Repository navigation
CI: add a bare-install smoke job (pip install ondine, no extras) #189
Description
Two bugs shipped in 1.11.0 (both fixed in #188) were invisible to CI purely because every CI job installs the dev environment, never the published package as a user gets it.
| Bug | Why CI couldn't see it |
|---|---|
enrich(polars_df, ...) → ModuleNotFoundError: No module named 'pyarrow' |
uv sync --all-extras installs pyarrow, so the missing declaration was masked. polars is a core dep; pyarrow is only in parquet/all. |
enrich(schema=...) → ValueError: Either dataframe or source_path must be provided |
test_enrich.py mocks Pipeline.execute, so the broken code path never ran. |
Both were caught in under five minutes by doing what a user does:
python -m venv /tmp/v && /tmp/v/bin/pip install --only-binary ondine ondine
/tmp/v/bin/python -c "import ondine, polars as pl; ondine.enrich(pl.DataFrame(...), ...)"Proposal
Add a CI job that:
- Creates a clean venv and installs the built wheel with no extras (
--only-binary ondine, nouv sync). - Exercises the public API for real against a cheap provider (or a stub server) — at minimum:
ondine.enrich(pandas_df, ...)ondine.enrich(polars_df, ...)← catches the pyarrow class of bugondine.enrich(..., schema=Model)← catches the rebuild class of bugondine.enrich(csv_path, ...)
- Repeats for the
[mcp]and[all]extras so extras-only imports are also checked.
The existing Install Profile jobs install extras and only run import ondine — they verify the package imports, not that it works. This job should call the API.
Related: the enrich() unit tests should also grow at least one test that runs a real pipeline end-to-end with a stub LLM client instead of mocking Pipeline.execute, so the loader stage is actually exercised.
Activity
test(ci): exercise the public API on a bare install (#189) (#211)