Cognee Code Graph

Build and query an architectural graph of a codebase (symbols, calls, imports, endpoints, dependencies) via Cognee's enola-backed pipeline. Requires a Cognee server >= 1.5.3. Indexing makes no LLM or embedding calls — it is fast, deterministic, and token-free.

Visible search progress

Before running a Cognee search, send one short user-visible progress sentence, for example: "Searching Cognee for the earlier decision…" or "Checking Cognee's code graph for callers of this function…". Name Cognee in the normal working text before the tool call; keep this to one update per related search batch. Afterwards, describe relevant findings, an empty result, or a failed search accurately. Do not claim Cognee improved the answer merely because a search ran.

Index a repository

"${CLAUDE_PLUGIN_ROOT}/scripts/cognee-index-repo.sh" <repo-path-or-git-url> [--dataset <name>] [--index-vectors] [--wait <seconds>]

• Local path (e.g. . or /path/to/repo): works when the Cognee server shares this filesystem — the default local plugin server does. Cloud servers reject local paths; pass a git URL instead.
• Git URL (https://github.com/org/repo): the server shallow-clones it. Freshness follows pushed commits — local edits are invisible to it.
• Default dataset is codebase-<repo-name>-<digest>, one per indexed repo — narrow datasets keep code searches fast, and the path digest keeps two checkouts that share a basename (~/work/a/service, ~/work/b/service) from landing in one graph. The name is printed on success; you rarely need it, because --code searches resolve it from the current checkout.
• --index-vectors: also embeds the code facts so semantic/hybrid search can see them (needs an embedding provider). Without it, code knowledge is reachable ONLY through code search below — conceptual questions about the code then rely on ingested docs/READMEs.
• The submit returns quickly (background pipeline). --wait 60 polls until the graph is queryable. Poll manually via GET /api/v1/datasets/status?dataset=<id>&pipeline=code_graph_pipeline.

Indexing writes enola's snapshot to <repo>/.enola/ (untracked) — add it to .gitignore or global excludes. The plugin's change detection ignores it.

Indexing is also the freshness opt-in: for local-path repos the plugin records a git fingerprint and automatically re-submits the repo in the background when a turn changed the working tree, so the graph tracks your edits. Re-indexing an unchanged repo is skipped server-side (content hashes), so re-running the command is always safe.

Automatic indexing

Opening the agent inside a git repository indexes it automatically in the background at session start — no setup step. What happens depends on the server, because indexing a repo means the code has to be readable by it:

Situation	Behavior
Local server (default), new repo	Indexed automatically — the server reads the working tree in place, so the code never leaves the machine
Cloud/remote server, new repo	Not indexed automatically; run the index command with a git URL, or set COGNEE_CODE_AUTOINDEX=always
Already-indexed repo	Always refreshed if the tree changed since the last index — regardless of the setting above
Not a git repo / no code files / very large repo (>3000 source files)	Skipped; index it explicitly if you want it

Set COGNEE_CODE_AUTOINDEX=off to disable automatic indexing of new repos (explicitly indexed repos keep refreshing). Explicit indexing has no size cap.

What the graph reflects (freshness)

The graph tracks different things depending on where the Cognee server runs. Both are normal, expected behavior — but the results look identical, so know which one applies:

Server	Graph reflects	Stays current via
Local (default)	The working tree, uncommitted and untracked changes included	Automatic re-index after any turn that changed a file
Cloud / remote	The last pushed commit — the server clones the repo and cannot read this machine's disk	Pushing, then re-indexing

On a cloud server, local edits are invisible to the graph until they are pushed. Nothing errors and nothing looks unusual: a question about a symbol you just renamed locally is answered from the pushed state, confidently. So when the answer matters and the work is in progress, either push first or say so in the answer. {"operation": "delta"} shows what the last index actually changed and is the fastest way to check what the graph currently knows.

Local-path indexing therefore requires a server that shares this filesystem; cloud servers reject local paths and need a git URL.

Query the code graph

"${CLAUDE_PLUGIN_ROOT}/scripts/cognee-search.sh" "<seed>" 10 --code [--code-query '<json>']

The repository's code dataset is resolved from the current directory, so --dataset is only needed to query a repo you are not standing in.

Without --code-query, the query text is a seed name (exact/suffix/substring match). With it, pick one exact operation:

Operation	Answers	Example
query_facts	filtered listing	{"operation":"query_facts","kind":"route","limit":50} — all API endpoints
explore	neighborhood of one node	{"operation":"explore","name":"UserService","max_depth":1}
traverse	follow edges from seeds	{"operation":"traverse","start":"main","direction":"forward","max_depth":3}
find_path	how A reaches B	{"operation":"find_path","source":"AuthMiddleware","target":"Database"}
impact_analysis	what breaks if X changes	{"operation":"impact_analysis","targets":["process_payment"]}
architecture (>=1.6.3)	module, route, storage and service map	{"operation":"architecture","diagram":"mermaid"}
insights (>=1.6.3)	explainer findings with cited facts	{"operation":"insights","diagram":"mermaid"}
delta	what the last index changed	{"operation":"delta"}

An ambiguous seed returns an error listing the candidates — pick one and retry with its exact id or a repo filter. A seed that doesn't resolve returns empty (not an error): the graph simply has no such symbol.

When to use which search

• Structural question naming a symbol/file ("what calls X", "what breaks if I change X", "list all endpoints") → --code. Exact, instant, no tokens.
• Conceptual question naming nothing ("how does auth work here?") → regular cognee-search.sh (hybrid/graph). Remember: graph-only code is invisible there unless indexed with --index-vectors or docs were ingested.
• Chain them: hybrid discovers the name ("the payment flow" → PaymentProcessor), then --code gives the exact structure around it.
• Verify freshness: after edits, {"operation":"delta"} shows what the last re-index added/updated/removed. Use it when facts look stale.

Treat results as a map, not ground truth — verify important claims against the actual files before editing.

Per-file ingestion (no repo index)

To store a single code file in normal memory under its real filename (routes down the zero-LLM code path server-side, no cross-file edges):

"${CLAUDE_PLUGIN_ROOT}/scripts/cognee-remember.sh" --file src/payments.py --node-set project_docs

Prefer the repo index above when you care about callers/imports across files.

Automatic recall

Once a repo is indexed from its checkout, the per-prompt memory recall adds a code lane automatically when your prompt mentions an identifier-shaped token (process_payment, UserService, billing/api.py) — the facts appear in the injected context as === Code graph facts ===. No action needed.