codegraph

Analyze any source-code project through its pre-built local code graph: index it, then answer graph-first codebase questions — symbol source + call paths, callers/callees, change impact, affected tests, file inventory. Use when asked how a codebase works, who calls a symbol, what a change would break, or which tests cover a change.

CodeGraph — Graph-First Codebase Analysis

Answer codebase questions from a pre-built local code graph instead of a grep+read crawl. CodeGraph parses a project with tree-sitter, stores symbols/edges/files in SQLite with FTS5, and returns the verbatim source an agent needs in one call.

Graph-first doctrine

Start with explore: when it finds the relevant symbols, it returns their verbatim, line-numbered source plus call paths — a completed Read, not a search result.

• Treat explore output as the read source itself: it is byte-for-byte the on-disk file, identical to a Read call. Using it directly is the point of the tool.
• For flow questions, name the symbols spanning the flow: codegraph explore 'mutateElement renderScene' surfaces the call path among them, riding dynamic-dispatch hops grep cannot follow.
• After edits, if a response opens with a ⚠️ Some files referenced below were edited since the last index sync banner, Read only the listed files; everything unlisted is fresh.

Indexing

CodeGraph is per-project and index-driven. The graph lives in <repo>/.codegraph/.

• Before querying, confirm the project is indexed with codegraph status.
• An unindexed project: stop calling codegraph for it this session and fall back to built-in Read/Grep. Indexing is the caller's decision — mention codegraph init exists, run it only when authorized.
• codegraph init in the project root builds .codegraph/ and the graph in one step. It respects .gitignore and default-excluded dirs (node_modules, dist, vendor, ...); files over 1 MB are skipped.

🔴 STOP — index-changing commands

init, index, sync, uninit, and unlock modify .codegraph/ or remove its lock. Do not run them without explicit authorization. uninit is destructive; never use it merely to recover from a query failure. If the graph is missing, stale, locked, or unhealthy and authorization is absent, fall back to built-in Read/Grep and state the limit.

Workflow

Step 1 — Confirm codegraph

text

codegraph --version

If codegraph is absent and installing it is authorized, follow the install section in references/cli-reference.md (standalone bundle or npm; Node >=20 <25; new shell needed after bundle install).

Done when: a version string prints.

Step 2 — Check the index

text

codegraph status <repo>

status takes a positional path, not --path (only explore/query/node/files/callers/callees/impact/affected use -p/--path). Running it inside the repo also works.

Done when: you know whether the project is indexed and the backend is healthy.

Step 3 — Establish the graph

🔴 STOP: If the index is absent, request authorization before codegraph init <repo>, then re-run status. Without authorization, fall back to built-in tools and state the limit.

Done when: codegraph status reports the graph initialized with sane counts — or you have a fallback in hand.

Step 4 — One-shot explore

text

codegraph explore --path <repo> '<question or symbol bag>'

Ask the whole question first. A bag of symbol names works better for flows than broad natural language.

Done when: you have the relevant source and call paths for the question — or an explicit unindexed/empty signal telling you to fall back or narrow.

If explore misses

Treat an empty result, mostly unrelated symbols, or a result that omits the requested path/symbol as a miss. Do not repeat the same broad natural-language query.

1. Locate the target with codegraph query --path <repo> --kind function <symbol>. Record its filePath, startLine, and endLine.
2. Read the exact source with codegraph node --path <repo> --file <filePath> --offset <startLine> --limit <lineCount>, where lineCount = endLine - startLine + 1.
3. Trace relationships with callers, callees, or impact for the selected symbol.
4. If a precise query is still empty, fall back to built-in Read/Grep and say that the graph could not establish the answer.

Step 5 — Narrow

Only when explore left ground uncovered: query for symbol lookup (kind/path/name fields), node to read one symbol or file, callers/callees for one-hop relationships, impact for change radius, affected for a complete test-file backstop, files for tree inventory. Prefer --json for scriptable output.

For focused test selection, begin with callers and symbol-level impact; use affected last to catch omissions. affected is file-level and unranked, so do not choose a focused test run from it alone.

Done when: every symbol, call path, and affected test the question needs is accounted for.

Step 6 — Answer from the graph

Answer directly from what codegraph returned, citing path anchors. Results come from a full AST parse — the graph outranks recall.

Done when: every material claim in the answer traces to returned source with a path anchor.

Reference

Full flag table, query syntax, env vars, ignore rules, language support, install, and performance caveats: references/cli-reference.md.

Do not

• Do not install, initialize, rebuild, sync, unlock, or remove an index without explicit authorization.
• Do not treat a successful but unrelated explore result as evidence for the question.
• Do not claim test-case-level coverage from affected; it reports affected files, not individual tests.