Contributing to Attocode Intelligence

Python 3.12/3.13 and Git are supported. The root pyproject.toml is a virtual workspace; packages/code-intel is the only publishable Python package. Keep attocode_intel imports and the attocode-code-intel command stable.

Repository layout

Path	Contents
packages/code-intel/	The product: src/attocode_intel, tests, scripts
packages/code-intel/evals/	Frozen study designs and results; not shipped in the wheel or sdist
eval/	Development and CI harness (eval.quality, eval.matrix); the product never imports it
frontend/	Dashboard, built into the service image
docker/code-intel/	Service image
docs/, mkdocs.yml	Documentation site
scripts/	CI helpers

Development

uv sync --locked --package attocode-code-intel --extra dev --extra service
uv run --no-sync python -m eval.quality local

The bounded local gate checks product contracts, scale and lifecycle behavior. For the preserved engine and HTTP/MCP integration coverage:

uv run --no-sync python -m pytest -c packages/code-intel/pyproject.toml --confcutdir=packages/code-intel/tests packages/code-intel/tests/unit packages/code-intel/tests/integration -p eval.quality.no_agent -q

Run the complete package suite with packages/code-intel/tests instead. The no_agent guard requires a clean environment without the retired attocode distribution and rejects agent imports. Existing developer environments may still contain that editable distribution; create a new environment with UV_PROJECT_ENVIRONMENT=/tmp/intelligence-dev uv sync --locked --package attocode-code-intel --extra dev --extra service and use the same environment variable on subsequent uv run commands. Do not remove shared environments or caches to make a check pass.

PostgreSQL checks use TEST_DATABASE_URL; CI runs the real database, migrations, locking and branch-isolation controls on both Python versions. The dashboard and Docker image have separate CI checks. Real Python/TypeScript/Go/Rust language servers and clean macOS/Linux wheels also gate delivery.

Optional embeddings, model-backed evaluations and broad corpus acquisition must be explicit. Default development setup does not install model extras. The quality runner uses deadlines, managed storage and scratch cleanup; follow the quality guide before broad or cloud runs.

Use fresh source judgments for ranking changes, retain failed results and report unsupported/limited evidence honestly. No agent-outcome or performance claim is established by a passing unit suite. Run scripts/check_pack_licenses.py when changing rule packs. Generate the tool catalog with packages/code-intel/scripts/generate_catalog.py after tool changes.

Docs: uv sync --locked --package attocode-code-intel --extra docs, then uv run --no-sync mkdocs build --strict. Published intelligence releases use intel-v* tags and the standalone release workflow. The retired agent and its development instructions are linked in the archive notes.