screenwright

The wedge is the eyes. Generating UI is solved; verifying the render against a brand + accessibility bar, in a loop is not. screenwright paints one self-contained HTML surface, then renders → screenshots → audits → fixes until it passes, and hands the result back. It does the painting and verification; it does not ship a runtime or convert to a framework — that's the host's job.

Agents write frontend blind: the markup looks right but misses the small obvious things — a 4px misalignment, a missing focus ring, text that overflows on a long name, a mobile layout that never got built. screenwright gives the agent eyes.

Two stacked gates, asymmetric:

• axe a11y — machine, deterministic, always blocks. The defensible core.
• Visual fidelity — judgment from the screenshot. Blocks only when a reference image exists; advisory otherwise (no reference = no ground truth to gate on).