code-graph

---

name: code-graph description: Use when you want an AI coding agent to answer structural questions about a repo (where is X, what calls Y, architecture) without grep-reading files — sets up a local codebase-memory-mcp graph per repo, isolated per clone/worktree, and watches for stale indexes. Triggers: "index this repo", "code graph", "give the agent a map of the codebase", "stop re-reading files", "which repos need re-indexing".

code-graph — codebase-memory graph, provisioned & kept fresh

Overview

Wires a local codebase-memory-mcp graph into a repo so agents query a structural map (functions, calls, routes, imports) instead of reading files one at a time — the single biggest per-task token lever alongside ctx-wire. One graph per repo path, so your parallel clones and worktrees stay isolated — a query in one never returns another branch's code. A companion watcher flags graphs that have gone stale.

Everything runs locally. No API key, no data leaves the machine. The binary is a single audited MIT C build; the only outbound call it makes is a background GitHub version check (harmless). Prefer building from source over the release binary if you want to erase all binary trust.

Prerequisites

• codebase-memory-mcp binary discoverable: on PATH, at ~/dev/codebase-memory-mcp/build/c/codebase-memory-mcp, or via CBM_BIN. If absent, setup.sh prints the build command and stops — it never auto-compiles (it's a ~279MB C build; opt in explicitly).
• git (optional) — used only to label the current HEAD in reports. Staleness itself compares the graph to the working tree, so it works without git too.
• jq (optional) — for merge-safe edits to an existing .mcp.json.

Setup flow (per repo)

1. Provision + index the current repo:

   bash "${CLAUDE_PLUGIN_ROOT}/skills/code-graph/setup.sh" "$PWD"
   

   This locates the binary, merges a codebase-memory server into the repo's .mcp.json (never clobbering other servers), indexes the tree in ~9s under a path-derived project name, and registers the repo in ~/.claude/code-graph/repos.txt for the watcher.
2. Never force --name. The path-derived name is what keeps clones and worktrees isolated. Overriding it makes two checkouts share (and clobber) one graph — the exact bug to avoid.
3. Approve the MCP server once in a Claude Code session rooted at the repo, then ask structural questions (search_graph, get_architecture) instead of grepping. .mcp.json is untracked; add it to .git/info/exclude if you don't want it in git status.

Keep-fresh watcher

A graph indexed on branch A returns stale answers after you switch branches or land commits. Schedule the drift check so you're told which graphs to refresh:

1. Guard against duplicates. List existing scheduled jobs first. If a 22,52 * * * * watcher running code-graph/drift.sh already exists, STOP — offer to replace it, never add a second.
2. Schedule the watcher. Use the prompt in watcher-prompt.md (bundled beside this file). Schedule it recurring at 22,52 * * * * (offset from the watch-limits watcher's 7,37 so they don't collide). If the host has no cron tool, run it as a /loop instead.
3. The watcher reads repos.txt, runs drift.sh per repo, and reports STALE / NOT_INDEXED graphs with the exact ~9s re-index command. It never auto-re-indexes — it surfaces the picks; you pull the trigger.

Honesty note (why there's no savings badge)

ctx-wire savings are metered — it emits a real saved_tokens counter. codebase-memory savings are counterfactual (tokens you didn't spend because the agent queried the graph instead of reading files); any figure is a model, not a measurement. Report it as an estimate if asked — do NOT publish it as a hard badge alongside the metered ctx-wire numbers.