rw-planner

name: rw-planner description: "Lite+Contract planner: hybrid askQuestions + subagent planning + DAG/task-graph generation with approval integrity. Bootstraps .ai/ workspace on first run. Input: feature request — resolves required fields first, asks only missing fields, then deep-dive if ambiguous."

RW Planner Skill

Bootstrap workspace + hybrid interview → feature normalization → task decomposition → DAG/task-graph generation.

When to Use

• Starting a new feature that needs structured decomposition
• Re-planning after loop escalation (NEXT_COMMAND=rw-planner)
• Reviewing or revising an existing plan

Procedure Overview

Phase 1: Bootstrap   → scaffold .ai/ dirs and templates (first run only)
Phase 2: Plan        → interview → feature → plan artifacts → atomic tasks
Output:  NEXT_COMMAND=rw-loop

Phase 1: Bootstrap

Scaffold the workspace for first-time use. Skip items that already exist.

1. Create directories:
   • .ai/features, .ai/tasks, .ai/notes, .ai/runtime, .ai/plans, .ai/memory
2. Create .ai/CONTEXT.md with the following default language policy:

   # CONTEXT
   - Response language: 한국어
   - Machine tokens: English (pending, in-progress, completed, blocked, VERIFICATION_EVIDENCE, etc.)
   - Section headers: English (unless overridden below)
   

3. Create .ai/memory/shared-memory.md (see memory-contract.md).
4. Copy feature-template.md to .ai/features/FEATURE-TEMPLATE.md.
5. Create .ai/PROGRESS.md with initial format:

   # Progress
   
   ## Task Status
   | Task | Title | Status | Commit |
   |------|-------|--------|--------|
   
   ## Phase Status
   Current Phase: Phase 1
   
   ## Log
   

6. Create .ai/PLAN.md with initial format:

   # Plan
   - PLAN_ID: (none)
   - Feature Key: (none)
   - Strategy: (none)
   - Task Range: (none)
   

7. Validate runSubagent is available. If not, print RW_ENV_UNSUPPORTED and stop.
8. Validate askQuestions is available. If not, print INTERVIEW_REQUIRED and stop.

Phase 2: Plan

Load full planner contract: planner-contract.md

Quick Reference

1. Step 0 Guard: Validate .ai/CONTEXT.md, check for .ai/PAUSE.md, verify runSubagent + askQuestions.
2. Hybrid Intake (via askQuestions):
   • Phase A: Resolve TARGET_KIND, USER_PATH, SCOPE_BOUNDARY, PLANNING_PROFILE, ACCEPTANCE_SIGNAL(Gate A/B), REFERENCE_BASELINE/PARITY_REQUIRED from request. Ask only for missing fields.
   • Phase B: Deep-dive (6–10 questions) if ambiguity remains.
   • Phase C: Confirmation gate (one yes/no).
   • Phase D: Ambiguity scoring (0–100 rubric) → select PLAN_STRATEGY (SINGLE or PARALLEL_AUTO).
   • Phase E: Subagent planning — generate plan candidates via runSubagent, confirm selection.
3. Feature File: Create/update under .ai/features/ with approval metadata + SHA-256 hash integrity.
4. Approval Gate: Feature must have Approval: APPROVED. Hash mismatch resets approval.
5. Plan Artifacts: plan-summary.yaml, task-graph.yaml, research_findings_*.yaml under .ai/plans/<PLAN_ID>/ (task-graph nodes must start with status: pending). If parity is required, reference-parity.md is mandatory (use assets/reference-parity-template.md).
6. Task Decomposition: Create 2–6 atomic TASK-XX-*.md in .ai/tasks/ with frontmatter status: pending, acceptance criteria, user path, and fast task-scoped verification commands. Same-phase tasks must be independent by default; add dependency edges only with explicit rationale. Put full regression commands in TASK-00 phase/final gate policy, define Environment Preflight Requirements + runtime evidence expectations for user-visible tasks, and map user-visible tasks to parity element IDs when parity is required.
7. Update Progress: Append task rows to .ai/PROGRESS.md as pending.

Planner Output Contract (success)

FEATURE_FILE=<path>
FEATURE_KEY=<JIRA-123|FEATURE-XX>
FEATURE_STATUS=PLANNED
PLAN_ID=<id>
PLAN_STRATEGY=<SINGLE|PARALLEL_AUTO>
AMBIGUITY_SCORE=<0-100>
AMBIGUITY_REASONS=<comma-separated-codes>
PLAN_MODE=<INITIAL|REPLAN|EXTENSION>
PLAN_TASK_RANGE=<TASK-XX~TASK-YY>
TASK_BOOTSTRAP_FILE=<path>
TASK_GRAPH_FILE=<path>
PLAN_RISK_LEVEL=<LOW|MEDIUM|HIGH>
PLAN_CONFIDENCE=<HIGH|MEDIUM|LOW>
OPEN_QUESTIONS_COUNT=<n>
NEXT_COMMAND=rw-loop

After emitting the success output contract, stop planner execution immediately. Do not start implementation, do not dispatch rw-loop internally, and do not modify product code.

Failure Handling

Language Policy

• .ai/CONTEXT.md 부트스트랩 기본값: prose는 한국어, 머신 토큰은 영어.
• 모든 .ai/** 아티팩트 작성 전 .ai/CONTEXT.md를 먼저 읽을 것.