@tanstack/ai-grok 0.18.5 → 0.18.7

package/dist/esm/adapters/video.js

@@ -150,7 +150,6 @@ var GrokVideoAdapter = class extends BaseVideoAdapter {
 		if (referenceAudioCount > 3) throw new Error(`${this.name}: ${model} accepts at most 3 reference voices; received ${referenceAudioCount}.`);
 		if (referenceImageCount > 7) throw new Error(`${this.name}: ${model} accepts at most 7 reference images; received ${referenceImageCount}.`);
 		const [startFrame] = startFrames;
-		if (startFrame && hasReference) throw new Error(`${this.name}: image-to-video and reference-to-video cannot be combined. Use a starting-frame image, or reference images / voices, not both.`);
 		const parsedSize = size !== void 0 ? parseGrokVideoSize(size) : void 0;
 		const resolvedResolution = generationOptions.resolution ?? parsedSize?.resolution;
 		if (hasReference && resolvedResolution === "1080p") throw new Error(`${this.name}: reference-to-video is capped at 720p on ${model}.`);

package/dist/esm/adapters/video.js.map

@@ -1 +1 @@
-{"version":3,"file":"video.js","names":[],"sources":["../../../src/adapters/video.ts"],"sourcesContent":["import { resolveMediaPrompt } from '@tanstack/ai'\nimport { BaseVideoAdapter, snapToDurationOption } from '@tanstack/ai/adapters'\nimport { toRunErrorPayload } from '@tanstack/ai/adapter-internals'\nimport { getGrokApiKeyFromEnv, withGrokDefaults } from '../utils/client'\nimport {\n  GROK_VIDEO_MAX_REFERENCE_AUDIOS,\n  GROK_VIDEO_MAX_REFERENCE_IMAGES,\n  getGrokVideoDurationOptions,\n  isGrokVideoReferenceModel,\n  isGrokVideoSourceModel,\n  parseGrokVideoSize,\n  validateVideoSize,\n} from '../video/video-provider-options'\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type {\n  ImagePart,\n  MediaInputMetadata,\n  TokenUsage,\n  VideoGenerationOptions,\n  VideoJobResult,\n  VideoPart,\n  VideoStatusResult,\n  VideoUrlResult,\n} from '@tanstack/ai'\nimport type { GrokVideoModel } from '../model-meta'\nimport type {\n  GrokVideoModelDurationByName,\n  GrokVideoModelInputModalitiesByName,\n  GrokVideoModelProviderOptionsByName,\n  GrokVideoModelSizeByName,\n  GrokVideoRuntimeOptions,\n} from '../video/video-provider-options'\nimport type { GrokClientConfig } from '../utils/client'\n\n/**\n * Configuration for Grok video adapter.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GrokVideoConfig extends GrokClientConfig {}\n\n/**\n * xAI bills video generation in \"USD ticks\": 10^10 ticks per US dollar\n * (e.g. one grok-imagine-video-1.5 second costs $0.08 = 800_000_000 ticks).\n */\nconst USD_TICKS_PER_DOLLAR = 10_000_000_000\n\n/** Response of the POST /v1/videos/{generations,edits,extensions} endpoints. */\ninterface GrokVideoCreateResponse {\n  request_id?: string\n}\n\n/** Response of GET /v1/videos/{request_id}. */\ninterface GrokVideoStatusResponse {\n  status?: string\n  progress?: number\n  model?: string\n  video?: {\n    url?: string\n    duration?: number\n  }\n  usage?: {\n    cost_in_usd_ticks?: number\n  }\n  error?: string\n}\n\n/**\n * Convert a TanStack image / video part to the URL string accepted by xAI's\n * Imagine video endpoints: public URLs pass through (fetched by xAI's\n * servers), data sources become base64 data URIs.\n */\nfunction mediaPartToUrl(\n  part: ImagePart<MediaInputMetadata> | VideoPart<MediaInputMetadata>,\n): string {\n  if (part.source.type === 'url') return part.source.value\n  return `data:${part.source.mimeType};base64,${part.source.value}`\n}\n\nfunction buildGrokVideoUsage(\n  response: GrokVideoStatusResponse,\n): TokenUsage | undefined {\n  const seconds = response.video?.duration\n  const ticks = response.usage?.cost_in_usd_ticks\n  if (seconds === undefined && ticks === undefined) return undefined\n  return {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n    ...(seconds !== undefined && {\n      billed: { quantity: seconds, unit: 'seconds' },\n      unitsBilled: seconds,\n    }),\n    ...(ticks !== undefined && { cost: ticks / USD_TICKS_PER_DOLLAR }),\n  }\n}\n\n/**\n * Grok Video Generation Adapter (xAI Imagine API)\n *\n * Tree-shakeable adapter for the grok-imagine video models using the\n * async jobs/polling architecture: create a generation request, poll it,\n * then read the completed video URL.\n *\n * Both models support text-to-video and image-to-video;\n * `grok-imagine-video-1.5` is xAI's documented default and adds native\n * 1080p generation plus reference-to-video inputs. Source-video edit\n * and extend are `grok-imagine-video` only.\n *\n * The Imagine video endpoints are not part of the OpenAI SDK surface (and\n * xAI rejects the SDK's multipart paths), so requests are plain JSON calls\n * issued with the configured `fetch` (or the global one).\n *\n * @experimental Video generation is an experimental feature and may change.\n *\n * Features:\n * - Async job-based video generation (1–15 second clips with audio)\n * - Aspect-ratio sizing via the \"aspectRatio_resolution\" size template\n *   (e.g. '16:9_720p'), consistent with the grok-imagine image models\n * - Image-to-video via an `image` prompt part (starting frame URL or data URI)\n * - Reference-to-video via image prompt parts with\n *   `metadata.role: 'reference'` or `'character'` (→ `reference_images`)\n *   and preset voices via `modelOptions.reference_audios`\n *   (grok-imagine-video-1.5 only)\n * - Video editing / extension on `grok-imagine-video` via a source\n *   `video` prompt part and `modelOptions.mode: 'edit' | 'extend'`\n *   (`/v1/videos/edits` / `/v1/videos/extensions`; in extend mode\n *   `duration` is the added tail)\n * - Usage reporting: billed seconds (`usage.billed`) and exact cost\n */\nexport class GrokVideoAdapter<\n  TModel extends GrokVideoModel,\n> extends BaseVideoAdapter<\n  TModel,\n  GrokVideoModelProviderOptionsByName[TModel],\n  GrokVideoModelProviderOptionsByName,\n  GrokVideoModelSizeByName,\n  GrokVideoModelInputModalitiesByName,\n  GrokVideoModelDurationByName\n> {\n  readonly name = 'grok' as const\n\n  private readonly clientConfig: GrokVideoConfig\n\n  constructor(config: GrokVideoConfig, model: TModel) {\n    super({}, model)\n    this.clientConfig = withGrokDefaults(config)\n  }\n\n  private async request(\n    path: string,\n    init?: Omit<RequestInit, 'headers'>,\n  ): Promise<Response> {\n    // workerd's fetch must be invoked with this === globalThis\n    const fetchFn = this.clientConfig.fetch ?? globalThis.fetch.bind(globalThis)\n    return await fetchFn(`${this.clientConfig.baseURL}${path}`, {\n      ...init,\n      headers: {\n        'Content-Type': 'application/json',\n        Authorization: `Bearer ${this.clientConfig.apiKey}`,\n      },\n    })\n  }\n\n  /**\n   * Reads the error message out of an Imagine API error body\n   * (`{\"code\": \"...\", \"error\": \"...\"}`), falling back to the raw text.\n   */\n  private async errorMessage(response: Response): Promise<string> {\n    const body = await response.text()\n    try {\n      const parsed: unknown = JSON.parse(body)\n      if (\n        typeof parsed === 'object' &&\n        parsed !== null &&\n        'error' in parsed &&\n        typeof parsed.error === 'string'\n      ) {\n        return parsed.error\n      }\n    } catch {\n      // not JSON — fall through to the raw body\n    }\n    return body\n  }\n\n  async createVideoJob(\n    options: VideoGenerationOptions<\n      GrokVideoModelProviderOptionsByName[TModel],\n      GrokVideoModelSizeByName[TModel],\n      GrokVideoModelDurationByName[TModel]\n    >,\n  ): Promise<VideoJobResult> {\n    const { model, size, modelOptions, logger } = options\n\n    // `mode` is a routing hint for this adapter, not an API field — strip it\n    // before the remaining options are spread onto the request body. The\n    // per-model map narrows what callers can pass, but modelOptions often\n    // arrives as deserialized JSON, so the adapter handles the widest option\n    // surface (the 1.5 shape) uniformly and gates by model at runtime.\n    const { mode, ...wireOptions } = (modelOptions ??\n      {}) as GrokVideoRuntimeOptions\n\n    // `mode` is typed 'edit' | 'extend' but reaches us untrusted from JSON\n    // callers. An unrecognised value must not fall through to the\n    // generations endpoint with a source-video body — that would silently\n    // run (and bill) a generation the caller never asked for.\n    if (mode !== undefined && mode !== 'edit' && mode !== 'extend') {\n      throw new Error(\n        `${this.name}: unknown modelOptions.mode '${String(mode)}'. ` +\n          `Expected 'edit' or 'extend'.`,\n      )\n    }\n\n    // The interleaved prompt decomposes into verbatim text plus typed media\n    // buckets. Reference audio is voice-id based (not an audio file), so\n    // audio prompt parts have no request field to land in.\n    const resolved = resolveMediaPrompt(options.prompt)\n    if (resolved.audios.length > 0) {\n      throw new Error(\n        `${this.name}.createVideoJob does not support audio prompt parts (model: ${model}). ` +\n          `To reference a preset voice, pass modelOptions.reference_audios ` +\n          `(e.g. [{ voice_id: 'eve' }]).`,\n      )\n    }\n\n    // A video prompt part is the source clip for edit / extension mode.\n    // Those endpoints are grok-imagine-video only — 1.5 has no video input.\n    if (\n      !isGrokVideoSourceModel(model) &&\n      (mode !== undefined || resolved.videos.length > 0)\n    ) {\n      throw new Error(\n        `${this.name}: ${model} does not support video editing or extension. ` +\n          `Use 'grok-imagine-video' for /v1/videos/edits and /v1/videos/extensions.`,\n      )\n    }\n\n    // The mode must be chosen explicitly because the two endpoints have\n    // different semantics (edit rewrites the clip, extend appends\n    // `duration` seconds).\n    if (resolved.videos.length > 1) {\n      throw new Error(\n        `${this.name}: ${model} accepts at most one source video; received ${resolved.videos.length}.`,\n      )\n    }\n    const [sourceVideo] = resolved.videos\n    if (sourceVideo && mode === undefined) {\n      throw new Error(\n        `${this.name}: a video prompt part needs modelOptions.mode set to ` +\n          `'edit' (rewrite the clip) or 'extend' (append to it).`,\n      )\n    }\n    if (!sourceVideo && mode !== undefined) {\n      throw new Error(\n        `${this.name}: modelOptions.mode '${mode}' requires a video prompt ` +\n          `part carrying the source clip.`,\n      )\n    }\n\n    if (mode !== undefined && sourceVideo) {\n      return await this.createSourceVideoJob({\n        model,\n        mode,\n        sourceVideo,\n        resolved,\n        wireOptions,\n        size,\n        genericDuration: options.duration,\n        logger,\n      })\n    }\n\n    validateVideoSize(model, size)\n\n    // Pull the specially-handled keys out of the wire options: `duration`\n    // is folded into the snapped value below, and the reference fields are\n    // re-added explicitly so a JSON-serialized `null` or empty array reads\n    // as \"unset\" instead of leaking onto the wire.\n    const {\n      duration: rawOptionDuration,\n      reference_images: explicitReferenceImages,\n      reference_audios: referenceAudios,\n      ...generationOptions\n    } = wireOptions\n\n    // Coerce the requested duration into the model's valid range (1–15s,\n    // integer) instead of rejecting it — `snapDuration` clamps and rounds.\n    // modelOptions wins over the generic `duration`, mirroring the size\n    // precedence below.\n    const rawDuration = rawOptionDuration ?? options.duration\n    const duration =\n      rawDuration != null ? this.snapDuration(rawDuration) : undefined\n\n    // Image parts split by role: un-roled / 'start_frame' images become the\n    // starting frame (image-to-video); 'reference' / 'character' images\n    // become reference_images (reference-to-video). The Imagine API has no\n    // mask / control / end-frame inputs. Unknown role strings (possible via\n    // JSON callers) throw rather than silently dropping the part.\n    const startFrames: Array<ImagePart<MediaInputMetadata>> = []\n    const referenceImages: Array<{ url: string }> = []\n    for (const part of resolved.images) {\n      const role = part.metadata?.role\n      switch (role) {\n        case 'mask':\n        case 'control':\n        case 'end_frame':\n          throw new Error(\n            `${this.name}: the Imagine video API has no '${role}' image ` +\n              `input on model ${model}. Use an un-roled / 'start_frame' ` +\n              `image as the starting frame, or 'reference' images.`,\n          )\n        case 'reference':\n        case 'character':\n          referenceImages.push({ url: mediaPartToUrl(part) })\n          break\n        case 'start_frame':\n        case undefined:\n          startFrames.push(part)\n          break\n        default:\n          throw new Error(\n            `${this.name}: unknown image metadata.role '${String(role)}'. ` +\n              `Expected 'start_frame', 'reference', or 'character'.`,\n          )\n      }\n    }\n    if (startFrames.length > 1) {\n      throw new Error(\n        `${this.name}: ${model} accepts at most one starting-frame image; received ${startFrames.length}. ` +\n          `Use metadata.role: 'reference' for reference-to-video inputs.`,\n      )\n    }\n    // Explicit modelOptions.reference_images replaces the part-derived list\n    // (an explicit empty array means \"none\").\n    const finalReferenceImages =\n      explicitReferenceImages ??\n      (referenceImages.length > 0 ? referenceImages : undefined)\n    const referenceImageCount = finalReferenceImages?.length ?? 0\n    const referenceAudioCount = referenceAudios?.length ?? 0\n    const hasReference = referenceImageCount > 0 || referenceAudioCount > 0\n\n    // Reference inputs are a grok-imagine-video-1.5 feature. The per-model\n    // options map already hides the fields from other models at compile\n    // time; this runtime gate covers prompt-part roles and untyped callers.\n    if (!isGrokVideoReferenceModel(model) && hasReference) {\n      throw new Error(\n        `${this.name}: ${model} does not support reference-to-video inputs. ` +\n          `Use 'grok-imagine-video-1.5' for reference_images / reference_audios.`,\n      )\n    }\n    if (referenceAudioCount > GROK_VIDEO_MAX_REFERENCE_AUDIOS) {\n      throw new Error(\n        `${this.name}: ${model} accepts at most ${GROK_VIDEO_MAX_REFERENCE_AUDIOS} reference voices; received ${referenceAudioCount}.`,\n      )\n    }\n    if (referenceImageCount > GROK_VIDEO_MAX_REFERENCE_IMAGES) {\n      throw new Error(\n        `${this.name}: ${model} accepts at most ${GROK_VIDEO_MAX_REFERENCE_IMAGES} reference images; received ${referenceImageCount}.`,\n      )\n    }\n\n    // Image-to-video: the single image prompt part becomes the starting frame\n    // and the prompt text describes the desired motion. URL sources are\n    // fetched by xAI's servers; data sources are sent as base64 data URIs.\n    const [startFrame] = startFrames\n\n    // xAI rejects `image` + `reference_images` / `reference_audios` as a\n    // 400: only one of image-to-video or reference-to-video can be active.\n    if (startFrame && hasReference) {\n      throw new Error(\n        `${this.name}: image-to-video and reference-to-video cannot be combined. ` +\n          `Use a starting-frame image, or reference images / voices, not both.`,\n      )\n    }\n\n    // The generic `size` option carries an \"aspectRatio_resolution\" template\n    // (e.g. '16:9_720p') and maps to the Imagine API's `aspect_ratio` /\n    // `resolution` parameters; explicit modelOptions win over the template\n    // (including `reference_images`, which replaces the part-derived list).\n    const parsedSize = size !== undefined ? parseGrokVideoSize(size) : undefined\n    const resolvedResolution =\n      generationOptions.resolution ?? parsedSize?.resolution\n    if (hasReference && resolvedResolution === '1080p') {\n      throw new Error(\n        `${this.name}: reference-to-video is capped at 720p on ${model}.`,\n      )\n    }\n    const request = {\n      model,\n      prompt: resolved.text,\n      ...(startFrame && { image: { url: mediaPartToUrl(startFrame) } }),\n      ...(referenceImageCount > 0 && {\n        reference_images: finalReferenceImages,\n      }),\n      ...(referenceAudioCount > 0 && {\n        reference_audios: referenceAudios,\n      }),\n      ...(parsedSize && {\n        aspect_ratio: parsedSize.aspectRatio,\n        ...(parsedSize.resolution !== undefined && {\n          resolution: parsedSize.resolution,\n        }),\n      }),\n      // The remaining options spread after the size template so explicit\n      // aspect_ratio / resolution win over it; duration and the reference\n      // fields were destructured out above and re-added normalized.\n      ...generationOptions,\n      ...(duration !== undefined && { duration }),\n    }\n\n    return await this.postVideoJob('/videos/generations', request, {\n      model,\n      logger,\n      logLine: `activity=video.create provider=${this.name} model=${model} mode=generate size=${size ?? 'default'} duration=${duration ?? 'default'}`,\n    })\n  }\n\n  /**\n   * Build and post an edit / extension request. Both endpoints take only\n   * `model`, `prompt`, and the source `video` (plus `duration` — the length\n   * of the added tail — for extensions): output geometry is inherited from\n   * the source clip, capped at 720p, and edit outputs also inherit the\n   * source length. Rather than sending fields the API documents as ignored,\n   * the inapplicable options are rejected with actionable errors.\n   */\n  private async createSourceVideoJob(args: {\n    model: string\n    mode: 'edit' | 'extend'\n    sourceVideo: VideoPart<MediaInputMetadata>\n    resolved: ReturnType<typeof resolveMediaPrompt>\n    wireOptions: Omit<GrokVideoRuntimeOptions, 'mode'>\n    size: string | undefined\n    genericDuration: number | undefined\n    logger: VideoGenerationOptions<GrokVideoRuntimeOptions>['logger']\n  }): Promise<VideoJobResult> {\n    const { model, mode, sourceVideo, resolved, wireOptions, logger } = args\n    const endpoint = mode === 'edit' ? '/videos/edits' : '/videos/extensions'\n\n    if (resolved.images.length > 0) {\n      throw new Error(\n        `${this.name}: '${mode}' mode takes only the source video — image ` +\n          `prompt parts are not supported by ${endpoint}.`,\n      )\n    }\n\n    // Pull every generation-only key out of the wire options so nothing can\n    // leak into the edit/extend body via the spread below. JSON-serialized\n    // `null` values (a common \"unset\" encoding) are treated as absent;\n    // actual values are rejected with actionable errors.\n    const {\n      aspect_ratio: aspectRatio,\n      resolution,\n      duration: modeDuration,\n      reference_images: referenceImagesOption,\n      reference_audios: referenceAudiosOption,\n      ...passthrough\n    } = wireOptions\n    if (\n      (referenceImagesOption?.length ?? 0) > 0 ||\n      (referenceAudiosOption?.length ?? 0) > 0\n    ) {\n      throw new Error(\n        `${this.name}: reference inputs are only supported by video ` +\n          `generation, not '${mode}' mode.`,\n      )\n    }\n    if (args.size !== undefined || aspectRatio != null || resolution != null) {\n      throw new Error(\n        `${this.name}: '${mode}' mode does not accept size / aspect_ratio / ` +\n          `resolution — the output inherits the source clip's geometry ` +\n          `(capped at 720p).`,\n      )\n    }\n    const rawDuration = modeDuration ?? args.genericDuration\n    if (mode === 'edit' && rawDuration != null) {\n      throw new Error(\n        `${this.name}: 'edit' mode does not accept a duration — the output ` +\n          `inherits the source clip's length. Use mode 'extend' to append ` +\n          `seconds to the clip.`,\n      )\n    }\n    // Extend: the snapped duration is the added-tail length (1–15s).\n    const duration =\n      rawDuration != null ? this.snapDuration(rawDuration) : undefined\n\n    const request = {\n      model,\n      prompt: resolved.text,\n      video: { url: mediaPartToUrl(sourceVideo) },\n      ...passthrough,\n      ...(duration !== undefined && { duration }),\n    }\n\n    return await this.postVideoJob(endpoint, request, {\n      model,\n      logger,\n      logLine: `activity=video.create provider=${this.name} model=${model} mode=${mode} duration=${duration ?? 'default'}`,\n    })\n  }\n\n  /**\n   * POST a create-job request body to one of the Imagine video endpoints\n   * (`/videos/generations`, `/videos/edits`, `/videos/extensions`) and read\n   * the `request_id` out of the shared response shape.\n   */\n  private async postVideoJob(\n    endpoint: string,\n    request: Record<string, unknown>,\n    context: {\n      model: string\n      logger: VideoGenerationOptions<GrokVideoRuntimeOptions>['logger']\n      logLine: string\n    },\n  ): Promise<VideoJobResult> {\n    const { model, logger, logLine } = context\n    try {\n      logger.request(logLine, { provider: this.name, model })\n\n      const response = await this.request(endpoint, {\n        method: 'POST',\n        body: JSON.stringify(request),\n      })\n      if (!response.ok) {\n        throw new Error(\n          `grok: ${endpoint} request failed (${response.status} ${response.statusText}): ${await this.errorMessage(response)}`,\n        )\n      }\n\n      const result = (await response.json()) as GrokVideoCreateResponse\n      if (!result.request_id) {\n        throw new Error(`grok: ${endpoint} response contained no request_id`)\n      }\n      return { jobId: result.request_id, model }\n    } catch (error: unknown) {\n      logger.errors(`${this.name}.createVideoJob fatal`, {\n        error: toRunErrorPayload(error, `${this.name}.createVideoJob failed`),\n        source: `${this.name}.createVideoJob`,\n      })\n      throw error\n    }\n  }\n\n  private async retrieveJob(jobId: string): Promise<GrokVideoStatusResponse> {\n    const response = await this.request(`/videos/${jobId}`)\n    if (!response.ok) {\n      const error = new Error(\n        `grok: video status request failed (${response.status} ${response.statusText}): ${await this.errorMessage(response)}`,\n      )\n      ;(error as { status?: number }).status = response.status\n      throw error\n    }\n    return (await response.json()) as GrokVideoStatusResponse\n  }\n\n  async getVideoStatus(jobId: string): Promise<VideoStatusResult> {\n    let response: GrokVideoStatusResponse\n    try {\n      response = await this.retrieveJob(jobId)\n    } catch (error) {\n      if ((error as { status?: number }).status === 404) {\n        return { jobId, status: 'failed', error: 'Job not found' }\n      }\n      throw error\n    }\n\n    return {\n      jobId,\n      status: this.mapStatus(response.status),\n      ...(response.progress !== undefined && { progress: response.progress }),\n      ...(response.error !== undefined && { error: response.error }),\n    }\n  }\n\n  async getVideoUrl(jobId: string): Promise<VideoUrlResult> {\n    let response: GrokVideoStatusResponse\n    try {\n      response = await this.retrieveJob(jobId)\n    } catch (error) {\n      if ((error as { status?: number }).status === 404) {\n        throw new Error(`Video job not found: ${jobId}`)\n      }\n      throw error\n    }\n\n    const status = this.mapStatus(response.status)\n    if (status === 'failed') {\n      throw new Error(\n        `Video generation failed${response.error ? `: ${response.error}` : ''}. Job ID: ${jobId}`,\n      )\n    }\n    const url = response.video?.url\n    if (!url) {\n      throw new Error(\n        `Video is not ready for download. Check status first. Job ID: ${jobId}`,\n      )\n    }\n\n    const usage = buildGrokVideoUsage(response)\n    return {\n      jobId,\n      url,\n      ...(usage && { usage }),\n    }\n  }\n\n  /**\n   * Maps Imagine API job statuses onto the generic video status set. The\n   * API reports 'pending' while queued/generating (with a numeric\n   * `progress`), then a terminal 'done' / 'failed' / 'expired'.\n   */\n  protected mapStatus(\n    apiStatus: string | undefined,\n  ): 'pending' | 'processing' | 'completed' | 'failed' {\n    switch (apiStatus) {\n      case 'pending':\n      case 'queued':\n        return 'pending'\n      case 'done':\n      case 'completed':\n      case 'succeeded':\n        return 'completed'\n      case 'failed':\n      case 'expired':\n      case 'error':\n      case 'cancelled':\n        return 'failed'\n      case undefined:\n      default:\n        return 'processing'\n    }\n  }\n\n  /**\n   * Both grok-imagine video models accept a continuous 1–15 integer-second\n   * range. Consumers can use this to render UI without provider knowledge.\n   */\n  override availableDurations(): DurationOptions<\n    GrokVideoModelDurationByName[TModel]\n  > {\n    return getGrokVideoDurationOptions(this.model)\n  }\n\n  /**\n   * Coerce a raw seconds value to the closest valid duration (clamped to\n   * [1, 15] and rounded to whole seconds).\n   */\n  override snapDuration(\n    seconds: number,\n  ): GrokVideoModelDurationByName[TModel] | undefined {\n    return snapToDurationOption(seconds, this.availableDurations())\n  }\n}\n\n/**\n * Creates a Grok video adapter with an explicit API key.\n * Type resolution happens here at the call site.\n *\n * @experimental Video generation is an experimental feature and may change.\n *\n * @param model - The model name (e.g., 'grok-imagine-video-1.5')\n * @param apiKey - Your xAI API key\n * @param config - Optional additional configuration\n * @returns Configured Grok video adapter instance with resolved types\n *\n * @example\n * ```typescript\n * const adapter = createGrokVideo('grok-imagine-video-1.5', 'xai-...');\n *\n * const { jobId } = await generateVideo({\n *   adapter,\n *   prompt: 'A beautiful sunset over the ocean',\n *   size: '16:9_720p',\n *   duration: 5\n * });\n * ```\n */\nexport function createGrokVideo<TModel extends GrokVideoModel>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GrokVideoConfig, 'apiKey'>,\n): GrokVideoAdapter<TModel> {\n  return new GrokVideoAdapter({ apiKey, ...config }, model)\n}\n\n/**\n * Creates a Grok video adapter with automatic API key detection from environment variables.\n * Type resolution happens here at the call site.\n *\n * Looks for `XAI_API_KEY` in:\n * - `process.env` (Node.js)\n * - `window.env` (Browser with injected env)\n *\n * @experimental Video generation is an experimental feature and may change.\n *\n * @param model - The model name (e.g., 'grok-imagine-video-1.5')\n * @param config - Optional configuration (excluding apiKey which is auto-detected)\n * @returns Configured Grok video adapter instance with resolved types\n * @throws Error if XAI_API_KEY is not found in environment\n *\n * @example\n * ```typescript\n * // Automatically uses XAI_API_KEY from environment\n * const adapter = grokVideo('grok-imagine-video-1.5');\n *\n * // Image-to-video: an optional image prompt part is the starting frame.\n * const { jobId } = await generateVideo({\n *   adapter,\n *   prompt: [\n *     { type: 'text', content: 'Make the cat start playing the piano' },\n *     { type: 'image', source: { type: 'url', value: 'https://example.com/cat.png' } },\n *   ],\n * });\n *\n * // Poll for status\n * const status = await getVideoJobStatus({ adapter, jobId });\n * ```\n */\nexport function grokVideo<TModel extends GrokVideoModel>(\n  model: TModel,\n  config?: Omit<GrokVideoConfig, 'apiKey'>,\n): GrokVideoAdapter<TModel> {\n  const apiKey = getGrokApiKeyFromEnv()\n  return createGrokVideo(model, apiKey, config)\n}\n"],"mappings":";;;;;;;;;;AA6CA,IAAM,uBAAuB;;;;;;AA2B7B,SAAS,eACP,MACQ;CACR,IAAI,KAAK,OAAO,SAAS,OAAO,OAAO,KAAK,OAAO;CACnD,OAAO,QAAQ,KAAK,OAAO,SAAS,UAAU,KAAK,OAAO;AAC5D;AAEA,SAAS,oBACP,UACwB;CACxB,MAAM,UAAU,SAAS,OAAO;CAChC,MAAM,QAAQ,SAAS,OAAO;CAC9B,IAAI,YAAY,KAAA,KAAa,UAAU,KAAA,GAAW,OAAO,KAAA;CACzD,OAAO;EACL,cAAc;EACd,kBAAkB;EAClB,aAAa;EACb,GAAI,YAAY,KAAA,KAAa;GAC3B,QAAQ;IAAE,UAAU;IAAS,MAAM;GAAU;GAC7C,aAAa;EACf;EACA,GAAI,UAAU,KAAA,KAAa,EAAE,MAAM,QAAQ,qBAAqB;CAClE;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCA,IAAa,mBAAb,cAEU,iBAOR;CACA,OAAgB;CAEhB;CAEA,YAAY,QAAyB,OAAe;EAClD,MAAM,CAAC,GAAG,KAAK;EACf,KAAK,eAAe,iBAAiB,MAAM;CAC7C;CAEA,MAAc,QACZ,MACA,MACmB;EAGnB,OAAO,OADS,KAAK,aAAa,SAAS,WAAW,MAAM,KAAK,UAAU,EAAA,CACtD,GAAG,KAAK,aAAa,UAAU,QAAQ;GAC1D,GAAG;GACH,SAAS;IACP,gBAAgB;IAChB,eAAe,UAAU,KAAK,aAAa;GAC7C;EACF,CAAC;CACH;;;;;CAMA,MAAc,aAAa,UAAqC;EAC9D,MAAM,OAAO,MAAM,SAAS,KAAK;EACjC,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,IAAI;GACvC,IACE,OAAO,WAAW,YAClB,WAAW,QACX,WAAW,UACX,OAAO,OAAO,UAAU,UAExB,OAAO,OAAO;EAElB,QAAQ,CAER;EACA,OAAO;CACT;CAEA,MAAM,eACJ,SAKyB;EACzB,MAAM,EAAE,OAAO,MAAM,cAAc,WAAW;EAO9C,MAAM,EAAE,MAAM,GAAG,gBAAiB,gBAChC,CAAC;EAMH,IAAI,SAAS,KAAA,KAAa,SAAS,UAAU,SAAS,UACpD,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,+BAA+B,OAAO,IAAI,EAAE,gCAE3D;EAMF,MAAM,WAAW,mBAAmB,QAAQ,MAAM;EAClD,IAAI,SAAS,OAAO,SAAS,GAC3B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,8DAA8D,MAAM,iGAGnF;EAKF,IACE,CAAC,uBAAuB,KAAK,MAC5B,SAAS,KAAA,KAAa,SAAS,OAAO,SAAS,IAEhD,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,uHAEzB;EAMF,IAAI,SAAS,OAAO,SAAS,GAC3B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,8CAA8C,SAAS,OAAO,OAAO,EAC9F;EAEF,MAAM,CAAC,eAAe,SAAS;EAC/B,IAAI,eAAe,SAAS,KAAA,GAC1B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,2GAEf;EAEF,IAAI,CAAC,eAAe,SAAS,KAAA,GAC3B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,uBAAuB,KAAK,yDAE3C;EAGF,IAAI,SAAS,KAAA,KAAa,aACxB,OAAO,MAAM,KAAK,qBAAqB;GACrC;GACA;GACA;GACA;GACA;GACA;GACA,iBAAiB,QAAQ;GACzB;EACF,CAAC;EAGH,kBAAkB,OAAO,IAAI;EAM7B,MAAM,EACJ,UAAU,mBACV,kBAAkB,yBAClB,kBAAkB,iBAClB,GAAG,sBACD;EAMJ,MAAM,cAAc,qBAAqB,QAAQ;EACjD,MAAM,WACJ,eAAe,OAAO,KAAK,aAAa,WAAW,IAAI,KAAA;EAOzD,MAAM,cAAoD,CAAC;EAC3D,MAAM,kBAA0C,CAAC;EACjD,KAAK,MAAM,QAAQ,SAAS,QAAQ;GAClC,MAAM,OAAO,KAAK,UAAU;GAC5B,QAAQ,MAAR;IACE,KAAK;IACL,KAAK;IACL,KAAK,aACH,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,kCAAkC,KAAK,yBAChC,MAAM,sFAE5B;IACF,KAAK;IACL,KAAK;KACH,gBAAgB,KAAK,EAAE,KAAK,eAAe,IAAI,EAAE,CAAC;KAClD;IACF,KAAK;IACL,KAAK,KAAA;KACH,YAAY,KAAK,IAAI;KACrB;IACF,SACE,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,iCAAiC,OAAO,IAAI,EAAE,wDAE7D;GACJ;EACF;EACA,IAAI,YAAY,SAAS,GACvB,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,sDAAsD,YAAY,OAAO,gEAElG;EAIF,MAAM,uBACJ,4BACC,gBAAgB,SAAS,IAAI,kBAAkB,KAAA;EAClD,MAAM,sBAAsB,sBAAsB,UAAU;EAC5D,MAAM,sBAAsB,iBAAiB,UAAU;EACvD,MAAM,eAAe,sBAAsB,KAAK,sBAAsB;EAKtE,IAAI,CAAC,0BAA0B,KAAK,KAAK,cACvC,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,mHAEzB;EAEF,IAAI,sBAAA,GACF,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,gDAAiF,oBAAoB,EAC9H;EAEF,IAAI,sBAAA,GACF,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,gDAAiF,oBAAoB,EAC9H;EAMF,MAAM,CAAC,cAAc;EAIrB,IAAI,cAAc,cAChB,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,gIAEf;EAOF,MAAM,aAAa,SAAS,KAAA,IAAY,mBAAmB,IAAI,IAAI,KAAA;EACnE,MAAM,qBACJ,kBAAkB,cAAc,YAAY;EAC9C,IAAI,gBAAgB,uBAAuB,SACzC,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,4CAA4C,MAAM,EACjE;EAEF,MAAM,UAAU;GACd;GACA,QAAQ,SAAS;GACjB,GAAI,cAAc,EAAE,OAAO,EAAE,KAAK,eAAe,UAAU,EAAE,EAAE;GAC/D,GAAI,sBAAsB,KAAK,EAC7B,kBAAkB,qBACpB;GACA,GAAI,sBAAsB,KAAK,EAC7B,kBAAkB,gBACpB;GACA,GAAI,cAAc;IAChB,cAAc,WAAW;IACzB,GAAI,WAAW,eAAe,KAAA,KAAa,EACzC,YAAY,WAAW,WACzB;GACF;GAIA,GAAG;GACH,GAAI,aAAa,KAAA,KAAa,EAAE,SAAS;EAC3C;EAEA,OAAO,MAAM,KAAK,aAAa,uBAAuB,SAAS;GAC7D;GACA;GACA,SAAS,kCAAkC,KAAK,KAAK,SAAS,MAAM,sBAAsB,QAAQ,UAAU,YAAY,YAAY;EACtI,CAAC;CACH;;;;;;;;;CAUA,MAAc,qBAAqB,MASP;EAC1B,MAAM,EAAE,OAAO,MAAM,aAAa,UAAU,aAAa,WAAW;EACpE,MAAM,WAAW,SAAS,SAAS,kBAAkB;EAErD,IAAI,SAAS,OAAO,SAAS,GAC3B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,KAAK,KAAK,+EACgB,SAAS,EAClD;EAOF,MAAM,EACJ,cAAc,aACd,YACA,UAAU,cACV,kBAAkB,uBAClB,kBAAkB,uBAClB,GAAG,gBACD;EACJ,KACG,uBAAuB,UAAU,KAAK,MACtC,uBAAuB,UAAU,KAAK,GAEvC,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,kEACS,KAAK,QAC7B;EAEF,IAAI,KAAK,SAAS,KAAA,KAAa,eAAe,QAAQ,cAAc,MAClE,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,KAAK,KAAK,2HAGzB;EAEF,MAAM,cAAc,gBAAgB,KAAK;EACzC,IAAI,SAAS,UAAU,eAAe,MACpC,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,0IAGf;EAGF,MAAM,WACJ,eAAe,OAAO,KAAK,aAAa,WAAW,IAAI,KAAA;EAEzD,MAAM,UAAU;GACd;GACA,QAAQ,SAAS;GACjB,OAAO,EAAE,KAAK,eAAe,WAAW,EAAE;GAC1C,GAAG;GACH,GAAI,aAAa,KAAA,KAAa,EAAE,SAAS;EAC3C;EAEA,OAAO,MAAM,KAAK,aAAa,UAAU,SAAS;GAChD;GACA;GACA,SAAS,kCAAkC,KAAK,KAAK,SAAS,MAAM,QAAQ,KAAK,YAAY,YAAY;EAC3G,CAAC;CACH;;;;;;CAOA,MAAc,aACZ,UACA,SACA,SAKyB;EACzB,MAAM,EAAE,OAAO,QAAQ,YAAY;EACnC,IAAI;GACF,OAAO,QAAQ,SAAS;IAAE,UAAU,KAAK;IAAM;GAAM,CAAC;GAEtD,MAAM,WAAW,MAAM,KAAK,QAAQ,UAAU;IAC5C,QAAQ;IACR,MAAM,KAAK,UAAU,OAAO;GAC9B,CAAC;GACD,IAAI,CAAC,SAAS,IACZ,MAAM,IAAI,MACR,SAAS,SAAS,mBAAmB,SAAS,OAAO,GAAG,SAAS,WAAW,KAAK,MAAM,KAAK,aAAa,QAAQ,GACnH;GAGF,MAAM,SAAU,MAAM,SAAS,KAAK;GACpC,IAAI,CAAC,OAAO,YACV,MAAM,IAAI,MAAM,SAAS,SAAS,kCAAkC;GAEtE,OAAO;IAAE,OAAO,OAAO;IAAY;GAAM;EAC3C,SAAS,OAAgB;GACvB,OAAO,OAAO,GAAG,KAAK,KAAK,wBAAwB;IACjD,OAAO,kBAAkB,OAAO,GAAG,KAAK,KAAK,uBAAuB;IACpE,QAAQ,GAAG,KAAK,KAAK;GACvB,CAAC;GACD,MAAM;EACR;CACF;CAEA,MAAc,YAAY,OAAiD;EACzE,MAAM,WAAW,MAAM,KAAK,QAAQ,WAAW,OAAO;EACtD,IAAI,CAAC,SAAS,IAAI;GAChB,MAAM,wBAAQ,IAAI,MAChB,sCAAsC,SAAS,OAAO,GAAG,SAAS,WAAW,KAAK,MAAM,KAAK,aAAa,QAAQ,GACpH;GACC,MAA+B,SAAS,SAAS;GAClD,MAAM;EACR;EACA,OAAQ,MAAM,SAAS,KAAK;CAC9B;CAEA,MAAM,eAAe,OAA2C;EAC9D,IAAI;EACJ,IAAI;GACF,WAAW,MAAM,KAAK,YAAY,KAAK;EACzC,SAAS,OAAO;GACd,IAAK,MAA8B,WAAW,KAC5C,OAAO;IAAE;IAAO,QAAQ;IAAU,OAAO;GAAgB;GAE3D,MAAM;EACR;EAEA,OAAO;GACL;GACA,QAAQ,KAAK,UAAU,SAAS,MAAM;GACtC,GAAI,SAAS,aAAa,KAAA,KAAa,EAAE,UAAU,SAAS,SAAS;GACrE,GAAI,SAAS,UAAU,KAAA,KAAa,EAAE,OAAO,SAAS,MAAM;EAC9D;CACF;CAEA,MAAM,YAAY,OAAwC;EACxD,IAAI;EACJ,IAAI;GACF,WAAW,MAAM,KAAK,YAAY,KAAK;EACzC,SAAS,OAAO;GACd,IAAK,MAA8B,WAAW,KAC5C,MAAM,IAAI,MAAM,wBAAwB,OAAO;GAEjD,MAAM;EACR;EAGA,IADe,KAAK,UAAU,SAAS,MACnC,MAAW,UACb,MAAM,IAAI,MACR,0BAA0B,SAAS,QAAQ,KAAK,SAAS,UAAU,GAAG,YAAY,OACpF;EAEF,MAAM,MAAM,SAAS,OAAO;EAC5B,IAAI,CAAC,KACH,MAAM,IAAI,MACR,gEAAgE,OAClE;EAGF,MAAM,QAAQ,oBAAoB,QAAQ;EAC1C,OAAO;GACL;GACA;GACA,GAAI,SAAS,EAAE,MAAM;EACvB;CACF;;;;;;CAOA,UACE,WACmD;EACnD,QAAQ,WAAR;GACE,KAAK;GACL,KAAK,UACH,OAAO;GACT,KAAK;GACL,KAAK;GACL,KAAK,aACH,OAAO;GACT,KAAK;GACL,KAAK;GACL,KAAK;GACL,KAAK,aACH,OAAO;GACT,KAAK,KAAA;GACL,SACE,OAAO;EACX;CACF;;;;;CAMA,qBAEE;EACA,OAAO,4BAA4B,KAAK,KAAK;CAC/C;;;;;CAMA,aACE,SACkD;EAClD,OAAO,qBAAqB,SAAS,KAAK,mBAAmB,CAAC;CAChE;AACF;;;;;;;;;;;;;;;;;;;;;;;;AAyBA,SAAgB,gBACd,OACA,QACA,QAC0B;CAC1B,OAAO,IAAI,iBAAiB;EAAE;EAAQ,GAAG;CAAO,GAAG,KAAK;AAC1D;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCA,SAAgB,UACd,OACA,QAC0B;CAE1B,OAAO,gBAAgB,OADR,qBACe,GAAQ,MAAM;AAC9C"}
+{"version":3,"file":"video.js","names":[],"sources":["../../../src/adapters/video.ts"],"sourcesContent":["import { resolveMediaPrompt } from '@tanstack/ai'\nimport { BaseVideoAdapter, snapToDurationOption } from '@tanstack/ai/adapters'\nimport { toRunErrorPayload } from '@tanstack/ai/adapter-internals'\nimport { getGrokApiKeyFromEnv, withGrokDefaults } from '../utils/client'\nimport {\n  GROK_VIDEO_MAX_REFERENCE_AUDIOS,\n  GROK_VIDEO_MAX_REFERENCE_IMAGES,\n  getGrokVideoDurationOptions,\n  isGrokVideoReferenceModel,\n  isGrokVideoSourceModel,\n  parseGrokVideoSize,\n  validateVideoSize,\n} from '../video/video-provider-options'\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type {\n  ImagePart,\n  MediaInputMetadata,\n  TokenUsage,\n  VideoGenerationOptions,\n  VideoJobResult,\n  VideoPart,\n  VideoStatusResult,\n  VideoUrlResult,\n} from '@tanstack/ai'\nimport type { GrokVideoModel } from '../model-meta'\nimport type {\n  GrokVideoModelDurationByName,\n  GrokVideoModelInputModalitiesByName,\n  GrokVideoModelProviderOptionsByName,\n  GrokVideoModelSizeByName,\n  GrokVideoRuntimeOptions,\n} from '../video/video-provider-options'\nimport type { GrokClientConfig } from '../utils/client'\n\n/**\n * Configuration for Grok video adapter.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GrokVideoConfig extends GrokClientConfig {}\n\n/**\n * xAI bills video generation in \"USD ticks\": 10^10 ticks per US dollar\n * (e.g. one grok-imagine-video-1.5 second costs $0.08 = 800_000_000 ticks).\n */\nconst USD_TICKS_PER_DOLLAR = 10_000_000_000\n\n/** Response of the POST /v1/videos/{generations,edits,extensions} endpoints. */\ninterface GrokVideoCreateResponse {\n  request_id?: string\n}\n\n/** Response of GET /v1/videos/{request_id}. */\ninterface GrokVideoStatusResponse {\n  status?: string\n  progress?: number\n  model?: string\n  video?: {\n    url?: string\n    duration?: number\n  }\n  usage?: {\n    cost_in_usd_ticks?: number\n  }\n  error?: string\n}\n\n/**\n * Convert a TanStack image / video part to the URL string accepted by xAI's\n * Imagine video endpoints: public URLs pass through (fetched by xAI's\n * servers), data sources become base64 data URIs.\n */\nfunction mediaPartToUrl(\n  part: ImagePart<MediaInputMetadata> | VideoPart<MediaInputMetadata>,\n): string {\n  if (part.source.type === 'url') return part.source.value\n  return `data:${part.source.mimeType};base64,${part.source.value}`\n}\n\nfunction buildGrokVideoUsage(\n  response: GrokVideoStatusResponse,\n): TokenUsage | undefined {\n  const seconds = response.video?.duration\n  const ticks = response.usage?.cost_in_usd_ticks\n  if (seconds === undefined && ticks === undefined) return undefined\n  return {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n    ...(seconds !== undefined && {\n      billed: { quantity: seconds, unit: 'seconds' },\n      unitsBilled: seconds,\n    }),\n    ...(ticks !== undefined && { cost: ticks / USD_TICKS_PER_DOLLAR }),\n  }\n}\n\n/**\n * Grok Video Generation Adapter (xAI Imagine API)\n *\n * Tree-shakeable adapter for the grok-imagine video models using the\n * async jobs/polling architecture: create a generation request, poll it,\n * then read the completed video URL.\n *\n * Both models support text-to-video and image-to-video;\n * `grok-imagine-video-1.5` is xAI's documented default and adds native\n * 1080p generation plus reference-to-video inputs. Source-video edit\n * and extend are `grok-imagine-video` only.\n *\n * The Imagine video endpoints are not part of the OpenAI SDK surface (and\n * xAI rejects the SDK's multipart paths), so requests are plain JSON calls\n * issued with the configured `fetch` (or the global one).\n *\n * @experimental Video generation is an experimental feature and may change.\n *\n * Features:\n * - Async job-based video generation (1–15 second clips with audio)\n * - Aspect-ratio sizing via the \"aspectRatio_resolution\" size template\n *   (e.g. '16:9_720p'), consistent with the grok-imagine image models\n * - Image-to-video via an `image` prompt part (starting frame URL or data URI)\n * - Reference-to-video via image prompt parts with\n *   `metadata.role: 'reference'` or `'character'` (→ `reference_images`)\n *   and preset voices via `modelOptions.reference_audios`\n *   (grok-imagine-video-1.5 only)\n * - Video editing / extension on `grok-imagine-video` via a source\n *   `video` prompt part and `modelOptions.mode: 'edit' | 'extend'`\n *   (`/v1/videos/edits` / `/v1/videos/extensions`; in extend mode\n *   `duration` is the added tail)\n * - Usage reporting: billed seconds (`usage.billed`) and exact cost\n */\nexport class GrokVideoAdapter<\n  TModel extends GrokVideoModel,\n> extends BaseVideoAdapter<\n  TModel,\n  GrokVideoModelProviderOptionsByName[TModel],\n  GrokVideoModelProviderOptionsByName,\n  GrokVideoModelSizeByName,\n  GrokVideoModelInputModalitiesByName,\n  GrokVideoModelDurationByName\n> {\n  readonly name = 'grok' as const\n\n  private readonly clientConfig: GrokVideoConfig\n\n  constructor(config: GrokVideoConfig, model: TModel) {\n    super({}, model)\n    this.clientConfig = withGrokDefaults(config)\n  }\n\n  private async request(\n    path: string,\n    init?: Omit<RequestInit, 'headers'>,\n  ): Promise<Response> {\n    // workerd's fetch must be invoked with this === globalThis\n    const fetchFn = this.clientConfig.fetch ?? globalThis.fetch.bind(globalThis)\n    return await fetchFn(`${this.clientConfig.baseURL}${path}`, {\n      ...init,\n      headers: {\n        'Content-Type': 'application/json',\n        Authorization: `Bearer ${this.clientConfig.apiKey}`,\n      },\n    })\n  }\n\n  /**\n   * Reads the error message out of an Imagine API error body\n   * (`{\"code\": \"...\", \"error\": \"...\"}`), falling back to the raw text.\n   */\n  private async errorMessage(response: Response): Promise<string> {\n    const body = await response.text()\n    try {\n      const parsed: unknown = JSON.parse(body)\n      if (\n        typeof parsed === 'object' &&\n        parsed !== null &&\n        'error' in parsed &&\n        typeof parsed.error === 'string'\n      ) {\n        return parsed.error\n      }\n    } catch {\n      // not JSON — fall through to the raw body\n    }\n    return body\n  }\n\n  async createVideoJob(\n    options: VideoGenerationOptions<\n      GrokVideoModelProviderOptionsByName[TModel],\n      GrokVideoModelSizeByName[TModel],\n      GrokVideoModelDurationByName[TModel]\n    >,\n  ): Promise<VideoJobResult> {\n    const { model, size, modelOptions, logger } = options\n\n    // `mode` is a routing hint for this adapter, not an API field — strip it\n    // before the remaining options are spread onto the request body. The\n    // per-model map narrows what callers can pass, but modelOptions often\n    // arrives as deserialized JSON, so the adapter handles the widest option\n    // surface (the 1.5 shape) uniformly and gates by model at runtime.\n    const { mode, ...wireOptions } = (modelOptions ??\n      {}) as GrokVideoRuntimeOptions\n\n    // `mode` is typed 'edit' | 'extend' but reaches us untrusted from JSON\n    // callers. An unrecognised value must not fall through to the\n    // generations endpoint with a source-video body — that would silently\n    // run (and bill) a generation the caller never asked for.\n    if (mode !== undefined && mode !== 'edit' && mode !== 'extend') {\n      throw new Error(\n        `${this.name}: unknown modelOptions.mode '${String(mode)}'. ` +\n          `Expected 'edit' or 'extend'.`,\n      )\n    }\n\n    // The interleaved prompt decomposes into verbatim text plus typed media\n    // buckets. Reference audio is voice-id based (not an audio file), so\n    // audio prompt parts have no request field to land in.\n    const resolved = resolveMediaPrompt(options.prompt)\n    if (resolved.audios.length > 0) {\n      throw new Error(\n        `${this.name}.createVideoJob does not support audio prompt parts (model: ${model}). ` +\n          `To reference a preset voice, pass modelOptions.reference_audios ` +\n          `(e.g. [{ voice_id: 'eve' }]).`,\n      )\n    }\n\n    // A video prompt part is the source clip for edit / extension mode.\n    // Those endpoints are grok-imagine-video only — 1.5 has no video input.\n    if (\n      !isGrokVideoSourceModel(model) &&\n      (mode !== undefined || resolved.videos.length > 0)\n    ) {\n      throw new Error(\n        `${this.name}: ${model} does not support video editing or extension. ` +\n          `Use 'grok-imagine-video' for /v1/videos/edits and /v1/videos/extensions.`,\n      )\n    }\n\n    // The mode must be chosen explicitly because the two endpoints have\n    // different semantics (edit rewrites the clip, extend appends\n    // `duration` seconds).\n    if (resolved.videos.length > 1) {\n      throw new Error(\n        `${this.name}: ${model} accepts at most one source video; received ${resolved.videos.length}.`,\n      )\n    }\n    const [sourceVideo] = resolved.videos\n    if (sourceVideo && mode === undefined) {\n      throw new Error(\n        `${this.name}: a video prompt part needs modelOptions.mode set to ` +\n          `'edit' (rewrite the clip) or 'extend' (append to it).`,\n      )\n    }\n    if (!sourceVideo && mode !== undefined) {\n      throw new Error(\n        `${this.name}: modelOptions.mode '${mode}' requires a video prompt ` +\n          `part carrying the source clip.`,\n      )\n    }\n\n    if (mode !== undefined && sourceVideo) {\n      return await this.createSourceVideoJob({\n        model,\n        mode,\n        sourceVideo,\n        resolved,\n        wireOptions,\n        size,\n        genericDuration: options.duration,\n        logger,\n      })\n    }\n\n    validateVideoSize(model, size)\n\n    // Pull the specially-handled keys out of the wire options: `duration`\n    // is folded into the snapped value below, and the reference fields are\n    // re-added explicitly so a JSON-serialized `null` or empty array reads\n    // as \"unset\" instead of leaking onto the wire.\n    const {\n      duration: rawOptionDuration,\n      reference_images: explicitReferenceImages,\n      reference_audios: referenceAudios,\n      ...generationOptions\n    } = wireOptions\n\n    // Coerce the requested duration into the model's valid range (1–15s,\n    // integer) instead of rejecting it — `snapDuration` clamps and rounds.\n    // modelOptions wins over the generic `duration`, mirroring the size\n    // precedence below.\n    const rawDuration = rawOptionDuration ?? options.duration\n    const duration =\n      rawDuration != null ? this.snapDuration(rawDuration) : undefined\n\n    // Image parts split by role: un-roled / 'start_frame' images become the\n    // starting frame (image-to-video); 'reference' / 'character' images\n    // become reference_images (reference-to-video). The Imagine API has no\n    // mask / control / end-frame inputs. Unknown role strings (possible via\n    // JSON callers) throw rather than silently dropping the part.\n    const startFrames: Array<ImagePart<MediaInputMetadata>> = []\n    const referenceImages: Array<{ url: string }> = []\n    for (const part of resolved.images) {\n      const role = part.metadata?.role\n      switch (role) {\n        case 'mask':\n        case 'control':\n        case 'end_frame':\n          throw new Error(\n            `${this.name}: the Imagine video API has no '${role}' image ` +\n              `input on model ${model}. Use an un-roled / 'start_frame' ` +\n              `image as the starting frame, or 'reference' images.`,\n          )\n        case 'reference':\n        case 'character':\n          referenceImages.push({ url: mediaPartToUrl(part) })\n          break\n        case 'start_frame':\n        case undefined:\n          startFrames.push(part)\n          break\n        default:\n          throw new Error(\n            `${this.name}: unknown image metadata.role '${String(role)}'. ` +\n              `Expected 'start_frame', 'reference', or 'character'.`,\n          )\n      }\n    }\n    if (startFrames.length > 1) {\n      throw new Error(\n        `${this.name}: ${model} accepts at most one starting-frame image; received ${startFrames.length}. ` +\n          `Use metadata.role: 'reference' for reference-to-video inputs.`,\n      )\n    }\n    // Explicit modelOptions.reference_images replaces the part-derived list\n    // (an explicit empty array means \"none\").\n    const finalReferenceImages =\n      explicitReferenceImages ??\n      (referenceImages.length > 0 ? referenceImages : undefined)\n    const referenceImageCount = finalReferenceImages?.length ?? 0\n    const referenceAudioCount = referenceAudios?.length ?? 0\n    const hasReference = referenceImageCount > 0 || referenceAudioCount > 0\n\n    // Reference inputs are a grok-imagine-video-1.5 feature. The per-model\n    // options map already hides the fields from other models at compile\n    // time; this runtime gate covers prompt-part roles and untyped callers.\n    if (!isGrokVideoReferenceModel(model) && hasReference) {\n      throw new Error(\n        `${this.name}: ${model} does not support reference-to-video inputs. ` +\n          `Use 'grok-imagine-video-1.5' for reference_images / reference_audios.`,\n      )\n    }\n    if (referenceAudioCount > GROK_VIDEO_MAX_REFERENCE_AUDIOS) {\n      throw new Error(\n        `${this.name}: ${model} accepts at most ${GROK_VIDEO_MAX_REFERENCE_AUDIOS} reference voices; received ${referenceAudioCount}.`,\n      )\n    }\n    if (referenceImageCount > GROK_VIDEO_MAX_REFERENCE_IMAGES) {\n      throw new Error(\n        `${this.name}: ${model} accepts at most ${GROK_VIDEO_MAX_REFERENCE_IMAGES} reference images; received ${referenceImageCount}.`,\n      )\n    }\n\n    // Image-to-video: the single image prompt part becomes the starting frame\n    // and the prompt text describes the desired motion. URL sources are\n    // fetched by xAI's servers; data sources are sent as base64 data URIs.\n    // On grok-imagine-video-1.5 the starting frame combines with reference\n    // inputs — `image` pins the first frame while reference_images /\n    // reference_audios steer subjects and voices. Classic grok-imagine-video\n    // rejects that mix, but it rejects reference inputs outright, so the\n    // model gate above already covers it.\n    const [startFrame] = startFrames\n\n    // The generic `size` option carries an \"aspectRatio_resolution\" template\n    // (e.g. '16:9_720p') and maps to the Imagine API's `aspect_ratio` /\n    // `resolution` parameters; explicit modelOptions win over the template\n    // (including `reference_images`, which replaces the part-derived list).\n    const parsedSize = size !== undefined ? parseGrokVideoSize(size) : undefined\n    const resolvedResolution =\n      generationOptions.resolution ?? parsedSize?.resolution\n    if (hasReference && resolvedResolution === '1080p') {\n      throw new Error(\n        `${this.name}: reference-to-video is capped at 720p on ${model}.`,\n      )\n    }\n    const request = {\n      model,\n      prompt: resolved.text,\n      ...(startFrame && { image: { url: mediaPartToUrl(startFrame) } }),\n      ...(referenceImageCount > 0 && {\n        reference_images: finalReferenceImages,\n      }),\n      ...(referenceAudioCount > 0 && {\n        reference_audios: referenceAudios,\n      }),\n      ...(parsedSize && {\n        aspect_ratio: parsedSize.aspectRatio,\n        ...(parsedSize.resolution !== undefined && {\n          resolution: parsedSize.resolution,\n        }),\n      }),\n      // The remaining options spread after the size template so explicit\n      // aspect_ratio / resolution win over it; duration and the reference\n      // fields were destructured out above and re-added normalized.\n      ...generationOptions,\n      ...(duration !== undefined && { duration }),\n    }\n\n    return await this.postVideoJob('/videos/generations', request, {\n      model,\n      logger,\n      logLine: `activity=video.create provider=${this.name} model=${model} mode=generate size=${size ?? 'default'} duration=${duration ?? 'default'}`,\n    })\n  }\n\n  /**\n   * Build and post an edit / extension request. Both endpoints take only\n   * `model`, `prompt`, and the source `video` (plus `duration` — the length\n   * of the added tail — for extensions): output geometry is inherited from\n   * the source clip, capped at 720p, and edit outputs also inherit the\n   * source length. Rather than sending fields the API documents as ignored,\n   * the inapplicable options are rejected with actionable errors.\n   */\n  private async createSourceVideoJob(args: {\n    model: string\n    mode: 'edit' | 'extend'\n    sourceVideo: VideoPart<MediaInputMetadata>\n    resolved: ReturnType<typeof resolveMediaPrompt>\n    wireOptions: Omit<GrokVideoRuntimeOptions, 'mode'>\n    size: string | undefined\n    genericDuration: number | undefined\n    logger: VideoGenerationOptions<GrokVideoRuntimeOptions>['logger']\n  }): Promise<VideoJobResult> {\n    const { model, mode, sourceVideo, resolved, wireOptions, logger } = args\n    const endpoint = mode === 'edit' ? '/videos/edits' : '/videos/extensions'\n\n    if (resolved.images.length > 0) {\n      throw new Error(\n        `${this.name}: '${mode}' mode takes only the source video — image ` +\n          `prompt parts are not supported by ${endpoint}.`,\n      )\n    }\n\n    // Pull every generation-only key out of the wire options so nothing can\n    // leak into the edit/extend body via the spread below. JSON-serialized\n    // `null` values (a common \"unset\" encoding) are treated as absent;\n    // actual values are rejected with actionable errors.\n    const {\n      aspect_ratio: aspectRatio,\n      resolution,\n      duration: modeDuration,\n      reference_images: referenceImagesOption,\n      reference_audios: referenceAudiosOption,\n      ...passthrough\n    } = wireOptions\n    if (\n      (referenceImagesOption?.length ?? 0) > 0 ||\n      (referenceAudiosOption?.length ?? 0) > 0\n    ) {\n      throw new Error(\n        `${this.name}: reference inputs are only supported by video ` +\n          `generation, not '${mode}' mode.`,\n      )\n    }\n    if (args.size !== undefined || aspectRatio != null || resolution != null) {\n      throw new Error(\n        `${this.name}: '${mode}' mode does not accept size / aspect_ratio / ` +\n          `resolution — the output inherits the source clip's geometry ` +\n          `(capped at 720p).`,\n      )\n    }\n    const rawDuration = modeDuration ?? args.genericDuration\n    if (mode === 'edit' && rawDuration != null) {\n      throw new Error(\n        `${this.name}: 'edit' mode does not accept a duration — the output ` +\n          `inherits the source clip's length. Use mode 'extend' to append ` +\n          `seconds to the clip.`,\n      )\n    }\n    // Extend: the snapped duration is the added-tail length (1–15s).\n    const duration =\n      rawDuration != null ? this.snapDuration(rawDuration) : undefined\n\n    const request = {\n      model,\n      prompt: resolved.text,\n      video: { url: mediaPartToUrl(sourceVideo) },\n      ...passthrough,\n      ...(duration !== undefined && { duration }),\n    }\n\n    return await this.postVideoJob(endpoint, request, {\n      model,\n      logger,\n      logLine: `activity=video.create provider=${this.name} model=${model} mode=${mode} duration=${duration ?? 'default'}`,\n    })\n  }\n\n  /**\n   * POST a create-job request body to one of the Imagine video endpoints\n   * (`/videos/generations`, `/videos/edits`, `/videos/extensions`) and read\n   * the `request_id` out of the shared response shape.\n   */\n  private async postVideoJob(\n    endpoint: string,\n    request: Record<string, unknown>,\n    context: {\n      model: string\n      logger: VideoGenerationOptions<GrokVideoRuntimeOptions>['logger']\n      logLine: string\n    },\n  ): Promise<VideoJobResult> {\n    const { model, logger, logLine } = context\n    try {\n      logger.request(logLine, { provider: this.name, model })\n\n      const response = await this.request(endpoint, {\n        method: 'POST',\n        body: JSON.stringify(request),\n      })\n      if (!response.ok) {\n        throw new Error(\n          `grok: ${endpoint} request failed (${response.status} ${response.statusText}): ${await this.errorMessage(response)}`,\n        )\n      }\n\n      const result = (await response.json()) as GrokVideoCreateResponse\n      if (!result.request_id) {\n        throw new Error(`grok: ${endpoint} response contained no request_id`)\n      }\n      return { jobId: result.request_id, model }\n    } catch (error: unknown) {\n      logger.errors(`${this.name}.createVideoJob fatal`, {\n        error: toRunErrorPayload(error, `${this.name}.createVideoJob failed`),\n        source: `${this.name}.createVideoJob`,\n      })\n      throw error\n    }\n  }\n\n  private async retrieveJob(jobId: string): Promise<GrokVideoStatusResponse> {\n    const response = await this.request(`/videos/${jobId}`)\n    if (!response.ok) {\n      const error = new Error(\n        `grok: video status request failed (${response.status} ${response.statusText}): ${await this.errorMessage(response)}`,\n      )\n      ;(error as { status?: number }).status = response.status\n      throw error\n    }\n    return (await response.json()) as GrokVideoStatusResponse\n  }\n\n  async getVideoStatus(jobId: string): Promise<VideoStatusResult> {\n    let response: GrokVideoStatusResponse\n    try {\n      response = await this.retrieveJob(jobId)\n    } catch (error) {\n      if ((error as { status?: number }).status === 404) {\n        return { jobId, status: 'failed', error: 'Job not found' }\n      }\n      throw error\n    }\n\n    return {\n      jobId,\n      status: this.mapStatus(response.status),\n      ...(response.progress !== undefined && { progress: response.progress }),\n      ...(response.error !== undefined && { error: response.error }),\n    }\n  }\n\n  async getVideoUrl(jobId: string): Promise<VideoUrlResult> {\n    let response: GrokVideoStatusResponse\n    try {\n      response = await this.retrieveJob(jobId)\n    } catch (error) {\n      if ((error as { status?: number }).status === 404) {\n        throw new Error(`Video job not found: ${jobId}`)\n      }\n      throw error\n    }\n\n    const status = this.mapStatus(response.status)\n    if (status === 'failed') {\n      throw new Error(\n        `Video generation failed${response.error ? `: ${response.error}` : ''}. Job ID: ${jobId}`,\n      )\n    }\n    const url = response.video?.url\n    if (!url) {\n      throw new Error(\n        `Video is not ready for download. Check status first. Job ID: ${jobId}`,\n      )\n    }\n\n    const usage = buildGrokVideoUsage(response)\n    return {\n      jobId,\n      url,\n      ...(usage && { usage }),\n    }\n  }\n\n  /**\n   * Maps Imagine API job statuses onto the generic video status set. The\n   * API reports 'pending' while queued/generating (with a numeric\n   * `progress`), then a terminal 'done' / 'failed' / 'expired'.\n   */\n  protected mapStatus(\n    apiStatus: string | undefined,\n  ): 'pending' | 'processing' | 'completed' | 'failed' {\n    switch (apiStatus) {\n      case 'pending':\n      case 'queued':\n        return 'pending'\n      case 'done':\n      case 'completed':\n      case 'succeeded':\n        return 'completed'\n      case 'failed':\n      case 'expired':\n      case 'error':\n      case 'cancelled':\n        return 'failed'\n      case undefined:\n      default:\n        return 'processing'\n    }\n  }\n\n  /**\n   * Both grok-imagine video models accept a continuous 1–15 integer-second\n   * range. Consumers can use this to render UI without provider knowledge.\n   */\n  override availableDurations(): DurationOptions<\n    GrokVideoModelDurationByName[TModel]\n  > {\n    return getGrokVideoDurationOptions(this.model)\n  }\n\n  /**\n   * Coerce a raw seconds value to the closest valid duration (clamped to\n   * [1, 15] and rounded to whole seconds).\n   */\n  override snapDuration(\n    seconds: number,\n  ): GrokVideoModelDurationByName[TModel] | undefined {\n    return snapToDurationOption(seconds, this.availableDurations())\n  }\n}\n\n/**\n * Creates a Grok video adapter with an explicit API key.\n * Type resolution happens here at the call site.\n *\n * @experimental Video generation is an experimental feature and may change.\n *\n * @param model - The model name (e.g., 'grok-imagine-video-1.5')\n * @param apiKey - Your xAI API key\n * @param config - Optional additional configuration\n * @returns Configured Grok video adapter instance with resolved types\n *\n * @example\n * ```typescript\n * const adapter = createGrokVideo('grok-imagine-video-1.5', 'xai-...');\n *\n * const { jobId } = await generateVideo({\n *   adapter,\n *   prompt: 'A beautiful sunset over the ocean',\n *   size: '16:9_720p',\n *   duration: 5\n * });\n * ```\n */\nexport function createGrokVideo<TModel extends GrokVideoModel>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GrokVideoConfig, 'apiKey'>,\n): GrokVideoAdapter<TModel> {\n  return new GrokVideoAdapter({ apiKey, ...config }, model)\n}\n\n/**\n * Creates a Grok video adapter with automatic API key detection from environment variables.\n * Type resolution happens here at the call site.\n *\n * Looks for `XAI_API_KEY` in:\n * - `process.env` (Node.js)\n * - `window.env` (Browser with injected env)\n *\n * @experimental Video generation is an experimental feature and may change.\n *\n * @param model - The model name (e.g., 'grok-imagine-video-1.5')\n * @param config - Optional configuration (excluding apiKey which is auto-detected)\n * @returns Configured Grok video adapter instance with resolved types\n * @throws Error if XAI_API_KEY is not found in environment\n *\n * @example\n * ```typescript\n * // Automatically uses XAI_API_KEY from environment\n * const adapter = grokVideo('grok-imagine-video-1.5');\n *\n * // Image-to-video: an optional image prompt part is the starting frame.\n * const { jobId } = await generateVideo({\n *   adapter,\n *   prompt: [\n *     { type: 'text', content: 'Make the cat start playing the piano' },\n *     { type: 'image', source: { type: 'url', value: 'https://example.com/cat.png' } },\n *   ],\n * });\n *\n * // Poll for status\n * const status = await getVideoJobStatus({ adapter, jobId });\n * ```\n */\nexport function grokVideo<TModel extends GrokVideoModel>(\n  model: TModel,\n  config?: Omit<GrokVideoConfig, 'apiKey'>,\n): GrokVideoAdapter<TModel> {\n  const apiKey = getGrokApiKeyFromEnv()\n  return createGrokVideo(model, apiKey, config)\n}\n"],"mappings":";;;;;;;;;;AA6CA,IAAM,uBAAuB;;;;;;AA2B7B,SAAS,eACP,MACQ;CACR,IAAI,KAAK,OAAO,SAAS,OAAO,OAAO,KAAK,OAAO;CACnD,OAAO,QAAQ,KAAK,OAAO,SAAS,UAAU,KAAK,OAAO;AAC5D;AAEA,SAAS,oBACP,UACwB;CACxB,MAAM,UAAU,SAAS,OAAO;CAChC,MAAM,QAAQ,SAAS,OAAO;CAC9B,IAAI,YAAY,KAAA,KAAa,UAAU,KAAA,GAAW,OAAO,KAAA;CACzD,OAAO;EACL,cAAc;EACd,kBAAkB;EAClB,aAAa;EACb,GAAI,YAAY,KAAA,KAAa;GAC3B,QAAQ;IAAE,UAAU;IAAS,MAAM;GAAU;GAC7C,aAAa;EACf;EACA,GAAI,UAAU,KAAA,KAAa,EAAE,MAAM,QAAQ,qBAAqB;CAClE;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCA,IAAa,mBAAb,cAEU,iBAOR;CACA,OAAgB;CAEhB;CAEA,YAAY,QAAyB,OAAe;EAClD,MAAM,CAAC,GAAG,KAAK;EACf,KAAK,eAAe,iBAAiB,MAAM;CAC7C;CAEA,MAAc,QACZ,MACA,MACmB;EAGnB,OAAO,OADS,KAAK,aAAa,SAAS,WAAW,MAAM,KAAK,UAAU,EAAA,CACtD,GAAG,KAAK,aAAa,UAAU,QAAQ;GAC1D,GAAG;GACH,SAAS;IACP,gBAAgB;IAChB,eAAe,UAAU,KAAK,aAAa;GAC7C;EACF,CAAC;CACH;;;;;CAMA,MAAc,aAAa,UAAqC;EAC9D,MAAM,OAAO,MAAM,SAAS,KAAK;EACjC,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,IAAI;GACvC,IACE,OAAO,WAAW,YAClB,WAAW,QACX,WAAW,UACX,OAAO,OAAO,UAAU,UAExB,OAAO,OAAO;EAElB,QAAQ,CAER;EACA,OAAO;CACT;CAEA,MAAM,eACJ,SAKyB;EACzB,MAAM,EAAE,OAAO,MAAM,cAAc,WAAW;EAO9C,MAAM,EAAE,MAAM,GAAG,gBAAiB,gBAChC,CAAC;EAMH,IAAI,SAAS,KAAA,KAAa,SAAS,UAAU,SAAS,UACpD,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,+BAA+B,OAAO,IAAI,EAAE,gCAE3D;EAMF,MAAM,WAAW,mBAAmB,QAAQ,MAAM;EAClD,IAAI,SAAS,OAAO,SAAS,GAC3B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,8DAA8D,MAAM,iGAGnF;EAKF,IACE,CAAC,uBAAuB,KAAK,MAC5B,SAAS,KAAA,KAAa,SAAS,OAAO,SAAS,IAEhD,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,uHAEzB;EAMF,IAAI,SAAS,OAAO,SAAS,GAC3B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,8CAA8C,SAAS,OAAO,OAAO,EAC9F;EAEF,MAAM,CAAC,eAAe,SAAS;EAC/B,IAAI,eAAe,SAAS,KAAA,GAC1B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,2GAEf;EAEF,IAAI,CAAC,eAAe,SAAS,KAAA,GAC3B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,uBAAuB,KAAK,yDAE3C;EAGF,IAAI,SAAS,KAAA,KAAa,aACxB,OAAO,MAAM,KAAK,qBAAqB;GACrC;GACA;GACA;GACA;GACA;GACA;GACA,iBAAiB,QAAQ;GACzB;EACF,CAAC;EAGH,kBAAkB,OAAO,IAAI;EAM7B,MAAM,EACJ,UAAU,mBACV,kBAAkB,yBAClB,kBAAkB,iBAClB,GAAG,sBACD;EAMJ,MAAM,cAAc,qBAAqB,QAAQ;EACjD,MAAM,WACJ,eAAe,OAAO,KAAK,aAAa,WAAW,IAAI,KAAA;EAOzD,MAAM,cAAoD,CAAC;EAC3D,MAAM,kBAA0C,CAAC;EACjD,KAAK,MAAM,QAAQ,SAAS,QAAQ;GAClC,MAAM,OAAO,KAAK,UAAU;GAC5B,QAAQ,MAAR;IACE,KAAK;IACL,KAAK;IACL,KAAK,aACH,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,kCAAkC,KAAK,yBAChC,MAAM,sFAE5B;IACF,KAAK;IACL,KAAK;KACH,gBAAgB,KAAK,EAAE,KAAK,eAAe,IAAI,EAAE,CAAC;KAClD;IACF,KAAK;IACL,KAAK,KAAA;KACH,YAAY,KAAK,IAAI;KACrB;IACF,SACE,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,iCAAiC,OAAO,IAAI,EAAE,wDAE7D;GACJ;EACF;EACA,IAAI,YAAY,SAAS,GACvB,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,sDAAsD,YAAY,OAAO,gEAElG;EAIF,MAAM,uBACJ,4BACC,gBAAgB,SAAS,IAAI,kBAAkB,KAAA;EAClD,MAAM,sBAAsB,sBAAsB,UAAU;EAC5D,MAAM,sBAAsB,iBAAiB,UAAU;EACvD,MAAM,eAAe,sBAAsB,KAAK,sBAAsB;EAKtE,IAAI,CAAC,0BAA0B,KAAK,KAAK,cACvC,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,mHAEzB;EAEF,IAAI,sBAAA,GACF,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,gDAAiF,oBAAoB,EAC9H;EAEF,IAAI,sBAAA,GACF,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,IAAI,MAAM,gDAAiF,oBAAoB,EAC9H;EAWF,MAAM,CAAC,cAAc;EAMrB,MAAM,aAAa,SAAS,KAAA,IAAY,mBAAmB,IAAI,IAAI,KAAA;EACnE,MAAM,qBACJ,kBAAkB,cAAc,YAAY;EAC9C,IAAI,gBAAgB,uBAAuB,SACzC,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,4CAA4C,MAAM,EACjE;EAEF,MAAM,UAAU;GACd;GACA,QAAQ,SAAS;GACjB,GAAI,cAAc,EAAE,OAAO,EAAE,KAAK,eAAe,UAAU,EAAE,EAAE;GAC/D,GAAI,sBAAsB,KAAK,EAC7B,kBAAkB,qBACpB;GACA,GAAI,sBAAsB,KAAK,EAC7B,kBAAkB,gBACpB;GACA,GAAI,cAAc;IAChB,cAAc,WAAW;IACzB,GAAI,WAAW,eAAe,KAAA,KAAa,EACzC,YAAY,WAAW,WACzB;GACF;GAIA,GAAG;GACH,GAAI,aAAa,KAAA,KAAa,EAAE,SAAS;EAC3C;EAEA,OAAO,MAAM,KAAK,aAAa,uBAAuB,SAAS;GAC7D;GACA;GACA,SAAS,kCAAkC,KAAK,KAAK,SAAS,MAAM,sBAAsB,QAAQ,UAAU,YAAY,YAAY;EACtI,CAAC;CACH;;;;;;;;;CAUA,MAAc,qBAAqB,MASP;EAC1B,MAAM,EAAE,OAAO,MAAM,aAAa,UAAU,aAAa,WAAW;EACpE,MAAM,WAAW,SAAS,SAAS,kBAAkB;EAErD,IAAI,SAAS,OAAO,SAAS,GAC3B,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,KAAK,KAAK,+EACgB,SAAS,EAClD;EAOF,MAAM,EACJ,cAAc,aACd,YACA,UAAU,cACV,kBAAkB,uBAClB,kBAAkB,uBAClB,GAAG,gBACD;EACJ,KACG,uBAAuB,UAAU,KAAK,MACtC,uBAAuB,UAAU,KAAK,GAEvC,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,kEACS,KAAK,QAC7B;EAEF,IAAI,KAAK,SAAS,KAAA,KAAa,eAAe,QAAQ,cAAc,MAClE,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,KAAK,KAAK,2HAGzB;EAEF,MAAM,cAAc,gBAAgB,KAAK;EACzC,IAAI,SAAS,UAAU,eAAe,MACpC,MAAM,IAAI,MACR,GAAG,KAAK,KAAK,0IAGf;EAGF,MAAM,WACJ,eAAe,OAAO,KAAK,aAAa,WAAW,IAAI,KAAA;EAEzD,MAAM,UAAU;GACd;GACA,QAAQ,SAAS;GACjB,OAAO,EAAE,KAAK,eAAe,WAAW,EAAE;GAC1C,GAAG;GACH,GAAI,aAAa,KAAA,KAAa,EAAE,SAAS;EAC3C;EAEA,OAAO,MAAM,KAAK,aAAa,UAAU,SAAS;GAChD;GACA;GACA,SAAS,kCAAkC,KAAK,KAAK,SAAS,MAAM,QAAQ,KAAK,YAAY,YAAY;EAC3G,CAAC;CACH;;;;;;CAOA,MAAc,aACZ,UACA,SACA,SAKyB;EACzB,MAAM,EAAE,OAAO,QAAQ,YAAY;EACnC,IAAI;GACF,OAAO,QAAQ,SAAS;IAAE,UAAU,KAAK;IAAM;GAAM,CAAC;GAEtD,MAAM,WAAW,MAAM,KAAK,QAAQ,UAAU;IAC5C,QAAQ;IACR,MAAM,KAAK,UAAU,OAAO;GAC9B,CAAC;GACD,IAAI,CAAC,SAAS,IACZ,MAAM,IAAI,MACR,SAAS,SAAS,mBAAmB,SAAS,OAAO,GAAG,SAAS,WAAW,KAAK,MAAM,KAAK,aAAa,QAAQ,GACnH;GAGF,MAAM,SAAU,MAAM,SAAS,KAAK;GACpC,IAAI,CAAC,OAAO,YACV,MAAM,IAAI,MAAM,SAAS,SAAS,kCAAkC;GAEtE,OAAO;IAAE,OAAO,OAAO;IAAY;GAAM;EAC3C,SAAS,OAAgB;GACvB,OAAO,OAAO,GAAG,KAAK,KAAK,wBAAwB;IACjD,OAAO,kBAAkB,OAAO,GAAG,KAAK,KAAK,uBAAuB;IACpE,QAAQ,GAAG,KAAK,KAAK;GACvB,CAAC;GACD,MAAM;EACR;CACF;CAEA,MAAc,YAAY,OAAiD;EACzE,MAAM,WAAW,MAAM,KAAK,QAAQ,WAAW,OAAO;EACtD,IAAI,CAAC,SAAS,IAAI;GAChB,MAAM,wBAAQ,IAAI,MAChB,sCAAsC,SAAS,OAAO,GAAG,SAAS,WAAW,KAAK,MAAM,KAAK,aAAa,QAAQ,GACpH;GACC,MAA+B,SAAS,SAAS;GAClD,MAAM;EACR;EACA,OAAQ,MAAM,SAAS,KAAK;CAC9B;CAEA,MAAM,eAAe,OAA2C;EAC9D,IAAI;EACJ,IAAI;GACF,WAAW,MAAM,KAAK,YAAY,KAAK;EACzC,SAAS,OAAO;GACd,IAAK,MAA8B,WAAW,KAC5C,OAAO;IAAE;IAAO,QAAQ;IAAU,OAAO;GAAgB;GAE3D,MAAM;EACR;EAEA,OAAO;GACL;GACA,QAAQ,KAAK,UAAU,SAAS,MAAM;GACtC,GAAI,SAAS,aAAa,KAAA,KAAa,EAAE,UAAU,SAAS,SAAS;GACrE,GAAI,SAAS,UAAU,KAAA,KAAa,EAAE,OAAO,SAAS,MAAM;EAC9D;CACF;CAEA,MAAM,YAAY,OAAwC;EACxD,IAAI;EACJ,IAAI;GACF,WAAW,MAAM,KAAK,YAAY,KAAK;EACzC,SAAS,OAAO;GACd,IAAK,MAA8B,WAAW,KAC5C,MAAM,IAAI,MAAM,wBAAwB,OAAO;GAEjD,MAAM;EACR;EAGA,IADe,KAAK,UAAU,SAAS,MACnC,MAAW,UACb,MAAM,IAAI,MACR,0BAA0B,SAAS,QAAQ,KAAK,SAAS,UAAU,GAAG,YAAY,OACpF;EAEF,MAAM,MAAM,SAAS,OAAO;EAC5B,IAAI,CAAC,KACH,MAAM,IAAI,MACR,gEAAgE,OAClE;EAGF,MAAM,QAAQ,oBAAoB,QAAQ;EAC1C,OAAO;GACL;GACA;GACA,GAAI,SAAS,EAAE,MAAM;EACvB;CACF;;;;;;CAOA,UACE,WACmD;EACnD,QAAQ,WAAR;GACE,KAAK;GACL,KAAK,UACH,OAAO;GACT,KAAK;GACL,KAAK;GACL,KAAK,aACH,OAAO;GACT,KAAK;GACL,KAAK;GACL,KAAK;GACL,KAAK,aACH,OAAO;GACT,KAAK,KAAA;GACL,SACE,OAAO;EACX;CACF;;;;;CAMA,qBAEE;EACA,OAAO,4BAA4B,KAAK,KAAK;CAC/C;;;;;CAMA,aACE,SACkD;EAClD,OAAO,qBAAqB,SAAS,KAAK,mBAAmB,CAAC;CAChE;AACF;;;;;;;;;;;;;;;;;;;;;;;;AAyBA,SAAgB,gBACd,OACA,QACA,QAC0B;CAC1B,OAAO,IAAI,iBAAiB;EAAE;EAAQ,GAAG;CAAO,GAAG,KAAK;AAC1D;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCA,SAAgB,UACd,OACA,QAC0B;CAE1B,OAAO,gBAAgB,OADR,qBACe,GAAQ,MAAM;AAC9C"}

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

@@ -166,7 +166,8 @@ export interface GrokVideoProviderOptions extends GrokVideoBaseProviderOptions {
      * `metadata.role: 'reference'` (or `'character'`); set explicitly to
      * replace the part-derived list. Reference images are addressed from the
      * prompt text as `<IMAGE_0>`, `<IMAGE_1>`, … in request order, and do not
-     * lock the first frame.
+     * lock the first frame — pair them with a starting-frame image prompt part
+     * when the opening frame has to be pinned.
      */
     reference_images?: Array<{
         url: string;
@@ -234,6 +235,8 @@ export type GrokVideoModelSizeByName = {
  * Both models support text-to-video and accept an optional `image` prompt
  * part as the starting frame; image parts with `metadata.role: 'reference'`
  * or `'character'` become `reference_images` (grok-imagine-video-1.5 only).
+ * On grok-imagine-video-1.5 the two combine: the starting frame pins the
+ * first frame while the reference images steer subjects and style.
  * A `video` prompt part carries the source clip for edit / extension mode
  * on grok-imagine-video only (`modelOptions.mode: 'edit' | 'extend'`).
  *

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

@@ -1 +1 @@
-{"version":3,"file":"video-provider-options.js","names":[],"sources":["../../../src/video/video-provider-options.ts"],"sourcesContent":["/**\n * Grok Video Generation Provider Options (xAI Imagine API)\n *\n * Based on https://docs.x.ai/developers/model-capabilities/video/generation\n * (plus the image-to-video, reference-to-video, editing, and extension pages\n * under the same section).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\n\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type { GrokVideoModel } from '../model-meta'\n\n/**\n * Aspect ratios accepted by the grok-imagine video models.\n *\n * Note: this is a narrower set than the grok-imagine image models — the\n * video endpoint rejects the phone-screen ratios ('9:19.5', '9:20', …) and\n * 'auto'.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoAspectRatio =\n  | '1:1'\n  | '16:9'\n  | '9:16'\n  | '4:3'\n  | '3:4'\n  | '3:2'\n  | '2:3'\n\n/**\n * Resolution tiers for the grok-imagine video models.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoResolution = '480p' | '720p' | '1080p'\n\n/**\n * Resolutions accepted by grok-imagine-video (v1.0). Native 1080p is a\n * grok-imagine-video-1.5 generation feature.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoResolutionV1 = '480p' | '720p'\n\n/**\n * Size strings for grok-imagine video models. The Imagine API is\n * aspect-ratio based rather than pixel-size based; like the grok-imagine\n * image models, the generic `size` option uses an\n * `aspectRatio_resolution` template (\"16:9_720p\") — the resolution suffix\n * is optional (\"16:9\" uses the API default).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoSize =\n  | GrokVideoAspectRatio\n  | `${GrokVideoAspectRatio}_${GrokVideoResolution}`\n\n/**\n * Size strings for grok-imagine-video (v1.0) — 1080p is not in the type.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoSizeV1 =\n  | GrokVideoAspectRatio\n  | `${GrokVideoAspectRatio}_${GrokVideoResolutionV1}`\n\nconst GROK_VIDEO_ASPECT_RATIOS: ReadonlyArray<string> = [\n  '1:1',\n  '16:9',\n  '9:16',\n  '4:3',\n  '3:4',\n  '3:2',\n  '2:3',\n]\n\nconst GROK_VIDEO_RESOLUTIONS: ReadonlyArray<string> = ['480p', '720p', '1080p']\n\n/**\n * Video duration limits enforced by the Imagine API (seconds).\n */\nexport const GROK_VIDEO_MIN_DURATION = 1\nexport const GROK_VIDEO_MAX_DURATION = 15\n\n/**\n * Parses a grok video size string into its components.\n * Format: \"aspectRatio\" or \"aspectRatio_resolution\",\n * e.g. \"16:9_720p\" → { aspectRatio: \"16:9\", resolution: \"720p\" }.\n * Returns undefined when the string doesn't match the template.\n */\nexport function parseGrokVideoSize(\n  size: string,\n): { aspectRatio: string; resolution?: string } | undefined {\n  const match = size.match(/^([\\d.]+:[\\d.]+)(?:_(.+))?$/)\n  const [, aspectRatio, resolution] = match ?? []\n  if (aspectRatio === undefined) return undefined\n  return { aspectRatio, ...(resolution !== undefined && { resolution }) }\n}\n\n/**\n * Models that accept native 1080p on text-to-video and image-to-video.\n * Reference-to-video stays capped at 720p even on these models.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function isGrokVideoNative1080pModel(model: string): boolean {\n  return model === 'grok-imagine-video-1.5'\n}\n\n/**\n * Validate the `size` template for a given grok video model.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function validateVideoSize(\n  model: string,\n  size?: string,\n): asserts size is GrokVideoSize | undefined {\n  if (size === undefined) return\n  const parsed = parseGrokVideoSize(size)\n  if (!parsed || !GROK_VIDEO_ASPECT_RATIOS.includes(parsed.aspectRatio)) {\n    throw new Error(\n      `Size \"${size}\" is not supported by model \"${model}\". Expected ` +\n        `\"aspectRatio\" or \"aspectRatio_resolution\" (e.g. \"16:9_720p\") with ` +\n        `aspect ratio one of: ${GROK_VIDEO_ASPECT_RATIOS.join(', ')}`,\n    )\n  }\n  if (\n    parsed.resolution !== undefined &&\n    !GROK_VIDEO_RESOLUTIONS.includes(parsed.resolution)\n  ) {\n    throw new Error(\n      `Resolution \"${parsed.resolution}\" is not supported by model \"${model}\". ` +\n        `Supported resolutions: ${GROK_VIDEO_RESOLUTIONS.join(', ')}`,\n    )\n  }\n  if (parsed.resolution === '1080p' && !isGrokVideoNative1080pModel(model)) {\n    throw new Error(\n      `Resolution \"1080p\" is not supported by model \"${model}\". ` +\n        `Use 'grok-imagine-video-1.5' for native 1080p text-to-video / image-to-video.`,\n    )\n  }\n}\n\n/**\n * Per-model duration type. The Imagine API accepts any integer second in the\n * 1–15 range, so this is a continuous range expressed as `number` (a literal\n * union can't represent it). `snapDuration()` coerces a raw seconds value into\n * the valid range at runtime.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoModelDurationByName = {\n  'grok-imagine-video': number\n  'grok-imagine-video-1.5': number\n}\n\n/**\n * Runtime duration table backing `availableDurations()` / `snapDuration()`.\n * Both grok-imagine video models accept the same continuous 1–15 integer-second\n * range.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport const GROK_VIDEO_DURATIONS: {\n  readonly [TModel in GrokVideoModel]: DurationOptions<\n    GrokVideoModelDurationByName[TModel]\n  >\n} = {\n  'grok-imagine-video': {\n    kind: 'range',\n    min: GROK_VIDEO_MIN_DURATION,\n    max: GROK_VIDEO_MAX_DURATION,\n    step: 1,\n    unit: 'seconds',\n  },\n  'grok-imagine-video-1.5': {\n    kind: 'range',\n    min: GROK_VIDEO_MIN_DURATION,\n    max: GROK_VIDEO_MAX_DURATION,\n    step: 1,\n    unit: 'seconds',\n  },\n}\n\n/**\n * Look up the duration options for a grok video model.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function getGrokVideoDurationOptions<TModel extends GrokVideoModel>(\n  model: TModel,\n): DurationOptions<GrokVideoModelDurationByName[TModel]> {\n  return GROK_VIDEO_DURATIONS[model]\n}\n\n/**\n * Request mode for a source-video job. `'edit'` posts to `/v1/videos/edits`\n * (modify the source clip in place); `'extend'` posts to\n * `/v1/videos/extensions` (continue the source clip — `duration` is the\n * length of the **added tail**, not the total). Both require exactly one\n * video prompt part carrying the source clip, and both are\n * `grok-imagine-video` (v1.0) only.\n *\n * Output geometry (aspect ratio / resolution) is inherited from the source\n * clip in both modes, capped at 720p, and edit outputs also inherit the\n * source length — the adapter rejects `size`, `aspect_ratio`, `resolution`,\n * and (in edit mode) `duration` rather than sending fields the API ignores.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoMode = 'edit' | 'extend'\n\n/**\n * Provider options shared by both grok-imagine video models. These map\n * directly onto the Imagine API request body and take precedence over the\n * generic `size` / `duration` options when both are provided.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GrokVideoBaseProviderOptions {\n  /**\n   * Output aspect ratio. Generation only — edit / extend outputs inherit\n   * the source clip's geometry and the adapter rejects this in those modes.\n   */\n  aspect_ratio?: GrokVideoAspectRatio\n\n  /**\n   * Output resolution tier. Generation only — edit / extend outputs inherit\n   * the source clip's geometry and the adapter rejects this in those modes.\n   * `1080p` is grok-imagine-video-1.5 generation only; reference-to-video\n   * is capped at 720p.\n   */\n  resolution?: GrokVideoResolution\n\n  /**\n   * Video duration in integer seconds (1–15). In `'extend'` mode this is\n   * the length of the added tail only, not the total output length. Not\n   * valid in `'edit'` mode (the output inherits the source clip's length).\n   */\n  duration?: number\n}\n\n/**\n * Provider options for grok-imagine-video (v1.0), which is the only model\n * that accepts a source-video edit / extend job.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GrokVideoSourceProviderOptions extends GrokVideoBaseProviderOptions {\n  /**\n   * Selects the request mode for a source-video prompt part: `'edit'`\n   * (`/v1/videos/edits`) or `'extend'` (`/v1/videos/extensions`). Required\n   * when the prompt carries a video part; not valid without one. Omit for\n   * plain generation (`/v1/videos/generations`). grok-imagine-video only.\n   */\n  mode?: GrokVideoMode\n}\n\n/**\n * Provider options for grok-imagine-video-1.5, which adds the\n * reference-to-video inputs on top of the shared options.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GrokVideoProviderOptions extends GrokVideoBaseProviderOptions {\n  /**\n   * Reference images for reference-to-video generation (output capped at\n   * 720p). Usually populated from image prompt parts with\n   * `metadata.role: 'reference'` (or `'character'`); set explicitly to\n   * replace the part-derived list. Reference images are addressed from the\n   * prompt text as `<IMAGE_0>`, `<IMAGE_1>`, … in request order, and do not\n   * lock the first frame.\n   */\n  reference_images?: Array<{ url: string }>\n\n  /**\n   * Preset TTS voices to reference for generated speech (max 3). Voice ids\n   * come from the xAI TTS voice roster (e.g. 'eve', 'rex') or a custom\n   * voice id, and are addressed from the prompt text as `<AUDIO_0>`,\n   * `<AUDIO_1>`, `<AUDIO_2>`.\n   */\n  reference_audios?: Array<{ voice_id: string }>\n}\n\n/**\n * Widest option surface. Used when `modelOptions` arrives as deserialized\n * JSON and the adapter must validate fields the per-model map already\n * hides at compile time.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoRuntimeOptions = GrokVideoSourceProviderOptions &\n  GrokVideoProviderOptions\n\n/**\n * Maximum reference voices accepted by the Imagine video endpoint.\n */\nexport const GROK_VIDEO_MAX_REFERENCE_AUDIOS = 3\n\n/**\n * Maximum reference images accepted by the Imagine video endpoint.\n */\nexport const GROK_VIDEO_MAX_REFERENCE_IMAGES = 7\n\n/**\n * Model names whose per-model options declare the reference fields. Keeps\n * the runtime set below provably in sync with\n * {@link GrokVideoModelProviderOptionsByName} — a typo or a new\n * reference-capable model missing from the set is a compile error.\n */\ntype GrokVideoReferenceModel = {\n  [TModel in GrokVideoModel]: 'reference_images' extends keyof GrokVideoModelProviderOptionsByName[TModel]\n    ? TModel\n    : never\n}[GrokVideoModel]\n\n/**\n * Models that support reference-to-video inputs (`reference_images` /\n * `reference_audios`). The per-model provider-options map hides the fields\n * from other models at compile time; this backs the runtime gate for\n * untyped callers (e.g. deserialized JSON) so they get a clear error\n * instead of a raw API 400.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nconst GROK_VIDEO_REFERENCE_MODELS: ReadonlySet<string> =\n  new Set<GrokVideoReferenceModel>(['grok-imagine-video-1.5'])\n\n/**\n * True when the model accepts reference-to-video inputs.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function isGrokVideoReferenceModel(model: string): boolean {\n  return GROK_VIDEO_REFERENCE_MODELS.has(model)\n}\n\n/**\n * Model names whose per-model options declare `mode`. Same\n * provably-in-sync construction as {@link GrokVideoReferenceModel}.\n */\ntype GrokVideoSourceModel = {\n  [TModel in GrokVideoModel]: 'mode' extends keyof GrokVideoModelProviderOptionsByName[TModel]\n    ? TModel\n    : never\n}[GrokVideoModel]\n\n/**\n * Models that accept a source-video prompt part for `/v1/videos/edits`\n * and `/v1/videos/extensions`. xAI lists video input only on\n * grok-imagine-video (v1.0).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nconst GROK_VIDEO_SOURCE_MODELS: ReadonlySet<string> =\n  new Set<GrokVideoSourceModel>(['grok-imagine-video'])\n\n/**\n * True when the model accepts edit / extend source-video jobs.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function isGrokVideoSourceModel(model: string): boolean {\n  return GROK_VIDEO_SOURCE_MODELS.has(model)\n}\n\n/**\n * Type-only map from model name to its specific provider options. Only\n * grok-imagine-video-1.5 exposes the reference-to-video fields. Only\n * grok-imagine-video (v1.0) exposes `mode` for edit / extend.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoModelProviderOptionsByName = {\n  'grok-imagine-video': GrokVideoSourceProviderOptions\n  'grok-imagine-video-1.5': GrokVideoProviderOptions\n}\n\n/**\n * Type-only map from model name to its supported `size` strings.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoModelSizeByName = {\n  'grok-imagine-video': GrokVideoSizeV1\n  'grok-imagine-video-1.5': GrokVideoSize\n}\n\n/**\n * Type-only map from model name to the non-text prompt modalities it accepts.\n * Both models support text-to-video and accept an optional `image` prompt\n * part as the starting frame; image parts with `metadata.role: 'reference'`\n * or `'character'` become `reference_images` (grok-imagine-video-1.5 only).\n * A `video` prompt part carries the source clip for edit / extension mode\n * on grok-imagine-video only (`modelOptions.mode: 'edit' | 'extend'`).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoModelInputModalitiesByName = {\n  'grok-imagine-video': readonly ['image', 'video']\n  'grok-imagine-video-1.5': readonly ['image']\n}\n"],"mappings":";AAoEA,IAAM,2BAAkD;CACtD;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,IAAM,yBAAgD;CAAC;CAAQ;CAAQ;AAAO;;;;;;;AAc9E,SAAgB,mBACd,MAC0D;CAE1D,MAAM,GAAG,aAAa,cADR,KAAK,MAAM,6BACW,KAAS,CAAC;CAC9C,IAAI,gBAAgB,KAAA,GAAW,OAAO,KAAA;CACtC,OAAO;EAAE;EAAa,GAAI,eAAe,KAAA,KAAa,EAAE,WAAW;CAAG;AACxE;;;;;;;AAQA,SAAgB,4BAA4B,OAAwB;CAClE,OAAO,UAAU;AACnB;;;;;;AAOA,SAAgB,kBACd,OACA,MAC2C;CAC3C,IAAI,SAAS,KAAA,GAAW;CACxB,MAAM,SAAS,mBAAmB,IAAI;CACtC,IAAI,CAAC,UAAU,CAAC,yBAAyB,SAAS,OAAO,WAAW,GAClE,MAAM,IAAI,MACR,SAAS,KAAK,+BAA+B,MAAM,qGAEzB,yBAAyB,KAAK,IAAI,GAC9D;CAEF,IACE,OAAO,eAAe,KAAA,KACtB,CAAC,uBAAuB,SAAS,OAAO,UAAU,GAElD,MAAM,IAAI,MACR,eAAe,OAAO,WAAW,+BAA+B,MAAM,4BAC1C,uBAAuB,KAAK,IAAI,GAC9D;CAEF,IAAI,OAAO,eAAe,WAAW,CAAC,4BAA4B,KAAK,GACrE,MAAM,IAAI,MACR,iDAAiD,MAAM,iFAEzD;AAEJ;;;;;;;;AAsBA,IAAa,uBAIT;CACF,sBAAsB;EACpB,MAAM;EACN,KAAA;EACA,KAAA;EACA,MAAM;EACN,MAAM;CACR;CACA,0BAA0B;EACxB,MAAM;EACN,KAAA;EACA,KAAA;EACA,MAAM;EACN,MAAM;CACR;AACF;;;;;;AAOA,SAAgB,4BACd,OACuD;CACvD,OAAO,qBAAqB;AAC9B;;;;;;;;;;AAoIA,IAAM,8CACJ,IAAI,IAA6B,CAAC,wBAAwB,CAAC;;;;;;AAO7D,SAAgB,0BAA0B,OAAwB;CAChE,OAAO,4BAA4B,IAAI,KAAK;AAC9C;;;;;;;;AAmBA,IAAM,2CACJ,IAAI,IAA0B,CAAC,oBAAoB,CAAC;;;;;;AAOtD,SAAgB,uBAAuB,OAAwB;CAC7D,OAAO,yBAAyB,IAAI,KAAK;AAC3C"}
+{"version":3,"file":"video-provider-options.js","names":[],"sources":["../../../src/video/video-provider-options.ts"],"sourcesContent":["/**\n * Grok Video Generation Provider Options (xAI Imagine API)\n *\n * Based on https://docs.x.ai/developers/model-capabilities/video/generation\n * (plus the image-to-video, reference-to-video, editing, and extension pages\n * under the same section).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\n\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type { GrokVideoModel } from '../model-meta'\n\n/**\n * Aspect ratios accepted by the grok-imagine video models.\n *\n * Note: this is a narrower set than the grok-imagine image models — the\n * video endpoint rejects the phone-screen ratios ('9:19.5', '9:20', …) and\n * 'auto'.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoAspectRatio =\n  | '1:1'\n  | '16:9'\n  | '9:16'\n  | '4:3'\n  | '3:4'\n  | '3:2'\n  | '2:3'\n\n/**\n * Resolution tiers for the grok-imagine video models.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoResolution = '480p' | '720p' | '1080p'\n\n/**\n * Resolutions accepted by grok-imagine-video (v1.0). Native 1080p is a\n * grok-imagine-video-1.5 generation feature.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoResolutionV1 = '480p' | '720p'\n\n/**\n * Size strings for grok-imagine video models. The Imagine API is\n * aspect-ratio based rather than pixel-size based; like the grok-imagine\n * image models, the generic `size` option uses an\n * `aspectRatio_resolution` template (\"16:9_720p\") — the resolution suffix\n * is optional (\"16:9\" uses the API default).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoSize =\n  | GrokVideoAspectRatio\n  | `${GrokVideoAspectRatio}_${GrokVideoResolution}`\n\n/**\n * Size strings for grok-imagine-video (v1.0) — 1080p is not in the type.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoSizeV1 =\n  | GrokVideoAspectRatio\n  | `${GrokVideoAspectRatio}_${GrokVideoResolutionV1}`\n\nconst GROK_VIDEO_ASPECT_RATIOS: ReadonlyArray<string> = [\n  '1:1',\n  '16:9',\n  '9:16',\n  '4:3',\n  '3:4',\n  '3:2',\n  '2:3',\n]\n\nconst GROK_VIDEO_RESOLUTIONS: ReadonlyArray<string> = ['480p', '720p', '1080p']\n\n/**\n * Video duration limits enforced by the Imagine API (seconds).\n */\nexport const GROK_VIDEO_MIN_DURATION = 1\nexport const GROK_VIDEO_MAX_DURATION = 15\n\n/**\n * Parses a grok video size string into its components.\n * Format: \"aspectRatio\" or \"aspectRatio_resolution\",\n * e.g. \"16:9_720p\" → { aspectRatio: \"16:9\", resolution: \"720p\" }.\n * Returns undefined when the string doesn't match the template.\n */\nexport function parseGrokVideoSize(\n  size: string,\n): { aspectRatio: string; resolution?: string } | undefined {\n  const match = size.match(/^([\\d.]+:[\\d.]+)(?:_(.+))?$/)\n  const [, aspectRatio, resolution] = match ?? []\n  if (aspectRatio === undefined) return undefined\n  return { aspectRatio, ...(resolution !== undefined && { resolution }) }\n}\n\n/**\n * Models that accept native 1080p on text-to-video and image-to-video.\n * Reference-to-video stays capped at 720p even on these models.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function isGrokVideoNative1080pModel(model: string): boolean {\n  return model === 'grok-imagine-video-1.5'\n}\n\n/**\n * Validate the `size` template for a given grok video model.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function validateVideoSize(\n  model: string,\n  size?: string,\n): asserts size is GrokVideoSize | undefined {\n  if (size === undefined) return\n  const parsed = parseGrokVideoSize(size)\n  if (!parsed || !GROK_VIDEO_ASPECT_RATIOS.includes(parsed.aspectRatio)) {\n    throw new Error(\n      `Size \"${size}\" is not supported by model \"${model}\". Expected ` +\n        `\"aspectRatio\" or \"aspectRatio_resolution\" (e.g. \"16:9_720p\") with ` +\n        `aspect ratio one of: ${GROK_VIDEO_ASPECT_RATIOS.join(', ')}`,\n    )\n  }\n  if (\n    parsed.resolution !== undefined &&\n    !GROK_VIDEO_RESOLUTIONS.includes(parsed.resolution)\n  ) {\n    throw new Error(\n      `Resolution \"${parsed.resolution}\" is not supported by model \"${model}\". ` +\n        `Supported resolutions: ${GROK_VIDEO_RESOLUTIONS.join(', ')}`,\n    )\n  }\n  if (parsed.resolution === '1080p' && !isGrokVideoNative1080pModel(model)) {\n    throw new Error(\n      `Resolution \"1080p\" is not supported by model \"${model}\". ` +\n        `Use 'grok-imagine-video-1.5' for native 1080p text-to-video / image-to-video.`,\n    )\n  }\n}\n\n/**\n * Per-model duration type. The Imagine API accepts any integer second in the\n * 1–15 range, so this is a continuous range expressed as `number` (a literal\n * union can't represent it). `snapDuration()` coerces a raw seconds value into\n * the valid range at runtime.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoModelDurationByName = {\n  'grok-imagine-video': number\n  'grok-imagine-video-1.5': number\n}\n\n/**\n * Runtime duration table backing `availableDurations()` / `snapDuration()`.\n * Both grok-imagine video models accept the same continuous 1–15 integer-second\n * range.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport const GROK_VIDEO_DURATIONS: {\n  readonly [TModel in GrokVideoModel]: DurationOptions<\n    GrokVideoModelDurationByName[TModel]\n  >\n} = {\n  'grok-imagine-video': {\n    kind: 'range',\n    min: GROK_VIDEO_MIN_DURATION,\n    max: GROK_VIDEO_MAX_DURATION,\n    step: 1,\n    unit: 'seconds',\n  },\n  'grok-imagine-video-1.5': {\n    kind: 'range',\n    min: GROK_VIDEO_MIN_DURATION,\n    max: GROK_VIDEO_MAX_DURATION,\n    step: 1,\n    unit: 'seconds',\n  },\n}\n\n/**\n * Look up the duration options for a grok video model.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function getGrokVideoDurationOptions<TModel extends GrokVideoModel>(\n  model: TModel,\n): DurationOptions<GrokVideoModelDurationByName[TModel]> {\n  return GROK_VIDEO_DURATIONS[model]\n}\n\n/**\n * Request mode for a source-video job. `'edit'` posts to `/v1/videos/edits`\n * (modify the source clip in place); `'extend'` posts to\n * `/v1/videos/extensions` (continue the source clip — `duration` is the\n * length of the **added tail**, not the total). Both require exactly one\n * video prompt part carrying the source clip, and both are\n * `grok-imagine-video` (v1.0) only.\n *\n * Output geometry (aspect ratio / resolution) is inherited from the source\n * clip in both modes, capped at 720p, and edit outputs also inherit the\n * source length — the adapter rejects `size`, `aspect_ratio`, `resolution`,\n * and (in edit mode) `duration` rather than sending fields the API ignores.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoMode = 'edit' | 'extend'\n\n/**\n * Provider options shared by both grok-imagine video models. These map\n * directly onto the Imagine API request body and take precedence over the\n * generic `size` / `duration` options when both are provided.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GrokVideoBaseProviderOptions {\n  /**\n   * Output aspect ratio. Generation only — edit / extend outputs inherit\n   * the source clip's geometry and the adapter rejects this in those modes.\n   */\n  aspect_ratio?: GrokVideoAspectRatio\n\n  /**\n   * Output resolution tier. Generation only — edit / extend outputs inherit\n   * the source clip's geometry and the adapter rejects this in those modes.\n   * `1080p` is grok-imagine-video-1.5 generation only; reference-to-video\n   * is capped at 720p.\n   */\n  resolution?: GrokVideoResolution\n\n  /**\n   * Video duration in integer seconds (1–15). In `'extend'` mode this is\n   * the length of the added tail only, not the total output length. Not\n   * valid in `'edit'` mode (the output inherits the source clip's length).\n   */\n  duration?: number\n}\n\n/**\n * Provider options for grok-imagine-video (v1.0), which is the only model\n * that accepts a source-video edit / extend job.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GrokVideoSourceProviderOptions extends GrokVideoBaseProviderOptions {\n  /**\n   * Selects the request mode for a source-video prompt part: `'edit'`\n   * (`/v1/videos/edits`) or `'extend'` (`/v1/videos/extensions`). Required\n   * when the prompt carries a video part; not valid without one. Omit for\n   * plain generation (`/v1/videos/generations`). grok-imagine-video only.\n   */\n  mode?: GrokVideoMode\n}\n\n/**\n * Provider options for grok-imagine-video-1.5, which adds the\n * reference-to-video inputs on top of the shared options.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GrokVideoProviderOptions extends GrokVideoBaseProviderOptions {\n  /**\n   * Reference images for reference-to-video generation (output capped at\n   * 720p). Usually populated from image prompt parts with\n   * `metadata.role: 'reference'` (or `'character'`); set explicitly to\n   * replace the part-derived list. Reference images are addressed from the\n   * prompt text as `<IMAGE_0>`, `<IMAGE_1>`, … in request order, and do not\n   * lock the first frame — pair them with a starting-frame image prompt part\n   * when the opening frame has to be pinned.\n   */\n  reference_images?: Array<{ url: string }>\n\n  /**\n   * Preset TTS voices to reference for generated speech (max 3). Voice ids\n   * come from the xAI TTS voice roster (e.g. 'eve', 'rex') or a custom\n   * voice id, and are addressed from the prompt text as `<AUDIO_0>`,\n   * `<AUDIO_1>`, `<AUDIO_2>`.\n   */\n  reference_audios?: Array<{ voice_id: string }>\n}\n\n/**\n * Widest option surface. Used when `modelOptions` arrives as deserialized\n * JSON and the adapter must validate fields the per-model map already\n * hides at compile time.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoRuntimeOptions = GrokVideoSourceProviderOptions &\n  GrokVideoProviderOptions\n\n/**\n * Maximum reference voices accepted by the Imagine video endpoint.\n */\nexport const GROK_VIDEO_MAX_REFERENCE_AUDIOS = 3\n\n/**\n * Maximum reference images accepted by the Imagine video endpoint.\n */\nexport const GROK_VIDEO_MAX_REFERENCE_IMAGES = 7\n\n/**\n * Model names whose per-model options declare the reference fields. Keeps\n * the runtime set below provably in sync with\n * {@link GrokVideoModelProviderOptionsByName} — a typo or a new\n * reference-capable model missing from the set is a compile error.\n */\ntype GrokVideoReferenceModel = {\n  [TModel in GrokVideoModel]: 'reference_images' extends keyof GrokVideoModelProviderOptionsByName[TModel]\n    ? TModel\n    : never\n}[GrokVideoModel]\n\n/**\n * Models that support reference-to-video inputs (`reference_images` /\n * `reference_audios`). The per-model provider-options map hides the fields\n * from other models at compile time; this backs the runtime gate for\n * untyped callers (e.g. deserialized JSON) so they get a clear error\n * instead of a raw API 400.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nconst GROK_VIDEO_REFERENCE_MODELS: ReadonlySet<string> =\n  new Set<GrokVideoReferenceModel>(['grok-imagine-video-1.5'])\n\n/**\n * True when the model accepts reference-to-video inputs.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function isGrokVideoReferenceModel(model: string): boolean {\n  return GROK_VIDEO_REFERENCE_MODELS.has(model)\n}\n\n/**\n * Model names whose per-model options declare `mode`. Same\n * provably-in-sync construction as {@link GrokVideoReferenceModel}.\n */\ntype GrokVideoSourceModel = {\n  [TModel in GrokVideoModel]: 'mode' extends keyof GrokVideoModelProviderOptionsByName[TModel]\n    ? TModel\n    : never\n}[GrokVideoModel]\n\n/**\n * Models that accept a source-video prompt part for `/v1/videos/edits`\n * and `/v1/videos/extensions`. xAI lists video input only on\n * grok-imagine-video (v1.0).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nconst GROK_VIDEO_SOURCE_MODELS: ReadonlySet<string> =\n  new Set<GrokVideoSourceModel>(['grok-imagine-video'])\n\n/**\n * True when the model accepts edit / extend source-video jobs.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function isGrokVideoSourceModel(model: string): boolean {\n  return GROK_VIDEO_SOURCE_MODELS.has(model)\n}\n\n/**\n * Type-only map from model name to its specific provider options. Only\n * grok-imagine-video-1.5 exposes the reference-to-video fields. Only\n * grok-imagine-video (v1.0) exposes `mode` for edit / extend.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoModelProviderOptionsByName = {\n  'grok-imagine-video': GrokVideoSourceProviderOptions\n  'grok-imagine-video-1.5': GrokVideoProviderOptions\n}\n\n/**\n * Type-only map from model name to its supported `size` strings.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoModelSizeByName = {\n  'grok-imagine-video': GrokVideoSizeV1\n  'grok-imagine-video-1.5': GrokVideoSize\n}\n\n/**\n * Type-only map from model name to the non-text prompt modalities it accepts.\n * Both models support text-to-video and accept an optional `image` prompt\n * part as the starting frame; image parts with `metadata.role: 'reference'`\n * or `'character'` become `reference_images` (grok-imagine-video-1.5 only).\n * On grok-imagine-video-1.5 the two combine: the starting frame pins the\n * first frame while the reference images steer subjects and style.\n * A `video` prompt part carries the source clip for edit / extension mode\n * on grok-imagine-video only (`modelOptions.mode: 'edit' | 'extend'`).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GrokVideoModelInputModalitiesByName = {\n  'grok-imagine-video': readonly ['image', 'video']\n  'grok-imagine-video-1.5': readonly ['image']\n}\n"],"mappings":";AAoEA,IAAM,2BAAkD;CACtD;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,IAAM,yBAAgD;CAAC;CAAQ;CAAQ;AAAO;;;;;;;AAc9E,SAAgB,mBACd,MAC0D;CAE1D,MAAM,GAAG,aAAa,cADR,KAAK,MAAM,6BACW,KAAS,CAAC;CAC9C,IAAI,gBAAgB,KAAA,GAAW,OAAO,KAAA;CACtC,OAAO;EAAE;EAAa,GAAI,eAAe,KAAA,KAAa,EAAE,WAAW;CAAG;AACxE;;;;;;;AAQA,SAAgB,4BAA4B,OAAwB;CAClE,OAAO,UAAU;AACnB;;;;;;AAOA,SAAgB,kBACd,OACA,MAC2C;CAC3C,IAAI,SAAS,KAAA,GAAW;CACxB,MAAM,SAAS,mBAAmB,IAAI;CACtC,IAAI,CAAC,UAAU,CAAC,yBAAyB,SAAS,OAAO,WAAW,GAClE,MAAM,IAAI,MACR,SAAS,KAAK,+BAA+B,MAAM,qGAEzB,yBAAyB,KAAK,IAAI,GAC9D;CAEF,IACE,OAAO,eAAe,KAAA,KACtB,CAAC,uBAAuB,SAAS,OAAO,UAAU,GAElD,MAAM,IAAI,MACR,eAAe,OAAO,WAAW,+BAA+B,MAAM,4BAC1C,uBAAuB,KAAK,IAAI,GAC9D;CAEF,IAAI,OAAO,eAAe,WAAW,CAAC,4BAA4B,KAAK,GACrE,MAAM,IAAI,MACR,iDAAiD,MAAM,iFAEzD;AAEJ;;;;;;;;AAsBA,IAAa,uBAIT;CACF,sBAAsB;EACpB,MAAM;EACN,KAAA;EACA,KAAA;EACA,MAAM;EACN,MAAM;CACR;CACA,0BAA0B;EACxB,MAAM;EACN,KAAA;EACA,KAAA;EACA,MAAM;EACN,MAAM;CACR;AACF;;;;;;AAOA,SAAgB,4BACd,OACuD;CACvD,OAAO,qBAAqB;AAC9B;;;;;;;;;;AAqIA,IAAM,8CACJ,IAAI,IAA6B,CAAC,wBAAwB,CAAC;;;;;;AAO7D,SAAgB,0BAA0B,OAAwB;CAChE,OAAO,4BAA4B,IAAI,KAAK;AAC9C;;;;;;;;AAmBA,IAAM,2CACJ,IAAI,IAA0B,CAAC,oBAAoB,CAAC;;;;;;AAOtD,SAAgB,uBAAuB,OAAwB;CAC7D,OAAO,yBAAyB,IAAI,KAAK;AAC3C"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tanstack/ai-grok",
-  "version": "0.18.5",
+  "version": "0.18.7",
   "description": "xAI Grok adapter for TanStack AI chat, image generation, realtime, and structured outputs.",
   "author": "Tanner Linsley",
   "license": "MIT",
@@ -58,18 +58,18 @@
   ],
   "dependencies": {
     "openai": "^6.41.0",
-    "@tanstack/openai-base": "^0.10.11",
-    "@tanstack/ai-utils": "^0.4.0"
+    "@tanstack/ai-utils": "^0.4.0",
+    "@tanstack/openai-base": "^0.10.12"
   },
   "devDependencies": {
     "@vitest/coverage-v8": "4.1.10",
     "google-auth-library": "^10.5.0",
     "vite": "^8.2.1",
-    "@tanstack/ai": "0.54.0"
+    "@tanstack/ai": "0.55.0"
   },
   "peerDependencies": {
     "google-auth-library": "^10.5.0",
-    "@tanstack/ai": "^0.54.0"
+    "@tanstack/ai": "^0.55.0"
   },
   "peerDependenciesMeta": {
     "google-auth-library": {

package/src/adapters/video.ts

@@ -363,17 +363,13 @@ export class GrokVideoAdapter<
     // Image-to-video: the single image prompt part becomes the starting frame
     // and the prompt text describes the desired motion. URL sources are
     // fetched by xAI's servers; data sources are sent as base64 data URIs.
+    // On grok-imagine-video-1.5 the starting frame combines with reference
+    // inputs — `image` pins the first frame while reference_images /
+    // reference_audios steer subjects and voices. Classic grok-imagine-video
+    // rejects that mix, but it rejects reference inputs outright, so the
+    // model gate above already covers it.
     const [startFrame] = startFrames
 
-    // xAI rejects `image` + `reference_images` / `reference_audios` as a
-    // 400: only one of image-to-video or reference-to-video can be active.
-    if (startFrame && hasReference) {
-      throw new Error(
-        `${this.name}: image-to-video and reference-to-video cannot be combined. ` +
-          `Use a starting-frame image, or reference images / voices, not both.`,
-      )
-    }
-
     // The generic `size` option carries an "aspectRatio_resolution" template
     // (e.g. '16:9_720p') and maps to the Imagine API's `aspect_ratio` /
     // `resolution` parameters; explicit modelOptions win over the template

package/src/video/video-provider-options.ts

@@ -272,7 +272,8 @@ export interface GrokVideoProviderOptions extends GrokVideoBaseProviderOptions {
    * `metadata.role: 'reference'` (or `'character'`); set explicitly to
    * replace the part-derived list. Reference images are addressed from the
    * prompt text as `<IMAGE_0>`, `<IMAGE_1>`, … in request order, and do not
-   * lock the first frame.
+   * lock the first frame — pair them with a starting-frame image prompt part
+   * when the opening frame has to be pinned.
    */
   reference_images?: Array<{ url: string }>
 
@@ -394,6 +395,8 @@ export type GrokVideoModelSizeByName = {
  * Both models support text-to-video and accept an optional `image` prompt
  * part as the starting frame; image parts with `metadata.role: 'reference'`
  * or `'character'` become `reference_images` (grok-imagine-video-1.5 only).
+ * On grok-imagine-video-1.5 the two combine: the starting frame pins the
+ * first frame while the reference images steer subjects and style.
  * A `video` prompt part carries the source clip for edit / extension mode
  * on grok-imagine-video only (`modelOptions.mode: 'edit' | 'extend'`).
  *