Codex Jev Router — tiandee

It routes OpenAI Codex CLI turns through Jev: a local loopback Responses API proxy sends routing context to Jev, which selects a suitable Codex model and reasoning effort for each fresh turn, then forwards the request to Codex. The bridge runs locally.

Codex continues with its current model and effort - the bridge does not block work. Without a Jev key at all, the wrapper still starts Codex and prints a fallback notice; add the key when you want automatic routing.

By default the bridge maps Jev's reasoning_required score: below 0.30 low, below 0.60 medium, below 0.85 high, otherwise max. Set JEV_CODEX_AUTO_EFFORT=0 to preserve the effort selected in Codex, and a model without the requested effort falls back to the strongest supported lower level.

Clone the repository, then npm install and npm link, then check with codex-jev --version. The package is not published to npm, and it needs Node.js 20 or newer plus an authenticated OpenAI Codex CLI available as codex on your PATH.

Store it outside the repository: write JEV_API_KEY to ~/.jev-codex.env, which the bridge loads automatically, and chmod 600 the file. You can also export JEV_API_KEY in the shell that starts Codex. Do not commit the key or place it in a tracked file.

Yes. For each fresh turn the bridge adds a Codex commentary item rendered with the same layout and colors as the rest of the conversation, for example: [Jev] routed this turn to gpt-5.6-sol (max reasoning, confidence 0.95).

tiandee publishes it, listed as a madewithjev build with the MIT badge in the README; the listing notes the repository ships without a description, so the README is the primary source. Normal Codex arguments such as --model and --sandbox pass through the wrapper.