Switchboard: can a System One model route coding agents?

TypeSafe had just released Jev, a "System One" model that answers typed questions in a single fast pass and gives each answer a confidence score, without generating any text. I was curious whether something that quick and cheap could work as a router for coding agents, reading a task and picking the right model and reasoning effort, so routine work goes to a smaller model and the hard problems get the strongest one.

Switchboard is the open-source TypeScript CLI I built to try it. It sits between Claude Code or Codex and the model provider as a local proxy. When a new conversation starts, it asks Jev a few typed questions about the task, like how much capability it needs and how much effort it deserves, and a routing policy turns the answers into a model and effort. It falls back to a safe default when Jev isn't confident, and it keeps the same choice for the rest of the conversation so follow-ups don't switch models and lose the cache. Each decision took about half a second and cost well under a cent.

After my experiments, I don't think a model like this works as a router. Coding tasks are relative to the codebase: the same prompt can be simple in one repo and hard in another, so you can't always tell how complex a task is from the prompt. Jev's context window is also limited, so you can't even send it much about the codebase to make up for that. The code is still on GitHub and npm.