Hang check

Purpose

A dispatched subagent can hang: it stops making progress but never returns. This skill installs a recurring check that catches a real hang while leaving a slow-but-working agent alone. Judge on evidence, not the clock: elapsed time alone is never proof of a hang, because a long compile, model load, or remote wait looks identical to a stuck agent if you only watch the timer.

When to use

• One or more background subagents are running and may run long.
• The user asks for a watchdog, hang check, or "tell me if an agent gets stuck."

Skip it when no background subagents are active, when driving a single foreground task turn by turn, or when the user wants live per-event log streaming (use the Monitor tool for that).

Timer

Use CronCreate for the recurring check and CronDelete to remove it.

• Arm exactly one recurring job when the first watched subagent starts. One watchdog covers all active agents; do not arm a second.
• Every 10-15 minutes fits most work; shorten only if agents finish fast.
• The job prompt re-enters this skill: "Run hang-check: review each active subagent's status, recent output, and file activity; investigate any agent that looks stuck; remove the timer if no subagents remain active."
• Record the returned job ID for cleanup.
• CronCreate jobs auto-expire after 7 days; tell the user if the watched work could run longer.

Evidence and decision

Each check, read up to three signals per active agent:

File activity is evidence only when the agent is expected to touch files. An agent that is researching, reading, or waiting on a remote call may make no file changes while healthy; do not treat missing file changes as a hang.

Advancing on any relevant signal means healthy: leave it alone. Quiet on every relevant signal for an unusually long stretch relative to the task's pace means suspect a hang: investigate, do not stop it yet. "Unusually long" is relative to the task, not a fixed number.

Investigate before stopping

Before touching a suspected-hung agent:

• Read its most recent output in full. A stack trace, a prompt waiting on input, or a "downloading..." line changes the diagnosis.
• Check whether it is legitimately blocked on a long operation.
• Check file mtimes when files are expected: a file touched seconds ago means work is happening even if stdout is quiet.

Only after the evidence confirms a genuine hang, TaskStop the agent. Do not automatically re-dispatch it: decide whether to resume, replace, or cancel the work. Re-dispatch is wrong when the task is obsolete, already done by another agent, unsafe to repeat, or depends on non-idempotent actions. Report what the evidence showed, not just the action taken.

Cleanup

After each check, if no subagents remain active, remove the timer with CronDelete using the recorded job ID and confirm with CronList so no orphaned watchdog keeps firing. Re-arm only when new subagents are dispatched.

Workflow

1. Confirm one or more background subagents are active. If none, stop.
2. Arm one recurring CronCreate watchdog; record the job ID.
3. On each fire, read the relevant evidence signals for every active agent.
4. Leave advancing agents alone; flag agents quiet on every relevant signal.
5. Investigate each flagged agent before acting.
6. For a confirmed hang, TaskStop, then decide resume, replace, or cancel.
7. When no subagents remain active, CronDelete the timer; confirm with CronList.