Testing

Testing

Four layers, each covering what the one before cannot.

Layer	Where	Proves
Unit	src/tests/XmlComments.UnitTests	The rules (ID normalisation, which IDs are removed, overlap detection) and both tasks, run in-process against a mocked IBuildEngine.
Integration	src/tests/XmlComments.IntegrationTests	Real consumer solutions, built with the real OpenAPI generator on net10.0 and net11.0, importing the props and targets from source. Each API requests its own /openapi/v1.json. The same consumers are diagnosed with the skill’s script, so its report stays accurate.
Repository	src/tests/Repository.UnitTests	Every authored skill conforms to the Agent Skills specification and carries the product prefix.
Packed consumer	scripts/test-packed-consumer.ps1	The .nupkg itself, through PackageReference alone: the buildTransitive import, the packed task path, and nothing left in the build output.

The integration suite includes Build_GivenDeduplicationDisabled_ReproducesTheDuplicateKeyFailure, which turns the package off and asserts the generator still fails. It is what shows the other tests would notice a regression.

just test               # every tier
just test-unit          # unit + repository
just test-integration
just smoke-packed-consumer

The integration tier needs pwsh on the path for the script tests; GitHub’s hosted runners have it.

Every consumer compiles the same documented public attribute into the API and each library, which is what a source generator’s post-initialization output does. The tests therefore depend on no particular generator.