From 4c02f6cd22f5e27ed9f3b958a52e53284e847bca Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?AI=E4=B8=8D=E6=AD=A2=E8=AF=AD?= <12096460+jnMetaCode@users.noreply.github.com> Date: Tue, 29 Sep 2026 07:20:38 +0800 Subject: [PATCH 1/5] =?UTF-8?q?docs(cheatsheet):=20=E6=8C=89=202026-09=20?= =?UTF-8?q?=E5=AE=9E=E5=86=B5=E6=9B=B4=E6=96=B0=E8=83=BD=E5=8A=9B=E7=9F=A9?= =?UTF-8?q?=E9=98=B5=E3=80=81=E9=85=8D=E7=BD=AE=E6=96=87=E4=BB=B6=E3=80=81?= =?UTF-8?q?=E5=91=BD=E4=BB=A4=E9=80=9F=E6=9F=A5?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- cheatsheet.md | 136 +++++++++++++++++++++++++++----------------------- 1 file changed, 73 insertions(+), 63 deletions(-) diff --git a/cheatsheet.md b/cheatsheet.md index 1f21fc3..d42e454 100644 --- a/cheatsheet.md +++ b/cheatsheet.md @@ -2,7 +2,7 @@ > 一页看完所有工具的关键参数、命令、快捷键。模型和定价变化快，具体以各官网为准。 > -> 信息截止：**2026-04** +> 信息截止：**2026-09** --- @@ -10,37 +10,37 @@ | 你的诉求 | 首选 | 备选 | |---------|------|------| -| 在 IDE 里按 Tab 补全 | **Cursor** | Copilot / Windsurf / Trae | +| 在 IDE 里按 Tab 补全 | **Cursor** | Copilot / Devin Desktop（原 Windsurf）/ Trae | | 终端里跑 Agent 做复杂任务 | **Claude Code** | **Codex CLI** / Aider / Gemini CLI | | 已订阅 ChatGPT，想顺手用 | **Codex CLI** | Cursor 接 GPT | -| 超大代码库一次性分析 | **Gemini CLI**（2M 上下文） | Aider + Map 模式 | -| 预算紧 / 零成本 | **Trae**（免费）或 **Gemini CLI**（免费额度） | Aider + 本地模型；或 `codex --oss --local-provider ollama` | +| 超大代码库一次性分析 | **Claude Code**（Opus/Sonnet 5.5 均 1M 上下文） | Gemini CLI（1M，个人用户需改用 Antigravity CLI）/ Aider + Map 模式 | +| 预算紧 / 零成本 | **Trae**（免费版，仅 Auto 模式）或 **Aider + 本地模型** | `codex --oss --local-provider ollama`；Copilot / Kiro 免费档 | | 国内直连，不用 VPN | **Trae** | OpenClaw + 本地模型 | | 团队协作，规格驱动 | **Kiro**（Spec） | Claude Code + plan mode | | Git 原生，多模型切换 | **Aider** | — | | AI Agent 自动化（非纯编程） | **OpenClaw** | — | -| 只有 VS Code、不想装新东西 | **Copilot** | Cursor/Windsurf/Trae 都是 VS Code 分叉 | +| 只有 VS Code、不想装新东西 | **Copilot** | Cursor / Devin Desktop / Trae 都是 VS Code 分叉 | --- ## 二、能力矩阵 -| 维度 | Claude Code | Codex CLI | Cursor | Copilot | Windsurf | Gemini CLI | Kiro | Aider | Trae | OpenClaw | +| 维度 | Claude Code | Codex CLI | Cursor | Copilot | Devin Desktop（原 Windsurf） | Gemini CLI | Kiro | Aider | Trae | OpenClaw | |------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| | **类型** | CLI | CLI | IDE | IDE 插件 | IDE | CLI | IDE | CLI | IDE | Agent 框架 | -| **出品方** | Anthropic | OpenAI | Cursor | GitHub | Codeium | Google | AWS | 开源 | 字节 | 开源 | +| **出品方** | Anthropic | OpenAI | Cursor（SpaceX 旗下） | GitHub | Cognition | Google | AWS | 开源 | 字节 | OpenClaw 基金会 | | **Tab 补全** | — | — | ★★★ | ★★★ | ★★★ | — | ★★ | — | ★★ | — | | **Agent 执行** | ★★★ | ★★★ | ★★★ | ★★★ | ★★★ | ★★ | ★★ | ★★ | ★★ | ★★★ | -| **终端内运行** | ★★★ | ★★★ | ★ | — | ★ | ★★★ | — | ★★★ | — | ★★★ | -| **上下文窗口** | 200K | 跟模型 | 跟模型 | 跟模型 | 跟模型 | **2M** | 跟模型 | 跟模型 | 跟模型 | 跟模型 | -| **MCP 支持** | ✅ | ✅ | ✅ | ✅ | ✅ | 扩展机制 | — | — | — | 原生 | -| **Hook 自动化** | ✅ | ✅（beta，复用 Claude schema） | — | — | — | — | ✅ | ✅（lint/test） | — | ✅（Cron） | -| **Subagent** | ✅ | ✅（TOML） | — | — | — | — | — | — | — | ✅（Agent 空间） | -| **沙箱机制** | App 层 + Hook | OS 内核（Seatbelt/Landlock） | — | — | — | — | — | — | — | — | +| **终端内运行** | ★★★ | ★★★ | ★ | ★★（Copilot CLI） | ★ | ★★★ | ★★（Kiro CLI） | ★★★ | — | ★★★ | +| **上下文窗口** | **1M**（Haiku 200K） | 跟模型 | 跟模型 | 跟模型 | 跟模型 | **1M** | 跟模型 | 跟模型 | 跟模型 | 跟模型 | +| **MCP 支持** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | — | ✅ | 原生 | +| **Hook 自动化** | ✅（30+ 事件） | ✅（12 事件，默认开启） | ✅ | ✅ | ✅ | ✅ | ✅ | ✅（lint/test） | — | ✅（Cron） | +| **Subagent** | ✅ | ✅（TOML） | ✅ | ✅ | ✅ | ✅ | ✅ | — | ✅（自定义 Agent） | ✅（Agent 空间） | +| **沙箱机制** | OS 级 Bash 沙箱（Seatbelt/bubblewrap） | OS 级（Seatbelt/bubblewrap） | — | — | — | — | — | — | — | — | | **多模型切换** | ★（仅 Claude） | ★（仅 OpenAI） | ★★★ | ★★ | ★★★ | ★（仅 Gemini） | ★★ | ★★★（几乎所有 LLM） | ★★ | ★★★ | -| **开源** | — | ✅ Apache-2.0 | — | — | — | — | — | ✅ MIT | — | ✅ MIT | -| **国内直连** | — | — | — | — | — | — | — | — | ✅ | ✅（需配本地模型） | -| **定价模式** | API / Pro 订阅 | ChatGPT 订阅 / API | 免费 / $20 Pro | 免费 / $10 Pro | 免费 / 订阅 | 慷慨免费额度 | 预览免费 | 按你选的 LLM | 免费（有额度） | 开源免费 | +| **开源** | — | ✅ Apache-2.0 | — | — | — | ✅ Apache-2.0 | — | ✅ Apache-2.0 | — | ✅ MIT | +| **国内直连** | — | — | — | — | — | — | — | — | ✅（国内版 trae.cn） | ✅（需配本地模型） | +| **定价模式** | Pro/Max 订阅 / API | ChatGPT 订阅 / API | 免费 / $20 Pro | 免费 / $10 Pro 起（AI Credits） | 免费 / $20 Pro | 企业版 / 付费 API Key（个人免费已停） | 免费 50 credits / $20 Pro | 按你选的 LLM | 免费（Auto）/ $20 Pro | 开源免费 | --- @@ -50,15 +50,15 @@ | 工具 | 主配置文件 | 位置 | 备注 | |------|-----------|------|------| -| Claude Code | `CLAUDE.md` + `.claude/` | 项目根 | 控制在 200 行以内，大项目拆到 `.claude/rules/` | +| Claude Code | `CLAUDE.md` + `.claude/` | 项目根 | 控制在 200 行以内，大项目拆到 `.claude/rules/`（`paths:` 按路径加载）；无 CLAUDE.md 时读 `AGENTS.md` | | Codex CLI | `AGENTS.md` + `.codex/config.toml` | 项目根 | `~/.codex/AGENTS.md` 全局；子目录 AGENTS.md 覆盖父级 | -| Cursor | `.cursor/rules/*.md` | 项目根 | 支持 `globs` 按文件类型加载 | -| Copilot | `.github/copilot-instructions.md` | 项目根 | 同目录 `agents/` `chatModes/` | -| Windsurf | `.windsurfrules` | 项目根 | 单文件，不支持拆分 | +| Cursor | `.cursor/rules/*.mdc` | 项目根 | 必须是 `.mdc`（纯 `.md` 会被忽略）；frontmatter `description` / `globs` / `alwaysApply`；也读 `AGENTS.md` | +| Copilot | `.github/copilot-instructions.md` | 项目根 | 另有 `.github/instructions/*.instructions.md`（`applyTo`）、`.github/agents/*.agent.md`；chatModes 已废弃 | +| Devin Desktop（原 Windsurf） | `.devin/rules/*.md` | 项目根 | `trigger: always_on / model_decision / glob / manual`；`.windsurfrules` 仅向后兼容 | | Gemini CLI | `GEMINI.md` | 项目根 | 结构类似 CLAUDE.md | -| Kiro | `.kiro/steering/*.md` | 项目根 | 三种模式：`always` / `globs` / `manual` | +| Kiro | `.kiro/steering/*.md` | 项目根 | `inclusion: always / fileMatch / manual / auto` 四种模式；全局 `~/.kiro/steering/` | | Aider | `.aider.conf.yml` | 项目根 | YAML，含模型/lint/test 配置 | -| Trae | `.trae/rules/project_rules.md` | 项目根 | 支持中文规则 | +| Trae | `.trae/rules/*.md` | 项目根 | 递归读取三层；可导入 `AGENTS.md` / `CLAUDE.md` | | OpenClaw | `~/.openclaw/openclaw.json` | 用户目录 | JSON5，热加载 | --- @@ -69,7 +69,7 @@ ```bash # 安装与启动 -npm install -g @anthropic-ai/claude-code +curl -fsSL https://claude.ai/install.sh | bash # 官方推荐（也可 brew / winget / npm） claude # 进入交互模式 claude --resume # 恢复上次对话 claude --model haiku # 简单任务用便宜模型 @@ -77,7 +77,10 @@ claude -p "任务" --output-format json # headless 模式 # 交互中 /compact # 压缩上下文 -/plan # 进入 plan 模式 +/context # 看上下文占用 +/effort # 调思考深度（low → max） +Shift+Tab # 切换权限模式（含 plan / auto） +/code-review # 审查当前改动 Esc # 打断当前生成 ``` @@ -85,15 +88,15 @@ Esc # 打断当前生成 ```bash # 安装与启动 -npm install -g @openai/codex +curl -fsSL https://chatgpt.com/codex/install.sh | sh # 官方推荐（也可 npm / brew） codex # 进入 TUI（首次会引导登录 ChatGPT） -codex --sandbox workspace-write # 默认搭配（v0.125.0 起 `--full-auto` 已废弃，用此替代） +codex --sandbox workspace-write # 默认搭配（`--full-auto` 已移除） codex --sandbox read-only # 只读探索 codex --add-dir ../sibling-repo # 不放开沙箱、只多加可写目录 codex --yolo # 跳过沙箱+审批（仅在外部已隔离的环境用） codex exec --json "..." | jq -c . # JSONL 输出给后续脚本 -codex --oss --local-provider ollama -m qwen2.5-coder # 本地零成本 -codex mcp-server # 把 Codex 暴露给其他 Agent 当工具 +codex --oss --local-provider ollama -m qwen3-coder # 本地零成本 +codex app-server # 供其他程序集成（原 mcp-server 已移除） codex resume # 恢复上次对话 # 交互中 @@ -102,7 +105,8 @@ codex resume # 恢复上次对话 /model # 切换模型 /review # 审查 diff/分支/commit /compact # 压缩上下文 -/agent # 切换 subagent thread +/multi-agents # 管理 subagent（别名 /subagents） +/import # 从 Claude Code / Cursor 迁移配置 /diff # 看 git diff（含未跟踪） /debug-config # 排查 config.toml 不生效 ``` @@ -111,14 +115,14 @@ codex resume # 恢复上次对话 ``` Tab — 接受补全 -Cmd+L — 打开 Chat（选中代码自动带入） -Cmd+I — 打开 Composer（Agent 模式） +Cmd+I / Cmd+L — 打开/关闭 Agent 侧栏（选中代码自动带入） +Shift+Tab — 在 Agent / Ask / Plan 模式间切换 +Cmd+E — 切换 Agent 布局 Cmd+K — 行内编辑 -Cmd+Shift+L — 把当前文件加入 Chat 上下文 Esc — 拒绝补全 -@file @folder @web @terminal — 引用 -@notepad:名称 — 引用 Notepad +@Files @Folders @Terminals @Chats @Branch @Browser — 引用 +（Notepads 已在 2.0 移除，改用 rules / skills） ``` ### GitHub Copilot @@ -126,51 +130,57 @@ Esc — 拒绝补全 ``` Tab — 接受补全 Esc — 拒绝补全 -Cmd+Shift+I — 打开 Chat -Cmd+I — 行内编辑 +Ctrl+Cmd+I — 打开 Chat 视图 +Shift+Cmd+I — 以 Agent 模式打开 Chat +Cmd+I — 行内 Chat Alt+] / Alt+[ — 切换补全建议 #file #selection #terminal #problems — 引用 -@workspace — 全项目上下文 +#codebase — 强制语义检索全项目（Agent 会自动搜索） +/agents — 选择自定义 Agent（.github/agents/） ``` -### Windsurf +### Devin Desktop（原 Windsurf） ``` -Write 模式 — 直接写代码（类 Composer） -Chat 模式 — 对话问答 +Code 模式（默认）— 直接改代码 +Plan 模式 — 先出方案 +Ask 模式 — 只读问答 +Cmd+. — 切换模式 -@file @folder @web — 引用 -Cascade 会自动追踪你的编辑流，不用手动贴上下文 +本地 Agent 已从 Cascade 换成 Devin Local（支持 subagent） +/工作流名 — 运行 .devin/workflows/ 里的工作流 ``` ### Gemini CLI ```bash -npm install -g @google/gemini-cli +# ⚠️ 2026-06-18 起个人/免费用户停止服务，个人请改用 Antigravity CLI；企业版与付费 API Key 仍可用 +npm install -g @google/gemini-cli # 或 brew install gemini-cli gemini # 进入交互 -# 利用 2M 上下文做大代码库分析（不适合小任务） +# 1M 上下文适合大代码库分析（不适合小任务） ``` ### Kiro ``` -1. 描述需求 → 2. Kiro 生成 Spec → 3. 审查确认 → 4. 自动实现 + 测试 +1. 描述需求 → 2. Kiro 生成 Spec（requirements.md + design.md + tasks.md）→ 3. 审查确认 → 4. 按 tasks 逐项实现 -Steering 加载模式： -- always 每次对话都加载 -- globs: ["*.java"] 按文件匹配 -- manual 手动激活 +Steering 加载模式（frontmatter `inclusion:`）： +- always 每次对话都加载 +- fileMatch + fileMatchPattern: "**/*.java" 按文件匹配 +- manual 手动激活 +- auto 按 description 自动判断 ``` ### Aider ```bash # 安装与启动 -pip install aider-chat -aider --model claude-sonnet-4-5 -aider --model deepseek/deepseek-chat # 便宜 -aider --model ollama/qwen2.5-coder # 本地免费 +python -m pip install aider-install && aider-install +aider --model sonnet # 内置别名，指向较新的 Claude Sonnet +aider --model deepseek/deepseek-v4-pro # 便宜（deepseek-chat 已于 2026-07 退役） +aider --model ollama/qwen3-coder # 本地免费 # 交互中 /add file1 file2 — 添加文件到上下文 @@ -183,11 +193,11 @@ aider --model ollama/qwen2.5-coder # 本地免费 ### Trae ``` -Builder 模式 — Agent，跨文件修改 +Agent 模式 — 跨文件修改（可自建 Agent：提示词 + 工具 + MCP） Chat 模式 — 对话 @file @folder @web — 引用 -免费模型：Claude / GPT，按任务复杂度自选 +免费版仅 Auto 模式；国际版已于 2025-09 下架 Claude 模型 ``` ### OpenClaw @@ -197,9 +207,9 @@ openclaw onboard # 初始化 openclaw gateway start # 启动网关 openclaw doctor # 诊断 -openclaw skills install # 安装 Skill -openclaw cron add "0 9 * * *" "..." # 定时任务 -openclaw models set default # 切换模型 +openclaw skills install @owner/ # 安装 ClawHub Skill +openclaw automations create ... # 定时任务（cron 仍是别名） +openclaw models set anthropic/claude-sonnet-5-5 # 切换模型（provider/model） openclaw channels add telegram # 添加消息频道 ``` @@ -211,14 +221,14 @@ openclaw channels add telegram # 添加消息频道 ┌─ 主要在终端工作？ │ ├─ 要最强 Agent 能力 / 大型重构 → Claude Code │ ├─ 已订阅 ChatGPT / 想要内核级沙箱 → Codex CLI -│ ├─ 要超大上下文 / 免费 → Gemini CLI +│ ├─ Google 生态 / 企业授权 → Gemini CLI（个人用户 → Antigravity CLI） │ └─ 要多模型灵活切换 / Git 原生 → Aider │ ├─ 主要在 IDE 里？ │ ├─ 不想装新 IDE → GitHub Copilot（VS Code/JetBrains 插件） │ ├─ 愿意换 IDE，预算足 → Cursor -│ ├─ 喜欢 AI 主动帮忙 → Windsurf -│ ├─ 要中文 / 国内网络 / 免费 → Trae +│ ├─ 喜欢 AI 主动帮忙 / 本地+云端 Agent 看板 → Devin Desktop（原 Windsurf） +│ ├─ 要中文 / 国内网络 → Trae（国内版 trae.cn） │ └─ 团队协作 / 规格驱动 → Kiro │ └─ 做编程以外的 AI 自动化？ @@ -235,8 +245,8 @@ openclaw channels add telegram # 添加消息频道 | **Codex CLI + Cursor** | ChatGPT 订阅者顺手用：CLI 做 Agent、IDE 做补全 | | **Codex CLI + Claude Code** | 双 CLI 互补：Codex 跑 CI/脚本、Claude Code 做大重构 | | **Claude Code + Copilot** | 纯 VS Code 用户的轻量选择 | -| **Gemini CLI + Cursor** | 预算敏感：2M 免费分析 + 20$ IDE | -| **Aider + 本地 LLM** | 零 API 成本：`ollama/qwen2.5-coder` + Aider | +| **Copilot 免费版 + Aider 本地模型** | 预算敏感：IDE 补全 + 终端 Agent 都零订阅 | +| **Aider + 本地 LLM** | 零 API 成本：`ollama/qwen3-coder` + Aider | | **Codex CLI --oss + Ollama** | 零 API 成本但要 Codex 的 Agent 体验：内核级沙箱 + 本地模型 | | **Claude Code + OpenClaw** | 编程 + 自动化：CC 写代码，OpenClaw 跑定时任务 | From 3aa00c9b54d43263d9de78bf523f28a81855cf19 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?AI=E4=B8=8D=E6=AD=A2=E8=AF=AD?= <12096460+jnMetaCode@users.noreply.github.com> Date: Tue, 29 Sep 2026 07:21:10 +0800 Subject: [PATCH 2/5] =?UTF-8?q?docs(ecosystem):=20=E5=90=8C=E6=AD=A5?= =?UTF-8?q?=E5=A7=90=E5=A6=B9=E9=A1=B9=E7=9B=AE=E6=9C=80=E6=96=B0=E5=AE=89?= =?UTF-8?q?=E8=A3=85=E6=96=B9=E5=BC=8F=E4=B8=8E=E6=95=B0=E5=AD=97=EF=BC=88?= =?UTF-8?q?shellward=20MCP/SDK=E3=80=81superpowers-zh=E3=80=81agency-agent?= =?UTF-8?q?s-zh=20=E8=B7=AF=E5=BE=84=E3=80=81ao=20demo=EF=BC=89?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- ecosystem.md | 44 ++++++++++++++++++++------------------------ 1 file changed, 20 insertions(+), 24 deletions(-) diff --git a/ecosystem.md b/ecosystem.md index 8bd73bf..ffa92ff 100644 --- a/ecosystem.md +++ b/ecosystem.md @@ -15,7 +15,7 @@ | 依赖 | 最低版本 | 检查命令 | 安装方式 | |------|---------|---------|---------| -| **Node.js** | 18+ | `node -v` | [nodejs.org](https://nodejs.org) | +| **Node.js** | 20+ | `node -v` | [nodejs.org](https://nodejs.org) | | **npm** | 9+ | `npm -v` | 随 Node.js 安装 | | **Git** | 2.0+ | `git -v` | [git-scm.com](https://git-scm.com) | | **AI 编程工具** | 任意一个 | — | 至少装一个：Claude Code / Cursor / Copilot / Windsurf 等 | @@ -38,7 +38,7 @@ cd /your/project npx superpowers-zh ``` -安装器会自动检测你项目中使用的工具（Claude Code / Cursor / Copilot / Windsurf 等 14 种），将 20 个 skills 安装到正确位置。 +安装器会自动检测你项目中使用的工具（Claude Code / Codex CLI / Cursor / Kiro / Gemini CLI 等 26 种），将 21 个 skills 安装到正确位置。 ### 验证安装成功 @@ -47,7 +47,7 @@ npx superpowers-zh ```bash # Claude Code 用户 ls .claude/skills/ -# 应该看到 brainstorming.md、systematic-debugging.md 等 20 个文件 +# 应该看到 brainstorming/、systematic-debugging/ 等 21 个 skill 目录（每个目录里一个 SKILL.md） # Cursor 用户 ls .cursor/skills/ @@ -97,7 +97,7 @@ AI：在开始实现之前，我需要了解几个关键问题： | 问题 | 解决 | |------|------| -| `npx superpowers-zh` 报错 | 检查 Node.js 版本（需要 18+）：`node -v` | +| `npx superpowers-zh` 报错 | 检查 Node.js 版本（需要 20+）：`node -v` | | 安装后 AI 行为没变化 | 重启 AI 工具（关掉再打开），让它重新加载 skills | | 想卸载 | 删除对应目录即可，如 `rm -rf .claude/skills/` | @@ -134,7 +134,7 @@ cd agency-agents-zh ```bash # Claude Code 用户 —— 检查是否有角色文件 -ls ~/.claude/skills/ | head -10 +ls ~/.claude/agents/ | head -10 # 应该看到 engineering-security-engineer.md 等文件 # 或者直接问 AI： @@ -210,9 +210,11 @@ AI：（激活 database-optimizer 角色） # 全局安装（推荐，之后可以直接用 ao 命令） npm install -g agency-orchestrator -# 初始化：下载 276 个角色定义到本地 +# 零配置先体验（不需要 init） +ao demo + +# （可选）把 276 个中文角色复制到本地，方便自己修改 ao init -# 角色文件会下载到 ~/.ao/roles/ 目录 ``` ### 快速体验（无需 API Key） @@ -349,7 +351,7 @@ npm install shellward | 防护 | 说明 | |------|------| -| 提示词注入检测 | 32 条规则（18 条中文 + 14 条英文），风险评分 | +| 提示词注入检测 | 37 条规则（20 条中文 + 17 条英文），风险评分；另含 MCP 工具投毒 / rug-pull 检测 | | 危险命令拦截 | `rm -rf`、反弹 shell、fork bomb、`chmod 777` 等 | | PII 检测 | 身份证、银行卡、手机号、SSN、信用卡、API Key、JWT | | 数据外泄链检测 | 读取敏感数据 → 发邮件/HTTP POST/curl = 拦截 | @@ -363,24 +365,17 @@ npm install shellward 适用于 Claude Desktop / Cursor / Claude Code 等支持 MCP 的工具。 -把以下配置加到你的 MCP 配置文件里： - -```bash -# 先找到 shellward 的安装路径 -npm list -g shellward # 全局安装的路径 -# 或 -ls node_modules/shellward/src/mcp-server.ts # 项目内安装的路径 -``` +把以下配置加到你的 MCP 配置文件里（`npx` 会自动拉取 `shellward-mcp`，不用找安装路径）： ```json // Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json // Cursor: .cursor/mcp.json -// Claude Code: .claude/settings.json 的 mcpServers 部分 +// Claude Code: 项目根目录 .mcp.json（或 claude mcp add） { "mcpServers": { "shellward": { "command": "npx", - "args": ["shellward", "--mcp"] + "args": ["-y", "-p", "shellward", "shellward-mcp"] } } } @@ -391,14 +386,15 @@ ls node_modules/shellward/src/mcp-server.ts # 项目内安装的路径 3 行代码集成到任何 AI Agent 框架（LangChain、AutoGPT 等）： ```javascript -const { ShellWard } = require('shellward'); -const guard = new ShellWard(); +import { ShellWard } from 'shellward'; +const guard = new ShellWard({ mode: 'enforce' }); -// 在 Agent 执行工具调用之前检查 -const result = guard.check(toolCall); +// 在 Agent 执行命令之前检查 +const result = guard.checkCommand(cmd); if (!result.allowed) { console.log(`已拦截：${result.reason}`); } +// 另有 checkInjection / scanData / checkOutbound ``` ### 效果对比 @@ -524,13 +520,13 @@ npm install shellward 然后配置 MCP，让 AI 工具在执行操作时经过 shellward 检查： ```json -// 加到 .claude/settings.json（Claude Code） +// 加到 .mcp.json（Claude Code，项目根目录） // 或 .cursor/mcp.json（Cursor） { "mcpServers": { "shellward": { "command": "npx", - "args": ["shellward", "--mcp"] + "args": ["-y", "-p", "shellward", "shellward-mcp"] } } } From 45fd82380f71bf943b1ca241d1a49301d8f15708 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?AI=E4=B8=8D=E6=AD=A2=E8=AF=AD?= <12096460+jnMetaCode@users.noreply.github.com> Date: Tue, 29 Sep 2026 07:21:49 +0800 Subject: [PATCH 3/5] =?UTF-8?q?docs(resources):=20=E6=9B=B4=E6=96=B0?= =?UTF-8?q?=E6=98=9F=E6=95=B0=E3=80=81=E6=94=B9=E5=90=8D/=E5=BD=92?= =?UTF-8?q?=E6=A1=A3=E4=BB=93=E5=BA=93=EF=BC=8C=E5=90=88=E5=B9=B6=20Datawh?= =?UTF-8?q?ale=20=E9=87=8D=E5=A4=8D=E6=9D=A1=E7=9B=AE?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- resources.en.md | 2 +- resources.md | 44 ++++++++++++++++++++------------------------ 2 files changed, 21 insertions(+), 25 deletions(-) diff --git a/resources.en.md b/resources.en.md index dd66adf..2b76c0e 100644 --- a/resources.en.md +++ b/resources.en.md @@ -127,7 +127,7 @@ - Phodal Huang's book "Building LLM Applications" - AI coding architect's perspective, engineering and team adoption -- [unit-mesh/auto-dev](https://github.com/phodal/auto-dev) · ⭐ 3k+ · ZH/EN +- [phodal/auto-dev](https://github.com/phodal/auto-dev) · ⭐ 3k+ · ZH/EN - **Domestic Chinese JetBrains AI coding plugin** with custom Agent + Chinese DevIns scripts - AI coding option for IDEA/PyCharm ecosystem diff --git a/resources.md b/resources.md index 4ea110f..a08cf8b 100644 --- a/resources.md +++ b/resources.md @@ -2,7 +2,7 @@ > 本项目讲"10 款工具怎么用好"。要继续深入，这里是精选外部资源。**不求全，只求高质量**——每个条目都说清楚"什么情况值得看"。 > -> 信息截止：**2026-04**。链接失效请提 Issue。 +> 信息截止：**2026-09**。链接失效请提 Issue。 --- @@ -25,15 +25,11 @@ - Anthropic 官方 **9 章交互式 Prompt 教程**，可运行 Jupyter - 系统入门 Prompt 工程最快的路径，2-3 小时能跑完 -- [dair-ai/Prompt-Engineering-Guide](https://github.com/dair-ai/Prompt-Engineering-Guide) · ⭐ 60k+ · 英（有中文翻译） +- [dair-ai/Prompt-Engineering-Guide](https://github.com/dair-ai/Prompt-Engineering-Guide) · ⭐ 75k+ · 英（有中文翻译） - Prompt 工程百科全书：CoT、ReAct、RAG、Tool-use 各种 pattern - 想深入原理、看各种技术名词对应什么场景时翻 -- [datawhalechina/prompt-engineering-for-developers](https://github.com/datawhalechina/prompt-engineering-for-developers) · ⭐ 13k+ · 中 - - Datawhale 翻译的**吴恩达 Prompt 工程系列**+ 中文注解 - - 中文用户看视频 + 跟代码的最佳组合 - -- [phodal/prompt-patterns](https://github.com/phodal/prompt-patterns) · ⭐ 1k+ · 中 +- [phodal/prompt-patterns](https://github.com/phodal/prompt-patterns) · ⭐ 3k+ · 中 - 国内知名架构师黄峰达写的 Prompt 模式与 DSL 设计指南 - 写复杂 AI 编程工作流、设计可复用 Prompt 时的进阶参考 @@ -41,9 +37,9 @@ ## 🛠️ 工具专项 Awesome 列表 -- [PatrickJS/awesome-cursorrules](https://github.com/PatrickJS/awesome-cursorrules) · ⭐ 30k+ · 英 - - 社区收集的 `.cursorrules` 文件大全，按技术栈分类 - - 新项目起手抄一份对应技术栈的 rules +- [PatrickJS/awesome-cursorrules](https://github.com/PatrickJS/awesome-cursorrules) · ⭐ 40k+ · 英 + - 社区收集的 Cursor 规则文件大全，按技术栈分类 + - 新项目起手抄一份对应技术栈的 rules（注意新版 Cursor 要放进 `.cursor/rules/*.mdc`） - [hesreallyhim/awesome-claude-code](https://github.com/hesreallyhim/awesome-claude-code) · ⭐ 高 · 英 - Claude Code 社区最全资源：技巧、Skills、Hooks、工作流 @@ -51,15 +47,15 @@ - [github/awesome-copilot](https://github.com/github/awesome-copilot) · ⭐ 高 · 英 - GitHub 官方维护的 Copilot 资源集合 - - 查 Copilot 的 instruction / Chat mode / Agent 示例 + - 查 Copilot 的 instructions / 自定义 Agent / Skills 示例 - [addyosmani/gemini-cli-tips](https://github.com/addyosmani/gemini-cli-tips) · ⭐ 中 · 英 - Google 开发者 Addy Osmani 整理的 **30 个 Gemini CLI 技巧** - - Gemini CLI 用户直接抄作业 + - Gemini CLI 用户直接抄作业（注：2025-10 后未更新；个人用户已需迁往 Antigravity CLI） -- [detailobsessed/awesome-windsurf](https://github.com/detailobsessed/awesome-windsurf) · ⭐ 中 · 英 - - Windsurf 社区资源集合（规则、工作流、MCP） - - 用 Windsurf Cascade 找社区模式 +- [detailobsessed/awesome-devin](https://github.com/detailobsessed/awesome-devin) · ⭐ 中 · 英 + - 原 awesome-windsurf，Devin Desktop（原 Windsurf）社区资源集合（规则、工作流、MCP） + - 找 Devin Desktop 的社区规则和工作流 --- @@ -69,9 +65,9 @@ - **官方仓库**，Issues 和 Discussions 是新功能 / 新技巧的一手信息源 - 追 Claude Code 新特性（slash command / hooks / skill 新能力） -- [anthropics/courses](https://github.com/anthropics/courses) · ⭐ 高 · 英 +- [anthropics/courses](https://github.com/anthropics/courses) · ⭐ 高 · 英（已归档） - Anthropic 官方课程合集：Prompt、Tool use、RAG、MCP 全套 - - 想系统吃透 Anthropic 全家桶 + - 仓库已归档不再更新，原理仍适用；新 API 用法以 cookbooks 为准 - [anthropics/claude-cookbooks](https://github.com/anthropics/claude-cookbooks) · ⭐ 高 · 英 - Claude API 实战示例：多模态、工具调用、RAG、Agent 模式 @@ -113,21 +109,21 @@ ## 🇨🇳 中文优秀资源 -- [datawhalechina/llm-cookbook](https://github.com/datawhalechina/llm-cookbook) · ⭐ 22k+ · 中 - - Datawhale 的 LLM 开发中文指南合集 +- [datawhalechina/llm-cookbook](https://github.com/datawhalechina/llm-cookbook) · ⭐ 25k+ · 中 + - Datawhale 的 LLM 开发中文指南合集（含**吴恩达 Prompt 工程系列**中文版，原 prompt-engineering-for-developers 已并入） - 把 LLM 能力嵌入自己项目时的中文参考手册 -- [liaokongVFX/LangChain-Chinese-Getting-Started-Guide](https://github.com/liaokongVFX/LangChain-Chinese-Getting-Started-Guide) · ⭐ 8k+ · 中 +- [liaokongVFX/LangChain-Chinese-Getting-Started-Guide](https://github.com/liaokongVFX/LangChain-Chinese-Getting-Started-Guide) · ⭐ 9k+ · 中 - LangChain 中文入门指南：Agent / Chain / Memory - 自己造 AI 编程工具、Agent 型助手的基础 -- [phodal/aigc](https://github.com/phodal/aigc) · ⭐ 1k+ · 中 +- [phodal/aigc](https://github.com/phodal/aigc) · ⭐ 1.5k+ · 中 - 黄峰达《构筑大语言模型应用》电子书 - AI 编程架构师视角，适合看工程化与团队落地 -- [unit-mesh/auto-dev](https://github.com/phodal/auto-dev) · ⭐ 3k+ · 中英 - - **国产 JetBrains AI 编程插件**，支持自定义 Agent + 中文 DevIns 脚本 - - IDEA/PyCharm 生态里的 AI 编程选择 +- [phodal/auto-dev](https://github.com/phodal/auto-dev) · ⭐ 4k+ · 中英 + - 黄峰达主导的 AI 编程平台，从 JetBrains 插件演进为 Kotlin 多平台多 Agent 平台 + - IDEA/PyCharm 生态里的国产 AI 编程选择 --- From b704f52e8d1499129da8acffad8e2ed341ae8657 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?AI=E4=B8=8D=E6=AD=A2=E8=AF=AD?= <12096460+jnMetaCode@users.noreply.github.com> Date: Tue, 29 Sep 2026 07:21:53 +0800 Subject: [PATCH 4/5] =?UTF-8?q?docs(README):=20=E5=B7=A5=E5=85=B7=E8=A1=A8?= =?UTF-8?q?=E6=8C=89=202026-09=20=E5=AE=9E=E5=86=B5=E6=9B=B4=E6=96=B0?= =?UTF-8?q?=EF=BC=88Devin=20Desktop=20/=20Gemini=20CLI=20=E8=BF=81?= =?UTF-8?q?=E7=A7=BB=20/=20Trae=20/=20Cursor=EF=BC=89?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- README.md | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/README.md b/README.md index dc57759..3aa54bf 100644 --- a/README.md +++ b/README.md @@ -35,7 +35,7 @@ | 后端 / 重构 / 大项目 | [Claude Code](claude-code/) → [需求拆解](common/task-decomposition.md) → [实战场景](workflows/scenarios.md) | | 从 Copilot 迁来 | [Copilot](copilot/) → [Claude Code](claude-code/)（对比差异）→ [多工具选型](workflows/tool-selection.md) | | 已订阅 ChatGPT Plus/Pro | [Codex CLI](codex/) → [Claude Code](claude-code/) 对比 → [多工具选型](workflows/tool-selection.md) | -| 控成本 / 独立开发者 | [速查表](cheatsheet.md) → [Gemini CLI](gemini-cli/) 或 [Aider](aider/) + 本地模型 | +| 控成本 / 独立开发者 | [速查表](cheatsheet.md) → [Aider](aider/) + 本地模型 或 [Codex CLI](codex/) `--oss` | | 团队协作 / 高质量交付 | [Kiro](kiro/) → [代码审查](common/code-review.md) → [测试策略](common/testing.md) | **已经在用了？** 直接看：[速查表](cheatsheet.md) · [进阶技巧](#-10-款工具教程) · [实战工作流](#-实战工作流) · [生态项目](#-相关项目) @@ -48,14 +48,14 @@ |------|------|------| | [**Claude Code**](claude-code/) | CLI Agent | 66 个技巧，Agent + Skill + Hook 完整工作流 | | [**Codex CLI**](codex/) | CLI Agent | OpenAI 官方开源（Rust），Sandbox + AGENTS.md，ChatGPT 订阅可用 | -| [**Cursor**](cursor/) | IDE | .cursorrules 配置，Composer Agent 模式 | -| [**GitHub Copilot**](copilot/) | IDE 插件 | 行内补全 + Agent 模式 + 自定义指令 | -| [**OpenClaw**](openclaw/) | AI Agent 框架 | 338k Stars，多平台连接 + Skills + Cron 自动化 | -| [Windsurf](windsurf/) | IDE | Cascade Agent，自动上下文 | -| [Gemini CLI](gemini-cli/) | CLI | Google 出品，大代码库分析 | -| [Kiro](kiro/) | IDE | AWS 出品，Spec 驱动开发 | +| [**Cursor**](cursor/) | IDE | `.cursor/rules/*.mdc` 规则，Agent / Plan 模式，并行 Agents Window | +| [**GitHub Copilot**](copilot/) | IDE 插件 | 行内补全 + Agent 模式 + 自定义 Agent / 指令 | +| [**OpenClaw**](openclaw/) | AI Agent 框架 | 39 万+ Stars，多平台连接 + Skills + 定时自动化 | +| [Devin Desktop（原 Windsurf）](windsurf/) | IDE | Devin Local Agent，本地 + 云端 Agent 看板 | +| [Gemini CLI](gemini-cli/) | CLI | Google 出品，1M 上下文；个人用户已转向 Antigravity CLI | +| [Kiro](kiro/) | IDE + CLI | AWS 出品，Spec 驱动开发 | | [Aider](aider/) | CLI | Git 原生，支持几乎所有 LLM | -| [Trae](trae/) | IDE | 字节出品，免费 Claude/GPT，国内直连 | +| [Trae](trae/) | IDE | 字节出品，有免费档，国内版直连 | > 每个工具都有：**核心概念 → 快速上手 → 提示词技巧 → 进阶用法 → 配置模板** From 16480a3ce2f9df24d1bf7e96a8ed8188ed142029 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?AI=E4=B8=8D=E6=AD=A2=E8=AF=AD?= <12096460+jnMetaCode@users.noreply.github.com> Date: Tue, 29 Sep 2026 07:23:26 +0800 Subject: [PATCH 5/5] =?UTF-8?q?docs:=20=E5=90=8C=E6=AD=A5=E8=8B=B1?= =?UTF-8?q?=E6=96=87=E7=89=88=EF=BC=88cheatsheet=20/=20ecosystem=20/=20res?= =?UTF-8?q?ources=20/=20README=EF=BC=89?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- README.en.md | 16 +++--- cheatsheet.en.md | 136 +++++++++++++++++++++++++---------------------- ecosystem.en.md | 44 +++++++-------- resources.en.md | 44 +++++++-------- 4 files changed, 121 insertions(+), 119 deletions(-) diff --git a/README.en.md b/README.en.md index 7d7bec0..05b4536 100644 --- a/README.en.md +++ b/README.en.md @@ -44,7 +44,7 @@ | Backend / refactoring / big projects | [Claude Code](claude-code/) → [Task Decomposition](common/task-decomposition.en.md) → [Scenarios](workflows/scenarios.en.md) | | Migrating from Copilot | [Copilot](copilot/) → [Claude Code](claude-code/) (compare) → [Tool Selection](workflows/tool-selection.en.md) | | Already on ChatGPT Plus/Pro | [Codex CLI](codex/) → [Claude Code](claude-code/) (compare) → [Tool Selection](workflows/tool-selection.en.md) | -| Cost-conscious / solo dev | [Cheatsheet](cheatsheet.en.md) → [Gemini CLI](gemini-cli/) or [Aider](aider/) + local models | +| Cost-conscious / solo dev | [Cheatsheet](cheatsheet.en.md) → [Aider](aider/) + local models or [Codex CLI](codex/) `--oss` | | Team / high-quality delivery | [Kiro](kiro/) → [Code Review](common/code-review.en.md) → [Testing](common/testing.en.md) | **Already using AI tools?** Jump to: [Cheatsheet](cheatsheet.en.md) · [Advanced Tips](#10-tool-guides) · [Real-World Workflows](#real-world-workflows) · [Ecosystem](#ecosystem) @@ -57,14 +57,14 @@ |------|------|-----------| | [**Claude Code**](claude-code/README.en.md) | CLI Agent | 66 tips, Agent + Skill + Hook workflows | | [**Codex CLI**](codex/README.en.md) | CLI Agent | OpenAI's open-source (Rust), Sandbox + AGENTS.md, ChatGPT-plan friendly | -| [**Cursor**](cursor/README.en.md) | IDE | .cursorrules config, Composer Agent mode | -| [**GitHub Copilot**](copilot/README.en.md) | IDE Plugin | Inline completion + Agent mode + custom instructions | -| [**OpenClaw**](openclaw/README.en.md) | AI Agent Framework | 338k Stars, multi-platform + Skills + Cron automation | -| [Windsurf](windsurf/README.en.md) | IDE | Cascade Agent, automatic context | -| [Gemini CLI](gemini-cli/README.en.md) | CLI | By Google, large codebase analysis | -| [Kiro](kiro/README.en.md) | IDE | By AWS, spec-driven development | +| [**Cursor**](cursor/README.en.md) | IDE | `.cursor/rules/*.mdc` rules, Agent / Plan modes, parallel Agents Window | +| [**GitHub Copilot**](copilot/README.en.md) | IDE Plugin | Inline completion + Agent mode + custom agents / instructions | +| [**OpenClaw**](openclaw/README.en.md) | AI Agent Framework | 390k+ Stars, multi-platform + Skills + scheduled automation | +| [Devin Desktop (formerly Windsurf)](windsurf/README.en.md) | IDE | Devin Local Agent, local + cloud agent board | +| [Gemini CLI](gemini-cli/README.en.md) | CLI | By Google, 1M context; individual users have moved to Antigravity CLI | +| [Kiro](kiro/README.en.md) | IDE + CLI | By AWS, spec-driven development | | [Aider](aider/README.en.md) | CLI | Git-native, supports almost any LLM | -| [Trae](trae/README.en.md) | IDE | By ByteDance, free Claude/GPT access | +| [Trae](trae/README.en.md) | IDE | By ByteDance, has a free tier; China edition connects directly | > Every guide follows the same structure: **Core Concepts → Quick Start → Prompt Tips → Advanced Usage → Config Templates** diff --git a/cheatsheet.en.md b/cheatsheet.en.md index bccf72a..0bbb308 100644 --- a/cheatsheet.en.md +++ b/cheatsheet.en.md @@ -4,7 +4,7 @@ > Key parameters, commands, and hotkeys for all 10 tools on one page. Models and pricing move fast — check each tool's website for the current truth. > -> Snapshot date: **2026-04** +> Snapshot date: **2026-09** --- @@ -12,37 +12,37 @@ | What you want | First pick | Alternatives | |---------------|------------|--------------| -| Tab completion inside an IDE | **Cursor** | Copilot / Windsurf / Trae | +| Tab completion inside an IDE | **Cursor** | Copilot / Devin Desktop (formerly Windsurf) / Trae | | Agent in the terminal for complex tasks | **Claude Code** | **Codex CLI** / Aider / Gemini CLI | | Already subscribed to ChatGPT | **Codex CLI** | Cursor with GPT | -| Analyze a huge codebase in one pass | **Gemini CLI** (2M context) | Aider + map mode | -| Tight budget / zero cost | **Trae** (free) or **Gemini CLI** (free tier) | Aider + local models; or `codex --oss --local-provider ollama` | +| Analyze a huge codebase in one pass | **Claude Code** (Opus/Sonnet 5.5 both have 1M context) | Gemini CLI (1M; individual users must switch to Antigravity CLI) / Aider + map mode | +| Tight budget / zero cost | **Trae** (free plan, Auto mode only) or **Aider + local models** | `codex --oss --local-provider ollama`; Copilot / Kiro free tiers | | Direct network in China (no VPN) | **Trae** | OpenClaw + local model | | Team collaboration, spec-driven | **Kiro** (Spec) | Claude Code plan mode | | Git-native, multi-model | **Aider** | — | | AI agent automation (beyond coding) | **OpenClaw** | — | -| Stick with VS Code, nothing new to install | **Copilot** | Cursor/Windsurf/Trae are VS Code forks | +| Stick with VS Code, nothing new to install | **Copilot** | Cursor / Devin Desktop / Trae are VS Code forks | --- ## 2. Capability Matrix -| Dimension | Claude Code | Codex CLI | Cursor | Copilot | Windsurf | Gemini CLI | Kiro | Aider | Trae | OpenClaw | +| Dimension | Claude Code | Codex CLI | Cursor | Copilot | Devin Desktop (formerly Windsurf) | Gemini CLI | Kiro | Aider | Trae | OpenClaw | |-----------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| | **Type** | CLI | CLI | IDE | IDE plugin | IDE | CLI | IDE | CLI | IDE | Agent framework | -| **Vendor** | Anthropic | OpenAI | Cursor | GitHub | Codeium | Google | AWS | OSS | ByteDance | OSS | +| **Vendor** | Anthropic | OpenAI | Cursor (owned by SpaceX) | GitHub | Cognition | Google | AWS | OSS | ByteDance | OpenClaw Foundation | | **Tab completion** | — | — | ★★★ | ★★★ | ★★★ | — | ★★ | — | ★★ | — | | **Agent execution** | ★★★ | ★★★ | ★★★ | ★★★ | ★★★ | ★★ | ★★ | ★★ | ★★ | ★★★ | -| **Runs in terminal** | ★★★ | ★★★ | ★ | — | ★ | ★★★ | — | ★★★ | — | ★★★ | -| **Context window** | 200K | per model | per model | per model | per model | **2M** | per model | per model | per model | per model | -| **MCP support** | ✅ | ✅ | ✅ | ✅ | ✅ | Extension-based | — | — | — | Native | -| **Hook automation** | ✅ | ✅ (beta, reuses Claude schema) | — | — | — | — | ✅ | ✅ (lint/test) | — | ✅ (Cron) | -| **Subagent** | ✅ | ✅ (TOML) | — | — | — | — | — | — | — | ✅ (workspaces) | -| **Sandbox** | App-layer + hooks | OS-kernel (Seatbelt/Landlock) | — | — | — | — | — | — | — | — | +| **Runs in terminal** | ★★★ | ★★★ | ★ | ★★ (Copilot CLI) | ★ | ★★★ | ★★ (Kiro CLI) | ★★★ | — | ★★★ | +| **Context window** | **1M** (Haiku 200K) | per model | per model | per model | per model | **1M** | per model | per model | per model | per model | +| **MCP support** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | — | ✅ | Native | +| **Hook automation** | ✅ (30+ events) | ✅ (12 events, on by default) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (lint/test) | — | ✅ (Cron) | +| **Subagent** | ✅ | ✅ (TOML) | ✅ | ✅ | ✅ | ✅ | ✅ | — | ✅ (custom agents) | ✅ (workspaces) | +| **Sandbox** | OS-level Bash sandbox (Seatbelt/bubblewrap) | OS-level (Seatbelt/bubblewrap) | — | — | — | — | — | — | — | — | | **Multi-model** | ★ (Claude only) | ★ (OpenAI only) | ★★★ | ★★ | ★★★ | ★ (Gemini only) | ★★ | ★★★ (almost any LLM) | ★★ | ★★★ | -| **Open source** | — | ✅ Apache-2.0 | — | — | — | — | — | ✅ MIT | — | ✅ MIT | -| **Direct access in China** | — | — | — | — | — | — | — | — | ✅ | ✅ (with local model) | -| **Pricing model** | API / Pro plan | ChatGPT plan / API | Free / $20 Pro | Free / $10 Pro | Free / paid plan | Generous free tier | Free preview | Whatever LLM you pick | Free (with quota) | OSS free | +| **Open source** | — | ✅ Apache-2.0 | — | — | — | ✅ Apache-2.0 | — | ✅ Apache-2.0 | — | ✅ MIT | +| **Direct access in China** | — | — | — | — | — | — | — | — | ✅ (China edition trae.cn) | ✅ (with local model) | +| **Pricing model** | Pro/Max plan / API | ChatGPT plan / API | Free / $20 Pro | Free / from $10 Pro (AI Credits) | Free / $20 Pro | Enterprise / paid API key (free personal tier discontinued) | Free 50 credits / $20 Pro | Whatever LLM you pick | Free (Auto) / $20 Pro | OSS free | --- @@ -52,15 +52,15 @@ Knowing where each tool keeps its config is half the onboarding battle. | Tool | Main config | Location | Notes | |------|------------|----------|-------| -| Claude Code | `CLAUDE.md` + `.claude/` | Project root | Keep under 200 lines; split to `.claude/rules/` for big projects | +| Claude Code | `CLAUDE.md` + `.claude/` | Project root | Keep under 200 lines; split to `.claude/rules/` for big projects (`paths:` for path-scoped loading); falls back to `AGENTS.md` when there's no CLAUDE.md | | Codex CLI | `AGENTS.md` + `.codex/config.toml` | Project root | `~/.codex/AGENTS.md` for global; nested AGENTS.md overrides parents | -| Cursor | `.cursor/rules/*.md` | Project root | Supports `globs` for file-type loading | -| Copilot | `.github/copilot-instructions.md` | Project root | Plus `agents/` and `chatModes/` in same dir | -| Windsurf | `.windsurfrules` | Project root | Single file, no splitting | +| Cursor | `.cursor/rules/*.mdc` | Project root | Must be `.mdc` (plain `.md` is ignored); frontmatter `description` / `globs` / `alwaysApply`; also reads `AGENTS.md` | +| Copilot | `.github/copilot-instructions.md` | Project root | Also `.github/instructions/*.instructions.md` (`applyTo`) and `.github/agents/*.agent.md`; chatModes is deprecated | +| Devin Desktop (formerly Windsurf) | `.devin/rules/*.md` | Project root | `trigger: always_on / model_decision / glob / manual`; `.windsurfrules` kept only for backward compatibility | | Gemini CLI | `GEMINI.md` | Project root | Structure similar to CLAUDE.md | -| Kiro | `.kiro/steering/*.md` | Project root | Three modes: `always` / `globs` / `manual` | +| Kiro | `.kiro/steering/*.md` | Project root | Four `inclusion:` modes: `always` / `fileMatch` / `manual` / `auto`; global `~/.kiro/steering/` | | Aider | `.aider.conf.yml` | Project root | YAML with model/lint/test settings | -| Trae | `.trae/rules/project_rules.md` | Project root | Chinese rules supported | +| Trae | `.trae/rules/*.md` | Project root | Read recursively, up to three levels; can import `AGENTS.md` / `CLAUDE.md` | | OpenClaw | `~/.openclaw/openclaw.json` | User home | JSON5, hot-reloaded | --- @@ -71,7 +71,7 @@ Knowing where each tool keeps its config is half the onboarding battle. ```bash # Install & run -npm install -g @anthropic-ai/claude-code +curl -fsSL https://claude.ai/install.sh | bash # Officially recommended (brew / winget / npm also work) claude # Enter interactive claude --resume # Resume last session claude --model haiku # Cheap model for simple tasks @@ -79,7 +79,10 @@ claude -p "task" --output-format json # Headless # Inside the session /compact # Compress context -/plan # Enter plan mode +/context # Show context usage +/effort # Adjust thinking depth (low → max) +Shift+Tab # Cycle permission modes (incl. plan / auto) +/code-review # Review current changes Esc # Interrupt current generation ``` @@ -87,15 +90,15 @@ Esc # Interrupt current generation ```bash # Install & run -npm install -g @openai/codex +curl -fsSL https://chatgpt.com/codex/install.sh | sh # Officially recommended (npm / brew also work) codex # Enter TUI (first-run guides ChatGPT login) -codex --sandbox workspace-write # Default combo (`--full-auto` removed in v0.125.0; use this) +codex --sandbox workspace-write # Default combo (`--full-auto` has been removed) codex --sandbox read-only # Read-only exploration codex --add-dir ../sibling-repo # Add writable dirs without opening the whole sandbox codex --yolo # Skip sandbox AND approvals (only when externally sandboxed) codex exec --json "..." | jq -c . # JSONL output for downstream scripts -codex --oss --local-provider ollama -m qwen2.5-coder # Free, fully local -codex mcp-server # Expose Codex as MCP server for other agents +codex --oss --local-provider ollama -m qwen3-coder # Free, fully local +codex app-server # For integrating Codex into other programs (old mcp-server removed) codex resume # Resume last session # Inside the TUI @@ -104,7 +107,8 @@ codex resume # Resume last session /model # Switch model /review # Review diff / branch / commit /compact # Compress conversation -/agent # Switch between subagent threads +/multi-agents # Manage subagents (alias /subagents) +/import # Migrate config from Claude Code / Cursor /diff # Git diff (including untracked) /debug-config # Diagnose config.toml not taking effect ``` @@ -113,14 +117,14 @@ codex resume # Resume last session ``` Tab — Accept completion -Cmd+L — Open Chat (selection auto-attached) -Cmd+I — Open Composer (agent mode) +Cmd+I / Cmd+L — Toggle the Agent sidebar (selection auto-attached) +Shift+Tab — Cycle Agent / Ask / Plan modes +Cmd+E — Switch Agent layout Cmd+K — Inline edit -Cmd+Shift+L — Add current file to Chat context Esc — Reject completion -@file @folder @web @terminal — References -@notepad:name — Reference a Notepad +@Files @Folders @Terminals @Chats @Branch @Browser — References +(Notepads were removed in 2.0 — use rules / skills instead) ``` ### GitHub Copilot @@ -128,51 +132,57 @@ Esc — Reject completion ``` Tab — Accept completion Esc — Reject completion -Cmd+Shift+I — Open Chat -Cmd+I — Inline edit +Ctrl+Cmd+I — Open the Chat view +Shift+Cmd+I — Open Chat in Agent mode +Cmd+I — Inline chat Alt+] / Alt+[ — Cycle completion suggestions #file #selection #terminal #problems — References -@workspace — Whole-project context +#codebase — Force semantic search over the whole project (Agent searches automatically anyway) +/agents — Pick a custom agent (.github/agents/) ``` -### Windsurf +### Devin Desktop (formerly Windsurf) ``` -Write mode — Direct code changes (like Composer) -Chat mode — Q&A +Code mode (default) — Edit code directly +Plan mode — Plan first +Ask mode — Read-only Q&A +Cmd+. — Switch modes -@file @folder @web — References -Cascade tracks your edit flow automatically — no need to paste context +The local agent has moved from Cascade to Devin Local (with subagent support) +/workflow-name — Run a workflow from .devin/workflows/ ``` ### Gemini CLI ```bash -npm install -g @google/gemini-cli +# ⚠️ Service ended for individual/free users on 2026-06-18 — individuals should switch to Antigravity CLI; Enterprise and paid API keys still work +npm install -g @google/gemini-cli # or brew install gemini-cli gemini # Enter interactive -# Use the 2M context for big-codebase analysis (overkill for small tasks) +# The 1M context suits big-codebase analysis (overkill for small tasks) ``` ### Kiro ``` -1. Describe need → 2. Kiro writes Spec → 3. You review → 4. Auto-implement + test +1. Describe need → 2. Kiro writes Spec (requirements.md + design.md + tasks.md) → 3. You review → 4. Implement task by task -Steering load modes: -- always Loaded every conversation -- globs: ["*.java"] Loaded for matching files -- manual Invoked on demand +Steering load modes (frontmatter `inclusion:`): +- always Loaded every conversation +- fileMatch + fileMatchPattern: "**/*.java" Loaded for matching files +- manual Invoked on demand +- auto Decided automatically from description ``` ### Aider ```bash # Install & run -pip install aider-chat -aider --model claude-sonnet-4-5 -aider --model deepseek/deepseek-chat # Cheap -aider --model ollama/qwen2.5-coder # Free, local +python -m pip install aider-install && aider-install +aider --model sonnet # Built-in alias for a recent Claude Sonnet +aider --model deepseek/deepseek-v4-pro # Cheap (deepseek-chat was retired in 2026-07) +aider --model ollama/qwen3-coder # Free, local # Inside the session /add file1 file2 — Add files to context @@ -185,11 +195,11 @@ aider --model ollama/qwen2.5-coder # Free, local ### Trae ``` -Builder mode — Agent, multi-file edits +Agent mode — Multi-file edits (build your own agents: prompt + tools + MCP) Chat mode — Q&A @file @folder @web — References -Free models: Claude / GPT, pick per task complexity +Free plan is Auto mode only; the international edition dropped Claude models in 2025-09 ``` ### OpenClaw @@ -199,9 +209,9 @@ openclaw onboard # Initial setup openclaw gateway start # Start the gateway openclaw doctor # Diagnostics -openclaw skills install # Install a Skill -openclaw cron add "0 9 * * *" "..." # Scheduled task -openclaw models set default # Switch model +openclaw skills install @owner/ # Install a ClawHub Skill +openclaw automations create ... # Scheduled task (cron still works as an alias) +openclaw models set anthropic/claude-sonnet-5-5 # Switch model (provider/model) openclaw channels add telegram # Add messaging channel ``` @@ -213,14 +223,14 @@ openclaw channels add telegram # Add messaging channel ┌─ Mostly working in a terminal? │ ├─ Strongest Agent / big refactors → Claude Code │ ├─ Already on ChatGPT / kernel-level sandbox → Codex CLI -│ ├─ Huge context / free → Gemini CLI +│ ├─ Google ecosystem / enterprise license → Gemini CLI (individuals → Antigravity CLI) │ └─ Multi-model / Git-native → Aider │ ├─ Mostly in an IDE? │ ├─ Don't want a new IDE → GitHub Copilot (VS Code/JetBrains plugin) │ ├─ Willing to switch, budget ok → Cursor -│ ├─ Want the AI to be proactive → Windsurf -│ ├─ Chinese UI / local network / free → Trae +│ ├─ Want the AI to be proactive / local + cloud agent board → Devin Desktop (formerly Windsurf) +│ ├─ Chinese UI / local network → Trae (China edition trae.cn) │ └─ Team / spec-driven → Kiro │ └─ Non-coding AI automation? @@ -237,8 +247,8 @@ openclaw channels add telegram # Add messaging channel | **Codex CLI + Cursor** | For ChatGPT subscribers: CLI as Agent, IDE for completions | | **Codex CLI + Claude Code** | Two CLIs, complementary: Codex for CI/scripts, Claude Code for big refactors | | **Claude Code + Copilot** | Lightweight for pure VS Code users | -| **Gemini CLI + Cursor** | Budget-conscious: 2M free analysis + $20 IDE | -| **Aider + local LLM** | Zero API cost: `ollama/qwen2.5-coder` + Aider | +| **Copilot Free + Aider with local model** | Budget-conscious: IDE completion + terminal agent, zero subscriptions | +| **Aider + local LLM** | Zero API cost: `ollama/qwen3-coder` + Aider | | **Codex CLI --oss + Ollama** | Zero API cost but you want Codex's agent experience: kernel-level sandbox + local model | | **Claude Code + OpenClaw** | Coding + automation: CC codes, OpenClaw runs scheduled tasks | diff --git a/ecosystem.en.md b/ecosystem.en.md index 43cb704..520ae5a 100644 --- a/ecosystem.en.md +++ b/ecosystem.en.md @@ -17,7 +17,7 @@ Before you begin, make sure your environment is ready: | Dependency | Minimum Version | Check Command | How to Install | |------------|----------------|---------------|----------------| -| **Node.js** | 18+ | `node -v` | [nodejs.org](https://nodejs.org) | +| **Node.js** | 20+ | `node -v` | [nodejs.org](https://nodejs.org) | | **npm** | 9+ | `npm -v` | Comes with Node.js | | **Git** | 2.0+ | `git -v` | [git-scm.com](https://git-scm.com) | | **AI Coding Tool** | Any one | -- | Install at least one: Claude Code / Cursor / Copilot / Windsurf, etc. | @@ -40,7 +40,7 @@ cd /your/project npx superpowers-zh ``` -The installer auto-detects which tools you use (Claude Code / Cursor / Copilot / Windsurf and 14 others), installing 20 skills to the correct locations. +The installer auto-detects which tools you use (Claude Code / Codex CLI / Cursor / Kiro / Gemini CLI and more — 26 in total), installing 21 skills to the correct locations. ### Verify Installation @@ -49,7 +49,7 @@ Check that the skill files were created: ```bash # Claude Code users ls .claude/skills/ -# You should see brainstorming.md, systematic-debugging.md, and 20 other files +# You should see brainstorming/, systematic-debugging/, and other skill directories — 21 in total (each containing a SKILL.md) # Cursor users ls .cursor/skills/ @@ -99,7 +99,7 @@ AI: Before I start implementing, I need to clarify a few things: | Issue | Solution | |-------|----------| -| `npx superpowers-zh` errors out | Check your Node.js version (requires 18+): `node -v` | +| `npx superpowers-zh` errors out | Check your Node.js version (requires 20+): `node -v` | | AI behavior unchanged after install | Restart your AI tool (close and reopen) so it reloads the skills | | Want to uninstall | Just delete the directory, e.g., `rm -rf .claude/skills/` | @@ -136,7 +136,7 @@ cd agency-agents-zh ```bash # Claude Code users — check for role files -ls ~/.claude/skills/ | head -10 +ls ~/.claude/agents/ | head -10 # You should see files like engineering-security-engineer.md # Or just ask the AI: @@ -212,9 +212,11 @@ AI: (activates database-optimizer role) # Global install (recommended — gives you the ao command) npm install -g agency-orchestrator -# Initialize: download 276 role definitions locally +# Try it with zero config first (no init needed) +ao demo + +# (Optional) Copy the 276 Chinese roles locally so you can customize them ao init -# Role files are downloaded to ~/.ao/roles/ ``` ### Quick Demo (No API Key Required) @@ -351,7 +353,7 @@ npm install shellward | Protection | Description | |------------|-------------| -| Prompt injection detection | 32 rules (18 Chinese + 14 English), with risk scoring | +| Prompt injection detection | 37 rules (20 Chinese + 17 English), with risk scoring; also detects MCP tool poisoning / rug-pulls | | Dangerous command interception | `rm -rf`, reverse shells, fork bombs, `chmod 777`, etc. | | PII detection | National IDs, bank cards, phone numbers, SSN, credit cards, API keys, JWTs | | Data exfiltration chain detection | Read sensitive data -> send email/HTTP POST/curl = blocked | @@ -365,24 +367,17 @@ npm install shellward Works with Claude Desktop / Cursor / Claude Code and other MCP-compatible tools. -Add this to your MCP configuration file: - -```bash -# Find shellward's install path -npm list -g shellward # If globally installed -# or -ls node_modules/shellward/src/mcp-server.ts # If project-installed -``` +Add this to your MCP configuration file (`npx` fetches `shellward-mcp` automatically — no need to find the install path): ```json // Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json // Cursor: .cursor/mcp.json -// Claude Code: .claude/settings.json under mcpServers +// Claude Code: .mcp.json in the project root (or claude mcp add) { "mcpServers": { "shellward": { "command": "npx", - "args": ["shellward", "--mcp"] + "args": ["-y", "-p", "shellward", "shellward-mcp"] } } } @@ -393,14 +388,15 @@ ls node_modules/shellward/src/mcp-server.ts # If project-installed 3 lines of code to integrate with any AI Agent framework (LangChain, AutoGPT, etc.): ```javascript -const { ShellWard } = require('shellward'); -const guard = new ShellWard(); +import { ShellWard } from 'shellward'; +const guard = new ShellWard({ mode: 'enforce' }); -// Check before the Agent executes a tool call -const result = guard.check(toolCall); +// Check before the Agent executes a command +const result = guard.checkCommand(cmd); if (!result.allowed) { console.log(`Blocked: ${result.reason}`); } +// Also available: checkInjection / scanData / checkOutbound ``` ### Before and After @@ -526,13 +522,13 @@ npm install shellward Then configure MCP so your AI tool routes operations through shellward: ```json -// Add to .claude/settings.json (Claude Code) +// Add to .mcp.json (Claude Code, project root) // or .cursor/mcp.json (Cursor) { "mcpServers": { "shellward": { "command": "npx", - "args": ["shellward", "--mcp"] + "args": ["-y", "-p", "shellward", "shellward-mcp"] } } } diff --git a/resources.en.md b/resources.en.md index 2b76c0e..8e80620 100644 --- a/resources.en.md +++ b/resources.en.md @@ -4,7 +4,7 @@ > This project covers "how to use the 10 tools well." To go deeper, here's a curated external resource list. **Quality over completeness** — every entry states **when it's worth your time**. > -> Snapshot date: **2026-04**. Report dead links via Issue. +> Snapshot date: **2026-09**. Report dead links via Issue. --- @@ -27,15 +27,11 @@ - Anthropic's official **9-chapter interactive prompt tutorial**, runnable Jupyter - Fastest systematic onboarding to prompt engineering (2-3 hours) -- [dair-ai/Prompt-Engineering-Guide](https://github.com/dair-ai/Prompt-Engineering-Guide) · ⭐ 60k+ · EN (with Chinese translation) +- [dair-ai/Prompt-Engineering-Guide](https://github.com/dair-ai/Prompt-Engineering-Guide) · ⭐ 75k+ · EN (with Chinese translation) - The prompt-engineering encyclopedia: CoT, ReAct, RAG, tool-use patterns - Go here to understand terminology and which technique fits which scenario -- [datawhalechina/prompt-engineering-for-developers](https://github.com/datawhalechina/prompt-engineering-for-developers) · ⭐ 13k+ · ZH - - Datawhale's Chinese translation of **DeepLearning.AI prompt engineering courses** with notes - - Best combo for Chinese learners: video + code - -- [phodal/prompt-patterns](https://github.com/phodal/prompt-patterns) · ⭐ 1k+ · ZH +- [phodal/prompt-patterns](https://github.com/phodal/prompt-patterns) · ⭐ 3k+ · ZH - Prompt patterns and DSL design guide by prominent Chinese architect Phodal Huang - Advanced reference for building complex AI coding workflows and reusable prompts @@ -43,9 +39,9 @@ ## 🛠️ Awesome Lists per Tool -- [PatrickJS/awesome-cursorrules](https://github.com/PatrickJS/awesome-cursorrules) · ⭐ 30k+ · EN - - Community collection of `.cursorrules` files, categorized by tech stack - - Copy a matching stack's rules when starting a new project +- [PatrickJS/awesome-cursorrules](https://github.com/PatrickJS/awesome-cursorrules) · ⭐ 40k+ · EN + - Community collection of Cursor rule files, categorized by tech stack + - Copy a matching stack's rules when starting a new project (note: newer Cursor versions expect them in `.cursor/rules/*.mdc`) - [hesreallyhim/awesome-claude-code](https://github.com/hesreallyhim/awesome-claude-code) · ⭐ high · EN - Most comprehensive Claude Code community resources: tips, Skills, Hooks, workflows @@ -53,15 +49,15 @@ - [github/awesome-copilot](https://github.com/github/awesome-copilot) · ⭐ high · EN - GitHub's official Copilot resource collection - - Look up instruction / Chat mode / Agent examples + - Look up instructions / custom agent / Skills examples - [addyosmani/gemini-cli-tips](https://github.com/addyosmani/gemini-cli-tips) · ⭐ medium · EN - Google DevRel Addy Osmani's **30 Gemini CLI tips** - - Gemini CLI users: just copy these + - Gemini CLI users: just copy these (note: not updated since 2025-10; individual users now need to move to Antigravity CLI) -- [detailobsessed/awesome-windsurf](https://github.com/detailobsessed/awesome-windsurf) · ⭐ medium · EN - - Windsurf community resources (rules, workflows, MCP) - - Find community patterns for Cascade +- [detailobsessed/awesome-devin](https://github.com/detailobsessed/awesome-devin) · ⭐ medium · EN + - Formerly awesome-windsurf; community resources for Devin Desktop (formerly Windsurf) (rules, workflows, MCP) + - Find community rules and workflows for Devin Desktop --- @@ -71,9 +67,9 @@ - **Official repo**. Issues and Discussions are the primary source for new features and techniques - Track new features (slash commands / hooks / skills) -- [anthropics/courses](https://github.com/anthropics/courses) · ⭐ high · EN +- [anthropics/courses](https://github.com/anthropics/courses) · ⭐ high · EN (archived) - Anthropic's official course collection: Prompt, Tool use, RAG, MCP - - The full Anthropic syllabus in one place + - The repo is archived and no longer updated; the concepts still apply, but follow the cookbooks for current API usage - [anthropics/claude-cookbooks](https://github.com/anthropics/claude-cookbooks) · ⭐ high · EN - Claude API practical examples: multimodal, tool use, RAG, Agent patterns @@ -115,21 +111,21 @@ ## 🇨🇳 Chinese Resources -- [datawhalechina/llm-cookbook](https://github.com/datawhalechina/llm-cookbook) · ⭐ 22k+ · ZH - - Datawhale's Chinese LLM development cookbook collection +- [datawhalechina/llm-cookbook](https://github.com/datawhalechina/llm-cookbook) · ⭐ 25k+ · ZH + - Datawhale's Chinese LLM development cookbook collection (includes the Chinese edition of the **DeepLearning.AI prompt engineering courses**; the former prompt-engineering-for-developers repo has been merged in) - Chinese reference manual for embedding LLM capabilities into your project -- [liaokongVFX/LangChain-Chinese-Getting-Started-Guide](https://github.com/liaokongVFX/LangChain-Chinese-Getting-Started-Guide) · ⭐ 8k+ · ZH +- [liaokongVFX/LangChain-Chinese-Getting-Started-Guide](https://github.com/liaokongVFX/LangChain-Chinese-Getting-Started-Guide) · ⭐ 9k+ · ZH - LangChain Chinese getting-started guide: Agent / Chain / Memory - Foundation for building your own AI coding tools / Agent assistants -- [phodal/aigc](https://github.com/phodal/aigc) · ⭐ 1k+ · ZH +- [phodal/aigc](https://github.com/phodal/aigc) · ⭐ 1.5k+ · ZH - Phodal Huang's book "Building LLM Applications" - AI coding architect's perspective, engineering and team adoption -- [phodal/auto-dev](https://github.com/phodal/auto-dev) · ⭐ 3k+ · ZH/EN - - **Domestic Chinese JetBrains AI coding plugin** with custom Agent + Chinese DevIns scripts - - AI coding option for IDEA/PyCharm ecosystem +- [phodal/auto-dev](https://github.com/phodal/auto-dev) · ⭐ 4k+ · ZH/EN + - AI coding platform led by Phodal Huang, evolved from a JetBrains plugin into a Kotlin Multiplatform multi-agent platform + - Chinese-made AI coding option for the IDEA/PyCharm ecosystem ---