donvito/codex-astra-luna-orchestrator

What it solves

This project provides a structured configuration for Codex, setting up a multi-agent orchestration system where a high-reasoning model (GPT-6 Astra) acts as the manager and a faster model (GPT-5.6 Luna) handles the execution tasks. It solves the problem of balancing high-quality architectural oversight with efficient, token-conscious execution of coding tasks.

How it works

The system uses a root orchestrator and specialized subagents. Depending on the user's plan (Pro or Plus), the root orchestrator is either Astra or Luna. Specialized roles—including an explorer, worker, researcher, and tester—are pinned to the Luna model for execution. A separate reviewer agent, pinned to Astra, performs the final independent review to ensure the code is verified by a different model than the one that wrote it.

Who it’s for

Developers using Codex who want to automate complex coding workflows through a coordinated team of AI agents with different reasoning capabilities and costs.

Highlights

• Multi-Model Topology: Uses a tiered system where Astra orchestrates and reviews while Luna executes.
• Plan-Aware Installation: Includes an installer that configures the root model based on whether the user has a Pro or Plus plan to optimize rate limits.
• Specialized Agent Roles: Pre-defined roles for exploration, research, implementation, and testing.
• Token Tracking: Includes a Python script to analyze Codex session logs and report token usage per thread and model.
• Customizable Reasoning: Allows tuning of reasoning effort (low, medium, high) for different agents to balance speed and and cost.