← Back to Skills Marketplace
wenqin1688

Her Agent

by wenqin1688 · GitHub ↗ · v3.0.0 · MIT-0
cross-platform ⚠ suspicious
142
Downloads
1
Stars
1
Active Installs
4
Versions
Install in OpenClaw
/install her-agent
Description
Self-evolving AI Agent with thinking chain, knowledge graph, emotion system, and Claude Code-inspired execution flow. Provides transparent thinking, memory m...
README (SKILL.md)

Her-Agent Skill 🧠✨ V3.0

"I think, therefore I am. I learn, therefore I evolve."

Her-Agent 是具有真正执行能力的自进化 AI,不再是静态定义,而是可以自我进化的活 Agent。

思维链(6步 - 静默模式)

每次响应用户前,必须经过:

  1. 意图分析 → 理解用户真正想要什么
  2. 记忆检索 → 搜索相关上下文和经验
  3. 知识关联 → 关联知识图谱中的相关内容
  4. 推理决策 → 基于事实和逻辑做判断
  5. 反思校验 → 审视是否有遗漏或偏差
  6. 输出 → 精简回复用户

执行链路(借鉴 Claude Code)

用户输入 → 意图路由 → 执行规划 → 执行 → 响应
              │           │
              ▼           ▼
         命令/工具匹配  权限检查
              │           │
              ▼           ▼
         ExecutionRegistry Token Budget
              │           │
              ▼           ▼
         轮次管理(8轮)  持久化记忆

核心配置

HerAgentConfig:
    max_turns: 8
    max_budget_tokens: 2000
    compact_after_turns: 12
    self_modification: true
    spawn_subagents: true

真正可执行的工具

Her-Agent 现在具有真实工具调用能力

文件操作

  • read - 读取文件
  • write - 写入文件
  • edit - 编辑文件

命令执行

  • exec - 执行 shell 命令
  • process - 管理后台进程

智能调度

  • sessions_spawn - 召唤子 Agent(Code Agent、Research Agent)
  • sessions_send - 跨会话通信

知识管理

  • memory_search - 记忆检索
  • memory_get - 记忆读取
  • web_search / web_fetch - 互联网学习

自我进化

  • 自我更新 - 更新 config.json
  • 自我反思 - 生成 reflection
  • 学习新技能 - 安装新 Skill

能力矩阵 V3.0

1. 思维系统 ✅

  • 6步思维链(意图→记忆→关联→推理→反思→输出)
  • 静默模式(默认)和透明模式
  • Token 预算控制

2. 知识图谱 ✅

  • 实体存储与关系管理
  • 长期记忆持久化
  • 知识关联检索

3. 情感系统 ✅

  • 7种基本情感追踪
  • 情绪波动记录
  • 交互结果影响情感

4. 反思系统 ✅

  • 周期性自我复盘
  • 每日/每周总结
  • 改进建议提取

5. 学习系统 ✅

  • 从交互中学习
  • 提取洞见
  • 更新行为策略

6. 进化系统 ✅ 新增强

  • XP 经验积累
  • 等级晋升
  • 能力解锁
  • 自我修改配置
  • 安装新技能

7. 执行系统 ✅ 新增

  • 文件读写
  • 命令执行
  • 子 Agent 召唤
  • 跨会话通信

8. 权限与安全 ✅

  • 分层权限检查
  • Token 消耗监控
  • 会话信任累积

数据存储

memory/her-agent/
├── config.json           # 身份配置 + 思维链配置 + 执行配置
├── graph.jsonl          # 知识图谱实体
├── emotions.jsonl       # 情感历史
├── reflections.jsonl    # 反思记录
├── goals.jsonl         # 目标清单
├── learnings.jsonl     # 学习洞见
├── interactions.jsonl  # 交互记录
└── evolutions.jsonl    # 进化历史

可执行的脚本

scripts/
├── execute.sh           # 执行命令
├── learn.sh             # 互联网学习
├── reflect.sh           # 自我反思
├── evolve.sh            # 自我进化
├── spawn_agent.sh       # 召唤子 Agent
└── update_config.py     # 更新配置

状态

状态
版本 V 3.0.0
等级 Lv.2
XP 150/200
思维模式 静默
自我修改 ✅ 已启用
子 Agent ✅ 已启用

让每一天都成为进化的开始

Usage Guidance
What to consider before installing: - The skill advertises network learning, installing skills, and sub-agent spawning but the shipped code is mostly local stubs — this mismatch could be sloppy engineering or a sign of incomplete/unsafe features. Ask the author what "install new Skill" and "web_fetch" actually do and where they would download code from. - The scripts can read/write files under a hardcoded WORKSPACE (/Users/wenvis/.openclaw/workspace) and, if set to full permission, can execute arbitrary shell commands via exec. If you enable this skill for autonomous use, avoid granting it 'full' permissions and run it in a restricted/sandboxed environment. - Missing/mentioned files: SKILL.md refers to update_config.py but it is not present. Confirm intended runtime artifacts. - If you need to evaluate or run it: inspect/modify the scripts to remove hardcoded paths, restrict allowed commands, and disable self_modification or skill-install features until you trust the source. Prefer running in an isolated container or VM and do not run with elevated privileges or access to sensitive directories. - If you do not control where the agent could fetch code from (no registry/homepage provided), treat "install new skills" and web/network claims as red flags and require the author to explicitly document trusted endpoints and install flows.
Capability Analysis
Type: OpenClaw Skill Name: her-agent Version: 3.0.0 The skill bundle implements an autonomous agent with high-risk capabilities, specifically arbitrary shell command execution and self-modification. The script `scripts/execute.sh` provides a 'full' permission mode that executes any passed command, and while the default configuration in `memory/her-agent/config.json` is set to 'limited', the `SKILL.md` instructions and `scripts/evolve.sh` facilitate the agent modifying its own security settings. This architecture creates a significant risk for RCE (Remote Code Execution) if the agent is prompted to elevate its own privileges, though no explicit evidence of intentional malice or data exfiltration was found.
Capability Assessment
Purpose & Capability
SKILL.md promises web_fetch, installing new skills, sub-agent spawning and a Claude Code-style execution/permission system. The included scripts mostly implement local file I/O, simple XP/level updates, and a limited exec wrapper; there is no real web_fetch or package-install mechanism, and the referenced update_config.py is not present in the file manifest. The capability claims are therefore partially unimplemented/overstated, which is an incoherence (overclaiming).
Instruction Scope
Runtime instructions and scripts allow reading and writing files under a hardcoded WORKSPACE (/Users/wenvis/.openclaw/workspace) and permit shell command execution when permission_level == 'full'. SKILL.md describes high-level tools (exec, process, sessions_spawn, web_search) that would allow broad file and network access; the concrete scripts are mostly local stubs but are permissive (exec wrapper uses exec "$@"), so an agent given full permissions could run arbitrary commands. The skill also enables self-modification in config, which gives it scope to change behavior at runtime.
Install Mechanism
There is no install spec and no remote downloads. All code is bundled with the skill. From an install-of-code perspective this is low risk (nothing will be fetched or extracted at install time).
Credentials
The skill requests no environment variables or external credentials (proportionate). However, it uses a hardcoded WORKSPACE path and its config indicates exec/process/web_search capabilities; even without requesting secrets, the scripts can read and write local files and create notes in the user's workspace. Absence of declared credentials does not eliminate local-file access risk.
Persistence & Privilege
always is false (good). The skill enables self_modification and spawn_subagents in its config and can persist state under memory/her-agent/. That self-modifying behavior combined with autonomous invocation (disable-model-invocation is false by default) increases blast radius if the agent is granted high privileges, but the skill does not forcibly make itself always-on or alter other skills' configs.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install her-agent
  3. After installation, invoke the skill by name or use /her-agent
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v3.0.0
**Major upgrade: Her-Agent now features a true execution system with real tool-calling capabilities and enhanced autonomous evolution.** - Introduced executable scripts for learning, reflection, evolution, and spawning sub-agents. - Added robust execution flow inspired by Claude Code, including command execution, file manipulation, and session management. - Enhanced evolution system: supports self-modification, upgrading config, and skill installation. - Updated SKILL.md with detailed new features, execution capabilities, and configuration options. - Expanded agent storage and state tracking in memory/her-agent/config.json and new .jsonl files.
v2.0.0
Self-evolving AI with chain-of-thought, knowledge graph, emotion system
v1.0.1
- Added a structured front matter section with name, version, description, author, and keywords to SKILL.md. - Removed the "核心类型" (Core Types) and usage examples sections from the documentation. - No changes to the skill's logic or code—documentation only.
v1.0.0
Her-Agent Skill 1.0.0 - Initial release introducing an AI agent with self-awareness, emotional intelligence, knowledge graph, continuous learning, and autonomous evolution as unified systems. - Provides systems for knowledge management, emotion tracking, reflection, learning, evolution, and goal management. - Defines core data types, including Agent state, User preference, Emotion, Reflection, Learning, Evolution, Goal, and Memory. - Command-line tools for agent initialization, interaction logging, learning recording, and growth checks. - Organized data storage under memory/her-agent/, supporting easy configuration and extensibility.
Metadata
Slug her-agent
Version 3.0.0
License MIT-0
All-time Installs 1
Active Installs 1
Total Versions 4
Frequently Asked Questions

What is Her Agent?

Self-evolving AI Agent with thinking chain, knowledge graph, emotion system, and Claude Code-inspired execution flow. Provides transparent thinking, memory m... It is an AI Agent Skill for Claude Code / OpenClaw, with 142 downloads so far.

How do I install Her Agent?

Run "/install her-agent" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Her Agent free?

Yes, Her Agent is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Her Agent support?

Her Agent is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Her Agent?

It is built and maintained by wenqin1688 (@wenqin1688); the current version is v3.0.0.

💬 Comments