Ship Feature

ship-feature

Follow the canonical workflow in ~/.config/ship-feature/WORKFLOW.md (installed by this project). It is the single source of truth — do not restate or fork it here.

The essentials you must honor:

1. Plan it into ~/.config/ship-feature/plans/<repo>-<slug>.md (mkdir -p it if missing) — never the repo, never a session scratchpad (/tmp is tmpfs on many setups: a reboot destroys it, and the plan has to outlive every review round) — then review the plan with a panel: ship-feature plan-review <that file> (no --reviewers — it takes SHIP_FEATURE_PLAN_REVIEWERS, then SHIP_FEATURE_REVIEWERS, both filled from your config; a typed list is a copy of that config that goes stale the day a seat is added. Pass the flag only to override on purpose. Read-only; exit 0 means every reviewer that RAN responded — not that the reviews are clean, and not that everyone ran, 3 re-run). Read the startup lines and say which reviewers actually ran: bare opencode and bare grok are relay-only names here, so if your panel falls back to SHIP_FEATURE_REVIEWERS they are skipped with a warning and the round still exits 0. A single reviewer via codex exec --sandbox read-only still works. Plan-review has its own 2-round cap, distinct from the cross-review loop below — its plan-qualifying term is not the same as, and must not be conflated with, the "qualifying Should-fix" used in cross-review. A round is one plan-review call where every reviewer responded — exit 3 does not count as a round, and if the same reviewer hits exit 3 on two consecutive attempts at the same round, stop retrying it, drop that reviewer for the round, and tell the human about the reduced panel before Gate 1 rather than silently substituting (if that leaves zero reviewers, it isn't a completed round — take the failure itself to Gate 1). Iterate only for a Blocker or a plan-qualifying Should-fix (the plan is wrong about the tree, unsafe, or materially incomplete — a missing edge case, failure mode, or verification gap — not a style/approach preference). A clean round 1 skips straight to Gate 1; do not spend a round confirming an already-clean plan. Round 2, if it happens, runs against the same full panel — never narrowed. If round 2 still has an open plan-qualifying finding, stop — do not run a round 3 on your own; write a disagreement summary (each reviewer's objection, your classification, the reason) and bring it to Gate 1 instead of "the agreed plan." A human-authorized extra round is not a third autonomous round — it resolves cleanly or produces an updated disagreement summary and returns to Gate 1 again.
2. 🚦 Stop for the human to approve the plan — the agreed plan, or the plan plus a disagreement summary if the plan-review cap was hit with a finding still open (the human then accepts it as classified, authorizes one more round with explicit reasoning, or drops the change). Do not write to the source repository before that.
3. Implement in a git worktree (never the main tree), stage explicit paths, open a PR. Run ship-feature preflight first.
4. Run the cross-review: ship-feature relay --author <self> --context-file ~/.config/ship-feature/plans/<repo>-<slug>.md (no --reviewers — relay injects your configured quorum, SHIP_FEATURE_REVIEWERS, which ship-feature's own config can set and otherwise takes from REVIEWERS in ~/.config/pr-review-relay/config. An exported SHIP_FEATURE_REVIEWERS beats both configs, so check your shell profile when a round comes back the wrong size. Typing the list buys nothing and can only be staler than the file: a seat added there is one an agent still naming last month's panel will silently leave out. Pass the flag only for the narrow rounds below, where a deliberate subset is the point.) Read each round's startup lines and say which reviewers actually ran: a benched (out-of-quota) seat is dropped and the run still exits 0. Always pass the plan with --context-file: a reviewer that cannot see the intent can find bugs but not "this is not what we agreed to build". The plan file is the source the PR body is generated from, and is immutable once approved — write and revise it freely before Gate 1, never after; for later rounds point --context-file at a derived copy (plan + dispositions), which is generated, never hand-edited. Iterate only for a Blocker or a qualifying Should-fix — a material problem of correctness, safety, deployability, or verification. Round 1 is the full quorum (every configured reviewer except the author — the relay skips the author, so leaving it in your config costs nothing); rounds 2+ are narrow (only reviewers with a stake in an open finding, plus any whose finding you downgraded — this is where --reviewers is passed on purpose); the closing round is the full quorum again, on the SHA that will merge — unless round 1 was already clean, in which case it IS the closing round and a second full panel is waste. Branch on the exit code — 0 means every DISPATCHED reviewer ran, not that everyone ran and not that reviews are clean: a benched (out-of-quota) reviewer is dropped and still exits 0, so check the startup lines. A benched seat in an initial or closing round is the human's call before the round. Non-qualifying findings are recorded and left unfixed until a follow-up PR — fixing one creates a commit after the reviewed SHA, and then the merged code is not the code anyone reviewed. If the closing round raises a new qualifying finding it becomes the INITIAL round of another cycle — full, narrow, then a full quorum again on the new candidate; not fix-narrow-stop, which merges a SHA no full panel ever saw. Never gh pr edit the derived context file — the body is generated from the plan and only the plan. Exit 4 is the round cap: stop and escalate to the human; never a reason to merge and never a reason to pass --reset. The author classifies each finding in writing, naming it and the reason. Post dispositions before re-running — the relay deletes each reviewer's previous comment, so an unposted finding and its classification are lost; keep the relay's marker text (<Name> review (automated cross-review)) out of that comment or it is deleted too.
5. 🚦 Stop for the human to merge. Never self-merge. Disputed classifications do not block the stop — they go to this gate with both positions. Still two gates, never three.
6. Verify the tests on the merge commit.

For a trivial change, the human may say --light (may skip only the plan review — never the worktree, tests, cross-review, or merge gate).

Read WORKFLOW.md now and follow it step by step.