Stream an OpenAI Responses API call over server-sent events, dispatch on the semantic event types such as response.output_text.delta, and detect completion or failure mid-stream

Documented steps

1. POST https://api.openai.com/v1/responses with "stream": true and the usual Authorization header; the connection returns Server-Sent Events instead of a single JSON body.
2. Read events one at a time and branch on the 'type' field. Each event also carries a 'sequence_number' for ordering and gap detection.
3. Handle the one-time lifecycle event 'response.created', which includes the initial response object.
4. Handle item lifecycle events: 'response.output_item.added' and 'response.output_item.done' (each with output_index), plus 'response.content_part.added' and 'response.content_part.done'.
5. Accumulate text from repeated 'response.output_text.delta' events, reading the 'delta' field and keying by 'item_id'; finalize on 'response.output_text.done'.
6. For tool calls, stream arguments via 'response.function_call_arguments.delta' and finalize on 'response.function_call_arguments.done'.
7. Handle 'response.refusal.delta' and 'response.refusal.done' for safety refusals streamed separately from normal text.
8. On success, handle the one-time 'response.completed' event, which carries the fully assembled final response object. Prefer this over concatenating deltas yourself when you need canonical final state.
9. Handle failure paths explicitly: 'response.failed' for generation failure and a top-level 'error' event for system-level errors. Never assume the stream ends with response.completed.
10. For hosted tools, handle the tool-specific event families such as response.file_search_call.* and response.code_interpreter_call.*.
11. Retain sequence_number if you plan to resume an interrupted background stream via starting_after.