Playwright Game Testing — giving the agent eyes

Our games are imperative Three.js drawing into a single <canvas>. Unit tests can't see pixels and the DOM is mostly an empty WebGL surface. This skill is how an agent observes the running game: boot it, drive it like a player, capture the canvas, read HUD state, and loop on the result until the change actually looks/behaves right. It is the load-bearing skill for autonomous visual dev.

Two ways to drive Playwright, use both:

• Playwright MCP server — interactive, inside your turn. Navigate, click, press keys, browser_take_screenshot, look at the image, decide, repeat. This is the self-correction loop.
• Playwright test files (@playwright/test) — committed regression specs in tests/. Use these to lock in behavior once you've verified it via MCP.

File layout