| name | ship |
| description | Take a finished local change to a merged PR through non-negotiable release gates. Use when the user says ship it, /ship, get this merged, release this, or when /coach has a READY local diff. Enforces simplify, verify, design applicability, PR evidence, independent review, CI, and merge babysitting while preserving Anthropic budget where safe. |
Ship
Goal
The change is merged to the default branch, or you stop with the exact unmet invariant and evidence. You may choose the order, but you may not waive invariants.
Invariants
These are preconditions, not a checklist. If any is unmet, achieving it becomes the next action.
- Diff understood.
git status and git diff have been inspected; every changed/untracked file is intentional.
- Simplified.
/simplify has run on the final shape, or explicitly returned UNCHANGED. Substantial code changes after simplify reopen this invariant.
- Verified.
/verify ended in VERIFIED with real entry-point evidence. Tests/builds alone do not satisfy this. Any code change after VERIFIED reopens this invariant.
- Design checked when user-visible.
/design-check passed, passed with notes, or returned DESIGN N/A.
- PR carries evidence.
/pr opened a PR whose body includes verification evidence, review/design notes, and risk/rollback.
- Independent review addressed. Prefer
/codex:review --wait for normal diffs and /codex:adversarial-review --wait for high-risk/design-heavy diffs when Codex is available. Otherwise use an independent Claude reviewer. Critical findings are fixed; important findings are fixed or explicitly accepted/rebutted. Non-trivial fixes reopen invariants 2-4.
- Green and uncontested. CI is green, required approvals/review comments are addressed, conflicts are resolved, and merge method matches repo convention.
- Merged or blocked honestly.
/babysit owns the final loop until merged or a real blocker is reported.
Budget policy
- Use deterministic tools and shell commands before model review where possible.
- Prefer Codex for independent review to preserve Anthropic budget, but do not skip review if Codex is unavailable.
- Do not launch redundant Claude subagents during ship; ship is a gate loop, not an implementation swarm.
- If a gate fails because credentials or budgets are exhausted, stop with
BLOCKED and the exact missing requirement.
State tracking
Maintain a visible invariant checklist. Report transitions tersely:
[ship] simplify: SIMPLIFIED → verify[ship] verify: NOT VERIFIED → fixing blocker[ship] review: critical fixed → re-verify[ship] ci: green + approved → merge
PR evidence minimum
The PR body must contain:
- changed behaviour (
What)
- reason/context (
Why)
- verification commands and representative output (
How verified)
- design screenshots/output or
DESIGN N/A
- independent review status and unresolved accepted risks
- rollback notes
Exit report
End with:
[ship] <MERGED | BLOCKED | NOT SHIPPED>
PR: <url if opened>
Merged: <yes/no + merge commit if available>
Invariant status:
- Diff understood: <evidence>
- Simplified: <SIMPLIFIED/UNCHANGED/BLOCKED>
- Verified: <VERIFIED/NOT VERIFIED>
- Design: <PASS/PASS WITH NOTES/DESIGN N/A/FAIL>
- Review: <tool used + unresolved findings>
- CI/merge: <status>
Next: <only if not merged>
Gotchas
- Never open a PR before verification unless the user explicitly asks for a draft/WIP PR; label it as not verified in the body.
- Never merge with unresolved critical review findings.
- Do not count a reviewer's prose summary as evidence that CI or verification passed.
- A late CI/review fix invalidates prior verification if it changes behaviour.
