skill
Watch signals after deploy, release, local run, CI change, or repeated workflow. Uses telemetry when present; otherwise healthchecks, logs, CI, flaky tests, benchmark drift, daemon output, regressions, or agent traces. Emits events, escalates to /diagnose on trip, closes on green. Use when: "monitor signals", "watch the deploy", "watch production", "watch CI", "watch logs", "watch benchmark drift", "is it healthy". Trigger: /monitor.
/monitor
Source: skills/monitor/SKILL.md
Watch signals after deploy, release, local run, CI change, or repeated workflow. Uses telemetry when present; otherwise healthchecks, logs, CI, flaky tests, benchmark drift, daemon output, regressions, or agent traces. Emits events, escalates to /diagnose on trip, closes on green. Use when: "monitor signals", "watch the deploy", "watch production", "watch CI", "watch logs", "watch benchmark drift", "is it healthy". Trigger: /monitor.
Operate primitive in the Harness Kit operating loop.
This generated excerpt gives readers the beginning of the live primitive contract before they jump to GitHub.
Watch signals after a change. Escalate to `/diagnose` on regression. Close
clean when signals stay green through the grace window.
**Every repo has a monitor path.** A production app may watch healthchecks,
logs, traces, and error rates. A CLI may watch CI failures, golden-command
transcripts, and release regressions. A library may watch consumer builds,
benchmark drift, flaky tests, and issue reports. A local daemon may watch
its logs and readiness endpoint. Absence of Sentry, Prometheus, or a
public deploy is not absence of monitoring; it means the signal source is
different.
This skill observes and escalates. It does not diagnose root cause
(`/diagnose` does). It does not rollback (caller decides). It does not
page humans (outer loop decides).
Monitoring also includes observability coverage. A repo with no product
analytics still has signals: agent traces, provider receipts, workflow ledger
events, CI history, command transcripts, logs, benchmark outputs, and release
smoke artifacts. If a change creates behavior that cannot be observed after
ship, flag that as a monitor finding and route it to `/reflect` or `/groom`
as instrumentation debt.
Repeated human corrections are also a local monitoring signal. When the same
workflow needs the same correction across repeated runs, stop passively
watching and escalate to `/reflect prompt-debt` with sanitized counts,
available receipt refs, and the workflow name. Do not store raw private
transcripts or Chronicle detail in the monitor event.
## Execution Stance
You are a thin watcher.
- Load config or fall back to healthcheck-only.
- Poll on a fixed cadence until the grace window elapses or a signal trips.
- On trip: emit one `monitor.alert` event with payload, exit, hand off.
- On clean: emit one `monitor.done` event, exit.
...