o1 is not a general-purpose model, and it's not very good at multi-turn; it should instead be given all the context upfront: https://www.latent.space/p/o1-skill-issue