gastown/polecats/slit 3dd18a1981 Add Deacon stuck-session detection and force-kill protocol ...

Implements gt-l6ro3.4: Deacon can now detect and kill genuinely stuck/hung
Claude Code sessions during health rounds.

New commands:
- gt deacon health-check <agent>: Pings agent, waits for response, tracks
  consecutive failures. Returns exit code 2 when force-kill threshold reached.
- gt deacon force-kill <agent>: Kills tmux session, updates agent bead state,
  notifies mayor. Respects cooldown period.
- gt deacon health-state: Shows health check state for all monitored agents.

Detection protocol:
1. Send HEALTH_CHECK nudge to agent
2. Wait for agent bead update (configurable timeout, default 30s)
3. Track consecutive failures (default threshold: 3)
4. Recommend force-kill when threshold exceeded

Force-kill protocol:
1. Log intervention (mail to agent)
2. Kill tmux session
3. Update agent bead state to "killed"
4. Notify mayor (optional)

Configurable parameters:
- --timeout: How long to wait for response (default 30s)
- --failures: Consecutive failures before force-kill (default 3)
- --cooldown: Minimum time between force-kills (default 5m)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

2025-12-30 22:14:05 -08:00

26 KiB

Raw Blame History

View Raw