@librechat/agents 3.0.776 → 3.1.0

package/src/scripts/thinking-bedrock.ts

@@ -0,0 +1,159 @@
+// src/scripts/thinking-bedrock.ts
+import { config } from 'dotenv';
+config();
+import { HumanMessage, BaseMessage } from '@langchain/core/messages';
+import type { UsageMetadata } from '@langchain/core/messages';
+import * as t from '@/types';
+import { ChatModelStreamHandler, createContentAggregator } from '@/stream';
+import { createCodeExecutionTool } from '@/tools/CodeExecutor';
+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 testBedrockThinking(): 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 = getLLMConfig(Providers.BEDROCK);
+
+  // Enable thinking with token budget for Bedrock
+  const llmConfig = {
+    ...baseLlmConfig,
+    model: 'us.anthropic.claude-3-7-sonnet-20250219-v1:0',
+    maxTokens: 5000,
+    additionalModelRequestFields: {
+      thinking: { type: 'enabled', budget_tokens: 2000 },
+    },
+  };
+
+  const run = await Run.create<t.IState>({
+    runId: 'test-bedrock-thinking-id',
+    graphConfig: {
+      instructions,
+      type: 'standard',
+      tools: [createCodeExecutionTool()],
+      llmConfig,
+    },
+    returnContent: true,
+    customHandlers: customHandlers as t.RunConfig['customHandlers'],
+  });
+
+  const config = {
+    configurable: {
+      thread_id: 'bedrock-thinking-test-thread',
+    },
+    streamMode: 'values',
+    version: 'v2' as const,
+  };
+
+  // Test 1: Regular thinking mode
+  console.log('\n\nTest 1: Bedrock Regular thinking mode');
+  const userMessage1 = `Please print 'hello world' in python`;
+  conversationHistory.push(new HumanMessage(userMessage1));
+
+  console.log('Running first query with Bedrock thinking enabled...');
+  const firstInputs = { messages: [...conversationHistory] };
+  await run.processStream(firstInputs, config);
+
+  // Extract and display thinking blocks
+  const finalMessages = run.getRunMessages();
+  console.log('\n\nFinal messages after Test 1:');
+  console.dir(finalMessages, { depth: null });
+
+  // Test 2: Try multi-turn conversation
+  console.log(
+    '\n\nTest 2: Multi-turn conversation with Bedrock 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 Bedrock thinking enabled...');
+  const secondInputs = { messages: [...conversationHistory] };
+  await run.processStream(secondInputs, config);
+
+  // Display thinking blocks for second response
+  const finalMessages2 = run.getRunMessages();
+  console.log('\n\nBedrock thinking feature test completed!');
+  console.dir(finalMessages2, { depth: null });
+
+  console.log('\n\nContent parts:');
+  console.dir(_contentParts, { depth: null });
+}
+
+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);
+});
+
+testBedrockThinking().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/thinking.ts

@@ -1,7 +1,11 @@
 // src/scripts/test-thinking.ts
 import { config } from 'dotenv';
 config();
-import { HumanMessage, SystemMessage, BaseMessage } from '@langchain/core/messages';
+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';
@@ -21,17 +25,23 @@ async function testThinking(): Promise<void> {
   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 => {
+      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 } });
-      }
+        aggregateContent({
+          event,
+          data: data as unknown as { result: t.ToolEndEvent },
+        });
+      },
     },
     [GraphEvents.ON_RUN_STEP]: {
       handle: (event: GraphEvents.ON_RUN_STEP, data: t.RunStep) => {
@@ -39,29 +49,38 @@ async function testThinking(): Promise<void> {
       },
     },
     [GraphEvents.ON_RUN_STEP_DELTA]: {
-      handle: (event: GraphEvents.ON_RUN_STEP_DELTA, data: t.RunStepDeltaEvent) => {
+      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) => {
+      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) => {
+      handle: (
+        event: GraphEvents.ON_REASONING_DELTA,
+        data: t.ReasoningDeltaEvent
+      ) => {
         aggregateContent({ event, data });
       },
     },
   };
 
   const baseLlmConfig: t.LLMConfig = getLLMConfig(Providers.ANTHROPIC);
-  
+
   // Enable thinking with token budget
   const llmConfig = {
     ...baseLlmConfig,
     model: 'claude-3-7-sonnet-latest',
-    thinking: { type: "enabled", budget_tokens: 2000 }
+    thinking: { type: 'enabled', budget_tokens: 2000 },
   };
 
   const run = await Run.create<t.IState>({
@@ -93,7 +112,7 @@ async function testThinking(): Promise<void> {
   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();
 
@@ -101,30 +120,32 @@ async function testThinking(): Promise<void> {
   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 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!');
+  console.dir(finalMessages3, { depth: null });
 }
 
 process.on('unhandledRejection', (reason, promise) => {
@@ -147,4 +168,4 @@ testThinking().catch((err) => {
   console.log('Content parts:');
   console.dir(_contentParts, { depth: null });
   process.exit(1);
-});
+});

package/src/scripts/tools.ts

@@ -18,9 +18,13 @@ async function testStandardStreaming(): Promise<void> {
   const { userName, location, provider, currentDate } = await getArgs();
   const { contentParts, aggregateContent } = createContentAggregator();
   const customHandlers = {
-    [GraphEvents.TOOL_END]: new ToolEndHandler(undefined, (name?: string) => {
-      return true;
-    }),
+    [GraphEvents.TOOL_END]: new ToolEndHandler(
+      undefined,
+      undefined,
+      (name?: string) => {
+        return true;
+      }
+    ),
     [GraphEvents.CHAT_MODEL_END]: {
       handle: (
         _event: string,

package/src/stream.ts

@@ -339,7 +339,8 @@ hasToolCallChunks: ${hasToolCallChunks}
         (c) =>
           (c.type?.startsWith(ContentTypes.THINKING) ?? false) ||
           (c.type?.startsWith(ContentTypes.REASONING) ?? false) ||
-          (c.type?.startsWith(ContentTypes.REASONING_CONTENT) ?? false)
+          (c.type?.startsWith(ContentTypes.REASONING_CONTENT) ?? false) ||
+          c.type === 'redacted_thinking'
       )
     ) {
       await graph.dispatchReasoningDelta(stepId, {
@@ -365,7 +366,8 @@ hasToolCallChunks: ${hasToolCallChunks}
       Array.isArray(chunk.content) &&
       (chunk.content[0]?.type === ContentTypes.THINKING ||
         chunk.content[0]?.type === ContentTypes.REASONING ||
-        chunk.content[0]?.type === ContentTypes.REASONING_CONTENT)
+        chunk.content[0]?.type === ContentTypes.REASONING_CONTENT ||
+        chunk.content[0]?.type === 'redacted_thinking')
     ) {
       reasoning_content = 'valid';
     } else if (

package/src/tools/ToolNode.ts

@@ -145,9 +145,9 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
 
       /**
        * Inject session context for code execution tools when available.
+       * Each file uses its own session_id (supporting multi-session file tracking).
        * Both session_id and _injected_files are injected directly to invokeParams
        * (not inside args) so they bypass Zod schema validation and reach config.toolCall.
-       * This avoids /files endpoint race conditions.
        */
       if (
         call.name === Constants.EXECUTE_CODE ||
@@ -156,14 +156,18 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
         const codeSession = this.sessions?.get(Constants.EXECUTE_CODE) as
           | t.CodeSessionContext
           | undefined;
-        if (codeSession?.session_id != null && codeSession.files.length > 0) {
-          /** Convert tracked files to CodeEnvFile format for the API */
+        if (codeSession?.files != null && codeSession.files.length > 0) {
+          /**
+           * Convert tracked files to CodeEnvFile format for the API.
+           * Each file uses its own session_id (set when file was created).
+           * This supports files from multiple parallel/sequential executions.
+           */
           const fileRefs: t.CodeEnvFile[] = codeSession.files.map((file) => ({
-            session_id: codeSession.session_id,
+            session_id: file.session_id ?? codeSession.session_id,
             id: file.id,
             name: file.name,
           }));
-          /** Inject session_id and files directly - bypasses Zod, reaches config.toolCall */
+          /** Inject latest session_id and files - bypasses Zod, reaches config.toolCall */
           invokeParams = {
             ...invokeParams,
             session_id: codeSession.session_id,

package/src/types/tools.ts

@@ -74,6 +74,8 @@ export type FileRef = {
   id: string;
   name: string;
   path?: string;
+  /** Session ID this file belongs to (for multi-session file tracking) */
+  session_id?: string;
 };
 
 export type FileRefs = FileRef[];