Support for OpenAI’s Flex processing

[jannikmaierhoefer]

Describe the feature or potential improvement

OpenAI’s Flex processing makes responses.create() return immediately and pushes the real completion (or failure) to a later responses.wait() / retrieve() call. Our current OpenAI integration only wraps create/parse, so the Langfuse observation ends before Flex actually finishes—no final output, usage, latency, or errors are captured.

Request: add Flex-aware instrumentation for Responses.wait, responses.retrieve, and their async counterparts (or otherwise defer closing the observation when Flex jobs are “in_progress”). That would let Langfuse emit correct results and error states for non-streaming Flex calls.

Additional information

No response

[bartlomiej-korpus]

Hello! It's not only related to flex processing but to responses with webhooks in general.