@tanstack/ai-gemini 0.19.1 → 0.20.1

package/dist/esm/realtime/utils.js

@@ -0,0 +1,350 @@
+const captureWorkletCode = `
+class AudioCaptureProcessor extends AudioWorkletProcessor {
+  constructor() {
+    super();
+    this.bufferSize = 512; // 32ms at 16kHz — per Gemini best practices (20-40ms chunks)
+    this.buffer = new Float32Array(this.bufferSize);
+    this.bufferIndex = 0;
+  }
+
+  process(inputs, outputs, parameters) {
+    const input = inputs[0];
+
+    if (input && input.length > 0) {
+      const inputChannel = input[0];
+
+      // Buffer the incoming audio
+      for (let i = 0; i < inputChannel.length; i++) {
+        this.buffer[this.bufferIndex++] = inputChannel[i];
+
+        // When buffer is full, send it to main thread
+        if (this.bufferIndex >= this.bufferSize) {
+          // Send the buffered audio to the main thread
+          this.port.postMessage({
+            type: "audio",
+            data: this.buffer.slice(),
+          });
+
+          // Reset buffer
+          this.bufferIndex = 0;
+        }
+      }
+    }
+
+    // Return true to keep the processor alive
+    return true;
+  }
+}
+
+// Register the processor
+registerProcessor("audio-capture-processor", AudioCaptureProcessor);`;
+const playbackWorkletCode = `
+class PCMProcessor extends AudioWorkletProcessor {
+  constructor() {
+    super();
+    this.audioQueue = [];
+    this.currentOffset = 0; // Track position in current buffer (avoids slice())
+
+    this.port.onmessage = (event) => {
+      if (event.data === "interrupt") {
+        // Clear the queue on interrupt
+        this.audioQueue = [];
+        this.currentOffset = 0;
+      } else if (event.data instanceof Float32Array) {
+        // Add audio data to the queue
+        this.audioQueue.push(event.data);
+      }
+    };
+  }
+
+  process(inputs, outputs, parameters) {
+    const output = outputs[0];
+    if (output.length === 0) return true;
+
+    const channel = output[0];
+    let outputIndex = 0;
+
+    // Fill the output buffer from the queue
+    while (outputIndex < channel.length && this.audioQueue.length > 0) {
+      const currentBuffer = this.audioQueue[0];
+
+      if (!currentBuffer || currentBuffer.length === 0) {
+        this.audioQueue.shift();
+        this.currentOffset = 0;
+        continue;
+      }
+
+      const remainingOutput = channel.length - outputIndex;
+      const remainingBuffer = currentBuffer.length - this.currentOffset;
+      const copyLength = Math.min(remainingOutput, remainingBuffer);
+
+      // Copy audio data to output using offset (no slice allocation)
+      for (let i = 0; i < copyLength; i++) {
+        channel[outputIndex++] = currentBuffer[this.currentOffset++];
+      }
+
+      // If we've consumed the entire buffer, move to the next one
+      if (this.currentOffset >= currentBuffer.length) {
+        this.audioQueue.shift();
+        this.currentOffset = 0;
+      }
+    }
+
+    // Fill remaining output with silence
+    while (outputIndex < channel.length) {
+      channel[outputIndex++] = 0;
+    }
+
+    return true;
+  }
+}
+
+registerProcessor("pcm-processor", PCMProcessor);`;
+function calculateLevel(analyser) {
+  const data = new Uint8Array(analyser.fftSize);
+  analyser.getByteTimeDomainData(data);
+  let maxDeviation = 0;
+  for (const sample of data) {
+    const deviation = Math.abs(sample - 128);
+    if (deviation > maxDeviation) {
+      maxDeviation = deviation;
+    }
+  }
+  const normalized = maxDeviation / 128;
+  return Math.min(1, normalized * 1.5);
+}
+function base64ToArrayBuffer(base64) {
+  const binary = atob(base64);
+  const bytes = Uint8Array.from(binary, (char) => char.charCodeAt(0));
+  return bytes.buffer;
+}
+const emptyFrequencyData = new Uint8Array(1024);
+const emptyTimeDomainData = new Uint8Array(2048).fill(128);
+class AudioStreamer {
+  audioContext = null;
+  audioWorklet = null;
+  mediaStream = null;
+  analyser = null;
+  isStreaming = false;
+  sampleRate = 16e3;
+  client = null;
+  constructor(client) {
+    this.client = client;
+  }
+  get inputLevel() {
+    if (!this.analyser) return 0;
+    return calculateLevel(this.analyser);
+  }
+  get inputFrequencyData() {
+    if (!this.analyser) return emptyFrequencyData;
+    const data = new Uint8Array(this.analyser.frequencyBinCount);
+    this.analyser.getByteFrequencyData(data);
+    return data;
+  }
+  get inputTimeDomainData() {
+    if (!this.analyser) return emptyTimeDomainData;
+    const data = new Uint8Array(this.analyser.fftSize);
+    this.analyser.getByteTimeDomainData(data);
+    return data;
+  }
+  get inputSampleRate() {
+    return this.sampleRate;
+  }
+  async start() {
+    try {
+      const audioConstraints = {
+        sampleRate: this.sampleRate,
+        echoCancellation: true,
+        noiseSuppression: true,
+        autoGainControl: true
+      };
+      this.mediaStream = await navigator.mediaDevices.getUserMedia({
+        audio: audioConstraints
+      });
+      const track = this.mediaStream.getAudioTracks()[0];
+      const settings = track?.getSettings();
+      if (settings?.autoGainControl) {
+        console.warn("Native AGC not supported.");
+      }
+      this.audioContext = new AudioContext({
+        sampleRate: this.sampleRate
+      });
+      if (this.audioContext.state === "suspended") {
+        await this.audioContext.resume();
+      }
+      const workletBlob = new Blob([captureWorkletCode], {
+        type: "application/javascript"
+      });
+      const workletUrl = URL.createObjectURL(workletBlob);
+      await this.audioContext.audioWorklet.addModule(workletUrl);
+      URL.revokeObjectURL(workletUrl);
+      this.audioWorklet = new AudioWorkletNode(
+        this.audioContext,
+        "audio-capture-processor"
+      );
+      this.audioWorklet.port.onmessage = (event) => {
+        if (!this.isStreaming) return;
+        if (event.data.type === "audio") {
+          const inputData = event.data.data;
+          const pcmData = this.convertToPCM16(inputData);
+          const base64Audio = this.arrayBufferToBase64(pcmData);
+          if (this.client?.isSetupComplete) {
+            this.client.sendAudioMessage(base64Audio);
+          }
+        }
+      };
+      this.analyser = this.audioContext.createAnalyser();
+      this.analyser.fftSize = 2048;
+      this.analyser.smoothingTimeConstant = 0.3;
+      const source = this.audioContext.createMediaStreamSource(this.mediaStream);
+      source.connect(this.analyser);
+      this.analyser.connect(this.audioWorklet);
+      this.isStreaming = true;
+    } catch (error) {
+      this.stop();
+      throw error;
+    }
+  }
+  stop() {
+    this.isStreaming = false;
+    if (this.audioWorklet) {
+      this.audioWorklet.disconnect();
+      this.audioWorklet.port.close();
+      this.audioWorklet = null;
+    }
+    if (this.audioContext) {
+      void this.audioContext.close();
+      this.audioContext = null;
+    }
+    if (this.mediaStream) {
+      this.mediaStream.getTracks().forEach((track) => track.stop());
+      this.mediaStream = null;
+    }
+  }
+  startAudioCapture() {
+    if (this.mediaStream) {
+      for (const track of this.mediaStream.getAudioTracks()) {
+        track.enabled = true;
+      }
+    }
+    this.isStreaming = true;
+  }
+  stopAudioCapture() {
+    if (this.mediaStream) {
+      for (const track of this.mediaStream.getAudioTracks()) {
+        track.enabled = false;
+      }
+    }
+    this.isStreaming = false;
+  }
+  convertToPCM16(float32Array) {
+    const int16Array = new Int16Array(float32Array.length);
+    for (let i = 0; i < float32Array.length; i++) {
+      const sample = Math.max(-1, Math.min(1, float32Array[i] ?? 0));
+      int16Array[i] = sample * 32767;
+    }
+    return int16Array.buffer;
+  }
+  arrayBufferToBase64(buffer) {
+    const bytes = new Uint8Array(buffer);
+    const binary = String.fromCharCode(...bytes);
+    return btoa(binary);
+  }
+}
+class AudioPlayer {
+  audioContext = null;
+  workletNode = null;
+  gainNode = null;
+  analyser = null;
+  isInitialized = false;
+  volume = 1;
+  sampleRate = 24e3;
+  get outputLevel() {
+    if (!this.analyser) return 0;
+    return calculateLevel(this.analyser);
+  }
+  get outputFrequencyData() {
+    if (!this.analyser) return emptyFrequencyData;
+    const data = new Uint8Array(this.analyser.frequencyBinCount);
+    this.analyser.getByteFrequencyData(data);
+    return data;
+  }
+  get outputTimeDomainData() {
+    if (!this.analyser) return emptyTimeDomainData;
+    const data = new Uint8Array(this.analyser.fftSize);
+    this.analyser.getByteTimeDomainData(data);
+    return data;
+  }
+  get outputSampleRate() {
+    return this.sampleRate;
+  }
+  async init() {
+    if (this.isInitialized) return;
+    try {
+      this.audioContext = new AudioContext({
+        sampleRate: this.sampleRate
+      });
+      const workletBlob = new Blob([playbackWorkletCode], {
+        type: "application/javascript"
+      });
+      const workletUrl = URL.createObjectURL(workletBlob);
+      await this.audioContext.audioWorklet.addModule(workletUrl);
+      URL.revokeObjectURL(workletUrl);
+      this.workletNode = new AudioWorkletNode(
+        this.audioContext,
+        "pcm-processor"
+      );
+      this.gainNode = this.audioContext.createGain();
+      this.gainNode.gain.value = this.volume;
+      this.analyser = this.audioContext.createAnalyser();
+      this.analyser.fftSize = 2048;
+      this.analyser.smoothingTimeConstant = 0.3;
+      this.workletNode.connect(this.gainNode);
+      this.gainNode.connect(this.analyser);
+      this.analyser.connect(this.audioContext.destination);
+      this.isInitialized = true;
+    } catch (error) {
+      this.destroy();
+      throw error;
+    }
+  }
+  async play(pcmData) {
+    if (!this.isInitialized) {
+      await this.init();
+    }
+    if (this.audioContext?.state === "suspended") {
+      await this.audioContext.resume();
+    }
+    const inputArray = new Int16Array(pcmData);
+    const float32Data = new Float32Array(inputArray.length);
+    for (let i = 0; i < inputArray.length; i++) {
+      float32Data[i] = (inputArray[i] ?? 0) / 32768;
+    }
+    this.workletNode?.port.postMessage(float32Data);
+  }
+  /* Interrupt playback */
+  interrupt() {
+    if (this.workletNode) {
+      this.workletNode.port.postMessage("interrupt");
+    }
+  }
+  setVolume(volume) {
+    this.volume = Math.max(0, Math.min(1, volume));
+    if (this.gainNode) {
+      this.gainNode.gain.value = this.volume;
+    }
+  }
+  destroy() {
+    if (this.audioContext) {
+      void this.audioContext.close();
+      this.audioContext = null;
+    }
+    this.isInitialized = false;
+  }
+}
+export {
+  AudioPlayer,
+  AudioStreamer,
+  base64ToArrayBuffer
+};
+//# sourceMappingURL=utils.js.map

package/dist/esm/realtime/utils.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"utils.js","sources":["../../../src/realtime/utils.ts"],"sourcesContent":["import type { GeminiLiveClient } from './client'\n\n/**\n * Audio Worklet Processor for capturing and processing audio\n */\nconst captureWorkletCode = `\nclass AudioCaptureProcessor extends AudioWorkletProcessor {\n  constructor() {\n    super();\n    this.bufferSize = 512; // 32ms at 16kHz — per Gemini best practices (20-40ms chunks)\n    this.buffer = new Float32Array(this.bufferSize);\n    this.bufferIndex = 0;\n  }\n\n  process(inputs, outputs, parameters) {\n    const input = inputs[0];\n\n    if (input && input.length > 0) {\n      const inputChannel = input[0];\n\n      // Buffer the incoming audio\n      for (let i = 0; i < inputChannel.length; i++) {\n        this.buffer[this.bufferIndex++] = inputChannel[i];\n\n        // When buffer is full, send it to main thread\n        if (this.bufferIndex >= this.bufferSize) {\n          // Send the buffered audio to the main thread\n          this.port.postMessage({\n            type: \"audio\",\n            data: this.buffer.slice(),\n          });\n\n          // Reset buffer\n          this.bufferIndex = 0;\n        }\n      }\n    }\n\n    // Return true to keep the processor alive\n    return true;\n  }\n}\n\n// Register the processor\nregisterProcessor(\"audio-capture-processor\", AudioCaptureProcessor);`\n\n/**\n * Audio Playback Worklet Processor for playing PCM audio.\n * Uses an offset tracker instead of slice() to avoid allocations\n * on the real-time audio thread.\n */\nconst playbackWorkletCode = `\nclass PCMProcessor extends AudioWorkletProcessor {\n  constructor() {\n    super();\n    this.audioQueue = [];\n    this.currentOffset = 0; // Track position in current buffer (avoids slice())\n\n    this.port.onmessage = (event) => {\n      if (event.data === \"interrupt\") {\n        // Clear the queue on interrupt\n        this.audioQueue = [];\n        this.currentOffset = 0;\n      } else if (event.data instanceof Float32Array) {\n        // Add audio data to the queue\n        this.audioQueue.push(event.data);\n      }\n    };\n  }\n\n  process(inputs, outputs, parameters) {\n    const output = outputs[0];\n    if (output.length === 0) return true;\n\n    const channel = output[0];\n    let outputIndex = 0;\n\n    // Fill the output buffer from the queue\n    while (outputIndex < channel.length && this.audioQueue.length > 0) {\n      const currentBuffer = this.audioQueue[0];\n\n      if (!currentBuffer || currentBuffer.length === 0) {\n        this.audioQueue.shift();\n        this.currentOffset = 0;\n        continue;\n      }\n\n      const remainingOutput = channel.length - outputIndex;\n      const remainingBuffer = currentBuffer.length - this.currentOffset;\n      const copyLength = Math.min(remainingOutput, remainingBuffer);\n\n      // Copy audio data to output using offset (no slice allocation)\n      for (let i = 0; i < copyLength; i++) {\n        channel[outputIndex++] = currentBuffer[this.currentOffset++];\n      }\n\n      // If we've consumed the entire buffer, move to the next one\n      if (this.currentOffset >= currentBuffer.length) {\n        this.audioQueue.shift();\n        this.currentOffset = 0;\n      }\n    }\n\n    // Fill remaining output with silence\n    while (outputIndex < channel.length) {\n      channel[outputIndex++] = 0;\n    }\n\n    return true;\n  }\n}\n\nregisterProcessor(\"pcm-processor\", PCMProcessor);`\n\nfunction calculateLevel(analyser: AnalyserNode): number {\n  const data = new Uint8Array(analyser.fftSize)\n  analyser.getByteTimeDomainData(data)\n\n  // Find peak deviation from center (128 is silence)\n  // This is more responsive than RMS for voice level meters\n  let maxDeviation = 0\n  for (const sample of data) {\n    const deviation = Math.abs(sample - 128)\n    if (deviation > maxDeviation) {\n      maxDeviation = deviation\n    }\n  }\n\n  // Normalize to 0-1 range (max deviation is 128)\n  // Scale by 1.5x so that ~66% amplitude reads as full scale\n  // This provides good visual feedback without pegging too early\n  const normalized = maxDeviation / 128\n  return Math.min(1, normalized * 1.5)\n}\n\nexport function base64ToArrayBuffer(base64: string): ArrayBuffer {\n  const binary = atob(base64)\n  const bytes = Uint8Array.from(binary, (char) => char.charCodeAt(0))\n  return bytes.buffer\n}\n\n// Empty arrays for when visualization isn't available\n// frequencyBinCount = fftSize / 2 = 1024\nconst emptyFrequencyData = new Uint8Array(1024)\nconst emptyTimeDomainData = new Uint8Array(2048).fill(128) // 128 is silence\n\nexport class AudioStreamer {\n  private audioContext: AudioContext | null = null\n  private audioWorklet: AudioWorkletNode | null = null\n  private mediaStream: MediaStream | null = null\n  private analyser: AnalyserNode | null = null\n  private isStreaming = false\n  private readonly sampleRate = 16000\n  private readonly client: GeminiLiveClient | null = null\n\n  constructor(client: GeminiLiveClient) {\n    this.client = client\n  }\n\n  get inputLevel() {\n    if (!this.analyser) return 0\n    return calculateLevel(this.analyser)\n  }\n\n  get inputFrequencyData() {\n    if (!this.analyser) return emptyFrequencyData\n    const data = new Uint8Array(this.analyser.frequencyBinCount)\n    this.analyser.getByteFrequencyData(data)\n    return data\n  }\n\n  get inputTimeDomainData() {\n    if (!this.analyser) return emptyTimeDomainData\n    const data = new Uint8Array(this.analyser.fftSize)\n    this.analyser.getByteTimeDomainData(data)\n    return data\n  }\n\n  get inputSampleRate() {\n    return this.sampleRate\n  }\n\n  async start() {\n    try {\n      const audioConstraints: MediaTrackConstraints = {\n        sampleRate: this.sampleRate,\n        echoCancellation: true,\n        noiseSuppression: true,\n        autoGainControl: true,\n      }\n\n      // Get microphone access\n      this.mediaStream = await navigator.mediaDevices.getUserMedia({\n        audio: audioConstraints,\n      })\n\n      // Check if native AGC is active\n      const track = this.mediaStream.getAudioTracks()[0]\n      const settings = track?.getSettings()\n\n      if (settings?.autoGainControl) {\n        console.warn('Native AGC not supported.')\n      }\n\n      // Create audio context\n      this.audioContext = new AudioContext({\n        sampleRate: this.sampleRate,\n      })\n\n      if (this.audioContext.state === 'suspended') {\n        await this.audioContext.resume()\n      }\n\n      const workletBlob = new Blob([captureWorkletCode], {\n        type: 'application/javascript',\n      })\n      const workletUrl = URL.createObjectURL(workletBlob)\n\n      // Load the audio worklet module, then release the blob URL.\n      await this.audioContext.audioWorklet.addModule(workletUrl)\n      URL.revokeObjectURL(workletUrl)\n\n      // Create the audio worklet node\n      this.audioWorklet = new AudioWorkletNode(\n        this.audioContext,\n        'audio-capture-processor',\n      )\n\n      // Set up message handling from the worklet\n      this.audioWorklet.port.onmessage = (event) => {\n        if (!this.isStreaming) return\n\n        if (event.data.type === 'audio') {\n          const inputData = event.data.data\n          const pcmData = this.convertToPCM16(inputData)\n          const base64Audio = this.arrayBufferToBase64(pcmData)\n\n          // Send to Gemini only if after setup complete\n          if (this.client?.isSetupComplete) {\n            this.client.sendAudioMessage(base64Audio)\n          }\n        }\n      }\n\n      // Create analyser for volume detection\n      this.analyser = this.audioContext.createAnalyser()\n      this.analyser.fftSize = 2048 // Larger size for more accurate level detection\n      this.analyser.smoothingTimeConstant = 0.3\n\n      // Connect the audio graph\n      const source = this.audioContext.createMediaStreamSource(this.mediaStream)\n      source.connect(this.analyser)\n      this.analyser.connect(this.audioWorklet)\n\n      // Start streaming\n      this.isStreaming = true\n    } catch (error) {\n      // Clean up the mic + audio context if setup failed partway through.\n      this.stop()\n      throw error\n    }\n  }\n\n  stop() {\n    this.isStreaming = false\n\n    if (this.audioWorklet) {\n      this.audioWorklet.disconnect()\n      this.audioWorklet.port.close()\n      this.audioWorklet = null\n    }\n\n    if (this.audioContext) {\n      void this.audioContext.close()\n      this.audioContext = null\n    }\n\n    if (this.mediaStream) {\n      this.mediaStream.getTracks().forEach((track) => track.stop())\n      this.mediaStream = null\n    }\n  }\n\n  startAudioCapture() {\n    if (this.mediaStream) {\n      for (const track of this.mediaStream.getAudioTracks()) {\n        track.enabled = true\n      }\n    }\n    this.isStreaming = true\n  }\n\n  stopAudioCapture() {\n    if (this.mediaStream) {\n      // Disable tracks rather than stopping them to allow re-enabling\n      for (const track of this.mediaStream.getAudioTracks()) {\n        track.enabled = false\n      }\n    }\n    this.isStreaming = false\n  }\n\n  private convertToPCM16(float32Array: Float32Array): ArrayBuffer {\n    const int16Array = new Int16Array(float32Array.length)\n    for (let i = 0; i < float32Array.length; i++) {\n      const sample = Math.max(-1, Math.min(1, float32Array[i] ?? 0))\n      int16Array[i] = sample * 0x7fff\n    }\n    return int16Array.buffer\n  }\n\n  private arrayBufferToBase64(buffer: ArrayBuffer): string {\n    const bytes = new Uint8Array(buffer)\n    const binary = String.fromCharCode(...bytes)\n    return btoa(binary)\n  }\n}\n\nexport class AudioPlayer {\n  private audioContext: AudioContext | null = null\n  private workletNode: AudioWorkletNode | null = null\n  private gainNode: GainNode | null = null\n  private analyser: AnalyserNode | null = null\n  private isInitialized = false\n  private volume = 1.0\n  private readonly sampleRate = 24000\n\n  get outputLevel() {\n    if (!this.analyser) return 0\n    return calculateLevel(this.analyser)\n  }\n\n  get outputFrequencyData() {\n    if (!this.analyser) return emptyFrequencyData\n    const data = new Uint8Array(this.analyser.frequencyBinCount)\n    this.analyser.getByteFrequencyData(data)\n    return data\n  }\n\n  get outputTimeDomainData() {\n    if (!this.analyser) return emptyTimeDomainData\n    const data = new Uint8Array(this.analyser.fftSize)\n    this.analyser.getByteTimeDomainData(data)\n    return data\n  }\n\n  get outputSampleRate() {\n    return this.sampleRate\n  }\n\n  async init() {\n    if (this.isInitialized) return\n\n    try {\n      // Create audio context at 24kHz to match Gemini\n      this.audioContext = new AudioContext({\n        sampleRate: this.sampleRate,\n      })\n\n      const workletBlob = new Blob([playbackWorkletCode], {\n        type: 'application/javascript',\n      })\n      const workletUrl = URL.createObjectURL(workletBlob)\n\n      // Load the audio worklet module, then release the blob URL.\n      await this.audioContext.audioWorklet.addModule(workletUrl)\n      URL.revokeObjectURL(workletUrl)\n\n      // Create worklet node\n      this.workletNode = new AudioWorkletNode(\n        this.audioContext,\n        'pcm-processor',\n      )\n\n      // Create gain node for volume control\n      this.gainNode = this.audioContext.createGain()\n      this.gainNode.gain.value = this.volume\n\n      // Create analyser for volume detection\n      this.analyser = this.audioContext.createAnalyser()\n      this.analyser.fftSize = 2048 // Larger size for more accurate level detection\n      this.analyser.smoothingTimeConstant = 0.3\n\n      // Connect nodes\n      this.workletNode.connect(this.gainNode)\n      this.gainNode.connect(this.analyser)\n      this.analyser.connect(this.audioContext.destination)\n\n      this.isInitialized = true\n    } catch (error) {\n      // Release the audio context if initialization failed partway through.\n      this.destroy()\n      throw error\n    }\n  }\n\n  async play(pcmData: ArrayBuffer) {\n    if (!this.isInitialized) {\n      await this.init()\n    }\n\n    // Resume audio context if suspended\n    if (this.audioContext?.state === 'suspended') {\n      await this.audioContext.resume()\n    }\n\n    // Convert PCM16 LE to Float32\n    const inputArray = new Int16Array(pcmData)\n    const float32Data = new Float32Array(inputArray.length)\n    for (let i = 0; i < inputArray.length; i++) {\n      float32Data[i] = (inputArray[i] ?? 0) / 32768\n    }\n\n    // Send to worklet for playback\n    this.workletNode?.port.postMessage(float32Data)\n  }\n\n  /* Interrupt playback */\n  interrupt() {\n    if (this.workletNode) {\n      this.workletNode.port.postMessage('interrupt')\n    }\n  }\n\n  setVolume(volume: number) {\n    this.volume = Math.max(0, Math.min(1, volume))\n    if (this.gainNode) {\n      this.gainNode.gain.value = this.volume\n    }\n  }\n\n  destroy() {\n    if (this.audioContext) {\n      void this.audioContext.close()\n      this.audioContext = null\n    }\n    this.isInitialized = false\n  }\n}\n"],"names":[],"mappings":"AAKA,MAAM,qBAAqB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AA8C3B,MAAM,sBAAsB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AA+D5B,SAAS,eAAe,UAAgC;AACtD,QAAM,OAAO,IAAI,WAAW,SAAS,OAAO;AAC5C,WAAS,sBAAsB,IAAI;AAInC,MAAI,eAAe;AACnB,aAAW,UAAU,MAAM;AACzB,UAAM,YAAY,KAAK,IAAI,SAAS,GAAG;AACvC,QAAI,YAAY,cAAc;AAC5B,qBAAe;AAAA,IACjB;AAAA,EACF;AAKA,QAAM,aAAa,eAAe;AAClC,SAAO,KAAK,IAAI,GAAG,aAAa,GAAG;AACrC;AAEO,SAAS,oBAAoB,QAA6B;AAC/D,QAAM,SAAS,KAAK,MAAM;AAC1B,QAAM,QAAQ,WAAW,KAAK,QAAQ,CAAC,SAAS,KAAK,WAAW,CAAC,CAAC;AAClE,SAAO,MAAM;AACf;AAIA,MAAM,qBAAqB,IAAI,WAAW,IAAI;AAC9C,MAAM,sBAAsB,IAAI,WAAW,IAAI,EAAE,KAAK,GAAG;AAElD,MAAM,cAAc;AAAA,EACjB,eAAoC;AAAA,EACpC,eAAwC;AAAA,EACxC,cAAkC;AAAA,EAClC,WAAgC;AAAA,EAChC,cAAc;AAAA,EACL,aAAa;AAAA,EACb,SAAkC;AAAA,EAEnD,YAAY,QAA0B;AACpC,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,IAAI,aAAa;AACf,QAAI,CAAC,KAAK,SAAU,QAAO;AAC3B,WAAO,eAAe,KAAK,QAAQ;AAAA,EACrC;AAAA,EAEA,IAAI,qBAAqB;AACvB,QAAI,CAAC,KAAK,SAAU,QAAO;AAC3B,UAAM,OAAO,IAAI,WAAW,KAAK,SAAS,iBAAiB;AAC3D,SAAK,SAAS,qBAAqB,IAAI;AACvC,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,sBAAsB;AACxB,QAAI,CAAC,KAAK,SAAU,QAAO;AAC3B,UAAM,OAAO,IAAI,WAAW,KAAK,SAAS,OAAO;AACjD,SAAK,SAAS,sBAAsB,IAAI;AACxC,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,kBAAkB;AACpB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,QAAQ;AACZ,QAAI;AACF,YAAM,mBAA0C;AAAA,QAC9C,YAAY,KAAK;AAAA,QACjB,kBAAkB;AAAA,QAClB,kBAAkB;AAAA,QAClB,iBAAiB;AAAA,MAAA;AAInB,WAAK,cAAc,MAAM,UAAU,aAAa,aAAa;AAAA,QAC3D,OAAO;AAAA,MAAA,CACR;AAGD,YAAM,QAAQ,KAAK,YAAY,eAAA,EAAiB,CAAC;AACjD,YAAM,WAAW,OAAO,YAAA;AAExB,UAAI,UAAU,iBAAiB;AAC7B,gBAAQ,KAAK,2BAA2B;AAAA,MAC1C;AAGA,WAAK,eAAe,IAAI,aAAa;AAAA,QACnC,YAAY,KAAK;AAAA,MAAA,CAClB;AAED,UAAI,KAAK,aAAa,UAAU,aAAa;AAC3C,cAAM,KAAK,aAAa,OAAA;AAAA,MAC1B;AAEA,YAAM,cAAc,IAAI,KAAK,CAAC,kBAAkB,GAAG;AAAA,QACjD,MAAM;AAAA,MAAA,CACP;AACD,YAAM,aAAa,IAAI,gBAAgB,WAAW;AAGlD,YAAM,KAAK,aAAa,aAAa,UAAU,UAAU;AACzD,UAAI,gBAAgB,UAAU;AAG9B,WAAK,eAAe,IAAI;AAAA,QACtB,KAAK;AAAA,QACL;AAAA,MAAA;AAIF,WAAK,aAAa,KAAK,YAAY,CAAC,UAAU;AAC5C,YAAI,CAAC,KAAK,YAAa;AAEvB,YAAI,MAAM,KAAK,SAAS,SAAS;AAC/B,gBAAM,YAAY,MAAM,KAAK;AAC7B,gBAAM,UAAU,KAAK,eAAe,SAAS;AAC7C,gBAAM,cAAc,KAAK,oBAAoB,OAAO;AAGpD,cAAI,KAAK,QAAQ,iBAAiB;AAChC,iBAAK,OAAO,iBAAiB,WAAW;AAAA,UAC1C;AAAA,QACF;AAAA,MACF;AAGA,WAAK,WAAW,KAAK,aAAa,eAAA;AAClC,WAAK,SAAS,UAAU;AACxB,WAAK,SAAS,wBAAwB;AAGtC,YAAM,SAAS,KAAK,aAAa,wBAAwB,KAAK,WAAW;AACzE,aAAO,QAAQ,KAAK,QAAQ;AAC5B,WAAK,SAAS,QAAQ,KAAK,YAAY;AAGvC,WAAK,cAAc;AAAA,IACrB,SAAS,OAAO;AAEd,WAAK,KAAA;AACL,YAAM;AAAA,IACR;AAAA,EACF;AAAA,EAEA,OAAO;AACL,SAAK,cAAc;AAEnB,QAAI,KAAK,cAAc;AACrB,WAAK,aAAa,WAAA;AAClB,WAAK,aAAa,KAAK,MAAA;AACvB,WAAK,eAAe;AAAA,IACtB;AAEA,QAAI,KAAK,cAAc;AACrB,WAAK,KAAK,aAAa,MAAA;AACvB,WAAK,eAAe;AAAA,IACtB;AAEA,QAAI,KAAK,aAAa;AACpB,WAAK,YAAY,YAAY,QAAQ,CAAC,UAAU,MAAM,MAAM;AAC5D,WAAK,cAAc;AAAA,IACrB;AAAA,EACF;AAAA,EAEA,oBAAoB;AAClB,QAAI,KAAK,aAAa;AACpB,iBAAW,SAAS,KAAK,YAAY,eAAA,GAAkB;AACrD,cAAM,UAAU;AAAA,MAClB;AAAA,IACF;AACA,SAAK,cAAc;AAAA,EACrB;AAAA,EAEA,mBAAmB;AACjB,QAAI,KAAK,aAAa;AAEpB,iBAAW,SAAS,KAAK,YAAY,eAAA,GAAkB;AACrD,cAAM,UAAU;AAAA,MAClB;AAAA,IACF;AACA,SAAK,cAAc;AAAA,EACrB;AAAA,EAEQ,eAAe,cAAyC;AAC9D,UAAM,aAAa,IAAI,WAAW,aAAa,MAAM;AACrD,aAAS,IAAI,GAAG,IAAI,aAAa,QAAQ,KAAK;AAC5C,YAAM,SAAS,KAAK,IAAI,IAAI,KAAK,IAAI,GAAG,aAAa,CAAC,KAAK,CAAC,CAAC;AAC7D,iBAAW,CAAC,IAAI,SAAS;AAAA,IAC3B;AACA,WAAO,WAAW;AAAA,EACpB;AAAA,EAEQ,oBAAoB,QAA6B;AACvD,UAAM,QAAQ,IAAI,WAAW,MAAM;AACnC,UAAM,SAAS,OAAO,aAAa,GAAG,KAAK;AAC3C,WAAO,KAAK,MAAM;AAAA,EACpB;AACF;AAEO,MAAM,YAAY;AAAA,EACf,eAAoC;AAAA,EACpC,cAAuC;AAAA,EACvC,WAA4B;AAAA,EAC5B,WAAgC;AAAA,EAChC,gBAAgB;AAAA,EAChB,SAAS;AAAA,EACA,aAAa;AAAA,EAE9B,IAAI,cAAc;AAChB,QAAI,CAAC,KAAK,SAAU,QAAO;AAC3B,WAAO,eAAe,KAAK,QAAQ;AAAA,EACrC;AAAA,EAEA,IAAI,sBAAsB;AACxB,QAAI,CAAC,KAAK,SAAU,QAAO;AAC3B,UAAM,OAAO,IAAI,WAAW,KAAK,SAAS,iBAAiB;AAC3D,SAAK,SAAS,qBAAqB,IAAI;AACvC,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,uBAAuB;AACzB,QAAI,CAAC,KAAK,SAAU,QAAO;AAC3B,UAAM,OAAO,IAAI,WAAW,KAAK,SAAS,OAAO;AACjD,SAAK,SAAS,sBAAsB,IAAI;AACxC,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,mBAAmB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,OAAO;AACX,QAAI,KAAK,cAAe;AAExB,QAAI;AAEF,WAAK,eAAe,IAAI,aAAa;AAAA,QACnC,YAAY,KAAK;AAAA,MAAA,CAClB;AAED,YAAM,cAAc,IAAI,KAAK,CAAC,mBAAmB,GAAG;AAAA,QAClD,MAAM;AAAA,MAAA,CACP;AACD,YAAM,aAAa,IAAI,gBAAgB,WAAW;AAGlD,YAAM,KAAK,aAAa,aAAa,UAAU,UAAU;AACzD,UAAI,gBAAgB,UAAU;AAG9B,WAAK,cAAc,IAAI;AAAA,QACrB,KAAK;AAAA,QACL;AAAA,MAAA;AAIF,WAAK,WAAW,KAAK,aAAa,WAAA;AAClC,WAAK,SAAS,KAAK,QAAQ,KAAK;AAGhC,WAAK,WAAW,KAAK,aAAa,eAAA;AAClC,WAAK,SAAS,UAAU;AACxB,WAAK,SAAS,wBAAwB;AAGtC,WAAK,YAAY,QAAQ,KAAK,QAAQ;AACtC,WAAK,SAAS,QAAQ,KAAK,QAAQ;AACnC,WAAK,SAAS,QAAQ,KAAK,aAAa,WAAW;AAEnD,WAAK,gBAAgB;AAAA,IACvB,SAAS,OAAO;AAEd,WAAK,QAAA;AACL,YAAM;AAAA,IACR;AAAA,EACF;AAAA,EAEA,MAAM,KAAK,SAAsB;AAC/B,QAAI,CAAC,KAAK,eAAe;AACvB,YAAM,KAAK,KAAA;AAAA,IACb;AAGA,QAAI,KAAK,cAAc,UAAU,aAAa;AAC5C,YAAM,KAAK,aAAa,OAAA;AAAA,IAC1B;AAGA,UAAM,aAAa,IAAI,WAAW,OAAO;AACzC,UAAM,cAAc,IAAI,aAAa,WAAW,MAAM;AACtD,aAAS,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;AAC1C,kBAAY,CAAC,KAAK,WAAW,CAAC,KAAK,KAAK;AAAA,IAC1C;AAGA,SAAK,aAAa,KAAK,YAAY,WAAW;AAAA,EAChD;AAAA;AAAA,EAGA,YAAY;AACV,QAAI,KAAK,aAAa;AACpB,WAAK,YAAY,KAAK,YAAY,WAAW;AAAA,IAC/C;AAAA,EACF;AAAA,EAEA,UAAU,QAAgB;AACxB,SAAK,SAAS,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,MAAM,CAAC;AAC7C,QAAI,KAAK,UAAU;AACjB,WAAK,SAAS,KAAK,QAAQ,KAAK;AAAA,IAClC;AAAA,EACF;AAAA,EAEA,UAAU;AACR,QAAI,KAAK,cAAc;AACrB,WAAK,KAAK,aAAa,MAAA;AACvB,WAAK,eAAe;AAAA,IACtB;AACA,SAAK,gBAAgB;AAAA,EACvB;AACF;"}

package/dist/esm/video/video-provider-options.d.ts

@@ -1,16 +1,27 @@
+import { GEMINI_INTERACTIONS_VIDEO_MODELS, GEMINI_VIDEO_MODELS } from '../model-meta.js';
 import { DurationOptions } from '@tanstack/ai/adapters';
-import { GenerateVideosConfig } from '@google/genai';
-import { GEMINI_VIDEO_MODELS } from '../model-meta.js';
+import { GenerateVideosConfig, Interactions } from '@google/genai';
 /**
- * Model type for Gemini Veo video generation.
+ * Model type for Gemini video generation (Veo + Omni Flash).
  * @experimental Video generation is an experimental feature and may change.
  */
 export type GeminiVideoModel = (typeof GEMINI_VIDEO_MODELS)[number];
 /**
- * Supported aspect ratios for Veo video generation. This is the `size` value
- * for the Gemini video adapter — Veo expresses output shape as an aspect
- * ratio (plus an optional `resolution` in `modelOptions`), not pixel
- * dimensions.
+ * Video models served by the Interactions API (Gemini Omni Flash) rather
+ * than Veo's `:predictLongRunning` operations flow.
+ * @experimental Omni video generation is an experimental feature and may change.
+ */
+export type GeminiInteractionsVideoModel = (typeof GEMINI_INTERACTIONS_VIDEO_MODELS)[number];
+/**
+ * Runtime guard for the Interactions-served video models.
+ * @experimental Omni video generation is an experimental feature and may change.
+ */
+export declare function isInteractionsVideoModel(model: GeminiVideoModel): model is GeminiInteractionsVideoModel;
+/**
+ * Supported aspect ratios for Gemini video generation. This is the `size`
+ * value for the Gemini video adapter — both Veo and Omni Flash express
+ * output shape as an aspect ratio (plus an optional `resolution` in Veo's
+ * `modelOptions`), not pixel dimensions.
  *
  * @experimental Video generation is an experimental feature and may change.
  */
@@ -30,13 +41,37 @@ export type GeminiVideoSize = '16:9' | '9:16';
  * @experimental Video generation is an experimental feature and may change.
  */
 export type GeminiVideoProviderOptions = Omit<GenerateVideosConfig, 'durationSeconds' | 'aspectRatio' | 'lastFrame' | 'referenceImages' | 'httpOptions' | 'abortSignal'>;
+/**
+ * Provider-specific options for Gemini Omni Flash video generation on the
+ * Interactions API.
+ *
+ * Derived from the SDK's `Interactions.CreateModelInteractionParamsNonStreaming`,
+ * minus the fields the adapter manages itself:
+ * - `model` / `input` — set from the adapter's model and the `prompt`
+ * - `stream` / `background` — the adapter always creates a background job
+ *   and polls it through the `generateVideo` jobs API
+ * - `response_modalities` / `response_format` — the adapter requests video
+ *   output and maps the top-level `size` option onto
+ *   `response_format.aspect_ratio`
+ * - `tools` / `response_mime_type` — not applicable to video generation
+ *
+ * Notable passthroughs:
+ * - `previous_interaction_id` — conversational video editing: chain a new
+ *   prompt onto a prior Omni interaction to refine its video
+ * - `generation_config.video_config.task` — pin the task mode
+ *   (`'text_to_video' | 'image_to_video' | 'reference_to_video' | 'edit'`)
+ *   instead of letting the model infer it
+ *
+ * @experimental Omni video generation is an experimental feature and may change.
+ */
+export type GeminiOmniVideoProviderOptions = Omit<Interactions.CreateModelInteractionParamsNonStreaming, 'model' | 'input' | 'stream' | 'background' | 'response_modalities' | 'response_format' | 'response_mime_type' | 'tools'>;
 /**
  * Model-specific provider options mapping.
  *
  * @experimental Video generation is an experimental feature and may change.
  */
 export type GeminiVideoModelProviderOptionsByName = {
-    [TModel in GeminiVideoModel]: GeminiVideoProviderOptions;
+    [TModel in GeminiVideoModel]: TModel extends GeminiInteractionsVideoModel ? GeminiOmniVideoProviderOptions : GeminiVideoProviderOptions;
 };
 /**
  * Model-specific size (aspect ratio) mapping.
@@ -49,16 +84,20 @@ export type GeminiVideoModelSizeByName = {
 /**
  * Per-model prompt input modalities. Every Veo model accepts image
  * conditioning inputs (first frame, last frame, reference images) alongside
- * the text prompt.
+ * the text prompt. Omni Flash additionally accepts video inputs (short
+ * reference clips / videos to edit).
  *
  * @experimental Video generation is an experimental feature and may change.
  */
 export type GeminiVideoModelInputModalitiesByName = {
-    [TModel in GeminiVideoModel]: readonly ['image'];
+    [TModel in GeminiVideoModel]: TModel extends GeminiInteractionsVideoModel ? readonly ['image', 'video'] : readonly ['image'];
 };
 /**
- * Per-model duration unions (seconds, as numbers — the API's
- * `parameters.durationSeconds` field is numeric).
+ * Per-model duration unions (seconds, as numbers — Veo's
+ * `parameters.durationSeconds` field is numeric; Omni Flash accepts a
+ * continuous 3–10 second range, fractional seconds included, so it stays
+ * `number` — the adapter rejects out-of-range values at job creation,
+ * against the range entry below).
  *
  * @experimental Video generation is an experimental feature and may change.
  */
@@ -66,14 +105,20 @@ export type GeminiVideoModelDurationByName = {
     'veo-3.1-generate-preview': 4 | 6 | 8;
     'veo-3.1-fast-generate-preview': 4 | 6 | 8;
     'veo-3.1-lite-generate-preview': 4 | 6 | 8;
+    'gemini-omni-flash-preview': number;
 };
 /**
  * Runtime duration table backing `availableDurations()` / `snapDuration()`.
  *
- * Curated from the official Veo docs
+ * Veo values are curated from the official docs
  * (https://ai.google.dev/gemini-api/docs/video) — the Gemini OpenAPI spec
  * types the `:predictLongRunning` request's `parameters` as unconstrained,
  * so it carries no per-model duration information to derive these from.
+ * Omni Flash's 3–10s range was verified against the live API
+ * (2026-07-02): `response_format.duration` takes a `"<seconds>s"` string,
+ * fractional values are accepted, out-of-range values are rejected with
+ * "minimum allowed 3s" / "maximum allowed 10s", and omitting it defaults
+ * to a 10-second clip.
  *
  * @experimental Video generation is an experimental feature and may change.
  */
@@ -81,7 +126,7 @@ export declare const GEMINI_VIDEO_DURATIONS: {
     readonly [TModel in GeminiVideoModel]: DurationOptions<GeminiVideoModelDurationByName[TModel]>;
 };
 /**
- * Look up the duration options for a Veo model.
+ * Look up the duration options for a Gemini video model.
  *
  * @experimental Video generation is an experimental feature and may change.
  */

package/dist/esm/video/video-provider-options.js

@@ -1,13 +1,26 @@
+import { GEMINI_INTERACTIONS_VIDEO_MODELS } from "../model-meta.js";
+function isInteractionsVideoModel(model) {
+  return GEMINI_INTERACTIONS_VIDEO_MODELS.includes(
+    model
+  );
+}
 const GEMINI_VIDEO_DURATIONS = {
   "veo-3.1-generate-preview": { kind: "discrete", values: [4, 6, 8] },
   "veo-3.1-fast-generate-preview": { kind: "discrete", values: [4, 6, 8] },
-  "veo-3.1-lite-generate-preview": { kind: "discrete", values: [4, 6, 8] }
+  "veo-3.1-lite-generate-preview": { kind: "discrete", values: [4, 6, 8] },
+  "gemini-omni-flash-preview": {
+    kind: "range",
+    min: 3,
+    max: 10,
+    unit: "seconds"
+  }
 };
 function getGeminiVideoDurationOptions(model) {
   return GEMINI_VIDEO_DURATIONS[model];
 }
 export {
   GEMINI_VIDEO_DURATIONS,
-  getGeminiVideoDurationOptions
+  getGeminiVideoDurationOptions,
+  isInteractionsVideoModel
 };
 //# sourceMappingURL=video-provider-options.js.map

package/dist/esm/video/video-provider-options.js.map

@@ -1 +1 @@
-{"version":3,"file":"video-provider-options.js","sources":["../../../src/video/video-provider-options.ts"],"sourcesContent":["/**\n * Gemini Veo Video Generation Provider Options\n *\n * Based on https://ai.google.dev/gemini-api/docs/video\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type { GenerateVideosConfig } from '@google/genai'\nimport type { GEMINI_VIDEO_MODELS } from '../model-meta'\n\n/**\n * Model type for Gemini Veo video generation.\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModel = (typeof GEMINI_VIDEO_MODELS)[number]\n\n/**\n * Supported aspect ratios for Veo video generation. This is the `size` value\n * for the Gemini video adapter — Veo expresses output shape as an aspect\n * ratio (plus an optional `resolution` in `modelOptions`), not pixel\n * dimensions.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoSize = '16:9' | '9:16'\n\n/**\n * Provider-specific options for Gemini Veo video generation.\n *\n * Derived from the SDK's `GenerateVideosConfig`, minus the fields the\n * adapter manages itself:\n * - `durationSeconds` — set via the typed top-level `duration` option\n *   (use `adapter.snapDuration(seconds)` to coerce raw seconds)\n * - `aspectRatio` — set via the top-level `size` option\n * - `lastFrame` / `referenceImages` — set via image parts in the `prompt`\n *   with `metadata.role: 'end_frame'` / `'reference'`\n * - `httpOptions` / `abortSignal` — client-level transport concerns\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoProviderOptions = Omit<\n  GenerateVideosConfig,\n  | 'durationSeconds'\n  | 'aspectRatio'\n  | 'lastFrame'\n  | 'referenceImages'\n  | 'httpOptions'\n  | 'abortSignal'\n>\n\n/**\n * Model-specific provider options mapping.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelProviderOptionsByName = {\n  [TModel in GeminiVideoModel]: GeminiVideoProviderOptions\n}\n\n/**\n * Model-specific size (aspect ratio) mapping.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelSizeByName = {\n  [TModel in GeminiVideoModel]: GeminiVideoSize\n}\n\n/**\n * Per-model prompt input modalities. Every Veo model accepts image\n * conditioning inputs (first frame, last frame, reference images) alongside\n * the text prompt.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelInputModalitiesByName = {\n  [TModel in GeminiVideoModel]: readonly ['image']\n}\n\n/**\n * Per-model duration unions (seconds, as numbers — the API's\n * `parameters.durationSeconds` field is numeric).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelDurationByName = {\n  'veo-3.1-generate-preview': 4 | 6 | 8\n  'veo-3.1-fast-generate-preview': 4 | 6 | 8\n  'veo-3.1-lite-generate-preview': 4 | 6 | 8\n}\n\n/**\n * Runtime duration table backing `availableDurations()` / `snapDuration()`.\n *\n * Curated from the official Veo docs\n * (https://ai.google.dev/gemini-api/docs/video) — the Gemini OpenAPI spec\n * types the `:predictLongRunning` request's `parameters` as unconstrained,\n * so it carries no per-model duration information to derive these from.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport const GEMINI_VIDEO_DURATIONS: {\n  readonly [TModel in GeminiVideoModel]: DurationOptions<\n    GeminiVideoModelDurationByName[TModel]\n  >\n} = {\n  'veo-3.1-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'veo-3.1-fast-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'veo-3.1-lite-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n}\n\n/**\n * Look up the duration options for a Veo model.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function getGeminiVideoDurationOptions<TModel extends GeminiVideoModel>(\n  model: TModel,\n): DurationOptions<GeminiVideoModelDurationByName[TModel]> {\n  return GEMINI_VIDEO_DURATIONS[model]\n}\n"],"names":[],"mappings":"AAsGO,MAAM,yBAIT;AAAA,EACF,4BAA4B,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AAAA,EAChE,iCAAiC,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AAAA,EACrE,iCAAiC,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AACvE;AAOO,SAAS,8BACd,OACyD;AACzD,SAAO,uBAAuB,KAAK;AACrC;"}
+{"version":3,"file":"video-provider-options.js","sources":["../../../src/video/video-provider-options.ts"],"sourcesContent":["/**\n * Gemini Video Generation Provider Options\n *\n * Covers two request paths behind the one video adapter:\n * - Veo models — long-running operations via `:predictLongRunning`\n *   (https://ai.google.dev/gemini-api/docs/video)\n * - Gemini Omni Flash — background jobs via the Interactions API\n *   (https://ai.google.dev/gemini-api/docs/omni)\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nimport { GEMINI_INTERACTIONS_VIDEO_MODELS } from '../model-meta'\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type { GenerateVideosConfig, Interactions } from '@google/genai'\nimport type { GEMINI_VIDEO_MODELS } from '../model-meta'\n\n/**\n * Model type for Gemini video generation (Veo + Omni Flash).\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModel = (typeof GEMINI_VIDEO_MODELS)[number]\n\n/**\n * Video models served by the Interactions API (Gemini Omni Flash) rather\n * than Veo's `:predictLongRunning` operations flow.\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport type GeminiInteractionsVideoModel =\n  (typeof GEMINI_INTERACTIONS_VIDEO_MODELS)[number]\n\n/**\n * Runtime guard for the Interactions-served video models.\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport function isInteractionsVideoModel(\n  model: GeminiVideoModel,\n): model is GeminiInteractionsVideoModel {\n  return (GEMINI_INTERACTIONS_VIDEO_MODELS as ReadonlyArray<string>).includes(\n    model,\n  )\n}\n\n/**\n * Supported aspect ratios for Gemini video generation. This is the `size`\n * value for the Gemini video adapter — both Veo and Omni Flash express\n * output shape as an aspect ratio (plus an optional `resolution` in Veo's\n * `modelOptions`), not pixel dimensions.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoSize = '16:9' | '9:16'\n\n/**\n * Provider-specific options for Gemini Veo video generation.\n *\n * Derived from the SDK's `GenerateVideosConfig`, minus the fields the\n * adapter manages itself:\n * - `durationSeconds` — set via the typed top-level `duration` option\n *   (use `adapter.snapDuration(seconds)` to coerce raw seconds)\n * - `aspectRatio` — set via the top-level `size` option\n * - `lastFrame` / `referenceImages` — set via image parts in the `prompt`\n *   with `metadata.role: 'end_frame'` / `'reference'`\n * - `httpOptions` / `abortSignal` — client-level transport concerns\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoProviderOptions = Omit<\n  GenerateVideosConfig,\n  | 'durationSeconds'\n  | 'aspectRatio'\n  | 'lastFrame'\n  | 'referenceImages'\n  | 'httpOptions'\n  | 'abortSignal'\n>\n\n/**\n * Provider-specific options for Gemini Omni Flash video generation on the\n * Interactions API.\n *\n * Derived from the SDK's `Interactions.CreateModelInteractionParamsNonStreaming`,\n * minus the fields the adapter manages itself:\n * - `model` / `input` — set from the adapter's model and the `prompt`\n * - `stream` / `background` — the adapter always creates a background job\n *   and polls it through the `generateVideo` jobs API\n * - `response_modalities` / `response_format` — the adapter requests video\n *   output and maps the top-level `size` option onto\n *   `response_format.aspect_ratio`\n * - `tools` / `response_mime_type` — not applicable to video generation\n *\n * Notable passthroughs:\n * - `previous_interaction_id` — conversational video editing: chain a new\n *   prompt onto a prior Omni interaction to refine its video\n * - `generation_config.video_config.task` — pin the task mode\n *   (`'text_to_video' | 'image_to_video' | 'reference_to_video' | 'edit'`)\n *   instead of letting the model infer it\n *\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport type GeminiOmniVideoProviderOptions = Omit<\n  Interactions.CreateModelInteractionParamsNonStreaming,\n  | 'model'\n  | 'input'\n  | 'stream'\n  | 'background'\n  | 'response_modalities'\n  | 'response_format'\n  | 'response_mime_type'\n  | 'tools'\n>\n\n/**\n * Model-specific provider options mapping.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelProviderOptionsByName = {\n  [TModel in GeminiVideoModel]: TModel extends GeminiInteractionsVideoModel\n    ? GeminiOmniVideoProviderOptions\n    : GeminiVideoProviderOptions\n}\n\n/**\n * Model-specific size (aspect ratio) mapping.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelSizeByName = {\n  [TModel in GeminiVideoModel]: GeminiVideoSize\n}\n\n/**\n * Per-model prompt input modalities. Every Veo model accepts image\n * conditioning inputs (first frame, last frame, reference images) alongside\n * the text prompt. Omni Flash additionally accepts video inputs (short\n * reference clips / videos to edit).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelInputModalitiesByName = {\n  [TModel in GeminiVideoModel]: TModel extends GeminiInteractionsVideoModel\n    ? readonly ['image', 'video']\n    : readonly ['image']\n}\n\n/**\n * Per-model duration unions (seconds, as numbers — Veo's\n * `parameters.durationSeconds` field is numeric; Omni Flash accepts a\n * continuous 3–10 second range, fractional seconds included, so it stays\n * `number` — the adapter rejects out-of-range values at job creation,\n * against the range entry below).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelDurationByName = {\n  'veo-3.1-generate-preview': 4 | 6 | 8\n  'veo-3.1-fast-generate-preview': 4 | 6 | 8\n  'veo-3.1-lite-generate-preview': 4 | 6 | 8\n  'gemini-omni-flash-preview': number\n}\n\n/**\n * Runtime duration table backing `availableDurations()` / `snapDuration()`.\n *\n * Veo values are curated from the official docs\n * (https://ai.google.dev/gemini-api/docs/video) — the Gemini OpenAPI spec\n * types the `:predictLongRunning` request's `parameters` as unconstrained,\n * so it carries no per-model duration information to derive these from.\n * Omni Flash's 3–10s range was verified against the live API\n * (2026-07-02): `response_format.duration` takes a `\"<seconds>s\"` string,\n * fractional values are accepted, out-of-range values are rejected with\n * \"minimum allowed 3s\" / \"maximum allowed 10s\", and omitting it defaults\n * to a 10-second clip.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport const GEMINI_VIDEO_DURATIONS: {\n  readonly [TModel in GeminiVideoModel]: DurationOptions<\n    GeminiVideoModelDurationByName[TModel]\n  >\n} = {\n  'veo-3.1-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'veo-3.1-fast-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'veo-3.1-lite-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'gemini-omni-flash-preview': {\n    kind: 'range',\n    min: 3,\n    max: 10,\n    unit: 'seconds',\n  },\n}\n\n/**\n * Look up the duration options for a Gemini video model.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function getGeminiVideoDurationOptions<TModel extends GeminiVideoModel>(\n  model: TModel,\n): DurationOptions<GeminiVideoModelDurationByName[TModel]> {\n  return GEMINI_VIDEO_DURATIONS[model]\n}\n"],"names":[],"mappings":";AAkCO,SAAS,yBACd,OACuC;AACvC,SAAQ,iCAA2D;AAAA,IACjE;AAAA,EAAA;AAEJ;AAwIO,MAAM,yBAIT;AAAA,EACF,4BAA4B,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AAAA,EAChE,iCAAiC,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AAAA,EACrE,iCAAiC,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AAAA,EACrE,6BAA6B;AAAA,IAC3B,MAAM;AAAA,IACN,KAAK;AAAA,IACL,KAAK;AAAA,IACL,MAAM;AAAA,EAAA;AAEV;AAOO,SAAS,8BACd,OACyD;AACzD,SAAO,uBAAuB,KAAK;AACrC;"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tanstack/ai-gemini",
-  "version": "0.19.1",
+  "version": "0.20.1",
   "description": "Google Gemini adapter for TanStack AI chat, images, speech, audio generation, and structured outputs.",
   "author": "Tanner Linsley",
   "license": "MIT",
@@ -54,18 +54,18 @@
     "text-to-speech"
   ],
   "dependencies": {
-    "@google/genai": "^2.8.0",
+    "@google/genai": "^2.10.0",
     "partial-json": "^0.1.7",
     "@tanstack/ai-utils": "0.3.1"
   },
   "peerDependencies": {
-    "@tanstack/ai": "^0.40.0"
+    "@tanstack/ai": "^0.42.0"
   },
   "devDependencies": {
     "@vitest/coverage-v8": "4.0.14",
     "vite": "^7.3.3",
     "zod": "^4.2.0",
-    "@tanstack/ai": "0.40.0"
+    "@tanstack/ai": "0.42.0"
   },
   "scripts": {
     "build": "vite build",

package/src/adapters/audio.ts

@@ -11,7 +11,7 @@ import type {
   AudioGenerationResult,
 } from '@tanstack/ai'
 import type { GoogleGenAI } from '@google/genai'
-import type { GeminiClientConfig } from '../utils'
+import type { GeminiClientConfig } from '../utils/client'
 
 /**
  * Provider options for Gemini Lyria music generation.