~/.claude/
└── skills/
    └── pm/                              # PM Agent Skill
        ├── implementation.md            # ~25KB - 全ワークフロー
        ├── modules/
        │   ├── git-status.md            # ~5KB - Git状態フォーマット
        │   ├── token-counter.md         # ~8KB - トークンカウント
        │   └── pm-formatter.md          # ~10KB - ステータス出力
        └── workflows/
            └── task-management.md       # ~15KB - タスク管理

superclaude/
├── agents/
│   └── pm-agent.md                      # ~50KB - Agent定義
├── commands/
│   └── pm.md                            # ~5KB - /sc:pm command
└── core/
    └── pm_init/                         # Python init hooks
        ├── __init__.py
        ├── context_contract.py          # ~10KB - Context管理
        ├── init_hook.py                 # ~10KB - Session start
        └── reflexion_memory.py          # ~12KB - Reflexion

Total: ~150KB ≈ 35K-40K tokens

src/superclaude/
└── pm_agent/                            # Python package
    ├── __init__.py                      # Package exports
    ├── confidence.py                    # ~8KB - Pre-execution
    ├── self_check.py                    # ~15KB - Post-validation
    ├── reflexion.py                     # ~12KB - Error learning
    └── token_budget.py                  # ~10KB - Budget management

tests/pm_agent/
├── test_confidence_check.py             # 18 tests
├── test_self_check_protocol.py          # 16 tests
├── test_reflexion_pattern.py            # 16 tests
└── test_token_budget.py                 # 29 tests

Total: ~45KB ≈ 10K-12K tokens (import時のみ)

Trigger: EVERY session start (自動)
Method: pm_init/init_hook.py

Actions:
  1. PARALLEL Read:
     - docs/memory/pm_context.md
     - docs/memory/last_session.md
     - docs/memory/next_actions.md
     - docs/memory/current_plan.json
  2. Confidence Check (200 tokens)
  3. Output: 🟢 [branch] | [n]M [n]D | [token]%

Token Cost: ~8K (memory files) + 200 (confidence)

# 自動実行なし - 手動で呼び出し
from superclaude.pm_agent.confidence import ConfidenceChecker

checker = ConfidenceChecker()
confidence = checker.assess(context)

Token Cost: ~2K (confidence moduleのみ)

# superclaude/agents/pm-agent.md より

Confidence Check (200 tokens):
  ❓ "全ファイル読めた？"
  ❓ "コンテキストに矛盾ない？"
  ❓ "次のアクション実行に十分な情報？"

Output: Markdown形式
Location: Agent definition内

# src/superclaude/pm_agent/confidence.py

class ConfidenceChecker:
    def assess(self, context: Dict[str, Any]) -> float:
        """
        Assess confidence (0.0-1.0)

        Checks:
        1. Documentation verified? (40%)
        2. Patterns identified? (30%)
        3. Implementation clear? (30%)

        Budget: 100-200 tokens
        """
        # Python実装
        return confidence_score

# agents/pm-agent.md より

Self-Evaluation Checklist:
  - [ ] Did I follow architecture patterns?
  - [ ] Did I read documentation first?
  - [ ] Did I check existing implementations?
  - [ ] Are all tasks complete?
  - [ ] What mistakes did I make?
  - [ ] What did I learn?

Token Budget:
  Simple: 200 tokens
  Medium: 1,000 tokens
  Complex: 2,500 tokens

Output: docs/pdca/[feature]/check.md

# src/superclaude/pm_agent/self_check.py

class SelfCheckProtocol:
    def validate(self, implementation: Dict[str, Any])
        -> Tuple[bool, List[str]]:
        """
        Four Questions Protocol:
        1. All tests pass?
        2. Requirements met?
        3. Assumptions verified?
        4. Evidence exists?

        7 Hallucination Red Flags detection

        Returns: (passed, issues)
        """
        # Python実装

# superclaude/core/pm_init/reflexion_memory.py

class ReflexionMemory:
    """
    Error learning with dual storage:
    1. Local JSONL: docs/memory/solutions_learned.jsonl
    2. Mindbase: Semantic search (if available)

    Lookup: mindbase → grep fallback
    """

# src/superclaude/pm_agent/reflexion.py

class ReflexionPattern:
    """
    Same dual storage strategy:
    1. Local JSONL: docs/memory/solutions_learned.jsonl
    2. Mindbase: Semantic search (optional)

    Methods:
    - get_solution(error_info) → past solution lookup
    - record_error(error_info) → save to memory
    - get_statistics() → recurrence rate
    """

# agents/pm-agent.md より

Token Budget (Complexity-Based):
  Simple Task (typo): 200 tokens
  Medium Task (bug): 1,000 tokens
  Complex Task (feature): 2,500 tokens

Implementation: Markdown定義のみ
Enforcement: 手動

# src/superclaude/pm_agent/token_budget.py

class TokenBudgetManager:
    BUDGETS = {
        "simple": 200,
        "medium": 1000,
        "complex": 2500,
    }

    def use(self, tokens: int) -> bool:
        """Track usage"""

    @property
    def remaining(self) -> int:
        """Get remaining budget"""

    def get_recommendation(self) -> str:
        """Suggest optimization"""

EVERY session start:
  - Auto-read memory files
  - Auto-restore context
  - Auto-output status

# Manual activation required
from superclaude.pm_agent.confidence import ConfidenceChecker
checker = ConfidenceChecker()

Auto-generate:
  - docs/pdca/[feature]/plan.md
  - docs/pdca/[feature]/do.md
  - docs/pdca/[feature]/check.md
  - docs/pdca/[feature]/act.md

# なし - ユーザーが手動で記録

# workflows/task-management.md
- TodoWrite auto-tracking
- Progress checkpoints
- Session continuity

# TodoWriteはClaude Codeネイティブツールとして利用可能
# PM Agent特有のワークフローなし

# Core PM Agent (This PR) - always installed
pip install -e .

# Skills PM Agent (Upstream) - optional
superclaude install-skill pm-agent

# 本家Skills版のみ使用
superclaude install-skill pm-agent

# Pytest fixturesは使わない

# This PRのみ
pip install -e .

# Skillsなし