codeskeleton

Reveal the skeleton of your codebase — turn any folder of code into a queryable knowledge graph. Single binary, zero runtime dependencies, blazing fast.

codeskeleton .

codeskeleton-out/
├── graph.html         interactive graph — click nodes, search, filter by community
├── GRAPH_REPORT.md    god nodes, surprising connections, suggested questions
├── graph.json         persistent graph — query later without re-reading
└── cache/             SHA256 cache — re-runs only process changed files

Install

Requires: Rust 1.70+

cargo install codeskeleton

Or build from source:

git clone https://github.com/DhanushNehru/codeskeleton.git
cd codeskeleton
cargo build --release
./target/release/codeskeleton .

Usage

codeskeleton .              # analyze current directory
codeskeleton ./src          # analyze a specific folder
codeskeleton . --no-cache   # force full re-extraction

Add a .cographignore file to exclude folders:

# .cographignore
vendor/
node_modules/
dist/
*.generated.py

Same syntax as .gitignore. Patterns match against file paths relative to the analyzed folder.

What You Get

God nodes — highest-degree concepts (what everything connects through)

Surprising connections — cross-community edges ranked by structural distance, with plain-English explanations

Communities — automatically detected clusters of related code with cohesion scores

Suggested questions — 4-5 questions the graph is uniquely positioned to answer

Interactive visualization — dark-themed vis.js graph with search, click-to-inspect, community coloring

Incremental builds — SHA256 file caching means re-runs only process changed files

Supported Languages

How It Works

codeskeleton runs a deterministic AST pass using tree-sitter. No LLM needed — pure structural extraction:

1. Detect — walks the directory tree respecting .gitignore and .cographignore
2. Cache — SHA256 hashes each file, skips unchanged files from previous runs
3. Extract — tree-sitter parses each file in parallel (Rayon), extracts classes/structs, functions/methods, imports, and call sites
4. Build — assembles all extractions into a petgraph knowledge graph
5. Cluster — label propagation community detection groups related nodes
6. Analyze — identifies god nodes (highest degree), surprising cross-community connections, generates questions
7. Export — writes graph.json, graph.html (vis.js), and GRAPH_REPORT.md

Every relationship is tagged EXTRACTED (found directly in source) or INFERRED (call-graph second pass). You always know what was found vs guessed.

Architecture

detect → cache-check → extract (parallel) → build_graph → cluster → analyze → report → export

Each stage is a pure function in its own module. No shared mutable state, no side effects outside codeskeleton-out/.

Performance

codeskeleton is written in Rust for maximum performance:

• Parallel extraction — Rayon processes all files across all CPU cores
• Zero-copy parsing — tree-sitter operates on raw bytes, no string allocation
• Incremental builds — SHA256 caching means only changed files are re-extracted
• Single binary — no Python, no Node.js, no runtime dependencies
• Native speed — compiled to optimized machine code with LTO

Contributing

Adding a language:

1. Add the tree-sitter grammar crate to Cargo.toml
2. Add a variant to SupportedLanguage in languages.rs
3. Define the LanguageSpec with AST node types
4. Add the extension mapping in from_extension()
5. Add an import extractor in extract.rs
6. Add test fixtures

License

MIT