Periodic liveness so a hung deploy differs from a slow one.

Deploy heartbeats + stale-worker detection

Intent — Periodic [heartbeat] phase=X elapsed=Ns emissions from long-running deploys, plus a stale-worker sweep, so a hung deploy or worker is distinguishable from a merely slow one.

Motivation — the failure it kills

A long deploy (25–40 min) or a worker that stalls is indistinguishable from progress without a liveness signal: you cannot tell "slow build" from "hung process." The failure is blind waiting and undetected hangs: the orchestrator either kills a healthy-but-slow deploy or waits forever on a dead one. It recurs on every long-running operation.

Why it's not just "check the process is alive / wait for it to finish"

Process-alive does not mean progressing; a deadlocked process is alive. Waiting-for-finish gives no mid-flight signal at all. Heartbeats emit phase + elapsed every 30 s, so both liveness and progress (is the phase advancing?) are observable, and the stale-worker sweep flags a worker that has stopped heartbeating. Can a deploy brief tell a slow build from a hung one by asking whether the process exists? No. A hung process exists too. It can only tell them apart by watching the phase advance, which is what the heartbeat gives it.

Mechanism

The deploy phase loop emits [heartbeat] phase=X elapsed=Ns to stderr every 30 s (grep the deploy log to confirm liveness). A startup concurrency guard refuses a second overlapping deploy (the blocking sibling of the signal). The stale-worker sweep detects workers that have stopped emitting and flags them for cleanup.

Prerequisites

• A phase loop with a periodic emit. The operation must be structured into observable phases.
• A parseable heartbeat format (phase= / elapsed=) a consumer can grep.
• A consumer that knows the expected cadence, so "no heartbeat for N×30 s" reads as stale.

Consequences & costs

• Liveness ≠ correctness. A deploy can heartbeat steadily while still failing — the signal proves it is moving, not that it is right.
• Cadence knowledge is required. Judging staleness needs the consumer to know the 30 s cadence; without that the signal is just noise.
• Granularity + noise. 30 s resolution can miss short hangs, and the heartbeat lines add log volume.

Known uses

• The [heartbeat] phase=X elapsed=Ns emissions from the deploy phase loop.
• The startup concurrency guard (refuses overlapping deploys).
• The stale-worker detection sweep.

Related mechanisms

• See also (sibling) — agent-registry, typed-event-bus: the fleet's other signal surfaces; heartbeats are the deploy-specific one.
• Layer — supports staged-deploy-gates: it is the liveness signal over the deploy those gates run.