langchain 1.0.0-alpha.7 → 1.0.0-alpha.9

package/dist/agents/middlewareAgent/middleware/promptCaching.js

@@ -1,197 +0,0 @@
-import { createMiddleware } from "../middleware.js";
-import { z } from "zod/v3";
-
-//#region src/agents/middlewareAgent/middleware/promptCaching.ts
-const DEFAULT_ENABLE_CACHING = true;
-const DEFAULT_TTL = "5m";
-const DEFAULT_MIN_MESSAGES_TO_CACHE = 3;
-const DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR = "warn";
-const contextSchema = z.object({
-	enableCaching: z.boolean().default(DEFAULT_ENABLE_CACHING),
-	ttl: z.enum(["5m", "1h"]).default(DEFAULT_TTL),
-	minMessagesToCache: z.number().default(DEFAULT_MIN_MESSAGES_TO_CACHE),
-	unsupportedModelBehavior: z.enum([
-		"ignore",
-		"warn",
-		"raise"
-	]).default(DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR)
-});
-var PromptCachingMiddlewareError = class extends Error {
-	constructor(message) {
-		super(message);
-		this.name = "PromptCachingMiddlewareError";
-	}
-};
-/**
-* Creates a prompt caching middleware for Anthropic models to optimize API usage.
-*
-* This middleware automatically adds cache control headers to the last messages when using Anthropic models,
-* enabling their prompt caching feature. This can significantly reduce costs for applications with repetitive
-* prompts, long system messages, or extensive conversation histories.
-*
-* ## How It Works
-*
-* The middleware intercepts model requests and adds cache control metadata that tells Anthropic's
-* API to cache processed prompt prefixes. On subsequent requests with matching prefixes, the
-* cached representations are reused, skipping redundant token processing.
-*
-* ## Benefits
-*
-* - **Cost Reduction**: Avoid reprocessing the same tokens repeatedly (up to 90% savings on cached portions)
-* - **Lower Latency**: Cached prompts are processed faster as embeddings are pre-computed
-* - **Better Scalability**: Reduced computational load enables handling more requests
-* - **Consistent Performance**: Stable response times for repetitive queries
-*
-* @param middlewareOptions - Configuration options for the caching behavior
-* @param middlewareOptions.enableCaching - Whether to enable prompt caching (default: `true`)
-* @param middlewareOptions.ttl - Cache time-to-live: `"5m"` for 5 minutes or `"1h"` for 1 hour (default: `"5m"`)
-* @param middlewareOptions.minMessagesToCache - Minimum number of messages required before caching is applied (default: `3`)
-* @param middlewareOptions.unsupportedModelBehavior - The behavior to take when an unsupported model is used (default: `"warn"`)
-*
-* @returns A middleware instance that can be passed to `createAgent`
-*
-* @throws {Error} If used with non-Anthropic models
-*
-* @example
-* Basic usage with default settings
-* ```typescript
-* import { createAgent } from "langchain";
-* import { anthropicPromptCachingMiddleware } from "langchain/middleware";
-*
-* const agent = createAgent({
-*   model: "anthropic:claude-3-5-sonnet",
-*   middleware: [
-*     anthropicPromptCachingMiddleware()
-*   ]
-* });
-* ```
-*
-* @example
-* Custom configuration for longer conversations
-* ```typescript
-* const cachingMiddleware = anthropicPromptCachingMiddleware({
-*   ttl: "1h",  // Cache for 1 hour instead of default 5 minutes
-*   minMessagesToCache: 5  // Only cache after 5 messages
-* });
-*
-* const agent = createAgent({
-*   model: "anthropic:claude-3-5-sonnet",
-*   systemPrompt: "You are a helpful assistant with deep knowledge of...", // Long system prompt
-*   middleware: [cachingMiddleware]
-* });
-* ```
-*
-* @example
-* Conditional caching based on runtime context
-* ```typescript
-* const agent = createAgent({
-*   model: "anthropic:claude-3-5-sonnet",
-*   middleware: [
-*     anthropicPromptCachingMiddleware({
-*       enableCaching: true,
-*       ttl: "5m"
-*     })
-*   ]
-* });
-*
-* // Disable caching for specific requests
-* await agent.invoke(
-*   { messages: [new HumanMessage("Process this without caching")] },
-*   {
-*     configurable: {
-*       middleware_context: { enableCaching: false }
-*     }
-*   }
-* );
-* ```
-*
-* @example
-* Optimal setup for customer support chatbot
-* ```typescript
-* const supportAgent = createAgent({
-*   model: "anthropic:claude-3-5-sonnet",
-*   systemPrompt: `You are a customer support agent for ACME Corp.
-*
-*     Company policies:
-*     - Always be polite and professional
-*     - Refer to knowledge base for product information
-*     - Escalate billing issues to human agents
-*     ... (extensive policies and guidelines)
-*   `,
-*   tools: [searchKnowledgeBase, createTicket, checkOrderStatus],
-*   middleware: [
-*     anthropicPromptCachingMiddleware({
-*       ttl: "1h",  // Long TTL for stable system prompt
-*       minMessagesToCache: 1  // Cache immediately due to large system prompt
-*     })
-*   ]
-* });
-* ```
-*
-* @remarks
-* - **Anthropic Only**: This middleware only works with Anthropic models and will throw an error if used with other providers
-* - **Automatic Application**: Caching is applied automatically when message count exceeds `minMessagesToCache`
-* - **Cache Scope**: Caches are isolated per API key and cannot be shared across different keys
-* - **TTL Options**: Only supports "5m" (5 minutes) and "1h" (1 hour) as TTL values per Anthropic's API
-* - **Best Use Cases**: Long system prompts, multi-turn conversations, repetitive queries, RAG applications
-* - **Cost Impact**: Cached tokens are billed at 10% of the base input token price, cache writes are billed at 25% of the base
-*
-* @see {@link createAgent} for agent creation
-* @see {@link https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching} Anthropic's prompt caching documentation
-* @public
-*/
-function anthropicPromptCachingMiddleware(middlewareOptions) {
-	return createMiddleware({
-		name: "PromptCachingMiddleware",
-		contextSchema,
-		modifyModelRequest: (options, state, runtime) => {
-			/**
-			* If the runtime values match the schema default values, use the middleware option
-			* values otherwise use the runtime values. This allows to apply general configurations
-			* for all invocations, and override them for specific invocations.
-			*/
-			const enableCaching = runtime.context.enableCaching === DEFAULT_ENABLE_CACHING ? middlewareOptions?.enableCaching ?? runtime.context.enableCaching : runtime.context.enableCaching ?? middlewareOptions?.enableCaching;
-			const ttl = runtime.context.ttl === DEFAULT_TTL ? middlewareOptions?.ttl ?? runtime.context.ttl : runtime.context.ttl ?? middlewareOptions?.ttl;
-			const minMessagesToCache = runtime.context.minMessagesToCache === DEFAULT_MIN_MESSAGES_TO_CACHE ? middlewareOptions?.minMessagesToCache ?? runtime.context.minMessagesToCache : runtime.context.minMessagesToCache ?? middlewareOptions?.minMessagesToCache ?? DEFAULT_MIN_MESSAGES_TO_CACHE;
-			const unsupportedModelBehavior = runtime.context.unsupportedModelBehavior === DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR ? middlewareOptions?.unsupportedModelBehavior ?? runtime.context.unsupportedModelBehavior : runtime.context.unsupportedModelBehavior ?? middlewareOptions?.unsupportedModelBehavior;
-			if (!enableCaching || !options.model) return void 0;
-			const isAnthropicModel = options.model.getName() === "ChatAnthropic" || options.model.getName() === "ConfigurableModel" && options.model._defaultConfig?.modelProvider === "anthropic";
-			if (!isAnthropicModel) {
-				const modelName = options.model.getName();
-				const modelInfo = options.model.getName() === "ConfigurableModel" ? `${modelName} (${options.model._defaultConfig?.modelProvider})` : modelName;
-				const baseMessage = `Unsupported model '${modelInfo}'. Prompt caching requires an Anthropic model`;
-				if (unsupportedModelBehavior === "raise") throw new PromptCachingMiddlewareError(`${baseMessage} (e.g., 'anthropic:claude-4-0-sonnet').`);
-				else if (unsupportedModelBehavior === "warn") console.warn(`PromptCachingMiddleware: Skipping caching for ${modelName}. Consider switching to an Anthropic model for caching benefits.`);
-				return void 0;
-			}
-			const messagesCount = state.messages.length + (options.systemPrompt ? 1 : 0);
-			if (messagesCount < minMessagesToCache) return options;
-			/**
-			* Add cache_control to the last message
-			*/
-			const lastMessage = options.messages.at(-1);
-			if (!lastMessage) return options;
-			if (Array.isArray(lastMessage.content)) lastMessage.content = [...lastMessage.content.slice(0, -1), {
-				...lastMessage.content.at(-1),
-				cache_control: {
-					type: "ephemeral",
-					ttl
-				}
-			}];
-			else if (typeof lastMessage.content === "string") lastMessage.content = [{
-				type: "text",
-				text: lastMessage.content,
-				cache_control: {
-					type: "ephemeral",
-					ttl
-				}
-			}];
-			else throw new PromptCachingMiddlewareError("Last message content is not a string or array");
-			return options;
-		}
-	});
-}
-
-//#endregion
-export { anthropicPromptCachingMiddleware };
-//# sourceMappingURL=promptCaching.js.map

package/dist/agents/middlewareAgent/middleware/promptCaching.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"promptCaching.js","names":["message: string","middlewareOptions?: PromptCachingMiddlewareConfig"],"sources":["../../../../src/agents/middlewareAgent/middleware/promptCaching.ts"],"sourcesContent":["import { z } from \"zod/v3\";\nimport { ContentBlock } from \"@langchain/core/messages\";\nimport { InferInteropZodInput } from \"@langchain/core/utils/types\";\n\nimport { ConfigurableModel } from \"../../../chat_models/universal.js\";\nimport { createMiddleware } from \"../middleware.js\";\n\nconst DEFAULT_ENABLE_CACHING = true;\nconst DEFAULT_TTL = \"5m\";\nconst DEFAULT_MIN_MESSAGES_TO_CACHE = 3;\nconst DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR = \"warn\";\n\nconst contextSchema = z.object({\n  /**\n   * Whether to enable prompt caching.\n   * @default true\n   */\n  enableCaching: z.boolean().default(DEFAULT_ENABLE_CACHING),\n  /**\n   * The time-to-live for the cached prompt.\n   * @default \"5m\"\n   */\n  ttl: z.enum([\"5m\", \"1h\"]).default(DEFAULT_TTL),\n  /**\n   * The minimum number of messages required before caching is applied.\n   * @default 3\n   */\n  minMessagesToCache: z.number().default(DEFAULT_MIN_MESSAGES_TO_CACHE),\n  /**\n   * The behavior to take when an unsupported model is used.\n   * - \"ignore\" will ignore the unsupported model and continue without caching.\n   * - \"warn\" will warn the user and continue without caching.\n   * - \"raise\" will raise an error and stop the agent.\n   * @default \"warn\"\n   */\n  unsupportedModelBehavior: z\n    .enum([\"ignore\", \"warn\", \"raise\"])\n    .default(DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR),\n});\nexport type PromptCachingMiddlewareConfig = Partial<\n  InferInteropZodInput<typeof contextSchema>\n>;\n\nclass PromptCachingMiddlewareError extends Error {\n  constructor(message: string) {\n    super(message);\n    this.name = \"PromptCachingMiddlewareError\";\n  }\n}\n\n/**\n * Creates a prompt caching middleware for Anthropic models to optimize API usage.\n *\n * This middleware automatically adds cache control headers to the last messages when using Anthropic models,\n * enabling their prompt caching feature. This can significantly reduce costs for applications with repetitive\n * prompts, long system messages, or extensive conversation histories.\n *\n * ## How It Works\n *\n * The middleware intercepts model requests and adds cache control metadata that tells Anthropic's\n * API to cache processed prompt prefixes. On subsequent requests with matching prefixes, the\n * cached representations are reused, skipping redundant token processing.\n *\n * ## Benefits\n *\n * - **Cost Reduction**: Avoid reprocessing the same tokens repeatedly (up to 90% savings on cached portions)\n * - **Lower Latency**: Cached prompts are processed faster as embeddings are pre-computed\n * - **Better Scalability**: Reduced computational load enables handling more requests\n * - **Consistent Performance**: Stable response times for repetitive queries\n *\n * @param middlewareOptions - Configuration options for the caching behavior\n * @param middlewareOptions.enableCaching - Whether to enable prompt caching (default: `true`)\n * @param middlewareOptions.ttl - Cache time-to-live: `\"5m\"` for 5 minutes or `\"1h\"` for 1 hour (default: `\"5m\"`)\n * @param middlewareOptions.minMessagesToCache - Minimum number of messages required before caching is applied (default: `3`)\n * @param middlewareOptions.unsupportedModelBehavior - The behavior to take when an unsupported model is used (default: `\"warn\"`)\n *\n * @returns A middleware instance that can be passed to `createAgent`\n *\n * @throws {Error} If used with non-Anthropic models\n *\n * @example\n * Basic usage with default settings\n * ```typescript\n * import { createAgent } from \"langchain\";\n * import { anthropicPromptCachingMiddleware } from \"langchain/middleware\";\n *\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   middleware: [\n *     anthropicPromptCachingMiddleware()\n *   ]\n * });\n * ```\n *\n * @example\n * Custom configuration for longer conversations\n * ```typescript\n * const cachingMiddleware = anthropicPromptCachingMiddleware({\n *   ttl: \"1h\",  // Cache for 1 hour instead of default 5 minutes\n *   minMessagesToCache: 5  // Only cache after 5 messages\n * });\n *\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   systemPrompt: \"You are a helpful assistant with deep knowledge of...\", // Long system prompt\n *   middleware: [cachingMiddleware]\n * });\n * ```\n *\n * @example\n * Conditional caching based on runtime context\n * ```typescript\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   middleware: [\n *     anthropicPromptCachingMiddleware({\n *       enableCaching: true,\n *       ttl: \"5m\"\n *     })\n *   ]\n * });\n *\n * // Disable caching for specific requests\n * await agent.invoke(\n *   { messages: [new HumanMessage(\"Process this without caching\")] },\n *   {\n *     configurable: {\n *       middleware_context: { enableCaching: false }\n *     }\n *   }\n * );\n * ```\n *\n * @example\n * Optimal setup for customer support chatbot\n * ```typescript\n * const supportAgent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   systemPrompt: `You are a customer support agent for ACME Corp.\n *\n *     Company policies:\n *     - Always be polite and professional\n *     - Refer to knowledge base for product information\n *     - Escalate billing issues to human agents\n *     ... (extensive policies and guidelines)\n *   `,\n *   tools: [searchKnowledgeBase, createTicket, checkOrderStatus],\n *   middleware: [\n *     anthropicPromptCachingMiddleware({\n *       ttl: \"1h\",  // Long TTL for stable system prompt\n *       minMessagesToCache: 1  // Cache immediately due to large system prompt\n *     })\n *   ]\n * });\n * ```\n *\n * @remarks\n * - **Anthropic Only**: This middleware only works with Anthropic models and will throw an error if used with other providers\n * - **Automatic Application**: Caching is applied automatically when message count exceeds `minMessagesToCache`\n * - **Cache Scope**: Caches are isolated per API key and cannot be shared across different keys\n * - **TTL Options**: Only supports \"5m\" (5 minutes) and \"1h\" (1 hour) as TTL values per Anthropic's API\n * - **Best Use Cases**: Long system prompts, multi-turn conversations, repetitive queries, RAG applications\n * - **Cost Impact**: Cached tokens are billed at 10% of the base input token price, cache writes are billed at 25% of the base\n *\n * @see {@link createAgent} for agent creation\n * @see {@link https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching} Anthropic's prompt caching documentation\n * @public\n */\nexport function anthropicPromptCachingMiddleware(\n  middlewareOptions?: PromptCachingMiddlewareConfig\n) {\n  return createMiddleware({\n    name: \"PromptCachingMiddleware\",\n    contextSchema,\n    modifyModelRequest: (options, state, runtime) => {\n      /**\n       * If the runtime values match the schema default values, use the middleware option\n       * values otherwise use the runtime values. This allows to apply general configurations\n       * for all invocations, and override them for specific invocations.\n       */\n      const enableCaching =\n        runtime.context.enableCaching === DEFAULT_ENABLE_CACHING\n          ? middlewareOptions?.enableCaching ?? runtime.context.enableCaching\n          : runtime.context.enableCaching ?? middlewareOptions?.enableCaching;\n      const ttl =\n        runtime.context.ttl === DEFAULT_TTL\n          ? middlewareOptions?.ttl ?? runtime.context.ttl\n          : runtime.context.ttl ?? middlewareOptions?.ttl;\n      const minMessagesToCache =\n        runtime.context.minMessagesToCache === DEFAULT_MIN_MESSAGES_TO_CACHE\n          ? middlewareOptions?.minMessagesToCache ??\n            runtime.context.minMessagesToCache\n          : runtime.context.minMessagesToCache ??\n            middlewareOptions?.minMessagesToCache ??\n            DEFAULT_MIN_MESSAGES_TO_CACHE;\n      const unsupportedModelBehavior =\n        runtime.context.unsupportedModelBehavior ===\n        DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR\n          ? middlewareOptions?.unsupportedModelBehavior ??\n            runtime.context.unsupportedModelBehavior\n          : runtime.context.unsupportedModelBehavior ??\n            middlewareOptions?.unsupportedModelBehavior;\n\n      // Skip if caching is disabled\n      if (!enableCaching || !options.model) {\n        return undefined;\n      }\n\n      const isAnthropicModel =\n        options.model.getName() === \"ChatAnthropic\" ||\n        (options.model.getName() === \"ConfigurableModel\" &&\n          (options.model as ConfigurableModel)._defaultConfig?.modelProvider ===\n            \"anthropic\");\n      if (!isAnthropicModel) {\n        // Get model name for better error context\n        const modelName = options.model.getName();\n        const modelInfo =\n          options.model.getName() === \"ConfigurableModel\"\n            ? `${modelName} (${\n                (options.model as ConfigurableModel)._defaultConfig\n                  ?.modelProvider\n              })`\n            : modelName;\n\n        const baseMessage = `Unsupported model '${modelInfo}'. Prompt caching requires an Anthropic model`;\n\n        if (unsupportedModelBehavior === \"raise\") {\n          throw new PromptCachingMiddlewareError(\n            `${baseMessage} (e.g., 'anthropic:claude-4-0-sonnet').`\n          );\n        } else if (unsupportedModelBehavior === \"warn\") {\n          console.warn(\n            `PromptCachingMiddleware: Skipping caching for ${modelName}. Consider switching to an Anthropic model for caching benefits.`\n          );\n        }\n        return undefined;\n      }\n\n      const messagesCount =\n        state.messages.length + (options.systemPrompt ? 1 : 0);\n\n      if (messagesCount < minMessagesToCache) {\n        return options;\n      }\n\n      /**\n       * Add cache_control to the last message\n       */\n      const lastMessage = options.messages.at(-1);\n      if (!lastMessage) {\n        return options;\n      }\n\n      if (Array.isArray(lastMessage.content)) {\n        lastMessage.content = [\n          ...lastMessage.content.slice(0, -1),\n          {\n            ...lastMessage.content.at(-1),\n            cache_control: {\n              type: \"ephemeral\",\n              ttl,\n            },\n          } as ContentBlock,\n        ];\n      } else if (typeof lastMessage.content === \"string\") {\n        lastMessage.content = [\n          {\n            type: \"text\",\n            text: lastMessage.content,\n            cache_control: {\n              type: \"ephemeral\",\n              ttl,\n            },\n          },\n        ];\n      } else {\n        throw new PromptCachingMiddlewareError(\n          \"Last message content is not a string or array\"\n        );\n      }\n\n      return options;\n    },\n  });\n}\n"],"mappings":";;;;AAOA,MAAM,yBAAyB;AAC/B,MAAM,cAAc;AACpB,MAAM,gCAAgC;AACtC,MAAM,qCAAqC;AAE3C,MAAM,gBAAgB,EAAE,OAAO;CAK7B,eAAe,EAAE,SAAS,CAAC,QAAQ,uBAAuB;CAK1D,KAAK,EAAE,KAAK,CAAC,MAAM,IAAK,EAAC,CAAC,QAAQ,YAAY;CAK9C,oBAAoB,EAAE,QAAQ,CAAC,QAAQ,8BAA8B;CAQrE,0BAA0B,EACvB,KAAK;EAAC;EAAU;EAAQ;CAAQ,EAAC,CACjC,QAAQ,mCAAmC;AAC/C,EAAC;AAKF,IAAM,+BAAN,cAA2C,MAAM;CAC/C,YAAYA,SAAiB;EAC3B,MAAM,QAAQ;EACd,KAAK,OAAO;CACb;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwHD,SAAgB,iCACdC,mBACA;AACA,QAAO,iBAAiB;EACtB,MAAM;EACN;EACA,oBAAoB,CAAC,SAAS,OAAO,YAAY;;;;;;GAM/C,MAAM,gBACJ,QAAQ,QAAQ,kBAAkB,yBAC9B,mBAAmB,iBAAiB,QAAQ,QAAQ,gBACpD,QAAQ,QAAQ,iBAAiB,mBAAmB;GAC1D,MAAM,MACJ,QAAQ,QAAQ,QAAQ,cACpB,mBAAmB,OAAO,QAAQ,QAAQ,MAC1C,QAAQ,QAAQ,OAAO,mBAAmB;GAChD,MAAM,qBACJ,QAAQ,QAAQ,uBAAuB,gCACnC,mBAAmB,sBACnB,QAAQ,QAAQ,qBAChB,QAAQ,QAAQ,sBAChB,mBAAmB,sBACnB;GACN,MAAM,2BACJ,QAAQ,QAAQ,6BAChB,qCACI,mBAAmB,4BACnB,QAAQ,QAAQ,2BAChB,QAAQ,QAAQ,4BAChB,mBAAmB;AAGzB,OAAI,CAAC,iBAAiB,CAAC,QAAQ,MAC7B,QAAO;GAGT,MAAM,mBACJ,QAAQ,MAAM,SAAS,KAAK,mBAC3B,QAAQ,MAAM,SAAS,KAAK,uBAC1B,QAAQ,MAA4B,gBAAgB,kBACnD;AACN,OAAI,CAAC,kBAAkB;IAErB,MAAM,YAAY,QAAQ,MAAM,SAAS;IACzC,MAAM,YACJ,QAAQ,MAAM,SAAS,KAAK,sBACxB,GAAG,UAAU,EAAE,EACZ,QAAQ,MAA4B,gBACjC,cACL,CAAC,CAAC,GACH;IAEN,MAAM,cAAc,CAAC,mBAAmB,EAAE,UAAU,6CAA6C,CAAC;AAElG,QAAI,6BAA6B,QAC/B,OAAM,IAAI,6BACR,GAAG,YAAY,uCAAuC,CAAC;aAEhD,6BAA6B,QACtC,QAAQ,KACN,CAAC,8CAA8C,EAAE,UAAU,gEAAgE,CAAC,CAC7H;AAEH,WAAO;GACR;GAED,MAAM,gBACJ,MAAM,SAAS,UAAU,QAAQ,eAAe,IAAI;AAEtD,OAAI,gBAAgB,mBAClB,QAAO;;;;GAMT,MAAM,cAAc,QAAQ,SAAS,GAAG,GAAG;AAC3C,OAAI,CAAC,YACH,QAAO;AAGT,OAAI,MAAM,QAAQ,YAAY,QAAQ,EACpC,YAAY,UAAU,CACpB,GAAG,YAAY,QAAQ,MAAM,GAAG,GAAG,EACnC;IACE,GAAG,YAAY,QAAQ,GAAG,GAAG;IAC7B,eAAe;KACb,MAAM;KACN;IACD;GACF,CACF;YACQ,OAAO,YAAY,YAAY,UACxC,YAAY,UAAU,CACpB;IACE,MAAM;IACN,MAAM,YAAY;IAClB,eAAe;KACb,MAAM;KACN;IACD;GACF,CACF;OAED,OAAM,IAAI,6BACR;AAIJ,UAAO;EACR;CACF,EAAC;AACH"}

package/dist/agents/middlewareAgent/middleware/summarization.cjs

@@ -1,274 +0,0 @@
-const require_rolldown_runtime = require('../../../_virtual/rolldown_runtime.cjs');
-const require_middleware = require('../middleware.cjs');
-const __langchain_core_messages = require_rolldown_runtime.__toESM(require("@langchain/core/messages"));
-const __langchain_langgraph = require_rolldown_runtime.__toESM(require("@langchain/langgraph"));
-const __langchain_core_utils_types = require_rolldown_runtime.__toESM(require("@langchain/core/utils/types"));
-const zod_v3 = require_rolldown_runtime.__toESM(require("zod/v3"));
-const uuid = require_rolldown_runtime.__toESM(require("uuid"));
-
-//#region src/agents/middlewareAgent/middleware/summarization.ts
-const DEFAULT_SUMMARY_PROMPT = `<role>
-Context Extraction Assistant
-</role>
-
-<primary_objective>
-Your sole objective in this task is to extract the highest quality/most relevant context from the conversation history below.
-</primary_objective>
-
-<objective_information>
-You're nearing the total number of input tokens you can accept, so you must extract the highest quality/most relevant pieces of information from your conversation history.
-This context will then overwrite the conversation history presented below. Because of this, ensure the context you extract is only the most important information to your overall goal.
-</objective_information>
-
-<instructions>
-The conversation history below will be replaced with the context you extract in this step. Because of this, you must do your very best to extract and record all of the most important context from the conversation history.
-You want to ensure that you don't repeat any actions you've already completed, so the context you extract from the conversation history should be focused on the most important information to your overall goal.
-</instructions>
-
-The user will message you with the full message history you'll be extracting context from, to then replace. Carefully read over it all, and think deeply about what information is most important to your overall goal that should be saved:
-
-With all of this in mind, please carefully read over the entire conversation history, and extract the most important and relevant context to replace it so that you can free up space in the conversation history.
-Respond ONLY with the extracted context. Do not include any additional information, or text before or after the extracted context.
-
-<messages>
-Messages to summarize:
-{messages}
-</messages>`;
-const SUMMARY_PREFIX = "## Previous conversation summary:";
-const DEFAULT_MESSAGES_TO_KEEP = 20;
-const DEFAULT_TRIM_TOKEN_LIMIT = 4e3;
-const DEFAULT_FALLBACK_MESSAGE_COUNT = 15;
-const SEARCH_RANGE_FOR_TOOL_PAIRS = 5;
-const contextSchema = zod_v3.z.object({
-	model: zod_v3.z.custom(),
-	maxTokensBeforeSummary: zod_v3.z.number().optional(),
-	messagesToKeep: zod_v3.z.number().default(DEFAULT_MESSAGES_TO_KEEP),
-	tokenCounter: zod_v3.z.function().args(zod_v3.z.array(zod_v3.z.any())).returns(zod_v3.z.union([zod_v3.z.number(), zod_v3.z.promise(zod_v3.z.number())])).optional(),
-	summaryPrompt: zod_v3.z.string().default(DEFAULT_SUMMARY_PROMPT),
-	summaryPrefix: zod_v3.z.string().default(SUMMARY_PREFIX)
-});
-/**
-* Default token counter that approximates based on character count
-* @param messages Messages to count tokens for
-* @returns Approximate token count
-*/
-function countTokensApproximately(messages) {
-	let totalChars = 0;
-	for (const msg of messages) {
-		let textContent;
-		if (typeof msg.content === "string") textContent = msg.content;
-		else if (Array.isArray(msg.content)) textContent = msg.content.map((item) => {
-			if (typeof item === "string") return item;
-			if (item.type === "text" && "text" in item) return item.text;
-			return "";
-		}).join("");
-		else textContent = "";
-		totalChars += textContent.length;
-	}
-	return Math.ceil(totalChars / 4);
-}
-/**
-* Summarization middleware that automatically summarizes conversation history when token limits are approached.
-*
-* This middleware monitors message token counts and automatically summarizes older
-* messages when a threshold is reached, preserving recent messages and maintaining
-* context continuity by ensuring AI/Tool message pairs remain together.
-*
-* @param options Configuration options for the summarization middleware
-* @returns A middleware instance
-*
-* @example
-* ```ts
-* import { summarizationMiddleware } from "langchain/middleware";
-* import { createAgent } from "langchain";
-*
-* const agent = createAgent({
-*   llm: model,
-*   tools: [getWeather],
-*   middleware: [
-*     summarizationMiddleware({
-*       model: new ChatOpenAI({ model: "gpt-4o" }),
-*       maxTokensBeforeSummary: 4000,
-*       messagesToKeep: 20,
-*     })
-*   ],
-* });
-*
-* ```
-*/
-function summarizationMiddleware(options) {
-	return require_middleware.createMiddleware({
-		name: "SummarizationMiddleware",
-		contextSchema,
-		beforeModel: async (state, runtime) => {
-			/**
-			* Parse user options to get their explicit values
-			*/
-			const userOptions = (0, __langchain_core_utils_types.interopParse)(contextSchema, options);
-			/**
-			* Merge context with user options, preferring user options when context has default values
-			*/
-			const config = {
-				model: userOptions.model,
-				maxTokensBeforeSummary: runtime.context.maxTokensBeforeSummary !== void 0 ? runtime.context.maxTokensBeforeSummary : userOptions.maxTokensBeforeSummary,
-				messagesToKeep: runtime.context.messagesToKeep === DEFAULT_MESSAGES_TO_KEEP ? userOptions.messagesToKeep : runtime.context.messagesToKeep ?? userOptions.messagesToKeep,
-				tokenCounter: runtime.context.tokenCounter !== void 0 ? runtime.context.tokenCounter : userOptions.tokenCounter,
-				summaryPrompt: runtime.context.summaryPrompt === DEFAULT_SUMMARY_PROMPT ? userOptions.summaryPrompt : runtime.context.summaryPrompt ?? userOptions.summaryPrompt,
-				summaryPrefix: runtime.context.summaryPrefix === SUMMARY_PREFIX ? userOptions.summaryPrefix : runtime.context.summaryPrefix ?? userOptions.summaryPrefix
-			};
-			const { messages } = state;
-			ensureMessageIds(messages);
-			const tokenCounter = config.tokenCounter || countTokensApproximately;
-			const totalTokens = await tokenCounter(messages);
-			if (config.maxTokensBeforeSummary == null || totalTokens < config.maxTokensBeforeSummary) return;
-			const { systemPrompt, conversationMessages } = splitSystemMessage(messages);
-			const cutoffIndex = findSafeCutoff(conversationMessages, config.messagesToKeep);
-			if (cutoffIndex <= 0) return;
-			const { messagesToSummarize, preservedMessages } = partitionMessages(systemPrompt, conversationMessages, cutoffIndex);
-			const summary = await createSummary(messagesToSummarize, config.model, config.summaryPrompt, tokenCounter);
-			const updatedSystemMessage = buildUpdatedSystemMessage(systemPrompt, summary, config.summaryPrefix);
-			return { messages: [
-				new __langchain_core_messages.RemoveMessage({ id: __langchain_langgraph.REMOVE_ALL_MESSAGES }),
-				updatedSystemMessage,
-				...preservedMessages
-			] };
-		}
-	});
-}
-/**
-* Ensure all messages have unique IDs
-*/
-function ensureMessageIds(messages) {
-	for (const msg of messages) if (!msg.id) msg.id = (0, uuid.v4)();
-}
-/**
-* Separate system message from conversation messages
-*/
-function splitSystemMessage(messages) {
-	if (messages.length > 0 && __langchain_core_messages.SystemMessage.isInstance(messages[0])) return {
-		systemPrompt: messages[0],
-		conversationMessages: messages.slice(1)
-	};
-	return {
-		systemPrompt: null,
-		conversationMessages: messages
-	};
-}
-/**
-* Partition messages into those to summarize and those to preserve
-*/
-function partitionMessages(systemPrompt, conversationMessages, cutoffIndex) {
-	const messagesToSummarize = conversationMessages.slice(0, cutoffIndex);
-	const preservedMessages = conversationMessages.slice(cutoffIndex);
-	if (systemPrompt) messagesToSummarize.unshift(systemPrompt);
-	return {
-		messagesToSummarize,
-		preservedMessages
-	};
-}
-/**
-* Build updated system message incorporating the summary
-*/
-function buildUpdatedSystemMessage(originalSystemMessage, summary, summaryPrefix) {
-	let originalContent = "";
-	if (originalSystemMessage) {
-		const { content: content$1 } = originalSystemMessage;
-		if (typeof content$1 === "string") originalContent = content$1.split(summaryPrefix)[0].trim();
-	}
-	const content = originalContent ? `${originalContent}\n${summaryPrefix}\n${summary}` : `${summaryPrefix}\n${summary}`;
-	return new __langchain_core_messages.SystemMessage({
-		content,
-		id: originalSystemMessage?.id || (0, uuid.v4)()
-	});
-}
-/**
-* Find safe cutoff point that preserves AI/Tool message pairs
-*/
-function findSafeCutoff(messages, messagesToKeep) {
-	if (messages.length <= messagesToKeep) return 0;
-	const targetCutoff = messages.length - messagesToKeep;
-	for (let i = targetCutoff; i >= 0; i--) if (isSafeCutoffPoint(messages, i)) return i;
-	return 0;
-}
-/**
-* Check if cutting at index would separate AI/Tool message pairs
-*/
-function isSafeCutoffPoint(messages, cutoffIndex) {
-	if (cutoffIndex >= messages.length) return true;
-	const searchStart = Math.max(0, cutoffIndex - SEARCH_RANGE_FOR_TOOL_PAIRS);
-	const searchEnd = Math.min(messages.length, cutoffIndex + SEARCH_RANGE_FOR_TOOL_PAIRS);
-	for (let i = searchStart; i < searchEnd; i++) {
-		if (!hasToolCalls(messages[i])) continue;
-		const toolCallIds = extractToolCallIds(messages[i]);
-		if (cutoffSeparatesToolPair(messages, i, cutoffIndex, toolCallIds)) return false;
-	}
-	return true;
-}
-/**
-* Check if message is an AI message with tool calls
-*/
-function hasToolCalls(message) {
-	return __langchain_core_messages.AIMessage.isInstance(message) && "tool_calls" in message && Array.isArray(message.tool_calls) && message.tool_calls.length > 0;
-}
-/**
-* Extract tool call IDs from an AI message
-*/
-function extractToolCallIds(aiMessage) {
-	const toolCallIds = /* @__PURE__ */ new Set();
-	if (aiMessage.tool_calls) for (const toolCall of aiMessage.tool_calls) {
-		const id = typeof toolCall === "object" && "id" in toolCall ? toolCall.id : null;
-		if (id) toolCallIds.add(id);
-	}
-	return toolCallIds;
-}
-/**
-* Check if cutoff separates an AI message from its corresponding tool messages
-*/
-function cutoffSeparatesToolPair(messages, aiMessageIndex, cutoffIndex, toolCallIds) {
-	for (let j = aiMessageIndex + 1; j < messages.length; j++) {
-		const message = messages[j];
-		if (__langchain_core_messages.ToolMessage.isInstance(message) && toolCallIds.has(message.tool_call_id)) {
-			const aiBeforeCutoff = aiMessageIndex < cutoffIndex;
-			const toolBeforeCutoff = j < cutoffIndex;
-			if (aiBeforeCutoff !== toolBeforeCutoff) return true;
-		}
-	}
-	return false;
-}
-/**
-* Generate summary for the given messages
-*/
-async function createSummary(messagesToSummarize, model, summaryPrompt, tokenCounter) {
-	if (!messagesToSummarize.length) return "No previous conversation history.";
-	const trimmedMessages = await trimMessagesForSummary(messagesToSummarize, tokenCounter);
-	if (!trimmedMessages.length) return "Previous conversation was too long to summarize.";
-	try {
-		const formattedPrompt = summaryPrompt.replace("{messages}", JSON.stringify(trimmedMessages, null, 2));
-		const response = await model.invoke(formattedPrompt);
-		const { content } = response;
-		return typeof content === "string" ? content.trim() : "Error generating summary: Invalid response format";
-	} catch (e) {
-		return `Error generating summary: ${e}`;
-	}
-}
-/**
-* Trim messages to fit within summary generation limits
-*/
-async function trimMessagesForSummary(messages, tokenCounter) {
-	try {
-		return await (0, __langchain_core_messages.trimMessages)(messages, {
-			maxTokens: DEFAULT_TRIM_TOKEN_LIMIT,
-			tokenCounter: async (msgs) => Promise.resolve(tokenCounter(msgs)),
-			strategy: "last",
-			allowPartial: true,
-			includeSystem: true
-		});
-	} catch {
-		return messages.slice(-DEFAULT_FALLBACK_MESSAGE_COUNT);
-	}
-}
-
-//#endregion
-exports.countTokensApproximately = countTokensApproximately;
-exports.summarizationMiddleware = summarizationMiddleware;
-//# sourceMappingURL=summarization.cjs.map

package/dist/agents/middlewareAgent/middleware/summarization.cjs.map

@@ -1 +0,0 @@
-{"version":3,"file":"summarization.cjs","names":["z","messages: BaseMessage[]","textContent: string","options: SummarizationMiddlewareConfig","createMiddleware","RemoveMessage","REMOVE_ALL_MESSAGES","SystemMessage","systemPrompt: SystemMessage | null","conversationMessages: BaseMessage[]","cutoffIndex: number","originalSystemMessage: SystemMessage | null","summary: string","summaryPrefix: string","content","messagesToKeep: number","message: BaseMessage","AIMessage","aiMessage: AIMessage","aiMessageIndex: number","toolCallIds: Set<string>","ToolMessage","messagesToSummarize: BaseMessage[]","model: BaseLanguageModel","summaryPrompt: string","tokenCounter: TokenCounter"],"sources":["../../../../src/agents/middlewareAgent/middleware/summarization.ts"],"sourcesContent":["import { z } from \"zod/v3\";\nimport { v4 as uuid } from \"uuid\";\nimport {\n  BaseMessage,\n  AIMessage,\n  SystemMessage,\n  ToolMessage,\n  RemoveMessage,\n  trimMessages,\n} from \"@langchain/core/messages\";\nimport { BaseLanguageModel } from \"@langchain/core/language_models/base\";\nimport {\n  interopParse,\n  InferInteropZodOutput,\n  InferInteropZodInput,\n} from \"@langchain/core/utils/types\";\nimport { REMOVE_ALL_MESSAGES } from \"@langchain/langgraph\";\nimport { createMiddleware } from \"../middleware.js\";\n\nconst DEFAULT_SUMMARY_PROMPT = `<role>\nContext Extraction Assistant\n</role>\n\n<primary_objective>\nYour sole objective in this task is to extract the highest quality/most relevant context from the conversation history below.\n</primary_objective>\n\n<objective_information>\nYou're nearing the total number of input tokens you can accept, so you must extract the highest quality/most relevant pieces of information from your conversation history.\nThis context will then overwrite the conversation history presented below. Because of this, ensure the context you extract is only the most important information to your overall goal.\n</objective_information>\n\n<instructions>\nThe conversation history below will be replaced with the context you extract in this step. Because of this, you must do your very best to extract and record all of the most important context from the conversation history.\nYou want to ensure that you don't repeat any actions you've already completed, so the context you extract from the conversation history should be focused on the most important information to your overall goal.\n</instructions>\n\nThe user will message you with the full message history you'll be extracting context from, to then replace. Carefully read over it all, and think deeply about what information is most important to your overall goal that should be saved:\n\nWith all of this in mind, please carefully read over the entire conversation history, and extract the most important and relevant context to replace it so that you can free up space in the conversation history.\nRespond ONLY with the extracted context. Do not include any additional information, or text before or after the extracted context.\n\n<messages>\nMessages to summarize:\n{messages}\n</messages>`;\n\nconst SUMMARY_PREFIX = \"## Previous conversation summary:\";\n\nconst DEFAULT_MESSAGES_TO_KEEP = 20;\nconst DEFAULT_TRIM_TOKEN_LIMIT = 4000;\nconst DEFAULT_FALLBACK_MESSAGE_COUNT = 15;\nconst SEARCH_RANGE_FOR_TOOL_PAIRS = 5;\n\ntype TokenCounter = (messages: BaseMessage[]) => number | Promise<number>;\n\nconst contextSchema = z.object({\n  model: z.custom<BaseLanguageModel>(),\n  maxTokensBeforeSummary: z.number().optional(),\n  messagesToKeep: z.number().default(DEFAULT_MESSAGES_TO_KEEP),\n  tokenCounter: z\n    .function()\n    .args(z.array(z.any()))\n    .returns(z.union([z.number(), z.promise(z.number())]))\n    .optional(),\n  summaryPrompt: z.string().default(DEFAULT_SUMMARY_PROMPT),\n  summaryPrefix: z.string().default(SUMMARY_PREFIX),\n});\n\nexport type SummarizationMiddlewareConfig = InferInteropZodInput<\n  typeof contextSchema\n>;\n\n/**\n * Default token counter that approximates based on character count\n * @param messages Messages to count tokens for\n * @returns Approximate token count\n */\nexport function countTokensApproximately(messages: BaseMessage[]): number {\n  let totalChars = 0;\n  for (const msg of messages) {\n    let textContent: string;\n    if (typeof msg.content === \"string\") {\n      textContent = msg.content;\n    } else if (Array.isArray(msg.content)) {\n      textContent = msg.content\n        .map((item) => {\n          if (typeof item === \"string\") return item;\n          if (item.type === \"text\" && \"text\" in item) return item.text;\n          return \"\";\n        })\n        .join(\"\");\n    } else {\n      textContent = \"\";\n    }\n    totalChars += textContent.length;\n  }\n  // Approximate 1 token = 4 characters\n  return Math.ceil(totalChars / 4);\n}\n\n/**\n * Summarization middleware that automatically summarizes conversation history when token limits are approached.\n *\n * This middleware monitors message token counts and automatically summarizes older\n * messages when a threshold is reached, preserving recent messages and maintaining\n * context continuity by ensuring AI/Tool message pairs remain together.\n *\n * @param options Configuration options for the summarization middleware\n * @returns A middleware instance\n *\n * @example\n * ```ts\n * import { summarizationMiddleware } from \"langchain/middleware\";\n * import { createAgent } from \"langchain\";\n *\n * const agent = createAgent({\n *   llm: model,\n *   tools: [getWeather],\n *   middleware: [\n *     summarizationMiddleware({\n *       model: new ChatOpenAI({ model: \"gpt-4o\" }),\n *       maxTokensBeforeSummary: 4000,\n *       messagesToKeep: 20,\n *     })\n *   ],\n * });\n *\n * ```\n */\nexport function summarizationMiddleware(\n  options: SummarizationMiddlewareConfig\n) {\n  return createMiddleware({\n    name: \"SummarizationMiddleware\",\n    contextSchema,\n    beforeModel: async (state, runtime) => {\n      /**\n       * Parse user options to get their explicit values\n       */\n      const userOptions = interopParse(contextSchema, options);\n\n      /**\n       * Merge context with user options, preferring user options when context has default values\n       */\n      const config = {\n        model: userOptions.model,\n        maxTokensBeforeSummary:\n          runtime.context.maxTokensBeforeSummary !== undefined\n            ? runtime.context.maxTokensBeforeSummary\n            : userOptions.maxTokensBeforeSummary,\n        messagesToKeep:\n          runtime.context.messagesToKeep === DEFAULT_MESSAGES_TO_KEEP\n            ? userOptions.messagesToKeep\n            : runtime.context.messagesToKeep ?? userOptions.messagesToKeep,\n        tokenCounter:\n          runtime.context.tokenCounter !== undefined\n            ? runtime.context.tokenCounter\n            : userOptions.tokenCounter,\n        summaryPrompt:\n          runtime.context.summaryPrompt === DEFAULT_SUMMARY_PROMPT\n            ? userOptions.summaryPrompt\n            : runtime.context.summaryPrompt ?? userOptions.summaryPrompt,\n        summaryPrefix:\n          runtime.context.summaryPrefix === SUMMARY_PREFIX\n            ? userOptions.summaryPrefix\n            : runtime.context.summaryPrefix ?? userOptions.summaryPrefix,\n      } as InferInteropZodOutput<typeof contextSchema>;\n      const { messages } = state;\n\n      // Ensure all messages have IDs\n      ensureMessageIds(messages);\n\n      const tokenCounter = config.tokenCounter || countTokensApproximately;\n      const totalTokens = await tokenCounter(messages);\n\n      if (\n        config.maxTokensBeforeSummary == null ||\n        totalTokens < config.maxTokensBeforeSummary\n      ) {\n        return;\n      }\n\n      const { systemPrompt, conversationMessages } =\n        splitSystemMessage(messages);\n      const cutoffIndex = findSafeCutoff(\n        conversationMessages,\n        config.messagesToKeep\n      );\n\n      if (cutoffIndex <= 0) {\n        return;\n      }\n\n      const { messagesToSummarize, preservedMessages } = partitionMessages(\n        systemPrompt,\n        conversationMessages,\n        cutoffIndex\n      );\n\n      const summary = await createSummary(\n        messagesToSummarize,\n        config.model,\n        config.summaryPrompt,\n        tokenCounter\n      );\n\n      const updatedSystemMessage = buildUpdatedSystemMessage(\n        systemPrompt,\n        summary,\n        config.summaryPrefix\n      );\n\n      return {\n        messages: [\n          new RemoveMessage({ id: REMOVE_ALL_MESSAGES }),\n          updatedSystemMessage,\n          ...preservedMessages,\n        ],\n      };\n    },\n  });\n}\n\n/**\n * Ensure all messages have unique IDs\n */\nfunction ensureMessageIds(messages: BaseMessage[]): void {\n  for (const msg of messages) {\n    if (!msg.id) {\n      msg.id = uuid();\n    }\n  }\n}\n\n/**\n * Separate system message from conversation messages\n */\nfunction splitSystemMessage(messages: BaseMessage[]): {\n  systemPrompt: SystemMessage | null;\n  conversationMessages: BaseMessage[];\n} {\n  if (messages.length > 0 && SystemMessage.isInstance(messages[0])) {\n    return {\n      systemPrompt: messages[0] as SystemMessage,\n      conversationMessages: messages.slice(1),\n    };\n  }\n  return {\n    systemPrompt: null,\n    conversationMessages: messages,\n  };\n}\n\n/**\n * Partition messages into those to summarize and those to preserve\n */\nfunction partitionMessages(\n  systemPrompt: SystemMessage | null,\n  conversationMessages: BaseMessage[],\n  cutoffIndex: number\n): { messagesToSummarize: BaseMessage[]; preservedMessages: BaseMessage[] } {\n  const messagesToSummarize = conversationMessages.slice(0, cutoffIndex);\n  const preservedMessages = conversationMessages.slice(cutoffIndex);\n\n  // Include system message in messages to summarize to capture previous summaries\n  if (systemPrompt) {\n    messagesToSummarize.unshift(systemPrompt);\n  }\n\n  return { messagesToSummarize, preservedMessages };\n}\n\n/**\n * Build updated system message incorporating the summary\n */\nfunction buildUpdatedSystemMessage(\n  originalSystemMessage: SystemMessage | null,\n  summary: string,\n  summaryPrefix: string\n): SystemMessage {\n  let originalContent = \"\";\n  if (originalSystemMessage) {\n    const { content } = originalSystemMessage;\n    if (typeof content === \"string\") {\n      originalContent = content.split(summaryPrefix)[0].trim();\n    }\n  }\n\n  const content = originalContent\n    ? `${originalContent}\\n${summaryPrefix}\\n${summary}`\n    : `${summaryPrefix}\\n${summary}`;\n\n  return new SystemMessage({\n    content,\n    id: originalSystemMessage?.id || uuid(),\n  });\n}\n\n/**\n * Find safe cutoff point that preserves AI/Tool message pairs\n */\nfunction findSafeCutoff(\n  messages: BaseMessage[],\n  messagesToKeep: number\n): number {\n  if (messages.length <= messagesToKeep) {\n    return 0;\n  }\n\n  const targetCutoff = messages.length - messagesToKeep;\n\n  for (let i = targetCutoff; i >= 0; i--) {\n    if (isSafeCutoffPoint(messages, i)) {\n      return i;\n    }\n  }\n\n  return 0;\n}\n\n/**\n * Check if cutting at index would separate AI/Tool message pairs\n */\nfunction isSafeCutoffPoint(\n  messages: BaseMessage[],\n  cutoffIndex: number\n): boolean {\n  if (cutoffIndex >= messages.length) {\n    return true;\n  }\n\n  const searchStart = Math.max(0, cutoffIndex - SEARCH_RANGE_FOR_TOOL_PAIRS);\n  const searchEnd = Math.min(\n    messages.length,\n    cutoffIndex + SEARCH_RANGE_FOR_TOOL_PAIRS\n  );\n\n  for (let i = searchStart; i < searchEnd; i++) {\n    if (!hasToolCalls(messages[i])) {\n      continue;\n    }\n\n    const toolCallIds = extractToolCallIds(messages[i] as AIMessage);\n    if (cutoffSeparatesToolPair(messages, i, cutoffIndex, toolCallIds)) {\n      return false;\n    }\n  }\n\n  return true;\n}\n\n/**\n * Check if message is an AI message with tool calls\n */\nfunction hasToolCalls(message: BaseMessage): boolean {\n  return (\n    AIMessage.isInstance(message) &&\n    \"tool_calls\" in message &&\n    Array.isArray(message.tool_calls) &&\n    message.tool_calls.length > 0\n  );\n}\n\n/**\n * Extract tool call IDs from an AI message\n */\nfunction extractToolCallIds(aiMessage: AIMessage): Set<string> {\n  const toolCallIds = new Set<string>();\n  if (aiMessage.tool_calls) {\n    for (const toolCall of aiMessage.tool_calls) {\n      const id =\n        typeof toolCall === \"object\" && \"id\" in toolCall ? toolCall.id : null;\n      if (id) {\n        toolCallIds.add(id);\n      }\n    }\n  }\n  return toolCallIds;\n}\n\n/**\n * Check if cutoff separates an AI message from its corresponding tool messages\n */\nfunction cutoffSeparatesToolPair(\n  messages: BaseMessage[],\n  aiMessageIndex: number,\n  cutoffIndex: number,\n  toolCallIds: Set<string>\n): boolean {\n  for (let j = aiMessageIndex + 1; j < messages.length; j++) {\n    const message = messages[j];\n    if (\n      ToolMessage.isInstance(message) &&\n      toolCallIds.has(message.tool_call_id)\n    ) {\n      const aiBeforeCutoff = aiMessageIndex < cutoffIndex;\n      const toolBeforeCutoff = j < cutoffIndex;\n      if (aiBeforeCutoff !== toolBeforeCutoff) {\n        return true;\n      }\n    }\n  }\n  return false;\n}\n\n/**\n * Generate summary for the given messages\n */\nasync function createSummary(\n  messagesToSummarize: BaseMessage[],\n  model: BaseLanguageModel,\n  summaryPrompt: string,\n  tokenCounter: TokenCounter\n): Promise<string> {\n  if (!messagesToSummarize.length) {\n    return \"No previous conversation history.\";\n  }\n\n  const trimmedMessages = await trimMessagesForSummary(\n    messagesToSummarize,\n    tokenCounter\n  );\n\n  if (!trimmedMessages.length) {\n    return \"Previous conversation was too long to summarize.\";\n  }\n\n  try {\n    const formattedPrompt = summaryPrompt.replace(\n      \"{messages}\",\n      JSON.stringify(trimmedMessages, null, 2)\n    );\n    const response = await model.invoke(formattedPrompt);\n    const { content } = response;\n    return typeof content === \"string\"\n      ? content.trim()\n      : \"Error generating summary: Invalid response format\";\n  } catch (e) {\n    return `Error generating summary: ${e}`;\n  }\n}\n\n/**\n * Trim messages to fit within summary generation limits\n */\nasync function trimMessagesForSummary(\n  messages: BaseMessage[],\n  tokenCounter: TokenCounter\n): Promise<BaseMessage[]> {\n  try {\n    return await trimMessages(messages, {\n      maxTokens: DEFAULT_TRIM_TOKEN_LIMIT,\n      tokenCounter: async (msgs) => Promise.resolve(tokenCounter(msgs)),\n      strategy: \"last\",\n      allowPartial: true,\n      includeSystem: true,\n    });\n  } catch {\n    // Fallback to last N messages if trimming fails\n    return messages.slice(-DEFAULT_FALLBACK_MESSAGE_COUNT);\n  }\n}\n"],"mappings":";;;;;;;;;AAmBA,MAAM,yBAAyB,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;WA0BrB,CAAC;AAEZ,MAAM,iBAAiB;AAEvB,MAAM,2BAA2B;AACjC,MAAM,2BAA2B;AACjC,MAAM,iCAAiC;AACvC,MAAM,8BAA8B;AAIpC,MAAM,gBAAgBA,SAAE,OAAO;CAC7B,OAAOA,SAAE,QAA2B;CACpC,wBAAwBA,SAAE,QAAQ,CAAC,UAAU;CAC7C,gBAAgBA,SAAE,QAAQ,CAAC,QAAQ,yBAAyB;CAC5D,cAAcA,SACX,UAAU,CACV,KAAKA,SAAE,MAAMA,SAAE,KAAK,CAAC,CAAC,CACtB,QAAQA,SAAE,MAAM,CAACA,SAAE,QAAQ,EAAEA,SAAE,QAAQA,SAAE,QAAQ,CAAC,AAAC,EAAC,CAAC,CACrD,UAAU;CACb,eAAeA,SAAE,QAAQ,CAAC,QAAQ,uBAAuB;CACzD,eAAeA,SAAE,QAAQ,CAAC,QAAQ,eAAe;AAClD,EAAC;;;;;;AAWF,SAAgB,yBAAyBC,UAAiC;CACxE,IAAI,aAAa;AACjB,MAAK,MAAM,OAAO,UAAU;EAC1B,IAAIC;AACJ,MAAI,OAAO,IAAI,YAAY,UACzB,cAAc,IAAI;WACT,MAAM,QAAQ,IAAI,QAAQ,EACnC,cAAc,IAAI,QACf,IAAI,CAAC,SAAS;AACb,OAAI,OAAO,SAAS,SAAU,QAAO;AACrC,OAAI,KAAK,SAAS,UAAU,UAAU,KAAM,QAAO,KAAK;AACxD,UAAO;EACR,EAAC,CACD,KAAK,GAAG;OAEX,cAAc;EAEhB,cAAc,YAAY;CAC3B;AAED,QAAO,KAAK,KAAK,aAAa,EAAE;AACjC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+BD,SAAgB,wBACdC,SACA;AACA,QAAOC,oCAAiB;EACtB,MAAM;EACN;EACA,aAAa,OAAO,OAAO,YAAY;;;;GAIrC,MAAM,6DAA2B,eAAe,QAAQ;;;;GAKxD,MAAM,SAAS;IACb,OAAO,YAAY;IACnB,wBACE,QAAQ,QAAQ,2BAA2B,SACvC,QAAQ,QAAQ,yBAChB,YAAY;IAClB,gBACE,QAAQ,QAAQ,mBAAmB,2BAC/B,YAAY,iBACZ,QAAQ,QAAQ,kBAAkB,YAAY;IACpD,cACE,QAAQ,QAAQ,iBAAiB,SAC7B,QAAQ,QAAQ,eAChB,YAAY;IAClB,eACE,QAAQ,QAAQ,kBAAkB,yBAC9B,YAAY,gBACZ,QAAQ,QAAQ,iBAAiB,YAAY;IACnD,eACE,QAAQ,QAAQ,kBAAkB,iBAC9B,YAAY,gBACZ,QAAQ,QAAQ,iBAAiB,YAAY;GACpD;GACD,MAAM,EAAE,UAAU,GAAG;GAGrB,iBAAiB,SAAS;GAE1B,MAAM,eAAe,OAAO,gBAAgB;GAC5C,MAAM,cAAc,MAAM,aAAa,SAAS;AAEhD,OACE,OAAO,0BAA0B,QACjC,cAAc,OAAO,uBAErB;GAGF,MAAM,EAAE,cAAc,sBAAsB,GAC1C,mBAAmB,SAAS;GAC9B,MAAM,cAAc,eAClB,sBACA,OAAO,eACR;AAED,OAAI,eAAe,EACjB;GAGF,MAAM,EAAE,qBAAqB,mBAAmB,GAAG,kBACjD,cACA,sBACA,YACD;GAED,MAAM,UAAU,MAAM,cACpB,qBACA,OAAO,OACP,OAAO,eACP,aACD;GAED,MAAM,uBAAuB,0BAC3B,cACA,SACA,OAAO,cACR;AAED,UAAO,EACL,UAAU;IACR,IAAIC,wCAAc,EAAE,IAAIC,0CAAqB;IAC7C;IACA,GAAG;GACJ,EACF;EACF;CACF,EAAC;AACH;;;;AAKD,SAAS,iBAAiBL,UAA+B;AACvD,MAAK,MAAM,OAAO,SAChB,KAAI,CAAC,IAAI,IACP,IAAI,mBAAW;AAGpB;;;;AAKD,SAAS,mBAAmBA,UAG1B;AACA,KAAI,SAAS,SAAS,KAAKM,wCAAc,WAAW,SAAS,GAAG,CAC9D,QAAO;EACL,cAAc,SAAS;EACvB,sBAAsB,SAAS,MAAM,EAAE;CACxC;AAEH,QAAO;EACL,cAAc;EACd,sBAAsB;CACvB;AACF;;;;AAKD,SAAS,kBACPC,cACAC,sBACAC,aAC0E;CAC1E,MAAM,sBAAsB,qBAAqB,MAAM,GAAG,YAAY;CACtE,MAAM,oBAAoB,qBAAqB,MAAM,YAAY;AAGjE,KAAI,cACF,oBAAoB,QAAQ,aAAa;AAG3C,QAAO;EAAE;EAAqB;CAAmB;AAClD;;;;AAKD,SAAS,0BACPC,uBACAC,SACAC,eACe;CACf,IAAI,kBAAkB;AACtB,KAAI,uBAAuB;EACzB,MAAM,EAAE,oBAAS,GAAG;AACpB,MAAI,OAAOC,cAAY,UACrB,kBAAkBA,UAAQ,MAAM,cAAc,CAAC,GAAG,MAAM;CAE3D;CAED,MAAM,UAAU,kBACZ,GAAG,gBAAgB,EAAE,EAAE,cAAc,EAAE,EAAE,SAAS,GAClD,GAAG,cAAc,EAAE,EAAE,SAAS;AAElC,QAAO,IAAIP,wCAAc;EACvB;EACA,IAAI,uBAAuB,oBAAY;CACxC;AACF;;;;AAKD,SAAS,eACPN,UACAc,gBACQ;AACR,KAAI,SAAS,UAAU,eACrB,QAAO;CAGT,MAAM,eAAe,SAAS,SAAS;AAEvC,MAAK,IAAI,IAAI,cAAc,KAAK,GAAG,IACjC,KAAI,kBAAkB,UAAU,EAAE,CAChC,QAAO;AAIX,QAAO;AACR;;;;AAKD,SAAS,kBACPd,UACAS,aACS;AACT,KAAI,eAAe,SAAS,OAC1B,QAAO;CAGT,MAAM,cAAc,KAAK,IAAI,GAAG,cAAc,4BAA4B;CAC1E,MAAM,YAAY,KAAK,IACrB,SAAS,QACT,cAAc,4BACf;AAED,MAAK,IAAI,IAAI,aAAa,IAAI,WAAW,KAAK;AAC5C,MAAI,CAAC,aAAa,SAAS,GAAG,CAC5B;EAGF,MAAM,cAAc,mBAAmB,SAAS,GAAgB;AAChE,MAAI,wBAAwB,UAAU,GAAG,aAAa,YAAY,CAChE,QAAO;CAEV;AAED,QAAO;AACR;;;;AAKD,SAAS,aAAaM,SAA+B;AACnD,QACEC,oCAAU,WAAW,QAAQ,IAC7B,gBAAgB,WAChB,MAAM,QAAQ,QAAQ,WAAW,IACjC,QAAQ,WAAW,SAAS;AAE/B;;;;AAKD,SAAS,mBAAmBC,WAAmC;CAC7D,MAAM,8BAAc,IAAI;AACxB,KAAI,UAAU,WACZ,MAAK,MAAM,YAAY,UAAU,YAAY;EAC3C,MAAM,KACJ,OAAO,aAAa,YAAY,QAAQ,WAAW,SAAS,KAAK;AACnE,MAAI,IACF,YAAY,IAAI,GAAG;CAEtB;AAEH,QAAO;AACR;;;;AAKD,SAAS,wBACPjB,UACAkB,gBACAT,aACAU,aACS;AACT,MAAK,IAAI,IAAI,iBAAiB,GAAG,IAAI,SAAS,QAAQ,KAAK;EACzD,MAAM,UAAU,SAAS;AACzB,MACEC,sCAAY,WAAW,QAAQ,IAC/B,YAAY,IAAI,QAAQ,aAAa,EACrC;GACA,MAAM,iBAAiB,iBAAiB;GACxC,MAAM,mBAAmB,IAAI;AAC7B,OAAI,mBAAmB,iBACrB,QAAO;EAEV;CACF;AACD,QAAO;AACR;;;;AAKD,eAAe,cACbC,qBACAC,OACAC,eACAC,cACiB;AACjB,KAAI,CAAC,oBAAoB,OACvB,QAAO;CAGT,MAAM,kBAAkB,MAAM,uBAC5B,qBACA,aACD;AAED,KAAI,CAAC,gBAAgB,OACnB,QAAO;AAGT,KAAI;EACF,MAAM,kBAAkB,cAAc,QACpC,cACA,KAAK,UAAU,iBAAiB,MAAM,EAAE,CACzC;EACD,MAAM,WAAW,MAAM,MAAM,OAAO,gBAAgB;EACpD,MAAM,EAAE,SAAS,GAAG;AACpB,SAAO,OAAO,YAAY,WACtB,QAAQ,MAAM,GACd;CACL,SAAQ,GAAG;AACV,SAAO,CAAC,0BAA0B,EAAE,GAAG;CACxC;AACF;;;;AAKD,eAAe,uBACbxB,UACAwB,cACwB;AACxB,KAAI;AACF,SAAO,kDAAmB,UAAU;GAClC,WAAW;GACX,cAAc,OAAO,SAAS,QAAQ,QAAQ,aAAa,KAAK,CAAC;GACjE,UAAU;GACV,cAAc;GACd,eAAe;EAChB,EAAC;CACH,QAAO;AAEN,SAAO,SAAS,MAAM,CAAC,+BAA+B;CACvD;AACF"}

package/dist/agents/middlewareAgent/middleware/summarization.d.cts

@@ -1,91 +0,0 @@
-import { AgentMiddleware } from "../types.cjs";
-import * as _langchain_core_language_models_base1 from "@langchain/core/language_models/base";
-import { BaseLanguageModel } from "@langchain/core/language_models/base";
-import { InferInteropZodInput } from "@langchain/core/utils/types";
-import { BaseMessage } from "@langchain/core/messages";
-import { z } from "zod/v3";
-
-//#region src/agents/middlewareAgent/middleware/summarization.d.ts
-declare const contextSchema: z.ZodObject<{
-  model: z.ZodType<BaseLanguageModel<any, _langchain_core_language_models_base1.BaseLanguageModelCallOptions>, z.ZodTypeDef, BaseLanguageModel<any, _langchain_core_language_models_base1.BaseLanguageModelCallOptions>>;
-  maxTokensBeforeSummary: z.ZodOptional<z.ZodNumber>;
-  messagesToKeep: z.ZodDefault<z.ZodNumber>;
-  tokenCounter: z.ZodOptional<z.ZodFunction<z.ZodTuple<[z.ZodArray<z.ZodAny, "many">], z.ZodUnknown>, z.ZodUnion<[z.ZodNumber, z.ZodPromise<z.ZodNumber>]>>>;
-  summaryPrompt: z.ZodDefault<z.ZodString>;
-  summaryPrefix: z.ZodDefault<z.ZodString>;
-}, "strip", z.ZodTypeAny, {
-  model: BaseLanguageModel<any, _langchain_core_language_models_base1.BaseLanguageModelCallOptions>;
-  maxTokensBeforeSummary?: number | undefined;
-  messagesToKeep: number;
-  tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;
-  summaryPrompt: string;
-  summaryPrefix: string;
-}, {
-  model: BaseLanguageModel<any, _langchain_core_language_models_base1.BaseLanguageModelCallOptions>;
-  maxTokensBeforeSummary?: number | undefined;
-  messagesToKeep?: number | undefined;
-  tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;
-  summaryPrompt?: string | undefined;
-  summaryPrefix?: string | undefined;
-}>;
-type SummarizationMiddlewareConfig = InferInteropZodInput<typeof contextSchema>;
-/**
- * Default token counter that approximates based on character count
- * @param messages Messages to count tokens for
- * @returns Approximate token count
- */
-declare function countTokensApproximately(messages: BaseMessage[]): number;
-/**
- * Summarization middleware that automatically summarizes conversation history when token limits are approached.
- *
- * This middleware monitors message token counts and automatically summarizes older
- * messages when a threshold is reached, preserving recent messages and maintaining
- * context continuity by ensuring AI/Tool message pairs remain together.
- *
- * @param options Configuration options for the summarization middleware
- * @returns A middleware instance
- *
- * @example
- * ```ts
- * import { summarizationMiddleware } from "langchain/middleware";
- * import { createAgent } from "langchain";
- *
- * const agent = createAgent({
- *   llm: model,
- *   tools: [getWeather],
- *   middleware: [
- *     summarizationMiddleware({
- *       model: new ChatOpenAI({ model: "gpt-4o" }),
- *       maxTokensBeforeSummary: 4000,
- *       messagesToKeep: 20,
- *     })
- *   ],
- * });
- *
- * ```
- */
-declare function summarizationMiddleware(options: SummarizationMiddlewareConfig): AgentMiddleware<undefined, z.ZodObject<{
-  model: z.ZodType<BaseLanguageModel<any, _langchain_core_language_models_base1.BaseLanguageModelCallOptions>, z.ZodTypeDef, BaseLanguageModel<any, _langchain_core_language_models_base1.BaseLanguageModelCallOptions>>;
-  maxTokensBeforeSummary: z.ZodOptional<z.ZodNumber>;
-  messagesToKeep: z.ZodDefault<z.ZodNumber>;
-  tokenCounter: z.ZodOptional<z.ZodFunction<z.ZodTuple<[z.ZodArray<z.ZodAny, "many">], z.ZodUnknown>, z.ZodUnion<[z.ZodNumber, z.ZodPromise<z.ZodNumber>]>>>;
-  summaryPrompt: z.ZodDefault<z.ZodString>;
-  summaryPrefix: z.ZodDefault<z.ZodString>;
-}, "strip", z.ZodTypeAny, {
-  model: BaseLanguageModel<any, _langchain_core_language_models_base1.BaseLanguageModelCallOptions>;
-  maxTokensBeforeSummary?: number | undefined;
-  messagesToKeep: number;
-  tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;
-  summaryPrompt: string;
-  summaryPrefix: string;
-}, {
-  model: BaseLanguageModel<any, _langchain_core_language_models_base1.BaseLanguageModelCallOptions>;
-  maxTokensBeforeSummary?: number | undefined;
-  messagesToKeep?: number | undefined;
-  tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;
-  summaryPrompt?: string | undefined;
-  summaryPrefix?: string | undefined;
-}>, any>;
-//#endregion
-export { SummarizationMiddlewareConfig, countTokensApproximately, summarizationMiddleware };
-//# sourceMappingURL=summarization.d.cts.map