@librechat/agents 3.0.776 → 3.1.0

package/dist/types/llm/bedrock/index.d.ts

@@ -1,5 +1,9 @@
 /**
  * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts
+ * and adds support for latest @langchain/aws features:
+ *
+ * - Application Inference Profiles (PR #9129)
+ * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+
  *
  * Bedrock sends the same contentBlockIndex for both text and tool_use content blocks,
  * causing LangChain's merge logic to fail with "field[contentBlockIndex] already exists"
@@ -11,19 +15,94 @@
  * the purpose of tracking tool call ordering.
  */
 import { ChatBedrockConverse } from '@langchain/aws';
+import { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';
+import type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';
 import type { ChatBedrockConverseInput } from '@langchain/aws';
 import type { BaseMessage } from '@langchain/core/messages';
-import { ChatGenerationChunk } from '@langchain/core/outputs';
-import type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';
+/**
+ * Service tier type for Bedrock invocations.
+ * Requires AWS SDK >= 3.966.0 to actually work.
+ * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
+ */
+export type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';
+/**
+ * Extended input interface with additional features:
+ * - applicationInferenceProfile: Use an inference profile ARN instead of model ID
+ * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)
+ */
+export interface CustomChatBedrockConverseInput extends ChatBedrockConverseInput {
+    /**
+     * Application Inference Profile ARN to use for the model.
+     * For example, "arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx"
+     * When provided, this ARN will be used for the actual inference calls instead of the model ID.
+     * Must still provide `model` as normal modelId to benefit from all the metadata.
+     * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html
+     */
+    applicationInferenceProfile?: string;
+    /**
+     * Service tier for model invocation.
+     * Specifies the processing tier type used for serving the request.
+     * Supported values are 'priority', 'default', 'flex', and 'reserved'.
+     *
+     * - 'priority': Prioritized processing for lower latency
+     * - 'default': Standard processing tier
+     * - 'flex': Flexible processing tier with lower cost
+     * - 'reserved': Reserved capacity for consistent performance
+     *
+     * If not provided, AWS uses the default tier.
+     * Note: Requires AWS SDK >= 3.966.0 to work.
+     * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
+     */
+    serviceTier?: ServiceTierType;
+}
+/**
+ * Extended call options with serviceTier override support.
+ */
+export interface CustomChatBedrockConverseCallOptions {
+    serviceTier?: ServiceTierType;
+}
 export declare class CustomChatBedrockConverse extends ChatBedrockConverse {
-    constructor(fields?: ChatBedrockConverseInput);
+    /**
+     * Application Inference Profile ARN to use instead of model ID.
+     */
+    applicationInferenceProfile?: string;
+    /**
+     * Service tier for model invocation.
+     */
+    serviceTier?: ServiceTierType;
+    constructor(fields?: CustomChatBedrockConverseInput);
     static lc_name(): string;
     /**
-     * Override _streamResponseChunks to strip contentBlockIndex from response_metadata
-     * This prevents LangChain's merge conflicts when the same index is used for
-     * different content types (text vs tool calls)
+     * Get the model ID to use for API calls.
+     * Returns applicationInferenceProfile if set, otherwise returns this.model.
+     */
+    protected getModelId(): string;
+    /**
+     * Override invocationParams to add serviceTier support.
+     */
+    invocationParams(options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions): ReturnType<ChatBedrockConverse['invocationParams']> & {
+        serviceTier?: {
+            type: ServiceTierType;
+        };
+    };
+    /**
+     * Override _generateNonStreaming to use applicationInferenceProfile as modelId.
+     * Uses the same model-swapping pattern as streaming for consistency.
+     */
+    _generateNonStreaming(messages: BaseMessage[], options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions, runManager?: CallbackManagerForLLMRun): Promise<ChatResult>;
+    /**
+     * Override _streamResponseChunks to:
+     * 1. Use applicationInferenceProfile as modelId (by temporarily swapping this.model)
+     * 2. Strip contentBlockIndex from response_metadata to prevent merge conflicts
+     *
+     * Note: We delegate to super._streamResponseChunks() to preserve @langchain/aws's
+     * internal chunk handling which correctly preserves array content for reasoning blocks.
+     */
+    _streamResponseChunks(messages: BaseMessage[], options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions, runManager?: CallbackManagerForLLMRun): AsyncGenerator<ChatGenerationChunk>;
+    /**
+     * Clean a chunk by removing contentBlockIndex from response_metadata.
      */
-    _streamResponseChunks(messages: BaseMessage[], options: this['ParsedCallOptions'], runManager?: CallbackManagerForLLMRun): AsyncGenerator<ChatGenerationChunk>;
+    private cleanChunk;
     /**
      * Check if contentBlockIndex exists at any level in the object
      */

package/dist/types/llm/bedrock/types.d.ts

@@ -0,0 +1,27 @@
+/**
+ * Type definitions for Bedrock Converse utilities.
+ */
+import type { Message as BedrockMessage, SystemContentBlock as BedrockSystemContentBlock, ContentBlock as BedrockContentBlock, ConverseResponse, ContentBlockDeltaEvent, ConverseStreamMetadataEvent, ContentBlockStartEvent, ReasoningContentBlock, ReasoningContentBlockDelta } from '@aws-sdk/client-bedrock-runtime';
+/**
+ * Reasoning content block type for LangChain messages.
+ */
+export interface MessageContentReasoningBlock {
+    type: 'reasoning_content';
+    reasoningText?: {
+        text?: string;
+        signature?: string;
+    };
+    redactedContent?: string;
+}
+export interface MessageContentReasoningBlockReasoningTextPartial {
+    type: 'reasoning_content';
+    reasoningText: {
+        text?: string;
+        signature?: string;
+    };
+}
+export interface MessageContentReasoningBlockRedacted {
+    type: 'reasoning_content';
+    redactedContent: string;
+}
+export type { BedrockMessage, BedrockSystemContentBlock, BedrockContentBlock, ConverseResponse, ContentBlockDeltaEvent, ConverseStreamMetadataEvent, ContentBlockStartEvent, ReasoningContentBlock, ReasoningContentBlockDelta, };

package/dist/types/llm/bedrock/utils/index.d.ts

@@ -0,0 +1,5 @@
+/**
+ * Bedrock Converse utility exports.
+ */
+export { convertToConverseMessages, extractImageInfo, langchainReasoningBlockToBedrockReasoningBlock, concatenateLangchainReasoningBlocks, } from './message_inputs';
+export { convertConverseMessageToLangChainMessage, handleConverseStreamContentBlockStart, handleConverseStreamContentBlockDelta, handleConverseStreamMetadata, bedrockReasoningBlockToLangchainReasoningBlock, bedrockReasoningDeltaToLangchainPartialReasoningBlock, } from './message_outputs';

package/dist/types/llm/bedrock/utils/message_inputs.d.ts

@@ -0,0 +1,31 @@
+/**
+ * Utility functions for converting LangChain messages to Bedrock Converse messages.
+ * Ported from @langchain/aws common.js
+ */
+import { type BaseMessage, MessageContentComplex } from '@langchain/core/messages';
+import type { BedrockMessage, BedrockSystemContentBlock, BedrockContentBlock, MessageContentReasoningBlock } from '../types';
+/**
+ * Convert a LangChain reasoning block to a Bedrock reasoning block.
+ */
+export declare function langchainReasoningBlockToBedrockReasoningBlock(content: MessageContentReasoningBlock): {
+    reasoningText?: {
+        text?: string;
+        signature?: string;
+    };
+    redactedContent?: Uint8Array;
+};
+/**
+ * Concatenate consecutive reasoning blocks in content array.
+ */
+export declare function concatenateLangchainReasoningBlocks(content: Array<MessageContentComplex | MessageContentReasoningBlock>): Array<MessageContentComplex | MessageContentReasoningBlock>;
+/**
+ * Extract image info from a base64 string or URL.
+ */
+export declare function extractImageInfo(base64: string): BedrockContentBlock;
+/**
+ * Convert LangChain messages to Bedrock Converse messages.
+ */
+export declare function convertToConverseMessages(messages: BaseMessage[]): {
+    converseMessages: BedrockMessage[];
+    converseSystem: BedrockSystemContentBlock[];
+};

package/dist/types/llm/bedrock/utils/message_outputs.d.ts

@@ -0,0 +1,33 @@
+/**
+ * Utility functions for converting Bedrock Converse responses to LangChain messages.
+ * Ported from @langchain/aws common.js
+ */
+import { AIMessage } 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 declare function bedrockReasoningDeltaToLangchainPartialReasoningBlock(reasoningContent: ReasoningContentBlockDelta): MessageContentReasoningBlockReasoningTextPartial | MessageContentReasoningBlockRedacted;
+/**
+ * Convert a Bedrock reasoning block to a LangChain reasoning block.
+ */
+export declare function bedrockReasoningBlockToLangchainReasoningBlock(reasoningContent: ReasoningContentBlock): MessageContentReasoningBlock;
+/**
+ * Convert a Bedrock Converse message to a LangChain message.
+ */
+export declare function convertConverseMessageToLangChainMessage(message: BedrockMessage, responseMetadata: Omit<ConverseResponse, 'output'>): AIMessage;
+/**
+ * Handle a content block delta event from Bedrock Converse stream.
+ */
+export declare function handleConverseStreamContentBlockDelta(contentBlockDelta: ContentBlockDeltaEvent): ChatGenerationChunk;
+/**
+ * Handle a content block start event from Bedrock Converse stream.
+ */
+export declare function handleConverseStreamContentBlockStart(contentBlockStart: ContentBlockStartEvent): ChatGenerationChunk | null;
+/**
+ * Handle a metadata event from Bedrock Converse stream.
+ */
+export declare function handleConverseStreamMetadata(metadata: ConverseStreamMetadataEvent, extra: {
+    streamUsage: boolean;
+}): ChatGenerationChunk;

package/dist/types/types/tools.d.ts

@@ -57,6 +57,8 @@ 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[];
 export type ExecuteResult = {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "3.0.776",
+  "version": "3.1.00",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",
@@ -51,12 +51,14 @@
     "image": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/image.ts --provider 'google' --name 'Jo' --location 'New York, NY'",
     "code_exec_files": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/code_exec_files.ts --provider 'openAI' --name 'Jo' --location 'New York, NY'",
     "code_exec_session": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/code_exec_session.ts --provider 'openAI' --name 'Jo' --location 'New York, NY'",
+    "code_exec_multi_session": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/code_exec_multi_session.ts --provider 'openAI' --name 'Jo' --location 'New York, NY'",
     "code_exec_simple": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/code_exec_simple.ts --provider 'openAI' --name 'Jo' --location 'New York, NY'",
     "simple": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/simple.ts --provider 'bedrock' --name 'Jo' --location 'New York, NY'",
     "caching": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/caching.ts --name 'Jo' --location 'New York, NY'",
     "thinking": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/thinking.ts --name 'Jo' --location 'New York, NY'",
+    "thinking:bedrock": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/thinking-bedrock.ts --name 'Jo' --location 'New York, NY'",
     "memory": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/memory.ts --provider 'openAI' --name 'Jo' --location 'New York, NY'",
-    "tool": "node --trace-warnings -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/tools.ts --provider 'openrouter' --name 'Jo' --location 'New York, NY'",
+    "tool": "node --trace-warnings -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/tools.ts --provider 'bedrock' --name 'Jo' --location 'New York, NY'",
     "search": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/search.ts --provider 'bedrock' --name 'Jo' --location 'New York, NY'",
     "tool_search": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/tool_search.ts",
     "programmatic_exec": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/programmatic_exec.ts",
@@ -110,6 +112,7 @@
     }
   },
   "dependencies": {
+    "@aws-sdk/client-bedrock-runtime": "^3.970.0",
     "@langchain/anthropic": "^0.3.26",
     "@langchain/aws": "^0.1.15",
     "@langchain/core": "^0.3.80",

package/src/graphs/Graph.ts

@@ -1042,8 +1042,8 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
 
     /**
      * Extract and store code execution session context from artifacts.
-     * Only update session_id when files are generated - this ensures we don't
-     * lose the original session that contains the files.
+     * Each file is stamped with its source session_id to support multi-session file tracking.
+     * When the same filename appears in a later execution, the newer version replaces the old.
      */
     const toolName = output.name;
     if (
@@ -1060,17 +1060,35 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
         artifact.session_id !== ''
       ) {
         /**
-         * Files were generated - update session with the new session_id.
-         * The new session_id is the one that contains these files.
+         * Stamp each new file with its source session_id.
+         * This enables files from different executions (parallel or sequential)
+         * to be tracked and passed to subsequent calls.
          */
+        const filesWithSession: t.FileRefs = newFiles.map((file) => ({
+          ...file,
+          session_id: artifact.session_id,
+        }));
+
         const existingSession = this.sessions.get(Constants.EXECUTE_CODE) as
           | t.CodeSessionContext
           | undefined;
         const existingFiles = existingSession?.files ?? [];
 
+        /**
+         * Merge files, preferring latest versions by name.
+         * If a file with the same name exists, replace it with the new version.
+         * This handles cases where files are edited/recreated in subsequent executions.
+         */
+        const newFileNames = new Set(filesWithSession.map((f) => f.name));
+        const filteredExisting = existingFiles.filter(
+          (f) => !newFileNames.has(f.name)
+        );
+
         this.sessions.set(Constants.EXECUTE_CODE, {
+          /** Keep latest session_id for reference/fallback */
           session_id: artifact.session_id,
-          files: [...existingFiles, ...newFiles],
+          /** Accumulated files with latest versions preferred */
+          files: [...filteredExisting, ...filesWithSession],
           lastUpdated: Date.now(),
         });
       }

package/src/llm/bedrock/index.ts

@@ -1,5 +1,9 @@
 /**
  * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts
+ * and adds support for latest @langchain/aws features:
+ *
+ * - Application Inference Profiles (PR #9129)
+ * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+
  *
  * Bedrock sends the same contentBlockIndex for both text and tool_use content blocks,
  * causing LangChain's merge logic to fail with "field[contentBlockIndex] already exists"
@@ -12,15 +16,74 @@
  */
 
 import { ChatBedrockConverse } from '@langchain/aws';
-import type { ChatBedrockConverseInput } from '@langchain/aws';
 import { AIMessageChunk } from '@langchain/core/messages';
-import type { BaseMessage } from '@langchain/core/messages';
-import { ChatGenerationChunk } from '@langchain/core/outputs';
+import { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';
 import type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';
+import type { ChatBedrockConverseInput } from '@langchain/aws';
+import type { BaseMessage } from '@langchain/core/messages';
+
+/**
+ * Service tier type for Bedrock invocations.
+ * Requires AWS SDK >= 3.966.0 to actually work.
+ * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
+ */
+export type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';
+
+/**
+ * Extended input interface with additional features:
+ * - applicationInferenceProfile: Use an inference profile ARN instead of model ID
+ * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)
+ */
+export interface CustomChatBedrockConverseInput
+  extends ChatBedrockConverseInput {
+  /**
+   * Application Inference Profile ARN to use for the model.
+   * For example, "arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx"
+   * When provided, this ARN will be used for the actual inference calls instead of the model ID.
+   * Must still provide `model` as normal modelId to benefit from all the metadata.
+   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html
+   */
+  applicationInferenceProfile?: string;
+
+  /**
+   * Service tier for model invocation.
+   * Specifies the processing tier type used for serving the request.
+   * Supported values are 'priority', 'default', 'flex', and 'reserved'.
+   *
+   * - 'priority': Prioritized processing for lower latency
+   * - 'default': Standard processing tier
+   * - 'flex': Flexible processing tier with lower cost
+   * - 'reserved': Reserved capacity for consistent performance
+   *
+   * If not provided, AWS uses the default tier.
+   * Note: Requires AWS SDK >= 3.966.0 to work.
+   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
+   */
+  serviceTier?: ServiceTierType;
+}
+
+/**
+ * Extended call options with serviceTier override support.
+ */
+export interface CustomChatBedrockConverseCallOptions {
+  serviceTier?: ServiceTierType;
+}
 
 export class CustomChatBedrockConverse extends ChatBedrockConverse {
-  constructor(fields?: ChatBedrockConverseInput) {
+  /**
+   * Application Inference Profile ARN to use instead of model ID.
+   */
+  applicationInferenceProfile?: string;
+
+  /**
+   * Service tier for model invocation.
+   */
+  serviceTier?: ServiceTierType;
+
+  constructor(fields?: CustomChatBedrockConverseInput) {
     super(fields);
+    this.applicationInferenceProfile = fields?.applicationInferenceProfile;
+    this.serviceTier = fields?.serviceTier;
   }
 
   static lc_name(): string {
@@ -28,52 +91,126 @@ export class CustomChatBedrockConverse extends ChatBedrockConverse {
   }
 
   /**
-   * Override _streamResponseChunks to strip contentBlockIndex from response_metadata
-   * This prevents LangChain's merge conflicts when the same index is used for
-   * different content types (text vs tool calls)
+   * Get the model ID to use for API calls.
+   * Returns applicationInferenceProfile if set, otherwise returns this.model.
+   */
+  protected getModelId(): string {
+    return this.applicationInferenceProfile ?? this.model;
+  }
+
+  /**
+   * Override invocationParams to add serviceTier support.
+   */
+  override invocationParams(
+    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions
+  ): ReturnType<ChatBedrockConverse['invocationParams']> & {
+    serviceTier?: { type: ServiceTierType };
+  } {
+    const baseParams = super.invocationParams(options);
+
+    /** Service tier from options or fall back to class-level setting */
+    const serviceTierType = options?.serviceTier ?? this.serviceTier;
+
+    return {
+      ...baseParams,
+      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,
+    };
+  }
+
+  /**
+   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.
+   * Uses the same model-swapping pattern as streaming for consistency.
    */
-  async *_streamResponseChunks(
+  override async _generateNonStreaming(
     messages: BaseMessage[],
-    options: this['ParsedCallOptions'],
+    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,
+    runManager?: CallbackManagerForLLMRun
+  ): Promise<ChatResult> {
+    // Temporarily swap model for applicationInferenceProfile support
+    const originalModel = this.model;
+    if (
+      this.applicationInferenceProfile != null &&
+      this.applicationInferenceProfile !== ''
+    ) {
+      this.model = this.applicationInferenceProfile;
+    }
+
+    try {
+      return await super._generateNonStreaming(messages, options, runManager);
+    } finally {
+      // Restore original model
+      this.model = originalModel;
+    }
+  }
+
+  /**
+   * Override _streamResponseChunks to:
+   * 1. Use applicationInferenceProfile as modelId (by temporarily swapping this.model)
+   * 2. Strip contentBlockIndex from response_metadata to prevent merge conflicts
+   *
+   * Note: We delegate to super._streamResponseChunks() to preserve @langchain/aws's
+   * internal chunk handling which correctly preserves array content for reasoning blocks.
+   */
+  override async *_streamResponseChunks(
+    messages: BaseMessage[],
+    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,
     runManager?: CallbackManagerForLLMRun
   ): AsyncGenerator<ChatGenerationChunk> {
-    const baseStream = super._streamResponseChunks(
-      messages,
-      options,
-      runManager
-    );
-
-    for await (const chunk of baseStream) {
-      // Only process if we have response_metadata
-      if (
-        chunk.message instanceof AIMessageChunk &&
-        (chunk.message as Partial<AIMessageChunk>).response_metadata &&
-        typeof chunk.message.response_metadata === 'object'
-      ) {
-        // Check if contentBlockIndex exists anywhere in response_metadata (top level or nested)
-        const hasContentBlockIndex = this.hasContentBlockIndex(
-          chunk.message.response_metadata
-        );
-
-        if (hasContentBlockIndex) {
-          const cleanedMetadata = this.removeContentBlockIndex(
-            chunk.message.response_metadata
-          ) as Record<string, unknown>;
-
-          yield new ChatGenerationChunk({
-            text: chunk.text,
-            message: new AIMessageChunk({
-              ...chunk.message,
-              response_metadata: cleanedMetadata,
-            }),
-            generationInfo: chunk.generationInfo,
-          });
-          continue;
-        }
+    // Temporarily swap model for applicationInferenceProfile support
+    const originalModel = this.model;
+    if (
+      this.applicationInferenceProfile != null &&
+      this.applicationInferenceProfile !== ''
+    ) {
+      this.model = this.applicationInferenceProfile;
+    }
+
+    try {
+      // Use parent's streaming logic which correctly handles reasoning content
+      const baseStream = super._streamResponseChunks(
+        messages,
+        options,
+        runManager
+      );
+
+      for await (const chunk of baseStream) {
+        // Clean contentBlockIndex from response_metadata to prevent merge conflicts
+        yield this.cleanChunk(chunk);
       }
+    } finally {
+      // Restore original model
+      this.model = originalModel;
+    }
+  }
 
-      yield chunk;
+  /**
+   * Clean a chunk by removing contentBlockIndex from response_metadata.
+   */
+  private cleanChunk(chunk: ChatGenerationChunk): ChatGenerationChunk {
+    const message = chunk.message;
+    if (!(message instanceof AIMessageChunk)) {
+      return chunk;
     }
+
+    const metadata = message.response_metadata as Record<string, unknown>;
+    const hasContentBlockIndex = this.hasContentBlockIndex(metadata);
+    if (!hasContentBlockIndex) {
+      return chunk;
+    }
+
+    const cleanedMetadata = this.removeContentBlockIndex(metadata) as Record<
+      string,
+      unknown
+    >;
+
+    return new ChatGenerationChunk({
+      text: chunk.text,
+      message: new AIMessageChunk({
+        ...message,
+        response_metadata: cleanedMetadata,
+      }),
+      generationInfo: chunk.generationInfo,
+    });
   }
 
   /**