that endpoint, so those models complete
normally instead of erroring out. This is already how the Opencode Go provider
handles GPT-5.6 models (PR #1443); extending the same capability to the generic
OpenAI-Compatible provider would let Responses-only GPT models be used in
Zoo Code.
Constraints / preferences (optional)
the safest shape, since not all OpenAI-compatible gateways implement
/v1/responses.
/chat/completions.
(abort signal support for openai-native and openai-compatible), which touches
the same code path.
Request checklist
Zoo Code Task Links (optional)
No response
Acceptance criteria (optional)
No response
Proposed approach (optional)
No response
Trade-offs / risks (optional)
No response