Menu
Post-create PR lifecycle automation — rerun infra failures, triage review comments adversarially, enqueue when green, verify merged main. Designed for /loop.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
| name | shepherd-prs |
| description | Post-create PR lifecycle automation — rerun infra failures, triage review comments adversarially, enqueue when green, verify merged main. Designed for /loop. |
| allowed-tools | Bash, Read, Grep, Glob, Task, Agent |
| argument-hint | [PR number to focus on, or empty for all open PRs] |
Run after /ship (or any PR creation) so authoring sessions never
babysit checks. Designed to be invoked repeatedly via /loop (e.g.
/loop /shepherd-prs) — each invocation is one idempotent sweep.
Bound by .claude/rules/orchestrator-rules.md; halts (reports, does
not improvise) on anything outside the playbook below.
List open PRs: gh pr list --json number,headRefName,headRefOid,isDraft,autoMergeRequest,mergeStateStatus
(or scope to $ARGUMENTS if a PR number was given). Skip drafts.
For each PR, fetch check state via REST (GraphQL has been
observed to 401 during GitHub auth incidents while REST stays up):
gh api repos/{owner}/{repo}/commits/<headRefOid>/check-runs
Classify and act per the playbook.
auto-merge (or the user pre-authorized
enqueueing for it): enqueue via gh pr merge <n> --auto --squash
(match the repo's merge method). Otherwise report "ready".Fetch the failing job's log tail (gh api .../jobs/<id>/logs).
Infrastructure class — rerun, don't touch code:
HTTP 401: Requires authentication (GitHub auth incident — check
https://www.githubstatus.com/api/v2/status.json to confirm)The runner has received a shutdown signal / lost communicationETIMEDOUT, ECONNRESETAction: gh run rerun <run_id> --failed. Respect a cap: if
.run_attempt is already ≥ 3, stop rerunning and surface to the
user instead — three infra failures in a row is a real signal.
Real-failure class — everything else (test failures, gate
failures, lint). Do NOT auto-fix blind. If the failure is the
coupling gate or index staleness, the fix is mechanical
(make pr-prep, commit, push — only while NOT enqueued). Anything
else: surface the failing output to the user (orchestrator rule 4)
and move to the next PR.
Triage adversarially — never apply findings unverified:
/code-review Stage 2.)make pr-prep,
commit, push.Verify on-disk truth, not just the MERGED status:
git fetch origin main && git log origin/main --oneline -5 — confirm
the squashed commit is present. If a fix was orphaned by a queue race,
surface immediately.
## shepherd: <n> open PRs
- #<num> <branch>: <state> — <action taken or "no action">
...
merged since last sweep: <list or none>
Run the local fast CI loop (`make ci`) and automatically fix discovered issues using concurrent agents
One-time contributor setup — run `make setup` and verify governed reads work end-to-end so `/init` can report lifecycle and structural counts.
Run dead code and duplicate detection, get categorized cleanup recommendations
Create a git commit with an impact-focused conventional commit message
Staged adversarial review — triage, decorrelated finders (warm + cold), per-finding refuters, evidence-ready report
Glue workflow — coupling gate, local review, conventional commit, and PR creation with waiver discipline in one governed sequence