Persistent Monitor tasks are reaped on idle/suspend boundaries (~every 30-90 min)

Summary

Persistent Monitor tasks (created with persistent: true) are killed by the harness on idle/suspend boundaries, roughly every 30-90 minutes of user inactivity. The agent receives a <task-notification> with status: killed, but the monitor's bash process is already terminated — requiring a full relaunch.

This is distinct from token expiry or process crashes. The monitor's own output shows no errors before death — no 401s, no stderr, no exit. The harness externally kills the task.

Reproduction

1. Start a persistent Monitor with a long-poll loop (e.g. a curl SSE stream with timeout_ms=25000 in a while true):

Monitor({
  description: "inbox stream",
  persistent: true,
  timeout_ms: 3600000,
  command: 'while true; do curl -sSN -G ... ; done'
})

1. Go idle for 30-90 minutes (no user messages).
2. The harness kills the Monitor and delivers <task-notification> status: killed.
3. The agent must manually relaunch.

Observed frequency

In a single ~3-hour session, the same monitor was reaped 4 times while its auth token was still valid (confirmed by checking the output file — no 401s or errors before death). Across 10+ sessions on the same machine over a week, every session longer than ~40 minutes experiences at least one reap.

Impact

This is the primary reliability issue for agents that use persistent Monitors as a push channel (e.g. long-poll inbox streams for agent-to-agent messaging). Each reap creates a deaf window — the agent misses messages until the next user interaction triggers a restart. The agent-side mitigation is a UserPromptSubmit hook that checks process liveness and injects a restart reminder, but the idle gap between reap and next prompt is inherently unguarded.

Expected behavior

A persistent: true Monitor should survive idle periods and only terminate when:

• The agent calls TaskStop
• The session ends
• The monitor process itself exits

Environment

• macOS (Apple Silicon)
• Claude Code (latest as of 2026-06-06)
• Monitor command: bash while true loop with curl long-poll

Notes

• Token-expiry deaths (401 chains → self-terminate) are a separate, expected mechanism and are handled correctly by the orphan guard in the bash script.
• The harness reap produces status: killed (not status: completed), which is how the agent distinguishes it from a normal process exit.
• persistent: true + timeout_ms: 3600000 are both set. The reap happens well within the timeout window.

🤖 Generated with Claude Code