@lobehub/lobehub 2.1.3 → 2.1.4

package/packages/model-runtime/src/core/openaiCompatibleFactory/index.ts

@@ -73,6 +73,7 @@ export interface OpenAICompatibleFactoryOptions<T extends Record<string, any> =
   baseURL?: string;
   chatCompletion?: {
     excludeUsage?: boolean;
+    forceImageBase64?: boolean;
     handleError?: (
       error: any,
       options: ConstructorOptions<T>,
@@ -387,7 +388,9 @@ export const createOpenAICompatibleRuntime = <T extends Record<string, any> = an
           this.baseURL = targetBaseURL;
         }
 
-        const messages = await convertOpenAIMessages(postPayload.messages);
+        const messages = await convertOpenAIMessages(postPayload.messages, {
+          forceImageBase64: chatCompletion?.forceImageBase64,
+        });
 
         let response: Stream<OpenAI.Chat.Completions.ChatCompletionChunk>;
 
@@ -877,7 +880,9 @@ export const createOpenAICompatibleRuntime = <T extends Record<string, any> = an
       delete res.frequency_penalty;
       delete res.presence_penalty;
 
-      const input = await convertOpenAIResponseInputs(messages as any);
+      const input = await convertOpenAIResponseInputs(messages as any, {
+        forceImageBase64: chatCompletion?.forceImageBase64,
+      });
 
       const isStreaming = payload.stream !== false;
       log(
@@ -1006,7 +1011,9 @@ export const createOpenAICompatibleRuntime = <T extends Record<string, any> = an
 
       if (shouldUseResponses) {
         log('calling responses.create for tool calling');
-        const input = await convertOpenAIResponseInputs(messages as any);
+        const input = await convertOpenAIResponseInputs(messages as any, {
+          forceImageBase64: chatCompletion?.forceImageBase64,
+        });
 
         const res = await this.client.responses.create(
           {

package/packages/model-runtime/src/core/streams/google/google-ai.test.ts

@@ -251,16 +251,16 @@ describe('GoogleGenerativeAIStream', () => {
       expect(chunks).toEqual(
         [
           'id: chat_1',
-          'event: content_part',
-          'data: {"content":"234","partType":"text"}\n',
+          'event: text',
+          'data: "234"\n',
 
           'id: chat_1',
           'event: text',
           'data: ""\n',
 
           'id: chat_1',
-          'event: content_part',
-          `data: {"content":"567890\\n","partType":"text"}\n`,
+          'event: text',
+          `data: "567890\\n"\n`,
           // stop
           'id: chat_1',
           'event: stop',
@@ -384,12 +384,12 @@ describe('GoogleGenerativeAIStream', () => {
           `data: {"content":"**Finalizing Interpretation**\\n\\n","inReasoning":true,"partType":"text"}\n`,
 
           'id: chat_1',
-          'event: content_part',
-          `data: {"content":"简单来说，","partType":"text"}\n`,
+          'event: text',
+          'data: "简单来说，"\n',
 
           'id: chat_1',
-          'event: content_part',
-          `data: {"content":"文本内容。","partType":"text"}\n`,
+          'event: text',
+          'data: "文本内容。"\n',
           // stop
           'id: chat_1',
           'event: stop',
@@ -471,12 +471,12 @@ describe('GoogleGenerativeAIStream', () => {
       expect(chunks).toEqual(
         [
           'id: chat_1',
-          'event: content_part',
-          'data: {"content":"234","partType":"text"}\n',
+          'event: text',
+          'data: "234"\n',
 
           'id: chat_1',
-          'event: content_part',
-          `data: {"content":"567890\\n","partType":"text"}\n`,
+          'event: text',
+          'data: "567890\\n"\n',
           // stop
           'id: chat_1',
           'event: stop',
@@ -1103,7 +1103,7 @@ describe('GoogleGenerativeAIStream', () => {
               content: {
                 parts: [
                   {
-                    text: '**Planning the Solution**\n\nI\'m solidifying my plan...',
+                    text: "**Planning the Solution**\n\nI'm solidifying my plan...",
                     thought: true,
                   },
                 ],
@@ -1901,5 +1901,46 @@ describe('GoogleGenerativeAIStream', () => {
         ].map((i) => i + '\n'),
       );
     });
+
+    it('should NOT use multimodal processing if only thoughtsTokenCount is present in metadata but no thought parts', async () => {
+      vi.spyOn(uuidModule, 'nanoid').mockReturnValueOnce('1');
+
+      const data = [
+        {
+          candidates: [
+            {
+              content: {
+                parts: [{ text: 'Hello world' }],
+                role: 'model',
+              },
+              index: 0,
+            },
+          ],
+          usageMetadata: {
+            promptTokenCount: 10,
+            candidatesTokenCount: 2,
+            totalTokenCount: 17,
+            thoughtsTokenCount: 5,
+          },
+          modelVersion: 'gemini-2.5-flash',
+        },
+      ];
+
+      const mockGoogleStream = new ReadableStream({
+        start(controller) {
+          data.forEach((item) => {
+            controller.enqueue(item);
+          });
+          controller.close();
+        },
+      });
+
+      const protocolStream = GoogleGenerativeAIStream(mockGoogleStream);
+      const chunks = await decodeStreamChunks(protocolStream);
+
+      // Should use 'text' event, not 'content_part'
+      expect(chunks).toContain('event: text\n');
+      expect(chunks).not.toContain('event: content_part\n');
+    });
   });
 });

package/packages/model-runtime/src/core/streams/google/index.ts

@@ -120,7 +120,6 @@ const transformGoogleGenerativeAIStream = (
     const hasReasoningParts = parts.some((p: any) => p.thought === true);
     const hasImageParts = parts.some((p: any) => p.inlineData);
     const hasThoughtSignature = parts.some((p: any) => p.thoughtSignature);
-    const hasThoughtsInMetadata = (usageMetadata as any)?.thoughtsTokenCount > 0;
 
     // Check model version to determine if new format should be used
     const modelVersion = (chunk as any).modelVersion || '';
@@ -144,8 +143,7 @@ const transformGoogleGenerativeAIStream = (
     // 1. There are reasoning parts in current chunk (thought: true)
     // 2. There are multiple parts with images (multimodal content)
     // 3. There are thoughtSignature in parts (reasoning metadata attached to content)
-    // 4. There is thoughtsTokenCount in metadata (indicates response contains reasoning)
-    // 5. This is Gemini 3 model with image generation (always use new format for consistency)
+    // 4. This is Gemini 3 model with image generation (always use new format for consistency)
     // BUT NOT for:
     // - The legacy single-image scenario
     // - Grounding metadata scenario (uses legacy text + grounding events)
@@ -153,7 +151,6 @@ const transformGoogleGenerativeAIStream = (
       (hasReasoningParts ||
         (hasImageParts && parts.length > 1) ||
         hasThoughtSignature ||
-        hasThoughtsInMetadata ||
         isGemini3Model) &&
       !isSingleImageWithFinish &&
       !hasGroundingMetadata;

package/packages/model-runtime/src/providers/moonshot/index.ts

@@ -15,8 +15,9 @@ export interface MoonshotModelCard {
 export const params = {
   baseURL: 'https://api.moonshot.cn/v1',
   chatCompletion: {
+    forceImageBase64: true,
     handlePayload: (payload: ChatStreamPayload) => {
-      const { enabledSearch, messages, temperature, tools, ...rest } = payload;
+      const { enabledSearch, messages, model, temperature, thinking, tools, ...rest } = payload;
 
       const filteredMessages = messages.map((message: any) => {
         let normalizedMessage = message;
@@ -51,12 +52,33 @@ export const params = {
           ]
         : tools;
 
-      // Resolve parameters with normalization
+      const isK25Model = model === 'kimi-k2.5';
+
+      if (isK25Model) {
+        const thinkingParam =
+          thinking?.type === 'disabled' ? { type: 'disabled' } : { type: 'enabled' };
+        const isThinkingEnabled = thinkingParam.type === 'enabled';
+
+        return {
+          ...rest,
+          frequency_penalty: 0,
+          messages: filteredMessages,
+          model,
+          presence_penalty: 0,
+          temperature: isThinkingEnabled ? 1 : 0.6,
+          thinking: thinkingParam,
+          tools: moonshotTools,
+          top_p: 0.95,
+        } as any;
+      }
+
+      // Resolve parameters with normalization for non-K2.5 models
       const resolvedParams = resolveParameters({ temperature }, { normalizeTemperature: true });
 
       return {
         ...rest,
         messages: filteredMessages,
+        model,
         temperature: resolvedParams.temperature,
         tools: moonshotTools,
       } as any;

package/packages/model-runtime/src/providers/qwen/index.ts

@@ -47,24 +47,25 @@ export const LobeQwenAI = createOpenAICompatibleRuntime({
         ...rest,
         ...(model.includes('-thinking')
           ? {
-              enable_thinking: true,
+            enable_thinking: true,
+            thinking_budget:
+              thinking?.budget_tokens === 0 ? 0 : thinking?.budget_tokens || undefined,
+          }
+          : [
+            'qwen3',
+            'qwen-turbo',
+            'qwen-plus',
+            'qwen-flash',
+            'deepseek-v3.1',
+            'deepseek-v3.2',
+            'glm',
+            'kimi-k2.5',
+          ].some((keyword) => model.toLowerCase().includes(keyword))
+            ? {
+              enable_thinking: thinking !== undefined ? thinking.type === 'enabled' : false,
               thinking_budget:
                 thinking?.budget_tokens === 0 ? 0 : thinking?.budget_tokens || undefined,
             }
-          : [
-                'qwen3',
-                'qwen-turbo',
-                'qwen-plus',
-                'qwen-flash',
-                'deepseek-v3.1',
-                'deepseek-v3.2',
-                'glm',
-              ].some((keyword) => model.toLowerCase().includes(keyword))
-            ? {
-                enable_thinking: thinking !== undefined ? thinking.type === 'enabled' : false,
-                thinking_budget:
-                  thinking?.budget_tokens === 0 ? 0 : thinking?.budget_tokens || undefined,
-              }
             : {}),
         frequency_penalty: undefined,
         model,

package/src/server/routers/lambda/__tests__/integration/aiAgent/execAgent.integration.test.ts

@@ -232,14 +232,15 @@ describe('execAgent', () => {
         })
         .returning();
 
-      const [thread] = await serverDB
+      const threadResult = (await serverDB
         .insert(threads)
         .values({
           topicId: topic.id,
           type: 'standalone',
           userId,
         })
-        .returning();
+        .returning()) as { id: string }[];
+      const thread = threadResult[0];
 
       const caller = aiAgentRouter.createCaller(createTestContext());