ship-it

name	ship-it
description	One-command post-merge close-out for an approved PR. Trigger on "ship it", "land this PR", "close out PR

Ship It — post-merge close-out

code-review-and-pr stops at "ready to merge?". This skill is the tail after that: merge → confirm close → clean → version-check → clone refresh → plugin update → (optionally) publish. Steps 1–6 run autonomously for PRs whose base is the integration branch (production); step 7 runs only on explicit request.

When NOT to run

• Prepping or reviewing a PR → code-review-and-pr.
• Under a wave supervisor / build orchestrator that serializes merges → open the PR and stop; the supervisor drives the merge (see wave-supervisor). Never ship-it around an orchestrator.
• Base branch is a release/publish branch or the merge triggers a public release → confirm with the user first.

Pipeline

Input: a PR number (or infer the current branch's open PR via gh pr view --json number).

1. Merge (with the transient-retry loop)

gh pr merge <N> --squash --delete-branch

GitHub's mergeability computation is eventually consistent: right after approval or a fresh push, the merge can fail with "Pull Request is not mergeable" even though nothing is wrong. That failure is usually transient:

• Wait ~10s, check gh pr view <N> --json mergeable,mergeStateStatus, retry. Up to 3 retries.
• mergeable: CONFLICTING is a real conflict — stop, report, don't loop.
• Any auth/permission error — stop and report.

2. Confirm the linked issue closed

gh pr view <N> --json closingIssuesReferences,body

• For each Closes #M linkage: gh issue view <M> --json state. If still open, close it with a comment linking the merged PR.
• If the trailer was Part of / Advances #M (multi-PR issue): the issue stays open — tick its completed checklist items instead. Do not close umbrella issues early.

3. Clean up worktree + branches

From the main clone (not inside the doomed worktree):

git worktree list                       # find the feature worktree, if any
git worktree remove <path>              # --force only for disposable leftovers
git branch -D <branch>                  # if the local branch persists
git fetch --prune                       # remote branch was deleted by --delete-branch

Superset workspace deletion is owned by the superset-launch skill — hand off there rather than reimplementing.

4. Version check + repair (the load-bearing step — #210)

Why: the per-PR auto-bump Action pushes its bump commit to the PR head branch asynchronously and can lose the race against gh pr merge --squash — the squash captures the pre-bump tree, the branch is deleted, and the plugin ships to production unbumped. Then claude plugin update reports "already at latest" and never delivers the fix to version-detecting consumers. This has happened for real (#198 shipped at the old version; #372 shipped unbumped at 2026.07.1203).

Detect — in the production clone after pulling the merge (for an older merge, substitute <merge-sha> for HEAD):

git diff --name-only HEAD^..HEAD -- plugins/          # which plugin dirs changed
git diff HEAD^..HEAD -- .claude-plugin/marketplace.json 'plugins/*/.claude-plugin/plugin.json'

If a plugin's files changed but its version fields did not → unbumped ship.

Repair (catch-up bump):

1. Branch off production HEAD.
2. Run make version-bump — for a merge that isn't HEAD, pass MERGE_DIFF_RANGE=<merge-sha>^..<merge-sha>. The script detects the changed plugins and writes both marketplace.json and each plugin's plugin.json in sync (a repo test enforces they match).
3. Never hand-edit version fields — the bump script is the single writer; hand edits drift the two manifests.
4. Commit chore(<plugin>): version catch-up <old> -> <new>, push, open a PR to production, squash-merge it (a catch-up bump is routine → autonomous). Then re-run this step's detect on the catch-up merge to confirm green.

If the bump landed inside the squash normally, report that and move on.

5. Clone refresh

In each local production clone the user works from:

git pull --ff-only

--ff-only always — never introduce merge/rebase commits into a tracking clone during close-out. If it refuses, the clone has local divergence: report it, don't force.

6. Plugin update

claude plugin marketplace update rpw-agent-marketplace
claude plugin update rpw-published@rpw-agent-marketplace   # repeat per affected plugin

• The full name@marketplace identifier is required — a bare claude plugin update rpw-published fails to resolve.
• Report the version delta (old → new). If it reports "already at latest" but step 4 said the version did bump, the marketplace cache didn't refresh — rerun the marketplace update first.
• Note to the user: running Claude sessions need a restart to pick up the updated plugin.

7. Publish (optional — gated on explicit go)

Never part of the autonomous tail. Only when the user explicitly asks to publish:

make publish-public

This filters the tree per .public-publish.yml, pushes a publish-staging branch to the public mirror, runs the sensitive-content check locally (CI can't — IP ACL), and opens a PR against the mirror's default branch. Stop there. The human reviews the actual public diff and their merge IS the publish — never auto-merge the mirror PR. The PUBLIC_REPO_RELEASE_CONFIRM gate applies. Consider make release-notes / make tag-release alongside, per the release flow in AGENTS.md.

Report shape

End with one tidy summary:

• Merged: PR #N → production @ <sha> (retries needed: n)
• Issue: #M closed (auto / manual) — or left open (Part of)
• Cleanup: worktree removed, branches pruned
• Version: bump landed in squash (old → new) / catch-up PR #X merged / no plugin files touched
• Clones: refreshed (--ff-only)
• Plugin update: old → new — restart sessions to apply
• Publish: not run / mirror PR awaiting the human's merge