I wonder if the simplest challenge that works would be to simply render a button in a canvas. That way the bare minimum that the crawler needs to do is run a vision pipeline to find the button.