dsh-builtin-browser 0.4.8 → 0.4.9

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/CHANGELOG.md CHANGED
@@ -4,6 +4,52 @@
4
4
 
5
5
  ---
6
6
 
7
+ # 第四十五轮(2026-10-15,系统浏览器载体的启动生命周期:多开空白页 + 重启后孤儿)
8
+
9
+ 报告者给了一份很硬的证据链 —— CDP `/json/list` 每 200ms 采样的时序、五处建议改法、以及**一条自我更正**(他撤回的第一条建议:去掉启动参数里的 URL 会让 Chrome 变成两个窗口,更糟)。两个问题都发生"把浏览器拉起来"的那一刻,除此之外毫无关系。
10
+
11
+ ## 一、启动那个页面被丢弃,多出来的空白页无法关闭
12
+
13
+ **现象**:每次载体启动都看到 2 个空白标签,其中一个被关掉或残留;残留的那个**不在插件的标签模型里**(`dsh_browser_list_tabs` 只报 1 个),所以 `dsh_browser_close_tab` 关不掉它,窗口长期不关时会累积。
14
+
15
+ **根因两处**:
16
+ 1. `start()` 的启动参数末尾带 `about:blank`,而 Chrome 自己的启动页也在 —— 实测**干净退出后的全新启动也是 2 个页面**,所以不是崩溃恢复,是 Chrome 的固有行为(报告者做了对照实验排除,并保留了那句参数:去掉它会变成 2 个**窗口**)。
17
+ 2. `buildSession()` 用 `Target.createTarget({url:'about:blank'})` **新建**,而不是领养已经存在的那一个 → 启动那个被永久丢弃。
18
+
19
+ **修法**:连上后 `Target.setDiscoverTargets({discover:true})` + `Target.getTargets`,把无人认领的空白页收成**最多 1 个**(只关 `about:blank` 与 `chrome://newtab*`,且**绝不碰已在 `sessions` 里的**);`buildSession()` 先领养那个空白页,没有再新建。**`setDiscoverTargets` 不能省** —— 不先开发现,`getTargets` 可能列不出握手前就存在的页面,而**那正是需要关掉的那个**;报告者第一版补丁就是因此静默失效。
20
+
21
+ ## 二、DSH 重启后旧的载体浏览器成为孤儿,下一次 open 必定失败
22
+
23
+ **现象**:重启 dsh 后,上次拉起的 Chrome 仍在跑;第一次调用直接报 `exited immediately after launch (exit code 0)`,必须手动杀掉用该 profile 的进程才能恢复。
24
+
25
+ **根因**:`start()` 里"杀掉上一个浏览器"只在 `this.child !== undefined` 时执行,而 **DSH 重启后新进程里 `this.child` 是空的** → 旧实例没人管 → 新实例把请求移交给它然后退出 0 → 而这个退出被判成"启动失败"。
26
+
27
+ **修法**:`start()` **在 spawn 之前**先 `adoptRunningBrowser()` —— 读本 profile 的 `DevToolsActivePort`,要求端点自报家门(`/json/version` 的 `Browser` 字段 + `Browser.getVersion` 的 `product`,都必须像浏览器),通过就接管,**不 spawn**;启动回来是 hand-off(exit 0)时再试一次接管。
28
+
29
+ **为什么"接管"是安全的**:启动循环刻意不信任上一次留下的端口文件,理由正当(端口号会被复用到非浏览器进程上)。但**这里多两个前提**:文件在**本插件自己的 profile 目录**里(只有用这个 `--user-data-dir` 启动的浏览器会写它),而且端点要**两次**自证是浏览器。
30
+
31
+ **顺带撤销一处有害的"保护"**:spawn 前删除 `DevToolsActivePort` 被去掉了。它保护不到任何东西 —— `freshPort()` 本来就只认 spawn 之后写入的文件 —— 却销毁了"那个浏览器还在跑"的唯一记录,而那正是接管所需要的。测试里那条"过期文件必须已被删除"的断言随之反正。
32
+
33
+ ## 三、teardown:先请它自己关,再杀
34
+
35
+ `dispose()` 原本固定超时后 `kill()`,而 Windows 上那是 `TerminateProcess`:浏览器来不及写 profile,Chrome 会把这次结束记成崩溃 —— **下一次启动因此触发会话恢复,用户看到一个多出来的窗口**。现在先发 CDP `Browser.close`,给 `CLOSE_GRACE_MS`(6s,`unref` 的定时器,不拖住进程退出)让它自己走,仍在才杀。
36
+
37
+ ## 四、这一轮最该记住的事:同一个位置之前碰过三次,每次只修了一层
38
+
39
+ 用户的疑问是"我记得之前改过类似的,怎么又犯了"。查证后:**不是回归,是一个决定从来没被做过**。
40
+
41
+ | 轮次 | 当时解决的 | 当时没解决的 |
42
+ | --- | --- | --- |
43
+ | 第三十三轮 | "只认本次启动写下的端口"(mtime),不再误连到别的进程 | 那个**真浏览器**被留在原地没人管 |
44
+ | 第三十六轮 | `'exited'` 的文案按退出码区分,不再把崩溃码也说成残留进程 | **它已经准确描述了 hand-off 这个现象,却仍然把它当失败报出来** |
45
+ | 第三十七轮 | `forgetBrowserState` 覆盖"进程已退出、`this.child` 被清空"的形态,视图不再指向死浏览器 | 孤儿进程**仍然占着 profile** |
46
+
47
+ 三次都走到了同一个岔路口,三次都只修了自己那一面 —— **没人问"那浏览器还在跑,我们为什么不接上它"**。第三十六轮那句文案尤其值得记:它把 hand-off 描述得足够清楚,于是读到它的人和报告者都得出"去手动杀掉残留进程"的结论,而正确的结论是"这不该是一个错误"。
48
+
49
+ **测试**:254 → 260(新增 6 条生命周期守卫:接管先于启动、hand-off 读作"已经在那儿"、发现先于列举、只保留一个空白、领养而非丢弃、teardown 先请后杀)。`tsc` 干净。
50
+
51
+ ---
52
+
7
53
  # 第四十四轮(2026-10-14,工具名空间:`browser_*` → `dsh_browser_*`)
8
54
 
9
55
  腾讯 BrowserSkill 的 DSH 插件(`@wxg-prc-cpg/browser-skill-dsh-plugin` @0.3.2)与本插件同装时,它六个工具一个都注册不上。根因不是它的 bug:双方都往**全局**工具层注册了 `browser_session`,宿主对重名直接抛错,而它把注册包在 `try/catch` 里只 `console.warn` —— 桌面版看不到 stdout,于是表现成"插件完全没反应"。用户先怀疑的是它坏了。
@@ -111,6 +111,16 @@ export declare class SystemBrowserViewHost implements ElectronBrowserViewHost {
111
111
  */
112
112
  private clientSuspect;
113
113
  private child;
114
+ /**
115
+ * True when the browser in use was adopted from a previous run rather than launched here.
116
+ *
117
+ * Worth recording because the two differ on teardown: a browser this host spawned is its to
118
+ * kill, while one it merely took over belongs to a process that is already gone — closing it is
119
+ * still right (it holds the profile nobody else can use), but it is a decision, not cleanup.
120
+ */
121
+ private adoptedRunning;
122
+ /** Whether this connection's surplus blank pages have been collapsed already. */
123
+ private blanksNormalized;
114
124
  private starting;
115
125
  /**
116
126
  * Browsers that ignored a kill.
@@ -167,6 +177,25 @@ export declare class SystemBrowserViewHost implements ElectronBrowserViewHost {
167
177
  */
168
178
  private forgetBrowserState;
169
179
  private ensureClient;
180
+ /**
181
+ * Take over a browser already serving this profile, instead of racing it.
182
+ *
183
+ * A previous DSH process leaves its browser running: the child handle died with that process, so
184
+ * nothing here knows about it, and `start()`'s cleanup branch is skipped because it keys on
185
+ * `this.child`. Launching then does something worse than fail — on Windows Chrome hands the
186
+ * request to the instance already holding the profile and exits 0, and that exit is read as a
187
+ * launch failure. The user is told to go kill the leftovers by hand, on every restart, which is
188
+ * exactly what was reported.
189
+ *
190
+ * The launch loop deliberately refuses to trust a `<profile>/DevToolsActivePort` left by a
191
+ * previous launch, and it is right to: a stale port number can be reused by something that is
192
+ * not a browser at all. What makes adoption safe HERE is that this runs BEFORE launching, on a
193
+ * file that lives inside this plugin's own profile directory — only a browser started with this
194
+ * `--user-data-dir` writes it — and the endpoint is then asked to identify itself.
195
+ * @param why - a short phrase for the log, naming which attempt this is.
196
+ * @returns a connected client for the running browser, or undefined when there is none.
197
+ */
198
+ private adoptRunningBrowser;
170
199
  /**
171
200
  * Launch the browser with a private profile and connect over CDP.
172
201
  *
@@ -236,6 +265,37 @@ export declare class SystemBrowserViewHost implements ElectronBrowserViewHost {
236
265
  private ensureSession;
237
266
  /** Create and attach the target for one view. Serialised by {@link ensureSession}. */
238
267
  private buildSession;
268
+ /**
269
+ * Take an unclaimed blank page, if the browser has one to spare.
270
+ *
271
+ * "Unclaimed" is what matters: a blank page the plugin is already driving belongs to a view, and
272
+ * handing the same target to a second view would make two views share one page. The page kept by
273
+ * {@link normalizeBlankPages} is exactly the unclaimed one this is looking for.
274
+ * @param client - the connected browser.
275
+ * @returns the target id to reuse, or undefined when a new page has to be created.
276
+ */
277
+ private adoptBlankTarget;
278
+ /** Create one page, refusing to go on without an id. */
279
+ private createTarget;
280
+ /**
281
+ * Collapse the blank pages a freshly started browser comes with down to one.
282
+ *
283
+ * Chrome opens more than one page when it is given a URL on top of its own start page, and this
284
+ * is the browser's behaviour rather than a session restore: measured on a cleanly exited profile
285
+ * as well as a crashed one, both pages appear within a second of launch. There is nothing to
286
+ * configure away — by the time a client can talk to the browser, both already exist.
287
+ *
288
+ * Left alone the surplus page is invisible to the plugin's tab model and therefore unclosable
289
+ * through the tools, so it survives every navigation and accumulates for as long as the window
290
+ * lives. One is kept deliberately: {@link buildSession} adopts it instead of creating a third.
291
+ *
292
+ * `Target.setDiscoverTargets` has to come first. Without it the browser may omit pages that
293
+ * existed before this client connected — which is precisely the page that needs closing, and the
294
+ * reason the first attempt at this fix failed silently.
295
+ * @param client - the connected browser.
296
+ * @returns how many surplus blank pages were closed.
297
+ */
298
+ private normalizeBlankPages;
239
299
  /** Views whose session is being created right now, so a second caller waits for the first. */
240
300
  private readonly sessionBuilding;
241
301
  /** sessionId -> targetId, so a destroyed view can close the right page. */
@@ -36,6 +36,15 @@ import { join } from 'node:path';
36
36
  * exited, and that exit was reported as a bad browser path.
37
37
  */
38
38
  const EXIT_GRACE_MS = 3_000;
39
+ /**
40
+ * How long a browser is given to close itself after being asked, before it is killed.
41
+ *
42
+ * Long enough for Chrome to flush its profile, which is what keeps the next launch from reading
43
+ * the profile as crashed and offering a session restore. Deliberately not the same knob as
44
+ * {@link EXIT_GRACE_MS}: that one waits for a browser to die because a NEW one cannot start until
45
+ * it does, this one merely prefers a graceful exit to a kill.
46
+ */
47
+ const CLOSE_GRACE_MS = 6_000;
39
48
  /**
40
49
  * Ceiling on one attempt to ask a discovered CDP port who it is.
41
50
  *
@@ -147,6 +156,21 @@ const INSTALL_LOCATIONS = {
147
156
  * @param env - the environment to resolve against.
148
157
  * @returns the path with %NAME% references replaced, or the entry unchanged when unknown.
149
158
  */
159
+ /**
160
+ * Whether a target URL is one of the blanks a browser opens by itself.
161
+ *
162
+ * Deliberately narrow: `about:blank` and Chrome's own new-tab pages, and nothing else. A wider
163
+ * test ("has no path", "starts with chrome://") would sooner or later match a page the user is
164
+ * reading, and the caller closes what this returns.
165
+ * @param url - the target's reported URL.
166
+ */
167
+ function isBlankUrl(url) {
168
+ if (typeof url !== 'string')
169
+ return false;
170
+ return url === 'about:blank'
171
+ || url.startsWith('chrome://newtab')
172
+ || url.startsWith('chrome://new-tab-page');
173
+ }
150
174
  function expandLocation(location, env) {
151
175
  return location.replace(/%([^%]+)%/g, (whole, name) => env[name] ?? env[name.toUpperCase()] ?? env[name.toLowerCase()] ?? whole);
152
176
  }
@@ -385,6 +409,16 @@ export class SystemBrowserViewHost {
385
409
  */
386
410
  clientSuspect = false;
387
411
  child;
412
+ /**
413
+ * True when the browser in use was adopted from a previous run rather than launched here.
414
+ *
415
+ * Worth recording because the two differ on teardown: a browser this host spawned is its to
416
+ * kill, while one it merely took over belongs to a process that is already gone — closing it is
417
+ * still right (it holds the profile nobody else can use), but it is a decision, not cleanup.
418
+ */
419
+ adoptedRunning = false;
420
+ /** Whether this connection's surplus blank pages have been collapsed already. */
421
+ blanksNormalized = false;
388
422
  starting;
389
423
  /**
390
424
  * Browsers that ignored a kill.
@@ -468,6 +502,10 @@ export class SystemBrowserViewHost {
468
502
  forgetBrowserState() {
469
503
  this.views.clear();
470
504
  this.sessions.clear();
505
+ // A new browser comes with its own surplus pages, so the tidy-up runs again for it. Without
506
+ // this, a restart would inherit the previous connection's "already handled" and leave the new
507
+ // browser's extra blank page in place for the rest of the session.
508
+ this.blanksNormalized = false;
471
509
  }
472
510
  async ensureClient() {
473
511
  // A disposed host must never start a browser: doing so would spawn a process
@@ -501,7 +539,78 @@ export class SystemBrowserViewHost {
501
539
  }
502
540
  // Concurrent first calls share one startup rather than racing two browsers.
503
541
  this.starting ??= this.start().finally(() => { this.starting = undefined; });
504
- return await this.starting;
542
+ const client = await this.starting;
543
+ // Once per connection. A browser opens more than one page when it starts, and the surplus is
544
+ // invisible to the plugin's tab model — see normalizeBlankPages for why that matters.
545
+ if (!this.blanksNormalized) {
546
+ this.blanksNormalized = true;
547
+ await this.normalizeBlankPages(client);
548
+ }
549
+ return client;
550
+ }
551
+ /**
552
+ * Take over a browser already serving this profile, instead of racing it.
553
+ *
554
+ * A previous DSH process leaves its browser running: the child handle died with that process, so
555
+ * nothing here knows about it, and `start()`'s cleanup branch is skipped because it keys on
556
+ * `this.child`. Launching then does something worse than fail — on Windows Chrome hands the
557
+ * request to the instance already holding the profile and exits 0, and that exit is read as a
558
+ * launch failure. The user is told to go kill the leftovers by hand, on every restart, which is
559
+ * exactly what was reported.
560
+ *
561
+ * The launch loop deliberately refuses to trust a `<profile>/DevToolsActivePort` left by a
562
+ * previous launch, and it is right to: a stale port number can be reused by something that is
563
+ * not a browser at all. What makes adoption safe HERE is that this runs BEFORE launching, on a
564
+ * file that lives inside this plugin's own profile directory — only a browser started with this
565
+ * `--user-data-dir` writes it — and the endpoint is then asked to identify itself.
566
+ * @param why - a short phrase for the log, naming which attempt this is.
567
+ * @returns a connected client for the running browser, or undefined when there is none.
568
+ */
569
+ async adoptRunningBrowser(why) {
570
+ const portFile = join(this.profileDir, 'DevToolsActivePort');
571
+ let port;
572
+ try {
573
+ if (!existsSync(portFile))
574
+ return undefined;
575
+ const value = Number(readFileSync(portFile, 'utf8').split('\n')[0]?.trim() ?? '');
576
+ if (!Number.isInteger(value) || value <= 0)
577
+ return undefined;
578
+ port = value;
579
+ }
580
+ catch {
581
+ return undefined;
582
+ }
583
+ try {
584
+ const version = await (await fetch(`http://127.0.0.1:${port}/json/version`, {
585
+ signal: AbortSignal.timeout(PORT_PROBE_MS),
586
+ })).json();
587
+ if (typeof version.webSocketDebuggerUrl !== 'string')
588
+ return undefined;
589
+ // The port number alone proves nothing — it can be reused once the browser that owned it
590
+ // exits. The endpoint's own answer is the evidence, and it has to name a browser.
591
+ if (typeof version.Browser !== 'string' || !/chrome|chromium|edge/i.test(version.Browser))
592
+ return undefined;
593
+ const client = new CdpClient(version.webSocketDebuggerUrl);
594
+ await client.whenReady();
595
+ // Asked through the protocol too, so a websocket that merely accepts connections is not
596
+ // mistaken for a browser that answers.
597
+ const info = await client.send('Browser.getVersion', {});
598
+ if (typeof info?.product !== 'string' || !/chrome|chromium|edge/i.test(info.product)) {
599
+ client.close();
600
+ return undefined;
601
+ }
602
+ if (this.disposed) {
603
+ client.close();
604
+ return undefined;
605
+ }
606
+ this.client = client;
607
+ this.adoptedRunning = true;
608
+ return client;
609
+ }
610
+ catch {
611
+ // No endpoint, nothing listening, or not a browser: all of them mean "launch one".
612
+ return undefined;
613
+ }
505
614
  }
506
615
  /**
507
616
  * Launch the browser with a private profile and connect over CDP.
@@ -512,6 +621,12 @@ export class SystemBrowserViewHost {
512
621
  * @returns the connected client.
513
622
  */
514
623
  async start() {
624
+ // Before spawning anything: is a browser for THIS profile already running? A DSH restart
625
+ // leaves one behind, and launching into it is the failure reported as "exited immediately
626
+ // after launch (exit code 0)" that required killing seven processes by hand.
627
+ const adopted = await this.adoptRunningBrowser('before launch');
628
+ if (adopted !== undefined)
629
+ return adopted;
515
630
  const args = [
516
631
  '--remote-debugging-port=0',
517
632
  `--user-data-dir=${this.profileDir}`,
@@ -569,15 +684,14 @@ export class SystemBrowserViewHost {
569
684
  });
570
685
  });
571
686
  }
572
- // The port file belongs to the browser that just died, and it is NOT removed when the
573
- // process is killed: nothing gives a terminated browser a chance to clean up. The launch
574
- // loop below reads it to learn the port, so a stale file means connecting to whatever else
575
- // holds that port — or to the previous instance — and publishing a client this host does
576
- // not own and dispose() can never kill.
577
- try {
578
- rmSync(join(this.profileDir, 'DevToolsActivePort'), { force: true });
579
- }
580
- catch { /* nothing to remove */ }
687
+ // The port file is deliberately NOT deleted here, and that is a reversal.
688
+ //
689
+ // It used to be removed so the launch loop below could not read a previous browser's port. But
690
+ // the loop already refuses any file older than the spawn (`freshPort` compares mtimes), so the
691
+ // deletion never protected it — while it DID destroy the only record of a browser that may
692
+ // still be running for this profile. That record is what `adoptRunningBrowser` needs when the
693
+ // launch comes back as a hand-off: the browser was alive the whole time, and without the file
694
+ // there is no way to find it again.
581
695
  const spawnedAt = Date.now();
582
696
  let child;
583
697
  try {
@@ -728,6 +842,14 @@ export class SystemBrowserViewHost {
728
842
  const code = child.exitCode;
729
843
  const signal = child.signalCode;
730
844
  const clean = code === 0 && signal === null;
845
+ // Exit code 0 IS the hand-off: the request went to an instance already holding this profile,
846
+ // so that instance is our browser and it is running right now. Adopting it turns the
847
+ // reported "kill the leftovers by hand, on every restart" into a call that just works.
848
+ if (clean) {
849
+ const recovered = await this.adoptRunningBrowser('after a hand-off');
850
+ if (recovered !== undefined)
851
+ return recovered;
852
+ }
731
853
  throw new Error(clean
732
854
  ? `dsh-builtin-browser: ${target} exited immediately after launch (exit code 0). `
733
855
  + 'That means it handed the request to an instance already using this profile '
@@ -892,20 +1014,39 @@ export class SystemBrowserViewHost {
892
1014
  if (this.disposed)
893
1015
  return;
894
1016
  this.disposed = true;
1017
+ const client = this.client;
895
1018
  this.forgetBrowserState();
896
- this.client?.close();
897
- try {
898
- this.child?.kill();
899
- }
900
- catch { /* already gone */ }
901
- // And anything that ignored an earlier kill. They hold this profile, so leaving them
902
- // running means the next launch contends with a browser nobody owns.
903
- for (const leftover of this.stubborn.splice(0)) {
1019
+ this.client = undefined;
1020
+ // Ask before killing. `kill()` on Windows is TerminateProcess, which gives the browser no
1021
+ // chance to flush its profile — and Chrome records how it ended in that profile, so a
1022
+ // hard kill leaves it marked as a crash and the NEXT launch offers to restore a session.
1023
+ // That is a visible symptom (an extra window) traced all the way back to this line.
1024
+ if (client !== undefined) {
904
1025
  try {
905
- leftover.kill();
1026
+ void client.send('Browser.close', {}).catch(() => { });
906
1027
  }
907
1028
  catch { /* already gone */ }
908
1029
  }
1030
+ // Both the browser we launched and anything that ignored an earlier kill. They hold this
1031
+ // profile, so leaving them running means the next launch contends with a browser nobody owns.
1032
+ const doomed = [this.child, ...this.stubborn.splice(0)].filter((entry) => entry !== undefined);
1033
+ this.child = undefined;
1034
+ for (const dying of doomed) {
1035
+ /**
1036
+ * A bounded reprieve, then the kill we used to do immediately.
1037
+ *
1038
+ * The timer is unref'd so it never holds the process open: a DSH shutdown that is itself
1039
+ * ending does not wait for this, and the browser — having been asked to close — generally
1040
+ * gets there on its own. If it does not, the kill still lands while the process lives.
1041
+ */
1042
+ const force = () => { try {
1043
+ dying.kill();
1044
+ }
1045
+ catch { /* already gone */ } };
1046
+ const timer = setTimeout(force, CLOSE_GRACE_MS);
1047
+ timer.unref?.();
1048
+ dying.once('exit', () => { clearTimeout(timer); });
1049
+ }
909
1050
  const ephemeral = this.ephemeralDir;
910
1051
  if (ephemeral !== undefined && this.child !== undefined) {
911
1052
  // The browser flushes its profile while shutting down, so deleting later avoids
@@ -957,10 +1098,15 @@ export class SystemBrowserViewHost {
957
1098
  /** Create and attach the target for one view. Serialised by {@link ensureSession}. */
958
1099
  async buildSession(viewId) {
959
1100
  const client = await this.ensureClient();
960
- const created = await client.send('Target.createTarget', { url: 'about:blank' });
961
- const targetId = created?.targetId;
962
- if (typeof targetId !== 'string')
963
- throw new Error('dsh-builtin-browser: the browser did not create a page');
1101
+ // Adopt a blank page the browser opened on its own before creating another one.
1102
+ //
1103
+ // Chrome starts with more than one page when handed a URL on top of its own start page, so
1104
+ // creating a target here discarded the original and left it behind as a tab that is not in
1105
+ // the plugin's tab model — invisible to `dsh_browser_list_tabs`, and therefore unclosable
1106
+ // through the tools. Taking the blank instead makes the page the browser already opened the
1107
+ // one the caller gets.
1108
+ const adopted = await this.adoptBlankTarget(client);
1109
+ const targetId = adopted ?? await this.createTarget(client);
964
1110
  const attached = await client.send('Target.attachToTarget', { targetId, flatten: true });
965
1111
  const sessionId = attached?.sessionId;
966
1112
  if (typeof sessionId !== 'string')
@@ -970,6 +1116,94 @@ export class SystemBrowserViewHost {
970
1116
  this.views.set(viewId, { session: sessionId, client });
971
1117
  return sessionId;
972
1118
  }
1119
+ /**
1120
+ * Take an unclaimed blank page, if the browser has one to spare.
1121
+ *
1122
+ * "Unclaimed" is what matters: a blank page the plugin is already driving belongs to a view, and
1123
+ * handing the same target to a second view would make two views share one page. The page kept by
1124
+ * {@link normalizeBlankPages} is exactly the unclaimed one this is looking for.
1125
+ * @param client - the connected browser.
1126
+ * @returns the target id to reuse, or undefined when a new page has to be created.
1127
+ */
1128
+ async adoptBlankTarget(client) {
1129
+ try {
1130
+ const listed = await client.send('Target.getTargets', {});
1131
+ const claimed = new Set(this.sessions.values());
1132
+ for (const entry of listed?.targetInfos ?? []) {
1133
+ if (entry.type !== 'page' || typeof entry.targetId !== 'string')
1134
+ continue;
1135
+ if (!isBlankUrl(entry.url))
1136
+ continue;
1137
+ if (claimed.has(entry.targetId))
1138
+ continue;
1139
+ return entry.targetId;
1140
+ }
1141
+ return undefined;
1142
+ }
1143
+ catch {
1144
+ // Discovery is best-effort; creating a page is the behaviour this replaced.
1145
+ return undefined;
1146
+ }
1147
+ }
1148
+ /** Create one page, refusing to go on without an id. */
1149
+ async createTarget(client) {
1150
+ const created = await client.send('Target.createTarget', { url: 'about:blank' });
1151
+ const targetId = created?.targetId;
1152
+ if (typeof targetId !== 'string')
1153
+ throw new Error('dsh-builtin-browser: the browser did not create a page');
1154
+ return targetId;
1155
+ }
1156
+ /**
1157
+ * Collapse the blank pages a freshly started browser comes with down to one.
1158
+ *
1159
+ * Chrome opens more than one page when it is given a URL on top of its own start page, and this
1160
+ * is the browser's behaviour rather than a session restore: measured on a cleanly exited profile
1161
+ * as well as a crashed one, both pages appear within a second of launch. There is nothing to
1162
+ * configure away — by the time a client can talk to the browser, both already exist.
1163
+ *
1164
+ * Left alone the surplus page is invisible to the plugin's tab model and therefore unclosable
1165
+ * through the tools, so it survives every navigation and accumulates for as long as the window
1166
+ * lives. One is kept deliberately: {@link buildSession} adopts it instead of creating a third.
1167
+ *
1168
+ * `Target.setDiscoverTargets` has to come first. Without it the browser may omit pages that
1169
+ * existed before this client connected — which is precisely the page that needs closing, and the
1170
+ * reason the first attempt at this fix failed silently.
1171
+ * @param client - the connected browser.
1172
+ * @returns how many surplus blank pages were closed.
1173
+ */
1174
+ async normalizeBlankPages(client) {
1175
+ try {
1176
+ await client.send('Target.setDiscoverTargets', { discover: true });
1177
+ const listed = await client.send('Target.getTargets', {});
1178
+ const blanks = (listed?.targetInfos ?? []).filter(entry => entry.type === 'page' && typeof entry.targetId === 'string' && isBlankUrl(entry.url));
1179
+ const claimed = new Set(this.sessions.values());
1180
+ let closed = 0;
1181
+ // Keep the first unclaimed blank (that is the one buildSession will adopt) and close the rest.
1182
+ let kept = false;
1183
+ for (const entry of blanks) {
1184
+ const targetId = entry.targetId;
1185
+ if (claimed.has(targetId))
1186
+ continue;
1187
+ if (!kept) {
1188
+ kept = true;
1189
+ continue;
1190
+ }
1191
+ try {
1192
+ await client.send('Target.closeTarget', { targetId });
1193
+ closed += 1;
1194
+ }
1195
+ catch {
1196
+ // It may already be gone; one failure must not stop the rest.
1197
+ }
1198
+ }
1199
+ return closed;
1200
+ }
1201
+ catch {
1202
+ // A browser that will not answer Target commands has a real problem, and the next command
1203
+ // will report it. Turning a tidy-up into a functional error here would hide that.
1204
+ return 0;
1205
+ }
1206
+ }
973
1207
  /** Views whose session is being created right now, so a second caller waits for the first. */
974
1208
  sessionBuilding = new Map();
975
1209
  /** sessionId -> targetId, so a destroyed view can close the right page. */
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "dsh-builtin-browser",
3
- "version": "0.4.8",
3
+ "version": "0.4.9",
4
4
  "description": "Shared real browser plugin for DeepSeek Harness: install-and-use — a visible native browser the human can take over, driven by the agent over CDP. DOM-level interaction (React/Vue safe), per-task session isolation, cookie persistence (dsh_browser_auth), CAPTCHA detection, batch form fill (dsh_browser_fill), authenticated download, history/replay, action restriction",
5
5
  "type": "module",
6
6
  "main": "lib/index.js",
@@ -38,6 +38,15 @@ import { join } from 'node:path'
38
38
  * exited, and that exit was reported as a bad browser path.
39
39
  */
40
40
  const EXIT_GRACE_MS = 3_000
41
+ /**
42
+ * How long a browser is given to close itself after being asked, before it is killed.
43
+ *
44
+ * Long enough for Chrome to flush its profile, which is what keeps the next launch from reading
45
+ * the profile as crashed and offering a session restore. Deliberately not the same knob as
46
+ * {@link EXIT_GRACE_MS}: that one waits for a browser to die because a NEW one cannot start until
47
+ * it does, this one merely prefers a graceful exit to a kill.
48
+ */
49
+ const CLOSE_GRACE_MS = 6_000
41
50
 
42
51
  /**
43
52
  * Ceiling on one attempt to ask a discovered CDP port who it is.
@@ -178,6 +187,21 @@ const INSTALL_LOCATIONS: Record<DetectedBrowser['kind'], Partial<Record<NodeJS.P
178
187
  * @param env - the environment to resolve against.
179
188
  * @returns the path with %NAME% references replaced, or the entry unchanged when unknown.
180
189
  */
190
+ /**
191
+ * Whether a target URL is one of the blanks a browser opens by itself.
192
+ *
193
+ * Deliberately narrow: `about:blank` and Chrome's own new-tab pages, and nothing else. A wider
194
+ * test ("has no path", "starts with chrome://") would sooner or later match a page the user is
195
+ * reading, and the caller closes what this returns.
196
+ * @param url - the target's reported URL.
197
+ */
198
+ function isBlankUrl(url: unknown): boolean {
199
+ if (typeof url !== 'string') return false
200
+ return url === 'about:blank'
201
+ || url.startsWith('chrome://newtab')
202
+ || url.startsWith('chrome://new-tab-page')
203
+ }
204
+
181
205
  function expandLocation(location: string, env: NodeJS.ProcessEnv): string {
182
206
  return location.replace(/%([^%]+)%/g, (whole, name: string) =>
183
207
  env[name] ?? env[name.toUpperCase()] ?? env[name.toLowerCase()] ?? whole)
@@ -415,6 +439,16 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
415
439
  */
416
440
  private clientSuspect = false
417
441
  private child: ChildProcess | undefined
442
+ /**
443
+ * True when the browser in use was adopted from a previous run rather than launched here.
444
+ *
445
+ * Worth recording because the two differ on teardown: a browser this host spawned is its to
446
+ * kill, while one it merely took over belongs to a process that is already gone — closing it is
447
+ * still right (it holds the profile nobody else can use), but it is a decision, not cleanup.
448
+ */
449
+ private adoptedRunning = false
450
+ /** Whether this connection's surplus blank pages have been collapsed already. */
451
+ private blanksNormalized = false
418
452
  private starting: Promise<CdpClient> | undefined
419
453
 
420
454
  /**
@@ -503,6 +537,10 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
503
537
  private forgetBrowserState(): void {
504
538
  this.views.clear()
505
539
  this.sessions.clear()
540
+ // A new browser comes with its own surplus pages, so the tidy-up runs again for it. Without
541
+ // this, a restart would inherit the previous connection's "already handled" and leave the new
542
+ // browser's extra blank page in place for the rest of the session.
543
+ this.blanksNormalized = false
506
544
  }
507
545
 
508
546
  private async ensureClient(): Promise<CdpClient> {
@@ -535,7 +573,70 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
535
573
  }
536
574
  // Concurrent first calls share one startup rather than racing two browsers.
537
575
  this.starting ??= this.start().finally(() => { this.starting = undefined })
538
- return await this.starting
576
+ const client = await this.starting
577
+ // Once per connection. A browser opens more than one page when it starts, and the surplus is
578
+ // invisible to the plugin's tab model — see normalizeBlankPages for why that matters.
579
+ if (!this.blanksNormalized) {
580
+ this.blanksNormalized = true
581
+ await this.normalizeBlankPages(client)
582
+ }
583
+ return client
584
+ }
585
+
586
+ /**
587
+ * Take over a browser already serving this profile, instead of racing it.
588
+ *
589
+ * A previous DSH process leaves its browser running: the child handle died with that process, so
590
+ * nothing here knows about it, and `start()`'s cleanup branch is skipped because it keys on
591
+ * `this.child`. Launching then does something worse than fail — on Windows Chrome hands the
592
+ * request to the instance already holding the profile and exits 0, and that exit is read as a
593
+ * launch failure. The user is told to go kill the leftovers by hand, on every restart, which is
594
+ * exactly what was reported.
595
+ *
596
+ * The launch loop deliberately refuses to trust a `<profile>/DevToolsActivePort` left by a
597
+ * previous launch, and it is right to: a stale port number can be reused by something that is
598
+ * not a browser at all. What makes adoption safe HERE is that this runs BEFORE launching, on a
599
+ * file that lives inside this plugin's own profile directory — only a browser started with this
600
+ * `--user-data-dir` writes it — and the endpoint is then asked to identify itself.
601
+ * @param why - a short phrase for the log, naming which attempt this is.
602
+ * @returns a connected client for the running browser, or undefined when there is none.
603
+ */
604
+ private async adoptRunningBrowser(why: string): Promise<CdpClient | undefined> {
605
+ const portFile = join(this.profileDir, 'DevToolsActivePort')
606
+ let port: number
607
+ try {
608
+ if (!existsSync(portFile)) return undefined
609
+ const value = Number(readFileSync(portFile, 'utf8').split('\n')[0]?.trim() ?? '')
610
+ if (!Number.isInteger(value) || value <= 0) return undefined
611
+ port = value
612
+ } catch {
613
+ return undefined
614
+ }
615
+ try {
616
+ const version = await (await fetch(`http://127.0.0.1:${port}/json/version`, {
617
+ signal: AbortSignal.timeout(PORT_PROBE_MS),
618
+ })).json() as { webSocketDebuggerUrl?: string; Browser?: string }
619
+ if (typeof version.webSocketDebuggerUrl !== 'string') return undefined
620
+ // The port number alone proves nothing — it can be reused once the browser that owned it
621
+ // exits. The endpoint's own answer is the evidence, and it has to name a browser.
622
+ if (typeof version.Browser !== 'string' || !/chrome|chromium|edge/i.test(version.Browser)) return undefined
623
+ const client = new CdpClient(version.webSocketDebuggerUrl)
624
+ await client.whenReady()
625
+ // Asked through the protocol too, so a websocket that merely accepts connections is not
626
+ // mistaken for a browser that answers.
627
+ const info = await client.send('Browser.getVersion', {}) as { product?: string }
628
+ if (typeof info?.product !== 'string' || !/chrome|chromium|edge/i.test(info.product)) {
629
+ client.close()
630
+ return undefined
631
+ }
632
+ if (this.disposed) { client.close(); return undefined }
633
+ this.client = client
634
+ this.adoptedRunning = true
635
+ return client
636
+ } catch {
637
+ // No endpoint, nothing listening, or not a browser: all of them mean "launch one".
638
+ return undefined
639
+ }
539
640
  }
540
641
 
541
642
  /**
@@ -547,6 +648,11 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
547
648
  * @returns the connected client.
548
649
  */
549
650
  private async start(): Promise<CdpClient> {
651
+ // Before spawning anything: is a browser for THIS profile already running? A DSH restart
652
+ // leaves one behind, and launching into it is the failure reported as "exited immediately
653
+ // after launch (exit code 0)" that required killing seven processes by hand.
654
+ const adopted = await this.adoptRunningBrowser('before launch')
655
+ if (adopted !== undefined) return adopted
550
656
  const args = [
551
657
  '--remote-debugging-port=0',
552
658
  `--user-data-dir=${this.profileDir}`,
@@ -598,12 +704,14 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
598
704
  })
599
705
  })
600
706
  }
601
- // The port file belongs to the browser that just died, and it is NOT removed when the
602
- // process is killed: nothing gives a terminated browser a chance to clean up. The launch
603
- // loop below reads it to learn the port, so a stale file means connecting to whatever else
604
- // holds that port — or to the previous instance — and publishing a client this host does
605
- // not own and dispose() can never kill.
606
- try { rmSync(join(this.profileDir, 'DevToolsActivePort'), { force: true }) } catch { /* nothing to remove */ }
707
+ // The port file is deliberately NOT deleted here, and that is a reversal.
708
+ //
709
+ // It used to be removed so the launch loop below could not read a previous browser's port. But
710
+ // the loop already refuses any file older than the spawn (`freshPort` compares mtimes), so the
711
+ // deletion never protected it — while it DID destroy the only record of a browser that may
712
+ // still be running for this profile. That record is what `adoptRunningBrowser` needs when the
713
+ // launch comes back as a hand-off: the browser was alive the whole time, and without the file
714
+ // there is no way to find it again.
607
715
  const spawnedAt = Date.now()
608
716
  let child: ReturnType<typeof spawn>
609
717
  try {
@@ -723,6 +831,13 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
723
831
  const code = child.exitCode
724
832
  const signal = child.signalCode
725
833
  const clean = code === 0 && signal === null
834
+ // Exit code 0 IS the hand-off: the request went to an instance already holding this profile,
835
+ // so that instance is our browser and it is running right now. Adopting it turns the
836
+ // reported "kill the leftovers by hand, on every restart" into a call that just works.
837
+ if (clean) {
838
+ const recovered = await this.adoptRunningBrowser('after a hand-off')
839
+ if (recovered !== undefined) return recovered
840
+ }
726
841
  throw new Error(
727
842
  clean
728
843
  ? `dsh-builtin-browser: ${target} exited immediately after launch (exit code 0). `
@@ -892,13 +1007,34 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
892
1007
  dispose(): void {
893
1008
  if (this.disposed) return
894
1009
  this.disposed = true
1010
+ const client = this.client
895
1011
  this.forgetBrowserState()
896
- this.client?.close()
897
- try { this.child?.kill() } catch { /* already gone */ }
898
- // And anything that ignored an earlier kill. They hold this profile, so leaving them
899
- // running means the next launch contends with a browser nobody owns.
900
- for (const leftover of this.stubborn.splice(0)) {
901
- try { leftover.kill() } catch { /* already gone */ }
1012
+ this.client = undefined
1013
+ // Ask before killing. `kill()` on Windows is TerminateProcess, which gives the browser no
1014
+ // chance to flush its profile — and Chrome records how it ended in that profile, so a
1015
+ // hard kill leaves it marked as a crash and the NEXT launch offers to restore a session.
1016
+ // That is a visible symptom (an extra window) traced all the way back to this line.
1017
+ if (client !== undefined) {
1018
+ try { void client.send('Browser.close', {}).catch(() => { /* it is closing, not talking */ }) } catch { /* already gone */ }
1019
+ }
1020
+ // Both the browser we launched and anything that ignored an earlier kill. They hold this
1021
+ // profile, so leaving them running means the next launch contends with a browser nobody owns.
1022
+ const doomed = [this.child, ...this.stubborn.splice(0)].filter(
1023
+ (entry): entry is ChildProcess => entry !== undefined,
1024
+ )
1025
+ this.child = undefined
1026
+ for (const dying of doomed) {
1027
+ /**
1028
+ * A bounded reprieve, then the kill we used to do immediately.
1029
+ *
1030
+ * The timer is unref'd so it never holds the process open: a DSH shutdown that is itself
1031
+ * ending does not wait for this, and the browser — having been asked to close — generally
1032
+ * gets there on its own. If it does not, the kill still lands while the process lives.
1033
+ */
1034
+ const force = (): void => { try { dying.kill() } catch { /* already gone */ } }
1035
+ const timer = setTimeout(force, CLOSE_GRACE_MS)
1036
+ timer.unref?.()
1037
+ dying.once('exit', () => { clearTimeout(timer) })
902
1038
  }
903
1039
  const ephemeral = this.ephemeralDir
904
1040
  if (ephemeral !== undefined && this.child !== undefined) {
@@ -945,9 +1081,15 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
945
1081
  /** Create and attach the target for one view. Serialised by {@link ensureSession}. */
946
1082
  private async buildSession(viewId: string): Promise<string> {
947
1083
  const client = await this.ensureClient()
948
- const created = await client.send('Target.createTarget', { url: 'about:blank' }) as { targetId?: string }
949
- const targetId = created?.targetId
950
- if (typeof targetId !== 'string') throw new Error('dsh-builtin-browser: the browser did not create a page')
1084
+ // Adopt a blank page the browser opened on its own before creating another one.
1085
+ //
1086
+ // Chrome starts with more than one page when handed a URL on top of its own start page, so
1087
+ // creating a target here discarded the original and left it behind as a tab that is not in
1088
+ // the plugin's tab model — invisible to `dsh_browser_list_tabs`, and therefore unclosable
1089
+ // through the tools. Taking the blank instead makes the page the browser already opened the
1090
+ // one the caller gets.
1091
+ const adopted = await this.adoptBlankTarget(client)
1092
+ const targetId = adopted ?? await this.createTarget(client)
951
1093
  const attached = await client.send('Target.attachToTarget', { targetId, flatten: true }) as { sessionId?: string }
952
1094
  const sessionId = attached?.sessionId
953
1095
  if (typeof sessionId !== 'string') throw new Error('dsh-builtin-browser: the browser did not attach to the new page')
@@ -957,6 +1099,92 @@ export class SystemBrowserViewHost implements ElectronBrowserViewHost {
957
1099
  return sessionId
958
1100
  }
959
1101
 
1102
+ /**
1103
+ * Take an unclaimed blank page, if the browser has one to spare.
1104
+ *
1105
+ * "Unclaimed" is what matters: a blank page the plugin is already driving belongs to a view, and
1106
+ * handing the same target to a second view would make two views share one page. The page kept by
1107
+ * {@link normalizeBlankPages} is exactly the unclaimed one this is looking for.
1108
+ * @param client - the connected browser.
1109
+ * @returns the target id to reuse, or undefined when a new page has to be created.
1110
+ */
1111
+ private async adoptBlankTarget(client: CdpClient): Promise<string | undefined> {
1112
+ try {
1113
+ const listed = await client.send('Target.getTargets', {}) as {
1114
+ targetInfos?: Array<{ targetId?: string; type?: string; url?: string }>
1115
+ }
1116
+ const claimed = new Set(this.sessions.values())
1117
+ for (const entry of listed?.targetInfos ?? []) {
1118
+ if (entry.type !== 'page' || typeof entry.targetId !== 'string') continue
1119
+ if (!isBlankUrl(entry.url)) continue
1120
+ if (claimed.has(entry.targetId)) continue
1121
+ return entry.targetId
1122
+ }
1123
+ return undefined
1124
+ } catch {
1125
+ // Discovery is best-effort; creating a page is the behaviour this replaced.
1126
+ return undefined
1127
+ }
1128
+ }
1129
+
1130
+ /** Create one page, refusing to go on without an id. */
1131
+ private async createTarget(client: CdpClient): Promise<string> {
1132
+ const created = await client.send('Target.createTarget', { url: 'about:blank' }) as { targetId?: string }
1133
+ const targetId = created?.targetId
1134
+ if (typeof targetId !== 'string') throw new Error('dsh-builtin-browser: the browser did not create a page')
1135
+ return targetId
1136
+ }
1137
+
1138
+ /**
1139
+ * Collapse the blank pages a freshly started browser comes with down to one.
1140
+ *
1141
+ * Chrome opens more than one page when it is given a URL on top of its own start page, and this
1142
+ * is the browser's behaviour rather than a session restore: measured on a cleanly exited profile
1143
+ * as well as a crashed one, both pages appear within a second of launch. There is nothing to
1144
+ * configure away — by the time a client can talk to the browser, both already exist.
1145
+ *
1146
+ * Left alone the surplus page is invisible to the plugin's tab model and therefore unclosable
1147
+ * through the tools, so it survives every navigation and accumulates for as long as the window
1148
+ * lives. One is kept deliberately: {@link buildSession} adopts it instead of creating a third.
1149
+ *
1150
+ * `Target.setDiscoverTargets` has to come first. Without it the browser may omit pages that
1151
+ * existed before this client connected — which is precisely the page that needs closing, and the
1152
+ * reason the first attempt at this fix failed silently.
1153
+ * @param client - the connected browser.
1154
+ * @returns how many surplus blank pages were closed.
1155
+ */
1156
+ private async normalizeBlankPages(client: CdpClient): Promise<number> {
1157
+ try {
1158
+ await client.send('Target.setDiscoverTargets', { discover: true })
1159
+ const listed = await client.send('Target.getTargets', {}) as {
1160
+ targetInfos?: Array<{ targetId?: string; type?: string; url?: string }>
1161
+ }
1162
+ const blanks = (listed?.targetInfos ?? []).filter(
1163
+ entry => entry.type === 'page' && typeof entry.targetId === 'string' && isBlankUrl(entry.url),
1164
+ )
1165
+ const claimed = new Set(this.sessions.values())
1166
+ let closed = 0
1167
+ // Keep the first unclaimed blank (that is the one buildSession will adopt) and close the rest.
1168
+ let kept = false
1169
+ for (const entry of blanks) {
1170
+ const targetId = entry.targetId as string
1171
+ if (claimed.has(targetId)) continue
1172
+ if (!kept) { kept = true; continue }
1173
+ try {
1174
+ await client.send('Target.closeTarget', { targetId })
1175
+ closed += 1
1176
+ } catch {
1177
+ // It may already be gone; one failure must not stop the rest.
1178
+ }
1179
+ }
1180
+ return closed
1181
+ } catch {
1182
+ // A browser that will not answer Target commands has a real problem, and the next command
1183
+ // will report it. Turning a tidy-up into a functional error here would hide that.
1184
+ return 0
1185
+ }
1186
+ }
1187
+
960
1188
  /** Views whose session is being created right now, so a second caller waits for the first. */
961
1189
  private readonly sessionBuilding = new Map<string, Promise<string>>()
962
1190