Disposable task
Use a harmless branch or sandbox task. The point is to test lifecycle behavior, not to risk production work.
Sleep/wake resilience
Separate recoverable pause from silent failure.
Local AI agents can survive a terminal close but still fail when the host sleeps, wakes, changes network, or runs closed-lid.
Two heartbeats
Record one host timestamp and one worker timestamp. If both stop, the host slept. If only the worker stops, the agent failed.
Classify clearly
Mark each run as continued, paused recoverably, failed loudly, or failed silently. Silent failure is the bug to eliminate.
Host heartbeat
Start with a timestamp that cannot bluff.
Run this next to the agent, then close the lid or trigger the sleep/wake case you want to test.
while true; do date >> ~/agent-sleep-wake-host.log; sleep 15; done
Receipt fields that make the result useful.
- Interruption tested: terminal close, lock, lid close, sleep/wake, network change, or transport.
- Host heartbeat range and worker heartbeat range.
- Battery, heat, network reachability, and whether awake settings were restored.
- Files changed, commands run, failed checks, stop reason, and next operator decision.
- Local stack host contract and caffeinate vs closed-lid MacBook checklist.
Want Codex to run the host mode?
Take A Coffee gives Codex temporary awake mode, host checks, restore step, and a receipt for local macOS and Windows agent work.