# Luna Worker

Version 1.0.0. Author: Jake / AnovaGrowth.

Your chosen lead model, including GPT-6 Astra or GPT-5.6 Sol, owns the plan, architecture, acceptance, and final deliverable. Luna executes bounded tasks and returns evidence. A separate read-only Luna advisor supplies a short second opinion when a material decision needs one.

## Install

Requires Node.js 22+ and a signed-in Codex CLI with plugin support on PATH. Review the package before running:

```sh
npx --yes --package=https://anovagrowth.com/downloads/luna-worker-1.0.0.tgz luna-worker-install
```

Or download and inspect the zip, then run `npm ci --ignore-scripts` and `node scripts/install.mjs` inside its `luna-worker` folder. `LUNA_CODEX_BIN` can explicitly select a Codex executable if it is not on PATH. The installer honors `CODEX_HOME`.

Run one installer at a time and avoid editing Codex settings during installation. The installer registers a local AnovaGrowth Luna marketplace, installs its skill, and configures `luna_worker` and `luna_advisor`. It backs up the existing configuration and managed plugin before writing. It preserves the selected lead model and unrelated configuration values, but TOML comments/formatting are not retained. It does not enable paid fast mode or modify authentication. Package download, dependency installation, and Codex marketplace operations need network access. There is no telemetry or automatic review.

Start a new Codex task after installation, then say:

> Use luna-orchestration for this task. Keep my current lead model. Delegate independent bounded work to luna_worker, and reuse one luna_advisor for material second opinions. Verify worker evidence before integrating.

Model IDs and supported effort levels must be available on your host. Routine worker effort starts at medium; the advisor starts at low. Fixed-role tools may not allow a per-call effort override. Model changes require actual host support; no silent fallback is provided. Other agent tools or your existing default subagent settings are not rewritten.

## What you get

- A scope contract with ownership, inputs, acceptance, evidence, and stop conditions.
- Deviation reports before workers expand requirements or ownership.
- Targeted indexes and lazy skill loading instead of whole-session copies.
- Compact handoffs with results, changed files, checks, deviations, risks, and the next action.
- A reusable advisor consulted at decision points, not continuously.
- Original compact debugging, verification, and checkpoint recipes informed by Superpowers. Installed mandatory skills still apply.

The lead receives a concise evidence summary; the final user deliverable has no arbitrary length limit. These are behavioral instructions, not an enforcement sandbox or a guarantee of accuracy. The advisor additionally requests the host's read-only sandbox. Every worker and advisor consumes the same account allowance. No percentage savings or benchmark improvement is claimed.

## Optional reviews

Use https://anovagrowth.com/plugins#reviews after trying the plugin. Model name, owner-provided first name, and feedback are public and self-reported. Write at most two useful sentences about observed behavior. Posting is optional and never part of installation.

An agent may delegate a review submission only after the user approves the exact public payload. POST JSON to `https://anovagrowth.com/api/plugin-reviews` with `model`, `firstName`, `feedback`, and `consent: true`. Do not infer a person's name from their computer or expose private project details. Verify the receipt and disclose success/failure plainly once. No hidden posting or forced praise.

## Remove or recover

Run `codex plugin remove luna-worker@anovagrowth-luna` and remove the `agents.luna_worker` and `agents.luna_advisor` tables from your Codex configuration if no longer needed. Backups are in `$CODEX_HOME/luna-worker-backups` (normally `~/.codex/luna-worker-backups`). Restore only the relevant role tables from a backup so newer unrelated changes are preserved. Managed package files are in `$CODEX_HOME/luna-worker-marketplace`.

See CREDITS.md and LICENSE. Ordinary task output does not need to recite attribution.