agent-pipeline

name	agent-pipeline
description	Multi-Agent Collaborative Pipeline for Claude Code — orchestrates 7 development phases with 6 specialized AI roles
version	2.0.0
author	JS-mark
tags	["multi-agent","pipeline","mcp","development-workflow","automation"]

/agent — Multi-Agent Collaborative Pipeline

Unified command for the multi-agent development workflow. Dispatches by subcommand.

Usage: /agent <subcommand> [args]

IMPORTANT: All MCP tool calls MUST include the root parameter set to the current working directory (project root). This ensures pipeline data stays per-project, not global.

Subcommand	Alias	Description
run <feature> [phase]	r	Run the 7-phase pipeline
status	s	Show pipeline progress
dashboard [port]	d	Start monitoring web UI
init	i	Initialize workflow in current project
(no args / help)	h	Show usage guide

---

Subcommand: help (default when no args)

Display:

/agent — Multi-Agent Collaborative Pipeline

Usage: /agent <subcommand> [args]

Subcommands:
  run <feature> [phase]   Run 7-phase pipeline (parallel Phase 2, fix-test loop)
  status                  Show pipeline progress
  dashboard [port]        Start monitoring dashboard (default port: 3210)
  init                    Initialize multi-agent workflow in current project

Pipeline Phases:
  1. PM           → PRD 需求文档
  2. 并行设计      → 架构师 + UI 设计师 + 测试计划 (并行)
  3. 全栈开发      → 代码实现
  4. 代码审查      → 审查报告
  5. 审查修复      → 修复 Blocker
  6. ��试执行      → 编写测试并执行
  7. 修复循环      → Bug 修复 + 重新测试 (max 5 rounds)

Quick Start:
  /agent init                    # Initialize (interactive)
  /agent run my-feature          # Full pipeline
  /agent run my-feature 3        # Resume from Phase 3
  /agent status                  # Check progress
  /agent dashboard               # Visual monitoring

CLI (any project dir):
  npx @aspect-mark/agent-pipeline status
  npx @aspect-mark/agent-pipeline dashboard [port]
  # Or via project shim:
  node scripts/pipeline.js status

---

Subcommand: run

Parse from $ARGUMENTS: skip first word ("run"/"r"), next word = feature name, optional next = start phase 1-7.

Step 0: Validate

1. If no feature name, ask the user
2. Use MCP tool pipeline_check_ready with root = current working directory to verify role files exist. If missing, tell user to run /agent init
3. Set: FEATURE, START_PHASE (default 1)

Step 1: Initialize status

Use MCP tool pipeline_init_status with parameters:

• root: current working directory
• feature: the feature name
• startPhase: the start phase (default 1)
• maxFixRounds: 5

This creates docs/.pipeline-status.json with the v2 schema.

Step 2: Confirm

Show plan to user and ask for confirmation.

Step 3: Execute phases

Before each phase: use pipeline_update_phase with root = current working directory → status "running" After each phase: use pipeline_update_phase with root = current working directory → status "done"

---

Phase 1: PM Agent

Launch one Agent (foreground):

Prompt: "你是产品经理。请阅读 .claude/roles/pm.md 了解你的角色职责和文档模板。
分析功能「<FEATURE>」的需求，编写 PRD 文档并输出到 docs/prd/feature-<FEATURE>.md。
请严格按照角色文件中的 PRD 模板格式编写，包含完整的功能需求、用户故事、验收标准等。"

Phase 2: Parallel Design (3 Agents)

CRITICAL: Launch all 3 Agent tool calls in a SINGLE message for parallel execution:

Before launching: use pipeline_update_phase phaseId=2, status="running"

Agent 2a - 架构师:

Prompt: "你是架构设计师。请阅读 .claude/roles/architect.md 了解职责和模板。
阅读 PRD docs/prd/feature-<FEATURE>.md，编写技术设计文档。
输出到 docs/architecture/feature-<FEATURE>.md。"

Agent 2b - UI 设计师:

Prompt: "你是 UI 设计师。请阅读 .claude/roles/ui-designer.md 了解职责和模板。
阅读 PRD docs/prd/feature-<FEATURE>.md，编写 UI 设计文档。
输出到 docs/ui-design/feature-<FEATURE>.md。"

Agent 2c - 测试工程师（仅计划）:

Prompt: "你是测试工程师。请阅读 .claude/roles/tester.md 了解职责和模板。
阅读 PRD docs/prd/feature-<FEATURE>.md，仅编写测试计划和测试用例文档，不写代码。
输出到 docs/test-plans/feature-<FEATURE>.md。"

As each agent completes, use pipeline_update_phase with phaseId=2 and subtaskRole to update its status. When all 3 done: use pipeline_update_phase phaseId=2, status="done"

Phase 3: Fullstack

Prompt: "你是全栈工程师。请阅读 .claude/roles/fullstack.md。
阅读上游文档（PRD + 架构 + UI + 测试计划），实现功能「<FEATURE>」代码到 src/。"

Phase 4: Code Review

Prompt: "你是代码审查员。请阅读 .claude/roles/reviewer.md。
审查功能「<FEATURE>」代码，输出到 docs/reviews/feature-<FEATURE>.md。
必须标注严重程度：🔴 Blocker / 🟡 Warning / 🟢 Info。
末尾添加「Blocker 汇总」。无 Blocker 写「无 Blocker 问题」。"

Phase 5: Fix Review

Prompt: "你是全栈工程师。阅读审查报告 docs/reviews/feature-<FEATURE>.md，
修复所有 🔴 Blocker 和 🟡 Warning。输出修复日志到 docs/reviews/feature-<FEATURE>-fix-log.md。"

Skip if review says "无 Blocker 问题".

Phase 6: Test Execution

Prompt: "你是测试工程师。根据测试计划编写测试代码并执行，
输出结果到 docs/test-plans/feature-<FEATURE>-results-round-1.md。
全部通过则在总结中写「全部通过」。"

Phase 7: Fix-Test Loop (max 5 rounds)

for round = 1 to 5:
  1. Read latest test results, check if "全部通过" → break
  2. Launch Fullstack Agent → fix bugs
  3. Launch Tester Agent → retest → results-round-<N+1>.md
  4. Use pipeline_update_fix_round with `root` = current working directory to record round result

Step 4: Finalize

Use MCP tool pipeline_finalize with root = current working directory and status "done". Display summary with all output files.

---

Subcommand: status

Step 1: Use MCP tool

Call pipeline_status with root = current working directory to get the current pipeline state.

Step 2: Display

Format the JSON response as a visual progress table:

Feature: <name>  |  Status: <running/done>  |  Started: <time>
─────────────────────────────────────────────────────
Phase 1  产品经理        ✅ done
Phase 2  并行设计        ✅ done
         ├─ 架构师       ✅ done    (29.9 KB)
         ├─ UI 设计师    ✅ done    (31.0 KB)
         └─ 测试工程师   ✅ done    (15.2 KB)
Phase 3  全栈开发        🔄 running (12 files)
Phase 4  代码审查        ⏳ pending
Phase 5  审查修复        ⏳ pending
Phase 6  测试执行        ⏳ pending
Phase 7  修复循环        ⏳ pending  [_ _ _ _ _]
─────────────────────────────────────────────────────

Step 3: Show next steps

Suggest /agent run <feature> <phase> to resume, or /agent dashboard to monitor.

---

Subcommand: dashboard

Use MCP tool pipeline_dashboard with action "start", root = current working directory, and optional port (default 3210).

Report the returned URL to the user.

---

Subcommand: init

Step 1: Ask the user

1. Project description — What is this project?
2. Tech stack — e.g. "Vue 3 + TypeScript + Hono"
3. Roles — Which to include? Default all 6: PM, Architect, UI Designer, Fullstack, Tester, Reviewer
4. Language — Documentation language? Default: 中文

Step 2: Initialize

Use MCP tool pipeline_init_project with:

• root: current working directory (MUST pass explicitly)
• techStack: user's answer
• language: user's answer

Step 3: Summary

Multi-Agent Workflow initialized!

Plugin:    @aspect-mark/agent-pipeline (MCP Server)
Project:   .claude/roles/{pm,architect,ui-designer,fullstack,tester,reviewer}.md
           scripts/pipeline.js  (CLI shim)

Quick start:
  /agent run <feature>                       # Full pipeline
  /agent run <feature> 3                     # From Phase 3
  /agent status                              # Check progress
  /agent dashboard                           # Visual monitoring