video-workflow/e2e/ip-operator-real-llm.spec.ts

import { expect, Page, test } from '@playwright/test';
import fs from 'node:fs';
import path from 'node:path';

declare global {
  interface Window {
    __seedIpOperatorScenario?: (scenario: 'empty') => unknown;
  }
}

const scenarioSeedPath = path.resolve(__dirname, '../scripts/validation/ip-operator-scenario-seed.js');
const sampleRecordPath = path.resolve(
  __dirname,
  '../docs/项目概括/20-实施方案/10-IP操盘工作台V1/IP操盘工作台V1-真实LLM样例验收记录.md',
);
const debugPlanPath = path.resolve(__dirname, '../test-results/ip-operator-real-llm-plan.json');

test.describe('IP操盘真实 LLM 样例验收', () => {
  test.skip(process.env.IP_OPERATOR_REAL_LLM !== '1', 'Set IP_OPERATOR_REAL_LLM=1 to run the paid/network real LLM validation.');

  test('can generate a real evaluable IP operator plan', async ({ page }) => {
    test.setTimeout(300_000);
    await seedEmptyLoggedInWorkbench(page);
    await page.goto('/');

    const profileCard = page.locator('.ip-card').first();
    await profileCard.locator('input').nth(0).fill('林川老板增长笔记');
    await profileCard.locator('input').nth(1).fill('传统行业增长顾问');
    await profileCard.locator('input').nth(2).fill('本地生活 / 家居建材 / 教育培训');
    await profileCard.locator('input').nth(3).fill('传统行业中小企业老板');
    await profileCard.locator('input').nth(4).fill('老板 IP 诊断与私域转化陪跑');
    await profileCard.locator('textarea').nth(0).fill('内容像广告、不知道如何表达专业判断、缺少真实案例支撑。');
    await profileCard.locator('textarea').nth(1).fill('服务过本地商家，做过短视频账号定位和私域转化。');
    await profileCard.locator('textarea').nth(2).fill('账号定位、选题策划、脚本诊断、私域转化路径设计。');
    await profileCard.locator('textarea').nth(3).fill('直接、务实、案例化。');
    await profileCard.locator('textarea').nth(4).fill('不夸大收益，不做成功学。');
    await profileCard.locator('textarea').nth(5).fill('生成可执行的 7 天内容测试方案。');
    await profileCard.locator('.btn-primary').click();

    const questionnaireCard = page.locator('.ip-card').first();
    const answers = [
      '老板想用内容建立信任，但担心拍出来像广告。',
      '我能把真实服务经验翻译成用户听得懂的判断。',
      '过去服务过本地门店和中小企业，有转化案例。',
      '目前通过咨询、诊断和陪跑服务承接。',
      '不想被误解成卖课或夸大结果的人。',
      '擅长直接讲问题、讲案例、给判断标准。',
      '对标本地商业顾问和老板 IP 账号。',
      '借鉴他们的问题切入、案例表达和评论区痛点捕捉。',
      '我的不同是更强调本地业务和私域转化。',
      '目前最缺稳定选题、脚本和拍摄节奏。',
    ];
    for (let index = 0; index < answers.length; index += 1) {
      await questionnaireCard.locator('textarea').nth(index).fill(answers[index]);
    }
    await questionnaireCard.locator('.btn-primary').click();

    const benchmarkCard = page.locator('.ip-card').first();
    await benchmarkCard.locator('input').nth(0).fill('商业小纸条');
    await benchmarkCard.locator('input').nth(1).fill('https://example.com/benchmark-account');
    await benchmarkCard.locator('input').nth(2).fill('https://www.douyin.com/video/7123456789012345678');
    await benchmarkCard.locator('textarea').nth(0).fill('问题切入直接，能让老板快速代入；希望 AI 判断爆款点如何迁移到本地老板 IP。');

    const optionalDetails = benchmarkCard.locator('.ip-benchmark-optional');
    await optionalDetails.locator('summary').click();
    await benchmarkCard.locator('input').nth(3).fill('为什么你的短视频越拍越像广告？');
    await benchmarkCard.locator('input').nth(4).fill('点赞 3.2 万，评论 1800+，收藏明显高于普通内容');
    await benchmarkCard.locator('textarea').nth(1).fill('开头直接戳中老板害怕广告感的痛点，中段用咨询案例拆解信任建立，结尾引导评论区自查。');
    await benchmarkCard.locator('textarea').nth(2).fill('评论区大量用户问“怎么表达才不像卖东西”“私域怎么接”，说明转化链路可以从诊断清单切入。');
    await benchmarkCard.locator('.btn-primary').click();

    const generateCard = page.locator('.ip-card').first();
    await expect(generateCard.locator('.btn-primary')).toBeEnabled();
    await generateCard.locator('.btn-primary').click();

    await expect(page.locator('.ip-result')).toBeVisible({ timeout: 240_000 });
    const plan = await page.evaluate(() => {
      const planIndex = JSON.parse(localStorage.getItem('tiktok.ipOperator.plans.local-guest') || '[]');
      const planId = planIndex[0];
      return JSON.parse(localStorage.getItem(`tiktok.ipOperator.plan.${planId}.local-guest`) || 'null');
    });

    writeDebugPlan(plan);

    expect(plan).toBeTruthy();
    expect(['ready', 'needs_input']).toContain(plan.status);
    expect(plan.diagnosis?.positioningStatement || '').not.toHaveLength(0);
    expect(plan.benchmarkAnalyses?.[0]?.borrowablePoints?.length || 0).toBeGreaterThanOrEqual(2);
    expect(plan.benchmarkAnalyses?.[0]?.nonCopyablePoints?.length || 0).toBeGreaterThanOrEqual(2);
    expect(plan.opportunityMap?.length || 0).toBeGreaterThanOrEqual(3);
    expect(plan.topics?.length || 0).toBeGreaterThanOrEqual(8);
    expect((plan.scripts || []).filter((script: any) => script.type === 'full').length).toBeGreaterThanOrEqual(3);
    expect(plan.sevenDayTestPlan?.length || 0).toBe(7);
    expect(plan.qualityCheck?.overallScore).toBeGreaterThanOrEqual(0);
    expect(plan.qualityCheck?.overallScore).toBeLessThanOrEqual(100);

    const topicsWithQuality = (plan.topics || []).filter((topic: any) => topic.qualityLevel && typeof topic.productionReadinessScore === 'number');
    const scriptsWithQuality = (plan.scripts || []).filter((script: any) => script.qualityLevel && script.scriptSummary);
    expect(topicsWithQuality.length).toBeGreaterThanOrEqual(8);
    expect(scriptsWithQuality.length).toBeGreaterThanOrEqual(3);
    expect(plan.qualityAudit?.overallScore).toBeGreaterThanOrEqual(0);
    expect(plan.qualityAudit?.overallScore).toBeLessThanOrEqual(100);
    expect(plan.qualityAudit?.dimensions?.length || 0).toBeGreaterThanOrEqual(6);
    expect(plan.qualityAudit?.productionGate?.reason || '').not.toHaveLength(0);

    writeRealLlmSampleRecord(plan);
  });
});

async function seedEmptyLoggedInWorkbench(page: Page) {
  await page.addInitScript({ path: scenarioSeedPath });
  await page.addInitScript(() => {
    window.__seedIpOperatorScenario?.('empty');
    localStorage.setItem('tiktok.currentTab', 'ip-operator');
  });
}

function writeDebugPlan(plan: any): void {
  fs.mkdirSync(path.dirname(debugPlanPath), { recursive: true });
  fs.writeFileSync(debugPlanPath, `${JSON.stringify(plan, null, 2)}\n`, 'utf8');
}

function writeRealLlmSampleRecord(plan: any): void {
  const fullScript = (plan.scripts || []).find((script: any) => script.type === 'full');
  const lines = [
    '# IP操盘工作台 V1 真实 LLM 样例验收记录',
    '',
    `验收时间：${new Date().toISOString()}`,
    '',
    '## 输入 IP 档案',
    '',
    '- IP 名称：林川老板增长笔记',
    '- 身份：传统行业增长顾问',
    '- 行业：本地生活 / 家居建材 / 教育培训',
    '- 目标用户：传统行业中小企业老板',
    '- 目标：生成可执行的 7 天内容测试方案',
    '',
    '## 输出摘要',
    '',
    `- 最终状态：${plan.status}`,
    `- 质量分：${plan.qualityCheck?.overallScore ?? '未返回'}/100`,
    `- P2A 质量审计分：${plan.qualityAudit?.overallScore ?? '未返回'}/100`,
    `- 选题质量分层数量：${(plan.topics || []).filter((topic: any) => topic.qualityLevel).length}`,
    `- 脚本质量分层数量：${(plan.scripts || []).filter((script: any) => script.qualityLevel).length}`,
    `- 制作门槛说明：${text(plan.qualityAudit?.productionGate?.reason)}`,
    `- 定位一句话：${text(plan.diagnosis?.positioningStatement)}`,
    `- 待补资料：${text((plan.missingInputs || []).map((item: any) => `${item.type}：${item.description}`))}`,
    '',
    '## 代表选题',
    '',
    ...(plan.topics || []).slice(0, 3).map((topic: any) => `- ${text(topic.title)}：${text(topic.fitReason)}；质量 ${text(topic.qualityLevel)}；可制作分 ${text(topic.productionReadinessScore)}`),
    '',
    '## 完整口播摘要',
    '',
    `- 标题：${text(fullScript?.title)}`,
    `- 脚本质量：${text(fullScript?.qualityLevel)}`,
    `- 脚本摘要：${text(fullScript?.scriptSummary)}`,
    `- 钩子：${text(fullScript?.hook)}`,
    `- 观点：${text(fullScript?.viewpoint)}`,
    '',
    '## 人工复核结论',
    '',
    '- 结论：自动结构验收通过；仍需人工复核内容是否符合具体客户语气、案例真实性和业务边界。',
    '',
  ];
  fs.mkdirSync(path.dirname(sampleRecordPath), { recursive: true });
  fs.writeFileSync(sampleRecordPath, `${lines.join('\n').trim()}\n`, 'utf8');
}

function text(value: unknown): string {
  if (value === null || value === undefined || value === '') return '待补充';
  if (Array.isArray(value)) return value.map(text).filter(Boolean).join('；') || '待补充';
  if (typeof value === 'object') return JSON.stringify(value);
  return String(value);
}