fix(desktop-browser): ask only about the user's own on-screen page, never the agent's

Author: waleedlatif1

apps/desktop/e2e/browser-page-dialogs.spec.ts

@@ -27,6 +27,7 @@ const FORM_FIXTURE = `<!doctype html><title>form</title>
 type Bridge = typeof globalThis & {
   simDesktop: SimDesktopApi
   pageDialog?: BrowserPageDialog | null
+  boundsTimer?: number
 }
 
 /**
@@ -79,7 +80,7 @@ test('page dialogs wait for the user on their page and stay automatic for the ag
     response.end(
       !isSite
         ? SHELL_FIXTURE
-        : path === '/form'
+        : path === '/form' || path === '/agent'
           ? FORM_FIXTURE
           : '<!doctype html><title>next</title>'
     )
@@ -117,7 +118,7 @@ test('page dialogs wait for the user on their page and stay automatic for the ag
           scope
         )
       updateBounds()
-      window.setInterval(updateBounds, 200)
+      bridge.boundsTimer = window.setInterval(updateBounds, 200)
       bridge.pageDialog = null
       api.onPageState((state) => {
         bridge.pageDialog = state.dialog ?? null
@@ -260,6 +261,53 @@ test('page dialogs wait for the user on their page and stay automatic for the ag
       await expect.poll(pageUrl).toBe(`${site}/next`)
       expect(await pageDialog()).toBeNull()
     })
+    /** Opens a page-initiated confirm (no user gesture) on the tab at `path`. */
+    const confirmFromPage = (path: string) =>
+      shellApp.evaluate(({ webContents }, url) => {
+        const contents = webContents.getAllWebContents().find((c) => c.getURL() === url)
+        if (!contents) throw new Error(`No page at ${url}`)
+        void contents.executeJavaScript(
+          "setTimeout(() => { document.title = 'confirm:' + confirm('Continue?') })"
+        )
+      }, `${site}${path}`)
+    const titleAt = (path: string) =>
+      shellApp.evaluate(
+        ({ webContents }, url) =>
+          webContents
+            .getAllWebContents()
+            .find((c) => c.getURL() === url)
+            ?.getTitle() ?? null,
+        `${site}${path}`
+      )
+
+    await check(
+      "a dialog on the agent's page stays automatic until the user takes it",
+      async () => {
+        await execute('browser_open_tab', { url: `${site}/agent` })
+        const tabs = await shell.evaluate(
+          async (scope) =>
+            (await (globalThis as Bridge).simDesktop.browserAgent.activateScope(scope)).tabs,
+          SCOPE
+        )
+        const agentTabId = tabs.find((tab) => tab.url === `${site}/agent`)?.tabId
+        // The resource strip mirrors the agent's tab on screen without the user claiming it.
+        await panelAction({ action: 'switch-tab', tabId: agentTabId, claim: false })
+        await confirmFromPage('/agent')
+        await expect.poll(() => titleAt('/agent')).toBe('confirm:false')
+        expect(await pageDialog()).toBeNull()
+      }
+    )
+
+    await check('a dialog while the browser is off screen stays automatic', async () => {
+      await shell.evaluate((scope) => {
+        const bridge = globalThis as Bridge
+        window.clearInterval(bridge.boundsTimer)
+        bridge.simDesktop.browserAgent.setPanelBounds(null, null, scope)
+      }, SCOPE)
+      await confirmFromPage('/agent')
+      await expect.poll(() => titleAt('/agent')).toBe('confirm:false')
+      expect(await pageDialog()).toBeNull()
+    })
     passed = true
   } finally {
     mkdirSync(dirname(reportPath), { recursive: true })

apps/desktop/src/main/browser-agent/session.ts

@@ -1881,13 +1881,18 @@ export function enablePageDialogs(): void {
 
 /**
  * Whether a dialog on this tab is the user's to answer: the renderer can show
- * it, the page is the one on screen, and no agent action is driving it.
+ * it, the page is on screen, and it is the user's page rather than the agent's.
+ * Everything else keeps the shell answering, so the user is never asked about
+ * work they did not start and a hidden page is never left blocked.
  */
 function userOwnsPageDialogs(tab: AgentTab): boolean {
   const contents = tab.view.webContents
   if (!currentScope.pageDialogsEnabled || contents.isDestroyed()) return false
-  if (tab.id !== currentScope.activeTabId || isDispatchingAgentInput(contents)) return false
-  return !(currentScope.automationActive && automationTab()?.id === tab.id)
+  if (tab.id !== currentScope.activeTabId || !isPanelVisible()) return false
+  if (getBrowserScopeId() !== getActiveBrowserScopeId()) return false
+  if (isDispatchingAgentInput(contents)) return false
+  if (automationTab()?.id !== tab.id) return true
+  return automationTabClaimedByUser() && !currentScope.automationActive
 }
 
 function holdPageDialogForUser(
@@ -3450,8 +3455,11 @@ export function switchTab(tabId: string, { claim = true }: { claim?: boolean } =
     revokeTabMediaPermissions(previousActiveTab, false)
     previousActiveTab.pendingUserFocus = false
   }
-  currentScope.activeTabId = tab.id
+  // The claim describes the page on screen, so a mirrored switch to another
+  // page leaves that page unclaimed rather than inheriting the last one's.
   if (claim) currentScope.visibleTabUserSelected = true
+  else if (currentScope.activeTabId !== tab.id) currentScope.visibleTabUserSelected = false
+  currentScope.activeTabId = tab.id
   promotePendingTabRestore(tab)
   // Visible selection does not move the automation exemption; the user may
   // inspect another page while a tool continues in its background tab.