miuuyy/Astra-Ares

What it solves

Astra-Ares reduces token usage and cost when using high-reasoning models (like GPT-6 Astra, Sol, and Luna) by dynamically adjusting the "reasoning effort" based on the complexity of the current task step. Instead of running at a constant high effort, it matches the thinking depth to the specific need of the next generation.

How it works

It integrates with a patched version of the Codex CLI to implement an adaptive loop. A decision model called "Jev" analyzes the current task context—including the goal, public progress, and recent tool results—to determine the appropriate reasoning effort (e.g., LOW to HIGH) and how many subsequent generations that setting should last (a "lease").

Because GPT-6 models allow changing reasoning effort without invalidating the prompt prefix used for caching, Astra-Ares can update these settings mid-conversation while maintaining cache efficiency. The system communicates directly with OpenAI, with Jev acting as an external evaluator that triggers updates at specific checkpoints.

Who it’s for

Technical users and developers who use Codex and GPT-6 models for agentic tasks and want to optimize token consumption and performance through adaptive reasoning.

Highlights

• Dynamic Effort Scaling: Automatically switches reasoning depth based on the immediate requirements of the task.
• Lease System: Reduces API overhead by applying a reasoning choice for a set number of generations (1, 2, 5, or 10) before re-evaluating.
• Cache Preservation: Leverages native GPT-6 capabilities to change effort without breaking prompt-cache reuse.
• Bounded Context: Sends a curated subset of task data (up to 28,000 tokens) to the evaluator to keep decision-making efficient.