Token导航 LogoToken导航TokenDH.com
研究检索敏感数据github未标认证来源可访问许可证需确认审计提醒

team-orchestration团队编排

Agent Skill

team-orchestration 用于查找、检索和筛选相关信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

总安装

259

周安装

11

GitHub Stars

24

下载量

91
CodexClaudeCursorGemini CLI

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

GitHub

来源数

2

许可证

unknown

最后核验

2026-05-01

来源状态

来源可访问

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:team-orchestration(团队编排)
来源仓库:https://github.com/noobygains/godmode
仓库路径:skills/team-orchestration
安装命令:
npx skills add https://github.com/noobygains/godmode --skill team-orchestration
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/noobygains/godmode --skill team-orchestration

简介

提供端到端的虚拟团队任务创建、分配与监控功能。

  • 支持跨前端/后端工程师协作完成特定软件模块开发。
  • 内置任务阻塞关系管理与进度可视化界面。team-orchestration 属于研究检索类 Skill,可作为该场景下的辅助能力补充。
  • 适用于 Cursor、Claude、Codex 等支持技能扩展的环境。
  • 需配置 API 密钥并授权访问本地工作树与测试工具。

SKILL.md

Team Orchestration

Overview

Agent Teams enable multiple Claude Code sessions to collaborate on a shared project with direct peer-to-peer messaging and shared task lists. Unlike subagents, teammates can communicate with each other, claim tasks dynamically, and coordinate on shared problems.

Core principle: Deploy teams when tasks benefit from collaboration, not merely parallelism. If teammates will never need to message each other, use parallel subagents instead.

No exceptions. No workarounds. No shortcuts.

The Prime Directive

NO TEAM WITHOUT A COLLABORATION REQUIREMENT

If teammates will never exchange messages, you do not need a team.

When to Use

digraph when_teams {
    "Multiple tasks?" [shape=diamond];
    "Need peer-to-peer collaboration?" [shape=diamond];
    "Tasks independent?" [shape=diamond];
    "Agent Teams" [shape=box style=filled fillcolor=lightgreen];
    "parallel-execution" [shape=box];
    "delegated-execution" [shape=box];
    "Single agent" [shape=box];

    "Multiple tasks?" -> "Need peer-to-peer collaboration?" [label="yes"];
    "Multiple tasks?" -> "Single agent" [label="no"];
    "Need peer-to-peer collaboration?" -> "Agent Teams" [label="yes - teammates must share discoveries"];
    "Need peer-to-peer collaboration?" -> "Tasks independent?" [label="no"];
    "Tasks independent?" -> "parallel-execution" [label="yes - fire and forget"];
    "Tasks independent?" -> "delegated-execution" [label="no - sequential"];
}

Deploy Teams when:

  • Parallel research where discoveries from one investigation redirect another
  • Multi-module features where frontend/backend/tests must coordinate
  • Cross-layer changes requiring interface negotiation
  • Debugging with competing hypotheses that need to share evidence
  • Large refactoring across multiple subsystems with shared conventions

Do not use when:

  • Simple independent tasks (use godmode:parallel-execution)
  • Sequential dependent tasks (use godmode:delegated-execution)
  • Single-file or single-module changes
  • Agent Teams feature not enabled
  • Fewer than 2 genuinely collaborative tasks

Prerequisites

Agent Teams is an experimental feature. It must be enabled:

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=true

Without this, TeamCreate, TaskCreate, and SendMessage tools are unavailable.

Deferred tools: TeamCreate, TaskCreate, TaskUpdate, TaskGet, TaskList, SendMessage, and EnterWorktree are deferred tools. They must be loaded via ToolSearch before first use:

// Load all team tools at once
ToolSearch: { "query": "select:SendMessage,TaskCreate,TaskUpdate,TaskList,TaskGet,EnterWorktree" }

The Entry Protocol

BEFORE forming a team:

1. ENUMERATE: What are all the tasks?
2. MAP: Which tasks need information from other tasks?
3. COUNT: How many task pairs require shared discoveries?
4. DECIDE:
   - 0 pairs need collaboration -> Use parallel-execution
   - 1 pair needs collaboration -> Consider single agent or subagents
   - 2+ pairs need collaboration -> Use Agent Teams
5. ONLY THEN: Form the team

Skip any step = unnecessary team overhead

The Workflow

Step 1: Evaluate Task-Team Fit

Before reaching for TeamCreate, answer these:

QuestionIf YesIf No
Can a single agent handle this?Do that insteadContinue
Can parallel subagents handle this?Use parallel-executionContinue
Do agents need to share discoveries mid-work?TeamsSubagents
Do agents need to negotiate interfaces?TeamsSubagents
Is coordination overhead justified?TeamsSimpler approach

Step 2: Architect the Team

Create the team, then spawn teammates using the Agent tool:

// 1. Create the team
TeamCreate: {
  "team_name": "auth-refactor-team",
  "description": "Refactor authentication across API and frontend modules"
}

// 2. Spawn each teammate (2-5 max)
Agent: {
  "name": "backend-eng",
  "team_name": "auth-refactor-team",
  "prompt": "You are the backend engineer. Own all files in src/api/auth/. Implement the new token refresh flow. Coordinate with frontend-eng on the session interface.",
  "description": "Backend auth implementation",
  "mode": "bypassPermissions"
}

Agent: {
  "name": "frontend-eng",
  "team_name": "auth-refactor-team",
  "prompt": "You are the frontend engineer. Own all files in src/ui/auth/. Update login flow to use new token refresh. Coordinate with backend-eng on the session interface.",
  "description": "Frontend auth implementation",
  "mode": "bypassPermissions"
}

Teammate count guidance:

CountWhenCoordination Cost
2Two distinct modules that must agree on interfaceLow
3Frontend + backend + tests, or 3 independent subsystemsMedium
4Cross-cutting refactor with 4 modulesHigh
5Maximum — only for genuinely large effortsVery High

Step 2b: Isolate Workspaces

Each teammate should work in an isolated git worktree to prevent merge conflicts at the filesystem level:

// Called by each teammate at the start of their work
EnterWorktree: { "name": "auth-backend" }

EnterWorktree creates a separate git worktree so teammates edit files in parallel without stepping on each other. The team lead runs godmode:merge-protocol to integrate worktrees at the end.

Step 3: Define Tasks with Clear Boundaries

// Create tasks with clear scope and success criteria
TaskCreate: {
  "subject": "Implement token refresh endpoint",
  "description": "Build POST /api/auth/refresh that validates expired access tokens against refresh tokens. Return new access+refresh token pair. Files: src/api/auth/refresh.ts, src/api/auth/refresh.test.ts",
  "activeForm": "Implementing token refresh endpoint"
}

TaskCreate: {
  "subject": "Update login UI for token refresh",
  "description": "Update login flow to store refresh token and auto-refresh on 401. Files: src/ui/auth/login.tsx, src/ui/auth/session.ts",
  "activeForm": "Updating login UI for token refresh"
}

// Set up dependencies and assign owners
TaskUpdate: { "taskId": "2", "addBlockedBy": ["1"] }
TaskUpdate: { "taskId": "1", "owner": "backend-eng" }
TaskUpdate: { "taskId": "2", "owner": "frontend-eng" }

Task design rules:

  • Clear scope: exactly which files/modules this teammate owns
  • Dependencies: use addBlockedBy for tasks that must complete first
  • Success criteria: what "done" looks like in the description
  • File ownership: NO overlap between teammates (prevents merge conflicts)

File ownership is non-negotiable. Two teammates editing the same file guarantees merge conflicts. If a shared file needs changes from multiple perspectives, assign it to ONE teammate who coordinates with others via messaging.

Step 4: Coordinate as Team Lead

digraph team_lead_loop {
    "Check TaskList" [shape=box];
    "All tasks completed?" [shape=diamond];
    "Any teammate blocked?" [shape=diamond];
    "Send clarifying message" [shape=box];
    "Resolve dependency" [shape=box];
    "Integrate results" [shape=box style=filled fillcolor=lightgreen];

    "Check TaskList" -> "All tasks completed?";
    "All tasks completed?" -> "Integrate results" [label="yes"];
    "All tasks completed?" -> "Any teammate blocked?" [label="no"];
    "Any teammate blocked?" -> "Send clarifying message" [label="needs info"];
    "Any teammate blocked?" -> "Resolve dependency" [label="blocked by task"];
    "Any teammate blocked?" -> "Check TaskList" [label="no - working"];
    "Send clarifying message" -> "Check TaskList";
    "Resolve dependency" -> "Check TaskList";
}

As team lead:

  • Track task completion via TaskList (no parameters needed — just call it)
  • Relay discoveries between teammates using direct messages:
// Direct message to relay a discovery
SendMessage: {
  "type": "message",
  "recipient": "frontend-eng",
  "content": "backend-eng finished the refresh endpoint. The response shape is { accessToken, refreshToken, expiresIn }. You can unblock now.",
  "summary": "Refresh endpoint contract ready"
}
  • Unblock dependencies by marking prerequisite tasks complete:
TaskUpdate: { "taskId": "1", "status": "completed" }
  • Arbitrate conflicts if teammates disagree on approach
  • Use broadcast ONLY for team-wide critical issues (expensive — sends N messages):
// Broadcast — use sparingly, sends N separate messages
SendMessage: {
  "type": "broadcast",
  "content": "Convention change: all auth endpoints now return camelCase keys",
  "summary": "Auth API convention change"
}

Step 5: Integrate Results

When all tasks complete:

  1. Review all changes together for consistency
  2. Confirm no file conflicts between teammates
  3. Run full test suite
  4. Verify that cross-module interfaces agree
  5. REQUIRED SUB-PROTOCOL: Use godmode:merge-protocol

Team Patterns

See team-patterns.md in this directory for five documented team patterns:

  1. Exploration Team — Parallel investigation of different dimensions
  2. Feature Team — Multi-module development (frontend + backend + tests)
  3. Diagnosis Team — Competing hypotheses tested concurrently
  4. Inspection Team — Multi-perspective code review
  5. Migration Team — Cross-cutting convention changes across codebase

Each pattern includes: when to use, team structure, task design, coordination flow, and example.

Tools at a Glance

ToolPurposeKey ParametersWhen
TeamCreateForm teamteam_name, descriptionOnce at start
AgentSpawn teammatename, team_name, prompt, description, modeOnce per teammate
EnterWorktreeIsolate workspacenameEach teammate at start
TaskCreateAdd tasksubject, description, activeFormDuring setup
TaskListView all tasks*(none)*Monitoring
TaskGetGet task detailstaskIdBefore starting work
TaskUpdateClaim/complete/updatetaskId, status, owner, addBlockedByThroughout
SendMessageDirect messagetype: "message", recipient, content, summaryCoordination
SendMessageBroadcasttype: "broadcast", content, summaryCritical issues only
SendMessageShutdowntype: "shutdown_request", recipient, contentWinding down

Teammate Behavior: What to Expect

Teammates are independent Claude sessions. Key behaviors:

  • Idle between turns — This is normal. Teammates go idle when waiting for messages or tasks. They reactivate when messaged or when tasks become available.
  • Claim tasks — Teammates use TaskUpdate to assign themselves before starting work.
  • Message each other — Teammates can and should message each other directly, not just the team lead.
  • Self-directed — Once given a task, teammates work autonomously. Avoid micromanagement.

Cognitive Traps

RationalizationTruth
"Teams are always superior to subagents"Teams add coordination overhead. Deploy only when collaboration is necessary.
"More teammates = faster delivery"More teammates = more coordination. 3 focused teammates outperform 6 scattered ones.
"I'll form the team and figure out tasks later"Tasks MUST be designed before team formation. No tasks = idle teammates burning tokens.
"Teammates can share files"Shared files = merge conflicts. Assign clear ownership.
"Broadcast is fine for routine updates"Broadcast sends N messages. Use SendMessage to specific teammates.
"Dependencies are implied"Undefined dependencies = teammates stepping on each other.

Guardrails

Prohibited:

  • Forming teams for simple tasks (overhead not justified)
  • Allowing teammates to edit the same files (merge conflicts)
  • Skipping integration review after team completion
  • Ignoring teammate messages (they contain discoveries)
  • Forming teams with more than 5 teammates (coordination explodes)
  • Using broadcast for routine updates (use direct messages)
  • Starting a team without clearly designed tasks
  • Leaving dependencies undefined between tasks

Mandatory:

  • Define clear file ownership boundaries
  • Set up task dependencies with blockedBy
  • Review all changes together after completion
  • Run full test suite after integration
  • Use SendMessage for teammate-to-teammate coordination
  • Mark tasks complete when done (teammates check TaskList for available work)

Tool Invocation Reference

Complete parameter specifications for every Agent Teams tool. Copy these directly — parameter names are exact.

TeamCreate

Creates a named team. Call once before spawning teammates.

{
  "team_name": "feature-billing-team",
  "description": "Build subscription billing feature"
}
ParameterTypeRequiredNotes
team_namestringyesKebab-case identifier
descriptionstringyesWhat the team is working on

Agent (spawn teammate)

Spawns a teammate into an existing team. Each teammate is an independent Claude session.

{
  "name": "backend-eng",
  "team_name": "feature-billing-team",
  "prompt": "You are the backend engineer on this team. Own src/api/billing/. Build the Stripe webhook handler and subscription CRUD endpoints. Use godmode:test-first for implementation. Coordinate with frontend-eng on API contracts via SendMessage.",
  "description": "Backend billing implementation",
  "mode": "bypassPermissions"
}
ParameterTypeRequiredNotes
namestringyesTeammate identifier (used in SendMessage recipient)
team_namestringyesMust match TeamCreate team_name
promptstringyesFull instructions — include file ownership, scope, coordination rules
descriptionstringyesShort description shown in UI
modestringno"bypassPermissions" lets teammate work without permission prompts

EnterWorktree

Creates an isolated git worktree for a teammate. Prevents filesystem conflicts when multiple teammates edit files in parallel.

{
  "name": "billing-feature"
}
ParameterTypeRequiredNotes
namestringyesWorktree identifier — becomes the branch/directory name

TaskCreate

Adds a task to the shared task list visible to all teammates.

{
  "subject": "Implement Stripe webhook handler",
  "description": "Build POST /webhooks/stripe endpoint that handles subscription.created, subscription.updated, and subscription.deleted events. Files: src/api/billing/webhooks.ts, src/api/billing/webhooks.test.ts",
  "activeForm": "Implementing Stripe webhook handler"
}
ParameterTypeRequiredNotes
subjectstringyesImperative form title (e.g., "Fix auth bug")
descriptionstringyesFull scope, file list, acceptance criteria
activeFormstringnoPresent continuous form for spinner (e.g., "Fixing auth bug")

TaskUpdate

Updates a task's status, owner, or dependencies. Use for claiming, completing, and wiring up blocked-by relationships.

// Assign owner
{ "taskId": "1", "owner": "backend-eng" }

// Set dependency — task 3 cannot start until tasks 1 and 2 complete
{ "taskId": "3", "addBlockedBy": ["1", "2"] }

// Mark in progress
{ "taskId": "1", "status": "in_progress" }

// Mark complete
{ "taskId": "1", "status": "completed" }
ParameterTypeRequiredNotes
taskIdstringyesTask ID from TaskCreate/TaskList
statusstringno"pending", "in_progress", "completed", or "deleted"
ownerstringnoTeammate name to assign
addBlockedBystring[]noTask IDs that must complete before this task
addBlocksstring[]noTask IDs that this task blocks
subjectstringnoUpdated title
descriptionstringnoUpdated description
activeFormstringnoUpdated spinner text

TaskList

Returns all tasks with their status, owner, and blocked-by info. No parameters.

// No parameters — just call it
TaskList: {}

TaskGet

Retrieves full details of a specific task.

{ "taskId": "1" }
ParameterTypeRequiredNotes
taskIdstringyesTask ID to retrieve

SendMessage

Three message types for team communication.

Direct message (default — use this for most communication):

{
  "type": "message",
  "recipient": "backend-eng",
  "content": "The profile API needs a PUT /profile/preferences endpoint. Frontend needs to update preferences without replacing the full profile.",
  "summary": "API contract update request"
}

Broadcast (sends to ALL teammates — use sparingly):

{
  "type": "broadcast",
  "content": "Convention change: all API responses now use camelCase keys",
  "summary": "API convention change"
}

Shutdown request (graceful teammate shutdown):

{
  "type": "shutdown_request",
  "recipient": "backend-eng",
  "content": "All tasks complete, wrapping up"
}
ParameterTypeRequiredNotes
typestringyes"message", "broadcast", or "shutdown_request"
recipientstringfor message/shutdownTeammate name
contentstringyesMessage body
summarystringfor message/broadcast5-10 word preview shown in UI

Complete Team Setup Example

End-to-end example wiring up a 2-person feature team:

// 1. Create team
TeamCreate: { "team_name": "billing-team", "description": "Build subscription billing" }

// 2. Create tasks
TaskCreate: {
  "subject": "Build Stripe webhook handler",
  "description": "POST /webhooks/stripe handling subscription events. Files: src/api/billing/webhooks.ts",
  "activeForm": "Building Stripe webhook handler"
}
TaskCreate: {
  "subject": "Build billing dashboard UI",
  "description": "Subscription management page showing plan, usage, invoices. Files: src/ui/billing/",
  "activeForm": "Building billing dashboard"
}

// 3. Wire dependencies and assign owners
TaskUpdate: { "taskId": "2", "addBlockedBy": ["1"] }
TaskUpdate: { "taskId": "1", "owner": "backend-eng" }
TaskUpdate: { "taskId": "2", "owner": "frontend-eng" }

// 4. Spawn teammates (they auto-claim their assigned tasks)
Agent: {
  "name": "backend-eng",
  "team_name": "billing-team",
  "prompt": "You are the backend engineer. Own src/api/billing/. Start with task #1. Use EnterWorktree to isolate your workspace. Use godmode:test-first. Message frontend-eng with the API contract when done.",
  "description": "Backend billing implementation",
  "mode": "bypassPermissions"
}
Agent: {
  "name": "frontend-eng",
  "team_name": "billing-team",
  "prompt": "You are the frontend engineer. Own src/ui/billing/. Task #2 is blocked by #1. Wait for backend-eng to share the API contract, then implement. Use EnterWorktree to isolate your workspace.",
  "description": "Frontend billing implementation",
  "mode": "bypassPermissions"
}

// 5. Monitor as team lead
TaskList: {}

// 6. When all tasks complete, shutdown teammates
SendMessage: { "type": "shutdown_request", "recipient": "backend-eng", "content": "All done, wrapping up" }
SendMessage: { "type": "shutdown_request", "recipient": "frontend-eng", "content": "All done, wrapping up" }

Integration

Related orchestration protocols:

  • godmode:parallel-execution — For independent tasks without collaboration requirement
  • godmode:delegated-execution — For sequential tasks with review checkpoints
  • godmode:task-runner — For plan execution in a separate session

Teammates should use:

  • godmode:test-first — For implementation tasks
  • godmode:fault-diagnosis — For diagnosis team investigations
  • godmode:completion-gate — Before marking tasks complete

Required for completion:

  • godmode:merge-protocol — After all team tasks complete

适合场景

01

用户想查找某类 Agent Skill 时

02

需要根据任务场景推荐可安装能力包时

03

需要对比不同来源的安装命令和来源信息时

能力概览

能力 1

按任务关键词查找相关 Skills

能力 2

展示可复制的安装命令

能力 3

保留来源站点、仓库和原始说明,方便继续核验

能力 4

展示第三方安全扫描或审计结果

安装后应在对应宿主中按原始 README 的触发条件使用;具体调用方式请以来源页面和 README 为准。

平台分布

Codex

37.65%
按下载量换算34

Claude

26.86%
按下载量换算24

Cursor

16.68%
按下载量换算15

Gemini CLI

10.23%
按下载量换算9

安全审计

Gen Agent Trust Hub

可疑

Socket

通过

Snyk

可疑

权限和风险

敏感数据

该 Skill 可能接触密钥、Token、环境变量或敏感配置,应进入高风险复核队列,默认不自动发布。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。来源安全扫描存在 warning/failed 结果,不能写成本站确认安全。当前只有一个来源,正式发布前建议补源仓库或其他目录站核验。

来源信息

继续浏览同类 Skills