Astra directs, Luna executes: tiered Codex agents

What it does

This repository is a bundle of Codex configuration templates and installer scripts that implement a tiered multi-agent coding topology. GPT-6 Astra serves as the root orchestrator and independent reviewer, while GPT-5.6 Luna handles the execution subagents—explorer, worker, tester, and researcher. Two profiles map to Codex subscription tiers: Pro keeps Astra at the top, while Plus pushes orchestration down to Luna to cut costs. It is essentially curated TOML presets and shell scripts copied into your target project.

The interesting bit

The setup treats model selection as resource scheduling rather than a capability arms race, deliberately pinning grunt-work roles to Luna while Astra handles direction and final review. A bundled scripts/token_usage.py reads Codex session logs and breaks down token spend per thread, role, and model against your rate-limit windows.

Key highlights

• Two ready-made profiles (Pro and Plus) keyed to Codex subscription tiers
• Explicit role pinning: Luna for explorer/worker/tester/researcher, Astra for reviewer
• Installer scripts for macOS, Linux, and Windows that copy configs without rewriting them
• Token usage tracker that parses Codex session logs by thread, role, and model
• Guides for tuning concurrency and reasoning effort for different codebase sizes

Caveats

• This is fundamentally configuration glue and shell scripts, not a standalone application
• Token costs depend heavily on repository size and task shape; the README notes there is “no single number” for expected usage
• Changing default_subagent_model does not override explicitly pinned role files

Verdict

Worth a look if you are already using OpenAI Codex and want to optimize token spend across a multi-agent workflow. Skip it if you are not on a Codex Pro or Plus plan, or if you prefer a single-model, single-agent setup.