feat(window): support multiple application windows for multi-monitor workspaces

src/main/agentManagement/AgentManagementBridge.test.ts

@@ -4,10 +4,16 @@ import type { ManagedAgentRendererDescriptor } from '@mcp/shared/agentManagement
 const sentRendererRequests: unknown[] = []
 const resolveProviderTranscriptPath = vi.fn(async () => '/tmp/provider-agent.jsonl')
 
-vi.mock('@main/window/mainWindow.js', () => ({
-  sendToMainWindow: (_channel: string, request: unknown) => {
+// See the note in OrchestrationBridge.test.ts: routing is stubbed to one
+// always-resolvable window so these tests can be about the bridge's own
+// serialization and inventory behavior.
+const sessionWindowOwner = vi.fn((_sessionId: string): string | null => 'test-window')
+
+vi.mock('@main/window/windowRegistry.js', () => ({
+  sendToWindow: (_windowId: string, _channel: string, request: unknown) => {
     sentRendererRequests.push(request)
   },
+  windowForSession: (sessionId: string) => sessionWindowOwner(sessionId),
 }))
 
 vi.mock('@main/providerSwitch/shared.js', () => ({

src/main/agentManagement/AgentManagementBridge.ts

@@ -1,7 +1,7 @@
 import { randomUUID } from 'node:crypto'
 import { stat } from 'node:fs/promises'
 
-import { sendToMainWindow } from '@main/window/mainWindow.js'
+import { sendToWindow, windowForSession } from '@main/window/windowRegistry.js'
 import {
   findCodexRolloutPathsBySessionIds,
   resolveProviderTranscriptPath,
@@ -366,7 +366,20 @@ export class AgentManagementBridge {
       // read-only: each crosses React's renderer-owned workspace model and bulk
       // reads may page durable history. Backpressure here prevents an MCP burst
       // from racing workspace mutations or starving the renderer event loop.
-      sendToMainWindow('agent-management:request', request)
+      //
+      // WHY the caller's window is resolved here and a miss THROWS: the request
+      // is answered from one window's workspace inventory, and this tool's whole
+      // contract is that it reports the caller's project. Throwing inside this
+      // try reuses the existing failure path (timer cleared, pending dropped,
+      // slot released), which is exactly the fail-closed behavior the timeout
+      // comment above describes.
+      const target = windowForSession(request.callerSessionId)
+      if (!target) {
+        throw new Error(
+          `No Agent Code window owns agent-management caller session ${request.callerSessionId}`,
+        )
+      }
+      sendToWindow(target, 'agent-management:request', request)
     } catch (error) {
       clearTimeout(timer)
       this.pending.delete(request.requestId)

src/main/dictation/hotkey.ts

@@ -1,12 +1,23 @@
 import { globalShortcut } from 'electron'
 
-import { sendToMainWindow } from '@main/window/mainWindow.js'
+import { sendToFocusedWindow } from '@main/window/windowRegistry.js'
 import { startMacDictationHotkeyHelper, stopMacDictationHotkeyHelper } from '@main/dictation/macHotkeyHelper.js'
 
 let currentBinding = ''
 let registeredElectronHotkey = ''
 let electronHotkeyRecording = false
 
+// WHY every edge here goes to the FOCUSED window rather than a fixed one:
+// the dictation hotkey is a global accelerator, so it fires no matter which
+// window the user is in — and with several windows open, "start listening" can
+// only sensibly mean the composer they are looking at. `sendToFocusedWindow`
+// falls back to the most recently focused window, which matters because a
+// global shortcut can arrive while Agent Code is not even frontmost, and
+// `BrowserWindow.getFocusedWindow()` returns null in exactly that case.
+//
+// Down and up edges are NOT re-resolved independently: see the release note on
+// `releaseElectronHotkeyRecording`.
+
 /**
  * Electron exposes an accelerator activation callback, not a physical key-up
  * callback. Pretending one activation was both down and up made every quiet
@@ -23,7 +34,7 @@ let electronHotkeyRecording = false
 function releaseElectronHotkeyRecording(binding: string): void {
   if (!electronHotkeyRecording) return
   electronHotkeyRecording = false
-  sendToMainWindow('dictation:hotkey-up', { binding })
+  sendToFocusedWindow('dictation:hotkey-up', { binding })
 }
 
 // WHY a binding-classifier lives here (packaged-app fix):
@@ -154,7 +165,7 @@ export async function configureDictationHotkey(binding: string): Promise<{
           return
         }
         electronHotkeyRecording = true
-        sendToMainWindow('dictation:hotkey-down', { binding: currentBinding })
+        sendToFocusedWindow('dictation:hotkey-down', { binding: currentBinding })
       })
       if (ok) registeredElectronHotkey = classification.accelerator
       return { ok, binding: currentBinding, native: false }
@@ -187,12 +198,12 @@ export async function configureDictationHotkey(binding: string): Promise<{
         // lifecycle problems without guessing from Deepgram logs.
         // eslint-disable-next-line no-console
         console.debug('[dictation:hotkey] down', { binding: currentBinding, at: Date.now() })
-        sendToMainWindow('dictation:hotkey-down', { binding: currentBinding })
+        sendToFocusedWindow('dictation:hotkey-down', { binding: currentBinding })
       },
       onRelease: () => {
         // eslint-disable-next-line no-console
         console.debug('[dictation:hotkey] up', { binding: currentBinding, at: Date.now() })
-        sendToMainWindow('dictation:hotkey-up', { binding: currentBinding })
+        sendToFocusedWindow('dictation:hotkey-up', { binding: currentBinding })
       },
     })
     return {

src/main/incident/installWindowIncidentHooks.ts

@@ -4,7 +4,7 @@ import { execFile } from 'node:child_process'
 import { monitorEventLoopDelay } from 'node:perf_hooks'
 
 import type { AppRunJournal } from '@main/incident/AppRunJournal.js'
-import { getOutboundIpcDiagnostics } from '@main/window/mainWindow.js'
+import { getOutboundIpcDiagnostics } from '@main/window/windowRegistry.js'
 import type { RendererFreezeHeartbeat } from '@shared/incident/rendererFreeze.js'
 
 const HEARTBEAT_STALL_MS = 4_000