langchain 1.0.2 → 1.0.3

package/dist/agents/middleware/summarization.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"summarization.d.ts","names":["__types_js0","z","BaseLanguageModel","InferInteropZodInput","contextSchema","_langchain_core_language_models_base0","BaseLanguageModelCallOptions","ZodTypeDef","ZodType","ZodNumber","ZodOptional","ZodDefault","ZodAny","ZodArray","ZodUnknown","ZodTuple","ZodPromise","ZodUnion","ZodFunction","ZodString","ZodTypeAny","Promise","ZodObject","SummarizationMiddlewareConfig","summarizationMiddleware","AgentMiddleware"],"sources":["../../../src/agents/middleware/summarization.d.ts"],"sourcesContent":["import { z } from \"zod/v3\";\nimport { BaseLanguageModel } from \"@langchain/core/language_models/base\";\nimport { InferInteropZodInput } from \"@langchain/core/utils/types\";\ndeclare const contextSchema: z.ZodObject<{\n    model: z.ZodType<BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>, z.ZodTypeDef, BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>>;\n    maxTokensBeforeSummary: z.ZodOptional<z.ZodNumber>;\n    messagesToKeep: z.ZodDefault<z.ZodNumber>;\n    tokenCounter: z.ZodOptional<z.ZodFunction<z.ZodTuple<[z.ZodArray<z.ZodAny, \"many\">], z.ZodUnknown>, z.ZodUnion<[z.ZodNumber, z.ZodPromise<z.ZodNumber>]>>>;\n    summaryPrompt: z.ZodDefault<z.ZodString>;\n    summaryPrefix: z.ZodDefault<z.ZodString>;\n}, \"strip\", z.ZodTypeAny, {\n    model: BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>;\n    maxTokensBeforeSummary?: number | undefined;\n    messagesToKeep: number;\n    tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;\n    summaryPrompt: string;\n    summaryPrefix: string;\n}, {\n    model: BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>;\n    maxTokensBeforeSummary?: number | undefined;\n    messagesToKeep?: number | undefined;\n    tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;\n    summaryPrompt?: string | undefined;\n    summaryPrefix?: string | undefined;\n}>;\nexport type SummarizationMiddlewareConfig = InferInteropZodInput<typeof contextSchema>;\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\";\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 declare function summarizationMiddleware(options: SummarizationMiddlewareConfig): import(\"./types.js\").AgentMiddleware<undefined, z.ZodObject<{\n    model: z.ZodType<BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>, z.ZodTypeDef, BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>>;\n    maxTokensBeforeSummary: z.ZodOptional<z.ZodNumber>;\n    messagesToKeep: z.ZodDefault<z.ZodNumber>;\n    tokenCounter: z.ZodOptional<z.ZodFunction<z.ZodTuple<[z.ZodArray<z.ZodAny, \"many\">], z.ZodUnknown>, z.ZodUnion<[z.ZodNumber, z.ZodPromise<z.ZodNumber>]>>>;\n    summaryPrompt: z.ZodDefault<z.ZodString>;\n    summaryPrefix: z.ZodDefault<z.ZodString>;\n}, \"strip\", z.ZodTypeAny, {\n    model: BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>;\n    maxTokensBeforeSummary?: number | undefined;\n    messagesToKeep: number;\n    tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;\n    summaryPrompt: string;\n    summaryPrefix: string;\n}, {\n    model: BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>;\n    maxTokensBeforeSummary?: number | undefined;\n    messagesToKeep?: number | undefined;\n    tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;\n    summaryPrompt?: string | undefined;\n    summaryPrefix?: string | undefined;\n}>, any>;\nexport {};\n"],"mappings":";;;;;;;cAGcI,eAAeH,CAAAA,CAAEqB;SACpBrB,CAAAA,CAAEO,QAAQN,uBAoBnBG,qCAAAA,CApByFC,4BAAAA,GAA+BL,CAAAA,CAAEM,YAAYL,uBAAFG,qCAAAA,CAAwEC,4BAAAA;0BAClLL,CAAAA,CAAES,YAAYT,CAAAA,CAAEQ;kBACxBR,CAAAA,CAAEU,WAAWV,CAAAA,CAAEQ;EAHrBL,YAAAA,EAIIH,CAAAA,CAAES,WAiBlB,CAjB8BT,CAAAA,CAAEiB,WAiBhC,CAjB4CjB,CAAAA,CAAEc,QAiB9C,CAAA,CAjBwDd,CAAAA,CAAEY,QAiB1D,CAjBmEZ,CAAAA,CAAEW,MAiBrE,EAAA,MAAA,CAAA,CAAA,EAjBuFX,CAAAA,CAAEa,UAiBzF,CAAA,EAjBsGb,CAAAA,CAAEgB,QAiBxG,CAAA,CAjBkHhB,CAAAA,CAAEQ,SAiBpH,EAjB+HR,CAAAA,CAAEe,UAiBjI,CAjB4If,CAAAA,CAAEQ,SAiB9I,CAAA,CAAA,CAAA,CAAA,CAAA;EAAA,aAAA,EAhBiBR,CAAAA,CAAEU,UAgBnB,CAhB8BV,CAAAA,CAAEkB,SAgBhC,CAAA;EAAA,aAAAd,EAfiBJ,CAAAA,CAAEU,UAenBN,CAf8BJ,CAAAA,CAAEkB,SAehCd,CAAAA;CApBqH,EAAA,OAAlGH,EAMTD,CAAAA,CAAEmB,UANOlB,EAAAA;EAAiB,KAAsFK,EAOjHL,iBAPiHK,CAAAA,GAAAA,EAMpGF,qCAAAA,CACyDC,4BAAAA,CAP2CC;EAAU,sBAAAF,CAAAA,EAAAA,MAAAA,GAAAA,SAAwEC;EAA4B,cAAlGJ,EAAAA,MAAAA;EAAiB,YAA5IM,CAAAA,EAAAA,CAAAA,CAAAA,MAAAA,EAAAA,GAAAA,EAAAA,EAAAA,GAAAA,IAAAA,EAAAA,OAAAA,EAAAA,EAAAA,GAAAA,MAAAA,GAUuDa,OAVvDb,CAAAA,MAAAA,CAAAA,CAAAA,GAAAA,SAAAA;EAAO,aACwBC,EAAAA,MAAAA;EAAS,aAAvBC,EAAAA,MAAAA;CAAW,EAAA;EACG,KAAtBC,EAYXT,iBAZWS,CAAAA,GAAAA,EAQqDN,qCAAAA,CAIMC,4BAAAA,CAZ3DK;EAAU,sBACuCC,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAM,cAAjBC,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAQ,YAAuBC,CAAAA,EAAAA,CAAAA,CAAAA,MAAAA,EAAAA,GAAAA,EAAAA,EAAAA,GAAAA,IAAAA,EAAAA,OAAAA,EAAAA,EAAAA,GAAAA,MAAAA,GAcvBO,OAduBP,CAAAA,MAAAA,CAAAA,CAAAA,GAAAA,SAAAA;EAAU,aAArDC,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAQ,aAA8DN,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;CAAS,CAAA;AAAER,KAkBrHsB,6BAAAA,GAAgCpB,oBAlBuFa,CAAAA,OAkB3DZ,aAlB2DY,CAAAA;;;;;;;;;;;;;;;;AAJ3F;AAsBxC;;;;AAAgE;AA8BhE;;;;;;;;AACWf,iBADauB,uBAAAA,CACXhB,OAAAA,EAD4Ce,6BAC5Cf,CAAAA,EADuI,eACvIA,CAAAA,SAAAA,EAD4HP,CAAAA,CAAEqB,SAC9Hd,CAAAA;EAAO,KACwBC,EADjCR,CAAAA,CAAEO,OAC+BC,CADvBP,iBACuBO,CAAAA,GAAAA,EAF0CJ,qCAAAA,CACKC,4BAAAA,CAC/CG,EAD8ER,CAAAA,CAAEM,UAChFE,EAD4FP,iBAC5FO,CAAAA,GAAAA,EAD0FJ,qCAAAA,CAAwEC,4BAAAA,CAClKG,CAAAA;EAAS,sBAAvBC,EAAFT,CAAAA,CAAES,WAAAA,CAAYT,CAAAA,CAAEQ,SAAdC,CAAAA;EAAW,cACND,EAAfR,CAAAA,CAAEU,UAAaF,CAAFR,CAAAA,CAAEQ,SAAAA,CAAAA;EAAS,YAAtBE,EACJV,CAAAA,CAAES,WADEC,CACUV,CAAAA,CAAEiB,WADZP,CACwBV,CAAAA,CAAEc,QAD1BJ,CAAAA,CACoCV,CAAAA,CAAEY,QADtCF,CAC+CV,CAAAA,CAAEW,MADjDD,EAAAA,MAAAA,CAAAA,CAAAA,EACmEV,CAAAA,CAAEa,UADrEH,CAAAA,EACkFV,CAAAA,CAAEgB,QADpFN,CAAAA,CAC8FV,CAAAA,CAAEQ,SADhGE,EAC2GV,CAAAA,CAAEe,UAD7GL,CACwHV,CAAAA,CAAEQ,SAD1HE,CAAAA,CAAAA,CAAAA,CAAAA,CAAAA;EAAU,aACuCC,EACpDX,CAAAA,CAAEU,UADkDC,CACvCX,CAAAA,CAAEkB,SADqCP,CAAAA;EAAM,aAAjBC,EAEzCZ,CAAAA,CAAEU,UAFuCE,CAE5BZ,CAAAA,CAAEkB,SAF0BN,CAAAA;CAAQ,EAAA,OAAuBC,EAG/Eb,CAAAA,CAAEmB,UAH6EN,EAAAA;EAAU,KAArDC,EAIrCb,iBAJqCa,CAAAA,GAAAA,EAGxBV,qCAAAA,CACyDC,4BAAAA,CAJjCS;EAAQ,sBAA8DN,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAS,cAAiBA,EAAAA,MAAAA;EAAS,YAAtBO,CAAAA,EAAAA,CAAAA,CAAAA,MAAAA,EAAAA,GAAAA,EAAAA,EAAAA,GAAAA,IAAAA,EAAAA,OAAAA,EAAAA,EAAAA,GAAAA,MAAAA,GAO/DK,OAP+DL,CAAAA,MAAAA,CAAAA,CAAAA,GAAAA,SAAAA;EAAU,aAAnCC,EAAAA,MAAAA;EAAQ,aAAhFC,EAAAA,MAAAA;CAAW,EAAA;EAAd,KACGC,EAUvBjB,iBAVuBiB,CAAAA,GAAAA,EAMyCd,qCAAAA,CAIMC,4BAAAA,CAV/Ca;EAAS,sBAAtBR,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAU,cACGQ,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAS,YAAtBR,CAAAA,EAAAA,CAAAA,CAAAA,MAAAA,EAAAA,GAAAA,EAAAA,EAAAA,GAAAA,IAAAA,EAAAA,OAAAA,EAAAA,EAAAA,GAAAA,MAAAA,GAY+CU,OAZ/CV,CAAAA,MAAAA,CAAAA,CAAAA,GAAAA,SAAAA;EAAU,aACjBS,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAU,aAAAf,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;CACqF,CAAA,EAAA,GAAlGH,CAAAA"}
+{"version":3,"file":"summarization.d.ts","names":["__types_js0","z","BaseLanguageModel","InferInteropZodInput","contextSchema","_langchain_core_language_models_base0","BaseLanguageModelCallOptions","ZodTypeDef","ZodType","ZodNumber","ZodOptional","ZodDefault","ZodAny","ZodArray","ZodUnknown","ZodTuple","ZodPromise","ZodUnion","ZodFunction","ZodString","ZodTypeAny","Promise","ZodObject","SummarizationMiddlewareConfig","summarizationMiddleware","AgentMiddleware"],"sources":["../../../src/agents/middleware/summarization.d.ts"],"sourcesContent":["import { z } from \"zod/v3\";\nimport { BaseLanguageModel } from \"@langchain/core/language_models/base\";\nimport { InferInteropZodInput } from \"@langchain/core/utils/types\";\ndeclare const contextSchema: z.ZodObject<{\n    model: z.ZodType<string | BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>, z.ZodTypeDef, string | BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>>;\n    maxTokensBeforeSummary: z.ZodOptional<z.ZodNumber>;\n    messagesToKeep: z.ZodDefault<z.ZodNumber>;\n    tokenCounter: z.ZodOptional<z.ZodFunction<z.ZodTuple<[z.ZodArray<z.ZodAny, \"many\">], z.ZodUnknown>, z.ZodUnion<[z.ZodNumber, z.ZodPromise<z.ZodNumber>]>>>;\n    summaryPrompt: z.ZodDefault<z.ZodString>;\n    summaryPrefix: z.ZodDefault<z.ZodString>;\n}, \"strip\", z.ZodTypeAny, {\n    model: string | BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>;\n    maxTokensBeforeSummary?: number | undefined;\n    messagesToKeep: number;\n    tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;\n    summaryPrompt: string;\n    summaryPrefix: string;\n}, {\n    model: string | BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>;\n    maxTokensBeforeSummary?: number | undefined;\n    messagesToKeep?: number | undefined;\n    tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;\n    summaryPrompt?: string | undefined;\n    summaryPrefix?: string | undefined;\n}>;\nexport type SummarizationMiddlewareConfig = InferInteropZodInput<typeof contextSchema>;\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\";\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 declare function summarizationMiddleware(options: SummarizationMiddlewareConfig): import(\"./types.js\").AgentMiddleware<undefined, z.ZodObject<{\n    maxTokensBeforeSummary: z.ZodOptional<z.ZodNumber>;\n    messagesToKeep: z.ZodDefault<z.ZodNumber>;\n    tokenCounter: z.ZodOptional<z.ZodFunction<z.ZodTuple<[z.ZodArray<z.ZodAny, \"many\">], z.ZodUnknown>, z.ZodUnion<[z.ZodNumber, z.ZodPromise<z.ZodNumber>]>>>;\n    summaryPrompt: z.ZodDefault<z.ZodString>;\n    summaryPrefix: z.ZodDefault<z.ZodString>;\n} & {\n    model: z.ZodOptional<z.ZodType<BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>, z.ZodTypeDef, BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions>>>;\n}, \"strip\", z.ZodTypeAny, {\n    maxTokensBeforeSummary?: number | undefined;\n    messagesToKeep: number;\n    tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;\n    summaryPrompt: string;\n    summaryPrefix: string;\n    model?: BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions> | undefined;\n}, {\n    maxTokensBeforeSummary?: number | undefined;\n    messagesToKeep?: number | undefined;\n    tokenCounter?: ((args_0: any[], ...args: unknown[]) => number | Promise<number>) | undefined;\n    summaryPrompt?: string | undefined;\n    summaryPrefix?: string | undefined;\n    model?: BaseLanguageModel<any, import(\"@langchain/core/language_models/base\").BaseLanguageModelCallOptions> | undefined;\n}>, any>;\nexport {};\n"],"mappings":";;;;;;;cAGcI,eAAeH,CAAAA,CAAEqB;SACpBrB,CAAAA,CAAEO,iBAAiBN,uBAoB5BG,qCAAAA,CApBkGC,4BAAAA,GAA+BL,CAAAA,CAAEM,qBAAqBL,uBAAXG,qCAAAA,CAAiFC,4BAAAA;0BACpML,CAAAA,CAAES,YAAYT,CAAAA,CAAEQ;kBACxBR,CAAAA,CAAEU,WAAWV,CAAAA,CAAEQ;EAHrBL,YAAAA,EAIIH,CAAAA,CAAES,WAiBlB,CAjB8BT,CAAAA,CAAEiB,WAiBhC,CAjB4CjB,CAAAA,CAAEc,QAiB9C,CAAA,CAjBwDd,CAAAA,CAAEY,QAiB1D,CAjBmEZ,CAAAA,CAAEW,MAiBrE,EAAA,MAAA,CAAA,CAAA,EAjBuFX,CAAAA,CAAEa,UAiBzF,CAAA,EAjBsGb,CAAAA,CAAEgB,QAiBxG,CAAA,CAjBkHhB,CAAAA,CAAEQ,SAiBpH,EAjB+HR,CAAAA,CAAEe,UAiBjI,CAjB4If,CAAAA,CAAEQ,SAiB9I,CAAA,CAAA,CAAA,CAAA,CAAA;EAAA,aAAA,EAhBiBR,CAAAA,CAAEU,UAgBnB,CAhB8BV,CAAAA,CAAEkB,SAgBhC,CAAA;EAAA,aAAAd,EAfiBJ,CAAAA,CAAEU,UAenBN,CAf8BJ,CAAAA,CAAEkB,SAehCd,CAAAA;CApB8H,EAAA,OAAlGH,EAMlBD,CAAAA,CAAEmB,UANgBlB,EAAAA;EAAiB,KAAsFK,EAAAA,MAAAA,GAOjHL,iBAPiHK,CAAAA,GAAAA,EAM7GF,qCAAAA,CACkEC,4BAAAA,CAP2CC;EAAU,sBAAAF,CAAAA,EAAAA,MAAAA,GAAAA,SAAiFC;EAA4B,cAAlGJ,EAAAA,MAAAA;EAAiB,YAA9JM,CAAAA,EAAAA,CAAAA,CAAAA,MAAAA,EAAAA,GAAAA,EAAAA,EAAAA,GAAAA,IAAAA,EAAAA,OAAAA,EAAAA,EAAAA,GAAAA,MAAAA,GAUuDa,OAVvDb,CAAAA,MAAAA,CAAAA,CAAAA,GAAAA,SAAAA;EAAO,aACwBC,EAAAA,MAAAA;EAAS,aAAvBC,EAAAA,MAAAA;CAAW,EAAA;EACG,KAAtBC,EAAAA,MAAAA,GAYFT,iBAZES,CAAAA,GAAAA,EAQqDN,qCAAAA,CAIeC,4BAAAA,CAZpEK;EAAU,sBACuCC,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAM,cAAjBC,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAQ,YAAuBC,CAAAA,EAAAA,CAAAA,CAAAA,MAAAA,EAAAA,GAAAA,EAAAA,EAAAA,GAAAA,IAAAA,EAAAA,OAAAA,EAAAA,EAAAA,GAAAA,MAAAA,GAcvBO,OAduBP,CAAAA,MAAAA,CAAAA,CAAAA,GAAAA,SAAAA;EAAU,aAArDC,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAQ,aAA8DN,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;CAAS,CAAA;AAAER,KAkBrHsB,6BAAAA,GAAgCpB,oBAlBuFa,CAAAA,OAkB3DZ,aAlB2DY,CAAAA;;;;;;;;;;;;;;;;AAJ3F;AAsBxC;;;;AAAgE;AA8BhE;;;;;;;;AAG0Df,iBAHlCuB,uBAAAA,CAGoCX,OAAAA,EAHHU,6BAGGV,CAAAA,EAHwF,eAGxFA,CAAAA,SAAAA,EAH6EZ,CAAAA,CAAEqB,SAG/ET,CAAAA;EAAQ,sBAAuBC,EAF/Db,CAAAA,CAAES,WAE6DI,CAFjDb,CAAAA,CAAEQ,SAE+CK,CAAAA;EAAU,cAArDC,EAD5Bd,CAAAA,CAAEU,UAC0BI,CADfd,CAAAA,CAAEQ,SACaM,CAAAA;EAAQ,YAA8DN,EAApGR,CAAAA,CAAES,WAAkGD,CAAtFR,CAAAA,CAAEiB,WAAoFT,CAAxER,CAAAA,CAAEc,QAAsEN,CAAAA,CAA5DR,CAAAA,CAAEY,QAA0DJ,CAAjDR,CAAAA,CAAEW,MAA+CH,EAAAA,MAAAA,CAAAA,CAAAA,EAA7BR,CAAAA,CAAEa,UAA2BL,CAAAA,EAAdR,CAAAA,CAAEgB,QAAYR,CAAAA,CAAFR,CAAAA,CAAEQ,SAAAA,EAAWR,CAAAA,CAAEe,UAAbP,CAAwBR,CAAAA,CAAEQ,SAA1BA,CAAAA,CAAAA,CAAAA,CAAAA,CAAAA;EAAS,aAAiBA,EAC7HR,CAAAA,CAAEU,UAD2HF,CAChHR,CAAAA,CAAEkB,SAD8GV,CAAAA;EAAS,aAAtBO,EAEhHf,CAAAA,CAAEU,UAF8GK,CAEnGf,CAAAA,CAAEkB,SAFiGH,CAAAA;CAAU,GAAA;EAA3B,KAAhFE,EAIvBjB,CAAAA,CAAES,WAJqBQ,CAITjB,CAAAA,CAAEO,OAJOU,CAIChB,iBAJDgB,CAAAA,GAAAA,EAEHb,qCAAAA,CAE0EC,4BAAAA,CAJvEY,EAIsGjB,CAAAA,CAAEM,UAJxGW,EAIoHhB,iBAJpHgB,CAAAA,GAAAA,EAIkHb,qCAAAA,CAAwEC,4BAAAA,CAJ1LY,CAAAA,CAAAA;CAAW,EAAA,OAAzBR,EAKRT,CAAAA,CAAEmB,UALMV,EAAAA;EAAW,sBACGS,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAS,cAAtBR,EAAAA,MAAAA;EAAU,YACGQ,CAAAA,EAAAA,CAAAA,CAAAA,MAAAA,EAAAA,GAAAA,EAAAA,EAAAA,GAAAA,IAAAA,EAAAA,OAAAA,EAAAA,EAAAA,GAAAA,MAAAA,GAMkCE,OANlCF,CAAAA,MAAAA,CAAAA,CAAAA,GAAAA,SAAAA;EAAS,aAAtBR,EAAAA,MAAAA;EAAU,aAAAN,EAAAA,MAAAA;EAEsG,KAAlGH,CAAAA,EAOvBA,iBAPuBA,CAAAA,GAAAA,EAIwCG,qCAAAA,CAGOC,4BAAAA,CAP/CJ,GAAAA,SAAAA;CAAiB,EAAA;EAAgG,sBAAAG,CAAAA,EAAAA,MAAAA,GAAAA,SAAwEC;EAA4B,cAAlGJ,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAiB,YAA5IM,CAAAA,EAAAA,CAAAA,CAAAA,MAAAA,EAAAA,GAAAA,EAAAA,EAAAA,GAAAA,IAAAA,EAAAA,OAAAA,EAAAA,EAAAA,GAAAA,MAAAA,GAWyCa,OAXzCb,CAAAA,MAAAA,CAAAA,CAAAA,GAAAA,SAAAA;EAAO,aAArBE,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAW,aACVU,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAU,KAG4CC,CAAAA,EAUxDnB,iBAVwDmB,CAAAA,GAAAA,EAOOhB,qCAAAA,CAGOC,4BAAAA,CAVde,GAAAA,SAAAA;CAAO,CAAA,EAAA,GAAAhB,CAAAA"}

package/dist/agents/middleware/summarization.js

@@ -1,3 +1,4 @@
+import { initChatModel } from "../../chat_models/universal.js";
 import { hasToolCalls } from "../utils.js";
 import { countTokensApproximately } from "./utils.js";
 import { createMiddleware } from "../middleware.js";
@@ -80,7 +81,7 @@ const contextSchema = z.object({
 function summarizationMiddleware(options) {
 	return createMiddleware({
 		name: "SummarizationMiddleware",
-		contextSchema,
+		contextSchema: contextSchema.extend({ model: z.custom().optional() }),
 		beforeModel: async (state, runtime) => {
 			/**
 			* Parse user options to get their explicit values
@@ -98,6 +99,7 @@ function summarizationMiddleware(options) {
 				summaryPrefix: runtime.context.summaryPrefix === SUMMARY_PREFIX ? userOptions.summaryPrefix : runtime.context.summaryPrefix ?? userOptions.summaryPrefix
 			};
 			const { messages } = state;
+			const model = typeof config.model === "string" ? await initChatModel(config.model) : config.model;
 			ensureMessageIds(messages);
 			const tokenCounter = config.tokenCounter || countTokensApproximately;
 			const totalTokens = await tokenCounter(messages);
@@ -106,7 +108,7 @@ function summarizationMiddleware(options) {
 			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 summary = await createSummary(messagesToSummarize, model, config.summaryPrompt, tokenCounter);
 			const updatedSystemMessage = buildUpdatedSystemMessage(systemPrompt, summary, config.summaryPrefix);
 			return { messages: [
 				new RemoveMessage({ id: REMOVE_ALL_MESSAGES }),

package/dist/agents/middleware/summarization.js.map

@@ -1 +1 @@
-{"version":3,"file":"summarization.js","names":["options: SummarizationMiddlewareConfig","messages: BaseMessage[]","uuid","systemPrompt: SystemMessage | null","conversationMessages: BaseMessage[]","cutoffIndex: number","originalSystemMessage: SystemMessage | null","summary: string","summaryPrefix: string","content","messagesToKeep: number","aiMessage: AIMessage","aiMessageIndex: number","toolCallIds: Set<string>","messagesToSummarize: BaseMessage[]","model: BaseLanguageModel","summaryPrompt: string","tokenCounter: TokenCounter"],"sources":["../../../src/agents/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\";\nimport { countTokensApproximately } from \"./utils.js\";\nimport { hasToolCalls } from \"../utils.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 * 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\";\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 * 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":";;;;;;;;;;AAqBA,MAAM,yBAAyB,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;WA0BrB,CAAC;AAEZ,MAAM,iBAAiB;AAEvB,MAAM,2BAA2B;AACjC,MAAM,2BAA2B;AACjC,MAAM,iCAAiC;AACvC,MAAM,8BAA8B;AAIpC,MAAM,gBAAgB,EAAE,OAAO;CAC7B,OAAO,EAAE,QAA2B;CACpC,wBAAwB,EAAE,QAAQ,CAAC,UAAU;CAC7C,gBAAgB,EAAE,QAAQ,CAAC,QAAQ,yBAAyB;CAC5D,cAAc,EACX,UAAU,CACV,KAAK,EAAE,MAAM,EAAE,KAAK,CAAC,CAAC,CACtB,QAAQ,EAAE,MAAM,CAAC,EAAE,QAAQ,EAAE,EAAE,QAAQ,EAAE,QAAQ,CAAC,AAAC,EAAC,CAAC,CACrD,UAAU;CACb,eAAe,EAAE,QAAQ,CAAC,QAAQ,uBAAuB;CACzD,eAAe,EAAE,QAAQ,CAAC,QAAQ,eAAe;AAClD,EAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCF,SAAgB,wBACdA,SACA;AACA,QAAO,iBAAiB;EACtB,MAAM;EACN;EACA,aAAa,OAAO,OAAO,YAAY;;;;GAIrC,MAAM,cAAc,aAAa,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,IAAI,cAAc,EAAE,IAAI,oBAAqB;IAC7C;IACA,GAAG;GACJ,EACF;EACF;CACF,EAAC;AACH;;;;AAKD,SAAS,iBAAiBC,UAA+B;AACvD,MAAK,MAAM,OAAO,SAChB,KAAI,CAAC,IAAI,IACP,IAAI,KAAKC,IAAM;AAGpB;;;;AAKD,SAAS,mBAAmBD,UAG1B;AACA,KAAI,SAAS,SAAS,KAAK,cAAc,WAAW,SAAS,GAAG,CAC9D,QAAO;EACL,cAAc,SAAS;EACvB,sBAAsB,SAAS,MAAM,EAAE;CACxC;AAEH,QAAO;EACL,cAAc;EACd,sBAAsB;CACvB;AACF;;;;AAKD,SAAS,kBACPE,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,IAAI,cAAc;EACvB;EACA,IAAI,uBAAuB,MAAMP,IAAM;CACxC;AACF;;;;AAKD,SAAS,eACPD,UACAS,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,kBACPT,UACAI,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,mBAAmBM,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,wBACPV,UACAW,gBACAP,aACAQ,aACS;AACT,MAAK,IAAI,IAAI,iBAAiB,GAAG,IAAI,SAAS,QAAQ,KAAK;EACzD,MAAM,UAAU,SAAS;AACzB,MACE,YAAY,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,uBACbhB,UACAgB,cACwB;AACxB,KAAI;AACF,SAAO,MAAM,aAAa,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"}
+{"version":3,"file":"summarization.js","names":["options: SummarizationMiddlewareConfig","messages: BaseMessage[]","uuid","systemPrompt: SystemMessage | null","conversationMessages: BaseMessage[]","cutoffIndex: number","originalSystemMessage: SystemMessage | null","summary: string","summaryPrefix: string","content","messagesToKeep: number","aiMessage: AIMessage","aiMessageIndex: number","toolCallIds: Set<string>","messagesToSummarize: BaseMessage[]","model: BaseLanguageModel","summaryPrompt: string","tokenCounter: TokenCounter"],"sources":["../../../src/agents/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\";\nimport { countTokensApproximately } from \"./utils.js\";\nimport { hasToolCalls } from \"../utils.js\";\nimport { initChatModel } from \"../../chat_models/universal.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<string | 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 * 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\";\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: contextSchema.extend({\n      /**\n       * `model` should be required when initializing the middleware,\n       * but can be omitted within context when invoking the middleware.\n       */\n      model: z.custom<BaseLanguageModel>().optional(),\n    }),\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      const model =\n        typeof config.model === \"string\"\n          ? await initChatModel(config.model)\n          : config.model;\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        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 * 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":";;;;;;;;;;;AAsBA,MAAM,yBAAyB,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;WA0BrB,CAAC;AAEZ,MAAM,iBAAiB;AAEvB,MAAM,2BAA2B;AACjC,MAAM,2BAA2B;AACjC,MAAM,iCAAiC;AACvC,MAAM,8BAA8B;AAIpC,MAAM,gBAAgB,EAAE,OAAO;CAC7B,OAAO,EAAE,QAAoC;CAC7C,wBAAwB,EAAE,QAAQ,CAAC,UAAU;CAC7C,gBAAgB,EAAE,QAAQ,CAAC,QAAQ,yBAAyB;CAC5D,cAAc,EACX,UAAU,CACV,KAAK,EAAE,MAAM,EAAE,KAAK,CAAC,CAAC,CACtB,QAAQ,EAAE,MAAM,CAAC,EAAE,QAAQ,EAAE,EAAE,QAAQ,EAAE,QAAQ,CAAC,AAAC,EAAC,CAAC,CACrD,UAAU;CACb,eAAe,EAAE,QAAQ,CAAC,QAAQ,uBAAuB;CACzD,eAAe,EAAE,QAAQ,CAAC,QAAQ,eAAe;AAClD,EAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCF,SAAgB,wBACdA,SACA;AACA,QAAO,iBAAiB;EACtB,MAAM;EACN,eAAe,cAAc,OAAO,EAKlC,OAAO,EAAE,QAA2B,CAAC,UAAU,CAChD,EAAC;EACF,aAAa,OAAO,OAAO,YAAY;;;;GAIrC,MAAM,cAAc,aAAa,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;GAErB,MAAM,QACJ,OAAO,OAAO,UAAU,WACpB,MAAM,cAAc,OAAO,MAAM,GACjC,OAAO;GAGb,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,OACA,OAAO,eACP,aACD;GAED,MAAM,uBAAuB,0BAC3B,cACA,SACA,OAAO,cACR;AAED,UAAO,EACL,UAAU;IACR,IAAI,cAAc,EAAE,IAAI,oBAAqB;IAC7C;IACA,GAAG;GACJ,EACF;EACF;CACF,EAAC;AACH;;;;AAKD,SAAS,iBAAiBC,UAA+B;AACvD,MAAK,MAAM,OAAO,SAChB,KAAI,CAAC,IAAI,IACP,IAAI,KAAKC,IAAM;AAGpB;;;;AAKD,SAAS,mBAAmBD,UAG1B;AACA,KAAI,SAAS,SAAS,KAAK,cAAc,WAAW,SAAS,GAAG,CAC9D,QAAO;EACL,cAAc,SAAS;EACvB,sBAAsB,SAAS,MAAM,EAAE;CACxC;AAEH,QAAO;EACL,cAAc;EACd,sBAAsB;CACvB;AACF;;;;AAKD,SAAS,kBACPE,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,IAAI,cAAc;EACvB;EACA,IAAI,uBAAuB,MAAMP,IAAM;CACxC;AACF;;;;AAKD,SAAS,eACPD,UACAS,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,kBACPT,UACAI,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,mBAAmBM,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,wBACPV,UACAW,gBACAP,aACAQ,aACS;AACT,MAAK,IAAI,IAAI,iBAAiB,GAAG,IAAI,SAAS,QAAQ,KAAK;EACzD,MAAM,UAAU,SAAS;AACzB,MACE,YAAY,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,uBACbhB,UACAgB,cACwB;AACxB,KAAI;AACF,SAAO,MAAM,aAAa,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/middleware/todoListMiddleware.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"todoListMiddleware.d.ts","names":["z","TODO_LIST_MIDDLEWARE_SYSTEM_PROMPT","stateSchema","ZodString","ZodEnum","ZodTypeAny","ZodObject","ZodArray","ZodDefault","TodoMiddlewareState","infer","TodoListMiddlewareOptions","todoListMiddleware","__types_js8","AgentMiddleware"],"sources":["../../../src/agents/middleware/todoListMiddleware.d.ts"],"sourcesContent":["import { z } from \"zod/v3\";\nexport declare const TODO_LIST_MIDDLEWARE_SYSTEM_PROMPT = \"## `write_todos`\\n\\nYou have access to the `write_todos` tool to help you manage and plan complex objectives. \\nUse this tool for complex objectives to ensure that you are tracking each necessary step and giving the user visibility into your progress.\\nThis tool is very helpful for planning complex objectives, and for breaking down these larger complex objectives into smaller steps.\\n\\nIt is critical that you mark todos as completed as soon as you are done with a step. Do not batch up multiple steps before marking them as completed.\\nFor simple objectives that only require a few steps, it is better to just complete the objective directly and NOT use this tool.\\nWriting todos takes time and tokens, use it when it is helpful for managing complex many-step problems! But not for simple few-step requests.\\n\\n## Important To-Do List Usage Notes to Remember\\n- The `write_todos` tool should never be called multiple times in parallel.\\n- Don't be afraid to revise the To-Do list as you go. New information may reveal new tasks that need to be done, or old tasks that are irrelevant.\";\ndeclare const stateSchema: z.ZodObject<{\n    todos: z.ZodDefault<z.ZodArray<z.ZodObject<{\n        content: z.ZodString;\n        status: z.ZodEnum<[\"pending\", \"in_progress\", \"completed\"]>;\n    }, \"strip\", z.ZodTypeAny, {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }, {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }>, \"many\">>;\n}, \"strip\", z.ZodTypeAny, {\n    todos: {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }[];\n}, {\n    todos?: {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }[] | undefined;\n}>;\nexport type TodoMiddlewareState = z.infer<typeof stateSchema>;\nexport interface TodoListMiddlewareOptions {\n    /**\n     * Custom system prompt to guide the agent on using the todo tool.\n     * If not provided, uses the default {@link PLANNING_MIDDLEWARE_SYSTEM_PROMPT}.\n     */\n    systemPrompt?: string;\n    /**\n     * Custom description for the {@link writeTodos} tool.\n     * If not provided, uses the default {@link WRITE_TODOS_DESCRIPTION}.\n     */\n    toolDescription?: string;\n}\n/**\n * Creates a middleware that provides todo list management capabilities to agents.\n *\n * This middleware adds a `write_todos` tool that allows agents to create and manage\n * structured task lists for complex multi-step operations. It's designed to help\n * agents track progress, organize complex tasks, and provide users with visibility\n * into task completion status.\n *\n * The middleware automatically injects system prompts that guide the agent on when\n * and how to use the todo functionality effectively.\n *\n * @example\n * ```typescript\n * import { todoListMiddleware, createAgent } from 'langchain';\n *\n * const agent = createAgent({\n *   model: \"openai:gpt-4o\",\n *   middleware: [todoListMiddleware()],\n * });\n *\n * // Agent now has access to write_todos tool and todo state tracking\n * const result = await agent.invoke({\n *   messages: [new HumanMessage(\"Help me refactor my codebase\")]\n * });\n *\n * console.log(result.todos); // Array of todo items with status tracking\n * ```\n *\n * @returns A configured middleware instance that provides todo management capabilities\n *\n * @see {@link TodoMiddlewareState} for the state schema\n * @see {@link writeTodos} for the tool implementation\n */\nexport declare function todoListMiddleware(options?: TodoListMiddlewareOptions): import(\"./types.js\").AgentMiddleware<z.ZodObject<{\n    todos: z.ZodDefault<z.ZodArray<z.ZodObject<{\n        content: z.ZodString;\n        status: z.ZodEnum<[\"pending\", \"in_progress\", \"completed\"]>;\n    }, \"strip\", z.ZodTypeAny, {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }, {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }>, \"many\">>;\n}, \"strip\", z.ZodTypeAny, {\n    todos: {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }[];\n}, {\n    todos?: {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }[] | undefined;\n}>, undefined, any>;\nexport {};\n"],"mappings":";;;;cACqBC,kCAAAA;UAwBJU,yBAAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6COC,kBAAAA,WAA6BD,4BAA4E,gBAAXX,CAAAA,CAAEM;SAC7GN,CAAAA,CAAEQ,WAAWR,CAAAA,CAAEO,SAASP,CAAAA,CAAEM;aACpBN,CAAAA,CAAEG;YACHH,CAAAA,CAAEI;cACFJ,CAAAA,CAAEK;;;;;;;YAONL,CAAAA,CAAEK"}
+{"version":3,"file":"todoListMiddleware.d.ts","names":["z","TODO_LIST_MIDDLEWARE_SYSTEM_PROMPT","stateSchema","ZodString","ZodEnum","ZodTypeAny","ZodObject","ZodArray","ZodDefault","TodoMiddlewareState","infer","TodoListMiddlewareOptions","todoListMiddleware","__types_js3","AgentMiddleware"],"sources":["../../../src/agents/middleware/todoListMiddleware.d.ts"],"sourcesContent":["import { z } from \"zod/v3\";\nexport declare const TODO_LIST_MIDDLEWARE_SYSTEM_PROMPT = \"## `write_todos`\\n\\nYou have access to the `write_todos` tool to help you manage and plan complex objectives. \\nUse this tool for complex objectives to ensure that you are tracking each necessary step and giving the user visibility into your progress.\\nThis tool is very helpful for planning complex objectives, and for breaking down these larger complex objectives into smaller steps.\\n\\nIt is critical that you mark todos as completed as soon as you are done with a step. Do not batch up multiple steps before marking them as completed.\\nFor simple objectives that only require a few steps, it is better to just complete the objective directly and NOT use this tool.\\nWriting todos takes time and tokens, use it when it is helpful for managing complex many-step problems! But not for simple few-step requests.\\n\\n## Important To-Do List Usage Notes to Remember\\n- The `write_todos` tool should never be called multiple times in parallel.\\n- Don't be afraid to revise the To-Do list as you go. New information may reveal new tasks that need to be done, or old tasks that are irrelevant.\";\ndeclare const stateSchema: z.ZodObject<{\n    todos: z.ZodDefault<z.ZodArray<z.ZodObject<{\n        content: z.ZodString;\n        status: z.ZodEnum<[\"pending\", \"in_progress\", \"completed\"]>;\n    }, \"strip\", z.ZodTypeAny, {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }, {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }>, \"many\">>;\n}, \"strip\", z.ZodTypeAny, {\n    todos: {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }[];\n}, {\n    todos?: {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }[] | undefined;\n}>;\nexport type TodoMiddlewareState = z.infer<typeof stateSchema>;\nexport interface TodoListMiddlewareOptions {\n    /**\n     * Custom system prompt to guide the agent on using the todo tool.\n     * If not provided, uses the default {@link PLANNING_MIDDLEWARE_SYSTEM_PROMPT}.\n     */\n    systemPrompt?: string;\n    /**\n     * Custom description for the {@link writeTodos} tool.\n     * If not provided, uses the default {@link WRITE_TODOS_DESCRIPTION}.\n     */\n    toolDescription?: string;\n}\n/**\n * Creates a middleware that provides todo list management capabilities to agents.\n *\n * This middleware adds a `write_todos` tool that allows agents to create and manage\n * structured task lists for complex multi-step operations. It's designed to help\n * agents track progress, organize complex tasks, and provide users with visibility\n * into task completion status.\n *\n * The middleware automatically injects system prompts that guide the agent on when\n * and how to use the todo functionality effectively.\n *\n * @example\n * ```typescript\n * import { todoListMiddleware, createAgent } from 'langchain';\n *\n * const agent = createAgent({\n *   model: \"openai:gpt-4o\",\n *   middleware: [todoListMiddleware()],\n * });\n *\n * // Agent now has access to write_todos tool and todo state tracking\n * const result = await agent.invoke({\n *   messages: [new HumanMessage(\"Help me refactor my codebase\")]\n * });\n *\n * console.log(result.todos); // Array of todo items with status tracking\n * ```\n *\n * @returns A configured middleware instance that provides todo management capabilities\n *\n * @see {@link TodoMiddlewareState} for the state schema\n * @see {@link writeTodos} for the tool implementation\n */\nexport declare function todoListMiddleware(options?: TodoListMiddlewareOptions): import(\"./types.js\").AgentMiddleware<z.ZodObject<{\n    todos: z.ZodDefault<z.ZodArray<z.ZodObject<{\n        content: z.ZodString;\n        status: z.ZodEnum<[\"pending\", \"in_progress\", \"completed\"]>;\n    }, \"strip\", z.ZodTypeAny, {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }, {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }>, \"many\">>;\n}, \"strip\", z.ZodTypeAny, {\n    todos: {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }[];\n}, {\n    todos?: {\n        content: string;\n        status: \"completed\" | \"in_progress\" | \"pending\";\n    }[] | undefined;\n}>, undefined, any>;\nexport {};\n"],"mappings":";;;;cACqBC,kCAAAA;UAwBJU,yBAAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6COC,kBAAAA,WAA6BD,4BAA4E,gBAAXX,CAAAA,CAAEM;SAC7GN,CAAAA,CAAEQ,WAAWR,CAAAA,CAAEO,SAASP,CAAAA,CAAEM;aACpBN,CAAAA,CAAEG;YACHH,CAAAA,CAAEI;cACFJ,CAAAA,CAAEK;;;;;;;YAONL,CAAAA,CAAEK"}

package/dist/agents/middleware/toolCallLimit.cjs

@@ -5,59 +5,50 @@ const zod_v3 = require_rolldown_runtime.__toESM(require("zod/v3"));
 
 //#region src/agents/middleware/toolCallLimit.ts
 /**
-* Count tool calls in a list of messages.
+* Build the error message content for ToolMessage when limit is exceeded.
 *
-* @param messages - List of messages to count tool calls in.
-* @param toolName - If specified, only count calls to this specific tool.
-*   If undefined, count all tool calls.
-* @returns The total number of tool calls (optionally filtered by toolName).
-*/
-function countToolCallsInMessages(messages, toolName) {
-	let count = 0;
-	for (const message of messages) if (__langchain_core_messages.AIMessage.isInstance(message) && message.tool_calls) if (toolName === void 0) count += message.tool_calls.length;
-	else count += message.tool_calls.filter((tc) => tc.name === toolName).length;
-	return count;
-}
-/**
-* Get messages from the current run (after the last HumanMessage).
+* This message is sent to the model, so it should not reference thread/run concepts
+* that the model has no notion of.
 *
-* @param messages - Full list of messages.
-* @returns Messages from the current run (after last HumanMessage).
+* @param toolName - Tool name being limited (if specific tool), or undefined for all tools.
+* @param threadCount - Current thread tool call count.
+* @param runCount - Current run tool call count.
+* @param threadLimit - Thread tool call limit (if set).
+* @param runLimit - Run tool call limit (if set).
+* @returns A concise message. If only run limit is exceeded (not thread limit),
+* returns a simple "limit exceeded" message without instructing model to stop.
+* If thread limit is exceeded, includes instruction not to call again.
 */
-function getRunMessages(messages) {
-	/**
-	* Find the last HumanMessage
-	*/
-	let lastHumanIndex = -1;
-	for (let i = messages.length - 1; i >= 0; i--) if (__langchain_core_messages.HumanMessage.isInstance(messages[i])) {
-		lastHumanIndex = i;
-		break;
+function buildToolMessageContent(toolName, threadCount, runCount, threadLimit, runLimit) {
+	const threadExceeded = threadLimit !== void 0 && threadCount > threadLimit;
+	const onlyRunExceeded = runLimit !== void 0 && runCount > runLimit && !threadExceeded;
+	if (onlyRunExceeded) {
+		if (toolName) return `Tool call limit exceeded for '${toolName}'.`;
+		return "Tool call limit exceeded.";
 	}
-	/**
-	* If no HumanMessage found, return all messages
-	*/
-	if (lastHumanIndex === -1) return messages;
-	/**
-	* Return messages after the last HumanMessage
-	*/
-	return messages.slice(lastHumanIndex + 1);
+	if (toolName) return `Tool call limit exceeded. Do not call '${toolName}' again.`;
+	return "Tool call limit exceeded. Do not make additional tool calls.";
 }
 /**
-* Build a message indicating which tool call limits were reached.
+* Build the final AI message content for 'end' behavior.
+*
+* This message is displayed to the user, so it should include detailed information
+* about which limits were exceeded.
 *
 * @param threadCount - Current thread tool call count.
 * @param runCount - Current run tool call count.
 * @param threadLimit - Thread tool call limit (if set).
 * @param runLimit - Run tool call limit (if set).
 * @param toolName - Tool name being limited (if specific tool), or undefined for all tools.
-* @returns A formatted message describing which limits were reached.
+* @returns A formatted message describing which limits were exceeded.
 */
-function buildToolLimitExceededMessage(threadCount, runCount, threadLimit, runLimit, toolName) {
-	const toolDesc = toolName ? `'${toolName}' tool call` : "Tool call";
+function buildFinalAIMessageContent(threadCount, runCount, threadLimit, runLimit, toolName) {
+	const toolDesc = toolName ? `'${toolName}' tool` : "Tool";
 	const exceededLimits = [];
-	if (threadLimit !== void 0 && threadCount >= threadLimit) exceededLimits.push(`thread limit reached (${threadCount}/${threadLimit})`);
-	if (runLimit !== void 0 && runCount >= runLimit) exceededLimits.push(`run limit reached (${runCount}/${runLimit})`);
-	return `${toolDesc} limit${exceededLimits.length > 1 ? "s" : ""}: ${exceededLimits.join(", ")}. Stopping to prevent further tool calls.`;
+	if (threadLimit !== void 0 && threadCount > threadLimit) exceededLimits.push(`thread limit exceeded (${threadCount}/${threadLimit} calls)`);
+	if (runLimit !== void 0 && runCount > runLimit) exceededLimits.push(`run limit exceeded (${runCount}/${runLimit} calls)`);
+	const limitsText = exceededLimits.join(" and ");
+	return `${toolDesc} call limit reached: ${limitsText}.`;
 }
 /**
 * Exception raised when tool call limits are exceeded.
@@ -87,7 +78,7 @@ var ToolCallLimitExceededError = class extends Error {
 	*/
 	toolName;
 	constructor(threadCount, runCount, threadLimit, runLimit, toolName = void 0) {
-		const message = buildToolLimitExceededMessage(threadCount, runCount, threadLimit, runLimit, toolName);
+		const message = buildFinalAIMessageContent(threadCount, runCount, threadLimit, runLimit, toolName);
 		super(message);
 		this.name = "ToolCallLimitExceededError";
 		this.threadCount = threadCount;
@@ -104,8 +95,20 @@ const ToolCallLimitOptionsSchema = zod_v3.z.object({
 	toolName: zod_v3.z.string().optional(),
 	threadLimit: zod_v3.z.number().optional(),
 	runLimit: zod_v3.z.number().optional(),
-	exitBehavior: zod_v3.z.enum(["end", "error"]).default("end")
+	exitBehavior: zod_v3.z.enum([
+		"continue",
+		"error",
+		"end"
+	]).default("continue")
+});
+/**
+* Middleware state schema to track the number of model calls made at the thread and run level.
+*/
+const stateSchema = zod_v3.z.object({
+	threadToolCallCount: zod_v3.z.record(zod_v3.z.string(), zod_v3.z.number()).default({}),
+	runToolCallCount: zod_v3.z.record(zod_v3.z.string(), zod_v3.z.number()).default({})
 });
+const DEFAULT_TOOL_COUNT_KEY = "__all__";
 /**
 * Middleware that tracks tool call counts and enforces limits.
 *
@@ -124,68 +127,66 @@ const ToolCallLimitOptionsSchema = zod_v3.z.object({
 * @param options.threadLimit - Maximum number of tool calls allowed per thread. undefined means no limit.
 * @param options.runLimit - Maximum number of tool calls allowed per run. undefined means no limit.
 * @param options.exitBehavior - What to do when limits are exceeded.
-*   - "end": Jump to the end of the agent execution and inject an artificial AI message indicating that the limit was exceeded.
-*   - "error": throws a ToolCallLimitExceededError
+*   - "continue": Block exceeded tools with error messages, let other tools continue. Model decides when to end. (default)
+*   - "error": Raise a ToolCallLimitExceededError exception
+*   - "end": Stop execution immediately with a ToolMessage + AI message for the single tool call that exceeded the limit. Raises NotImplementedError if there are multiple tool calls.
 *
-* @throws {Error} If both limits are undefined or if exitBehavior is invalid.
+* @throws {Error} If both limits are undefined.
+* @throws {NotImplementedError} If exitBehavior is "end" and there are multiple tool calls.
 *
-* @example Limit all tool calls globally
+* @example Continue execution with blocked tools (default)
 * ```ts
 * import { toolCallLimitMiddleware } from "@langchain/langchain/agents/middleware";
 * import { createAgent } from "@langchain/langchain/agents";
 *
-* const globalLimiter = toolCallLimitMiddleware({
+* // Block exceeded tools but let other tools and model continue
+* const limiter = toolCallLimitMiddleware({
 *   threadLimit: 20,
 *   runLimit: 10,
-*   exitBehavior: "end"
+*   exitBehavior: "continue", // default
 * });
 *
 * const agent = createAgent({
 *   model: "openai:gpt-4o",
-*   middleware: [globalLimiter]
+*   middleware: [limiter]
 * });
 * ```
 *
-* @example Limit a specific tool
+* @example Stop immediately when limit exceeded
 * ```ts
-* import { toolCallLimitMiddleware } from "@langchain/langchain/agents/middleware";
-* import { createAgent } from "@langchain/langchain/agents";
-*
-* const searchLimiter = toolCallLimitMiddleware({
-*   toolName: "search",
-*   threadLimit: 5,
-*   runLimit: 3,
+* // End execution immediately with an AI message
+* const limiter = toolCallLimitMiddleware({
+*   runLimit: 5,
 *   exitBehavior: "end"
 * });
 *
 * const agent = createAgent({
 *   model: "openai:gpt-4o",
-*   middleware: [searchLimiter]
+*   middleware: [limiter]
 * });
 * ```
 *
-* @example Use both in the same agent
+* @example Raise exception on limit
 * ```ts
-* import { toolCallLimitMiddleware } from "@langchain/langchain/agents/middleware";
-* import { createAgent } from "@langchain/langchain/agents";
-*
-* const globalLimiter = toolCallLimitMiddleware({
-*   threadLimit: 20,
-*   runLimit: 10,
-*   exitBehavior: "end"
-* });
-*
-* const searchLimiter = toolCallLimitMiddleware({
+* // Strict limit with exception handling
+* const limiter = toolCallLimitMiddleware({
 *   toolName: "search",
 *   threadLimit: 5,
-*   runLimit: 3,
-*   exitBehavior: "end"
+*   exitBehavior: "error"
 * });
 *
 * const agent = createAgent({
 *   model: "openai:gpt-4o",
-*   middleware: [globalLimiter, searchLimiter]
+*   middleware: [limiter]
 * });
+*
+* try {
+*   const result = await agent.invoke({ messages: [new HumanMessage("Task")] });
+* } catch (error) {
+*   if (error instanceof ToolCallLimitExceededError) {
+*     console.log(`Search limit exceeded: ${error}`);
+*   }
+* }
 * ```
 */
 function toolCallLimitMiddleware(options) {
@@ -194,47 +195,159 @@ function toolCallLimitMiddleware(options) {
 	*/
 	if (options.threadLimit === void 0 && options.runLimit === void 0) throw new Error("At least one limit must be specified (threadLimit or runLimit)");
 	/**
-	* Apply default for exitBehavior and validate
+	* Apply default for exitBehavior
 	*/
-	const exitBehavior = options.exitBehavior ?? "end";
-	if (exitBehavior !== "end" && exitBehavior !== "error") throw new Error(`Invalid exit behavior: ${exitBehavior}. Must be 'end' or 'error'`);
+	const exitBehavior = options.exitBehavior ?? "continue";
 	/**
 	* Generate the middleware name based on the tool name
 	*/
 	const middlewareName = options.toolName ? `ToolCallLimitMiddleware[${options.toolName}]` : "ToolCallLimitMiddleware";
 	return require_middleware.createMiddleware({
 		name: middlewareName,
-		beforeModel: {
+		stateSchema,
+		afterModel: {
 			canJumpTo: ["end"],
 			hook: (state) => {
-				const messages = state.messages;
 				/**
-				* Count tool calls in entire thread
+				* Get the last AI message to check for tool calls
+				*/
+				const lastAIMessage = [...state.messages].reverse().find(__langchain_core_messages.AIMessage.isInstance);
+				if (!lastAIMessage || !lastAIMessage.tool_calls) return void 0;
+				/**
+				* Helper to check if limit would be exceeded by one more call
+				*/
+				const wouldExceedLimit = (threadCount, runCount) => {
+					return options.threadLimit !== void 0 && threadCount + 1 > options.threadLimit || options.runLimit !== void 0 && runCount + 1 > options.runLimit;
+				};
+				/**
+				* Helper to check if a tool call matches our filter
+				*/
+				const matchesToolFilter = (toolCall) => {
+					return options.toolName === void 0 || toolCall.name === options.toolName;
+				};
+				/**
+				* Separate tool calls into allowed and blocked based on limits
+				*/
+				const separateToolCalls = (toolCalls, threadCount, runCount) => {
+					const allowed$1 = [];
+					const blocked$1 = [];
+					let tempThreadCount = threadCount;
+					let tempRunCount = runCount;
+					for (const toolCall of toolCalls) {
+						if (!matchesToolFilter(toolCall)) continue;
+						if (wouldExceedLimit(tempThreadCount, tempRunCount)) blocked$1.push(toolCall);
+						else {
+							allowed$1.push(toolCall);
+							tempThreadCount += 1;
+							tempRunCount += 1;
+						}
+					}
+					return {
+						allowed: allowed$1,
+						blocked: blocked$1,
+						finalThreadCount: tempThreadCount + blocked$1.length,
+						finalRunCount: tempRunCount + blocked$1.length
+					};
+				};
+				/**
+				* Get the count key for this middleware instance
+				*/
+				const countKey = options.toolName ?? DEFAULT_TOOL_COUNT_KEY;
+				/**
+				* Get current counts
+				*/
+				const threadCounts = { ...state.threadToolCallCount ?? {} };
+				const runCounts = { ...state.runToolCallCount ?? {} };
+				const currentThreadCount = threadCounts[countKey] ?? 0;
+				const currentRunCount = runCounts[countKey] ?? 0;
+				/**
+				* Separate tool calls into allowed and blocked
+				*/
+				const { allowed, blocked, finalThreadCount, finalRunCount } = separateToolCalls(lastAIMessage.tool_calls, currentThreadCount, currentRunCount);
+				/**
+				* Update counts to include ALL tool call attempts (both allowed and blocked)
+				*/
+				threadCounts[countKey] = finalThreadCount;
+				runCounts[countKey] = finalRunCount;
+				/**
+				* If no tool calls are blocked, just update counts
+				*/
+				if (blocked.length === 0) {
+					if (allowed.length > 0) return {
+						threadToolCallCount: threadCounts,
+						runToolCallCount: runCounts
+					};
+					return void 0;
+				}
+				/**
+				* Handle different exit behaviors
 				*/
-				const threadCount = countToolCallsInMessages(messages, options.toolName);
+				if (exitBehavior === "error") throw new ToolCallLimitExceededError(finalThreadCount, finalRunCount, options.threadLimit, options.runLimit, options.toolName);
 				/**
-				* Count tool calls in current run (after last HumanMessage)
+				* Build tool message content (sent to model - no thread/run details)
 				*/
-				const runMessages = getRunMessages(messages);
-				const runCount = countToolCallsInMessages(runMessages, options.toolName);
+				const toolMsgContent = buildToolMessageContent(options.toolName, finalThreadCount, finalRunCount, options.threadLimit, options.runLimit);
 				/**
-				* Check if any limits are exceeded
+				* Inject artificial error ToolMessages for blocked tool calls
 				*/
-				const threadLimitExceeded = options.threadLimit !== void 0 && threadCount >= options.threadLimit;
-				const runLimitExceeded = options.runLimit !== void 0 && runCount >= options.runLimit;
-				if (!threadLimitExceeded && !runLimitExceeded) return void 0;
-				if (exitBehavior === "error") throw new ToolCallLimitExceededError(threadCount, runCount, options.threadLimit, options.runLimit, options.toolName);
+				const artificialMessages = blocked.map((toolCall) => new __langchain_core_messages.ToolMessage({
+					content: toolMsgContent,
+					tool_call_id: toolCall.id,
+					name: toolCall.name,
+					status: "error"
+				}));
+				if (exitBehavior === "end") {
+					/**
+					* Check if there are tool calls to other tools that would continue executing
+					* For tool-specific limiters: check for calls to other tools
+					* For global limiters: check if there are multiple different tool types
+					*/
+					let otherTools = [];
+					if (options.toolName !== void 0)
+ /**
+					* Tool-specific limiter: check for calls to other tools
+					*/
+					otherTools = lastAIMessage.tool_calls.filter((tc) => tc.name !== options.toolName);
+					else {
+						/**
+						* Global limiter: check if there are multiple different tool types
+						* If there are allowed calls, those would execute
+						* But even if all are blocked, we can't handle multiple tool types with "end"
+						*/
+						const uniqueToolNames = new Set(lastAIMessage.tool_calls.map((tc) => tc.name).filter(Boolean));
+						if (uniqueToolNames.size > 1)
+ /**
+						* Multiple different tool types - use allowed calls to show which ones
+						*/
+						otherTools = allowed.length > 0 ? allowed : lastAIMessage.tool_calls;
+					}
+					if (otherTools.length > 0) {
+						const toolNames = Array.from(new Set(otherTools.map((tc) => tc.name).filter(Boolean))).join(", ");
+						throw new Error(`Cannot end execution with other tool calls pending. Found calls to: ${toolNames}. Use 'continue' or 'error' behavior instead.`);
+					}
+					/**
+					* Build final AI message content (displayed to user - includes thread/run details)
+					*/
+					const finalMsgContent = buildFinalAIMessageContent(finalThreadCount, finalRunCount, options.threadLimit, options.runLimit, options.toolName);
+					artificialMessages.push(new __langchain_core_messages.AIMessage(finalMsgContent));
+					return {
+						threadToolCallCount: threadCounts,
+						runToolCallCount: runCounts,
+						jumpTo: "end",
+						messages: artificialMessages
+					};
+				}
 				/**
-				* Create a message indicating the limit was exceeded
+				* For exit_behavior="continue", return error messages to block exceeded tools
 				*/
-				const limitMessage = buildToolLimitExceededMessage(threadCount, runCount, options.threadLimit, options.runLimit, options.toolName);
-				const limitAiMessage = new __langchain_core_messages.AIMessage(limitMessage);
 				return {
-					jumpTo: "end",
-					messages: [limitAiMessage]
+					threadToolCallCount: threadCounts,
+					runToolCallCount: runCounts,
+					messages: artificialMessages
 				};
 			}
-		}
+		},
+		afterAgent: () => ({ runToolCallCount: {} })
 	});
 }