Pr Maintenance Loop

Hourly PR Maintenance Loop

Schedules the fix-all skill to run every hour via Claude Code's native /loop primitive (CronCreate/ScheduleWakeup) — not GitHub Actions, no new external service, no secrets beyond the local git/gh session already in use. All the actual check/fix logic (sync, CI, CodeRabbit threads, lint, coverage, code-hygiene — the security model and audible notifications that govern it) lives in fix-all; this skill only owns the scheduling.

Known limitations (surface these, don't hide them)

• Session-only, and there's no cloud-compatible alternative. The loop lives in this Claude Code process, using this exact worktree and this local gh/git session (signing key, gh auth, checkout state) — that's the whole point, per the intro above ("no new external service, no secrets beyond the local git/gh session already in use"). It dies when the process ends and auto-expires after 7 days. It only fires while idle. There is no cross-session persistence — that's why this skill re-establishes the loop every session rather than assuming one is already running. A cloud-scheduled agent runs in a fresh, isolated environment with none of that local state, so it cannot run this particular loop — there's no cloud path to fall back to. That's why Step 1 below calls CronCreate directly instead of going through the generic /loop skill: /loop would ask whether to use a cloud schedule for anything ≥60 minutes, and that question has no useful answer here.
• Codex parity is unsolved. This skill only covers Claude Code sessions. If the user runs a Codex workspace on the same branch, tell them explicitly that this loop does not cover it.

Step 0 — preconditions

1. gh pr view --json number,state,mergeStateStatus for the current branch. If there's no open PR, tell the user and stop — don't create one.
2. Check whether a loop is already running this session (list scheduled jobs; look for the sentinel first line [PR-MAINTENANCE-LOOP pr=#<number> repo=cloudposse/atmos started=<ts>]). If found for this PR, don't start a second one — report it's already active. A fresh session never inherits a stale job, since the old one died with the old process, so this only guards against double-starting within one session.

Step 1 — announce and start

Print a one-line announcement before starting — never start this silently:

Starting hourly PR-maintenance loop for PR #<number>...

Call CronCreate directly — cron: "0 * * * *" (every hour), recurring: true, prompt set to the hourly prompt template below with <number>/<repo> filled in and started=<current timestamp> for the self-expiry sentinel. Do not route this through the generic /loop skill: /loop asks whether to use a cloud schedule for any interval ≥60 minutes, but per "Known limitations" above there is no cloud-compatible way to run this loop, so that question has no useful answer — don't ask it, go straight to CronCreate. If a user explicitly insists on trying /loop anyway and it offers the cloud option, tell them plainly it won't work for this loop and point back to this skill instead of proceeding down that path.

After scheduling, immediately run the hourly prompt's steps once now (don't wait for the first cron fire) — same as /loop's own fixed-interval mode would have done automatically, but since this skill bypasses /loop it must do that step itself.

Hourly prompt template

This is the literal text re-enqueued every cycle via /loop 60m:

[PR-MAINTENANCE-LOOP pr=#<number> repo=cloudposse/atmos started=<ts>]

1. Self-expiry: if more than ~6.5 days have passed since `started=`, say so in the summary
  (approaching the 7-day CronCreate expiry) but keep running.

2. `gh pr view <number> --json state,mergeStateStatus`. If state != OPEN, report it and cancel
  this recurring job — don't keep firing no-ops after merge/close.

3. Invoke the `fix-all` skill (`Skill({skill: "fix-all"})`) — it owns steps 3-10 of what used to
  be inlined here: sync, CI check, CodeRabbit threads, lint, coverage, code-hygiene, and the
  security model / audible-notification rules governing all of it.

4. Always end with a one-line cycle summary, even on the no-op path, for auditability.

Related

• fix-all skill — owns the actual per-cycle work this loop schedules: security model, audible notifications, and the sync/CI/threads/lint/coverage sequence. Also invocable directly for a one-shot run without a recurring loop.
• say skill — audible human-attention nudges; see fix-all for the trigger list.
• pull-request skill — owns the label decision tree fix-all applies autonomously each cycle, plus the still human-attended parts of the PR workflow (blog posts, signing setup).
• .claude/settings.json — the permissions allowlist that enforces fix-all's hard prohibitions when it runs unattended from this loop.