为什么值得关注

它将经过实战检验的智能体工作流提炼为可重用模板，自动实现多智能体委派（协调者、架构师、构建者群、审查者），强制执行对抗性审查关卡，并通过一条 npx 命令适用于任何仓库。

适合谁使用

使用 Claude 进行 AI 辅助编程的开发者
需要结构化多智能体 AI 工作流的团队
寻求自动化 PR 审查的开源维护者
探索 AI 驱动开发流程的工程管理者

典型使用场景

使用 AI 管理并行任务来自动化功能实现
通过参考规范的等同模式重构或移植遗留代码
运行跨多个独立审查者的对抗性代码审查
为 AI 生成的变更采用安全的分支工作流

项目优势

通过 npx contractor-kit 无需安装即可在任何仓库上直接运行
强制分离职责（架构师设计、构建者实现、审查者把关）
包含秘密扫描、危险命令拦截和文件保护等安全钩子
支持可配置的模型层级以匹配可用 AI 模型

使用前须知

设计上会覆盖已有的 CLAUDE.md 和 .claude/ 文件（但有提示和备份）
需要手动配置占位符（contractor.config）和模型设置
复杂流水线对于小型或单文件修改可能过于繁琐

README 快速开始

=16" />

npx contractor-kit — install a senior-contractor workflow into any repo in one command.

Contractor

A disciplined, delegation-driven workflow for Claude. Drop it into any repo and Claude starts working like a senior engineering contractor: it researches before it builds, plans before it executes, delegates grunt work to a builder swarm, reviews everything adversarially, and never touches your default branch directly.

Contractor is a project-agnostic distillation of a battle-tested CLAUDE.md and .claude/ setup — the personal paths, names, and one-project specifics stripped out and replaced with {{PLACEHOLDERS}} you fill once.

What's inside

bin/cli.js                      # the npx installer (`contractor` / `contractor fill`)
package.json                    # npm package manifest (bin, files, MIT)
template/                       # the payload copied into a target repo:
  CLAUDE.md                     #   Claude's instructions (Research → Plan → Execute → Review)
  contractor.config.example     #   the placeholders to fill, once per project
  .claude/
    agents/                     #   task-manager, architect, builder, reviewer, code-reviewer,
                                #   security-reviewer, pr-test-analyzer, performance-reviewer, silent-failure-hunter
    hooks/                      #   secret scan, dangerous-command block, file protection, session context
    rules/code-quality.md       #   naming, markers, anti-defaults
    settings.json               #   wires the hooks + a safe permission allow/deny list

Core ideas (full detail in template/CLAUDE.md):

Research → Plan → Execute → Review — never execute first.
Branch Safety — every change on its own branch, PR + independent review before merge, no direct commits to the default branch.
The Delegation Org — orchestrator (owns strategy + git) → task-manager → architect (design only) → builder swarm (5+, parallel) → independent reviewers. No agent both writes and self-approves.
Spec Engine — a named source of truth is the executable spec; nobody invents behavior.
Model tiers — map orchestrator / thinking / builder to whatever models you run.

Features

Delegation org, not a lone agent. Work flows through a fixed chain — orchestrator (owns strategy + git) → task-manager (backlog) → architect (design only) → builder swar

项目描述

A disciplined, delegation-driven operating system for AI coding agents. npx contractor-kit.

contractor