HackerNews中文版

Claude Code 有一个计划模式，在编码之前会编写实施计划。我开发了一个工具，可以拦截这些计划，并将其发送给一个竞争的人工智能（Codex CLI 或 Gemini CLI）进行对抗性审查，然后才允许 Claude 继续进行编码。当 Claude 调用 `ExitPlanMode` 时，会触发 `PreToolUse` 钩子，将计划发送给审查者，并返回 `{ permissionDecision: "deny", permissionDecisionReason: "<feedback>" }`。Claude 收到拒绝和反馈后，会修改计划并再次尝试。在可配置的轮次（默认：2）后，计划将通过审查。最后，您将获得原始计划与最终计划的差异对比。不同的模型有不同的盲点。其价值在于第二个视角，而不是某个模型“更好”。 ``` npm install -g claude-plan-reviewer claude-plan-reviewer setup ``` 该工具需要 Node.js 18 及以上版本，以及 Codex CLI 或 Gemini CLI。约 400 行 JavaScript，零依赖。采用 MIT 许可证。 <a href="https://github.com/yuuichieguchi/claude-plan-reviewer" rel="nofollow">https://github.com/yuuichieguchi/claude-plan-reviewer</a>

查看原文

Claude Code has a plan mode where it writes implementation plans before coding. I built a tool that intercepts those plans and sends them to a competing AI (Codex CLI or Gemini CLI) for adversarial review before Claude is allowed to proceed.When Claude calls `ExitPlanMode`, a `PreToolUse` hook fires, sends the plan to the reviewer, and returns `{ permissionDecision: "deny", permissionDecisionReason: "<feedback>" }`. Claude receives the denial with the feedback, revises the plan, and tries again. After a configurable number of rounds (default: 2), it passes through. At the end you get a diff of the original vs. final plan.Different models have different blind spots. The value is in the second perspective, not in one model being "better."``` npm install -g claude-plan-reviewer claude-plan-reviewer setup ```Requires Node.js 18+ and Codex CLI or Gemini CLI. ~400 lines of JS, zero dependencies. MIT licensed.<a href="https://github.com/yuuichieguchi/claude-plan-reviewer" rel="nofollow">https://github.com/yuuichieguchi/claude-plan-reviewer</a>

展示HN：Claude-plan-reviewer – 竞争AI评审Claude代码的计划