Use Python 3.12 or newer and install the checkout with development dependencies:
python -m pip install -e '.[dev]'
python -c 'import regolo; print(regolo.__file__)'The module path should point into this checkout's src/regolo directory. For
PyCharm, open the directory containing pyproject.toml, select the same virtual
environment, and mark src as a Sources Root if it is not detected automatically.
With uv, use uv pip install -e '.[dev]'.
requirements.txt contains runtime dependencies only. To regenerate it:
python -m piptools compile --strip-extras --output-file requirements.txt pyproject.tomlpython -m ruff check .
python -m ruff format --check .
python -m mypy
python -m pytest --cov=regolo --cov-report=term-missing
python scripts/check_devmid_contract.py tests/fixtures/devmid-3.9-public-openapi.jsonThe test suite uses HTTPX MockTransport and rejects real HTTP requests. It uses test credentials and temporary token files. Tests cover management requests, authentication, pagination, billing, exports, CLI behavior, inference, streaming, tools, structured replies, media inputs and transport ownership.
Schema discovery tests use simulated API extensions. These test documents are
not server exports. The SDK's py.typed marker and overloads support typed
consumer code; examples are in tests/typing_api.py.
GitHub CI runs lint, formatting, type checks, tests, contract checks and builds on Python 3.12, 3.13 and 3.14. Package, bumpver and module versions must agree.
tests/fixtures/devmid-3.9-public-openapi.json contains the 3.9 public user API
contract. Compare another public export against it with:
python scripts/check_devmid_contract.py public-openapi.jsonAdded or removed operations, parameter changes and request/response schema changes produce a nonzero exit status. Descriptions, titles and examples do not affect the comparison. Update the implementation, tests and documentation for API changes before replacing the fixture. Include only public user operations and their schemas; use test data instead of production records or credentials.
Live inference tests make paid API requests and run separately from pytest:
REGOLO_API_KEY='<KEY>' python scripts/smoke_inference.py --run-paid --max-requests 16 --report live-smoke-report.jsonOmit REGOLO_API_KEY to enter the key at a password prompt. The runner selects
accessible catalog models, caps requests at 16, generates at most one 1024×1024
image and limits chat calls to 128 output tokens. It makes no automatic retries
and excludes credentials from its report. Audio inputs can use generated silence
to check upload/response formats. The runner does not exercise custom deployments
or video input.
Use --only structured_json function_tools to select workflows. Incomplete or
truncated replies are failures. Test custom deployment lifecycles separately
with authorized test accounts and explicitly unload billable resources afterward.
For a focused reranking check, use --only rerank --max-requests 2. This makes
one catalog request and one reranking request. Qwen3 rerankers use their required
instruction/query/document tags; the check still requires the relevant document
to rank first.
python -m buildBuilds are written to dist/. Wheels include runtime modules and the CLI entry
point. Source distributions also include tests, the public schema fixture,
scripts and documentation. Rebuild release files after changing the source.