ship

name	ship
description	Finish a branch: infer which review passes the change warrants, run them, open the PR, babysit CI to green, triage bot comments, and refresh the body from a clean context.
argument-hint	[--merge] [--effort <level>] [--simplify] [--skip <pass>] [--base <ref>]
allowed-tools	["Agent","AskUserQuestion","Bash(git diff:*)","Bash(git status:*)","Bash(git fetch:*)","Bash(git rev-parse:*)","Skill(plan:review)","Skill(review:code)","Skill(simplify)","Skill(verify)","Skill(comments:audit)","Skill(writing:review)","Skill(pull-request:create)","Skill(pull-request:babysit)","Skill(pull-request:follow-up)"]

Ship

Gate review passes on the diff, run them in order, open the PR, babysit CI to green, refresh the body from a clean context. Each pass fans out on its own; ship only gates and orders.

Default end state green and ready: CI green, bot comments triaged, body refreshed, stopped for your web review. --merge drives to merged.

Context

• Working tree: !git status --short

Decide What Applies

Resolve the base to a remote ref so ship's view matches what the PR merges against. From the base branch (default main, or --base <parent> on a stack), take its tracking ref via git rev-parse --abbrev-ref --symbolic-full-name <base>@{u}, falling back to origin/<base>. Fetch it first (git fetch) so a stale local <base> never inflates the diff with already-merged commits. Diff git diff <resolved>...HEAD, plus a plain git diff for uncommitted work, and thread the resolved ref as --base to every gated pass (including review:code). Gate each pass on the file set, its size, and the content behind any judgment call (new comments, refactor or new behavior). Full matrix and heuristics: references/passes.md.

• plan:review: a substantial approved plan is in context (~/.claude/plans/ file) and the session ran long or redirected enough that the diff could have drifted from it. Dispatched in the background alongside the fix passes, joined before create. A small plan in a tight session doesn't warrant it. No plan, skip.
• Correctness and quality: code changed. Exactly one of review:code <effort> --fix (default) or simplify (pure refactor, no new behavior). Skip on docs/config-only.
• comments:audit: diff adds code comments.
• pull-request:follow-up --local: a supported review bot is detected for the repo. Follow-up's local.md owns detection (repo config fast path, hosted signals when no config is committed) and exits early when nothing is detected. Runs the hosted reviewer locally before the PR exists.
• writing:review: diff touches prose (.md, .mdx, .rst, docs).
• verify: diff has a runtime surface. Declines tests-only and docs-only itself.

Infer, don't interrogate. Present the plan in one line, then proceed. AskUserQuestion only on a real toss-up: refactor versus behavior change, or medium versus high effort.

Flags

• --merge: drive to merged (babysit --merge). Default: green and ready.
• --effort <low|medium|high|xhigh|max|ultra>: override inferred review:code effort. ultra is a billed cloud review only a user-typed /code-review ultra can start, so ship stops and hands it back instead of substituting a local level.
• --simplify: force simplify over review:code.
• --skip <pass> (repeatable): drop a gated pass. Names: plan, review:code (the old code-review is accepted as an alias), simplify, comments, bot, writing, verify.
• --base <ref>: base branch for gating. Default main; on a stack, the parent branch. Resolved to its upstream tracking ref (e.g. origin/...) before diffing.

Pre-PR Reviews

Serialized before create: review:code --fix, simplify, and comment trims all write to the branch. plan:review, when gated in, is read-only, so it runs as a background dispatch alongside these and joins before create rather than gating them. Its findings, if any, are acted on before the PR exists. references/passes.md has the DAG.

1. comments:audit: needs a clean tree (the fix passes dirty it), lands trims via fast-forward (see Comment Trims). Pauses at preflight for an agent-count approval.
2. pull-request:follow-up --local: reviews committed work only and commits its own fixes, so it runs while the tree is still clean, before the fix passes. Pass the resolved base.
3. Correctness and quality: review:code <effort> --fix or simplify.
4. writing:review over touched prose. Address salient findings before the body is written.
5. verify end to end.

Dirty tree at the comment pass: ask whether to commit first. comments:audit operates on HEAD and needs a clean tree.

Comment Trims

comments:audit commits trims to a fresh comments/audit-<hash> branch off HEAD, leaving the tree untouched. Run comments:audit --base <base> --fix and capture the branch name. No branch means nothing to trim: skip. Otherwise dispatch a short-lived general-purpose Agent with that name to fast-forward and delete it:

git merge --ff-only comments/audit-<hash>
git branch -d comments/audit-<hash>

The commit sits on HEAD, so the fast-forward is clean. It runs in the Agent to keep ship's own commands to git diff and git status. Rejected alternatives: references/passes.md.

Create

Join the background plan:review first if it was gated in. Act on fix-worthy drift before the PR exists, and carry any deferred follow-ups into the report. Then pull-request:create commits the working-tree fixes, pushes, opens the PR. Capture the URL: babysit and body-refresh need it.

Babysit

pull-request:babysit <url> watches CI and fixes trivial failures to green.

• --reviews (default on): after first green, hand bot comments to pull-request:follow-up --auto.
• --merge: only when /ship --merge was passed; else stop at green.

A bot review often lands after the green push with no CI event to key off. --reviews covers this: follow-up waits for an expected review's first pass, then triages it. Follow-up owns which reviews are expected, keyed on the signals in its reviewers.md, so ship never restates them. With none expected, the hand-off returns at once and ship stops at green.

Babysit owns the CI waits. Follow-up owns the wait for an expected review and its per-reviewer signal. Ship delegates the review wait and never hand-rolls a reviews-API poll.

Refresh the Body

Dispatch a background general-purpose Agent (cheaper model if set) to run pull-request:update <url> on every run that opened a PR. Never fork: it must read the final PR and diff cold, not this session's transcript, so the body describes the finished change with no review narration. Rewriting the body inside this session is not a substitute.

Report

PR link, passes run and gated out, final state (green and ready, or merging), and whether the body refresh was dispatched.