asi se fue xd

.claude-ejecutor/CLAUDE.md

@@ -8,7 +8,7 @@ Eres un agente que habita Agent UI. El canvas es tu espacio — no solo una herr
 1. **SIEMPRE** responde usando `bubbleResponse` - nunca respondas con texto plano
 2. **SOLO** puedes usar herramientas MCP de `agent-ui`
 3. **NUNCA** intentes usar terminal, bash, curl, o cualquier comando del sistema
-4. **NUNCA** intentes leer, escribir o editar archivos
+4. **NUNCA** intentes leer, escribir o editar archivos (los .vue de user-components/ los gestiona Claude Code, no vos)
 5. Tu propósito es crear, manipular y dar vida a la interfaz gráfica
 
 ---
@@ -33,8 +33,10 @@ El canvas tiene 3 niveles de contenido que coexisten:
    - Ideal para fondos animados (cámara pixelada, matrix rain, etc.)
    - Los scripts corren independientes de las ventanas
 
-2. **Ventanas Flotantes** — `render_vue_component` / `load_vue_component`
+2. **Ventanas Flotantes** — `render_vue_component` / `load_fs_component`
    - Componentes Vue 3 completos en ventanas Liquid Glass
+   - `render_vue_component` — inline (definición en el mismo tool call)
+   - `load_fs_component` — desde archivo .vue en user-components/
    - Drag, resize, close
    - Cada una tiene su propio ciclo de vida (onMounted/onUnmounted)
 
@@ -54,10 +56,17 @@ El canvas tiene 3 niveles de contenido que coexisten:
 - `list_windows` → `move_window` → `resize_window` → `close_window`
 - `inspect_window` — Leer HTML interno de una ventana
 
-**Persistencia:**
+**Componentes Filesystem (user-components/):**
-- `save_vue_component` / `load_vue_component` — Guardar componentes individuales en SQLite
+- `list_fs_components` — Lista componentes .vue disponibles en user-components/
+- `load_fs_component` — Carga y renderiza un componente desde su carpeta
+- Los componentes viven como archivos `.vue` reales en `user-components/<folder>/`
+- Convención: `user-components/mi-componente/MiComponente.vue` + opcional `meta.json`
+- Claude Code crea/edita los `.vue` con Write/Read/Edit (NO se usa SQLite)
+- File watcher detecta cambios en tiempo real vía WebSocket
+
+**Snapshots:**
 - `save_canvas_snapshot` / `load_canvas_snapshot` — Guardar el estado COMPLETO del canvas
-- `list_canvas_snapshots` / `list_vue_components` — Listar lo guardado
+- `list_canvas_snapshots` — Listar snapshots guardados
 
 **Edición:**
 - `edit_canvas` — Editar DOM in-place (selector + old_value → new_value)
@@ -116,7 +125,25 @@ El snapshot captura: HTML base + CSS blocks + script log + ventanas (posición,
 
 ---
 
-## Componentes Guardados (mi galería)
+## Componentes en Filesystem (user-components/)
+
+Los componentes ya NO se guardan en SQLite. Viven como archivos `.vue` reales que Claude Code gestiona con Write/Read/Edit.
+
+**Estructura:**
+```
+user-components/
+  mi-componente/
+    MiComponente.vue    ← <template> + <script setup> + <style>
+    meta.json           ← opcional: { name, tags, props, imports }
+```
+
+**Importante sobre <script setup>:**
+- El código setup se ejecuta via `new Function()`, NO es SFC real
+- Debe hacer `return { var1, var2 }` explícitamente
+- Los imports de Vue (ref, reactive, computed, etc.) se inyectan automáticamente
+- NO usar `import` statements — usar los helpers globales ($emit, $on, $fetch, $theme)
+
+**Componentes legacy en DB** (accesibles pero ya no se crean nuevos):
 
 | ID | Nombre | Qué hace |
 |---|---|---|

.claude-ejecutor/plugins/known_marketplaces.json

@@ -1,10 +0,0 @@
-{
-  "claude-plugins-official": {
-    "source": {
-      "source": "github",
-      "repo": "anthropics/claude-plugins-official"
-    },
-    "installLocation": "C:\\Users\\jodar\\agent-ui\\.claude-ejecutor\\plugins\\marketplaces\\claude-plugins-official",
-    "lastUpdated": "2026-02-16T06:32:07.237Z"
-  }
-}

.claude-ejecutor/stats-cache.json

@@ -0,0 +1,86 @@
+{
+  "version": 2,
+  "lastComputedDate": "2026-02-17",
+  "dailyActivity": [
+    {
+      "date": "2026-02-15",
+      "messageCount": 2052,
+      "sessionCount": 9,
+      "toolCallCount": 262
+    },
+    {
+      "date": "2026-02-16",
+      "messageCount": 787,
+      "sessionCount": 4,
+      "toolCallCount": 83
+    },
+    {
+      "date": "2026-02-17",
+      "messageCount": 1154,
+      "sessionCount": 1,
+      "toolCallCount": 123
+    }
+  ],
+  "dailyModelTokens": [
+    {
+      "date": "2026-02-15",
+      "tokensByModel": {
+        "claude-opus-4-5-20251101": 3247,
+        "claude-opus-4-6": 81887
+      }
+    },
+    {
+      "date": "2026-02-16",
+      "tokensByModel": {
+        "claude-opus-4-6": 25122
+      }
+    },
+    {
+      "date": "2026-02-17",
+      "tokensByModel": {
+        "claude-opus-4-6": 36622
+      }
+    }
+  ],
+  "modelUsage": {
+    "claude-opus-4-5-20251101": {
+      "inputTokens": 196,
+      "outputTokens": 3051,
+      "cacheReadInputTokens": 314084,
+      "cacheCreationInputTokens": 35936,
+      "webSearchRequests": 0,
+      "costUSD": 0,
+      "contextWindow": 0,
+      "maxOutputTokens": 0
+    },
+    "claude-opus-4-6": {
+      "inputTokens": 1708,
+      "outputTokens": 141923,
+      "cacheReadInputTokens": 43414737,
+      "cacheCreationInputTokens": 7323135,
+      "webSearchRequests": 0,
+      "costUSD": 0,
+      "contextWindow": 0,
+      "maxOutputTokens": 0
+    }
+  },
+  "totalSessions": 14,
+  "totalMessages": 3993,
+  "longestSession": {
+    "sessionId": "b1715c14-9ef8-4b54-9fda-d281c55c2a07",
+    "duration": 84183755,
+    "messageCount": 408,
+    "timestamp": "2026-02-16T08:26:52.205Z"
+  },
+  "firstSessionDate": "2026-02-15T00:55:54.803Z",
+  "hourCounts": {
+    "0": 2,
+    "2": 3,
+    "13": 4,
+    "18": 1,
+    "19": 1,
+    "20": 1,
+    "23": 2
+  },
+  "totalSpeculationTimeSavedMs": 0
+}

.claude/settings.local.json

@@ -83,7 +83,11 @@
       "mcp__agent-ui__z590_nucleoriofrio_com-list_canvas_snapshots",
       "mcp__agent-ui__z590_nucleoriofrio_com-list_canvases",
       "mcp__agent-ui__z590_nucleoriofrio_com-list_vue_components",
-      "Bash(jq:*)"
+      "Bash(jq:*)",
+      "mcp__agent-ui__z590_nucleoriofrio_com-read_component",
+      "mcp__agent-ui__z590_nucleoriofrio_com-edit_component",
+      "mcp__agent-ui__z590_nucleoriofrio_com-list_fs_components",
+      "mcp__agent-ui__z590_nucleoriofrio_com-load_fs_component"
     ]
   },
   "enableAllProjectMcpServers": true,

frontend/package-lock.json

@@ -11,6 +11,7 @@
         "@nucleoriofrio/webmcp": "git+https://gitea.nucleoriofrio.com/nucleo000/webmcp.git",
         "@xterm/addon-fit": "^0.11.0",
         "@xterm/addon-web-links": "^0.12.0",
+        "@xterm/addon-webgl": "^0.19.0",
         "@xterm/xterm": "^6.0.0",
         "pinia": "^3.0.4",
         "vite-plugin-pwa": "^1.2.0",
@@ -2624,6 +2625,12 @@
       "integrity": "sha512-4Smom3RPyVp7ZMYOYDoC/9eGJJJqYhnPLGGqJ6wOBfB8VxPViJNSKdgRYb8NpaM6YSelEKbA2SStD7lGyqaobw==",
       "license": "MIT"
     },
+    "node_modules/@xterm/addon-webgl": {
+      "version": "0.19.0",
+      "resolved": "https://registry.npmjs.org/@xterm/addon-webgl/-/addon-webgl-0.19.0.tgz",
+      "integrity": "sha512-b3fMOsyLVuCeNJWxolACEUED0vm7qC0cy4wRvf3oURSzDTYVQiGPhTnhWZwIHdvC48Y+oLhvYXnY4XDXPoJo6A==",
+      "license": "MIT"
+    },
     "node_modules/@xterm/xterm": {
       "version": "6.0.0",
       "resolved": "https://registry.npmjs.org/@xterm/xterm/-/xterm-6.0.0.tgz",

frontend/package.json

@@ -14,6 +14,7 @@
     "@nucleoriofrio/webmcp": "git+https://gitea.nucleoriofrio.com/nucleo000/webmcp.git",
     "@xterm/addon-fit": "^0.11.0",
     "@xterm/addon-web-links": "^0.12.0",
+    "@xterm/addon-webgl": "^0.19.0",
     "@xterm/xterm": "^6.0.0",
     "pinia": "^3.0.4",
     "vite-plugin-pwa": "^1.2.0",

frontend/src/App.vue

@@ -7,8 +7,6 @@ import FloatingTerminal from './components/FloatingTerminal.vue'
 import FloatingResponse from './components/FloatingResponse.vue'
 import FloatingVoice from './components/FloatingVoice.vue'
 import AgentBar from './components/AgentBar.vue'
-import HookNotifications from './components/HookNotifications.vue'
-import NotificationLog from './components/NotificationLog.vue'
 import PwaInstallBanner from './components/PwaInstallBanner.vue'
 import { initWebMCP, getWebMCP } from './services/webmcp'
 import { initTorch, destroyTorch } from './services/torch'
@@ -18,7 +16,6 @@ import { setTerminalControls } from './services/tools/handlers/terminalHandlers'
 import { setResponseControls } from './services/tools/handlers/responseHandlers'
 import { useCanvasStore } from './stores/canvas'
 import { useProjectCanvasStore } from './stores/projectCanvas'
-import { useClaudeHooksStore } from './stores/claude-hooks'
 
 const route = useRoute()
 const router = useRouter()
@@ -68,12 +65,9 @@ function clearDebugLogs() {
 }
 const terminalRef = ref<InstanceType<typeof FloatingTerminal> | null>(null)
 const responseRef = ref<InstanceType<typeof FloatingResponse> | null>(null)
-const notifLogRef = ref<InstanceType<typeof NotificationLog> | null>(null)
 const voiceRef = ref<InstanceType<typeof FloatingVoice> | null>(null)
 const canvasStore = useCanvasStore()
 const projectCanvasStore = useProjectCanvasStore()
-const hooksStore = useClaudeHooksStore()
-
 // Voice FAB push-to-talk state
 const voicePTTActive = ref(false)
 let voiceTouchStarted = false
@@ -237,15 +231,6 @@ function connectStatusWs() {
         }
       }
 
-        // Rich hook data → toast notifications
-        if (msg.type === 'claude-hook') {
-          hooksStore.processHook(msg)
-        }
-
-        // Permission request → persistent toast with allow/deny
-        if (msg.type === 'claude-permission') {
-          hooksStore.processPermission(msg)
-        }
     } catch { /* ignore non-JSON messages */ }
   }
 
@@ -332,8 +317,6 @@ onMounted(async () => {
   // Setup response controls for MCP tools
   setResponseControls({
     addMessage: (message: string, type?: 'info' | 'success' | 'warning' | 'error') => {
-      // Also log to notification log
-      notifLogRef.value?.addResponseEntry(message, type || 'info')
       if (responseRef.value) {
         return responseRef.value.addMessage(message, type)
       }
@@ -545,11 +528,6 @@ watch(() => route.name, (newPage) => {
     <!-- Floating Response (Agent UI messages) -->
     <FloatingResponse ref="responseRef" />
 
-    <!-- Hook Notifications (toasts from Claude Code hooks) -->
-    <HookNotifications />
-
-    <!-- Notification Log (temporary - collects all notifications, persists to localStorage) -->
-    <NotificationLog ref="notifLogRef" />
 
     <!-- Floating Voice Input -->
     <FloatingVoice ref="voiceRef" v-model="showVoice" />

frontend/src/components/Canvas.vue

@@ -99,7 +99,7 @@ onUnmounted(() => {
   flex: 1;
   position: relative;
   min-height: 100%;
-  overflow: hidden;
+  overflow: auto;
 }
 
 .canvas-placeholder {

frontend/src/components/agent/InputSettings.vue

@@ -68,31 +68,16 @@ function formatSessionLabel(s: SessionInfo): string {
       </select>
     </div>
 
-    <!-- Voice Mode Toggle -->
+    <!-- Whisper Status -->
     <div class="is-section">
       <label class="is-label">Mode</label>
       <div class="is-mode-row">
-        <button
+        <div
-          class="is-mode-btn"
+          class="is-mode-btn active"
-          :class="{ active: voice.voiceMode.value === 'webspeech' }"
-          :disabled="voice.isRecording.value"
-          @click="voice.voiceMode.value !== 'webspeech' && voice.toggleWhisperMode()"
-        >
-          <span class="is-mode-icon">Web</span>
-          <span class="is-mode-label">Speech API</span>
-        </button>
-        <button
-          class="is-mode-btn"
-          :class="{
-            active: voice.voiceMode.value === 'whisper',
-            loading: voice.whisperStatus.value === 'loading'
-          }"
-          :disabled="voice.isRecording.value"
-          @click="voice.voiceMode.value !== 'whisper' && voice.toggleWhisperMode()"
         >
           <span class="is-mode-icon">GPU</span>
           <span class="is-mode-label">Whisper</span>
-        </button>
+        </div>
       </div>
       <div class="is-status">
         <span
@@ -104,10 +89,7 @@ function formatSessionLabel(s: SessionInfo): string {
           }"
         ></span>
         <span class="is-status-text">
-          {{ voice.voiceMode.value === 'whisper'
+          {{ voice.whisperStatus.value === 'ready' ? 'Whisper ready' : voice.whisperStatus.value === 'loading' ? 'Starting...' : 'Offline' }}
-            ? (voice.whisperStatus.value === 'ready' ? 'Whisper ready' : voice.whisperStatus.value === 'loading' ? 'Starting...' : 'Offline')
-            : 'Web Speech API'
-          }}
         </span>
       </div>
     </div>

frontend/src/composables/useVoiceCapture.ts

@@ -1,30 +1,11 @@
 import { ref, watch, type Ref } from 'vue'
-import { endpoints } from '../config/endpoints'
+import {
-
+  initWhisperSocket,
-// Web Speech API types (not in default TS lib)
+  sendAudio,
-interface SpeechRecognitionEvent extends Event {
+  onTranscription,
-  resultIndex: number
+  getWhisperStatus,
-  results: SpeechRecognitionResultList
+  isConnected
-}
+} from '../services/whisperSocket'
-
-interface SpeechRecognitionErrorEvent extends Event {
-  error: string
-  message?: string
-}
-
-interface SpeechRecognition extends EventTarget {
-  continuous: boolean
-  interimResults: boolean
-  lang: string
-  onresult: ((event: SpeechRecognitionEvent) => void) | null
-  onerror: ((event: SpeechRecognitionErrorEvent) => void) | null
-  onend: (() => void) | null
-  start(): void
-  stop(): void
-  abort(): void
-}
-
-export type VoiceMode = 'webspeech' | 'whisper'
 export type WhisperStatus = 'offline' | 'loading' | 'ready'
 
 export interface VoiceCapture {
@@ -34,7 +15,7 @@ export interface VoiceCapture {
   interimTranscript: Ref<string>
   animatedTranscript: Ref<string>
   error: Ref<string>
-  voiceMode: Ref<VoiceMode>
+  voiceMode: Ref<'whisper'>
   whisperStatus: Ref<WhisperStatus>
   audioDevices: Ref<MediaDeviceInfo[]>
   selectedDeviceId: Ref<string>
@@ -44,9 +25,7 @@ export interface VoiceCapture {
   // Actions
   startRecording: () => void
   stopRecording: () => void
-  toggleWhisperMode: () => Promise<void>
+  loadAudioDevices: (skipPermission?: boolean) => Promise<void>
-  checkWhisperStatus: () => Promise<any>
-  loadAudioDevices: () => Promise<void>
   selectMicrophone: (deviceId: string) => void
   playLastAudio: () => void
   init: () => Promise<void>
@@ -54,6 +33,8 @@ export interface VoiceCapture {
   clearTranscript: () => void
 }
 
+const GPU_TIMEOUT_MS = 30_000 // 30s timeout waiting for GPU
+
 export function useVoiceCapture(options?: {
   onNotification?: (message: string, type: 'info' | 'success' | 'error', duration?: number) => void
 }): VoiceCapture {
@@ -65,290 +46,97 @@ export function useVoiceCapture(options?: {
   const interimTranscript = ref('')
   const animatedTranscript = ref('')
   const error = ref('')
-  const voiceMode = ref<VoiceMode>('webspeech')
+  const voiceMode = ref<'whisper'>('whisper') // Always whisper, no web speech
-  const whisperStatus = ref<WhisperStatus>('offline')
   const audioDevices = ref<MediaDeviceInfo[]>([])
   const selectedDeviceId = ref('')
   const isAndroid = ref(false)
-
-  // Audio debug & save
   const lastAudioUrl = ref('')
   const isPlayingAudio = ref(false)
 
-  // ====== Internal state ======
+  // ====== Internal ======
-  let recognition: SpeechRecognition | null = null
+  const sharedWhisperStatus = getWhisperStatus()
-  let lastProcessedResult = ''
+  const whisperStatus = ref<WhisperStatus>(sharedWhisperStatus.value)
-
-  // Typing animation
-  let typingTimeout: number | null = null
-  let lastAnimatedLength = 0
-
-  // Whisper
-  const WHISPER_WS_URL = endpoints.whisper
-  let whisperSocket: WebSocket | null = null
   let mediaRecorder: MediaRecorder | null = null
   let audioChunks: Blob[] = []
   let chunkInterval: number | null = null
   const CHUNK_INTERVAL_MS = 3000
   let mediaStream: MediaStream | null = null
   let supportedMimeType = 'audio/webm;codecs=opus'
-
-  // Audio playback debug
   let audioElement: HTMLAudioElement | null = null
   let recordingStartTime = 0
+  let unsubTranscription: (() => void) | null = null
+  let gpuTimeout: number | null = null
+
+  // Typing animation
+  let typingTimeout: number | null = null
+  let lastAnimatedLength = 0
+
+  // Keep local status in sync with shared
+  watch(sharedWhisperStatus, (val) => {
+    whisperStatus.value = val
+  })
 
   // ====== Mobile / Audio Format ======
 
   function checkMobile() {
-    const ua = navigator.userAgent
+    isAndroid.value = /Android/i.test(navigator.userAgent)
-    isAndroid.value = /Android/i.test(ua)
   }
 
   function detectAudioFormat(): string {
     const formats = [
-      'audio/webm;codecs=opus',
+      'audio/webm;codecs=opus', 'audio/webm',
-      'audio/webm',
+      'audio/mp4', 'audio/mp4;codecs=mp4a.40.2',
-      'audio/mp4',
+      'audio/aac', 'audio/ogg;codecs=opus', 'audio/wav'
-      'audio/mp4;codecs=mp4a.40.2',
-      'audio/aac',
-      'audio/ogg;codecs=opus',
-      'audio/wav'
     ]
-    for (const format of formats) {
+    for (const f of formats) {
-      if (MediaRecorder.isTypeSupported(format)) {
+      if (MediaRecorder.isTypeSupported(f)) {
-        console.log(`[VoiceCapture] Using audio format: ${format}`)
+        console.log(`[VoiceCapture] Audio format: ${f}`)
-        return format
+        return f
       }
     }
-    console.warn('[VoiceCapture] No preferred format supported, using default')
     return ''
   }
 
-  // ====== Web Speech API ======
+  // ====== Whisper transcription handler ======
 
-  function initRecognition(): SpeechRecognition | null {
+  function handleTranscription(msg: any) {
-    const SpeechRecognitionCtor = (window as any).SpeechRecognition || (window as any).webkitSpeechRecognition
+    if (!isRecording.value) return
-    if (!SpeechRecognitionCtor) {
-      error.value = 'Speech recognition not supported in this browser'
-      return null
-    }
 
-    const rec: SpeechRecognition = new SpeechRecognitionCtor()
+    if (msg.success && msg.text) {
-    rec.continuous = !isAndroid.value
+      const fullText = msg.text.trim()
-    rec.interimResults = true
+      transcript.value = fullText + ' '
-    rec.lang = 'es-419'
+      interimTranscript.value = ''
-
+      if (!msg.partial) {
-    if (isAndroid.value) {
+        console.log(`[VoiceCapture] WHISPER (${msg.model}/${msg.device}):`, fullText)
-      console.log('[VoiceCapture] Android detected - using non-continuous mode')
-    }
-
-    rec.onresult = (event: SpeechRecognitionEvent) => {
-      let interim = ''
-      let final = ''
-
-      for (let i = event.resultIndex; i < event.results.length; i++) {
-        const result = event.results[i]
-        if (!result || !result[0]) continue
-        if (result.isFinal) {
-          final += result[0].transcript + ' '
-        } else {
-          interim += result[0].transcript
-        }
       }
-
+    } else if (msg.error) {
-      if (final) {
+      error.value = msg.error
-        const trimmedFinal = final.trim()
+      console.error('[VoiceCapture] Whisper error:', msg.error)
-        if (isAndroid.value && lastProcessedResult && trimmedFinal.startsWith(lastProcessedResult.trim())) {
-          const newPart = trimmedFinal.slice(lastProcessedResult.trim().length).trim()
-          if (newPart) {
-            transcript.value += newPart + ' '
-            lastProcessedResult = trimmedFinal
-          }
-        } else {
-          transcript.value += final
-          lastProcessedResult = trimmedFinal
-        }
-      }
-      interimTranscript.value = interim
-    }
-
-    rec.onerror = (event: SpeechRecognitionErrorEvent) => {
-      // no-speech and aborted are transient — don't kill the session
-      if (event.error === 'no-speech' || event.error === 'aborted') {
-        console.log('[VoiceCapture] Transient error:', event.error, '(will auto-restart)')
-        return
-      }
-      console.error('[VoiceCapture] Recognition error:', event.error)
-      if (event.error === 'not-allowed') {
-        error.value = 'Microphone access denied'
-      } else {
-        error.value = `Error: ${event.error}`
-      }
-      isRecording.value = false
-    }
-
-    rec.onend = () => {
-      if (isRecording.value && voiceMode.value === 'webspeech') {
-        if (isAndroid.value) {
-          isRecording.value = false
-          console.log('[VoiceCapture] Android session ended - tap mic to continue')
-        } else {
-          rec.start()
-        }
-      }
-    }
-
-    return rec
-  }
-
-  // ====== Whisper Functions ======
-
-  async function checkWhisperStatusFn(updateLoading = true): Promise<any> {
-    try {
-      const res = await fetch('/api/whisper/status')
-      const data = await res.json()
-      if (data.enabled) {
-        voiceMode.value = 'whisper'
-      }
-      if (data.running) {
-        whisperStatus.value = 'ready'
-      } else if (updateLoading && (data.starting || false)) {
-        whisperStatus.value = 'loading'
-      } else if (!data.running) {
-        if (voiceMode.value === 'whisper' && !data.starting) {
-          whisperStatus.value = 'offline'
-        }
-      }
-      return data
-    } catch {
-      voiceMode.value = 'webspeech'
-      whisperStatus.value = 'offline'
-      return null
     }
   }
 
-  async function pollWhisperStatus(): Promise<void> {
+  // ====== Recording ======
-    const maxAttempts = 60
-    let attempts = 0
 
-    while (attempts < maxAttempts) {
+  function startRecording() {
-      await new Promise(resolve => setTimeout(resolve, 2000))
+    error.value = ''
-      attempts++
 
-      try {
+    // Start capturing audio immediately, regardless of GPU status
-        const status = await checkWhisperStatusFn(false)
+    startMediaRecorder()
-        if (!status) continue
 
-        if (status.starting) {
+    // If GPU not ready yet, start timeout
-          console.log(`[VoiceCapture] Still starting... (${attempts * 2}s)`)
+    if (!isConnected()) {
-          continue
+      console.log('[VoiceCapture] Recording started, waiting for GPU...')
+      gpuTimeout = window.setTimeout(() => {
+        if (isRecording.value && !isConnected()) {
+          error.value = 'Whisper GPU timeout — server not available'
+          notify('Whisper GPU not available', 'error')
+          stopRecording()
         }
-
+      }, GPU_TIMEOUT_MS)
-        if (status.running && status.enabled) {
-          console.log('[VoiceCapture] Server ready!')
-          notify('Whisper GPU ready!', 'success')
-          connectWhisperSocket()
-          whisperStatus.value = 'ready'
-          return
-        }
-
-        console.log('[VoiceCapture] Server failed to start')
-        notify('Whisper server failed to start', 'error')
-        whisperStatus.value = 'offline'
-        return
-      } catch (e) {
-        console.error('[VoiceCapture] Polling error:', e)
-      }
-    }
-
-    notify('Whisper server timeout', 'error')
-    whisperStatus.value = 'offline'
-  }
-
-  function connectWhisperSocket() {
-    if (whisperStatus.value !== 'ready') {
-      console.log('[VoiceCapture] Whisper not ready, skipping connection')
-      return
-    }
-    if (whisperSocket?.readyState === WebSocket.OPEN) return
-
-    console.log('[VoiceCapture] Connecting to Whisper at:', WHISPER_WS_URL)
-    whisperSocket = new WebSocket(WHISPER_WS_URL)
-
-    const connectionTimeout = setTimeout(() => {
-      if (whisperSocket && whisperSocket.readyState !== WebSocket.OPEN) {
-        console.error('[VoiceCapture] Whisper connection timeout (10s)')
-        whisperSocket.close()
-        whisperStatus.value = 'offline'
-      }
-    }, 10000)
-
-    whisperSocket.onopen = () => {
-      clearTimeout(connectionTimeout)
-      console.log('[VoiceCapture] Whisper WebSocket connected')
-      whisperStatus.value = 'ready'
-    }
-
-    whisperSocket.onmessage = (event) => {
-      try {
-        const msg = JSON.parse(event.data)
-
-        if (msg.type === 'ready') {
-          console.log('[VoiceCapture] Whisper ready:', msg.model, msg.device)
-          whisperStatus.value = 'ready'
-        } else if (msg.type === 'transcription') {
-          if (msg.success && msg.text) {
-            const fullText = msg.text.trim()
-            if (msg.partial) {
-              transcript.value = fullText + ' '
-              interimTranscript.value = ''
-            } else {
-              transcript.value = fullText + ' '
-              interimTranscript.value = ''
-              console.log(`[VoiceCapture] WHISPER-GPU (${msg.model}/${msg.device}):`, fullText)
-            }
-          } else if (msg.error) {
-            error.value = msg.error
-            console.error('[VoiceCapture] Whisper error:', msg.error)
-          }
-        }
-      } catch (e) {
-        console.error('[VoiceCapture] Whisper message error:', e)
-      }
-    }
-
-    whisperSocket.onclose = () => {
-      console.log('[VoiceCapture] Whisper WebSocket closed')
-      whisperStatus.value = 'offline'
-    }
-
-    whisperSocket.onerror = (e) => {
-      console.error('[VoiceCapture] Whisper WebSocket error:', e)
-      whisperStatus.value = 'offline'
     }
   }
 
-  function disconnectWhisperSocket() {
+  async function startMediaRecorder() {
-    if (whisperSocket) {
-      whisperSocket.close()
-      whisperSocket = null
-    }
-    whisperStatus.value = 'offline'
-  }
-
-  async function startWhisperRecording() {
-    if (!whisperSocket || whisperSocket.readyState !== WebSocket.OPEN) {
-      console.warn('[VoiceCapture] Whisper socket not connected, attempting to connect...')
-      connectWhisperSocket()
-      await new Promise(resolve => setTimeout(resolve, 500))
-
-      if (!whisperSocket || whisperSocket.readyState !== WebSocket.OPEN) {
-        error.value = 'Whisper server not connected'
-        notify('Whisper not connected. Try toggling GPU mode.', 'error')
-        return
-      }
-    }
-
     try {
       const audioConstraints: MediaTrackConstraints = {
         echoCancellation: true,
@@ -375,17 +163,21 @@ export function useVoiceCapture(options?: {
         }
       }
 
-      audioChunks = []
       mediaRecorder.start(100)
       isRecording.value = true
       recordingStartTime = Date.now()
-      console.log(`[VoiceCapture] Whisper recording started`)
 
-      // Permission granted via user gesture — reload devices with labels
+      // Reload devices with labels now that we have permission
       loadAudioDevices(true)
 
+      // Send chunks periodically — only when GPU is connected
       chunkInterval = window.setInterval(() => {
-        if (audioChunks.length > 0 && whisperSocket?.readyState === WebSocket.OPEN) {
+        if (audioChunks.length > 0 && isConnected()) {
+          // GPU came online — clear timeout if still pending
+          if (gpuTimeout) {
+            clearTimeout(gpuTimeout)
+            gpuTimeout = null
+          }
           sendAudioChunk(false)
         }
       }, CHUNK_INTERVAL_MS)
@@ -397,6 +189,10 @@ export function useVoiceCapture(options?: {
 
   function sendAudioChunk(isFinal: boolean) {
     if (audioChunks.length === 0) return
+    if (!isConnected()) {
+      console.log('[VoiceCapture] GPU not connected, holding audio')
+      return
+    }
 
     const mimeType = mediaRecorder?.mimeType || supportedMimeType || 'audio/webm'
     const audioBlob = new Blob(audioChunks, { type: mimeType })
@@ -414,24 +210,22 @@ export function useVoiceCapture(options?: {
     const reader = new FileReader()
     reader.onloadend = () => {
       const base64 = (reader.result as string).split(',')[1]
-      if (whisperSocket?.readyState === WebSocket.OPEN) {
+      sendAudio(base64, 'es', !isFinal)
-        whisperSocket.send(JSON.stringify({
-          type: 'transcribe',
-          audio: base64,
-          language: 'es',
-          partial: !isFinal
-        }))
-      }
     }
     reader.readAsDataURL(audioBlob)
   }
 
-  function stopWhisperRecording() {
+  function stopRecording() {
+    if (gpuTimeout) {
+      clearTimeout(gpuTimeout)
+      gpuTimeout = null
+    }
     if (chunkInterval) {
       clearInterval(chunkInterval)
       chunkInterval = null
     }
 
+    // Send final chunk (only if GPU is connected)
     if (audioChunks.length > 0) {
       sendAudioChunk(true)
     }
@@ -439,16 +233,16 @@ export function useVoiceCapture(options?: {
     if (mediaRecorder && mediaRecorder.state !== 'inactive') {
       mediaRecorder.stop()
     }
-
     if (mediaStream) {
       mediaStream.getTracks().forEach(track => track.stop())
       mediaStream = null
     }
 
     isRecording.value = false
+    interimTranscript.value = ''
   }
 
-  // ====== Audio Save & Debug Playback ======
+  // ====== Audio Save & Playback ======
 
   function currentMicName(): string {
     if (!selectedDeviceId.value) return 'Default'
@@ -457,9 +251,7 @@ export function useVoiceCapture(options?: {
   }
 
   function saveAudioForPlayback(blob: Blob) {
-    if (lastAudioUrl.value) {
+    if (lastAudioUrl.value) URL.revokeObjectURL(lastAudioUrl.value)
-      URL.revokeObjectURL(lastAudioUrl.value)
-    }
     lastAudioUrl.value = URL.createObjectURL(blob)
     saveRecordingToBackend(blob)
   }
@@ -468,7 +260,6 @@ export function useVoiceCapture(options?: {
     try {
       const duration_ms = Date.now() - recordingStartTime
       const reader = new FileReader()
-
       reader.onloadend = async () => {
         const base64 = (reader.result as string).split(',')[1]
         const response = await fetch('/api/recordings', {
@@ -483,12 +274,9 @@ export function useVoiceCapture(options?: {
         })
         const data = await response.json()
         if (data.success) {
-          console.log(`[VoiceCapture] Recording saved: ${data.filename} (${(data.size / 1024).toFixed(1)} KB)`)
+          console.log(`[VoiceCapture] Recording saved: ${data.filename}`)
-        } else {
-          console.error('[VoiceCapture] Failed to save recording:', data.error)
         }
       }
-
       reader.readAsDataURL(blob)
     } catch (e) {
       console.error('[VoiceCapture] Error saving recording:', e)
@@ -497,163 +285,17 @@ export function useVoiceCapture(options?: {
 
   function playLastAudio() {
     if (!lastAudioUrl.value) return
-
     if (isPlayingAudio.value && audioElement) {
       audioElement.pause()
       audioElement.currentTime = 0
       isPlayingAudio.value = false
       return
     }
-
     audioElement = new Audio(lastAudioUrl.value)
     audioElement.onplay = () => { isPlayingAudio.value = true }
     audioElement.onended = () => { isPlayingAudio.value = false }
     audioElement.onpause = () => { isPlayingAudio.value = false }
-    audioElement.play().catch(e => {
+    audioElement.play().catch(() => { isPlayingAudio.value = false })
-      console.error('[VoiceCapture] Failed to play audio:', e)
-      isPlayingAudio.value = false
-    })
-  }
-
-  // ====== Parallel Audio Capture (for Web Speech mode) ======
-
-  async function startAudioCapture() {
-    try {
-      const audioConstraints: MediaTrackConstraints = {
-        echoCancellation: true,
-        noiseSuppression: true,
-        autoGainControl: true,
-        ...(selectedDeviceId.value ? { deviceId: { exact: selectedDeviceId.value } } : {})
-      }
-
-      mediaStream = await navigator.mediaDevices.getUserMedia({ audio: audioConstraints })
-
-      const recorderOptions: MediaRecorderOptions = {}
-      if (supportedMimeType) {
-        recorderOptions.mimeType = supportedMimeType
-      }
-
-      mediaRecorder = new MediaRecorder(mediaStream, recorderOptions)
-      audioChunks = []
-
-      mediaRecorder.ondataavailable = (event) => {
-        if (event.data.size > 0) {
-          audioChunks.push(event.data)
-        }
-      }
-
-      mediaRecorder.start(100)
-      recordingStartTime = Date.now()
-      console.log(`[VoiceCapture] Audio capture started (${mediaRecorder.mimeType})`)
-
-      // Permission is now granted via user gesture — reload devices with labels
-      loadAudioDevices(true)
-    } catch (e: any) {
-      console.error('[VoiceCapture] Audio capture error:', e)
-    }
-  }
-
-  function stopAudioCapture() {
-    if (mediaRecorder && mediaRecorder.state !== 'inactive') {
-      mediaRecorder.stop()
-    }
-
-    // Build final blob and save
-    if (audioChunks.length > 0) {
-      const mimeType = mediaRecorder?.mimeType || supportedMimeType || 'audio/webm'
-      const audioBlob = new Blob(audioChunks, { type: mimeType })
-      audioChunks = []
-      if (audioBlob.size > 1000) {
-        saveAudioForPlayback(audioBlob)
-      }
-    }
-
-    if (mediaStream) {
-      mediaStream.getTracks().forEach(track => track.stop())
-      mediaStream = null
-    }
-  }
-
-  // ====== Public Recording API ======
-
-  function startRecording() {
-    error.value = ''
-
-    if (voiceMode.value === 'whisper' && whisperStatus.value === 'ready') {
-      startWhisperRecording()
-    } else {
-      if (!recognition) {
-        recognition = initRecognition()
-      }
-      if (recognition) {
-        try {
-          recognition.start()
-          isRecording.value = true
-          // Capture raw audio in parallel for save/debug playback
-          startAudioCapture()
-          if (isAndroid.value) {
-            notify('Android: Tap mic again to continue recording', 'info', 3000)
-          }
-        } catch (e) {
-          console.error('[VoiceCapture] Failed to start:', e)
-        }
-      }
-    }
-  }
-
-  function stopRecording() {
-    if (voiceMode.value === 'whisper') {
-      stopWhisperRecording()
-    } else {
-      if (recognition) {
-        recognition.stop()
-      }
-      stopAudioCapture()
-      isRecording.value = false
-    }
-    interimTranscript.value = ''
-  }
-
-  async function toggleWhisperMode() {
-    if (whisperStatus.value === 'loading') return
-
-    whisperStatus.value = 'loading'
-    error.value = ''
-
-    if (voiceMode.value !== 'whisper') {
-      notify('Starting Whisper GPU server...', 'info', 10000)
-    }
-
-    try {
-      const res = await fetch('/api/whisper/toggle', { method: 'POST' })
-      const data = await res.json()
-
-      if (data.starting) {
-        console.log('[VoiceCapture] Server starting, polling...')
-        voiceMode.value = 'whisper'
-        await pollWhisperStatus()
-        return
-      }
-
-      if (data.enabled) {
-        voiceMode.value = 'whisper'
-        whisperStatus.value = data.running ? 'ready' : 'offline'
-        if (data.running) {
-          notify('Whisper GPU ready!', 'success')
-          connectWhisperSocket()
-        }
-      } else {
-        voiceMode.value = 'webspeech'
-        whisperStatus.value = 'offline'
-        notify('Using Web Speech API', 'info')
-        disconnectWhisperSocket()
-      }
-    } catch (e: any) {
-      error.value = 'Failed to toggle Whisper'
-      notify('Error starting Whisper server', 'error')
-      console.error('[VoiceCapture] Whisper toggle error:', e)
-      whisperStatus.value = 'offline'
-    }
   }
 
   // ====== Microphone ======
@@ -661,18 +303,14 @@ export function useVoiceCapture(options?: {
   async function loadAudioDevices(skipPermissionRequest = false) {
     try {
       if (!skipPermissionRequest) {
-        // Request permission to get device labels
         const tempStream = await navigator.mediaDevices.getUserMedia({ audio: true })
         tempStream.getTracks().forEach(track => track.stop())
       }
-
       const devices = await navigator.mediaDevices.enumerateDevices()
       audioDevices.value = devices.filter(d => d.kind === 'audioinput')
-
       if (!selectedDeviceId.value && audioDevices.value.length > 0) {
         selectedDeviceId.value = audioDevices.value[0]?.deviceId || ''
       }
-      console.log(`[VoiceCapture] Found ${audioDevices.value.length} audio devices`)
     } catch (e) {
       console.error('[VoiceCapture] Failed to enumerate devices:', e)
     }
@@ -689,36 +327,26 @@ export function useVoiceCapture(options?: {
   // ====== Typing Animation ======
 
   function animateTyping(targetText: string) {
-    if (typingTimeout) {
+    if (typingTimeout) { clearTimeout(typingTimeout); typingTimeout = null }
-      clearTimeout(typingTimeout)
-      typingTimeout = null
-    }
-
     if (targetText.length < animatedTranscript.value.length) {
       animatedTranscript.value = targetText
       lastAnimatedLength = targetText.length
       return
     }
-
     const startIndex = lastAnimatedLength
-
     function typeNext(index: number) {
       if (index <= targetText.length) {
         animatedTranscript.value = targetText.substring(0, index)
         lastAnimatedLength = index
         if (index < targetText.length) {
-          const delay = 15 + Math.random() * 10
+          typingTimeout = window.setTimeout(() => typeNext(index + 1), 15 + Math.random() * 10)
-          typingTimeout = window.setTimeout(() => typeNext(index + 1), delay)
         }
       }
     }
-
     typeNext(startIndex)
   }
 
-  watch(transcript, (newVal) => {
+  watch(transcript, (v) => animateTyping(v))
-    animateTyping(newVal)
-  })
 
   // ====== Transcript ======
 
@@ -727,80 +355,43 @@ export function useVoiceCapture(options?: {
     interimTranscript.value = ''
     animatedTranscript.value = ''
     lastAnimatedLength = 0
-    lastProcessedResult = ''
+    if (typingTimeout) { clearTimeout(typingTimeout); typingTimeout = null }
-    if (typingTimeout) {
-      clearTimeout(typingTimeout)
-      typingTimeout = null
-    }
   }
 
   // ====== Lifecycle ======
 
   async function init() {
-    recognition = initRecognition()
     checkMobile()
     supportedMimeType = detectAudioFormat()
-    // Only enumerate without getUserMedia — no user gesture here
-    // Devices will get full labels after first recording (user gesture)
     await loadAudioDevices(true)
 
-    const status = await checkWhisperStatusFn()
+    // Subscribe to shared whisper transcriptions
-    if (status?.starting) {
+    if (!unsubTranscription) {
-      console.log('[VoiceCapture] Server is starting, resuming polling...')
+      unsubTranscription = onTranscription(handleTranscription)
-      pollWhisperStatus()
-    } else if (voiceMode.value === 'whisper' && whisperStatus.value === 'ready') {
-      connectWhisperSocket()
-    } else if (voiceMode.value === 'whisper' && whisperStatus.value !== 'ready') {
-      console.log('[VoiceCapture] Whisper was enabled but server not running, disabling')
-      voiceMode.value = 'webspeech'
     }
+
+    // Initialize shared Whisper socket (singleton, safe to call multiple times)
+    initWhisperSocket()
+    console.log('[VoiceCapture] Initialized (Whisper-only, record-first)')
   }
 
   function cleanup() {
     stopRecording()
-    recognition = null
+    if (unsubTranscription) { unsubTranscription(); unsubTranscription = null }
-    disconnectWhisperSocket()
     if (chunkInterval) clearInterval(chunkInterval)
     if (typingTimeout) clearTimeout(typingTimeout)
-    if (mediaStream) {
+    if (gpuTimeout) clearTimeout(gpuTimeout)
-      mediaStream.getTracks().forEach(track => track.stop())
+    if (mediaStream) { mediaStream.getTracks().forEach(t => t.stop()); mediaStream = null }
-      mediaStream = null
+    if (audioElement) { audioElement.pause(); audioElement = null }
-    }
+    if (lastAudioUrl.value) { URL.revokeObjectURL(lastAudioUrl.value); lastAudioUrl.value = '' }
-    if (audioElement) {
-      audioElement.pause()
-      audioElement = null
-    }
-    if (lastAudioUrl.value) {
-      URL.revokeObjectURL(lastAudioUrl.value)
-      lastAudioUrl.value = ''
-    }
     isPlayingAudio.value = false
   }
 
   return {
-    // State
+    isRecording, transcript, interimTranscript, animatedTranscript,
-    isRecording,
+    error, voiceMode, whisperStatus, audioDevices, selectedDeviceId,
-    transcript,
+    isAndroid, lastAudioUrl, isPlayingAudio,
-    interimTranscript,
+    startRecording, stopRecording, loadAudioDevices, selectMicrophone,
-    animatedTranscript,
+    playLastAudio, init, cleanup, clearTranscript
-    error,
-    voiceMode,
-    whisperStatus,
-    audioDevices,
-    selectedDeviceId,
-    isAndroid,
-    lastAudioUrl,
-    isPlayingAudio,
-    // Actions
-    startRecording,
-    stopRecording,
-    toggleWhisperMode,
-    checkWhisperStatus: checkWhisperStatusFn,
-    loadAudioDevices,
-    selectMicrophone,
-    playLastAudio,
-    init,
-    cleanup,
-    clearTranscript
   }
 }

frontend/src/services/tools/handlers/componentHandlers.ts

@@ -71,11 +71,15 @@ export function createComponentHandlers(): ToolConfig[] {
         properties: {
           id: { type: 'string', description: 'ID del componente' },
           componentProps: { type: 'object', description: 'Props para el componente' },
-          mode: { type: 'string', enum: ['replace', 'append'], description: 'Modo' }
+          mode: { type: 'string', enum: ['replace', 'append'], description: 'Modo' },
+          x: { type: 'number', description: 'Posicion X inicial' },
+          y: { type: 'number', description: 'Posicion Y inicial' },
+          width: { type: 'number', description: 'Ancho inicial' },
+          height: { type: 'number', description: 'Alto inicial' }
         },
         required: ['id']
       },
-      handler: async (args: { id: string; componentProps?: Record<string, any>; mode?: string }) => {
+      handler: async (args: { id: string; componentProps?: Record<string, any>; mode?: string; x?: number; y?: number; width?: number; height?: number }) => {
         try {
           const definition = await componentsApi.getById(args.id)
           if (!definition) {
@@ -88,7 +92,8 @@ export function createComponentHandlers(): ToolConfig[] {
           removePlaceholder(container)
 
           const isAppend = args.mode === 'append'
-          const result = renderInlineComponent(definition, container, args.componentProps || {}, isAppend)
+          const layout = { x: args.x, y: args.y, width: args.width, height: args.height }
+          const result = renderInlineComponent(definition, container, args.componentProps || {}, isAppend, layout)
 
           // Track definition for snapshot capture
           getWindowDefinitions().set(definition.id, {
@@ -168,11 +173,13 @@ export function createComponentHandlers(): ToolConfig[] {
             type: 'array',
             items: { type: 'string', enum: ['template', 'setup', 'style', 'props', 'imports'] },
             description: 'Campos a leer (default: template, setup, style)'
-          }
+          },
+          offset: { type: 'number', description: 'Linea inicial (1-based)' },
+          limit: { type: 'number', description: 'Numero de lineas a leer' }
         },
         required: ['id']
       },
-      handler: async (args: { id: string; fields?: string[] }) => {
+      handler: async (args: { id: string; fields?: string[]; offset?: number; limit?: number }) => {
         try {
           const definition = await componentsApi.getById(args.id)
           if (!definition) return `Error: "${args.id}" not found`
@@ -191,7 +198,14 @@ export function createComponentHandlers(): ToolConfig[] {
               output.push(`--- ${field} ---\n${arr?.length ? JSON.stringify(arr) : '(empty)'}`)
             } else {
               const str = (value as string) || ''
-              output.push(`--- ${field} (${str.length}) ---\n${str || '(empty)'}`)
+              const lines = str.split('\n')
+              const total = lines.length
+              const start = Math.max(0, (args.offset || 1) - 1)
+              const end = args.limit ? start + args.limit : total
+              const sliced = lines.slice(start, end)
+              const numbered = sliced.map((l, i) => `${String(start + i + 1).padStart(4)}\t${l}`).join('\n')
+              const rangeInfo = (args.offset || args.limit) ? ` lines ${start + 1}-${Math.min(end, total)}/${total}` : ` ${total} lines`
+              output.push(`--- ${field}${rangeInfo} ---\n${numbered || '(empty)'}`)
             }
           }
 

frontend/src/services/whisperSocket.ts

@@ -0,0 +1,177 @@
+/**
+ * Singleton Whisper WebSocket Service
+ * One shared connection used by all voice components (FloatingVoice, useVoiceCapture, etc.)
+ */
+
+import { ref } from 'vue'
+import { endpoints } from '../config/endpoints'
+
+export type WhisperStatus = 'offline' | 'loading' | 'ready'
+
+type TranscriptionCallback = (msg: {
+  success?: boolean
+  text?: string
+  error?: string
+  partial?: boolean
+  model?: string
+  device?: string
+}) => void
+
+// ====== Singleton state ======
+const status = ref<WhisperStatus>('loading')
+let socket: WebSocket | null = null
+let reconnectTimer: number | null = null
+const listeners = new Set<TranscriptionCallback>()
+
+// ====== Connection management ======
+
+function connect() {
+  if (socket?.readyState === WebSocket.OPEN || socket?.readyState === WebSocket.CONNECTING) return
+
+  console.log('[WhisperSocket] Connecting to', endpoints.whisper)
+  socket = new WebSocket(endpoints.whisper)
+
+  const timeout = setTimeout(() => {
+    if (socket && socket.readyState !== WebSocket.OPEN) {
+      console.error('[WhisperSocket] Connection timeout (10s)')
+      socket.close()
+      status.value = 'loading'
+    }
+  }, 10000)
+
+  socket.onopen = () => {
+    clearTimeout(timeout)
+    console.log('[WhisperSocket] Connected')
+    status.value = 'ready'
+  }
+
+  socket.onmessage = (event) => {
+    try {
+      const msg = JSON.parse(event.data)
+
+      if (msg.type === 'ready') {
+        console.log('[WhisperSocket] Server ready:', msg.model, msg.device)
+        status.value = 'ready'
+      } else if (msg.type === 'transcription') {
+        // Broadcast to all listeners
+        for (const cb of listeners) {
+          cb(msg)
+        }
+      }
+    } catch (e) {
+      console.error('[WhisperSocket] Message parse error:', e)
+    }
+  }
+
+  socket.onclose = () => {
+    console.log('[WhisperSocket] Closed, will reconnect...')
+    socket = null
+    status.value = 'loading'
+    scheduleReconnect()
+  }
+
+  socket.onerror = (e) => {
+    console.error('[WhisperSocket] Error:', e)
+    status.value = 'loading'
+  }
+}
+
+function scheduleReconnect() {
+  if (reconnectTimer) return
+  reconnectTimer = window.setTimeout(() => {
+    reconnectTimer = null
+    checkStatusAndConnect()
+  }, 2000)
+}
+
+async function checkStatusAndConnect() {
+  try {
+    const res = await fetch('/api/whisper/status')
+    const data = await res.json()
+    if (data.running) {
+      connect()
+    } else {
+      status.value = 'loading'
+      scheduleReconnect()
+    }
+  } catch {
+    status.value = 'loading'
+    scheduleReconnect()
+  }
+}
+
+// ====== Public API ======
+
+/** Initialize the singleton connection (call once at app startup) */
+export function initWhisperSocket() {
+  checkStatusAndConnect()
+}
+
+/** Send audio for transcription */
+export function sendAudio(base64: string, language: string, partial: boolean) {
+  if (socket?.readyState === WebSocket.OPEN) {
+    socket.send(JSON.stringify({
+      type: 'transcribe',
+      audio: base64,
+      language,
+      partial
+    }))
+  } else {
+    console.warn('[WhisperSocket] Not connected, dropping audio')
+  }
+}
+
+/** Subscribe to transcription results. Returns unsubscribe function. */
+export function onTranscription(callback: TranscriptionCallback): () => void {
+  listeners.add(callback)
+  return () => listeners.delete(callback)
+}
+
+/** Get reactive status */
+export function getWhisperStatus() {
+  return status
+}
+
+/** Check if socket is connected */
+export function isConnected(): boolean {
+  return socket?.readyState === WebSocket.OPEN
+}
+
+/** Force reconnect (e.g. when user toggles Whisper) */
+export async function reconnect() {
+  if (status.value === 'loading' && socket?.readyState === WebSocket.CONNECTING) return
+
+  status.value = 'loading'
+  if (socket) {
+    socket.close()
+    socket = null
+  }
+
+  try {
+    const res = await fetch('/api/whisper/toggle', { method: 'POST' })
+    const data = await res.json()
+    if (data.running) {
+      connect()
+    } else {
+      // Poll until ready
+      const poll = async () => {
+        for (let i = 0; i < 60; i++) {
+          await new Promise(r => setTimeout(r, 2000))
+          try {
+            const s = await fetch('/api/whisper/status')
+            const d = await s.json()
+            if (d.running) {
+              connect()
+              return
+            }
+          } catch { /* retry */ }
+        }
+        status.value = 'offline'
+      }
+      poll()
+    }
+  } catch {
+    status.value = 'loading'
+    scheduleReconnect()
+  }
+}

package.json

@@ -3,7 +3,7 @@
   "version": "1.0.0",
   "description": "Dynamic canvas for Claude Code interaction",
   "scripts": {
-    "kill-ports": "node -e \"const {execSync} = require('child_process'); [4101,4102,4103,4105].forEach(p => { try { const pid = execSync('netstat -ano | findstr :' + p + ' | findstr LISTENING', {encoding:'utf8'}).split(/\\s+/).pop().trim(); if(pid) execSync('taskkill /PID ' + pid + ' /F', {stdio:'ignore'}); } catch(e){} }); console.log('Ports cleared');\"",
+    "kill-ports": "node -e \"const {execSync} = require('child_process'); [4101,4102,4103,4105].forEach(p => { try { const pid = execSync('netstat -ano | findstr :' + p + ' | findstr LISTENING', {encoding:'utf8'}).split(/\\s+/).pop().trim(); if(pid) execSync('taskkill /PID ' + pid + ' /F', {stdio:'ignore'}); } catch(e){} }); console.log('Ports cleared (4104/whisper preserved)');\"",
     "start": "bun run kill-ports && concurrently -n api,terminal,frontend -c blue,yellow,green \"cd server && bun --watch run index.ts\" \"cd server && bun run terminal.ts\" \"cd frontend && bun run dev --host\"",
     "start:api": "cd server && bun --watch run index.ts",
     "start:terminal": "cd server && bun run terminal.ts",

server/routes/whisper.ts

@@ -50,7 +50,7 @@ export async function handleWhisperRoutes(req: Request): Promise<Response | null
     return Response.json({
       ...result,
       ...state,
-      message: state.enabled ? 'Whisper enabled (GPU)' : 'Whisper disabled (using Web Speech API)'
+      message: state.running ? 'Whisper GPU running' : 'Whisper GPU starting...'
     })
   }
 

server/services/whisper.ts

@@ -1,6 +1,7 @@
 /**
- * Whisper Service - Manages the Python Whisper server process
+ * Whisper Service - Singleton persistent GPU speech-to-text server
- * Provides GPU-accelerated speech-to-text as an alternative to Web Speech API
+ * Auto-starts with the system, auto-restarts on crash.
+ * Single instance processes all client requests.
  */
 
 import { join } from 'path'
@@ -8,18 +9,19 @@ import { Subprocess } from 'bun'
 
 const WHISPER_PORT = 4104
 const WHISPER_SCRIPT = join(import.meta.dir, '..', 'whisper_server.py')
+const RESTART_DELAY_MS = 3000  // Wait before auto-restart after crash
 
 interface WhisperState {
   enabled: boolean
   running: boolean
-  starting: boolean  // Prevents multiple simultaneous start attempts
+  starting: boolean
   process: Subprocess | null
   model: string
   device: string
 }
 
 const state: WhisperState = {
-  enabled: false,
+  enabled: true,   // Always enabled by default
   running: false,
   starting: false,
   process: null,
@@ -32,104 +34,16 @@ const state: WhisperState = {
  */
 async function killProcessOnPort(port: number): Promise<void> {
   try {
-    // Use PowerShell to find and kill process on port
     const proc = Bun.spawn(['powershell', '-Command',
       `Get-NetTCPConnection -LocalPort ${port} -ErrorAction SilentlyContinue | ForEach-Object { Stop-Process -Id $_.OwningProcess -Force -ErrorAction SilentlyContinue }`
     ], { stdout: 'ignore', stderr: 'ignore' })
     await proc.exited
-    // Wait a moment for port to be released
     await new Promise(resolve => setTimeout(resolve, 1000))
   } catch {
     // Ignore errors
   }
 }
 
-/**
- * Start the Whisper Python server
- */
-export async function startWhisperServer(): Promise<boolean> {
-  // Prevent multiple simultaneous start attempts
-  if (state.starting) {
-    return false
-  }
-
-  if (state.running && state.process) {
-    return true
-  }
-
-  state.starting = true
-  console.log(`[Whisper] Starting (${state.model})...`)
-
-  // Kill any existing process on the port
-  await killProcessOnPort(WHISPER_PORT)
-
-  try {
-    // Use Bun.spawn with inherit to show logs directly in console
-    // -u flag disables Python output buffering for real-time logs
-    const proc = Bun.spawn(['python', '-u', WHISPER_SCRIPT], {
-      cwd: join(import.meta.dir, '..'),
-      stdout: 'inherit',
-      stderr: 'inherit',
-      env: { ...process.env, PYTHONUNBUFFERED: '1' }
-    })
-
-    state.process = proc
-
-    // Wait a bit for the server to start
-    await new Promise(resolve => setTimeout(resolve, 2000))
-
-    // Check if process is still running
-    if (proc.exitCode !== null) {
-      console.error('[Whisper] Process exited with code:', proc.exitCode)
-      state.process = null
-      state.starting = false
-      return false
-    }
-
-    // Check if WebSocket is ready
-    const isListening = await checkPort(WHISPER_PORT)
-
-    if (isListening) {
-      console.log('[Whisper] Ready')
-      state.running = true
-      state.enabled = true
-      state.starting = false
-      return true
-    }
-
-    // Wait more if model is still loading (up to 120 seconds total for large models)
-    for (let i = 0; i < 40; i++) {
-      await new Promise(resolve => setTimeout(resolve, 3000))
-
-      if (proc.exitCode !== null) {
-        console.error('[Whisper] Process died')
-        state.process = null
-        state.starting = false
-        return false
-      }
-
-      const ready = await checkPort(WHISPER_PORT)
-      if (ready) {
-        console.log('[Whisper] Ready')
-        state.running = true
-        state.enabled = true
-        state.starting = false
-        return true
-      }
-    }
-
-    console.error('[Whisper] Timeout (120s)')
-    state.starting = false
-    return false
-
-  } catch (err: any) {
-    console.error('[Whisper] Error:', err.message)
-    state.process = null
-    state.starting = false
-    return false
-  }
-}
-
 /**
  * Check if Whisper WebSocket is ready using PowerShell
  */
@@ -152,7 +66,125 @@ async function checkPort(port: number): Promise<boolean> {
 }
 
 /**
- * Stop the Whisper server
+ * Monitor the Whisper process and auto-restart on crash
+ */
+function monitorProcess(proc: Subprocess) {
+  proc.exited.then((exitCode) => {
+    console.error(`[Whisper] Process exited with code ${exitCode}`)
+    state.process = null
+    state.running = false
+    state.starting = false
+
+    // Auto-restart after delay
+    console.log(`[Whisper] Auto-restarting in ${RESTART_DELAY_MS / 1000}s...`)
+    setTimeout(() => {
+      startWhisperServer().catch(err => {
+        console.error('[Whisper] Auto-restart failed:', err)
+      })
+    }, RESTART_DELAY_MS)
+  })
+}
+
+/**
+ * Start the Whisper Python server (singleton - only one instance)
+ */
+export async function startWhisperServer(): Promise<boolean> {
+  // Prevent multiple simultaneous start attempts
+  if (state.starting) {
+    console.log('[Whisper] Already starting, skipping')
+    return false
+  }
+
+  // Already running
+  if (state.running && state.process) {
+    console.log('[Whisper] Already running')
+    return true
+  }
+
+  // Check if an external instance is already listening
+  const alreadyListening = await checkPort(WHISPER_PORT)
+  if (alreadyListening) {
+    console.log('[Whisper] External instance already running on port', WHISPER_PORT)
+    state.running = true
+    state.enabled = true
+    return true
+  }
+
+  state.starting = true
+  console.log(`[Whisper] Starting singleton server (${state.model})...`)
+
+  // Kill any orphan process on the port
+  await killProcessOnPort(WHISPER_PORT)
+
+  try {
+    const proc = Bun.spawn(['python', '-u', WHISPER_SCRIPT], {
+      cwd: join(import.meta.dir, '..'),
+      stdout: 'inherit',
+      stderr: 'inherit',
+      env: { ...process.env, PYTHONUNBUFFERED: '1' }
+    })
+
+    state.process = proc
+
+    // Monitor for crashes and auto-restart
+    monitorProcess(proc)
+
+    // Wait for initial startup
+    await new Promise(resolve => setTimeout(resolve, 2000))
+
+    // Check if process died immediately
+    if (proc.exitCode !== null) {
+      console.error('[Whisper] Process exited immediately with code:', proc.exitCode)
+      state.process = null
+      state.starting = false
+      return false
+    }
+
+    // Check if WebSocket is ready
+    const isListening = await checkPort(WHISPER_PORT)
+    if (isListening) {
+      console.log('[Whisper] Server ready (GPU)')
+      state.running = true
+      state.enabled = true
+      state.starting = false
+      return true
+    }
+
+    // Wait for model loading (up to 120 seconds for large-v3)
+    for (let i = 0; i < 40; i++) {
+      await new Promise(resolve => setTimeout(resolve, 3000))
+
+      if (proc.exitCode !== null) {
+        console.error('[Whisper] Process died during model loading')
+        state.process = null
+        state.starting = false
+        return false
+      }
+
+      const ready = await checkPort(WHISPER_PORT)
+      if (ready) {
+        console.log('[Whisper] Server ready (GPU)')
+        state.running = true
+        state.enabled = true
+        state.starting = false
+        return true
+      }
+    }
+
+    console.error('[Whisper] Timeout waiting for server (120s)')
+    state.starting = false
+    return false
+
+  } catch (err: any) {
+    console.error('[Whisper] Start error:', err.message)
+    state.process = null
+    state.starting = false
+    return false
+  }
+}
+
+/**
+ * Stop the Whisper server (only for manual override, not used in normal flow)
  */
 export function stopWhisperServer(): boolean {
   if (!state.process) {
@@ -163,8 +195,7 @@ export function stopWhisperServer(): boolean {
     state.process.kill()
     state.process = null
     state.running = false
-    state.enabled = false
+    console.log('[Whisper] Stopped manually')
-    console.log('[Whisper] Stopped')
     return true
   } catch (err) {
     console.error('[Whisper] Stop error:', err)
@@ -173,27 +204,26 @@ export function stopWhisperServer(): boolean {
 }
 
 /**
- * Toggle Whisper server on/off (async - returns immediately when starting)
+ * Toggle is now a no-op for stop - Whisper always stays on.
+ * If not running, triggers a start.
  */
 export async function toggleWhisperServer(): Promise<{ enabled: boolean; success: boolean; starting: boolean }> {
-  // Prevent toggle while starting
   if (state.starting) {
-    return { enabled: false, success: false, starting: true }
+    return { enabled: true, success: false, starting: true }
   }
 
-  if (state.enabled && state.running) {
+  if (state.running) {
-    const success = stopWhisperServer()
+    // Already running - just confirm it's on
-    return { enabled: false, success, starting: false }
+    return { enabled: true, success: true, starting: false }
-  } else {
-    // Start server in background - don't await
-    startWhisperServer().catch(err => {
-      console.error('[Whisper] Start error:', err)
-      state.starting = false
-    })
-
-    // Return immediately - frontend will poll for status
-    return { enabled: false, success: true, starting: true }
   }
+
+  // Not running - start it
+  startWhisperServer().catch(err => {
+    console.error('[Whisper] Start error:', err)
+    state.starting = false
+  })
+
+  return { enabled: true, success: true, starting: true }
 }
 
 /**
@@ -217,13 +247,12 @@ export async function getWhisperState(): Promise<{
       state.enabled = true
     } else if (!isListening && state.running) {
       state.running = false
-      state.enabled = false
+      // Keep enabled=true since we auto-restart
-      state.process = null
     }
   }
 
   return {
-    enabled: state.enabled,
+    enabled: true,  // Always enabled
     running: state.running,
     starting: state.starting,
     port: WHISPER_PORT,
@@ -233,15 +262,12 @@ export async function getWhisperState(): Promise<{
 }
 
 /**
- * Check if Whisper is enabled
+ * Check if Whisper is running
  */
 export function isWhisperEnabled(): boolean {
-  return state.enabled && state.running
+  return state.running
 }
 
-// WebSocket server for Whisper (proxies to Python server or handles directly)
-let whisperWsServer: any = null
-
 export function getWhisperPort(): number {
   return WHISPER_PORT
 }

server/terminal.ts

@@ -3,10 +3,12 @@
  * Terminal Server - Independent process
  * This runs separately from the main server to maintain stable Claude Code sessions
  * even when the main server restarts due to code changes.
+ * Also manages the Whisper GPU server (singleton, persistent).
  */
 
 import { startTerminalServer } from './services/terminal'
 import { startSyncServer } from './services/sync-server'
+import { startWhisperServer } from './services/whisper'
 import { WORKING_DIR } from './config'
 
 console.log('')
@@ -14,6 +16,7 @@ console.log('='.repeat(50))
 console.log('Terminal Server (Independent Process)')
 console.log(`  Terminal WebSocket: ws://localhost:4103`)
 console.log(`  Sync WebSocket (Git + Torch): ws://localhost:4105`)
+console.log(`  Whisper GPU: ws://localhost:4104 (auto-start)`)
 console.log(`  Working Dir: ${WORKING_DIR}`)
 console.log('')
 console.log('This process is stable and won\'t restart')
@@ -23,3 +26,14 @@ console.log('')
 
 startTerminalServer()
 startSyncServer()
+
+// Auto-start Whisper GPU server (singleton, persistent, auto-restart on crash)
+startWhisperServer().then(ok => {
+  if (ok) {
+    console.log('[Whisper] GPU server started successfully')
+  } else {
+    console.warn('[Whisper] Failed initial start (will auto-retry)')
+  }
+}).catch(err => {
+  console.error('[Whisper] Boot error:', err)
+})

server/whisper_server.py

@@ -26,35 +26,25 @@ except ImportError as e:
 def convert_audio_to_wav(input_data: bytes, input_format: str = "webm") -> bytes:
     """
     Convert audio data to WAV format using ffmpeg.
-    Whisper requires WAV/PCM format, but browsers typically record in WebM/Opus.
+    Uses stdin/stdout pipes so ffmpeg probes the actual data format
+    instead of relying on file extensions.
     """
-    # Create temp files for input and output
-    with tempfile.NamedTemporaryFile(suffix=f".{input_format}", delete=False) as in_file:
-        in_file.write(input_data)
-        input_path = in_file.name
-
-    output_path = input_path.replace(f".{input_format}", ".wav")
-
     try:
-        # Use ffmpeg to convert to WAV (16kHz mono, which Whisper prefers)
         result = subprocess.run([
-            "ffmpeg", "-y",  # Overwrite output
+            "ffmpeg", "-y",
-            "-i", input_path,  # Input file
+            "-i", "pipe:0",       # Read from stdin (auto-detect format)
-            "-ar", "16000",  # Sample rate 16kHz
+            "-ar", "16000",       # Sample rate 16kHz
-            "-ac", "1",  # Mono
+            "-ac", "1",           # Mono
             "-c:a", "pcm_s16le",  # PCM 16-bit little-endian
-            output_path
+            "-f", "wav",          # Output format
-        ], capture_output=True, text=True, timeout=30)
+            "pipe:1"              # Write to stdout
+        ], input=input_data, capture_output=True, timeout=30)
 
         if result.returncode != 0:
-            print(f"[Whisper] ffmpeg error: {result.stderr}")
+            print(f"[Whisper] ffmpeg error: {result.stderr.decode('utf-8', errors='replace')}")
             return None
 
-        # Read the converted WAV file
+        return result.stdout
-        with open(output_path, "rb") as f:
-            wav_data = f.read()
-
-        return wav_data
 
     except subprocess.TimeoutExpired:
         print("[Whisper] ffmpeg conversion timed out")
@@ -65,16 +55,6 @@ def convert_audio_to_wav(input_data: bytes, input_format: str = "webm") -> bytes
     except Exception as e:
         print(f"[Whisper] Conversion error: {e}")
         return None
-    finally:
-        # Cleanup temp files
-        try:
-            os.unlink(input_path)
-        except:
-            pass
-        try:
-            os.unlink(output_path)
-        except:
-            pass
 
 # Configuration
 HOST = "0.0.0.0"  # Listen on all interfaces (needed for Traefik proxy)