fix(ai): route VIP98K reference edits through task API

lib/ai/providers/vip98k.ts

@@ -108,6 +108,25 @@ function inferKind(modelId: string, config: Vip98kModelRuntimeConfig) {
   return "chat";
 }
 
+const VIP98K_OPENAI_IMAGE_ALIASES = new Set([
+  "nano-banana-2",
+  "nano-banana-pro",
+  "gpt-image-2-vip",
+]);
+
+function resolveVip98kReferenceModelId(modelId: string) {
+  switch (modelId.trim().toLowerCase()) {
+    case "nano-banana-2":
+      return "google/nano-banana-2/reference-to-image";
+    case "nano-banana-pro":
+      return "google/nano-banana-pro/edit";
+    case "gpt-image-2-vip":
+      return "openai/gpt-image-2/edit";
+    default:
+      return modelId;
+  }
+}
+
 function delay(ms: number) {
   return new Promise((resolve) => setTimeout(resolve, ms));
 }
@@ -746,10 +765,6 @@ export function createVip98k({ apiKey, baseURL }: Vip98kConfig) {
       const modelConfig = parseRuntimeConfig(defaultModelConfig);
       const kind = inferKind(modelId, modelConfig);
 
-      if (kind === "image") {
-        return openAIImageProvider.chat(modelId) as any;
-      }
-
       if (kind === "chat") {
         return openAITextProvider.chat(modelId) as any;
       }
@@ -774,14 +789,25 @@ export function createVip98k({ apiKey, baseURL }: Vip98kConfig) {
           if (!activeModelId) throw new Error("VIP98K model ID is not configured");
           const activeKind = inferKind(activeModelId, runtimeConfig);
           const { prompt, imageUrls, videoUrls } = extractPromptAndMedia(options.messages);
+          if (
+            activeKind === "image" &&
+            imageUrls.length === 0 &&
+            VIP98K_OPENAI_IMAGE_ALIASES.has(activeModelId.toLowerCase())
+          ) {
+            return openAIImageProvider.chat(activeModelId).doGenerate(options as any);
+          }
           const [publicImageUrls, publicVideoUrls] = await Promise.all([
             Promise.all(imageUrls.map(ensurePublicMediaUrl)),
             Promise.all(videoUrls.map(ensurePublicMediaUrl)),
           ]);
+          const taskModelId =
+            activeKind === "image" && publicImageUrls.length > 0
+              ? resolveVip98kReferenceModelId(activeModelId)
+              : activeModelId;
           const text = await runTask({
             apiKey,
             baseURL: rootBaseURL,
-            modelId: activeModelId,
+            modelId: taskModelId,
             kind: activeKind,
             prompt,
             imageUrls: publicImageUrls,

scripts/test-vip98k-capability.ts

@@ -49,7 +49,11 @@ const DEFAULT_MODEL_IDS: Record<Capability, string[]> = {
   chat: ["gpt-5.4-mini", "gpt-5.4", "claude-sonnet-4-6"],
   multimodal: ["gpt-5.4-mini", "gpt-5.4", "claude-sonnet-4-6"],
   image: ["nano-banana-2", "nano-banana-pro", "gpt-image-2-vip"],
-  "multi-image-edit": ["nano-banana-2", "nano-banana-pro", "gpt-image-2-vip"],
+  "multi-image-edit": [
+    "google/nano-banana-2/reference-to-image",
+    "google/nano-banana-pro/edit",
+    "openai/gpt-image-2/edit",
+  ],
   rmbg: ["remove-background"],
   enhance: ["enhance", "image-upscaler"],
   extend: ["image-expansion"],
@@ -302,18 +306,27 @@ async function runExternalTask(params: {
   capability: Capability;
 }) {
   const imageUrl = "https://dummyimage.com/512x512/2563eb/ffffff.png&text=VIP98K";
+  const secondImageUrl = "https://dummyimage.com/512x512/dc2626/ffffff.png&text=REFERENCE";
   const body: Record<string, unknown> = {
     model: params.modelId,
     displayName: `provider-smoke-${params.capability}`,
     prompt: "Create a minimal test result with one simple object",
   };
-  if (["rmbg", "enhance", "extend", "image-to-3d"].includes(params.capability)) {
+  if (
+    ["multi-image-edit", "rmbg", "enhance", "extend", "image-to-3d"].includes(
+      params.capability,
+    )
+  ) {
+    const images =
+      params.capability === "multi-image-edit" ? [imageUrl, secondImageUrl] : [imageUrl];
     Object.assign(body, {
       image: imageUrl,
       imageUrl,
       image_url: imageUrl,
-      images: [imageUrl],
-      imageUrls: [imageUrl],
+      images,
+      imageUrls: images,
+      referenceImages: images,
+      reference_images: images,
     });
   }
   if (params.capability === "extend") {
@@ -424,12 +437,12 @@ async function main() {
         modelId,
         multimodal: capability === "multimodal",
       });
-    } else if (capability === "image" || capability === "multi-image-edit") {
+    } else if (capability === "image") {
       outputs = await runOpenAIImage({
         baseUrl,
         apiKey,
         modelId,
-        multiImage: capability === "multi-image-edit",
+        multiImage: false,
       });
     } else {
       outputs = await runExternalTask({ baseUrl, apiKey, modelId, capability });