LLM Wiki

Skill Factory Layers

7分钟阅读

Skill Factory Layers

Definition

定义

The Skill Factory 7-layer architecture is a reference model for building, testing, and distributing agent skills at enterprise scale. Each layer addresses a specific concern in the skill lifecycle.

The Skill Factory 7-layer architecture 是一个用于在企业级规模下构建、测试和分发 Agent skills 的参考模型。每一层都解决了 Skill 生命周期中的特定问题。

Details

详情

The 7-layer model separates concerns from infrastructure to delivery:

7 层模型将关注点从基础设施层到交付层进行了分离：

Layer 1 - Infrastructure

  • Runtime environments: Docker, Firecracker microVMs, Kubernetes
  • Sandboxing: Isolated execution to prevent skill interference
  • Resource limits: CPU, memory, disk I/O caps per skill
  • Networking: Controlled egress for external API calls
**第一层 - 基础设施** - **运行环境**：Docker、Firecracker microVM、Kubernetes - **沙箱机制**：隔离执行环境，防止 Skill 之间相互干扰 - **资源限制**：针对每个 Skill 设定 CPU、内存及磁盘 I/O 上限 - **网络管控**：针对外部 API 调用实施受控的出站流量管理

Layer 2 - Skill Definition

  • SKILL.md specification: Standardized format for skill documentation
  • Metadata schema: Name, version, permissions, dependencies
  • Versioning: Semantic versioning with compatibility declarations
  • Progressive disclosure: ~100 tokens metadata, <5000 tokens body
**第 2 层 - Skill 定义** - **SKILL.md 规范**：Skill 文档的标准化格式 - **Metadata schema**：包括名称、版本、权限、依赖项 - **版本控制**：语义化版本控制，包含兼容性声明 - **渐进式披露**：约 100 tokens 的 metadata，正文 < 5000 tokens

Layer 3 - Orchestration

  • Skill composition: Chaining multiple skills for complex tasks
  • Dependency resolution: Installing required dependencies
  • Execution planning: Determining skill invocation order
  • State management: Passing data between skills
**第 3 层 - Orchestration** - **Skill 组合**：链式调用多个 Skill 以处理复杂任务 - **依赖解析**：安装所需的依赖项 - **执行规划**：确定 Skill 的调用顺序 - **状态管理**：在 Skill 之间传递数据

Layer 4 - Secure Execution

  • Permission enforcement: Check ACLs before skill execution
  • Audit logging: Record all skill invocations with context
  • Rollback mechanisms: Undo actions when errors detected
  • Timeout controls: Prevent runaway skills
**第 4 层 - 安全执行** - **权限强制执行**：在 skill 执行前检查 ACL - **审计日志**：记录所有 skill 调用及其上下文 - **回滚机制**：检测到错误时撤销操作 - **超时控制**：防止 skill 失控运行

Layer 5 - Behavioral Observability

  • OpenTelemetry integration: Distributed tracing for skill execution
  • Performance metrics: Latency, throughput, error rates
  • Error tracking: Structured logging for debugging
  • Usage analytics: Which skills are used, by whom, how often
**第 5 层 - 行为可观测性** - **OpenTelemetry 集成**：针对 Skill 执行的分布式追踪 - **性能指标**：延迟、吞吐量、错误率 - **错误追踪**：用于调试的结构化日志 - **使用分析**：Skill 使用情况、使用者及使用频率

Layer 6 - Governance

  • Approval workflows: Skills require review before publication
  • Compliance checks: Validate skills meet security policies
  • Access control: Role-based permissions for skill usage
  • Deprecation: Sunset old skills with migration paths
**第 6 层 - 治理** - **审批工作流**：Skill 发布前需经过审查 - **合规性检查**：验证 Skill 是否符合安全策略 - **访问控制**：基于角色的 Skill 使用权限管理 - **废弃**：按迁移路径逐步废弃旧 Skill

Layer 7 - Delivery

  • Registry APIs: Publish and discover skills
  • Package management: Install, update, uninstall skills
  • Update distribution: Push updates to deployed agents
  • Rollback: Revert to previous skill versions
**第 7 层 - 交付** - **Registry APIs**：发布和发现技能 - **Package management**：安装、更新和卸载技能 - **Update distribution**：向已部署的 Agent 推送更新 - **Rollback**：回退到之前的技能版本

6-Phase Build Workflow

**六阶段构建 Workflow**

The factory uses a standardized workflow for skill development:

工厂使用标准化的工作流进行技能开发：

Phase 1 - Spec

  • Define requirements and interface
  • Identify permissions needed
  • Document expected inputs/outputs
**阶段 1 - Spec** - 定义需求与接口 - 确定所需的权限 - 记录预期的输入/输出

Phase 2 - Scaffold

  • Generate boilerplate from templates
  • Set up testing infrastructure
  • Configure CI/CD pipeline
**阶段 2 - Scaffold** - 基于模板生成脚手架代码 - 搭建测试基础设施 - 配置 CI/CD 流水线

Phase 3 - Implement

  • Write core logic with security boundaries
  • Handle errors and edge cases
  • Optimize for token efficiency
**第 3 阶段 - 实施** - 编写具有安全边界的核心逻辑 - 处理错误和边缘情况 - 针对 token 效率进行优化

Phase 4 - Test

  • Unit tests for individual functions
  • Integration tests in sandbox
  • Security tests for permission enforcement
**阶段 4 - 测试** - 针对单个函数的单元测试 - 在沙箱环境中进行的集成测试 - 针对权限执行的安全测试

Phase 5 - Document

  • Generate SKILL.md with examples
  • Write usage guides
  • Document known limitations
**第 5 阶段 - 文档** - 生成包含示例的 SKILL.md - 编写使用指南 - 记录已知限制

Phase 6 - Publish

  • Push to registry with semantic versioning
  • Announce to users
  • Monitor adoption and errors
**第 6 阶段 - 发布** - 推送至 Registry 并采用语义化版本控制 - 向用户发布公告 - 监控采用率及错误

Progressive Disclosure Principle

**渐进式披露原则**

Skills follow a token budget hierarchy:

  • Metadata: ~100 tokens (name, description, permissions)
  • Body: <5000 tokens (full implementation)
  • References: On-demand (detailed docs, examples)
Skill 遵循 Token 预算层级结构： - **Metadata**：约 100 Token（名称、描述、权限） - **Body**：<5000 Token（完整实现） - **References**：按需加载（详细文档、示例）

This ensures skills remain discoverable without overwhelming agent context windows.

这确保了 skills 保持可发现性，同时不会耗尽 Agent 的上下文窗口。

China-Specific Adaptations

  • Domestic registries: Skills hosted on Alibaba Cloud, Tencent Cloud
  • Compliance layer: MLPS 2.0 and PIPL checks built into Layer 6
  • Platform integration: Skills packaged for DingTalk, Feishu, WeChat
  • Localization: Documentation in Chinese, examples use domestic services
**中国区专属适配**
  • 国内镜像仓库：托管于阿里云、腾讯云的 Skills
  • 合规层：内置于 Layer 6 的 MLPS 2.0 和 PIPL 合规检查
  • 平台集成：适配钉钉、飞书、微信的 Skills 封装
  • 本地化：中文文档，示例采用国内服务

Connections

连接
- 相关内容：[[enterprise-agent-china/concepts/skill-registry|Skill Registry]]，[[enterprise-agent-china/entities/agentskills-io|agentskills.io]] - 提及于：[[enterprise-agent-china/sources/skill-factory-framework|Skill Factory Framework]]

关系图谱

反向链接