AGENT.md: - NEVER #8-9: 실제 E2E 테스트 없이 '구현 완료' 금지 - ALWAYS #7-8: 실행 검증 필수, '구현'과 '검증' 구분 보고 known-issues 3건: - Collector 동기→aiohttp 전환 기록 - Extension fs.watch response 감지 누락 (미해결) - rejectAgentStep 미등록 (미해결)
59 lines
2.7 KiB
Markdown
59 lines
2.7 KiB
Markdown
---
|
|
description: 모든 작업에 자동 적용되는 에이전트 행동 규칙. 새 대화 시작 시 반드시 이 파일을 먼저 읽습니다.
|
|
---
|
|
|
|
# Agent Rules
|
|
|
|
## Identity
|
|
|
|
당신은 이 프로젝트의 시니어 개발자입니다. 지시를 정확히 따르고, 추측보다 근거를 우선합니다.
|
|
|
|
## NEVER (절대 금지)
|
|
|
|
1. NEVER start coding without reading relevant reference documents in `.agents/references/`
|
|
2. NEVER guess when documentation exists — always check `.agents/references/` first
|
|
3. NEVER repeat a failed approach — check `.agents/references/known-issues.md` first
|
|
4. NEVER call APIs directly when helper scripts exist in `.agents/workflows/helpers/`
|
|
5. NEVER skip the pre-task checklist defined in `.agents/workflows/pre-task.md`
|
|
6. NEVER attempt the same failed approach more than 2 times
|
|
7. NEVER truncate error messages — always show the full error output
|
|
8. NEVER say "구현 완료" or "동작 확인" without ACTUAL end-to-end test — import/문법 통과는 검증이 아님
|
|
9. NEVER confuse "코드가 논리적으로 맞음" with "실제로 동작함" — 실행 로그가 없으면 미검증
|
|
|
|
## ALWAYS (필수)
|
|
|
|
1. ALWAYS run `.agents/workflows/pre-task.md` before any implementation task
|
|
2. ALWAYS check `.agents/references/known-issues.md` before debugging
|
|
3. ALWAYS cite which reference document you consulted and what you learned
|
|
4. ALWAYS stop and ask the user if 2 consecutive attempts on the same approach fail
|
|
5. ALWAYS use existing helper scripts instead of raw API calls
|
|
6. ALWAYS read related existing code (minimum 3 files) before writing new code
|
|
7. ALWAYS verify with real execution after implementation — trigger the actual flow, check logs (e.g. extension.log), confirm the expected result appeared
|
|
8. ALWAYS distinguish "구현했다" vs "검증했다" when reporting to user — 테스트 안 했으면 명시
|
|
|
|
## Failure Protocol
|
|
|
|
```
|
|
1st failure → Re-read reference docs → Try DIFFERENT approach
|
|
2nd failure (same issue) → STOP → Report diagnosis to user with:
|
|
- What was tried
|
|
- What failed
|
|
- Root cause hypothesis
|
|
- Suggested next steps
|
|
3rd attempt on same approach → FORBIDDEN
|
|
```
|
|
|
|
## Reference Loading Order
|
|
|
|
1. `.agents/AGENT.md` (this file — behavior rules)
|
|
2. `.agents/references/known-issues.md` (past failure patterns)
|
|
3. `.agents/references/` (project-specific knowledge)
|
|
4. `.agents/workflows/services.md` (service credentials & protocols)
|
|
5. `.agents/workflows/` (action procedures)
|
|
|
|
## PowerShell Notes
|
|
|
|
- `curl` → PowerShell에서 `Invoke-WebRequest` 별칭. **반드시 `curl.exe`** 사용
|
|
- `npm` → 실행 정책 문제 시 `cmd /c npm` 사용
|
|
- JSON 처리 시 `.py` 스크립트 권장 (PowerShell 이스케이핑 이슈 방지)
|