| name | arib-graph |
| argument-hint | [build | refresh | query <entity>] |
| description | Memory | Code-graph subsystem — a native lightweight IMPORT graph (which file imports which, god-node candidates) built with ripgrep/grep, so a large monorepo can be navigated by structure instead of re-grepping every session. build/refresh/query. Honest scope: structural import graph, NOT semantic (no call-graph/type resolution); no Graphify dependency. Loads ON DEMAND only (zero always-on tokens). ADR-034. |
/arib-graph — the code-graph subsystem
CCM's memory remembers decisions and state; the code-graph remembers structure. It's a
native, lightweight import graph — built with ripgrep/grep, git-diffable JSON — so on a
large monorepo you navigate by "what imports QoLService?" instead of grepping the whole
tree every session.
Honest scope (ADR-034): this is an import/symbol graph (file → file edges +
in/out degree), NOT Graphify's semantic analysis — no call-graph, no type
resolution. It is dependency-free (rg preferred, grep fallback) and loads on demand
only — nothing about it is always-on, so the 8K session budget is untouched.
Files (in memory/code-graph/)
| File | Committed? | What |
|---|
graph.json | No (gitignored) | the generated graph — nodes (files) + edges (imports) + stats |
GRAPH_REPORT.md | Yes (seed) | readable summary — counts, god-node candidates, busiest importers |
graph-manifest.json | Yes (seed) | freshness — built, last_build_commit, file_count, stale |
Modes
build
Full graph generation — run once per repo, or after a major refactor:
scripts/build-code-graph.sh
scripts/build-code-graph.sh --root apps/api --max 8000
Writes graph.json + refreshes GRAPH_REPORT.md + graph-manifest.json. On CCM itself
(markdown + shell) the graph is empty — that's honest; it's meant for downstream
TS/JS/Python/Go codebases.
refresh
Re-run build to pick up changes since the last build commit (the MVP refresh is a full
rebuild — cheap on rg). Updates the manifest's last_build_commit so staleness resets.
query <entity>
Ask the built graph directly, e.g. "what imports b?" — read graph.json and filter:
jq --arg e "<entity>" '.edges | map(select(.to|test($e))) | map(.from) | unique' memory/code-graph/graph.json
jq --arg f "<file>" '.edges | map(select(.from==$f)) | map(.to)' memory/code-graph/graph.json
If the graph isn't built (graph-manifest.json built:false), say so and offer build.
How it surfaces (advisory only)
graph-consult.sh (PreToolUse, advisory exit-0) — before a Grep/Glob search,
if the graph exists it surfaces a one-line structural hint pointing at /arib-graph query; pure no-op when the graph is absent, and it never blocks (the blocking
gate is pre-tool-use.sh)./arib-session-start — when graph-manifest.json is stale (or missing in a code
project), it recommends /arib-graph refresh; it never auto-runs (autonomy guard).
Anti-patterns
Claiming semantic/call-graph capability (it's an import graph) · committing the generated
graph.json (it's gitignored, per-project) · adding a graph digest to always-on context
(it loads on demand only) · the consult hook ever blocking (advisory exit-0 only).
