@librechat/agents 2.1.2 → 2.1.4

package/src/scripts/thinking.ts

@@ -0,0 +1,152 @@
+// src/scripts/test-thinking.ts
+import { config } from 'dotenv';
+config();
+import { HumanMessage, SystemMessage, BaseMessage } from '@langchain/core/messages';
+import type { UsageMetadata } from '@langchain/core/messages';
+import * as t from '@/types';
+import { ChatModelStreamHandler, createContentAggregator } from '@/stream';
+import { ToolEndHandler, ModelEndHandler } from '@/events';
+import { GraphEvents, Providers } from '@/common';
+import { getLLMConfig } from '@/utils/llmConfig';
+import { getArgs } from '@/scripts/args';
+import { Run } from '@/run';
+
+const conversationHistory: BaseMessage[] = [];
+let _contentParts: t.MessageContentComplex[] = [];
+const collectedUsage: UsageMetadata[] = [];
+
+async function testThinking(): Promise<void> {
+  const { userName } = await getArgs();
+  const instructions = `You are a helpful AI assistant for ${userName}. When answering questions, be thorough in your reasoning.`;
+  const { contentParts, aggregateContent } = createContentAggregator();
+  _contentParts = contentParts as t.MessageContentComplex[];
+  
+  // Set up event handlers
+  const customHandlers = {
+    [GraphEvents.TOOL_END]: new ToolEndHandler(),
+    [GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(collectedUsage),
+    [GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
+    [GraphEvents.ON_RUN_STEP_COMPLETED]: {
+      handle: (event: GraphEvents.ON_RUN_STEP_COMPLETED, data: t.StreamEventData): void => {
+        console.log('====== ON_RUN_STEP_COMPLETED ======');
+        aggregateContent({ event, data: data as unknown as { result: t.ToolEndEvent } });
+      }
+    },
+    [GraphEvents.ON_RUN_STEP]: {
+      handle: (event: GraphEvents.ON_RUN_STEP, data: t.RunStep) => {
+        aggregateContent({ event, data });
+      },
+    },
+    [GraphEvents.ON_RUN_STEP_DELTA]: {
+      handle: (event: GraphEvents.ON_RUN_STEP_DELTA, data: t.RunStepDeltaEvent) => {
+        aggregateContent({ event, data });
+      },
+    },
+    [GraphEvents.ON_MESSAGE_DELTA]: {
+      handle: (event: GraphEvents.ON_MESSAGE_DELTA, data: t.MessageDeltaEvent) => {
+        aggregateContent({ event, data });
+      },
+    },
+    [GraphEvents.ON_REASONING_DELTA]: {
+      handle: (event: GraphEvents.ON_REASONING_DELTA, data: t.ReasoningDeltaEvent) => {
+        aggregateContent({ event, data });
+      },
+    },
+  };
+
+  const baseLlmConfig: t.LLMConfig = getLLMConfig(Providers.ANTHROPIC);
+  
+  if (baseLlmConfig.provider !== 'anthropic') {
+    console.error('This test requires Anthropic as the LLM provider. Please specify provider=anthropic');
+    process.exit(1);
+  }
+  
+  // Enable thinking with token budget
+  const llmConfig = {
+    ...baseLlmConfig,
+    model: 'claude-3-7-sonnet-latest',
+    thinking: { type: "enabled", budget_tokens: 2000 }
+  };
+
+  const run = await Run.create<t.IState>({
+    runId: 'test-thinking-id',
+    graphConfig: {
+      instructions,
+      type: 'standard',
+      llmConfig,
+    },
+    returnContent: true,
+    customHandlers: customHandlers as t.RunConfig['customHandlers'],
+  });
+
+  const config = {
+    configurable: {
+      thread_id: 'thinking-test-thread',
+    },
+    streamMode: 'values',
+    version: 'v2' as const,
+  };
+
+  // Test 1: Regular thinking mode
+  console.log('\n\nTest 1: Regular thinking mode');
+  const userMessage1 = `What would be the environmental and economic impacts if all cars globally were replaced by electric vehicles overnight?`;
+  conversationHistory.push(new HumanMessage(userMessage1));
+
+  console.log('Running first query with thinking enabled...');
+  const firstInputs = { messages: [...conversationHistory] };
+  await run.processStream(firstInputs, config);
+  
+  // Extract and display thinking blocks
+  const finalMessages = run.getRunMessages();
+
+  // Test 2: Try multi-turn conversation
+  console.log('\n\nTest 2: Multi-turn conversation with thinking enabled');
+  const userMessage2 = `Given your previous analysis, what would be the most significant technical challenges in making this transition?`;
+  conversationHistory.push(new HumanMessage(userMessage2));
+  
+  console.log('Running second query with thinking enabled...');
+  const secondInputs = { messages: [...conversationHistory] };
+  await run.processStream(secondInputs, config);
+  
+  // Display thinking blocks for second response
+  const finalMessages2 = run.getRunMessages();
+
+  // Test 3: Redacted thinking mode
+  console.log('\n\nTest 3: Redacted thinking mode');
+  const magicString = "ANTHROPIC_MAGIC_STRING_TRIGGER_REDACTED_THINKING_46C9A13E193C177646C7398A98432ECCCE4C1253D5E2D82641AC0E52CC2876CB";
+  const userMessage3 = `${magicString}\n\nExplain how quantum computing works in simple terms.`;
+  
+  // Reset conversation for clean test
+  conversationHistory.length = 0;
+  conversationHistory.push(new HumanMessage(userMessage3));
+  
+  console.log('Running query with redacted thinking...');
+  const thirdInputs = { messages: [...conversationHistory] };
+  await run.processStream(thirdInputs, config);
+  
+  // Display redacted thinking blocks
+  const finalMessages3 = run.getRunMessages();
+  console.log('\n\nThinking feature test completed!');
+}
+
+process.on('unhandledRejection', (reason, promise) => {
+  console.error('Unhandled Rejection at:', promise, 'reason:', reason);
+  console.log('Conversation history:');
+  console.dir(conversationHistory, { depth: null });
+  console.log('Content parts:');
+  console.dir(_contentParts, { depth: null });
+  process.exit(1);
+});
+
+process.on('uncaughtException', (err) => {
+  console.error('Uncaught Exception:', err);
+});
+
+testThinking().catch((err) => {
+  console.error(err);
+  console.log('Conversation history:');
+  console.dir(conversationHistory, { depth: null });
+  console.log('Content parts:');
+  console.dir(_contentParts, { depth: null });
+  process.exit(1);
+});

package/src/scripts/tools.ts

@@ -6,7 +6,7 @@ import { HumanMessage, BaseMessage } from '@langchain/core/messages';
 import { TavilySearchResults } from '@langchain/community/tools/tavily_search';
 import type * as t from '@/types';
 import { ChatModelStreamHandler, createContentAggregator } from '@/stream';
-import { ToolEndHandler } from '@/events';
+import { ToolEndHandler, ModelEndHandler } from '@/events';
 
 
 import { getArgs } from '@/scripts/args';
@@ -20,7 +20,7 @@ async function testStandardStreaming(): Promise<void> {
   const { contentParts, aggregateContent } = createContentAggregator();
   const customHandlers = {
     [GraphEvents.TOOL_END]: new ToolEndHandler(),
-    // [GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(),
+    [GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(),
     [GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
     [GraphEvents.ON_RUN_STEP_COMPLETED]: {
       handle: (event: GraphEvents.ON_RUN_STEP_COMPLETED, data: t.StreamEventData): void => {

package/src/splitStream.ts

@@ -141,14 +141,19 @@ export class SplitStreamHandler {
       });
     }
   };
+  getDeltaContent(chunk: t.CustomChunk): string {
+    return chunk.choices?.[0]?.delta.content ?? '';
+  }
+  getReasoningDelta(chunk: t.CustomChunk): string {
+    return chunk.choices?.[0]?.delta[this.reasoningKey] ?? '';
+  }
   handle(chunk?: t.CustomChunk): void {
     if (!chunk) {
       return;
     }
 
-    const content = chunk.choices?.[0]?.delta.content ?? '';
-    const reasoning_content = chunk.choices?.[0]?.delta[this.reasoningKey] ?? '';
-
+    const content = this.getDeltaContent(chunk);
+    const reasoning_content = this.getReasoningDelta(chunk);
     if (!content.length && !reasoning_content.length) {
       return;
     }

package/src/stream.ts

@@ -213,6 +213,12 @@ hasToolCallChunks: ${hasToolCallChunks}
       graph.dispatchMessageDelta(stepId, {
         content,
       });
+    } else if (content.every((c) => c.type?.startsWith(ContentTypes.THINKING))) {
+      graph.dispatchReasoningDelta(stepId, {
+        content: content.map((c) => ({
+          type: ContentTypes.THINK,
+          think: (c as t.ThinkingContentText).thinking,
+      }))});
     }
   }
   handleToolCallChunks = ({
@@ -271,8 +277,11 @@ hasToolCallChunks: ${hasToolCallChunks}
     });
   };
   handleReasoning(chunk: Partial<AIMessageChunk>, graph: Graph): void {
-    const reasoning_content = chunk.additional_kwargs?.[graph.reasoningKey] as string | undefined;
-    if (reasoning_content != null && reasoning_content && (chunk.content == null || chunk.content === '')) {
+    let reasoning_content = chunk.additional_kwargs?.[graph.reasoningKey] as string | undefined;
+    if (Array.isArray(chunk.content) && chunk.content[0]?.type === 'thinking') {
+      reasoning_content = 'valid';
+    }
+    if (reasoning_content != null && reasoning_content && (chunk.content == null || chunk.content === '' || reasoning_content === 'valid')) {
       graph.currentTokenType = ContentTypes.THINK;
       graph.tokenTypeSwitch = 'reasoning';
       return;

package/src/types/stream.ts

@@ -216,9 +216,16 @@ export type ReasoningContentText = {
   think: string;
 };
 
+/** Anthropic's Reasoning Content Block Format */
+export type ThinkingContentText = {
+  type: ContentTypes.THINKING;
+  index?: number;
+  thinking: string;
+};
+
 // eslint-disable-next-line @typescript-eslint/no-explicit-any
-export type MessageContentComplex = (ReasoningContentText | MessageContentText | MessageContentImageUrl | (Record<string, any> & {
-  type?: 'text' | 'image_url' | 'think' | string;
+export type MessageContentComplex = (ThinkingContentText | ReasoningContentText | MessageContentText | MessageContentImageUrl | (Record<string, any> & {
+  type?: 'text' | 'image_url' | 'think' | 'thinking' | string;
 // eslint-disable-next-line @typescript-eslint/no-explicit-any
 }) | (Record<string, any> & {
   type?: never;