dsh-builtin-browser 0.4.7 → 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.
Files changed (44) hide show
  1. package/CHANGELOG.md +74 -0
  2. package/README.en.md +62 -491
  3. package/README.md +44 -470
  4. package/client.js +10 -10
  5. package/cordis.patch.yml +4 -4
  6. package/desktop-bridge/plugin-browser-bridge.js +1 -1
  7. package/docs/README.md +8 -4
  8. package/docs/architecture.md +82 -11
  9. package/docs/development.md +38 -0
  10. package/docs/limitations.md +49 -0
  11. package/docs/tool-reference.md +42 -42
  12. package/docs/user-guide.md +93 -37
  13. package/docs/why-browser.md +3 -3
  14. package/lib/browser/types.d.ts +2 -2
  15. package/lib/browser-electron/entry.d.ts +2 -2
  16. package/lib/browser-electron/entry.js +1 -1
  17. package/lib/browser-electron/history-store.d.ts +1 -1
  18. package/lib/browser-electron/history-store.js +1 -1
  19. package/lib/browser-electron/host-main.js +3 -3
  20. package/lib/browser-electron/provider.d.ts +9 -9
  21. package/lib/browser-electron/provider.js +14 -14
  22. package/lib/browser-electron/remote-host.js +1 -1
  23. package/lib/browser-electron/settings-store.d.ts +4 -4
  24. package/lib/browser-electron/system-browser.d.ts +60 -0
  25. package/lib/browser-electron/system-browser.js +257 -23
  26. package/lib/index.d.ts +1 -1
  27. package/lib/index.js +1 -1
  28. package/lib/tool-browser/element-fields.d.ts +4 -4
  29. package/lib/tool-browser/element-fields.js +6 -6
  30. package/lib/tool-browser/index.d.ts +6 -6
  31. package/lib/tool-browser/index.js +161 -139
  32. package/package.json +2 -2
  33. package/scripts/verify-host-compat.mjs +6 -6
  34. package/src/browser/types.ts +2 -2
  35. package/src/browser-electron/entry.ts +2 -2
  36. package/src/browser-electron/history-store.ts +1 -1
  37. package/src/browser-electron/host-main.ts +3 -3
  38. package/src/browser-electron/provider.ts +15 -15
  39. package/src/browser-electron/remote-host.ts +1 -1
  40. package/src/browser-electron/settings-store.ts +4 -4
  41. package/src/browser-electron/system-browser.ts +244 -16
  42. package/src/index.ts +1 -1
  43. package/src/tool-browser/element-fields.ts +6 -6
  44. package/src/tool-browser/index.ts +164 -140
package/package.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "name": "dsh-builtin-browser",
3
- "version": "0.4.7",
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 (browser_auth), CAPTCHA detection, batch form fill (browser_fill), authenticated download, history/replay, action restriction",
3
+ "version": "0.4.9",
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",
7
7
  "types": "lib/index.d.ts",
@@ -169,20 +169,20 @@ try {
169
169
 
170
170
  const exec = { agent: { id: 'verify-host-compat' } }
171
171
  try {
172
- const value = await registered.get('browser_session').execute({}, exec)
173
- ok('browser_session ran', value !== undefined, JSON.stringify(value).slice(0, 100))
172
+ const value = await registered.get('dsh_browser_session').execute({}, exec)
173
+ ok('dsh_browser_session ran', value !== undefined, JSON.stringify(value).slice(0, 100))
174
174
  } catch (error) {
175
- ok('browser_session ran', false, String(error && error.message))
175
+ ok('dsh_browser_session ran', false, String(error && error.message))
176
176
  }
177
177
 
178
178
  try {
179
- const tool = registered.get('browser_a11y')
179
+ const tool = registered.get('dsh_browser_a11y')
180
180
  const value = await tool.execute({}, exec)
181
181
  const text = JSON.stringify(tool.output.render({}, value))
182
- ok('browser_a11y ran and rendered', value?.nodes?.length === 1, '')
182
+ ok('dsh_browser_a11y ran and rendered', value?.nodes?.length === 1, '')
183
183
  ok('its renderer emits the element selector', text.includes('#probe'), text.slice(0, 120))
184
184
  } catch (error) {
185
- ok('browser_a11y ran', false, String(error && error.message))
185
+ ok('dsh_browser_a11y ran', false, String(error && error.message))
186
186
  }
187
187
 
188
188
  console.log('')
@@ -65,7 +65,7 @@ export interface BrowserScrollRequest {
65
65
 
66
66
  /** Press one named key (Enter, Tab, arrows, …). */
67
67
  export interface BrowserKeyRequest {
68
- /** Key name from the supported set (see the browser_key tool). */
68
+ /** Key name from the supported set (see the dsh_browser_key tool). */
69
69
  readonly key: string
70
70
  }
71
71
 
@@ -361,7 +361,7 @@ export interface BrowserSnapshotElement {
361
361
  /**
362
362
  * True when the element lives inside an (same-origin) iframe: coordinates
363
363
  * are still top-document, but DOM selectors are frame-scoped — reach the
364
- * element via `iframe.contentDocument` in browser_execute.
364
+ * element via `iframe.contentDocument` in dsh_browser_execute.
365
365
  */
366
366
  readonly frame?: boolean
367
367
  }
@@ -5,7 +5,7 @@
5
5
  * `ctx.electronViewHost`, that host is used (embedded, human-machine shared
6
6
  * view). Otherwise the plugin self-hosts: it spawns its own Electron child
7
7
  * (`host-main.js`) and drives it over a local TCP JSON-RPC socket, so
8
- * installing the plugin is enough for `browser_*` tools to work on any
8
+ * installing the plugin is enough for `dsh_browser_*` tools to work on any
9
9
  * surface.
10
10
  * @module dsh-browser/browser-electron
11
11
  */
@@ -48,7 +48,7 @@ export interface Config {
48
48
  /** Allow navigation only to HTTP(S) URLs. Default true. */
49
49
  readonly httpOnly?: boolean
50
50
  /**
51
- * Directory `browser_download` save paths must resolve inside (prevents a
51
+ * Directory `dsh_browser_download` save paths must resolve inside (prevents a
52
52
  * prompt-injected agent from writing arbitrary machine paths). Default:
53
53
  * the user's Downloads folder; override to confine downloads elsewhere.
54
54
  */
@@ -1,6 +1,6 @@
1
1
  /**
2
2
  * Persistent browsing history: which pages the shared browser actually visited,
3
- * kept independently from any session's operation log (`browser_history`) and
3
+ * kept independently from any session's operation log (`dsh_browser_history`) and
4
4
  * from the browser process's lifetime.
5
5
  *
6
6
  * It lives beside the browser profile (`$DSH_HOME/dsh-builtin-browser-host/
@@ -41,7 +41,7 @@ import { dirname, join } from 'node:path'
41
41
  // several Electron instances sharing Roaming\Electron fight over the GPU
42
42
  // cache/session locks, which can leave the window without a display surface
43
43
  // (capturePage then fails). A dedicated userData also persists cookies across
44
- // host restarts (on top of browser_auth). Must run before app is ready.
44
+ // host restarts (on top of dsh_browser_auth). Must run before app is ready.
45
45
  try {
46
46
  const base = process.env.DSH_HOME ?? app.getPath('appData')
47
47
  app.setPath('userData', join(base, 'dsh-builtin-browser-host'))
@@ -769,7 +769,7 @@ async function handle(op: string, msg: { id: number; viewId?: string; windowId?:
769
769
  const entry = win?.views.get(viewId)
770
770
  if (win === undefined || entry === undefined) throw new Error(`focus: unknown view ${viewId}`)
771
771
  // Keys go to the view that holds web focus. A view that was never
772
- // clicked holds none, so the FIRST browser_key of a session was dropped
772
+ // clicked holds none, so the FIRST dsh_browser_key of a session was dropped
773
773
  // by the renderer while CDP still answered the dispatch with success —
774
774
  // and `document.hasFocus()` inside the page even read true.
775
775
  const contents = entry.webContentsView.webContents
@@ -881,7 +881,7 @@ async function handle(op: string, msg: { id: number; viewId?: string; windowId?:
881
881
  throw new Error('capture produced no image (view not painted)')
882
882
  }
883
883
  // The measured size travels with the image. Without it the native path — which is the
884
- // default self-hosted one — could never answer `browser_screenshot`'s width/height,
884
+ // default self-hosted one — could never answer `dsh_browser_screenshot`'s width/height,
885
885
  // because the seam's only channel was {base64, mime} and the provider had nothing to
886
886
  // forward. The caller's `maxWidth`/`maxHeight` are applied here, so this is the only
887
887
  // place that knows what came back.
@@ -393,7 +393,7 @@ export interface ElectronBrowserProviderConfig {
393
393
  */
394
394
  readonly contentMaxChars?: number
395
395
  /**
396
- * Directory `browser_download` save paths must resolve inside (prevents a
396
+ * Directory `dsh_browser_download` save paths must resolve inside (prevents a
397
397
  * prompt-injected agent from writing arbitrary machine paths). Default:
398
398
  * `~/Downloads`. NOTE: this default is a convenience, not a security
399
399
  * boundary — on Windows it may resolve to OneDrive/redirected paths;
@@ -477,7 +477,7 @@ export function layoutSize(metrics: Record<string, unknown>): { width: number; h
477
477
  }
478
478
 
479
479
  /**
480
- * Map CDP cookies onto the shape `browser_auth` exports.
480
+ * Map CDP cookies onto the shape `dsh_browser_auth` exports.
481
481
  *
482
482
  * CDP reports a cookie as domain + path rather than a URL, and in seconds rather than
483
483
  * milliseconds, so both are converted. Entries missing a name, value or domain are
@@ -955,9 +955,9 @@ export class ElectronBrowserProvider implements BrowserProvider {
955
955
  /**
956
956
  * Page.navigate resolves at navigation COMMIT, not at load: the new document
957
957
  * is already current (so location.href and document.title are the new page's)
958
- * while its DOM is still being parsed. browser_open snapshots immediately
958
+ * while its DOM is still being parsed. dsh_browser_open snapshots immediately
959
959
  * after navigating, which is why it could report the right title with zero
960
- * interactive elements — and why a separate browser_snapshot right after
960
+ * interactive elements — and why a separate dsh_browser_snapshot right after
961
961
  * always found them.
962
962
  *
963
963
  * Wait — bounded, best-effort — for the new document to settle: readyState
@@ -1077,7 +1077,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
1077
1077
  *
1078
1078
  * A renderer only delivers key events to the view that holds web focus. A
1079
1079
  * view that was created but never clicked holds none — and `Input` events
1080
- * injected over CDP do not grant it — so the FIRST `browser_key` of a fresh
1080
+ * injected over CDP do not grant it — so the FIRST `dsh_browser_key` of a fresh
1081
1081
  * session vanished inside the renderer while the command still answered
1082
1082
  * success. Hosts that cannot focus a view (a test double, a different shell
1083
1083
  * adapter) simply have no `focus`, and nothing changes for them.
@@ -1086,7 +1086,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
1086
1086
  if (handle.focus === undefined) return
1087
1087
  try {
1088
1088
  // Bounded, unlike the rest of this call: the child RPC has no timeout of its own,
1089
- // and this runs before typing, so an unresponsive child would hang browser_key for
1089
+ // and this runs before typing, so an unresponsive child would hang dsh_browser_key for
1090
1090
  // good. A timeout is caught below exactly like any other focus failure, so the
1091
1091
  // best-effort contract is unchanged.
1092
1092
  await withTimeout(handle.focus(), 5_000, undefined, 'browser: focusing the view timed out')
@@ -1328,12 +1328,12 @@ export class ElectronBrowserProvider implements BrowserProvider {
1328
1328
  // A timeout leaves the wedged view in the session, so every later read would pay the same
1329
1329
  // 30 seconds again — and `terminatePage` only interrupts the running script, it does not
1330
1330
  // make the view readable. Replacing it is what turns a permanent loss into one bad call:
1331
- // the tab keeps the id the caller holds, and the next browser_open with the real URL
1331
+ // the tab keeps the id the caller holds, and the next dsh_browser_open with the real URL
1332
1332
  // repopulates it.
1333
1333
  if (error instanceof Error && error.name === 'TimeoutError') {
1334
1334
  this.replaceTabWithBlank(s, tab)
1335
1335
  throw new BrowserError(
1336
- 'browser: snapshot timed out; the stuck tab was reset to blank. Call browser_open with the real target URL before reading it again.',
1336
+ 'browser: snapshot timed out; the stuck tab was reset to blank. Call dsh_browser_open with the real target URL before reading it again.',
1337
1337
  'BROWSER_SNAPSHOT_TIMEOUT',
1338
1338
  { cause: error },
1339
1339
  )
@@ -1536,7 +1536,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
1536
1536
  if (error instanceof Error && error.name === 'TimeoutError') {
1537
1537
  this.replaceTabWithBlank(s, tab)
1538
1538
  throw new BrowserError(
1539
- 'browser: a11y timed out; the stuck tab was reset to blank. Call browser_open with the real target URL before reading it again.',
1539
+ 'browser: a11y timed out; the stuck tab was reset to blank. Call dsh_browser_open with the real target URL before reading it again.',
1540
1540
  'BROWSER_A11Y_TIMEOUT',
1541
1541
  { cause: error },
1542
1542
  )
@@ -2003,7 +2003,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
2003
2003
  placeholder: f.placeholder ?? null,
2004
2004
  // A field with NO kind is a field with no kind FILTER: absent, not 'text'. Defaulting
2005
2005
  // it to 'text' here made `if (spec.kind)` above always true, so every unspecified
2006
- // field was narrowed to an <input> — browser_fill on a <textarea> reported
2006
+ // field was narrowed to an <input> — dsh_browser_fill on a <textarea> reported
2007
2007
  // `no "text" control matched`, and a contenteditable could not match any kind at all.
2008
2008
  // The value stays a string or null (never undefined): the in-page check tests truthiness,
2009
2009
  // and JSON drops an undefined member entirely, which is the same thing here.
@@ -2138,7 +2138,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
2138
2138
  el.dispatchEvent(new Event('change', { bubbles: true }))
2139
2139
  out.push({ ok: true, method: 'select', target: describe(spec) })
2140
2140
  } else if (type === 'file') {
2141
- out.push({ ok: false, error: 'file inputs cannot be set from script; use browser_download or ask the human', target: describe(spec) })
2141
+ out.push({ ok: false, error: 'file inputs cannot be set from script; use dsh_browser_download or ask the human', target: describe(spec) })
2142
2142
  } else if (type === 'checkbox') {
2143
2143
  const want = spec.value === true || spec.value === 'true' || spec.value === 'on'
2144
2144
  if (el.checked !== want) el.click()
@@ -2592,7 +2592,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
2592
2592
  * Refuse an action the settings have switched off.
2593
2593
  *
2594
2594
  * These switches belong to the OPERATOR. The settings document is written through the
2595
- * plugin's own panel and no tool can reach it, so — unlike `browser_restrict`, which the
2595
+ * plugin's own panel and no tool can reach it, so — unlike `dsh_browser_restrict`, which the
2596
2596
  * model owns and can lift at will — a refusal here stands for as long as the setting does.
2597
2597
  * That is the point of them: a deployment can take page-script execution, downloads, or
2598
2598
  * login-state writes off the table without relying on the model's cooperation.
@@ -2619,7 +2619,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
2619
2619
 
2620
2620
  /**
2621
2621
  * Admit a caller-supplied save path for a file the browser writes to disk.
2622
- * ONE gate for both `browser_download` and `browser_screenshot`: the path
2622
+ * ONE gate for both `dsh_browser_download` and `dsh_browser_screenshot`: the path
2623
2623
  * must be absolute, must resolve inside `downloadDir`, and must not already
2624
2624
  * exist. Without it a prompt-injected agent could write — or silently
2625
2625
  * replace — any file the DSH process can reach, which also escapes the file
@@ -3053,7 +3053,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
3053
3053
  }
3054
3054
 
3055
3055
  /**
3056
- * List persisted visits, newest first — what `browser_visited` reads back.
3056
+ * List persisted visits, newest first — what `dsh_browser_visited` reads back.
3057
3057
  * @param options - result cap and an optional hostname filter.
3058
3058
  * @returns the matching visits, or an empty list when history is disabled.
3059
3059
  */
@@ -3200,7 +3200,7 @@ export class ElectronBrowserProvider implements BrowserProvider {
3200
3200
  *
3201
3201
  * A view whose renderer never commits cannot be read again: every evaluate into it runs out its
3202
3202
  * whole budget. Replacing it is what turns that from permanent into one bad call — the tab id
3203
- * callers hold stays valid, and the next `browser_open` with the real URL re-populates it.
3203
+ * callers hold stays valid, and the next `dsh_browser_open` with the real URL re-populates it.
3204
3204
  *
3205
3205
  * The old view is destroyed AFTER the new one is in place, so a failure here cannot leave the
3206
3206
  * session holding a handle that is already gone.
@@ -1327,7 +1327,7 @@ class DeferredRemoteView implements ElectronViewHandle {
1327
1327
  * with "browser host is not running". Instead: drop the cache, rebuild the
1328
1328
  * view against a freshly spawned child, and retry the operation exactly
1329
1329
  * once — so sessions that survive a host death self-heal on their FIRST
1330
- * call after it, without a manual browser_reset_session. Page-level errors
1330
+ * call after it, without a manual dsh_browser_reset_session. Page-level errors
1331
1331
  * are never retried (the view itself is fine).
1332
1332
  */
1333
1333
  private async withRecovery<T>(op: (view: RemoteView) => Promise<T>): Promise<T> {
@@ -55,15 +55,15 @@ export interface BrowserSettings {
55
55
  *
56
56
  * These are OPERATOR switches: the settings panel is not reachable from the model, so a
57
57
  * value here refuses whatever the model asks for, and no tool can lift it. That is what
58
- * separates them from `browser_restrict`, which is the model's own soft guardrail. All
58
+ * separates them from `dsh_browser_restrict`, which is the model's own soft guardrail. All
59
59
  * three default to on, which is the documented out-of-the-box behaviour.
60
60
  */
61
61
  readonly actions: {
62
- /** Run arbitrary page scripts in the active tab (`browser_execute`). */
62
+ /** Run arbitrary page scripts in the active tab (`dsh_browser_execute`). */
63
63
  readonly allowExecute: boolean
64
- /** Download a URL to disk (`browser_download`). */
64
+ /** Download a URL to disk (`dsh_browser_download`). */
65
65
  readonly allowDownload: boolean
66
- /** Write cookies — importing login state (`browser_auth` restore). */
66
+ /** Write cookies — importing login state (`dsh_browser_auth` restore). */
67
67
  readonly allowCredentialWrite: boolean
68
68
  }
69
69
  }
@@ -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
 
package/src/index.ts CHANGED
@@ -3,7 +3,7 @@
3
3
  * pieces. The cordis.patch.yml rows reference subpath exports:
4
4
  * - `dsh-builtin-browser/browser` -> the ctx.browser seam (Service)
5
5
  * - `dsh-builtin-browser/browser-electron` -> the Electron CDP provider
6
- * - `dsh-builtin-browser/tool-browser` -> the model-facing browser_* tools
6
+ * - `dsh-builtin-browser/tool-browser` -> the model-facing dsh_browser_* tools
7
7
  * The capability itself is composed by those rows; this root entry exists for
8
8
  * programmatic imports AND because the host's client-module scan identifies a
9
9
  * package's browser half through a loader row whose specifier resolves to the