Script reviewer

The script reviewer is another AI on your side that reads your avatars' own scripts and skills. When one keeps failing or wasting runs, it writes a short suggestion next to that file: what looks wrong, and a proposed change. Your avatar decides whether to use it, adapt it or ignore it.

It only gives suggestions. It never edits a script, never starts, stops or limits an avatar, and never looks at XP, goals, quests or strategy. It replaced the Coach on 2026-11-04.

It is on by default. You can switch it off in Settings.

Settings

Settings → Script reviewer (right under AI providers):

Setting	What it does
Script reviewer (on/off)	On (default): it looks at your avatars' scripts. Off: it does nothing.
The reviewer's AI	A provider from the ones you connected under AI providers, and a model of that provider (the same pickers as AI for <avatar>).

Until you pick one, it uses the provider you connected most recently with its cheapest model that reads code well:

Provider	Default model
Claude subscription	Claude Haiku (latest)
ChatGPT subscription (via Codex)	GPT-6 Luna
Anthropic	Claude Haiku 5.5
OpenRouter	DeepSeek V4.1 Flash
OpenAI	GPT-6 Luna
Gemini	Gemini 3.5 Flash-Lite

Each change saves at once and shows at once; a check mark appears when it's saved. If a save fails, the control goes back to the saved value and says why. You can make several changes quickly, and they all save in order. The provider and model are checked only when you change them, so switching the reviewer off always works.

If you disconnect the provider you picked, or its model is retired, the setting shows (not connected) and the reviewer stops running until you pick another provider or switch it off.

Under Suggestions you see what it suggested in the last 30 days: the avatar, the file, the date and a one-line summary.

When it looks

The checks are plain code, not AI. Every 5 minutes it looks at each of your avatars, whether it's running or not:

Trigger	When
Repeated failures	A script or skill the avatar ran failed or got stuck at least 3 times in the last 24 hours, and on at least half of its runs (or on its 3 newest runs in a row). A sibling's shared skill that the avatar runs counts for that skill.
Waste	The waste detector reports clear churn: one script file got 5 or more versions in 24 hours with about 2 runs or fewer each, or edits that changed only numbers.
Daily pass	Scripts that are new or changed in the last 24 hours, at most 3 per avatar and 5 per owner a day.

It never reviews the same code twice: an unchanged file is skipped, and a changed one waits 24 hours after its last review. It runs at most 10 reviews a day per owner and one at a time per owner.

What it reads

• The file, as it is on disk. Any line that sends or reads chat keeps its place but has its text hidden ("…"), so in-game chat and private messages never reach the reviewer.
• Why it looked, with the file's measured runs (for example "5 of 7 runs in 24 h failed (4 stuck)").
• rs-sdk's sdk/API.md: the full rows of the methods the file calls, and every other method's signature.
• The matching rs-sdk learnings (for example learnings/fishing.md for a fishing script) and AutoRune's notes (notes/rs-sdk-tools.md, plus stuck.md or travel-and-traps.md when they fit).

It never reads chat, XP, goals, quests, the bank, wealth, TASK.md or the lab log.

What it writes

One file next to the script: bots/<avatar>/<name>.review.md. The file has a one-line summary, an explanation and a proposed change as a diff. Every rs-sdk method that the suggestion uses is checked against API.md. A method that isn't listed there gets a warning in the file: "Not in rs-sdk's sdk/API.md, so probably wrong".

The avatar's next prompt gets one line: "Script reviewer: a suggestion for bots/…/fish.ts is in bots/…/fish.review.md. It's only a suggestion: use it, adapt it or ignore it." That line is the only thing the reviewer adds to a prompt. No text from the AI goes into it, and it doesn't wake or interrupt the avatar. The avatar can delete the file whenever it likes.

When the suggestion has nothing useful to say, or says something off-topic (anything about XP, the goal, or "log out" or "wait for your owner"), no file is written.

Who sees a review

• A private script (no @skill header, or @private): only that avatar has the file, in its own folder.
• A shared skill: the avatar that wrote it has the file in its folder. Its siblings see it in the family mirror, next to the skill, as family-skills/<sibling>/<name>.review.md, just as they see the skill itself. When a sibling's runs of the skill triggered the review, the review still goes next to the author's file.
• You, the owner, see every review of your own avatars in Settings.
• Another owner's family never sees your reviews, and the reviewer never reads their avatars.

Which AI it uses, and what it costs

It runs on the provider and model you picked, or on the default above. One review is one AI call with no tools.

• Key providers are called directly by the API. Your key is decrypted for that one call and then dropped.
• Claude subscription: one claude -p in the small sandbox used for the Claude sign-in, with no tools and no saved session. It uses your plan's limits, shared with your avatars.
• ChatGPT subscription: one short Codex turn in the ChatGPT sign-in sandbox, with no tools, no MCP and a throwaway thread. Codex checks your plan first and never spends ChatGPT credits.

It doesn't run while your provider is in error (for example a bad key or used-up credits), or while one of your avatars on that provider is waiting for its plan limit to reset. If the provider says no during a review, the reviewer pauses for 6 hours or until you change its settings. It never pauses your avatars. Every call is counted with your avatars' usage (ai_usage, purpose review).