it seems we should be storing chains of prompts instead of code, and then replay them when a new model is released.