Agent Skills: /smith-ship — ship the current change end-to-end

/smith-ship — ship the current change end-to-end

Drive the change in the current worktree from review to merged, following smith conventions. Run phases in order; stop and surface only on the must-ask triggers in @smith-gh-pr/SKILL.md (Review Convergence Protocol). Argument (if given) is the intended PR scope/title.

Live state

• Status: !git status -s
• Branch: !git branch --show-current
• Open PR (if any): !gh pr status 2>/dev/null | head -20

Procedure

Load and follow @smith-gh-pr/SKILL.md, @smith-git/SKILL.md, @smith-style/SKILL.md, @smith-worktree/SKILL.md, and @smith-subagents/SKILL.md.

0. Preflight — run /smith-preflight to see which invariants the change already violates. Exactly one result routes instead of blocking: a branch-first FAIL caused by the checkout sitting on a protected branch, and only when it is the ONLY failing check — paired with any other FAIL it stops here like the rest, or the companion failure rides along unaddressed. A branch-first FAIL from a detached HEAD is never that case. Every other FAIL stops here — a secret-scan failure in particular must not reach review or commit. The binding refusal on push stays at step 4.
1. Isolate — a branch-first FAIL from step 0 is not cleared by being clean: if the checkout sits on a protected branch with nothing modified, branch or EnterWorktree before any edit, then re-run the gate. Never take path (a) below while still on a protected branch — committing there is what @smith-guidance Harmless forbids; branch first, then commit. Otherwise check git status --porcelain: if the checkout is dirty, STOP and pick ONE path: (a) commit or stash the changes, then EnterWorktree; or (b) branch in place (git switch -c …) and continue in the current checkout — no EnterWorktree at all (the tree is still dirty and the worktree-dirty-guard hook would block it; a new worktree starts clean, stranding uncommitted changes). worktree.baseRef: head only preserves already-committed unpushed work, never uncommitted changes. Then, if on the EnterWorktree path in a background session, enter the worktree (see @smith-worktree) and rename the branch to the @smith-style convention before any push.
2. Review to convergence — run the /smith-review loop; it owns the tool marshaling, per-round receipts, and convergence criteria. Do not ship until it reports converged.
3. Commit — logically and semantically atomic; conventional subject ≤72, body ≤72/line, Assisted-by: trailer (see @smith-style). One concern per commit.
4. Push & PR — re-run /smith-preflight on the finished commit first and do NOT push on a NO-GO (step 0's verdict was taken before the branch and commit existed, so it does not carry). Then push the renamed branch; gh pr create --base <default> --assignee @me with a What/Why/Testing body ending with the Assisted-by: line (@smith-style). Link issues only if real. The body and title are content: show them and create on an explicit yes (@smith-guidance Harmless).
5. Address review — follow @smith-gh-pr (Code Review Cycle, Posting Review Findings) for fetching, replying, attribution, suggestion blocks, and the content-vs-mechanics gate. Ship-specific: fix high-confidence findings with fix + amend (not new commits); re-run /smith-preflight before each of those pushes too — an amended commit is content step 4's verdict never saw, so a token introduced while fixing a review finding would otherwise reach the remote unscanned; after each push rerun the /smith-review loop to a fresh converged verdict (full receipt plus criteria — a mechanics-only recheck is not convergence); confirm CodeRabbit actually ran before trusting 0 (@smith-gh-pr "CodeRabbit fails OPEN").
6. Merge — re-run /smith-preflight once more first: the PR title, body and every review reply are external writes authored AFTER step 4 gated, so this is the only point that measures them. Do NOT merge on a NO-GO: the refusal binds here exactly as it does at step 4, and a gate that runs without being able to stop the merge is decoration. Then, on both a GO and a converged verdict for the LATEST pushed commit (earlier convergence does not carry across new commits), gh pr merge --squash --delete-branch (targets the current branch's PR). If this branch has an open child PR stacked on it, OMIT --delete-branch and follow @smith-gh-pr "Stacked PRs" instead.
7. Sync & clean up — ExitWorktree (remove), then ff-only pull the repo's DEFAULT branch in the primary checkout; clear any squash-merge orphan branch (@smith-worktree Sync-After-Squash-Merge).

Stacked mode — /smith-ship stack

For multi-part work that should ship as a dependent stack of PRs. Also load @smith-gh-pr "Stacked PRs" (+ its references/STACKS.md when operating on the stack) and @smith-worktree.

• Decompose first — split into logically AND semantically atomic units, each one branch/PR. Present the numbered decomposition and get scope approval BEFORE any branch exists (@smith-guidance Operating Discipline; this is a distinct, earlier gate than the existing-branch scope verification in @smith-gh-pr Stacked PRs).
• Per unit, run phases 0-4 above in that unit's own worktree, with two stack overrides: the PR base is the parent branch (not the default branch), and the PR body carries Depends on: / Blocks:. Branch names follow @smith-style. The stack's PR titles and bodies are shown together and opened on one explicit yes (batched consent).
• Phases 5-7 are stack-wide, not per unit — address review across the stack, then merge bottom-up with retarget-before-merge and manual parent branch deletion per @smith-gh-pr Stacked PRs (cli/cli#1168); cascade rebases; ff-only sync once per merge.
• Report per branch — verify the stack scope (no omissions) and state status per branch in-band.

State results in your own message (in-band); end with a result: line.