@shop-os/modelkit 0.2.15 → 0.2.17

package/dist/index.d.cts

@@ -110,8 +110,14 @@ interface VideoRequest {
     watermark?: boolean;
     /** Provider-side safety checker toggle (fal MiniMax). Provider default when unset. */
     safetyChecker?: boolean;
-    /** Seedance omni-reference subtask hint. `edit` forces duration -1 and adaptive ratio. */
+    /** Seedance omni-reference subtask hint. `edit` forces duration -1 and adaptive ratio.
+     * `extend` on a `videoExtension` row continues the clip by `durationSeconds` more. */
     taskType?: "reference" | "edit" | "extend";
+    /** Continue a provider-side session instead of re-sending the clip (Gemini Omni
+     * Interactions `previous_interaction_id`). The prior call's `VideoResult.providerRef`
+     * carries it. The provider rejects it together with an explicit video task, so the
+     * adapter sends the prompt alone; `taskType: "extend"` still selects the extend path. */
+    previousInteractionId?: string;
     mode?: "standard" | "pro";
 }
 /**
@@ -163,7 +169,9 @@ interface MediaPart {
     cleanup?: () => Promise<void>;
 }
 type RequiredInput = "firstFrame" | "sourceVideo" | "references";
-type SupportedOption = "seed" | "quality" | "negativePrompt" | "cfgScale" | "generateAudio" | "promptExpansion" | "lastFrame" | "count" | "outputFormat" | "scaleFactor" | "safetySettings" | "temperature" | "upscale" | "watermark" | "safetyChecker" | "taskType" | "instrumental";
+type SupportedOption = "seed" | "quality" | "negativePrompt" | "cfgScale" | "generateAudio" | "promptExpansion" | "lastFrame" | "count" | "outputFormat" | "scaleFactor" | "safetySettings" | "temperature" | "upscale" | "watermark" | "safetyChecker" | "taskType" | "instrumental"
+/** Video rows that extend a clip in fixed segments up to a cumulative cap (see `ModelConstraints.extension`). */
+ | "videoExtension";
 interface VideoSecondsLimits {
     /** Minimum length of each reference video, in seconds. */
     min?: number;
@@ -213,6 +221,12 @@ interface ModelConstraints {
         /** The provider sends the source as a reference video, so it counts toward `references.videoSeconds.total` (Seedance). */
         inReferenceTotal?: boolean;
     };
+    /** Video extension (rows with `supports: ["videoExtension"]`): each `taskType: "extend"`
+     * call adds up to `segmentSeconds` and the clip can grow to `maxTotalSeconds`. */
+    extension?: {
+        segmentSeconds: number;
+        maxTotalSeconds: number;
+    };
 }
 /** JSON-compatible value. Provider options and raw payloads carry these, never `unknown`. */
 type JsonPrimitive = string | number | boolean | null;
@@ -354,6 +368,10 @@ interface VideoResult {
     };
     usage: UsageEvent;
     raw?: unknown;
+    /** Provider handle for chaining a follow-up call (`VideoRequest.previousInteractionId`). */
+    providerRef?: {
+        interactionId: string;
+    };
 }
 interface VideoCallJob extends VideoJob {
     warnings: Warning[];
@@ -640,6 +658,7 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
             safetyChecker: "safetyChecker";
             taskType: "taskType";
             instrumental: "instrumental";
+            videoExtension: "videoExtension";
         }>>>;
         ratioIgnoredWith: z.ZodOptional<z.ZodArray<z.ZodEnum<{
             firstFrame: "firstFrame";
@@ -651,6 +670,10 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
             editMin: z.ZodOptional<z.ZodNumber>;
             inReferenceTotal: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$loose>>;
+        extension: z.ZodOptional<z.ZodObject<{
+            segmentSeconds: z.ZodNumber;
+            maxTotalSeconds: z.ZodNumber;
+        }, z.core.$loose>>;
     }, z.core.$loose>;
     defaults: z.ZodObject<{
         references: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodObject<{

package/dist/index.d.ts

@@ -110,8 +110,14 @@ interface VideoRequest {
     watermark?: boolean;
     /** Provider-side safety checker toggle (fal MiniMax). Provider default when unset. */
     safetyChecker?: boolean;
-    /** Seedance omni-reference subtask hint. `edit` forces duration -1 and adaptive ratio. */
+    /** Seedance omni-reference subtask hint. `edit` forces duration -1 and adaptive ratio.
+     * `extend` on a `videoExtension` row continues the clip by `durationSeconds` more. */
     taskType?: "reference" | "edit" | "extend";
+    /** Continue a provider-side session instead of re-sending the clip (Gemini Omni
+     * Interactions `previous_interaction_id`). The prior call's `VideoResult.providerRef`
+     * carries it. The provider rejects it together with an explicit video task, so the
+     * adapter sends the prompt alone; `taskType: "extend"` still selects the extend path. */
+    previousInteractionId?: string;
     mode?: "standard" | "pro";
 }
 /**
@@ -163,7 +169,9 @@ interface MediaPart {
     cleanup?: () => Promise<void>;
 }
 type RequiredInput = "firstFrame" | "sourceVideo" | "references";
-type SupportedOption = "seed" | "quality" | "negativePrompt" | "cfgScale" | "generateAudio" | "promptExpansion" | "lastFrame" | "count" | "outputFormat" | "scaleFactor" | "safetySettings" | "temperature" | "upscale" | "watermark" | "safetyChecker" | "taskType" | "instrumental";
+type SupportedOption = "seed" | "quality" | "negativePrompt" | "cfgScale" | "generateAudio" | "promptExpansion" | "lastFrame" | "count" | "outputFormat" | "scaleFactor" | "safetySettings" | "temperature" | "upscale" | "watermark" | "safetyChecker" | "taskType" | "instrumental"
+/** Video rows that extend a clip in fixed segments up to a cumulative cap (see `ModelConstraints.extension`). */
+ | "videoExtension";
 interface VideoSecondsLimits {
     /** Minimum length of each reference video, in seconds. */
     min?: number;
@@ -213,6 +221,12 @@ interface ModelConstraints {
         /** The provider sends the source as a reference video, so it counts toward `references.videoSeconds.total` (Seedance). */
         inReferenceTotal?: boolean;
     };
+    /** Video extension (rows with `supports: ["videoExtension"]`): each `taskType: "extend"`
+     * call adds up to `segmentSeconds` and the clip can grow to `maxTotalSeconds`. */
+    extension?: {
+        segmentSeconds: number;
+        maxTotalSeconds: number;
+    };
 }
 /** JSON-compatible value. Provider options and raw payloads carry these, never `unknown`. */
 type JsonPrimitive = string | number | boolean | null;
@@ -354,6 +368,10 @@ interface VideoResult {
     };
     usage: UsageEvent;
     raw?: unknown;
+    /** Provider handle for chaining a follow-up call (`VideoRequest.previousInteractionId`). */
+    providerRef?: {
+        interactionId: string;
+    };
 }
 interface VideoCallJob extends VideoJob {
     warnings: Warning[];
@@ -640,6 +658,7 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
             safetyChecker: "safetyChecker";
             taskType: "taskType";
             instrumental: "instrumental";
+            videoExtension: "videoExtension";
         }>>>;
         ratioIgnoredWith: z.ZodOptional<z.ZodArray<z.ZodEnum<{
             firstFrame: "firstFrame";
@@ -651,6 +670,10 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
             editMin: z.ZodOptional<z.ZodNumber>;
             inReferenceTotal: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$loose>>;
+        extension: z.ZodOptional<z.ZodObject<{
+            segmentSeconds: z.ZodNumber;
+            maxTotalSeconds: z.ZodNumber;
+        }, z.core.$loose>>;
     }, z.core.$loose>;
     defaults: z.ZodObject<{
         references: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodObject<{

package/dist/index.js

@@ -1402,7 +1402,14 @@ var VIDEO_MODELS = [
       durations: [3, 4, 5, 6, 7, 8, 9, 10],
       references: { images: 6, videos: 1, videoSeconds: { max: 10 } },
       sourceVideoSeconds: { max: 10 },
-      supports: []
+      // Extension (probe 2026-09-24, 360p): each `taskType: "extend"` call
+      // renders `durationSeconds` more and returns the FULL cumulative clip
+      // (3 s -> 6 s -> 9 s); the provider caps the clip at 40 s. A first
+      // extension can upload a <= 10 s source with task "extend"; later ones
+      // chain by `previousInteractionId` (the API rejects the id together
+      // with a video task, so those calls carry the prompt alone).
+      supports: ["videoExtension"],
+      extension: { segmentSeconds: 10, maxTotalSeconds: 40 }
     },
     defaults: {},
     usage: { ...VIDEO_USAGE },
@@ -1410,7 +1417,7 @@ var VIDEO_MODELS = [
     display: {
       provider: "google",
       logo: GOOGLE_LOGO2,
-      description: "Google Gemini Omni Flash 1.1 conversational video generation and editing; text, image, and up to 10-second video input, 3-10 second output at up to 4K and 24 FPS."
+      description: "Google Gemini Omni Flash 1.1 conversational video generation, editing and extension; text, image, and up to 10-second video input, 3-10 second output at up to 4K and 24 FPS, extendable to 40 seconds."
     },
     capabilities: { acceptsSourceVideo: true }
   },
@@ -2144,6 +2151,7 @@ var SupportedOptionSchema = z2.enum([
   "watermark",
   "safetyChecker",
   "taskType",
+  "videoExtension",
   "instrumental"
 ]);
 var UsageUnitSchema = z2.enum([
@@ -2191,7 +2199,8 @@ var ModelConstraintsSchema = z2.looseObject({
   requires: z2.array(RequiredInputSchema).optional(),
   supports: z2.array(SupportedOptionSchema).optional(),
   ratioIgnoredWith: z2.array(z2.enum(["firstFrame", "sourceVideo"])).optional(),
-  sourceVideoSeconds: VideoSecondsLimitsSchema.omit({ total: true }).extend({ inReferenceTotal: z2.boolean().optional() }).optional()
+  sourceVideoSeconds: VideoSecondsLimitsSchema.omit({ total: true }).extend({ inReferenceTotal: z2.boolean().optional() }).optional(),
+  extension: z2.looseObject({ segmentSeconds: z2.number(), maxTotalSeconds: z2.number() }).optional()
 });
 var MediaRefSchema = z2.union([
   z2.looseObject({
@@ -4460,12 +4469,34 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
   const modelId = ctx.model.id;
   const providerModelId = ctx.model.providerModelId;
   const duration = req.durationSeconds ?? 10;
-  const images = req.referenceImages ?? [];
-  const videos = req.referenceVideos ?? [];
-  const task = req.sourceVideo ? "edit" : req.firstFrame ? "image_to_video" : images.length > 0 || videos.length > 0 ? "reference_to_video" : "text_to_video";
+  const extend = req.taskType === "extend";
+  const previousId = extend ? req.previousInteractionId : void 0;
+  const chained = previousId !== void 0;
+  for (const field of ["referenceImages", "referenceVideos", "firstFrame"]) {
+    const value = req[field];
+    if (!chained || value === void 0 || Array.isArray(value) && value.length === 0)
+      continue;
+    ctx.warnings.push({
+      code: "option_ignored",
+      field,
+      from: Array.isArray(value) ? value.length : "set",
+      message: `${modelId} chained extension sends the prompt alone; ${field} ignored`
+    });
+  }
+  const images = chained ? [] : req.referenceImages ?? [];
+  const videos = chained ? [] : req.referenceVideos ?? [];
+  const firstFrame = chained ? void 0 : req.firstFrame;
+  if (extend && !chained && req.sourceVideo === void 0) {
+    throw new ModelKitError({
+      code: "missing_input",
+      message: `${modelId} extend needs a source video or a previousInteractionId`,
+      model: modelId
+    });
+  }
+  const task = extend ? "extend" : req.sourceVideo ? "edit" : firstFrame ? "image_to_video" : images.length > 0 || videos.length > 0 ? "reference_to_video" : "text_to_video";
   const isOmniV1 = providerModelId === "gemini-omni-flash-preview";
   const responseFormat = { type: "video", delivery: "uri" };
-  if (req.sourceVideo === void 0 && req.aspectRatio !== void 0 && req.aspectRatio !== "adaptive") {
+  if (req.sourceVideo === void 0 && !extend && req.aspectRatio !== void 0 && req.aspectRatio !== "adaptive") {
     responseFormat.aspect_ratio = toOmniAspect(req.aspectRatio);
   }
   if (!isOmniV1) {
@@ -4497,18 +4528,26 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
   const cleanupUploads = async () => {
     for (const fileName of uploadedNames) await deleteGoogleFile(fileName, apiKey, ctx.fetch);
   };
-  const callCreate = (input) => client2.interactions.create({
-    model: providerModelId,
-    input,
-    response_format: responseFormat,
-    generation_config: { video_config: { task } },
-    store: true
-  });
+  const callCreate = (input) => client2.interactions.create(
+    previousId !== void 0 ? {
+      model: providerModelId,
+      input,
+      response_format: responseFormat,
+      store: true,
+      previous_interaction_id: previousId
+    } : {
+      model: providerModelId,
+      input,
+      response_format: responseFormat,
+      generation_config: { video_config: { task } },
+      store: true
+    }
+  );
   const createRetryOptions = mediaRetryOptions(ctx.call, ctx.signal);
   let interaction;
   try {
     const parts = [];
-    if (req.sourceVideo) {
+    if (req.sourceVideo && !chained) {
       const media = await resolveMedia(
         req.sourceVideo,
         ctx.fetch,
@@ -4531,11 +4570,11 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
         mime_type: media.mime
       });
     }
-    if (req.firstFrame) {
-      if (req.firstFrame.label !== void 0)
-        parts.push({ type: "text", text: req.firstFrame.label });
+    if (firstFrame) {
+      if (firstFrame.label !== void 0)
+        parts.push({ type: "text", text: firstFrame.label });
       const media = await resolveMedia(
-        req.firstFrame,
+        firstFrame,
         ctx.fetch,
         ctx.signal,
         "image/png",
@@ -4594,7 +4633,7 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
       }
     }
   } finally {
-    await cleanupUploads();
+    if (!(extend && !chained)) await cleanupUploads();
   }
   const jobId = interaction.id ?? crypto.randomUUID();
   const usage2 = videoUsage2(ctx, req, duration);
@@ -4638,6 +4677,7 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
         return void 0;
       }, pollOptions);
       const video = extractOmniVideo(done, modelId);
+      const providerRef = { interactionId: done.id ?? jobId };
       if (video.data) {
         return {
           video: {
@@ -4645,14 +4685,16 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
             mime: video.mime_type ?? "video/mp4"
           },
           usage: usage2,
-          raw: done
+          raw: done,
+          providerRef
         };
       }
       if (video.uri)
         return {
           video: await downloadVideo(ctx, video.uri, apiKey, signal),
           usage: usage2,
-          raw: done
+          raw: done,
+          providerRef
         };
       throw classifyHttpError(
         void 0,
@@ -7544,6 +7586,7 @@ function resolveVideoOptions(model, request, onUnsupported = "error") {
   setOrDelete(merged, "referenceImages", images);
   setOrDelete(merged, "referenceVideos", videos);
   setOrDelete(merged, "referenceAudios", audios);
+  const keepExtend = merged.taskType === "extend" && (model.constraints.supports ?? []).includes("videoExtension");
   dropUnsupported(merged, model, warnings, [
     "seed",
     "negativePrompt",
@@ -7553,8 +7596,17 @@ function resolveVideoOptions(model, request, onUnsupported = "error") {
     "lastFrame",
     "watermark",
     "safetyChecker",
-    "taskType"
+    ...keepExtend ? [] : ["taskType"]
   ]);
+  if (merged.previousInteractionId !== void 0 && !keepExtend) {
+    warnings.push({
+      code: "option_ignored",
+      field: "previousInteractionId",
+      from: merged.previousInteractionId,
+      message: `${model.id} does not support previousInteractionId; ignored`
+    });
+    delete merged.previousInteractionId;
+  }
   return { request: merged, warnings };
 }
 function resolveAudioOptions(model, request, onUnsupported = "error") {