Thanks for putting this together. I independently hit the same compatibility gap from the OpenAI Python SDK chat.completions.parse(...) path.

Concrete downstream symptom:

• client sends /v1/chat/completions
• request includes response_format: { "type": "json_schema", "json_schema": { ... "strict": true } }
• Codex generates completion tokens
• the OpenAI-compatible response has empty/null choices[0].message.content
• clients relying on parsed structured output fail with "no parsed content"

I verified current HEAD (aa526920af322568968a30fe820b2b9d55545f8a) does not read or translate response_format in chat-completions.ts.

Your PR's generateObject(...) approach is the cleaner upstream-sized fix if it works consistently with the Codex provider. In a local self-hosted deployment I used a lower-level variant that translates chat response_format to Responses API text.format and returns the resulting JSON string as chat message.content. That path was needed for Honcho's deriver because the Python OpenAI SDK only needs valid JSON content back; client-side .parse() then populates .message.parsed.

I also have a working Vitest regression for the chat-to-Responses translation path: a structured chat completion maps to Codex Responses text format and returns populated chat content. I would be happy to contribute that test to this branch if useful.

One thing to consider before merge: the OpenAI SDK can send both:

• response_format.type = "json_schema"
• response_format.type = "json_object"

If json_object remains unsupported, it would be useful to reject it explicitly or add a small pass-through test so callers do not get silent unstructured output.

The skeptical-maintainer question I would ask is: "Does this promise OpenAI-grade schema enforcement even if Codex does not support the exact same semantics?" I do not think it has to. The compatibility contract that matters for many clients is narrower: honor the structured-output request well enough to return JSON in message.content, or fail clearly rather than ignoring response_format.