SKILL: Confirm Plan (Plan Confirmation Workflow)

Purpose: Extract plan from conversation, create plan file, run auto-review with Interactive Recovery Target: Plan-Reviewer Agent confirming plans after /00_plan

Quick Start

When to Use This Skill

• Confirm plan after /00_plan completes
• Create plan file in .pilot/plan/draft/
• Run auto-review with Interactive Recovery for BLOCKING findings
• Verify 100% requirements coverage

Quick Reference

# TaskList Pattern: TaskUpdate (analyzing → completed) + TaskCreate (confirming)
TaskUpdate "analyzing" "completed"
TaskCreate "confirming" "in_progress"

# Extract plan from conversation
PLAN_CONTENT=$(extract_from_conversation "$CONVERSATION")

# Generate plan file name
TS="$(date +%Y%m%d_%H%M%S)"
PLAN_FILE="$PROJECT_ROOT/.pilot/plan/draft/${TS}_{work_name}.md"

# Requirements verification (BLOCKING if incomplete)
verify_requirements_coverage "$PLAN_CONTENT"

# Auto-review with Interactive Recovery
invoke_plan-reviewer "$PLAN_FILE"
resolve_blocking_findings "$PLAN_FILE"

What This Skill Covers

In Scope

• Dual-source extraction: Load decisions from draft file + scan conversation
• Cross-check verification: Compare draft vs conversation, detect omissions
• Plan file creation with full template structure
• Requirements verification (100% coverage required)
• Conversation highlights extraction (code examples, diagrams)
• Auto-review with Interactive Recovery for BLOCKING findings
• GPT delegation for large plans (5+ SCs)

Out of Scope

• Plan creation → /00_plan command
• Plan execution → /02_execute command
• TDD methodology → @.claude/skills/tdd/SKILL.md

Core Philosophy

No Execution: Only creates plan file and reviews | Context-Driven: Extract from conversation | English Only: Plan MUST be in English | Strict Mode Default: BLOCKING → Interactive Recovery

Definition of Done

Functional Completion:

• All SC implemented
• All tests passing
• Coverage >= 80%

Quality Gates:

• Type check: 0 errors
• Lint: 0 violations
• E2E verification passed

Delivery Contract:

• No partial completion allowed
• No "demo/skeleton/basic" versions
• Only explicit deferral with user approval permitted

⚠️ EXECUTION DIRECTIVE

IMPORTANT: Execute ALL steps below IMMEDIATELY and AUTOMATICALLY without waiting for user input.

• Do NOT pause between steps
• Do NOT ask "should I continue?" or wait for "keep going"
• Execute Step 1 → 2 → 2.5 → 3 → 4 in sequence
• Only stop for BLOCKING findings that require Interactive Recovery

Execution Steps (Summary)

Step 1: Dual-Source Extraction

1. Load draft file from .pilot/plan/draft/ (reuse existing or create new)
2. Scan conversation for User Requirements (UR-1, UR-2, ...) and Decisions (D-1, D-2, ...)
3. Cross-check draft vs conversation, flag MISSING items
4. Resolve omissions using AskUserQuestion
5. Extract conversation highlights (code examples, diagrams, CLI commands)
6. Verify 100% requirements coverage (UR → SC mapping)
7. Verify scope completeness (scope vs SC mapping, assumptions verified)
8. Run Self-Contained verification (9-point checklist)
9. SC Granularity Check: Verify each SC follows Atomic SC Principle ("One SC = One File OR One Concern")

BLOCKING if: Requirements incomplete, scope gaps, or self-contained check fails

Step 2: Create or Update Plan File

• Use absolute path based on Claude Code's initial working directory
• Reuse existing draft or create new one
• Apply full plan template (User Requirements, Context Pack, Success Criteria, PRP Analysis, etc.)
• Include Context Pack Formats and Zero-Knowledge TODO Format sections

Step 2.5: GPT Delegation Check (Mandatory Oracle Consultation)

• Trigger: Large plans (5+ Success Criteria) OR architecture keywords in plan
• Architecture keywords: architecture, tradeoff, design, scalability, pattern, choice
• Oracle Role: Auditor - Plan quality audit
• Delegate to GPT Plan Reviewer using codex CLI
• Graceful fallback if Codex not installed

Mandatory Oracle Consultation: All plans receive quality audit via GPT Auditor to verify requirements coverage, SC clarity, dependencies, and risks before confirmation.

Step 3: Auto-Review & Auto-Apply

• Invoke plan-reviewer agent
• Review criteria: requirements coverage, SC clarity, dependencies, risks
• Output: <PLAN_COMPLETE> or <PLAN_BLOCKED>
• BLOCKING → Interactive Recovery loop (max 5 iterations)
• Auto-apply Critical, Warning, Suggestion findings

Step 4: Move to pending

• Move plan file to .pilot/plan/pending/
• STOP: Do NOT proceed to /02_execute automatically

Step 4.5: Cleanup Intermediate Files

• Delete matching *_draft.md and *_context_pack.md files from draft/
• Match by work_name extracted from confirmed plan filename
• Log deleted files for audit trail

⛔ MAIN ORCHESTRATOR RESTRICTIONS (ABSOLUTE)

FORBIDDEN (orchestrator direct use prohibited):

• Direct plan validation without Task tool
• Direct BLOCKING gate resolution without agent delegation

MANDATORY (must delegate via Task tool):

• Step 1.6/1.7/1.9 verification: Task: subagent_type: plan-reviewer before user escalation
• Step 2.5 GPT review: Via gpt-delegation skill
• Step 3 auto-review: Task: subagent_type: plan-reviewer

TRIVIAL EXCEPTIONS (no delegation needed):

• Dual-source extraction scanning, file operations, conversation parsing

WHY: Subagent isolation provides 50-80% context savings (CLAUDE.md:58-59)

Argument Parsing

Parse $ARGUMENTS from command invocation:

• [work_name]: Optional work name for plan file
• --lenient: Bypass BLOCKING findings
• --no-review: Skip all review steps

Further Reading

Internal: @.claude/skills/confirm-plan/REFERENCE.md - Detailed implementation, Context Pack formats, Zero-Knowledge TODO format, Self-Contained verification checklist | @.claude/skills/spec-driven-workflow/SKILL.md - SPEC-First methodology (Problem-Requirements-Plan)

External: Specification by Example | User Stories Applied

⚠️ MANDATORY: This skill only creates plan. Run /02_execute to implement.