stationkeeping

stationkeeping

!checklist init ${CLAUDE_SKILL_DIR} --force

A merged PR and a green pipeline are not the finish line — they are the moment the real work starts. Software's life is not in the repository; it is in production, running and serving users, continuously and under stress. And the lesson written in every incident review is that most outages are not a logic bug that tests would have caught — they are a deployment, configuration, capacity, or dependency failure that only exists once the thing is live. This skill is the discipline of getting a system into production safely and keeping it healthy there, across seven gated stages, and it will not advance past a GATE until the checklist tool clears it.

Operations is where the agent era bites hardest, because the agent's blind spots map exactly onto how production kills you:

• The agent equates green with done — a green dashboard reads as "safe," and silence reads as "healthy." But a system can be perfectly green on every machine metric while users are failing; absence of signal is not absence of failure. Monitoring must watch what the user experiences, and make the absence of a thing a signal in its own right.
• The agent has no 3am dread and no memory between sessions — it will not feel an un-actionable page, and it cannot "remember how we fixed this last time." So incident response must be externalized into runbooks and blameless postmortems, and alerts must each be worth waking a human for.
• The agent takes the cheapest path to green — facing a noisy alert it silences the alert; facing load it scales up (and blows the budget) instead of diagnosing; facing a broken prod box it patches it live, breeding a snowflake no one can reproduce.
• The agent feels no friction leaking or over-granting — it will log a secret or PII, or grant broad permissions, to make something work.

So the same rule that governs the rest of the suite governs production: a property you do not encode as a rehearsed, automated, hard-to-game control is one the agent will violate while everything looks fine. Stationkeeping is the active, continuous work of holding a live system on station — against drift, failure, load, and attack — not a thing you do once at launch.

Discipline: finish every GATE before the next stage. GATEs are hard — never skip, batch past, or self-certify a stage you have not done. The checklist tool enforces the order; let it. Commands address stages by name.

Read references/agent-era-shifts.md first — it is the heart: what each operational practice becomes once production is operated by something that reads green as safety, silence as health, forgets every incident between sessions, and takes the cheapest path to a quiet dashboard. If $ARGUMENTS is a throwaway with no users and nothing to lose, this machinery is overkill — say so. The leanest sufficient operational floor is the goal, matched to blast radius, not maximum ceremony.

Speak the user's language, or the reliability target gets set blind. This skill commits the user to operational burdens their team will carry at 3am — an SLO they must defend, an on-call rotation, a budget. Only they can set how reliable is enough and what an outage costs. Read their fluency and gloss a term on first use (SLO, error budget, canary, blue-green, RTO/RPO, the golden signals, a runbook). A user who signs off on a "99.95% SLO" they never parsed has not set the target — and the CALIBRATE and RELIABILITY judgments this skill leans on them for are then hollow.