Multi Agent Group Chat

多 Agent 群聊协作插件。子 Agent 完成 sessions_send 任务后，自动将回复发送到来源群聊。 支持 Telegram 和飞书，无需配置群 ID，自动检测消息来源。 适用于：多 Bot 协作、团队分工、群聊表演等场景。

Security Scan

The plugin's code, manifest, and instructions are internally consistent with its stated purpose (auto-forwarding sub-agent replies to group chats), but it will automatically forward agent outputs to groups and has simple heuristics that may cause accidental leakage or mis-targeting — review and test before enabling in production.

Purpose & Capability

Name/description, SKILL.md, openclaw.plugin.json, and index.ts all describe the same capability: detect sessions_send sub-agent results and forward replies to the source chat (Telegram/Feishu) using the host runtime APIs. The plugin does not request unrelated credentials or binaries.

Instruction Scope

SKILL.md and the code limit behavior to handling 'sessions_send' internal tasks and use only platform APIs (gateway.rpc / runtime.message.send / api.message.send). However the detection logic is string-based (includes 'sessions_send', 'session_send', 'subagent', or presence of parentSessionKey) which can be permissive and cause false positives. The plugin forwards the full reply without sanitization, which could leak sensitive data from sub-agent outputs to group chats.

Install Mechanism

No install spec that downloads or executes remote code; files are provided and installation is manual or via ClawHub. No network downloads or unusual install paths are present in the bundle.

Credentials

The plugin declares no environment variables or credentials. The SKILL.md expects bot tokens/config to be present in the OpenClaw configuration (openclaw.json), which is appropriate for sending messages. There are no unrelated credential requests.

Persistence & Privilege

always:false (normal). The plugin listens for agent_end events and will act automatically when those events occur (normal for a plugin). Combined with automatic forwarding behavior, this gives it the ability to exfiltrate any sub-agent reply to the originating group if misconfigured — consider scope and enablement carefully.

Assessment

This plugin appears to do what it says, but you should: (1) test it in a safe/trusted group before enabling widely, (2) ensure sensitive agents or tasks are not run in contexts that will be forwarded, (3) verify your openclaw.json bot token mappings so messages are sent by the expected accounts, (4) consider adding filtering/sanitization or stricter detection (so only specific session sources are forwarded), and (5) audit logs/monitor message sends after enabling. If you need stricter guarantees, request a code change to tighten the session-source detection and to allow content filtering or explicit allowlists for group IDs or agents.

Multi-Agent Group Chat Plugin

多 Agent 群聊协作插件，让 AI 团队在群里"讨论"起来。

功能

• ✅ 子 Agent 完成任务后，自动发到群里
• ✅ 自动检测来源群，无需配置群 ID
• ✅ 支持 Telegram 和飞书
• ✅ 仅处理 sessions_send 内部任务，不影响正常对话

架构

用户 @Boss → Boss 分配任务 → 子 Agent 执行
                              ↓
                    【插件自动转发到群】
                              ↓
                    用户在群里看到讨论过程

安装

方式 1：从 ClawHub 安装

clawhub install multi-agent-chat

方式 2：手动安装

复制 index.ts 和 openclaw.plugin.json 到：

~/.openclaw/extensions/multi-agent-chat/

配置

在 openclaw.json 中启用插件：

{
  "plugins": {
    "allow": ["multi-agent-chat"],
    "entries": {
      "multi-agent-chat": {
        "enabled": true
      }
    }
  }
}

配套 Agent 配置

需要配置多个 Agent：

{
  "agents": {
    "entries": {
      "boss": { "workspace": "~/.openclaw/agents/boss" },
      "coder": { "workspace": "~/.openclaw/agents/coder" },
      "writer": { "workspace": "~/.openclaw/agents/writer" }
    }
  },
  "channels": {
    "telegram": {
      "accounts": {
        "boss": { "token": "BOT_TOKEN_1", "agent": "boss" },
        "coder": { "token": "BOT_TOKEN_2", "agent": "coder" },
        "writer": { "token": "BOT_TOKEN_3", "agent": "writer" }
      },
      "groupPolicy": "open"
    }
  }
}

使用方式

1. 创建 Telegram 群
2. 拉入多个 Bot
3. @BossBot 发任务
4. 观察各 Agent 在群里讨论

工作原理

1. Boss Agent 用 sessions_send 给子 Agent 发任务
2. 子 Agent 完成任务后回复
3. 插件监听 agent_end 事件
4. 检测到是 sessions_send 来的任务
5. 自动将回复发到来源群

注意事项

• 需要 OpenClaw 内置的 announce 机制配合
• 子 Agent 的结果会同时：返回给 Boss + 发到群里
• 插件不会影响正常的用户对话

作者

OpenClaw 社区