Plugins

A separate package can add adapters and classifiers without changing dataveil. It registers them as entry points, and the built-in ones are registered the same way:

# the plugin package's pyproject.toml
[project.entry-points."dataveil.adapters"]
acme = "acme_dataveil:make_adapter"

[project.entry-points."dataveil.classifiers"]
acme = "acme_dataveil:CLASSIFIERS"

# acme_dataveil.py
from dataveil.core.config import CustomClassifier

CLASSIFIERS = [CustomClassifier(tag="ACME:ACCOUNT_ID", pattern="^AC[0-9]{8}$", name_hints=("account",))]

def make_adapter(options):  # options: the [adapter.options] table, as strings
    return AcmeAdapter(options["account"])

After pip install acme-dataveil, every classification includes ACME:ACCOUNT_ID, and DATAVEIL_ADAPTER=acme or [adapter] kind = "acme" selects the adapter.

A classifier plugin provides definitions, not code: a CustomClassifier or Classifier, or a list of them. dataveil runs each one as the same aggregate COUNT as a built-in classifier, so a plugin can only produce counts. Its patterns follow the same rules as custom classifiers. A broken plugin, a pattern that fails those rules, or a tag that is already taken stops classification with an error instead of being skipped. An adapter plugin's factory must return an Adapter, and it must not reuse a built-in adapter name.

Testing a plugin

dataveil's own plugin tests write a package with its own .dist-info metadata to a temporary directory on sys.path. importlib.metadata finds it exactly as it finds a package installed with pip. See tests/test_plugins.py.