fix(agent): use remote image urls for doubao multimodal vision

lib/ai/auto-agent-graph.ts

@@ -147,8 +147,19 @@ async function toAgentVisionDataUrl(imageUrl: string): Promise<string | null> {
   }
 }
 
-async function buildAgentVisionInputs(imageUrls: string[]) {
+async function buildAgentVisionInputs(
+  imageUrls: string[],
+  options?: {
+    preferRemoteUrls?: boolean;
+  },
+) {
   if (imageUrls.length === 0) return [] as Array<{ type: "image"; image: string }>;
+  if (options?.preferRemoteUrls) {
+    return imageUrls.map((url) => ({
+      type: "image" as const,
+      image: url,
+    }));
+  }
 
   const convertedResults = await Promise.all(imageUrls.map((url) => toAgentVisionDataUrl(url)));
   const inlineCount = convertedResults.filter(Boolean).length;
@@ -492,7 +503,7 @@ function isLikelyVisionCapableModel(item: EnabledModelWithProvider) {
   if (item.model.supportsVision) return true;
   if (isAiModelMultimodal(item.model)) return true;
 
-  const providerType = item.provider.type || "";
+  const providerType = String(item.provider.type || "");
   const modelText = `${item.model.displayName || ""} ${item.model.modelId || ""}`.toLowerCase();
 
   if (providerType === "volcano-multimodal") return true;
@@ -523,6 +534,24 @@ function isLikelyVisionCapableModel(item: EnabledModelWithProvider) {
   return false;
 }
 
+function shouldPreferRemoteVisionUrls(item?: EnabledModelWithProvider) {
+  if (!item) return false;
+  const providerType = String(item.provider.type || "");
+  const baseUrl = String((item.provider as { baseUrl?: string | null })?.baseUrl || "").toLowerCase();
+  const modelText = `${item.model.displayName || ""} ${item.model.modelId || ""}`.toLowerCase();
+
+  if (providerType === "volcano-multimodal" || providerType === "doubao-video" || providerType === "volcano-video") {
+    return true;
+  }
+  if (baseUrl.includes("ark.cn-beijing.volces.com")) {
+    return true;
+  }
+  if (/(doubao[-_.]?seed[-_.]?2[-_.]?0|seed[-_.]?2[-_.]?0)/i.test(modelText)) {
+    return true;
+  }
+  return false;
+}
+
 function pickQwenVisionModel(models: EnabledModelWithProvider[]) {
   const visionModels = models.filter((item) => isLikelyVisionCapableModel(item));
   const zhipuVision = visionModels.find(
@@ -651,7 +680,10 @@ async function generateAgentConversation(
 2) 只写客观可见信息，不要编造品牌、尺寸、镜头参数。
 3) 全部输出控制在 6 行内。`;
 
-        const visionInputs = await buildAgentVisionInputs(validImageUrls);
+        const preferRemoteVisionUrls = shouldPreferRemoteVisionUrls(visionModelConfig);
+        const visionInputs = await buildAgentVisionInputs(validImageUrls, {
+          preferRemoteUrls: preferRemoteVisionUrls,
+        });
         const { text: visionText } = await generateText({
           model: visionModel,
           messages: [
@@ -666,7 +698,7 @@ async function generateAgentConversation(
 
         visualAnalysis = visionText.trim();
         console.log(
-          `[Auto Agent] Vision analysis model: ${visionModelConfig.model.displayName} (${visionModelConfig.provider.type})`
+          `[Auto Agent] Vision analysis model: ${visionModelConfig.model.displayName} (${visionModelConfig.provider.type}), preferRemote=${preferRemoteVisionUrls}`
         );
       } catch (error) {
         console.warn("[Auto Agent] Vision analysis failed, continue without visual summary");

lib/ai/smart-analyzer.ts

@@ -324,19 +324,68 @@ async function buildAnalyzerVisionImages(imageUrls?: string[]) {
   });
 }
 
+function shouldPreferRemoteVisionUrls(modelWithProvider?: {
+  model?: {
+    displayName?: string | null;
+    modelId?: string | null;
+  } | null;
+  provider?: {
+    type?: string | null;
+    baseUrl?: string | null;
+  } | null;
+}): boolean {
+  if (!modelWithProvider) return false;
+  const providerType = String(modelWithProvider.provider?.type || "");
+  const baseUrl = String(modelWithProvider.provider?.baseUrl || "").toLowerCase();
+  const modelText = `${modelWithProvider.model?.displayName || ""} ${modelWithProvider.model?.modelId || ""}`.toLowerCase();
+
+  if (providerType === "volcano-multimodal" || providerType === "doubao-video" || providerType === "volcano-video") {
+    return true;
+  }
+  if (baseUrl.includes("ark.cn-beijing.volces.com")) {
+    return true;
+  }
+  if (/(doubao[-_.]?seed[-_.]?2[-_.]?0|seed[-_.]?2[-_.]?0)/i.test(modelText)) {
+    return true;
+  }
+  return false;
+}
+
+async function buildAnalyzerVisionImagesByMode(params: {
+  imageUrls?: string[];
+  preferRemoteUrls: boolean;
+}) {
+  const urls = (params.imageUrls || [])
+    .filter((url) => isHttpUrl(url))
+    .slice(0, ANALYZER_MAX_IMAGE_COUNT);
+  if (urls.length === 0) return [] as Array<{ type: "image"; image: string }>;
+  if (params.preferRemoteUrls) {
+    return urls.map((url) => ({ type: "image" as const, image: url }));
+  }
+  return buildAnalyzerVisionImages(urls);
+}
+
 function parseAnalysisResultFromModelText(
   responseText: string,
   userText: string,
   hasImage: boolean,
 ): AnalysisResult | null {
   const jsonMatch = responseText.match(/\{[\s\S]*\}/);
-  if (!jsonMatch) return null;
+  if (!jsonMatch) {
+    console.warn("[Smart Analyzer] Model response missing JSON:", responseText.slice(0, 240));
+    return null;
+  }
 
   try {
     const result = JSON.parse(jsonMatch[0]) as AnalysisResult;
     return normalizeIntentForImageContext(result, userText, hasImage);
   } catch (error) {
-    console.error("[Smart Analyzer] Failed to parse model response:", error);
+    console.error(
+      "[Smart Analyzer] Failed to parse model response:",
+      error,
+      "raw=",
+      jsonMatch[0].slice(0, 240),
+    );
     return null;
   }
 }
@@ -780,11 +829,17 @@ ${historyText}
     ? `${contextMessage}用户上传了图片，请求内容："${text}"`
     : `${contextMessage}用户请求（无图片）："${text}"`;
 
-  const visionImages = hasImage ? await buildAnalyzerVisionImages(imageUrls) : [];
+  const preferRemoteVisionUrls = shouldPreferRemoteVisionUrls(selectedAnalyzerModel);
+  const visionImages = hasImage
+    ? await buildAnalyzerVisionImagesByMode({
+        imageUrls,
+        preferRemoteUrls: preferRemoteVisionUrls,
+      })
+    : [];
   if (visionImages.length > 0) {
     const inlineCount = visionImages.filter((item) => item.image.startsWith("data:image/")).length;
     console.log(
-      `[Smart Analyzer] Prepared analyzer images: total=${visionImages.length}, inline=${inlineCount}, remote=${visionImages.length - inlineCount}`,
+      `[Smart Analyzer] Prepared analyzer images: total=${visionImages.length}, inline=${inlineCount}, remote=${visionImages.length - inlineCount}, preferRemote=${preferRemoteVisionUrls}`,
     );
   }