Code Graph

Code Graph Skill

Purpose: Use the code graph (codebase-memory-mcp) for sub-millisecond symbol lookup, function search, dependency analysis, and blast radius detection. This replaces brute-force grep and file reading for code navigation.

Core Principle

Graph first, file second. Before reading files or grepping, query the code graph. Only read full files when you need to modify them or need context beyond what the graph provides.

Consider graph when planning. When planning any change — feature, refactor, bug fix — start by querying the graph to understand scope, dependencies, and blast radius. This applies to thinking and planning phases, not just implementation. Grep is still the right tool for searching string literals, log messages, config values, and content that lives outside code structure.

Copy code
┌────────────────────────────────────────────────────────────────┐
│  GRAPH FIRST, FILE SECOND                                      │
│  ─────────────────────────────────────────────────────────────│
│  The code graph indexes your entire codebase as a persistent   │
│  knowledge graph. Claude queries it via MCP for instant         │
│  symbol lookup, dependency chains, and blast radius — instead   │
│  of reading hundreds of files.                                 │
│                                                                │
│  14 MCP tools │ 64 languages │ sub-ms queries │ zero deps      │
│  ~99% fewer tokens for navigation vs brute-force file reads    │
├────────────────────────────────────────────────────────────────┤
│  AUTO-UPDATED                                                  │
│  ─────────────────────────────────────────────────────────────│
│  File watcher keeps graph in sync. Post-commit hook ensures    │
│  freshness. No manual rebuild needed.                          │
└────────────────────────────────────────────────────────────────┘

When to Use Graph vs Direct Read

Rule: If a graph tool can answer the question, use it. Only open files when you need the full source to make edits.

Available MCP Tools

Indexing & Status

Querying & Navigation

Analysis

Management

Workflow: Before Any Code Change

Copy code
0. PLAN       → get_architecture + search_graph to understand scope before planning
1. LOCATE     → search_graph to find the symbol
2. UNDERSTAND → get_code_snippet for context
3. BLAST      → detect_changes to assess impact
4. TRACE      → trace_call_path to find all affected callers
5. CHANGE     → Read file, make edit
6. VERIFY     → detect_changes again to confirm scope

Step 0 applies to planning, not just coding. When the user asks you to plan a feature, refactor, or fix — query the graph first to understand what exists, what depends on what, and what the scope looks like. This prevents plans based on wrong assumptions about the codebase.

Never skip step 3. Blast radius analysis prevents unexpected breakage from changes to shared code.

Graph Data & Freshness

The graph stays fresh automatically through 3 layers — no manual rebuild needed:

You do NOT need to manually re-index unless you do a major restructure (rename entire directories, switch branches with massive diffs). In that case: index_repository once, then the 3 layers keep it fresh.

• Storage: .code-graph/ directory (auto-created, gitignored)
• MCP config: .mcp.json at project root (committed, shared with team)

MCP Configuration

The code graph MCP server is configured in .mcp.json at project root:

jsonCopy code
{
  "mcpServers": {
    "codebase-memory": {
      "command": "codebase-memory-mcp",
      "args": []
    }
  }
}

Installation: ~/.claude/install-graph-tools.sh

Decision Framework

Copy code
Need to find a symbol/function?
  → search_graph (sub-ms, structured result)
  → NOT: grep -r "functionName" (slow, unstructured)

Need to understand dependencies?
  → query_graph or trace_call_path (complete, traversable)
  → NOT: manually reading import statements

Need to assess change impact?
  → detect_changes (comprehensive, instant)
  → NOT: searching for usages manually across files

Need to understand architecture?
  → get_architecture (high-level overview)
  → NOT: reading every directory listing

Need to read/modify code?
  → search_graph to locate, then Read the specific file
  → NOT: reading entire directories hoping to find it

Anti-Patterns