Jev Auto Router Guide

How to route Codex GPT models per call with Jev

A coding task mixes hard reasoning with routine follow-ups. Per-call routing chooses model and reasoning effort inside the same Codex session — not by spawning a new worker for each task.

Jev Auto Router is an experimental Codex model router. TypeSafe's Jev makes one typed Choice over host-supported (model, reasoning effort) pairs. A local Responses proxy keeps the Codex session and tool loop continuous while the route can change call by call.

Status honesty first.

Architecture is approved; runtime is a validation prototype. Cross-model switching in a real Codex tool loop, the complete verification path, and savings have not been proven end to end. This guide explains the design and prerequisites — not an install-and-auto-route promise.

Jev Auto Router per-call architecture diagram

What you need

A Jev key alone does not make live per-call routing available.

What happens on each call

  1. Local Responses proxy receives the call, checks whether routing is off or the call is infrastructure-only, and builds only pairs the host can request.
  2. Jev Choice selects model and effort together. Ordinary candidates are Luna Max, Terra, and Sol. GPT-6 is ineligible by default.
  3. Execution guard validates ALLOW or DENY. It does not re-pick a different plan. Timeout, low confidence, or invalid answers fall back to Terra/medium with a recorded reason.
  4. Native Responses executes with the Codex session and tool loop intact. Logs record the actual model, effort, usage, and fallback reason.

What this is not

Local development commands

These validate current prototype code; they do not connect Codex to a production proxy:

npm ci
npm test
npm run typecheck

Related guides