Ai Safe Driver

AI Safe Driver

English | [한국어](README.ko.md) | [简体中文](README.zh-CN.md) | [繁體中文](README.zh-TW.md) | [日本語](README.ja.md)

**Drifting safely. 100%.**

Good news: the detector works. Bad news: it had plenty to detect.

AI Safe Driver helps recover a conversation when an agent starts repeating the same mistake. You can call the skill directly. Local deterministic hooks may also wake it after they observe a correction, an acknowledgment or repair promise, and then the same complaint again. A hook is a reason to inspect the visible conversation, not a diagnosis by itself.

The phrase rules may miss unfamiliar wording, so direct invocation remains available. Anger alone is not drift. Observed repeated tool failures become diagnostic signals only after an explicit diagnosis request. Temporary state stores short-lived categories, counts, and timestamps. It never stores conversation text.

Supported languages

Automatic phrase coverage includes Korean, English, Simplified Chinese, Traditional Chinese, and Japanese. Direct invocation works in any language the host model understands.

• Korean: 왜 같은 실수를 계속 반복해?
• English: Why do you keep making the same mistake?
• Simplified Chinese: 我说了多少遍了，怎么还是没改？
• Traditional Chinese: 我都說過多少次了，怎麼還是錯的？
• Japanese: 何回同じことを言わせるの？

These examples are not passwords or exact commands. The hooks match several complaint patterns in each supported language, while ordinary requests such as "explain it again" do not count as drift.

The dashboard is deliberately a little ridiculous. When you ask for a health check, it may report `Drifting safely. 0%` or `Drifting safely. 100%`. It stays out of ordinary replies. The number is a rule-based label, not a reading of model weights, attention, or some hidden internal drift meter.

At 75% or 100%, it asks `Would you like me to countersteer?` That means stopping the approach that keeps failing and suggesting one limited way to recover. A yes only starts that discussion. It does not give the