Jev Auto Router

Validation prototype · architecture approved

Per-call GPT model routing for Codex.

TypeSafe Jev chooses the model and reasoning effort for each call. A local Responses proxy keeps the Codex session and tool loop continuous. Independent verification checks the finished task. Router Compass puts route, usage, and acceptance in one place.

View on GitHub Read the README Browse guides
Honest status: architecture is set; runtime is still a validation prototype. Cross-model switching in a real Codex tool loop, the full verification path, and savings have not been proven end to end. This is not an install-and-auto-route promise.

Why per call

Hard reasoning and routine follow-ups do not need the same model.

One Choice per meaningful call

Jev selects model and effort together over host-supported pairs. Local code adds no task-kind table or second semantic router.

Same Codex session

The local Responses proxy preserves session continuity and the tool loop while the route can change call by call.

Independent verification

Task completion is checked against the original request. The executing model's claim of success is not evidence.

Architecture

Codex call → local proxy → Jev Choice → Responses → task verification.

The proxy builds only the `(model, reasoning effort)` pairs the host can actually request. Jev makes one Choice. An execution guard validates ALLOW or DENY. Native Responses keeps streaming, tool IDs, and continuation intact. After the task, verification and Router Compass connect route choices to real usage and acceptance.

Jev Auto Router per-call architecture: Codex session through local proxy, Jev, execution guard, native Responses, and independent verification

Capability tiers

Four tiers. GPT-6 stays scarce by default.

Tier Typical role Ordinary calls
Luna Max Explicit, mechanical follow-up steps Eligible
Terra Everyday work and the fallback baseline Eligible
Sol Harder reasoning, implementation, and correction Eligible
GPT-6 Scarce escalation supported by evidence Ineligible by default
  • Fallback default is Terra/medium with a recorded reason.
  • Turning the router off restores the host's original model.
  • Tier names must match pairs the host actually supports.

Needs

What you need before any live routing experiment.

TypeSafe / Jev API key

The proxy reads JEV_API_KEY. TypeSafe examples use TYPESAFE_API_KEY; both can hold the same key. Never commit secrets.

Local host stack

Node.js 22+, a signed-in Codex CLI, and access to the model and reasoning-effort pairs you want to route. A Jev key alone is not enough.

Guides

Search-oriented how-tos for Codex model routing.

Short technical pages for the queries people actually search: per-call routing, verification, and a Chinese-language walkthrough of the same design.

How to route Codex GPT models per call with Jev

Why per-call routing exists, what Jev chooses, and what the local proxy preserves.

Read the guide
How Codex model routing verification works

Independent acceptance, Router Compass evidence, and why savings are not assumed.

Read the guide
用 Jev 做 Codex 逐调用模型路由（中文）

面向中文社区讲清逐调用架构、前置条件和当前原型边界。

阅读指南
All guides

Browse the full collection from one index page.

Open guides hub

Try the source

Prototype code is open. Production routing is not claimed yet.

Local development validates the current proxy, routing decisions, verification, and Compass structures with npm ci, npm test, and npm run typecheck. Those commands do not connect Codex to a production proxy. License: Apache-2.0.

Star on GitHub Read ADR 0017 Blog write-up 掘金