jev-codex-router

All projects · Developer tools

Per-turn Codex model and reasoning routing through a local server that asks Jev to pick a tier, then relays Responses traffic through an existing Codex Router install.

At a glance	Details
Source	Source
Maintainer	0xNatoshi (Thibault Saint-Jean). Independently curated; this entry is not an upstream submission or endorsement.
Format	Python local HTTP server (server/jev_server.py) plus Codex Router generic-provider registration; optional hook/PoC and backtest tools. Upstream repository is archived on GitHub as of this review (still publicly readable under MIT).
Requirements	macOS-oriented Codex desktop setup with a working Codex Router checkout (bin/codex-router), Python 3.11+, and TYPESAFE_API_KEY (default env file ~/.hermes/.env). Offline source inspection needs no account.
License	MIT. ChatGPT / OpenCode tandem usage and TypeSafe inference have separate account and cost requirements.

When to use

Use it when Codex already runs through Codex Router and you want each turn (including tool-step continuations) classified by Jev so mechanical work can stay on cheaper tiers while hard steps keep a frontier model. Prefer a simpler fixed model when you do not want a local classification server or Codex Router dependency. Because the upstream repo is archived, expect no further upstream fixes unless the author un-archives or publishes a successor.

It is distinct from catalogued jev-router (CLI launchers/proxies for Claude Code and Codex) and Jev Model Router (Claude Code function-hook mod): this project plugs into Codex Router's generic provider and curated jev/auto model rather than wrapping the CLI itself.

How it works

Codex → Codex Router (:4202)
          └─ "jev/auto" → local jev_server.py (:4319)
                → Jev classifies tier + thinking depth
                → policy picks luna / sol / astra (or Codex-dry tandem)
                → relay Responses SSE via the router caller edge

The server posts to https://api.typesafe.ai/v1/systemone with default model jev-latest. Default policy maps mechanical work to gpt-5.6-luna (max thinking, priority), standard work to gpt-5.6-sol, and hard/ambiguous work to gpt-6-astra. Confidence below 0.5 holds to the middle tier rather than downgrading further. Failures fall open to astra; a kill-switch file skips Jev entirely. When native ChatGPT usage is exhausted, an optional tandem substitutes OpenCode GLM/DeepSeek flash models. Decisions are logged under ~/.codex/codex-router/jev-router-live.jsonl.

Upstream BACKTEST.md reports measured savings on a replay protocol; those figures are upstream research, not catalog-verified quality or cost guarantees.

Get started

Setup clones source and starts a local server; live routing needs a TypeSafe key and an existing Codex Router session.

git clone https://github.com/0xNatoshi/jev-codex-router.git
cd jev-codex-router
git checkout 8292b519659280884627a962c826ac7721136a64
# provide TYPESAFE_API_KEY via the environment or ~/.hermes/.env
python3 server/jev_server.py
curl -s http://127.0.0.1:4319/health

Then register the generic provider and curated model with your Codex Router checkout as documented in the upstream Quickstart and AGENTS.md. Live turns send prompt/tool digests to TypeSafe and model traffic to the ChatGPT (or tandem) backends; provider charges or subscription limits can apply.

Examples and demos

• BACKTEST.md and poc/ replay tools document the savings protocol and sample JSON; not re-run for this listing.
• CI workflow is present; this catalog review did not execute it.
• No separate automated unit-test suite was found in the reviewed tree.

Limits and data handling

Jev receives turn/tool digests as constructed by the server (including tool-output digests for continuations). Full coding traffic still flows through Codex Router and the selected backend. Thresholds and reported savings are operational/policy research, not independently validated accuracy. The Codex-dry tandem and native model IDs can change with provider availability. macOS + Codex Router prerequisites were not reproduced on the Linux review host.

Review and maintenance

Reviewed on 2026-09-20 at commit 8292b51: MIT. Upstream GitHub repository is archived (read-only); listing remains for discoverability of the last reviewed revision. AI-assisted source review of server/jev_server.py, README, BACKTEST notes, license, and repository layout. No automated offline test suite was run; no live TypeSafe calls, Codex Router install, or end-to-end routing session were performed.

Related: jev-router, Jev Model Router.