Author here. Remember Chip's Challenge? The 1992 Windows one from the Microsoft Entertainment Pack. I wanted to play it again so I ported it to the browser. Well, Claude did. I was too lazy to write the code, so I made a rule that I wouldn't write any and just bossed it around.
Between us we've solved 80 of the 149 levels. It didn't come free: plenty I had to hand-hold or play through myself, and for the nasty ones I had Claude build a solver that watches YouTube speedruns and rebuilds the moves frame by frame (oddly satisfying to watch: https://www.youtube.com/watch?v=6wndAf4EXNc).
Obligatory IP note: this is someone else's game. I'm assuming it's effectively abandonware but I honestly don't know, so the site might have to come down at some point. The repo is stripped of all the original art and assets, code only.
Nice job - but it's definitely not abandonware having been re-released on Steam [1] (along with a sequel) back in 2015.
Regarding the verifier that plays against the live engine, I’ve approached the problem from a similar angle by having LLM agents effectively borrow a page from the speedrunning community in the form of tool-assisted speedruns, allowing the LLM access only to a virtualized game controller.
Even if you do pull the game itself I would still definitely leave all the post mortem stuff up. I think it's just as interesting and worth keeping around - especially the YT vids demonstrating the harness.
I don't have a GH repo up for the TAS system yet - it's a bespoke mess right now since it was built with the old game "Castle of the Winds" in mind but I'll definitely consider it in the future!
Regarding the verifier that plays against the live engine, I’ve approached the problem from a similar angle by having LLM agents effectively borrow a page from the speedrunning community in the form of tool-assisted speedruns, allowing the LLM access only to a virtualized game controller.
[1] - https://store.steampowered.com/app/346850/Chips_Challenge_1