Run a long OpenAI Responses API call in background mode, poll or stream its progress, resume an interrupted stream, and cancel it

Documented steps

1. Create the response with background:true so the call returns immediately instead of blocking until generation finishes. The docs frame this for long-running tasks on models such as GPT-5.6, GPT-5.2, and GPT-5.2 Pro.
2. Capture the returned response id immediately; it is the only handle to the job.
3. Poll with GET https://api.openai.com/v1/responses/{response_id}. Continue polling while status is 'queued' or 'in_progress' — those are the only non-terminal states.
4. Stop polling as soon as the status leaves those two states and branch on the terminal status.
5. To stream a background response, pass stream:true alongside background:true at creation.
6. Track the 'sequence_number' on every streamed event so you can resume after a dropped connection.
7. Resume an interrupted stream with GET https://api.openai.com/v1/responses/{response_id}?stream=true&starting_after={sequence_number}.
8. Cancel with POST https://api.openai.com/v1/responses/{response_id}/cancel.
9. Treat cancel as safe to retry: 'Cancelling twice is idempotent - subsequent calls simply return the final Response object.'
10. Retrieve results through the same GET /v1/responses/{response_id} endpoint used for ordinary responses; there is no separate background-only retrieval endpoint.