claw-orchestra

OpenClaw native multi-agent orchestrator. Based on AOrchestra 4-tuple (I,C,T,M) abstraction. Dynamically creates sub-agents, parallel execution, smart routin...

Security Scan

The skill's code, instructions, and requirements are internally consistent with its stated purpose as a multi-agent orchestrator; it requests no credentials or external installs, but it does include file-write behavior and names tools like exec/read/write that you should review in the context of your OpenClaw runtime policy.

Purpose & Capability

Name/description (multi-agent orchestrator) aligns with the provided code and SKILL.md: the package implements AgentTuple, Orchestrator, OpenClawAdapter, routing, a learner (cost/experience), and explicit integration points for sessions_spawn and sessions_history. No unrelated credentials, binaries, or external cloud access are requested.

Instruction Scope

SKILL.md instructs the orchestrator to spawn subagents, poll subagent status via 'subagents list' / 'sessions_history', integrate results immediately, and deliver a Markdown file to the user. These steps are coherent with orchestration behavior, but the policy of auto-generating and sending files (message(filePath=...)) is explicit and may be surprising — it causes the skill to write a file and transmit it to the user without further prompting.

Install Mechanism

No install spec is provided (instruction-only skill with bundled source). All code is included in the skill bundle; there are no external downloads, remote installers, or third-party package fetches that would increase risk.

Credentials

The skill does not request environment variables or credentials. However, runtime components do perform local file I/O: CostTracker defaults to '/workspace/projects/claw-orchestra/costs.json' and will create directories/files there. The AgentTuple 'coder' factory lists tools ['exec','read','write','edit'] (tool names only); if your OpenClaw environment exposes such tools to spawned subagents, child agents could execute commands or access files. These behaviors are proportionate to a local orchestrator but you should confirm platform tool permissions and file path policies.

Persistence & Privilege

The skill does not set always=true and does not request persistent system-wide privileges. It keeps its state within provided objects and writes to its own configured files. There is no code attempting to modify other skills or global agent settings.

Assessment

This skill appears to implement exactly what it claims: an OpenClaw multi-agent orchestrator that spawns and integrates subagents. It requires no secrets and has no external installers, which is good. Before installing, check two things in your OpenClaw environment: (1) what tools the platform exposes to spawned subagents — if tools named exec/read/write are available, child agents could run commands or access files, so restrict those tools or audit prompts you send to the orchestrator; (2) the skill's default cost-tracker path (/workspace/projects/claw-orchestra/costs.json) will create and write files — change the path or ensure the skill cannot overwrite sensitive locations. Also be aware the SKILL.md mandates auto-generating and sending a Markdown report file; if you don't want automatic file transmission, modify that behavior or supervise first runs. Overall the package is coherent and not requesting unexpected privileges, but review runtime tool permissions and file-write locations before use.

OpenClaw 原生的多智能体编排器 —— 用乐团指挥的方式协调多个 Agent 完成复杂任务。

---

⚠️ 核心规则（必读）

1. 结果交付方式

必须生成 MD 文件发送给用户，不要在聊天中输出长文本！

2. 子Agent完成后必须主动整合

最后一个子Agent完成后，主Agent必须立即整合结果并发送报告，不能等待用户催促！

3. 动态四元组必须突出

每次编排都要展示四元组的动态生成过程，让用户感受到"智能编排"的价值。

---

🎯 编排流程

步骤1：启动公告

🎼 ClawOrchestra 动态编排器已激活

📋 任务: [用户任务简述]
🎯 类型: 调研 | 复杂度: 中等 | 模式: 🔀 并行

步骤2：动态四元组生成

⚡ 动态四元组生成中...

┌─────────────────────────────────────────────────────┐
│  Φ = (I, C, T, M)  动态子Agent创建                   │
├─────────────────────────────────────────────────────┤
│  Agent-1: 🔍 Swarm专家                              │
│  ├── I: "搜索 Agent Swarm 最新框架和论文"            │
│  ├── C: [主任务上下文]                               │
│  ├── T: [web_search, web_fetch]                     │
│  └── M: GLM (快速搜索，成本低)                       │
├─────────────────────────────────────────────────────┤
│  Agent-2: 🎯 编排专家                               │
│  ├── I: "搜索 Agent 编排器技术和趋势"                │
│  ├── C: [主任务上下文]                               │
│  ├── T: [web_search, web_fetch]                     │
│  └── M: GLM (快速搜索)                               │
├─────────────────────────────────────────────────────┤
│  Agent-3: 📊 协作专家                               │
│  ├── I: "搜索多智能体协作模式"                       │
│  ├── C: [主任务上下文]                               │
│  ├── T: [web_search, web_fetch]                     │
│  └── M: Kimi (深度分析，长上下文)                    │
└─────────────────────────────────────────────────────┘

✨ 智能决策: 3个子Agent并行执行，预计节省 60% 时间

步骤3：派遣子Agent

🚀 子Agent小队出发 (并行模式)

[1] 🔍 Swarm专家  → GLM  → Agent Swarm 调研
[2] 🎯 编排专家   → GLM  → 编排器技术调研
[3] 📊 协作专家   → Kimi → 多智能体协作分析

⏳ 执行中... (预计 60-90s)

步骤4：等待并整合（关键！）

主Agent必须轮询等待所有子Agent完成，然后立即整合！

⏳ 监控子Agent状态...

✅ Swarm专家 完成 (65s, 90k tokens)
✅ 编排专家 完成 (93s, 71k tokens)
✅ 协作专家 完成 (65s, 117k tokens)

📊 全部完成！正在整合结果...

步骤5：生成并发送报告

📝 生成调研报告...

📊 执行统计:
| Agent | 模型 | 耗时 | Tokens |
|-------|------|------|--------|
| 🔍 Swarm专家 | GLM | 65s | 90k |
| 🎯 编排专家 | GLM | 93s | 71k |
| 📊 协作专家 | Kimi | 65s | 117k |

⚡ 并行节省: 67% 时间
💰 总成本: ~278k tokens

📤 正在发送报告文件...

---

🔧 技术实现要点

1. 并行执行

同一轮 sessions_spawn 调用 = 真并行

2. 结果等待

使用 subagents list 轮询状态，直到所有子Agent完成

3. 结果获取

使用 sessions_history 获取子Agent输出

4. 报告生成

整合结果 → 生成 MD 文件 → message(filePath=...) 发送

---

🤖 模型选择

任务类型	推荐模型	原因
搜索、收集信息	GLM	便宜、快速、中文好
代码、分析、长文	Kimi	长上下文、代码强
复杂推理	GLM	均衡

注意：目前仅支持 lixiang 内部模型（GLM、Kimi）

---

📚 参考文献

• AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration
• GitHub: https://github.com/zcyynl/claw-orchestra

---

🗓️ 更新日志

v0.1.1 (2026-03-16)

• ✅ 四元组抽象 (I, C, T, M)
• ✅ 编排器骨架 (Delegate/Finish)
• ✅ OpenClaw 适配器 (sessions_spawn)
• ✅ 智能路由 (上下文/模型/工具)
• ✅ 经验库 + 成本追踪
• ✅ 交互优化（动态四元组展示）
• ✅ 自动结果整合（不再卡住）