skills/nightshift/references/compose/shifts/runbook-verification.md

# Runbook verification — finite — safe procedure replay with existing observability

Use when the owner names a runbook or operational procedure and supplies logs, metrics, traces,
or failure-surface evidence to verify it in a declared safe environment — not vendor shopping,
not assumed production telemetry, and not destructive emergency steps.

Write receipts from `$NIGHTSHIFT_PLUGIN_ROOT/skills/nightshift/references/receipts/cycle-specialist-evidence.md`.
The model writes the receipt. Unparsed tool output is `unavailable`, never "no findings".

Supported when a named procedure and safe/disposable environment are declared. Observability and
diagnostics work as modes below before separate catalog entries. Requires repository mode. Never
select this entry in artifact mode. Typical hours: 1–3.

## Observability/diagnostics mode

When the primary gap is missing logs, metrics, traces, or failure-surface visibility, inventory
the signals that exist into a `mode: observability-surface` receipt from `receipts/cycle-specialist-evidence.md`
first. Record absent signals honestly; never assume emitted telemetry reached production.

```text
- [ ] **Runbook verification — replay a named safe procedure with supplied evidence.**
  - Discovery: record the procedure name, declared safe/disposable environment, and supplied
    logs/metrics/traces or failure-surface samples. For observability gaps, write the
    signal inventory before proposing instrumentation. Write a `mode: runbook-verify` receipt from
    `receipts/cycle-specialist-evidence.md` for each step cluster.
  - Never select this entry when work mode is artifact.
  - Verify one step cluster per cycle in the named safe environment; fix repository automation or
    docs when a declared step fails; park production-only steps. Run the item gate, commit.
  - Never impose a vendor, assume telemetry reached production, execute destructive emergency
    steps, or run production-only steps without explicit owner authority.
  - Never certify full operational readiness from one script — record verified, refused, and
    production-only steps separately.
  - Finish with one receipt row per signal and step, marked verified, refused, or production-only.
  - Dedupe against snag-log.md (ALL seen — fixed and rejected).
  - Ends when every in-scope step is verified, refused with reason, or parked as production-only,
    and observability gaps are recorded absent — not inferred.
  - Verify: the item gate is green at every commit; every verified step replays clean in the named
    safe environment, and the receipt states measured versus unmeasured for each surface.
```