智能体驱动的 E2E 测试：不是替代，而是新的一层

# 传统确定性测试——每一步都硬编码
from playwright.sync_api import sync_playwright

def test_login_flow():
    with sync_playwright() as p:
        browser = p.chromium.launch()
        page = browser.new_page()
        page.goto("https://app.example.com/login")
        page.fill("#email", "user@test.com")
        page.fill("#password", "secret123")
        page.click("button[type='submit']")
        assert page.url == "https://app.example.com/dashboard"
        browser.close()

// MCP 工具调用示例——Agent 通过工具协议直接操控浏览器
{
  "tool": "playwright_navigate",
  "arguments": { "url": "https://app.example.com/login" }
}
// Agent 观察页面后自主决定下一步
{
  "tool": "playwright_click",
  "arguments": { "selector": "button[data-testid='submit']" }
}

# Agent 生成并执行的 CLI 命令序列
npx playwright codegen --target=python https://app.example.com

# Agent 也可以直接运行已有测试并分析结果
npx playwright test --reporter=json tests/smoke/ 2>&1 | \
  jq '.suites[].tests[] | select(.status=="failed") | .name'

# Agent 自动生成的 Playwright 测试——写完后人审、CI 跑
import re
from playwright.sync_api import Page, expect

def test_checkout_with_discount(page: Page):
    """Agent 生成：验证折扣码在结算流程中正确生效"""
    page.goto("/cart")
    # Agent 学会了用 data-testid 而非脆弱的 CSS selector
    page.click("[data-testid='proceed-to-checkout']")
    page.fill("[data-testid='discount-code']", "SAVE20")
    page.click("[data-testid='apply-discount']")
    # 断言折扣金额出现且数值合理
    discount_text = page.locator("[data-testid='discount-amount']").text_content()
    discount_value = float(re.search(r"\d+\.\d{2}", discount_text).group())
    assert discount_value > 0, "折扣金额应为正数"
    subtotal = float(page.locator("[data-testid='subtotal']").text_content().split("$")[1])
    assert discount_value <= subtotal * 0.3, "折扣不应超过小计的30%"

# 项目目录结构
tests/
├── deterministic/          # 传统 + Agent 生成的确定性测试，进 CI
│   ├── login.spec.py
│   ├── checkout.spec.py    # Agent 生成，人审核后提交
│   └── smoke/
├── exploratory/            # Agent 实时探索，不进 CI，本地或定时跑
│   ├── agent_mcp_configs/  # MCP 工具配置
│   └── exploration_prompts/
└── shared/
    └── test_data.py        # 非生产测试数据工厂

# .github/workflows/e2e-deterministic.yml
name: E2E Deterministic Tests
on: [push, pull_request]
jobs:
  test:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - run: pip install playwright pytest
      - run: playwright install chromium
      - run: pytest tests/deterministic/ --timeout=60s
        env:
          TEST_ENV: staging  # 只用非生产数据

# .github/workflows/e2e-exploratory.yml
name: E2E Exploratory Agent Scan
on:
  schedule:
    - cron: "0 6 * * 1-5"  # 工作日早上跑一次
jobs:
  explore:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Run agentic exploration
        run: |
          npx @anthropic-ai/claude-code \
            --prompt "用 Playwright MCP 工具探索 staging 环境的完整用户流程，记录所有异常行为" \
            --allowedTools "playwright_navigate,playwright_click,playwright_fill,playwright_screenshot" \
            --maxTurns 50
        env:
          MCP_ENDPOINT: ${{ secrets.STAGING_MCP_URL }}
      - name: Parse agent findings
        run: python scripts/parse_exploration_report.py
      - name: Create issues for bugs found
        uses: actions/create-issue@v1
        with:
          title-prefix: "[Exploratory] "

# shared/test_data.py——非生产数据工厂
import faker

fake = faker.Faker()

def create_test_user(env="staging"):
    """只在 staging 创建测试用户，绝不碰生产"""
    return {
        "email": f"e2e_test_{fake.uuid4()}@test.example.com",
        "password": "TestPass!2024",
        "env": env,  # 确认环境标记
    }

# Agent 生成的测试必须引用这个工厂，不能硬编码真实用户