Vibe Coding DiscoverAstra-Ares
View on GitHubAdaptive reasoning effort for GPT-6 during Codex tasks, powered by Jev to reduce token usage.
Astra-Ares is an experimental patched Codex CLI that lets an LLM evaluator (Jev) pick GPT-6 reasoning effort per generation, aiming to cut token usage while preserving prompt-cache prefixes. Requires Node 22+, Rust, and a ~10GB local Codex build.
Use Cases
Dynamically lower reasoning effort on easy Codex steps to save tokensRun coding tasks with a patched Codex CLI that mid-conversation changes effortEvaluate next-step reasoning depth from bounded tool and task contextKeep prompt-cache prefix valid while switching reasoning effortInspect effort-change decisions via JSONL decision logsValidate and debug provider config with ares doctor
Built With
- Language
- JavaScript
- Frameworks
- Codex CLI · Node.js · Bun · Rust · OpenRouter API
Tags
adaptive-reasoning · reasoning-effort · token-optimization · codex-cli · cli · prompt-caching · llm-routing · agent-harness · cost-reduction · patched-cli · nodejs · openrouter · experimental