Kitchen Loop: User-Spec-Driven Development로 만드는 자가 진화 코드베이스

# Kitchen Loop UAT Gate 핵심 패턴 예시

## 1. Sealed Test Card 형식 (구현 에이전트가 작성)
```
## Test Card: Backtest Service API

Step 1:
  command: curl -X POST http://localhost:8000/api/v1/backtest -H 'Content-Type: application/json' -d '{"strategy": "simple_ma"}'
  expected_exit_code: 0
  expected_output_contains: '"job_id"'

Step 2 (poll for completion):
  command: curl http://localhost:8000/api/v1/backtest/{job_id}
  expected_exit_code: 0
  expected_output_contains: '"status": "completed"'
  expected_output_contains: '"pnl"'

Step 3 (verify rejection of bad input):
  command: curl -X POST http://localhost:8000/api/v1/backtest -d '{"strategy": ""}'
  expected_exit_code: 0
  expected_output_contains: '"error"'
```

## 2. Fresh Evaluator 실행 (Haiku 같은 약한 모델, 제로 컨텍스트)
```python
def run_uat_gate(test_card_path: str, worktree_path: str) -> dict:
    """
    - Information wall: evaluator gets ONLY the test card
    - Weak model: use cheapest/weakest available
    - Read-only: cannot modify product files
    """
    evaluator_prompt = f"""
    You are a fresh user with zero context about the implementation.
    Execute EVERY step in this test card exactly as written.
    Do NOT modify any product files.
    Report exact command output and exit codes.
    
    Test Card:
    {open(test_card_path).read()}
    """
    
    result = run_agent(
        model="claude-haiku",  # weakest available
        prompt=evaluator_prompt,
        working_dir=worktree_path,
        read_only=True
    )
    
    # Anti-cheat: verify no product files were modified
    git_diff = subprocess.run(
        ["git", "diff", "--name-only"],
        cwd=worktree_path, capture_output=True, text=True
    ).stdout.strip()
    
    if git_diff:
        return {"verdict": "EVAL_CHEAT_FAIL", "modified_files": git_diff}
    
    return parse_verdict(result)

# 가능한 verdict:
# PASS → PR 머지 진행
# PRODUCT_FAIL → 티켓 오픈 유지, PR에 uat-failed 태그
# UAT_SPEC_FAIL → 테스트 카드 자체가 불명확
# EVAL_CHEAT_FAIL → 평가자가 파일 수정 시도 (심각한 문제)
```

## 3. 3-Tier 시나리오 배분 프롬프트
```
System: You are the Kitchen Loop ideation agent.
Select the next test scenario using this distribution:
- 30% Foundation (T1): Single feature, happy path, must always work
- 50% Composition (T2): Combine 2+ features, find seam bugs  
- 20% Frontier (T3): Reach beyond current capabilities, produce gap analysis

Current spec surface: {spec_matrix}
Blocked combos (known broken, skip): {blocked_combos}
Recent iterations: {last_5_scenarios}

Select ONE scenario and document:
1. Usage scenario (what a real user would try)
2. Expected behavior
3. How to verify it worked (state deltas)
```