Skill Extractor

从复杂任务中提取可复用的技能文档，参考 Hermes Agent 的 Skill Documents 设计。用于当任务完成后，识别值得保留的步骤流程，并存储为可搜索的技能文件，下次遇到类似任务时自动检索并建议使用。

Security Scan

The skill's code, runtime instructions, and requested resources are consistent with its stated purpose of extracting, storing, and searching reusable skill documents; it performs local file and SQLite operations but does not request external credentials or network endpoints.

Purpose & Capability

Name/description (extract, store, search, suggest skills) align with the included scripts. The scripts create and use a local skills directory (~/.openclaw/.../skills) and a local SQLite DB to index/search skill markdown files — exactly what you'd expect for this feature. No unrelated credentials, binaries, or external services are requested.

Instruction Scope

SKILL.md and the scripts limit actions to analyzing provided conversation text, generating SKILL.md files, saving them under ~/.openclaw/workspace/skills/skill-extractor/skills, and indexing/searching via a local SQLite FTS5 DB. This is within scope, but important runtime behaviors: the skill will store copies of conversation text (potentially containing sensitive data) on disk and create/modify a local skills.db. The search script uses an FTS5 virtual table and references an ORDER BY rank clause that may be a bug on some sqlite builds (functional issue, not a security indicator).

Install Mechanism

There is no install spec (instruction-only deployment), which is low risk. However, the package includes runnable Python scripts that create directories and files under the user's home directory; running them will write content to disk and create a SQLite DB. No remote downloads, installers, or external code fetches are present.

Credentials

The skill declares no required environment variables, credentials, or config paths beyond creating and using its own directory under the user's home. It does not request unrelated secrets or system-wide configuration access.

Persistence & Privilege

always is false and the skill only persists data under its own ~/.openclaw workspace. It does not modify other skills' configs or system-wide settings. The skill can be invoked autonomously by the agent (disable-model-invocation is false) — this is the platform default and acceptable here, but users should be aware the agent could suggest extraction automatically.

Assessment

This skill appears coherent and localized: it analyzes conversation text, generates SKILL.md files, and stores/indexes them under ~/.openclaw/workspace/skills/skill-extractor/. Before using/installing, consider: (1) privacy — conversations (including any secrets) may be written to disk in plaintext; if you handle sensitive data, do not enable automatic saving or consider encrypting that directory; (2) review/confirm saves — prefer manual confirmation rather than automatic "save" so you control what gets persisted; (3) runtime requirements — the search script uses SQLite FTS5 (your Python's sqlite must have FTS5 enabled) and there are minor implementation issues (e.g., ORDER BY rank) that may cause runtime errors; (4) audit usage — because the agent can invoke the script autonomously, limit or monitor automated suggestions if you don't want the agent to store conversational data without explicit approval. Overall the code is local and does not attempt network exfiltration or request unrelated credentials.

Skill Extractor — 任务技能提取器

参考 Hermes Agent 的 Skill Documents 机制，在 OpenClaw 中实现"从经验中学习"的闭环。

核心功能

1. 提取（Extract） — 从对话历史中识别多步骤流程，生成标准化 SKILL.md
2. 存储（Store） — 保存到 ~/.openclaw/workspace/skills/skill-extractor/skills/
3. 检索（Search） — FTS5 全文检索，在新任务中匹配相关技能
4. 建议（Suggest） — 任务完成后主动提示是否提取技能

工作流程

复杂任务完成
    ↓
触发技能提取检测（多步骤 / 超过N轮 / 成功率高的任务）
    ↓
LLM 回顾对话 → 提取关键步骤
    ↓
生成 SKILL.md 格式的技能文档
    ↓
展示给用户确认是否保存
    ↓
保存 → 下次遇到类似任务自动建议使用

触发条件

当检测到以下情况时，建议提取技能：

• 任务超过 5 轮对话完成
• 包含多个工具调用
• 任务执行成功（有明确输出）
• 涉及罕见问题或特殊解决方案

技能文档格式

遵循 agentskills.io 标准：

---
name: <skill-name>
description: <简短描述：何时使用>
---

# <Skill Name>

## When to Use
- 触发条件描述

## Procedure
1. 步骤1
2. 步骤2
3. 步骤3（含错误处理）

## Notes
- 注意事项或变体

使用方式

手动触发

当用户完成复杂任务后，说"提取技能"或"保存为技能"

自动建议

系统检测到适合提取的任务后，主动提示： "这个任务完成得很好，要把它保存为可复用的技能吗？"

检索技能

当用户描述新任务时，自动检索相关技能：

# 当前任务: <用户描述>
# 匹配技能:
- skill-name: <名称>
  relevance: <相关性>
  trigger: <触发条件>