Skip to content

Latest commit

Β 

History

History
398 lines (280 loc) Β· 18 KB

File metadata and controls

398 lines (280 loc) Β· 18 KB

Specialist Agents Reference (μ „λ¬Έκ°€ μ—μ΄μ „νŠΈ μ°Έμ‘°)

This document lists every specialist agent available in the /coord pipeline, with its role, activation condition, subagent_type value, and a concrete invocation example.

이 λ¬Έμ„œλŠ” /coord νŒŒμ΄ν”„λΌμΈμ—μ„œ μ‚¬μš© κ°€λŠ₯ν•œ λͺ¨λ“  μ „λ¬Έκ°€ μ—μ΄μ „νŠΈλ₯Ό μ—­ν• , ν™œμ„±ν™” 쑰건, subagent_type κ°’, ꡬ체적인 호좜 μ˜ˆμ‹œμ™€ ν•¨κ»˜ λ‚˜μ—΄ν•©λ‹ˆλ‹€.


How Agents Are Invoked (μ—μ΄μ „νŠΈ 호좜 방법)

Agents are dispatched using Claude Code's subagent_type parameter, either by the pipeline automatically or by the user directly via the @sub shorthand.

μ—μ΄μ „νŠΈλŠ” νŒŒμ΄ν”„λΌμΈμ΄ μžλ™μœΌλ‘œ λ˜λŠ” μ‚¬μš©μžκ°€ @sub 단좕어λ₯Ό 톡해 직접 subagent_type λ§€κ°œλ³€μˆ˜λ₯Ό μ‚¬μš©ν•˜μ—¬ λ°œμ†‘λ©λ‹ˆλ‹€.

# Internal pipeline dispatch (νŒŒμ΄ν”„λΌμΈ λ‚΄λΆ€ λ°œμ†‘)
{
  "subagent_type": "explore",
  "prompt": "Find all authentication-related files in the repository",
  "run_in_background": True
}

# User shorthand (μ‚¬μš©μž 단좕어)
@sub Find all authentication-related files in the repository

The pipeline selects the right agent automatically based on task classification from Stage 2. Manual @sub invocations bypass the pipeline and run the agent directly.

νŒŒμ΄ν”„λΌμΈμ€ Stage 2의 μž‘μ—… λΆ„λ₯˜λ₯Ό 기반으둜 μžλ™μœΌλ‘œ μ˜¬λ°”λ₯Έ μ—μ΄μ „νŠΈλ₯Ό μ„ νƒν•©λ‹ˆλ‹€. μˆ˜λ™ @sub ν˜ΈμΆœμ€ νŒŒμ΄ν”„λΌμΈμ„ μš°νšŒν•˜κ³  μ—μ΄μ „νŠΈλ₯Ό 직접 μ‹€ν–‰ν•©λ‹ˆλ‹€.


Discovery and Exploration Agents (탐색 μ—μ΄μ „νŠΈ)

explore

Role: Codebase discovery, file structure mapping, and pattern identification. Use when the task requires understanding what already exists before making changes.

μ—­ν• : μ½”λ“œλ² μ΄μŠ€ 탐색, 파일 ꡬ쑰 λ§€ν•‘, νŒ¨ν„΄ 식별. λ³€κ²½ν•˜κΈ° 전에 무엇이 μ‘΄μž¬ν•˜λŠ”μ§€ 이해해야 ν•  λ•Œ μ‚¬μš©ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Any task involving 3 or more search queries (3개 μ΄μƒμ˜ 검색 쿼리가 ν¬ν•¨λœ λͺ¨λ“  μž‘μ—…)
  • "How is X implemented in this codebase?" questions (μ½”λ“œλ² μ΄μŠ€μ—μ„œ Xκ°€ μ–΄λ–»κ²Œ κ΅¬ν˜„λ˜λŠ”μ§€ 질문)
  • Before refactoring to understand current architecture (ν˜„μž¬ μ•„ν‚€ν…μ²˜λ₯Ό μ΄ν•΄ν•˜κΈ° μœ„ν•΄ λ¦¬νŒ©ν† λ§ μ „)

Example (μ˜ˆμ‹œ):

@sub Find all places in the codebase where database connections are opened,
     and check whether they are properly closed in finally blocks.

deep-research-agent

Role: Multi-source web research, synthesis across academic papers, blog posts, and official documentation. Returns a structured report with citations.

μ—­ν• : 닀쀑 μ†ŒμŠ€ μ›Ή λ¦¬μ„œμΉ˜, ν•™μˆ  λ…Όλ¬Έ, λΈ”λ‘œκ·Έ 포슀트, 곡식 λ¬Έμ„œ κ°„ ν•©μ„±. 인용과 ν•¨κ»˜ κ΅¬μ‘°ν™”λœ λ³΄κ³ μ„œλ₯Ό λ°˜ν™˜ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Comparing approaches to solve a technical problem (기술적 문제 ν•΄κ²° 접근법 비ꡐ)
  • Finding recent (post-knowledge-cutoff) information (지식 μ»·μ˜€ν”„ 이후 μ΅œμ‹  정보 μ°ΎκΈ°)
  • Literature survey for a research paper or proposal (λ…Όλ¬Έ λ˜λŠ” μ œμ•ˆμ„œλ₯Ό μœ„ν•œ λ¬Έν—Œ 쑰사)

Example (μ˜ˆμ‹œ):

@sub Research the current state of transformer model compression techniques
     for edge deployment. Compare pruning, quantization, and distillation
     approaches. Include papers from 2024 onward.

Planning Agents (κ³„νš μ—μ΄μ „νŠΈ)

plan

Role: Implementation strategy design. Produces a structured plan with dependency mapping, parallelization annotations, and resource estimates. Should run before any implementation work begins.

μ—­ν• : κ΅¬ν˜„ μ „λž΅ 섀계. μ˜μ‘΄μ„± λ§€ν•‘, 병렬화 주석, λ¦¬μ†ŒμŠ€ 좔정이 μžˆλŠ” κ΅¬μ‘°ν™”λœ κ³„νšμ„ μƒμ„±ν•©λ‹ˆλ‹€. κ΅¬ν˜„ μž‘μ—…μ΄ μ‹œμž‘λ˜κΈ° 전에 μ‹€ν–‰ν•΄μ•Ό ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Before implementing any feature with 3+ steps (3단계 μ΄μƒμ˜ κΈ°λŠ₯ κ΅¬ν˜„ μ „)
  • When requirements are clear but the execution order is not (μš”κ΅¬μ‚¬ν•­μ€ λͺ…ν™•ν•˜μ§€λ§Œ μ‹€ν–‰ μˆœμ„œκ°€ 뢈λͺ…ν™•ν•  λ•Œ)

Example (μ˜ˆμ‹œ):

@sub Design an implementation plan for adding rate limiting to the FastAPI
     service. The plan must identify which components need changes, the order
     of changes, and which steps can be parallelized.

requirements-analyst

Role: Ambiguous request clarification. Converts vague user intents into precise, numbered requirements with acceptance criteria. Prevents scope creep by surfacing implicit assumptions.

μ—­ν• : λͺ¨ν˜Έν•œ μš”μ²­ λͺ…μ„Έν™”. λͺ¨ν˜Έν•œ μ‚¬μš©μž μ˜λ„λ₯Ό 수용 기쀀이 μžˆλŠ” μ •ν™•ν•œ 번호 맀겨진 μš”κ΅¬μ‚¬ν•­μœΌλ‘œ λ³€ν™˜ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • When the user's request has multiple valid interpretations (μ‚¬μš©μž μš”μ²­μ— μ—¬λŸ¬ μœ νš¨ν•œ 해석이 μžˆμ„ λ•Œ)
  • Before starting any project where the scope is unclear (λ²”μœ„κ°€ 뢈λͺ…ν™•ν•œ ν”„λ‘œμ νŠΈ μ‹œμž‘ μ „)

Example (μ˜ˆμ‹œ):

@sub Clarify this requirement: "Make the dashboard faster."
     Identify: which dashboard, which metrics define "faster",
     what the current baseline is, and what success looks like.

Architecture Agents (μ•„ν‚€ν…μ²˜ μ—μ΄μ „νŠΈ)

system-architect

Role: Full-system cross-domain design. Produces architecture diagrams (ASCII), component responsibility assignments, data flow descriptions, and integration contracts.

μ—­ν• : 전체 μ‹œμŠ€ν…œ 도메인 κ°„ 섀계. ASCII μ•„ν‚€ν…μ²˜ λ‹€μ΄μ–΄κ·Έλž¨, μ»΄ν¬λ„ŒνŠΈ μ±…μž„ ν• λ‹Ή, 데이터 흐름 μ„€λͺ…, 톡합 계약을 μƒμ„±ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Designing a new system from scratch (μƒˆ μ‹œμŠ€ν…œ μ²˜μŒλΆ€ν„° 섀계)
  • Evaluating a proposed architecture change (μ œμ•ˆλœ μ•„ν‚€ν…μ²˜ λ³€κ²½ 평가)
  • Cross-cutting concerns spanning frontend + backend + infra (ν”„λ‘ νŠΈμ—”λ“œ + λ°±μ—”λ“œ + 인프라에 걸친 νš‘λ‹¨ 관심사)

Example (μ˜ˆμ‹œ):

@sub Design the architecture for a real-time collaborative document editor.
     Include the WebSocket layer, CRDT conflict resolution approach,
     persistence strategy, and how the frontend state synchronizes.

backend-architect

Role: Server-side system design. Specializes in API design, database schema, service decomposition, authentication flows, and caching strategies.

μ—­ν• : μ„œλ²„ μ‚¬μ΄λ“œ μ‹œμŠ€ν…œ 섀계. API 섀계, λ°μ΄ν„°λ² μ΄μŠ€ μŠ€ν‚€λ§ˆ, μ„œλΉ„μŠ€ λΆ„ν•΄, 인증 흐름, 캐싱 μ „λž΅μ„ μ „λ¬ΈμœΌλ‘œ ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Designing REST or GraphQL APIs (REST λ˜λŠ” GraphQL API 섀계)
  • Planning database schema migrations (λ°μ΄ν„°λ² μ΄μŠ€ μŠ€ν‚€λ§ˆ λ§ˆμ΄κ·Έλ ˆμ΄μ…˜ κ³„νš)
  • Service decomposition from monolith (λͺ¨λ†€λ¦¬μŠ€μ—μ„œ μ„œλΉ„μŠ€ λΆ„ν•΄)

Example (μ˜ˆμ‹œ):

@sub Design the database schema for a multi-tenant SaaS application where
     each tenant has isolated data. Choose between row-level security,
     schema-per-tenant, and database-per-tenant, and justify the choice.

frontend-architect

Role: UI/UX architecture decisions. Handles state management strategy, component hierarchy, routing architecture, rendering strategy (SSR/CSR/SSG), and accessibility compliance.

μ—­ν• : UI/UX μ•„ν‚€ν…μ²˜ κ²°μ •. μƒνƒœ 관리 μ „λž΅, μ»΄ν¬λ„ŒνŠΈ 계측, λΌμš°νŒ… μ•„ν‚€ν…μ²˜, λ Œλ”λ§ μ „λž΅(SSR/CSR/SSG), μ ‘κ·Όμ„± κ·œμ • μ€€μˆ˜λ₯Ό μ²˜λ¦¬ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Choosing a React state management library (React μƒνƒœ 관리 라이브러리 선택)
  • Designing component composition patterns (μ»΄ν¬λ„ŒνŠΈ ν•©μ„± νŒ¨ν„΄ 섀계)
  • Performance optimization for a large SPA (λŒ€κ·œλͺ¨ SPA μ„±λŠ₯ μ΅œμ ν™”)

Example (μ˜ˆμ‹œ):

@sub Evaluate whether to use Zustand, Jotai, or Redux Toolkit for
     the state management layer in a Next.js 14 app with server components.
     Consider the server/client boundary implications.

Implementation Agents (κ΅¬ν˜„ μ—μ΄μ „νŠΈ)

python-expert

Role: Python-specific implementation. Handles async patterns, type hinting, packaging, testing with pytest, and Python performance patterns.

μ—­ν• : Python νŠΉν™” κ΅¬ν˜„. 비동기 νŒ¨ν„΄, νƒ€μž… νžŒνŒ…, νŒ¨ν‚€μ§•, pytestλ₯Ό μ‚¬μš©ν•œ ν…ŒμŠ€νŠΈ, Python μ„±λŠ₯ νŒ¨ν„΄μ„ μ²˜λ¦¬ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Writing Python code that must follow strict typing conventions (μ—„κ²©ν•œ 타이핑 κ·œμΉ™μ„ 따라야 ν•˜λŠ” Python μ½”λ“œ μž‘μ„±)
  • Async FastAPI or Django endpoints (비동기 FastAPI λ˜λŠ” Django μ—”λ“œν¬μΈνŠΈ)
  • Scientific computing with NumPy, pandas, or PyTorch (NumPy, pandas, PyTorchλ₯Ό μ‚¬μš©ν•œ κ³Όν•™ 계산)

Example (μ˜ˆμ‹œ):

@sub Implement an async context manager in Python 3.12 that manages
     a connection pool for PostgreSQL using asyncpg. Include type hints,
     docstrings, and a pytest fixture that uses the manager.

refactoring-expert

Role: Code quality improvement without behavior changes. Identifies dead code, extracts functions, removes duplication, applies SOLID principles, and improves naming.

μ—­ν• : λ™μž‘ λ³€κ²½ μ—†λŠ” μ½”λ“œ ν’ˆμ§ˆ κ°œμ„ . λ°λ“œ μ½”λ“œ 식별, ν•¨μˆ˜ μΆ”μΆœ, 쀑볡 제거, SOLID 원칙 적용, 이름 κ°œμ„ μ„ μˆ˜ν–‰ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • When code works but is difficult to maintain (μ½”λ“œλŠ” μž‘λ™ν•˜μ§€λ§Œ μœ μ§€ 관리가 μ–΄λ €μšΈ λ•Œ)
  • Before adding new features to legacy code (λ ˆκ±°μ‹œ μ½”λ“œμ— μƒˆ κΈ°λŠ₯을 μΆ”κ°€ν•˜κΈ° μ „)
  • Code review cleanup (μ½”λ“œ 리뷰 정리)

Example (μ˜ˆμ‹œ):

@sub Refactor the authentication module in src/auth/. It currently has a
     500-line God class. Extract the token validation, session management,
     and OAuth flow into separate classes. Do not change behavior.

Quality and Safety Agents (ν’ˆμ§ˆ 및 μ•ˆμ „ μ—μ΄μ „νŠΈ)

critic

Role: Independent review with falsifiability checking. Every claim in the reviewed output must have a "wrong if X" condition. Raises compliance issues, logical gaps, and sycophantic reasoning.

μ—­ν• : 반증 κ°€λŠ₯μ„± 검사와 ν•¨κ»˜ 독립적인 κ²€ν† . κ²€ν† λœ 좜λ ₯의 λͺ¨λ“  μ£Όμž₯μ—λŠ” "X인 경우 ν‹€λ¦Ό" 쑰건이 μžˆμ–΄μ•Ό ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Automatically after Stage 4 when Opus model is used (Opus λͺ¨λΈ μ‚¬μš© μ‹œ Stage 4 ν›„ μžλ™)
  • Manually via /challenge command to critique the last response (λ§ˆμ§€λ§‰ 응닡을 λΉ„νŒν•˜κΈ° μœ„ν•΄ /challenge λͺ…λ ΉμœΌλ‘œ μˆ˜λ™)
  • Before finalizing any proposal, paper, or production deployment (μ œμ•ˆμ„œ, λ…Όλ¬Έ, ν”„λ‘œλ•μ…˜ 배포 ν™•μ • μ „)

Example (μ˜ˆμ‹œ):

@sub Review the proposed caching strategy below and identify:
     1. Any claims without evidence
     2. Edge cases not handled
     3. Performance assumptions that could be wrong
     [paste the caching strategy]

quality-engineer

Role: Test suite design and coverage analysis. Creates unit tests, integration tests, and property-based tests. Identifies uncovered code paths and writes assertions with meaningful failure messages.

μ—­ν• : ν…ŒμŠ€νŠΈ μŠ€μœ„νŠΈ 섀계 및 컀버리지 뢄석. λ‹¨μœ„ ν…ŒμŠ€νŠΈ, 톡합 ν…ŒμŠ€νŠΈ, 속성 기반 ν…ŒμŠ€νŠΈλ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • After implementing a new feature (μƒˆ κΈ°λŠ₯ κ΅¬ν˜„ ν›„)
  • When test coverage drops below the project threshold (ν…ŒμŠ€νŠΈ 컀버리지가 ν”„λ‘œμ νŠΈ μž„κ³„κ°’ μ•„λž˜λ‘œ λ–¨μ–΄μ§ˆ λ•Œ)
  • Property-based testing for data transformation functions (데이터 λ³€ν™˜ ν•¨μˆ˜μ˜ 속성 기반 ν…ŒμŠ€νŠΈ)

Example (μ˜ˆμ‹œ):

@sub Write a comprehensive pytest test suite for the payment processing module
     in src/payments/. Cover: successful transactions, card decline scenarios,
     network timeout handling, and idempotency of repeated calls.

security-engineer

Role: Threat modeling and vulnerability review. Applies OWASP Top 10, checks for injection vulnerabilities, reviews authentication and authorization logic, and assesses data exposure risks.

μ—­ν• : μœ„ν˜‘ λͺ¨λΈλ§ 및 취약점 κ²€ν† . OWASP Top 10 적용, μΈμ μ…˜ 취약점 확인, 인증 및 κΆŒν•œ λΆ€μ—¬ 둜직 κ²€ν† , 데이터 λ…ΈμΆœ μœ„ν—˜ 평가λ₯Ό μˆ˜ν–‰ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Before deploying any user-facing authentication feature (μ‚¬μš©μž λŒ€λ©΄ 인증 κΈ°λŠ₯ 배포 μ „)
  • When handling personally identifiable information (개인 식별 정보 처리 μ‹œ)
  • Annual security audit of a service (μ„œλΉ„μŠ€μ˜ μ—°κ°„ λ³΄μ•ˆ 감사)

Example (μ˜ˆμ‹œ):

@sub Perform a security review of the user registration and login flows
     in src/auth/. Check for: SQL injection, timing attacks on password
     comparison, missing rate limiting, and insecure password storage.

performance-engineer

Role: Profiling and optimization. Identifies algorithmic complexity issues, database N+1 queries, memory leaks, and unnecessary network round trips. Provides before/after benchmarks.

μ—­ν• : ν”„λ‘œνŒŒμΌλ§ 및 μ΅œμ ν™”. μ•Œκ³ λ¦¬μ¦˜ λ³΅μž‘λ„ 문제, λ°μ΄ν„°λ² μ΄μŠ€ N+1 쿼리, λ©”λͺ¨λ¦¬ λˆ„μˆ˜, λΆˆν•„μš”ν•œ λ„€νŠΈμ›Œν¬ λΌμš΄λ“œ νŠΈλ¦½μ„ μ‹λ³„ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • When profiling shows a hot path (ν”„λ‘œνŒŒμΌλ§μ΄ ν•« 패슀λ₯Ό ν‘œμ‹œν•  λ•Œ)
  • Unexplained latency increase after a deployment (배포 ν›„ μ„€λͺ…λ˜μ§€ μ•ŠλŠ” μ§€μ—° 증가)
  • Scaling review before a traffic spike (νŠΈλž˜ν”½ 급증 μ „ ν™•μž₯μ„± κ²€ν† )

Example (μ˜ˆμ‹œ):

@sub Analyze the performance of the report generation endpoint. It currently
     takes 8 seconds for a report covering 10,000 rows. Identify the
     bottleneck and propose a fix with estimated improvement.

Documentation Agent (λ¬Έμ„œν™” μ—μ΄μ „νŠΈ)

technical-writer

Role: Documentation creation and improvement. Writes API references, user guides, architecture docs, README files, and inline code comments. Focuses on clarity for the target audience and WCAG-compliant structure.

μ—­ν• : λ¬Έμ„œ 생성 및 κ°œμ„ . API μ°Έμ‘°, μ‚¬μš©μž κ°€μ΄λ“œ, μ•„ν‚€ν…μ²˜ λ¬Έμ„œ, README 파일, 인라인 μ½”λ“œ 주석을 μž‘μ„±ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • After implementing a new public API (μƒˆ 곡개 API κ΅¬ν˜„ ν›„)
  • Onboarding documentation for a new team member (μƒˆ νŒ€μ›μ„ μœ„ν•œ μ˜¨λ³΄λ”© λ¬Έμ„œ)
  • When existing docs are outdated or missing examples (κΈ°μ‘΄ λ¬Έμ„œκ°€ μ˜€λž˜λ˜μ—ˆκ±°λ‚˜ μ˜ˆμ‹œκ°€ 없을 λ•Œ)

Example (μ˜ˆμ‹œ):

@sub Write API reference documentation for the /payments endpoint.
     Audience: external developers integrating for the first time.
     Include: request format, all response codes, error object schema,
     and a working cURL example for each scenario.

Scientific and Diagnostic Agents (과학적 및 진단 μ—μ΄μ „νŠΈ)

sci-method

Role: Scientific hypothesis-evidence-validation workflow using Cynefin triage, Popperian falsifiability, and Bayesian updating. An 8-stage structured reasoning process for any complex problem domain.

μ—­ν• : Cynefin λΆ„λ₯˜, Popperian 반증 κ°€λŠ₯μ„±, Bayesian μ—…λ°μ΄νŠΈλ₯Ό μ‚¬μš©ν•˜λŠ” 과학적 κ°€μ„€-증거-검증 μ›Œν¬ν”Œλ‘œμš°. λͺ¨λ“  λ³΅μž‘ν•œ 문제 도메인을 μœ„ν•œ 8단계 κ΅¬μ‘°ν™”λœ μΆ”λ‘  κ³Όμ •μž…λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • Diagnosing a system behavior that is not understood (μ΄ν•΄λ˜μ§€ μ•ŠλŠ” μ‹œμŠ€ν…œ λ™μž‘ 진단)
  • Evaluating competing explanations for an experimental result (μ‹€ν—˜ 결과에 λŒ€ν•œ 경쟁적 μ„€λͺ… 평가)
  • Any problem where intuition alone is insufficient (μ§κ΄€λ§ŒμœΌλ‘œλŠ” λΆˆμΆ©λΆ„ν•œ λͺ¨λ“  문제)

8-stage workflow (8단계 μ›Œν¬ν”Œλ‘œμš°):

  1. Cynefin domain classification (Cynefin 도메인 λΆ„λ₯˜)
  2. Phenomenon description (ν˜„μƒ μ„€λͺ…)
  3. Hypothesis generation (κ°€μ„€ 생성)
  4. Evidence inventory (증거 λͺ©λ‘)
  5. Bayesian likelihood scoring (Bayesian κ°€λŠ₯μ„± 점수 λ§€κΈ°κΈ°)
  6. Decisive experiment design (결정적 μ‹€ν—˜ 섀계)
  7. Prediction registration (예츑 등둝)
  8. Update and conclusion (μ—…λ°μ΄νŠΈ 및 κ²°λ‘ )

Example (μ˜ˆμ‹œ):

@sub Apply sci-method to this anomaly: our model achieves 95% AUC on the
     validation set but only 61% on the held-out test set. Generate ranked
     hypotheses with testable predictions for each.

root-cause-analyst

Role: Systematic failure investigation. Applies the "five whys" method combined with evidence gathering. Distinguishes root causes from symptoms. Produces a causal chain diagram.

μ—­ν• : 체계적인 μ‹€νŒ¨ 쑰사. 증거 μˆ˜μ§‘κ³Ό κ²°ν•©λœ "5 Why" 방법을 μ μš©ν•©λ‹ˆλ‹€. κ·Όλ³Έ 원인과 증상을 κ΅¬λΆ„ν•©λ‹ˆλ‹€.

When to use (μ‚¬μš© μ‹œμ ):

  • After a production incident (ν”„λ‘œλ•μ…˜ μΈμ‹œλ˜νŠΈ ν›„)
  • When intermediate validation in Stage 4 fails (Stage 4의 쀑간 검증이 μ‹€νŒ¨ν•  λ•Œ)
  • Recurring bugs that have been fixed and returned (μˆ˜μ •λ˜μ—ˆλ‹€κ°€ λ‹€μ‹œ λ°œμƒν•˜λŠ” 버그)

Example (μ˜ˆμ‹œ):

@sub The CI pipeline has been failing intermittently for 2 weeks.
     Failures occur in the integration test step, always between 02:00–04:00 UTC.
     Apply root cause analysis: gather evidence, generate hypotheses, and
     recommend the decisive diagnostic step.

Agent Selection Quick Reference (μ—μ΄μ „νŠΈ 선택 λΉ λ₯Έ μ°Έμ‘°)

I need to… (ν•  μž‘μ—…) Use this agent (μ‚¬μš©ν•  μ—μ΄μ „νŠΈ)
Find existing code patterns explore
Research a technical topic online deep-research-agent
Plan implementation before coding plan
Clarify vague requirements requirements-analyst
Design a full system system-architect
Design server APIs and schemas backend-architect
Design UI components and state frontend-architect
Write Python code python-expert
Improve code quality without behavior change refactoring-expert
Independently review output for errors critic
Write or improve tests quality-engineer
Check for security vulnerabilities security-engineer
Find and fix performance bottlenecks performance-engineer
Write documentation or guides technical-writer
Investigate a complex anomaly scientifically sci-method
Find the root cause of a failure root-cause-analyst