@shop-os/modelkit 0.2.26 → 0.3.0

package/dist/index.d.cts

@@ -42,7 +42,13 @@ interface UpscaleOptions {
     faceEnhancementCreativity?: number;
     faceEnhancementStrength?: number;
 }
-type ModelType = "text" | "image" | "video" | "audio" | "upscale" | "bg-remove";
+type ModelType = "text" | "image" | "video" | "audio" | "upscale" | "bg-remove"
+/** Video in, video out, no prompt: upscale, frame interpolation, SDR to HDR. */
+ | "video-tool";
+/** What a `video-tool` row does to its source clip. */
+type VideoToolKind = "upscale" | "interpolate" | "hdr";
+/** Container/codec a video-tool row can write. */
+type VideoToolOutputFormat = "h265" | "prores";
 type Lifecycle = "active" | "deprecated" | "retired";
 type Surface = "cowork" | "space_builder";
 /** @deprecated use Surface[] */
@@ -133,6 +139,61 @@ interface VideoCallOptions extends Omit<VideoRequest, "model"> {
     /** Reaches the provider only for AI SDK model instances; catalog rows carry their own. */
     providerOptions?: ProviderOptions;
 }
+/** What the caller measured about a video-tool source; the kit never measures. */
+interface VideoToolSource {
+    url: string;
+    width: number;
+    height: number;
+    durationSeconds: number;
+    fps?: number;
+}
+interface VideoToolRequest {
+    model: string;
+    source: VideoToolSource;
+    /** Upscale rows. */
+    upscaleFactor?: number;
+    /** Interpolate rows. */
+    targetFps?: number;
+    /** HDR rows; the row's first `outputFormats` entry when unset. */
+    outputFormat?: VideoToolOutputFormat;
+}
+/** The output a video-tool request will produce, from the request alone. */
+interface VideoToolOutputSpec {
+    width: number;
+    height: number;
+    durationSeconds: number;
+    fps?: number;
+    resolution: VideoResolution;
+    format: "h264" | VideoToolOutputFormat;
+}
+interface VideoToolCallContext extends CallContext {
+    /** Where the provider posts the result when the job ends. */
+    webhookUrl?: string;
+}
+/**
+ * A submitted video-tool job. Plain JSON so a caller can store it and pass it back
+ * to `videoToolStatus` / `videoToolResult` / `cancelVideoTool` from another process.
+ */
+interface VideoToolSubmission {
+    model: string;
+    provider: "fal";
+    requestId: string;
+    statusUrl: string;
+    responseUrl: string;
+    cancelUrl: string;
+    output: VideoToolOutputSpec;
+}
+type VideoToolStatus = "queued" | "running" | "completed";
+interface VideoToolResult {
+    video: {
+        url: string;
+        mime?: string;
+        bytes?: number;
+    };
+    output: VideoToolOutputSpec;
+    usage: UsageEvent;
+    raw?: unknown;
+}
 interface AudioRequest {
     model: string;
     prompt: string;
@@ -193,6 +254,33 @@ interface ReferenceLimits {
     audioOnly?: boolean;
     /** Per-video length caps (min/max each) and combined total for reference videos. */
     videoSeconds?: VideoSecondsLimits;
+    /** Per-clip length caps (min/max each) and combined total for reference audio, in seconds. */
+    audioSeconds?: {
+        min?: number;
+        max?: number;
+        total?: number;
+    };
+    /** File types the provider takes for reference audio, as lowercase extensions. */
+    audioFormats?: string[];
+    /** Largest reference audio file the provider takes, in bytes. */
+    audioMaxBytes?: number;
+}
+interface VideoToolConstraints {
+    kind: VideoToolKind;
+    /** Upscale: allowed factor range. */
+    upscaleFactor?: {
+        min: number;
+        max: number;
+    };
+    /** Upscale: the output's long edge may not pass this, in pixels. */
+    maxOutputLongEdge?: number;
+    /** Interpolate: allowed target frame rate range. */
+    fps?: {
+        min: number;
+        max: number;
+    };
+    /** HDR: the formats the row writes, first is the default. */
+    outputFormats?: VideoToolOutputFormat[];
 }
 /** Whether a video row's output has sound: always, switchable (`supports: generateAudio`), or silent. */
 type AudioOutput = "always" | "optional" | "never";
@@ -235,7 +323,12 @@ interface ModelConstraints {
     extension?: {
         segmentSeconds: number;
         maxTotalSeconds: number;
+        /** `joined` (absent): the provider returns the whole clip so far (Omni).
+         * `new`: it returns only the added footage and the caller joins it (Seedance). */
+        returns?: "new" | "joined";
     };
+    /** What a `video-tool` row accepts. Read it through `videoToolConstraints(row)`. */
+    tool?: VideoToolConstraints;
     /** Whether the output has sound. Read it through `outputAudio(row)`, which tolerates bad data. */
     audio?: AudioOutput;
     /** Prompt conventions. Read them through `promptConventions(row)`, which fills the default. */
@@ -296,6 +389,11 @@ interface CatalogDisplay {
 interface CatalogCapabilities {
     /** Video rows: takes an existing video as the thing to edit or extend. */
     acceptsSourceVideo?: boolean;
+    /**
+     * Video rows: offered for prompt-driven edits of a source clip (`taskType: "edit"`). Absent
+     * means the caller falls back to `acceptsSourceVideo`; `false` opts a row out.
+     */
+    videoEdit?: boolean;
     /** Video rows: first/last frames and reference images are mutually exclusive inputs. */
     framesExcludeReferences?: boolean;
     /** Video rows: a last frame is accepted only together with a first frame. */
@@ -414,6 +512,8 @@ interface UsageEvent {
     imageQuality?: ImageQuality;
     resolution?: VideoResolution;
     durationSeconds?: number;
+    /** Output frame rate (video-tool rows price by output frames). */
+    fps?: number;
     inputTokens?: number;
     outputTokens?: number;
     reportedCostUsd?: number;
@@ -580,6 +680,7 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
         audio: "audio";
         upscale: "upscale";
         "bg-remove": "bg-remove";
+        "video-tool": "video-tool";
     }>;
     adapter: z.ZodEnum<{
         "ai-sdk-text": "ai-sdk-text";
@@ -652,6 +753,13 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
                 total: z.ZodOptional<z.ZodNumber>;
                 editMin: z.ZodOptional<z.ZodNumber>;
             }, z.core.$loose>>;
+            audioSeconds: z.ZodOptional<z.ZodObject<{
+                min: z.ZodOptional<z.ZodNumber>;
+                max: z.ZodOptional<z.ZodNumber>;
+                total: z.ZodOptional<z.ZodNumber>;
+            }, z.core.$loose>>;
+            audioFormats: z.ZodOptional<z.ZodArray<z.ZodString>>;
+            audioMaxBytes: z.ZodOptional<z.ZodNumber>;
         }, z.core.$loose>>;
         requires: z.ZodOptional<z.ZodArray<z.ZodEnum<{
             references: "references";
@@ -691,6 +799,30 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
         extension: z.ZodOptional<z.ZodObject<{
             segmentSeconds: z.ZodNumber;
             maxTotalSeconds: z.ZodNumber;
+            returns: z.ZodOptional<z.ZodEnum<{
+                new: "new";
+                joined: "joined";
+            }>>;
+        }, z.core.$loose>>;
+        tool: z.ZodOptional<z.ZodObject<{
+            kind: z.ZodEnum<{
+                upscale: "upscale";
+                interpolate: "interpolate";
+                hdr: "hdr";
+            }>;
+            upscaleFactor: z.ZodOptional<z.ZodObject<{
+                min: z.ZodNumber;
+                max: z.ZodNumber;
+            }, z.core.$loose>>;
+            maxOutputLongEdge: z.ZodOptional<z.ZodNumber>;
+            fps: z.ZodOptional<z.ZodObject<{
+                min: z.ZodNumber;
+                max: z.ZodNumber;
+            }, z.core.$loose>>;
+            outputFormats: z.ZodOptional<z.ZodArray<z.ZodEnum<{
+                h265: "h265";
+                prores: "prores";
+            }>>>;
         }, z.core.$loose>>;
         audio: z.ZodOptional<z.ZodUnknown>;
         prompt: z.ZodOptional<z.ZodUnknown>;
@@ -879,6 +1011,7 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
     }, z.core.$loose>>;
     capabilities: z.ZodOptional<z.ZodObject<{
         acceptsSourceVideo: z.ZodOptional<z.ZodBoolean>;
+        videoEdit: z.ZodOptional<z.ZodBoolean>;
         framesExcludeReferences: z.ZodOptional<z.ZodBoolean>;
         lastFrameRequiresFirstFrame: z.ZodOptional<z.ZodBoolean>;
         reasoning: z.ZodOptional<z.ZodBoolean>;
@@ -923,6 +1056,20 @@ declare function validateInputSpec(raw: ProviderOptionValue, row: ValidateRow):
 declare const ANTHROPIC_TRANSPORTS: readonly ["direct", "bedrock"];
 type AnthropicTransport = (typeof ANTHROPIC_TRANSPORTS)[number];
 
+/** fal's webhook body (https://fal.ai/docs/model-apis/model-endpoints/webhooks). */
+declare const falWebhookBodySchema: z.ZodObject<{
+    request_id: z.ZodString;
+    gateway_request_id: z.ZodOptional<z.ZodNullable<z.ZodString>>;
+    status: z.ZodEnum<{
+        OK: "OK";
+        ERROR: "ERROR";
+    }>;
+    payload: z.ZodOptional<z.ZodNullable<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    error: z.ZodOptional<z.ZodNullable<z.ZodString>>;
+    payload_error: z.ZodOptional<z.ZodNullable<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strip>;
+type FalWebhookBody = z.infer<typeof falWebhookBodySchema>;
+
 /** Everything an adapter needs for one call. Resolved by the core before dispatch. */
 interface AdapterContext {
     model: CatalogModel;
@@ -947,6 +1094,13 @@ interface Adapter {
     generateImage?(req: ImageRequest, ctx: AdapterContext): Promise<ImageResult>;
     generateVideo?(req: VideoRequest, ctx: AdapterContext): Promise<VideoJob>;
     generateAudio?(req: AudioRequest, ctx: AdapterContext): Promise<AudioResult>;
+    /** Video tools run as provider jobs the caller tracks; see ModelKit.submitVideoTool. */
+    submitVideoTool?(req: VideoToolRequest, ctx: AdapterContext, opts: {
+        webhookUrl?: string;
+    }): Promise<VideoToolSubmission>;
+    videoToolStatus?(sub: VideoToolSubmission, ctx: AdapterContext): Promise<VideoToolStatus>;
+    videoToolResult?(sub: VideoToolSubmission, ctx: AdapterContext, webhook?: FalWebhookBody): Promise<VideoToolResult>;
+    cancelVideoTool?(sub: VideoToolSubmission, ctx: AdapterContext): Promise<boolean>;
     /**
      * Turn a media reference into an AI SDK file part the provider's language
      * model accepts. Google uploads large videos through the Files API.
@@ -1163,6 +1317,21 @@ interface ModelKit extends TextMethods {
     prepareMediaPart(modelId: string, ref: MediaRef, mediaType: string, call?: CallContext): Promise<MediaPart>;
     /** Which options a caller may show for this model. */
     describe(modelId: string): CatalogModel;
+    /**
+     * Start a video-tool job (`type: "video-tool"`) and return without waiting. Store the
+     * submission; the provider posts to `call.webhookUrl` when the job ends.
+     */
+    submitVideoTool(req: VideoToolRequest, call?: VideoToolCallContext): Promise<VideoToolSubmission>;
+    videoToolStatus(sub: VideoToolSubmission, call?: CallContext): Promise<VideoToolStatus>;
+    /**
+     * The finished output. Pass the verified webhook body when there is one; without it the
+     * kit fetches the provider's stored result. Emits the usage event.
+     */
+    videoToolResult(sub: VideoToolSubmission, call?: CallContext & {
+        webhook?: FalWebhookBody;
+    }): Promise<VideoToolResult>;
+    /** True when the provider accepted the cancel, false when the job had already ended. */
+    cancelVideoTool(sub: VideoToolSubmission, call?: CallContext): Promise<boolean>;
 }
 interface LanguageModelHandle {
     model: LanguageModelV3;
@@ -1209,6 +1378,29 @@ declare const xaiVideoAdapter: Adapter;
 
 declare const DEFAULT_ADAPTERS: Record<AdapterId, Adapter>;
 
+/** A video-tool row's constraints, or undefined when the row has none or they are invalid. */
+declare function videoToolConstraints(row: CatalogModel): VideoToolConstraints | undefined;
+/** The pricing tier a frame lands in, by its short edge. */
+declare function resolutionTier(width: number, height: number): VideoResolution;
+type VideoToolPlan = {
+    ok: true;
+    output: VideoToolOutputSpec;
+} | {
+    ok: false;
+    error: string;
+};
+/**
+ * What a video-tool request produces, checked against the row. Pure: callers use it to
+ * validate and price before submitting, and the adapter uses it to build the usage event.
+ */
+declare function planVideoTool(row: CatalogModel, req: Omit<VideoToolRequest, "model">): VideoToolPlan;
+
+/**
+ * Topaz Labs video tools on fal. `providerOptions.fal.model` names the Topaz model the
+ * endpoint runs; the adapter sends it as `model`. Exposed on cowork only (Creative Studio).
+ */
+declare const VIDEO_TOOL_MODELS: CatalogModel[];
+
 /**
  * All adapters registered. Keys default to process.env (pro = bare names,
  * enterprise = ENTERPRISE_ prefix, sloosh = SLOOSH_ prefix) when `keys` is omitted.
@@ -1244,4 +1436,4 @@ declare function describeThrown(cause: unknown): ThrownErrorInfo;
 /** Add the AI SDK `GeneratedFile` accessors to adapter image bytes. base64 is computed on first read. */
 declare function toGeneratedImage(image: ImageBytes): GeneratedImage;
 
-export { ADAPTER_IDS, ADAPTER_KEYS, AUDIO_INPUT_SPEC_KEYS, type Adapter, type AdapterContext, type AdapterEndpoints, type AdapterId, type AdapterRegistry, type AiTextProviderOptions, type AspectRatio, type AudioOutput, type AudioRequest, type AudioResult, type CallContext, Catalog, type CatalogAuth, type CatalogCache, type CatalogCapabilities, type CatalogDisplay, type CatalogEnvelope, type CatalogModel, CatalogModelSchema, CatalogModelSchemaBase, type CatalogOrigin, CatalogSchema, type CatalogSource, type CatalogTextModel, DEFAULT_ADAPTERS, type DataOnlyRow, type DataOnlyVerdict, type DefaultObjectOutput, type Exposure, ExposureSchema, type GeneratedImage, IMAGE_INPUT_SPEC_KEYS, INPUT_SPEC_SUPPORTS_KEYS, type ImageBytes, type ImageCallOptions, type ImageCallResult, type ImageQuality, type ImageRequest, type ImageResult, type ImageSafetySetting, type ImageSize, type InputSpec, type InputSpecIssue, type InputSpecKey, InputSpecSchema, type JsonObject, JsonObjectSchema, type JsonPrimitive, type JsonValue, JsonValueSchema, type KeyTier, type LanguageModelHandle, type Lifecycle, type LiveCatalogState, MODELS, type MediaPart, type MediaRef, type ModelConstraints, type ModelKit, type ModelKitConfig, ModelKitError, type ModelKitErrorCode, type ModelKitErrorInit, type ModelType, type ModelkitGenerateObjectOptions, type ModelkitGenerateTextOptions, type ModelkitStreamObjectOptions, type ModelkitStreamTextOptions, type ObjectOutput, type ObjectResult, type OnUnsupported, type OutputFormat, type PricingKind, type PrimeCatalogResult, type PromptConventions, type ProviderErrorBody, type ProviderKeyName, type ProviderKeys, type ProviderOptionValue, type ProviderOptions, type QuarantineResult, type ReferenceLimits, type ReferenceSyntax, type RequiredInput, type RequiredKeyName, type ResolveResult, type Routing, SURFACES, type SupportedOption, type Surface, SurfaceSchema, type TextMethods, type TextModelInstance, type UpscaleOptions, type UsageEvent, type UsageModality, type UsageUnit, VIDEO_INPUT_SPEC_KEYS, type VideoCallJob, type VideoCallOptions, type VideoJob, type VideoModel, type VideoRequest, type VideoResolution, type VideoResult, type VideoSecondsLimits, type Warning, adapterServesType, aiSdkTextAdapter, arkAdapter, aspectRatioToSize, classifyHttpError, compareCatalogOrder, createDefaultModelKit, createModelKit, dataOnlyCapable, describe, describeThrown, generateImage as experimental_generateImage, falAdapter, fileCatalogCache, generateAudio, generateImage, generateObject, generateText, generateVideo, googleGenaiAdapter, googleImageAdapter, hasAwsCredentials, higgsfieldAdapter, isInputSpecKey, isListed, isListedOn, isModelKitError, keysFromEnv, languageModel, liveCatalog, liveState, mergeTextProviderOptions, missingTierKeys, nearestAspectRatio, normalizeExposure, openaiImageAdapter, passthroughTextModel, pollUntil, prepareMediaPart, primeCatalog, prodiaAdapter, readErrorBody, replicateAdapter, resetLiveCatalog, resolveAudioOptions, resolveImageOptions, resolveKeys, resolveVideoOptions, shouldRetry, sleep, sortCatalogRows, streamObject, streamText, textUsageEvent, toGeneratedImage, validateInputSpec, videoLengthViolation, withRetry, xaiVideoAdapter };
+export { ADAPTER_IDS, ADAPTER_KEYS, AUDIO_INPUT_SPEC_KEYS, type Adapter, type AdapterContext, type AdapterEndpoints, type AdapterId, type AdapterRegistry, type AiTextProviderOptions, type AspectRatio, type AudioOutput, type AudioRequest, type AudioResult, type CallContext, Catalog, type CatalogAuth, type CatalogCache, type CatalogCapabilities, type CatalogDisplay, type CatalogEnvelope, type CatalogModel, CatalogModelSchema, CatalogModelSchemaBase, type CatalogOrigin, CatalogSchema, type CatalogSource, type CatalogTextModel, DEFAULT_ADAPTERS, type DataOnlyRow, type DataOnlyVerdict, type DefaultObjectOutput, type Exposure, ExposureSchema, type FalWebhookBody, type GeneratedImage, IMAGE_INPUT_SPEC_KEYS, INPUT_SPEC_SUPPORTS_KEYS, type ImageBytes, type ImageCallOptions, type ImageCallResult, type ImageQuality, type ImageRequest, type ImageResult, type ImageSafetySetting, type ImageSize, type InputSpec, type InputSpecIssue, type InputSpecKey, InputSpecSchema, type JsonObject, JsonObjectSchema, type JsonPrimitive, type JsonValue, JsonValueSchema, type KeyTier, type LanguageModelHandle, type Lifecycle, type LiveCatalogState, MODELS, type MediaPart, type MediaRef, type ModelConstraints, type ModelKit, type ModelKitConfig, ModelKitError, type ModelKitErrorCode, type ModelKitErrorInit, type ModelType, type ModelkitGenerateObjectOptions, type ModelkitGenerateTextOptions, type ModelkitStreamObjectOptions, type ModelkitStreamTextOptions, type ObjectOutput, type ObjectResult, type OnUnsupported, type OutputFormat, type PricingKind, type PrimeCatalogResult, type PromptConventions, type ProviderErrorBody, type ProviderKeyName, type ProviderKeys, type ProviderOptionValue, type ProviderOptions, type QuarantineResult, type ReferenceLimits, type ReferenceSyntax, type RequiredInput, type RequiredKeyName, type ResolveResult, type Routing, SURFACES, type SupportedOption, type Surface, SurfaceSchema, type TextMethods, type TextModelInstance, type UpscaleOptions, type UsageEvent, type UsageModality, type UsageUnit, VIDEO_INPUT_SPEC_KEYS, VIDEO_TOOL_MODELS, type VideoCallJob, type VideoCallOptions, type VideoJob, type VideoModel, type VideoRequest, type VideoResolution, type VideoResult, type VideoSecondsLimits, type VideoToolCallContext, type VideoToolConstraints, type VideoToolKind, type VideoToolOutputFormat, type VideoToolOutputSpec, type VideoToolPlan, type VideoToolRequest, type VideoToolResult, type VideoToolSource, type VideoToolStatus, type VideoToolSubmission, type Warning, adapterServesType, aiSdkTextAdapter, arkAdapter, aspectRatioToSize, classifyHttpError, compareCatalogOrder, createDefaultModelKit, createModelKit, dataOnlyCapable, describe, describeThrown, generateImage as experimental_generateImage, falAdapter, falWebhookBodySchema, fileCatalogCache, generateAudio, generateImage, generateObject, generateText, generateVideo, googleGenaiAdapter, googleImageAdapter, hasAwsCredentials, higgsfieldAdapter, isInputSpecKey, isListed, isListedOn, isModelKitError, keysFromEnv, languageModel, liveCatalog, liveState, mergeTextProviderOptions, missingTierKeys, nearestAspectRatio, normalizeExposure, openaiImageAdapter, passthroughTextModel, planVideoTool, pollUntil, prepareMediaPart, primeCatalog, prodiaAdapter, readErrorBody, replicateAdapter, resetLiveCatalog, resolutionTier, resolveAudioOptions, resolveImageOptions, resolveKeys, resolveVideoOptions, shouldRetry, sleep, sortCatalogRows, streamObject, streamText, textUsageEvent, toGeneratedImage, validateInputSpec, videoLengthViolation, videoToolConstraints, withRetry, xaiVideoAdapter };

package/dist/index.d.ts

@@ -42,7 +42,13 @@ interface UpscaleOptions {
     faceEnhancementCreativity?: number;
     faceEnhancementStrength?: number;
 }
-type ModelType = "text" | "image" | "video" | "audio" | "upscale" | "bg-remove";
+type ModelType = "text" | "image" | "video" | "audio" | "upscale" | "bg-remove"
+/** Video in, video out, no prompt: upscale, frame interpolation, SDR to HDR. */
+ | "video-tool";
+/** What a `video-tool` row does to its source clip. */
+type VideoToolKind = "upscale" | "interpolate" | "hdr";
+/** Container/codec a video-tool row can write. */
+type VideoToolOutputFormat = "h265" | "prores";
 type Lifecycle = "active" | "deprecated" | "retired";
 type Surface = "cowork" | "space_builder";
 /** @deprecated use Surface[] */
@@ -133,6 +139,61 @@ interface VideoCallOptions extends Omit<VideoRequest, "model"> {
     /** Reaches the provider only for AI SDK model instances; catalog rows carry their own. */
     providerOptions?: ProviderOptions;
 }
+/** What the caller measured about a video-tool source; the kit never measures. */
+interface VideoToolSource {
+    url: string;
+    width: number;
+    height: number;
+    durationSeconds: number;
+    fps?: number;
+}
+interface VideoToolRequest {
+    model: string;
+    source: VideoToolSource;
+    /** Upscale rows. */
+    upscaleFactor?: number;
+    /** Interpolate rows. */
+    targetFps?: number;
+    /** HDR rows; the row's first `outputFormats` entry when unset. */
+    outputFormat?: VideoToolOutputFormat;
+}
+/** The output a video-tool request will produce, from the request alone. */
+interface VideoToolOutputSpec {
+    width: number;
+    height: number;
+    durationSeconds: number;
+    fps?: number;
+    resolution: VideoResolution;
+    format: "h264" | VideoToolOutputFormat;
+}
+interface VideoToolCallContext extends CallContext {
+    /** Where the provider posts the result when the job ends. */
+    webhookUrl?: string;
+}
+/**
+ * A submitted video-tool job. Plain JSON so a caller can store it and pass it back
+ * to `videoToolStatus` / `videoToolResult` / `cancelVideoTool` from another process.
+ */
+interface VideoToolSubmission {
+    model: string;
+    provider: "fal";
+    requestId: string;
+    statusUrl: string;
+    responseUrl: string;
+    cancelUrl: string;
+    output: VideoToolOutputSpec;
+}
+type VideoToolStatus = "queued" | "running" | "completed";
+interface VideoToolResult {
+    video: {
+        url: string;
+        mime?: string;
+        bytes?: number;
+    };
+    output: VideoToolOutputSpec;
+    usage: UsageEvent;
+    raw?: unknown;
+}
 interface AudioRequest {
     model: string;
     prompt: string;
@@ -193,6 +254,33 @@ interface ReferenceLimits {
     audioOnly?: boolean;
     /** Per-video length caps (min/max each) and combined total for reference videos. */
     videoSeconds?: VideoSecondsLimits;
+    /** Per-clip length caps (min/max each) and combined total for reference audio, in seconds. */
+    audioSeconds?: {
+        min?: number;
+        max?: number;
+        total?: number;
+    };
+    /** File types the provider takes for reference audio, as lowercase extensions. */
+    audioFormats?: string[];
+    /** Largest reference audio file the provider takes, in bytes. */
+    audioMaxBytes?: number;
+}
+interface VideoToolConstraints {
+    kind: VideoToolKind;
+    /** Upscale: allowed factor range. */
+    upscaleFactor?: {
+        min: number;
+        max: number;
+    };
+    /** Upscale: the output's long edge may not pass this, in pixels. */
+    maxOutputLongEdge?: number;
+    /** Interpolate: allowed target frame rate range. */
+    fps?: {
+        min: number;
+        max: number;
+    };
+    /** HDR: the formats the row writes, first is the default. */
+    outputFormats?: VideoToolOutputFormat[];
 }
 /** Whether a video row's output has sound: always, switchable (`supports: generateAudio`), or silent. */
 type AudioOutput = "always" | "optional" | "never";
@@ -235,7 +323,12 @@ interface ModelConstraints {
     extension?: {
         segmentSeconds: number;
         maxTotalSeconds: number;
+        /** `joined` (absent): the provider returns the whole clip so far (Omni).
+         * `new`: it returns only the added footage and the caller joins it (Seedance). */
+        returns?: "new" | "joined";
     };
+    /** What a `video-tool` row accepts. Read it through `videoToolConstraints(row)`. */
+    tool?: VideoToolConstraints;
     /** Whether the output has sound. Read it through `outputAudio(row)`, which tolerates bad data. */
     audio?: AudioOutput;
     /** Prompt conventions. Read them through `promptConventions(row)`, which fills the default. */
@@ -296,6 +389,11 @@ interface CatalogDisplay {
 interface CatalogCapabilities {
     /** Video rows: takes an existing video as the thing to edit or extend. */
     acceptsSourceVideo?: boolean;
+    /**
+     * Video rows: offered for prompt-driven edits of a source clip (`taskType: "edit"`). Absent
+     * means the caller falls back to `acceptsSourceVideo`; `false` opts a row out.
+     */
+    videoEdit?: boolean;
     /** Video rows: first/last frames and reference images are mutually exclusive inputs. */
     framesExcludeReferences?: boolean;
     /** Video rows: a last frame is accepted only together with a first frame. */
@@ -414,6 +512,8 @@ interface UsageEvent {
     imageQuality?: ImageQuality;
     resolution?: VideoResolution;
     durationSeconds?: number;
+    /** Output frame rate (video-tool rows price by output frames). */
+    fps?: number;
     inputTokens?: number;
     outputTokens?: number;
     reportedCostUsd?: number;
@@ -580,6 +680,7 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
         audio: "audio";
         upscale: "upscale";
         "bg-remove": "bg-remove";
+        "video-tool": "video-tool";
     }>;
     adapter: z.ZodEnum<{
         "ai-sdk-text": "ai-sdk-text";
@@ -652,6 +753,13 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
                 total: z.ZodOptional<z.ZodNumber>;
                 editMin: z.ZodOptional<z.ZodNumber>;
             }, z.core.$loose>>;
+            audioSeconds: z.ZodOptional<z.ZodObject<{
+                min: z.ZodOptional<z.ZodNumber>;
+                max: z.ZodOptional<z.ZodNumber>;
+                total: z.ZodOptional<z.ZodNumber>;
+            }, z.core.$loose>>;
+            audioFormats: z.ZodOptional<z.ZodArray<z.ZodString>>;
+            audioMaxBytes: z.ZodOptional<z.ZodNumber>;
         }, z.core.$loose>>;
         requires: z.ZodOptional<z.ZodArray<z.ZodEnum<{
             references: "references";
@@ -691,6 +799,30 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
         extension: z.ZodOptional<z.ZodObject<{
             segmentSeconds: z.ZodNumber;
             maxTotalSeconds: z.ZodNumber;
+            returns: z.ZodOptional<z.ZodEnum<{
+                new: "new";
+                joined: "joined";
+            }>>;
+        }, z.core.$loose>>;
+        tool: z.ZodOptional<z.ZodObject<{
+            kind: z.ZodEnum<{
+                upscale: "upscale";
+                interpolate: "interpolate";
+                hdr: "hdr";
+            }>;
+            upscaleFactor: z.ZodOptional<z.ZodObject<{
+                min: z.ZodNumber;
+                max: z.ZodNumber;
+            }, z.core.$loose>>;
+            maxOutputLongEdge: z.ZodOptional<z.ZodNumber>;
+            fps: z.ZodOptional<z.ZodObject<{
+                min: z.ZodNumber;
+                max: z.ZodNumber;
+            }, z.core.$loose>>;
+            outputFormats: z.ZodOptional<z.ZodArray<z.ZodEnum<{
+                h265: "h265";
+                prores: "prores";
+            }>>>;
         }, z.core.$loose>>;
         audio: z.ZodOptional<z.ZodUnknown>;
         prompt: z.ZodOptional<z.ZodUnknown>;
@@ -879,6 +1011,7 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
     }, z.core.$loose>>;
     capabilities: z.ZodOptional<z.ZodObject<{
         acceptsSourceVideo: z.ZodOptional<z.ZodBoolean>;
+        videoEdit: z.ZodOptional<z.ZodBoolean>;
         framesExcludeReferences: z.ZodOptional<z.ZodBoolean>;
         lastFrameRequiresFirstFrame: z.ZodOptional<z.ZodBoolean>;
         reasoning: z.ZodOptional<z.ZodBoolean>;
@@ -923,6 +1056,20 @@ declare function validateInputSpec(raw: ProviderOptionValue, row: ValidateRow):
 declare const ANTHROPIC_TRANSPORTS: readonly ["direct", "bedrock"];
 type AnthropicTransport = (typeof ANTHROPIC_TRANSPORTS)[number];
 
+/** fal's webhook body (https://fal.ai/docs/model-apis/model-endpoints/webhooks). */
+declare const falWebhookBodySchema: z.ZodObject<{
+    request_id: z.ZodString;
+    gateway_request_id: z.ZodOptional<z.ZodNullable<z.ZodString>>;
+    status: z.ZodEnum<{
+        OK: "OK";
+        ERROR: "ERROR";
+    }>;
+    payload: z.ZodOptional<z.ZodNullable<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    error: z.ZodOptional<z.ZodNullable<z.ZodString>>;
+    payload_error: z.ZodOptional<z.ZodNullable<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strip>;
+type FalWebhookBody = z.infer<typeof falWebhookBodySchema>;
+
 /** Everything an adapter needs for one call. Resolved by the core before dispatch. */
 interface AdapterContext {
     model: CatalogModel;
@@ -947,6 +1094,13 @@ interface Adapter {
     generateImage?(req: ImageRequest, ctx: AdapterContext): Promise<ImageResult>;
     generateVideo?(req: VideoRequest, ctx: AdapterContext): Promise<VideoJob>;
     generateAudio?(req: AudioRequest, ctx: AdapterContext): Promise<AudioResult>;
+    /** Video tools run as provider jobs the caller tracks; see ModelKit.submitVideoTool. */
+    submitVideoTool?(req: VideoToolRequest, ctx: AdapterContext, opts: {
+        webhookUrl?: string;
+    }): Promise<VideoToolSubmission>;
+    videoToolStatus?(sub: VideoToolSubmission, ctx: AdapterContext): Promise<VideoToolStatus>;
+    videoToolResult?(sub: VideoToolSubmission, ctx: AdapterContext, webhook?: FalWebhookBody): Promise<VideoToolResult>;
+    cancelVideoTool?(sub: VideoToolSubmission, ctx: AdapterContext): Promise<boolean>;
     /**
      * Turn a media reference into an AI SDK file part the provider's language
      * model accepts. Google uploads large videos through the Files API.
@@ -1163,6 +1317,21 @@ interface ModelKit extends TextMethods {
     prepareMediaPart(modelId: string, ref: MediaRef, mediaType: string, call?: CallContext): Promise<MediaPart>;
     /** Which options a caller may show for this model. */
     describe(modelId: string): CatalogModel;
+    /**
+     * Start a video-tool job (`type: "video-tool"`) and return without waiting. Store the
+     * submission; the provider posts to `call.webhookUrl` when the job ends.
+     */
+    submitVideoTool(req: VideoToolRequest, call?: VideoToolCallContext): Promise<VideoToolSubmission>;
+    videoToolStatus(sub: VideoToolSubmission, call?: CallContext): Promise<VideoToolStatus>;
+    /**
+     * The finished output. Pass the verified webhook body when there is one; without it the
+     * kit fetches the provider's stored result. Emits the usage event.
+     */
+    videoToolResult(sub: VideoToolSubmission, call?: CallContext & {
+        webhook?: FalWebhookBody;
+    }): Promise<VideoToolResult>;
+    /** True when the provider accepted the cancel, false when the job had already ended. */
+    cancelVideoTool(sub: VideoToolSubmission, call?: CallContext): Promise<boolean>;
 }
 interface LanguageModelHandle {
     model: LanguageModelV3;
@@ -1209,6 +1378,29 @@ declare const xaiVideoAdapter: Adapter;
 
 declare const DEFAULT_ADAPTERS: Record<AdapterId, Adapter>;
 
+/** A video-tool row's constraints, or undefined when the row has none or they are invalid. */
+declare function videoToolConstraints(row: CatalogModel): VideoToolConstraints | undefined;
+/** The pricing tier a frame lands in, by its short edge. */
+declare function resolutionTier(width: number, height: number): VideoResolution;
+type VideoToolPlan = {
+    ok: true;
+    output: VideoToolOutputSpec;
+} | {
+    ok: false;
+    error: string;
+};
+/**
+ * What a video-tool request produces, checked against the row. Pure: callers use it to
+ * validate and price before submitting, and the adapter uses it to build the usage event.
+ */
+declare function planVideoTool(row: CatalogModel, req: Omit<VideoToolRequest, "model">): VideoToolPlan;
+
+/**
+ * Topaz Labs video tools on fal. `providerOptions.fal.model` names the Topaz model the
+ * endpoint runs; the adapter sends it as `model`. Exposed on cowork only (Creative Studio).
+ */
+declare const VIDEO_TOOL_MODELS: CatalogModel[];
+
 /**
  * All adapters registered. Keys default to process.env (pro = bare names,
  * enterprise = ENTERPRISE_ prefix, sloosh = SLOOSH_ prefix) when `keys` is omitted.
@@ -1244,4 +1436,4 @@ declare function describeThrown(cause: unknown): ThrownErrorInfo;
 /** Add the AI SDK `GeneratedFile` accessors to adapter image bytes. base64 is computed on first read. */
 declare function toGeneratedImage(image: ImageBytes): GeneratedImage;
 
-export { ADAPTER_IDS, ADAPTER_KEYS, AUDIO_INPUT_SPEC_KEYS, type Adapter, type AdapterContext, type AdapterEndpoints, type AdapterId, type AdapterRegistry, type AiTextProviderOptions, type AspectRatio, type AudioOutput, type AudioRequest, type AudioResult, type CallContext, Catalog, type CatalogAuth, type CatalogCache, type CatalogCapabilities, type CatalogDisplay, type CatalogEnvelope, type CatalogModel, CatalogModelSchema, CatalogModelSchemaBase, type CatalogOrigin, CatalogSchema, type CatalogSource, type CatalogTextModel, DEFAULT_ADAPTERS, type DataOnlyRow, type DataOnlyVerdict, type DefaultObjectOutput, type Exposure, ExposureSchema, type GeneratedImage, IMAGE_INPUT_SPEC_KEYS, INPUT_SPEC_SUPPORTS_KEYS, type ImageBytes, type ImageCallOptions, type ImageCallResult, type ImageQuality, type ImageRequest, type ImageResult, type ImageSafetySetting, type ImageSize, type InputSpec, type InputSpecIssue, type InputSpecKey, InputSpecSchema, type JsonObject, JsonObjectSchema, type JsonPrimitive, type JsonValue, JsonValueSchema, type KeyTier, type LanguageModelHandle, type Lifecycle, type LiveCatalogState, MODELS, type MediaPart, type MediaRef, type ModelConstraints, type ModelKit, type ModelKitConfig, ModelKitError, type ModelKitErrorCode, type ModelKitErrorInit, type ModelType, type ModelkitGenerateObjectOptions, type ModelkitGenerateTextOptions, type ModelkitStreamObjectOptions, type ModelkitStreamTextOptions, type ObjectOutput, type ObjectResult, type OnUnsupported, type OutputFormat, type PricingKind, type PrimeCatalogResult, type PromptConventions, type ProviderErrorBody, type ProviderKeyName, type ProviderKeys, type ProviderOptionValue, type ProviderOptions, type QuarantineResult, type ReferenceLimits, type ReferenceSyntax, type RequiredInput, type RequiredKeyName, type ResolveResult, type Routing, SURFACES, type SupportedOption, type Surface, SurfaceSchema, type TextMethods, type TextModelInstance, type UpscaleOptions, type UsageEvent, type UsageModality, type UsageUnit, VIDEO_INPUT_SPEC_KEYS, type VideoCallJob, type VideoCallOptions, type VideoJob, type VideoModel, type VideoRequest, type VideoResolution, type VideoResult, type VideoSecondsLimits, type Warning, adapterServesType, aiSdkTextAdapter, arkAdapter, aspectRatioToSize, classifyHttpError, compareCatalogOrder, createDefaultModelKit, createModelKit, dataOnlyCapable, describe, describeThrown, generateImage as experimental_generateImage, falAdapter, fileCatalogCache, generateAudio, generateImage, generateObject, generateText, generateVideo, googleGenaiAdapter, googleImageAdapter, hasAwsCredentials, higgsfieldAdapter, isInputSpecKey, isListed, isListedOn, isModelKitError, keysFromEnv, languageModel, liveCatalog, liveState, mergeTextProviderOptions, missingTierKeys, nearestAspectRatio, normalizeExposure, openaiImageAdapter, passthroughTextModel, pollUntil, prepareMediaPart, primeCatalog, prodiaAdapter, readErrorBody, replicateAdapter, resetLiveCatalog, resolveAudioOptions, resolveImageOptions, resolveKeys, resolveVideoOptions, shouldRetry, sleep, sortCatalogRows, streamObject, streamText, textUsageEvent, toGeneratedImage, validateInputSpec, videoLengthViolation, withRetry, xaiVideoAdapter };
+export { ADAPTER_IDS, ADAPTER_KEYS, AUDIO_INPUT_SPEC_KEYS, type Adapter, type AdapterContext, type AdapterEndpoints, type AdapterId, type AdapterRegistry, type AiTextProviderOptions, type AspectRatio, type AudioOutput, type AudioRequest, type AudioResult, type CallContext, Catalog, type CatalogAuth, type CatalogCache, type CatalogCapabilities, type CatalogDisplay, type CatalogEnvelope, type CatalogModel, CatalogModelSchema, CatalogModelSchemaBase, type CatalogOrigin, CatalogSchema, type CatalogSource, type CatalogTextModel, DEFAULT_ADAPTERS, type DataOnlyRow, type DataOnlyVerdict, type DefaultObjectOutput, type Exposure, ExposureSchema, type FalWebhookBody, type GeneratedImage, IMAGE_INPUT_SPEC_KEYS, INPUT_SPEC_SUPPORTS_KEYS, type ImageBytes, type ImageCallOptions, type ImageCallResult, type ImageQuality, type ImageRequest, type ImageResult, type ImageSafetySetting, type ImageSize, type InputSpec, type InputSpecIssue, type InputSpecKey, InputSpecSchema, type JsonObject, JsonObjectSchema, type JsonPrimitive, type JsonValue, JsonValueSchema, type KeyTier, type LanguageModelHandle, type Lifecycle, type LiveCatalogState, MODELS, type MediaPart, type MediaRef, type ModelConstraints, type ModelKit, type ModelKitConfig, ModelKitError, type ModelKitErrorCode, type ModelKitErrorInit, type ModelType, type ModelkitGenerateObjectOptions, type ModelkitGenerateTextOptions, type ModelkitStreamObjectOptions, type ModelkitStreamTextOptions, type ObjectOutput, type ObjectResult, type OnUnsupported, type OutputFormat, type PricingKind, type PrimeCatalogResult, type PromptConventions, type ProviderErrorBody, type ProviderKeyName, type ProviderKeys, type ProviderOptionValue, type ProviderOptions, type QuarantineResult, type ReferenceLimits, type ReferenceSyntax, type RequiredInput, type RequiredKeyName, type ResolveResult, type Routing, SURFACES, type SupportedOption, type Surface, SurfaceSchema, type TextMethods, type TextModelInstance, type UpscaleOptions, type UsageEvent, type UsageModality, type UsageUnit, VIDEO_INPUT_SPEC_KEYS, VIDEO_TOOL_MODELS, type VideoCallJob, type VideoCallOptions, type VideoJob, type VideoModel, type VideoRequest, type VideoResolution, type VideoResult, type VideoSecondsLimits, type VideoToolCallContext, type VideoToolConstraints, type VideoToolKind, type VideoToolOutputFormat, type VideoToolOutputSpec, type VideoToolPlan, type VideoToolRequest, type VideoToolResult, type VideoToolSource, type VideoToolStatus, type VideoToolSubmission, type Warning, adapterServesType, aiSdkTextAdapter, arkAdapter, aspectRatioToSize, classifyHttpError, compareCatalogOrder, createDefaultModelKit, createModelKit, dataOnlyCapable, describe, describeThrown, generateImage as experimental_generateImage, falAdapter, falWebhookBodySchema, fileCatalogCache, generateAudio, generateImage, generateObject, generateText, generateVideo, googleGenaiAdapter, googleImageAdapter, hasAwsCredentials, higgsfieldAdapter, isInputSpecKey, isListed, isListedOn, isModelKitError, keysFromEnv, languageModel, liveCatalog, liveState, mergeTextProviderOptions, missingTierKeys, nearestAspectRatio, normalizeExposure, openaiImageAdapter, passthroughTextModel, planVideoTool, pollUntil, prepareMediaPart, primeCatalog, prodiaAdapter, readErrorBody, replicateAdapter, resetLiveCatalog, resolutionTier, resolveAudioOptions, resolveImageOptions, resolveKeys, resolveVideoOptions, shouldRetry, sleep, sortCatalogRows, streamObject, streamText, textUsageEvent, toGeneratedImage, validateInputSpec, videoLengthViolation, videoToolConstraints, withRetry, xaiVideoAdapter };