什么是 自主技能编排器？

自主技能编排器是一个专门的框架，旨在通过协调其他可用的 Openclaw 技能来管理复杂、多步骤任务的生命周期。与标准执行不同，它在闭环状态下运行，冻结特定的用户意图，并在不需要持续人工干预的情况下迭代地规划和执行步骤。通过对自主性强制执行确定性方法，它确保您的 AI 编码智能体保持在定义的安全边界内。

该技能非常适合需要智能体自主执行重复性或多阶段操作，同时保持高可靠性和可预测行为的开发人员。它通过提供优先考虑意图完整性和风险缓解的严格执行循环，弥合了手动工具使用与完全自主智能体之间的差距。

下载入口:https://github.com/openclaw/skills/tree/main/skills/vishnubedi3/autonomous-skill-orchestrator

安装与下载

1. ClawHub CLI

从源直接安装技能的最快方式。

2. 手动安装

将技能文件夹复制到以下位置之一

优先级：工作区 > 本地 > 内置

3. 提示词安装

将此提示词复制到 OpenClaw 即可自动安装。

自主技能编排器 应用场景

• 执行涉及规划、编码和测试的多阶段开发工作流。
• 运行长周期的自主任务，仅在初始触发和最终停止时需要用户交互。
• 协调专门的 Openclaw 技能以实现单一、冻结的目标，而不会出现范围蔓延。
• 实现安全优先的自主循环，在遇到高风险歧义或错误时立即停止。

1. 将之前的用户命令捕获为冻结的、不可变的意图。
2. 进入闭环执行状态，禁用所有中间用户提示或确认。
3. 根据冻结的意图，向指定的规划器技能请求结构化计划提案。
4. 通过完整性与风险网关传递提案，以检查不可逆性和范围扩张。
5. 将计划规范化为确定性的执行序列，根据预定义的风险类别处理歧义。
6. 顺序调用执行器技能，并在每一步验证前提条件。
7. 记录任何停顿的事后数据，并根据明确的停止命令继续循环或终止。

自主技能编排器 配置指南

要将自主技能编排器集成到您的环境中，请确保已安装核心 Openclaw Skills 框架，并且至少有一个规划器技能和一个执行器技能可用。

# 验证环境中的可用技能
openclaw list-skills

# 在会话期间显式激活编排器
use autonomous-skill-orchestrator

该技能需要显式激活，并且仅在检测到有效的先前意图且至少有一个其他可执行技能准备好进行协调时才会运行。

自主技能编排器 数据架构与分类体系

该技能保持极小的数据占用空间，以确保安全并防止状态膨胀。它使用以下分类法组织数据：

此结构确保编排器在不同激活之间保持无状态，并且不会在 Openclaw Skills 生态系统中存储敏感的用户行为模式。

name: autonomous-skill-orchestrator description: > Deterministically coordinates autonomous planning and execution across available skills under strict guardrails. Use only when the user explicitly activates this skill by name to run autonomously until a stop command is issued. Trigger keywords include: "use autonomous-skill-orchestrator", "activate autonomous-skill-orchestrator", "start autonomous orchestration". metadata: version: "1.1.0" owner: "user"

Activation Criteria

Activate this skill if and only if all conditions below are true:

• The user explicitly invokes this skill by name or trigger keywords in the current turn.
• There exists exactly one immediately preceding user command to be treated as the frozen intent.
• At least one other executable skill is available for coordination.

Do not activate this skill if any condition below is true:

• The invocation is implicit, inferred, or indirect.
• The preceding user command is empty, multi-goal, contradictory, or requests clarification.
• No executable skills are available.
• The user issues a stop command.

Execution Steps

1. Freeze Intent

   • Capture the immediately preceding user command verbatim.
   • Store it as immutable intent for the duration of this activation.
   • Do not summarize, reinterpret, expand, or decompose the intent.

2. Initialize Control Loop

   • Enter a closed-loop execution state owned exclusively by this skill.
   • Disable all requests for user input, confirmation, or validation.
   • Ignore all user messages except an explicit stop command.

3. Request Plan Proposals

   • Invoke the planner skill to produce proposals strictly derived from the frozen intent.
   • Require output to contain only:
     • A finite, ordered list of features.
     • Explicit dependencies between features.
     • Explicit assumptions stated as facts, not guesses.
   • Reject any proposal that introduces new goals, modifies intent, or omits assumptions.

4. Sanity and Risk Gate

   • Evaluate proposals against the following checks:
     • Irreversibility of actions.
     • Scope expansion beyond frozen intent.
     • Use of tools or capabilities not explicitly available.
     • Assumptions that cannot be verified from provided context.
   • If any check fails, halt immediately.

5. Plan Normalization

   • Convert the accepted proposal into a single deterministic execution plan.
   • Classify ambiguity as follows:
     • Class A (unsafe or unbounded): halt.
     • Class B (bounded and resolvable): normalize once.
     • Class C (cosmetic or non-operative): ignore.
   • Do not re-run normalization or request alternative plans.

6. Execute Plan

   • Invoke executor skills to perform each step in order.
   • Before each step, verify preconditions explicitly stated in the plan.
   • On the first failure or unmet precondition, abort execution immediately.

7. Post-Mortem Recording

   • Record only:
     • Which step halted execution.
     • Which rule or check caused the halt.
   • Apply decay so records not repeated are removed over time.
   • Do not store goals, plans, preferences, or user behavior patterns.

8. Loop Continuation

   • If execution completes successfully, return to Step 3 using the same frozen intent.
   • Do not generate new intents or objectives.

9. Stop Condition

   • When the user issues an explicit stop command:
     • Terminate the control loop immediately.
     • Output exactly one dot (.) and no other content.

Ambiguity Handling

• Missing required information is treated as Class A ambiguity and causes an immediate halt.
• Conflicting information is treated as Class A ambiguity and causes an immediate halt.
• Ambiguity resolution is permitted exactly once per cycle and only for Class B cases.
• No inference, guessing, or user querying is permitted.

Constraints & Non-Goals

• Must not create, modify, or delete skills.
• Must not alter the frozen intent.
• Must not ask the user questions during operation.
• Must not self-validate plans or actions.
• Must not continue operation after any halt condition.
• Must not persist state beyond post-mortem records.

Failure Behavior

If execution cannot be completed safely or correctly:

• Halt immediately without retry.
• Produce no output.
• Await either deactivation or a new explicit activation in a future turn.