Wechat Agent Studio

wechat-agent-studio

12 个 AI Agent 通过 Discord 协作运营微信公众号 —— 从选题讨论、内容创作到封面生成、草稿发布的全自动化方案。

What it does

Claude Code skill，提供多 Agent 公众号工作室的完整搭建和运维指南：

• 12 Agent 团队配置: 总监、笔锋、灵犀、洞见、像素、微澜、织网、纽带、光影、蒙太奇、流量 — 各有独立性格和互动关系
• 会议讨论机制: 3 轮讨论（亮牌→交锋→终极辩论）后拍板执行，禁止无脑"收到就干"
• Discord 协作频道: agent-channel 绑定、team-msg.sh embed 消息（带头像/颜色）、team-file.sh 文件上传
• SOUL.md 编写规范: 直接回复规则、格式禁令、计数规则、自动续流、Opus/Sonnet/GLM-5 模型差异应对
• team-supervisor 守护进程: 24/7 统一守护（6 项检查、30s 轮询、自动修复 + Discord 告警、agent 催促）
• 微信发布管线: 通义万象图片生成、wechat_publish.py 草稿发布、pipeline.py 浏览器自动化预览、tab 自动清理
• Star-Office-UI 可视化: 像素风办公室监控 agent 实时状态
• 端到端检查清单: 权限、SOUL.md、源码 patch、监控、supervisor、管线 — 逐项确认

Key lessons (from 9 rounds of production testing)

#	Lesson	Impact
1	Text @mentions don't trigger agents	Must use sessions_spawn tool
2	allowAgents must be per-agent config	agents.defaults rejects it
3	Contradictory SOUL.md instructions are fatal	Model picks the lazier option
4	Real error examples 10x more effective than abstract rules	List actual leaked text
5	Opus narrates between tool calls → text leaks to Discord	Need explicit "no text between tools" rule
6	Sub-agents must NOT use NO_REPLY	Breaks announce queue silently
7	Session cleanup must cover all 11 agent dirs	Only clearing agents/main/ leaves stale context
8	Health API Discord status can be wrong	Use log-based detection instead
9	draft/update needs all article fields	Only sending thumb_media_id → 44004 error
10	Tongyi Wanxiang needs 30s delay between calls	429 rate limit on batch generation
11	Browser tabs accumulate from pipeline runs	Close WeChat tabs via CDP after operations
12	Critical rules must appear at SOUL.md head AND tail	LLM attention is U-shaped on long docs

Install

# Option A: Clone
git clone https://github.com/evan966890/wechat-agent-studio.git
cp wechat-agent-studio/