Activity Monitor Skill

PM2 guardian service that monitors the active runtime agent's (Claude or Codex) activity state and ensures it's always running.

When to Use

This is a PM2 service (not directly invoked by Claude). It runs continuously in the background.

What It Does

1. Activity Monitoring: Tracks the agent's busy/idle state every second
2. Status File: Writes ~/zylos/activity-monitor/agent-status.json with current state (busy/idle, idle_seconds, health)
3. Guardian Mode: Automatically restarts the agent if it stops or crashes
4. Maintenance Awareness: Waits for restart/upgrade scripts to complete before starting the agent
5. Heartbeat Liveness Detection: Periodically sends heartbeat probes via the C4 control queue to verify the agent is responsive, triggering recovery when probes fail
6. Health Check: Periodically enqueues system health checks (PM2, disk, memory) via the C4 control queue
7. Daily Upgrade: Enqueues a Claude Code upgrade via the C4 control queue at 5:00 AM local time daily (disabled by default; enable with zylos config set daily_upgrade_enabled true)
8. Context Monitoring: Receives context usage data after every turn; triggers early memory sync and new-session handoff at configurable thresholds (Claude default 70%, Codex default 75%)

Status File Format

{
  "state": "idle",
  "last_activity": 1738675200,
  "last_check": 1738675210,
  "last_check_human": "2026-02-04 20:30:10",
  "idle_seconds": 5,
  "inactive_seconds": 10,
  "source": "conv_file",
  "health": "ok"
}

• state: "busy" | "idle" | "stopped" | "offline"
• idle_seconds: Time since entering idle state (0 when busy)
• source: "conv_file" (reliable) | "tmux_activity" (fallback)
• health: "ok" | "recovering" | "down" — liveness health from the heartbeat engine

PM2 Management

# Start
pm2 start ~/zylos/.claude/skills/activity-monitor/scripts/activity-monitor.js --name activity-monitor

# Restart
pm2 restart activity-monitor

# View logs
pm2 logs activity-monitor

# Check status
pm2 list

Guardian Behavior

• Detection: Checks if the agent is running every second
• Restart Delay: Waits 5 seconds of continuous "not running" before restarting
• Maintenance Wait: Detects restart/upgrade scripts and waits for completion
• Recovery Prompt: Sends catch-up message via C4 after restart

How It Works

1. Monitor Loop (every 1s):

   • Check if tmux session exists
   • Check if agent process is running
   • Detect activity (conversation file mtime for Claude; tmux activity for Codex)
   • Calculate idle/busy state
   • Write status to ~/zylos/activity-monitor/agent-status.json

2. Guardian Logic:

   • If agent not running for 5+ seconds → restart
   • Wait for maintenance scripts to complete
   • Send recovery prompt via C4 after successful restart

3. Activity Detection:

   • Claude primary: Conversation file modification time (reliable)
   • Codex / fallback: tmux window activity
   • Threshold: 3 seconds without activity = idle

Heartbeat Liveness Detection

The heartbeat engine runs inside the activity monitor and uses the C4 control queue to verify Claude is actually responsive (not just process-alive).

State Machine

          heartbeat interval elapsed
  ┌─────────────────────────────────────┐
  │                                     ▼
  │  ok ──[primary fails]──► ok (verify phase)
  │  ▲                              │
  │  │                     [verify fails]
  │  │                              ▼
  │  │                         recovering ──[recovery fails]──► recovering
  │  │                              │                               │
  │  │                    [ack received]              [max failures reached]
  │  │                              │                               │
  │  └──────────────────────────────┘                               ▼
  │                                                               down
  │                                                                 │
  │                                          [ack received after manual fix]
  └─────────────────────────────────────────────────────────────────┘

Phases

Ack Deadline

Each heartbeat probe is enqueued with an ack deadline. If Claude does not acknowledge the probe before the deadline expires, the control record transitions to timeout status and the engine treats it as a failure.

Recovery Behavior

When health transitions back to ok, the engine reads ~/zylos/activity-monitor/pending-channels.jsonl and sends a recovery notification to each recorded channel/endpoint via C4, then clears the file.

Health Check

The activity monitor periodically enqueues system health checks via the C4 control queue.

• Interval: Every 24 hours (86400 seconds)
• Persisted state: ~/zylos/activity-monitor/health-check-state.json (survives restarts)
• Priority: 3 (normal)
• Gated by health: Only enqueued when health === 'ok'
• Gated by agent: Only enqueued when the agent process is running

The health check control message instructs Claude to:

1. Check PM2 services via pm2 jlist
2. Check disk space via df -h
3. Check memory via free -m
4. If issues found, notify the most recent communication channel
5. Log results to ~/zylos/logs/health.log
6. Acknowledge the control message

Daily Upgrade

The activity monitor enqueues a daily Claude Code upgrade via the C4 control queue.

• Enabled: Off by default. Enable with zylos config set daily_upgrade_enabled true
• Schedule: 5:00 AM local time (configured by DAILY_UPGRADE_HOUR)
• Timezone: Loaded from ~/zylos/.env TZ field, falls back to process.env.TZ, then UTC
• Persisted state: ~/zylos/activity-monitor/daily-upgrade-state.json (tracks last upgrade date)
• Priority: 3 (normal)
• Gated by health: Only enqueued when health === 'ok'
• Gated by config: Only enqueued when daily_upgrade_enabled is true in config
• Once per day: Checks local date to avoid duplicate enqueues

The control message instructs Claude to use the upgrade-claude skill, which handles idle detection, /exit, upgrade, and automatic restart.

Context Monitoring

Event-driven context monitoring via Claude Code's statusLine feature, replacing the old hourly polling.

• Mechanism: context-monitor.js runs as a statusLine command — Claude Code pipes context data to it via stdin after every turn (zero turn cost)
• Status file: Writes ~/zylos/activity-monitor/statusline.json with full context data (used_percentage, remaining_percentage, cost, session_id)
• Two thresholds:
  • Early memory sync at 80% of threshold (e.g. 56% for Claude default 70%): Enqueues a prompt for Claude to run memory sync as a background task, so it completes well before the session switch. Only triggers when unsummarized conversation count exceeds the checkpoint threshold (30).
  • New-session handoff at threshold (Claude default 70%, Codex default 75%, configurable via new_session_threshold / codex_new_session_threshold in config.json): Enqueues the new-session trigger. Priority 1, 5-minute cooldown.
• Delivery: Enqueues via C4 control queue with bypass_state — ensures the trigger reaches Claude even during long tasks
• Two-stage design: The trigger message instructs Claude to start the new-session handoff flow; the actual /clear is gated by block-queue-until-idle in the new-session skill's final step
• Log: ~/zylos/activity-monitor/context-monitor.log

The early memory sync decouples memory sync from the session switch. Memory sync is also triggered by the new session's startup hook if unsummarized conversations exceed the threshold, so data is never lost.

Daily Memory Commit

The activity monitor runs a daily git commit of the memory/ directory.

• Schedule: 3:00 AM local time
• Timezone: Same as daily upgrade (from ~/zylos/.env TZ)
• Persisted state: ~/zylos/activity-monitor/daily-memory-commit-state.json
• Script: Calls zylos-memory/scripts/daily-commit.js directly (no C4, no Claude needed)
• Once per day: Date-based deduplication
• Idempotent: If no memory changes exist, the script skips the commit

Timing Guarantees

Both daily tasks (upgrade, memory commit) use the same DailySchedule class:

• Window: Each target hour provides a ~3600s window; with ~1s loop interval it is virtually impossible to miss entirely
• Dedup: Date-based (last_date === today) ensures exactly-once per day, even with imprecise timing
• Persistence: State files survive activity monitor restarts

Permission Auto-Approve

When Claude Code shows a permission prompt (PermissionRequest event), the hook-auth-prompt.js hook automatically sends an Enter keystroke to approve it.

This is intentional by design. Zylos operates in a full-delegation model where the machine's permissions are entirely granted to the AI agent. Auto-approving permission prompts is the expected behavior, not a security gap.

How It Works

1. Claude Code fires a PermissionRequest hook when a tool call requires user approval
2. hook-auth-prompt.js logs the event to hook-timing.log
3. If auto_approve_permission is true (default), it enqueues a [KEYSTROKE]Enter control at priority 0 with bypass-state and 1-second delay
4. The C4 dispatcher delivers the Enter keystroke to the tmux session, auto-confirming the prompt

Configuration

# Disable auto-approve (require manual confirmation)
zylos config set auto_approve_permission false

# Re-enable (default)
zylos config set auto_approve_permission true

Config key: auto_approve_permission in ~/.zylos/config.json (default: true).

Log File

Activity log: ~/zylos/activity-monitor/activity.log

• Auto-truncates to 500 lines daily
• Logs state changes and guardian actions