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

bash

# 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

•Load draft file from .pilot/plan/draft/ (reuse existing or create new)
•Scan conversation for User Requirements (UR-1, UR-2, ...) and Decisions (D-1, D-2, ...)
•Cross-check draft vs conversation, flag MISSING items
•Resolve omissions using AskUserQuestion
•Extract conversation highlights (code examples, diagrams, CLI commands)
•Verify 100% requirements coverage (UR → SC mapping)
•Verify scope completeness (scope vs SC mapping, assumptions verified)
•Run Self-Contained verification (9-point checklist)
•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

confirm-plan