Architecture

Architecture

CodeGraph is a Rust workspace: 15 library crates under crates/* plus the codegraph CLI under bin/. The workspace uses edition 2021 and a pinned Rust 1.95 toolchain, with dependencies centralized in the root Cargo.toml.

The pipeline

A normal extract run flows through these stages:

detect  ->  extract  ->  graph build  ->  cluster + analyze  ->  output
                                   \-> (optional) semantic pass

1. detect - walk the directory, classify files, apply ignore rules, and skip sensitive files. See Extraction.
2. extract - run tree-sitter (or a regex extractor) per file to produce nodes and edges. Results are cached per file so unchanged files skip re-parsing. See Languages.
3. graph build - assemble nodes/edges into a graph, resolve symbols across files, and deduplicate.
4. cluster + analyze - detect communities, compute betweenness and god nodes, find import cycles and surprising connections. See Analysis and Reports.
5. output - write graph.json and the report, visualizations, and exports. See Output Formats.

The optional semantic pass (--semantic) sends documents and papers to an LLM to add concept nodes and to break ties during dedup. See Semantic Analysis.

Querying and serving read graph.json back; they do not re-extract. See Querying and MCP Server.

Crates

Crate	Responsibility
codegraph-core	The shared data contract: NodeId, FileType, Confidence, Node, Edge, Hyperedge, the graph.json node-link DTO, id generation, and schema validation
codegraph-detect	File discovery, classification, ignore handling (.codegraphignore / .gitignore), manifest building, and sensitive-file detection
codegraph-extract	Tree-sitter (and regex) extractors that turn source files into core nodes/edges; languages gated behind lang-* features; per-file AST cache
codegraph-graph	Graph assembly: build, symbol resolution, dedup (MinHash/LSH), clustering and community detection, betweenness, analysis
codegraph-semantic	The LLM semantic pass: documents and papers to concept nodes, plus the optional dedup tiebreaker
codegraph-llm	Pluggable LLM client layer: provider registry with env auto-detect, response cache, JSON repair, token-budget chunking, adaptive retry
codegraph-query	Query (IDF-scored subgraph retrieval), shortest path, node explanation, and reverse impact
codegraph-output	Output writers: graph.json, HTML viewers, SVG, GraphML, Cypher, DOT, Mermaid call-flow, D3 tree, Obsidian, wiki, and live database push
codegraph-report	The GRAPH_REPORT.md generator
codegraph-ingest	External-source ingestion: URL (SSRF-guarded), MCP config, Cargo, Postgres, SCIP, office, media
codegraph-server	The MCP server (read-only graph and PR tools) over stdio and HTTP, plus a small REST surface
codegraph-prs	Graph-aware PR dashboard: classification, CI rollup, blast radius, conflict grouping
codegraph-incremental	Changed-files rebuild engine plus git integration (hooks, merge driver, watch, concurrency lock)
codegraph-workspace	Multi-repo / monorepo federation: member discovery, namespacing, cross-repo resolution, global store, merge-graphs
codegraph-skillgen	Generates and installs the host-assistant integration: the Claude skill file + .claude/settings.json hooks, the always-on instruction blocks (AGENTS.md/GEMINI.md/etc.), and the Codex MCP server + SessionStart hook config (project .codex/ or global ~/.codex/)
bin/codegraph	The CLI that wires the crates into commands

Data model

The graph is a node-link structure serialized to graph.json (NetworkX-compatible shape).

• Nodes have an id, a label, a file_type (one of code, document, paper, image, rationale, concept), and a source_file. Optional fields include source_location, community, and repo (for federated graphs).
• Edges (serialized under links) have a source, target, relation (for example calls, imports, imports_from, inherits, implements, references, contains, depends_on, reads_from), and a confidence of EXTRACTED, INFERRED, or AMBIGUOUS. Cross-repo edges are flagged with cross_repo.

The exact JSON shape and every export format are documented in Output Formats.

Determinism

Extraction parallelizes across files but collects results in a stable order, so graph.json is byte-identical for the same input regardless of thread scheduling. Community numbers are assigned deterministically (community 0 is the largest) and kept stable across incremental rebuilds.

Where to go next

• Configuration knobs and files: Configuration
• Building, testing, and CI: Development