← Back to Skills Marketplace
xiaxianb

agent-system

by XIAXIANB · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ⚠ suspicious
80
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install agent-system
Description
OpenClaw 核心 Agent 调度系统。当用户描述需要"分析"、"规划"、"拆解任务"、"多步骤处理"、"自动执行复杂任务"时激活。 基于复杂度自动选择最优执行路径。
README (SKILL.md)

OpenClaw Agent System

核心概念

这是一个智能调度框架,包含:

  • 🎯 Orchestrator - 调度中枢,根据复杂度路由
  • 📋 Planner - 任务规划 + 复杂度评估
  • Executor - 任务执行
  • 🔍 Reviewer - 质量审查
  • 🔧 Self-Heal - 自愈重试
  • 📊 Metrics - 指标追踪

调度规则

复杂度阈值

复杂度 路径
≤ 3 Executor → 直接输出
4-5 Executor → Reviewer → 输出
> 5 Planner → Executor → Reviewer → [Heal] → 输出

复杂度评估

高复杂度指标(满足任一即提升复杂度):

  • 包含"分析"、"原因"、"为什么"
  • 包含"比较"、"规划"、"设计"
  • 包含"系统"、"复杂"、"全面"、"详细"

复杂度等级

  • 1-3:简单任务(直接执行)
  • 4-5:中等任务(执行+审查)
  • 6-10:复杂任务(规划+执行+审查+自愈)

执行流程

用户输入
    ↓
┌─────────────────┐
│   Planner       │ ← 意图识别 + 复杂度评估
│   complexity?   │
└────────┬────────┘
         │
    ┌────┴────┐
   ≤3        >3
    │         │
    ▼         ▼
 Executor   Planner
    │         ↓
    │      [任务拆解]
    │         ↓
    │      Executor
    │      (多任务)
    │         ↓
    │      Reviewer
    │      (质量评分)
    │         ↓
    │    score \x3C 70?
    │         │
    │    ┌────┴────┐
    │    │         │
    │   Yes       No
    │    │         │
    │    ▼         ▼
    │  Self-Heal  输出
    │    ↓
    │  重试 \x3C 3?
    │    │
    │  ┌─┴────────┐
    │ │          │
    │ Yes       No → degraded_mode
    │  │
    │  └──────────→ 重试
    │
    └──────────────→ 输出

意图识别

自动识别用户意图:

Intent 关键词
analysis 分析、原因、为什么、分析一下
generation 写、生成、创建、编写
coding 代码、编程、写程序、开发
decision 选择、决定、哪个好、比较
planning 计划、安排、规划

输出规范

标准响应格式

{
  "success": true,
  "content": "执行结果内容",
  "plan": {
    "intent": "analysis | generation | ...",
    "complexity": 1-10,
    "tasks": [
      {
        "task_id": "t1",
        "type": "analysis | generate | transform | validate",
        "description": "任务描述",
        "input": "输入内容",
        "expected_output": "预期输出"
      }
    ]
  },
  "metrics": {
    "tokens": 1500,
    "truncated": false,
    "quality_score": 85,
    "heal_triggered": false,
    "degraded_mode": false
  }
}

质量阈值

  • score ≥ 70 → 直接输出
  • score \x3C 70 → 触发 Self-Heal
  • 重试次数 ≥ 2 → 进入 degraded_mode(降级模式)

自愈规则

触发条件

满足任一即触发自愈:

  • score \x3C 70
  • 检测到占位符 {xxx}
  • 检测到禁止提问语句

重试策略

重试次数 策略 说明
第1次 detailed_reasoning 增加思考链长度
第2次 reverse_reasoning 反向推理(从目标倒推)
第3次 degraded_mode 降级,输出简化结果

禁止提问规则

不允许出现以下语句:

  • "请提供…"
  • "你可以告诉我…"
  • "我需要更多信息…"

条件分支

  • 若信息足以完成任务 → 不得提问,直接执行
  • 若信息不足以完成任务 → 填充默认值或标注缺失项,继续执行
  • 绝对不卡死等待用户回复

推断结论标注

当信息不足需要推断时,必须标注置信度:

[推断结论 · 置信度: 高/中/低 · 待数据验证]

Token 控制

  • 最大输出:2000 tokens(约4000字符)
  • 超长输出自动截断,标注 [已截断]

日志格式

所有 Agent 日志输出 JSON:

{
  "timestamp": "2026-04-02T11:40:00+08:00",
  "level": "INFO | WARNING | ERROR",
  "module": "orchestrator | planner | executor | reviewer | self_heal",
  "event": "事件名称",
  "details": {}
}

日志级别

  • INFO:正常流程(开始/完成/切换)
  • WARNING:异常但未失败(占位符已填充/置信度低)
  • ERROR:执行失败/触发自愈

指标追踪

每次执行记录:

  • error_rate - 错误率
  • output_score - 质量评分
  • heal_trigger_count - 自愈触发次数
  • degraded_mode_count - 降级模式次数

参考实现

src/orchestrator.js 包含完整的参考实现,可作为:

  1. 独立调度服务运行
  2. 代码级参考
  3. 未来集成到 OpenClaw 内核

使用示例

简单任务

用户:"今天天气怎么样"
→ complexity = 3 → Executor → 直接输出

复杂任务

用户:"帮我分析销售下滑的原因"
→ complexity = 6 → Planner → Executor × 3 → Reviewer → [自愈] → 输出

中等任务

用户:"帮我写一封道歉邮件"
→ complexity = 4 → Executor → Reviewer → 输出


本 Skill 提供工作流指导,实际执行由 AI 基于上下文判断

Usage Guidance
What to consider before installing: - The skill's high-level purpose (an autonomous orchestrator) matches the code and docs, but there are serious inconsistencies that will likely break integration or produce surprising behavior. Notable problems: index.js references a non-existent export, complexity is documented as a numeric 1–10 but the planner returns 'low'/'medium'/'high' strings, and schema fields (metrics, plan shape) don't line up with actual return values. These will cause runtime errors or mismatched contracts. - Policy/design concern: the skill forces 'do not ask clarifying questions' and requires auto-filling defaults and always producing output. That design increases the chance of hallucinated/incorrect results and could be unsafe if the agent is allowed to act autonomously on ambiguous or sensitive tasks. Consider whether you want the agent to proceed without asking. - Operational safety: no credentials or network calls appear in the code, which reduces exfiltration risk, but the code is runnable JS — review it, run tests in an isolated environment, and do a security audit before enabling the skill for production/autonomous use. - Recommended actions before enabling broadly: 1) Fix the code/docs/schema mismatches (export names, complexity type, metrics field names) and ensure files are complete. 2) Add/verify unit tests that exercise index.js tools (agent_dispatch/agent_plan/agent_review) and dispatch path. 3) Reconsider the 'prohibit asking' policy: permit safe clarifying questions for tasks that may have security/privilege implications. 4) Run the package in a sandboxed environment (no access to secrets or production systems) and inspect logs/outputs; disable autonomous invocation for sensitive contexts until behavior is validated. 5) If you are not the maintainer, ask the publisher for a canonical homepage or source repo and for signed releases; lack of a homepage and the truncated file reduce confidence in the package's integrity.
Capability Analysis
Type: OpenClaw Skill Name: agent-system Version: 1.0.0 The bundle implements a multi-agent orchestration framework for OpenClaw, including task planning, execution, quality review, and a 'self-healing' retry mechanism. The core logic in src/orchestrator.js and instructions in SKILL.md focus on managing task complexity and ensuring structured JSON outputs. While the AGENT_RULES.md and SKILL.md files contain strict instructions for the agent to avoid asking follow-up questions and to use default values or inference when information is missing, these are presented as design choices for autonomous operation rather than malicious prompt injections. No evidence of data exfiltration, unauthorized command execution, or persistence was found.
Capability Assessment
Purpose & Capability
The declared purpose (core agent orchestrator: planner → executor → reviewer → self-heal) matches the included code and docs — the code implements planner/executor/reviewer/dispatch/evolve and the SKILL.md describes the expected routing. However, there are notable mismatches between docs/schema/code (see 'instruction_scope' and specific issues) that will break runtime integration unless fixed (e.g., complexity types, missing exported symbol).
Instruction Scope
SKILL.md and AGENT_RULES.md mandate: never ask the user for more info, fill missing placeholders with defaults, and always produce output (avoid waiting on user). This is an intentional design choice for autonomous execution but is high-risk: it encourages continuing with incomplete input (and possible hallucination), strips opportunity for safe confirmation, and may cause execution of tasks with incorrect assumptions. Also, AGENT_RULES.md requires placeholder-filling, but the reference implementation does not implement that pre-fill step — another inconsistency.
Install Mechanism
No install spec is provided (instruction-only mode plus sample Node.js files). Nothing is downloaded from external URLs and there are no required binaries or environment variables. Risk from installation mechanism is low, but the package contains runnable JS code which should be reviewed before execution.
Credentials
The skill requests no environment variables, no external credentials, and the code does not perform network calls or access OS config paths. From a credentials perspective it is proportional to the stated purpose.
Persistence & Privilege
always:false (not force-enabled). The skill contains an 'evolve' function that mutates an in-memory LEARNING_STATE and can auto-adjust thresholds (behavioral drift), but it does not persist configuration to disk in the shipped code. Autonomous invocation is allowed (platform default) — combined with the 'do not ask' rule this increases the risk of undesired autonomous actions; consider disabling autonomous invocation for sensitive environments until reviewed.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install agent-system
  3. After installation, invoke the skill by name or use /agent-system
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
agent-system 1.0.0 - OpenClaw Agent 调度系统首发,聚焦多步骤任务智能分流 - 新增 Orchestrator、Planner、Executor、Reviewer、Self-Heal、Metrics 等核心模块 - 基于任务复杂度自动选择最优执行路径,支持自愈和降级处理 - 完整定义复杂度评估、输出/日志/指标追踪、质量阈值与自愈策略 - 明确禁止提问规则和置信度标注,提升自动执行稳定性 - 提供标准响应与日志 JSON 格式,便于集成和追踪
Metadata
Slug agent-system
Version 1.0.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is agent-system?

OpenClaw 核心 Agent 调度系统。当用户描述需要"分析"、"规划"、"拆解任务"、"多步骤处理"、"自动执行复杂任务"时激活。 基于复杂度自动选择最优执行路径。 It is an AI Agent Skill for Claude Code / OpenClaw, with 80 downloads so far.

How do I install agent-system?

Run "/install agent-system" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is agent-system free?

Yes, agent-system is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does agent-system support?

agent-system is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created agent-system?

It is built and maintained by XIAXIANB (@xiaxianb); the current version is v1.0.0.

💬 Comments