Astra-Ares
Patched Codex CLI where Jev reads task context mid-run and picks the next reasoning effort level and how long it should last.
Overview
Astra-Ares is an experimental reference implementation for adaptive reasoning effort in Codex. While a Codex task runs, Jev reads bounded task context and chooses how much the selected GPT-6 model (Astra, Sol or Luna) should think next and for how many generations that effort should last; Codex applies the change while work continues. The aim is to cut token usage by matching reasoning depth to each step.
It relies on GPT-6 models being able to change reasoning effort without invalidating the cached prompt prefix, so the same model, conversation and direct OpenAI connection are kept. Setup downloads a pinned Codex version and applies the included patch; you then pick an Ares entry in /model. Your Codex login is used for the model and a separate key for Jev; other OpenRouter decision models or a local decision service can replace Jev. The bridge is MIT and the patched Codex source Apache-2.0.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | How much reasoning effort does the next step need?(inferred) | LOW · MEDIUM · HIGH · ... |
| Choice | For how many generations should this effort last?(inferred) | — |
How it works
- Codex task runs on an Ares model entry
- Jev reads bounded task context
- Jev picks effort level and duration
- Codex applies the change without breaking the prompt cache
Strengths
- Aims to reduce token usage
- Keeps the cached prefix and direct model connection
- Decision backend is swappable
Caveats
- Runs a separate patched Codex CLI
- Early preview; reference implementation
- Savings not quantified in the excerpt
Getting started
Run the setup to install the patched Codex CLI, add your Jev key, then choose Astra Ares, Sol Ares or Luna Ares in /model.At a glance
- Type
- CLI
- Status
- prototype
- Runtime
- patched Codex CLI + bridge
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Adjusts a Codex task's reasoning effort mid-run by asking Jev how hard the next step looks. Runs a patched Codex CLI and says it is a reference implementation rather than an app.
MrJev/awesome-jev
Frequently asked questions
What does Jev decide?
The reasoning effort level for the next steps and how long to keep it.
Does it change models?
No, the model stays fixed; only reasoning effort changes.
Is it ready for daily use?
The README calls it an early-preview reference implementation.
Related projects
- jev-router★ 593
Per-turn model routing for Claude Code and Codex: wraps the real CLI and lets Jev send simple turns to fast models and hard ones to strong.
ProxybetaJavaScript - jev-codex-router★ 274
Per-call router for Codex where Jev picks the model and thinking effort for every model call, including tool continuations.
ProjectbetaPython TypeScript/Bun launcher that asks Jev which model and reasoning effort each Codex, Claude Code or Pi turn should use.
ProjectbetaTypeScriptGo CLI model router that uses one Jev analysis to recommend a model, harness and effort level, then selects deterministically.
CLIbetaGoLocal bridge where Jev selects model and reasoning effort for Codex Desktop and CLI, with a Windows service and cache-aware routing.
CLIbetaTypeScriptmacOS wrapper for Codex that lets Jev pick the model and reasoning effort for a thread, then pins that route.
Projectbeta