dio/serabut ? reverse-engineered prompt

Build me a TypeScript tool that can extract records from one public website at a time using an evidence driven approach.

I want it to start from a URL, inspect the page and network traffic, prefer stable JSON or GraphQL requests when it finds them, and otherwise fall back to structured data or visible page content. It should keep track of why it chose each approach, and save the extracted records with provenance so I can see where each field came from.

Use a browser based loop for the agent, and make sure it only works on public pages. It should not use credentials, should not touch private network targets, and should only replay read only requests it has observed. I also want a simple command line experience, useful limits like max pages and max turns, and outputs written to a run folder with a report, findings, records, and CSV if possible.

If you need to check current docs for the browser or agent tooling, go ahead and look them up online.