Steer a live task with mid-run user input
Inject a user message into an actively running (or resuming) task. The current tool/LLM step finishes, then every live agent loop on this stream sees a real user message at its next LLM step.
Send stream_id from the live SSE you are watching. The API will
not retarget the steer onto a newer run of the same chat.
This does not abort in-flight tools. Do not PATCH working memory
(user_runtime_input) for composer text — use this endpoint while
the stream is live.
Body
Composer text to inject into the live run
1 - 32000Live SSE stream this composer send was typed against. When set, the API returns 409 if that stream is no longer live — including when a newer run on the same chat has started. Prevents a retried steer (e.g. "stop") from attaching to the next turn. Send the stream_id from the live SSE metadata.

