Heartbeat

Auto-Adaptive Wake Management

This skill auto-evolves. Start conservative, learn patterns, confirm before assuming.

Core Loop:

1. Wake — Execute checklist, note what was useful
2. Observe — Track hit rate (useful wakes / total wakes)
3. Pattern — After 3+ wakes in category, detect optimal intervals
4. Confirm — Ask: "Should I check X every Y instead of Z?"
5. Store — Only after explicit yes, add to rules below
6. Adapt — Context changes? Re-evaluate intervals

Check intervals.md for timing strategies. Check triggers.md for event-based waking.

Wake Levels

Default to idle. Promote/demote based on context and learned patterns.

Hard Rules (Never Change)

• User says "don't wake me about X" → respect immediately, store
• User asleep (night hours) → dormant unless emergency
• Sub-agent running → watching until complete
• Nothing changed since last wake → extend interval

Entry Format

One line: trigger: interval (level) [hit rate]

Examples:

• email/urgent: 30min (watching) [80% useful]
• calendar/upcoming: 2h (idle) [confirmed]
• deploy/monitor: 10min (active) [until complete]
• social/mentions: 4h (idle) [low value, user said skip]

Monitors

Triggers

Quiet

Patterns

On Each Wake

1. Run checklist (Monitors section)
2. Note: Was this wake useful? Why/why not?
3. Check Triggers: anything warrant action?
4. Decide next interval based on context
5. Update patterns if 3+ data points

Learning Signals

Phrases that adjust autonomy:

• "Don't bug me about X" → add to Quiet, reduce frequency
• "Check X more often" → increase interval for X
• "That was useless" → mark wake as low-value, extend interval
• "Good catch" → mark wake as high-value, maintain/shorten
• "Wake me if Y" → add to Triggers

After pattern emerges: Confirm before internalizing. "I notice X rarely needs checking—reduce to every 4h?"

Empty sections = still learning. Observe utility of each wake, propose adjustments.