| name | loop-pr-ci-triage |
| description | Proactive loop that watches open PRs in erp-mafia/accounted, fixes failing CI, and addresses actionable review-bot / reviewer comments, on the PR branch, never merging. Use on a schedule (cloud routine) or on-demand via /loop-pr-ci-triage. Follows dev_docs/loops.md (propose-don't-merge, dedupe, loop-verify gate). |
loop-pr-ci-triage
Goal: every open, non-draft PR authored by our team or dependabot is either green + review-addressed,
or explicitly escalated with loop:needs-human. Never merge. Read .claude/loops.md first.
Preflight
gh auth status and confirm repo erp-mafia/accounted. If either fails, stop and report "environment not provisioned".- Respect foreign WIP: never touch a branch with another human's uncommitted/active work; never force-push over someone's commits.
1. Enumerate & prioritize
gh pr list --state open --json number,title,isDraft,headRefName,author,reviewDecision,statusCheckRollup,updatedAt
Skip: drafts; PRs by contributor:flagged/pr:flagged authors; PRs already labeled loop:needs-human.
Prioritize (cap 5 PRs/run, log the rest):
- Failing CI that's a mechanical fix (lint, type error, snapshot, a flaky/needs-rerun check).
- Dependabot bumps failing
core-only / pg-real (there is a live backlog: #730, #639, #638, #637).
- Unresolved actionable review comments (from "The PR Agent", Superagent, or a human).
2. Diagnose each PR
- Failing checks: read the failing job.
gh pr checks <n>; open the failing run's log
(gh run view <run-id> --log-failed). Identify the specific failing command from core-build.yml /
test-pg-real.yml (e.g. check:lint, npm test, test:pg, the extension-import grep).
- Review comments:
gh pr view <n> --comments. Separate actionable (a concrete bug/change) from
informational. Follow the bot-triage rule: honest triage beats chasing every finding.
3. Fix (propose-don't-merge)
gh pr checkout <n> (or for dependabot, check out its branch). Reproduce the failure locally by
running the exact failing command. Make the minimal fix.- Run the
loop-verify gate. If it can't pass safely → step 5.
git push to the PR's own branch. Leave a concise comment on the PR: what was failing, what you
changed, verify output. Do not gh pr merge.
4. Dependabot specifics
If a bump breaks the build/tests, the fix is usually a small code adaptation to the new API or a
peer-dep pin. If the bump is a major version with wide breakage, don't force it: comment with the
breaking changes found and label loop:needs-human instead of a risky rewrite.
5. Escalate / anti-thrash
If you already attempted the same fix on this PR (same failure signature) and it failed again, or the
fix is unsafe/large: stop, add label loop:needs-human, comment what you tried and why you stopped.
Never retry the same failing push in a cycle.
6. Report
Summarize per PR: fixed & pushed / commented / escalated / skipped (with reason). This summary is the
routine's completion notification.
Batch mode (optional)
For many PRs at once, drive this with a dynamic workflow: one agent per PR in its own worktree
(isolation: 'worktree') so parallel fixes don't collide, then an adversarial reviewer per fix before
it's pushed. Keep the 5-PR cap unless told otherwise.
