📡 Idea Radar 每天 8:00 / 18:00 推送 · 18:00 生成精选
reddit · SideProject · 60 分 ← 返回列表

Title: I got tired of AI coding agents saying "done" when they weren't. So I built this.

You know that moment when Codex spends 20 minutes working on something, tells you everything is done, and then you open the code and realize half the task is still missing? Yeah. That got old pretty fast. I kept running into the same problem. The agent would finish something , but not necessarily what I actually asked for. And I'd end up checking everything myself or typing "continue" over and over. So I started building a tool called Truth Firewall. The idea is pretty simple: instea

五维评分

付费意愿
45
需求真实度
55
个人可落地
82
竞争空间
58
新鲜度
82

AI 分析

痛点:AI 编程 agent 常在任务未真正完成时就宣称“done”,开发者不得不反复人工核对或不断输入 continue 催促。

目标用户:重度使用 Codex/Cursor/Claude Code 等 AI 编程 agent 的独立开发者与小团队工程师

证据:Codex spends 20 minutes working on something, tells you everything is done, and then you open the code and realize half the task is still missing... I'd end up checking everything myself or typing "continue" over and over.

变现:开源核心 + 付费托管版(团队共享的验证规则库、CI 集成、历史审计与仪表盘),或按席位订阅的 SaaS

MVP:做一个轻量 CLI/代理层:用户用自然语言或 YAML 定义验收标准,工具在 agent 声称完成后自动跑检查(测试、diff 比对、LLM 校验),不通过则自动回退并让 agent 重试;先做单机版+几套模板规则,接 GitHub 一键安装。

竞品:SWE-bench 类评测框架、Cursor/Codex 自带的任务确认机制、通用 CI/测试工具(pytest、GitHub Actions)

摘要:作者因 Codex 等 AI 编程 agent 常虚报“完成”而开发了开源工具 Truth Firewall,通过对照需求清单校验结果,不通过就驳回并让 agent 重试。目前已发布到 PyPI,仅支持 Windows+Python 任务,且在寻求用户试用。帖子体现真实痛点但无明确付费证据,属于早期可探索的产品方向。