@lobehub/chat 1.64.3 → 1.65.1

package/src/libs/agent-runtime/utils/streams/openai.test.ts

@@ -1576,5 +1576,186 @@ describe('OpenAIStream', () => {
         ].map((i) => `${i}\n`),
       );
     });
+
+    it('should handle claude reasoning in litellm openai mode', async () => {
+      const data = [
+        {
+          id: '1',
+          created: 1740505568,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [
+            {
+              index: 0,
+              delta: {
+                thinking_blocks: [
+                  { type: 'thinking', thinking: '我需要找94的所有质', ignature_delta: null },
+                ],
+                reasoning_content: '我需要找出394的所有质',
+                content: '',
+                role: 'assistant',
+              },
+            },
+          ],
+          thinking_blocks: [
+            { type: 'thinking', thinking: '我需要找94的所有质', signature_delta: null },
+          ],
+        },
+        {
+          id: '1',
+          created: 1740505569,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [
+            {
+              index: 0,
+              delta: {
+                thinking_blocks: [
+                  { type: 'thinking', thinking: '因数。\n质因数是', signature_delta: null },
+                ],
+                reasoning_content: '因数。\n\n质因数是',
+                content: '',
+              },
+            },
+          ],
+          thinking_blocks: [
+            { type: 'thinking', thinking: '因数。\n\n质因数是', signature_delta: null },
+          ],
+        },
+        {
+          id: '1',
+          created: 1740505569,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [
+            {
+              index: 0,
+              delta: {
+                thinking_blocks: [
+                  { type: 'thinking', thinking: '÷ 2 = 197', signature_delta: null },
+                ],
+                reasoning_content: '÷ 2 = 197',
+                content: '',
+              },
+            },
+          ],
+          thinking_blocks: [{ type: 'thinking', thinking: '÷ 2 = 197', signature_delta: null }],
+        },
+        {
+          id: '1',
+          created: 1740505571,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [
+            {
+              index: 0,
+              delta: {
+                thinking_blocks: [
+                  { type: 'thinking', thinking: '197。n394 = 2 ', signature_delta: null },
+                ],
+                reasoning_content: '197。\n394 = 2 ',
+                content: '',
+              },
+            },
+          ],
+          thinking_blocks: [{ type: 'thinking', thinking: '\n394 = 2 ', signature_delta: null }],
+        },
+        {
+          id: '1',
+          created: 1740505571,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [
+            {
+              index: 0,
+              delta: {
+                content: '',
+                tool_calls: [{ function: { arguments: '{}' }, type: 'function', index: -1 }],
+              },
+            },
+          ],
+        },
+        {
+          id: '1',
+          created: 1740505571,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [{ index: 0, delta: { content: '要找出394的质因数，我需要将' } }],
+        },
+        {
+          id: '1',
+          created: 1740505571,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [{ index: 0, delta: { content: '394分解为质数的乘积' } }],
+        },
+        {
+          id: '1',
+          created: 1740505573,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [{ index: 0, delta: { content: '2和197。' } }],
+        },
+        {
+          id: '1',
+          created: 1740505573,
+          model: 'claude-3-7-sonnet-latest',
+          object: 'chat.completion.chunk',
+          choices: [{ finish_reason: 'stop', index: 0, delta: {} }],
+        },
+      ];
+
+      const mockOpenAIStream = new ReadableStream({
+        start(controller) {
+          data.forEach((chunk) => {
+            controller.enqueue(chunk);
+          });
+
+          controller.close();
+        },
+      });
+
+      const protocolStream = OpenAIStream(mockOpenAIStream);
+
+      const decoder = new TextDecoder();
+      const chunks = [];
+
+      // @ts-ignore
+      for await (const chunk of protocolStream) {
+        chunks.push(decoder.decode(chunk, { stream: true }));
+      }
+
+      expect(chunks).toEqual(
+        [
+          'id: 1',
+          'event: reasoning',
+          `data: "我需要找出394的所有质"\n`,
+          'id: 1',
+          'event: reasoning',
+          `data: "因数。\\n\\n质因数是"\n`,
+          'id: 1',
+          'event: reasoning',
+          `data: "÷ 2 = 197"\n`,
+          'id: 1',
+          'event: reasoning',
+          `data: "197。\\n394 = 2 "\n`,
+          'id: 1',
+          'event: text',
+          `data: ""\n`,
+          'id: 1',
+          'event: text',
+          `data: "要找出394的质因数，我需要将"\n`,
+          'id: 1',
+          'event: text',
+          `data: "394分解为质数的乘积"\n`,
+          'id: 1',
+          'event: text',
+          `data: "2和197。"\n`,
+          'id: 1',
+          'event: stop',
+          `data: "stop"\n`,
+        ].map((i) => `${i}\n`),
+      );
+    });
   });
 });

package/src/libs/agent-runtime/utils/streams/openai.ts

@@ -46,36 +46,46 @@ export const transformOpenAIStream = (
 
     // tools calling
     if (typeof item.delta?.tool_calls === 'object' && item.delta.tool_calls?.length > 0) {
-      return {
-        data: item.delta.tool_calls.map((value, index): StreamToolCallChunkData => {
-          if (streamContext && !streamContext.tool) {
-            streamContext.tool = { id: value.id!, index: value.index, name: value.function!.name! };
-          }
-
-          return {
-            function: {
-              arguments: value.function?.arguments ?? '{}',
-              name: value.function?.name ?? null,
-            },
-            id:
-              value.id ||
-              streamContext?.tool?.id ||
-              generateToolCallId(index, value.function?.name),
-
-            // mistral's tool calling don't have index and function field, it's data like:
-            // [{"id":"xbhnmTtY7","function":{"name":"lobe-image-designer____text2image____builtin","arguments":"{\"prompts\": [\"A photo of a small, fluffy dog with a playful expression and wagging tail.\", \"A watercolor painting of a small, energetic dog with a glossy coat and bright eyes.\", \"A vector illustration of a small, adorable dog with a short snout and perky ears.\", \"A drawing of a small, scruffy dog with a mischievous grin and a wagging tail.\"], \"quality\": \"standard\", \"seeds\": [123456, 654321, 111222, 333444], \"size\": \"1024x1024\", \"style\": \"vivid\"}"}}]
-
-            // minimax's tool calling don't have index field, it's data like:
-            // [{"id":"call_function_4752059746","type":"function","function":{"name":"lobe-image-designer____text2image____builtin","arguments":"{\"prompts\": [\"一个流浪的地球，背景是浩瀚"}}]
-
-            // so we need to add these default values
-            index: typeof value.index !== 'undefined' ? value.index : index,
-            type: value.type || 'function',
-          };
-        }),
-        id: chunk.id,
-        type: 'tool_calls',
-      } as StreamProtocolToolCallChunk;
+      const tool_calls = item.delta.tool_calls.filter(
+        (value) => value.index >= 0 || typeof value.index === 'undefined',
+      );
+
+      if (tool_calls.length > 0) {
+        return {
+          data: item.delta.tool_calls.map((value, index): StreamToolCallChunkData => {
+            if (streamContext && !streamContext.tool) {
+              streamContext.tool = {
+                id: value.id!,
+                index: value.index,
+                name: value.function!.name!,
+              };
+            }
+
+            return {
+              function: {
+                arguments: value.function?.arguments ?? '{}',
+                name: value.function?.name ?? null,
+              },
+              id:
+                value.id ||
+                streamContext?.tool?.id ||
+                generateToolCallId(index, value.function?.name),
+
+              // mistral's tool calling don't have index and function field, it's data like:
+              // [{"id":"xbhnmTtY7","function":{"name":"lobe-image-designer____text2image____builtin","arguments":"{\"prompts\": [\"A photo of a small, fluffy dog with a playful expression and wagging tail.\", \"A watercolor painting of a small, energetic dog with a glossy coat and bright eyes.\", \"A vector illustration of a small, adorable dog with a short snout and perky ears.\", \"A drawing of a small, scruffy dog with a mischievous grin and a wagging tail.\"], \"quality\": \"standard\", \"seeds\": [123456, 654321, 111222, 333444], \"size\": \"1024x1024\", \"style\": \"vivid\"}"}}]
+
+              // minimax's tool calling don't have index field, it's data like:
+              // [{"id":"call_function_4752059746","type":"function","function":{"name":"lobe-image-designer____text2image____builtin","arguments":"{\"prompts\": [\"一个流浪的地球，背景是浩瀚"}}]
+
+              // so we need to add these default values
+              index: typeof value.index !== 'undefined' ? value.index : index,
+              type: value.type || 'function',
+            };
+          }),
+          id: chunk.id,
+          type: 'tool_calls',
+        } as StreamProtocolToolCallChunk;
+      }
     }
 
     // 给定结束原因

package/src/libs/agent-runtime/utils/streams/protocol.ts

@@ -12,6 +12,10 @@ export interface StreamContext {
    * this flag is used to check if the pplx citation is returned,and then not return it again
    */
   returnedPplxCitation?: boolean;
+  thinking?: {
+    id: string;
+    name: string;
+  };
   tool?: {
     id: string;
     index: number;
@@ -29,6 +33,10 @@ export interface StreamProtocolChunk {
     | 'tool_calls'
     // Model Thinking
     | 'reasoning'
+    // use for reasoning signature, maybe only anthropic
+    | 'reasoning_signature'
+    // flagged reasoning signature
+    | 'flagged_reasoning_signature'
     // Search or Grounding
     | 'grounding'
     // stop signal

package/src/locales/default/chat.ts

@@ -32,7 +32,13 @@ export default {
   },
   duplicateTitle: '{{title}} 副本',
   emptyAgent: '暂无助手',
-  extendControls: {
+  extendParams: {
+    enableReasoning: {
+      title: '开启深度思考',
+    },
+    reasoningBudgetToken: {
+      title: '思考消耗 Token',
+    },
     title: '模型扩展功能',
   },
   historyRange: '历史范围',

package/src/services/__tests__/chat.test.ts

@@ -635,7 +635,7 @@ describe('ChatService', () => {
     });
   });
 
-  describe('processMessage', () => {
+  describe('reorderToolMessages', () => {
     it('should reorderToolMessages', () => {
       const input: OpenAIChatMessage[] = [
         {
@@ -746,7 +746,9 @@ describe('ChatService', () => {
         },
       ]);
     });
+  });
 
+  describe('processMessage', () => {
     describe('handle with files content in server mode', () => {
       it('should includes files', async () => {
         // 重新模拟模块，设置 isServerMode 为 true
@@ -833,46 +835,45 @@ describe('ChatService', () => {
           },
         ]);
       });
-    });
 
-    it('should include image files in server mode', async () => {
-      // 重新模拟模块，设置 isServerMode 为 true
-      vi.doMock('@/const/version', () => ({
-        isServerMode: true,
-        isDeprecatedEdition: true,
-      }));
+      it('should include image files in server mode', async () => {
+        // 重新模拟模块，设置 isServerMode 为 true
+        vi.doMock('@/const/version', () => ({
+          isServerMode: true,
+          isDeprecatedEdition: true,
+        }));
 
-      // 需要在修改模拟后重新导入相关模块
-      const { chatService } = await import('../chat');
-      const messages = [
-        {
-          content: 'Hello',
-          role: 'user',
-          imageList: [
-            {
-              id: 'file1',
-              url: 'http://example.com/image.jpg',
-              alt: 'abc.png',
-            },
-          ],
-        }, // Message with files
-        { content: 'Hey', role: 'assistant' }, // Regular user message
-      ] as ChatMessage[];
+        // 需要在修改模拟后重新导入相关模块
+        const { chatService } = await import('../chat');
+        const messages = [
+          {
+            content: 'Hello',
+            role: 'user',
+            imageList: [
+              {
+                id: 'file1',
+                url: 'http://example.com/image.jpg',
+                alt: 'abc.png',
+              },
+            ],
+          }, // Message with files
+          { content: 'Hey', role: 'assistant' }, // Regular user message
+        ] as ChatMessage[];
 
-      const getChatCompletionSpy = vi.spyOn(chatService, 'getChatCompletion');
-      await chatService.createAssistantMessage({
-        messages,
-        plugins: [],
-        model: 'gpt-4-vision-preview',
-      });
+        const getChatCompletionSpy = vi.spyOn(chatService, 'getChatCompletion');
+        await chatService.createAssistantMessage({
+          messages,
+          plugins: [],
+          model: 'gpt-4-vision-preview',
+        });
 
-      expect(getChatCompletionSpy).toHaveBeenCalledWith(
-        {
-          messages: [
-            {
-              content: [
-                {
-                  text: `Hello
+        expect(getChatCompletionSpy).toHaveBeenCalledWith(
+          {
+            messages: [
+              {
+                content: [
+                  {
+                    text: `Hello
 
 <!-- SYSTEM CONTEXT (NOT PART OF USER QUERY) -->
 <context.instruction>following part contains context information injected by the system. Please follow these instructions:
@@ -888,24 +889,61 @@ describe('ChatService', () => {
 
 </files_info>
 <!-- END SYSTEM CONTEXT -->`,
-                  type: 'text',
-                },
-                {
-                  image_url: { detail: 'auto', url: 'http://example.com/image.jpg' },
-                  type: 'image_url',
-                },
-              ],
-              role: 'user',
+                    type: 'text',
+                  },
+                  {
+                    image_url: { detail: 'auto', url: 'http://example.com/image.jpg' },
+                    type: 'image_url',
+                  },
+                ],
+                role: 'user',
+              },
+              {
+                content: 'Hey',
+                role: 'assistant',
+              },
+            ],
+            model: 'gpt-4-vision-preview',
+          },
+          undefined,
+        );
+      });
+    });
+
+    it('should handle assistant messages with reasoning correctly', () => {
+      const messages = [
+        {
+          role: 'assistant',
+          content: 'The answer is 42.',
+          reasoning: {
+            content: 'I need to calculate the answer to life, universe, and everything.',
+            signature: 'thinking_process',
+          },
+        },
+      ] as ChatMessage[];
+
+      const result = chatService['processMessages']({
+        messages,
+        model: 'gpt-4',
+        provider: 'openai',
+      });
+
+      expect(result).toEqual([
+        {
+          content: [
+            {
+              signature: 'thinking_process',
+              thinking: 'I need to calculate the answer to life, universe, and everything.',
+              type: 'thinking',
             },
             {
-              content: 'Hey',
-              role: 'assistant',
+              text: 'The answer is 42.',
+              type: 'text',
             },
           ],
-          model: 'gpt-4-vision-preview',
+          role: 'assistant',
         },
-        undefined,
-      );
+      ]);
     });
   });
 });
@@ -917,6 +955,7 @@ describe('ChatService', () => {
 vi.mock('../_auth', async (importOriginal) => {
   return importOriginal();
 });
+
 describe('AgentRuntimeOnClient', () => {
   describe('initializeWithClientStore', () => {
     describe('should initialize with options correctly', () => {

package/src/services/chat.ts

@@ -214,9 +214,35 @@ class ChatService {
 
     const tools = shouldUseTools ? filterTools : undefined;
 
+    // ============  3. process extend params   ============ //
+
+    let extendParams: Record<string, any> = {};
+
+    const isModelHasExtendParams = aiModelSelectors.isModelHasExtendParams(
+      payload.model,
+      payload.provider!,
+    )(useAiInfraStore.getState());
+
+    // model
+    if (isModelHasExtendParams) {
+      const modelExtendParams = aiModelSelectors.modelExtendParams(
+        payload.model,
+        payload.provider!,
+      )(useAiInfraStore.getState());
+      // if model has extended params, then we need to check if the model can use reasoning
+
+      if (modelExtendParams!.includes('enableReasoning') && chatConfig.enableReasoning) {
+        extendParams.thinking = {
+          budget_tokens: chatConfig.reasoningBudgetToken || 1024,
+          type: 'enabled',
+        };
+      }
+    }
+
     return this.getChatCompletion(
       {
         ...params,
+        ...extendParams,
         enabledSearch: enabledSearch && isModelHasBuiltinSearch ? true : undefined,
         messages: oaiMessages,
         tools,
@@ -459,8 +485,20 @@ class ChatService {
         }
 
         case 'assistant': {
+          // signature is a signal of anthropic thinking mode
+          const shouldIncludeThinking = m.reasoning && !!m.reasoning?.signature;
+
           return {
-            content: m.content,
+            content: shouldIncludeThinking
+              ? [
+                  {
+                    signature: m.reasoning!.signature,
+                    thinking: m.reasoning!.content,
+                    type: 'thinking',
+                  } as any,
+                  { text: m.content, type: 'text' },
+                ]
+              : m.content,
             role: m.role,
             tool_calls: m.tools?.map(
               (tool): MessageToolCall => ({

package/src/store/agent/slices/chat/__snapshots__/selectors.test.ts.snap

@@ -8,7 +8,9 @@ exports[`agentSelectors > defaultAgentConfig > should merge DEFAULT_AGENT_CONFIG
     "enableAutoCreateTopic": true,
     "enableCompressHistory": true,
     "enableHistoryCount": true,
+    "enableReasoning": true,
     "historyCount": 8,
+    "reasoningBudgetToken": 1024,
     "searchMode": "off",
   },
   "model": "gpt-3.5-turbo",

package/src/store/aiInfra/slices/aiModel/selectors.ts

@@ -70,14 +70,14 @@ const modelContextWindowTokens = (id: string, provider: string) => (s: AIProvide
   return model?.contextWindowTokens;
 };
 
-const modelExtendControls = (id: string, provider: string) => (s: AIProviderStoreState) => {
+const modelExtendParams = (id: string, provider: string) => (s: AIProviderStoreState) => {
   const model = getEnabledModelById(id, provider)(s);
 
-  return model?.settings?.extendControls;
+  return model?.settings?.extendParams;
 };
 
-const isModelHasExtendControls = (id: string, provider: string) => (s: AIProviderStoreState) => {
-  const controls = modelExtendControls(id, provider)(s);
+const isModelHasExtendParams = (id: string, provider: string) => (s: AIProviderStoreState) => {
+  const controls = modelExtendParams(id, provider)(s);
 
   return !!controls && controls.length > 0;
 };
@@ -119,13 +119,13 @@ export const aiModelSelectors = {
   isModelHasBuiltinSearch,
   isModelHasBuiltinSearchConfig,
   isModelHasContextWindowToken,
-  isModelHasExtendControls,
+  isModelHasExtendParams,
   isModelLoading,
   isModelSupportReasoning,
   isModelSupportToolUse,
   isModelSupportVision,
   modelBuiltinSearchImpl,
   modelContextWindowTokens,
-  modelExtendControls,
+  modelExtendParams,
   totalAiProviderModelList,
 };

package/src/store/chat/slices/aiChat/actions/generateAIChat.ts

@@ -472,7 +472,7 @@ export const generateAIChat: StateCreator<
         // update the content after fetch result
         await internal_updateMessageContent(messageId, content, {
           toolCalls,
-          reasoning: !!reasoning ? { content: reasoning, duration } : undefined,
+          reasoning: !!reasoning ? { ...reasoning, duration } : undefined,
           search: !!grounding?.citations ? grounding : undefined,
         });
       },

package/src/store/user/slices/settings/selectors/__snapshots__/settings.test.ts.snap

@@ -75,7 +75,9 @@ exports[`settingsSelectors > defaultAgent > should merge DEFAULT_AGENT and s.set
       "enableAutoCreateTopic": true,
       "enableCompressHistory": true,
       "enableHistoryCount": true,
+      "enableReasoning": true,
       "historyCount": 8,
+      "reasoningBudgetToken": 1024,
       "searchMode": "off",
     },
     "model": "gpt-3.5-turbo",

package/src/types/agent/index.ts

@@ -55,33 +55,45 @@ export interface LobeAgentConfig {
   tts: LobeAgentTTSConfig;
 }
 
+/* eslint-disable sort-keys-fix/sort-keys-fix, typescript-sort-keys/interface */
+
 export interface LobeAgentChatConfig {
-  autoCreateTopicThreshold: number;
   displayMode?: 'chat' | 'docs';
+
   enableAutoCreateTopic?: boolean;
-  /**
-   * 历史消息长度压缩阈值
-   */
-  enableCompressHistory?: boolean;
-  /**
-   * 开启历史记录条数
-   */
-  enableHistoryCount?: boolean;
+  autoCreateTopicThreshold: number;
+
   enableMaxTokens?: boolean;
 
+  /**
+   * 是否开启推理
+   */
+  enableReasoning?: boolean;
   /**
    * 自定义推理强度
    */
   enableReasoningEffort?: boolean;
+  reasoningBudgetToken?: number;
 
   /**
    * 历史消息条数
    */
   historyCount?: number;
+  /**
+   * 开启历史记录条数
+   */
+  enableHistoryCount?: boolean;
+  /**
+   * 历史消息长度压缩阈值
+   */
+  enableCompressHistory?: boolean;
+
   inputTemplate?: string;
+
   searchMode?: SearchMode;
   useModelBuiltinSearch?: boolean;
 }
+/* eslint-enable */
 
 export const AgentChatConfigSchema = z.object({
   autoCreateTopicThreshold: z.number().default(2),
@@ -90,8 +102,10 @@ export const AgentChatConfigSchema = z.object({
   enableCompressHistory: z.boolean().optional(),
   enableHistoryCount: z.boolean().optional(),
   enableMaxTokens: z.boolean().optional(),
+  enableReasoning: z.boolean().optional(),
   enableReasoningEffort: z.boolean().optional(),
   historyCount: z.number().optional(),
+  reasoningBudgetToken: z.number().optional(),
   searchMode: z.enum(['off', 'on', 'auto']).optional(),
 });