| name | pr-land |
| description | Review a PR with a dedicated agent and land it UP the tree (leaf→dev, dev→integration). Spawns an isolated Sonnet /pr-review agent that posts findings as PR comments, patches what it finds, replies with the resolution, and updates the PR description; then — once review-clean and green — merges the PR into its base tier. The terminal merge to `main` is held for the maintainer (use /land for that squash). The agent-driven per-PR review loop, parameterized by PR number + base tier. |
| when_to_use | Use to land a working/leaf or kickoff PR onto `dev` (or `dev` onto `integration`) the reviewed, collision-safe way — for each PR you bring up in a concurrent wave. NOT for the final squash onto `main` (that terminal, maintainer-gated step is /land or /wave-land). |
| allowed-tools | Bash(git fetch:*), Bash(git diff:*), Bash(git merge:*), Bash(gh pr:*), Bash(scripts/checks/branch-guard.sh:*), Bash(scripts/checks/worktree-guard.sh:*) |
pr-land
The merge gate below main is the agent's, not a human's (CLAUDE.md §Autonomous PR workflow). This
skill operationalizes the per-PR agent-review loop: every PR worked up the tree gets a dedicated
reviewer, is driven clean through PR comment threads, and is merged into its base tier — with the
merge to main the one terminal checkpoint held for the maintainer.
Parameters
PR — the pull-request number.BASE — the base tier it lands on: dev (from a leaf/working/kickoff branch) or integration
(from dev). For main, stop and use /land (the curated squash + maintainer checkpoint).MODEL — the reviewer model (default Sonnet, per the hybrid swarm assignment).
The loop
- Spawn an isolated
/pr-review reviewer (isolation:"worktree" — mitigation #11; never the
shared main tree). Brief it to apply .claude/skills/pr-review/SKILL.md + the shared rubric to the
PR diff and audit the house rules: the transparency rule (per-op tags never upgraded without a
checked basis), append-only decisions, grounding, never-silent G2, and a hallucination/consistency
pass. Honest and non-sycophantic (house rule #4) — only approve on merit; do not manufacture
findings; flag anything architecturally significant rather than guessing.
- Post findings as PR comments (frugal, severity-ranked). If nothing material, one short approving
comment beats nitpicks.
- Patch what's found (the same reviewer if it has context, else a fresh isolated agent): fix on
the PR branch, re-run the change-scoped gate for what changed (not the full
just check — that
tightens at integration), reply to each comment with the resolution applied, and update the
PR description to match the net change. Commit + push as separate commands (mitigation #12 —
no compound commit && push; keep main/integration/dev out of commit-message text); never
push a protected branch (branch-guard is armed).
- Green, then merge up the tree. When the review is resolved and CI/gates are green, pull the
base down first (
update_pull_request_branch / merge BASE in — mitigation #6, so the diff is
only this branch's net change and the merge is conflict-free), then merge the PR into BASE with
the merge method (--no-ff, lineage preserved) — squash is only for the main landing.
- Shared-file ordering. When several PRs touch the same shared file (e.g.
issues.yaml), land
them sequentially, pulling the freshly-merged base down before the next (mitigation #6).
Boundaries (never-silent)
- Stops at
main. This skill merges only onto dev/integration. The integration → main step
is a curated squash PR, gated by the maintainer / /pr-review — that is /land (or
/wave-land for a head), never this skill.
- Leaves do not finalize
CHANGELOG/Doc-Index/api-index or close out issues — they record
only their own issue and FLAG the rest up; that reconciliation is the integration tier's
(CLAUDE.md §Concurrent-PR development).
- Reviewer + patcher run isolated (
/worktree-guard --leaf); the orchestrator's main tree stays a
clean pointer (/worktree-guard).
Composes with /wave (which calls this per work-item), /pr-review (the rubric it applies),
/worktree-guard and /branch-guard (the safeguards). Full rationale: CLAUDE.md §Concurrent-PR
development + §Autonomous PR workflow.
