Loop Heartbeat

Scheduled discovery automation that runs scans, finds issues, and routes findings to goals or triage. The heartbeat is what makes a loop a loop — not a one-shot run.

Based on Addy Osmani's Loop Engineering: "Automations are the thing that makes a loop an actual loop. Run prompts on a schedule. Findings go to a triage inbox; empty runs archive themselves."

When to Activate

• User invokes /heartbeat start to begin scheduled scans
• User invokes /heartbeat run for a one-shot scan
• A goal converges and follow-up scanning is needed
• Project onboarding includes quality monitoring setup

Core Concept

The heartbeat is the "discovery layer" of a loop. It answers: "What's broken right now?" on a recurring basis without human prompting.

Every 30 minutes:
  ┌─ Run test suite ─── PASS → skip
  ├─ Run linter ─────── FAIL → create goal (auto-fix)
  ├─ Audit deps ─────── FAIL → triage inbox (human decision)
  └─ Type check ─────── PASS → skip

Configuration

Create .claude/heartbeat.yaml in the project root:

heartbeat:
  interval: "30m"
  scans:
    - name: "test-health"
      command: "npm test 2>&1 | tail -10"
      onFailure: "auto-goal"
      description: "Test suite health"
    - name: "lint-drift"
      command: "npm run lint -- --quiet 2>&1 | wc -l"
      threshold: 0
      onFailure: "triage"
      description: "Lint error count"
    - name: "type-check"
      command: "npx tsc --noEmit 2>&1; echo EXIT:$?"
      onFailure: "auto-goal"
      description: "TypeScript type errors"
  budget:
    maxDollarsPerHour: 2.0
    pauseOnExhaust: true

Scan Result Classification

onFailure	Behavior	Use When
auto-goal	Creates /goal automatically	Machine can fix it (tests, lint, types)
triage	Adds to /triage inbox	Human judgment needed (deps, security)
notify	Desktop notification only	Informational, no action required
ignore	Log silently	Monitoring only, aggregate later

Budget Controls

• maxDollarsPerHour: Pause heartbeat if scanning costs exceed this
• pauseOnExhaust: If true, pause (resumable); if false, stop entirely
• Goal creation inherits the default goal budget from /goal settings

Empty Run Handling

When all scans pass:

• No goals or triage items are created
• Run is logged to ~/.claude/heartbeat-last-run.json
• Next run proceeds on schedule
• This is the "archive themselves silently" behavior

Integration Points

• /goal: auto-goal failures create goals via completion-oracle.js
• /triage: triage failures append to ~/.claude/triage/inbox.jsonl
• CronCreate: Scheduling uses the Claude Code CronCreate primitive
• ScheduleWakeup: Alternative for dynamic-interval loops
• Hooks: Runs as a scheduled task, not a synchronous hook

Hard Bans

• Heartbeat MUST NOT modify code (it's discovery-only)
• Heartbeat MUST NOT create goals without a stopping condition
• Budget limits MUST be enforced (no unbounded scan spending)
• Failed scans MUST be classified (no silent failures)
• Empty runs MUST NOT create noise (pass = skip)

Example Session

User: /heartbeat start

System:
  Loaded .claude/heartbeat.yaml (3 scans configured)
  Registered CronCreate: every 30m
  Next run: 30m from now
  Budget: $2.00/hour

[30 minutes later, heartbeat fires]

Scan results:
  ✓ test-health: all tests pass
  ✗ lint-drift: 4 lint errors found → created goal-a1b2c3d4
  ✓ type-check: no type errors

Summary: 2/3 scans passed, 1 goal created