dsh-browser-view 0.8.0 → 0.8.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/lib/index.js CHANGED
@@ -178,8 +178,8 @@ const BROWSER_DEFAULT_CONFIG = {
178
178
  };
179
179
  //#endregion
180
180
  //#region lib/types/lib/prompts.js
181
- const FULL_TOOL_PROMPT = "浏览器工具操作一个共享浏览器会话,支持多标签页与多配置文件。定位操作元素最省事的方式是意图语义定位:browser_click/browser_type/browser_select 直接传 what(元素名称/label,如 what:'登录'),页内打分匹配并自动注册 ref,一次调用替代 快照→查找→点击 三轮;零命中报错会附最相近候选。需要精确控制时再用 selector(CSS)/ref(aria 快照)。打开页面后可先 browser_snapshot 读文本与链接、browser_elements 读表单结构(含关联 label 与展开的下拉选项);browser_find(query 传名称,只返回命中 ref+区域上下文,大页面最省 token;零命中会给相近候选)适合多目标排查,需要全页结构时再用 browser_snapshot { mode: \"aria\" }(可交互元素带 ref=eN,click/type/select/press 直接传 ref 代替 CSS 选择器);下拉选择(原生 select 与 antd/element 组件库)用 browser_select 传 selector/ref/what + option 文本一步完成;固定小序列(点击→等待→输入→提交)用 browser_batch 一次执行减少往返;标签管理用 browser_tabs;下载用 browser_download(wait=true 可等新下载);登录态用 browser_cookies。做自动化测试时:用 browser_wait_for 显式等待页面稳定/元素出现,用 browser_assert 断言(失败自动截图;visual={baseline} 可做视觉回归,先用 browser_screenshot { baseline } 建基线),用 browser_network 检查接口调用与 4xx/5xx,跑完一轮可用 browser_record save 保存操作序列、browser_replay 回放做回归(选择器失效自动按指纹自愈;vars/data 支持参数化与数据驱动)。每次 replay/batch 都记入运行历史,browser_report list/export 可查看与导出回归报告。抓取结构化数据用 browser_extract(fields 定义 CSS 字段,可翻页,可导出 CSV/JSONL)。遇到需要登录的网站:先 browser_open 打开登录页,告知用户请在浏览器窗口中完成登录,再调用 browser_wait_for_login 等待(可给 successSelector/successUrl);验证码/二次确认等更广的「人补一步」场景用 browser_pause_for_human(提示语会显示在实时画面面板);配置 userDataDir 时登录态会持久化。用户可通过界面右下角面板的实时画面查看浏览器。";
182
- const CORE_TOOL_PROMPT = "浏览器工具操作一个共享浏览器会话(toolset=core:已精简工具集)。定位操作元素最省事的方式是意图语义定位:browser_click/browser_type/browser_select 直接传 what(元素名称/label,如 what:'登录'),一次调用替代快照→查找→点击;零命中报错附最相近候选。需要精确控制时再用 selector/ref:打开页面后用 browser_snapshot 读内容(scope 限定区域、since 增量重读省 token),browser_find(query 传名称,返回 ref),点击/输入/按键/下拉用 browser_click/browser_type/browser_press/browser_select(selector 或 ref);固定小序列用 browser_batch;等待用 browser_wait/browser_wait_for,断言用 browser_assert;接口检查用 browser_network;标签 browser_tabs;下载/上传 browser_download/browser_upload。需要人工:登录用 browser_wait_for_login,验证码/二次确认用 browser_pause_for_human,敏感动作(提交/删除/支付)前用 browser_confirm。环境异常先跑 browser_doctor 诊断。录制/回放/报告/抽取/profile/隔离 context 等工具当前未注册(toolset=core),需要时把配置改为 toolset: all 并重启 DSH。用户可通过界面右下角面板的实时画面查看浏览器。";
181
+ const FULL_TOOL_PROMPT = "浏览器工具操作一个共享浏览器会话,支持多标签页与多配置文件。定位操作元素最省事的方式是意图语义定位:browser_click/browser_type/browser_select 直接传 what(元素名称/label,如 what:'登录'),页内打分匹配并自动注册 ref,一次调用替代 快照→查找→点击 三轮;零命中报错会附最相近候选。需要精确控制时再用 selector(CSS)/ref(aria 快照)。打开页面后可先 browser_snapshot 读文本与链接、browser_elements 读表单结构(含关联 label 与展开的下拉选项);browser_find(query 传名称,只返回命中 ref+区域上下文,大页面最省 token;零命中会给相近候选)适合多目标排查,需要全页结构时再用 browser_snapshot { mode: \"aria\" }(可交互元素带 ref=eN,click/type/select/press 直接传 ref 代替 CSS 选择器);下拉选择(原生 select 与 antd/element 组件库)用 browser_select 传 selector/ref/what + option 文本一步完成;固定小序列(点击→等待→输入→提交)用 browser_batch 一次执行减少往返;标签管理用 browser_tabs;下载用 browser_download(wait=true 可等新下载);登录态用 browser_cookies。做自动化测试时:用 browser_wait_for 显式等待页面稳定/元素出现,用 browser_assert 断言(失败自动截图;visual={baseline} 可做视觉回归,先用 browser_screenshot { baseline } 建基线),用 browser_network 检查接口调用与 4xx/5xx,跑完一轮可用 browser_record save 保存操作序列、browser_replay 回放做回归(选择器失效自动按指纹自愈;vars/data 支持参数化与数据驱动)。每次 replay/batch 都记入运行历史,browser_report list/export 可查看与导出回归报告。抓取结构化数据用 browser_extract(fields 定义 CSS 字段,可翻页,可导出 CSV/JSONL)。遇到需要登录的网站:先 browser_open 打开登录页,告知用户在浏览器窗口完成登录后点面板登录卡上的「我已完成登录」,再调用 browser_wait_for_login 等待(用户点按钮即返回,不自动猜测登录状态;也可给 successSelector/successUrl 作确定性自动条件;无面板环境必须给其一,否则调用会报错);验证码/二次确认等更广的「人补一步」场景用 browser_pause_for_human(提示语会显示在实时画面面板);配置 userDataDir 时登录态会持久化。用户可通过界面右下角面板的实时画面查看浏览器。";
182
+ const CORE_TOOL_PROMPT = "浏览器工具操作一个共享浏览器会话(toolset=core:已精简工具集)。定位操作元素最省事的方式是意图语义定位:browser_click/browser_type/browser_select 直接传 what(元素名称/label,如 what:'登录'),一次调用替代快照→查找→点击;零命中报错附最相近候选。需要精确控制时再用 selector/ref:打开页面后用 browser_snapshot 读内容(scope 限定区域、since 增量重读省 token),browser_find(query 传名称,返回 ref),点击/输入/按键/下拉用 browser_click/browser_type/browser_press/browser_select(selector 或 ref);固定小序列用 browser_batch;等待用 browser_wait/browser_wait_for,断言用 browser_assert;接口检查用 browser_network;标签 browser_tabs;下载/上传 browser_download/browser_upload。需要人工:登录用 browser_wait_for_login(用户在面板点「我已完成登录」即返回;无面板环境必须给 successSelector/successUrl,否则调用报错),验证码/二次确认用 browser_pause_for_human,敏感动作(提交/删除/支付)前用 browser_confirm。环境异常先跑 browser_doctor 诊断。录制/回放/报告/抽取/profile/隔离 context 等工具当前未注册(toolset=core),需要时把配置改为 toolset: all 并重启 DSH。用户可通过界面右下角面板的实时画面查看浏览器。";
183
183
  /** confirmMode 附加指引(与 DSH 入口同款语义,manual 面向宿主审批、auto 说明硬闸)。 */
184
184
  function confirmGuidanceOf(confirmMode) {
185
185
  if (confirmMode === "manual") return "本环境 confirmMode 为 manual:凡提交/删除/支付等不可逆或影响线上数据的敏感动作,执行前必须先调用 browser_confirm(reason 说明操作与影响,等待用户在实时画面面板点允许/拒绝;拒绝或超时即放弃该动作,不得换路径绕过)。";
@@ -534,6 +534,33 @@ function snapshotVersionEntry(cache, version) {
534
534
  if (cache === void 0) return void 0;
535
535
  return cache.history[cache.history.length - 1]?.version === version ? cache.history[cache.history.length - 1] : cache.history.find((e) => e.version === version);
536
536
  }
537
+ const lastStableNavSeqs = /* @__PURE__ */ new WeakMap();
538
+ /** 记录"最近一次成功操作"的导航代际(pageIdentity 打点,零成本)。 */
539
+ function markStableNav(page) {
540
+ lastStableNavSeqs.set(page, pageNavSeq(page));
541
+ }
542
+ /** 当前代际距最近成功操作代际的差值(≥1 = 期间发生过导航)。 */
543
+ function navSinceStable(page) {
544
+ return pageNavSeq(page) - (lastStableNavSeqs.get(page) ?? pageNavSeq(page));
545
+ }
546
+ const externalNavs = /* @__PURE__ */ new WeakMap();
547
+ /** 记录一次疑似外部导航(非工具窗口内发生的 framenavigated)。 */
548
+ function noteExternalNav(page, url) {
549
+ const list = externalNavs.get(page) ?? [];
550
+ list.push({
551
+ url,
552
+ at: Date.now()
553
+ });
554
+ if (list.length > 5) list.shift();
555
+ externalNavs.set(page, list);
556
+ }
557
+ /** 读取并清空疑似外部导航记录,拼装 warning 文案;无记录返回 undefined。 */
558
+ function drainExternalNavWarning(page) {
559
+ const list = externalNavs.get(page);
560
+ if (list === void 0 || list.length === 0) return void 0;
561
+ externalNavs.delete(page);
562
+ return `检测到非工具触发的页面导航(页面自动跳转或人工操作:${list.map((n) => n.url.slice(0, 80)).join(" → ")}),此前的定位结果可能基于被切换后的页面,请核实当前页面状态`;
563
+ }
537
564
  //#endregion
538
565
  //#region lib/types/browser/session/collaboration.js
539
566
  /**
@@ -573,6 +600,22 @@ function resolveConfirm(id, decision) {
573
600
  pendingConfirm.decision = decision;
574
601
  return true;
575
602
  }
603
+ let pendingLoginWait = null;
604
+ function setPendingLoginWait(entry) {
605
+ pendingLoginWait = entry;
606
+ }
607
+ function getPendingLoginWait() {
608
+ return pendingLoginWait;
609
+ }
610
+ function clearPendingLoginWait() {
611
+ pendingLoginWait = null;
612
+ }
613
+ /** 面板提交「用户已完成登录」;id 不匹配(已被新请求取代/超时清理)时返回 false。 */
614
+ function resolveLoginWait(id) {
615
+ if (pendingLoginWait === null || pendingLoginWait.id !== id) return false;
616
+ pendingLoginWait.done = true;
617
+ return true;
618
+ }
576
619
  let pendingPick = null;
577
620
  function setPendingPick(entry) {
578
621
  pendingPick = entry;
@@ -904,6 +947,53 @@ function injectWalker(entry) {
904
947
  const src = `"use strict"; return function (payload) { var __name = (t, v) => t; const w = (${CORE_SRC})(); return (${entry.toString()})(payload, w); };`;
905
948
  return new Function(src)();
906
949
  }
950
+ const pickTopDialog = injectWalker((payload, w) => {
951
+ const effectiveZ = (el) => {
952
+ let node = el;
953
+ while (node !== null) {
954
+ const cs = getComputedStyle(node);
955
+ if (cs.position !== "static" && cs.zIndex !== "auto") {
956
+ const z = Number.parseInt(cs.zIndex, 10);
957
+ if (Number.isFinite(z)) return z;
958
+ }
959
+ node = w.composedParent(node);
960
+ }
961
+ return 0;
962
+ };
963
+ const visible = (el) => {
964
+ if (!(el instanceof HTMLElement)) return false;
965
+ const cs = getComputedStyle(el);
966
+ if (cs.display === "none" || cs.visibility === "hidden") return false;
967
+ const r = el.getBoundingClientRect();
968
+ return r.width > 0 && r.height > 0;
969
+ };
970
+ const collect = (sel) => {
971
+ try {
972
+ return Array.from(document.querySelectorAll(sel)).filter(visible);
973
+ } catch {
974
+ return [];
975
+ }
976
+ };
977
+ let tier = "precise";
978
+ let nodes = collect(payload.precise);
979
+ if (nodes.length === 0) {
980
+ tier = "fuzzy";
981
+ nodes = collect(payload.fuzzy);
982
+ }
983
+ if (nodes.length === 0) return null;
984
+ const zs = nodes.map((n) => effectiveZ(n));
985
+ const maxZ = Math.max(...zs);
986
+ const top = nodes.filter((n, i) => zs[i] === maxZ);
987
+ const outermost = top.filter((n) => !top.some((m) => m !== n && m.contains(n)));
988
+ const pool = outermost.length > 0 ? outermost : top;
989
+ const winner = pool[pool.length - 1];
990
+ if (winner === void 0) return null;
991
+ return {
992
+ selector: w.cssPath(winner),
993
+ z: maxZ,
994
+ tier
995
+ };
996
+ });
907
997
  //#endregion
908
998
  //#region lib/types/browser/session/recorder.js
909
999
  /** 已保存的录制脚本:name → 步骤数组。 */
@@ -1151,15 +1241,109 @@ function cleanupScreenshots(dir, opts) {
1151
1241
  return removed;
1152
1242
  }
1153
1243
  //#endregion
1244
+ //#region lib/types/browser/primitives/ui-library.js
1245
+ const GENERIC = {
1246
+ dialogWords: [
1247
+ "modal",
1248
+ "drawer",
1249
+ "dialog"
1250
+ ],
1251
+ dialogExcludedParts: [
1252
+ "mask",
1253
+ "wrap",
1254
+ "root",
1255
+ "header",
1256
+ "footer",
1257
+ "title",
1258
+ "close",
1259
+ "body",
1260
+ "btn"
1261
+ ],
1262
+ /** 下拉触发器内部件(点它展开) */
1263
+ dropdownTriggerParts: "[class*='selector'], [role='combobox']",
1264
+ /** 下拉浮层容器的子串词(portal 挂载的面板) */
1265
+ dropdownFloatWords: ["select-dropdown", "popper"],
1266
+ /** 通用下拉选项行(select 状态机消费;库精确层另补) */
1267
+ dropdownOptionRows: "[class*='item-option']:not([class*='empty']), [class*='select-dropdown'] li:not([class*='empty'])",
1268
+ /** 可交互「选项」枚举(elements 工具下拉选项清单消费) */
1269
+ dropdownOptionEnumeration: [
1270
+ "[role='option']",
1271
+ "[role='menuitem']",
1272
+ "[class*='select'][class*='dropdown'] [class*='option']",
1273
+ "[class*='select'][class*='dropdown'] [class*='item']",
1274
+ "[role='listbox'] [role='option']"
1275
+ ].join(", "),
1276
+ /** 多选标签清除按钮 */
1277
+ dropdownClearButtons: "[class*='clear'], [aria-label*='clear']",
1278
+ /** 多选标签移除按钮 */
1279
+ dropdownRemoveButtons: "[class*='remove'], [class*='close'], [aria-label*='close']",
1280
+ /** 空态指示器 */
1281
+ emptyStates: "[class*='empty'], [class*='no-data'], [role='status'], [class*='item-empty']",
1282
+ /** 表单项容器(宽松:BEM/-item 子串 + group + form 元素;label 反查主用) */
1283
+ formItemContainers: "[class*='form-item'], [class*='form__item'], [role='group'], form, .form-group",
1284
+ /**
1285
+ * 表单项容器(严格:带分隔符前缀,避免 platform-item 之类无关门类误命中;
1286
+ * elements 工具页内 labelOf 消费)
1287
+ */
1288
+ formItemContainersStrict: "[class*='-form-item'], [class*='__form-item'], form, .form-group",
1289
+ /** 表单项宿主判定(目标是否处于表单语境;含 label 本身) */
1290
+ formItemHosts: "label, [class*='form-item'], [class*='form__item'], [class*='form-group']",
1291
+ /** 表单项 label(dom 原语 label 反查消费) */
1292
+ formItemLabels: "[class*='form-item-label'], [class*='form-item__label'], [class*='label'], label",
1293
+ /** 表单 label(简化集;elements 工具页内 labelOf 消费) */
1294
+ formLabelSimple: "[class*='label'], label"
1295
+ };
1296
+ /**
1297
+ * 注册新库(示例:naive-ui):
1298
+ * const naiveUi: UILibrarySpec = {
1299
+ * name: "naive-ui",
1300
+ * modalContainers: [".n-modal-container"],
1301
+ * drawerContainers: [".n-drawer-container"],
1302
+ * dropdownContainers: [".n-base-select-menu"],
1303
+ * dropdownOptions: [".n-base-select-option"],
1304
+ * };
1305
+ * 追加到 UI_LIBRARIES 即可,弹窗顶层仲裁与下拉浮层发现自动获得该库精确层支持;
1306
+ * 未注册的库仍由通用子串层兜底。配套要求:补一个该库真页面的 e2e
1307
+ * (playground 模式照抄 test:enterprise),防止想当然的类名污染仲裁。
1308
+ */
1309
+ const UI_LIBRARIES = [{
1310
+ name: "antd",
1311
+ modalContainers: [".ant-modal", ".ant-modal-content"],
1312
+ drawerContainers: [".ant-drawer", ".ant-drawer-content-wrapper"],
1313
+ dropdownContainers: [".ant-select-dropdown"],
1314
+ dropdownOptions: [".ant-select-item-option"]
1315
+ }, {
1316
+ name: "element-plus",
1317
+ modalContainers: [".el-dialog"],
1318
+ drawerContainers: [".el-drawer"],
1319
+ dropdownContainers: [".el-select-dropdown"],
1320
+ dropdownOptions: [".el-select-dropdown__item"]
1321
+ }];
1322
+ const flatOf = (libs, key) => libs.flatMap((lib) => lib[key]);
1323
+ /** 弹窗/抽屉容器(库精确层原料,未加 display:none 预过滤)。 */
1324
+ function dialogContainersOf(libs = UI_LIBRARIES) {
1325
+ return [...flatOf(libs, "modalContainers"), ...flatOf(libs, "drawerContainers")];
1326
+ }
1327
+ /** 弹窗模糊层选择器串:子串词 + 子节点排除(:not 双命名)+ display:none 预过滤。 */
1328
+ function composeDialogFuzzy(words = GENERIC.dialogWords, excludedParts = GENERIC.dialogExcludedParts) {
1329
+ return words.map((word) => [`[class*='${word}']:not([style*='display: none'])`, ...excludedParts.map((part) => `:not([class*='${word}-${part}']):not([class*='${word}__${part}'])`)].join("")).join(", ");
1330
+ }
1331
+ /** 下拉浮层容器选择器列表(库精确层 + 通用子串层,均带 display:none 预过滤)。 */
1332
+ function dropdownFloatSelectors(libs = UI_LIBRARIES) {
1333
+ return [...flatOf(libs, "dropdownContainers"), ...GENERIC.dropdownFloatWords.map((w) => `[class*='${w}']`)].map((sel) => `${sel}:not([style*='display: none']):not([class*='hidden'])`);
1334
+ }
1335
+ /** 下拉选项行选择器列表(标准层 role=option 由消费方自带 + 库精确层 + 通用子串层)。 */
1336
+ function dropdownOptionRowsOf(libs = UI_LIBRARIES) {
1337
+ return [...flatOf(libs, "dropdownOptions").map((sel) => `${sel}:not([class*='empty'])`), GENERIC.dropdownOptionRows];
1338
+ }
1339
+ /** 预组合聚合常量(消费方直接 import)。 */
1340
+ const LIBRARY_DIALOG_CONTAINERS = dialogContainersOf();
1341
+ const LIBRARY_DROPDOWN_CONTAINERS = dropdownFloatSelectors();
1342
+ const LIBRARY_DROPDOWN_OPTIONS = dropdownOptionRowsOf();
1343
+ //#endregion
1154
1344
  //#region lib/types/browser/tools/common.js
1155
- /** 组件库下拉框的可见选项选择器并集 (基于通用 WAI-ARIA 与主流组件库规范)。 */
1156
- const DROPDOWN_OPTION_SELECTOR = [
1157
- "[role='option']",
1158
- "[role='menuitem']",
1159
- "[class*='select'][class*='dropdown'] [class*='option']",
1160
- "[class*='select'][class*='dropdown'] [class*='item']",
1161
- "[role='listbox'] [role='option']"
1162
- ].join(", ");
1345
+ /** 组件库下拉框的可见选项枚举(ui-library 注册表单一来源,值与抽取前等价)。 */
1346
+ const DROPDOWN_OPTION_SELECTOR = GENERIC.dropdownOptionEnumeration;
1163
1347
  /** 读取 PNG 文件头(IHDR)中的宽高,不解码像素(导出供测试)。 */
1164
1348
  function pngSize(buf) {
1165
1349
  if (buf.length < 24 || buf.readUInt32BE(0) !== 2303741511) throw new Error("not a PNG file");
@@ -1294,10 +1478,216 @@ function workspaceFilePath(cwd, file) {
1294
1478
  if (p !== root && !p.startsWith(root + "/") && !p.startsWith(root + "\\")) throw new Error(`文件路径越出工作区: ${file}`);
1295
1479
  return p;
1296
1480
  }
1481
+ const AsyncFunctionConstructor = Object.getPrototypeOf(async function() {}).constructor;
1482
+ /** 编译校验(只编译不执行):语法合法返回 true。 */
1483
+ function compiles(body) {
1484
+ try {
1485
+ new AsyncFunctionConstructor(body);
1486
+ return true;
1487
+ } catch {
1488
+ return false;
1489
+ }
1490
+ }
1491
+ /** `/` 出现在这些字符(或关键字)之后按正则字面量处理,否则按除号。 */
1492
+ const REGEX_ALLOWED_AFTER_CHARS = /* @__PURE__ */ new Set("([{,;=:?&|!%^~*+<>-/");
1493
+ const REGEX_ALLOWED_AFTER_WORDS = /* @__PURE__ */ new Set([
1494
+ "return",
1495
+ "typeof",
1496
+ "instanceof",
1497
+ "in",
1498
+ "of",
1499
+ "new",
1500
+ "delete",
1501
+ "void",
1502
+ "throw",
1503
+ "case",
1504
+ "do",
1505
+ "else",
1506
+ "yield",
1507
+ "await"
1508
+ ]);
1509
+ /**
1510
+ * 词法扫描(为 ② 服务,fail-safe):定位最后一个「顶层显式 ;」的结束位置,
1511
+ * 并检测顶层 return 关键字(有则说明调用方自己管理返回,不自动改写)。
1512
+ * 顶层 = 不在字符串/模板/注释/正则内,且 ()[]{} 合并深度为 0(含模板 ${} 插值内)。
1513
+ */
1514
+ function scanTopLevel(script) {
1515
+ let lastSemicolonEnd = null;
1516
+ let topLevelReturn = false;
1517
+ let state = "code";
1518
+ let depth = 0;
1519
+ const tplDepths = [];
1520
+ let prevSig = "";
1521
+ let prevWord = "";
1522
+ const isWordChar = (c) => /[A-Za-z0-9_$]/.test(c);
1523
+ for (let i = 0; i < script.length; i++) {
1524
+ const ch = script[i];
1525
+ switch (state) {
1526
+ case "line":
1527
+ if (ch === "\n") state = "code";
1528
+ continue;
1529
+ case "block":
1530
+ if (ch === "*" && script[i + 1] === "/") {
1531
+ state = "code";
1532
+ i++;
1533
+ }
1534
+ continue;
1535
+ case "sq":
1536
+ case "dq":
1537
+ if (ch === "\\") i++;
1538
+ else if (state === "sq" && ch === "'" || state === "dq" && ch === "\"") state = "code";
1539
+ continue;
1540
+ case "tpl":
1541
+ if (ch === "\\") i++;
1542
+ else if (ch === "`") {
1543
+ state = "code";
1544
+ prevSig = "`";
1545
+ prevWord = "";
1546
+ } else if (ch === "$" && script[i + 1] === "{") {
1547
+ tplDepths.push(depth);
1548
+ state = "code";
1549
+ i++;
1550
+ }
1551
+ continue;
1552
+ case "regex":
1553
+ if (ch === "\\") i++;
1554
+ else if (ch === "[") state = "regexClass";
1555
+ else if (ch === "/") {
1556
+ state = "code";
1557
+ prevSig = "/";
1558
+ prevWord = "";
1559
+ }
1560
+ continue;
1561
+ case "regexClass":
1562
+ if (ch === "\\") i++;
1563
+ else if (ch === "]") state = "regex";
1564
+ continue;
1565
+ default:
1566
+ if (ch === "/" && script[i + 1] === "/") {
1567
+ state = "line";
1568
+ i++;
1569
+ continue;
1570
+ }
1571
+ if (ch === "/" && script[i + 1] === "*") {
1572
+ state = "block";
1573
+ i++;
1574
+ continue;
1575
+ }
1576
+ if (ch === "/") {
1577
+ if (prevSig === "" || REGEX_ALLOWED_AFTER_CHARS.has(prevSig) || REGEX_ALLOWED_AFTER_WORDS.has(prevWord)) state = "regex";
1578
+ else {
1579
+ prevSig = "/";
1580
+ prevWord = "";
1581
+ }
1582
+ continue;
1583
+ }
1584
+ if (ch === "'" || ch === "\"") {
1585
+ state = ch === "'" ? "sq" : "dq";
1586
+ prevSig = ch;
1587
+ prevWord = "";
1588
+ continue;
1589
+ }
1590
+ if (ch === "`") {
1591
+ state = "tpl";
1592
+ prevSig = ch;
1593
+ prevWord = "";
1594
+ continue;
1595
+ }
1596
+ if (ch === "(" || ch === "[" || ch === "{") {
1597
+ depth++;
1598
+ prevSig = ch;
1599
+ prevWord = "";
1600
+ continue;
1601
+ }
1602
+ if (ch === ")" || ch === "]") {
1603
+ depth = Math.max(0, depth - 1);
1604
+ prevSig = ch;
1605
+ prevWord = "";
1606
+ continue;
1607
+ }
1608
+ if (ch === "}") {
1609
+ if (tplDepths.length > 0 && depth === tplDepths[tplDepths.length - 1]) {
1610
+ tplDepths.pop();
1611
+ state = "tpl";
1612
+ } else depth = Math.max(0, depth - 1);
1613
+ prevSig = ch;
1614
+ prevWord = "";
1615
+ continue;
1616
+ }
1617
+ if (ch === ";") {
1618
+ if (depth === 0 && tplDepths.length === 0) lastSemicolonEnd = i + 1;
1619
+ prevSig = ch;
1620
+ prevWord = "";
1621
+ continue;
1622
+ }
1623
+ if (isWordChar(ch)) {
1624
+ let j = i;
1625
+ while (j < script.length && isWordChar(script[j])) j++;
1626
+ const word = script.slice(i, j);
1627
+ if (depth === 0 && tplDepths.length === 0 && word === "return" && prevSig !== "." && !isWordChar(prevSig)) topLevelReturn = true;
1628
+ prevWord = word;
1629
+ prevSig = script[j - 1];
1630
+ i = j - 1;
1631
+ continue;
1632
+ }
1633
+ if (!/\s/.test(ch)) {
1634
+ prevSig = ch;
1635
+ prevWord = "";
1636
+ }
1637
+ continue;
1638
+ }
1639
+ }
1640
+ return {
1641
+ lastSemicolonEnd,
1642
+ topLevelReturn
1643
+ };
1644
+ }
1645
+ /**
1646
+ * 把 eval 脚本包装成可直接交给 page.evaluate 的代码串,并标明实际执行形态。
1647
+ * 语法非法时在 Node 侧抛出引擎原始错误(消息与页内 V8 同源,提前暴露)。
1648
+ */
1649
+ function buildEvalScript(script) {
1650
+ if (compiles(`return (\n${script}\n)`)) return {
1651
+ code: `(async () => (\n${script}\n))()`,
1652
+ mode: "expression"
1653
+ };
1654
+ try {
1655
+ new AsyncFunctionConstructor(script);
1656
+ } catch (err) {
1657
+ throw new Error(`eval 脚本语法错误: ${err.message}`);
1658
+ }
1659
+ const { lastSemicolonEnd, topLevelReturn } = scanTopLevel(script);
1660
+ if (!topLevelReturn && lastSemicolonEnd !== null) {
1661
+ const head = script.slice(0, lastSemicolonEnd);
1662
+ const tail = script.slice(lastSemicolonEnd);
1663
+ if (tail.trim() !== "") {
1664
+ const candidate = `${head}\nreturn (\n${tail}\n);`;
1665
+ if (compiles(candidate)) return {
1666
+ code: `(async () => {\n${candidate}\n})()`,
1667
+ mode: "statements-auto-return"
1668
+ };
1669
+ }
1670
+ }
1671
+ return {
1672
+ code: `(async () => {\n${script}\n})()`,
1673
+ mode: "statements"
1674
+ };
1675
+ }
1297
1676
  /**
1298
1677
  * 规则化抽取:fields 定义抽取结构,next 选择器驱动翻页(maxPages 上限),
1299
1678
  * 行数超 maxRows 截断;format/file 控制序列化与落盘(json/csv/jsonl/markdown)。
1300
1679
  */
1680
+ async function assertExpectedUrl(page, expectUrl, action) {
1681
+ if (typeof expectUrl !== "string" || expectUrl === "") return;
1682
+ const current = page.url();
1683
+ let hit;
1684
+ try {
1685
+ hit = new RegExp(expectUrl).test(current);
1686
+ } catch {
1687
+ hit = current.includes(expectUrl);
1688
+ }
1689
+ if (!hit) throw new Error(`${action} 前置校验失败: expectUrl "${expectUrl}" 与当前页面不符(当前: ${current})——页面可能被切换/回退,请先 browser_snapshot 确认后再操作`);
1690
+ }
1301
1691
  //#endregion
1302
1692
  //#region lib/types/browser/primitives/dom.js
1303
1693
  /**
@@ -1315,9 +1705,9 @@ function extractElementLabel(el) {
1315
1705
  const text = labelledBy.split(/\s+/).map((id) => document.getElementById(id)).filter((n) => !!n).map((n) => (n.innerText || n.textContent || "").trim()).filter(Boolean).join(" ");
1316
1706
  if (text) return text;
1317
1707
  }
1318
- const formItem = el.closest("[class*='form-item'], [class*='form__item'], [role='group'], form, .form-group");
1708
+ const formItem = el.closest(GENERIC.formItemContainers);
1319
1709
  if (formItem) {
1320
- const labelEl = formItem.querySelector("[class*='form-item-label'], [class*='form-item__label'], [class*='label'], label");
1710
+ const labelEl = formItem.querySelector(GENERIC.formItemLabels);
1321
1711
  if (labelEl) {
1322
1712
  const txt = (labelEl.innerText || labelEl.textContent || "").trim();
1323
1713
  if (txt) return txt;
@@ -1512,29 +1902,63 @@ async function endChangeWatch(page, options = {}) {
1512
1902
  }
1513
1903
  //#endregion
1514
1904
  //#region lib/types/browser/primitives/dialog.js
1905
+ /** 标准层:WAI-ARIA 语义,任何库与原生实现通用。 */
1906
+ const ARIA_DIALOG_SELECTORS = [
1907
+ "[role='dialog']",
1908
+ "[role='alertdialog']",
1909
+ "[aria-modal='true']",
1910
+ "dialog[open]"
1911
+ ];
1912
+ /** display:none 内联样式预过滤(动画关闭中的弹窗/keep-alive 残留)。 */
1913
+ const hiddenGuard = (sel) => `${sel}:not([style*='display: none'])`;
1914
+ /** 精确层 = 标准层 + 库精确层(ui-library 注册表单一来源,新增 UI 库 = 追加 spec)。 */
1915
+ const DIALOG_PRECISE_SELECTORS = [...ARIA_DIALOG_SELECTORS, ...LIBRARY_DIALOG_CONTAINERS].map(hiddenGuard).join(", ");
1916
+ /** 通用模糊层(未注册库兜底):子串词 + :not() 排除子节点。 */
1917
+ const DIALOG_FUZZY_SELECTORS = composeDialogFuzzy();
1515
1918
  /**
1516
- * 通用活动弹窗/对话框/抽屉选择器列表
1517
- * 基于 W3C WAI-ARIA 规范 (role="dialog", aria-modal="true") 与现代 Web 通用命名规范
1919
+ * 历史并集 = 精确层 + 模糊层。semantic.ts 弹窗上下文检测与 getActiveDialog 的
1920
+ * domOrder 兜底消费(行为与抽取前的并集等价)。
1518
1921
  */
1519
- const ACTIVE_DIALOG_SELECTORS = [
1520
- "[role='dialog']:not([style*='display: none'])",
1521
- "[role='alertdialog']:not([style*='display: none'])",
1522
- "[aria-modal='true']:not([style*='display: none'])",
1523
- "dialog[open]",
1524
- ".ant-modal:not([style*='display: none'])",
1525
- ".el-dialog:not([style*='display: none'])",
1526
- ".ant-drawer:not([style*='display: none'])",
1527
- ".el-drawer:not([style*='display: none'])",
1528
- "[class*='modal']:not([style*='display: none'])",
1529
- "[class*='drawer']:not([style*='display: none'])",
1530
- "[class*='dialog']:not([style*='display: none'])"
1531
- ].join(", ");
1922
+ const ACTIVE_DIALOG_SELECTORS = [DIALOG_PRECISE_SELECTORS, DIALOG_FUZZY_SELECTORS].join(", ");
1532
1923
  /**
1533
- * 获取当前页面上处于活动状态的最顶层弹窗容器 (Locator)
1924
+ * 获取当前页面上处于活动状态的最顶层弹窗容器;无弹窗返回 null。
1925
+ * ① 快速通道——并集零命中(绝大多数无弹窗的工具调用)直接返回 null,与旧行为
1926
+ * 「空 locator + 下游 count 0」等价,零额外 evaluate 往返;
1927
+ * ② 有弹窗时走分层选择器 + 页内 z-index/可见性仲裁(walker pickTopDialog,
1928
+ * 页内逻辑单一来源),单次 round-trip;
1929
+ * ③ evaluate 失败(极端 CSP/销毁竞态)或复核计数为 0 时回退文档序 last()
1930
+ * 旧行为并标 via: "domOrder",不阻断主流程。
1534
1931
  */
1535
- function getActiveDialog(page) {
1932
+ async function getActiveDialog(page) {
1933
+ try {
1934
+ const unionLoc = page.locator(ACTIVE_DIALOG_SELECTORS);
1935
+ if (typeof unionLoc.count === "function") {
1936
+ if (await withTimeout(unionLoc.count(), 800, "弹窗快速探测").catch(() => 0) === 0) return null;
1937
+ }
1938
+ } catch {}
1939
+ try {
1940
+ const pick = await withTimeout(page.evaluate(pickTopDialog, {
1941
+ precise: DIALOG_PRECISE_SELECTORS,
1942
+ fuzzy: DIALOG_FUZZY_SELECTORS
1943
+ }), 1500, "活动弹窗仲裁");
1944
+ if (pick !== null && typeof pick === "object" && typeof pick.selector === "string" && pick.selector !== "") {
1945
+ const hit = pick;
1946
+ const loc = page.locator(hit.selector);
1947
+ if (typeof loc.count === "function") {
1948
+ if (await withTimeout(loc.count(), 800, "活动弹窗复核").catch(() => 0) > 0) return {
1949
+ locator: typeof loc.first === "function" ? loc.first() : loc,
1950
+ via: "zIndex",
1951
+ z: hit.z,
1952
+ tier: hit.tier
1953
+ };
1954
+ }
1955
+ }
1956
+ } catch {}
1536
1957
  const dialogLoc = page.locator(ACTIVE_DIALOG_SELECTORS);
1537
- return typeof dialogLoc.last === "function" ? dialogLoc.last() : dialogLoc.first();
1958
+ return {
1959
+ locator: typeof dialogLoc.last === "function" ? dialogLoc.last() : dialogLoc.first(),
1960
+ via: "domOrder"
1961
+ };
1538
1962
  }
1539
1963
  //#endregion
1540
1964
  //#region lib/types/browser/primitives/select.js
@@ -1553,14 +1977,14 @@ async function executeUniversalSelect(ctx) {
1553
1977
  throw new Error(`原生下拉无选项 "${want}"; 可用选项: ${options.map((t) => t.trim()).join(" / ")}`);
1554
1978
  }
1555
1979
  let resolvedTarget = target;
1556
- if (await withTimeout(target.evaluate((el) => {
1980
+ if (await withTimeout(target.evaluate((el, hosts) => {
1557
1981
  if (el.tagName.toLowerCase() === "label") return true;
1558
1982
  const cls = typeof el.className === "string" ? el.className : "";
1559
1983
  if (/(?:form-item|form__item|form-group)/i.test(cls)) return true;
1560
- if (el.closest("label, [class*='form-item'], [class*='form__item'], [class*='form-group']")) return !el.matches("[role='combobox'], select, [class*='select-selector']");
1984
+ if (el.closest(hosts)) return !el.matches("[role='combobox'], select, [class*='select-selector']");
1561
1985
  if (!el.matches("[role='combobox'], select, [class*='select'], [class*='selector']")) return true;
1562
1986
  return false;
1563
- }), 1500, "探测 label 或非标文本容器").catch(() => false)) {
1987
+ }, GENERIC.formItemHosts), 1500, "探测 label 或非标文本容器").catch(() => false)) {
1564
1988
  const descendantSelect = target.locator("xpath=.//*[@role='combobox' or self::select or contains(@class, 'select-selector') or contains(@class, 'select')][not(self::label)]").first();
1565
1989
  if (await descendantSelect.count().catch(() => 0) > 0) resolvedTarget = descendantSelect;
1566
1990
  else {
@@ -1585,15 +2009,15 @@ async function executeUniversalSelect(ctx) {
1585
2009
  const hostLocator = resolvedTarget.locator("xpath=ancestor-or-self::*[@role='combobox' or contains(@class, 'select') or contains(@class, 'selector')][last()]");
1586
2010
  const actualTarget = await hostLocator.count().catch(() => 0) > 0 ? hostLocator.first() : resolvedTarget;
1587
2011
  if (clear) try {
1588
- const clearBtn = actualTarget.locator("[class*='clear'], [aria-label*='clear']").first();
2012
+ const clearBtn = actualTarget.locator(GENERIC.dropdownClearButtons).first();
1589
2013
  if (await clearBtn.count().catch(() => 0) > 0 && await clearBtn.isVisible().catch(() => false)) await clearBtn.click({ timeout: 1e3 }).catch(() => {});
1590
2014
  else {
1591
- const removeBtns = actualTarget.locator("[class*='remove'], [class*='close'], [aria-label*='close']");
2015
+ const removeBtns = actualTarget.locator(GENERIC.dropdownRemoveButtons);
1592
2016
  const removeCount = await removeBtns.count().catch(() => 0);
1593
2017
  for (let i = 0; i < removeCount; i++) await removeBtns.nth(0).click({ timeout: 500 }).catch(() => {});
1594
2018
  }
1595
2019
  } catch {}
1596
- if (!await withTimeout(actualTarget.evaluate((el) => {
2020
+ const probeExpanded = () => withTimeout(actualTarget.evaluate((el) => {
1597
2021
  if (el.getAttribute("aria-expanded") === "true") return true;
1598
2022
  const host = el.closest("[role='combobox'], [class*='select']") || el;
1599
2023
  if (host && host.getAttribute("aria-expanded") === "true") return true;
@@ -1606,8 +2030,11 @@ async function executeUniversalSelect(ctx) {
1606
2030
  if (dd && dd.offsetParent !== null) return true;
1607
2031
  }
1608
2032
  return false;
1609
- }), 2e3, "检查下拉展开状态").catch(() => false)) {
1610
- const triggerChild = actualTarget.locator("[class*='selector'], [role='combobox']").first();
2033
+ }), 2e3, "检查下拉展开状态").catch(() => false);
2034
+ const isExpanded = await probeExpanded();
2035
+ let expandSignalMissed = false;
2036
+ if (!isExpanded) {
2037
+ const triggerChild = actualTarget.locator(GENERIC.dropdownTriggerParts).first();
1611
2038
  let clicked = false;
1612
2039
  if (await triggerChild.count().catch(() => 0) > 0) {
1613
2040
  await triggerChild.click({ timeout: 2e3 }).catch(async () => {
@@ -1625,7 +2052,9 @@ async function executeUniversalSelect(ctx) {
1625
2052
  });
1626
2053
  });
1627
2054
  await page.waitForTimeout(120);
2055
+ expandSignalMissed = !await probeExpanded();
1628
2056
  }
2057
+ const optionWaitMs = expandSignalMissed ? 1500 : Math.min(timeout, 5e3);
1629
2058
  const controlsId = await withTimeout(actualTarget.evaluate((el) => {
1630
2059
  const direct = el.getAttribute("aria-controls") || el.getAttribute("aria-owns");
1631
2060
  if (direct) return direct;
@@ -1633,8 +2062,8 @@ async function executeUniversalSelect(ctx) {
1633
2062
  const inner = host?.querySelector("input[role='combobox'], [aria-controls], [aria-owns]");
1634
2063
  return inner?.getAttribute("aria-controls") || inner?.getAttribute("aria-owns") || host?.getAttribute("aria-controls") || "";
1635
2064
  }), 2e3, "解析下拉关联 controlsId").catch(() => "");
1636
- const optionSelector = "[role='option']:not([style*='height: 0px']):not([class*='empty']), [class*='item-option']:not([class*='empty']), [class*='select-dropdown'] li:not([class*='empty'])";
1637
- const dropdownSelector = "[role='listbox']:not([style*='display: none']), [class*='select-dropdown']:not([style*='display: none']):not([class*='hidden']), [class*='popper']:not([style*='display: none'])";
2065
+ const optionSelector = ["[role='option']:not([style*='height: 0px']):not([class*='empty'])", ...LIBRARY_DROPDOWN_OPTIONS].join(", ");
2066
+ const dropdownSelector = ["[role='listbox']:not([style*='display: none'])", ...LIBRARY_DROPDOWN_CONTAINERS].join(", ");
1638
2067
  let options;
1639
2068
  let targetDropdown = null;
1640
2069
  if (controlsId) {
@@ -1652,7 +2081,7 @@ async function executeUniversalSelect(ctx) {
1652
2081
  }
1653
2082
  await options.first().waitFor({
1654
2083
  state: "visible",
1655
- timeout: Math.min(timeout, 5e3)
2084
+ timeout: optionWaitMs
1656
2085
  }).catch(() => {});
1657
2086
  const visibleTexts = (await options.evaluateAll((els) => els.map((el) => {
1658
2087
  return {
@@ -1662,10 +2091,12 @@ async function executeUniversalSelect(ctx) {
1662
2091
  };
1663
2092
  })).catch(() => [])).map((e) => e.text || e.title || e.aria).filter(Boolean);
1664
2093
  if (visibleTexts.length === 0) {
1665
- const isEmpty = targetDropdown !== null ? await targetDropdown.locator("[class*='empty'], [class*='no-data'], [role='status'], [class*='item-empty']").first().isVisible().catch(() => false) : false;
2094
+ const emptySelector = GENERIC.emptyStates;
2095
+ const isEmpty = targetDropdown !== null ? await targetDropdown.locator(emptySelector).first().isVisible().catch(() => false) : false;
1666
2096
  await actualTarget.click({ timeout: 1e3 }).catch(() => {});
1667
- if (isEmpty) throw new Error(`下拉浮层当前无选项(数据为空态/无可选数据): ${label}`);
1668
- throw new Error(`点击 ${label} 后未出现下拉选项(选项列表为空态)`);
2097
+ const attribution = expandSignalMissed ? ";触发器点击疑似未生效(展开信号缺失,可能被透明遮罩拦截,建议 browser_snapshot 核实或配合 click effect 验证)" : "";
2098
+ if (isEmpty) throw new Error(`下拉浮层当前无选项(数据为空态/无可选数据): ${label}${attribution}`);
2099
+ throw new Error(`点击 ${label} 后未出现下拉选项(选项列表为空态)${attribution}`);
1669
2100
  }
1670
2101
  const wantStripped = want.replace(/\s+/g, "");
1671
2102
  let index = visibleTexts.findIndex((t) => t === want);
@@ -1811,7 +2242,7 @@ async function fillFormFieldHelper(page, field, scope) {
1811
2242
  const root = scope ? page.locator(scope) : page;
1812
2243
  const { key, value } = field;
1813
2244
  try {
1814
- const formItems = root.locator("[class*='form-item'], [class*='form__item'], [role='group'], .form-group");
2245
+ const formItems = root.locator(GENERIC.formItemContainers);
1815
2246
  const count = await formItems.count().catch(() => 0);
1816
2247
  for (let i = 0; i < count; i++) {
1817
2248
  const item = formItems.nth(i);
@@ -1873,6 +2304,45 @@ async function fillFormFieldHelper(page, field, scope) {
1873
2304
  };
1874
2305
  }
1875
2306
  //#endregion
2307
+ //#region lib/types/browser/primitives/drag.js
2308
+ /**
2309
+ * 核心:在两个已解析的 Locator 之间执行拖拽。
2310
+ * 优先 Playwright dragTo(覆盖 HTML5 DnD 与多数组件库);
2311
+ * 失败或调用方指定 preferMouse(如跨 Frame,dragTo 的已知限制)时
2312
+ * 降级为物理鼠标序列。返回实际使用的通道。失败抛错(带可定位的诊断信息)。
2313
+ */
2314
+ async function dragAndDropLocators(page, source, target, opts = {}) {
2315
+ const timeoutMs = opts.timeoutMs ?? 3e3;
2316
+ if (!opts.preferMouse) try {
2317
+ if (typeof source.dragTo === "function") {
2318
+ await source.dragTo(target, { timeout: timeoutMs });
2319
+ return "dragTo";
2320
+ }
2321
+ } catch {}
2322
+ let srcErr = null;
2323
+ const srcBox = await source.boundingBox({ timeout: timeoutMs }).catch((e) => {
2324
+ srcErr = e;
2325
+ return null;
2326
+ });
2327
+ if (!srcBox) throw new Error(`drag 失败: 源元素不可见或未命中(等待 ${timeoutMs}ms)${srcErr instanceof Error ? `: ${srcErr.message}` : ""}`);
2328
+ let tgtErr = null;
2329
+ const tgtBox = await target.boundingBox({ timeout: timeoutMs }).catch((e) => {
2330
+ tgtErr = e;
2331
+ return null;
2332
+ });
2333
+ if (!tgtBox) throw new Error(`drag 失败: 目标元素不可见或未命中(等待 ${timeoutMs}ms)${tgtErr instanceof Error ? `: ${tgtErr.message}` : ""}`);
2334
+ const viewport = typeof page.viewportSize === "function" ? page.viewportSize() : null;
2335
+ const outsideViewport = viewport !== null && (srcBox.x + srcBox.width < 0 || srcBox.x > viewport.width || tgtBox.x + tgtBox.width < 0 || tgtBox.x > viewport.width || srcBox.y + srcBox.height < 0 || srcBox.y > viewport.height || tgtBox.y + tgtBox.height < 0 || tgtBox.y > viewport.height);
2336
+ if (outsideViewport && typeof target.scrollIntoViewIfNeeded === "function") await target.scrollIntoViewIfNeeded({ timeout: timeoutMs }).catch(() => {});
2337
+ const sx = (outsideViewport ? await source.boundingBox({ timeout: timeoutMs }).catch(() => null) : null) ?? srcBox;
2338
+ const tx = (outsideViewport ? await target.boundingBox({ timeout: timeoutMs }).catch(() => null) : null) ?? tgtBox;
2339
+ await page.mouse.move(sx.x + sx.width / 2, sx.y + sx.height / 2);
2340
+ await page.mouse.down();
2341
+ await page.mouse.move(tx.x + tx.width / 2, tx.y + tx.height / 2, { steps: 8 });
2342
+ await page.mouse.up();
2343
+ return "mouse";
2344
+ }
2345
+ //#endregion
1876
2346
  //#region lib/types/browser/session/semantic.js
1877
2347
  /**
1878
2348
  * 页内语义扫描:按 query 打分收集可交互元素(弹窗内优先、可见优先、高分优先)。
@@ -2231,6 +2701,37 @@ const failureContextProbe = injectWalker(() => {
2231
2701
  hasShadow
2232
2702
  };
2233
2703
  });
2704
+ /**
2705
+ * selector 零命中时的相似属性探针(best-effort):提取首个属性片段(或 #id)
2706
+ * 做包含匹配计数——直接暴露「锚点属性全失效」与「组合选择器其他部分写错」两种病因。
2707
+ * 纯提示用途,失败/无属性片段返回空串。
2708
+ */
2709
+ const similarAttrProbe = injectWalker((rawSelector) => {
2710
+ let attr = "";
2711
+ let val = "";
2712
+ const m = rawSelector.match(/\[([\w-]+)(?:\^|\$|\*|~)?=['"]([^'"]+)['"]\]/);
2713
+ if (m !== null) {
2714
+ attr = m[1] ?? "";
2715
+ val = m[2] ?? "";
2716
+ } else {
2717
+ const idm = rawSelector.match(/^#([\w-]+)/);
2718
+ if (idm !== null) {
2719
+ attr = "id";
2720
+ val = idm[1] ?? "";
2721
+ }
2722
+ }
2723
+ if (attr === "" || val === "") return "";
2724
+ try {
2725
+ const safe = val.replace(/["\\]/g, "");
2726
+ if (safe === "") return "";
2727
+ const hits = Array.from(document.querySelectorAll(`[${attr}*="${safe}"]`)).slice(0, 50);
2728
+ if (hits.length === 0) return `页面不存在任何 ${attr} 含 "${safe}" 的元素(锚点属性可能已失效或页面已切换)`;
2729
+ const samples = hits.slice(0, 3).map((el) => el.getAttribute(attr) || "").filter((s) => s !== "").join(" | ").slice(0, 120);
2730
+ return `页面存在 ${hits.length} 个 ${attr} 含 "${safe}" 的元素(组合选择器的其他部分可能写错)${samples !== "" ? `,如: ${samples}` : ""}`;
2731
+ } catch {
2732
+ return "";
2733
+ }
2734
+ });
2234
2735
  /** ref 坐标兜底的名称相似判定:去空白互相包含,或 LCS 占比 ≥ 0.6。 */
2235
2736
  function nameSimilar(a, b) {
2236
2737
  if (a === "" || b === "") return true;
@@ -2312,9 +2813,9 @@ async function resolveTargetLocator(page, args, opts = {}) {
2312
2813
  };
2313
2814
  }
2314
2815
  if (!isIFrame) try {
2315
- const activeDialog = getActiveDialog(page);
2316
- if (activeDialog && typeof activeDialog.getByRole === "function") {
2317
- const inDialog = entry.name !== "" ? activeDialog.getByRole(entry.role, { name: buildFlexibleTextPattern(entry.name) }).first() : activeDialog.getByRole(entry.role).first();
2816
+ const dlg = (await getActiveDialog(page))?.locator;
2817
+ if (dlg && typeof dlg.getByRole === "function") {
2818
+ const inDialog = entry.name !== "" ? dlg.getByRole(entry.role, { name: buildFlexibleTextPattern(entry.name) }).first() : dlg.getByRole(entry.role).first();
2318
2819
  if (await withTimeout(inDialog.count(), 1e3, "弹窗内 role 查找").catch(() => 0) > 0) return {
2319
2820
  locator: inDialog,
2320
2821
  container,
@@ -2436,10 +2937,29 @@ async function resolveTargetLocator(page, args, opts = {}) {
2436
2937
  }
2437
2938
  const what = args.what;
2438
2939
  if (typeof what === "string" && what.trim() !== "" && args.selector === void 0 && args.ref === void 0 && args.mark === void 0 && rowPattern === void 0) {
2940
+ const navSeqBefore = pageNavSeq(page);
2941
+ let urlBefore = null;
2942
+ try {
2943
+ urlBefore = page.url();
2944
+ } catch {}
2439
2945
  const res = await resolveSemanticTarget(page, what, {
2440
2946
  scope: typeof args.scope === "string" && args.scope !== "" ? args.scope : void 0,
2441
2947
  preferredRoles: opts.preferredRoles
2442
2948
  });
2949
+ let urlAfter = urlBefore;
2950
+ try {
2951
+ urlAfter = page.url();
2952
+ } catch {}
2953
+ if (pageNavSeq(page) !== navSeqBefore || urlAfter !== urlBefore) {
2954
+ const raceErr = /* @__PURE__ */ new Error(`页面在语义定位过程中发生变化(${urlBefore} → ${page.url()}),本次扫描结果已作废;请重试或重新 browser_snapshot`);
2955
+ raceErr.locationFailure = {
2956
+ phase: "navRace",
2957
+ candidates: 0,
2958
+ dialogActive: false,
2959
+ hasShadow: false
2960
+ };
2961
+ throw raceErr;
2962
+ }
2443
2963
  if (!("candidates" in res)) return res;
2444
2964
  for (const role of [
2445
2965
  "button",
@@ -2496,6 +3016,14 @@ async function resolveTargetLocator(page, args, opts = {}) {
2496
3016
  ...ctx
2497
3017
  };
2498
3018
  } catch {}
3019
+ const navDelta = navSinceStable(page);
3020
+ if (navDelta >= 1) {
3021
+ let route = "";
3022
+ try {
3023
+ route = page.url().slice(0, 80);
3024
+ } catch {}
3025
+ err.message = `${err.message};页面自上次成功操作后已发生 ${navDelta} 次导航,what 目标可能在另一页面${route !== "" ? `(当前: ${route})` : ""}`;
3026
+ }
2499
3027
  throw err;
2500
3028
  }
2501
3029
  if (args.selector === void 0) throw new Error("需要 selector 或 ref 之一 (或指定 row 与 action 进行表格行级定界,或用 what 语义定位)");
@@ -2513,9 +3041,9 @@ async function resolveTargetLocator(page, args, opts = {}) {
2513
3041
  }
2514
3042
  }
2515
3043
  try {
2516
- const activeDialog = getActiveDialog(page);
2517
- if (activeDialog && typeof activeDialog.locator === "function") {
2518
- const inDialog = activeDialog.locator(rawSelector);
3044
+ const dlg = (await getActiveDialog(page))?.locator;
3045
+ if (dlg && typeof dlg.locator === "function") {
3046
+ const inDialog = dlg.locator(rawSelector);
2519
3047
  if (typeof inDialog.count === "function") {
2520
3048
  const inDialogCount = await withTimeout(inDialog.count(), 1e3, "弹窗内 selector 探测").catch(() => 0);
2521
3049
  if (inDialogCount > 0) {
@@ -2536,7 +3064,43 @@ async function resolveTargetLocator(page, args, opts = {}) {
2536
3064
  let visibleCount;
2537
3065
  let target = typeof baseLocator.first === "function" ? baseLocator.first() : baseLocator;
2538
3066
  if (typeof baseLocator.count === "function") {
2539
- totalCount = await withTimeout(baseLocator.count(), 2e3, "selector 计数").catch(() => 0);
3067
+ try {
3068
+ totalCount = await withTimeout(baseLocator.count(), 2e3, "selector 计数");
3069
+ } catch {
3070
+ const err = /* @__PURE__ */ new Error(`selector 计数超时: ${rawSelector}(页面可能正忙或 DOM 过大,可重试;若持续超时建议 browser_snapshot 核实页面状态)`);
3071
+ err.locationFailure = {
3072
+ phase: "countTimeout",
3073
+ candidates: 0,
3074
+ dialogActive: false,
3075
+ hasShadow: false
3076
+ };
3077
+ throw err;
3078
+ }
3079
+ if (totalCount === 0) {
3080
+ const zeroMatchErr = /* @__PURE__ */ new Error(`selector 未命中: ${rawSelector}(元素未渲染/选择器写错/页面已切换;可用 browser_elements 探查表单结构、browser_wait_for 等元素出现,或 browser_snapshot 取 ref)`);
3081
+ zeroMatchErr.locationFailure = {
3082
+ phase: "zeroMatch",
3083
+ candidates: 0,
3084
+ dialogActive: false,
3085
+ hasShadow: false
3086
+ };
3087
+ try {
3088
+ const ctx = await withTimeout(page.evaluate(failureContextProbe, void 0), 1500, "失败特征探测");
3089
+ if (ctx !== null && typeof ctx === "object") zeroMatchErr.locationFailure = {
3090
+ phase: "zeroMatch",
3091
+ candidates: 0,
3092
+ ...ctx
3093
+ };
3094
+ } catch {}
3095
+ try {
3096
+ const hint = await withTimeout(page.evaluate(similarAttrProbe, rawSelector), 1500, "相似属性探测");
3097
+ if (typeof hint === "string" && hint !== "") {
3098
+ zeroMatchErr.similarHint = hint;
3099
+ zeroMatchErr.message = `${zeroMatchErr.message};${hint}`;
3100
+ }
3101
+ } catch {}
3102
+ throw zeroMatchErr;
3103
+ }
2540
3104
  if (totalCount > 0 && typeof baseLocator.locator === "function") {
2541
3105
  const visibleLocator = baseLocator.locator("visible=true");
2542
3106
  if (typeof visibleLocator.count === "function") {
@@ -2951,6 +3515,8 @@ const DOM_COMPAT_SCRIPT = `
2951
3515
  };
2952
3516
  })();
2953
3517
  `;
3518
+ /** 最近一次工具触碰时间戳(getPage 打点):外部导航判定基准。 */
3519
+ let lastToolTouchAt = 0;
2954
3520
  function attachPage(page, config) {
2955
3521
  const timeoutMs = config.timeoutMs ?? 3e4;
2956
3522
  page.setDefaultTimeout(timeoutMs);
@@ -2974,6 +3540,7 @@ function attachPage(page, config) {
2974
3540
  const url = frame.url();
2975
3541
  if (!url || url === "about:blank") return;
2976
3542
  recordNav(profile, url);
3543
+ if (Date.now() - lastToolTouchAt > 1e4) noteExternalNav(page, url);
2977
3544
  });
2978
3545
  page.on("load", () => {
2979
3546
  const list = historyOf(profile);
@@ -3199,6 +3766,7 @@ async function newPage(config) {
3199
3766
  }
3200
3767
  /** 返回当前活动页面(没有则新建),并保证浏览器已启动。 */
3201
3768
  async function getPage(config) {
3769
+ lastToolTouchAt = Date.now();
3202
3770
  const state = stateOf();
3203
3771
  if (state.browser === null && state.context === null) await launchProfile(currentProfile, config);
3204
3772
  pruneClosedPages(state);
@@ -3262,6 +3830,7 @@ async function closeBrowser() {
3262
3830
  }
3263
3831
  /** Best-effort current page identity, safe when nothing is open.title 有界防页面忙时挂起。 */
3264
3832
  async function pageIdentity(page) {
3833
+ markStableNav(page);
3265
3834
  try {
3266
3835
  const title = await new Promise((resolve, reject) => {
3267
3836
  const t = setTimeout(() => reject(/* @__PURE__ */ new Error("title timeout")), 2e3);
@@ -3866,15 +4435,20 @@ async function snapshotTool(config, args) {
3866
4435
  const isDialogScope = /modal|dialog|drawer/i.test(scope);
3867
4436
  let hint = isDialogScope ? "。目标容器可能已消失(弹窗/抽屉已关闭、已提交或页面已导航)" : "";
3868
4437
  if (isDialogScope && typeof page.locator === "function") try {
3869
- const activeDialogLoc = getActiveDialog(page);
3870
- if ((typeof activeDialogLoc?.count === "function" ? await withTimeout(activeDialogLoc.count(), 1e3, "活动弹窗探测").catch(() => 0) : 0) > 0 && typeof activeDialogLoc.first === "function") {
4438
+ const activeDialogLoc = (await getActiveDialog(page))?.locator;
4439
+ const count = activeDialogLoc !== void 0 && typeof activeDialogLoc.count === "function" ? await withTimeout(activeDialogLoc.count(), 1e3, "活动弹窗探测").catch(() => 0) : 0;
4440
+ if (activeDialogLoc !== void 0 && count > 0 && typeof activeDialogLoc.first === "function") {
3871
4441
  const alternate = await withTimeout(activeDialogLoc.first().evaluate((el) => {
3872
- if (el.getAttribute("role") === "dialog") return "[role='dialog']";
4442
+ const out = [];
4443
+ const role = el.getAttribute("role");
4444
+ if (role === "dialog" || role === "alertdialog") out.push(`[role='${role}']`);
3873
4445
  const cls = el.className;
3874
4446
  if (typeof cls === "string" && cls.trim()) {
3875
- const parts = cls.trim().split(/\s+/);
3876
- return `.${parts.find((c) => /modal|dialog|drawer/i.test(c)) || parts[0]}`;
4447
+ const tokens = cls.trim().split(/\s+/);
4448
+ const relevant = tokens.find((c) => /modal|dialog|drawer/i.test(c)) || tokens[0];
4449
+ out.push(`.${relevant}`);
3877
4450
  }
4451
+ if (out.length > 0) return out.join("");
3878
4452
  return el.tagName.toLowerCase();
3879
4453
  }), 1e3, "解析活动弹窗标识").catch(() => null);
3880
4454
  if (alternate) hint += ` (检测到当前页面存在「${alternate}」容器,可能混淆了弹窗与抽屉类型)`;
@@ -4194,7 +4768,7 @@ async function somFallbackShot(config, exec) {
4194
4768
  async function elementsTool(config, args) {
4195
4769
  const page = await getPage(config);
4196
4770
  const max = Math.min(Math.max(Number(args.max ?? 60) || 60, 1), 200);
4197
- const data = await page.evaluate(([maxN, optionSelector]) => {
4771
+ const data = await page.evaluate(([maxN, optionSelector, formItemSel, formLabelSel]) => {
4198
4772
  /** 提取元素的关联 label:aria-label → label[for] → form-item 标签(antd/element)。 */
4199
4773
  function labelOf(el) {
4200
4774
  const aria = el.getAttribute("aria-label");
@@ -4203,7 +4777,7 @@ async function elementsTool(config, args) {
4203
4777
  const text = document.querySelector(`label[for="${CSS.escape(el.id)}"]`)?.textContent?.trim();
4204
4778
  if (text) return text;
4205
4779
  }
4206
- const formLabel = el.closest("[class*='-form-item'], [class*='__form-item'], form, .form-group")?.querySelector("[class*='label'], label")?.textContent?.trim();
4780
+ const formLabel = el.closest(formItemSel)?.querySelector(formLabelSel)?.textContent?.trim();
4207
4781
  if (formLabel) return formLabel.replace(/[*::]\s*$/, "");
4208
4782
  return "";
4209
4783
  }
@@ -4235,7 +4809,12 @@ async function elementsTool(config, args) {
4235
4809
  elements: out,
4236
4810
  dropdownOptions: Array.from(document.querySelectorAll(optionSelector)).map((el) => (el.textContent ?? "").trim()).filter(Boolean).slice(0, 50)
4237
4811
  };
4238
- }, [max, DROPDOWN_OPTION_SELECTOR]);
4812
+ }, [
4813
+ max,
4814
+ DROPDOWN_OPTION_SELECTOR,
4815
+ GENERIC.formItemContainersStrict,
4816
+ GENERIC.formLabelSimple
4817
+ ]);
4239
4818
  return {
4240
4819
  count: data.elements.length,
4241
4820
  elements: data.elements,
@@ -4412,6 +4991,8 @@ async function enrichSemanticMiss(e, config, exec) {
4412
4991
  }
4413
4992
  async function clickTool(config, args, exec) {
4414
4993
  const page = await getPage(config);
4994
+ await assertExpectedUrl(page, args.expectUrl, "click");
4995
+ const extWarning = drainExternalNavWarning(page);
4415
4996
  let targetRes;
4416
4997
  try {
4417
4998
  targetRes = await resolveTargetLocator(page, args);
@@ -4514,6 +5095,7 @@ async function clickTool(config, args, exec) {
4514
5095
  clicked: String(args.ref ?? args.selector ?? args.mark ?? args.what ?? (args.row ? `row: ${args.row}` : "target")),
4515
5096
  effect,
4516
5097
  ...quiet ? { effectHint: "点击后未观察到任何 DOM/URL/网络变化,可能未生效(被遮挡/事件被吞/前置条件未满足);建议 browser_snapshot 核实或换目标" } : {},
5098
+ ...extWarning !== void 0 ? { warning: extWarning } : {},
4517
5099
  ...ambiguityFields(targetRes),
4518
5100
  ...targetRes.viaBridge === true ? { viaBridge: true } : {},
4519
5101
  ...await pageIdentity(page)
@@ -4521,6 +5103,7 @@ async function clickTool(config, args, exec) {
4521
5103
  }
4522
5104
  return {
4523
5105
  clicked: String(args.ref ?? args.selector ?? args.mark ?? args.what ?? (args.row ? `row: ${args.row}` : "target")),
5106
+ ...extWarning !== void 0 ? { warning: extWarning } : {},
4524
5107
  ...ambiguityFields(targetRes),
4525
5108
  ...targetRes.viaBridge === true ? { viaBridge: true } : {},
4526
5109
  ...await pageIdentity(page)
@@ -4528,6 +5111,8 @@ async function clickTool(config, args, exec) {
4528
5111
  }
4529
5112
  async function typeTool(config, args) {
4530
5113
  const page = await getPage(config);
5114
+ await assertExpectedUrl(page, args.expectUrl, "type");
5115
+ const extWarning = drainExternalNavWarning(page);
4531
5116
  const targetRes = await resolveTargetLocator(page, args, { preferredRoles: [
4532
5117
  "textbox",
4533
5118
  "searchbox",
@@ -4542,27 +5127,54 @@ async function typeTool(config, args) {
4542
5127
  const selector = String(args.ref ?? args.selector ?? "");
4543
5128
  if (await fillEditorHelper(page, selector, text)) return {
4544
5129
  typed: true,
4545
- editor: true,
5130
+ typedVia: "editor",
4546
5131
  selector,
4547
5132
  ...ambiguityFields(targetRes),
4548
5133
  ...await pageIdentity(page)
4549
5134
  };
4550
5135
  }
5136
+ let typedVia = "clicked";
5137
+ let focusedElement;
4551
5138
  if (args.clear === true) {
4552
5139
  if (isHuman) {
4553
5140
  await target.fill("");
4554
5141
  await humanType(page, text, { speed: config.humanizeSpeed });
4555
5142
  } else await target.fill(text, { timeout: args.timeoutMs ?? config.timeoutMs });
5143
+ typedVia = "fill";
4556
5144
  } else {
4557
5145
  const ambiguous = args.strict === true || (targetRes.matchedCount ?? 0) > 1;
5146
+ let clickOk = false;
4558
5147
  try {
4559
5148
  await target.click({ timeout: Math.min(args.timeoutMs ?? config.timeoutMs, 5e3) });
5149
+ clickOk = true;
4560
5150
  } catch (e) {
4561
5151
  if (ambiguous) throw new Error(`type 定位目标存在歧义(命中 ${targetRes.matchedCount ?? "?"} 个元素)且首个不可交互,已中止;请用 scope 限定或 browser_snapshot 取 ref 消歧;原始错误: ${e instanceof Error ? e.message : String(e)}`);
4562
- await target.click({
4563
- timeout: 2e3,
4564
- force: true
4565
- }).catch(() => {});
5152
+ try {
5153
+ await target.click({
5154
+ timeout: 2e3,
5155
+ force: true
5156
+ });
5157
+ clickOk = true;
5158
+ } catch {}
5159
+ }
5160
+ if (!clickOk) {
5161
+ const probe = await withTimeout(target.evaluate((el) => {
5162
+ const ae = document.activeElement;
5163
+ if (ae === null) return {
5164
+ owns: false,
5165
+ desc: null
5166
+ };
5167
+ return {
5168
+ owns: ae === el || el.contains(ae),
5169
+ desc: ae.outerHTML.slice(0, 200)
5170
+ };
5171
+ }), 2e3, "焦点归属探测").catch(() => ({
5172
+ owns: false,
5173
+ desc: null
5174
+ }));
5175
+ focusedElement = probe.desc ?? void 0;
5176
+ if (!probe.owns) throw new Error(`type 目标不可点击且焦点不在目标上,已中止输入(避免盲打到错误输入框)。当前焦点: ${focusedElement ?? "(无法探测/无焦点)"};目标: ${String(args.ref ?? args.selector ?? args.what ?? "")}。建议 browser_snapshot 核实目标状态,或 browser_wait_for 等元素可交互`);
5177
+ typedVia = "focus-kept";
4566
5178
  }
4567
5179
  if (isHuman) await humanType(page, text, { speed: config.humanizeSpeed });
4568
5180
  else await page.keyboard.type(text, { delay: args.delayMs ?? 0 });
@@ -4570,6 +5182,9 @@ async function typeTool(config, args) {
4570
5182
  if (args.submit === true) await page.keyboard.press("Enter");
4571
5183
  return {
4572
5184
  typed: true,
5185
+ typedVia,
5186
+ ...focusedElement !== void 0 ? { focusedElement } : {},
5187
+ ...extWarning !== void 0 ? { warning: extWarning } : {},
4573
5188
  selector: String(args.ref ?? args.selector),
4574
5189
  submit: args.submit === true,
4575
5190
  ...ambiguityFields(targetRes),
@@ -4580,6 +5195,8 @@ async function pressTool(config, args) {
4580
5195
  const page = await getPage(config);
4581
5196
  const key = String(args.key);
4582
5197
  let warning;
5198
+ const extWarning = drainExternalNavWarning(page);
5199
+ if (extWarning !== void 0) warning = warning !== void 0 ? `${warning};${extWarning}` : extWarning;
4583
5200
  if (key === "Escape" && args.selector === void 0 && args.ref === void 0) {
4584
5201
  const checkState = await page.evaluate(() => {
4585
5202
  return {
@@ -4623,6 +5240,8 @@ async function pressTool(config, args) {
4623
5240
  */
4624
5241
  async function selectTool(config, args) {
4625
5242
  const page = await getPage(config);
5243
+ await assertExpectedUrl(page, args.expectUrl, "select");
5244
+ const extWarning = drainExternalNavWarning(page);
4626
5245
  const targetRes = await resolveTargetLocator(page, args, { preferredRoles: [
4627
5246
  "combobox",
4628
5247
  "select",
@@ -4645,6 +5264,7 @@ async function selectTool(config, args) {
4645
5264
  clear: args.clear === true
4646
5265
  })).selected,
4647
5266
  selector: label,
5267
+ ...extWarning !== void 0 ? { warning: extWarning } : {},
4648
5268
  ...ambiguityFields(targetRes),
4649
5269
  ...await pageIdentity(page)
4650
5270
  };
@@ -4652,8 +5272,8 @@ async function selectTool(config, args) {
4652
5272
  async function evalTool(config, args) {
4653
5273
  const page = await getPage(config);
4654
5274
  const { script, unwrapped } = unwrapDoubleEncodedScript(String(args.script ?? ""));
4655
- const wrapped = /[\n;]|^\s*return\b/.test(script) ? `(async () => {\n${script}\n})()` : `(${script})`;
4656
- const raw = await page.evaluate(wrapped);
5275
+ const { code, mode } = buildEvalScript(script);
5276
+ const raw = await page.evaluate(code);
4657
5277
  let result;
4658
5278
  try {
4659
5279
  result = JSON.parse(JSON.stringify(raw ?? null));
@@ -4662,6 +5282,7 @@ async function evalTool(config, args) {
4662
5282
  }
4663
5283
  return {
4664
5284
  result,
5285
+ mode,
4665
5286
  ...unwrapped ? { scriptUnwrapped: true } : {}
4666
5287
  };
4667
5288
  }
@@ -4862,6 +5483,41 @@ async function formFillTool(config, args) {
4862
5483
  ...submitted !== void 0 ? { submitted } : {}
4863
5484
  };
4864
5485
  }
5486
+ /**
5487
+ * browser_drag:在页面里把源元素拖到目标元素上。
5488
+ * 源/目标各自走统一定位管线(selector / ref / what+scope / mark 四选一前缀),
5489
+ * 支持 HTML5 原生拖放(dragTo)与鼠标拖动类控件(物理鼠标序列降级);
5490
+ * 跨 Frame 拖拽自动绕过 dragTo 的同 Frame 限制。
5491
+ */
5492
+ async function dragTool(config, args) {
5493
+ const pickSide = (side) => {
5494
+ const remapped = {};
5495
+ if (typeof args[side] === "string" && args[side] !== "") remapped.selector = args[side];
5496
+ if (typeof args[`${side}Ref`] === "string" && args[`${side}Ref`] !== "") remapped.ref = args[`${side}Ref`];
5497
+ if (typeof args[`${side}What`] === "string" && args[`${side}What`].trim() !== "") remapped.what = args[`${side}What`];
5498
+ if (typeof args[`${side}Mark`] === "number") remapped.mark = args[`${side}Mark`];
5499
+ if (typeof args[`${side}Scope`] === "string" && args[`${side}Scope`] !== "") remapped.scope = args[`${side}Scope`];
5500
+ if (Object.keys(remapped).filter((k) => k !== "scope").length === 0) throw new Error(`drag 需要 ${side} 定位: ${side}(CSS 选择器)/${side}Ref(aria ref)/${side}What(意图语义)/${side}Mark(SoM 标徽)至少给一个`);
5501
+ return remapped;
5502
+ };
5503
+ const describeSide = (remapped) => remapped.what !== void 0 ? `what:"${remapped.what}"` : remapped.ref !== void 0 ? `ref:${remapped.ref}` : remapped.mark !== void 0 ? `mark:${remapped.mark}` : String(remapped.selector);
5504
+ const srcArgs = pickSide("source");
5505
+ const tgtArgs = pickSide("target");
5506
+ const page = await getPage(config);
5507
+ const srcRes = await resolveTargetLocator(page, srcArgs);
5508
+ const tgtRes = await resolveTargetLocator(page, tgtArgs);
5509
+ const crossFrame = srcRes.container !== tgtRes.container;
5510
+ return {
5511
+ dragged: true,
5512
+ via: await dragAndDropLocators(page, srcRes.locator, tgtRes.locator, {
5513
+ timeoutMs: typeof args.timeoutMs === "number" ? args.timeoutMs : void 0,
5514
+ preferMouse: crossFrame
5515
+ }),
5516
+ source: describeSide(srcArgs),
5517
+ target: describeSide(tgtArgs),
5518
+ ...await pageIdentity(page)
5519
+ };
5520
+ }
4865
5521
  //#endregion
4866
5522
  //#region lib/types/browser/tools/testing.js
4867
5523
  async function waitTool(config, args) {
@@ -5583,57 +6239,49 @@ function previewHighlight(page, selector, label) {
5583
6239
  }).catch(() => {});
5584
6240
  }
5585
6241
  /**
5586
- * 常见认证/登录相关的 URL 特征正则,用于防止跳板页提前误判。
5587
- */
5588
- const AUTH_URL_PATTERN = /login|signin|sign-in|oauth|passport|sso|cas|challenge|captcha|verify|auth\b/i;
5589
- /**
5590
6242
  * 等待人工登录底层执行(独立导出供测试注入 fake page)。
5591
- * 协同机制:
5592
- * 1. 挂起等待期间置入 instruction 至实时画面面板(setPauseInstruction),结束后统一清理;
5593
- * 2. 存活标签感知:感知 context 内所有存活页面,原标签关闭时自动迁移至存活页,新标签出现时通知外层注册;
5594
- * 3. 稳态判定:
5595
- * - 显式 successSelector / successUrl 优先;
5596
- * - 缺省时,检测 URL 离开登录页(非 AUTH_URL_PATTERN)、登录密码框消除、或 SPA 内容签名稳态变化,防范跳板页误判与 SPA 弹窗漏判。
6243
+ * 完成判定(按优先级):
6244
+ * 1. **用户主动告知(主通道)**:pendingLoginWait 单槽状态机,面板常驻登录等待卡
6245
+ * 「我已完成登录」按钮 → resolveLoginWait(id),500ms 确认节拍;
6246
+ * 2. 显式确定性条件:successSelector(任一存活标签可见)/ successUrl(URL 正则,
6247
+ * 任一存活标签命中),pollIntervalMs 节拍多标签扫描,命中即把目标页切为命中页;
6248
+ * 3. 超时返回 loggedIn:false(附 waitedMs 与 detail,不抛错)。
6249
+ *
6250
+ * 登录状态不再自动猜测:URL 认证特征/密码框消除/SPA 签名启发式对「登录表单
6251
+ * 隐藏不卸载(keep-alive/微前端)」「登录后 URL 仍含 auth 子串」的 SPA 会永久
6252
+ * 漏判(用户已登录仍空转到超时),已整体移除——登录完成由用户说了算。
6253
+ * 保留:存活标签感知(原标签关闭自愈迁移、新标签经 onNewPageDetected 注册)。
5597
6254
  */
5598
6255
  async function runWaitForLoginOnPage(initialPage, args, onNewPageDetected) {
5599
6256
  const instruction = typeof args.instruction === "string" && args.instruction.trim() ? args.instruction.trim() : "请在浏览器窗口中完成登录";
5600
6257
  const timeoutMs = Math.min(Math.max(Number(args.timeoutMs ?? 12e4) || 12e4, 1e3), 6e5);
5601
6258
  const pollIntervalMs = Math.min(Math.max(Number(args.pollIntervalMs ?? 2e3) || 2e3, 100), 5e3);
6259
+ const successSelector = typeof args.successSelector === "string" ? args.successSelector : "";
6260
+ const successUrl = typeof args.successUrl === "string" ? args.successUrl : "";
5602
6261
  let currentPage = initialPage;
5603
- let startUrl = "";
5604
- try {
5605
- startUrl = initialPage.url();
5606
- } catch {}
5607
- const checkHasPassword = async (p) => {
5608
- try {
5609
- return await p.evaluate(() => !!document.querySelector("input[type=\"password\"]"));
5610
- } catch {
5611
- return false;
5612
- }
5613
- };
5614
- const getPageSig = async (p) => {
5615
- try {
5616
- return await p.evaluate(() => {
5617
- const body = document.body;
5618
- return `${location.href}|${body ? body.innerText.length : 0}|${document.querySelectorAll("input,button,a").length}`;
5619
- });
5620
- } catch {
5621
- try {
5622
- return p.url();
5623
- } catch {
5624
- return "";
5625
- }
5626
- }
5627
- };
5628
- const startHasPassword = await checkHasPassword(initialPage);
5629
- const startSig = args.successSelector || args.successUrl ? "" : await getPageSig(initialPage);
6262
+ const id = `login-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`;
6263
+ setPendingLoginWait({
6264
+ id,
6265
+ instruction,
6266
+ startedAt: Date.now()
6267
+ });
5630
6268
  const startedAt = Date.now();
5631
6269
  let loggedIn = false;
5632
- let detail = `等待登录超时(${timeoutMs}ms)`;
5633
- setPauseInstruction(instruction);
6270
+ let detail = `等待登录超时(${timeoutMs}ms),用户未在面板确认登录完成`;
6271
+ let lastExplicitCheck = 0;
5634
6272
  try {
5635
6273
  const deadline = startedAt + timeoutMs;
5636
6274
  while (Date.now() < deadline) {
6275
+ const pending = getPendingLoginWait();
6276
+ if (pending !== null && pending.id !== id) {
6277
+ detail = "登录等待已被新的等待请求取代";
6278
+ break;
6279
+ }
6280
+ if (pending !== null && pending.done) {
6281
+ loggedIn = true;
6282
+ detail = "用户已在面板确认完成登录";
6283
+ break;
6284
+ }
5637
6285
  let candidatePages = [];
5638
6286
  try {
5639
6287
  const context = typeof currentPage.context === "function" ? currentPage.context() : null;
@@ -5658,77 +6306,35 @@ async function runWaitForLoginOnPage(initialPage, args, onNewPageDetected) {
5658
6306
  isCurrentClosed = false;
5659
6307
  }
5660
6308
  }
5661
- try {
5662
- if (args.successSelector) {
5663
- const targets = candidatePages.length > 0 ? candidatePages : isCurrentClosed ? [] : [currentPage];
5664
- for (const p of targets) try {
5665
- if (await p.locator(String(args.successSelector)).first().isVisible()) {
5666
- loggedIn = true;
5667
- currentPage = p;
5668
- detail = `命中 successSelector: ${args.successSelector}`;
5669
- break;
5670
- }
5671
- } catch {}
5672
- } else if (args.successUrl) {
5673
- const re = new RegExp(String(args.successUrl));
5674
- const targets = candidatePages.length > 0 ? candidatePages : isCurrentClosed ? [] : [currentPage];
5675
- for (const p of targets) try {
5676
- const u = p.url();
5677
- if (re.test(u)) {
6309
+ if ((successSelector !== "" || successUrl !== "") && Date.now() - lastExplicitCheck >= pollIntervalMs) {
6310
+ lastExplicitCheck = Date.now();
6311
+ const targets = candidatePages.length > 0 ? candidatePages : isCurrentClosed ? [] : [currentPage];
6312
+ for (const p of targets) try {
6313
+ if (successSelector !== "") {
6314
+ if (await p.locator(successSelector).first().isVisible()) {
5678
6315
  loggedIn = true;
5679
6316
  currentPage = p;
5680
- detail = `命中 successUrl: ${args.successUrl}`;
6317
+ detail = `命中 successSelector: ${successSelector}`;
5681
6318
  break;
5682
6319
  }
5683
- } catch {}
5684
- } else {
5685
- const target = !isCurrentClosed ? currentPage : candidatePages[0];
5686
- if (target) {
5687
- let curUrl = "";
5688
- try {
5689
- curUrl = target.url();
5690
- } catch {}
5691
- if (curUrl && curUrl !== "about:blank") {
5692
- if (curUrl !== startUrl) {
5693
- if (!AUTH_URL_PATTERN.test(curUrl)) {
5694
- loggedIn = true;
5695
- currentPage = target;
5696
- detail = `URL 离开登录页跳转至 ${curUrl}`;
5697
- } else {
5698
- const hasPwd = await checkHasPassword(target);
5699
- if (startHasPassword && !hasPwd) {
5700
- loggedIn = true;
5701
- currentPage = target;
5702
- detail = `登录密码输入框已消除,页面跳转至 ${curUrl}`;
5703
- }
5704
- }
5705
- } else {
5706
- const hasPwd = await checkHasPassword(target);
5707
- if (startHasPassword && !hasPwd) {
5708
- loggedIn = true;
5709
- currentPage = target;
5710
- detail = "登录密码输入框已消除(SPA 登录完成)";
5711
- } else if (await getPageSig(target) !== startSig && (!startHasPassword || !hasPwd)) {
5712
- loggedIn = true;
5713
- currentPage = target;
5714
- detail = "页面内容签名变更(SPA 登录完成)";
5715
- }
5716
- }
5717
- }
6320
+ } else if (new RegExp(successUrl).test(p.url())) {
6321
+ loggedIn = true;
6322
+ currentPage = p;
6323
+ detail = `命中 successUrl: ${successUrl}`;
6324
+ break;
5718
6325
  }
5719
- }
5720
- } catch (e) {
5721
- detail = `检查登录状态出错: ${e instanceof Error ? e.message : String(e)}`;
6326
+ } catch {}
5722
6327
  }
5723
6328
  if (loggedIn) break;
5724
6329
  try {
5725
- await currentPage.waitForTimeout(pollIntervalMs);
6330
+ await currentPage.waitForTimeout(500);
5726
6331
  } catch {
5727
- await sleep$1(pollIntervalMs);
6332
+ await sleep$1(500);
5728
6333
  }
5729
6334
  }
5730
6335
  } finally {
5731
- clearPauseInstruction();
6336
+ const pending = getPendingLoginWait();
6337
+ if (pending !== null && pending.id === id) clearPendingLoginWait();
5732
6338
  }
5733
6339
  return {
5734
6340
  loggedIn,
@@ -5739,11 +6345,15 @@ async function runWaitForLoginOnPage(initialPage, args, onNewPageDetected) {
5739
6345
  }
5740
6346
  /**
5741
6347
  * 等待人工登录完成(配合 headless:false 使用):agent 打开登录页后调用本工具,
5742
- * 提示用户在浏览器窗口中完成登录(输入/扫码/双因素)。支持 instruction 提示语
5743
- * 与右下角实时画面面板联动显示横幅,支持多标签感知与 SPA 签名自适应判定。
5744
- * 超时返回 loggedIn:false(附 waitedMs 与 detail,不抛错)。
6348
+ * 用户在浏览器窗口完成登录(输入/扫码/双因素)后,在实时画面面板的**常驻登录
6349
+ * 等待卡**上点「我已完成登录」主动告知,工具立即返回(500ms 确认节拍)。
6350
+ * 亦支持显式确定性条件 successSelector/successUrl(多标签扫描自动判定)。
6351
+ * 不做任何登录状态启发式猜测。超时返回 loggedIn:false(附 waitedMs 与
6352
+ * detail,不抛错)。无面板环境(MCP 独立模式)必须显式给条件,否则 fail-fast。
5745
6353
  */
5746
6354
  async function waitForLoginTool(config, args) {
6355
+ const hasExplicit = typeof args.successSelector === "string" && args.successSelector !== "" || typeof args.successUrl === "string" && args.successUrl !== "";
6356
+ if (!isInteractivePanelPresent() && !hasExplicit) throw new Error("等待人工登录失败: 当前环境无交互面板,用户无法点「我已完成登录」确认。请显式提供 successSelector(登录成功后出现的元素,如 \"#avatar\")或 successUrl(登录成功后 URL 应匹配的正则)后重试");
5747
6357
  const page = await getPage(config);
5748
6358
  const state = stateOf();
5749
6359
  const result = await runWaitForLoginOnPage(page, args, (newPage) => {
@@ -7748,7 +8358,11 @@ function getInteractionTools(config, timeout, stepImpls) {
7748
8358
  type: "boolean",
7749
8359
  description: "多命中即抛错(对齐 Playwright strict),要求定位唯一;默认 false 操作第 1 个并在返回中提示命中数。"
7750
8360
  },
7751
- timeoutMs: { type: "integer" }
8361
+ timeoutMs: { type: "integer" },
8362
+ expectUrl: {
8363
+ type: "string",
8364
+ description: "页面预期校验(opt-in):正则(非法时按子串);定位前当前 URL 不匹配直接报错,把「页面被切换/回退」从诡异失败变为一句话可诊断的错误。"
8365
+ }
7752
8366
  },
7753
8367
  additionalProperties: false
7754
8368
  },
@@ -7776,6 +8390,10 @@ function getInteractionTools(config, timeout, stepImpls) {
7776
8390
  viaBridge: {
7777
8391
  type: "boolean",
7778
8392
  description: "what 未命中时经页面 __DSH_BRIDGE__(vite-plugin-dsh-bridge)兜底命中。"
8393
+ },
8394
+ warning: {
8395
+ type: "string",
8396
+ description: "执行过程中检测到的环境异常提示(如非工具触发的页面导航)。"
7779
8397
  }
7780
8398
  },
7781
8399
  required: [
@@ -7790,7 +8408,7 @@ function getInteractionTools(config, timeout, stepImpls) {
7790
8408
  }),
7791
8409
  definePlainTool({
7792
8410
  name: "browser_type",
7793
- description: "向输入框输入文本:selector(CSS)、ref 或 what(意图描述,如 '用户名'/'搜索')三选一定位目标(同时传多个会报参数错误;容器内定位用 what + scope)。clear=true 时先清空再输入(推荐用于表单);submit=true 时输入后按回车;editor=true 时适配 Monaco/CodeMirror/富文本编辑器。",
8411
+ description: "向输入框输入文本:selector(CSS)、ref 或 what(意图描述,如 '用户名'/'搜索')三选一定位目标(同时传多个会报参数错误;容器内定位用 what + scope)。clear=true 时先清空再输入(推荐用于表单);submit=true 时输入后按回车;editor=true 时适配 Monaco/CodeMirror/富文本编辑器。定位与点击均失败时做焦点归属校验:焦点不在目标内(含子元素)则中止输入并附当前焦点线索,不会盲打到错误输入框;返回 typedVia 标明实际输入通道。expectUrl 传入时先校验当前页面 URL,不符直接报错(防页面被切换后打错位置)。",
7794
8412
  parameters: {
7795
8413
  type: "object",
7796
8414
  properties: {
@@ -7826,7 +8444,11 @@ function getInteractionTools(config, timeout, stepImpls) {
7826
8444
  description: "多命中即抛错(对齐 Playwright strict);多命中或 strict 时不再 force 降级点击,遮罩拦截会直接暴露。"
7827
8445
  },
7828
8446
  delayMs: { type: "integer" },
7829
- timeoutMs: { type: "integer" }
8447
+ timeoutMs: { type: "integer" },
8448
+ expectUrl: {
8449
+ type: "string",
8450
+ description: "页面预期校验(opt-in):正则(非法时按子串);定位前当前 URL 不匹配直接报错,把「页面被切换/回退」从诡异失败变为一句话可诊断的错误。"
8451
+ }
7830
8452
  },
7831
8453
  required: ["text"],
7832
8454
  additionalProperties: false
@@ -7836,6 +8458,20 @@ function getInteractionTools(config, timeout, stepImpls) {
7836
8458
  additionalProperties: false,
7837
8459
  properties: {
7838
8460
  typed: { type: "boolean" },
8461
+ typedVia: {
8462
+ type: "string",
8463
+ enum: [
8464
+ "fill",
8465
+ "clicked",
8466
+ "focus-kept",
8467
+ "editor"
8468
+ ],
8469
+ description: "实际输入通道:fill=clear 清空重填;clicked=点击获得焦点;focus-kept=点击超时但焦点已在目标内续打;editor=编辑器适配。"
8470
+ },
8471
+ focusedElement: {
8472
+ type: "string",
8473
+ description: "focus-kept 通道下当前焦点元素 outerHTML 片段(调试线索)。"
8474
+ },
7839
8475
  selector: { type: "string" },
7840
8476
  submit: { type: "boolean" },
7841
8477
  url: { type: "string" },
@@ -7851,6 +8487,10 @@ function getInteractionTools(config, timeout, stepImpls) {
7851
8487
  hint: {
7852
8488
  type: "string",
7853
8489
  description: "matchedCount>1 时的消歧指引。"
8490
+ },
8491
+ warning: {
8492
+ type: "string",
8493
+ description: "执行过程中检测到的环境异常提示(如非工具触发的页面导航)。"
7854
8494
  }
7855
8495
  },
7856
8496
  required: [
@@ -7965,7 +8605,11 @@ function getInteractionTools(config, timeout, stepImpls) {
7965
8605
  type: "boolean",
7966
8606
  description: "多命中即抛错(对齐 Playwright strict)。"
7967
8607
  },
7968
- timeoutMs: { type: "integer" }
8608
+ timeoutMs: { type: "integer" },
8609
+ expectUrl: {
8610
+ type: "string",
8611
+ description: "页面预期校验(opt-in):正则(非法时按子串);定位前当前 URL 不匹配直接报错,把「页面被切换/回退」从诡异失败变为一句话可诊断的错误。"
8612
+ }
7969
8613
  },
7970
8614
  required: ["option"],
7971
8615
  additionalProperties: false
@@ -7989,6 +8633,10 @@ function getInteractionTools(config, timeout, stepImpls) {
7989
8633
  hint: {
7990
8634
  type: "string",
7991
8635
  description: "matchedCount>1 时的消歧指引。"
8636
+ },
8637
+ warning: {
8638
+ type: "string",
8639
+ description: "执行过程中检测到的环境异常提示(如非工具触发的页面导航)。"
7992
8640
  }
7993
8641
  },
7994
8642
  required: [
@@ -8002,9 +8650,95 @@ function getInteractionTools(config, timeout, stepImpls) {
8002
8650
  timeoutMs: timeout(),
8003
8651
  execute: (args) => selectTool(config, args)
8004
8652
  }),
8653
+ definePlainTool({
8654
+ name: "browser_drag",
8655
+ description: "把源元素拖到目标元素上(拖拽/拖放):源与目标各自定位——source(CSS)/sourceRef(aria ref)/sourceWhat(意图语义,+sourceScope)/sourceMark(SoM 标徽)四选一,target/targetRef/targetWhat(+targetScope)/targetMark 同理。优先 Playwright dragTo(覆盖 HTML5 原生拖放),失败或跨 Frame 时自动降级为物理鼠标序列(覆盖看板卡片、滑块等鼠标拖动类控件),返回 via 标明实际通道。注意:dragged=true 表示拖拽序列已执行,不保证业务侧生效(drop 区拒绝、控件无响应仍为 true),需要验证请配合 browser_assert。源/目标不可见或零命中时抛错并附诊断。",
8656
+ parameters: {
8657
+ type: "object",
8658
+ properties: {
8659
+ source: {
8660
+ type: "string",
8661
+ description: "源元素 CSS 选择器;与 sourceRef/sourceWhat/sourceMark 互斥。"
8662
+ },
8663
+ sourceRef: {
8664
+ type: "string",
8665
+ description: "源元素 aria 快照 ref(如 'e3');与 source/sourceWhat/sourceMark 互斥。"
8666
+ },
8667
+ sourceWhat: {
8668
+ type: "string",
8669
+ description: "源元素意图描述(如 '任务卡片 A'):页内打分语义定位;与 source/sourceRef/sourceMark 互斥;配合 sourceScope。"
8670
+ },
8671
+ sourceMark: {
8672
+ type: "integer",
8673
+ description: "源元素 SoM 标徽编号(browser_screenshot { som: true });与 source/sourceRef/sourceWhat 互斥。"
8674
+ },
8675
+ sourceScope: {
8676
+ type: "string",
8677
+ description: "配合 sourceWhat 限定扫描范围的容器选择器。"
8678
+ },
8679
+ target: {
8680
+ type: "string",
8681
+ description: "目标元素 CSS 选择器;与 targetRef/targetWhat/targetMark 互斥。"
8682
+ },
8683
+ targetRef: {
8684
+ type: "string",
8685
+ description: "目标元素 aria 快照 ref;与 target/targetWhat/targetMark 互斥。"
8686
+ },
8687
+ targetWhat: {
8688
+ type: "string",
8689
+ description: "目标元素意图描述(如 '待办列'):页内打分语义定位;与 target/targetRef/targetMark 互斥;配合 targetScope。"
8690
+ },
8691
+ targetMark: {
8692
+ type: "integer",
8693
+ description: "目标元素 SoM 标徽编号;与 target/targetRef/targetWhat 互斥。"
8694
+ },
8695
+ targetScope: {
8696
+ type: "string",
8697
+ description: "配合 targetWhat 限定扫描范围的容器选择器。"
8698
+ },
8699
+ timeoutMs: { type: "integer" }
8700
+ },
8701
+ additionalProperties: false
8702
+ },
8703
+ outputSchema: {
8704
+ type: "object",
8705
+ additionalProperties: false,
8706
+ properties: {
8707
+ dragged: {
8708
+ type: "boolean",
8709
+ description: "拖拽序列已执行(true 不保证业务侧生效,drop 被拒绝也为 true)。"
8710
+ },
8711
+ via: {
8712
+ type: "string",
8713
+ description: "实际执行通道:'dragTo'(HTML5/组件库)或 'mouse'(物理鼠标,含跨 Frame 与降级)。"
8714
+ },
8715
+ source: {
8716
+ type: "string",
8717
+ description: "实际使用的源定位描述(what:/ref:/mark:/selector)。"
8718
+ },
8719
+ target: {
8720
+ type: "string",
8721
+ description: "实际使用的目标定位描述。"
8722
+ },
8723
+ url: { type: "string" },
8724
+ title: { type: "string" }
8725
+ },
8726
+ required: [
8727
+ "dragged",
8728
+ "via",
8729
+ "source",
8730
+ "target",
8731
+ "url",
8732
+ "title"
8733
+ ]
8734
+ },
8735
+ render: (_args, value) => `Dragged ${value.source} → ${value.target} (via ${value.via})`,
8736
+ timeoutMs: timeout(),
8737
+ execute: (args) => dragTool(config, args)
8738
+ }),
8005
8739
  definePlainTool({
8006
8740
  name: "browser_eval",
8007
- description: "在页面上下文执行一段 JavaScript。表达式直接求值(如 'document.title');含换行/分号的语句会被包进 async IIFE,可用 return 返回。结果必须是可 JSON 序列化的值。",
8741
+ description: "在页面上下文执行一段 JavaScript。表达式(含多行对象字面量/箭头函数 IIFE/顶层 await)直接返回值;语句脚本自动 return 最后一个顶层表达式的值,也可显式 return。结果必须是可 JSON 序列化的值,返回 mode 标明实际执行形态。",
8008
8742
  parameters: {
8009
8743
  type: "object",
8010
8744
  properties: { script: {
@@ -8019,6 +8753,15 @@ function getInteractionTools(config, timeout, stepImpls) {
8019
8753
  additionalProperties: false,
8020
8754
  properties: {
8021
8755
  result: {},
8756
+ mode: {
8757
+ type: "string",
8758
+ enum: [
8759
+ "expression",
8760
+ "statements-auto-return",
8761
+ "statements"
8762
+ ],
8763
+ description: "实际执行形态:expression=表达式直取值(支持顶层 await 与多行对象字面量);statements-auto-return=语句体自动 return 最后一个顶层表达式;statements=语句体(显式 return 或纯副作用,无返回值时 result 为 null)。"
8764
+ },
8022
8765
  scriptUnwrapped: {
8023
8766
  type: "boolean",
8024
8767
  description: "入参 script 疑似被 JSON 双重序列化(以引号包裹且含 \\n 字面量),已自动解码一层。"
@@ -8819,13 +9562,13 @@ function getCollaborationTools(config, timeout, stepImpls) {
8819
9562
  return [
8820
9563
  definePlainTool({
8821
9564
  name: "browser_wait_for_login",
8822
- description: "等待人工登录完成。agent 在登录页调用本工具后,提示用户在浏览器窗口中完成登录(有头模式 headless:false 时用户可直接操作;配 userDataDir 登录态会持久化)。支持 instruction 提示语(将显示在右下角实时画面面板);成功条件:优先取 successSelector(登录后出现的元素,如 '#avatar')或 successUrl(登录后 URL 正则);两者未指定时,结合 URL 变化、登录特征消除(密码框等)与页面签名稳态综合判定。轮询间隔 2 秒,超时返回 loggedIn:false(附 waitedMs 与 detail)。",
9565
+ description: "等待人工登录完成。agent 在登录页调用本工具后,用户在浏览器窗口完成登录(有头模式 headless:false;配 userDataDir 登录态会持久化),然后在实时画面面板的常驻登录等待卡上点「我已完成登录」主动告知,工具立即返回(500ms 确认节拍)——不猜测登录状态,完成与否由用户说了算。亦可显式给 successSelector(登录后出现的元素,如 '#avatar')或 successUrl(登录后 URL 正则)作为确定性自动条件(多标签扫描,轮询间隔 2 秒)。超时返回 loggedIn:false(附 waitedMs 与 detail,不抛错)。注意:无面板环境(独立 MCP 模式)用户无法点按钮确认,必须显式给 successSelector 或 successUrl,否则调用直接报错。",
8823
9566
  parameters: {
8824
9567
  type: "object",
8825
9568
  properties: {
8826
9569
  instruction: {
8827
9570
  type: "string",
8828
- description: "可选:给用户看的提示说明(默认「请在浏览器窗口中完成登录」)。"
9571
+ description: "可选:给用户看的提示说明,展示在面板登录等待卡上(默认「请在浏览器窗口中完成登录」)。"
8829
9572
  },
8830
9573
  timeoutMs: {
8831
9574
  type: "integer",
@@ -8833,11 +9576,11 @@ function getCollaborationTools(config, timeout, stepImpls) {
8833
9576
  },
8834
9577
  successSelector: {
8835
9578
  type: "string",
8836
- description: "登录成功后应出现的 CSS 选择器。"
9579
+ description: "可选:登录成功后应出现的 CSS 选择器(确定性自动条件,任一标签可见即完成)。"
8837
9580
  },
8838
9581
  successUrl: {
8839
9582
  type: "string",
8840
- description: "登录成功后 URL 应匹配的正则。"
9583
+ description: "可选:登录成功后 URL 应匹配的正则(确定性自动条件)。"
8841
9584
  }
8842
9585
  },
8843
9586
  additionalProperties: false
@@ -9497,14 +10240,16 @@ function registerTools(ctx, config) {
9497
10240
  "browser_upload",
9498
10241
  "browser_form",
9499
10242
  "browser_form_fill",
9500
- "browser_eval"
10243
+ "browser_eval",
10244
+ "browser_drag"
9501
10245
  ]);
9502
10246
  /** 操作预告:这些工具执行前把目标元素在实时画面上高亮(异步不阻塞)。 */
9503
10247
  const PREVIEW_TOOLS = /* @__PURE__ */ new Set([
9504
10248
  "browser_click",
9505
10249
  "browser_type",
9506
10250
  "browser_select",
9507
- "browser_press"
10251
+ "browser_press",
10252
+ "browser_drag"
9508
10253
  ]);
9509
10254
  /**
9510
10255
  * auto 审批闸(confirmMode:"auto" 硬约束):live 调用在 register 包装层拦,
@@ -9559,8 +10304,8 @@ function registerTools(ctx, config) {
9559
10304
  sessionTracer.startStep(currentProfile, def.name, args, activePage());
9560
10305
  if (PREVIEW_TOOLS.has(def.name)) {
9561
10306
  const page = activePage();
9562
- const selector = typeof args?.selector === "string" && args.selector ? args.selector : typeof args?.ref === "string" && args.ref && page !== null ? pageRef(page, args.ref)?.selector ?? "" : "";
9563
- if (selector !== "") previewHighlight(page, selector, `${def.name.replace(/^browser_/, "")} ${args?.ref ? `[${args.ref}] ` : ""}${args?.key ?? args?.value ?? args?.text ?? ""}`.trim());
10307
+ const selector = typeof args?.selector === "string" && args.selector ? args.selector : def.name === "browser_drag" && typeof args?.source === "string" && args.source ? args.source : typeof args?.ref === "string" && args.ref && page !== null ? pageRef(page, args.ref)?.selector ?? "" : "";
10308
+ if (selector !== "") previewHighlight(page, selector, `${def.name.replace(/^browser_/, "")} ${args?.ref ? `[${args.ref}] ` : ""}${args?.key ?? args?.value ?? args?.text ?? args?.source ?? ""}`.trim());
9564
10309
  }
9565
10310
  try {
9566
10311
  const result = await original(args, exec);
@@ -9605,6 +10350,7 @@ function registerTools(ctx, config) {
9605
10350
  browser_type: (a) => typeTool(config, a),
9606
10351
  browser_press: (a) => pressTool(config, a),
9607
10352
  browser_select: (a) => selectTool(config, a),
10353
+ browser_drag: (a) => dragTool(config, a),
9608
10354
  browser_eval: (a) => evalTool(config, a),
9609
10355
  browser_screenshot: (a) => screenshotTool(config, a),
9610
10356
  browser_wait: (a) => waitTool(config, a),
@@ -9812,6 +10558,24 @@ function registerBrowserApi(ctx, config, base) {
9812
10558
  sendJson(res, 200, { confirm: getPendingConfirm() });
9813
10559
  return;
9814
10560
  }
10561
+ if (path === `${base}/login`) {
10562
+ if ((req.method ?? "GET") === "POST") {
10563
+ const body = await readJsonBody(req).catch(() => void 0);
10564
+ const id = typeof body?.id === "string" ? body.id : "";
10565
+ if (!id) {
10566
+ sendText(res, 400, "id required");
10567
+ return;
10568
+ }
10569
+ const resolved = resolveLoginWait(id);
10570
+ sendJson(res, resolved ? 200 : 409, {
10571
+ resolved,
10572
+ id
10573
+ });
10574
+ return;
10575
+ }
10576
+ sendJson(res, 200, { loginWait: getPendingLoginWait() });
10577
+ return;
10578
+ }
9815
10579
  if (path === `${base}/pick`) {
9816
10580
  if ((req.method ?? "GET") === "POST") {
9817
10581
  const body = await readJsonBody(req).catch(() => void 0);