Greenroom

Greenroom is QA for the screens a pull request actually changed. It reads the diff, works out which screens are in scope, and walks them the way a user would: a real browser for web, a real iOS Simulator for apps. Then it posts the handoff on the PR as a GitHub Check, with screenshots, repro steps, and a pointer to the likely line of code.

There are no test scripts to write. Scoping comes from the code first and a model second, and when the impact of a change is unclear, the pass gets wider, not quieter. The driver takes one action at a time, and each one is checked against a deterministic policy before it runs.

The driver never grades its own homework. Separate judges, from a different model provider than the one that drove, decide what broke. Every role has a fallback on another provider, and the final report is assembled by a template with no model calls at all. It is the boring-contracts argument applied to QA: models make the judgment calls, and code owns everything else.

It is advisory on purpose. A clean pass is a green Check. Anything else is neutral, never a blocked merge. I would rather earn a blocking gate than assume one.

Its first real run outside the eval harness walked its own landing page and found three real defects. My own apps are its first customers.