Session heartbeat

A long autonomous run rarely fails at the limit. It fails at the last dispatch before the limit — the one nobody could finish. A heartbeat is a self-paced recurring wake that turns the ceiling from a crash into a scheduled pause: every beat reads standing before spending anything, and ends with enough state on disk for another session to take the run over.

This is the mechanism. unattended-operation is the policy layered on it for runs a human hands over; it inherits everything here and restates none of it.

When to use

• An orchestrator is rolling through a multi-slice plan over hours.
• The run will cross a usage-limit reset — or may, and you don't know.
• Resuming after a limit, a crash, or a compaction, with work in flight.
• About to dispatch, without knowing if the window can finish that agent.
• A run is producing wakeups but no landed commits.

When NOT to use

• One task inside one sitting. A beat costs a full context read; below the horizon where a limit or handoff is plausible it buys nothing.
• Waiting on a single result your harness will notify you about. That is a notification, not a schedule.
• The run is unattended by human handover — load unattended-operation too. This skill has no rules about what may not happen while nobody watches.