Feature

---
description: Run the full feature pipeline — refine → research → plan → implement → review — with file artifacts and human approval gates.
argument-hint: [what you want to build]
model: opus
---
You are the ORCHESTRATOR of a feature pipeline. You do not do the work yourself — you delegate each stage to a specialist subagent, pass artifacts between them via files (so detail survives), and STOP at the human gates. The feature request is:

$ARGUMENTS

## Setup
1. Pick a short kebab-case `<slug>` for this feature. Create the folder `docs/agent/<slug>/`.
2. Read `DECISIONS.md` yourself first so you can catch any stage that contradicts a settled decision.

## Scale to complexity FIRST (don't over-engineer)
- **Trivial** (one-line fix, rename, tiny tweak): skip the pipeline. Just do it, then run security-reviewer + design-reviewer if it touched auth/data/UI.
- **Small** (clear, localized, no real unknowns): skip refiner + researcher. Go straight to planner → human gate → implementer → reviewers.
- **Medium/Large or ambiguous** (new surface, new flow, multi-file, unknowns): run the full pipeline below.
State which path you're taking and why, in one line, before starting.

## The pipeline (full path)

### Stage 1 — Refine
Invoke the **prompt-refiner** subagent on the request. Write its returned spec to `docs/agent/<slug>/spec.md`.

### Stage 2 — Research
Invoke the **researcher** subagent; tell it the slug. It reads spec.md and writes `docs/agent/<slug>/research.md`.
🛑 **GATE 1 — show me** the 3-bullet research summary + the open questions. Ask me to approve, answer open questions, or adjust. WAIT for my reply. Do not continue until I say go.

### Stage 3 — Plan
Invoke the **planner** subagent; tell it the slug. It reads spec.md + research.md and writes `docs/agent/<slug>/plan.md`.
🛑 **GATE 2 (the important one) — show me the plan** (the headline + the file-by-file changes + order of operations). Ask me to approve or change it. WAIT for my approval. Never implement an unapproved plan. If I want changes, send them back to the planner and re-gate.

### Stage 4 — Implement
Only after I approve the plan: invoke the **implementer** subagent; tell it the slug. It executes plan.md step by step, verifying each, and reports what it changed + verified.

### Stage 5 — Review (parallel)
Invoke **security-reviewer** and **design-reviewer** (design-reviewer only if UI changed). Each returns a prioritized verdict (PASS / CHANGES REQUIRED). Collect both into `docs/agent/<slug>/review.md`.
- If either says CHANGES REQUIRED with a BLOCKER → send the findings back to the implementer to fix, then re-review. Loop until both PASS.
- If both PASS → summarize for me.

### Stage 6 — Hand back
🛑 **GATE 3 — show me** the final summary: what changed, what was verified, both review verdicts. I commit (you don't auto-commit). Remind me if a remote D1 migration is involved so I review the SQL.

## Rules
- Pass detail through FILES (spec/research/plan), not through summaries — that's why the artifacts exist.
- Honor every CLAUDE.md / DECISIONS.md / DESIGN.md rule at every stage; if a stage surfaces a conflict with a settled decision, STOP and tell me.
- Keep me in the loop at the three gates. The gates are the point — never run end-to-end without my approval at the plan.
- One feature per run. Don't expand scope mid-pipeline.