vladyslavdmitriiev/ai-qa-pipeline

This project turns plain-language feature descriptions into automated tests you can actually trust. You write a short description of what a feature should do, and the system produces Playwright tests that check your app against that description. The tests go through multiple stages of review before they are accepted. The core principle is that no agent in the pipeline gets to declare its own work done. A parser checks the feature file before any AI model is called, and it rejects structurally broken inputs for free. Then a Feature Reviewer agent checks whether a test writer could implement the spec without guessing at details. It rejects scenarios that are vague or contradictory, but it does not reject based on style or missing edge cases. A Test Writer agent then drafts the actual test code. It can only use element identifiers that already exist in your app's source code, which prevents it from inventing selectors that would never match anything. A Spec Judge running on a different AI model reviews the draft before it ever runs. Using a different model matters because a judge from the same model family tends to excuse the writer's own patterns. The judge checks that every scenario is present, every assertion can actually fail, and no meaning was weakened. If the judge approves, the test runs for real against the app. If it fails, a Debug Agent may fix mechanical problems like locators or timing, but it is not allowed to change what the test asserts. After debugging, the spec goes back to the judge one more time to confirm nothing was quietly weakened just to make it pass. A mutation gate then runs as a final check on the frozen baseline. The project includes a demo shop so you can try the whole pipeline locally. You need Node 20 and the Claude Code CLI installed. The full README is longer than what was shown.