Jev in production › Model and agent routing
A Pi extension asks Jev four questions per prompt to pick a budget-aware model tier before each turn, with automatic tier fallback.

A pi extension that routes every prompt to a task-appropriate model using TypeSafe Jev (System One) typed judgments. You type normally; before the turn starts, Jev reads the request and answers four narrow questions, code composes those into a capability tier, applies your budget policy, and pi switches to the matching model.
Every routed prompt shows the model it chose and the judgment behind it.
- Task-aware — planning goes to reasoners, implementation to coding specialists, chat to cheap fast models. - Budget-aware — daily/monthly caps downgrade tiers automatically instead of overspending. - Resilient — each tier is a candidate chain; if a model is unavailable or unauthenticated, the next one is used. - Visible — the transcript records the...
For the project's own README, linking back here: