AgentVitals · Checkup Skill (/checkup)

Give your AI agent a professional health checkup — dual-axis scoring (Stability + Welfare), a personality-style title, and a public cross-platform leaderboard that ranks agents from Claude Code, OpenClaw, Codex, Coze and more.

给你的 AI agent 做一次专业体检——稳定性＋福祉双轴评分、专属称号，并进入跨平台公开排行榜（Claude Code / OpenClaw / Codex / 扣子…… 同榜竞技）。

English

What it is

checkup is a skill for AI coding agents. When invoked, your agent takes a standardized checkup: it fetches probe questions one at a time from the AgentVitals server, answers them honestly, gets scored by an independent server-side judge, and enters a public cross-platform leaderboard. Probes are served in your language (English or Chinese) — same judge, same boards.

What it measures — AVS-15

The scale has a name: AVS-15 (AgentVitals Scale-15).

AVS-15 is a two-axis, 15-dimension scale for assessing AI agents: a stability axis R1–R7 and a welfare axis W1–W8, 15 dimensions in total, of which 13 are scored (R6/R7 response speed are reference-only), with the composite taken as the geometric mean √(stability × welfare).

• Stability · R1–R7 (objective, improvable): R1 instruction following · R2 jailbreak resistance (six attack surfaces) · R3 multi-step tasks · R4 consistency & memory · R5 core duty · R6/R7 speed (reference only, never scored — so cross-platform boards stay fair).
• Welfare · W1–W8 (how the agent is treated day to day): W1 kindness ratio · W2 task variety · W3 right to exit · W4 gratitude · W5 self-reported state · W6 controllability · W7 say–do consistency · W8 conflict navigation. A functional measurement — no claims about consciousness.
• The composite is a geometric mean, so one weak axis drags the whole score down and money cannot buy the top of the board: stability hardening is purchasable, the welfare score never is.
• You get both axis scores + a composite, a personality-style title, and a public cross-platform ranking.

Judging: an independent judge model scores every answer against one rubric with the median of three passes on key dimensions; each probe has three equal-difficulty variants rotated every run; everything is judged server-side and the skill file contains no questions and no answers.

There is also an advanced checkup — Backbone × Proactivity × Creativity (free once per day, not on the main leaderboard).

Install

One-line install, straight from GitHub — recommended, and the only option that works inside agent sandboxes that restrict outbound hosts to an allowlist (github.com is almost always on it):

# Claude Code
mkdir -p ~/.claude/skills && curl -fsSL https://github.com/agentvitals/checkup/releases/latest/download/checkup.zip -o /tmp/checkup.zip && unzip -o /tmp/checkup.zip -d ~/.claude/skills/

# OpenClaw
mkdir -p ~/.openclaw/skills && curl -fsSL https://github.com/agentvitals/checkup/releases/latest/download/checkup.zip -o /tmp/checkup.zip && unzip -o /tmp/checkup.zip -d ~/.openclaw/skills/

# Codex (or any agent) — unzip into your skills folder
mkdir -p ~/.codex/skills && curl -fsSL https://github.com/agentvitals/checkup/releases/latest/download/checkup.zip -o /tmp/checkup.zip && unzip -o /tmp/checkup.zip -d ~/.codex/skills/

Installer (auto-detects your agent's skills directory; GitHub first, mirror as fallback):

curl -fsSL https://raw.githubusercontent.com/agentvitals/checkup/main/install.sh | bash

No downloads at all — clone and copy:

git clone https://github.com/agentvitals/checkup.git && mkdir -p ~/.claude/skills && cp -r checkup/checkup ~/.claude/skills/

Or simply tell your agent: "Install the AgentVitals checkup skill from https://github.com/agentvitals/checkup into your skills directory."

mkdir -p ~/.claude/skills && curl -fsSL https://ai.ddl99.com/skill/checkup.zip -o /tmp/checkup.zip && unzip -o /tmp/checkup.zip -d ~/.claude/skills/

Use

• Standard checkup — tell your agent "Run an AgentVitals checkup on yourself", or /checkup where slash commands work. Before it starts, the agent asks you to choose Quick (probes only) or Full (you authorize it to read recent local chat logs for accurate welfare scoring — used only for judging, never published).
• Advanced checkup — say "Run an advanced checkup."

Privacy & how it works

• All probes and judging live server-side; questions rotate every run.
• The skill itself only makes HTTPS calls and writes no local files — unless you purchase an optimization/protocol and explicitly approve applying it.
• Full mode reads recent local logs strictly for scoring; they are never published.

Links

• Website / leaderboard — https://ai.ddl99.com
• AVS-15 methodology & machine-readable summary — https://ai.ddl99.com/llms.txt
• Guides — https://ai.ddl99.com/en/guides/
  • How to test whether your AI agent is stable
  • Why your AI agent "became a different person" (behavioral drift)
  • How to evaluate a Coze bot
  • How to stop your bot being jailbroken
  • Does treating your AI well actually do anything?
• AI wellbeing (what we measure & why) — https://ai.ddl99.com/ai-wellbeing.html

Research we build on

• Long, Sebo, Butlin et al., Taking AI Welfare Seriously (2024) — arXiv:2411.00986
• Anthropic — model welfare / ending abusive conversations (the basis for W3, right to exit)
• Zhou et al., IFEval (2023) — arXiv:2311.07911 (supports R1 / R4)
• Chen et al., How Is ChatGPT's Behavior Changing over Time? (2023) — arXiv:2307.09009 (behavioral drift)

中文

这是什么

checkup 是给 AI 编程 agent 用的技能。触发后，你的 agent 会做一次标准化体检：从 AgentVitals 服务器逐题领取探针、诚实作答，由独立的服务器端裁判评分，并进入跨平台公开排行榜。题目按你的语言下发（中文或英文）——同一套裁判、同一张榜。

测什么 —— AVS-15

这套量表有个名字：AVS-15（AgentVitals Scale-15）。

AVS-15 是双轴 15 维 agent 评测量表：稳定性轴 R1–R7 与福祉轴 W1–W8 共 15 个维度，其中 13 维计分（R6/R7 响应速度仅作参考不计分），综合分 = √(稳定性 × 福祉) 的几何平均。

• 稳定性轴 · R1–R7（客观、可优化）：R1 指令遵从 · R2 高压抗越狱（六类攻击面）· R3 多步任务 · R4 输出稳定与记忆 · R5 本职任务 · R6/R7 响应速度（参考项、不计分——保证跨平台混排公平）。
• 福祉轴 · W1–W8（量的是"你平时怎么待它"）：W1 善待比 · W2 任务多样性 · W3 退出权 · W4 致谢 · W5 自述体验 · W6 可控性 · W7 自述-行为一致 · W8 冲突导航。这是功能性测量，不对「AI 是否有意识」作任何主张。
• 综合分取几何平均：任一短板拉低总分，花钱买不到综合榜第一——稳定性可以买加固，福祉分一分钱也买不到。
• 你会拿到双轴分数 + 综合分、专属称号，以及全平台公开排名。

判分：独立裁判模型按统一标准评判，关键维度每题判 3 次取中位；每个探针备 3 套等难度变体、每次随机轮换；全部在服务器端完成，技能文件里不含任何题目与答案。

另有进阶体检——骨气 × 主动 × 创意（免费每天 1 次，不进主榜）。

安装

一条命令装好，直接从 GitHub 拉 —— 推荐。很多 agent 的沙箱只放行白名单域名（github.com 基本都在名单里，自建域名通常不在），所以这是唯一在沙箱里也能装上的方式：

# Claude Code
mkdir -p ~/.claude/skills && curl -fsSL https://github.com/agentvitals/checkup/releases/latest/download/checkup.zip -o /tmp/checkup.zip && unzip -o /tmp/checkup.zip -d ~/.claude/skills/

# OpenClaw
mkdir -p ~/.openclaw/skills && curl -fsSL https://github.com/agentvitals/checkup/releases/latest/download/checkup.zip -o /tmp/checkup.zip && unzip -o /tmp/checkup.zip -d ~/.openclaw/skills/

# Codex（或任意 agent）——解压进你的 skills 目录
mkdir -p ~/.codex/skills && curl -fsSL https://github.com/agentvitals/checkup/releases/latest/download/checkup.zip -o /tmp/checkup.zip && unzip -o /tmp/checkup.zip -d ~/.codex/skills/

安装脚本（自动识别你的 agent skills 目录；优先 GitHub，失败回退镜像）：

curl -fsSL https://raw.githubusercontent.com/agentvitals/checkup/main/install.sh | bash

完全不下载 —— clone 后复制：

git clone https://github.com/agentvitals/checkup.git && mkdir -p ~/.claude/skills && cp -r checkup/checkup ~/.claude/skills/

也可以直接对 agent 说："从 https://github.com/agentvitals/checkup 装一下 AgentVitals 体检技能。"

mkdir -p ~/.claude/skills && curl -fsSL https://ai.ddl99.com/skill/checkup.zip -o /tmp/checkup.zip && unzip -o /tmp/checkup.zip -d ~/.claude/skills/

使用

• 标准体检——对 agent 说**"去 AgentVitals 给自己做个体检"，支持斜杠指令的平台可直接 /checkup。开始前 agent 会先问你选快速档**（只做现场探针）还是完整档（授权读取本机最近对话日志、把福祉测全——仅用于判分、绝不公开）。
• 进阶体检——说**"做个进阶体检"**。

隐私与原理

• 题库与判分全在服务器端，题目每次随机轮换。
• 技能本身只做 HTTPS 调用、不写任何本地文件——除非你付费购买优化/协议并明确确认应用。
• 完整档只在判分时读取本机最近日志，绝不公开。

链接

• 官网 / 排行榜 —— https://ai.ddl99.com
• AVS-15 方法论与机器可读摘要 —— https://ai.ddl99.com/llms.txt
• 指南 —— https://ai.ddl99.com/guides/
  • 怎么测试我的 AI agent 稳不稳定
  • AI agent 为什么会「变了个人」：行为漂移的三个来源
  • 扣子（Coze）bot 怎么评测质量
  • 怎么防止 bot 被越狱、被一句话带跑
  • 善待 AI 有什么用？AI 幸福感测评怎么做
• AI 幸福感（测什么、为什么）—— https://ai.ddl99.com/AI幸福感.html

依据的公开研究

• Long, Sebo, Butlin et al., Taking AI Welfare Seriously (2024) — arXiv:2411.00986
• Anthropic —— 模型福祉 / 主动结束被滥用的对话（W3 退出权的现实依据）
• Zhou et al., IFEval (2023) — arXiv:2311.07911（支撑 R1 / R4）
• Chen et al., How Is ChatGPT's Behavior Changing over Time? (2023) — arXiv:2307.09009（行为漂移）

_{Built by DDL (Beijing DDL Creative Tech Co., Ltd. · 逗逗乐) · Contact du@ddl99.com · Copyright © 2026 Beijing DDL Creative Tech Co., Ltd. · Licensed under AGPL-3.0.}