$ protest run tests:session -n 4

A colleague and I kept talking about the magic side of pytest. We love the framework, but fixtures resolved by name, no types, no Ctrl+Click.

I wanted something more declarative, closer to what FastAPI does with DI.

Since v0.2.0, ProTest treats LLM evals as first-class citizens: an eval is just a test that returns a value — scored, not asserted. Same fixtures, same DI, same parallelism; judge, scoring and short-circuit built in. Your evals live right next to the tests they ship with, not in a separate framework.

The project grew fast. Thread pools, async exit stacks, event bus, tree-based scoping... Low-level async stuff you never touch when building APIs.

To validate the approach, I rewrote large parts of pydantic, httpx and starlette test suites with ProTest. Result: on httpx and starlette, tests run 20-30% faster than the official suites, thanks to native async.

v0.2.0 released (release-please, CI, docs). Most importantly, continuously dogfooded: ProTest is the eval harness of Felix↗, my continuity assistant — every behavior of its LLM engine is a replayable eval.