games-development-ai/spikes/playtest-agent/playtest-agent-spike.cjs

/**
 * playtest-agent-spike.cjs —— 「测试 agent 替 E/G/H」spike 原型(2026-07-09 创始人拍板:拆着杀,spike 先行)。
 *
 * 职责:一个多模态 agent 按 brief 真玩一款游戏——每步看截图(+运行日志尾巴,均为人类可见的诚实证据,
 * 不读 _forensicsView 契约状态)→ 决定动作(tap/key/wait)→ CDP 派发 → 最终给结构化裁决与修复反馈。
 * 考卷 = 今日基线 10 局已知真相游戏(5 过关 + 冤案 r2 + 真哑 puzzle-r3 + 真空壳 trpg-r3 等),
 * 判对冤案与真哑 = 证明它比旧 E/G/H 强。
 *
 * 用法: node playtest-agent-spike.cjs <gameId> [--model=glm-5.2] [--steps=12] [--port=4998] [--cdp=9331]
 * 输出: stdout JSON 一行(裁决) + 转写落 transcripts/playtest-<gid>.transcript.json(与考卷存档同目录)
 * 依赖: 静态服务与 headless Chrome 已起(同 probe-tap 的起法);NEWAPI_KEY 从环境注入(runner 负责)。
 */
'use strict';
const fs = require('node:fs');
const http = require('node:http');
const path = require('node:path');
// 仓相对路径(本档位于 spikes/playtest-agent/,仓根 = ../../):复用 harness 的 CDP 会话件与 game-runtime 的 ws 依赖
const REPO = path.join(__dirname, '..', '..');
const H = require(path.join(REPO, 'game-runtime', 'test', 'harness', 'browser-evidence.cjs'));
const { CdpSession } = H;
const WebSocket = require(path.join(REPO, 'game-runtime', 'node_modules', 'ws'));

const args = process.argv.slice(2);
const GID = args.find((a) => !a.startsWith('--'));
const opt = (k, d) => { const m = args.find((a) => a.startsWith(`--${k}=`)); return m ? m.split('=')[1] : d; };
const MODEL = opt('model', 'glm-5.2');
const MAX_STEPS = parseInt(opt('steps', '12'), 10);
const PORT = opt('port', '4998');
const CDP = `http://localhost:${opt('cdp', '9331')}`;
const BASE_URL = process.env.NEWAPI_BASE_URL || 'http://100.64.0.8:3000';
const KEY = process.env.NEWAPI_KEY;
if (!KEY) { console.error('缺 NEWAPI_KEY(runner 应从凭据档注入)'); process.exit(2); }
// 转写落 transcripts/(入仓后的固定落点;复跑产物与考卷存档同目录)
const SCRATCH = path.join(__dirname, 'transcripts');
fs.mkdirSync(SCRATCH, { recursive: true });
const delay = (ms) => new Promise((r) => setTimeout(r, ms));

function httpJson(method, urlStr) {
  return new Promise((resolve, reject) => {
    const u = new URL(urlStr);
    const req = http.request({ hostname: u.hostname, port: u.port, path: u.pathname + u.search, method, timeout: 15000 },
      (res) => { let b = ''; res.on('data', (c) => (b += c)); res.on('end', () => { try { resolve(b ? JSON.parse(b) : {}); } catch (e) { reject(e); } }); });
    req.on('error', reject); req.on('timeout', () => req.destroy(new Error('timeout'))); req.end();
  });
}

/** 调 new-api /v1/chat/completions(OpenAI 兼容,image_url data-URI;trust_env 语义=Node 原生 fetch 不走系统代理需显式 no proxy——用 http 模块直连,天然绕代理)。 */
function chat(messages, maxTokens) {
  const body = JSON.stringify({ model: MODEL, max_tokens: maxTokens || 4000, messages });
  const u = new URL(BASE_URL + '/v1/chat/completions');
  return new Promise((resolve, reject) => {
    const req = http.request({ hostname: u.hostname, port: u.port, path: u.pathname, method: 'POST',
      headers: { 'Authorization': `Bearer ${KEY}`, 'Content-Type': 'application/json', 'Content-Length': Buffer.byteLength(body) }, timeout: 120000 },
      (res) => { let b = ''; res.on('data', (c) => (b += c)); res.on('end', () => {
        try {
          const d = JSON.parse(b);
          if (res.statusCode !== 200) return reject(new Error(`HTTP ${res.statusCode}: ${b.slice(0, 200)}`));
          const msg = (d.choices || [{}])[0].message || {};
          resolve({ text: msg.content || '', usage: d.usage || {} });
        } catch (e) { reject(new Error('响应非 JSON: ' + b.slice(0, 200))); }
      }); });
    req.on('error', reject); req.on('timeout', () => req.destroy(new Error('LLM 调用超时'))); req.write(body); req.end();
  });
}

const SYSTEM = `你是小游戏【真人视角测试员】。你只依赖两样人类可见的证据:游戏截图 + 运行日志尾巴。绝不臆测截图里没有的东西。
任务:按 brief 把这款游戏当真人一样玩起来(点按钮/点目标/等待),验证三件事:①玩得通(能从菜单进游戏、操作有响应、能推进到有结果)②有决策层(点哪有区别、有输赢/取舍)③与 brief 是一款游戏。
【坐标必读】截图上烧了粉色坐标尺:细格 50px、粗格 100px,左缘 y100..y800、上缘 x100..x300 是刻度标签。
报 tap 坐标前,先对着刻度读出目标的位置(如:按钮中心在 y400 与 y500 粗线正中间 → y=450),绝不凭感觉估。
【落点回显】截图上的亮黄色圆环+十字 = 你上一步 tap 的真实落点。每一步先看它:落点不在你想点的目标上,
就按偏差校正坐标(如圆环在按钮下方 60px → 下一步 y 减 60);落点在目标上而画面没反应,才是游戏的问题。
每一步你只输出一个 JSON(不要 markdown 围栏):
  {"act":"tap","x":195,"y":489,"why":"点开始按钮(按钮上沿贴 y500 线上方约半格)"}   —— 坐标系 390×844,原点左上
  {"act":"key","key":"ArrowLeft","why":"向左"}
  {"act":"wait","ms":800,"why":"等动画/观察窗"}
  {"act":"verdict","pass":true|false,"canSeeScreenshot":true|false,"problems":["…"],"feedback":"给写游戏的 agent 的修复指引(具体到现象与复现步骤)","summary":"一句话"}
规则:每次 tap 后对照上一步截图判断画面变没变;没变就【必须换坐标】(先沿 y 轴上下各挪 50 试)或 wait,同一坐标最多连点 2 次;若换过 ≥3 处坐标画面仍毫无响应,判 fail 并把「点了哪些坐标都没反应」写进 problems;≤${'${MAX_STEPS}'} 步内必须给 verdict;看不到截图时如实报 canSeeScreenshot=false 并判 fail(problems 注明「测试员看不到画面」)。`;

(async () => {
  // 读 brief(诚实来源:生成时落盘的 brief.json)
  const briefPath = `/Users/lili/Project/games-development-ai/game-runtime/games/amgen-${GID}/evidence/brief.json`;
  const brief = JSON.parse(fs.readFileSync(briefPath, 'utf-8')).brief;

  // 连页面(照 play.cdp.cjs:DPR=1 省 token,agent 看 390×844 原比例)
  const created = await httpJson('PUT', `${CDP}/json/new?${encodeURI('about:blank')}`);
  const ws = await new Promise((res, rej) => { const s = new WebSocket(created.webSocketDebuggerUrl, { maxPayload: 256 * 1024 * 1024 }); s.on('open', () => res(s)); s.on('error', rej); });
  const cdp = new CdpSession(ws, created.id);
  await cdp.send('Page.enable'); await cdp.send('Runtime.enable');
  await cdp.send('Emulation.setDeviceMetricsOverride', { width: 390, height: 844, deviceScaleFactor: 1, mobile: true });
  try { await cdp.send('Emulation.setTouchEmulationEnabled', { enabled: true, maxTouchPoints: 1 }); } catch (_) {}
  await cdp.send('Page.navigate', { url: `http://localhost:${PORT}/games/_wg1-gen/${GID}/index.html` });
  const deadline = Date.now() + 30000; await delay(800);
  for (;;) {
    const st = await cdp.evaluate('({ b: !!window.__genBooted, e: !!window.__gameHostEngineInitFired, err: (window.__genBootError||window.__gameHostBootError||null) })');
    if (st && st.err) { console.log(JSON.stringify({ gid: GID, verdict: 'boot-dead', err: st.err })); process.exit(0); } // A 门职责,测试员不接
    if (st && st.b && st.e) break;
    if (Date.now() > deadline) throw new Error('装载超时');
    await delay(300);
  }

  // 坐标尺覆盖层(spike v2:治 vision-LLM 坐标接地弱——M3 实测把 y=489 的按钮估到 650~720,连点六次全空)。
  //   细网格 50px、粗线+刻度标签 100px;pointer-events:none 不挡输入;只进截图,游戏逻辑无感知。
  await cdp.evaluate(`(function(){
    if (document.getElementById('__testgrid')) return;
    var d = document.createElement('div');
    d.id = '__testgrid';
    d.style.cssText = 'position:fixed;inset:0;pointer-events:none;z-index:99999;' +
      'background-image:linear-gradient(rgba(255,0,90,.28) 1px,transparent 1px),linear-gradient(90deg,rgba(255,0,90,.28) 1px,transparent 1px),' +
      'linear-gradient(rgba(255,0,90,.14) 1px,transparent 1px),linear-gradient(90deg,rgba(255,0,90,.14) 1px,transparent 1px);' +
      'background-size:100px 100px,100px 100px,50px 50px,50px 50px;';
    for (var y = 100; y < 844; y += 100) {
      var t = document.createElement('span');
      t.textContent = 'y' + y;
      t.style.cssText = 'position:absolute;left:2px;top:' + (y - 14) + 'px;font:12px monospace;color:#ff005a;background:rgba(255,255,255,.75);padding:0 2px;';
      d.appendChild(t);
    }
    for (var x = 100; x < 390; x += 100) {
      var s = document.createElement('span');
      s.textContent = 'x' + x;
      s.style.cssText = 'position:absolute;top:2px;left:' + (x - 14) + 'px;font:12px monospace;color:#ff005a;background:rgba(255,255,255,.75);padding:0 2px;';
      d.appendChild(s);
    }
    document.body.appendChild(d);
  })()`);
  const shot = async () => {
    const r = await cdp.send('Page.captureScreenshot', { format: 'jpeg', quality: 62 });
    return 'data:image/jpeg;base64,' + r.data;
  };
  const logTail = async () => cdp.evaluate('(function(){ var lg=window.__gameLog||[]; return lg.slice(-6).map(function(e){ return (e.scope||"")+"/"+(e.tag||"")+":"+String(e.msg||"").slice(0,60); }); })()');

  // 落点回显:tap 后在页面放亮黄圆环+十字(pointer-events:none),下一帧截图里 agent 能看到自己点在哪。
  const placeTapMarker = (x, y) => cdp.evaluate(`(function(){
    var m = document.getElementById('__tapmark');
    if (!m) { m = document.createElement('div'); m.id='__tapmark'; document.body.appendChild(m); }
    m.style.cssText = 'position:fixed;left:${x - 14}px;top:${y - 14}px;width:28px;height:28px;pointer-events:none;z-index:100000;' +
      'border:3px solid #ffe000;border-radius:50%;box-shadow:0 0 0 1px #000;' +
      'background:radial-gradient(circle,rgba(255,224,0,.9) 2px,transparent 3px);';
  })()`);

  const transcript = [];
  const history = [];  // 文本化历史(旧截图不重发,只留动作与观察摘要)
  let usageIn = 0, usageOut = 0, verdict = null;
  let lastTap = null, sameTapStreak = 0, pushedBack = false;
  const tapPoints = new Set();

  for (let step = 0; step < MAX_STEPS && !verdict; step++) {
    const img = await shot();
    const lg = await logTail();
    // 防误导硬提示(runner 侧确定性注入,不靠模型自觉)
    let hint = '';
    if (sameTapStreak >= 2) hint = `\n⚠ 你已连续 ${sameTapStreak} 次点同一坐标(${lastTap});规则:立即换点(先 y±50)或 wait,不许再点原处。`;
    const userParts = [
      { type: 'text', text: `brief:${brief}\n已执行历史:${history.length ? history.join(' → ') : '(第一步)'}\n运行日志尾:${JSON.stringify(lg)}${hint}\n当前截图(第 ${step + 1} 步,共 ${MAX_STEPS} 步预算;黄圆环=你上一步落点):` },
      { type: 'image_url', image_url: { url: img } },
      { type: 'text', text: '输出下一个动作 JSON(或 verdict):' },
    ];
    const resp = await chat([{ role: 'system', content: SYSTEM.replace('${MAX_STEPS}', String(MAX_STEPS)) }, { role: 'user', content: userParts }], 3000);
    usageIn += resp.usage.prompt_tokens || 0; usageOut += resp.usage.completion_tokens || 0;
    let act;
    try { act = JSON.parse(resp.text.replace(/^```json?\s*|```\s*$/g, '').trim()); }
    catch (_) { history.push(`step${step}:输出非JSON(截断记录)`); transcript.push({ step, raw: resp.text.slice(0, 300) }); continue; }
    transcript.push({ step, act, log: lg });
    if (act.act === 'verdict') {
      // 反早退:fail 裁决若探索不足(独立落点 <3),驳回一次逼它继续测(narrative-r1 两拍即弃案的确定性补丁)
      if (act.pass === false && act.canSeeScreenshot !== false && tapPoints.size < 3 && !pushedBack) {
        pushedBack = true;
        history.push(`裁决被驳回(仅试过 ${tapPoints.size} 处落点,规则要求 ≥3 处;继续测试)`);
        continue;
      }
      verdict = act; break;
    }
    if (act.act === 'tap' && typeof act.x === 'number') {
      await cdp.send('Input.dispatchTouchEvent', { type: 'touchStart', touchPoints: [{ x: act.x, y: act.y }] });
      await cdp.send('Input.dispatchTouchEvent', { type: 'touchEnd', touchPoints: [] });
      await placeTapMarker(act.x, act.y);
      const key = `${Math.round(act.x)},${Math.round(act.y)}`;
      sameTapStreak = (lastTap === key) ? sameTapStreak + 1 : 1;
      lastTap = key;
      tapPoints.add(key);
      history.push(`tap(${key})[${act.why || ''}]`);
      await delay(450);
    } else if (act.act === 'key' && act.key) {
      const code = act.key.startsWith('Arrow') ? act.key : act.key;
      await cdp.send('Input.dispatchKeyEvent', { type: 'keyDown', code, key: act.key });
      await cdp.send('Input.dispatchKeyEvent', { type: 'keyUp', code, key: act.key });
      history.push(`key(${act.key})`);
      sameTapStreak = 0;
      await delay(300);
    } else if (act.act === 'wait') {
      history.push(`wait(${act.ms || 500})`);
      sameTapStreak = 0;
      await delay(Math.min(act.ms || 500, 3000));
    } else {
      history.push(`step${step}:未知动作`);
    }
  }
  if (!verdict) verdict = { act: 'verdict', pass: false, problems: ['步数预算内未给出裁决(测试员超时)'], feedback: '', summary: 'no-verdict', degraded: true };

  const out = { gid: GID, model: MODEL, pass: verdict.pass, canSee: verdict.canSeeScreenshot, problems: verdict.problems || [], feedback: verdict.feedback || '', summary: verdict.summary || '', steps: transcript.length, tokens: { in: usageIn, out: usageOut } };
  fs.writeFileSync(path.join(SCRATCH, `playtest-${GID}.transcript.json`), JSON.stringify({ out, transcript }, null, 1));
  console.log(JSON.stringify(out));
  process.exit(0);
})().catch((e) => { console.error('SPIKE-ERR:', e.message); process.exit(2); });