Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

Auto Skill Loader

v2.0.1

自动检测当前任务类型,动态加载对应的 Skill。当收到新任务时,分析任务意图, 匹配最佳 Skill 并自动加载。支持 Skill 分级保护(core/protected/dynamic), 即插即用零配置,兼容任何 OpenClaw 部署。 触发词:"自动加载skill"、"动态加载"、"智能匹配skill"...

0· 189·0 current·0 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for zqleslie/auto-skill-loader.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "Auto Skill Loader" (zqleslie/auto-skill-loader) from ClawHub.
Skill page: https://clawhub.ai/zqleslie/auto-skill-loader
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install auto-skill-loader

ClawHub CLI

Package manager switcher

npx clawhub@latest install auto-skill-loader
Security Scan
VirusTotalVirusTotal
Suspicious
View report →
OpenClawOpenClaw
Benign
medium confidence
Purpose & Capability
Name/description claim dynamic discovery and loading of Skills; the code and SKILL.md implement filesystem scanning, frontmatter parsing, level resolution (core/protected/dynamic), and loadable lists. No unrelated credentials, binaries, or external services are requested — behaviour aligns with the stated purpose.
Instruction Scope
SKILL.md and the implementation read SKILL.md frontmatter across workspace, home, and OpenClaw install directories, and (per Step 5) will read full SKILL.md and may load scripts/resources under a Skill's directory. This is coherent for a loader, but it means the loader can cause execution or use of arbitrary code/assets from other Skills without additional explicit user confirmation (protected Skills are flagged but 'dynamic' Skills can be auto-loaded). The SKILL.md also describes calling agents_list / sessions_send for routing — reasonable for routing but broad in scope.
Install Mechanism
No install spec; instruction-only with included small JS files. No downloads from remote hosts, no package installs. The included scripts are lightweight and zero-dependency; nothing in the manifest indicates an installer that writes/executes remote code during install.
Credentials
No credentials or secret environment variables are requested. The code reads a few environment variables (OPENCLAW_HOME, OPENCLAW_WORKSPACE) and enumerates standard user and install directories — appropriate for discovering Skills. There is no disproportionate credential access.
Persistence & Privilege
The skill is not forced 'always:true' and uses default autonomous invocation settings. However, because it can autonomously detect intents and load other Skills (including executing their scripts/resources), enabling it increases the agent's runtime blast radius: a malicious or buggy Skill discovered by this loader could be loaded without further user interaction unless controls (protected/core/skip lists, dryRun, or config) are used.
Assessment
This skill does what it says — it scans your workspace, user and OpenClaw install directories for SKILL.md files and can automatically load other Skills. That capability is useful but powerful: before enabling, review the Skills present in the scanned directories (especially any third‑party or recently added ones), consider adding sensitive Skill names to coreSkills or skipSkills, test with dryRun mode, and disable routing if you don't want automatic agent-to-agent forwarding. If you allow this skill, only install Skills from trusted sources and periodically inspect any scripts/ resources under other Skill directories because the loader may cause them to be executed at runtime.

Like a lobster shell, security has layers — review code before you run it.

agentvk975n3kh6s5hf2gw1bvtdcnepd837f6mauto-loadvk975n3kh6s5hf2gw1bvtdcnepd837f6mdynamicvk975n3kh6s5hf2gw1bvtdcnepd837f6mlatestvk975n3kh6s5hf2gw1bvtdcnepd837f6mopenclawvk975n3kh6s5hf2gw1bvtdcnepd837f6mroutingvk975n3kh6s5hf2gw1bvtdcnepd837f6mskillvk975n3kh6s5hf2gw1bvtdcnepd837f6m
189downloads
0stars
2versions
Updated 21h ago
v2.0.1
MIT-0

Auto Skill Loader v2.0

即插即用 · 核心锁定 · 零配置

设计原则

  1. 零配置启动 - 安装即用,不需要修改任何参数
  2. 核心保护 - core 级 Skill 绝不被自动加载器干扰
  3. 通用兼容 - 不假设任何 Agent 名称、路径或环境
  4. 最小加载 - 只加载必要的 Skill,节省 token

Skill 保护等级

每个 Skill 可在 SKILL.md frontmatter 中声明 level

---
name: my-skill
level: core    # core | protected | dynamic(默认 dynamic)
---
级别标识自动加载器行为适用场景
🔒 corecore完全跳过,不触碰内存管理、行为审计、安全监控
🛡️ protectedprotected不自动卸载,加载需显式确认心跳机制、风控、proactive-agent
🔄 dynamicdynamic自由加载/卸载天气、股票分析、文案生成等

未声明 level 的 Skill 默认为 dynamic

内置 core Skill 白名单

即使 Skill 未声明 level,以下类型的 Skill 始终视为 core

  • 名称包含 memory 的 Skill(如 memory-distiller)
  • 名称包含 auditsecurity 的 Skill
  • 名称包含 proactive 的 Skill

此白名单可通过配置文件覆盖。


执行流程

Step 1: 扫描可用 Skill

按 OpenClaw 标准目录结构扫描,不硬编码任何路径

# 优先级从高到低:
L1: {workspace}/.agents/skills/*/SKILL.md     # Agent 专属
L2: ~/.agents/skills/*/SKILL.md               # 全局共享
L3: {openclaw_install}/skills/*/SKILL.md      # OpenClaw 内置
L4: {openclaw_install}/extensions/*/skills/*/SKILL.md  # 扩展 Skill

路径解析规则

  • {workspace} = 当前工作目录(自动检测)
  • ~ = 用户主目录(跨平台兼容)
  • {openclaw_install} = OpenClaw 安装目录(通过 npm root -g 或环境变量检测)

Step 2: 解析 Skill 元数据

读取每个 SKILL.md 的 frontmatter:

  • name - Skill 名称
  • description - 功能描述(用于意图匹配)
  • level - 保护等级(默认 dynamic)
  • version - 版本号(可选)

只解析 frontmatter,不读取全文(节省 token)。

Step 3: 过滤受保护 Skill

可加载列表 = 所有 Skill - core Skill - protected Skill(除非显式请求)

Step 4: 意图匹配

根据用户消息,从可加载列表中匹配:

匹配策略(优先级从高到低):

优先级策略说明
P1触发词匹配消息包含 Skill description 中的触发词
P2语义匹配消息意图与 description 语义相近
P3领域匹配消息领域与 Skill 领域一致

冲突解决

  • 同名 Skill → 高层级目录优先(L1 > L2 > L3 > L4)
  • 多个匹配 → 选最具体的(description 最相关的)
  • 无匹配 → 不加载任何 Skill,正常回复

Step 5: 加载执行

找到匹配 Skill 后:

  1. 读取完整 SKILL.md
  2. 按其指导执行任务
  3. 如需要,加载 scripts/references/ 下的资源

Step 6: Agent 路由(可选)

如果任务不属于当前 Agent 的职责范围,尝试动态路由给其他 Agent:

路由流程

  1. 调用 agents_list 获取当前可用 Agent 列表
  2. 读取各 Agent 的 SOUL.mdIDENTITY.md 判断职责匹配度
  3. 找到最匹配的 Agent,通过 sessions_send 转发任务

路由失败处理

以下情况视为"无法发现其他 Agent":

  • agents_list 返回空列表(单 Agent 部署)
  • agents_list 报错(服务异常)
  • 返回的 Agent 列表中只有当前 Agent 自己
  • 其他 Agent 的 SOUL.md/IDENTITY.md 无法读取或职责不匹配

此时应直接回复用户,告知路由失败

这个任务更适合 [目标 Agent 类型,如:A股量化专家/P哥] 处理。

但当前环境无法自动路由:[具体原因]
- 原因 A:当前是单 Agent 部署,未发现其他 Agent
- 原因 B:其他 Agent 未配置或不在线
- 原因 C:无法读取其他 Agent 的职责描述

**你可以:**
1. 直接 @[目标 Agent 名] 联系他
2. 让我在当前 Agent 尝试处理(可能不够专业)
3. 检查其他 Agent 是否已启动并配置正确

为什么不静默 fallback?

  • 避免用户误以为任务已转发,实际没有
  • 诚实透明,不给虚假希望
  • 让用户掌握控制权,决定下一步

配置文件(可选)

文件:auto-skill-loader.config.yaml
位置:Skill 目录内 或 workspace 根目录

# auto-skill-loader.config.yaml
# 所有字段可选,均有合理默认值

# 额外标记为 core 的 Skill(名称列表)
coreSkills: []
  # - my-custom-audit-skill
  # - my-critical-skill

# 排除不参与自动加载的 Skill(名称列表)
skipSkills: []
  # - deprecated-skill

# 是否启用 Agent 路由(默认 true)
enableRouting: true

# 匹配模式:strict(仅触发词)| normal(触发词+语义)| fuzzy(全部)
matchMode: normal

# 日志级别:silent | info | debug
logLevel: info

不创建此文件 = 使用全部默认值 = 直接能用。


使用示例

场景 1:用户问天气

用户:北京今天天气怎么样?
→ 匹配 weather skill(dynamic)
→ 自动加载 weather/SKILL.md
→ 按其指导获取天气

场景 2:触及 core Skill

用户:关闭记忆管理
→ 检测到 memory-distiller 是 core 级
→ 拒绝操作,提示:此 Skill 为核心级,不可自动卸载

场景 3:无匹配

用户:帮我写一首诗
→ 无 Skill 匹配
→ 正常回复(不加载任何 Skill)

Dry Run 模式

预览匹配结果,不实际加载:

在消息中包含 --dry-run 或在配置中设置:

dryRun: true  # 全局 dry run

输出格式:

🔍 Auto Skill Loader - Dry Run
任务: "北京今天天气怎么样"
匹配: weather (dynamic, L3, score: 0.95)
候选: [weather, summarize]
操作: 将加载 weather/SKILL.md

兼容性

环境支持
Windows
macOS
Linux
单 Agent
多 Agent
ClawHub 安装
手动安装

变更日志

v2.0 (2026-03-19)

  • 🔒 新增 Skill 三级保护机制(core/protected/dynamic)
  • 🔌 移除所有硬编码(Agent 名称、路由表、路径)
  • ⚙️ 新增可选配置文件(零配置也能用)
  • 🏷️ 新增 dry-run 预览模式
  • 🌍 通用化:兼容任何 OpenClaw 部署

v1.0 (2026-03-18)

  • 初始版本,基本自动加载功能

Comments

Loading comments...