Everything Claude Code (ECC) — 智能体框架性能系统 (Agent Harness Performance System)

技能来自 ara.so — Daily 2026 Skills 合集。

Everything Claude Code (ECC) 是一个面向 AI 智能体框架的生产就绪型性能优化系统。它提供专门的子智能体、可复用的技能、自定义斜杠命令、持久化记忆的钩子、安全扫描以及语言特定的规则——所有这些都源自 10 多个月的日常实际使用。适用于 Claude Code、Cursor、Codex、OpenCode 和 Antigravity。

安装 (Installation)

选项 1：插件市场（推荐）(Option 1: Plugin Marketplace – Recommended)

# 在 Claude Code 内部运行： (Inside Claude Code, run:)
/plugin marketplace add affaan-m/everything-claude-code
/plugin install everything-claude-code@everything-claude-code

选项 2：手动克隆 (Option 2: Manual Clone)

git clone https://github.com/affaan-m/everything-claude-code.git
cd everything-claude-code

# 为你的技术栈安装规则 (Install rules for your language stack)
./install.sh typescript
# 多种语言 (Multiple languages):
./install.sh typescript python golang swift
# 指定目标 IDE (Target a specific IDE):
./install.sh --target cursor typescript

安装规则（始终需要）(Install Rules – Always Required)

Claude Code 插件无法自动分发规则——请通过 ./install.sh 手动安装，或从 rules/ 复制到项目的 .claude/rules/ 目录中。

目录结构 (Directory Structure)

everything-claude-code/
├── .claude-plugin/         # 插件和市场清单 (Plugin and marketplace manifests)
│   ├── plugin.json
│   └── marketplace.json
├── agents/                 # 专门的子智能体（规划器、架构师等）(Specialized subagents - planner, architect, etc.)
├── commands/               # 斜杠命令 (/plan, /security-scan, 等) (Slash commands)
├── skills/                 # 可复用的技能模块 (Reusable skill modules)
├── hooks/                  # 生命周期钩子（会话启动、停止、编辑后等）(Lifecycle hooks - SessionStart, Stop, PostEdit, etc.)
├── rules/
│   ├── common/             # 语言无关规则 (Language-agnostic rules)
│   ├── typescript/
│   ├── python/
│   ├── golang/
│   └── swift/
├── scripts/                # 安装和工具脚本 (Setup and utility scripts)
└── install.sh              # 交互式安装器 (Interactive installer)

关键命令 (Key Commands)

安装后，使用带命名空间的形式（插件安装）或短形式（手动安装）：

# 规划与架构 (Planning & architecture)
/everything-claude-code:plan "添加 OAuth2 登录流程 (Add OAuth2 login flow)"
/everything-claude-code:architect "设计多租户 SaaS 系统 (Design a multi-tenant SaaS system)"

# 研究优先的开发 (Research-first development)
/everything-claude-code:research "Node.js 中限流的最佳方法 (Best approach for rate limiting in Node.js)"

# 安全 (Security)
/everything-claude-code:security-scan
/everything-claude-code:harness-audit

# 智能体循环与编排 (Agent loops and orchestration)
/everything-claude-code:loop-start
/everything-claude-code:loop-status
/everything-claude-code:quality-gate
/everything-claude-code:model-route

# 多智能体工作流 (Multi-agent workflows)
/everything-claude-code:multi-plan
/everything-claude-code:multi-execute
/everything-claude-code:multi-backend
/everything-claude-code:multi-frontend

# 会话与记忆 (Session and memory)
/everything-claude-code:sessions
/everything-claude-code:instinct-import

# PM2 编排 (PM2 orchestration)
/everything-claude-code:pm2

# 包管理器设置 (Package manager setup)
/everything-claude-code:setup-pm

手动安装时，去掉 everything-claude-code: 前缀：/plan、/sessions 等。

钩子运行时控制 (Hook Runtime Controls)

ECC 钩子在智能体生命周期事件时触发。在运行时控制严格程度，无需编辑文件：

# 设置钩子严格级别 (Set hook strictness profile)
export ECC_HOOK_PROFILE=minimal    # 最小侵入 (Least intrusive)
export ECC_HOOK_PROFILE=standard   # 默认 (Default)
export ECC_HOOK_PROFILE=strict     # 最大强制执行 (Maximum enforcement)

# 按 ID 禁用特定钩子（逗号分隔）(Disable specific hooks by ID - comma-separated)
export ECC_DISABLED_HOOKS="pre:bash:tmux-reminder,post:edit:typecheck"

涵盖的钩子事件：SessionStart、Stop、PostEdit、PreBash、PostBash 等。

包管理器检测 (Package Manager Detection)

ECC 按以下优先级链自动检测包管理器：

1. CLAUDE_PACKAGE_MANAGER 环境变量
2. .claude/package-manager.json（项目级）
3. package.json → packageManager 字段
4. 锁文件检测（package-lock.json、yarn.lock、pnpm-lock.yaml、bun.lockb）
5. ~/.claude/package-manager.json（全局）
6. 回退到第一个可用的管理器

# 通过环境变量设置 (Set via environment)
export CLAUDE_PACKAGE_MANAGER=pnpm

# 全局设置 (Set globally)
node scripts/setup-package-manager.js --global pnpm

# 按项目设置 (Set per-project)
node scripts/setup-package-manager.js --project bun

# 检测当前设置 (Detect current setting)
node scripts/setup-package-manager.js --detect

技能系统 (Skills System)

技能是智能体加载以获得领域专业知识的 Markdown 模块。可以单独或批量安装。

使用技能 (Using a Skill)

# 在提示中显式引用技能 (Reference a skill explicitly in your prompt)
"在实现之前，使用搜索优先技能找到正确的缓存方法 (Use the search-first skill to find the right caching approach before implementing)"

# 或通过斜杠命令触发 (Or trigger via slash command)
/everything-claude-code:research "API 响应的内容哈希策略 (content hashing strategies for API responses)"

内置技能 (Notable Built-in Skills)

编写自定义技能 (Writing a Custom Skill)

创建 skills/my-skill.md：

---
name: my-skill
description: 此技能的作用 (What this skill does)
triggers:
  - "激活此技能的短语 (phrase that activates this skill)"
---

# 我的技能 (My Skill)

## 何时使用 (When to Use)
...

## 模式 (Pattern)
\`\`\`typescript
// 具体示例 (concrete example)
\`\`\`

## 规则 (Rules)
- 规则一 (Rule one)
- 规则二 (Rule two)

直觉系统（持续学习）(Instincts System – Continuous Learning)

直觉是从会话中提取并保存以供重复使用的模式。它们带有置信度分数，并随时间演变。

导出直觉 (Export an Instinct)

/everything-claude-code:instinct-import

直觉文件格式 (Instinct File Format)

---
name: prefer-zod-for-validation
confidence: 0.92
extracted_from: session-2026-02-14
---

# 行动 (Action)
在 TypeScript 项目中始终使用 Zod 进行运行时模式验证。
(Always use Zod for runtime schema validation in TypeScript projects.)

# 证据 (Evidence)
在会话期间捕获了 3 个 TypeScript 单独遗漏的运行时类型错误。
(Caught 3 runtime type errors that TypeScript alone missed during session.)

# 示例 (Examples)
\`\`\`typescript
import { z } from 'zod'

const UserSchema = z.object({
  id: z.string().uuid(),
  email: z.string().email(),
  role: z.enum(['admin', 'user'])
})

type User = z.infer<typeof UserSchema>
\`\`\`

规则架构 (Rules Architecture)

规则按语言强制执行编码标准。只安装你的技术栈需要的规则。

# TypeScript + Python
./install.sh typescript python

# 检查已安装的内容 (Check what's installed)
ls .claude/rules/

规则目录布局 (Rule Directory Layout)

rules/
├── common/         # 适用于所有语言 (Applies to all languages)
│   ├── research-first.md
│   ├── security-baseline.md
│   └── verification-loops.md
├── typescript/
│   ├── no-any.md
│   ├── zod-validation.md
│   └── strict-mode.md
├── python/
│   ├── type-hints.md
│   └── django-patterns.md
└── golang/
    └── error-wrapping.md

智能体（子智能体委派）(Agents – Subagent Delegation)

智能体是编排器委派的专门角色：

# 在提示中显式引用智能体 (In your prompt, reference an agent explicitly)
"将架构决策委派给架构师智能体 (Delegate architecture decisions to the architect agent)"
"使用规划器智能体将此功能分解为任务 (Use the planner agent to break this feature into tasks)"

可用智能体包括：planner、architect、researcher、verifier、security-auditor 等。每个都位于 agents/<name>.md 中，带有自己的系统提示、工具列表和约束条件。

AgentShield 安全扫描 (AgentShield Security Scanning)

直接从 Claude Code 运行安全扫描：

/everything-claude-code:security-scan

这将针对你的代码库调用 AgentShield 扫描器（1282 个测试，102 条规则），并发现：

• 硬编码的密钥 (Hardcoded secrets)
• 注入漏洞 (Injection vulnerabilities)
• 不安全的依赖 (Insecure dependencies)
• 智能体提示注入模式 (Agent prompt injection patterns)

记忆持久化钩子 (Memory Persistence Hooks)

ECC 钩子自动保存和恢复会话上下文：

// hooks/session-start.js — 在新会话中加载之前的上下文 (loads prior context on new session)
const fs = require('fs')
const path = require('path')

const memoryPath = path.join(process.env.HOME, '.claude', 'session-memory.json')

if (fs.existsSync(memoryPath)) {
  const memory = JSON.parse(fs.readFileSync(memoryPath, 'utf8'))
  console.log('已恢复会话上下文：', memory.summary)
}

// hooks/stop.js — 退出时保存会话摘要 (saves session summary on exit)
const summary = {
  timestamp: new Date().toISOString(),
  summary: process.env.ECC_SESSION_SUMMARY || '',
  skills_used: (process.env.ECC_SKILLS_USED || '').split(',')
}

fs.writeFileSync(memoryPath, JSON.stringify(summary, null, 2))

跨平台支持 (Cross-Platform Support)

Full (via --target antigravity installer flag)

常见模式 (Common Patterns)

研究优先的开发 (Research-First Development)

"在实现支付 webhook 处理程序之前，使用搜索优先技能验证当前 Stripe webhook 验证的最佳实践。"
(Before implementing the payment webhook handler, use the search-first skill to 
verify current Stripe webhook verification best practices.)

Token 优化 (Token Optimization)

# 将简单任务路由到更便宜的模型 (Route to cheaper model for simple tasks)
/everything-claude-code:model-route "为此纯函数编写单元测试 (Write a unit test for this pure function)"

# 对长时间分析使用后台进程 (Use background processes for long analysis)
/everything-claude-code:harness-audit

使用 Git Worktrees 进行并行化 (Parallelization with Git Worktrees)

# 为并行的智能体任务创建隔离的工作树 (Create isolated worktrees for parallel agent tasks)
git worktree add ../feature-auth -b feature/auth
git worktree add ../feature-payments -b feature/payments

# 每个 Claude Code 会话在自己的工作树中运行 (Each Claude Code session operates in its own worktree)
# 两者都完成后合并 (Merge when both complete)

验证循环 (Verification Loop)

/everything-claude-code:loop-start    # 开始跟踪循环 (Begin tracked loop)
# ... 智能体执行工作 (agent does work) ...
/everything-claude-code:loop-status   # 检查进度 (Check progress)
/everything-claude-code:quality-gate  # 在合并前强制执行通过标准 (Enforce pass criteria before merge)

故障排除 (Troubleshooting)

安装后插件命令未找到 (Plugin commands not found after install)

/plugin list everything-claude-code@everything-claude-code
# 如果为空，重新运行： (If empty, re-run:) /plugin install everything-claude-code@everything-claude-code

规则未应用 (Rules not applied)

# 规则需要手动安装 — 插件系统无法分发它们 (Rules require manual install — plugin system cannot distribute them)
cd everything-claude-code && ./install.sh typescript
# 验证 (Verify):
ls ~/.claude/rules/   # 或项目根目录下的 .claude/rules/ (or .claude/rules/ in project root)

钩子未触发 (Hooks not firing)

# 检查配置文件 (Check profile setting)
echo $ECC_HOOK_PROFILE
# 检查禁用列表 (Check disabled list)
echo $ECC_DISABLED_HOOKS
# 重置为默认值 (Reset to defaults)
unset ECC_HOOK_PROFILE
unset ECC_DISABLED_HOOKS

直觉导入内容丢失 (Instinct import drops content)
确保你使用的是 v1.4.1+ 版本。早期版本有一个错误，parse_instinct_file() 会静默丢弃 Action/Evidence/Examples 部分。拉取最新版本并重新运行。

使用了错误的包管理器 (Wrong package manager used)

node scripts/setup-package-manager.js --detect
export CLAUDE_PACKAGE_MANAGER=pnpm   # 显式覆盖 (Override explicitly)

资源 (Resources)

• 主页 (Homepage): https://ecc.tools
• GitHub: https://github.com/affaan-m/everything-claude-code
• GitHub App（市场）(GitHub App – Marketplace): https://github.com/marketplace/ecc-tools
• npm（通用）(npm – universal): ecc-universal
• npm（安全）(npm – security): ecc-agentshield
• 简略指南 (Shorthand Guide): https://x.com/affaanmustafa/status/2012378465664745795
• 详细指南 (Longform Guide): https://x.com/affaanmustafa/status/2014040193557471352