BefikerG/agent-qa ? reverse-engineered prompt

Build me an evidence first evaluation app for comparing Browser Agent runs with WebMCP runs on deterministic commerce tasks.

I want it to keep structured traces, save evidence and final state snapshots in SQLite, and only mark a task as successful when the observed state actually proves it. If something is missing or unclear, it should stay unknown instead of being counted as a pass. Add a safety layer that can block or stop risky actions like payment or order related steps, and make those stops show up clearly in the results. The browser side should run against a local fixture storefront with Playwright, and the app should be able to replay scenarios, compare outcomes, and give a simple diagnosis or recommendation at the end.

Please keep the codebase in TypeScript, use the existing Next.js app shell, and wire up the tests so the main evaluation flow is reliable. Look up current docs online if you need to.