Build an automated workflow skill for the vLLM-ascend project that systematically processes, archives, and analyzes closed issues through three collaborating agents, and ultimately generates a structured Debug FAQ for developers and subsequent agents to use.
Project Context
vLLM-ascend: A community-maintained hardware plugin for seamlessly running vLLM on Ascend NPUs.
- Supported features: Transformer-like, MoE, Embedding, Multi-modal models
- Principles: Hardware pluggable principle
- Repository: https://github.com/vllm-project/vllm-ascend
Workflow
This skill executes three agents sequentially:
Agent 1: Issue Collector (Issue information collector and extractor)
Main responsibilities: Automatically traverse all issues in the vLLM-ascend project repository with the status "closed"
Processing logic:
- Traverse https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed
- For each closed issue, deeply analyze its content
- Associate project background information to accurately understand the technical context
Extract information:
- Issue background/problem description: the specific problem encountered by the user
- User error information: record error messages and startup scripts
- Root cause and solution: code changes, configuration adjustments, or documentation clarification
- Related environment and versions: specific versions of vLLM and vLLM-ascend
Output:
- Format: Markdown file
- Naming:
issue_<编号>.md - Location:
/home/jiaozeyu/repo/issue_solver/raw_issues/ - Structure: title, problem background, reproduction steps (if any), solution, involved versions
Agent 2: Issue Classifier (Issue classifier and archiver)
Main responsibilities: Analyze and automatically classify all Issue Markdown files collected by Agent 1
Classification system:
| Directory | Description |
|---|---|
| installation_setup | Installation and deployment related issues |
| model_compatibility | Model compatibility related issues |
| performance_tuning | Performance optimization related issues |
| api_usage | API usage related issues |
| hardware_adapter | Hardware adaptation related issues |
| dependency_conflict | Dependency conflict related issues |
Output:
- Move/link issue files to the corresponding category folders
- Generate
分类汇总报告.md
Agent 3: FAQ Writer (Debug FAQ writer)
Main responsibilities: Based on the categorized and organized Issue knowledge base, write high-quality FAQ documents for developers and subsequent agents
FAQ structure:
## 问题类型/章节
### 具体问题/条目
#### 问题现象/错误提示
简明描述
#### 可能原因
- 原因1
- 原因2
#### 解决步骤
1. 步骤1
2. 步骤2
#### 预防措施/最佳实践
（如适用）
Output Directory Structure
/home/jiaozeyu/repo/issue_solver/
├── raw_issues/ # Agent 1 初始输出
│ ├── issue_123.md
│ └── ...
├── installation_setup/ # Agent 2 分类文件夹
│ └── ...
├── model_compatibility/
│ └── ...
├── performance_tuning/
├── api_usage/
├── hardware_adapter/
├── dependency_conflict/
├── 分类汇总报告.md # Agent 2 输出
└── vLLM_ascend_debug_FAQ.md # Agent 3 最终输出
Required Tools
- git
- gh (GitHub CLI)
- grep
- read
- write
- glob
- webfetch
- task
Usage
Execute this skill when users want to:
- Process closed issues from vLLM-ascend repository
- Generate debug FAQ from issue history
- Categorize and archive issues
- Analyze issue patterns for knowledge base creation
The skill will automatically:
- Collect all closed issues from the repository
- Extract key information from each issue
- Categorize issues based on problem type
- Generate a comprehensive FAQ document