@agimon-ai/po-autostop

Shuts a Po session down once the agent settles and stays idle.

A scripted run wants the session to exit on its own when the work is done. Waiting for agent_settled alone is not enough: the agent can settle and immediately pick up a queued message, and it can settle while the last response is still streaming. So this package settles, waits, looks again, and only then stops.

The policy

Moment	Session state	Outcome
agent_settled	messages queued	stand down
agent_settled	queue empty	look again in 5s
the scheduled look	messages queued	stand down
the scheduled look	background runner or agent active	look again in 5s
the scheduled look	still streaming	look again in 100ms
the scheduled look	idle, queue and background empty	context.shutdown()

An agent_start event disarms a pending stop. The cooldown is the grace period; the 100 ms recheck waits out a stream that has not finished draining. Supervised Runner processes (po-runner) hold while running or awaiting asynchronous completion admission (published without a status). Running commands stop holding after five minutes of settled background-work waiting. A new turn resets that clock. Shutdown uses Runner's existing cleanup to stop the commands. A runner listed as completed or failed no longer counts. Owned subagents (team-direct-runs, po-task), pending completion admission and holding-provider errors still hold after the deadline. The holding rule is shared with inline workflow step supervision through core's background-work predicates. Workflow runs (workflow-mcp) do not hold a session open.

Each wait is recorded as po_autostop.waiting_on_background_work telemetry naming the work holding the session, along with po_autostop.stood_down and po_autostop.shutdown_requested.

The idle shutdown watch adds the authoritative background-work gate before the policy may stop the session.

License

Business Source License 1.1: free for personal use, organisations under USD 1M revenue, non-profits and education. See LICENSE and https://agimon.ai/licensing.