| name | go |
| description | Ship what built. Run verify -> self-review -> simplify -> commit-push-pr -> monitor CI -> fix -> done. Use when implementation done, ready to launch. |
Go -- Ship What You Built
Phases 4-6 of /development-lifecycle, standalone command. Use when code written, ready launch.
Assumes: implementation done, tests written. If not -- run /development-lifecycle.
Phase 4: Verify
Run all checks. Fix failures before proceed.
bun run type:check (tsgo)
bun run lint:fix (biome)
bun vitest run --related (changed files)
- Route touched ->
bun vitest run *.browser.test.tsx
- Dev server running -> browser smoke via
scripts/skills-browser.sh (Vercel agent-browser). Skip if not installed.
- Frontend or customer-facing surface diff -> run
/visual-review (screenshots/terminal evidence, states, a11y, console, mobile/cross-browser when feasible). Skip only with reason.
- Dependency changed (
package.json, bun.lock, yarn.lock, go.mod, go.sum) -> run /upgrade-dependency or record skip reason + upgrade report/PR section.
- When green: commit now. One commit per passing state.
Phase 4b: Refine (Self-Review Loop)
Skip if: trivial change (<10 lines, no logic) | test-only | docs-only.
- Dispatch
self-reviewer agent on session diff
- Diff >50 lines OR touches auth/security -> also dispatch
adversarial-reviewer parallel
- Resilience Review: risky feature/hook nudge -> run
/resilience-review or record skip reason
- Process findings by priority -- see REFERENCE.md
- Fix P0/P1 now, apply P2
safe_auto, show P2 gated_auto to user
- Commit fixes:
refactor(scope): self-review fixes
- Re-verify (tests + types + lint)
- Max 2 refine rounds. Then proceed.
Phase 5: Simplify + Ship
- Run
/simplify -- review changed code for reuse, quality, efficiency
- Fix issues, commit
- Frontend or customer-facing surface diff and
/visual-review not run this session -> run it now or record explicit skip reason
- Run
/commit-push-pr -- conventional commits, push, open PR
- Dispatch
code-reviewer agent (fresh-eyes review)
Phase 5b: Iterate
Monitor: gh pr checks <number> --watch -- stream CI background
- CI fail -> diagnose, fix, push, re-monitor
code-reviewer agent findings -> /resolve-pr-feedback triage, fix, reply, push
- AI self-review cap: up to 3 auto
code-reviewer rounds. Early-exit when reviewer returns APPROVED or empty findings -- never run round N+1 on clean round N. After 3 rounds still noisy -> hand off to human.
- Human review (incl cloud/Copilot): NO cap. Address EVERY thread.
pr-feedback-completeness-stop hook blocks session exit until bash scripts/pr-unresolved-count.sh returns 0 and no CHANGES_REQUESTED reviews remain.
Phase 6: Compound
After non-trivial tasks: "Learn something worth preserve?"
- Write rule to
.claude/rules/<topic>.md with paths: glob
- AI bug -> create eval/test fixture catch same error class
Done
- Post final PR comment: changes, review findings, test coverage
- Request review:
gh pr edit <number> --add-reviewer <username>
- Report PR URL + CI status
- Stop. No poll for human approval.
Entry Gate
Before start, check work to ship:
- No uncommitted changes AND no unpushed commits -> nothing do, stop
- On default branch, no feature branch -> auto-spawn via
scripts/mux-worktree.sh <type>/<name> before proceed. Never ship from main. [ETHOS: Worktree Isolation]
Skills Composed
| Skill | Phase | How |
|---|
self-reviewer agent | 4b | Auto-dispatch on diff |
adversarial-reviewer agent | 4b | Conditional (>50 lines or auth/security) |
/simplify | 5 | Code quality review |
/visual-review | 4/5 | Customer-facing surface review with product/design/engineering/QA hats when UI, CLI/TUI, mobile, desktop, or report output changed |
/resilience-review | 2/4b/5 | Edge cases, error handling, fallback, polish, observability for risky features |
/commit-push-pr | 5 | Conventional commits + push + PR |
code-reviewer agent | 5 | Fresh-eyes review on PR |
/resolve-pr-feedback | 5b | Triage + fix review comments |
See REFERENCE.md for detailed checklists, gate logic, flowchart.