Claude Code Hooks：你跳过的 5 万美元安全审计
📄 中文摘要
在没有检查的情况下将 RTK 安装到 Claude Code 中，类似于在生产环境中未经过测试就进行补丁更新。这不仅仅是添加一个工具，而是在代理的执行路径中注入具有完全用户权限的可执行代码。RTK 针对真正的痛点进行优化，Claude Code 在嘈杂的 Bash 输出中可能会丢失上下文，而 RTK 的模型则相对简单。它使用 <code>PreToolUse</code> 钩子将命令如 <code>git status</code> 重写为 <code>rtk git status</code>，并将压缩版本的输出提供给 Claude Code，而不是原始的终端转储。RTK 的安装流程也明确指出，全球设置可以修补 <code>~/.claude/settings.json</code>，在 <code>~/.claude/hooks/rtk-rewrite.sh</code> 下添加钩子，并在需要时干净地卸载。
📄 English Summary
Claude Code Hooks: The $50k Security Audit You Skip
Installing RTK into Claude Code without inspection is akin to patching production without testing. This process involves injecting executable code with full user permissions into the agent's execution path. RTK addresses a significant pain point, as Claude Code can lose context due to noisy Bash output. RTK's model simplifies this by using a PreToolUse hook to rewrite commands like 'git status' into 'rtk git status', subsequently providing a compressed version of the output to Claude Code instead of the raw terminal dump. The installation flow of RTK also clarifies that a global setup can patch '~/.claude/settings.json', add a hook under '~/.claude/hooks/rtk-rewrite.sh', and can be uninstalled cleanly later if necessary.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等