技能验证
- 作者仓库星标 216
- 作者仓库 kagenti
- 领域
- 通用
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @kagenti · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 即装即用
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- 未声明(默认跨平台)
- 运行环境要求
- 未检出明确要求
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
# Validate Skill
## When to Use
- After creating or editing a skill
- Before committing skill changes
- When auditing all skills for consistency
## Validation Checks
### Required
- [ ] **Frontmatter**: Has `name:` and `description:` fields
- [ ] **Colon naming**: `name:` uses colon notation (e.g., `tdd:ci` not `tdd-ci`)
- [ ] **Directory match**: Directory name matches frontmatter `name:` field
- [ ] **Title**: Has `# Skill Name` as first heading
- [ ] **When to Use**: Has "When to Use" or "Overview" section
- [ ] **Related Skills**: Has "Related Skills" section at the end
- [ ] **Mermaid diagram**: Workflow/router skills have an embedded mermaid diagram
- [ ] **Diagram colors**: classDef colors match README color legend
### Command Format (Required)
- [ ] **Sandbox classification**: Skill is classified as sandbox or management (see below)
- [ ] **Single commands**: Sandbox skills use one command per code block (no `&&` chaining)
- [ ] **Auto-approve coverage**: All commands in sandbox skills match a pattern in `.claude/settings.json`
- [ ] **No multiline bash**: Sandbox skills avoid heredocs, multiline pipes, or `for` loops in commands
### Recommended
- [ ] **TOC**: Table of Contents present if skill > 50 lines
- [ ] **Placeholders**: All commands are copy-pasteable (no unexplained placeholders)
- [ ] **Task tracking**: TDD/RCA skills have "Task Tracking" section
- [ ] **Parent ref**: Parent category `SKILL.md` references this skill
- [ ] **Imperative voice**: Uses "Run X" not "You should run X"
- [ ] **Length**: Leaf skills are 80-200 lines (300 max)
- [ ] **Diagram-text match**: Diagram nodes correspond to textual flow steps
## Sandbox vs Management Classification
Skills operate on either **sandbox** (safe) or **management** (requires approval) targets:
| Type | Target | Auto-approve? | Command format |
|------|--------|---------------|----------------|
| **Sandbox** | Local Kind cluster, custom HyperShift hosted cluster | YES | Single commands, one per step |
| **Management** | Management cluster, AWS resources, git push, destructive ops | NO | Can chain commands (user approves anyway) |
### Sandbox skills (auto-approved)
Commands target `localtest.me`, `KUBECONFIG=~/clusters/hcp/kagenti-hypershift-custom-*`, or Kind clusters.
**IMPORTANT**: Run each command separately — not chained with `&&`. Chained or multiline commands break Claude Code's auto-approve pattern matching.
```markdown
## GOOD (each command runs separately, matches auto-approve patterns)
Check pod status:
```bash
kubectl get pods -n kagenti-system
```
Check logs:
```bash
kubectl logs -n kagenti-system deployment/mlflow
```
## BAD (chained commands won't match auto-approve patterns)
```bash
kubectl get pods -n kagenti-system && kubectl logs -n kagenti-system deployment/mlflow
```
```
### Management skills (require approval)
Commands target management clusters, AWS APIs, or perform destructive operations. These can use any command format since the user must approve each one.
## How to Validate
### Single Skill
```bash
# Check frontmatter
head -5 .claude/skills/<skill>/SKILL.md
# Check name matches directory
DIR_NAME=$(basename $(dirname .claude/skills/<skill>/SKILL.md))
SKILL_NAME=$(grep '^name:' .claude/skills/<skill>/SKILL.md | sed 's/name: //')
[ "$DIR_NAME" = "$SKILL_NAME" ] && echo "OK" || echo "MISMATCH: dir=$DIR_NAME name=$SKILL_NAME"
```
### All Skills
```bash
# Check all frontmatter name-vs-directory
for f in .claude/skills/*/SKILL.md; do
dir=$(basename $(dirname "$f"))
name=$(grep '^name:' "$f" | sed 's/name: //' | tr -d ' ')
[ "$dir" = "$name" ] || echo "MISMATCH: $dir != $name"
done
```
### Check Command Format (sandbox skills)
```bash
# Find chained commands in sandbox skills (potential auto-approve issues)
grep -n ' && ' .claude/skills/k8s:*/SKILL.md .claude/skills/kagenti:*/SKILL.md .claude/skills/kind:*/SKILL.md .claude/skills/local:*/SKILL.md .claude/skills/tdd:*/SKILL.md .claude/skills/rca:*/SKILL.md
```
### Check Mermaid Diagram Presence
```bash
for f in .claude/skills/*/SKILL.md; do
dir=$(basename $(dirname "$f"))
case "$dir" in git|k8s|auth|openshift|local|kind|helm|meta|genai|repo|testing) continue ;; esac
if ! grep -q '```mermaid' "$f"; then
echo "MISSING DIAGRAM: $dir"
fi
done
```
### Verify settings.json Coverage
For each command in a sandbox skill, verify it matches a pattern in `.claude/settings.json`:
| Command prefix | settings.json pattern |
|----------------|----------------------|
| `kubectl get` | `Bash(kubectl get:*)` |
| `kubectl describe` | `Bash(kubectl describe:*)` |
| `kubectl logs` | `Bash(kubectl logs:*)` |
| `helm list` | `Bash(helm list:*)` |
| `KUBECONFIG=~/clusters/hcp/... kubectl` | `Bash(KUBECONFIG=*/clusters/hcp/kagenti-hypershift-custom-*/auth/kubeconfig kubectl:*)` |
| `uv run pytest` | `Bash(uv run pytest:*)` |
If a command is NOT covered, add the pattern to `.claude/settings.json` in the `allow` array.
## Task Tracking
When validating multiple skills:
```
TaskCreate: "kagenti | skills | <category> | Verify | Validate <skill-name>"
```
## Related Skills
- `skills:write` - Create new skills following the standard
- `skills:retrospective` - Review session to identify skill improvements
章节与文件索引
按说明文件中的顺序排列。
何时使用
必填
命令格式(必填)
推荐
沙箱与管理分类
文件/命令 -> name: · description: · tdd:ci · tdd-ci · # Skill Name · .claude/settings.json · for · localtest.me
内容校验值 -> 5fd01ce5437c
当前原文的自动摘译,尚未人工审校。
何时使用
创建或编辑技能之后 提交技能更改之前 审核所有技能的一致性时
必填
[ ] Frontmatter:具有
name:和description:字段 [ ] 冒号命名:name:使用冒号表示法(例如,tdd:ci不是tdd-ci) [ ] 目录匹配:目录名称与 frontmattername:字段匹配 [ ] 标题:将# Skill Name作为第一个标题 [ ]何时使用:有“何时使用”或“概述”部分命令格式(必填)
[ ] 沙盒分类:技能被分类为沙盒或管理(见下文) [ ] 单命令:沙盒技能每个代码块使用一个命令(无
&&链接) [ ] 自动批准覆盖范围:沙盒技能中的所有命令都与.claude/settings.json中的模式匹配 [ ] 无多行 bash:沙盒技能避免命令中的heredocs、多行管道或for循环推荐
[ ] TOC:如果技能 > 50 行,则显示目录 [ ] 占位符:所有命令均可复制粘贴(没有未解释的占位符) [ ] 任务跟踪:TDD/RCA 技能有“任务跟踪”部分 [ ] 父参考:父类别
SKILL.md引用此技能 [ ] 命令式语音:使用“运行 X”而不是“你应该运行 X”沙箱与管理分类
技能在沙箱(安全)或管理(需要批准)目标上运行:类型·目标·自动批准? · 命令格式 沙箱 · 本地类集群、自定义 HyperShift 托管集群 · 是 · 单个命令,每步一个 管理 · 管理集群、AWS 资源、git 推送、破坏性操作 · 否 · 可以链接命令(无论如何都要经过用户批准)