dvcgen 0.9.0

Generate DVC pipeline files from Python declarations.

dvcgen

Write your DVC pipeline once, in Python.

dvcgen is an early-stage command-line tool for generating DVC pipeline files from lightweight declarations embedded in Python pipeline scripts.

Current Status

Implemented:

• A Python package named dvcgen
• A dvcgen console command
• CLI argument parsing for pipeline script paths
• CLI input validation and overwrite protection
• Public declaration helpers: stage(), dep(), out(), and param()
• Python script inspection for top-level literal declarations
• dvc.yaml generation
• params.yaml generation

Installation

uv tool install dvcgen

Or run without installing:

uvx dvcgen --help

Usage

Show CLI help:

dvcgen --help

Generate DVC files from one or more Python pipeline scripts:

dvcgen pipeline/*.py

The command writes dvc.yaml and params.yaml in the current directory. Stage names are derived from input Python filenames. For example, pipeline/train.py becomes the train stage.

By default, dvcgen refuses to overwrite existing dvc.yaml or params.yaml files. Use --force when you intentionally want to replace them:

dvcgen --force pipeline/*.py

Write files to another directory with --output-dir:

dvcgen --output-dir generated pipeline/*.py

Bad inputs fail with an error message and a non-zero exit code. Successful runs print the files that were written.

Inspect declarations from Python without executing the pipeline script:

from dvcgen.inspect import inspect_file

declarations = inspect_file("pipeline/train.py")
print(declarations.deps)
print(declarations.outs)
print(declarations.params)

Release

Publishing is intentionally manual while the project is early stage. Build and validate artifacts before uploading anything:

uv run python -m build
uv run twine check dist/*

Use TestPyPI first when rehearsing a release. Create a TestPyPI API token, then upload with the token as the password:

uv run twine upload --repository testpypi dist/*

Use the production PyPI repository only when the version, changelog, and package name decision are ready:

uv run twine upload dist/*

For both repositories, use __token__ as the username and the repository API token as the password. Avoid committing tokens or storing them in project files.

Before the first production upload, decide whether to publish the current minimal release to reserve the dvcgen package name on PyPI. Once a version is uploaded to PyPI or TestPyPI, that exact version cannot be uploaded again; bump the version before retrying with changed artifacts.

Planned MVP

The intended MVP is:

1. Pipeline scripts declare dependencies, outputs, and parameters in Python.
2. dvcgen inspects those declarations without executing the scripts.
3. dvcgen writes dvc.yaml and params.yaml.

Example API:

from dvcgen import dep, out, param, stage

stage(
    cmd="python -m pipeline.train",
    wdir=".",
    desc="Train model",
    frozen=False,
    always_changed=False,
)

TRAIN_DATA = dep("data/processed.csv")
MODEL = out("models/model.pkl")

LR = param("train.lr", 0.001)

stage() declares metadata for the generated DVC stage. It is optional; when it is omitted, dvcgen keeps the default command:

"cmd": "python pipeline/train.py"

Supported stage fields are:

• cmd: override the command DVC runs for this stage
• wdir: stage working directory
• desc: human-readable stage description
• frozen: protect the stage from reproduction
• always_changed: always consider the stage changed

wdir, desc, frozen, and always_changed are emitted only when explicitly provided. Each pipeline script may declare at most one stage().

out() also accepts DVC output options as keyword arguments:

MODEL = out("models/model.pkl", cache=False, persist=True)

Supported output options are cache, remote, persist, desc, and push. Plain out("path") declarations continue to generate simple string entries. Metrics and plots are separate DVC stage metadata and are not modeled as out() options.

Running:

dvcgen pipeline/train.py

Generates dvc.yaml:

"stages":
  "train":
    "cmd": "python pipeline/train.py"
    "deps":
      - "pipeline/train.py"
      - "data/processed.csv"
    "outs":
      - "models/model.pkl"
    "params":
      - "train.lr"

And params.yaml:

"train":
  "lr": 0.001