Heartbeat

Heartbeat Orchestrator

The Heartbeat skill acts as the background processor for the Agentic OS. It reads the .agent/state/tasks.json queue, identifies pending tasks, routes them to the appropriate skills, and updates the task status upon completion.

How it works

Use the heartbeat.py script in this skill's scripts/ directory to manage the queue:

1. Discover scoped queues: Run heartbeat.py scopes first if you don't know what to work on. It lists every poppable queue under _<batch>_extension.<queue> paths with ready/in-progress/completed counts.
2. Pop Task: Run heartbeat.py pop (legacy mode) or heartbeat.py pop --scope _round3_extension.ag_implementation_queue (scoped mode) to get the next task. If _*_extension queues exist in tasks.json, pop without --scope will refuse with exit 11 — pass the matching --scope.
3. Execute: Read the task details and use the appropriate skill (e.g., osint, deep-research) to fulfill the task.
4. Resilience & Retry: If a task fails unexpectedly, autonomous agents MUST attempt to retry the task logic up to two times before formally failing.
5. Complete/Fail: Run heartbeat.py complete <task_id> --scope <path> --outcome '{"result": "..."}' (scoped) or heartbeat.py complete <task_id> --outcome '...' (legacy). Failing logs to .agent/state/errors.json.

Path varies by agent:

• Gemini / Antigravity: python3 ~/.gemini/skills/heartbeat/scripts/heartbeat.py …
• Claude: python3 ~/.claude/skills/heartbeat/scripts/heartbeat.py …

(Both paths resolve through symlinks to the same canonical script in ~/github/claude-skills, so behavior is identical.)

Scoped queues (added in 1.2.0)

When tasks.json contains a batch metadata block like _round3_extension, treat each list-of-dicts under it where every item has a task_id field as a scoped queue. The script enforces three rules from the project's AGENTS.md § Agentic OS Operating Discipline:

• Honor explicit queues: pop without --scope refuses (exit 11) when scoped queues exist. The broader pending list is off-limits while a scoped queue is active.
• Skip blocked items: any queue item whose gh_issue or task_id appears in any sibling blocked_on_* array is skipped automatically. The script tells you what was skipped and why.
• Stop on drained queue: when no ready items remain in the scoped queue, exit 10 with a reminder to fill the evidence report and stop, not trawl pending.

Example:

heartbeat.py scopes
# _round3_extension.ag_implementation_queue  (3 ready / 0 in_progress / 0 completed / 3 total)

heartbeat.py pop --scope _round3_extension.ag_implementation_queue
# pops the first ready item, marks status=in_progress in place

heartbeat.py complete round3-edit-item-33-southern-qualifier \
  --scope _round3_extension.ag_implementation_queue \
  --outcome '{"pr": 502, "merged": true}'

Stale in_progress handling (added in 1.2.0)

If tasks.json has items in in_progress from a prior session that crashed or handed off, pop refuses with exit 12 and prints the stale items. Resolve with one of:

heartbeat.py clear-stale          # move them all back to pending
heartbeat.py pop --accept-stale --scope <path>   # adopt and proceed

This prevents a fresh agent from silently inheriting another session's half-finished work — a documented failure mode.

Exit codes

Wrappers (heartbeat-runner shells, AG launch scripts) should branch on these — exit 10 means stop, exit 11/12 mean the operator needs to make a decision, exit 14 is a config bug.

Task Format

tasks.json legacy shape (still supported):

{
  "pending": [
    {
      "id": "task-123",
      "priority": "high",
      "description": "Deep research the current state of local LLM orchestration.",
      "assigned_skill": "deep-research",
      "project_id": "claude-skills",
      "agent_id": "antigravity",
      "user_id": "matthias",
      "created_at": "2026-04-25T12:00:00Z",
      "completed_at": null
    }
  ],
  "in_progress": [],
  "completed": [],
  "failed": []
}

Scoped-queue shape (alongside the legacy fields):

{
  "pending": [...],
  "in_progress": [],
  "completed": [],
  "failed": [],
  "_round3_extension": {
    "ag_implementation_queue": [
      {
        "order": 1,
        "task_id": "round3-edit-item-33-southern-qualifier",
        "gh_issue": 496,
        "estimate": "frontmatter only on 2 providers; ~10 min",
        "risk": "minimal"
      }
    ],
    "blocked_on_client": [
      { "items": "Spanish PDFs", "issues": [390, 391], "blocker": "translator pending" }
    ]
  }
}

The script treats anything under _*_extension whose items all have task_id as a poppable queue. Keys starting with blocked_on_ are filter inputs, not queues. Metadata lists like ag_pushed_branches (where items don't have task_id) are ignored by scopes discovery.

Execution

When invoked (or automatically on agent startup), the Heartbeat should process one task at a time to maintain stability and context length, unless explicitly asked to drain the queue. In a project with scoped queues, agents must pass --scope and respect the drained-queue stop signal — picking outside-queue work without authorization is the documented failure mode this version was written to prevent.

Agentic OS Integration

As a core OS component, this skill MUST log its own execution to .agent/state/last-run.json, noting which task was processed and what the outcome was. The 1.2.0 last-run schema adds a scope field so wrappers can tell scoped from legacy invocations.