envelope-team

Envelope Team

设计并生成 .envelope.json AI Agent 团队定义文件——这是在 Envelope 上构建和部署多 Agent 团队的开放标准。

何时使用此技能

• 用户要求构建、设计或生成 Envelope 团队
• 用户描述了可通过 AI Agent 团队自动化的工作流
• 用户想要创建用于部署的 .envelope.json 文件
• 用户需要有关 Agent 层级、访问策略或关卡的帮助

此技能的功能

1. 接收团队用途的自然语言描述
2. 设计 Agent 层级（监督者 → 经理 → 工作人员）
3. 生成符合架构规范的 .envelope.json 文件
4. 根据需要包含访问策略、变量、密钥和关卡
5. 解释用户询问的架构概念

使用方法

基本用法

Build me an Envelope team for customer support triage — one supervisor,
two agents that read Zendesk tickets, and a Slack notifier for escalations.

Create an Envelope team definition for outbound sales — an SDR manager
overseeing two SDRs who work HubSpot leads and a RevOps analyst.

使用有人参与的关卡

Build an Envelope team that drafts marketing emails and pauses for
human approval before sending.

示例

用户：“为我构建一个内容审核 Envelope 团队，由一名监督者和两名审核 Agent 组成，用于根据社区准则检查帖子。”

输出：

{
  "$schema": "https://schema.openenvelope.org/team/v1.json",
  "name": "Content Moderation Team",
  "slug": "content-moderation",
  "version": "1.0.0",
  "description": "A supervisor and two reviewer agents that check posts against community guidelines.",
  "visibility": "team",
  "category": "ops",
  "requiredVariables": ["companyName"],
  "requiredSecrets": ["MODERATION_API_KEY"],
  "metadata": { "generatedBy": "Claude Code · openenvelope.org" },
  "agents": [
    {
      "key": "supervisor",
      "name": "Moderation Supervisor",
      "role": "supervisor",
      "capabilities": ["Route content to reviewer agents", "Aggregate decisions", "Escalate edge cases"],
      "model": "anthropic:claude-sonnet-4-5",
      "systemPrompt": "You supervise the content moderation team at {{companyName}}. Delegate each item to a reviewer and consolidate their verdicts."
    },
    {
      "key": "policy-reviewer",
      "name": "Policy Reviewer",
      "role": "reviewer",
      "capabilities": ["Check content against community guidelines", "Flag policy violations"],
      "model": "anthropic:claude-haiku-3-5",
      "systemPrompt": "You review content for policy violations at {{companyName}}. Return a verdict of approve, flag, or remove with a reason.",
      "reportsToKey": "supervisor",
      "accessPolicy": {
        "accessPolicyVersion": "1",
        "rules": [
          { "host": "api.moderation-service.com", "action": "allow" },
          { "host": "*", "action": "deny" }
        ]
      }
    },
    {
      "key": "spam-reviewer",
      "name": "Spam Reviewer",
      "role": "reviewer",
      "capabilities": ["Detect spam, bot activity, and duplicate content"],
      "model": "anthropic:claude-haiku-3-5",
      "systemPrompt": "You detect spam and bot activity at {{companyName}}. Return a verdict of approve, flag, or remove with a confidence score.",
      "reportsToKey": "supervisor"
    }
  ]
}

架构参考

顶层字段

字段	必填	说明
$schema	是	始终为 https://schema.openenvelope.org/team/v1.json
name	是	易于理解的团队名称
slug	是	URL 安全格式，仅限小写字母和连字符
version	是	Semver，例如 "1.0.0"
description	是	该团队的用途
visibility	是	"public" · "team" · "private"
requiredVariables	否	非敏感配置，通过 {{varName}} 插值
requiredSecrets	否	API 密钥，加密存储并作为 ${SECRET_NAME} 注入
schedule	否	Cron 计划 { cron, timezone, task }
agents	是	Agent 定义
gates	否	有人参与的检查点

Agent 字段

字段	必填	说明
key	是	此文件中的唯一标识符
name	是	显示名称
role	是	自由填写，例如 "supervisor"、"analyst"、"sdr"
capabilities	是	此 Agent 能执行的操作，用于委派路由
model	是	例如 anthropic:claude-sonnet-4-5、anthropic:claude-haiku-3-5
systemPrompt	是	Agent 指令。使用 {{varName}} 和 ${SECRET_NAME}
reportsToKey	否	经理 Agent 的 key。顶层监督者省略此项
accessPolicy	否	出站 HTTP 允许列表

层级规则

应恰好有一个 Agent 不设置 reportsToKey——这就是接收初始任务的监督者。其他 Agent 都应引用其经理的 key。

访问策略

"accessPolicy": {
  "accessPolicyVersion": "1",
  "rules": [
    { "host": "api.example.com", "action": "allow" },
    { "host": "*", "action": "deny" }
  ]
}

有人参与的关卡

"gates": [
  {
    "name": "approval-check",
    "type": "approval",
    "trigger": { "afterAgent": "drafter" },
    "onApprove": "continue",
    "onReject": "halt"
  }
]

提示

• 监督者和经理使用 anthropic:claude-sonnet-4-5；执行重复性工作的叶节点 Agent 使用 anthropic:claude-haiku-3-5
• 将组织特有的信息（公司名称、支持邮箱、Slack 频道）放入 requiredVariables
• 严格限制允许列表时，请在访问策略规则末尾添加 { "host": "*", "action": "deny" }
• 在会发送邮件、公开发帖或进行购买的 Agent 之前添加关卡
• 部署前进行验证：npx @openenvelope/schema validate ./team.envelope.json

链接

• openenvelope.org — 部署和运行团队
• schema.openenvelope.org/team/v1.json — JSON 架构
• npmjs.com/package/@openenvelope/schema — TypeScript 类型和验证器
• openenvelope.org/mcp — 通过 MCP 从 Claude.ai 或 ChatGPT 运行团队