API

Streaming

Partner server-sent events, and the OpenAI-style stream.

Partner messages

POST the message JSON to /api/partner/v1/messages/stream, or to /api/partner/v1/conversations/{id}/messages/stream. The response is text/event-stream with Cache-Control: no-cache.

The turn finishes first. The reply text is then split into events of at most 80 characters. This is not token-by-token output from the model.

event: start with data {"ok": true}.
event: delta with data {"text": "..."}.
event: done with conversation_id, message, and credits.
event: error with error and code when the turn fails.

text/event-stream

event: start
data: {"ok": true}

event: delta
data: {"text": "The MOT is booked"}

event: done
data: {"ok": true, "conversation_id": "conv_…", "credits": 0}

Chat completions

Set stream true on POST /api/v1/chat/completions. The body is text/event-stream. Each line is data: {json} and the stream ends with data: [DONE]. A comment line is sent about every 15 seconds so the connection stays open. That shape follows the OpenAI stream, not the partner event names.