I have the urge to fork this, modify the score coding to give a smoother gradient between total failure and pure success, and then solve the game with a genetic algorithm.

Except I'm at work, so I can't do it. So I'll just post my idea on Hacker News and settle for the idea that someone, somewhere, might.

One alternative is to just use nethack (or a variant like crawl) as a testing ground for your AI agent.

It's using the file system as UI! The only place I've seen this before is version control systems, and I'm now sold that it's a great paradigm.

A game like this would be a great way to get children into programming.