Astra-Ares

Checked, not fully reviewed. This one declares its own limits clearly enough that a pretend hands-on would have been dishonest: it installs a separate, patched Codex CLI, built from a pinned source with Rust and a C toolchain, about 10 GB of disk, and integrates over a Unix socket. Its own platform table says macOS Apple Silicon is “built and tested locally”, Linux is “build paths provided; not yet acceptance-tested”, and Windows is unsupported. We ran its tests and read the integration; we did not build the patched Codex.

The idea is the interesting part, and as far as we know it is the only project in this directory doing it: GPT-6 models can change reasoning effort without invalidating the prompt prefix that caching depends on, so a cheap typed decision between steps can raise effort for the hard turn and drop it for the easy one, inside a task that is already running. Everything else in this category picks a model or an effort level once, at the start.

The README is honest in a way that is worth rewarding: it opens by calling itself an experimental reference implementation, says plainly that it runs a patched Codex, and separates what the author has tested from what he has only written down.

If you are on an Apple Silicon Mac and willing to run a patched agent, this is the experiment to watch. If you are not, read it for the mechanism rather than installing it.

For effort and model selection made once per turn instead, see jev-router and compact-adviser.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 27, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.