heartbeats

heartbeats

Dynamic peer-runtime federation. Workers register, take work via agent.delegate, send heartbeats, and deregister cleanly. Heartbeat loss reroutes the in-flight task to another worker — deduped by idempotency_key.

Before ARCP

Static worker pools with bespoke RPCs. The supervisor's "is this worker alive?" answer comes from a TCP keepalive (lies during GC) or a custom heartbeat that re-dispatch logic doesn't actually trust — so re-dispatch either fires too eagerly (duplicate execution) or not at all (stuck pipeline).

With ARCP

safeSetInterval(async () => {
  for (const w of staleWorkers(roster, DEADLINE_S)) {
    const task = jobsToTasks.get(w.inFlightJob ?? "");
    if (task) await dispatch(client, { task, roster, jobsToTasks }); // same idempotency_key
    roster.remove(w.workerId);
  }
}, HEARTBEAT_INTERVAL_SECONDS * 1000);

idempotency_key makes re-dispatch safe: a worker that survived the network blip will see the duplicate agent.delegate and dedupe.

ARCP primitives

• Capability negotiation (per-role extension) — RFC §7, §21.
• agent.delegate — §14.
• Job lifecycle (accepted → started → heartbeat → terminal) — §10.
• Heartbeat loss recovery — §10.3 (heartbeat_recovery: "block").
• idempotency_key for safe re-dispatch — §6.4.
• Trust levels — §15.3.

File tour

• main.ts — boots supervisor + small worker pool in-process. dispatch + reaper is the supervisor; execute + heartbeatLoop is the worker.
• work.ts — doWork stub.

Variations

• Priority queues by tagging tasks with envelope priority.
• Per-worker quota tracked via tokens.used metrics emitted from worker sessions (RFC §17.3.1).
• Replace the in-process workers with separate processes on real hosts; the protocol shape doesn't change.