@librechat/agents 3.0.79 → 3.0.80

package/dist/cjs/llm/bedrock/index.cjs

@@ -3,9 +3,6 @@
 var aws = require('@langchain/aws');
 var messages = require('@langchain/core/messages');
 var outputs = require('@langchain/core/outputs');
-var clientBedrockRuntime = require('@aws-sdk/client-bedrock-runtime');
-var message_inputs = require('./utils/message_inputs.cjs');
-var message_outputs = require('./utils/message_outputs.cjs');
 
 /**
  * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts
@@ -52,7 +49,7 @@ class CustomChatBedrockConverse extends aws.ChatBedrockConverse {
      */
     invocationParams(options) {
         const baseParams = super.invocationParams(options);
-        // Get serviceTier from options or fall back to class-level setting
+        /** Service tier from options or fall back to class-level setting */
         const serviceTierType = options?.serviceTier ?? this.serviceTier;
         return {
             ...baseParams,
@@ -61,86 +58,50 @@ class CustomChatBedrockConverse extends aws.ChatBedrockConverse {
     }
     /**
      * Override _generateNonStreaming to use applicationInferenceProfile as modelId.
+     * Uses the same model-swapping pattern as streaming for consistency.
      */
-    async _generateNonStreaming(messages, options, _runManager) {
-        const { converseMessages, converseSystem } = message_inputs.convertToConverseMessages(messages);
-        const params = this.invocationParams(options);
-        const command = new clientBedrockRuntime.ConverseCommand({
-            modelId: this.getModelId(),
-            messages: converseMessages,
-            system: converseSystem,
-            requestMetadata: options.requestMetadata,
-            ...params,
-        });
-        const response = await this.client.send(command, {
-            abortSignal: options.signal,
-        });
-        const { output, ...responseMetadata } = response;
-        if (!output?.message) {
-            throw new Error('No message found in Bedrock response.');
+    async _generateNonStreaming(messages, options, runManager) {
+        // 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;
         }
-        const message = message_outputs.convertConverseMessageToLangChainMessage(output.message, responseMetadata);
-        return {
-            generations: [
-                {
-                    text: typeof message.content === 'string' ? message.content : '',
-                    message,
-                },
-            ],
-        };
     }
     /**
      * Override _streamResponseChunks to:
-     * 1. Use applicationInferenceProfile as modelId
-     * 2. Include serviceTier in request
-     * 3. Strip contentBlockIndex from response_metadata to prevent merge conflicts
+     * 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.
      */
     async *_streamResponseChunks(messages, options, runManager) {
-        const { converseMessages, converseSystem } = message_inputs.convertToConverseMessages(messages);
-        const params = this.invocationParams(options);
-        let { streamUsage } = this;
-        if (options.streamUsage !== undefined) {
-            streamUsage = options.streamUsage;
+        // Temporarily swap model for applicationInferenceProfile support
+        const originalModel = this.model;
+        if (this.applicationInferenceProfile != null &&
+            this.applicationInferenceProfile !== '') {
+            this.model = this.applicationInferenceProfile;
         }
-        const command = new clientBedrockRuntime.ConverseStreamCommand({
-            modelId: this.getModelId(),
-            messages: converseMessages,
-            system: converseSystem,
-            requestMetadata: options.requestMetadata,
-            ...params,
-        });
-        const response = await this.client.send(command, {
-            abortSignal: options.signal,
-        });
-        if (response.stream) {
-            for await (const event of response.stream) {
-                if (event.contentBlockStart != null) {
-                    const chunk = message_outputs.handleConverseStreamContentBlockStart(event.contentBlockStart);
-                    if (chunk !== undefined) {
-                        const cleanedChunk = this.cleanChunk(chunk);
-                        yield cleanedChunk;
-                        await runManager?.handleLLMNewToken(cleanedChunk.text || '');
-                    }
-                }
-                else if (event.contentBlockDelta != null) {
-                    const chunk = message_outputs.handleConverseStreamContentBlockDelta(event.contentBlockDelta);
-                    if (chunk !== undefined) {
-                        const cleanedChunk = this.cleanChunk(chunk);
-                        yield cleanedChunk;
-                        await runManager?.handleLLMNewToken(cleanedChunk.text || '');
-                    }
-                }
-                else if (event.metadata != null) {
-                    const chunk = message_outputs.handleConverseStreamMetadata(event.metadata, {
-                        streamUsage,
-                    });
-                    if (chunk !== undefined) {
-                        const cleanedChunk = this.cleanChunk(chunk);
-                        yield cleanedChunk;
-                    }
-                }
+        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;
+        }
     }
     /**
      * Clean a chunk by removing contentBlockIndex from response_metadata.

package/dist/cjs/llm/bedrock/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts\n * and adds support for latest @langchain/aws features:\n *\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock sends the same contentBlockIndex for both text and tool_use content blocks,\n * causing LangChain's merge logic to fail with \"field[contentBlockIndex] already exists\"\n * errors. This wrapper simply strips contentBlockIndex from response_metadata to avoid\n * the conflict.\n *\n * The contentBlockIndex field is only used internally by Bedrock's streaming protocol\n * and isn't needed by application logic - the index field on tool_call_chunks serves\n * the purpose of tracking tool call ordering.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport {\n  ConverseCommand,\n  ConverseStreamCommand,\n} from '@aws-sdk/client-bedrock-runtime';\nimport {\n  convertToConverseMessages,\n  convertConverseMessageToLangChainMessage,\n  handleConverseStreamContentBlockStart,\n  handleConverseStreamContentBlockDelta,\n  handleConverseStreamMetadata,\n} from './utils';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\n/**\n * Extended input interface with additional features:\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  /**\n   * Application Inference Profile ARN to use instead of model ID.\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   */\n  serviceTier?: ServiceTierType;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n  }\n\n  static lc_name(): string {\n    return 'LibreChatBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to add serviceTier support.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const baseParams = super.invocationParams(options);\n\n    // Get serviceTier from options or fall back to class-level setting\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...baseParams,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    _runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    const { converseMessages, converseSystem } =\n      convertToConverseMessages(messages);\n    const params = this.invocationParams(options);\n\n    const command = new ConverseCommand({\n      modelId: this.getModelId(),\n      messages: converseMessages,\n      system: converseSystem,\n      requestMetadata: options.requestMetadata,\n      ...params,\n    });\n\n    const response = await this.client.send(command, {\n      abortSignal: options.signal,\n    });\n\n    const { output, ...responseMetadata } = response;\n    if (!output?.message) {\n      throw new Error('No message found in Bedrock response.');\n    }\n\n    const message = convertConverseMessageToLangChainMessage(\n      output.message,\n      responseMetadata\n    );\n\n    return {\n      generations: [\n        {\n          text: typeof message.content === 'string' ? message.content : '',\n          message,\n        },\n      ],\n    };\n  }\n\n  /**\n   * Override _streamResponseChunks to:\n   * 1. Use applicationInferenceProfile as modelId\n   * 2. Include serviceTier in request\n   * 3. Strip contentBlockIndex from response_metadata to prevent merge conflicts\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const { converseMessages, converseSystem } =\n      convertToConverseMessages(messages);\n    const params = this.invocationParams(options);\n\n    let { streamUsage } = this;\n    if (options.streamUsage !== undefined) {\n      streamUsage = options.streamUsage;\n    }\n\n    const command = new ConverseStreamCommand({\n      modelId: this.getModelId(),\n      messages: converseMessages,\n      system: converseSystem,\n      requestMetadata: options.requestMetadata,\n      ...params,\n    });\n\n    const response = await this.client.send(command, {\n      abortSignal: options.signal,\n    });\n\n    if (response.stream) {\n      for await (const event of response.stream) {\n        if (event.contentBlockStart != null) {\n          const chunk = handleConverseStreamContentBlockStart(\n            event.contentBlockStart\n          ) as ChatGenerationChunk | undefined;\n          if (chunk !== undefined) {\n            const cleanedChunk = this.cleanChunk(chunk);\n            yield cleanedChunk;\n            await runManager?.handleLLMNewToken(cleanedChunk.text || '');\n          }\n        } else if (event.contentBlockDelta != null) {\n          const chunk = handleConverseStreamContentBlockDelta(\n            event.contentBlockDelta\n          ) as ChatGenerationChunk | undefined;\n          if (chunk !== undefined) {\n            const cleanedChunk = this.cleanChunk(chunk);\n            yield cleanedChunk;\n            await runManager?.handleLLMNewToken(cleanedChunk.text || '');\n          }\n        } else if (event.metadata != null) {\n          const chunk = handleConverseStreamMetadata(event.metadata, {\n            streamUsage,\n          }) as ChatGenerationChunk | undefined;\n          if (chunk !== undefined) {\n            const cleanedChunk = this.cleanChunk(chunk);\n            yield cleanedChunk;\n          }\n        }\n      }\n    }\n  }\n\n  /**\n   * Clean a chunk by removing contentBlockIndex from response_metadata.\n   */\n  private cleanChunk(chunk: ChatGenerationChunk): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const metadata = message.response_metadata as Record<string, unknown>;\n    const hasContentBlockIndex = this.hasContentBlockIndex(metadata);\n    if (!hasContentBlockIndex) {\n      return chunk;\n    }\n\n    const cleanedMetadata = this.removeContentBlockIndex(metadata) as Record<\n      string,\n      unknown\n    >;\n\n    return new ChatGenerationChunk({\n      text: chunk.text,\n      message: new AIMessageChunk({\n        ...message,\n        response_metadata: cleanedMetadata,\n      }),\n      generationInfo: chunk.generationInfo,\n    });\n  }\n\n  /**\n   * Check if contentBlockIndex exists at any level in the object\n   */\n  private hasContentBlockIndex(obj: unknown): boolean {\n    if (obj === null || obj === undefined || typeof obj !== 'object') {\n      return false;\n    }\n\n    if ('contentBlockIndex' in obj) {\n      return true;\n    }\n\n    for (const value of Object.values(obj)) {\n      if (typeof value === 'object' && value !== null) {\n        if (this.hasContentBlockIndex(value)) {\n          return true;\n        }\n      }\n    }\n\n    return false;\n  }\n\n  /**\n   * Recursively remove contentBlockIndex from all levels of an object\n   */\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"names":["ChatBedrockConverse","convertToConverseMessages","ConverseCommand","convertConverseMessageToLangChainMessage","ConverseStreamCommand","handleConverseStreamContentBlockStart","handleConverseStreamContentBlockDelta","handleConverseStreamMetadata","AIMessageChunk","ChatGenerationChunk"],"mappings":";;;;;;;;;AAAA;;;;;;;;;;;;;;;AAeG;AAmEG,MAAO,yBAA0B,SAAQA,uBAAmB,CAAA;AAChE;;AAEG;AACH,IAAA,2BAA2B;AAE3B;;AAEG;AACH,IAAA,WAAW;AAEX,IAAA,WAAA,CAAY,MAAuC,EAAA;QACjD,KAAK,CAAC,MAAM,CAAC;AACb,QAAA,IAAI,CAAC,2BAA2B,GAAG,MAAM,EAAE,2BAA2B;AACtE,QAAA,IAAI,CAAC,WAAW,GAAG,MAAM,EAAE,WAAW;;AAGxC,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,0BAA0B;;AAGnC;;;AAGG;IACO,UAAU,GAAA;AAClB,QAAA,OAAO,IAAI,CAAC,2BAA2B,IAAI,IAAI,CAAC,KAAK;;AAGvD;;AAEG;AACM,IAAA,gBAAgB,CACvB,OAA0E,EAAA;QAI1E,MAAM,UAAU,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;;QAGlD,MAAM,eAAe,GAAG,OAAO,EAAE,WAAW,IAAI,IAAI,CAAC,WAAW;QAEhE,OAAO;AACL,YAAA,GAAG,UAAU;AACb,YAAA,WAAW,EAAE,eAAe,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE,GAAG,SAAS;SACrE;;AAGH;;AAEG;AACM,IAAA,MAAM,qBAAqB,CAClC,QAAuB,EACvB,OAAyE,EACzE,WAAsC,EAAA;QAEtC,MAAM,EAAE,gBAAgB,EAAE,cAAc,EAAE,GACxCC,wCAAyB,CAAC,QAAQ,CAAC;QACrC,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC;AAE7C,QAAA,MAAM,OAAO,GAAG,IAAIC,oCAAe,CAAC;AAClC,YAAA,OAAO,EAAE,IAAI,CAAC,UAAU,EAAE;AAC1B,YAAA,QAAQ,EAAE,gBAAgB;AAC1B,YAAA,MAAM,EAAE,cAAc;YACtB,eAAe,EAAE,OAAO,CAAC,eAAe;AACxC,YAAA,GAAG,MAAM;AACV,SAAA,CAAC;QAEF,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,EAAE;YAC/C,WAAW,EAAE,OAAO,CAAC,MAAM;AAC5B,SAAA,CAAC;QAEF,MAAM,EAAE,MAAM,EAAE,GAAG,gBAAgB,EAAE,GAAG,QAAQ;AAChD,QAAA,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE;AACpB,YAAA,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC;;QAG1D,MAAM,OAAO,GAAGC,wDAAwC,CACtD,MAAM,CAAC,OAAO,EACd,gBAAgB,CACjB;QAED,OAAO;AACL,YAAA,WAAW,EAAE;AACX,gBAAA;AACE,oBAAA,IAAI,EAAE,OAAO,OAAO,CAAC,OAAO,KAAK,QAAQ,GAAG,OAAO,CAAC,OAAO,GAAG,EAAE;oBAChE,OAAO;AACR,iBAAA;AACF,aAAA;SACF;;AAGH;;;;;AAKG;IACM,OAAO,qBAAqB,CACnC,QAAuB,EACvB,OAAyE,EACzE,UAAqC,EAAA;QAErC,MAAM,EAAE,gBAAgB,EAAE,cAAc,EAAE,GACxCF,wCAAyB,CAAC,QAAQ,CAAC;QACrC,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC;AAE7C,QAAA,IAAI,EAAE,WAAW,EAAE,GAAG,IAAI;AAC1B,QAAA,IAAI,OAAO,CAAC,WAAW,KAAK,SAAS,EAAE;AACrC,YAAA,WAAW,GAAG,OAAO,CAAC,WAAW;;AAGnC,QAAA,MAAM,OAAO,GAAG,IAAIG,0CAAqB,CAAC;AACxC,YAAA,OAAO,EAAE,IAAI,CAAC,UAAU,EAAE;AAC1B,YAAA,QAAQ,EAAE,gBAAgB;AAC1B,YAAA,MAAM,EAAE,cAAc;YACtB,eAAe,EAAE,OAAO,CAAC,eAAe;AACxC,YAAA,GAAG,MAAM;AACV,SAAA,CAAC;QAEF,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,EAAE;YAC/C,WAAW,EAAE,OAAO,CAAC,MAAM;AAC5B,SAAA,CAAC;AAEF,QAAA,IAAI,QAAQ,CAAC,MAAM,EAAE;YACnB,WAAW,MAAM,KAAK,IAAI,QAAQ,CAAC,MAAM,EAAE;AACzC,gBAAA,IAAI,KAAK,CAAC,iBAAiB,IAAI,IAAI,EAAE;oBACnC,MAAM,KAAK,GAAGC,qDAAqC,CACjD,KAAK,CAAC,iBAAiB,CACW;AACpC,oBAAA,IAAI,KAAK,KAAK,SAAS,EAAE;wBACvB,MAAM,YAAY,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;AAC3C,wBAAA,MAAM,YAAY;wBAClB,MAAM,UAAU,EAAE,iBAAiB,CAAC,YAAY,CAAC,IAAI,IAAI,EAAE,CAAC;;;AAEzD,qBAAA,IAAI,KAAK,CAAC,iBAAiB,IAAI,IAAI,EAAE;oBAC1C,MAAM,KAAK,GAAGC,qDAAqC,CACjD,KAAK,CAAC,iBAAiB,CACW;AACpC,oBAAA,IAAI,KAAK,KAAK,SAAS,EAAE;wBACvB,MAAM,YAAY,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;AAC3C,wBAAA,MAAM,YAAY;wBAClB,MAAM,UAAU,EAAE,iBAAiB,CAAC,YAAY,CAAC,IAAI,IAAI,EAAE,CAAC;;;AAEzD,qBAAA,IAAI,KAAK,CAAC,QAAQ,IAAI,IAAI,EAAE;AACjC,oBAAA,MAAM,KAAK,GAAGC,4CAA4B,CAAC,KAAK,CAAC,QAAQ,EAAE;wBACzD,WAAW;AACZ,qBAAA,CAAoC;AACrC,oBAAA,IAAI,KAAK,KAAK,SAAS,EAAE;wBACvB,MAAM,YAAY,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;AAC3C,wBAAA,MAAM,YAAY;;;;;;AAO5B;;AAEG;AACK,IAAA,UAAU,CAAC,KAA0B,EAAA;AAC3C,QAAA,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO;AAC7B,QAAA,IAAI,EAAE,OAAO,YAAYC,uBAAc,CAAC,EAAE;AACxC,YAAA,OAAO,KAAK;;AAGd,QAAA,MAAM,QAAQ,GAAG,OAAO,CAAC,iBAA4C;QACrE,MAAM,oBAAoB,GAAG,IAAI,CAAC,oBAAoB,CAAC,QAAQ,CAAC;QAChE,IAAI,CAAC,oBAAoB,EAAE;AACzB,YAAA,OAAO,KAAK;;QAGd,MAAM,eAAe,GAAG,IAAI,CAAC,uBAAuB,CAAC,QAAQ,CAG5D;QAED,OAAO,IAAIC,2BAAmB,CAAC;YAC7B,IAAI,EAAE,KAAK,CAAC,IAAI;YAChB,OAAO,EAAE,IAAID,uBAAc,CAAC;AAC1B,gBAAA,GAAG,OAAO;AACV,gBAAA,iBAAiB,EAAE,eAAe;aACnC,CAAC;YACF,cAAc,EAAE,KAAK,CAAC,cAAc;AACrC,SAAA,CAAC;;AAGJ;;AAEG;AACK,IAAA,oBAAoB,CAAC,GAAY,EAAA;AACvC,QAAA,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;AAChE,YAAA,OAAO,KAAK;;AAGd,QAAA,IAAI,mBAAmB,IAAI,GAAG,EAAE;AAC9B,YAAA,OAAO,IAAI;;QAGb,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE;YACtC,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,EAAE;AAC/C,gBAAA,IAAI,IAAI,CAAC,oBAAoB,CAAC,KAAK,CAAC,EAAE;AACpC,oBAAA,OAAO,IAAI;;;;AAKjB,QAAA,OAAO,KAAK;;AAGd;;AAEG;AACK,IAAA,uBAAuB,CAAC,GAAY,EAAA;QAC1C,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,EAAE;AACrC,YAAA,OAAO,GAAG;;AAGZ,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AACtB,YAAA,OAAO,GAAG,CAAC,GAAG,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,uBAAuB,CAAC,IAAI,CAAC,CAAC;;AAG9D,QAAA,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;YAC3B,MAAM,OAAO,GAA4B,EAAE;AAC3C,YAAA,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AAC9C,gBAAA,IAAI,GAAG,KAAK,mBAAmB,EAAE;oBAC/B,OAAO,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,uBAAuB,CAAC,KAAK,CAAC;;;AAGtD,YAAA,OAAO,OAAO;;AAGhB,QAAA,OAAO,GAAG;;AAEb;;;;"}
+{"version":3,"file":"index.cjs","sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts\n * and adds support for latest @langchain/aws features:\n *\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock sends the same contentBlockIndex for both text and tool_use content blocks,\n * causing LangChain's merge logic to fail with \"field[contentBlockIndex] already exists\"\n * errors. This wrapper simply strips contentBlockIndex from response_metadata to avoid\n * the conflict.\n *\n * The contentBlockIndex field is only used internally by Bedrock's streaming protocol\n * and isn't needed by application logic - the index field on tool_call_chunks serves\n * the purpose of tracking tool call ordering.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport type { BaseMessage } from '@langchain/core/messages';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\n/**\n * Extended input interface with additional features:\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  /**\n   * Application Inference Profile ARN to use instead of model ID.\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   */\n  serviceTier?: ServiceTierType;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n  }\n\n  static lc_name(): string {\n    return 'LibreChatBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to add serviceTier support.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const baseParams = super.invocationParams(options);\n\n    /** Service tier from options or fall back to class-level setting */\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...baseParams,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   * Uses the same model-swapping pattern as streaming for consistency.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    // Temporarily swap model for applicationInferenceProfile support\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      return await super._generateNonStreaming(messages, options, runManager);\n    } finally {\n      // Restore original model\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Override _streamResponseChunks to:\n   * 1. Use applicationInferenceProfile as modelId (by temporarily swapping this.model)\n   * 2. Strip contentBlockIndex from response_metadata to prevent merge conflicts\n   *\n   * Note: We delegate to super._streamResponseChunks() to preserve @langchain/aws's\n   * internal chunk handling which correctly preserves array content for reasoning blocks.\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    // Temporarily swap model for applicationInferenceProfile support\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      // Use parent's streaming logic which correctly handles reasoning content\n      const baseStream = super._streamResponseChunks(\n        messages,\n        options,\n        runManager\n      );\n\n      for await (const chunk of baseStream) {\n        // Clean contentBlockIndex from response_metadata to prevent merge conflicts\n        yield this.cleanChunk(chunk);\n      }\n    } finally {\n      // Restore original model\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Clean a chunk by removing contentBlockIndex from response_metadata.\n   */\n  private cleanChunk(chunk: ChatGenerationChunk): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const metadata = message.response_metadata as Record<string, unknown>;\n    const hasContentBlockIndex = this.hasContentBlockIndex(metadata);\n    if (!hasContentBlockIndex) {\n      return chunk;\n    }\n\n    const cleanedMetadata = this.removeContentBlockIndex(metadata) as Record<\n      string,\n      unknown\n    >;\n\n    return new ChatGenerationChunk({\n      text: chunk.text,\n      message: new AIMessageChunk({\n        ...message,\n        response_metadata: cleanedMetadata,\n      }),\n      generationInfo: chunk.generationInfo,\n    });\n  }\n\n  /**\n   * Check if contentBlockIndex exists at any level in the object\n   */\n  private hasContentBlockIndex(obj: unknown): boolean {\n    if (obj === null || obj === undefined || typeof obj !== 'object') {\n      return false;\n    }\n\n    if ('contentBlockIndex' in obj) {\n      return true;\n    }\n\n    for (const value of Object.values(obj)) {\n      if (typeof value === 'object' && value !== null) {\n        if (this.hasContentBlockIndex(value)) {\n          return true;\n        }\n      }\n    }\n\n    return false;\n  }\n\n  /**\n   * Recursively remove contentBlockIndex from all levels of an object\n   */\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"names":["ChatBedrockConverse","AIMessageChunk","ChatGenerationChunk"],"mappings":";;;;;;AAAA;;;;;;;;;;;;;;;AAeG;AAwDG,MAAO,yBAA0B,SAAQA,uBAAmB,CAAA;AAChE;;AAEG;AACH,IAAA,2BAA2B;AAE3B;;AAEG;AACH,IAAA,WAAW;AAEX,IAAA,WAAA,CAAY,MAAuC,EAAA;QACjD,KAAK,CAAC,MAAM,CAAC;AACb,QAAA,IAAI,CAAC,2BAA2B,GAAG,MAAM,EAAE,2BAA2B;AACtE,QAAA,IAAI,CAAC,WAAW,GAAG,MAAM,EAAE,WAAW;;AAGxC,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,0BAA0B;;AAGnC;;;AAGG;IACO,UAAU,GAAA;AAClB,QAAA,OAAO,IAAI,CAAC,2BAA2B,IAAI,IAAI,CAAC,KAAK;;AAGvD;;AAEG;AACM,IAAA,gBAAgB,CACvB,OAA0E,EAAA;QAI1E,MAAM,UAAU,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;;QAGlD,MAAM,eAAe,GAAG,OAAO,EAAE,WAAW,IAAI,IAAI,CAAC,WAAW;QAEhE,OAAO;AACL,YAAA,GAAG,UAAU;AACb,YAAA,WAAW,EAAE,eAAe,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE,GAAG,SAAS;SACrE;;AAGH;;;AAGG;AACM,IAAA,MAAM,qBAAqB,CAClC,QAAuB,EACvB,OAAyE,EACzE,UAAqC,EAAA;;AAGrC,QAAA,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK;AAChC,QAAA,IACE,IAAI,CAAC,2BAA2B,IAAI,IAAI;AACxC,YAAA,IAAI,CAAC,2BAA2B,KAAK,EAAE,EACvC;AACA,YAAA,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,2BAA2B;;AAG/C,QAAA,IAAI;YACF,OAAO,MAAM,KAAK,CAAC,qBAAqB,CAAC,QAAQ,EAAE,OAAO,EAAE,UAAU,CAAC;;gBAC/D;;AAER,YAAA,IAAI,CAAC,KAAK,GAAG,aAAa;;;AAI9B;;;;;;;AAOG;IACM,OAAO,qBAAqB,CACnC,QAAuB,EACvB,OAAyE,EACzE,UAAqC,EAAA;;AAGrC,QAAA,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK;AAChC,QAAA,IACE,IAAI,CAAC,2BAA2B,IAAI,IAAI;AACxC,YAAA,IAAI,CAAC,2BAA2B,KAAK,EAAE,EACvC;AACA,YAAA,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,2BAA2B;;AAG/C,QAAA,IAAI;;AAEF,YAAA,MAAM,UAAU,GAAG,KAAK,CAAC,qBAAqB,CAC5C,QAAQ,EACR,OAAO,EACP,UAAU,CACX;AAED,YAAA,WAAW,MAAM,KAAK,IAAI,UAAU,EAAE;;AAEpC,gBAAA,MAAM,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;;;gBAEtB;;AAER,YAAA,IAAI,CAAC,KAAK,GAAG,aAAa;;;AAI9B;;AAEG;AACK,IAAA,UAAU,CAAC,KAA0B,EAAA;AAC3C,QAAA,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO;AAC7B,QAAA,IAAI,EAAE,OAAO,YAAYC,uBAAc,CAAC,EAAE;AACxC,YAAA,OAAO,KAAK;;AAGd,QAAA,MAAM,QAAQ,GAAG,OAAO,CAAC,iBAA4C;QACrE,MAAM,oBAAoB,GAAG,IAAI,CAAC,oBAAoB,CAAC,QAAQ,CAAC;QAChE,IAAI,CAAC,oBAAoB,EAAE;AACzB,YAAA,OAAO,KAAK;;QAGd,MAAM,eAAe,GAAG,IAAI,CAAC,uBAAuB,CAAC,QAAQ,CAG5D;QAED,OAAO,IAAIC,2BAAmB,CAAC;YAC7B,IAAI,EAAE,KAAK,CAAC,IAAI;YAChB,OAAO,EAAE,IAAID,uBAAc,CAAC;AAC1B,gBAAA,GAAG,OAAO;AACV,gBAAA,iBAAiB,EAAE,eAAe;aACnC,CAAC;YACF,cAAc,EAAE,KAAK,CAAC,cAAc;AACrC,SAAA,CAAC;;AAGJ;;AAEG;AACK,IAAA,oBAAoB,CAAC,GAAY,EAAA;AACvC,QAAA,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;AAChE,YAAA,OAAO,KAAK;;AAGd,QAAA,IAAI,mBAAmB,IAAI,GAAG,EAAE;AAC9B,YAAA,OAAO,IAAI;;QAGb,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE;YACtC,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,EAAE;AAC/C,gBAAA,IAAI,IAAI,CAAC,oBAAoB,CAAC,KAAK,CAAC,EAAE;AACpC,oBAAA,OAAO,IAAI;;;;AAKjB,QAAA,OAAO,KAAK;;AAGd;;AAEG;AACK,IAAA,uBAAuB,CAAC,GAAY,EAAA;QAC1C,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,EAAE;AACrC,YAAA,OAAO,GAAG;;AAGZ,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AACtB,YAAA,OAAO,GAAG,CAAC,GAAG,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,uBAAuB,CAAC,IAAI,CAAC,CAAC;;AAG9D,QAAA,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;YAC3B,MAAM,OAAO,GAA4B,EAAE;AAC3C,YAAA,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AAC9C,gBAAA,IAAI,GAAG,KAAK,mBAAmB,EAAE;oBAC/B,OAAO,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,uBAAuB,CAAC,KAAK,CAAC;;;AAGtD,YAAA,OAAO,OAAO;;AAGhB,QAAA,OAAO,GAAG;;AAEb;;;;"}

package/dist/esm/llm/bedrock/index.mjs

@@ -1,9 +1,6 @@
 import { ChatBedrockConverse } from '@langchain/aws';
 import { AIMessageChunk } from '@langchain/core/messages';
 import { ChatGenerationChunk } from '@langchain/core/outputs';
-import { ConverseCommand, ConverseStreamCommand } from '@aws-sdk/client-bedrock-runtime';
-import { convertToConverseMessages } from './utils/message_inputs.mjs';
-import { convertConverseMessageToLangChainMessage, handleConverseStreamContentBlockStart, handleConverseStreamContentBlockDelta, handleConverseStreamMetadata } from './utils/message_outputs.mjs';
 
 /**
  * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts
@@ -50,7 +47,7 @@ class CustomChatBedrockConverse extends ChatBedrockConverse {
      */
     invocationParams(options) {
         const baseParams = super.invocationParams(options);
-        // Get serviceTier from options or fall back to class-level setting
+        /** Service tier from options or fall back to class-level setting */
         const serviceTierType = options?.serviceTier ?? this.serviceTier;
         return {
             ...baseParams,
@@ -59,86 +56,50 @@ class CustomChatBedrockConverse extends ChatBedrockConverse {
     }
     /**
      * Override _generateNonStreaming to use applicationInferenceProfile as modelId.
+     * Uses the same model-swapping pattern as streaming for consistency.
      */
-    async _generateNonStreaming(messages, options, _runManager) {
-        const { converseMessages, converseSystem } = convertToConverseMessages(messages);
-        const params = this.invocationParams(options);
-        const command = new ConverseCommand({
-            modelId: this.getModelId(),
-            messages: converseMessages,
-            system: converseSystem,
-            requestMetadata: options.requestMetadata,
-            ...params,
-        });
-        const response = await this.client.send(command, {
-            abortSignal: options.signal,
-        });
-        const { output, ...responseMetadata } = response;
-        if (!output?.message) {
-            throw new Error('No message found in Bedrock response.');
+    async _generateNonStreaming(messages, options, runManager) {
+        // 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;
         }
-        const message = convertConverseMessageToLangChainMessage(output.message, responseMetadata);
-        return {
-            generations: [
-                {
-                    text: typeof message.content === 'string' ? message.content : '',
-                    message,
-                },
-            ],
-        };
     }
     /**
      * Override _streamResponseChunks to:
-     * 1. Use applicationInferenceProfile as modelId
-     * 2. Include serviceTier in request
-     * 3. Strip contentBlockIndex from response_metadata to prevent merge conflicts
+     * 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.
      */
     async *_streamResponseChunks(messages, options, runManager) {
-        const { converseMessages, converseSystem } = convertToConverseMessages(messages);
-        const params = this.invocationParams(options);
-        let { streamUsage } = this;
-        if (options.streamUsage !== undefined) {
-            streamUsage = options.streamUsage;
+        // Temporarily swap model for applicationInferenceProfile support
+        const originalModel = this.model;
+        if (this.applicationInferenceProfile != null &&
+            this.applicationInferenceProfile !== '') {
+            this.model = this.applicationInferenceProfile;
         }
-        const command = new ConverseStreamCommand({
-            modelId: this.getModelId(),
-            messages: converseMessages,
-            system: converseSystem,
-            requestMetadata: options.requestMetadata,
-            ...params,
-        });
-        const response = await this.client.send(command, {
-            abortSignal: options.signal,
-        });
-        if (response.stream) {
-            for await (const event of response.stream) {
-                if (event.contentBlockStart != null) {
-                    const chunk = handleConverseStreamContentBlockStart(event.contentBlockStart);
-                    if (chunk !== undefined) {
-                        const cleanedChunk = this.cleanChunk(chunk);
-                        yield cleanedChunk;
-                        await runManager?.handleLLMNewToken(cleanedChunk.text || '');
-                    }
-                }
-                else if (event.contentBlockDelta != null) {
-                    const chunk = handleConverseStreamContentBlockDelta(event.contentBlockDelta);
-                    if (chunk !== undefined) {
-                        const cleanedChunk = this.cleanChunk(chunk);
-                        yield cleanedChunk;
-                        await runManager?.handleLLMNewToken(cleanedChunk.text || '');
-                    }
-                }
-                else if (event.metadata != null) {
-                    const chunk = handleConverseStreamMetadata(event.metadata, {
-                        streamUsage,
-                    });
-                    if (chunk !== undefined) {
-                        const cleanedChunk = this.cleanChunk(chunk);
-                        yield cleanedChunk;
-                    }
-                }
+        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;
+        }
     }
     /**
      * Clean a chunk by removing contentBlockIndex from response_metadata.

package/dist/esm/llm/bedrock/index.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts\n * and adds support for latest @langchain/aws features:\n *\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock sends the same contentBlockIndex for both text and tool_use content blocks,\n * causing LangChain's merge logic to fail with \"field[contentBlockIndex] already exists\"\n * errors. This wrapper simply strips contentBlockIndex from response_metadata to avoid\n * the conflict.\n *\n * The contentBlockIndex field is only used internally by Bedrock's streaming protocol\n * and isn't needed by application logic - the index field on tool_call_chunks serves\n * the purpose of tracking tool call ordering.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport {\n  ConverseCommand,\n  ConverseStreamCommand,\n} from '@aws-sdk/client-bedrock-runtime';\nimport {\n  convertToConverseMessages,\n  convertConverseMessageToLangChainMessage,\n  handleConverseStreamContentBlockStart,\n  handleConverseStreamContentBlockDelta,\n  handleConverseStreamMetadata,\n} from './utils';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\n/**\n * Extended input interface with additional features:\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  /**\n   * Application Inference Profile ARN to use instead of model ID.\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   */\n  serviceTier?: ServiceTierType;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n  }\n\n  static lc_name(): string {\n    return 'LibreChatBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to add serviceTier support.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const baseParams = super.invocationParams(options);\n\n    // Get serviceTier from options or fall back to class-level setting\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...baseParams,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    _runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    const { converseMessages, converseSystem } =\n      convertToConverseMessages(messages);\n    const params = this.invocationParams(options);\n\n    const command = new ConverseCommand({\n      modelId: this.getModelId(),\n      messages: converseMessages,\n      system: converseSystem,\n      requestMetadata: options.requestMetadata,\n      ...params,\n    });\n\n    const response = await this.client.send(command, {\n      abortSignal: options.signal,\n    });\n\n    const { output, ...responseMetadata } = response;\n    if (!output?.message) {\n      throw new Error('No message found in Bedrock response.');\n    }\n\n    const message = convertConverseMessageToLangChainMessage(\n      output.message,\n      responseMetadata\n    );\n\n    return {\n      generations: [\n        {\n          text: typeof message.content === 'string' ? message.content : '',\n          message,\n        },\n      ],\n    };\n  }\n\n  /**\n   * Override _streamResponseChunks to:\n   * 1. Use applicationInferenceProfile as modelId\n   * 2. Include serviceTier in request\n   * 3. Strip contentBlockIndex from response_metadata to prevent merge conflicts\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const { converseMessages, converseSystem } =\n      convertToConverseMessages(messages);\n    const params = this.invocationParams(options);\n\n    let { streamUsage } = this;\n    if (options.streamUsage !== undefined) {\n      streamUsage = options.streamUsage;\n    }\n\n    const command = new ConverseStreamCommand({\n      modelId: this.getModelId(),\n      messages: converseMessages,\n      system: converseSystem,\n      requestMetadata: options.requestMetadata,\n      ...params,\n    });\n\n    const response = await this.client.send(command, {\n      abortSignal: options.signal,\n    });\n\n    if (response.stream) {\n      for await (const event of response.stream) {\n        if (event.contentBlockStart != null) {\n          const chunk = handleConverseStreamContentBlockStart(\n            event.contentBlockStart\n          ) as ChatGenerationChunk | undefined;\n          if (chunk !== undefined) {\n            const cleanedChunk = this.cleanChunk(chunk);\n            yield cleanedChunk;\n            await runManager?.handleLLMNewToken(cleanedChunk.text || '');\n          }\n        } else if (event.contentBlockDelta != null) {\n          const chunk = handleConverseStreamContentBlockDelta(\n            event.contentBlockDelta\n          ) as ChatGenerationChunk | undefined;\n          if (chunk !== undefined) {\n            const cleanedChunk = this.cleanChunk(chunk);\n            yield cleanedChunk;\n            await runManager?.handleLLMNewToken(cleanedChunk.text || '');\n          }\n        } else if (event.metadata != null) {\n          const chunk = handleConverseStreamMetadata(event.metadata, {\n            streamUsage,\n          }) as ChatGenerationChunk | undefined;\n          if (chunk !== undefined) {\n            const cleanedChunk = this.cleanChunk(chunk);\n            yield cleanedChunk;\n          }\n        }\n      }\n    }\n  }\n\n  /**\n   * Clean a chunk by removing contentBlockIndex from response_metadata.\n   */\n  private cleanChunk(chunk: ChatGenerationChunk): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const metadata = message.response_metadata as Record<string, unknown>;\n    const hasContentBlockIndex = this.hasContentBlockIndex(metadata);\n    if (!hasContentBlockIndex) {\n      return chunk;\n    }\n\n    const cleanedMetadata = this.removeContentBlockIndex(metadata) as Record<\n      string,\n      unknown\n    >;\n\n    return new ChatGenerationChunk({\n      text: chunk.text,\n      message: new AIMessageChunk({\n        ...message,\n        response_metadata: cleanedMetadata,\n      }),\n      generationInfo: chunk.generationInfo,\n    });\n  }\n\n  /**\n   * Check if contentBlockIndex exists at any level in the object\n   */\n  private hasContentBlockIndex(obj: unknown): boolean {\n    if (obj === null || obj === undefined || typeof obj !== 'object') {\n      return false;\n    }\n\n    if ('contentBlockIndex' in obj) {\n      return true;\n    }\n\n    for (const value of Object.values(obj)) {\n      if (typeof value === 'object' && value !== null) {\n        if (this.hasContentBlockIndex(value)) {\n          return true;\n        }\n      }\n    }\n\n    return false;\n  }\n\n  /**\n   * Recursively remove contentBlockIndex from all levels of an object\n   */\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"names":[],"mappings":";;;;;;;AAAA;;;;;;;;;;;;;;;AAeG;AAmEG,MAAO,yBAA0B,SAAQ,mBAAmB,CAAA;AAChE;;AAEG;AACH,IAAA,2BAA2B;AAE3B;;AAEG;AACH,IAAA,WAAW;AAEX,IAAA,WAAA,CAAY,MAAuC,EAAA;QACjD,KAAK,CAAC,MAAM,CAAC;AACb,QAAA,IAAI,CAAC,2BAA2B,GAAG,MAAM,EAAE,2BAA2B;AACtE,QAAA,IAAI,CAAC,WAAW,GAAG,MAAM,EAAE,WAAW;;AAGxC,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,0BAA0B;;AAGnC;;;AAGG;IACO,UAAU,GAAA;AAClB,QAAA,OAAO,IAAI,CAAC,2BAA2B,IAAI,IAAI,CAAC,KAAK;;AAGvD;;AAEG;AACM,IAAA,gBAAgB,CACvB,OAA0E,EAAA;QAI1E,MAAM,UAAU,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;;QAGlD,MAAM,eAAe,GAAG,OAAO,EAAE,WAAW,IAAI,IAAI,CAAC,WAAW;QAEhE,OAAO;AACL,YAAA,GAAG,UAAU;AACb,YAAA,WAAW,EAAE,eAAe,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE,GAAG,SAAS;SACrE;;AAGH;;AAEG;AACM,IAAA,MAAM,qBAAqB,CAClC,QAAuB,EACvB,OAAyE,EACzE,WAAsC,EAAA;QAEtC,MAAM,EAAE,gBAAgB,EAAE,cAAc,EAAE,GACxC,yBAAyB,CAAC,QAAQ,CAAC;QACrC,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC;AAE7C,QAAA,MAAM,OAAO,GAAG,IAAI,eAAe,CAAC;AAClC,YAAA,OAAO,EAAE,IAAI,CAAC,UAAU,EAAE;AAC1B,YAAA,QAAQ,EAAE,gBAAgB;AAC1B,YAAA,MAAM,EAAE,cAAc;YACtB,eAAe,EAAE,OAAO,CAAC,eAAe;AACxC,YAAA,GAAG,MAAM;AACV,SAAA,CAAC;QAEF,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,EAAE;YAC/C,WAAW,EAAE,OAAO,CAAC,MAAM;AAC5B,SAAA,CAAC;QAEF,MAAM,EAAE,MAAM,EAAE,GAAG,gBAAgB,EAAE,GAAG,QAAQ;AAChD,QAAA,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE;AACpB,YAAA,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC;;QAG1D,MAAM,OAAO,GAAG,wCAAwC,CACtD,MAAM,CAAC,OAAO,EACd,gBAAgB,CACjB;QAED,OAAO;AACL,YAAA,WAAW,EAAE;AACX,gBAAA;AACE,oBAAA,IAAI,EAAE,OAAO,OAAO,CAAC,OAAO,KAAK,QAAQ,GAAG,OAAO,CAAC,OAAO,GAAG,EAAE;oBAChE,OAAO;AACR,iBAAA;AACF,aAAA;SACF;;AAGH;;;;;AAKG;IACM,OAAO,qBAAqB,CACnC,QAAuB,EACvB,OAAyE,EACzE,UAAqC,EAAA;QAErC,MAAM,EAAE,gBAAgB,EAAE,cAAc,EAAE,GACxC,yBAAyB,CAAC,QAAQ,CAAC;QACrC,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC;AAE7C,QAAA,IAAI,EAAE,WAAW,EAAE,GAAG,IAAI;AAC1B,QAAA,IAAI,OAAO,CAAC,WAAW,KAAK,SAAS,EAAE;AACrC,YAAA,WAAW,GAAG,OAAO,CAAC,WAAW;;AAGnC,QAAA,MAAM,OAAO,GAAG,IAAI,qBAAqB,CAAC;AACxC,YAAA,OAAO,EAAE,IAAI,CAAC,UAAU,EAAE;AAC1B,YAAA,QAAQ,EAAE,gBAAgB;AAC1B,YAAA,MAAM,EAAE,cAAc;YACtB,eAAe,EAAE,OAAO,CAAC,eAAe;AACxC,YAAA,GAAG,MAAM;AACV,SAAA,CAAC;QAEF,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,EAAE;YAC/C,WAAW,EAAE,OAAO,CAAC,MAAM;AAC5B,SAAA,CAAC;AAEF,QAAA,IAAI,QAAQ,CAAC,MAAM,EAAE;YACnB,WAAW,MAAM,KAAK,IAAI,QAAQ,CAAC,MAAM,EAAE;AACzC,gBAAA,IAAI,KAAK,CAAC,iBAAiB,IAAI,IAAI,EAAE;oBACnC,MAAM,KAAK,GAAG,qCAAqC,CACjD,KAAK,CAAC,iBAAiB,CACW;AACpC,oBAAA,IAAI,KAAK,KAAK,SAAS,EAAE;wBACvB,MAAM,YAAY,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;AAC3C,wBAAA,MAAM,YAAY;wBAClB,MAAM,UAAU,EAAE,iBAAiB,CAAC,YAAY,CAAC,IAAI,IAAI,EAAE,CAAC;;;AAEzD,qBAAA,IAAI,KAAK,CAAC,iBAAiB,IAAI,IAAI,EAAE;oBAC1C,MAAM,KAAK,GAAG,qCAAqC,CACjD,KAAK,CAAC,iBAAiB,CACW;AACpC,oBAAA,IAAI,KAAK,KAAK,SAAS,EAAE;wBACvB,MAAM,YAAY,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;AAC3C,wBAAA,MAAM,YAAY;wBAClB,MAAM,UAAU,EAAE,iBAAiB,CAAC,YAAY,CAAC,IAAI,IAAI,EAAE,CAAC;;;AAEzD,qBAAA,IAAI,KAAK,CAAC,QAAQ,IAAI,IAAI,EAAE;AACjC,oBAAA,MAAM,KAAK,GAAG,4BAA4B,CAAC,KAAK,CAAC,QAAQ,EAAE;wBACzD,WAAW;AACZ,qBAAA,CAAoC;AACrC,oBAAA,IAAI,KAAK,KAAK,SAAS,EAAE;wBACvB,MAAM,YAAY,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;AAC3C,wBAAA,MAAM,YAAY;;;;;;AAO5B;;AAEG;AACK,IAAA,UAAU,CAAC,KAA0B,EAAA;AAC3C,QAAA,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO;AAC7B,QAAA,IAAI,EAAE,OAAO,YAAY,cAAc,CAAC,EAAE;AACxC,YAAA,OAAO,KAAK;;AAGd,QAAA,MAAM,QAAQ,GAAG,OAAO,CAAC,iBAA4C;QACrE,MAAM,oBAAoB,GAAG,IAAI,CAAC,oBAAoB,CAAC,QAAQ,CAAC;QAChE,IAAI,CAAC,oBAAoB,EAAE;AACzB,YAAA,OAAO,KAAK;;QAGd,MAAM,eAAe,GAAG,IAAI,CAAC,uBAAuB,CAAC,QAAQ,CAG5D;QAED,OAAO,IAAI,mBAAmB,CAAC;YAC7B,IAAI,EAAE,KAAK,CAAC,IAAI;YAChB,OAAO,EAAE,IAAI,cAAc,CAAC;AAC1B,gBAAA,GAAG,OAAO;AACV,gBAAA,iBAAiB,EAAE,eAAe;aACnC,CAAC;YACF,cAAc,EAAE,KAAK,CAAC,cAAc;AACrC,SAAA,CAAC;;AAGJ;;AAEG;AACK,IAAA,oBAAoB,CAAC,GAAY,EAAA;AACvC,QAAA,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;AAChE,YAAA,OAAO,KAAK;;AAGd,QAAA,IAAI,mBAAmB,IAAI,GAAG,EAAE;AAC9B,YAAA,OAAO,IAAI;;QAGb,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE;YACtC,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,EAAE;AAC/C,gBAAA,IAAI,IAAI,CAAC,oBAAoB,CAAC,KAAK,CAAC,EAAE;AACpC,oBAAA,OAAO,IAAI;;;;AAKjB,QAAA,OAAO,KAAK;;AAGd;;AAEG;AACK,IAAA,uBAAuB,CAAC,GAAY,EAAA;QAC1C,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,EAAE;AACrC,YAAA,OAAO,GAAG;;AAGZ,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AACtB,YAAA,OAAO,GAAG,CAAC,GAAG,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,uBAAuB,CAAC,IAAI,CAAC,CAAC;;AAG9D,QAAA,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;YAC3B,MAAM,OAAO,GAA4B,EAAE;AAC3C,YAAA,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AAC9C,gBAAA,IAAI,GAAG,KAAK,mBAAmB,EAAE;oBAC/B,OAAO,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,uBAAuB,CAAC,KAAK,CAAC;;;AAGtD,YAAA,OAAO,OAAO;;AAGhB,QAAA,OAAO,GAAG;;AAEb;;;;"}
+{"version":3,"file":"index.mjs","sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts\n * and adds support for latest @langchain/aws features:\n *\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock sends the same contentBlockIndex for both text and tool_use content blocks,\n * causing LangChain's merge logic to fail with \"field[contentBlockIndex] already exists\"\n * errors. This wrapper simply strips contentBlockIndex from response_metadata to avoid\n * the conflict.\n *\n * The contentBlockIndex field is only used internally by Bedrock's streaming protocol\n * and isn't needed by application logic - the index field on tool_call_chunks serves\n * the purpose of tracking tool call ordering.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport type { BaseMessage } from '@langchain/core/messages';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\n/**\n * Extended input interface with additional features:\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  /**\n   * Application Inference Profile ARN to use instead of model ID.\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   */\n  serviceTier?: ServiceTierType;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n  }\n\n  static lc_name(): string {\n    return 'LibreChatBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to add serviceTier support.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const baseParams = super.invocationParams(options);\n\n    /** Service tier from options or fall back to class-level setting */\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...baseParams,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   * Uses the same model-swapping pattern as streaming for consistency.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    // Temporarily swap model for applicationInferenceProfile support\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      return await super._generateNonStreaming(messages, options, runManager);\n    } finally {\n      // Restore original model\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Override _streamResponseChunks to:\n   * 1. Use applicationInferenceProfile as modelId (by temporarily swapping this.model)\n   * 2. Strip contentBlockIndex from response_metadata to prevent merge conflicts\n   *\n   * Note: We delegate to super._streamResponseChunks() to preserve @langchain/aws's\n   * internal chunk handling which correctly preserves array content for reasoning blocks.\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    // Temporarily swap model for applicationInferenceProfile support\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      // Use parent's streaming logic which correctly handles reasoning content\n      const baseStream = super._streamResponseChunks(\n        messages,\n        options,\n        runManager\n      );\n\n      for await (const chunk of baseStream) {\n        // Clean contentBlockIndex from response_metadata to prevent merge conflicts\n        yield this.cleanChunk(chunk);\n      }\n    } finally {\n      // Restore original model\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Clean a chunk by removing contentBlockIndex from response_metadata.\n   */\n  private cleanChunk(chunk: ChatGenerationChunk): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const metadata = message.response_metadata as Record<string, unknown>;\n    const hasContentBlockIndex = this.hasContentBlockIndex(metadata);\n    if (!hasContentBlockIndex) {\n      return chunk;\n    }\n\n    const cleanedMetadata = this.removeContentBlockIndex(metadata) as Record<\n      string,\n      unknown\n    >;\n\n    return new ChatGenerationChunk({\n      text: chunk.text,\n      message: new AIMessageChunk({\n        ...message,\n        response_metadata: cleanedMetadata,\n      }),\n      generationInfo: chunk.generationInfo,\n    });\n  }\n\n  /**\n   * Check if contentBlockIndex exists at any level in the object\n   */\n  private hasContentBlockIndex(obj: unknown): boolean {\n    if (obj === null || obj === undefined || typeof obj !== 'object') {\n      return false;\n    }\n\n    if ('contentBlockIndex' in obj) {\n      return true;\n    }\n\n    for (const value of Object.values(obj)) {\n      if (typeof value === 'object' && value !== null) {\n        if (this.hasContentBlockIndex(value)) {\n          return true;\n        }\n      }\n    }\n\n    return false;\n  }\n\n  /**\n   * Recursively remove contentBlockIndex from all levels of an object\n   */\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"names":[],"mappings":";;;;AAAA;;;;;;;;;;;;;;;AAeG;AAwDG,MAAO,yBAA0B,SAAQ,mBAAmB,CAAA;AAChE;;AAEG;AACH,IAAA,2BAA2B;AAE3B;;AAEG;AACH,IAAA,WAAW;AAEX,IAAA,WAAA,CAAY,MAAuC,EAAA;QACjD,KAAK,CAAC,MAAM,CAAC;AACb,QAAA,IAAI,CAAC,2BAA2B,GAAG,MAAM,EAAE,2BAA2B;AACtE,QAAA,IAAI,CAAC,WAAW,GAAG,MAAM,EAAE,WAAW;;AAGxC,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,0BAA0B;;AAGnC;;;AAGG;IACO,UAAU,GAAA;AAClB,QAAA,OAAO,IAAI,CAAC,2BAA2B,IAAI,IAAI,CAAC,KAAK;;AAGvD;;AAEG;AACM,IAAA,gBAAgB,CACvB,OAA0E,EAAA;QAI1E,MAAM,UAAU,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;;QAGlD,MAAM,eAAe,GAAG,OAAO,EAAE,WAAW,IAAI,IAAI,CAAC,WAAW;QAEhE,OAAO;AACL,YAAA,GAAG,UAAU;AACb,YAAA,WAAW,EAAE,eAAe,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE,GAAG,SAAS;SACrE;;AAGH;;;AAGG;AACM,IAAA,MAAM,qBAAqB,CAClC,QAAuB,EACvB,OAAyE,EACzE,UAAqC,EAAA;;AAGrC,QAAA,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK;AAChC,QAAA,IACE,IAAI,CAAC,2BAA2B,IAAI,IAAI;AACxC,YAAA,IAAI,CAAC,2BAA2B,KAAK,EAAE,EACvC;AACA,YAAA,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,2BAA2B;;AAG/C,QAAA,IAAI;YACF,OAAO,MAAM,KAAK,CAAC,qBAAqB,CAAC,QAAQ,EAAE,OAAO,EAAE,UAAU,CAAC;;gBAC/D;;AAER,YAAA,IAAI,CAAC,KAAK,GAAG,aAAa;;;AAI9B;;;;;;;AAOG;IACM,OAAO,qBAAqB,CACnC,QAAuB,EACvB,OAAyE,EACzE,UAAqC,EAAA;;AAGrC,QAAA,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK;AAChC,QAAA,IACE,IAAI,CAAC,2BAA2B,IAAI,IAAI;AACxC,YAAA,IAAI,CAAC,2BAA2B,KAAK,EAAE,EACvC;AACA,YAAA,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,2BAA2B;;AAG/C,QAAA,IAAI;;AAEF,YAAA,MAAM,UAAU,GAAG,KAAK,CAAC,qBAAqB,CAC5C,QAAQ,EACR,OAAO,EACP,UAAU,CACX;AAED,YAAA,WAAW,MAAM,KAAK,IAAI,UAAU,EAAE;;AAEpC,gBAAA,MAAM,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;;;gBAEtB;;AAER,YAAA,IAAI,CAAC,KAAK,GAAG,aAAa;;;AAI9B;;AAEG;AACK,IAAA,UAAU,CAAC,KAA0B,EAAA;AAC3C,QAAA,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO;AAC7B,QAAA,IAAI,EAAE,OAAO,YAAY,cAAc,CAAC,EAAE;AACxC,YAAA,OAAO,KAAK;;AAGd,QAAA,MAAM,QAAQ,GAAG,OAAO,CAAC,iBAA4C;QACrE,MAAM,oBAAoB,GAAG,IAAI,CAAC,oBAAoB,CAAC,QAAQ,CAAC;QAChE,IAAI,CAAC,oBAAoB,EAAE;AACzB,YAAA,OAAO,KAAK;;QAGd,MAAM,eAAe,GAAG,IAAI,CAAC,uBAAuB,CAAC,QAAQ,CAG5D;QAED,OAAO,IAAI,mBAAmB,CAAC;YAC7B,IAAI,EAAE,KAAK,CAAC,IAAI;YAChB,OAAO,EAAE,IAAI,cAAc,CAAC;AAC1B,gBAAA,GAAG,OAAO;AACV,gBAAA,iBAAiB,EAAE,eAAe;aACnC,CAAC;YACF,cAAc,EAAE,KAAK,CAAC,cAAc;AACrC,SAAA,CAAC;;AAGJ;;AAEG;AACK,IAAA,oBAAoB,CAAC,GAAY,EAAA;AACvC,QAAA,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;AAChE,YAAA,OAAO,KAAK;;AAGd,QAAA,IAAI,mBAAmB,IAAI,GAAG,EAAE;AAC9B,YAAA,OAAO,IAAI;;QAGb,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE;YACtC,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,EAAE;AAC/C,gBAAA,IAAI,IAAI,CAAC,oBAAoB,CAAC,KAAK,CAAC,EAAE;AACpC,oBAAA,OAAO,IAAI;;;;AAKjB,QAAA,OAAO,KAAK;;AAGd;;AAEG;AACK,IAAA,uBAAuB,CAAC,GAAY,EAAA;QAC1C,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,EAAE;AACrC,YAAA,OAAO,GAAG;;AAGZ,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AACtB,YAAA,OAAO,GAAG,CAAC,GAAG,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,uBAAuB,CAAC,IAAI,CAAC,CAAC;;AAG9D,QAAA,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;YAC3B,MAAM,OAAO,GAA4B,EAAE;AAC3C,YAAA,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AAC9C,gBAAA,IAAI,GAAG,KAAK,mBAAmB,EAAE;oBAC/B,OAAO,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,uBAAuB,CAAC,KAAK,CAAC;;;AAGtD,YAAA,OAAO,OAAO;;AAGhB,QAAA,OAAO,GAAG;;AAEb;;;;"}

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

@@ -87,13 +87,16 @@ export declare class CustomChatBedrockConverse extends ChatBedrockConverse {
     };
     /**
      * 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>;
+    _generateNonStreaming(messages: BaseMessage[], options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions, runManager?: CallbackManagerForLLMRun): Promise<ChatResult>;
     /**
      * Override _streamResponseChunks to:
-     * 1. Use applicationInferenceProfile as modelId
-     * 2. Include serviceTier in request
-     * 3. Strip contentBlockIndex from response_metadata to prevent merge conflicts
+     * 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>;
     /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "3.0.79",
+  "version": "3.0.80",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",
@@ -56,6 +56,7 @@
     "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 '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'",

package/src/llm/bedrock/index.ts

@@ -21,17 +21,6 @@ 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 {
-  ConverseCommand,
-  ConverseStreamCommand,
-} from '@aws-sdk/client-bedrock-runtime';
-import {
-  convertToConverseMessages,
-  convertConverseMessageToLangChainMessage,
-  handleConverseStreamContentBlockStart,
-  handleConverseStreamContentBlockDelta,
-  handleConverseStreamMetadata,
-} from './utils';
 
 /**
  * Service tier type for Bedrock invocations.
@@ -119,7 +108,7 @@ export class CustomChatBedrockConverse extends ChatBedrockConverse {
   } {
     const baseParams = super.invocationParams(options);
 
-    // Get serviceTier from options or fall back to class-level setting
+    /** Service tier from options or fall back to class-level setting */
     const serviceTierType = options?.serviceTier ?? this.serviceTier;
 
     return {
@@ -130,110 +119,67 @@ export class CustomChatBedrockConverse extends ChatBedrockConverse {
 
   /**
    * Override _generateNonStreaming to use applicationInferenceProfile as modelId.
+   * Uses the same model-swapping pattern as streaming for consistency.
    */
   override async _generateNonStreaming(
     messages: BaseMessage[],
     options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,
-    _runManager?: CallbackManagerForLLMRun
+    runManager?: CallbackManagerForLLMRun
   ): Promise<ChatResult> {
-    const { converseMessages, converseSystem } =
-      convertToConverseMessages(messages);
-    const params = this.invocationParams(options);
-
-    const command = new ConverseCommand({
-      modelId: this.getModelId(),
-      messages: converseMessages,
-      system: converseSystem,
-      requestMetadata: options.requestMetadata,
-      ...params,
-    });
-
-    const response = await this.client.send(command, {
-      abortSignal: options.signal,
-    });
-
-    const { output, ...responseMetadata } = response;
-    if (!output?.message) {
-      throw new Error('No message found in Bedrock response.');
+    // Temporarily swap model for applicationInferenceProfile support
+    const originalModel = this.model;
+    if (
+      this.applicationInferenceProfile != null &&
+      this.applicationInferenceProfile !== ''
+    ) {
+      this.model = this.applicationInferenceProfile;
     }
 
-    const message = convertConverseMessageToLangChainMessage(
-      output.message,
-      responseMetadata
-    );
-
-    return {
-      generations: [
-        {
-          text: typeof message.content === 'string' ? message.content : '',
-          message,
-        },
-      ],
-    };
+    try {
+      return await super._generateNonStreaming(messages, options, runManager);
+    } finally {
+      // Restore original model
+      this.model = originalModel;
+    }
   }
 
   /**
    * Override _streamResponseChunks to:
-   * 1. Use applicationInferenceProfile as modelId
-   * 2. Include serviceTier in request
-   * 3. Strip contentBlockIndex from response_metadata to prevent merge conflicts
+   * 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 { converseMessages, converseSystem } =
-      convertToConverseMessages(messages);
-    const params = this.invocationParams(options);
-
-    let { streamUsage } = this;
-    if (options.streamUsage !== undefined) {
-      streamUsage = options.streamUsage;
+    // Temporarily swap model for applicationInferenceProfile support
+    const originalModel = this.model;
+    if (
+      this.applicationInferenceProfile != null &&
+      this.applicationInferenceProfile !== ''
+    ) {
+      this.model = this.applicationInferenceProfile;
     }
 
-    const command = new ConverseStreamCommand({
-      modelId: this.getModelId(),
-      messages: converseMessages,
-      system: converseSystem,
-      requestMetadata: options.requestMetadata,
-      ...params,
-    });
-
-    const response = await this.client.send(command, {
-      abortSignal: options.signal,
-    });
-
-    if (response.stream) {
-      for await (const event of response.stream) {
-        if (event.contentBlockStart != null) {
-          const chunk = handleConverseStreamContentBlockStart(
-            event.contentBlockStart
-          ) as ChatGenerationChunk | undefined;
-          if (chunk !== undefined) {
-            const cleanedChunk = this.cleanChunk(chunk);
-            yield cleanedChunk;
-            await runManager?.handleLLMNewToken(cleanedChunk.text || '');
-          }
-        } else if (event.contentBlockDelta != null) {
-          const chunk = handleConverseStreamContentBlockDelta(
-            event.contentBlockDelta
-          ) as ChatGenerationChunk | undefined;
-          if (chunk !== undefined) {
-            const cleanedChunk = this.cleanChunk(chunk);
-            yield cleanedChunk;
-            await runManager?.handleLLMNewToken(cleanedChunk.text || '');
-          }
-        } else if (event.metadata != null) {
-          const chunk = handleConverseStreamMetadata(event.metadata, {
-            streamUsage,
-          }) as ChatGenerationChunk | undefined;
-          if (chunk !== undefined) {
-            const cleanedChunk = this.cleanChunk(chunk);
-            yield cleanedChunk;
-          }
-        }
+    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;
     }
   }