Delimit Agent Poll
delimit_agent_pollExplicitly poll a dispatched worker's transcript to check completion and retrieve final output, surfacing uncertain outcomes instead of silent failures.
Instructions
Read a tracked worker's output without inference (explicit poll).
When to use: after delimit_agent_dispatch with launch starts a muse worker, to check whether it finished. Call it explicitly — there is no polling loop in the server. A finished worker exposes its final text here; a worker that died mid-run (e.g. model stream idle timeout) surfaces as "uncertain", never as a silent completion. When the final text carries a QUESTION: block from the worker, this call binds it to the task with a handoff receipt — answer it with delimit_agent_answer. When NOT to use: to read audit-only dispatches (use delimit_agent_status), to close a task (use delimit_agent_complete), or in a tight loop — one explicit call per check; the worker transcript is re-read each time.
Sibling contrast: delimit_agent_status reads the stored task record without touching worker output; this reads the worker's JSONL transcript and advances running → completed/uncertain. delimit_agent_complete accepts and closes; delimit_agent_cancel aborts.
Side effects: reads the worker's JSONL transcript plus its rc sidecar (no model inference) and writes the observed outcome to the task record via ai.agent_dispatch.poll_worker, with an audit entry per call. Recording a worker QUESTION: block additionally creates a handoff receipt addressed to the lease owner.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| task_id | Yes | Task id of a launched worker (AGT-xxx). Required. |
Output Schema
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||