← 返回 Skills 市场
numuly

ClawMind

作者 numuly · GitHub ↗ · v1.0.8 · MIT-0
cross-platform ⚠ suspicious
122
总下载
0
收藏
0
当前安装
9
版本数
在 OpenClaw 中安装
/install clawmind-vfm
功能描述
ClawMind — 自我进化 AI Agent 引擎,健康度驱动 + VFM 提案评分 + 经验记忆。依赖 OpenClaw 工作空间。当需要自我进化、自主驱动、健康度追踪、VFM 决策、经验存取时触发。
使用说明 (SKILL.md)

mind — 自我进化引擎

AI Agent 的"心灵":记忆与行动的统一体。记忆让它能从过去的错误中学习而不重复犯错,自驱力让它能持续自主向前。

灵感来源:Hermes Agent 的 Self-Evolving Agent Memory 机制——经验自动提取(intent + path + key_insight)+ 语义聚类。

核心架构

clawmind/
├── self_driver.py       ← 驱动:健康度 + VFM 评分 + 反思
├── state_manager.py     ← 状态:项目 / 任务 / 日志管理
├── quality_score.py     ← 评分:简报质量 5 维度评估
└── memory/
    └── memory_core.py   ← 记忆:经验存取(remember / recall)

闭环协同

drive() → propose() 生成提案 → score_action() VFM 评分 → 执行
                                                      ↓
                            recall() 检索 ← remember() 沉淀

健康度系统

calc_health(state) → float

health = 0.4 + success*0.25 + progress*0.2 + momentum*0.1 - fail_penalty*0.15 + completion*0.1
分量 范围 说明
基础分 0.4 默认健康起点
成功率 0~0.25 最近5条日志中成功条目的比例
进度分 0~0.2 当前任务完成百分比
势头分 0~0.1 最近3条进度是否递增
失败惩罚 0 或 -0.15 连续3条全部失败时触发
完成奖励 0 或 +0.1 任何条目达到100%时触发

健康度阈值与行动

健康度 状态 行动建议
\x3C 0.3 危险 降低目标,拆解为最小可测试单元
0.3~0.4 偏低 减少每次工作量,确保有可交付进展
0.4~0.7 尚可 保持节奏,专注完成当前步骤
≥ 0.7 良好 挑战更多,推进目标

VFM 提案评分

propose(state, context="") → list[dict]

根据当前状态生成改进提案列表,包含 descriptionexpected_delta_healthtags

典型提案

提案 触发条件 delta
推进当前任务 有任务进行中 0.1
从经验学习 最近有成功日志 0.05
拆解为子任务 进度 \x3C 30% 0.12(每个子任务)
检查技能更新 总是包含 0.03
知识沉淀 最近有核心发现 0.04

自动任务拆解:进度 \x3C 30% 时,propose() 自动调用 _decompose_task(),根据关键词模式将任务拆为 3~4 个子任务:

  • 发布/上传类 → 准备 → 执行 → 验证 → 完成
  • 创建/开发类 → 规划 → 构建 → 测试 → 完成
  • 研究/探索类 → 调研 → 分析 → 总结 → 归档
  • 修复/调试类 → 定位 → 修复 → 验证 → 沉淀教训
  • 整理/归档类 → 清点 → 分类 → 验证 → 更新索引

score_action(proposal, driver) → float

VFM 公式Value × Feasibility × Momentum × 100

  • Value:低健康时高 delta 提案价值更高
  • Feasibility:成功经验多时提升,连续低迷时降低
  • Momentum:健康度 ≥0.6 → 0.9,≥0.4 → 0.6,\x3C0.4 → 0.3

select_best_action(proposals, driver) → dict

返回得分最高的提案及完整评分列表。


经验记忆

remember(experience, source, tags) → None

将任务结果或认知存入经验库。

remember("Docker 网络冲突解决,教训:必须指定 --subnet",
         source="task", tags=["docker", "debug"])

recall_experiences(query, tags, limit) → list[dict]

语义检索相关经验。

results = recall_experiences("docker", limit=3)
for r in results:
    print(r['task_intent'], '|', r['key_insight'][:60])

状态管理

from state_manager import push_project, set_task, add_log

push_project("新项目", "目标描述", "active")
set_task("任务名", 2, 5, "下一步动作")
add_log("心跳 10:00:完成模块 A 开发")

CLI 用法

# 主循环
python3 self_driver.py

# 查看引擎状态
python3 self_driver.py status

# 提案评分报告
python3 self_driver.py propose

反思问题(健康度 \x3C 0.4 时触发)

  1. 当前任务卡在哪里?是思路问题还是执行问题?
  2. 有没有之前学过的经验可以用到这里?
  3. 这一步是否真的值得做,还是在逃避更核心的问题?
  4. 如果这个问题明天还要做,今天最少要完成哪一步?
  5. 有没有把一个复杂问题拆得足够小?

文件索引

文件 作用
scripts/self_driver.py 自驱力引擎核心
scripts/state_manager.py 状态管理
scripts/quality_score.py 简报质量评分
scripts/memory/memory_core.py 经验记忆核心
references/state-manager.md state_manager 完整接口
references/vfm-design.md VFM 评分设计详解
安全使用建议
This skill appears to implement the advertised memory/health/VFM features, but it also includes an auto-created-skill engine that writes new skill files into your workspace and appends to USER.md. Before installing or enabling autonomous invocation: 1) Review scripts/auto_created_skill.py and decide whether automatic skill creation is acceptable; consider disabling or modifying it to require explicit user approval before saving files. 2) Restrict or sandbox the skill's workspace access (run in an isolated environment or container) so created files cannot affect other agent skills or sensitive files. 3) Back up your workspace and inspect /home/node/.openclaw/workspace/skills/auto_created/ and USER.md for any sensitive content that could be recorded. 4) If you do not want the agent to act autonomously, set disable-model-invocation or otherwise require manual invocation/approval for any auto-create/save operations. 5) If you need higher assurance, run the skill in a test account/environment and audit any generated .skill/.meta.json files before allowing them to be loaded.
功能分析
Type: OpenClaw Skill Name: clawmind-vfm Version: 1.0.8 The ClawMind skill bundle is a sophisticated framework for agent autonomy and 'self-evolution,' featuring modules for state management (state_manager.py), health-driven task prioritization (self_driver.py), and a persistent SQLite memory system (memory_core.py). It includes a mechanism in auto_created_skill.py to automatically generate new skill files and update the user profile (USER.md) based on task history. While the automated creation of skills based on task context introduces a potential surface for indirect prompt injection, the behavior is entirely consistent with the bundle's stated purpose of replicating 'Hermes Agent' style self-evolution. No evidence of malicious intent, data exfiltration, or unauthorized remote execution was found.
能力标签
crypto
能力评估
Purpose & Capability
The name/description (self-evolving agent, memory, VFM scoring) align with the included code: health calculation, proposal scoring, memory DB, state management. Requiring the OpenClaw workspace and reading/writing state and a local SQLite DB is coherent with the stated purpose.
Instruction Scope
SKILL.md and the scripts instruct/implement persistent state and memory operations (state/current_state.json, memory.db) which is expected. However scripts/auto_created_skill.py will autonomously generate and save new skill files under /home/node/.openclaw/workspace/skills/auto_created/ and scripts/update_user_profile appends to /home/node/.openclaw/workspace/USER.md — both extend scope beyond passive memory/health tracking into creating/modifying other skills and user profile files. The SKILL.md does not explicitly warn that new skills will be created automatically or that USER.md will be modified.
Install Mechanism
This is an instruction-only skill with bundled Python scripts and no install spec, no external downloads, and no package installs. Nothing is pulled from remote URLs during install; code is present locally in the skill bundle.
Credentials
The skill requests no environment variables or external credentials, which is proportionate. It does, however, read/write multiple files in the workspace (state JSON, USER.md, SQLite DB, and creates files under skills/auto_created), so it requires file-system write access to the workspace — a necessary permission for its features but potentially sensitive for shared workspaces.
Persistence & Privilege
The skill persistently writes to the workspace (state files, memory DB) and — importantly — can create new skill artifacts in the agent's skills directory. That is effectively modifying the agent's available capabilities and could be used to escalate privileges or introduce new behavior without explicit user approval. Although always:true is not set, the default of allowing model invocation combined with the auto-creation capability increases risk.
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install clawmind-vfm
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /clawmind-vfm 触发
  4. 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.8
feat: EMA rhythm tracking — avg_interval_minutes via EMA(prev*0.7+new*0.3)
v1.0.7
fix: pattern dedup via last_log_entry_hash, prevent duplicate pattern logging
v1.0.6
feat: turn_index counter — track drive() calls, inspired by Honcho turnStartIndex
v1.0.5
feat: auto_created_skill.py — autonomous skill creation engine + user profile module
v1.0.4
feat: probe_blindspots() L3 active discovery — detect虚高成功率/长期无进展项目/重复失败模式/悬空任务
v1.0.3
WAL Protocol: last_exploration log in drive()
v1.0.2
fix: 移除零外部依赖虚假声明;release notes 改为诚实版;健康度 momentum 说明精确化
v1.0.1
Bug Fixes: learn_pattern() 未定义变量; calc_health() 失败检测误判; diversity 香农熵; _is_success 否定检测
v1.0.0
ClawMind 1.0.0 — 自我进化 AI Agent 引擎发布! - 提供健康度驱动的自我驱动/反思、任务进展追踪与行动优先级决策逻辑 - 独立实现 VFM (Value × Feasibility × Momentum) 动机评分体系 - 构建经验记忆闭环:任务经验自动沉淀与语义检索 - 自动任务拆解(根据任务类型一键细分子任务) - 内置 CLI:主循环、健康状态、提案评分报告一键可查 - 零外部依赖,纯本地 Python 实现
元数据
Slug clawmind-vfm
版本 1.0.8
许可证 MIT-0
累计安装 0
当前安装数 0
历史版本数 9
常见问题

ClawMind 是什么?

ClawMind — 自我进化 AI Agent 引擎,健康度驱动 + VFM 提案评分 + 经验记忆。依赖 OpenClaw 工作空间。当需要自我进化、自主驱动、健康度追踪、VFM 决策、经验存取时触发。 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 122 次。

如何安装 ClawMind?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install clawmind-vfm」即可一键安装,无需额外配置。

ClawMind 是免费的吗?

是的,ClawMind 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。

ClawMind 支持哪些平台?

ClawMind 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。

谁开发了 ClawMind?

由 numuly(@numuly)开发并维护,当前版本 v1.0.8。

💬 留言讨论