illuma-agents 1.0.37 → 1.0.39

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

@@ -6,7 +6,11 @@ var outputs = require('@langchain/core/outputs');
 
 /**
  * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts
- * and adds prompt caching support for Bedrock Converse API.
+ * and adds support for:
+ *
+ * - Prompt caching support for Bedrock Converse API (Illuma feature)
+ * - 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"
@@ -27,19 +31,34 @@ var outputs = require('@langchain/core/outputs');
  * response. This wrapper adds input_token_details to usage_metadata with cache information.
  */
 class CustomChatBedrockConverse extends aws.ChatBedrockConverse {
+    /** Enable Bedrock prompt caching for tool definitions */
     promptCache;
+    /** Application Inference Profile ARN to use instead of model ID */
+    applicationInferenceProfile;
+    /** Service tier for model invocation */
+    serviceTier;
     constructor(fields) {
         super(fields);
         this.promptCache = fields?.promptCache ?? false;
+        this.applicationInferenceProfile = fields?.applicationInferenceProfile;
+        this.serviceTier = fields?.serviceTier;
     }
     static lc_name() {
         return 'IllumaBedrockConverse';
     }
     /**
-     * Override invocationParams to add cachePoint to tools when promptCache is enabled.
-     * This enables Bedrock prompt caching for tool definitions.
+     * Get the model ID to use for API calls.
+     * Returns applicationInferenceProfile if set, otherwise returns this.model.
+     */
+    getModelId() {
+        return this.applicationInferenceProfile ?? this.model;
+    }
+    /**
+     * Override invocationParams to:
+     * 1. Add cachePoint to tools when promptCache is enabled
+     * 2. Add serviceTier support
      *
-     * STRATEGY: Separate cachePoints for core tools and MCP tools
+     * CACHING STRATEGY: Separate cachePoints for core tools and MCP tools
      * - Core tools (web_search, execute_code, etc.) are stable → cache first
      * - MCP tools (have '_mcp_' in name) are dynamic → cache separately after
      * - This allows core tools to stay cached when MCP selection changes
@@ -90,69 +109,117 @@ class CustomChatBedrockConverse extends aws.ChatBedrockConverse {
             }
             params.toolConfig.tools = toolsWithCache;
         }
-        return params;
+        // Add serviceTier support
+        const serviceTierType = options?.serviceTier ?? this.serviceTier;
+        return {
+            ...params,
+            serviceTier: serviceTierType ? { type: serviceTierType } : undefined,
+        };
+    }
+    /**
+     * Override _generateNonStreaming to use applicationInferenceProfile as modelId.
+     * Uses the same model-swapping pattern as streaming for consistency.
+     */
+    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;
+        }
     }
     /**
      * Override _streamResponseChunks to:
-     * 1. Strip contentBlockIndex from response_metadata to prevent merge conflicts
-     * 2. Extract cacheReadInputTokens/cacheWriteInputTokens and add to usage_metadata
+     * 1. Use applicationInferenceProfile as modelId (by temporarily swapping this.model)
+     * 2. Strip contentBlockIndex from response_metadata to prevent merge conflicts
+     * 3. Extract cacheReadInputTokens/cacheWriteInputTokens and add to usage_metadata
+     *
+     * 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$1, options, runManager) {
-        const baseStream = super._streamResponseChunks(messages$1, options, runManager);
-        for await (const chunk of baseStream) {
-            // Only process if we have response_metadata
-            if (chunk.message instanceof messages.AIMessageChunk &&
-                chunk.message.response_metadata &&
-                typeof chunk.message.response_metadata === 'object') {
-                const responseMetadata = chunk.message.response_metadata;
-                let needsModification = false;
-                let cleanedMetadata = responseMetadata;
-                // Check if contentBlockIndex exists anywhere in response_metadata
-                const hasContentBlockIndex = this.hasContentBlockIndex(responseMetadata);
-                if (hasContentBlockIndex) {
-                    cleanedMetadata = this.removeContentBlockIndex(responseMetadata);
-                    needsModification = true;
-                }
-                // Extract cache tokens from metadata.usage (Bedrock streaming format)
-                // The metadata chunk contains usage with cacheReadInputTokens/cacheWriteInputTokens
-                const metadata = responseMetadata.metadata;
-                const usage = (metadata?.usage ?? responseMetadata.usage);
-                let enhancedUsageMetadata = chunk.message.usage_metadata;
-                if (usage) {
-                    const cacheRead = usage.cacheReadInputTokens ?? 0;
-                    const cacheWrite = usage.cacheWriteInputTokens ?? 0;
-                    const inputTokens = usage.inputTokens ?? 0;
-                    const outputTokens = usage.outputTokens ?? 0;
-                    if (cacheRead > 0 || cacheWrite > 0) {
-                        needsModification = true;
-                        enhancedUsageMetadata = {
-                            input_tokens: chunk.message.usage_metadata?.input_tokens ?? inputTokens,
-                            output_tokens: chunk.message.usage_metadata?.output_tokens ?? outputTokens,
-                            total_tokens: chunk.message.usage_metadata?.total_tokens ??
-                                usage.totalTokens ??
-                                0,
-                            input_token_details: {
-                                cache_read: cacheRead,
-                                cache_creation: cacheWrite,
-                            },
-                        };
-                    }
-                }
-                if (needsModification) {
-                    yield new outputs.ChatGenerationChunk({
-                        text: chunk.text,
-                        message: new messages.AIMessageChunk({
-                            ...chunk.message,
-                            response_metadata: cleanedMetadata,
-                            usage_metadata: enhancedUsageMetadata,
-                        }),
-                        generationInfo: chunk.generationInfo,
-                    });
-                    continue;
-                }
+    async *_streamResponseChunks(messages, options, runManager) {
+        // 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 and enhance chunk
+                yield this.processChunk(chunk);
+            }
+        }
+        finally {
+            // Restore original model
+            this.model = originalModel;
+        }
+    }
+    /**
+     * Process a chunk by:
+     * 1. Removing contentBlockIndex from response_metadata
+     * 2. Extracting cache token information from Bedrock's usage data
+     */
+    processChunk(chunk) {
+        const message = chunk.message;
+        if (!(message instanceof messages.AIMessageChunk)) {
+            return chunk;
+        }
+        const responseMetadata = message.response_metadata;
+        let needsModification = false;
+        let cleanedMetadata = responseMetadata;
+        // Check if contentBlockIndex exists anywhere in response_metadata
+        const hasContentBlockIndex = this.hasContentBlockIndex(responseMetadata);
+        if (hasContentBlockIndex) {
+            cleanedMetadata = this.removeContentBlockIndex(responseMetadata);
+            needsModification = true;
+        }
+        // Extract cache tokens from metadata.usage (Bedrock streaming format)
+        // The metadata chunk contains usage with cacheReadInputTokens/cacheWriteInputTokens
+        const metadata = responseMetadata.metadata;
+        const usage = (metadata?.usage ?? responseMetadata.usage);
+        let enhancedUsageMetadata = message.usage_metadata;
+        if (usage) {
+            const cacheRead = usage.cacheReadInputTokens ?? 0;
+            const cacheWrite = usage.cacheWriteInputTokens ?? 0;
+            const inputTokens = usage.inputTokens ?? 0;
+            const outputTokens = usage.outputTokens ?? 0;
+            if (cacheRead > 0 || cacheWrite > 0) {
+                needsModification = true;
+                enhancedUsageMetadata = {
+                    input_tokens: message.usage_metadata?.input_tokens ?? inputTokens,
+                    output_tokens: message.usage_metadata?.output_tokens ?? outputTokens,
+                    total_tokens: message.usage_metadata?.total_tokens ??
+                        usage.totalTokens ??
+                        0,
+                    input_token_details: {
+                        cache_read: cacheRead,
+                        cache_creation: cacheWrite,
+                    },
+                };
             }
-            yield chunk;
         }
+        if (needsModification) {
+            return new outputs.ChatGenerationChunk({
+                text: chunk.text,
+                message: new messages.AIMessageChunk({
+                    ...message,
+                    response_metadata: cleanedMetadata,
+                    usage_metadata: enhancedUsageMetadata,
+                }),
+                generationInfo: chunk.generationInfo,
+            });
+        }
+        return chunk;
     }
     /**
      * Check if contentBlockIndex exists at any level in the object

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 prompt caching support for Bedrock Converse API.\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 * PROMPT CACHING:\n * When promptCache: true is set, this wrapper adds cachePoint markers to the tools array\n * to enable Bedrock prompt caching for tool definitions. This allows tool schemas to be\n * cached and reused across requests, reducing latency and costs.\n *\n * CACHE TOKEN EXTRACTION:\n * LangChain AWS doesn't extract cacheReadInputTokens/cacheWriteInputTokens from Bedrock's\n * response. This wrapper adds input_token_details to usage_metadata with cache information.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport type { BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\n\n/** Extended input type with promptCache option */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  promptCache?: boolean;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  promptCache: boolean;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.promptCache = fields?.promptCache ?? false;\n  }\n\n  static lc_name(): string {\n    return 'IllumaBedrockConverse';\n  }\n\n  /**\n   * Override invocationParams to add cachePoint to tools when promptCache is enabled.\n   * This enables Bedrock prompt caching for tool definitions.\n   *\n   * STRATEGY: Separate cachePoints for core tools and MCP tools\n   * - Core tools (web_search, execute_code, etc.) are stable → cache first\n   * - MCP tools (have '_mcp_' in name) are dynamic → cache separately after\n   * - This allows core tools to stay cached when MCP selection changes\n   *\n   * NOTE: Only Claude models support cachePoint - Nova and other models will reject it.\n   */\n  invocationParams(\n    options?: this['ParsedCallOptions']\n  ): ReturnType<ChatBedrockConverse['invocationParams']> {\n    const params = super.invocationParams(options);\n\n    // Add cachePoint to tools array if promptCache is enabled and tools exist\n    // Only Claude models support cachePoint - check model name\n    const modelId = this.model.toLowerCase();\n    const isClaudeModel =\n      modelId.includes('claude') || modelId.includes('anthropic');\n\n    if (\n      this.promptCache &&\n      isClaudeModel &&\n      params.toolConfig?.tools &&\n      Array.isArray(params.toolConfig.tools) &&\n      params.toolConfig.tools.length > 0\n    ) {\n      // Separate core tools from MCP tools\n      // MCP tools have '_mcp_' in their name (e.g., 'search_emails_mcp_Google-Workspace')\n      const coreTools: typeof params.toolConfig.tools = [];\n      const mcpTools: typeof params.toolConfig.tools = [];\n      const coreToolNames: string[] = [];\n      const mcpToolNames: string[] = [];\n\n      for (const tool of params.toolConfig.tools) {\n        // Check if tool has a name property with '_mcp_' pattern\n        const toolName =\n          (tool as { toolSpec?: { name?: string } }).toolSpec?.name ?? '';\n        if (toolName.includes('_mcp_')) {\n          mcpTools.push(tool);\n          mcpToolNames.push(toolName);\n        } else {\n          coreTools.push(tool);\n          coreToolNames.push(toolName);\n        }\n      }\n\n      // Build tools array with strategic cachePoints:\n      // [CoreTool1, CoreTool2, cachePoint] + [MCPTool1, MCPTool2, cachePoint]\n      const toolsWithCache: typeof params.toolConfig.tools = [];\n\n      // Add core tools with cachePoint (if any)\n      if (coreTools.length > 0) {\n        toolsWithCache.push(...coreTools);\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      // Add MCP tools with their own cachePoint (if any)\n      if (mcpTools.length > 0) {\n        toolsWithCache.push(...mcpTools);\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      // If no tools at all (shouldn't happen but safety check)\n      if (toolsWithCache.length === 0) {\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      params.toolConfig.tools = toolsWithCache;\n    }\n\n    return params;\n  }\n\n  /**\n   * Override _streamResponseChunks to:\n   * 1. Strip contentBlockIndex from response_metadata to prevent merge conflicts\n   * 2. Extract cacheReadInputTokens/cacheWriteInputTokens and add to usage_metadata\n   */\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const baseStream = super._streamResponseChunks(\n      messages,\n      options,\n      runManager\n    );\n\n    for await (const chunk of baseStream) {\n      // Only process if we have response_metadata\n      if (\n        chunk.message instanceof AIMessageChunk &&\n        (chunk.message as Partial<AIMessageChunk>).response_metadata &&\n        typeof chunk.message.response_metadata === 'object'\n      ) {\n        const responseMetadata = chunk.message.response_metadata as Record<\n          string,\n          unknown\n        >;\n        let needsModification = false;\n        let cleanedMetadata = responseMetadata;\n\n        // Check if contentBlockIndex exists anywhere in response_metadata\n        const hasContentBlockIndex =\n          this.hasContentBlockIndex(responseMetadata);\n        if (hasContentBlockIndex) {\n          cleanedMetadata = this.removeContentBlockIndex(\n            responseMetadata\n          ) as Record<string, unknown>;\n          needsModification = true;\n        }\n\n        // Extract cache tokens from metadata.usage (Bedrock streaming format)\n        // The metadata chunk contains usage with cacheReadInputTokens/cacheWriteInputTokens\n        const metadata = responseMetadata.metadata as\n          | Record<string, unknown>\n          | undefined;\n        const usage = (metadata?.usage ?? responseMetadata.usage) as\n          | Record<string, unknown>\n          | undefined;\n\n        let enhancedUsageMetadata: UsageMetadata | undefined =\n          chunk.message.usage_metadata;\n\n        if (usage) {\n          const cacheRead =\n            (usage.cacheReadInputTokens as number | undefined) ?? 0;\n          const cacheWrite =\n            (usage.cacheWriteInputTokens as number | undefined) ?? 0;\n          const inputTokens = (usage.inputTokens as number | undefined) ?? 0;\n          const outputTokens = (usage.outputTokens as number | undefined) ?? 0;\n\n          if (cacheRead > 0 || cacheWrite > 0) {\n            needsModification = true;\n            enhancedUsageMetadata = {\n              input_tokens:\n                chunk.message.usage_metadata?.input_tokens ?? inputTokens,\n              output_tokens:\n                chunk.message.usage_metadata?.output_tokens ?? outputTokens,\n              total_tokens:\n                chunk.message.usage_metadata?.total_tokens ??\n                (usage.totalTokens as number | undefined) ??\n                0,\n              input_token_details: {\n                cache_read: cacheRead,\n                cache_creation: cacheWrite,\n              },\n            };\n          }\n        }\n\n        if (needsModification) {\n          yield new ChatGenerationChunk({\n            text: chunk.text,\n            message: new AIMessageChunk({\n              ...chunk.message,\n              response_metadata: cleanedMetadata,\n              usage_metadata: enhancedUsageMetadata,\n            }),\n            generationInfo: chunk.generationInfo,\n          });\n          continue;\n        }\n      }\n\n      yield chunk;\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","messages","AIMessageChunk","ChatGenerationChunk"],"mappings":";;;;;;AAAA;;;;;;;;;;;;;;;;;;;;;AAqBG;AAeG,MAAO,yBAA0B,SAAQA,uBAAmB,CAAA;AAChE,IAAA,WAAW;AAEX,IAAA,WAAA,CAAY,MAAuC,EAAA;QACjD,KAAK,CAAC,MAAM,CAAC;QACb,IAAI,CAAC,WAAW,GAAG,MAAM,EAAE,WAAW,IAAI,KAAK;;AAGjD,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,uBAAuB;;AAGhC;;;;;;;;;;AAUG;AACH,IAAA,gBAAgB,CACd,OAAmC,EAAA;QAEnC,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;;;QAI9C,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,EAAE;AACxC,QAAA,MAAM,aAAa,GACjB,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAC;QAE7D,IACE,IAAI,CAAC,WAAW;YAChB,aAAa;YACb,MAAM,CAAC,UAAU,EAAE,KAAK;YACxB,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;YACtC,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAClC;;;YAGA,MAAM,SAAS,GAAmC,EAAE;YACpD,MAAM,QAAQ,GAAmC,EAAE;YAInD,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,UAAU,CAAC,KAAK,EAAE;;gBAE1C,MAAM,QAAQ,GACX,IAAyC,CAAC,QAAQ,EAAE,IAAI,IAAI,EAAE;AACjE,gBAAA,IAAI,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,EAAE;AAC9B,oBAAA,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC;;qBAEd;AACL,oBAAA,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC;;;;;YAOxB,MAAM,cAAc,GAAmC,EAAE;;AAGzD,YAAA,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE;AACxB,gBAAA,cAAc,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC;AACjC,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;;AAI1D,YAAA,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE;AACvB,gBAAA,cAAc,CAAC,IAAI,CAAC,GAAG,QAAQ,CAAC;AAChC,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;;AAI1D,YAAA,IAAI,cAAc,CAAC,MAAM,KAAK,CAAC,EAAE;AAC/B,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;AAG1D,YAAA,MAAM,CAAC,UAAU,CAAC,KAAK,GAAG,cAAc;;AAG1C,QAAA,OAAO,MAAM;;AAGf;;;;AAIG;IACH,OAAO,qBAAqB,CAC1BC,UAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;AAErC,QAAA,MAAM,UAAU,GAAG,KAAK,CAAC,qBAAqB,CAC5CA,UAAQ,EACR,OAAO,EACP,UAAU,CACX;AAED,QAAA,WAAW,MAAM,KAAK,IAAI,UAAU,EAAE;;AAEpC,YAAA,IACE,KAAK,CAAC,OAAO,YAAYC,uBAAc;gBACtC,KAAK,CAAC,OAAmC,CAAC,iBAAiB;gBAC5D,OAAO,KAAK,CAAC,OAAO,CAAC,iBAAiB,KAAK,QAAQ,EACnD;AACA,gBAAA,MAAM,gBAAgB,GAAG,KAAK,CAAC,OAAO,CAAC,iBAGtC;gBACD,IAAI,iBAAiB,GAAG,KAAK;gBAC7B,IAAI,eAAe,GAAG,gBAAgB;;gBAGtC,MAAM,oBAAoB,GACxB,IAAI,CAAC,oBAAoB,CAAC,gBAAgB,CAAC;gBAC7C,IAAI,oBAAoB,EAAE;AACxB,oBAAA,eAAe,GAAG,IAAI,CAAC,uBAAuB,CAC5C,gBAAgB,CACU;oBAC5B,iBAAiB,GAAG,IAAI;;;;AAK1B,gBAAA,MAAM,QAAQ,GAAG,gBAAgB,CAAC,QAErB;gBACb,MAAM,KAAK,IAAI,QAAQ,EAAE,KAAK,IAAI,gBAAgB,CAAC,KAAK,CAE3C;AAEb,gBAAA,IAAI,qBAAqB,GACvB,KAAK,CAAC,OAAO,CAAC,cAAc;gBAE9B,IAAI,KAAK,EAAE;AACT,oBAAA,MAAM,SAAS,GACZ,KAAK,CAAC,oBAA2C,IAAI,CAAC;AACzD,oBAAA,MAAM,UAAU,GACb,KAAK,CAAC,qBAA4C,IAAI,CAAC;AAC1D,oBAAA,MAAM,WAAW,GAAI,KAAK,CAAC,WAAkC,IAAI,CAAC;AAClE,oBAAA,MAAM,YAAY,GAAI,KAAK,CAAC,YAAmC,IAAI,CAAC;oBAEpE,IAAI,SAAS,GAAG,CAAC,IAAI,UAAU,GAAG,CAAC,EAAE;wBACnC,iBAAiB,GAAG,IAAI;AACxB,wBAAA,qBAAqB,GAAG;4BACtB,YAAY,EACV,KAAK,CAAC,OAAO,CAAC,cAAc,EAAE,YAAY,IAAI,WAAW;4BAC3D,aAAa,EACX,KAAK,CAAC,OAAO,CAAC,cAAc,EAAE,aAAa,IAAI,YAAY;AAC7D,4BAAA,YAAY,EACV,KAAK,CAAC,OAAO,CAAC,cAAc,EAAE,YAAY;AACzC,gCAAA,KAAK,CAAC,WAAkC;gCACzC,CAAC;AACH,4BAAA,mBAAmB,EAAE;AACnB,gCAAA,UAAU,EAAE,SAAS;AACrB,gCAAA,cAAc,EAAE,UAAU;AAC3B,6BAAA;yBACF;;;gBAIL,IAAI,iBAAiB,EAAE;oBACrB,MAAM,IAAIC,2BAAmB,CAAC;wBAC5B,IAAI,EAAE,KAAK,CAAC,IAAI;wBAChB,OAAO,EAAE,IAAID,uBAAc,CAAC;4BAC1B,GAAG,KAAK,CAAC,OAAO;AAChB,4BAAA,iBAAiB,EAAE,eAAe;AAClC,4BAAA,cAAc,EAAE,qBAAqB;yBACtC,CAAC;wBACF,cAAc,EAAE,KAAK,CAAC,cAAc;AACrC,qBAAA,CAAC;oBACF;;;AAIJ,YAAA,MAAM,KAAK;;;AAIf;;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:\n *\n * - Prompt caching support for Bedrock Converse API (Illuma feature)\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 * PROMPT CACHING:\n * When promptCache: true is set, this wrapper adds cachePoint markers to the tools array\n * to enable Bedrock prompt caching for tool definitions. This allows tool schemas to be\n * cached and reused across requests, reducing latency and costs.\n *\n * CACHE TOKEN EXTRACTION:\n * LangChain AWS doesn't extract cacheReadInputTokens/cacheWriteInputTokens from Bedrock's\n * response. This wrapper adds input_token_details to usage_metadata with cache information.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport type { BaseMessage, UsageMetadata } 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';\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 * - promptCache: Enable Bedrock prompt caching for tool definitions\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   * Enable Bedrock prompt caching for tool definitions.\n   * When true, adds cachePoint markers to tools array.\n   */\n  promptCache?: boolean;\n\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  /** Enable Bedrock prompt caching for tool definitions */\n  promptCache: boolean;\n\n  /** Application Inference Profile ARN to use instead of model ID */\n  applicationInferenceProfile?: string;\n\n  /** Service tier for model invocation */\n  serviceTier?: ServiceTierType;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.promptCache = fields?.promptCache ?? false;\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n  }\n\n  static lc_name(): string {\n    return 'IllumaBedrockConverse';\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:\n   * 1. Add cachePoint to tools when promptCache is enabled\n   * 2. Add serviceTier support\n   *\n   * CACHING STRATEGY: Separate cachePoints for core tools and MCP tools\n   * - Core tools (web_search, execute_code, etc.) are stable → cache first\n   * - MCP tools (have '_mcp_' in name) are dynamic → cache separately after\n   * - This allows core tools to stay cached when MCP selection changes\n   *\n   * NOTE: Only Claude models support cachePoint - Nova and other models will reject it.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const params = super.invocationParams(options);\n\n    // Add cachePoint to tools array if promptCache is enabled and tools exist\n    // Only Claude models support cachePoint - check model name\n    const modelId = this.model.toLowerCase();\n    const isClaudeModel =\n      modelId.includes('claude') || modelId.includes('anthropic');\n\n    if (\n      this.promptCache &&\n      isClaudeModel &&\n      params.toolConfig?.tools &&\n      Array.isArray(params.toolConfig.tools) &&\n      params.toolConfig.tools.length > 0\n    ) {\n      // Separate core tools from MCP tools\n      // MCP tools have '_mcp_' in their name (e.g., 'search_emails_mcp_Google-Workspace')\n      const coreTools: typeof params.toolConfig.tools = [];\n      const mcpTools: typeof params.toolConfig.tools = [];\n\n      for (const tool of params.toolConfig.tools) {\n        // Check if tool has a name property with '_mcp_' pattern\n        const toolName =\n          (tool as { toolSpec?: { name?: string } }).toolSpec?.name ?? '';\n        if (toolName.includes('_mcp_')) {\n          mcpTools.push(tool);\n        } else {\n          coreTools.push(tool);\n        }\n      }\n\n      // Build tools array with strategic cachePoints:\n      // [CoreTool1, CoreTool2, cachePoint] + [MCPTool1, MCPTool2, cachePoint]\n      const toolsWithCache: typeof params.toolConfig.tools = [];\n\n      // Add core tools with cachePoint (if any)\n      if (coreTools.length > 0) {\n        toolsWithCache.push(...coreTools);\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      // Add MCP tools with their own cachePoint (if any)\n      if (mcpTools.length > 0) {\n        toolsWithCache.push(...mcpTools);\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      // If no tools at all (shouldn't happen but safety check)\n      if (toolsWithCache.length === 0) {\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      params.toolConfig.tools = toolsWithCache;\n    }\n\n    // Add serviceTier support\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...params,\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   * 3. Extract cacheReadInputTokens/cacheWriteInputTokens and add to usage_metadata\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 and enhance chunk\n        yield this.processChunk(chunk);\n      }\n    } finally {\n      // Restore original model\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Process a chunk by:\n   * 1. Removing contentBlockIndex from response_metadata\n   * 2. Extracting cache token information from Bedrock's usage data\n   */\n  private processChunk(chunk: ChatGenerationChunk): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const responseMetadata = message.response_metadata as Record<\n      string,\n      unknown\n    >;\n    let needsModification = false;\n    let cleanedMetadata = responseMetadata;\n\n    // Check if contentBlockIndex exists anywhere in response_metadata\n    const hasContentBlockIndex = this.hasContentBlockIndex(responseMetadata);\n    if (hasContentBlockIndex) {\n      cleanedMetadata = this.removeContentBlockIndex(\n        responseMetadata\n      ) as Record<string, unknown>;\n      needsModification = true;\n    }\n\n    // Extract cache tokens from metadata.usage (Bedrock streaming format)\n    // The metadata chunk contains usage with cacheReadInputTokens/cacheWriteInputTokens\n    const metadata = responseMetadata.metadata as\n      | Record<string, unknown>\n      | undefined;\n    const usage = (metadata?.usage ?? responseMetadata.usage) as\n      | Record<string, unknown>\n      | undefined;\n\n    let enhancedUsageMetadata: UsageMetadata | undefined =\n      message.usage_metadata;\n\n    if (usage) {\n      const cacheRead = (usage.cacheReadInputTokens as number | undefined) ?? 0;\n      const cacheWrite =\n        (usage.cacheWriteInputTokens as number | undefined) ?? 0;\n      const inputTokens = (usage.inputTokens as number | undefined) ?? 0;\n      const outputTokens = (usage.outputTokens as number | undefined) ?? 0;\n\n      if (cacheRead > 0 || cacheWrite > 0) {\n        needsModification = true;\n        enhancedUsageMetadata = {\n          input_tokens: message.usage_metadata?.input_tokens ?? inputTokens,\n          output_tokens: message.usage_metadata?.output_tokens ?? outputTokens,\n          total_tokens:\n            message.usage_metadata?.total_tokens ??\n            (usage.totalTokens as number | undefined) ??\n            0,\n          input_token_details: {\n            cache_read: cacheRead,\n            cache_creation: cacheWrite,\n          },\n        };\n      }\n    }\n\n    if (needsModification) {\n      return new ChatGenerationChunk({\n        text: chunk.text,\n        message: new AIMessageChunk({\n          ...message,\n          response_metadata: cleanedMetadata,\n          usage_metadata: enhancedUsageMetadata,\n        }),\n        generationInfo: chunk.generationInfo,\n      });\n    }\n\n    return chunk;\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;;;;;;;;;;;;;;;;;;;;;;;;;AAyBG;AA+DG,MAAO,yBAA0B,SAAQA,uBAAmB,CAAA;;AAEhE,IAAA,WAAW;;AAGX,IAAA,2BAA2B;;AAG3B,IAAA,WAAW;AAEX,IAAA,WAAA,CAAY,MAAuC,EAAA;QACjD,KAAK,CAAC,MAAM,CAAC;QACb,IAAI,CAAC,WAAW,GAAG,MAAM,EAAE,WAAW,IAAI,KAAK;AAC/C,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,uBAAuB;;AAGhC;;;AAGG;IACO,UAAU,GAAA;AAClB,QAAA,OAAO,IAAI,CAAC,2BAA2B,IAAI,IAAI,CAAC,KAAK;;AAGvD;;;;;;;;;;;AAWG;AACM,IAAA,gBAAgB,CACvB,OAA0E,EAAA;QAI1E,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;;;QAI9C,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,EAAE;AACxC,QAAA,MAAM,aAAa,GACjB,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAC;QAE7D,IACE,IAAI,CAAC,WAAW;YAChB,aAAa;YACb,MAAM,CAAC,UAAU,EAAE,KAAK;YACxB,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;YACtC,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAClC;;;YAGA,MAAM,SAAS,GAAmC,EAAE;YACpD,MAAM,QAAQ,GAAmC,EAAE;YAEnD,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,UAAU,CAAC,KAAK,EAAE;;gBAE1C,MAAM,QAAQ,GACX,IAAyC,CAAC,QAAQ,EAAE,IAAI,IAAI,EAAE;AACjE,gBAAA,IAAI,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,EAAE;AAC9B,oBAAA,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC;;qBACd;AACL,oBAAA,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC;;;;;YAMxB,MAAM,cAAc,GAAmC,EAAE;;AAGzD,YAAA,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE;AACxB,gBAAA,cAAc,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC;AACjC,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;;AAI1D,YAAA,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE;AACvB,gBAAA,cAAc,CAAC,IAAI,CAAC,GAAG,QAAQ,CAAC;AAChC,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;;AAI1D,YAAA,IAAI,cAAc,CAAC,MAAM,KAAK,CAAC,EAAE;AAC/B,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;AAG1D,YAAA,MAAM,CAAC,UAAU,CAAC,KAAK,GAAG,cAAc;;;QAI1C,MAAM,eAAe,GAAG,OAAO,EAAE,WAAW,IAAI,IAAI,CAAC,WAAW;QAEhE,OAAO;AACL,YAAA,GAAG,MAAM;AACT,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;;;;;;;;AAQG;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,YAAY,CAAC,KAAK,CAAC;;;gBAExB;;AAER,YAAA,IAAI,CAAC,KAAK,GAAG,aAAa;;;AAI9B;;;;AAIG;AACK,IAAA,YAAY,CAAC,KAA0B,EAAA;AAC7C,QAAA,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO;AAC7B,QAAA,IAAI,EAAE,OAAO,YAAYC,uBAAc,CAAC,EAAE;AACxC,YAAA,OAAO,KAAK;;AAGd,QAAA,MAAM,gBAAgB,GAAG,OAAO,CAAC,iBAGhC;QACD,IAAI,iBAAiB,GAAG,KAAK;QAC7B,IAAI,eAAe,GAAG,gBAAgB;;QAGtC,MAAM,oBAAoB,GAAG,IAAI,CAAC,oBAAoB,CAAC,gBAAgB,CAAC;QACxE,IAAI,oBAAoB,EAAE;AACxB,YAAA,eAAe,GAAG,IAAI,CAAC,uBAAuB,CAC5C,gBAAgB,CACU;YAC5B,iBAAiB,GAAG,IAAI;;;;AAK1B,QAAA,MAAM,QAAQ,GAAG,gBAAgB,CAAC,QAErB;QACb,MAAM,KAAK,IAAI,QAAQ,EAAE,KAAK,IAAI,gBAAgB,CAAC,KAAK,CAE3C;AAEb,QAAA,IAAI,qBAAqB,GACvB,OAAO,CAAC,cAAc;QAExB,IAAI,KAAK,EAAE;AACT,YAAA,MAAM,SAAS,GAAI,KAAK,CAAC,oBAA2C,IAAI,CAAC;AACzE,YAAA,MAAM,UAAU,GACb,KAAK,CAAC,qBAA4C,IAAI,CAAC;AAC1D,YAAA,MAAM,WAAW,GAAI,KAAK,CAAC,WAAkC,IAAI,CAAC;AAClE,YAAA,MAAM,YAAY,GAAI,KAAK,CAAC,YAAmC,IAAI,CAAC;YAEpE,IAAI,SAAS,GAAG,CAAC,IAAI,UAAU,GAAG,CAAC,EAAE;gBACnC,iBAAiB,GAAG,IAAI;AACxB,gBAAA,qBAAqB,GAAG;AACtB,oBAAA,YAAY,EAAE,OAAO,CAAC,cAAc,EAAE,YAAY,IAAI,WAAW;AACjE,oBAAA,aAAa,EAAE,OAAO,CAAC,cAAc,EAAE,aAAa,IAAI,YAAY;AACpE,oBAAA,YAAY,EACV,OAAO,CAAC,cAAc,EAAE,YAAY;AACnC,wBAAA,KAAK,CAAC,WAAkC;wBACzC,CAAC;AACH,oBAAA,mBAAmB,EAAE;AACnB,wBAAA,UAAU,EAAE,SAAS;AACrB,wBAAA,cAAc,EAAE,UAAU;AAC3B,qBAAA;iBACF;;;QAIL,IAAI,iBAAiB,EAAE;YACrB,OAAO,IAAIC,2BAAmB,CAAC;gBAC7B,IAAI,EAAE,KAAK,CAAC,IAAI;gBAChB,OAAO,EAAE,IAAID,uBAAc,CAAC;AAC1B,oBAAA,GAAG,OAAO;AACV,oBAAA,iBAAiB,EAAE,eAAe;AAClC,oBAAA,cAAc,EAAE,qBAAqB;iBACtC,CAAC;gBACF,cAAc,EAAE,KAAK,CAAC,cAAc;AACrC,aAAA,CAAC;;AAGJ,QAAA,OAAO,KAAK;;AAGd;;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/cjs/main.cjs

@@ -17,9 +17,10 @@ var Calculator = require('./tools/Calculator.cjs');
 var CodeExecutor = require('./tools/CodeExecutor.cjs');
 var BrowserTools = require('./tools/BrowserTools.cjs');
 var ProgrammaticToolCalling = require('./tools/ProgrammaticToolCalling.cjs');
-var ToolSearchRegex = require('./tools/ToolSearchRegex.cjs');
+var ToolSearch = require('./tools/ToolSearch.cjs');
 var handlers = require('./tools/handlers.cjs');
 var tool = require('./tools/search/tool.cjs');
+var validate = require('./schemas/validate.cjs');
 var _enum = require('./common/enum.cjs');
 var graph = require('./utils/graph.cjs');
 var llm = require('./utils/llm.cjs');
@@ -93,17 +94,31 @@ exports.formatCompletedResponse = ProgrammaticToolCalling.formatCompletedRespons
 exports.makeRequest = ProgrammaticToolCalling.makeRequest;
 exports.normalizeToPythonIdentifier = ProgrammaticToolCalling.normalizeToPythonIdentifier;
 exports.unwrapToolResponse = ProgrammaticToolCalling.unwrapToolResponse;
-exports.countNestedGroups = ToolSearchRegex.countNestedGroups;
-exports.createToolSearchRegexTool = ToolSearchRegex.createToolSearchRegexTool;
-exports.escapeRegexSpecialChars = ToolSearchRegex.escapeRegexSpecialChars;
-exports.hasNestedQuantifiers = ToolSearchRegex.hasNestedQuantifiers;
-exports.isDangerousPattern = ToolSearchRegex.isDangerousPattern;
-exports.sanitizeRegex = ToolSearchRegex.sanitizeRegex;
+exports.countNestedGroups = ToolSearch.countNestedGroups;
+exports.createToolSearch = ToolSearch.createToolSearch;
+exports.escapeRegexSpecialChars = ToolSearch.escapeRegexSpecialChars;
+exports.extractMcpServerName = ToolSearch.extractMcpServerName;
+exports.formatServerListing = ToolSearch.formatServerListing;
+exports.getAvailableMcpServers = ToolSearch.getAvailableMcpServers;
+exports.getBaseToolName = ToolSearch.getBaseToolName;
+exports.getDeferredToolsListing = ToolSearch.getDeferredToolsListing;
+exports.hasNestedQuantifiers = ToolSearch.hasNestedQuantifiers;
+exports.isDangerousPattern = ToolSearch.isDangerousPattern;
+exports.isFromAnyMcpServer = ToolSearch.isFromAnyMcpServer;
+exports.isFromMcpServer = ToolSearch.isFromMcpServer;
+exports.normalizeServerFilter = ToolSearch.normalizeServerFilter;
+exports.performLocalSearch = ToolSearch.performLocalSearch;
+exports.sanitizeRegex = ToolSearch.sanitizeRegex;
 exports.handleServerToolResult = handlers.handleServerToolResult;
 exports.handleToolCallChunks = handlers.handleToolCallChunks;
 exports.handleToolCalls = handlers.handleToolCalls;
 exports.toolResultTypes = handlers.toolResultTypes;
 exports.createSearchTool = tool.createSearchTool;
+exports.createValidationErrorMessage = validate.createValidationErrorMessage;
+exports.isValidJsonSchema = validate.isValidJsonSchema;
+exports.normalizeJsonSchema = validate.normalizeJsonSchema;
+exports.validateStructuredOutput = validate.validateStructuredOutput;
+exports.zodToJsonSchema = validate.zodToJsonSchema;
 Object.defineProperty(exports, "Callback", {
 	enumerable: true,
 	get: function () { return _enum.Callback; }

package/dist/cjs/main.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"main.cjs","sources":[],"sourcesContent":[],"names":[],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;"}
+{"version":3,"file":"main.cjs","sources":[],"sourcesContent":[],"names":[],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;"}

package/dist/cjs/messages/cache.cjs

@@ -9,13 +9,97 @@ const debugCache = (message, data) => {
         console.log(`[Cache] ${message}`, data !== undefined ? JSON.stringify(data, null, 2) : '');
     }
 };
+/**
+ * Deep clones a message's content to prevent mutation of the original.
+ */
+function deepCloneContent(content) {
+    if (typeof content === 'string') {
+        return content;
+    }
+    if (Array.isArray(content)) {
+        return content.map((block) => ({ ...block }));
+    }
+    return content;
+}
+/**
+ * Clones a message with deep-cloned content, explicitly excluding LangChain
+ * serialization metadata to prevent coercion issues.
+ * Keeps lc_kwargs in sync with content to prevent LangChain serialization issues.
+ */
+function cloneMessage(message, content) {
+    const { lc_kwargs: _lc_kwargs, lc_serializable: _lc_serializable, lc_namespace: _lc_namespace, ...rest } = message;
+    const cloned = { ...rest, content };
+    // Sync lc_kwargs.content with the new content to prevent LangChain coercion issues
+    const lcKwargs = message.lc_kwargs;
+    if (lcKwargs != null) {
+        cloned.lc_kwargs = {
+            ...lcKwargs,
+            content: content,
+        };
+    }
+    // LangChain messages don't have a direct 'role' property - derive it from getType()
+    if ('getType' in message &&
+        typeof message.getType === 'function' &&
+        !('role' in cloned)) {
+        const msgType = message.getType();
+        const roleMap = {
+            human: 'user',
+            ai: 'assistant',
+            system: 'system',
+            tool: 'tool',
+        };
+        cloned.role = roleMap[msgType] || msgType;
+    }
+    return cloned;
+}
+/**
+ * Checks if a content block is a cache point
+ */
+function isCachePoint(block) {
+    return 'cachePoint' in block && !('type' in block);
+}
+/**
+ * Checks if a message's content needs cache control stripping.
+ * Returns true if content has cachePoint blocks or cache_control fields.
+ */
+function needsCacheStripping(content) {
+    for (let i = 0; i < content.length; i++) {
+        const block = content[i];
+        if (isCachePoint(block))
+            return true;
+        if ('cache_control' in block)
+            return true;
+    }
+    return false;
+}
+/**
+ * Checks if a message's content has Anthropic cache_control fields.
+ */
+function hasAnthropicCacheControl(content) {
+    for (let i = 0; i < content.length; i++) {
+        if ('cache_control' in content[i])
+            return true;
+    }
+    return false;
+}
+/**
+ * Checks if a message's content has Bedrock cachePoint blocks.
+ */
+function hasBedrockCachePoint(content) {
+    for (let i = 0; i < content.length; i++) {
+        if (isCachePoint(content[i]))
+            return true;
+    }
+    return false;
+}
 /**
  * Anthropic API: Adds cache control to the appropriate user messages in the payload.
  * Strips ALL existing cache control (both Anthropic and Bedrock formats) from all messages,
  * then adds fresh cache control to the last 2 user messages in a single backward pass.
  * This ensures we don't accumulate stale cache points across multiple turns.
+ * Returns a new array - only clones messages that require modification.
  * @param messages - The array of message objects.
- * @returns - The updated array of message objects with cache control added.
+ * @returns - A new array of message objects with cache control added.
  */
 function addCacheControl(messages) {
     if (!Array.isArray(messages) || messages.length < 2) {
@@ -24,55 +108,59 @@ function addCacheControl(messages) {
     const updatedMessages = [...messages];
     let userMessagesModified = 0;
     for (let i = updatedMessages.length - 1; i >= 0; i--) {
-        const message = updatedMessages[i];
-        const isUserMessage = ('getType' in message && message.getType() === 'human') ||
-            ('role' in message && message.role === 'user');
-        if (Array.isArray(message.content)) {
-            message.content = message.content.filter((block) => !isCachePoint(block));
-            for (let j = 0; j < message.content.length; j++) {
-                const block = message.content[j];
+        const originalMessage = updatedMessages[i];
+        const content = originalMessage.content;
+        const isUserMessage = ('getType' in originalMessage && originalMessage.getType() === 'human') ||
+            ('role' in originalMessage && originalMessage.role === 'user');
+        const hasArrayContent = Array.isArray(content);
+        const needsStripping = hasArrayContent &&
+            needsCacheStripping(content);
+        const needsCacheAdd = userMessagesModified < 2 &&
+            isUserMessage &&
+            (typeof content === 'string' || hasArrayContent);
+        if (!needsStripping && !needsCacheAdd) {
+            continue;
+        }
+        let workingContent;
+        if (hasArrayContent) {
+            workingContent = deepCloneContent(content).filter((block) => !isCachePoint(block));
+            for (let j = 0; j < workingContent.length; j++) {
+                const block = workingContent[j];
                 if ('cache_control' in block) {
                     delete block.cache_control;
                 }
             }
         }
+        else if (typeof content === 'string') {
+            workingContent = [
+                { type: 'text', text: content },
+            ];
+        }
+        else {
+            workingContent = [];
+        }
         if (userMessagesModified >= 2 || !isUserMessage) {
+            updatedMessages[i] = cloneMessage(originalMessage, workingContent);
             continue;
         }
-        if (typeof message.content === 'string') {
-            message.content = [
-                {
-                    type: 'text',
-                    text: message.content,
-                    cache_control: { type: 'ephemeral' },
-                },
-            ];
-            userMessagesModified++;
-        }
-        else if (Array.isArray(message.content)) {
-            for (let j = message.content.length - 1; j >= 0; j--) {
-                const contentPart = message.content[j];
-                if ('type' in contentPart && contentPart.type === 'text') {
-                    contentPart.cache_control = {
-                        type: 'ephemeral',
-                    };
-                    userMessagesModified++;
-                    break;
-                }
+        for (let j = workingContent.length - 1; j >= 0; j--) {
+            const contentPart = workingContent[j];
+            if ('type' in contentPart && contentPart.type === 'text') {
+                contentPart.cache_control = {
+                    type: 'ephemeral',
+                };
+                userMessagesModified++;
+                break;
             }
         }
+        updatedMessages[i] = cloneMessage(originalMessage, workingContent);
     }
     return updatedMessages;
 }
-/**
- * Checks if a content block is a cache point
- */
-function isCachePoint(block) {
-    return 'cachePoint' in block && !('type' in block);
-}
 /**
  * Removes all Anthropic cache_control fields from messages
  * Used when switching from Anthropic to Bedrock provider
+ * Returns a new array - only clones messages that require modification.
  */
 function stripAnthropicCacheControl(messages) {
     if (!Array.isArray(messages)) {
@@ -80,22 +168,26 @@ function stripAnthropicCacheControl(messages) {
     }
     const updatedMessages = [...messages];
     for (let i = 0; i < updatedMessages.length; i++) {
-        const message = updatedMessages[i];
-        const content = message.content;
-        if (Array.isArray(content)) {
-            for (let j = 0; j < content.length; j++) {
-                const block = content[j];
-                if ('cache_control' in block) {
-                    delete block.cache_control;
-                }
+        const originalMessage = updatedMessages[i];
+        const content = originalMessage.content;
+        if (!Array.isArray(content) || !hasAnthropicCacheControl(content)) {
+            continue;
+        }
+        const clonedContent = deepCloneContent(content);
+        for (let j = 0; j < clonedContent.length; j++) {
+            const block = clonedContent[j];
+            if ('cache_control' in block) {
+                delete block.cache_control;
             }
         }
+        updatedMessages[i] = cloneMessage(originalMessage, clonedContent);
     }
     return updatedMessages;
 }
 /**
  * Removes all Bedrock cachePoint blocks from messages
  * Used when switching from Bedrock to Anthropic provider
+ * Returns a new array - only clones messages that require modification.
  */
 function stripBedrockCacheControl(messages) {
     if (!Array.isArray(messages)) {
@@ -103,11 +195,13 @@ function stripBedrockCacheControl(messages) {
     }
     const updatedMessages = [...messages];
     for (let i = 0; i < updatedMessages.length; i++) {
-        const message = updatedMessages[i];
-        const content = message.content;
-        if (Array.isArray(content)) {
-            message.content = content.filter((block) => !isCachePoint(block));
+        const originalMessage = updatedMessages[i];
+        const content = originalMessage.content;
+        if (!Array.isArray(content) || !hasBedrockCachePoint(content)) {
+            continue;
         }
+        const clonedContent = deepCloneContent(content).filter((block) => !isCachePoint(block));
+        updatedMessages[i] = cloneMessage(originalMessage, clonedContent);
     }
     return updatedMessages;
 }