donvito/codex-astra-luna-orchestrator

What it is

Codex Astra/Luna Orchestrator is a helper package that adds a ready‑made orchestration configuration for the Codex CLI tool. It bundles a set of sub‑agents (explorer, worker, tester, researcher, reviewer) that are each tied to a specific GPT‑6 model variant (Astra, Sol, Luna). By installing one of the supplied profiles into a target code‑base, developers get a pre‑wired multi‑agent workflow that can automatically plan, implement, test and review changes.

Who it’s for

Anyone who already uses the Codex command‑line interface and wants a plug‑and‑play multi‑agent setup for larger coding tasks—e.g. building new features, refactoring, or generating end‑to‑end implementations.

Core concepts

Concept	Meaning
Orchestrator	The root agent (Astra or Sol) that coordinates the whole job.
Sub‑agents	Individual GPT‑6 Luna agents that perform specific roles: explorer (maps the problem), worker (writes code), tester (runs checks), researcher (gathers info).
Reviewer	A separate agent (Astra or Sol) that validates the output before it is merged.
Profile	A folder under profiles/ that contains a complete set of configuration (.codex/) and skill files (.agents/). Selecting a profile copies these into the target repository.
Skill	The astra-orchestrator skill definition (SKILL.md) that tells Codex how to invoke the orchestrated workflow.

How to get it

1. Clone the repo.
2. Run the platform‑specific installer (setup.sh or setup.ps1).
3. When the script asks, point it at the other repository you want to augment and pick a profile (1‑6). The installer copies:
   • .codex/config.toml and the per‑role .toml files into the target’s .codex/ directory.
   • The skill files into the target’s .agents/skills/astra‑orchestrator/.
   • An AGENTS.md file (or appends to an existing one) that documents the available agents.

Using the orchestrator

From the target project directory run the Codex CLI:

codex

Codex will automatically load the installed skill. You can also call it explicitly:

$astra-orchestrator
Implement the invoice export endpoint. Use the explorer to map the path, a worker to implement it, and the tester and reviewer to verify it.

The orchestrator then spins up the four Luna sub‑agents, lets them collaborate, and finally hands the result to the reviewer agent.

Profiles at a glance

#	Name	Root model	Execution model(s)	Reviewer model	Concurrent sub‑agents
1 (default)	pro	Astra (medium)	Luna (max)	Astra (low)	4
2	plus	Luna (max)	Luna (medium)	Astra (low)	4
3	pro-max-2-subagents	Astra (medium)	Luna (max)	Astra (low)	2
4	plus-max-2-subagents	Luna (max)	Luna (medium)	Astra (low)	2
5	GPT6‑SolMax‑LunaMax	Sol (max)	Luna (max)	Sol (max)	4
6	GPT6‑SolMedium‑LunaMax	Sol (medium)	Luna (max)	Sol (medium)	4

Choose the profile that matches the model access you have (Astra, Sol, or Luna) and the desired level of parallelism.

Documentation & helpers

• Guides – step‑by‑step markdown files in guides/ covering full orchestration, plus‑profile setup, fast iteration, routine coding, complex repo work, and token‑usage measurement.
• Scripts – a small Python utility (scripts/token_usage.py) for tracking token consumption.
• Tests – unit tests (tests/) that verify profile files and token‑usage logic.

License

Apache 2.0 – free to use, modify, and redistribute.

---

All details above are taken directly from the repository’s README; no additional features have been inferred.