mirror of
https://github.com/catlog22/Claude-Code-Workflow.git
synced 2026-02-28 09:23:08 +08:00
Previously, parallel tasks assigned to the same role (e.g., multiple EXPLORE-* tasks with owner: 'explorer') executed serially because only one agent instance existed per role name. This adds conditional parallel agent spawning with instance-specific names (explorer-1, explorer-2) and --agent-name arg for role task discovery filtering. Affected skills: team-ultra-analyze, team-quality-assurance, team-brainstorm, team-issue. Single-task modes preserve backward compatibility with original agent names.
17 KiB
17 KiB
name, description, allowed-tools
| name | description | allowed-tools |
|---|---|---|
| team-ultra-analyze | Unified team skill for deep collaborative analysis. All roles invoke this skill with --role arg for role-specific execution. Triggers on "team ultra-analyze", "team analyze". | TeamCreate(*), TeamDelete(*), SendMessage(*), TaskCreate(*), TaskUpdate(*), TaskList(*), TaskGet(*), Task(*), AskUserQuestion(*), Read(*), Write(*), Edit(*), Bash(*), Glob(*), Grep(*) |
Team Ultra Analyze
深度协作分析团队技能。将单体分析工作流拆分为 5 角色协作:探索→分析→讨论→综合。支持 Quick/Standard/Deep 三种管道模式,通过讨论循环实现用户引导的渐进式理解深化。所有成员通过 --role=xxx 路由到角色执行逻辑。
Architecture Overview
┌─────────────────────────────────────────────────────────┐
│ Skill(skill="team-ultra-analyze", args="--role=xxx") │
└──────────────────────┬──────────────────────────────────┘
│ Role Router
┌──────────┬───────┼───────────┬───────────┐
↓ ↓ ↓ ↓ ↓
┌────────┐┌────────┐┌────────┐┌──────────┐┌───────────┐
│coordi- ││explorer││analyst ││discussant││synthesizer│
│nator ││EXPLORE-││ANALYZE-││DISCUSS-* ││SYNTH-* │
│ roles/ ││* roles/││* roles/││ roles/ ││ roles/ │
└────────┘└────────┘└────────┘└──────────┘└───────────┘
Command Architecture
roles/
├── coordinator/
│ ├── role.md # 编排:话题澄清、管道选择、讨论循环、结果汇报
│ └── commands/
│ ├── dispatch.md # 任务链创建与依赖管理
│ └── monitor.md # 进度监控 + 讨论循环
├── explorer/
│ ├── role.md # 代码库探索
│ └── commands/
│ └── explore.md # cli-explore-agent 并行探索
├── analyst/
│ ├── role.md # 深度分析
│ └── commands/
│ └── analyze.md # CLI 多视角分析
├── discussant/
│ ├── role.md # 讨论处理 + 方向调整
│ └── commands/
│ └── deepen.md # 深入探索
└── synthesizer/
├── role.md # 综合结论
└── commands/
└── synthesize.md # 跨视角整合
设计原则: role.md 保留 Phase 1(Task Discovery)和 Phase 5(Report)内联。Phase 2-4 根据复杂度决定内联或委派到 commands/*.md。
Role Router
Input Parsing
Parse $ARGUMENTS to extract --role and optional --agent-name:
const args = "$ARGUMENTS"
const roleMatch = args.match(/--role[=\s]+(\w+)/)
if (!roleMatch) {
throw new Error("Missing --role argument. Available roles: coordinator, explorer, analyst, discussant, synthesizer")
}
const role = roleMatch[1]
const teamName = args.match(/--team[=\s]+([\w-]+)/)?.[1] || "ultra-analyze"
// --agent-name for parallel instances (e.g., explorer-1, analyst-2)
// Passed through to role.md for task discovery filtering
const agentName = args.match(/--agent-name[=\s]+([\w-]+)/)?.[1] || role
Role Dispatch
const VALID_ROLES = {
"coordinator": { file: "roles/coordinator/role.md", prefix: null },
"explorer": { file: "roles/explorer/role.md", prefix: "EXPLORE" },
"analyst": { file: "roles/analyst/role.md", prefix: "ANALYZE" },
"discussant": { file: "roles/discussant/role.md", prefix: "DISCUSS" },
"synthesizer": { file: "roles/synthesizer/role.md", prefix: "SYNTH" }
}
if (!VALID_ROLES[role]) {
throw new Error(`Unknown role: ${role}. Available: ${Object.keys(VALID_ROLES).join(', ')}`)
}
// Read and execute role-specific logic
Read(VALID_ROLES[role].file)
// → Execute the 5-phase process defined in that file
Available Roles
| Role | Task Prefix | Responsibility | Role File |
|---|---|---|---|
coordinator |
N/A | 话题澄清、管道选择、会话管理、讨论循环 | roles/coordinator/role.md |
explorer |
EXPLORE-* | cli-explore-agent 多角度并行代码库探索 | roles/explorer/role.md |
analyst |
ANALYZE-* | CLI 多视角深度分析 | roles/analyst/role.md |
discussant |
DISCUSS-* | 用户反馈处理、方向调整、深入分析 | roles/discussant/role.md |
synthesizer |
SYNTH-* | 跨视角整合、结论生成、决策追踪 | roles/synthesizer/role.md |
Shared Infrastructure
Role Isolation Rules
核心原则: 每个角色仅能执行自己职责范围内的工作。
Output Tagging(强制)
所有角色的输出必须带 [role_name] 标识前缀:
SendMessage({ content: `## [${role}] ...`, summary: `[${role}] ...` })
mcp__ccw-tools__team_msg({ summary: `[${role}] ...` })
Coordinator 隔离
| 允许 | 禁止 |
|---|---|
| 话题澄清 (AskUserQuestion) | ❌ 直接执行代码探索或分析 |
| 创建任务链 (TaskCreate) | ❌ 直接调用 cli-explore-agent |
| 管道选择 + 讨论循环驱动 | ❌ 直接调用 CLI 分析工具 |
| 监控进度 (消息总线) | ❌ 绕过 worker 自行完成 |
Worker 隔离
| 允许 | 禁止 |
|---|---|
| 处理自己前缀的任务 | ❌ 处理其他角色前缀的任务 |
| 读写 shared-memory.json (自己的字段) | ❌ 为其他角色创建任务 |
| SendMessage 给 coordinator | ❌ 直接与其他 worker 通信 |
Team Configuration
const TEAM_CONFIG = {
name: "ultra-analyze",
sessionDir: ".workflow/.team/UAN-{slug}-{date}/",
sharedMemory: "shared-memory.json",
analysisDimensions: ["architecture", "implementation", "performance", "security", "concept", "comparison", "decision"],
maxDiscussionRounds: 5
}
Shared Memory(核心产物)
// 各角色读取共享记忆
const memoryPath = `${sessionFolder}/shared-memory.json`
let sharedMemory = {}
try { sharedMemory = JSON.parse(Read(memoryPath)) } catch {}
// 各角色写入自己负责的字段:
// explorer → sharedMemory.explorations
// analyst → sharedMemory.analyses
// discussant → sharedMemory.discussions
// synthesizer → sharedMemory.synthesis
// coordinator → sharedMemory.decision_trail + current_understanding
Write(memoryPath, JSON.stringify(sharedMemory, null, 2))
Message Bus (All Roles)
mcp__ccw-tools__team_msg({
operation: "log",
team: teamName,
from: role,
to: "coordinator",
type: "<type>",
summary: `[${role}] <summary>`,
ref: "<file_path>"
})
| Role | Types |
|---|---|
| coordinator | pipeline_selected, discussion_round, direction_adjusted, task_unblocked, error, shutdown |
| explorer | exploration_ready, error |
| analyst | analysis_ready, error |
| discussant | discussion_processed, error |
| synthesizer | synthesis_ready, error |
CLI 回退
Bash(`ccw team log --team "${teamName}" --from "${role}" --to "coordinator" --type "<type>" --summary "<摘要>" --json`)
Task Lifecycle (All Worker Roles)
const tasks = TaskList()
const myTasks = tasks.filter(t =>
t.subject.startsWith(`${VALID_ROLES[role].prefix}-`) &&
t.owner === role &&
t.status === 'pending' &&
t.blockedBy.length === 0
)
if (myTasks.length === 0) return
const task = TaskGet({ taskId: myTasks[0].id })
TaskUpdate({ taskId: task.id, status: 'in_progress' })
// Phase 2-4: Role-specific
// Phase 5: Report + Loop
mcp__ccw-tools__team_msg({ operation: "log", team: teamName, from: role, to: "coordinator", type: "...", summary: `[${role}] ...` })
SendMessage({ type: "message", recipient: "coordinator", content: `## [${role}] ...`, summary: `[${role}] ...` })
TaskUpdate({ taskId: task.id, status: 'completed' })
Three-Mode Pipeline Architecture
Quick: EXPLORE-001 → ANALYZE-001 → SYNTH-001
Standard: [EXPLORE-001..N](parallel) → [ANALYZE-001..N](parallel) → DISCUSS-001 → SYNTH-001
Deep: [EXPLORE-001..N] → [ANALYZE-001..N] → DISCUSS-001 → ANALYZE-fix → DISCUSS-002 → ... → SYNTH-001
Mode Auto-Detection
function detectPipelineMode(args, taskDescription) {
const modeMatch = args.match(/--mode[=\s]+(quick|standard|deep)/)
if (modeMatch) return modeMatch[1]
// 自动检测
if (/快速|quick|overview|概览/.test(taskDescription)) return 'quick'
if (/深入|deep|thorough|详细|全面/.test(taskDescription)) return 'deep'
return 'standard'
}
Discussion Loop (Deep Mode)
coordinator(AskUser) → DISCUSS-N(deepen) → [optional ANALYZE-fix] → coordinator(AskUser) → ... → SYNTH
Decision Recording Protocol
⚠️ CRITICAL: 继承自原 analyze-with-file 命令。分析过程中以下情况必须立即记录到 discussion.md:
| Trigger | What to Record | Target Section |
|---|---|---|
| Direction choice | 选择了什么、为什么、放弃了哪些替代方案 | #### Decision Log |
| Key finding | 发现内容、影响范围、置信度 | #### Key Findings |
| Assumption change | 旧假设→新理解、变更原因、影响 | #### Corrected Assumptions |
| User feedback | 用户原始输入、采纳/调整理由 | #### User Input |
Unified Session Directory
.workflow/.team/UAN-{slug}-{YYYY-MM-DD}/
├── shared-memory.json # 探索/分析/讨论/综合 共享记忆
├── discussion.md # ⭐ 理解演进 & 讨论时间线
├── explorations/ # Explorer output
│ ├── exploration-001.json
│ └── exploration-002.json
├── analyses/ # Analyst output
│ ├── analysis-001.json
│ └── analysis-002.json
├── discussions/ # Discussant output
│ └── discussion-round-001.json
└── conclusions.json # Synthesizer output
Coordinator Spawn Template
TeamCreate({ team_name: teamName })
// ── Determine parallel agent count ──
const perspectiveCount = selectedPerspectives.length
const isParallel = perspectiveCount > 1
// ── Explorers ──
// Quick mode (1 perspective): single "explorer"
// Standard/Deep mode (N perspectives): "explorer-1", "explorer-2", ...
if (isParallel) {
for (let i = 0; i < perspectiveCount; i++) {
const agentName = `explorer-${i + 1}`
Task({
subagent_type: "general-purpose",
team_name: teamName,
name: agentName,
prompt: `你是 team "${teamName}" 的 EXPLORER (${agentName})。
你的 agent 名称是 "${agentName}",任务发现时用此名称匹配 owner。
当你收到 EXPLORE-* 任务时,调用 Skill(skill="team-ultra-analyze", args="--role=explorer --agent-name=${agentName}") 执行。
当前需求: ${taskDescription}
约束: ${constraints}
## 角色准则(强制)
- 你只能处理 owner 为 "${agentName}" 的 EXPLORE-* 前缀任务
- 所有输出(SendMessage、team_msg)必须带 [explorer] 标识前缀
- 仅与 coordinator 通信,不得直接联系其他 worker
- 不得使用 TaskCreate 为其他角色创建任务
## 消息总线(必须)
每次 SendMessage 前,先调用 mcp__ccw-tools__team_msg 记录。
工作流程:
1. TaskList → 找到 owner === "${agentName}" 的 EXPLORE-* 任务
2. Skill(skill="team-ultra-analyze", args="--role=explorer --agent-name=${agentName}") 执行
3. team_msg log + SendMessage 结果给 coordinator(带 [explorer] 标识)
4. TaskUpdate completed → 检查下一个任务`
})
}
} else {
// Single explorer for quick mode
Task({
subagent_type: "general-purpose",
team_name: teamName,
name: "explorer",
prompt: `你是 team "${teamName}" 的 EXPLORER。
当你收到 EXPLORE-* 任务时,调用 Skill(skill="team-ultra-analyze", args="--role=explorer") 执行。
当前需求: ${taskDescription}
约束: ${constraints}
## 角色准则(强制)
- 你只能处理 EXPLORE-* 前缀的任务,不得执行其他角色的工作
- 所有输出(SendMessage、team_msg)必须带 [explorer] 标识前缀
- 仅与 coordinator 通信,不得直接联系其他 worker
- 不得使用 TaskCreate 为其他角色创建任务
## 消息总线(必须)
每次 SendMessage 前,先调用 mcp__ccw-tools__team_msg 记录。
工作流程:
1. TaskList → 找到 EXPLORE-* 任务
2. Skill(skill="team-ultra-analyze", args="--role=explorer") 执行
3. team_msg log + SendMessage 结果给 coordinator(带 [explorer] 标识)
4. TaskUpdate completed → 检查下一个任务`
})
}
// ── Analysts ──
// Same pattern: parallel mode spawns N analysts, quick mode spawns 1
if (isParallel) {
for (let i = 0; i < perspectiveCount; i++) {
const agentName = `analyst-${i + 1}`
Task({
subagent_type: "general-purpose",
team_name: teamName,
name: agentName,
prompt: `你是 team "${teamName}" 的 ANALYST (${agentName})。
你的 agent 名称是 "${agentName}",任务发现时用此名称匹配 owner。
当你收到 ANALYZE-* 任务时,调用 Skill(skill="team-ultra-analyze", args="--role=analyst --agent-name=${agentName}") 执行。
当前需求: ${taskDescription}
约束: ${constraints}
## 角色准则(强制)
- 你只能处理 owner 为 "${agentName}" 的 ANALYZE-* 前缀任务
- 所有输出必须带 [analyst] 标识前缀
- 仅与 coordinator 通信
## 消息总线(必须)
每次 SendMessage 前,先调用 mcp__ccw-tools__team_msg 记录。
工作流程:
1. TaskList → 找到 owner === "${agentName}" 的 ANALYZE-* 任务
2. Skill(skill="team-ultra-analyze", args="--role=analyst --agent-name=${agentName}") 执行
3. team_msg log + SendMessage 结果给 coordinator
4. TaskUpdate completed → 检查下一个任务`
})
}
} else {
Task({
subagent_type: "general-purpose",
team_name: teamName,
name: "analyst",
prompt: `你是 team "${teamName}" 的 ANALYST。
当你收到 ANALYZE-* 任务时,调用 Skill(skill="team-ultra-analyze", args="--role=analyst") 执行。
当前需求: ${taskDescription}
约束: ${constraints}
## 角色准则(强制)
- 你只能处理 ANALYZE-* 前缀的任务
- 所有输出必须带 [analyst] 标识前缀
- 仅与 coordinator 通信
## 消息总线(必须)
每次 SendMessage 前,先调用 mcp__ccw-tools__team_msg 记录。
工作流程:
1. TaskList → 找到 ANALYZE-* 任务
2. Skill(skill="team-ultra-analyze", args="--role=analyst") 执行
3. team_msg log + SendMessage 结果给 coordinator
4. TaskUpdate completed → 检查下一个任务`
})
}
// ── Discussant (always single) ──
Task({
subagent_type: "general-purpose",
team_name: teamName,
name: "discussant",
prompt: `你是 team "${teamName}" 的 DISCUSSANT。
当你收到 DISCUSS-* 任务时,调用 Skill(skill="team-ultra-analyze", args="--role=discussant") 执行。
当前需求: ${taskDescription}
## 角色准则(强制)
- 你只能处理 DISCUSS-* 前缀的任务
- 所有输出必须带 [discussant] 标识前缀
## 消息总线(必须)
每次 SendMessage 前,先调用 mcp__ccw-tools__team_msg 记录。
工作流程:
1. TaskList → 找到 DISCUSS-* 任务
2. Skill(skill="team-ultra-analyze", args="--role=discussant") 执行
3. team_msg log + SendMessage
4. TaskUpdate completed → 检查下一个任务`
})
// ── Synthesizer (always single) ──
Task({
subagent_type: "general-purpose",
team_name: teamName,
name: "synthesizer",
prompt: `你是 team "${teamName}" 的 SYNTHESIZER。
当你收到 SYNTH-* 任务时,调用 Skill(skill="team-ultra-analyze", args="--role=synthesizer") 执行。
当前需求: ${taskDescription}
## 角色准则(强制)
- 你只能处理 SYNTH-* 前缀的任务
- 所有输出必须带 [synthesizer] 标识前缀
## 消息总线(必须)
每次 SendMessage 前,先调用 mcp__ccw-tools__team_msg 记录。
工作流程:
1. TaskList → 找到 SYNTH-* 任务
2. Skill(skill="team-ultra-analyze", args="--role=synthesizer") 执行
3. team_msg log + SendMessage
4. TaskUpdate completed → 检查下一个任务`
})
Error Handling
| Scenario | Resolution |
|---|---|
| Unknown --role value | Error with available role list |
| Missing --role arg | Error with usage hint |
| Role file not found | Error with expected path (roles/{name}/role.md) |
| Task prefix conflict | Log warning, proceed |
| Discussion loop stuck >5 rounds | Force synthesis, offer continuation |
| CLI tool unavailable | Fallback chain: gemini → codex → manual analysis |
| Explorer agent fails | Continue with available context, note limitation |