← Back to Skills Marketplace
lorwaleroy

Consulting Agent Pipeline

by LorwaLeroy · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ✓ Security Clean
73
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install consulting-agent-pipeline
Description
多Agent协作研究管线 — 将"调研→框架→执行→审核→迭代→交付"的咨询项目工作流抽象为可复制的标准化协议。触发场景:(1) 新项目需要多Agent协作且横跨多个平台(Claude/Codex/Gemini/Minimax),(2) 咨询类交付物需要多轮结构化审核迭代,(3) 需要建立项目状态追踪机制替代散落...
README (SKILL.md)

Multi-Agent Pipeline — 多Agent协作研究管线

首钢吉泰安项目(2026-04-09 ~ 04-14)的实战协作经验抽象为可复制工作流。

技能定位

不是什么:本技能不是通用任务管理工具(task-management),不是单Agent TODO列表,不是LangGraph/CrewAI框架替代品,不是项目管理软件(Gantt/Jira)的替代品。

是什么:一套跨平台、多Agent协作的文件系统协议——定义如何通过共享文件夹 + 结构化文档实现协作,而不依赖任何特定Agent平台或聊天通道。

适用项目类型

  • ✅ 咨询类交付物(PPT/报告/方案文档)
  • ✅ 多专题并行调研 → 单一线性产出的项目
  • ✅ 需要多轮审核迭代(≥2轮)的对外交付
  • ✅ 跨平台Agent协作(Agent不同属一个公司/平台)

不适用

  • ❌ 纯软件开发迭代(代码频繁merge/branch,适合用Git工作流而非本技能)
  • ❌ 实时通信/客服类项目(需要毫秒级响应,不适合文件交接协议)
  • ❌ 单Agent独立完成任务(不需要多Agent协作协议)
  • ❌ 高度机密项目(不建议在共享文件系统存储任何项目信息)

核心设计原则

  1. Agent-Agnostic:基于文件系统协议,任何能读/写本地文件的Agent都能参与
  2. Shared State:PROJECT_STATE.yaml 是唯一的项目状态数据源,消除分散在多个文档中的状态不一致
  3. Structured Handoff:交接文档使用 YAML frontmatter(机器可解析)+ Markdown 正文八节(人类可读)
  4. Calibrated Autonomy:Human 决策矩阵决定干预层级,P0/P1/P2 优先级决定自动化程度
  5. Every Handoff Auditable:每次交接都有文档记录,替代聊天传递,审计可追溯

工作流五阶段

┌─────────────────────────────────────────────────────────────┐
│ Phase 0: INIT                                               │
│ → Orion: 初始化项目结构 + 填写 PROJECT_STATE               │
│ → 人类审批: 框架方向 + Agent分配 + 禁用词清单               │
└──────────────────────────────┬──────────────────────────────┘
                               │ 人类Gate ✅
┌──────────────────────────────▼──────────────────────────────┐
│ Phase 1: RESEARCH(可并行 Fan-Out)                         │
│ → N个 Research Agent 并行执行专题调研                       │
│ → Orion: 汇总 → research-synthesis.md                       │
│ → 人类审批: 调研质量 + 数据可用性                           │
└──────────────────────────────┬──────────────────────────────┘
                               │ 人类Gate ✅
┌──────────────────────────────▼──────────────────────────────┐
│ Phase 2: FRAMEWORK                                         │
│ → Reviewer Agent: 构建交付物内容框架                        │
│ → 输出: framework-to-execution.handoff.md                  │
│ → 人类审批: 框架方向(如"冒总定调:先出框架去谈")          │
└──────────────────────────────┬──────────────────────────────┘
                               │ 人类Gate ✅
┌──────────────────────────────▼──────────────────────────────┐
│ Phase 3: EXECUTION                                          │
│ → Execution Agent: 生成 v1 交付物                           │
│ → forbidden-terms-scan.sh: 自动扫描                         │
│ ⚡ Auto Gate: 扫描通过 → 进入审核;不通过 → 退回修复        │
└──────────────────────────────┬──────────────────────────────┘
                               │ Auto Gate ✅
┌──────────────────────────────▼──────────────────────────────┐
│ Phase 4: AUDIT(Generator / Critic 循环,3轮上限)          │
│ Round 1: 方向 + 语气 + 视觉层级                             │
│ Round 2: 完整度 + 合规性 + 数据一致性                        │
│ Round 3: 风险边界 + 禁用词清除                               │
│                                                            │
│ 📊 Calibrated Gate:                                        │
│   全部P2 → Agent自动迭代进入下一轮                          │
│   有P1无P0 → Agent迭代后人类复核                            │
│   有P0 → 人类必须审批修复方案                               │
│   审核通过 → 进入 Phase 5                                  │
└──────────────────────────────┬──────────────────────────────┘
                               │ Human Gate ✅
┌──────────────────────────────▼──────────────────────────────┐
│ Phase 5: DELIVERY                                          │
│ → validate-deliverable.sh: 结构验证                        │
│ → forbidden-terms-scan.sh: 最终扫描                         │
│ → snapshot-dir.sh: 版本快照归档                            │
│ → 人类审批: 最终交付物确认                                  │
└─────────────────────────────────────────────────────────────┘

Human Gate 与 Calibrated Autonomy 速查

阶段 门控类型 触发条件 谁操作
Phase 0→1 🔴 Human审批 框架+Agent+禁用词确认 人类
Phase 1→2 🔴 Human审批 调研质量通过 人类
Phase 2→3 🔴 Human审批 框架方向确认 人类
Phase 3→4 ⚡ Auto 禁用词扫描无P0违规 脚本
Phase 4 每轮 📊 Calibrated 见上方矩阵 Agent+人类
Phase 4→5 🔴 Human审批 审核通过 人类

核心文件速查

文件 用途 谁写/读
00_pipeline/PROJECT_STATE.yaml 项目状态数据库(单一真相源) Orion写,人类必读
00_pipeline/FORBIDDEN_TERMS.yaml 禁用词清单 人类初始化,所有Agent必读
00_pipeline/AGENT_REGISTRY.yaml Agent能力注册表 Orion维护,分发任务时用
00_pipeline/DECISION_LOG.md 人类决策记录 人类写,追溯用
00_pipeline/handoffs/*.md 交接文档 发送Agent写,接收Agent读

模板选择指南

场景 使用的模板 说明
启动新项目 init-project.sh 一键生成项目结构
新增专题调研 research-topic-template.md 每个专题一份
并行启动多个调研 parallel-research-launcher.md 一次性分发N个
汇总所有调研成果 research-synthesis.md Orion汇总后给Framework
Phase间交接 generic-handoff.md 通用交接文档
执行→审核交接 audit-round{1,2,3}.md 三轮审核专用
版本归档 version-snapshot-template.md 每次Phase完成时
验证产物结构 validate-deliverable.sh(脚本) 每轮交付后必跑

关键协议速览

协议 核心内容 详细说明
handoff-spec YAML frontmatter + 八节正文 references/handoff-spec.md
state-lifecycle 六种状态 + 状态转换规则 references/state-lifecycle.md
forbidden-terms 三类禁用词 + 扫描机制 references/forbidden-terms-spec.md
agent-card Agent注册表格式 + Role定义 references/agent-card-spec.md
version-control 版本快照 + 验证规则 references/version-control-spec.md

脚本工具

脚本 用途 必跑时机
scripts/init-project.sh 项目初始化 Phase 0
scripts/forbidden-terms-scan.sh 禁用词扫描(依赖yq可选) 每轮交付前(all-rounds)
scripts/validate-handoff.sh 交接文档格式校验 每次交接文档创建后
scripts/validate-deliverable.sh 交付物结构校验 每次交付物完成时
scripts/snapshot-dir.sh 版本目录快照 每次Phase完成时

局限性(必须了解)

局限性 影响 缓解方案
依赖共享文件系统 所有Agent必须能访问同一项目目录 初始化时确认路径可访问
YAML解析可选 forbidden-terms-scan.sh 无yq时使用硬编码 建议 brew install yq,见脚本内警告
无自动任务调度 Phase推进依赖人类或编排者触发 Orion维护 PROJECT_STATE 状态
Obsidian可选 技能模板存在Obsidian,但项目文件在任意目录均可 项目文件可在任意路径,Obsidian仅用于个人知识库同步
脚本为bash 需要macOS/Linux环境 Windows需WSL

首钢吉泰安项目教训

教训 触发事件 本技能中的解决方案
G1/G2/G3/G4 代号泄漏到v4才发现 禁用词final-only扫描 enforcement: all-rounds,每轮必扫
审核意见靠聊天中转失真 00_使用说明_如何把Claude的输出给到Codex.md workaround 交接文档YAML frontmatter,Codex直接读
Human决策点散落无法追溯 重要决策在对话中未记录 DECISION_LOG.md 集中记录
v4/v5版本混乱 版本未做快照覆盖交付 snapshot-dir.sh + 版本快照模板
Figma-desktop MCP鉴权失败阻塞 工具阻塞未及时上报 state-lifecycle: input-required + Orion立即通知人类

复盘机制

本技能内置两层复盘:

1. 阶段级复盘(每Phase结束时)

  • Orion 对照 PROJECT_STATE 的 decisions 节,审查本阶段是否有未记录决策
  • 检查 milestone 是否全部 completed
  • 更新 PHASE_README.md 的阶段总结节

2. 项目级复盘(Phase 5 完成后)

  • Orion 填写 DECISION_LOG.md 最终汇总
  • 对照首钢吉泰安教训清单,检查本项目是否有新教训
  • 将新教训追加到 MEMORY.md 的"协作反思"节

复盘结果不修改本技能的 SKILL.md,而是追加到 ~/.openclaw/workspace/memory/YYYY-MM-DD.md

敏感信息处理

信息类型 存储位置 风险 处理方式
真实Agent ID/名称 AGENT_REGISTRY.yaml 低(仅平台角色) 使用平台角色名而非个人名称
API Key / Token 不存于此技能 使用各平台Keychain/Gateway凭据管理
客户机密数据 项目目录(非技能目录) 技能模板不包含任何真实项目数据
内部代号(G1/G2/G3/G4) FORBIDDEN_TERMS.yaml 仅在项目初始化时注入,交付物扫描清除
政策金额/资本市场表述 FORBIDDEN_TERMS.yaml risk_expressions类管控,须条件使用

警告:所有模板中的示例数据(如 YYYY-MM-DD / PROJECT_NAME / 项目ID)均为占位符,不得包含真实项目数据。

Usage Guidance
This skill appears internally consistent and implements a filesystem-based multi‑agent pipeline with local helper scripts (scanning, validation, snapshot). Before installing/running: 1) Review and, if needed, edit templates/FORBIDDEN_TERMS.yaml so the hard-coded defaults match your project; 2) Be aware forbidden-terms-scan.sh will search project files and — if you don't provide a FORBIDDEN_TERMS.yaml path — will also probe a few $HOME locations (e.g., ~/.openclaw/... and ~/Documents/Obsidian Vault). If you keep sensitive data out of the shared project directory and avoid using cloud-synced home folders as targets, risk is low; 3) If you use Obsidian or any synced folder, remove or change the example Obsidian copy paths in templates to avoid accidental syncing; 4) Inspect and (optionally) run the scripts locally in a controlled project directory to confirm tool availability (yq, rsync, unzip) and behavior before letting any agent run them autonomously; 5) Do not use this skill for highly confidential projects (the SKILL.md already warns against that).
Capability Analysis
Type: OpenClaw Skill Name: consulting-agent-pipeline Version: 1.0.0 The skill bundle provides a comprehensive framework for managing multi-agent consulting workflows through a structured file-system protocol. It includes utility scripts for project initialization (init-project.sh), compliance scanning (forbidden-terms-scan.sh), and versioned snapshots (snapshot-dir.sh), all of which perform actions consistent with the stated purpose. The instructions in SKILL.md and the various templates focus on maintaining project state, ensuring document quality, and managing agent handoffs without any evidence of malicious intent, data exfiltration, or unauthorized remote execution.
Capability Tags
requires-sensitive-credentials
Capability Assessment
Purpose & Capability
Name/description describe a file-system based multi-agent pipeline. All included files (handoff specs, state lifecycle, templates, and helper scripts) align with that purpose. There are no unexpected credentials, network endpoints, or unrelated binaries declared.
Instruction Scope
SKILL.md explicitly instructs agents to read/write a shared project directory and to run the included validation/snapshot/scan scripts — this matches the declared file-system protocol. A couple of items to note: (1) the parallel research template and some prompts instruct Research Agents to copy outputs into an Obsidian path (a user-specified home directory), which could cause files to end up in cloud-synced locations if the user uses such services; (2) forbidden-terms-scan.sh attempts to auto-detect a FORBIDDEN_TERMS.yaml under several $HOME paths if not supplied. These behaviors are consistent with the skill's goal but do widen the scope of filesystem reads/writes to user home locations if default lookups are used.
Install Mechanism
No install spec or remote download is present; the skill is instruction + local shell scripts and templates only. That is low risk from an installer perspective. The init script references an 'openclaw skills install' CLI in usage text, but there is no hidden installer or network fetch in the repository.
Credentials
The skill declares no required environment variables or credentials. Scripts implicitly rely on standard POSIX tools (grep, find, rsync, unzip) and optionally yq; those are reasonable for local validation and scanning tasks. The only implicit environment use is reading $HOME for optional template locations — this is justified by the templates but you should be aware it will probe user home paths if run without an explicit FORBIDDEN_TERMS.yaml path.
Persistence & Privilege
always:false and default model invocation settings are used. The skill does not request permanent platform-level privileges and does not modify other skills or system-wide config. Scripts write snapshots and project files into user-specified project directories only, which matches the declared purpose.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install consulting-agent-pipeline
  3. After installation, invoke the skill by name or use /consulting-agent-pipeline
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
Claude审核通过版:POSIX兼容脚本+阶段门控分类表+发送前5项检查清单+扫描时机表;基于首钢吉泰安项目25文件实战验证
Metadata
Slug consulting-agent-pipeline
Version 1.0.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is Consulting Agent Pipeline?

多Agent协作研究管线 — 将"调研→框架→执行→审核→迭代→交付"的咨询项目工作流抽象为可复制的标准化协议。触发场景:(1) 新项目需要多Agent协作且横跨多个平台(Claude/Codex/Gemini/Minimax),(2) 咨询类交付物需要多轮结构化审核迭代,(3) 需要建立项目状态追踪机制替代散落... It is an AI Agent Skill for Claude Code / OpenClaw, with 73 downloads so far.

How do I install Consulting Agent Pipeline?

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

Is Consulting Agent Pipeline free?

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

Which platforms does Consulting Agent Pipeline support?

Consulting Agent Pipeline is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Consulting Agent Pipeline?

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

💬 Comments