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.

┌────────────────────────────────────────────────────────────────┐
│  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

Task	Use Graph Tool	Use Direct Read?
Find function/class definition	`search_graph`	No
Get function signature + docs	`get_code_snippet`	No
Find all callers of a function	`trace_call_path`	No
Trace dependency chain	`query_graph`	No
Determine blast radius of change	`detect_changes`	No
Understand project architecture	`get_architecture`	No
Search for code patterns	`search_code`	No
Read full implementation to modify	`search_graph` to locate, then Read file	Yes
Understand business logic context	`get_code_snippet` for overview, then Read	Yes

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

Tool	Purpose	When to Use
`index_repository`	Build/rebuild graph for a project	First setup, or after major restructure
`index_status`	Check if graph is current	Before querying, if unsure of freshness
`list_projects`	List all indexed projects	Multi-project navigation

Querying & Navigation

Tool	Purpose	When to Use
`search_graph`	Find symbols by name (fuzzy)	"Find auth-related functions"
`search_code`	Text search across indexed codebase	"Find TODO comments", pattern matching
`get_code_snippet`	Get source code for a specific symbol	Need signature, docstring, implementation
`get_graph_schema`	Understand graph structure and relationships	Exploring what data is available
`query_graph`	Run structured graph queries	Complex dependency/relationship queries

Analysis

Tool	Purpose	When to Use
`trace_call_path`	Trace caller/callee chains	"Who calls sendEmail?", "What does init() trigger?"
`detect_changes`	Identify changed files and blast radius	Before/after code changes, PR review
`get_architecture`	High-level module/package structure	Onboarding, understanding project layout

Management

Tool	Purpose	When to Use
`delete_project`	Remove a project from the graph	Cleanup, project restructure
`manage_adr`	Architecture decision records	Document architectural decisions
`ingest_traces`	Import runtime traces	Performance analysis, dead code detection

Workflow: Before Any Code Change

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:

Layer	Trigger	What Happens
File watcher	Every file save	codebase-memory-mcp detects changes and re-indexes affected files in real-time
Auto-index	Session start	`auto_index: true` ensures graph is current when Claude Code starts
Post-commit hook	Every `git commit`	Touches `.code-graph/.needs-update` marker — file watcher picks it up (~10ms, non-blocking)

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:

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

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

Decision Framework

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

Anti-Pattern	Do This Instead
Grepping for function names	`search_graph` with the function name
Reading entire files to find a signature	`get_code_snippet` for the specific symbol
Manually tracing import chains	`trace_call_path` or `query_graph`
Making changes without checking impact	`detect_changes` before every edit to shared code
Reading all files in a directory	`get_architecture` for structure, `search_graph` for specifics
Ignoring graph staleness warnings	Check `index_status`, re-index if needed
Re-indexing on every query	Trust the file watcher; only manual re-index after major restructure

code-graph