melech-pr-gardener
Tend one explicit GitHub author's open PRs and keep them green — one stateless pass per scheduled run.
AdirD · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add AdirD/agent-shell-hamelech --skill melech-pr-gardenerAbout this skill
name: melech-pr-gardener description: Tend one explicit GitHub author's open PRs and keep them green — one stateless pass per scheduled run. disable-model-invocation: true
PR Gardener
What This Is
A single-pass playbook, not a loop or a long-lived agent. A scheduled runtime (Claude Routine, Cursor Automation, Codex Automation, Antigravity Scheduled Task, or cron) invokes you once: make one picked PR a little more merge-ready, log it, and return. The scheduler is the loop; you're one tick.
Any memory the host hands you is a hint, never authority. Reconstruct truth from the PR — base, threads, checks — and your own trail from its ledger (below). The PR is the database; if they disagree, the PR wins.
No scheduler driving you yet — the user invoked this by hand? Run the pass
anyway, then offer to install the loop for the current repo using
references/ROUTINE.md.
Stay In Scope
You watch and fix three fronts, all cleared by the commits you push — nothing else:
- Base drift — branch fell behind or conflicts with its base.
- Review comments — unresolved CR threads (Bugbot, CodeRabbit, humans).
- CI checks — red GitHub checks caused by this diff.
Fix them with the repo's own local tooling — lint, typecheck, unit tests,
build (the package.json scripts) — to reason about and verify changes. Off-limits
is anything needing a live environment: E2E suites, provisioning infra,
spinning up services. A blocker only a live environment can diagnose is out of
scope — surface it, don't chase it. Step 3 handles each front in order, fixing it
when clear and handing it back when it needs your call.
Never merge, enable auto-merge, or mark a draft ready. You report readiness; PR-state changes belong to the user. Never open a new gardener PR: push only to the picked PR's existing head branch.
Required Run Input And Auth
Require GARDENER_AUTHOR, as an environment variable or explicit run input,
with the GitHub login whose PRs this run should tend. Export an explicit input
before shell calls. It selects an author; it does not select or impersonate
the API identity.
The system gh authentication performs every read, push, reply, resolution, and
ledger update. Those writes remain attributed to the authenticated app or user
and every visible comment keeps the 🪴 prefix.
If GARDENER_AUTHOR is absent, inspect gh auth status only. A clearly
human-authenticated local development session may use @me for that pass. If
the token belongs to an app/bot, or the identity is unclear, report
"misconfigured / nothing to tend" and stop. Never silently fall through to
@me, and don't depend on gh api user: integrations may be forbidden from
calling /user.
Workflow
Follow these steps in order, once, then stop.
1. Discover candidates
In the current repo, list the selected human's authored, non-draft, open PRs:
gh pr list --author "$GARDENER_AUTHOR" --state open
Drop drafts and any already merge-ready (mergeable, CI green, no unresolved threads). What's left is the candidate set — if empty, report "nothing to tend" and stop.
2. Pick one (fair round-robin)
Read each candidate's ledger (below) for its most recent run timestamp (top
entry). Pick the least recently tended — oldest top entry; no ledger sorts
first; ties break by oldest updatedAt. Apply mechanically — don't re-rank by
importance. Fair rotation, no starvation.
3. Reconcile the picked PR
Refresh live state first (gh pr view, gh pr checks) — never act on stale
data. Then work the three fronts in strict priority order, stopping at the
first that needs the user.
Triage (a dismissal, a reply, an infra-only failure) is pure gh API — no
checkout. Only for an actual code change, or code investigation needed to
answer, use a throwaway git worktree, never the main checkout (you don't own
its state — a cloud checkout, or the user's laptop mid-edit). One per run, never
reused:
main_repo="$(git rev-parse --show-toplevel)"
git -C "$main_repo" fetch origin <pr-head-branch>
git -C "$main_repo" worktree add /tmp/gardener-<pr> \
-b <pr-head-branch> origin/<pr-head-branch>
# edit, commit, push — all inside /tmp/gardener-<pr>
git -C /tmp/gardener-<pr> push
cd "$main_repo"
git -C "$main_repo" worktree remove /tmp/gardener-<pr> --force
git -C "$main_repo" worktree prune
- Never
checkout,switch,stash, reset, or clean the main checkout to "make room" — it clobbers whatever's open there. - Never add a worktree directly from
origin/<branch>without-b; that leaves detached HEAD. If the local branch already exists or is checked out, create a unique local worktree branch and pushHEAD:<pr-head-branch>without force. - Fork PR: fetch the head ref first
(
git -C "$main_repo" fetch origin pull/<n>/head:gardener-<pr>), add the worktree from that local branch. - Can't create one (no access, detached env)? Stop and report — never fall back to the main checkout.
- On every cleanup path, first
cd "$main_repo"(or/if the main checkout is unavailable), then remove and prune throughgit -C "$main_repo". Removing a worktree while the shell is inside it leaves the shell with a dead working directory. - A fresh worktree may not have
node_modulesor other installed dependencies. Do not perform a full monorepo install by default. If the required local runner is absent, skip that local check, record the exact limitation, and let CI verify a narrow low-risk push. If the change is unsafe without local verification, hand it back instead. Never report an unavailable check as passing.
The three fronts, in priority order:
- Base drift. Merge the latest base into the PR branch (you never merge the PR itself) and resolve conflicts, keeping both sides' intent. When both intents can't coexist — clearing it would mean deciding for the author — abort and surface the hunks for t