/coord is a 6-stage pipeline built on top of Claude Code that orchestrates memory retrieval, planning, parallel research, specialist agent delegation, and anti-sycophancy verification for complex engineering and research tasks.
/coordλ λ©λͺ¨λ¦¬ κ²μ, κ³ν μ립, λ³λ ¬ 리μμΉ, μ λ¬Έκ° μμ΄μ νΈ μμ, μ첨 λ°©μ§ κ²μ¦μ μ‘°μ¨νλ 6λ¨κ³ νμ΄νλΌμΈμ
λλ€.
# Step 1 β Clone this repository (μ μ₯μ 볡μ¬)
git clone https://github.com/Transconnectome/claude-code-coord.git
cd claude-code-coord
# Step 2 β Copy framework files (νλ μμν¬ νμΌ λ³΅μ¬)
cp -r .claude/ ~/.claude/
# Step 3 β Install MCP servers, then verify (MCP μλ² μ€μΉ λ° κ²μ¦)
npm install -g @modelcontextprotocol/server-sequential-thinking
claude code --print "/coord 'Hello, is this working?'"Full installation details: docs/SETUP_GUIDE.md
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β /coord β 6-Stage Pipeline β
ββββββββββββ¬βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Stage 0 β BOOTSTRAP β detect task complexity, choose mode β
β μ€λΉ β (μμ
볡μ‘λ μΈ‘μ , μ€ν λͺ¨λ μ ν) β
ββββββββββββΌβββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Stage 1 β MEMORY PRE-CHECK β parallel 3-source scan β
β λ©λͺ¨λ¦¬ β Auto-memory β Serena MCP β Graphiti knowledge graph β
ββββββββββββΌβββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Stage 2 β PLAN β sequential-thinking MCP + parallelization map β
β κ³ν β (λ€λ¨κ³ μΆλ‘ + λ³λ ¬ν κ°λ₯ν μμ
λΆλ¦¬) β
ββββββββββββΌβββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Stage 3 β RESEARCH β parallel subagents (Explore β web β docs) β
β 리μμΉ β (νμ/μΉ/곡μλ¬Έμ μλΈμμ΄μ νΈ λ³λ ¬ μ€ν) β
ββββββββββββΌβββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Stage 4 β EXECUTE β specialist agents + intermediate validation β
β μ€ν β (μ λ¬Έκ° μμ΄μ νΈ + λ¨κ³λ³ κ²μ¦) β
ββββββββββββΌβββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Stage 5 β REVIEW β critic agent + optional premium model review β
β 리뷰 β (critic μμ΄μ νΈ + μ νμ ν리미μ λͺ¨λΈ κ²ν ) β
ββββββββββββΌβββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Stage 6 β MEMORY UPDATE β route lessons to correct store β
β λ©λͺ¨λ¦¬ β Auto-memory > Graphiti > Serena, update MEMORY.md β
ββββββββββββ΄βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
The pipeline always reads from three memory sources in parallel before planning. Each source serves a distinct purpose.
νμ΄νλΌμΈμ κ³ν μ μ νμ μΈ κ°μ§ λ©λͺ¨λ¦¬ μμ€λ₯Ό λ³λ ¬λ‘ μ½μ΅λλ€.
| Source | What it stores (μ μ₯ λ΄μ©) | Persistence (μ§μμ±) |
|---|---|---|
Auto-memory (~/.claude/projects/β¦/memory/) |
User preferences, session lessons, project state (μ¬μ©μ μ νΈ, μΈμ κ΅ν, νλ‘μ νΈ μν) | Cross-session (μΈμ κ°) |
| Serena MCP | Code symbols, file locations, workspace structure (μ½λ μ¬λ³Ό, νμΌ μμΉ, μν¬μ€νμ΄μ€ ꡬ쑰) | Within session + project (μΈμ λ΄ + νλ‘μ νΈ) |
Graphiti (graphiti-memory) |
Domain knowledge, research decisions, meeting records (λλ©μΈ μ§μ, μ°κ΅¬ κ²°μ , νμλ‘) | Long-term knowledge graph (μ₯κΈ° μ§μ κ·Έλν) |
| Task type (μμ μ ν) | Primary MCP | Fallback |
|---|---|---|
| Multi-step reasoning (λ€λ¨κ³ μΆλ‘ ) | sequential-thinking |
Native Claude |
| Web search (μΉ κ²μ) | tavily |
WebSearch |
| Official library docs (곡μ λ¬Έμ) | context7 |
Tavily |
| UI component generation (UI μμ±) | magic |
Manual HTML |
| Browser automation (λΈλΌμ°μ μλν) | playwright |
Manual testing |
| Symbol operations (μ¬λ³Ό μ‘°μ) | serena |
Grep + Edit |
| Bulk pattern edits (μΌκ΄ ν¨ν΄ νΈμ§) | morphllm |
MultiEdit |
| Document analysis (λ¬Έμ λΆμ) | notebooklm |
Read + summarize |
| Knowledge graph memory (μ§μ κ·Έλν) | graphiti-memory |
Auto-memory |
- Python 3.10 or higher (Python 3.10 μ΄μ)
- Node.js 18 or higher (Node.js 18 μ΄μ)
- Claude Code installed and authenticated (μ€μΉ λ° μΈμ¦ μλ£)
- API keys:
TAVILY_API_KEY,OPENROUTER_API_KEY(optional for premium review) (μ νμ β ν리미μ 리뷰μ©)
1. Clone the repository (μ μ₯μ 볡μ¬)
git clone https://github.com/Transconnectome/claude-code-coord.git
cd claude-code-coord2. Copy the .claude/ framework directory (νλ μμν¬ λλ ν 리 볡μ¬)
# Back up your existing ~/.claude/ first (κΈ°μ‘΄ μ€μ λ°±μ
κΆμ₯)
cp -r ~/.claude/ ~/.claude.backup/
cp -r .claude/ ~/.claude/3. Install MCP servers (MCP μλ² μ€μΉ)
# Required servers (νμ μλ²)
npm install -g @modelcontextprotocol/server-sequential-thinking
npm install -g @modelcontextprotocol/server-memory
# Recommended servers (κΆμ₯ μλ²)
npm install -g @tavily/mcp-server
npm install -g @21st-dev/magicFull MCP installation instructions: docs/MCP_SERVERS.md
4. Copy and configure settings (μ€μ νμΌ κ΅¬μ±)
cp settings/settings.json.template ~/.claude/settings.json
# Edit ~/.claude/settings.json β fill in API keys and hook paths
# ~/.claude/settings.json νΈμ§ β API ν€ λ° hook κ²½λ‘ μ
λ ₯5. Verify the installation (μ€μΉ κ²μ¦)
cd tests/
python -m pytest test_hooks.py -v# Explicit invocation (λͺ
μμ νΈμΆ)
/coord "Implement JWT authentication for the FastAPI service"
# With @sub shorthand (λ°±κ·ΈλΌμ΄λ μμ΄μ νΈ λ¨μΆ νΈμΆ)
@sub Implement JWT authentication for the FastAPI service/coord activates automatically when the user message contains signals of complexity. You do not need to type /coord explicitly.
볡μ‘μ± μ νΈκ° ν¬ν¨λ λ©μμ§μμλ /coordλ₯Ό λͺ
μμ μΌλ‘ μ
λ ₯νμ§ μμλ μλ νμ±νλ©λλ€.
| Signal category (μ νΈ λ²μ£Ό) | Example keywords (μμ ν€μλ) |
|---|---|
| Scope ambiguity (λ²μ λͺ¨νΈ) | "not sure", "thinking about", "explore" |
| Multi-domain (λ€μ€ λλ©μΈ) | "frontend + backend", "infra and code" |
| Long pipelines (κΈ΄ νμ΄νλΌμΈ) | "step by step", "refactor the whole", "migrate" |
| Debugging (λλ²κΉ ) | "figure out why", "root cause", "broken since" |
| Architecture (μν€ν μ²) | "design a system", "how should I structure" |
| Research (리μμΉ) | "find the best", "compare approaches", "survey" |
- Simple one-file edits (λ¨μ λ¨μΌ νμΌ μμ )
- Quick factual questions (λΉ λ₯Έ μ¬μ€ μ§λ¬Έ)
- Single-tool operations under 3 steps (3λ¨κ³ λ―Έλ§ λ¨μΌ λꡬ μμ )
- Casual conversation (μΌμμ μΈ λν)
For detailed use-case guidance: docs/WHEN_TO_USE.md
| Server | Role (μν ) | API key required (API ν€ νμ) |
|---|---|---|
sequential-thinking |
Multi-step structured reasoning (λ€λ¨κ³ ꡬ쑰μ μΆλ‘ ) | No |
memory |
Key-value session persistence (μΈμ λ΄ ν€-κ° μ μ₯) | No |
tavily |
Real-time web search (μ€μκ° μΉ κ²μ) | Yes β TAVILY_API_KEY |
context7 |
Official library documentation (곡μ λΌμ΄λΈλ¬λ¦¬ λ¬Έμ) | No |
magic |
UI component generation via 21st.dev (UI μ»΄ν¬λνΈ μμ±) | Yes β MAGIC_API_KEY |
playwright |
Browser automation and E2E testing (λΈλΌμ°μ μλν) | No |
morphllm |
Bulk pattern-based code edits (μΌκ΄ ν¨ν΄ κΈ°λ° μ½λ μμ ) | No |
serena |
Semantic code navigation and memory (μλ§¨ν± μ½λ νμ) | No |
notebooklm |
Document analysis and knowledge base (λ¬Έμ λΆμ + μ§μ λ² μ΄μ€) | No (Google auth) |
Full per-server documentation: docs/MCP_SERVERS.md
The hook system runs Python scripts at four Claude Code lifecycle events. It catches agreement bias before responses are generated and after they complete.
ν μμ€ν μ Claude Code μλͺ μ£ΌκΈ°μ λ€ μ΄λ²€νΈμμ Python μ€ν¬λ¦½νΈλ₯Ό μ€νν©λλ€. μλ΅μ΄ μμ±λκΈ° μ κ³Ό ν λͺ¨λμμ λμ νΈν₯μ κ°μ§ν©λλ€.
| Hook event | Script | Purpose (μν ) |
|---|---|---|
SessionStart |
chavis_session_init.py |
Load sycophancy lessons from prior sessions (μ΄μ μΈμ κ΅ν λ‘λ) |
UserPromptSubmit |
chavis_prompt_classify.py |
Classify risk level of incoming prompt (ν둬ννΈ μν μμ€ λΆλ₯) |
UserPromptSubmit |
chavis_strategic_challenge.py |
Fire Strategic Challenge Template on scope pivots (λ²μ νΌλ΄ μ μ λ΅ λμ ν νλ¦Ώ μμ±) |
Stop |
chavis_stop_audit.py |
Audit completed response for compliance drift (μλ£λ μλ΅μ κ·μ μ€μ νΈμ°¨ κ°μ¬) |
Stop |
chavis_persistent_logger.py |
Append evaluation to audit trail (κ°μ¬ μΆμ μ νκ° μΆκ°) |
Full hook documentation: docs/HOOKS.md
Specialist agents are delegated via subagent_type. The pipeline selects the right agent automatically based on task classification.
μ λ¬Έκ° μμ΄μ νΈλ subagent_typeμ ν΅ν΄ μμλ©λλ€. νμ΄νλΌμΈμ μμ
λΆλ₯λ₯Ό κΈ°λ°μΌλ‘ μλμΌλ‘ μ μ ν μμ΄μ νΈλ₯Ό μ νν©λλ€.
| Agent | When activated (νμ±ν μμ ) |
|---|---|
explore |
Codebase discovery, 3+ search queries (μ½λλ² μ΄μ€ νμ) |
plan |
Implementation strategy before coding (μ½λ© μ ꡬν μ λ΅) |
deep-research-agent |
Multi-source web research + synthesis (λ€μ€ μμ€ μΉ λ¦¬μμΉ) |
critic |
Review and falsifiability challenge (κ²ν λ° λ°μ¦ κ°λ₯μ± λμ ) |
backend-architect |
Server-side system design (μλ² μ¬μ΄λ μμ€ν μ€κ³) |
frontend-architect |
UI/UX architecture decisions (UI/UX μν€ν μ² κ²°μ ) |
python-expert |
Python-specific implementation (Python νΉν ꡬν) |
quality-engineer |
Test coverage and quality gates (ν μ€νΈ 컀λ²λ¦¬μ§ λ° νμ§ κ²μ΄νΈ) |
security-engineer |
Threat modeling and vulnerability review (μν λͺ¨λΈλ§ λ° μ·¨μ½μ κ²ν ) |
performance-engineer |
Profiling and optimization (νλ‘νμΌλ§ λ° μ΅μ ν) |
technical-writer |
Documentation and user guides (λ¬Έμν λ° μ¬μ©μ κ°μ΄λ) |
requirements-analyst |
Ambiguous request clarification (λͺ¨νΈν μμ² λͺ μΈν) |
system-architect |
Full system design across domains (μ 체 μμ€ν μ€κ³) |
sci-method |
Scientific hypothesis-evidence-validation workflow (κ³Όνμ κ°μ€-μ¦κ±°-κ²μ¦ μν¬νλ‘μ°) |
Full agent documentation: docs/AGENTS.md
cd tests/
python -m pytest test_hooks.py -v # Hook system unit tests (ν
μμ€ν
λ¨μ ν
μ€νΈ)
python -m pytest test_pipeline.py -v # 6-stage pipeline integration tests (νμ΄νλΌμΈ ν΅ν© ν
μ€νΈ)
python -m pytest test_agents.py -v # Agent routing tests (μμ΄μ νΈ λΌμ°ν
ν
μ€νΈ)| File | Contents |
|---|---|
docs/WORKFLOW.md |
6-stage pipeline deep dive (6λ¨κ³ νμ΄νλΌμΈ μμΈ) |
docs/AGENTS.md |
All specialist agents with examples (μ λ¬Έκ° μμ΄μ νΈ λͺ©λ‘ λ° μμ) |
docs/HOOKS.md |
Anti-sycophancy hook architecture (μ첨 λ°©μ§ ν μν€ν μ²) |
docs/WHEN_TO_USE.md |
Use-case guide with scenarios (μ¬μ© μΌμ΄μ€ κ°μ΄λ) |
docs/SETUP_GUIDE.md |
Step-by-step installation for beginners (μ΄λ³΄μμ© μ€μΉ κ°μ΄λ) |
docs/MCP_SERVERS.md |
Per-server configuration reference (μλ²λ³ μ€μ μ°Έμ‘°) |
MIT License. See LICENSE for full text.