reddit · SideProject · 10-07 09:50 61 分 ← 返回列表

goalpost - a free Claude Code plugin that won't let the agent say "done" until real checks pass

I use Claude Code's /goal a lot: you say what "done" means and the agent keeps working until a small judge model says the goal is met. Then I noticed the judge only reads the chat and never runs anything. I measured it on 4 tasks with hidden graders: it said "met" in 17 of 17 plain runs, and 8 of those were broken. So I built goalpost as a side project. It's a hooks plugin, you keep typing /goal like before. It freezes the goal into checkable criteria, blocks the stop until ea

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：AI 编程助手（Claude Code 等）不运行任何验证就宣称任务「完成」，导致开发者拿到实际是坏的代码，浪费大量排查时间。

目标用户：使用 Claude Code / AI 编码代理的个人开发者和工程团队，尤其是把 agent 接入 CI 或自动化流程的人。

证据：it said "met" in 17 of 17 plain runs, and 8 of those were broken. | 28/34 hidden tests without it, 34/34 with it. | It's not magic: it roughly doubles time and cost

变现：开源核心免费（MIT）+ 付费增值：团队版（集中式审计日志、CI 集成、多 agent/多仓库支持）、企业合规报告，或把「验证即服务」做成托管 SaaS 按用量收费。

MVP：2 周内做一个 Claude Code hooks 插件：把目标冻结为可检查的验收条目，每条必须有新跑的通过检查才允许 stop；再加一个测试文件防篡改守卫（检测 agent 悄悄改测试）和结束时的一次独立审计。

竞品：Claude Code 内置 /goal、CI/CD 流水线、TDD 工具链、其他 AI agent 验证 hooks 插件

摘要：作者发现 Claude Code 的 /goal 判定模型只看聊天记录、从不运行验证，实测 17 次全判「达成」，其中 8 次实际是坏的；同一隐藏测试任务下无插件 28/34、有插件 34/34。他自建开源插件 goalpost 强制每条验收标准都有新通过的检查才允许结束，代价是时间和成本大约翻倍。作者免费开源（MIT），没有直接付费证据，但暴露了 AI 编码代理「假完成」这一真实且可量化的痛点。