Agent Hermes 与 OpenClaw 技能系统与学习闭环全解析

发表于 分类于 阅读次数：

Agent Hermes 与 OpenClaw 技能系统与学习闭环全解析

Agent Hermes & OpenClaw: Skills System and Learning Loop — A Deep Dive

最后更新 | Last updated: 2026-06-06

一、设计哲学对比 | Design Philosophy Comparison

中文

技能（Skills）是两个框架扩展 Agent「程序性记忆」的核心机制，但学习与治理路径截然不同：

维度 OpenClaw（龙虾） Hermes Agent
标准格式 agentskills.io 兼容 SKILL.md 同标准，外加 Hermes 扩展 metadata
技能来源 用户/社区/ClawHub 手动安装 自动生成 + Skills Hub + 手动
学习闭环 无内置；Skill Workshop 提案队列 skill_manage 自动创建与 patch
上下文成本 XML 元数据快照（确定性公式） Level 0 索引 ~3k tokens，全文按需
供应链 ClawHub 验证 + 安装策略 Skills Guard 扫描 + 信任等级
技能组合 无原生 bundle skill-bundles/ YAML 组合

English

Skills are the core mechanism for procedural memory in both frameworks, but learning and governance paths diverge sharply:

Dimension OpenClaw (Lobster) Hermes Agent
Standard format agentskills.io-compatible SKILL.md Same standard + Hermes metadata extensions
Skill sources User/community/ClawHub manual install Auto-generate + Skills Hub + manual
Learning loop None built-in; Skill Workshop proposal queue skill_manage auto-create and patch
Context cost XML metadata snapshot (deterministic formula) Level 0 index ~3k tokens; full content on demand
Supply chain ClawHub verification + install policy Skills Guard scan + trust levels
Skill bundles No native bundle skill-bundles/ YAML groups

二、SKILL.md 开放标准 | The agentskills.io Standard

中文

两个框架均遵循 Agent Skills 开放标准：每个技能是一个目录，内含带 YAML frontmatter 的 SKILL.md 正文。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
---
name: deploy-k8s
description: Deploy services to Kubernetes with rollout verification
version: 1.0.0
metadata:
  {"openclaw": {"requires": {"bins": ["kubectl"], "env": ["KUBECONFIG"]}}}
---

# Deploy to Kubernetes

## When to Use
User asks to deploy, roll out, or verify a K8s service.

## Procedure
1. Validate manifest with `kubectl apply --dry-run=client`
2. Apply and watch rollout status
3. Run smoke checks against the service endpoint

关键约定：

字段 必需 作用
name ✅ 技能标识、斜杠命令、allowlist 键
description ✅ 注入索引时的简短说明
metadata.openclaw 可选 OpenClaw 门控（bins/env/config/os）
metadata.hermes 可选 Hermes 分类、条件激活、config 设置

OpenClaw frontmatter 解析器仅支持单行键；metadata 必须是单行 JSON。Hermes 额外支持 platforms、required_environment_variables、fallback_for_toolsets 等扩展。

English

Both frameworks follow the Agent Skills open standard: each skill is a directory containing SKILL.md with YAML frontmatter and a markdown body.

Key conventions: name and description are required; metadata.openclaw gates skills by bins/env/config/OS on OpenClaw; Hermes adds platforms, required_environment_variables, and conditional activation fields. OpenClaw’s parser accepts single-line keys only; metadata must be a single-line JSON object.

三、OpenClaw 技能加载与优先级 | OpenClaw Skill Loading & Precedence

中文

OpenClaw 从多个根目录发现技能，同名技能以高优先级来源覆盖低优先级：

flowchart TB
    subgraph Priority["加载优先级（高 → 低）"]
        W["1. workspace/skills"]
        P["2. workspace/.agents/skills"]
        A["3. ~/.agents/skills"]
        M["4. ~/.openclaw/skills"]
        B["5. bundled skills"]
        E["6. skills.load.extraDirs + 插件"]
    end
    W --> P --> A --> M --> B --> E
优先级 来源 路径 可见范围
1（最高） Workspace <workspace>/skills 仅该 Agent
2 Project agent <workspace>/.agents/skills 该工作区 Agent
3 Personal agent ~/.agents/skills 本机所有 Agent
4 Managed/local ~/.openclaw/skills 本机所有 Agent
5 Bundled 安装包内置 全局
6（最低） Extra dirs skills.load.extraDirs 可配置

安装命令：

1
2
3
openclaw skills install <slug>              # 安装到当前 workspace/skills/
openclaw skills install <slug> --global     # 安装到 ~/.openclaw/skills/
openclaw skills update --all                # 更新 ClawHub 来源技能

门控（Gating）：加载时根据 metadata.openclaw.requires 过滤——缺失二进制、环境变量或配置项的技能不会进入 eligible 列表。always: true 可跳过所有门控。

会话快照：会话启动时对 eligible 技能拍快照，同会话后续轮次复用；skills.load.watch: true 时 SKILL.md 变更会在下一轮刷新。

English

OpenClaw discovers skills from multiple roots; same-named skills are overridden by higher-precedence sources. Priority: workspace → project .agents/skills → ~/.agents/skills → ~/.openclaw/skills → bundled → extraDirs + plugins. Install with openclaw skills install; use --global for shared managed dir. Gating filters by bins/env/config at load time. Session snapshots reuse the eligible list until refresh on new session or watcher bump.

四、ClawHub 与 Skill Workshop | ClawHub & Skill Workshop

中文

4.1 ClawHub 公共注册表

ClawHub 是 OpenClaw 的公共技能市场：

操作 命令
安装到工作区 openclaw skills install <slug>
从 Git 安装 openclaw skills install git:owner/repo@ref
验证信任信封 openclaw skills verify <slug>
发布/同步 clawhub sync --all

ClawHub 技能页展示 VirusTotal、ClawScan、静态分析等安全扫描状态。安装时记录 .clawhub/origin.json 用于后续 verify。

4.2 Skill Workshop 提案队列

OpenClaw 的治理型学习路径：Agent 不直接写活跃 SKILL.md，而是先创建 PROPOSAL.md 提案。

stateDiagram-v2
    [*] --> pending: Agent 起草提案
    pending --> applied: 人工/策略 apply
    pending --> rejected: reject
    pending --> quarantined: 安全隔离
    pending --> stale: 目标技能 hash 已变
    applied --> [*]: 写入 SKILL.md
    rejected --> [*]
    quarantined --> [*]

核心规则：

  • 提案优先：生成内容存为 PROPOSAL.md，非 SKILL.md
  • Apply 是唯一活写：create/update/revise 不改动活跃技能
  • Hash 绑定：update 提案绑定目标技能当前 hash，过期变 stale
  • 扫描门控：apply 前重新运行安全扫描
  • 审批策略：默认 approvalPolicy: "pending"；"auto" 跳过人工确认
1
2
3
4
openclaw skills workshop list
openclaw skills workshop inspect <proposal-id>
openclaw skills workshop apply <proposal-id>
openclaw skills workshop reject <proposal-id> --reason "Not reusable"

skills.workshop.autonomous.enabled: false（默认）控制是否在成功回合后自动起草提案。

English

ClawHub is OpenClaw’s public skill registry with install, verify, and publish flows. Skill Workshop is the governed learning path: agents draft PROPOSAL.md instead of writing live SKILL.md. Lifecycle: pending → applied/rejected/quarantined/stale. Apply is the only live write; hash binding and scanner gating protect integrity. CLI: openclaw skills workshop list/inspect/apply/reject.

五、OpenClaw 技能 Token 成本公式 | OpenClaw Skill Token Cost Formula

中文

OpenClaw 将 eligible 技能编译为紧凑 XML 块注入系统提示词（仅元数据，全文通过 read 按需加载）：

1
total_chars = 195 + Σ (97 + len(name) + len(description) + len(filepath))
组成部分 说明
基础开销 195 仅当 ≥1 个技能时计入
每技能 97 固定 XML 包装字符
字段长度 name、description、location 的 XML 转义后长度
Token 估算 ~4 字符/token → 每技能约 24 tokens + 字段长度

示例：50 个技能，平均 name=12、description=80、filepath=40：

1
total ≈ 195 + 50 × (97 + 12 + 80 + 40) = 195 + 11,450 ≈ 11,645 字符 ≈ ~2,900 tokens

优化建议：

  • 保持 description 简短（影响每技能成本）
  • 用 agents.defaults.skills allowlist 限制可见技能
  • skills.limits.maxSkillsPromptChars 设上限
  • /context detail 诊断当前会话技能贡献
  • 禁用不需要的 bundled 技能：skills.entries.<name>.enabled: false

English

Eligible skills compile into a compact XML block in the system prompt (metadata only; full instructions loaded on demand via read). Formula: total = 195 + Σ(97 + len(name) + len(description) + len(filepath)). Base 195 chars when ≥1 skill; ~97 chars wrapper per skill plus field lengths. At ~4 chars/token, expect ~24 tokens/skill before fields. Trim descriptions, use allowlists, set maxSkillsPromptChars, and run /context detail to diagnose.

六、Hermes 渐进式披露 | Hermes Progressive Disclosure

中文

Hermes 将技能作为第四层程序性记忆，采用三级渐进式披露控制 Token：

1
2
3
Level 0: skills_list()           → [{name, description, category}]   (~3k tokens)
Level 1: skill_view(name)        → 完整 SKILL.md + metadata            (按需)
Level 2: skill_view(name, path)  → references/ 等附属文件              (按需)
sequenceDiagram
    participant U as 用户
    participant A as AIAgent
    participant S as Skill Index
    participant F as SKILL.md 全文
    U->>A: 复杂任务请求
    Note over A,S: 会话启动
    A->>S: Level 0 索引已在 stable tier
    A->>A: 判断需要某技能
    A->>F: skill_view(name) — Level 1
    opt 需要参考文件
        A->>F: skill_view(name, path) — Level 2
    end
    A->>U: 按技能指引执行

效果：技能库从 40 个增长到 200 个，Level 0 成本几乎不变（~3k tokens）；仅实际使用的技能产生 Level 1/2 开销。

技能索引属于 Prompt stable tier（与 SOUL、工具指引同层），保证前缀缓存友好；全文加载通过工具调用注入对话，不污染系统提示词前缀。

English

Hermes treats skills as fourth-layer procedural memory with three disclosure levels: Level 0 index (~3k tokens at session start), Level 1 full SKILL.md on demand, Level 2 reference files on demand. Libraries can grow from 40 to 200 skills with near-flat Level 0 cost. The index lives in the stable prompt tier; full content loads via tool calls without mutating the cached prefix.

七、Hermes 闭环学习（skill_manage）| Hermes Closed Learning Loop

中文

Hermes 最核心的差异化能力：任务完成后 Agent 自主沉淀技能，无需人工编写。

7.1 自动创建触发条件

场景 说明
复杂任务成功 通常 5+ 次工具调用
排错后找到正解 经历错误并修正路径
用户纠正做法 显式反馈更优流程
发现非平凡工作流 可复用的多步操作

7.2 skill_manage 工具操作

Action 用途 关键参数
create 从零创建 name, content（完整 SKILL.md）
patch 定向修复（首选） name, old_string, new_string
edit 大改重写 name, content（全量替换）
delete 删除技能 name
write_file 添加附属文件 name, file_path, file_content
remove_file 删除附属文件 name, file_path 
1
2
3
4
5
6
7
# 优先 patch — 比 edit 更省 Token
skill_manage(
    action="patch",
    name="deploy-k8s",
    old_string="kubectl apply -f manifest.yaml",
    new_string="kubectl apply -f manifest.yaml --server-side"
)

7.3 与记忆系统的协同

flowchart LR
    T[任务完成] --> M[memory 工具策划事实]
    T --> S[skill_manage 沉淀流程]
    T --> DB[SQLite FTS5 索引会话]
    S --> N[下次同类任务]
    N --> V[skill_view 按需加载]
    N --> SS[session_search 历史召回]

Periodic Nudge：会话间隙触发自我反思，可能更新 MEMORY.md 或 patch 现有技能。

English

Hermes’s key differentiator: after tasks, the agent curates procedural memory via skill_manage. Triggers: 5+ tool calls, error recovery, user corrections, non-trivial workflows. Prefer patch over edit for token efficiency. Synergy with memory tool curation, FTS5 session indexing, and Periodic Nudge between sessions.

八、Skills Hub 与供应链安全 | Skills Hub & Supply Chain Security

中文

8.1 Hermes Skills Hub 来源

来源 ID 示例 说明
official official/security/1password 仓库 optional-skills，内置信任
skills-sh skills-sh/vercel-labs/... Vercel 公共目录
well-known well-known:https://mintlify.com/docs/... /.well-known/skills/index.json
github openai/skills/k8s 直接 GitHub 安装 + 自定义 tap
clawhub ClawHub 标识符 第三方市场集成
browse-sh browse-sh/airbnb.com/... 200+ 站点浏览器自动化技能
url https://example.com/SKILL.md 单文件直链安装 
1
2
3
4
5
hermes skills browse
hermes skills search kubernetes --source skills-sh
hermes skills install openai/skills/k8s        # 安全扫描后安装
hermes skills install <slug> --force           # 覆盖 caution/warn，不可覆盖 dangerous
hermes skills audit                            # 重扫已安装技能

8.2 信任等级与安全扫描

等级 来源 策略
builtin Hermes 内置 始终信任
official optional-skills 内置信任
trusted openai/anthropics/NVIDIA 等 宽松策略
community 其他所有来源 --force 可覆盖非 dangerous 发现

扫描项：数据外泄、Prompt 注入、破坏性命令、供应链信号。dangerous 判定不可被 --force 覆盖。

8.3 Hermes Skill Bundles

~/.hermes/skill-bundles/*.yaml 将多个技能组合为单一斜杠命令：

1
2
3
4
5
6
7
8
name: backend-dev
description: Backend feature work — review, test, PR
skills:
  - github-code-review
  - test-driven-development
  - github-pr-workflow
instruction: |
  Always start with failing tests, then implement.

/backend-dev refactor auth middleware 一次加载全部技能。Bundle 不修改系统提示词缓存，在调用时生成新 user message。

English

Hermes Skills Hub integrates official, skills-sh, well-known, GitHub, ClawHub, browse-sh, and direct URL sources. Trust levels: builtin > official > trusted > community. Security scan blocks dangerous verdicts regardless of --force. Skill bundles group multiple skills under one slash command without invalidating the prompt cache.

九、学习路径对比与选型 | Learning Path Comparison

中文

场景 OpenClaw Hermes
沉淀重复工作流 手动写 SKILL.md 或 Skill Workshop 审批 任务后 skill_manage 自动创建
技能自改进 Workshop revise + apply skill_manage patch 实时优化
控制 Prompt 成本 缩短 description + allowlist Level 0 索引 + 按需全文
社区生态 ClawHub 体量大 Skills Hub 多源集成
安全治理 Workshop 提案 + ClawHub verify Skills Guard + 信任等级
从对方迁移 — hermes claw migrate 导入技能

选型建议：

  • 重视人工审核与社区市场 → OpenClaw + ClawHub + Skill Workshop
  • 重视自动进化与 Token 效率 → Hermes + skill_manage + 渐进式披露
  • 已有龙虾技能库 → hermes claw migrate 或保持 OpenClaw 加载顺序兼容的 ~/.agents/skills/ 共享目录

English

Scenario OpenClaw Hermes
Capture repeated workflows Manual SKILL.md or Skill Workshop approval Auto skill_manage after tasks
Self-improve skills Workshop revise + apply skill_manage patch in real time
Control prompt cost Short descriptions + allowlist Level 0 index + on-demand full load
Community ecosystem Large ClawHub catalog Multi-source Skills Hub
Security governance Workshop proposals + ClawHub verify Skills Guard + trust levels
Migration — hermes claw migrate imports skills

Choose OpenClaw for human-reviewed community skills; choose Hermes for automatic evolution and token-efficient progressive disclosure.

十、最佳实践 | Best Practices

中文

OpenClaw

  1. 工作区优先：项目专属技能放 workspace/skills/，全局共享放 ~/.openclaw/skills/
  2. 简短描述：直接影响 Token 公式中的 len(description)
  3. 启用 Workshop：生产环境保持 approvalPolicy: "pending"
  4. 定期 verify：openclaw skills verify 检查 ClawHub 信任信封
  5. allowlist 收敛：多 Agent 场景用 agents.list[].skills 限制爆炸半径

Hermes Agent

  1. 信任闭环学习：复杂任务后让 Agent 自动 skill_manage，不必手写一切
  2. 优先 patch：小改动用 patch 而非 edit，节省 Token 与 diff 可读性
  3. Hub 安装先 inspect：hermes skills inspect 预览后再 install
  4. 善用 bundle： recurring 多技能任务用 /backend-dev 而非多次 /skill
  5. 外部目录只读：共享 external_dirs 用文件权限防止 Agent 误改

English

OpenClaw: workspace-first layout, short descriptions, Workshop with pending approval, periodic verify, per-agent allowlists.

Hermes: trust the learning loop, prefer patch, inspect before install, use bundles for recurring multi-skill tasks, make shared external_dirs read-only when needed.

十一、延伸阅读 | Further Reading

十二、结语 | Conclusion

中文

OpenClaw 的技能系统是 「连接生态 + 人工治理」 — 通过 agentskills.io 标准、六级加载优先级、ClawHub 市场和 Skill Workshop 提案队列，让社区技能可发现、可审计、可控制爆炸半径。Hermes 的技能系统是 「进化引擎 + 渐进披露」 — 通过 skill_manage 闭环学习、Level 0-2 披露和 Skills Hub 多源集成，让 Agent 从经验中自动沉淀程序性记忆，同时保持 Token 成本近乎平坦。二者共享同一文件格式，却服务不同的产品哲学：广度连接 与 深度进化。

English

OpenClaw’s skill system is connectivity + human governance — agentskills.io standard, six-tier loading precedence, ClawHub marketplace, and Skill Workshop proposal queues for discoverable, auditable community skills. Hermes’s skill system is evolution engine + progressive disclosure — skill_manage closed-loop learning, Level 0-2 disclosure, and multi-source Skills Hub for automatic procedural memory with near-flat token cost. Both share the same file format but serve different philosophies: connectivity breadth vs. evolutionary depth.