code-graph（代码图）

Code Graph

Build a persistent .claude/codegraph.md index so any future task can read one file and know exactly where to look. Scripts do all structural extraction — the agent only writes purpose summaries.

Quick Start

# 1. Run the scanner (does all file discovery, export + import extraction)
node <skill-dir>/scripts/scan.js > .claude/codegraph.draft.md

# 2. Check if existing graph is still fresh
bash <skill-dir>/scripts/check-stale.sh   # exits 0 if fresh, 1 if stale

# 3. Agent annotates the draft with purpose summaries → saves as codegraph.md

Workflow

1. Check freshness first

bash <skill-dir>/scripts/check-stale.sh

Exits 0 (fresh) → read .claude/codegraph.md directly and skip all scanning. Exits 1 (stale or missing) → proceed to step 2.

2. Run the scanner

node <skill-dir>/scripts/scan.js > .claude/codegraph.draft.md

The script outputs a structured draft containing: file list, detected exports, local import graph, entry points, and "used-by" counts — all extracted via grep without reading full file content. No AI tokens spent.

For a partial update (only changed files since last graph):

node <skill-dir>/scripts/scan.js --since <recorded-commit> > .claude/codegraph.draft.md

3. Agent annotates — purpose summaries only

Read .claude/codegraph.draft.md. For each file node it will contain Purpose:???. Fill in one sentence only describing what the file does. Do not re-extract exports or imports — the script already did that.

Use grep to clarify a file's role if the exports alone aren't enough:

grep -n "export\|class\|function" src/services/userService.ts | head -20

Write the annotated result to .claude/codegraph.md (overwrite draft).

4. Using the graph

Before any task: read .claude/codegraph.md, identify relevant files from the index, read only those. Use grep to navigate within them:

grep -rn "functionName" src/          # find where something is defined
grep -rn "import.*userService" src/   # find all callers of a module

5. Update after changes

node <skill-dir>/scripts/scan.js --since HEAD~1 > .claude/codegraph.draft.md
# Agent re-annotates only the changed nodes, merges with existing graph

Guardrails

• Scripts run first — never ask the agent to discover files or extract exports manually.
• Never store line numbers in the graph — they go stale immediately.
• Graph must stay under 300 lines. If it exceeds that, the script switches to folder-level summaries.
• The graph is a navigation aid — always read the actual file before editing it.

References

• REFERENCE.md — graph file format, node template, folder-level summary, token budget.