ship

name	ship
description	Run the gate battery (verify, coverage, review, security-review), block on CRITICAL/HIGH, print a Handoff Report for sign-off, then deploy the epic to the sandbox and smoke-test it live.
allowed-tools	Bash, Read, Task
argument-hint	[--skip-deploy]

Reused across epics and callable standalone for re-verification, independent of factory (this is why ship is its own skill rather than a step inside factory — see the plan's reuse rule).

Gate battery (Tier 2-equivalent — all blocking)

Pass --epic <epic-id> to steps 1, 2, and 4 — each records its own result into .claude/telemetry/<epic-id>.json via telemetry.mjs, in the same shape as EvalRunResult (Article 10). Review is deliberately not passed --epic and never recorded there — it's the one judgment stage EVALUATION-FRAMEWORK.md itself keeps outside that closed schema, and the dogfood telemetry preserves that boundary rather than blurring it.

1. node .claude/gates/verify/verify.mjs --epic <epic-id> — build + test. Deterministic; exit code is the verdict (Article 6).
2. node .claude/gates/coverage/coverage.mjs --epic <epic-id> — every FR-n/NFR-n resolves to an automated, passing scenario (Article 3).
3. Review: dispatch one isolated reviewer subagent via Task, giving it only the diff since the epic's spec commit and these three criteria (Article 6) — never the implementer's own reasoning:
   • domain-vocabulary drift against docs/vocabulary.md (Article 4)
   • unjustified complexity relative to this epic's stated FR/NFR in docs/plan.md (Article 5)
   • report/CLI clarity, where relevant to this epic Have it write structured findings to .claude/findings/review.json (shape: see review.mjs's header comment), then run node .claude/gates/review/review.mjs to tally.
4. node .claude/gates/security-review/security-review.mjs --epic <epic-id> — secrets, dependency audit, subprocess-injection. Mandatory, never advisory (Article 7).

Any CRITICAL/HIGH finding from steps 2–4, or a nonzero exit from step 1, blocks — fix and re-run, don't proceed past a failed check.

Handoff Report

Once the battery is green, print a short summary (what changed, gate results, any advisory-only findings) and wait for the user's explicit sign-off before deploying — a human checkpoint on top of the machine gate, not instead of it.

Deploy + post-deploy verification

1. npx nx run <project>:sandbox (project name from .claude/epics.json). If oc/gh/podman preflight fails, report the exact fix the executor prints and stop — don't retry blindly. Record it: node .claude/lib/telemetry-cli.mjs record-stage --epic <epic-id> --category deployment --passed <0|1> --duration <seconds>.
2. Re-run this epic's designated smoke scenario(s) — the ones tagged as post-deploy in docs/requirements.md — against the live sandbox URL, not the local test env. This is post-deploy functional verification (Article 8); it's real from epic 1. Record it the same way, with --category post_deploy_verification.
3. node .claude/lib/telemetry-cli.mjs finalize --epic <epic-id> --build-passed <0|1> --security-gate-passed <0|1> --deployed <0|1> --post-deploy-verified <0|1> — closes out this epic's EvalRunResult-shaped record with overall and cycle_time_s (computed automatically from the start timestamp).
4. Update .claude/epics.json: mark the epic status: "shipped".