Really interesting project. I'm working on something similar, and I had a lot of trouble making native CLIs work with tracking their state. So I'm curious how you handle it: how do you tell when a task is complete or a session is stuck, and how do you avoid burning tokens, for example in loops? I tried quite a few approaches before finding one that works and is efficient enough.

I had the same problem with messages between agents. Getting them to talk to each other isn't that simple, especially when they use different CLIs. How do you handle that?