Pre-action boundary checking — validates agent tool calls against declared capabilities and task contracts
日本語の概要は準備中です。原文の説明を表示しています。
Semble (MinishLab) MCP wrapper — semantic code search via embeddings (98% token reduction vs grep+read, NDCG@10=0.854)
インストール方法を見るインストールする前に、エージェントに与えられる指示の中身を確認できます。
MinishLab/semble MCP 서버의 의미 기반 코드 검색 도구를 노출하는 wrapper 스킬. 코드 인식 청킹 + 정적 임베딩 + BM25 + RRF 조합으로 transformer 수준 정확도를 CPU에서 달성한다.
Semble과 CRG는 서로 다른 검색 축을 담당하며 상호 보완적이다.
| 도구 | 검색 유형 | 강점 | 약점 |
|---|---|---|---|
crg-integration | 구조 (AST) | 호출 그래프, 영향 분석, recall-우선 | 의미 검색 약함 |
semble-integration | 의미 (임베딩) | 자연어 쿼리, "비슷한 코드 찾기" | 구조 정보 없음 |
사용 가이드:
query_graph)get_impact_radius)Semble MCP 서버가 제공하는 도구를 통해 다음 기능을 수행한다. 구체적인 tool 이름은 MCP 서버 설치 후 Claude Code MCP introspection으로 확인한다.
| 기능 | 설명 |
|---|---|
| 의미 검색 | 자연어 쿼리로 코드베이스 내 관련 청크 반환 |
| 인덱싱 | 코드베이스를 code-aware 청킹 + 정적 임베딩으로 인덱싱 (CPU ~250ms) |
| 결과 ranking | BM25 + RRF + embedding fusion으로 최종 랭킹 결정 |
자세한 tool 사양은
guides/token-efficiency/semble.md참조.
R001 정책: auto-install 스크립트 금지. 아래 명령어는 사용자가 직접 실행한다.
# uv 기반 설치 (권장)
uv tool install semble
# 방법 A: claude mcp 명령어 (권장)
claude mcp add semble -- semble mcp
# 방법 B: .mcp.json 수동 편집
# 프로젝트 루트 또는 ~/.claude/.mcp.json 직접 수정
.mcp.json 수동 설정 예시:
{
"mcpServers": {
"semble": {
"command": "semble",
"args": ["mcp"]
}
}
}
auto-install hook 또는 스크립트로 .mcp.json을 자동 수정하는 방식은 R001 위반이다.
# 명시적 인덱싱 (선택 — 첫 MCP 호출 시 lazy 인덱싱도 가능)
semble index
자세한 설치/설정 절차는 guides/token-efficiency/semble.md 참조.
"인증 미들웨어 관련 코드 찾아줘"
→ Semble 의미 검색 → 관련 청크 반환 → 전체 파일 읽기 없이 컨텍스트 주입
"에러 핸들링 컨벤션이 어떻게 되어 있어?"
→ Semble "error handling convention" 쿼리 → 패턴 예시 청크 반환
1. Semble로 유사 패턴 사전 조사
2. /dev-review {focused_files} → 범위를 좁힌 리뷰
1. Semble "authentication related code" → 관련 파일 목록 식별
2. CRG get_impact_radius(identified_files) → 영향 범위 확인
3. 두 결과를 조합하여 완전한 컨텍스트 구성
R013 ecomode가 자동 활성화되는 시점 (context 사용량 ≥ 60%)에서 Semble을 적극 활용한다:
| Ecomode 상황 | Semble 활용법 |
|---|---|
| 광범위 grep 대신 | Semble 의미 검색 → 관련 청크만 추출 |
| 파일 전체 읽기 대신 | Semble으로 관련 섹션만 수신 |
| 패턴 파악 시 | Semble 쿼리 → 대표 예시만 수집 |
| 자산 | 조합 방식 |
|---|---|
crg-integration | Semble (의미) + CRG (구조) 결합으로 완전한 코드 이해 |
dev-review | Semble로 유사 패턴 사전 조사 → dev-review 입력 최적화 |
adversarial-review | Semble로 관련 보안 패턴 수집 → adversarial-review 컨텍스트 보강 |
| 기존 방식 | Semble 대체 가능 여부 |
|---|---|
claude-mem:smart-explore 의 자연어 탐색 | 일부 대체 가능 — 구조 탐색은 CRG 유지 |
| grep + Read 조합의 패턴 검색 | 대체 권장 (98% 토큰 절감) |
| 증상 | 원인 | 해결 |
|---|---|---|
| MCP 도구가 보이지 않음 | .mcp.json 미설정 또는 claude mcp add 미실행 | MCP 등록 확인, Claude Code 재시작 |
command not found: semble | uv tool install 미실행 또는 PATH 누락 | uv tool install semble 재실행, which semble 확인 |
| 인덱싱 오류 | 초기 인덱싱 미완료 | semble index 실행 후 재시도 |
| 검색 결과 없음 | 인덱싱 범위 외 또는 쿼리 불명확 | 인덱싱 재실행, 쿼리 구체화 |
| 서버 연결 실패 | Semble MCP 서버 미실행 | MCP 서버 로그 확인, Claude Code 재시작 |
| uv 미설치 | uv 패키지 매니저 없음 | curl -LsSf https://astral.sh/uv/install.sh | sh (사용자 직접 실행) |
guides/token-efficiency/semble.mdcrg-integration (구조 기반 검색)まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Pre-action boundary checking — validates agent tool calls against declared capabilities and task contracts
日本語の概要は準備中です。原文の説明を表示しています。
Auto-detect project context and optimize harness — deactivate unused agents/skills, suggest missing experts, generate project profile
日本語の概要は準備中です。原文の説明を表示しています。
Adversarial code review using attacker mindset — trust boundary, attack surface, business logic, and defense evaluation
日本語の概要は準備中です。原文の説明を表示しています。
Apache Airflow best practices for DAG authoring, testing, and production deployment
日本語の概要は準備中です。原文の説明を表示しています。
Alembic migration patterns for naming conventions, safety checks, expand-contract, env.py configuration, and CI integration
日本語の概要は準備中です。原文の説明を表示しています。
Pre-routing ambiguity analysis — scores request clarity and asks clarifying questions when needed (inspired by ouroboros)
日本語の概要は準備中です。原文の説明を表示しています。