Files
metona-ai-desktop/electron/harness/tools/built-in/browser.ts
T
thzxx 80cf5b482c
CI / 类型检查 + Lint + 单元测试 (push) Failing after 5m41s
CI / 全量测试 (Electron ABI) (push) Failing after 5m21s
CI / 产物编译验证 (push) Successful in 10m1s
fix: v0.6.1 修复回复/推理期间偶发崩溃 — 浏览器回退窗口竞态 + 崩溃可观测性
【根因(实证归因,非猜测)】
分析 userData/logs/main.log 全部 33 次启动会话,定位 3 处异常终止点
(07-25 ×2 / 08-22 ×1,启动标记前无 Database closed)。三处 100% 共享
同一模式:web_search 并行抓取 → 多个 web_fetch 同时进入浏览器回退 →
共享单例 BrowserWindowManager 中后到 open() 销毁前一个正在加载/执行
JS 的窗口。关键统计:56 次浏览器回退中 ERR_ABORTED(并发互毁的直接
证据)仅 3 次,而这 3 次恰好全部对应 3 个崩溃点;无并发销毁的 53 次
回退从未崩溃 —— 触发条件完全收敛。

缺陷链(三层叠加):
1. browserFetch 直接 open/evaluate 共享单例,无跨调用序列化 — 并发
   回退互相销毁窗口(ERR_ABORTED / "Object has been destroyed")
2. destroy() 对仍在使用中的 partition fire-and-forget
   clearStorageData/clearCache,与紧随其后的新窗口创建并发 —
   原生存储层竞态(崩溃引爆点)
3. ensureReady 检查与实际 executeJavaScript/loadURL 之间存在竞态窗口;
   loadURLWithTimeout 的 Race 落败方 rejection 无人处理

【修复(browser-window-manager.ts + web-fetch.ts + browser.ts)】
- 新增 fetchPageText:排队版页面抓取,串行化完整 open→等待→evaluate
  序列(与 open 共用单一操作链,destroy 只会在链上发生,跨链互毁彻底
  消除);web_fetch 浏览器回退改走此入口
- open() 拆分 openInternal(链内直调);open 与 fetchPageText 共用
  单一串行链,排队不分死锁
- destroy() 移除 session 存储清理(终态清理迁移至 close(),await 执行,
  不再与窗口创建并发)
- safeWebContents() 即时校验替代 racy 的 ensureReady;evaluate/extract/
  screenshot/click/type/scroll/waitForSelector 全部加固,消除对已销毁
  webContents 的调用
- loadURLWithTimeout 落败方 rejection 兜底(防 unhandledRejection)
- cleanupBrowser/cleanup/close 异步化适配(main.ts 退出链路 await)

【崩溃可观测性(此前崩溃无迹可查 — 日志无声截断)】
- process.on(uncaughtException/unhandledRejection) → [FATAL] 落盘
- app.on(render-process-gone/child-process-gone) → [FATAL] 落盘
- WindowManager: 每窗口 render-process-gone 日志 + 自动 reload 自愈
  (渲染进程 OOM/崩溃不再白屏卡死,可自动恢复)

【验证】
- lint 0/0;typecheck 双工程 0 错误;test:electron 252/252;build 通过
2026-08-22 19:16:41 +08:00

286 lines
9.9 KiB
TypeScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
/**
* web_browser — 统一浏览器交互工具
*
* 将 9 个浏览器操作合并为单个工具,通过 `action` 参数路由:
* 1. open — 打开 URL
* 2. screenshot — 截图(视口/元素/全页三模式)
* 3. evaluate — 执行 JavaScript
* 4. extract — 提取页面文本与链接
* 5. click — 点击元素
* 6. type — 输入文本
* 7. scroll — 滚动页面
* 8. wait — 等待条件
* 9. close — 关闭窗口
*
* 基于单例 BrowserWindowManager 实现。
*
* @see docs/Agent网络工具通用设计-v2.md — 第 4 章 browser 浏览器设计
*/
import type { IMetonaTool, ToolExecutionContext } from '../../types/metona-tool';
import type { MetonaToolDef } from '../../../harness/types';
import { MetonaToolCategory, MetonaRiskLevel } from '../../../harness/types';
import { BrowserWindowManager } from './browser-window-manager';
import { logTool } from './network-utils';
// ===== 单例 Manager =====
let managerInstance: BrowserWindowManager | null = null;
function getManager(): BrowserWindowManager {
if (!managerInstance) {
managerInstance = new BrowserWindowManager();
}
return managerInstance;
}
/** 获取共享浏览器管理器单例(供 web_fetch 等工具复用) */
export function getBrowserManager(): BrowserWindowManager {
return getManager();
}
/** 应用退出时清理(由 main.ts 调用;close 含 session 存储终态清理,需 await */
export async function cleanupBrowser(): Promise<void> {
await BrowserWindowManager.cleanup(managerInstance);
managerInstance = null;
}
// ===== WebBrowserTool =====
export class WebBrowserTool implements IMetonaTool {
readonly definition: MetonaToolDef = {
name: 'web_browser',
description: [
'Control a hidden browser window to navigate, inspect, and interact with web pages.',
'Use the "action" parameter to specify the operation:',
' open — Load a URL; optionally wait for a CSS selector before returning. Returns page title and URL.',
' screenshot — Capture the viewport (default), full page, or a specific element by CSS selector. Returns base64 PNG.',
' evaluate — Execute JavaScript in the page context. Returns the result of the last expression.',
' extract — Extract clean text and up to 50 links from the page or a specific element.',
' click — Click an element by CSS selector. Scrolls into view first. Can optionally wait for the selector.',
' type — Type text into an input element. Optionally clear first and submit the form.',
' scroll — Scroll the page: down/up (500px), top/bottom, or scroll to an element by selector.',
' wait — Wait for a CSS selector to appear (with timeout) or a fixed duration.',
' close — Close the browser window and release resources.',
].join('\n'),
parameters: {
type: 'object',
properties: {
action: {
type: 'string',
enum: [
'open',
'screenshot',
'evaluate',
'extract',
'click',
'type',
'scroll',
'wait',
'close',
],
description: 'Browser operation to perform',
},
url: {
type: 'string',
description: '[open] Target URL (http/https only)',
},
wait_selector: {
type: 'string',
description: '[open] CSS selector to wait for before returning (optional)',
},
script: {
type: 'string',
description: '[evaluate] JavaScript code to execute (must return a value)',
},
selector: {
type: 'string',
description: '[screenshot|extract|click|type|scroll] CSS selector of the target element',
},
full_page: {
type: 'boolean',
description: '[screenshot] Capture entire scrollable page (default false)',
},
text: {
type: 'string',
description: '[type] Text to type into the element',
},
clear: {
type: 'boolean',
description: '[type] Clear the field before typing (default true)',
},
submit: {
type: 'boolean',
description: '[type] Submit the form after typing (default false)',
},
direction: {
type: 'string',
enum: ['down', 'up', 'top', 'bottom'],
description: '[scroll] Scroll direction (default down). Ignored if selector is provided.',
},
wait: {
type: 'boolean',
description: '[click] Wait for selector to appear before clicking (default false)',
},
time_ms: {
type: 'number',
description:
'[wait] Fixed wait time in milliseconds (default 1000). Used as timeout when selector is also provided.',
},
},
required: ['action'],
},
category: MetonaToolCategory.NETWORK,
riskLevel: MetonaRiskLevel.HIGH,
requiresPermission: true,
timeoutMs: 60_000,
};
async execute(args: Record<string, unknown>, _context: ToolExecutionContext): Promise<unknown> {
const action = args.action as string;
if (!action) {
return { success: false, error: 'Missing required parameter: action' };
}
logTool('web_browser', `action=${action}`);
switch (action) {
// ===== open =====
case 'open': {
const url = args.url as string;
if (!url || !/^https?:\/\//i.test(url)) {
return { success: false, error: 'URL must start with http:// or https://' };
}
const waitSelector = args.wait_selector as string | undefined;
try {
const result = await getManager().open({ url, waitSelector });
return { success: true, action, ...result };
} catch (err) {
return { success: false, action, error: (err as Error).message };
}
}
// ===== screenshot =====
case 'screenshot': {
const fullPage = (args.full_page as boolean) ?? false;
const selector = args.selector as string | undefined;
try {
const result = await getManager().screenshot({ fullPage, selector });
return {
success: true,
action,
image: result.data,
width: result.width,
height: result.height,
mime_type: 'image/png',
};
} catch (err) {
return { success: false, action, error: (err as Error).message };
}
}
// ===== evaluate =====
case 'evaluate': {
const script = args.script as string;
if (!script) {
return { success: false, action, error: 'No script provided' };
}
try {
const result = await getManager().evaluate(script);
return { success: true, action, result };
} catch (err) {
return { success: false, action, error: (err as Error).message };
}
}
// ===== extract =====
case 'extract': {
const selector = args.selector as string | undefined;
try {
const result = await getManager().extract(selector);
return {
success: true,
action,
text: result.text,
links: result.links,
link_count: result.links.length,
};
} catch (err) {
return { success: false, action, error: (err as Error).message };
}
}
// ===== click =====
case 'click': {
const selector = args.selector as string;
const wait = (args.wait as boolean) ?? false;
if (!selector) {
return { success: false, action, error: 'No selector provided' };
}
try {
await getManager().click(selector, wait);
return { success: true, action, selector, clicked: true };
} catch (err) {
return { success: false, action, error: (err as Error).message };
}
}
// ===== type =====
case 'type': {
const selector = args.selector as string;
const text = args.text as string;
const clear = (args.clear as boolean) ?? true;
const submit = (args.submit as boolean) ?? false;
if (!selector) {
return { success: false, action, error: 'No selector provided' };
}
if (text === undefined || text === null) {
return { success: false, action, error: 'No text provided' };
}
try {
await getManager().type(selector, text, { clear, submit });
return { success: true, action, selector, typed: text.length };
} catch (err) {
return { success: false, action, error: (err as Error).message };
}
}
// ===== scroll =====
case 'scroll': {
const direction = (args.direction as string) ?? 'down';
const selector = args.selector as string | undefined;
try {
await getManager().scroll({
direction: direction as 'down' | 'up' | 'top' | 'bottom',
selector,
});
return { success: true, action, direction, selector };
} catch (err) {
return { success: false, action, error: (err as Error).message };
}
}
// ===== wait =====
case 'wait': {
const selector = args.selector as string | undefined;
const timeMs = (args.time_ms as number) ?? 1_000;
try {
await getManager().wait({ selector, timeMs });
return { success: true, action, waited_for: selector ?? `${timeMs}ms` };
} catch (err) {
return { success: false, action, error: (err as Error).message };
}
}
// ===== close =====
case 'close': {
await getManager().close();
return { success: true, action, closed: true };
}
default:
return { success: false, error: `Unknown action: ${action}` };
}
}
}