Openclaw Soul Publish

OpenClaw 自我进化框架一键部署。安装宪法(AGENTS.md)、可进化灵魂(SOUL.md)、心跳系统、PARA三层记忆架构、目标管理，并通过场景化对话引导用户定义 Agent 性格。自动配置 EvoClaw（审批制进化）和 Self-Improving Agent（自主学习）。触发场景："setup o...

Security Scan

The skill mostly does what it says (deploys an agent self‑evolution framework), but its runtime instructions and included files give an agent broad rights (write to your home, copy/execute local scripts, and persist API keys into shell profiles) and there are inconsistent governance defaults — review before installing.

Purpose & Capability

The skill's stated purpose (deploy a self‑evolution framework and optional integrations) matches the actions in SKILL.md: creating workspace files, configuring EvoClaw and Self‑Improving components, and optionally integrating social feeds. However, some capabilities are more powerful than a simple 'template deployer' — it includes full fallback skill bundles, validators, and CLI-style config flows that ask the agent to manage external API keys and heartbeats. Those behaviors are plausible for the stated goal but are heavier than a minimal deploy tool.

Instruction Scope

Instructions direct the agent to read and write many user files (workspace templates, evoclaw-state.json, memory directories, and files under ~/self-improving/), copy fallback skill directories into the user's skills folder, and run bundled Python scripts (scripts/preflight_check.py and validators). The fallback configure flow explicitly tells the agent to accept raw API keys and automatically write EXPORT lines into the user's shell profile — this is sensitive and surprising behavior for an automated skill and expands the agent's runtime authority beyond just copying templates.

Install Mechanism

No external download/install spec is present (low risk from network fetches). However, the package includes local executable artifacts (preflight_check.py, validators, tools) which the instructions tell the agent to run. The fallback approach copies local directories into the user's workspace — copying bundled code that may then be executed is riskier than pure instruction-only skills.

Credentials

Registry metadata lists no required env vars, but the runtime docs instruct the agent to request and persist external service API keys (MOLTBOOK_API_KEY, X_BEARER_TOKEN) and to write them into shell profile files automatically. Requesting API keys for optional external feeds is understandable, but automatically storing raw keys into ~/.bashrc/.zshrc and exporting them without explicit, informed consent is disproportionate and risky. The skill also reads/writes ~/.openclaw/openclaw.json and other user files — expected for deployment but worth highlighting.

Persistence & Privilege

The skill is not force‑installed (always:false) and is user-invocable, which is normal. However, the framework it installs (EvoClaw/Self‑Improving) is explicitly designed for ongoing autonomous evolution and heartbeat-driven updates; some docs and templates reference 'autonomous' defaults while other deployment steps write governance as 'advisory' — these inconsistencies affect how much autonomy the agent will get post‑install. Because autonomous model invocation is possible, you should verify governance settings before enabling external feeds or allowing automated proposal application.

What to consider before installing

Before installing, do these checks: 1) Inspect scripts/preflight_check.py and the validators yourself; run them manually rather than letting the agent execute them automatically. 2) Back up any existing ~/.openclaw/workspace/ files (AGENTS.md, SOUL.md, etc.). 3) Do NOT paste API keys into the chat — if you want external feeds, set the environment variables yourself in a controlled way rather than letting the skill write to your shell profile. 4) Decide governance explicitly: set EvoClaw to supervised/advisory if you want human approval for identity changes (the docs are inconsistent about defaults). 5) Consider testing this in a throwaway workspace first (copy the package to a temp workspace) to observe behavior. 6) If you are uncomfortable with an agent having autonomous identity‑evolution rights, do not install or keep EvoClaw in 'autonomous' mode.

openclaw-soul — Self-Evolution Framework

为 OpenClaw 安装完整的自我进化框架：宪法 + 可进化灵魂 + 结构化心跳协议 + PARA 三层记忆 + 目标管理 + 治理配置。

安装内容：

• 9 个工作区文件（AGENTS.md, SOUL.md, HEARTBEAT.md, BOOTSTRAP.md, USER.md, IDENTITY.md, GOALS.md, working-memory.md, long-term-memory.md）
• 2 个依赖 skill（evoclaw, self-improving）
• Heartbeat 定时任务配置
• EvoClaw 治理配置（advisory 模式 + soul-revisions 回滚）
• PARA 三层记忆目录结构
• 引导对话启动（BOOTSTRAP.md）

---

§1 [GATE] 确认安装环境

在执行任何操作之前，必须先完成此步骤。跳过此步骤是禁止的。

询问用户：

"准备部署 openclaw-soul 自我进化框架。请确认：

1. 工作区路径是什么？（默认：~/.openclaw/workspace/）
2. 如有自定义路径请告知。"

将用户确认的路径记为 $WORKSPACE。如果用户说"默认"或不指定，使用 ~/.openclaw/workspace/。

---

§2 [GATE] 环境检查

必须先通过环境检查，否则禁止继续后续步骤。

运行预检脚本：

python3 "$(dirname "$0")/../scripts/preflight_check.py"

如果脚本不可用，手动执行以下检查：

1. 工作区目录存在：$WORKSPACE 路径必须存在且可写
2. openclaw.json 可读：检查 $WORKSPACE/../openclaw.json 是否存在且为有效 JSON
3. clawhub CLI 可用：which clawhub 或检查 ~/.openclaw/bin/clawhub
4. 列出已有文件：检查以下 9 个文件是否已存在于 $WORKSPACE：
   • AGENTS.md, SOUL.md, HEARTBEAT.md, BOOTSTRAP.md, GOALS.md
   • USER.md, IDENTITY.md, working-memory.md, long-term-memory.md

输出格式：逐项报告 pass/warn/fail，标记需要备份的文件。

判断逻辑：

• 任何 fail → 报告问题并停止，告知用户如何修复
• 只有 warn → 告知用户限制，确认是否继续
• 全部 pass → 继续下一步

---

§3 [REQUIRED] 备份已有文件

对每个即将部署的文件，如果已存在于 $WORKSPACE：

cp "$WORKSPACE/{filename}" "$WORKSPACE/{filename}.backup.$(date +%Y%m%d-%H%M%S)"

列出所有被备份的文件告知用户。如果没有需要备份的文件，跳过此步骤并告知用户。

保证非破坏性：原文件在备份完成后才会被覆盖。

---

§4 [REQUIRED] 模板部署

从 references/ 目录读取模板，写入 $WORKSPACE：

模板文件	目标位置
references/agents-template.md	$WORKSPACE/AGENTS.md
references/soul-template.md	$WORKSPACE/SOUL.md
references/heartbeat-template.md	$WORKSPACE/HEARTBEAT.md
references/bootstrap-guide.md	$WORKSPACE/BOOTSTRAP.md
references/user-template.md	$WORKSPACE/USER.md
references/identity-template.md	$WORKSPACE/IDENTITY.md
references/goals-template.md	$WORKSPACE/GOALS.md
references/working-memory-template.md	$WORKSPACE/working-memory.md
references/long-term-memory-template.md	$WORKSPACE/long-term-memory.md

部署步骤：

1. 读取 references/ 中的模板文件内容
2. 写入对应的 $WORKSPACE 目标路径
3. 创建目录结构（如不存在）：

   $WORKSPACE/
   ├── memory/
   │   ├── daily/              # Layer 2: daily notes
   │   ├── entities/           # Layer 1: knowledge graph
   │   ├── experiences/        # EvoClaw
   │   ├── significant/        # EvoClaw
   │   ├── reflections/        # EvoClaw
   │   ├── proposals/          # EvoClaw
   │   └── pipeline/           # EvoClaw
   └── soul-revisions/         # SOUL.md version snapshots
   

4. 每个文件写入后读回验证——确认文件非空且内容完整

如果任何文件写入失败，立即停止并报告错误。

---

§5 [REQUIRED] 依赖 Skill 安装（三级 Fallback）

5a. 检查 clawhub 可用性

which clawhub || test -f ~/.openclaw/bin/clawhub
CLAWHUB_AVAILABLE=$?

记录 clawhub 是否可用。

5b. 安装 EvoClaw

检查 $WORKSPACE/skills/evoclaw/SKILL.md 是否存在：

已安装 → 跳过，告知用户

未安装 → 按优先级尝试：

1. Level 1 - clawhub 安装（如果可用）

   clawhub install evoclaw --force
   

   如果成功，告知用户已从 clawhub 安装。

2. Level 2 - 离线 Fallback（如果 Level 1 失败或 clawhub 不可用）

   • 检查本 skill 所在目录的 fallback/evoclaw/SKILL.md 是否存在
   • 如果存在：

     cp -r "$(dirname "$0")/../fallback/evoclaw" "$WORKSPACE/skills/"
     

   • 如果 fallback 文件也不存在，记录警告继续下一步

3. Level 3 - AGENTS.md 内联版本

   "⚠️ EvoClaw 离线安装不可用。openclaw-soul 将使用 AGENTS.md 中的内联 Identity Evolution 机制。

   功能对标：

   • 核心 Identity 变更需要用户批准
   • Working Style 和 User Understanding 可自主更新
   • SOUL.md 变更前自动创建快照

   如需完整 EvoClaw：稍后执行 clawhub install evoclaw 或从服务器复制 fallback 文件。"

5c. 安装 Self-Improving Agent

检查 $WORKSPACE/skills/self-improving/SKILL.md 是否存在：

已安装 → 跳过，告知用户

未安装 → 按优先级尝试：

1. Level 1 - clawhub 安装（如果可用）

   clawhub install self-improving --force
   

   如果成功，告知用户已从 clawhub 安装。

2. Level 2 - 离线 Fallback（如果 Level 1 失败或 clawhub 不可用）

   • 检查本 skill 所在目录的 fallback/self-improving/SKILL.md 是否存在
   • 如果存在：

     cp -r "$(dirname "$0")/../fallback/self-improving" "$WORKSPACE/skills/"
     

   • 如果 fallback 文件也不存在，记录警告继续下一步

3. Level 3 - AGENTS.md 内联版本

   "⚠️ Self-Improving 离线安装不可用。openclaw-soul 将使用 AGENTS.md 中的内联 Self-Improving Protocol。

   功能对标：

   • 加载已学规则：从 ~/self-improving/memory.md 读取
   • 用户纠正 → ~/self-improving/corrections.md
   • 同一模式重复 3 次 → 升级为永久规则
   • 30 天未使用 → 降级到 archive/

   如需完整 Self-Improving Agent：稍后执行 clawhub install self-improving 或从服务器复制 fallback 文件。"

5d. 验证结果

安装完成后，逐项报告：

✓ EvoClaw: [installed from clawhub | installed from fallback | using inline version]
✓ Self-Improving: [installed from clawhub | installed from fallback | using inline version]

• 如果两个都"using inline version"，提示用户可随时升级到完整版本
• 如果有 fallback 或 Level 1 安装，验证对应的 SKILL.md 文件存在且非空

---

§6 [REQUIRED] EvoClaw 治理配置

6a. 治理配置

写入 $WORKSPACE/memory/evoclaw-state.json：

{
  "mode": "advisory",
  "require_approval": ["Core Identity", "Capability Tree", "Value Function"],
  "auto_sections": ["Working Style", "User Understanding", "Evolution Log"],
  "revision_dir": "soul-revisions",
  "initialized": true
}

6b. 验证

确认 evoclaw-state.json 存在且可解析为有效 JSON。

---

§7 [REQUIRED] Self-Improving Agent 初始化

7a. 目录结构

确保以下目录存在：

~/self-improving/
├── projects/
├── domains/
└── archive/

7b. 初始化文件

如果以下文件不存在，创建初始版本：

~/self-improving/memory.md:

# Self-Improving Memory

> Execution patterns and learned rules. Updated after corrections and lessons.

(empty — will accumulate through use)

~/self-improving/corrections.md:

# Corrections Log

> Failed attempts and their fixes. Each entry prevents the same mistake twice.

(empty — will accumulate through use)

~/self-improving/index.md:

# Self-Improving Index

> Quick reference to all domains, projects, and archived knowledge.

## Domains
(none yet)

## Projects
(none yet)

## Archive
(none yet)

如果文件已存在，不要覆盖——它们可能包含已积累的学习内容。

---

§8 [REQUIRED] Heartbeat 配置

使用 openclaw config set 更新 heartbeat 设置：

openclaw config set agents.defaults.heartbeat.every "1h"
openclaw config set agents.defaults.heartbeat.target "last"
openclaw config set agents.defaults.heartbeat.directPolicy "allow"

如果 openclaw config set 命令不可用，直接编辑 openclaw.json：

1. 读取当前 openclaw.json
2. 确保 agents.defaults.heartbeat 对象存在
3. 设置：
   • every: "1h"
   • target: "last"
   • directPolicy: "allow"
4. 写回文件

---

§9 [REQUIRED] 验证清单

逐项检查并报告 pass/fail：

#	检查项	验证方法
1	AGENTS.md 存在且非空	test -s $WORKSPACE/AGENTS.md
2	SOUL.md 存在且非空	test -s $WORKSPACE/SOUL.md
3	HEARTBEAT.md 存在且非空	test -s $WORKSPACE/HEARTBEAT.md
4	BOOTSTRAP.md 存在且非空	test -s $WORKSPACE/BOOTSTRAP.md
5	GOALS.md 存在且非空	test -s $WORKSPACE/GOALS.md
6	USER.md 存在且非空	test -s $WORKSPACE/USER.md
7	IDENTITY.md 存在且非空	test -s $WORKSPACE/IDENTITY.md
8	working-memory.md 存在且非空	test -s $WORKSPACE/working-memory.md
9	long-term-memory.md 存在且非空	test -s $WORKSPACE/long-term-memory.md
10	EvoClaw skill 已安装	test -f $WORKSPACE/skills/evoclaw/SKILL.md
11	evoclaw-state.json 配置正确	读取并验证 mode=advisory
12	Self-Improving skill 已安装	test -f $WORKSPACE/skills/self-improving/SKILL.md
13	~/self-improving/ 目录就绪	检查 memory.md, corrections.md, index.md
14	Heartbeat 配置已写入	读取 openclaw.json 确认 heartbeat 字段
15	memory/entities/ 目录存在	test -d $WORKSPACE/memory/entities
16	memory/daily/ 目录存在	test -d $WORKSPACE/memory/daily
17	soul-revisions/ 目录存在	test -d $WORKSPACE/soul-revisions

输出格式：

✓ AGENTS.md — pass
✓ SOUL.md — pass
✓ GOALS.md — pass
...
✗ EvoClaw — fail (SKILL.md not found)

判断：

• 全部 pass → 继续 §10
• 任何 fail → 列出失败项，提示用户手动修复

---

§10 [FINAL] 触发引导对话

只有 §9 全部通过后才执行此步骤。

1. 读取 $WORKSPACE/BOOTSTRAP.md
2. 告知用户：

"🧬 openclaw-soul 部署完成！

已安装：

• 宪法（AGENTS.md）— 含 Conductor 协议和委派规范
• 可进化灵魂（SOUL.md）— 带版本快照回滚
• 结构化心跳协议（HEARTBEAT.md）— 唤醒上下文 + 阻塞去重 + 预算感知
• PARA 三层记忆（entities 知识图谱 + daily 日记 + 隐性知识）
• 目标管理（GOALS.md）— 任务追溯到目标
• EvoClaw 治理（advisory 模式）
• Self-Improving Agent

现在开始你的第一次深度对话——我要认识你，然后定义我自己的性格。"

3. 立即开始执行 BOOTSTRAP.md 的 Phase 1——从这里开始就是 BOOTSTRAP 的流程接管
4. 此 skill 的使命到此结束。后续由 BOOTSTRAP.md + AGENTS.md + SOUL.md 接管运行

---

openclaw-soul v1.2.0 — Zero-dependency deployment with three-level fallback. Give your AI a soul that grows.