While a Codex task runs, Jev looks at what is coming next and decides how hard the model should think about it, then adjusts the reasoning effort on the fly. The model and conversation stay the same, so nothing already cached is lost.
The goal is fewer tokens spent on simple steps without weakening the hard ones.
An experimental reference project. It builds its own patched copy of the Codex command-line tool, needs about 10 GB of free space for the first build, and is tested on Apple Silicon Macs.
What you get
- Full JavaScript source code, MIT licensed
- Codex patch and automated build setup
- Configuration and troubleshooting docs
Jev
JavaScript
Codex
Rust
Node.js
Claude Code
Python
OpenCode