Feature Shipper

Turn "I want to build a feature" into a fast execution chain.

Input (Pass Paths Only)

• feature.md: Requirements description (acceptance criteria + non-goals)
• repo_root
• run_dir

Optional flags (recommended inside feature.md)

• mode: plan-only | execute (default: execute)
• feature_slug: short slug for artifact naming (default: derived from title + timestamp)
• quality_bar: demo-ready | functional-only (default: demo-ready for user-facing UI features)

Output (Persisted)

• evidence/features/<feature_slug>-plan.md (checklist plan: tasks + verification)
• evidence/parallel/features/<feature_slug>/ (if implementation is split)
• evidence/features/<feature_slug>-summary.md

Process

0. Hooks doctor (required check; non-blocking): Run tool-hooks-doctor once at the start of the session to verify skill-evolution hooks are enabled. If missing, offer to install project-level hooks; continue either way.
1. Read feature.md, normalize into: acceptance criteria, boundaries, risks, rollback.
2. Prototype UI rule (default): if this feature affects user-facing UI and quality_bar isn’t functional-only, propose 1 “demo moment” (animation/micro-interaction) and add it to acceptance criteria. Must respect prefers-reduced-motion.
3. Produce 2 options (A: minimal; B: cleaner but slower), default to A. If user cares about “demo feel”, offer A-demo-ready vs A-functional-only as explicit sub-options.
4. Split into PR-sized small steps (each independently runnable + rollback-able).
5. Write plan to evidence/features/<feature_slug>-plan.md.
6. If mode: plan-only, stop here and ask for confirmation before implementing.
7. Implement (batch execution + checkpoints):
   • UI visual/layout/animation changes → First call tool-design-style-selector to load the project’s design-system.md, then strictly follow it. If tool-ui-ux-pro-max is installed, use it to ground motion/UX constraints (search “animation” + “accessibility”). For complex visual/animation/responsive design, delegate to /gemini frontend UI/UX senior design agent.
   • Business logic/data flow/integration → Implement directly.
   • Default batch rhythm: 3 small tasks per batch → run verification → report and wait for feedback; stop immediately for help when blocked/verification fails.
   • After each batch (or before merge), recommend using review-quality for a conclusive review + verdict.
     • review-quality is the single entry point and will auto-triage: if React/Next.js performance risk is detected, it will also run review-react-best-practices.
     • If the user explicitly wants only a React/Next.js perf audit, run review-react-best-practices directly.
8. Verification: can run, can build (and existing tests pass).
   • If verification fails (tests/build/runtime error): run tool-systematic-debugging before attempting more fixes.
   • Persist debugging artifacts to:
     • evidence/features/<feature_slug>-debug.md (repro steps, hypotheses, root cause, fix + re-verify)
9. Write evidence/features/<feature_slug>-summary.md: what was done, how verified, next steps.
10. Wrap up: Do a skill-evolution Evolution checkpoint (3 questions); if user chooses "want to optimize", run skill-improver based on this run_dir to produce minimal patch suggestions

Delivery Requirements

• No "big bang" refactoring
• Don't introduce new complexity (unless it significantly reduces future cost, and user confirms)