Responses API returns 400 upstream_error while models catalog succeeds

[coolbossov]

Summary

An authenticated OpenAI-compatible client can list models successfully, including the configured model, but minimal POST /v1/responses requests fail with HTTP 400 and error.type: upstream_error.

Redacted reproduction

1. GET /v1/models returns 200; the configured model appears in the catalog.
2. POST /v1/responses with a minimal plain-text prompt and no customer content returns 400.
3. The JSON response reports error.type: upstream_error; no stable error code is present.
4. Adding JSON-schema structured output produces the same 400.

Impact

This blocks a Portal relationship-research worker after it has safely fetched public candidates, before it can create the required cited review packet. The worker fails closed: no CRM fact, task, email, outreach, or identity change is made.

Notes

• No API keys, customer data, raw request payloads, or free-form upstream error text are included.
• The configured model is available in the catalog, so this appears specific to the Responses execution path or its upstream mapping.
• A minimal /v1/responses compatibility canary would help confirm the fix.