Use Luna Subagents

Sol High orchestrates. Luna Max executes.

Use Luna Subagents pairs GPT-5.6 Sol High as the parent orchestrator with GPT-5.6 Luna Max workers for routine, bounded implementation work.

Sol keeps responsibility for planning, judgment, synthesis, and final verification. Luna handles focused research, repetitive processing, independent implementation slices, testing, review, and high-volume transformations.

Codex does not provide this reusable Sol-to-Luna worker setup by default. This free and open-source skill safely adds the missing custom-agent configuration, shows the diff before making changes, and teaches the parent when—and when not—to delegate.

Illustrative workflow showing two bounded Luna tasks returning to the parent agent for consolidation and verification.

Why this pairing

• GPT-5.6 Sol High provides orchestration, judgment, synthesis, and parent-agent control.
• GPT-5.6 Luna Max provides efficient execution for clear, routine, and independently verifiable tasks.
• Use Luna Subagents bridges the default configuration gap while preserving explicit boundaries and parent verification.

What it does

• Installs or reconciles ~/.codex/agents/luna-worker.toml
• Pins the worker to gpt-5.6-luna with Max reasoning
• Shows a unified diff before changing configuration
• Backs up a conflicting agent file before replacement
• Preserves the parent session's sandbox and permissions
• Prevents overlapping write scopes between concurrent workers
• Refuses to silently substitute a different model
• Verifies results with the parent agent before completion

Install

Ask Codex:

$skill-installer Install the skill from https://github.com/aitransformationdirector/use-luna-subagents/tree/main/skills/use-luna-subagents

Start a fresh Codex thread after installation, then run:

$use-luna-subagents Configure my Luna worker and delegate suitable bounded work.

How it works

1. Inspect the current Codex and multi-agent configuration.
2. Show the exact Luna worker configuration diff.
3. Install or repair the standalone custom-agent file after approval.
4. Use Luna selectively for bounded work that benefits from isolation or parallelism.
5. Wait for workers and have the parent agent verify the consolidated result.

Requirements and limitations

• Requires a Codex release that supports custom agents and subagent workflows.
• Requires account access to gpt-5.6-luna and the requested reasoning level.
• A newly installed custom agent may require restarting Codex or opening a fresh thread.
• Configuration proves what Codex should request; runtime availability must still be verified.
• The workflow image above is an illustrative mockup, not a captured production run.

Brand

Part of the AI Transformation Director Skills collection: practical, inspectable tools for building reliable AI workflows.

Licence

MIT © 2026 AI Transformation Director

This is an independent community project and is not affiliated with or endorsed by OpenAI.