I've been fine tuning my own decision models lately and I've been interested in how there are no real multi-turn harnesses out there for decision models. I decided to build my own. The standard claude code style interface doesn't quite make sense with a decision model since it isn't a conversation so I built this out as a sort-of REPL inspired interface. You can also use it with the command line.

You just type facts as you get them and ask questions with a ?. The questions stick around so every new fact gets asked again and you can see the probabilities move. You can use this to chain together different series of decisions to power decision networks and graphs with nouli. You can bring your own Typesafe, OpenAI, OpenRouter, or local model to run with it.

There's some walkthroughs with real runs here: https://standing-desks.github.io/nouli/examples/

Would love feedback.