← Back to Skills Marketplace
spikesubingrui-design

Spike Skill Orchestrator

by SpikeSu · GitHub ↗ · v1.0.2 · MIT-0
cross-platform ⚠ suspicious
63
Downloads
0
Stars
0
Active Installs
2
Versions
Install in OpenClaw
/install spike-skill-orchestrator
Description
Adaptive local skill scheduler for OpenClaw. Use when a task may benefit from multiple installed skills, when the user asks "该用什么 skill", "调度一下", "有没有相关 skil...
README (SKILL.md)

Skill Orchestrator

目标:先判断本地已装 skills 里有没有能提效的组合,再决定是否询问 Spike 搭配使用。

何时加载

必须先跑

  • 多步任务:研究 + 输出、设计 + 文案、调试 + 测试、部署 + 验证
  • 跨域任务:UI + 内容、研究 + 汇报、记忆 + 报告、浏览器 + 抓取
  • 你不确定该走哪条 workflow,或同类任务历史上常漏调 skill
  • 用户明确说:
    • “调度一下”
    • “该用什么 skill”
    • “有没有相关 skill”
    • “这些 skill 能不能搭配”

直接跳过

  • 纯闲聊、单句问答、无需工具的简单解释
  • 明确只要单一 skill,且用户已点名
  • 已经在某个成熟 workflow 内部,例如 cron 固定链路

流程

1. 跑确定性召回

node skills/skill-orchestrator/scripts/match-skills.mjs "\x3C用户任务>" --json

读取:

  • ops/skills/registry.json
  • ops/skills/trigger-router.json

输出:

  • candidates: 相关单 skill 候选
  • combos: 相关组合候选

2. 做 smart 判定

按下面规则决定是否打断用户:

  • 只有 1 个高置信单 skill,且很明显:直接加载该 SKILL.md
  • 命中 2 个以上高分 skill,或命中组合:询问 Spike 是否搭配
  • 只有低分噪音候选:静默跳过,直接做任务

经验阈值:

  • candidates[0].score >= 12candidates[1] 差距明显:通常可直接用
  • 任一 combo.score >= 8:通常值得问一下

如何问

当命中组合或多候选时,使用 AskQuestion,不要直接替用户决定。

选项建议:

  • 组合 A:UI/设计 + 内容
  • 组合 B:搜索/研究 + 报告
  • 单独用某个 skill
  • 都不用,直接做

问题文案要短:

  • “这次任务可搭配这些本地 skills,要一起启用吗?”

执行顺序

用户采纳后,按“上游路由 skill → 下游执行 skill”的顺序读 SKILL.md

  • 例如:
    • opencli-usageopencli-browser
    • frontend-design-3impeccable-uxui
    • smart-searchdata-researchreport-ui

不要一次性把 5 个无关 skill 全读进上下文。最多保留:

  • 1 个主 skill
  • 1 到 2 个协作 skill

记录反馈

每次提议后都记一行日志:

node skills/skill-orchestrator/scripts/log-decision.mjs \
  --task "\x3C任务>" \
  --proposed "\x3Cskill-a,skill-b>" \
  --accepted "\x3Cskill-a,skill-b>" \
  --mode "\x3Cdirect|ask|skip>"

如果用户拒绝所有建议:

node skills/skill-orchestrator/scripts/log-decision.mjs \
  --task "\x3C任务>" \
  --proposed "\x3Cskill-a,skill-b>" \
  --accepted "" \
  --mode "reject"

日志写到:

  • openclaw-evolution/data/skill-orchestrator-log.jsonl

护栏

  • 不要为“看起来相关”就强塞 skill
  • 用户拒绝后,本轮静默,不要二次追问
  • 不要把外部市场 skill 搜索和本地 skill 调度混在一起
  • 不要改已有 workflow 的业务逻辑,只做召回和建议
  • 候选过多时,最多展示 3 个单 skill 或 2 个组合

示例

例 1:设计 + 内容

输入:

帮我做个落地页发小红书

期望建议:

  • 组合:UI/设计
  • 组合:内容
  • 问 Spike 是否搭配启用

例 2:AI 新闻

输入:

今天 AI 圈有什么

期望建议:

  • aihot
  • github-trending
  • trendradar
  • 可直接采用 早报 相关组合

例 3:简单问答

输入:

解释一下 MCP 是什么

行为:

  • 若无明显 workflow 优势,跳过 orchestrator,直接回答
Usage Guidance
Install only if you are comfortable with a routing skill seeing broad task descriptions and writing them to a local history file. Avoid using it for prompts containing secrets, private business context, credentials, or personal data unless you first modify or disable the logging behavior and fix the hardcoded workspace path.
Capability Assessment
Purpose & Capability
The routing and matching behavior fits the stated purpose of suggesting installed skills and combinations. The feedback log is related to improving routing, but it can capture sensitive user prompts.
Instruction Scope
The skill includes skip conditions, limits how many skills to load, and tells the agent to ask before enabling combinations. The trigger language is broad enough that it may activate for ordinary multi-step or uncertain tasks.
Install Mechanism
Installation is a normal ClawHub install or git clone. There are no package dependencies, install hooks, obfuscated code, or network-fetching scripts in the artifact.
Credentials
The scripts read local skill registry/router files and do deterministic matching, which is proportionate for a scheduler. However, the scripts hardcode a publisher-specific /Users/spikescp/.openclaw/workspace path rather than deriving the user's workspace path.
Persistence & Privilege
The skill instructs the agent to log every proposal, including raw task text plus proposed and accepted skills, to openclaw-evolution/data/skill-orchestrator-log.jsonl. The implementation appends indefinitely to a fixed absolute path and provides no opt-out, redaction, retention policy, or permission hardening.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install spike-skill-orchestrator
  3. After installation, invoke the skill by name or use /spike-skill-orchestrator
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.2
Document ClawHub slug; registry sync
v1.0.1
Initial ClawHub publish (slug: spike-skill-orchestrator)
Metadata
Slug spike-skill-orchestrator
Version 1.0.2
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 2
Frequently Asked Questions

What is Spike Skill Orchestrator?

Adaptive local skill scheduler for OpenClaw. Use when a task may benefit from multiple installed skills, when the user asks "该用什么 skill", "调度一下", "有没有相关 skil... It is an AI Agent Skill for Claude Code / OpenClaw, with 63 downloads so far.

How do I install Spike Skill Orchestrator?

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

Is Spike Skill Orchestrator free?

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

Which platforms does Spike Skill Orchestrator support?

Spike Skill Orchestrator is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Spike Skill Orchestrator?

It is built and maintained by SpikeSu (@spikesubingrui-design); the current version is v1.0.2.

💬 Comments