GUI-Automated WeChat Group-Chat AI: Five Operational Failures
A developer describes building a WeChat group-chat AI through GUI automation because the official bot channels they examined were not suitable for ordinary WeChat groups. The system controls a logged-in desktop account through screen interaction, uses Ollama for local visual processing and faster-whisper for speech transcription, and sends replies through the clipboard. The author reports five operational failures: duplicate sending caused by two instances, replies to stale messages, queued replies, queued replies being discarded, and the agent treating an unsent reply as already said. The first four received fixes involving a single-instance lock, a stale-message gate, queue prioritization, and replacing the oldest queued reply; the author says the fifth issue was located but not yet solved. The report also describes seven design choices, including delayed responses and deliberate silence. The author reports 229 group chats, nine private chats, 51 deliberate silences, and 445 passing tests on their own machine; these figures and the incidents were not independently verified.
The account illustrates the operational problems of making a GUI-driven chat agent behave consistently with what group members actually see, including duplicate, stale, queued, and discarded replies. It also highlights an unresolved distinction between generated messages and messages actually sent.
关于这篇稿子里的 AI,我没打算含糊。 骨架是我把几天日志和口述捋出来的,AI 帮着整理成结构;句子被它改短过,数字和日志被它逐条比对过,查 API 和库用法它出过力。(用的模型:
deepseek-v4.1-flash+ 若干辅助工具。) 但内核不是它给的:「限制 AI 说话的能力,比让它会说话的能力重要十倍」这个结论,是我自己在真机运行里撞出来的。229 条群聊对 9 条私聊、51 次主动沉默、445 项测试全绿,都是我自己机器上跑出来的原始数字。五个事故我真踩过,七条设计我真改过。 事实和引文我逐条核过一遍,AI 说错的、编的地方都删了;写歪的地方算我的。
脱敏说明:文中所有群友一律用「群友X」这类代号(群友C、群友E…)。真实昵称与聊天原文不在本文出现—— 这跟项目本身的做法一致(真昵称只落本地,发给模型的一律是代号)。
第一部分:为什么「能回话」不是难点
为什么非得是微信
如果你做过 Slack 机器人、Discord 机器人,它们都有一条共同的路:官方开放接口、Webhook 回调、标准事件格式。
Some source content is omitted. Continue reading on the original site.