Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Ten Silent Runtime Failures in a DeepSeek Coding Agent

AI summary

A developer describes six months of operating a custom terminal-based coding agent backed by DeepSeek, with 359 commits reported, including 129 bug-fix commits. The article details ten runtime failures that appeared normal, including missing completion signals, discarded input, false success, context compression losing task anchors, and cache-prefix invalidation. It recommends explicit failures, machine-readable command results, queued input, stable context anchors, and deterministic cache prefixes, but the usage and performance figures are self-reported.

In-depth insight

The first-hand failure catalog shows that coding-agent reliability often depends on runtime plumbing and observability rather than model capability alone.

agent 不报错才最吓人:自研编码工具半年,我踩了 10 个"看着正常"的运行时坑

前言

今年 6 月起,我给自己造一个 AI 编码工具:终端里跟大模型对话,它能读代码、改文件、跑命令、自己多轮干活(类似 Claude Code 的形态),底层接 DeepSeek。半年下来 359 次提交,129 次在修 bug。

修到后来我最大的体会是一句话:agent 出问题,八成不会报错。

听着有点反直觉。它照样流式输出、照样调工具、照样礼貌地说"完成了"——但根子上已经歪了。崩溃好查,安静地坏才难查。这篇文章就把这半年攒下的运行时坑挑 10 个最典型的讲透:每一个当时的表象都"看着正常",深挖下去全是问题。先上速览表:

#表象(看着正常)真相
坑一工具"永远转圈"像死机我忘了发"干完了"信号
坑二输入框清空了 = 发出去了?消息被两端静默吞掉
坑三语气流畅地说"基本完成了"活只干了一半(早收尾)
坑四命令失败,模型说成功输出像成功 ≠ 真成功
坑五每个动作都正常合起来 99 次工具调用
坑六引导生效了,结果还是 0 命中模型照抄出错形态
坑七工具输出丰富、格式正常目录是假的,谎言链三层
坑八审批卡弹个不停,看着很安全名字没对齐,安全在空转
坑九长任务中途"记性变差"压缩把任务锚点压碎了
坑十无报错无卡顿,功能一切正常账单在悄悄流血

#DeepSeek#deepseeker-code

Original · 12 min readRead the original
Group chat

No comments yet. Start the conversation.