One implementation thought I've had while doing something a bit similar (having GPT-3 play Zork): instead of mucking through the game state, you could also just

One implementation thought I've had while doing something a bit similar (having GPT-3 play Zork): instead of mucking through the game state, you could also just run "look" and "inventory" regularly. Mostly these are no-ops, but it should be simple enough to execute them and then revert the state (though I haven't tried). In some ways this would be more general and perhaps more robust (or maybe not). Though it's also pretty neat to see that internal state of Zork.