Workflow Decomposer

工作流任务拆解与模型编排技能。使用场景：(1) 收到复杂工作任务需要拆解为可执行步骤，(2) 需要为不同步骤选择最合适的模型，(3) 需要跟踪工作流进度和模型使用情况，(4) 长时间任务卡住需要问题诊断和解决方案。

Security Scan

This skill appears to do what it claims (decompose workflows and track progress) and does not request secrets or reach out to external endpoints; it only reads/writes local JSON state files and contains model-selection guidance.

Purpose & Capability

The name/description (workflow decomposition + model orchestration) aligns with the included templates and two Python scripts that manage workflow state. Nothing in the bundle requests unrelated credentials or system access. Minor mismatch: README and some wording imply the skill can '自动触发' (auto-trigger), but the published flags do not set always:true and there is no install-time trigger mechanism—this is an inconsistency in documentation rather than a technical risk.

Instruction Scope

SKILL.md stays focused on decomposition, model selection, progress tracking and stuck-step diagnosis. It does not instruct reading secrets or contacting external endpoints. The included scripts read/write workflow-state.json (or workspace/memory/workflow-state.json) — so the skill will persist state to the agent/workspace filesystem. The docs also reference web_fetch/web_search and runtime model availability as capabilities to choose models, but the skill does not require or configure network access; this is a functional assumption in the instructions that may not hold in all runtimes.

Install Mechanism

No install spec; this is primarily instruction + small local scripts. No downloads or third-party package installs are performed by the skill bundle itself.

Credentials

The skill declares no required environment variables, no credentials, and no special config paths beyond writing/reading local state files. The lack of secret access is proportional to the stated purpose.

Persistence & Privilege

The skill is not always-enabled and allows model invocation (normal). It does persist workflow state to files in the working directory (workflow-state.json) and to workspace/memory/workflow-state.json when using workflow_tracker.py. Writing to the workspace is expected for progress tracking but users should be aware files will be created/modified.

Assessment

This skill appears coherent and non-malicious, but review these before installing: (1) it writes state files (workflow-state.json or workspace/memory/workflow-state.json) to your agent/workspace — if you need to avoid writes run it in an isolated workspace or inspect/redirect those paths; (2) the skill prefers Alibaba Qwen models in its templates — ensure the runtime has those models available or edit templates to target models you trust; (3) documentation mentions auto-triggering and web_fetch capabilities but there is no install-time trigger or network integration in the bundle — if you expect automatic background runs or network fetches, verify the runtime behavior; (4) no secrets or external URLs are requested by the skill. If you want higher assurance, open and read the two Python scripts (they are short) and confirm the file paths used are acceptable, or run them in a sandboxed environment. If the publisher or homepage are unknown and you require provenance, prefer skills with a known source or more metadata.

Workflow Decomposer - 工作流任务拆解与模型编排

核心功能

本技能负责将复杂工作任务拆解为详细、可执行的步骤，并为每个步骤选择最合适的模型进行执行。

模型选择策略

推理模型优先级

1. 首选: 当前可用的最强推理模型
2. 次选: 如果有多个强推理模型，优先使用最新模型
3. 再次: 如果仍有多个，优先使用阿里系模型 (Qwen 系列)
4. 最后: 任意选择一个

模型输出要求

每次任务拆解后必须输出:

1. 任务拆解使用的模型: 明确告知使用的是哪个模型
2. 当前工作流进度: 清晰标识进行到了哪一步 (如: 步骤 2/8)
3. 当前步骤使用的模型: 说明该步骤由哪个模型执行

任务拆解流程

步骤 1: 接收任务

• 理解用户的核心需求
• 识别任务的复杂度和范围

步骤 2: 任务分析

• 分析任务需要的技能类型 (编码、写作、分析、创作等)
• 识别潜在的依赖关系和前置条件

步骤 3: 步骤拆解

• 将任务拆解为详细、可执行的子步骤
• 确保每个步骤都是可实现的，不胡编乱造
• 为每个步骤估算所需时间和复杂度

步骤 4: 模型匹配

• 根据每个步骤的特性选择最合适的模型
• 考虑因素: 模型专长、上下文长度、推理能力、速度

步骤 5: 执行与跟踪

• 按顺序执行每个步骤
• 跟踪进度并记录使用的模型
• 生成适合当前模型理解的内容格式

步骤 6: 问题处理

• 如果某步骤卡住超过合理时间，诊断问题
• 提供问题分析和至少 2 个解决方案
• 必要时调整后续步骤

输出格式模板

## 📋 任务拆解报告

**拆解模型**: [模型名称]
**任务**: [任务简述]
**总步骤数**: N

### 步骤概览
| 步骤 | 内容 | 使用模型 | 状态 |
|------|------|----------|------|
| 1/5 | [步骤描述] | [模型名] | ✅/⏳/❌ |
| 2/5 | [步骤描述] | [模型名] | ✅/⏳/❌ |
...

### 当前进度
**进行到**: 步骤 X/N
**当前步骤**: [详细描述]
**使用模型**: [模型名]
**预计耗时**: [时间]

### 步骤详情

#### 步骤 X: [步骤名称]
**目标**: [要完成什么]
**输入**: [需要什么信息/文件]
**输出**: [产生什么结果]
**模型提示**: [为该模型优化的执行指令]

问题诊断与解决

当工作流卡在某一步时:

1. 识别问题类型:

   • 模型理解错误
   • 工具/资源不可用
   • 依赖缺失
   • 超时/性能问题

2. 提供解决方案:

   • 方案 A: [直接解决方法]
   • 方案 B: [替代路径]
   • 方案 C: [降级方案]

3. 记录与学习:

   • 记录问题原因
   • 更新后续步骤的预防措施

模型内容优化

为不同模型生成适合其理解的内容:

• 强推理模型: 提供详细逻辑链和推理步骤
• 快速模型: 提供清晰、简洁的指令
• 代码模型: 提供明确的输入输出格式
• 创作模型: 提供风格参考和约束条件

使用示例

示例 1: 开发任务

用户: 帮我创建一个待办事项 Web 应用

拆解后:
- 步骤 1/6: 需求分析 (Qwen3.5-Plus)
- 步骤 2/6: 技术栈选择 (Qwen3.5-Plus)
- 步骤 3/6: 项目结构创建 (Claude Code)
- 步骤 4/6: 前端开发 (Claude Code)
- 步骤 5/6: 后端开发 (Claude Code)
- 步骤 6/6: 测试与部署 (Qwen3.5-Plus)

示例 2: 分析任务

用户: 分析这个项目的代码质量

拆解后:
- 步骤 1/4: 代码库扫描 (Qwen3.5-Plus)
- 步骤 2/4: 静态分析 (专用工具)
- 步骤 3/4: 问题分类 (Qwen3.5-Plus)
- 步骤 4/4: 报告生成 (Qwen3.5-Plus)

注意事项

1. 步骤粒度: 每个步骤应该在 5-30 分钟内可完成
2. 依赖检查: 确保前置步骤完成后才能执行后续步骤
3. 灵活调整: 根据实际情况动态调整步骤和模型选择
4. 透明沟通: 始终让用户知道当前进度和使用的模型

相关文件

• references/model-capabilities.md - 各模型能力对比
• references/workflow-templates.md - 常见工作流模板
• scripts/progress-tracker.py - 进度跟踪脚本 (可选)