CodeAtlas

Rust-native architecture maps for codebases. CodeAtlas scans a repository, extracts symbols and relationships with deterministic static analysis (tree-sitter), groups code into components, and documents the architecture as analysis.json, Markdown/Mermaid, HTML (Cytoscape), MDX, or reStructuredText. An optional LLM step names and describes components.

What works today

• codeatlas full --local <repo> produces .codeatlas/analysis.json + fingerprint.json (schema-compatible output) with deterministic, directory-derived component names.
• codeatlas-render (or --render md|html|mdx|rst) renders overview + per-component docs, with a manifest that reconciles stale files and refuses to overwrite unowned files.
• Optional component naming/description via an OpenAI-compatible provider (OpenAI, Ollama, OpenRouter, DeepSeek, LiteLLM, LM Studio). Without a provider, analysis still succeeds with deterministic names.
• Languages: Python, Rust, Go, JavaScript/TypeScript, Java, PHP, and C# (tree-sitter).

Quick start

pip install codeatlas-rs
codeatlas full --local /path/to/repo
codeatlas full --local /path/to/repo --render md
codeatlas-render /path/to/repo/.codeatlas/analysis.json --format html
codeatlas diff old/.codeatlas/analysis.json new/.codeatlas/analysis.json --fail-on-change

Output is written to <repo>/.codeatlas/. Configure an optional provider via OPENAI_API_KEY (+ OPENAI_BASE_URL, AGENT_MODEL); see codeatlas_rs.llm.

Wheels: codeatlas-rs ships abi3-py39 wheels, so one wheel covers CPython 3.9+. The release workflow builds manylinux (x86_64/aarch64), musllinux (x86_64/aarch64), macOS universal2, and Windows x86_64 wheels plus an sdist. Each wheel is installed and imported before upload. PyPI 0.8.0 ships a macOS arm64 wheel + sdist; the remaining targets are built and smoke-tested by the release workflow and reach PyPI when its upload job runs.

PR architecture diff

codeatlas diff compares two architecture maps, and .github/actions/architecture-diff runs it as a PR check. See docs/PR_DIFF.md.

Query the call graph

codeatlas full also writes .codeatlas/call_graph.json with the merged symbols and call/reference edges. Query it locally without re-analysis or an LLM:

codeatlas query /path/to/repo helper
codeatlas query /path/to/repo helper --direction callers --format json

analysis.json keeps its established schema; the call graph is a sidecar so local agents and editors can fetch precise symbol context. See docs/CALL_GRAPH_QUERY.md for the schema and the Python API.

Audit findings

The 2026-10-19 implementation audit — including the Java extractor work, verification results, and remaining gaps — is recorded in AUDIT_FINDINGS.md.

Status & roadmap

The deterministic core is implemented end-to-end (discovery → tree-sitter extraction → cross-file resolution → clustering → analysis.json → rendering). See docs/FEATURE_PARITY.md for what is implemented vs. deferred, and docs/IMPLEMENTATION_ROADMAP.md for the milestone plan. Documented limitations: tree-sitter resolution is intentionally simpler than an LSP pipeline (docs/audit/KNOWN_GAPS.md) and clustering is a simplified deterministic directory partition.

Bundle into any repository

codeatlas-bundle versions the architecture map inside a repository: it runs the analysis, renders the docs, and (optionally) commits them so the map evolves with the code. See packages/codeatlas-bundle.

pip install codeatlas-rs codeatlas-bundle
cd my-repo && codeatlas-bundle init && codeatlas-bundle build --commit

Layout

• crates/core — pure-Rust analysis engine (no Python, no network, no LLM).
• crates/py — PyO3 bindings (codeatlas_rs._core).
• python/codeatlas_rs — the Python wrapper (CLI, rendering, optional LLM).
• packages/codeatlas-bundle — pure-Python package that bundles the docs into other repositories.

Development

# Rust core (no Python needed)
cargo test -p codeatlas-core

# Rust workspace (needs a Python with a shared library, e.g. `export PYO3_PYTHON=$(which python3.12)`)
cargo build --workspace && cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --all --check

# Python package
maturin build --release --interpreter python3.12 --out dist
maturin develop --interpreter python3.12
python -m pytest tests

Requires Rust (stable) and Python 3.9+ (3.12 recommended), with maturin for the Python package.

Known issues and deferred work

Blocked / failed:

• Kotlin extractor (grammar incompatible with tree-sitter 0.25); Linux/Windows wheels.

Rejected / deferred:

• Chat UI; LSP-level resolution now.

Full shipped/blocked/rejected detail: QUEUE_STATUS.md.

License

MIT. See LICENSE and NOTICE.

Incremental workflow cache

codeatlas full --cache fingerprints repository source content (git HEAD/status plus file metadata) and reuses the previous analysis when nothing changed:

codeatlas full --local . --cache --render md

The first run stores analysis/fingerprint/coverage under .codeatlas/cache/; subsequent unchanged runs skip the Rust pipeline and reuse the artifacts. This is a whole-repository workflow cache, not per-file re-analysis.