工具厂
Claude Code工具的自我修复生命周期系统。大规模构建、评分、翻新和管理技能、代理、钩子、命令和MCP服务器。
问题
Claude Code的工具生态系统发展迅速。经过数月的构建，你最终会得到数百个工具——技能、代理、钩子、命令、MCP服务器、插件——而且无法知道哪些工具有效，哪些是重复的，哪些是无用的。质量不一致。工具悄无声息地腐烂。
解决方案
工具厂将工具视为产品。生来就有标准，经过标准测试，漂移时翻新，过时时退役。制造机器并对其进行维护的机器。
建筑
两层，三个空间：```
Layer 1: Factory (builds and maintains tools)
Layer 2: Tools (the skills/agents/hooks that do actual work)
Factory Floor (generators/) -- Creates new tools from templates
Range (range/) -- Scores tools against fitness criteria
Workshop (workshop/) -- Auto-retrofits failing tools
### 目录结构```
tool-factory/
generators/ # Tool creators (skill, agent, command, hook, MCP)
range/ # Quality scoring (14 criteria, 100-point scale)
workshop/ # Auto-retrofit for failing tools
composer/ # Pipeline chaining (YAML compositions)
registry/ # Tool inventory + intelligence + lifecycle
index.json # Single source of truth for all tools
intelligence.sh # Fleet health, usage, trends, dormant detection
lifecycle.sh # 4-phase maintenance cycle
score-history.jsonl # Append-only score snapshots
track-usage.sh # Invocation logger
lexicon/ # Constraint documents (best practices per vendor)
scripts/ # Automation (scheduled lifecycle wrapper)
lib/ # Shared utilities (colors, etc.)
validator/ # Structural validation
VISION.md # Where we're going (north star)
SPEC.md # What exists today (living spec)
BUILDING.md # How we got here (build log)
快速开始
创建工具```bash
Generate a new skill with frontmatter, test fixtures, and registry entry
bash generators/skill.sh my-new-skill
Generate other tool types
bash generators/agent.sh my-agent
bash generators/command.sh my-command
bash generators/hook.sh my-hook
bash generators/mcp.sh my-mcp-server
### 为你的工具打分```bash
# Score a single tool
bash range/runner.sh my-skill
# Score all tools (fleet-wide)
bash range/runner.sh --all
# Report mode (no changes)
bash range/runner.sh --report
评分以100分制对3个类别的14个标准进行评估。结果：通过（70+），部分通过（40-69），未通过（\<40）。
改装故障工具```bash
Fix a single tool
bash workshop/retrofit.sh my-skill
Fix all failing tools
bash workshop/retrofit.sh --all
研讨会增加了缺失的结构（前体、触发器、工作流），但没有触及使工具有价值的内容。
### 舰队情报```bash
# Fleet health overview
bash registry/intelligence.sh --health
# Usage analysis (top tools, ghost tools, decay risk)
bash registry/intelligence.sh --usage
# Score trends and regression detection
bash registry/intelligence.sh --trends
# Full intelligence report
bash registry/intelligence.sh
运行生命周期```bash
Full 4-phase maintenance: score -> retrofit -> dormant check -> decay alerts
bash registry/lifecycle.sh
Report mode (dry run, no changes)
bash registry/lifecycle.sh --report
Individual phases
bash registry/lifecycle.sh --score
bash registry/lifecycle.sh --fix
bash registry/lifecycle.sh --flag-dormant
### 组合管道```bash
# Create a new pipeline
bash composer/compose.sh new my-pipeline
# List all pipelines
bash composer/compose.sh list
# Validate a pipeline
bash composer/compose.sh validate my-pipeline
# Run a pipeline
bash composer/compose.sh run my-pipeline
组合是将工具链接到顺序工作流中的YAML文件：```yaml
name: deploy-pipeline
description: Score, fix, and report
steps:
- type: shell
name: preflight
command: "echo 'Starting...'"
-
type: shell
name: get-version
command: "cat package.json | jq -r .version"
export_as: VERSION -
type: shell
name: test
command: "npm test"
on_fail: fallback-lint -
type: shell
name: deploy
command: "echo Deploying v$VERSION" -
type: shell
name: fallback-lint
command: "npm run lint"
skip: true
特点：环境变量注射(`export_as`)，条件分支(`on_fail`)，只跳几步(`skip: true`)，以及 `continue_on_fail` 对于非关键步骤。
### 计划自动维护（macOS）
计划的生命周期每天通过launchd运行——工作日的模拟运行报告，周日的实时修复：```bash
# Test the wrapper manually
bash scripts/scheduled-lifecycle.sh
# Force a live run (regardless of day)
bash scripts/scheduled-lifecycle.sh --force
看 scripts/ 用于启动plist模板。
三合会
工具厂使用三文档系统而不是PRD：
- 视觉.md --我们要去哪里。七个支柱，每个支柱都有一个实现百分比。
- 规格.md --今天存在什么。能力、架构、验证界面。
- 建筑.md --我们是如何到达这里的。建立包含决策和理由的日志。
VISION和SPEC之间的差异就是路线图。任何两个文件都可以重建第三个文件。
需求
- macOS或Linux
- Bash 3.2+
- Python 3.6+
- Claude Code（用于工具本身）
许可证
麻省理工学院——见 许可证.
作者
建造于 id8Labs 作为Claude Code工具生态系统的一部分。
