watchdog

Watchdog

Continuity checks are timed or evidence-triggered interventions that keep an agent workflow from going idle, drifting, or entering needless bureaucracy. They operationalize the refocus doctrine.

Use this when

Authoring a watchdog policy for a long-running rig (artifact-pool edges, stale-owner detection, periodic reminders).
Choosing the intervention level for a stale-owner situation: is this a wake, a refocus, or an alignment checkpoint?
Drafting wake/refocus/alignment-checkpoint message text that prevents cadence pollution and bureaucracy theater.
Deciding cadence — scan vs wake intervals, conservative vs aggressive nudging.

Don't use this when

The agent is actively working and closing artifacts. Prefer no-op.
The intervention is masking bad startup or workflow design — fix the underlying cause, don't compensate via watchdog.
The work has a clear next-action handoff already in queue. Use queue nudges instead of a watchdog policy.

The 3-level intervention stack

Level	Goal	When	What it does
Wake	Restart motion	Owner appears idle, stale, blocked, or missing a next handoff	Small liveness nudge — does NOT reframe the work
Refocus	Correct drift	Output shows mode drift, approval regression, weak stop-condition reasoning	Medium alignment nudge — re-centers on role, north star, current approved workstream, coordination mode, stop conditions; does NOT interrupt valid work
Alignment checkpoint	Rebuild shared map	Phase boundary, lifecycle mutation, product-intent decision, confusing contradiction	Larger deliberate pause — agent runs full refocus before proceeding

cron, timers, and rig watchdog are scheduling substrates. They should NOT imply that every tick means the same semantic action. The intervention level is chosen by evidence, not cadence.

Current best practice for refocus message text

A well-shaped refocus message:

Starts with: finish current action first.
Names the intervention kind: wake, refocus, or alignment checkpoint.
Names the current approved workstream or says none is known.
Says what continuity means: continue, verify, hand off, or explicitly park.
Names stop conditions: scope/risk/posture changes, failed gate, contradictory evidence, or no continuity chain naming a sensible next step.
Does NOT convert itself into a fresh approval gate.
Does NOT wake delivery seats unless an approved workflow exists and the owner is stale.

Failure modes (7; verbatim)

Cadence is too frequent and pollutes the workstream.
Watchdog wakes the wrong seat instead of the stale owner.
Refocus text is too rule-shaped and creates brittle behavior.
Watchdog compensates for bad startup or workflow design instead of revealing it.
Refocus is misread as a new top-priority task and interrupts the current action.
Refocus becomes bureaucracy theater: an already-approved workflow stops for re-approval.
Repeated static nudges teach agents to answer the reminder instead of progressing the artifact.

Artifact-pool loop edges (the first evidence-aware pattern)

The first evidence-aware watchdog use that should remain small and explicit:

Consumer-pool wake: if a ready artifact exists, wake the consumer loop head.
Producer-edge repair: if upstream completion exists but the downstream artifact is missing, wake the producer loop head.

Scan cadence and wake cadence are separate. Worked example (the numbers are illustrative; the scan-vs-wake separation is the durable rule): scan every 30 seconds, wake at most every 600 seconds while work remains actionable.

This is not a generic workflow engine. It's a guardrail against cold pools and missing edge artifacts.

Active surface

Daemon rig watchdog is the active host coordination surface for new work. History records only loud evaluations (sent / terminal); quiet skip reasons are NOT recorded — POC parity so agents are not woken about scheduler polls.

Future shape (current limit)

The mature primitive should be evidence-aware: inspect durable queue/workflow state, pane activity, transcript/context growth, and known workstream frontier; classify the intervention as no-op / wake / refocus / alignment-checkpoint; prefer no-op for active owners; render message text from the active orientation graph instead of hard-coding static policy.

Until queue substrate, activity state, and lifecycle evidence are fully productized at this level of granularity: keep config-layer use small and explicit, prefer one workflow watchdog plus targeted exception handling over many per-seat nag loops.

About this skill