cherimoya: Instructions file for Claude Code

Cherimoya

Compact sequence-to-function deep learning model (DNA → genomic profile / counts) with custom Triton kernels and an end-to-end CLI.

Layout

• cherimoya/ — model + kernels. Public API is only Cherimoya, CheriBlock, EMA (see __init__.py); everything else is private and may break between versions.
• cherimoya_cli/ — cherimoya console script. Subcommands in commands/; all CLI defaults live in cherimoya_cli/defaults.py — edit there, not in individual subcommand files.
• tests/ — pytest suite. CPU-only by default; CUDA/Triton tests are auto-skipped when a GPU isn't visible.
• docs/ — Sphinx (furo). Built on Read the Docs.

Ignore .ipynb_checkpoints/ directories anywhere in the tree — stale Jupyter autosaves, gitignored, not authoritative. Never read or edit them.

Commands

pip install -e .                       # dev install
pip install -e .[docs]                 # add Sphinx + furo + copybutton

pytest                                 # full suite; CUDA/triton auto-skip
pytest -m "not cuda and not triton"    # force CPU-only
pytest tests/test_cheri.py::test_name -x -q   # one test, fail-fast, quiet

cd docs && sphinx-build -W -b html . _build/html   # strict docs build

cherimoya --help                       # CLI entry (see docs/cli.rst)

Subcommands: pipeline-json, pipeline, negatives, fit, evaluate, attribute, seqlets, marginalize, batch.

Conventions

• Indentation: tabs, not spaces. Don't let formatters convert.
• Commit messages: UPPERCASE verb prefix — ADD, FIX, UPDATE, REWRITE, RELEASE, REMOVE. Match the existing git log style.
• Docs/README accuracy is a hard rule: every concrete detail (defaults, versions, flags, behaviors) must be confirmable from source. If you can't verify it, omit it — don't assume.