Friday, Jul 31, 2026
OpenAI responses transport and response format handling
Anurag merged a major refactor of OpenAI integration in luca-py, moving to the /v1/responses endpoint and making response_format a real, opaque data structure that survives model switches.
1 person shipped this
The core work landed in quantized-ai/luca-py with a new `OpenAIResponsesTransport` that runs alongside the existing `OpenAITransport` [1] [2]. The new transport swaps `messages` for `input` items, `choices` for `output` items, and replaces delta chunks with typed SSE events. It also surfaces `status` and `incomplete_details` instead of `finish_reason`, aligning with OpenAI's current API shape. A five-reviewer pass surfaced three merge blockers, all confirmed against the live OpenAI and Anthropic APIs, which Anurag fixed in the final commit [1]. The work closes issue #4 by making provider-owned opaque data survive a model switch, the shared problem between the responses endpoint move and response_format handling. A single commit also landed to anurag629/anurag629 updating the README.
Sources
- Run OpenAI on /v1/responses and make response_format real · quantized-ai/luca-py
- Run OpenAI on /v1/responses and make response_format real (#18) · quantized-ai/luca-py