← Back to Skills Marketplace
tuobadaidai

Orchestration Reviewer Skill

by tuobadaidai · GitHub ↗ · v1.0.2 · MIT-0
cross-platform ✓ Security Clean
50
Downloads
0
Stars
0
Active Installs
2
Versions
Install in OpenClaw
/install orchestration-reviewer-skill
Description
对现有工作流计划(Plan IR)进行结构化评审,输出评审结论、问题清单和修订建议。 Use when user asks to 评审工作流计划、审查DAG、审核编排方案、review plan IR、 检查计划合理性、对比多个编排版本、自动评审工作流. 不适用于从零生成新计划/单步操作/纯文本任务。
README (SKILL.md)

Orchestration Reviewer Skill v1.0

对已有工作流计划(Plan IR)进行结构化评审,判断其是否满足业务目标、 是否具备可执行性与治理合理性,输出明确评审结论和修订建议。

核心定位:不从零生成计划,而是面向已有编排结果进行审查、评分和修订。

功能范围

  • 评审 Plan IR 与 business_goal 的对齐度
  • 检查拓扑、节点职责、技能绑定、作用域声明
  • 输出明确结论:pass / conditional_pass / reject / insufficient_information
  • 输出评分卡(scorecard)和问题清单(issues)
  • 可选修订:生成 revised_plan_ir + diff_report

不覆盖:从零生成计划、运行时调度/重试/补偿、执行结果担保。

使用

决策路由

用户说的 是否适用 说明
"帮我审一下这个执行计划" 调用本 Skill
"这个DAG编排合理吗" 调用本 Skill
"对比两个版本的计划哪个更好" 调用本 Skill
"从零帮我规划一个流程" 用 workflow-orchestration-skill
"直接执行这个任务" 单步操作,不需要评审

调用输入

{
  "business_goal": "需要编排的业务目标",
  "existing_plan_ir": { "nodes": [...], "edges": [...], ... },
  "available_skills_manifest": [...],
  "review_mode": "review_only",
  "constraints": { "max_nodes": 20, "max_depth": 5 }
}

可选字段:review_preferences(strictness_level / require_scorecard / prefer_minimal_change)。

五阶段流程

阶段一:输入规范化

  • 解析 business_goal、existing_plan_ir、available_skills_manifest

阶段二:静态结构校验

python3 scripts/validate_schema.py --plan plan.json
python3 scripts/validate_topology.py --plan plan.json
python3 scripts/validate_scope.py --plan plan.json

阶段三:语义评审(LLM 主导) 从 7 个维度评审:目标一致性、拓扑合理性、技能绑定、作用域安全、 可执行性、效率与复杂度、风险与治理。

阶段四:结论生成 输出 overall_decision + scorecard + issues + suggested_actions。

阶段五:可选修订review_mode = "review_and_revise" 时生成 revised_plan_ir + diff_report。

完整调用示例

# 1. 评审(仅 review)
python3 scripts/review_plan.py \
  --goal "新员工入职流程" \
  --plan plan.json \
  --manifest skills_manifest.json \
  --mode review_only \
  --output review_result.json

# 2. 评审 + 修订
python3 scripts/review_plan.py \
  --goal "新员工入职流程" \
  --plan plan.json \
  --manifest skills_manifest.json \
  --mode review_and_revise \
  --output review_result.json

# 3. 独立评分
python3 scripts/score_plan.py --plan plan.json --manifest skills.json --output score.json

# 4. 差异对比
python3 scripts/diff_plan.py --original v1.json --revised v2.json --output diff.json

补充说明

评审维度(7 维)

维度 检查内容
目标一致性 计划是否覆盖 business_goal 所需关键步骤
拓扑合理性 依赖关系是否合理,有无冗余串行、错误前置、不可达节点
技能绑定 target_skill 是否与节点职责匹配
作用域安全 scoped_state_keys 是否遵循最小权限
可执行性 计划是否可被工作流引擎稳定消费
效率与复杂度 有无冗余节点、可并行未并行、复杂度过高
风险与治理 高风险技能、缺少人工确认点、敏感数据暴露

结论判定

结论 含义
pass 计划合理,可进入执行
conditional_pass 基本合理,需按建议修正后执行
reject 计划不合理,不建议执行
insufficient_information 信息不足,无法判断

问题严重级别

级别 含义
critical 阻断性问题,必须修复(如循环依赖、技能不在白名单)
major 重要问题,强烈建议修复(如可并行未并行)
minor 次要问题,建议优化(如命名不规范)
info 信息提示(如全局状态引用)

前置条件

  • 已提供 business_goalexisting_plan_ir
  • 已提供 available_skills_manifest 或等价技能元数据
  • 不满足前置条件时返回结构化错误

安全规则

  • 评审结果仅作为治理与决策依据,不自动触发执行
  • 高风险修订建议要求人工复核
  • 修订结果必须再次通过静态校验

依赖

pip install jsonschema networkx pydantic deepdiff

目录结构

orchestration-reviewer-skill/
├── SKILL.md
├── references/
│   ├── review-manual.md
│   ├── review-rules.md
│   ├── scoring-policy.md
│   └── security-guardrails.md
├── scripts/
│   ├── review_plan.py
│   ├── validate_schema.py
│   ├── validate_topology.py
│   ├── validate_scope.py
│   ├── score_plan.py
│   └── diff_plan.py
└── assets/
    ├── input_schema.json
    ├── output_schema.json
    ├── sample_review_input.json
    └── sample_review_output.json
Usage Guidance
Use this skill for reviewing workflow plans, not for executing them. Provide only the plan and skill manifest data needed for review, and manually inspect any suggested revisions before using them in a real workflow.
Capability Assessment
Purpose & Capability
The stated purpose is to review existing Plan IR workflow plans; the artifacts implement schema, topology, scope, scoring, diff, and review helpers aligned with that purpose.
Instruction Scope
Runtime instructions are scoped to user-provided plan and manifest files, explicitly say review results do not automatically trigger execution, and require human review for high-risk revisions.
Install Mechanism
The README-style instructions mention installing Python dependencies such as jsonschema, networkx, pydantic, and deepdiff; this is disclosed and proportionate, with no install hook or automatic startup behavior found.
Credentials
Scripts read local JSON inputs and write local JSON reports; no network calls, credential/session/profile access, broad local indexing, or unrelated file access were found.
Persistence & Privilege
No persistence mechanism, privilege escalation, background worker, cron/system service, or automatic mutation of installed skills or agent behavior is present.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install orchestration-reviewer-skill
  3. After installation, invoke the skill by name or use /orchestration-reviewer-skill
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.2
P0: 修复 reviewer 输入契约 (existing_plan_ir vs plan_ir 解包兼容); P0: 修复 validate_schema 错误校验 review 请求整体而非 plan 对象; P1: 修复 validate_topology/scope 空 plan 假阳性通过; P2: 修复 review_plan 全部通过时无法输出 pass; P2: 改进 goal_alignment 评分 (基于关键词语义覆盖率而非节点数)
v1.0.1
P0: 修复 reviewer 输入契约 (existing_plan_ir vs plan_ir 解包兼容); P0: 修复 validate_schema 错误校验 review 请求整体而非 plan 对象; P1: 修复 validate_topology/scope 空 plan 假阳性通过; P2: 修复 review_plan 全部通过时无法输出 pass; P2: 改进 goal_alignment 评分 (基于关键词语义覆盖率而非节点数)
Metadata
Slug orchestration-reviewer-skill
Version 1.0.2
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 2
Frequently Asked Questions

What is Orchestration Reviewer Skill?

对现有工作流计划(Plan IR)进行结构化评审,输出评审结论、问题清单和修订建议。 Use when user asks to 评审工作流计划、审查DAG、审核编排方案、review plan IR、 检查计划合理性、对比多个编排版本、自动评审工作流. 不适用于从零生成新计划/单步操作/纯文本任务。 It is an AI Agent Skill for Claude Code / OpenClaw, with 50 downloads so far.

How do I install Orchestration Reviewer Skill?

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

Is Orchestration Reviewer Skill free?

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

Which platforms does Orchestration Reviewer Skill support?

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

Who created Orchestration Reviewer Skill?

It is built and maintained by tuobadaidai (@tuobadaidai); the current version is v1.0.2.

💬 Comments