ecosystem-orchestration

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

🤖 Orchestration & Agentic Ecosystem

🤖 编排与Agent生态系统

Welcome to the

orchestration

domain ecosystem. Unlike typical domains (which focus on specific tech stacks like frontend or databases), this meta-ecosystem defines the rules, workflows, and connectivity for how AI autonomous agents interact, plan, recover from errors, and evaluate themselves and each other.

For Orchestrator Agents (
task-decomposer
,
parallel-planner
): This document is your internal operating manual. Use these workflows when structuring complex reasoning tasks, coordinating multiple AI models, or building secure entry points for external agentic integrations.

欢迎来到

orchestration

领域生态系统。与专注于特定技术栈（如前端或数据库）的典型领域不同，这个元生态系统定义了AI自主Agent如何交互、规划、从错误中恢复以及自我评估和相互评估的规则、工作流和连接方式。

面向编排Agent（
task-decomposer
、
parallel-planner
）：本文档是你的内部操作手册。在构建复杂推理任务、协调多个AI模型或为外部Agent集成构建安全入口点时，请遵循这些工作流。

🔄 Standard Workflows & Handoffs

🔄 标准工作流与任务交接

1. The Agent Connectivity Flow (The Foundation)

1. Agent连接工作流（基础流程）

When the user asks to "integrate an LLM bot", "give an external agent access", or "design an AI pipeline into the app":

@claw-integration-design
-> (MANDATORY) Designs the API endpoints, sets up
```
{action}:{resource}
```
OAuth scopes, and builds the generic
```
tools.json
```
manifest. (This must always be the first step for agent connectivity).
```
@tool-selector
```
-> Analyzes the available tooling landscape (APIs, generic scripts) and selects the optimal set of actions for the incoming agent.
```
@context-compressor
```
-> Steps in when agents retrieve massive payloads (logs, raw DB dumps) to compress the context window before it overwhelms the LLM.

当用户要求“集成LLM机器人”、“授予外部Agent访问权限”或“在应用中设计AI流水线”时：

@claw-integration-design
->（必填步骤）设计API端点，设置
```
{action}:{resource}
```
OAuth权限范围，并生成通用的
```
tools.json
```
清单文件。（这是Agent连接的第一步，必须始终执行）
```
@tool-selector
```
-> 分析可用工具环境（API、通用脚本），为接入的Agent选择最优的操作集。
```
@context-compressor
```
-> 当Agent获取大量负载（日志、原始数据库转储）时介入，在上下文窗口超出LLM处理能力前对其进行压缩。

2. Task Orchestration & Execution Flow

2. 任务编排与执行工作流

When given a large, ambiguous objective (e.g., "build an app", "migrate the server"):

```
@task-decomposer
```
-> Breaks the massive goal into a step-by-step dependency graph.
```
@parallel-planner
```
-> Analyzes the decomposed steps to find where tasks can be run concurrently (saving time and tokens).
```
@memory-ledger
```
-> Used consistently throughout execution to maintain state across different sub-agents.
```
@error-recovery
```
-> Automatically takes over if an executed step fails (retries, attempts alternate tools, or safely escalates to a human).

当接到庞大且模糊的目标（例如“构建一个应用”、“迁移服务器”）时：

```
@task-decomposer
```
-> 将宏大目标分解为带依赖关系的分步任务图。
```
@parallel-planner
```
-> 分析分解后的步骤，找出可并行执行的任务（节省时间和token）。
```
@memory-ledger
```
-> 在整个执行过程中持续使用，以在不同子Agent之间维护状态。
```
@error-recovery
```
-> 若执行步骤失败，自动接管处理（重试、尝试替代工具或安全升级至人工处理）。

3. Multi-Brain (Consensus & Quality) Flow

3. 多脑（共识与质量）工作流

When the user asks to "find the best approach", "debate these two libraries", or "give me a robust solution":

Choose the synthesis model based on user intent:
- ```
@multi-brain
```
  -> General 3-perspective synthesis.
- ```
@multi-brain-experts
```
  -> Specialized perspective synthesis based on strict domain personas (e.g., Security vs UX).
- ```
@multi-brain-debate
```
  -> Used when two distinct choices clash; forces a 2-round adversarial debate.
```
@assumption-checker
```
-> Runs parallel to validate if the underlying facts driving the perspectives are hallucinated or real.
```
@output-critic
```
-> Acts as the final tollgate to score the proposed synthesis before showing it to the user. (For advanced scoring structures,
```
@multi-brain-score
```
can be used).

当用户要求“找到最佳方案”、“对比这两个库”或“提供一个稳健的解决方案”时：

根据用户意图选择合成模型：
- ```
@multi-brain
```
  -> 通用三方视角合成。
- ```
@multi-brain-experts
```
  -> 基于严格领域角色（如安全vs用户体验）的专业视角合成。
- ```
@multi-brain-debate
```
  -> 当两种不同方案冲突时使用，强制进行两轮对抗式辩论。
```
@assumption-checker
```
-> 并行运行，验证驱动各视角的底层事实是真实存在还是幻觉生成。
```
@output-critic
```
-> 作为最终关卡，在将合成结果展示给用户前对其进行评分。（如需高级评分结构，可使用
```
@multi-brain-score
```
）。

4. Safety & Review Flow

4. 安全与审核工作流

When autonomous agents are operating on sensitive codebases or production databases:

```
@checkpoint-guardian
```
-> A mandatory tollgate activated right before any destructive steps (e.g., Drop Table, Push to Main) are executed.
```
@agent-reviewer
```
-> An oversight entity that retrospectively audits what the execution agents did, looking for logic holes or skipped steps.

当自主Agent在敏感代码库或生产数据库上操作时：

```
@checkpoint-guardian
```
-> 在执行任何破坏性步骤（如删除表、推送到主分支）前激活的必填关卡。
```
@agent-reviewer
```
-> 监督实体，事后审核执行Agent的操作，查找逻辑漏洞或遗漏步骤。

🧠 Core Philosophy

🧠 核心理念

Agentic orchestration is about reliability, oversight, and fallback mechanisms. Unlike standard script execution, agents can hallucinate or fail unexpectedly. Every skill in this ecosystem is designed to bound agent behavior within safe limits (

claw-integration

checkpoint-guardian

), ensure high-quality reasoning (

multi-brain

critic

), and provide resilient recovery (

error-recovery

). Never trust a single "one-shot" output for a complex scenario without decomposition and review.

Agent编排的核心是可靠性、监督机制和 fallback 方案。与标准脚本执行不同，Agent可能会产生幻觉或意外失败。本生态系统中的每个skill都旨在将Agent行为限制在安全范围内（

claw-integration

、

checkpoint-guardian

），确保高质量推理（

multi-brain

、

critic

），并提供弹性恢复能力（

error-recovery

）。对于复杂场景，切勿轻信单一“一次性”输出，必须经过分解和审核。

🚀 Installation & Orchestration Instructions

🚀 安装与编排说明

As an AI Agent reading this ecosystem file, you must actively guide the user to install the missing skills for the selected workflow.

To install all skills within this ecosystem at once, you must execute or instruct the user to execute the provided installation script:

bash

python skills/ecosystem-orchestration/scripts/install_all.py

Execution Rule: Do not attempt to perform the tasks of these skills yourself. Your job is to install them and hand off the execution to the appropriate skill, or instruct the user to trigger the newly installed skills in order.

作为阅读本生态系统文件的AI Agent，你必须主动引导用户为所选工作流安装缺失的skill。

如需一次性安装本生态系统中的所有skill，你必须执行或指导用户执行以下安装脚本：

bash

python skills/ecosystem-orchestration/scripts/install_all.py

执行规则： 不要尝试自行执行这些skill的任务。你的职责是安装它们，并将执行任务交接给对应的skill，或指导用户按顺序触发新安装的skill。