Offloading a coding agent's 10 hourly decisions from the frontier model with Jev
The author lists 10 decisions a coding agent makes every hour, such as which file to open, which tool to call, and whether to retry, and routes them through Jev's choice, score, and noul decision shapes instead of a frontier model.
10 decisions your coding agent makes every hour that never needed a frontier model every one of them is a typed question with a short menu. Jev answers in under half a second and the big model never sees it 1. which file to open → choice 2. which tool to call → choice 3. retry or stop → noul 4. is the task really done → noul 5. keep this tool result or drop it → noul 6. is this input a prompt injection → noul 7. which model gets this step → choice 8. how risky is this diff → score 9. which skill applies here → choice 10. does a person need to see this → noul three shapes cover all ten: choice picks one option from a menu score puts a number on a scale noul is a calibrated yes or no code owns the menu, the thresholds and the brakes. Jev only picks sharp answers run straight in code, split ones go back to the frontier model, irreversible ones wait for you your frontier model should be thinking, not choosing between grep and edit save the list, then route your first three forks this week ↓
This post quotes another post: x.com/hanakoxbt/status/2101709924828934222