SID-2 Swarm
Reconnecting
Resume a run after a dropped connection, and acknowledge events you have processed.
Swarms can tolerate socket drops. Agents that aren't waiting for tool responses continue generating.
Resuming a run
Open a new connection with the same headers and send response.resume instead of response.create, with the run's ID and the last sequence_number you processed:
async with client.beta.responses.connect(extra_headers={"OpenAI-Beta": "responses_multi_agent=v1"}) as conn:
await conn.send({
"type": "response.resume",
"model": MODEL,
"response_id": response_id,
"last_sequence_number": last_sequence_number,
})
# Re-send every inject that was never acknowledged (see below), then continue the loop.Every event after last_sequence_number is sent again, with its original sequence number. Omit last_sequence_number to receive every event the server still holds.
- Re-send unacknowledged injects. An inject sent just before the drop may or may not have reached the server. Keep each inject until its
response.inject.createdorresponse.inject.failedarrives, and re-send the rest after resuming. One the server already applied is refused as already answered and changes nothing. - Answer calls you received but didn't answer. They aren't sent again if you processed them before the drop.
- If the resume is refused with
run_owned_elsewhere, the run is being moved on the server side. Wait a few seconds and resume again. - If it's refused with
response_not_found, the run is unknown or has expired.
Acknowledging events
Send response.ack with the last sequence_number you have processed. The server then drops earlier events and won't replay them.
await conn.send({"type": "response.ack", "response_id": response_id, "sequence_number": last_sequence_number})Time limits
- A swarm that receives no responses for 30 minutes ends with
response.failed, coderun_abandoned. Unanswered tool calls count as no progress. - A finished swarm can still be resumed for 5 minutes, to collect missed events.