monitor-deploylisted

# monitor-deploy The first minutes after a rollout are the most likely failure window. This skill watches them. ## Premise Declare GREEN only from observed signal, never from "the deploy command exited 0." Every verdict cites the probe that ran, the value it returned, and the threshold it was compared against — `/health → 200`, `error_rate = 0.4% (threshold 0.5%)`, `p95 = 145ms (baseline 152ms, +0% drift)`. A health-check that was never polled is not a passing health-check. The rollback decision is mechanical: a breach that persists past its `for:` debounce triggers `/rollback-deploy`, it is not a judgment call surfaced for discussion mid-incident. ## When to use - Immediately after `/deploy-stage` (it dispatches this skill unless `--no-monitor`). - As the recovery-confirmation step inside `/rollback-deploy` (watch the previous revision back to GREEN). - Standalone to watch an in-flight deploy you didn't start ("monitor without deploying"). - To extend an observation window past the default (canary soak, slow-burn watch). ## Halt conditions - Refuse to report GREEN without at least one successful health/readiness probe in the window — no probe, no pass. - Refuse to report an error rate or latency number you didn't pull from the live signal source; don't infer "looks fine" from logs alone. - Halt if no health endpoint / readiness signal is resolvable — ask for it rather than declaring success against nothing. - Halt if the baseline for latency/error-rate comparison can'