@librechat/agents 3.0.78 → 3.0.80

package/src/llm/bedrock/utils/message_outputs.ts

@@ -0,0 +1,310 @@
+/**
+ * Utility functions for converting Bedrock Converse responses to LangChain messages.
+ * Ported from @langchain/aws common.js
+ */
+import { AIMessage, AIMessageChunk } from '@langchain/core/messages';
+import { ChatGenerationChunk } from '@langchain/core/outputs';
+import type {
+  BedrockMessage,
+  ConverseResponse,
+  ContentBlockDeltaEvent,
+  ConverseStreamMetadataEvent,
+  ContentBlockStartEvent,
+  ReasoningContentBlock,
+  ReasoningContentBlockDelta,
+  MessageContentReasoningBlock,
+  MessageContentReasoningBlockReasoningTextPartial,
+  MessageContentReasoningBlockRedacted,
+} from '../types';
+
+/**
+ * Convert a Bedrock reasoning block delta to a LangChain partial reasoning block.
+ */
+export function bedrockReasoningDeltaToLangchainPartialReasoningBlock(
+  reasoningContent: ReasoningContentBlockDelta
+):
+  | MessageContentReasoningBlockReasoningTextPartial
+  | MessageContentReasoningBlockRedacted {
+  const { text, redactedContent, signature } =
+    reasoningContent as ReasoningContentBlockDelta & {
+      text?: string;
+      redactedContent?: Uint8Array;
+      signature?: string;
+    };
+
+  if (typeof text === 'string') {
+    return {
+      type: 'reasoning_content',
+      reasoningText: { text },
+    };
+  }
+  if (signature != null) {
+    return {
+      type: 'reasoning_content',
+      reasoningText: { signature },
+    };
+  }
+  if (redactedContent != null) {
+    return {
+      type: 'reasoning_content',
+      redactedContent: Buffer.from(redactedContent).toString('base64'),
+    };
+  }
+  throw new Error('Invalid reasoning content');
+}
+
+/**
+ * Convert a Bedrock reasoning block to a LangChain reasoning block.
+ */
+export function bedrockReasoningBlockToLangchainReasoningBlock(
+  reasoningContent: ReasoningContentBlock
+): MessageContentReasoningBlock {
+  const { reasoningText, redactedContent } =
+    reasoningContent as ReasoningContentBlock & {
+      reasoningText?: { text?: string; signature?: string };
+      redactedContent?: Uint8Array;
+    };
+
+  if (reasoningText != null) {
+    return {
+      type: 'reasoning_content',
+      reasoningText: reasoningText,
+    };
+  }
+  if (redactedContent != null) {
+    return {
+      type: 'reasoning_content',
+      redactedContent: Buffer.from(redactedContent).toString('base64'),
+    };
+  }
+  throw new Error('Invalid reasoning content');
+}
+
+/**
+ * Convert a Bedrock Converse message to a LangChain message.
+ */
+export function convertConverseMessageToLangChainMessage(
+  message: BedrockMessage,
+  responseMetadata: Omit<ConverseResponse, 'output'>
+): AIMessage {
+  if (message.content == null) {
+    throw new Error('No message content found in response.');
+  }
+  if (message.role !== 'assistant') {
+    throw new Error(
+      `Unsupported message role received in ChatBedrockConverse response: ${message.role}`
+    );
+  }
+
+  let requestId: string | undefined;
+  if (
+    '$metadata' in responseMetadata &&
+    responseMetadata.$metadata != null &&
+    typeof responseMetadata.$metadata === 'object' &&
+    'requestId' in responseMetadata.$metadata
+  ) {
+    requestId = responseMetadata.$metadata.requestId as string;
+  }
+
+  let tokenUsage:
+    | { input_tokens: number; output_tokens: number; total_tokens: number }
+    | undefined;
+  if (responseMetadata.usage != null) {
+    const input_tokens = responseMetadata.usage.inputTokens ?? 0;
+    const output_tokens = responseMetadata.usage.outputTokens ?? 0;
+    tokenUsage = {
+      input_tokens,
+      output_tokens,
+      total_tokens:
+        responseMetadata.usage.totalTokens ?? input_tokens + output_tokens,
+    };
+  }
+
+  if (
+    message.content.length === 1 &&
+    'text' in message.content[0] &&
+    typeof message.content[0].text === 'string'
+  ) {
+    return new AIMessage({
+      content: message.content[0].text,
+      response_metadata: responseMetadata,
+      usage_metadata: tokenUsage,
+      id: requestId,
+    });
+  } else {
+    const toolCalls: Array<{
+      id?: string;
+      name: string;
+      args: Record<string, unknown>;
+      type: 'tool_call';
+    }> = [];
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    const content: any[] = [];
+
+    message.content.forEach((c) => {
+      if (
+        'toolUse' in c &&
+        c.toolUse != null &&
+        c.toolUse.name != null &&
+        c.toolUse.name !== '' &&
+        c.toolUse.input != null &&
+        typeof c.toolUse.input === 'object'
+      ) {
+        toolCalls.push({
+          id: c.toolUse.toolUseId,
+          name: c.toolUse.name,
+          args: c.toolUse.input as Record<string, unknown>,
+          type: 'tool_call',
+        });
+      } else if ('text' in c && typeof c.text === 'string') {
+        content.push({ type: 'text', text: c.text });
+      } else if ('reasoningContent' in c && c.reasoningContent != null) {
+        content.push(
+          bedrockReasoningBlockToLangchainReasoningBlock(c.reasoningContent)
+        );
+      } else {
+        content.push(c);
+      }
+    });
+
+    return new AIMessage({
+      content: content.length ? content : '',
+      tool_calls: toolCalls.length ? toolCalls : undefined,
+      response_metadata: responseMetadata,
+      usage_metadata: tokenUsage,
+      id: requestId,
+    });
+  }
+}
+
+/**
+ * Handle a content block delta event from Bedrock Converse stream.
+ */
+export function handleConverseStreamContentBlockDelta(
+  contentBlockDelta: ContentBlockDeltaEvent
+): ChatGenerationChunk {
+  if (contentBlockDelta.delta == null) {
+    throw new Error('No delta found in content block.');
+  }
+
+  if (typeof contentBlockDelta.delta.text === 'string') {
+    return new ChatGenerationChunk({
+      text: contentBlockDelta.delta.text,
+      message: new AIMessageChunk({
+        content: contentBlockDelta.delta.text,
+        response_metadata: {
+          contentBlockIndex: contentBlockDelta.contentBlockIndex,
+        },
+      }),
+    });
+  } else if (contentBlockDelta.delta.toolUse != null) {
+    const index = contentBlockDelta.contentBlockIndex;
+    return new ChatGenerationChunk({
+      text: '',
+      message: new AIMessageChunk({
+        content: '',
+        tool_call_chunks: [
+          {
+            args: contentBlockDelta.delta.toolUse.input as string,
+            index,
+            type: 'tool_call_chunk',
+          },
+        ],
+        response_metadata: {
+          contentBlockIndex: contentBlockDelta.contentBlockIndex,
+        },
+      }),
+    });
+  } else if (contentBlockDelta.delta.reasoningContent != null) {
+    const reasoningBlock =
+      bedrockReasoningDeltaToLangchainPartialReasoningBlock(
+        contentBlockDelta.delta.reasoningContent
+      );
+    // Extract the text for additional_kwargs.reasoning_content (for stream handler compatibility)
+    const reasoningText =
+      'reasoningText' in reasoningBlock
+        ? (reasoningBlock.reasoningText.text ??
+          reasoningBlock.reasoningText.signature ??
+          ('redactedContent' in reasoningBlock
+            ? reasoningBlock.redactedContent
+            : ''))
+        : '';
+    return new ChatGenerationChunk({
+      text: '',
+      message: new AIMessageChunk({
+        content: [reasoningBlock],
+        additional_kwargs: {
+          // Set reasoning_content for stream handler to detect reasoning mode
+          reasoning_content: reasoningText,
+        },
+        response_metadata: {
+          contentBlockIndex: contentBlockDelta.contentBlockIndex,
+        },
+      }),
+    });
+  } else {
+    throw new Error(
+      `Unsupported content block type(s): ${JSON.stringify(contentBlockDelta.delta, null, 2)}`
+    );
+  }
+}
+
+/**
+ * Handle a content block start event from Bedrock Converse stream.
+ */
+export function handleConverseStreamContentBlockStart(
+  contentBlockStart: ContentBlockStartEvent
+): ChatGenerationChunk | null {
+  const index = contentBlockStart.contentBlockIndex;
+
+  if (contentBlockStart.start?.toolUse != null) {
+    return new ChatGenerationChunk({
+      text: '',
+      message: new AIMessageChunk({
+        content: '',
+        tool_call_chunks: [
+          {
+            name: contentBlockStart.start.toolUse.name,
+            id: contentBlockStart.start.toolUse.toolUseId,
+            index,
+            type: 'tool_call_chunk',
+          },
+        ],
+        response_metadata: {
+          contentBlockIndex: index,
+        },
+      }),
+    });
+  }
+
+  // Return null for non-tool content block starts (text blocks don't need special handling)
+  return null;
+}
+
+/**
+ * Handle a metadata event from Bedrock Converse stream.
+ */
+export function handleConverseStreamMetadata(
+  metadata: ConverseStreamMetadataEvent,
+  extra: { streamUsage: boolean }
+): ChatGenerationChunk {
+  const inputTokens = metadata.usage?.inputTokens ?? 0;
+  const outputTokens = metadata.usage?.outputTokens ?? 0;
+  const usage_metadata = {
+    input_tokens: inputTokens,
+    output_tokens: outputTokens,
+    total_tokens: metadata.usage?.totalTokens ?? inputTokens + outputTokens,
+  };
+
+  return new ChatGenerationChunk({
+    text: '',
+    message: new AIMessageChunk({
+      content: '',
+      usage_metadata: extra.streamUsage ? usage_metadata : undefined,
+      response_metadata: {
+        // Use the same key as returned from the Converse API
+        metadata,
+      },
+    }),
+  });
+}

package/src/messages/core.ts

@@ -41,7 +41,7 @@ User: ${userMessage[1]}
 const _allowedTypes = ['image_url', 'text', 'tool_use', 'tool_result'];
 const allowedTypesByProvider: Record<string, string[]> = {
   default: _allowedTypes,
-  [Providers.ANTHROPIC]: [..._allowedTypes, 'thinking'],
+  [Providers.ANTHROPIC]: [..._allowedTypes, 'thinking', 'redacted_thinking'],
   [Providers.BEDROCK]: [..._allowedTypes, 'reasoning_content'],
   [Providers.OPENAI]: _allowedTypes,
 };

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 (