Dev tools

donvito's codex-astra-luna-orchestrator lets Codex run GPT-6 Astra as the boss and GPT-5.6 Luna as the worker

September 11, 2026 at 11:35 AM PT

GitHub social card for the donvito/codex-astra-luna-orchestrator repository

Image: GitHub

Why it mattersA team paying for Codex on either the Pro or the Plus plan now has a shared, documented pattern for splitting the expensive planning model from the cheaper execution model, and a script that reports what each thread actually cost.

donvito/codex-astra-luna-orchestrator, a Codex configuration package that installs a hierarchical setup with GPT-6 Astra as the root orchestrator and GPT-5.6 Luna as the default subagent, hit 971 stars on GitHub in the six days since it was created on 5 September, per the repository's public star count and creation date on the GitHub API. The author describes it as "a configurable Codex setup where GPT-6 Astra is the root/orchestrator and reviewer, while GPT-5.6 Luna is the default and pinned model for execution subagents."

Two plans, one pattern

The package ships two profiles, matched to the two paid Codex plans. On the Pro plan, Astra is pinned to the root at medium reasoning effort and Luna picks up the explorer, worker, tester and researcher subagent roles at maximum effort. On the Plus plan, Luna runs the root at maximum effort as well, because the author's note reads: "Plus users: the root thread is the largest line item, so running it on Luna saves the most." A separate reviewer role runs at low effort in both plans, on Astra, and stays in the loop as a final independent check on subagent output.

The pattern rests on a claim the author sets out directly in the README: "orchestration is not free: the root stays in the loop for the whole task and every subagent carries its own context." That is the reason for holding the cheap model in the worker slot and putting the expensive one only where it decides what happens, or in the reviewer seat where it looks at what did happen.

The token accounting is the part worth reading

Under scripts/token_usage.py the repository ships a Python analysis script that reads the rollout logs Codex already writes under ~/.codex/sessions and reports usage per thread, per role, and per model, plus the change in the account's 5-hour and 7-day rate limit windows. That is the raw material for deciding whether the split is actually paying for itself on a given task shape: without it, the choice between Pro and Plus is a story about which model is smarter rather than a story about which pattern spends fewer tokens for the same outcome.

What it does and does not do

The package installs skills, agent role profiles as TOML files, and a AGENTS.md orchestration guide, so it is a configuration and skill drop, not a new agent runtime. It runs on top of Codex, which does the work of routing tool calls, holding sessions and issuing prompts. That is deliberate: the codex-astra-luna-orchestrator repo does not try to replace Codex, it tries to be the shared, versioned answer to "how should I set up Codex with these two models".

For a team already on Codex, that is the practical value. Two engineers reading the same short README can stand up the same orchestrator/worker split and look at the same numbers to decide whether the reviewer step or the max-effort subagents earned their tokens. That is a cheaper answer to reach than each team writing the split from scratch.

Source

The repository is donvito/codex-astra-luna-orchestrator on GitHub. Star count and creation date verified against the public repository API at the time this went out.

Source: GitHub

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

Engrim ships a local SQLite memory store that survives switching between Claude Code, Cursor, Antigravity and Codex

Engrim is a local-first SQLite memory engine for coding agents that carries architectural decisions and project state across sessions and across models, from Claude Code to Cursor to Google Antigravity to Codex CLI.

Source: Hacker NewsDev tools

fable-orchestrator, a Codex skill where Claude Fable 5.1 plans and cheaper models write the code, reached 555 GitHub stars in five days

fable-orchestrator is a routing skill for the Codex harness that has Claude Fable 5.1 plan and adjudicate while GPT-5.6 Luna and DeepSeek V4 Flash do the implementation, and it reached 555 GitHub stars in five days.

Source: GitHubDev tools

Pizza Bot, a new open-source app from AWS engineers, gives long-running AI agents an email-style inbox

Pizza Bot launched on 10 September under Apache-2.0 as a self-hosted desktop app where finished background agent work arrives in an Unread queue, anything paused for approval lands in an Action queue, and runs keep going after the client disconnects.

Source: PressDev tools