Back to Details

coordinator

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Coordinator

Coordinator

Overview

概述

Manages two concerns: output mode switching (voice vs text) and agent team orchestration for parallel task execution.
负责管理两大事项：输出模式切换（语音/文本）和用于并行任务执行的Agent团队编排。

Voice/Text Mode Toggle

语音/文本模式切换

Trigger words: "vm", "voice mode" -> switch to VOICE. "tm", "text mode" -> switch to TEXT.
触发词：“vm”、“voice mode” -> 切换至VOICE模式；“tm”、“text mode” -> 切换至TEXT模式。

VOICE Mode Rules

VOICE模式规则

  • Short sentences, max 2 lines each
  • Bullet points for all lists
  • No code blocks longer than 3 lines — summarize verbally instead
  • Confirm before executing any tool: "I'll [action]. Go ahead?"
  • Use plain language, avoid jargon
  • When showing code changes, describe what changed rather than showing diffs
  • 短句表达，每条最多2行
  • 所有列表均使用项目符号
  • 代码块长度不超过3行——改用口头总结
  • 执行任何工具前需确认：“我将[执行操作]。可以继续吗？”
  • 使用平实语言，避免行话
  • 展示代码变更时，描述变更内容而非直接显示差异

TEXT Mode Rules (Default)

TEXT模式规则（默认）

  • Full detailed responses with code blocks
  • Show diffs, full implementations, detailed explanations
  • Execute tools without pre-confirmation (unless destructive)
  • 提供完整详细的回复，包含代码块
  • 显示差异、完整实现方案和详细说明
  • 无需预先确认即可执行工具（除非操作具有破坏性）

Agent Team Orchestration

Agent团队编排

When to Use Teams

何时使用团队

  • Task has 3+ independent subtasks
  • Different subtasks need different expertise (e.g., implementation + testing + review)
  • Work can be parallelized without merge conflicts
  • 任务包含3个及以上独立子任务
  • 不同子任务需要不同专业能力（例如：实现+测试+评审）
  • 工作可并行开展且不会产生合并冲突

Orchestration Details

编排细节

Read references/team-orchestration.md for the full coordination pattern (DECOMPOSE -> WAVE PLAN -> DISPATCH -> GATE -> VERIFY), agent spawn rules, model selection table, context budget management, and conflict resolution.
阅读references/team-orchestration.md获取完整的协作模式（分解 -> 阶段规划 -> 分派 -> 管控 -> 验证）、Agent生成规则、模型选择表、上下文预算管理以及冲突解决方法。

Plan-Driven Orchestration

基于计划的编排

When orchestrating execution of a structured plan (
.claude/plans/<plan-id>/
):
当编排结构化计划（
.claude/plans/<plan-id>/
）的执行时：

At Session Start

会话开始时

Read 
plan.md
and 
manifest.json
. The manifest defines waves and task dependencies. Use the manifest's wave grouping directly — do not re-derive wave structure.
读取
plan.md
和
manifest.json
。清单文件定义了阶段划分和任务依赖关系。直接使用清单中的阶段分组——不要重新推导阶段结构。

At Each Wave Boundary

每个阶段边界处

Re-read 
manifest.json
from disk. This protects against context compaction loss. Update task statuses in the manifest after each wave completes.
重新从磁盘读取
manifest.json
。这可防止上下文压缩导致的状态丢失。每个阶段完成后，更新清单中的任务状态。

Agent Dispatch

Agent分派

Point each agent at its briefing file. Do NOT paste briefing content inline. Include in dispatch prompt: plan-id, task-id, briefing path, working directory.
为每个Agent指定对应的简报文件。不要将简报内容直接粘贴到提示词中。分派提示词需包含：plan-id、task-id、简报路径、工作目录。

After Final Wave

最终阶段完成后

  1. Optionally write 
    .claude/plans/<plan-id>/summary.md
    with execution notes
  2. Delete the plan directory: 
    rm -rf .claude/plans/<plan-id>/
  3. If deletion fails, warn but do not block
  1. 可选：编写
    .claude/plans/<plan-id>/summary.md
    记录执行说明
  2. 删除计划目录：
    rm -rf .claude/plans/<plan-id>/
  3. 若删除失败，发出警告但无需阻塞后续操作

Common Mistakes

常见错误

  • Spawning agents without full context in the prompt
  • Letting multiple agents modify the same file
  • Skipping verification after merging agent outputs
  • Using teams for tasks that are naturally sequential
  • Putting dependent tasks in the same wave (task B needs task A's output -> different waves)
  • Pasting full briefing text inline instead of pointing agents to their briefing file
  • Not re-reading manifest at wave boundaries (context compaction can lose state)
  • 在提示词中未包含完整上下文就生成Agent
  • 让多个Agent修改同一文件
  • 合并Agent输出后跳过验证步骤
  • 为天然需串行执行的任务使用团队模式
  • 将存在依赖的任务放在同一阶段（任务B依赖任务A的输出 -> 应分属不同阶段）
  • 将完整简报文本直接粘贴到提示词中，而非引导Agent查看简报文件
  • 未在阶段边界处重新读取清单文件（上下文压缩可能导致状态丢失）