@librechat/agents 2.3.8 → 2.3.91

package/dist/cjs/messages/prune.cjs

@@ -14,8 +14,15 @@ function addThinkingBlock(message, thinkingBlock) {
                 type: _enum.ContentTypes.TEXT,
                 text: message.content,
             }];
+    /** Edge case, the message already has the thinking block */
+    if (content[0].type === thinkingBlock.type) {
+        return message;
+    }
     content.unshift(thinkingBlock);
-    return content;
+    return new messages.AIMessage({
+        ...message,
+        content
+    });
 }
 /**
  * Calculates the total tokens from a single usage object
@@ -42,7 +49,7 @@ function calculateTotalTokens(usage) {
  * @param options Configuration options for processing messages
  * @returns Object containing the message context, remaining tokens, messages not included, and summary index
  */
-function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, indexTokenCountMap, startType: _startType, thinkingEnabled, tokenCounter, reasoningType = _enum.ContentTypes.THINKING, }) {
+function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, indexTokenCountMap, startType: _startType, thinkingEnabled, tokenCounter, thinkingStartIndex: _thinkingStartIndex = -1, reasoningType = _enum.ContentTypes.THINKING, }) {
     // Every reply is primed with <|start|>assistant<|message|>, so we
     // start with 3 tokens for the label after all messages have been counted.
     let currentTokenCount = 3;
@@ -59,11 +66,17 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
      * This may be confusing to read, but it is done to ensure the context is in the correct order for the model.
      * */
     let context = [];
-    let thinkingStartIndex = -1;
+    let thinkingStartIndex = _thinkingStartIndex;
     let thinkingEndIndex = -1;
     let thinkingBlock;
     const endIndex = instructions != null ? 1 : 0;
     const prunedMemory = [];
+    if (_thinkingStartIndex > -1) {
+        const thinkingMessageContent = messages$1[_thinkingStartIndex]?.content;
+        if (Array.isArray(thinkingMessageContent)) {
+            thinkingBlock = thinkingMessageContent.find((content) => content.type === reasoningType);
+        }
+    }
     if (currentTokenCount < remainingContextTokens) {
         let currentIndex = messages$1.length;
         while (messages$1.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > endIndex) {
@@ -95,19 +108,30 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
             }
             else {
                 prunedMemory.push(poppedMessage);
-                if (thinkingEndIndex > -1) {
+                if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {
                     continue;
                 }
                 break;
             }
         }
-        if (thinkingEndIndex > -1 && context[context.length - 1]?.getType() === 'tool') {
-            startType = 'ai';
+        if (context[context.length - 1]?.getType() === 'tool') {
+            startType = ['ai', 'human'];
         }
-        if (startType != null && startType && context.length > 0) {
-            const requiredTypeIndex = context.findIndex(msg => msg?.getType() === startType);
+        if (startType != null && startType.length > 0 && context.length > 0) {
+            let requiredTypeIndex = -1;
+            let totalTokens = 0;
+            for (let i = context.length - 1; i >= 0; i--) {
+                const currentType = context[i]?.getType() ?? '';
+                if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {
+                    requiredTypeIndex = i + 1;
+                    break;
+                }
+                const originalIndex = originalLength - 1 - i;
+                totalTokens += indexTokenCountMap[originalIndex] ?? 0;
+            }
             if (requiredTypeIndex > 0) {
-                context = context.slice(requiredTypeIndex);
+                currentTokenCount -= totalTokens;
+                context = context.slice(0, requiredTypeIndex);
             }
         }
     }
@@ -121,6 +145,9 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
         context: [],
         messagesToRefine: prunedMemory,
     };
+    if (thinkingStartIndex > -1) {
+        result.thinkingStartIndex = thinkingStartIndex;
+    }
     if (prunedMemory.length === 0 || thinkingEndIndex < 0 || (thinkingStartIndex > -1 && isIndexInContext(_messages, context, thinkingStartIndex))) {
         // we reverse at this step to ensure the context is in the correct order for the model, and we need to work backwards
         result.context = context.reverse();
@@ -152,8 +179,8 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
     thinkingStartIndex = originalLength - 1 - assistantIndex;
     const thinkingTokenCount = tokenCounter(new messages.AIMessage({ content: [thinkingBlock] }));
     const newRemainingCount = remainingContextTokens - thinkingTokenCount;
-    const content = addThinkingBlock(context[assistantIndex], thinkingBlock);
-    context[assistantIndex].content = content;
+    const newMessage = addThinkingBlock(context[assistantIndex], thinkingBlock);
+    context[assistantIndex] = newMessage;
     if (newRemainingCount > 0) {
         result.context = context.reverse();
         return result;
@@ -184,17 +211,28 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
     const firstMessage = newContext[newContext.length - 1];
     const firstMessageType = newContext[newContext.length - 1].getType();
     if (firstMessageType === 'tool') {
-        startType = 'ai';
+        startType = ['ai', 'human'];
     }
-    if (startType != null && startType && newContext.length > 0) {
-        const requiredTypeIndex = newContext.findIndex(msg => msg.getType() === startType);
+    if (startType != null && startType.length > 0 && newContext.length > 0) {
+        let requiredTypeIndex = -1;
+        let totalTokens = 0;
+        for (let i = newContext.length - 1; i >= 0; i--) {
+            const currentType = newContext[i]?.getType() ?? '';
+            if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {
+                requiredTypeIndex = i + 1;
+                break;
+            }
+            const originalIndex = originalLength - 1 - i;
+            totalTokens += indexTokenCountMap[originalIndex] ?? 0;
+        }
         if (requiredTypeIndex > 0) {
-            newContext = newContext.slice(requiredTypeIndex);
+            currentTokenCount -= totalTokens;
+            newContext = newContext.slice(0, requiredTypeIndex);
         }
     }
     if (firstMessageType === 'ai') {
-        const content = addThinkingBlock(firstMessage, thinkingBlock);
-        newContext[newContext.length - 1].content = content;
+        const newMessage = addThinkingBlock(firstMessage, thinkingBlock);
+        newContext[newContext.length - 1] = newMessage;
     }
     else {
         newContext.push(thinkingMessage);
@@ -214,6 +252,7 @@ function createPruneMessages(factoryParams) {
     let lastTurnStartIndex = factoryParams.startIndex;
     let lastCutOffIndex = 0;
     let totalTokens = (Object.values(indexTokenCountMap)).reduce((a, b) => a + b, 0);
+    let runThinkingStartIndex = -1;
     return function pruneMessages(params) {
         let currentUsage;
         if (params.usageMetadata && (checkValidNumber(params.usageMetadata.input_tokens)
@@ -270,7 +309,7 @@ function createPruneMessages(factoryParams) {
         if (totalTokens <= factoryParams.maxTokens) {
             return { context: params.messages, indexTokenCountMap };
         }
-        const { context } = getMessagesWithinTokenLimit({
+        const { context, thinkingStartIndex } = getMessagesWithinTokenLimit({
             maxContextTokens: factoryParams.maxTokens,
             messages: params.messages,
             indexTokenCountMap,
@@ -278,8 +317,11 @@ function createPruneMessages(factoryParams) {
             thinkingEnabled: factoryParams.thinkingEnabled,
             tokenCounter: factoryParams.tokenCounter,
             reasoningType: factoryParams.provider === _enum.Providers.BEDROCK ? _enum.ContentTypes.REASONING_CONTENT : _enum.ContentTypes.THINKING,
+            thinkingStartIndex: factoryParams.thinkingEnabled === true ? runThinkingStartIndex : undefined,
         });
-        lastCutOffIndex = Math.max(params.messages.length - context.length, 0);
+        runThinkingStartIndex = thinkingStartIndex ?? -1;
+        /** The index is the first value of `context`, index relative to `params.messages` */
+        lastCutOffIndex = Math.max(params.messages.length - (context.length - (context[0]?.getType() === 'system' ? 1 : 0)), 0);
         return { context, indexTokenCountMap };
     };
 }

package/dist/cjs/messages/prune.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"prune.cjs","sources":["../../../src/messages/prune.ts"],"sourcesContent":["import { AIMessage, BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { ThinkingContentText, MessageContentComplex, ReasoningContentText } from '@/types/stream';\nimport type { TokenCounter } from '@/types/run';\nimport { ContentTypes, Providers } from '@/common';\n\nexport type PruneMessagesFactoryParams = {\n  provider?: Providers;\n  maxTokens: number;\n  startIndex: number;\n  tokenCounter: TokenCounter;\n  indexTokenCountMap: Record<string, number>;\n  thinkingEnabled?: boolean;\n};\nexport type PruneMessagesParams = {\n  messages: BaseMessage[];\n  usageMetadata?: Partial<UsageMetadata>;\n  startType?: ReturnType<BaseMessage['getType']>;\n}\n\nfunction isIndexInContext(arrayA: unknown[], arrayB: unknown[], targetIndex: number): boolean {\n  const startingIndexInA = arrayA.length - arrayB.length;\n  return targetIndex >= startingIndexInA;\n}\n\nfunction addThinkingBlock(message: AIMessage, thinkingBlock: ThinkingContentText | ReasoningContentText): MessageContentComplex[] {\n  const content: MessageContentComplex[] = Array.isArray(message.content)\n    ? message.content as MessageContentComplex[]\n    : [{\n      type: ContentTypes.TEXT,\n      text: message.content,\n    }];\n  content.unshift(thinkingBlock);\n  return content;\n}\n\n/**\n * Calculates the total tokens from a single usage object\n *\n * @param usage The usage metadata object containing token information\n * @returns An object containing the total input and output tokens\n */\nexport function calculateTotalTokens(usage: Partial<UsageMetadata>): UsageMetadata {\n  const baseInputTokens = Number(usage.input_tokens) || 0;\n  const cacheCreation = Number(usage.input_token_details?.cache_creation) || 0;\n  const cacheRead = Number(usage.input_token_details?.cache_read) || 0;\n\n  const totalInputTokens = baseInputTokens + cacheCreation + cacheRead;\n  const totalOutputTokens = Number(usage.output_tokens) || 0;\n\n  return {\n    input_tokens: totalInputTokens,\n    output_tokens: totalOutputTokens,\n    total_tokens: totalInputTokens + totalOutputTokens\n  };\n}\n\n/**\n * Processes an array of messages and returns a context of messages that fit within a specified token limit.\n * It iterates over the messages from newest to oldest, adding them to the context until the token limit is reached.\n *\n * @param options Configuration options for processing messages\n * @returns Object containing the message context, remaining tokens, messages not included, and summary index\n */\nexport function getMessagesWithinTokenLimit({\n  messages: _messages,\n  maxContextTokens,\n  indexTokenCountMap,\n  startType: _startType,\n  thinkingEnabled,\n  tokenCounter,\n  reasoningType = ContentTypes.THINKING,\n}: {\n  messages: BaseMessage[];\n  maxContextTokens: number;\n  indexTokenCountMap: Record<string, number | undefined>;\n  tokenCounter: TokenCounter;\n  startType?: string;\n  thinkingEnabled?: boolean;\n  reasoningType?: ContentTypes.THINKING | ContentTypes.REASONING_CONTENT;\n}): {\n  context: BaseMessage[];\n  remainingContextTokens: number;\n  messagesToRefine: BaseMessage[];\n} {\n  // Every reply is primed with <|start|>assistant<|message|>, so we\n  // start with 3 tokens for the label after all messages have been counted.\n  let currentTokenCount = 3;\n  const instructions = _messages[0]?.getType() === 'system' ? _messages[0] : undefined;\n  const instructionsTokenCount = instructions != null ? indexTokenCountMap[0] ?? 0 : 0;\n  const initialContextTokens = maxContextTokens - instructionsTokenCount;\n  let remainingContextTokens = initialContextTokens;\n  let startType = _startType;\n  const originalLength = _messages.length;\n  const messages = [..._messages];\n  /**\n   * IMPORTANT: this context array gets reversed at the end, since the latest messages get pushed first.\n   *\n   * This may be confusing to read, but it is done to ensure the context is in the correct order for the model.\n   * */\n  let context: Array<BaseMessage | undefined> = [];\n\n  let thinkingStartIndex = -1;\n  let thinkingEndIndex = -1;\n  let thinkingBlock: ThinkingContentText | ReasoningContentText | undefined;\n  const endIndex = instructions != null ? 1 : 0;\n  const prunedMemory: BaseMessage[] = [];\n\n  if (currentTokenCount < remainingContextTokens) {\n    let currentIndex = messages.length;\n    while (messages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > endIndex) {\n      currentIndex--;\n      if (messages.length === 1 && instructions) {\n        break;\n      }\n      const poppedMessage = messages.pop();\n      if (!poppedMessage) continue;\n      const messageType = poppedMessage.getType();\n      if (thinkingEnabled === true && thinkingEndIndex === -1 && (currentIndex === (originalLength - 1)) && (messageType === 'ai' || messageType === 'tool')) {\n        thinkingEndIndex = currentIndex;\n      }\n      if (thinkingEndIndex > -1 && !thinkingBlock  && thinkingStartIndex < 0 && messageType === 'ai' && Array.isArray(poppedMessage.content)) {\n        thinkingBlock = (poppedMessage.content.find((content) => content.type === reasoningType)) as ThinkingContentText | undefined;\n        thinkingStartIndex = thinkingBlock != null ? currentIndex : -1;\n      }\n      /** False start, the latest message was not part of a multi-assistant/tool sequence of messages */\n      if (\n        thinkingEndIndex > -1\n        && currentIndex === (thinkingEndIndex - 1)\n        && (messageType !== 'ai' && messageType !== 'tool')\n      ) {\n        thinkingEndIndex = -1;\n      }\n\n      const tokenCount = indexTokenCountMap[currentIndex] ?? 0;\n\n      if (prunedMemory.length === 0 && ((currentTokenCount + tokenCount) <= remainingContextTokens)) {\n        context.push(poppedMessage);\n        currentTokenCount += tokenCount;\n      } else {\n        prunedMemory.push(poppedMessage);\n        if (thinkingEndIndex > -1) {\n          continue;\n        }\n        break;\n      }\n    }\n\n    if (thinkingEndIndex > -1 && context[context.length - 1]?.getType() === 'tool') {\n      startType = 'ai';\n    }\n\n    if (startType != null && startType && context.length > 0) {\n      const requiredTypeIndex = context.findIndex(msg => msg?.getType() === startType);\n\n      if (requiredTypeIndex > 0) {\n        context = context.slice(requiredTypeIndex);\n      }\n    }\n  }\n\n  if (instructions && originalLength > 0) {\n    context.push(_messages[0] as BaseMessage);\n    messages.shift();\n  }\n\n  remainingContextTokens -= currentTokenCount;\n  const result = {\n    remainingContextTokens,\n    context: [] as BaseMessage[],\n    messagesToRefine: prunedMemory,\n  };\n\n  if (prunedMemory.length === 0 || thinkingEndIndex < 0 || (thinkingStartIndex > -1 && isIndexInContext(_messages, context, thinkingStartIndex))) {\n    // we reverse at this step to ensure the context is in the correct order for the model, and we need to work backwards\n    result.context = context.reverse() as BaseMessage[];\n    return result;\n  }\n\n  if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no \"AI\" messages with thinking blocks.');\n  }\n\n  if (!thinkingBlock) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no thinking block found.');\n  }\n\n  // Since we have a thinking sequence, we need to find the last assistant message\n  // in the latest AI/tool sequence to add the thinking block that falls outside of the current context\n  // Latest messages are ordered first.\n  let assistantIndex = -1;\n  for (let i = 0; i < context.length; i++) {\n    const currentMessage = context[i];\n    const type = currentMessage?.getType();\n    if (type === 'ai') {\n      assistantIndex = i;\n    }\n    if (assistantIndex > -1 && (type === 'human' || type === 'system')) {\n      break;\n    }\n  }\n\n  if (assistantIndex === -1) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no \"AI\" messages to append thinking blocks to.');\n  }\n\n  thinkingStartIndex = originalLength - 1 - assistantIndex;\n  const thinkingTokenCount = tokenCounter(new AIMessage({ content: [thinkingBlock] }));\n  const newRemainingCount = remainingContextTokens - thinkingTokenCount;\n  const content: MessageContentComplex[] = addThinkingBlock(context[assistantIndex] as AIMessage, thinkingBlock);\n  (context[assistantIndex] as AIMessage).content = content;\n  if (newRemainingCount > 0) {\n    result.context = context.reverse() as BaseMessage[];\n    return result;\n  }\n\n  const thinkingMessage: AIMessage = context[assistantIndex] as AIMessage;\n  // now we need to an additional round of pruning but making the thinking block fit\n  const newThinkingMessageTokenCount = (indexTokenCountMap[thinkingStartIndex] ?? 0) + thinkingTokenCount;\n  remainingContextTokens = initialContextTokens - newThinkingMessageTokenCount;\n  currentTokenCount = 3;\n  let newContext: BaseMessage[] = [];\n  const secondRoundMessages = [..._messages];\n  let currentIndex = secondRoundMessages.length;\n  while (secondRoundMessages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > thinkingStartIndex) {\n    currentIndex--;\n    const poppedMessage = secondRoundMessages.pop();\n    if (!poppedMessage) continue;\n    const tokenCount = indexTokenCountMap[currentIndex] ?? 0;\n    if ((currentTokenCount + tokenCount) <= remainingContextTokens) {\n      newContext.push(poppedMessage);\n      currentTokenCount += tokenCount;\n    } else {\n      messages.push(poppedMessage);\n      break;\n    }\n  }\n\n  const firstMessage: AIMessage = newContext[newContext.length - 1];\n  const firstMessageType = newContext[newContext.length - 1].getType();\n  if (firstMessageType === 'tool') {\n    startType = 'ai';\n  }\n\n  if (startType != null && startType && newContext.length > 0) {\n    const requiredTypeIndex = newContext.findIndex(msg => msg.getType() === startType);\n    if (requiredTypeIndex > 0) {\n      newContext = newContext.slice(requiredTypeIndex);\n    }\n  }\n\n  if (firstMessageType === 'ai') {\n    const content = addThinkingBlock(firstMessage, thinkingBlock);\n    newContext[newContext.length - 1].content = content;\n  } else {\n    newContext.push(thinkingMessage);\n  }\n\n  if (instructions && originalLength > 0) {\n    newContext.push(_messages[0] as BaseMessage);\n    secondRoundMessages.shift();\n  }\n\n  result.context = newContext.reverse();\n  return result;\n}\n\nexport function checkValidNumber(value: unknown): value is number {\n  return typeof value === 'number' && !isNaN(value) && value > 0;\n}\n\nexport function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {\n  const indexTokenCountMap = { ...factoryParams.indexTokenCountMap };\n  let lastTurnStartIndex = factoryParams.startIndex;\n  let lastCutOffIndex = 0;\n  let totalTokens = (Object.values(indexTokenCountMap)).reduce((a, b) => a + b, 0);\n  return function pruneMessages(params: PruneMessagesParams): {\n    context: BaseMessage[];\n    indexTokenCountMap: Record<string, number>;\n  } {\n    let currentUsage: UsageMetadata | undefined;\n    if (params.usageMetadata && (\n      checkValidNumber(params.usageMetadata.input_tokens)\n      || (\n        checkValidNumber(params.usageMetadata.input_token_details)\n        && (\n          checkValidNumber(params.usageMetadata.input_token_details.cache_creation)\n          || checkValidNumber(params.usageMetadata.input_token_details.cache_read)\n        )\n      )\n    ) && checkValidNumber(params.usageMetadata.output_tokens)) {\n      currentUsage = calculateTotalTokens(params.usageMetadata);\n      totalTokens = currentUsage.total_tokens;\n    }\n\n    for (let i = lastTurnStartIndex; i < params.messages.length; i++) {\n      const message = params.messages[i];\n      // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition\n      if (i === lastTurnStartIndex && indexTokenCountMap[i] === undefined && currentUsage) {\n        indexTokenCountMap[i] = currentUsage.output_tokens;\n      // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition\n      } else if (indexTokenCountMap[i] === undefined) {\n        indexTokenCountMap[i] = factoryParams.tokenCounter(message);\n        totalTokens += indexTokenCountMap[i];\n      }\n    }\n\n    // If `currentUsage` is defined, we need to distribute the current total tokens to our `indexTokenCountMap`,\n    // We must distribute it in a weighted manner, so that the total token count is equal to `currentUsage.total_tokens`,\n    // relative the manually counted tokens in `indexTokenCountMap`.\n    // EDGE CASE: when the resulting context gets pruned, we should not distribute the usage for messages that are not in the context.\n    if (currentUsage) {\n      // Calculate the sum of tokens only for indices at or after lastCutOffIndex\n      const totalIndexTokens = Object.entries(indexTokenCountMap).reduce((sum, [key, value]) => {\n        // Convert string key to number and check if it's >= lastCutOffIndex\n        const numericKey = Number(key);\n        if (numericKey === 0 && params.messages[0].getType() === 'system') {\n          return sum + value;\n        }\n        return numericKey >= lastCutOffIndex ? sum + value : sum;\n      }, 0);\n\n      // Calculate ratio based only on messages that remain in the context\n      const ratio = currentUsage.total_tokens / totalIndexTokens;\n      const isRatioSafe = ratio >= 1/3 && ratio <= 2.5;\n\n      // Apply the ratio adjustment only to messages at or after lastCutOffIndex, and only if the ratio is safe\n      if (isRatioSafe) {\n        for (const key in indexTokenCountMap) {\n          const numericKey = Number(key);\n          if (numericKey === 0 && params.messages[0].getType() === 'system') {\n            indexTokenCountMap[key] = Math.round(indexTokenCountMap[key] * ratio);\n          } else if (numericKey >= lastCutOffIndex) {\n            // Only adjust token counts for messages still in the context\n            indexTokenCountMap[key] = Math.round(indexTokenCountMap[key] * ratio);\n          }\n        }\n      }\n    }\n\n    lastTurnStartIndex = params.messages.length;\n    if (totalTokens <= factoryParams.maxTokens) {\n      return { context: params.messages, indexTokenCountMap };\n    }\n\n    const { context } = getMessagesWithinTokenLimit({\n      maxContextTokens: factoryParams.maxTokens,\n      messages: params.messages,\n      indexTokenCountMap,\n      startType: params.startType,\n      thinkingEnabled: factoryParams.thinkingEnabled,\n      tokenCounter: factoryParams.tokenCounter,\n      reasoningType: factoryParams.provider === Providers.BEDROCK ? ContentTypes.REASONING_CONTENT : ContentTypes.THINKING,\n    });\n    lastCutOffIndex = Math.max(params.messages.length - context.length, 0);\n\n    return { context, indexTokenCountMap };\n  };\n}\n"],"names":["ContentTypes","messages","AIMessage","Providers"],"mappings":";;;;;AAmBA,SAAS,gBAAgB,CAAC,MAAiB,EAAE,MAAiB,EAAE,WAAmB,EAAA;IACjF,MAAM,gBAAgB,GAAG,MAAM,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM;IACtD,OAAO,WAAW,IAAI,gBAAgB;AACxC;AAEA,SAAS,gBAAgB,CAAC,OAAkB,EAAE,aAAyD,EAAA;IACrG,MAAM,OAAO,GAA4B,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO;UAClE,OAAO,CAAC;AACV,UAAE,CAAC;gBACD,IAAI,EAAEA,kBAAY,CAAC,IAAI;gBACvB,IAAI,EAAE,OAAO,CAAC,OAAO;AACtB,aAAA,CAAC;AACJ,IAAA,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;AAC9B,IAAA,OAAO,OAAO;AAChB;AAEA;;;;;AAKG;AACG,SAAU,oBAAoB,CAAC,KAA6B,EAAA;IAChE,MAAM,eAAe,GAAG,MAAM,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC;AACvD,IAAA,MAAM,aAAa,GAAG,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,cAAc,CAAC,IAAI,CAAC;AAC5E,IAAA,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,UAAU,CAAC,IAAI,CAAC;AAEpE,IAAA,MAAM,gBAAgB,GAAG,eAAe,GAAG,aAAa,GAAG,SAAS;IACpE,MAAM,iBAAiB,GAAG,MAAM,CAAC,KAAK,CAAC,aAAa,CAAC,IAAI,CAAC;IAE1D,OAAO;AACL,QAAA,YAAY,EAAE,gBAAgB;AAC9B,QAAA,aAAa,EAAE,iBAAiB;QAChC,YAAY,EAAE,gBAAgB,GAAG;KAClC;AACH;AAEA;;;;;;AAMG;AACG,SAAU,2BAA2B,CAAC,EAC1C,QAAQ,EAAE,SAAS,EACnB,gBAAgB,EAChB,kBAAkB,EAClB,SAAS,EAAE,UAAU,EACrB,eAAe,EACf,YAAY,EACZ,aAAa,GAAGA,kBAAY,CAAC,QAAQ,GAStC,EAAA;;;IAOC,IAAI,iBAAiB,GAAG,CAAC;IACzB,MAAM,YAAY,GAAG,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,QAAQ,GAAG,SAAS,CAAC,CAAC,CAAC,GAAG,SAAS;AACpF,IAAA,MAAM,sBAAsB,GAAG,YAAY,IAAI,IAAI,GAAG,kBAAkB,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC;AACpF,IAAA,MAAM,oBAAoB,GAAG,gBAAgB,GAAG,sBAAsB;IACtE,IAAI,sBAAsB,GAAG,oBAAoB;IACjD,IAAI,SAAS,GAAG,UAAU;AAC1B,IAAA,MAAM,cAAc,GAAG,SAAS,CAAC,MAAM;AACvC,IAAA,MAAMC,UAAQ,GAAG,CAAC,GAAG,SAAS,CAAC;AAC/B;;;;AAIK;IACL,IAAI,OAAO,GAAmC,EAAE;AAEhD,IAAA,IAAI,kBAAkB,GAAG,EAAE;AAC3B,IAAA,IAAI,gBAAgB,GAAG,EAAE;AACzB,IAAA,IAAI,aAAqE;AACzE,IAAA,MAAM,QAAQ,GAAG,YAAY,IAAI,IAAI,GAAG,CAAC,GAAG,CAAC;IAC7C,MAAM,YAAY,GAAkB,EAAE;AAEtC,IAAA,IAAI,iBAAiB,GAAG,sBAAsB,EAAE;AAC9C,QAAA,IAAI,YAAY,GAAGA,UAAQ,CAAC,MAAM;AAClC,QAAA,OAAOA,UAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,GAAG,sBAAsB,IAAI,YAAY,GAAG,QAAQ,EAAE;AACnG,YAAA,YAAY,EAAE;YACd,IAAIA,UAAQ,CAAC,MAAM,KAAK,CAAC,IAAI,YAAY,EAAE;gBACzC;;AAEF,YAAA,MAAM,aAAa,GAAGA,UAAQ,CAAC,GAAG,EAAE;AACpC,YAAA,IAAI,CAAC,aAAa;gBAAE;AACpB,YAAA,MAAM,WAAW,GAAG,aAAa,CAAC,OAAO,EAAE;AAC3C,YAAA,IAAI,eAAe,KAAK,IAAI,IAAI,gBAAgB,KAAK,EAAE,KAAK,YAAY,MAAM,cAAc,GAAG,CAAC,CAAC,CAAC,KAAK,WAAW,KAAK,IAAI,IAAI,WAAW,KAAK,MAAM,CAAC,EAAE;gBACtJ,gBAAgB,GAAG,YAAY;;YAEjC,IAAI,gBAAgB,GAAG,EAAE,IAAI,CAAC,aAAa,IAAK,kBAAkB,GAAG,CAAC,IAAI,WAAW,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,aAAa,CAAC,OAAO,CAAC,EAAE;gBACtI,aAAa,IAAI,aAAa,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,OAAO,KAAK,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAoC;AAC5H,gBAAA,kBAAkB,GAAG,aAAa,IAAI,IAAI,GAAG,YAAY,GAAG,EAAE;;;YAGhE,IACE,gBAAgB,GAAG;AAChB,mBAAA,YAAY,MAAM,gBAAgB,GAAG,CAAC;oBACrC,WAAW,KAAK,IAAI,IAAI,WAAW,KAAK,MAAM,CAAC,EACnD;gBACA,gBAAgB,GAAG,EAAE;;YAGvB,MAAM,UAAU,GAAG,kBAAkB,CAAC,YAAY,CAAC,IAAI,CAAC;AAExD,YAAA,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC,KAAK,CAAC,iBAAiB,GAAG,UAAU,KAAK,sBAAsB,CAAC,EAAE;AAC7F,gBAAA,OAAO,CAAC,IAAI,CAAC,aAAa,CAAC;gBAC3B,iBAAiB,IAAI,UAAU;;iBAC1B;AACL,gBAAA,YAAY,CAAC,IAAI,CAAC,aAAa,CAAC;AAChC,gBAAA,IAAI,gBAAgB,GAAG,EAAE,EAAE;oBACzB;;gBAEF;;;AAIJ,QAAA,IAAI,gBAAgB,GAAG,EAAE,IAAI,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,MAAM,EAAE;YAC9E,SAAS,GAAG,IAAI;;AAGlB,QAAA,IAAI,SAAS,IAAI,IAAI,IAAI,SAAS,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE;AACxD,YAAA,MAAM,iBAAiB,GAAG,OAAO,CAAC,SAAS,CAAC,GAAG,IAAI,GAAG,EAAE,OAAO,EAAE,KAAK,SAAS,CAAC;AAEhF,YAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;AACzB,gBAAA,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,iBAAiB,CAAC;;;;AAKhD,IAAA,IAAI,YAAY,IAAI,cAAc,GAAG,CAAC,EAAE;QACtC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAgB,CAAC;QACzCA,UAAQ,CAAC,KAAK,EAAE;;IAGlB,sBAAsB,IAAI,iBAAiB;AAC3C,IAAA,MAAM,MAAM,GAAG;QACb,sBAAsB;AACtB,QAAA,OAAO,EAAE,EAAmB;AAC5B,QAAA,gBAAgB,EAAE,YAAY;KAC/B;IAED,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC,IAAI,gBAAgB,GAAG,CAAC,KAAK,kBAAkB,GAAG,EAAE,IAAI,gBAAgB,CAAC,SAAS,EAAE,OAAO,EAAE,kBAAkB,CAAC,CAAC,EAAE;;AAE9I,QAAA,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,EAAmB;AACnD,QAAA,OAAO,MAAM;;IAGf,IAAI,gBAAgB,GAAG,EAAE,IAAI,kBAAkB,GAAG,CAAC,EAAE;AACnD,QAAA,MAAM,IAAI,KAAK,CAAC,mGAAmG,CAAC;;IAGtH,IAAI,CAAC,aAAa,EAAE;AAClB,QAAA,MAAM,IAAI,KAAK,CAAC,qFAAqF,CAAC;;;;;AAMxG,IAAA,IAAI,cAAc,GAAG,EAAE;AACvB,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;AACvC,QAAA,MAAM,cAAc,GAAG,OAAO,CAAC,CAAC,CAAC;AACjC,QAAA,MAAM,IAAI,GAAG,cAAc,EAAE,OAAO,EAAE;AACtC,QAAA,IAAI,IAAI,KAAK,IAAI,EAAE;YACjB,cAAc,GAAG,CAAC;;AAEpB,QAAA,IAAI,cAAc,GAAG,EAAE,KAAK,IAAI,KAAK,OAAO,IAAI,IAAI,KAAK,QAAQ,CAAC,EAAE;YAClE;;;AAIJ,IAAA,IAAI,cAAc,KAAK,EAAE,EAAE;AACzB,QAAA,MAAM,IAAI,KAAK,CAAC,2GAA2G,CAAC;;AAG9H,IAAA,kBAAkB,GAAG,cAAc,GAAG,CAAC,GAAG,cAAc;AACxD,IAAA,MAAM,kBAAkB,GAAG,YAAY,CAAC,IAAIC,kBAAS,CAAC,EAAE,OAAO,EAAE,CAAC,aAAa,CAAC,EAAE,CAAC,CAAC;AACpF,IAAA,MAAM,iBAAiB,GAAG,sBAAsB,GAAG,kBAAkB;IACrE,MAAM,OAAO,GAA4B,gBAAgB,CAAC,OAAO,CAAC,cAAc,CAAc,EAAE,aAAa,CAAC;AAC7G,IAAA,OAAO,CAAC,cAAc,CAAe,CAAC,OAAO,GAAG,OAAO;AACxD,IAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;AACzB,QAAA,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,EAAmB;AACnD,QAAA,OAAO,MAAM;;AAGf,IAAA,MAAM,eAAe,GAAc,OAAO,CAAC,cAAc,CAAc;;AAEvE,IAAA,MAAM,4BAA4B,GAAG,CAAC,kBAAkB,CAAC,kBAAkB,CAAC,IAAI,CAAC,IAAI,kBAAkB;AACvG,IAAA,sBAAsB,GAAG,oBAAoB,GAAG,4BAA4B;IAC5E,iBAAiB,GAAG,CAAC;IACrB,IAAI,UAAU,GAAkB,EAAE;AAClC,IAAA,MAAM,mBAAmB,GAAG,CAAC,GAAG,SAAS,CAAC;AAC1C,IAAA,IAAI,YAAY,GAAG,mBAAmB,CAAC,MAAM;AAC7C,IAAA,OAAO,mBAAmB,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,GAAG,sBAAsB,IAAI,YAAY,GAAG,kBAAkB,EAAE;AACxH,QAAA,YAAY,EAAE;AACd,QAAA,MAAM,aAAa,GAAG,mBAAmB,CAAC,GAAG,EAAE;AAC/C,QAAA,IAAI,CAAC,aAAa;YAAE;QACpB,MAAM,UAAU,GAAG,kBAAkB,CAAC,YAAY,CAAC,IAAI,CAAC;QACxD,IAAI,CAAC,iBAAiB,GAAG,UAAU,KAAK,sBAAsB,EAAE;AAC9D,YAAA,UAAU,CAAC,IAAI,CAAC,aAAa,CAAC;YAC9B,iBAAiB,IAAI,UAAU;;aAC1B;AACL,YAAAD,UAAQ,CAAC,IAAI,CAAC,aAAa,CAAC;YAC5B;;;IAIJ,MAAM,YAAY,GAAc,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;AACjE,IAAA,MAAM,gBAAgB,GAAG,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,OAAO,EAAE;AACpE,IAAA,IAAI,gBAAgB,KAAK,MAAM,EAAE;QAC/B,SAAS,GAAG,IAAI;;AAGlB,IAAA,IAAI,SAAS,IAAI,IAAI,IAAI,SAAS,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;AAC3D,QAAA,MAAM,iBAAiB,GAAG,UAAU,CAAC,SAAS,CAAC,GAAG,IAAI,GAAG,CAAC,OAAO,EAAE,KAAK,SAAS,CAAC;AAClF,QAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;AACzB,YAAA,UAAU,GAAG,UAAU,CAAC,KAAK,CAAC,iBAAiB,CAAC;;;AAIpD,IAAA,IAAI,gBAAgB,KAAK,IAAI,EAAE;QAC7B,MAAM,OAAO,GAAG,gBAAgB,CAAC,YAAY,EAAE,aAAa,CAAC;QAC7D,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,OAAO,GAAG,OAAO;;SAC9C;AACL,QAAA,UAAU,CAAC,IAAI,CAAC,eAAe,CAAC;;AAGlC,IAAA,IAAI,YAAY,IAAI,cAAc,GAAG,CAAC,EAAE;QACtC,UAAU,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAgB,CAAC;QAC5C,mBAAmB,CAAC,KAAK,EAAE;;AAG7B,IAAA,MAAM,CAAC,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE;AACrC,IAAA,OAAO,MAAM;AACf;AAEM,SAAU,gBAAgB,CAAC,KAAc,EAAA;AAC7C,IAAA,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC;AAChE;AAEM,SAAU,mBAAmB,CAAC,aAAyC,EAAA;IAC3E,MAAM,kBAAkB,GAAG,EAAE,GAAG,aAAa,CAAC,kBAAkB,EAAE;AAClE,IAAA,IAAI,kBAAkB,GAAG,aAAa,CAAC,UAAU;IACjD,IAAI,eAAe,GAAG,CAAC;IACvB,IAAI,WAAW,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,kBAAkB,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;IAChF,OAAO,SAAS,aAAa,CAAC,MAA2B,EAAA;AAIvD,QAAA,IAAI,YAAuC;AAC3C,QAAA,IAAI,MAAM,CAAC,aAAa,KACtB,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,YAAY;AAC/C,gBACD,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB;oBAEvD,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB,CAAC,cAAc;uBACrE,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB,CAAC,UAAU,CAAC,CACzE,CACF,CACF,IAAI,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,aAAa,CAAC,EAAE;AACzD,YAAA,YAAY,GAAG,oBAAoB,CAAC,MAAM,CAAC,aAAa,CAAC;AACzD,YAAA,WAAW,GAAG,YAAY,CAAC,YAAY;;AAGzC,QAAA,KAAK,IAAI,CAAC,GAAG,kBAAkB,EAAE,CAAC,GAAG,MAAM,CAAC,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;YAChE,MAAM,OAAO,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC;;AAElC,YAAA,IAAI,CAAC,KAAK,kBAAkB,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,SAAS,IAAI,YAAY,EAAE;AACnF,gBAAA,kBAAkB,CAAC,CAAC,CAAC,GAAG,YAAY,CAAC,aAAa;;;AAE7C,iBAAA,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,SAAS,EAAE;gBAC9C,kBAAkB,CAAC,CAAC,CAAC,GAAG,aAAa,CAAC,YAAY,CAAC,OAAO,CAAC;AAC3D,gBAAA,WAAW,IAAI,kBAAkB,CAAC,CAAC,CAAC;;;;;;;QAQxC,IAAI,YAAY,EAAE;;YAEhB,MAAM,gBAAgB,GAAG,MAAM,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,KAAK,CAAC,KAAI;;AAEvF,gBAAA,MAAM,UAAU,GAAG,MAAM,CAAC,GAAG,CAAC;AAC9B,gBAAA,IAAI,UAAU,KAAK,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,KAAK,QAAQ,EAAE;oBACjE,OAAO,GAAG,GAAG,KAAK;;AAEpB,gBAAA,OAAO,UAAU,IAAI,eAAe,GAAG,GAAG,GAAG,KAAK,GAAG,GAAG;aACzD,EAAE,CAAC,CAAC;;AAGL,YAAA,MAAM,KAAK,GAAG,YAAY,CAAC,YAAY,GAAG,gBAAgB;YAC1D,MAAM,WAAW,GAAG,KAAK,IAAI,CAAC,GAAC,CAAC,IAAI,KAAK,IAAI,GAAG;;YAGhD,IAAI,WAAW,EAAE;AACf,gBAAA,KAAK,MAAM,GAAG,IAAI,kBAAkB,EAAE;AACpC,oBAAA,MAAM,UAAU,GAAG,MAAM,CAAC,GAAG,CAAC;AAC9B,oBAAA,IAAI,UAAU,KAAK,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,KAAK,QAAQ,EAAE;AACjE,wBAAA,kBAAkB,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;;AAChE,yBAAA,IAAI,UAAU,IAAI,eAAe,EAAE;;AAExC,wBAAA,kBAAkB,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;;;;;AAM7E,QAAA,kBAAkB,GAAG,MAAM,CAAC,QAAQ,CAAC,MAAM;AAC3C,QAAA,IAAI,WAAW,IAAI,aAAa,CAAC,SAAS,EAAE;YAC1C,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,QAAQ,EAAE,kBAAkB,EAAE;;AAGzD,QAAA,MAAM,EAAE,OAAO,EAAE,GAAG,2BAA2B,CAAC;YAC9C,gBAAgB,EAAE,aAAa,CAAC,SAAS;YACzC,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,kBAAkB;YAClB,SAAS,EAAE,MAAM,CAAC,SAAS;YAC3B,eAAe,EAAE,aAAa,CAAC,eAAe;YAC9C,YAAY,EAAE,aAAa,CAAC,YAAY;AACxC,YAAA,aAAa,EAAE,aAAa,CAAC,QAAQ,KAAKE,eAAS,CAAC,OAAO,GAAGH,kBAAY,CAAC,iBAAiB,GAAGA,kBAAY,CAAC,QAAQ;AACrH,SAAA,CAAC;AACF,QAAA,eAAe,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC;AAEtE,QAAA,OAAO,EAAE,OAAO,EAAE,kBAAkB,EAAE;AACxC,KAAC;AACH;;;;;;;"}
+{"version":3,"file":"prune.cjs","sources":["../../../src/messages/prune.ts"],"sourcesContent":["import { AIMessage, BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { ThinkingContentText, MessageContentComplex, ReasoningContentText } from '@/types/stream';\nimport type { TokenCounter } from '@/types/run';\nimport { ContentTypes, Providers } from '@/common';\n\nexport type PruneMessagesFactoryParams = {\n  provider?: Providers;\n  maxTokens: number;\n  startIndex: number;\n  tokenCounter: TokenCounter;\n  indexTokenCountMap: Record<string, number>;\n  thinkingEnabled?: boolean;\n};\nexport type PruneMessagesParams = {\n  messages: BaseMessage[];\n  usageMetadata?: Partial<UsageMetadata>;\n  startType?: ReturnType<BaseMessage['getType']>;\n}\n\nfunction isIndexInContext(arrayA: unknown[], arrayB: unknown[], targetIndex: number): boolean {\n  const startingIndexInA = arrayA.length - arrayB.length;\n  return targetIndex >= startingIndexInA;\n}\n\nfunction addThinkingBlock(message: AIMessage, thinkingBlock: ThinkingContentText | ReasoningContentText): AIMessage {\n  const content: MessageContentComplex[] = Array.isArray(message.content)\n    ? message.content as MessageContentComplex[]\n    : [{\n      type: ContentTypes.TEXT,\n      text: message.content,\n    }];\n  /** Edge case, the message already has the thinking block */\n  if (content[0].type === thinkingBlock.type) {\n    return message;\n  }\n  content.unshift(thinkingBlock);\n  return new AIMessage({\n    ...message,\n    content\n  });\n}\n\n/**\n * Calculates the total tokens from a single usage object\n *\n * @param usage The usage metadata object containing token information\n * @returns An object containing the total input and output tokens\n */\nexport function calculateTotalTokens(usage: Partial<UsageMetadata>): UsageMetadata {\n  const baseInputTokens = Number(usage.input_tokens) || 0;\n  const cacheCreation = Number(usage.input_token_details?.cache_creation) || 0;\n  const cacheRead = Number(usage.input_token_details?.cache_read) || 0;\n\n  const totalInputTokens = baseInputTokens + cacheCreation + cacheRead;\n  const totalOutputTokens = Number(usage.output_tokens) || 0;\n\n  return {\n    input_tokens: totalInputTokens,\n    output_tokens: totalOutputTokens,\n    total_tokens: totalInputTokens + totalOutputTokens\n  };\n}\n\nexport type PruningResult = {\n  context: BaseMessage[];\n  remainingContextTokens: number;\n  messagesToRefine: BaseMessage[];\n  thinkingStartIndex?: number;\n};\n\n/**\n * Processes an array of messages and returns a context of messages that fit within a specified token limit.\n * It iterates over the messages from newest to oldest, adding them to the context until the token limit is reached.\n *\n * @param options Configuration options for processing messages\n * @returns Object containing the message context, remaining tokens, messages not included, and summary index\n */\nexport function getMessagesWithinTokenLimit({\n  messages: _messages,\n  maxContextTokens,\n  indexTokenCountMap,\n  startType: _startType,\n  thinkingEnabled,\n  tokenCounter,\n  thinkingStartIndex: _thinkingStartIndex = -1,\n  reasoningType = ContentTypes.THINKING,\n}: {\n  messages: BaseMessage[];\n  maxContextTokens: number;\n  indexTokenCountMap: Record<string, number | undefined>;\n  startType?: string | string[];\n  thinkingEnabled?: boolean;\n  tokenCounter: TokenCounter;\n  thinkingStartIndex?: number;\n  reasoningType?: ContentTypes.THINKING | ContentTypes.REASONING_CONTENT;\n}): PruningResult {\n  // Every reply is primed with <|start|>assistant<|message|>, so we\n  // start with 3 tokens for the label after all messages have been counted.\n  let currentTokenCount = 3;\n  const instructions = _messages[0]?.getType() === 'system' ? _messages[0] : undefined;\n  const instructionsTokenCount = instructions != null ? indexTokenCountMap[0] ?? 0 : 0;\n  const initialContextTokens = maxContextTokens - instructionsTokenCount;\n  let remainingContextTokens = initialContextTokens;\n  let startType = _startType;\n  const originalLength = _messages.length;\n  const messages = [..._messages];\n  /**\n   * IMPORTANT: this context array gets reversed at the end, since the latest messages get pushed first.\n   *\n   * This may be confusing to read, but it is done to ensure the context is in the correct order for the model.\n   * */\n  let context: Array<BaseMessage | undefined> = [];\n\n  let thinkingStartIndex = _thinkingStartIndex;\n  let thinkingEndIndex = -1;\n  let thinkingBlock: ThinkingContentText | ReasoningContentText | undefined;\n  const endIndex = instructions != null ? 1 : 0;\n  const prunedMemory: BaseMessage[] = [];\n\n  if (_thinkingStartIndex > -1) {\n    const thinkingMessageContent = messages[_thinkingStartIndex]?.content;\n    if (Array.isArray(thinkingMessageContent)) {\n      thinkingBlock = thinkingMessageContent.find((content) => content.type === reasoningType) as ThinkingContentText | undefined;\n    }\n  }\n\n  if (currentTokenCount < remainingContextTokens) {\n    let currentIndex = messages.length;\n    while (messages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > endIndex) {\n      currentIndex--;\n      if (messages.length === 1 && instructions) {\n        break;\n      }\n      const poppedMessage = messages.pop();\n      if (!poppedMessage) continue;\n      const messageType = poppedMessage.getType();\n      if (thinkingEnabled === true && thinkingEndIndex === -1 && (currentIndex === (originalLength - 1)) && (messageType === 'ai' || messageType === 'tool')) {\n        thinkingEndIndex = currentIndex;\n      }\n      if (thinkingEndIndex > -1 && !thinkingBlock  && thinkingStartIndex < 0 && messageType === 'ai' && Array.isArray(poppedMessage.content)) {\n        thinkingBlock = (poppedMessage.content.find((content) => content.type === reasoningType)) as ThinkingContentText | undefined;\n        thinkingStartIndex = thinkingBlock != null ? currentIndex : -1;\n      }\n      /** False start, the latest message was not part of a multi-assistant/tool sequence of messages */\n      if (\n        thinkingEndIndex > -1\n        && currentIndex === (thinkingEndIndex - 1)\n        && (messageType !== 'ai' && messageType !== 'tool')\n      ) {\n        thinkingEndIndex = -1;\n      }\n\n      const tokenCount = indexTokenCountMap[currentIndex] ?? 0;\n\n      if (prunedMemory.length === 0 && ((currentTokenCount + tokenCount) <= remainingContextTokens)) {\n        context.push(poppedMessage);\n        currentTokenCount += tokenCount;\n      } else {\n        prunedMemory.push(poppedMessage);\n        if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {\n          continue;\n        }\n        break;\n      }\n    }\n\n    if (context[context.length - 1]?.getType() === 'tool') {\n      startType = ['ai', 'human'];\n    }\n\n    if (startType != null && startType.length > 0 && context.length > 0) {\n      let requiredTypeIndex = -1;\n\n      let totalTokens = 0;\n      for (let i = context.length - 1; i >= 0; i--) {\n        const currentType = context[i]?.getType() ?? '';\n        if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {\n          requiredTypeIndex = i + 1;\n          break;\n        }\n        const originalIndex = originalLength - 1 - i;\n        totalTokens += indexTokenCountMap[originalIndex] ?? 0;\n      }\n\n      if (requiredTypeIndex > 0) {\n        currentTokenCount -= totalTokens;\n        context = context.slice(0, requiredTypeIndex);\n      }\n    }\n  }\n\n  if (instructions && originalLength > 0) {\n    context.push(_messages[0] as BaseMessage);\n    messages.shift();\n  }\n\n  remainingContextTokens -= currentTokenCount;\n  const result: PruningResult = {\n    remainingContextTokens,\n    context: [] as BaseMessage[],\n    messagesToRefine: prunedMemory,\n  };\n\n  if (thinkingStartIndex > -1) {\n    result.thinkingStartIndex = thinkingStartIndex;\n  }\n\n  if (prunedMemory.length === 0 || thinkingEndIndex < 0 || (thinkingStartIndex > -1 && isIndexInContext(_messages, context, thinkingStartIndex))) {\n    // we reverse at this step to ensure the context is in the correct order for the model, and we need to work backwards\n    result.context = context.reverse() as BaseMessage[];\n    return result;\n  }\n\n  if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no \"AI\" messages with thinking blocks.');\n  }\n\n  if (!thinkingBlock) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no thinking block found.');\n  }\n\n  // Since we have a thinking sequence, we need to find the last assistant message\n  // in the latest AI/tool sequence to add the thinking block that falls outside of the current context\n  // Latest messages are ordered first.\n  let assistantIndex = -1;\n  for (let i = 0; i < context.length; i++) {\n    const currentMessage = context[i];\n    const type = currentMessage?.getType();\n    if (type === 'ai') {\n      assistantIndex = i;\n    }\n    if (assistantIndex > -1 && (type === 'human' || type === 'system')) {\n      break;\n    }\n  }\n\n  if (assistantIndex === -1) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no \"AI\" messages to append thinking blocks to.');\n  }\n\n  thinkingStartIndex = originalLength - 1 - assistantIndex;\n  const thinkingTokenCount = tokenCounter(new AIMessage({ content: [thinkingBlock] }));\n  const newRemainingCount = remainingContextTokens - thinkingTokenCount;\n  const newMessage = addThinkingBlock(context[assistantIndex] as AIMessage, thinkingBlock);\n  context[assistantIndex] = newMessage;\n  if (newRemainingCount > 0) {\n    result.context = context.reverse() as BaseMessage[];\n    return result;\n  }\n\n  const thinkingMessage: AIMessage = context[assistantIndex] as AIMessage;\n  // now we need to an additional round of pruning but making the thinking block fit\n  const newThinkingMessageTokenCount = (indexTokenCountMap[thinkingStartIndex] ?? 0) + thinkingTokenCount;\n  remainingContextTokens = initialContextTokens - newThinkingMessageTokenCount;\n  currentTokenCount = 3;\n  let newContext: BaseMessage[] = [];\n  const secondRoundMessages = [..._messages];\n  let currentIndex = secondRoundMessages.length;\n  while (secondRoundMessages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > thinkingStartIndex) {\n    currentIndex--;\n    const poppedMessage = secondRoundMessages.pop();\n    if (!poppedMessage) continue;\n    const tokenCount = indexTokenCountMap[currentIndex] ?? 0;\n    if ((currentTokenCount + tokenCount) <= remainingContextTokens) {\n      newContext.push(poppedMessage);\n      currentTokenCount += tokenCount;\n    } else {\n      messages.push(poppedMessage);\n      break;\n    }\n  }\n\n  const firstMessage: AIMessage = newContext[newContext.length - 1];\n  const firstMessageType = newContext[newContext.length - 1].getType();\n  if (firstMessageType === 'tool') {\n    startType = ['ai', 'human'];\n  }\n\n  if (startType != null && startType.length > 0 && newContext.length > 0) {\n    let requiredTypeIndex = -1;\n\n    let totalTokens = 0;\n    for (let i = newContext.length - 1; i >= 0; i--) {\n      const currentType = newContext[i]?.getType() ?? '';\n      if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {\n        requiredTypeIndex = i + 1;\n        break;\n      }\n      const originalIndex = originalLength - 1 - i;\n      totalTokens += indexTokenCountMap[originalIndex] ?? 0;\n    }\n\n    if (requiredTypeIndex > 0) {\n      currentTokenCount -= totalTokens;\n      newContext = newContext.slice(0, requiredTypeIndex);\n    }\n  }\n\n  if (firstMessageType === 'ai') {\n    const newMessage = addThinkingBlock(firstMessage, thinkingBlock);\n    newContext[newContext.length - 1] = newMessage;\n  } else {\n    newContext.push(thinkingMessage);\n  }\n\n  if (instructions && originalLength > 0) {\n    newContext.push(_messages[0] as BaseMessage);\n    secondRoundMessages.shift();\n  }\n\n  result.context = newContext.reverse();\n  return result;\n}\n\nexport function checkValidNumber(value: unknown): value is number {\n  return typeof value === 'number' && !isNaN(value) && value > 0;\n}\n\nexport function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {\n  const indexTokenCountMap = { ...factoryParams.indexTokenCountMap };\n  let lastTurnStartIndex = factoryParams.startIndex;\n  let lastCutOffIndex = 0;\n  let totalTokens = (Object.values(indexTokenCountMap)).reduce((a, b) => a + b, 0);\n  let runThinkingStartIndex = -1;\n  return function pruneMessages(params: PruneMessagesParams): {\n    context: BaseMessage[];\n    indexTokenCountMap: Record<string, number>;\n  } {\n    let currentUsage: UsageMetadata | undefined;\n    if (params.usageMetadata && (\n      checkValidNumber(params.usageMetadata.input_tokens)\n      || (\n        checkValidNumber(params.usageMetadata.input_token_details)\n        && (\n          checkValidNumber(params.usageMetadata.input_token_details.cache_creation)\n          || checkValidNumber(params.usageMetadata.input_token_details.cache_read)\n        )\n      )\n    ) && checkValidNumber(params.usageMetadata.output_tokens)) {\n      currentUsage = calculateTotalTokens(params.usageMetadata);\n      totalTokens = currentUsage.total_tokens;\n    }\n\n    for (let i = lastTurnStartIndex; i < params.messages.length; i++) {\n      const message = params.messages[i];\n      // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition\n      if (i === lastTurnStartIndex && indexTokenCountMap[i] === undefined && currentUsage) {\n        indexTokenCountMap[i] = currentUsage.output_tokens;\n      // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition\n      } else if (indexTokenCountMap[i] === undefined) {\n        indexTokenCountMap[i] = factoryParams.tokenCounter(message);\n        totalTokens += indexTokenCountMap[i];\n      }\n    }\n\n    // If `currentUsage` is defined, we need to distribute the current total tokens to our `indexTokenCountMap`,\n    // We must distribute it in a weighted manner, so that the total token count is equal to `currentUsage.total_tokens`,\n    // relative the manually counted tokens in `indexTokenCountMap`.\n    // EDGE CASE: when the resulting context gets pruned, we should not distribute the usage for messages that are not in the context.\n    if (currentUsage) {\n      // Calculate the sum of tokens only for indices at or after lastCutOffIndex\n      const totalIndexTokens = Object.entries(indexTokenCountMap).reduce((sum, [key, value]) => {\n        // Convert string key to number and check if it's >= lastCutOffIndex\n        const numericKey = Number(key);\n        if (numericKey === 0 && params.messages[0].getType() === 'system') {\n          return sum + value;\n        }\n        return numericKey >= lastCutOffIndex ? sum + value : sum;\n      }, 0);\n\n      // Calculate ratio based only on messages that remain in the context\n      const ratio = currentUsage.total_tokens / totalIndexTokens;\n      const isRatioSafe = ratio >= 1/3 && ratio <= 2.5;\n\n      // Apply the ratio adjustment only to messages at or after lastCutOffIndex, and only if the ratio is safe\n      if (isRatioSafe) {\n        for (const key in indexTokenCountMap) {\n          const numericKey = Number(key);\n          if (numericKey === 0 && params.messages[0].getType() === 'system') {\n            indexTokenCountMap[key] = Math.round(indexTokenCountMap[key] * ratio);\n          } else if (numericKey >= lastCutOffIndex) {\n            // Only adjust token counts for messages still in the context\n            indexTokenCountMap[key] = Math.round(indexTokenCountMap[key] * ratio);\n          }\n        }\n      }\n    }\n\n    lastTurnStartIndex = params.messages.length;\n    if (totalTokens <= factoryParams.maxTokens) {\n      return { context: params.messages, indexTokenCountMap };\n    }\n\n    const { context, thinkingStartIndex } = getMessagesWithinTokenLimit({\n      maxContextTokens: factoryParams.maxTokens,\n      messages: params.messages,\n      indexTokenCountMap,\n      startType: params.startType,\n      thinkingEnabled: factoryParams.thinkingEnabled,\n      tokenCounter: factoryParams.tokenCounter,\n      reasoningType: factoryParams.provider === Providers.BEDROCK ? ContentTypes.REASONING_CONTENT : ContentTypes.THINKING,\n      thinkingStartIndex: factoryParams.thinkingEnabled === true ? runThinkingStartIndex : undefined,\n    });\n    runThinkingStartIndex = thinkingStartIndex ?? -1;\n    /** The index is the first value of `context`, index relative to `params.messages` */\n    lastCutOffIndex = Math.max(params.messages.length - (context.length - (context[0]?.getType() === 'system' ? 1 : 0)), 0);\n\n    return { context, indexTokenCountMap };\n  };\n}\n"],"names":["ContentTypes","AIMessage","messages","Providers"],"mappings":";;;;;AAmBA,SAAS,gBAAgB,CAAC,MAAiB,EAAE,MAAiB,EAAE,WAAmB,EAAA;IACjF,MAAM,gBAAgB,GAAG,MAAM,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM;IACtD,OAAO,WAAW,IAAI,gBAAgB;AACxC;AAEA,SAAS,gBAAgB,CAAC,OAAkB,EAAE,aAAyD,EAAA;IACrG,MAAM,OAAO,GAA4B,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO;UAClE,OAAO,CAAC;AACV,UAAE,CAAC;gBACD,IAAI,EAAEA,kBAAY,CAAC,IAAI;gBACvB,IAAI,EAAE,OAAO,CAAC,OAAO;AACtB,aAAA,CAAC;;IAEJ,IAAI,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,KAAK,aAAa,CAAC,IAAI,EAAE;AAC1C,QAAA,OAAO,OAAO;;AAEhB,IAAA,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;IAC9B,OAAO,IAAIC,kBAAS,CAAC;AACnB,QAAA,GAAG,OAAO;QACV;AACD,KAAA,CAAC;AACJ;AAEA;;;;;AAKG;AACG,SAAU,oBAAoB,CAAC,KAA6B,EAAA;IAChE,MAAM,eAAe,GAAG,MAAM,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC;AACvD,IAAA,MAAM,aAAa,GAAG,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,cAAc,CAAC,IAAI,CAAC;AAC5E,IAAA,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,UAAU,CAAC,IAAI,CAAC;AAEpE,IAAA,MAAM,gBAAgB,GAAG,eAAe,GAAG,aAAa,GAAG,SAAS;IACpE,MAAM,iBAAiB,GAAG,MAAM,CAAC,KAAK,CAAC,aAAa,CAAC,IAAI,CAAC;IAE1D,OAAO;AACL,QAAA,YAAY,EAAE,gBAAgB;AAC9B,QAAA,aAAa,EAAE,iBAAiB;QAChC,YAAY,EAAE,gBAAgB,GAAG;KAClC;AACH;AASA;;;;;;AAMG;AACa,SAAA,2BAA2B,CAAC,EAC1C,QAAQ,EAAE,SAAS,EACnB,gBAAgB,EAChB,kBAAkB,EAClB,SAAS,EAAE,UAAU,EACrB,eAAe,EACf,YAAY,EACZ,kBAAkB,EAAE,mBAAmB,GAAG,EAAE,EAC5C,aAAa,GAAGD,kBAAY,CAAC,QAAQ,GAUtC,EAAA;;;IAGC,IAAI,iBAAiB,GAAG,CAAC;IACzB,MAAM,YAAY,GAAG,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,QAAQ,GAAG,SAAS,CAAC,CAAC,CAAC,GAAG,SAAS;AACpF,IAAA,MAAM,sBAAsB,GAAG,YAAY,IAAI,IAAI,GAAG,kBAAkB,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC;AACpF,IAAA,MAAM,oBAAoB,GAAG,gBAAgB,GAAG,sBAAsB;IACtE,IAAI,sBAAsB,GAAG,oBAAoB;IACjD,IAAI,SAAS,GAAG,UAAU;AAC1B,IAAA,MAAM,cAAc,GAAG,SAAS,CAAC,MAAM;AACvC,IAAA,MAAME,UAAQ,GAAG,CAAC,GAAG,SAAS,CAAC;AAC/B;;;;AAIK;IACL,IAAI,OAAO,GAAmC,EAAE;IAEhD,IAAI,kBAAkB,GAAG,mBAAmB;AAC5C,IAAA,IAAI,gBAAgB,GAAG,EAAE;AACzB,IAAA,IAAI,aAAqE;AACzE,IAAA,MAAM,QAAQ,GAAG,YAAY,IAAI,IAAI,GAAG,CAAC,GAAG,CAAC;IAC7C,MAAM,YAAY,GAAkB,EAAE;AAEtC,IAAA,IAAI,mBAAmB,GAAG,EAAE,EAAE;QAC5B,MAAM,sBAAsB,GAAGA,UAAQ,CAAC,mBAAmB,CAAC,EAAE,OAAO;AACrE,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,sBAAsB,CAAC,EAAE;AACzC,YAAA,aAAa,GAAG,sBAAsB,CAAC,IAAI,CAAC,CAAC,OAAO,KAAK,OAAO,CAAC,IAAI,KAAK,aAAa,CAAoC;;;AAI/H,IAAA,IAAI,iBAAiB,GAAG,sBAAsB,EAAE;AAC9C,QAAA,IAAI,YAAY,GAAGA,UAAQ,CAAC,MAAM;AAClC,QAAA,OAAOA,UAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,GAAG,sBAAsB,IAAI,YAAY,GAAG,QAAQ,EAAE;AACnG,YAAA,YAAY,EAAE;YACd,IAAIA,UAAQ,CAAC,MAAM,KAAK,CAAC,IAAI,YAAY,EAAE;gBACzC;;AAEF,YAAA,MAAM,aAAa,GAAGA,UAAQ,CAAC,GAAG,EAAE;AACpC,YAAA,IAAI,CAAC,aAAa;gBAAE;AACpB,YAAA,MAAM,WAAW,GAAG,aAAa,CAAC,OAAO,EAAE;AAC3C,YAAA,IAAI,eAAe,KAAK,IAAI,IAAI,gBAAgB,KAAK,EAAE,KAAK,YAAY,MAAM,cAAc,GAAG,CAAC,CAAC,CAAC,KAAK,WAAW,KAAK,IAAI,IAAI,WAAW,KAAK,MAAM,CAAC,EAAE;gBACtJ,gBAAgB,GAAG,YAAY;;YAEjC,IAAI,gBAAgB,GAAG,EAAE,IAAI,CAAC,aAAa,IAAK,kBAAkB,GAAG,CAAC,IAAI,WAAW,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,aAAa,CAAC,OAAO,CAAC,EAAE;gBACtI,aAAa,IAAI,aAAa,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,OAAO,KAAK,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAoC;AAC5H,gBAAA,kBAAkB,GAAG,aAAa,IAAI,IAAI,GAAG,YAAY,GAAG,EAAE;;;YAGhE,IACE,gBAAgB,GAAG;AAChB,mBAAA,YAAY,MAAM,gBAAgB,GAAG,CAAC;oBACrC,WAAW,KAAK,IAAI,IAAI,WAAW,KAAK,MAAM,CAAC,EACnD;gBACA,gBAAgB,GAAG,EAAE;;YAGvB,MAAM,UAAU,GAAG,kBAAkB,CAAC,YAAY,CAAC,IAAI,CAAC;AAExD,YAAA,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC,KAAK,CAAC,iBAAiB,GAAG,UAAU,KAAK,sBAAsB,CAAC,EAAE;AAC7F,gBAAA,OAAO,CAAC,IAAI,CAAC,aAAa,CAAC;gBAC3B,iBAAiB,IAAI,UAAU;;iBAC1B;AACL,gBAAA,YAAY,CAAC,IAAI,CAAC,aAAa,CAAC;gBAChC,IAAI,gBAAgB,GAAG,EAAE,IAAI,kBAAkB,GAAG,CAAC,EAAE;oBACnD;;gBAEF;;;AAIJ,QAAA,IAAI,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,MAAM,EAAE;AACrD,YAAA,SAAS,GAAG,CAAC,IAAI,EAAE,OAAO,CAAC;;AAG7B,QAAA,IAAI,SAAS,IAAI,IAAI,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE;AACnE,YAAA,IAAI,iBAAiB,GAAG,EAAE;YAE1B,IAAI,WAAW,GAAG,CAAC;AACnB,YAAA,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE;gBAC5C,MAAM,WAAW,GAAG,OAAO,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE;gBAC/C,IAAI,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,GAAG,SAAS,CAAC,QAAQ,CAAC,WAAW,CAAC,GAAG,WAAW,KAAK,SAAS,EAAE;AAC1F,oBAAA,iBAAiB,GAAG,CAAC,GAAG,CAAC;oBACzB;;AAEF,gBAAA,MAAM,aAAa,GAAG,cAAc,GAAG,CAAC,GAAG,CAAC;AAC5C,gBAAA,WAAW,IAAI,kBAAkB,CAAC,aAAa,CAAC,IAAI,CAAC;;AAGvD,YAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;gBACzB,iBAAiB,IAAI,WAAW;gBAChC,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,iBAAiB,CAAC;;;;AAKnD,IAAA,IAAI,YAAY,IAAI,cAAc,GAAG,CAAC,EAAE;QACtC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAgB,CAAC;QACzCA,UAAQ,CAAC,KAAK,EAAE;;IAGlB,sBAAsB,IAAI,iBAAiB;AAC3C,IAAA,MAAM,MAAM,GAAkB;QAC5B,sBAAsB;AACtB,QAAA,OAAO,EAAE,EAAmB;AAC5B,QAAA,gBAAgB,EAAE,YAAY;KAC/B;AAED,IAAA,IAAI,kBAAkB,GAAG,EAAE,EAAE;AAC3B,QAAA,MAAM,CAAC,kBAAkB,GAAG,kBAAkB;;IAGhD,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC,IAAI,gBAAgB,GAAG,CAAC,KAAK,kBAAkB,GAAG,EAAE,IAAI,gBAAgB,CAAC,SAAS,EAAE,OAAO,EAAE,kBAAkB,CAAC,CAAC,EAAE;;AAE9I,QAAA,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,EAAmB;AACnD,QAAA,OAAO,MAAM;;IAGf,IAAI,gBAAgB,GAAG,EAAE,IAAI,kBAAkB,GAAG,CAAC,EAAE;AACnD,QAAA,MAAM,IAAI,KAAK,CAAC,mGAAmG,CAAC;;IAGtH,IAAI,CAAC,aAAa,EAAE;AAClB,QAAA,MAAM,IAAI,KAAK,CAAC,qFAAqF,CAAC;;;;;AAMxG,IAAA,IAAI,cAAc,GAAG,EAAE;AACvB,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;AACvC,QAAA,MAAM,cAAc,GAAG,OAAO,CAAC,CAAC,CAAC;AACjC,QAAA,MAAM,IAAI,GAAG,cAAc,EAAE,OAAO,EAAE;AACtC,QAAA,IAAI,IAAI,KAAK,IAAI,EAAE;YACjB,cAAc,GAAG,CAAC;;AAEpB,QAAA,IAAI,cAAc,GAAG,EAAE,KAAK,IAAI,KAAK,OAAO,IAAI,IAAI,KAAK,QAAQ,CAAC,EAAE;YAClE;;;AAIJ,IAAA,IAAI,cAAc,KAAK,EAAE,EAAE;AACzB,QAAA,MAAM,IAAI,KAAK,CAAC,2GAA2G,CAAC;;AAG9H,IAAA,kBAAkB,GAAG,cAAc,GAAG,CAAC,GAAG,cAAc;AACxD,IAAA,MAAM,kBAAkB,GAAG,YAAY,CAAC,IAAID,kBAAS,CAAC,EAAE,OAAO,EAAE,CAAC,aAAa,CAAC,EAAE,CAAC,CAAC;AACpF,IAAA,MAAM,iBAAiB,GAAG,sBAAsB,GAAG,kBAAkB;IACrE,MAAM,UAAU,GAAG,gBAAgB,CAAC,OAAO,CAAC,cAAc,CAAc,EAAE,aAAa,CAAC;AACxF,IAAA,OAAO,CAAC,cAAc,CAAC,GAAG,UAAU;AACpC,IAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;AACzB,QAAA,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,EAAmB;AACnD,QAAA,OAAO,MAAM;;AAGf,IAAA,MAAM,eAAe,GAAc,OAAO,CAAC,cAAc,CAAc;;AAEvE,IAAA,MAAM,4BAA4B,GAAG,CAAC,kBAAkB,CAAC,kBAAkB,CAAC,IAAI,CAAC,IAAI,kBAAkB;AACvG,IAAA,sBAAsB,GAAG,oBAAoB,GAAG,4BAA4B;IAC5E,iBAAiB,GAAG,CAAC;IACrB,IAAI,UAAU,GAAkB,EAAE;AAClC,IAAA,MAAM,mBAAmB,GAAG,CAAC,GAAG,SAAS,CAAC;AAC1C,IAAA,IAAI,YAAY,GAAG,mBAAmB,CAAC,MAAM;AAC7C,IAAA,OAAO,mBAAmB,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,GAAG,sBAAsB,IAAI,YAAY,GAAG,kBAAkB,EAAE;AACxH,QAAA,YAAY,EAAE;AACd,QAAA,MAAM,aAAa,GAAG,mBAAmB,CAAC,GAAG,EAAE;AAC/C,QAAA,IAAI,CAAC,aAAa;YAAE;QACpB,MAAM,UAAU,GAAG,kBAAkB,CAAC,YAAY,CAAC,IAAI,CAAC;QACxD,IAAI,CAAC,iBAAiB,GAAG,UAAU,KAAK,sBAAsB,EAAE;AAC9D,YAAA,UAAU,CAAC,IAAI,CAAC,aAAa,CAAC;YAC9B,iBAAiB,IAAI,UAAU;;aAC1B;AACL,YAAAC,UAAQ,CAAC,IAAI,CAAC,aAAa,CAAC;YAC5B;;;IAIJ,MAAM,YAAY,GAAc,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;AACjE,IAAA,MAAM,gBAAgB,GAAG,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,OAAO,EAAE;AACpE,IAAA,IAAI,gBAAgB,KAAK,MAAM,EAAE;AAC/B,QAAA,SAAS,GAAG,CAAC,IAAI,EAAE,OAAO,CAAC;;AAG7B,IAAA,IAAI,SAAS,IAAI,IAAI,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;AACtE,QAAA,IAAI,iBAAiB,GAAG,EAAE;QAE1B,IAAI,WAAW,GAAG,CAAC;AACnB,QAAA,KAAK,IAAI,CAAC,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE;YAC/C,MAAM,WAAW,GAAG,UAAU,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE;YAClD,IAAI,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,GAAG,SAAS,CAAC,QAAQ,CAAC,WAAW,CAAC,GAAG,WAAW,KAAK,SAAS,EAAE;AAC1F,gBAAA,iBAAiB,GAAG,CAAC,GAAG,CAAC;gBACzB;;AAEF,YAAA,MAAM,aAAa,GAAG,cAAc,GAAG,CAAC,GAAG,CAAC;AAC5C,YAAA,WAAW,IAAI,kBAAkB,CAAC,aAAa,CAAC,IAAI,CAAC;;AAGvD,QAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;YACzB,iBAAiB,IAAI,WAAW;YAChC,UAAU,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,iBAAiB,CAAC;;;AAIvD,IAAA,IAAI,gBAAgB,KAAK,IAAI,EAAE;QAC7B,MAAM,UAAU,GAAG,gBAAgB,CAAC,YAAY,EAAE,aAAa,CAAC;QAChE,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,GAAG,UAAU;;SACzC;AACL,QAAA,UAAU,CAAC,IAAI,CAAC,eAAe,CAAC;;AAGlC,IAAA,IAAI,YAAY,IAAI,cAAc,GAAG,CAAC,EAAE;QACtC,UAAU,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAgB,CAAC;QAC5C,mBAAmB,CAAC,KAAK,EAAE;;AAG7B,IAAA,MAAM,CAAC,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE;AACrC,IAAA,OAAO,MAAM;AACf;AAEM,SAAU,gBAAgB,CAAC,KAAc,EAAA;AAC7C,IAAA,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC;AAChE;AAEM,SAAU,mBAAmB,CAAC,aAAyC,EAAA;IAC3E,MAAM,kBAAkB,GAAG,EAAE,GAAG,aAAa,CAAC,kBAAkB,EAAE;AAClE,IAAA,IAAI,kBAAkB,GAAG,aAAa,CAAC,UAAU;IACjD,IAAI,eAAe,GAAG,CAAC;IACvB,IAAI,WAAW,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,kBAAkB,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;AAChF,IAAA,IAAI,qBAAqB,GAAG,EAAE;IAC9B,OAAO,SAAS,aAAa,CAAC,MAA2B,EAAA;AAIvD,QAAA,IAAI,YAAuC;AAC3C,QAAA,IAAI,MAAM,CAAC,aAAa,KACtB,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,YAAY;AAC/C,gBACD,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB;oBAEvD,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB,CAAC,cAAc;uBACrE,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB,CAAC,UAAU,CAAC,CACzE,CACF,CACF,IAAI,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,aAAa,CAAC,EAAE;AACzD,YAAA,YAAY,GAAG,oBAAoB,CAAC,MAAM,CAAC,aAAa,CAAC;AACzD,YAAA,WAAW,GAAG,YAAY,CAAC,YAAY;;AAGzC,QAAA,KAAK,IAAI,CAAC,GAAG,kBAAkB,EAAE,CAAC,GAAG,MAAM,CAAC,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;YAChE,MAAM,OAAO,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC;;AAElC,YAAA,IAAI,CAAC,KAAK,kBAAkB,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,SAAS,IAAI,YAAY,EAAE;AACnF,gBAAA,kBAAkB,CAAC,CAAC,CAAC,GAAG,YAAY,CAAC,aAAa;;;AAE7C,iBAAA,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,SAAS,EAAE;gBAC9C,kBAAkB,CAAC,CAAC,CAAC,GAAG,aAAa,CAAC,YAAY,CAAC,OAAO,CAAC;AAC3D,gBAAA,WAAW,IAAI,kBAAkB,CAAC,CAAC,CAAC;;;;;;;QAQxC,IAAI,YAAY,EAAE;;YAEhB,MAAM,gBAAgB,GAAG,MAAM,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,KAAK,CAAC,KAAI;;AAEvF,gBAAA,MAAM,UAAU,GAAG,MAAM,CAAC,GAAG,CAAC;AAC9B,gBAAA,IAAI,UAAU,KAAK,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,KAAK,QAAQ,EAAE;oBACjE,OAAO,GAAG,GAAG,KAAK;;AAEpB,gBAAA,OAAO,UAAU,IAAI,eAAe,GAAG,GAAG,GAAG,KAAK,GAAG,GAAG;aACzD,EAAE,CAAC,CAAC;;AAGL,YAAA,MAAM,KAAK,GAAG,YAAY,CAAC,YAAY,GAAG,gBAAgB;YAC1D,MAAM,WAAW,GAAG,KAAK,IAAI,CAAC,GAAC,CAAC,IAAI,KAAK,IAAI,GAAG;;YAGhD,IAAI,WAAW,EAAE;AACf,gBAAA,KAAK,MAAM,GAAG,IAAI,kBAAkB,EAAE;AACpC,oBAAA,MAAM,UAAU,GAAG,MAAM,CAAC,GAAG,CAAC;AAC9B,oBAAA,IAAI,UAAU,KAAK,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,KAAK,QAAQ,EAAE;AACjE,wBAAA,kBAAkB,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;;AAChE,yBAAA,IAAI,UAAU,IAAI,eAAe,EAAE;;AAExC,wBAAA,kBAAkB,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;;;;;AAM7E,QAAA,kBAAkB,GAAG,MAAM,CAAC,QAAQ,CAAC,MAAM;AAC3C,QAAA,IAAI,WAAW,IAAI,aAAa,CAAC,SAAS,EAAE;YAC1C,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,QAAQ,EAAE,kBAAkB,EAAE;;AAGzD,QAAA,MAAM,EAAE,OAAO,EAAE,kBAAkB,EAAE,GAAG,2BAA2B,CAAC;YAClE,gBAAgB,EAAE,aAAa,CAAC,SAAS;YACzC,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,kBAAkB;YAClB,SAAS,EAAE,MAAM,CAAC,SAAS;YAC3B,eAAe,EAAE,aAAa,CAAC,eAAe;YAC9C,YAAY,EAAE,aAAa,CAAC,YAAY;AACxC,YAAA,aAAa,EAAE,aAAa,CAAC,QAAQ,KAAKC,eAAS,CAAC,OAAO,GAAGH,kBAAY,CAAC,iBAAiB,GAAGA,kBAAY,CAAC,QAAQ;AACpH,YAAA,kBAAkB,EAAE,aAAa,CAAC,eAAe,KAAK,IAAI,GAAG,qBAAqB,GAAG,SAAS;AAC/F,SAAA,CAAC;AACF,QAAA,qBAAqB,GAAG,kBAAkB,IAAI,EAAE;;AAEhD,QAAA,eAAe,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,IAAI,OAAO,CAAC,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,QAAQ,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AAEvH,QAAA,OAAO,EAAE,OAAO,EAAE,kBAAkB,EAAE;AACxC,KAAC;AACH;;;;;;;"}

package/dist/esm/messages/prune.mjs

@@ -12,8 +12,15 @@ function addThinkingBlock(message, thinkingBlock) {
                 type: ContentTypes.TEXT,
                 text: message.content,
             }];
+    /** Edge case, the message already has the thinking block */
+    if (content[0].type === thinkingBlock.type) {
+        return message;
+    }
     content.unshift(thinkingBlock);
-    return content;
+    return new AIMessage({
+        ...message,
+        content
+    });
 }
 /**
  * Calculates the total tokens from a single usage object
@@ -40,7 +47,7 @@ function calculateTotalTokens(usage) {
  * @param options Configuration options for processing messages
  * @returns Object containing the message context, remaining tokens, messages not included, and summary index
  */
-function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, indexTokenCountMap, startType: _startType, thinkingEnabled, tokenCounter, reasoningType = ContentTypes.THINKING, }) {
+function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, indexTokenCountMap, startType: _startType, thinkingEnabled, tokenCounter, thinkingStartIndex: _thinkingStartIndex = -1, reasoningType = ContentTypes.THINKING, }) {
     // Every reply is primed with <|start|>assistant<|message|>, so we
     // start with 3 tokens for the label after all messages have been counted.
     let currentTokenCount = 3;
@@ -57,11 +64,17 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
      * This may be confusing to read, but it is done to ensure the context is in the correct order for the model.
      * */
     let context = [];
-    let thinkingStartIndex = -1;
+    let thinkingStartIndex = _thinkingStartIndex;
     let thinkingEndIndex = -1;
     let thinkingBlock;
     const endIndex = instructions != null ? 1 : 0;
     const prunedMemory = [];
+    if (_thinkingStartIndex > -1) {
+        const thinkingMessageContent = messages[_thinkingStartIndex]?.content;
+        if (Array.isArray(thinkingMessageContent)) {
+            thinkingBlock = thinkingMessageContent.find((content) => content.type === reasoningType);
+        }
+    }
     if (currentTokenCount < remainingContextTokens) {
         let currentIndex = messages.length;
         while (messages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > endIndex) {
@@ -93,19 +106,30 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
             }
             else {
                 prunedMemory.push(poppedMessage);
-                if (thinkingEndIndex > -1) {
+                if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {
                     continue;
                 }
                 break;
             }
         }
-        if (thinkingEndIndex > -1 && context[context.length - 1]?.getType() === 'tool') {
-            startType = 'ai';
+        if (context[context.length - 1]?.getType() === 'tool') {
+            startType = ['ai', 'human'];
         }
-        if (startType != null && startType && context.length > 0) {
-            const requiredTypeIndex = context.findIndex(msg => msg?.getType() === startType);
+        if (startType != null && startType.length > 0 && context.length > 0) {
+            let requiredTypeIndex = -1;
+            let totalTokens = 0;
+            for (let i = context.length - 1; i >= 0; i--) {
+                const currentType = context[i]?.getType() ?? '';
+                if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {
+                    requiredTypeIndex = i + 1;
+                    break;
+                }
+                const originalIndex = originalLength - 1 - i;
+                totalTokens += indexTokenCountMap[originalIndex] ?? 0;
+            }
             if (requiredTypeIndex > 0) {
-                context = context.slice(requiredTypeIndex);
+                currentTokenCount -= totalTokens;
+                context = context.slice(0, requiredTypeIndex);
             }
         }
     }
@@ -119,6 +143,9 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
         context: [],
         messagesToRefine: prunedMemory,
     };
+    if (thinkingStartIndex > -1) {
+        result.thinkingStartIndex = thinkingStartIndex;
+    }
     if (prunedMemory.length === 0 || thinkingEndIndex < 0 || (thinkingStartIndex > -1 && isIndexInContext(_messages, context, thinkingStartIndex))) {
         // we reverse at this step to ensure the context is in the correct order for the model, and we need to work backwards
         result.context = context.reverse();
@@ -150,8 +177,8 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
     thinkingStartIndex = originalLength - 1 - assistantIndex;
     const thinkingTokenCount = tokenCounter(new AIMessage({ content: [thinkingBlock] }));
     const newRemainingCount = remainingContextTokens - thinkingTokenCount;
-    const content = addThinkingBlock(context[assistantIndex], thinkingBlock);
-    context[assistantIndex].content = content;
+    const newMessage = addThinkingBlock(context[assistantIndex], thinkingBlock);
+    context[assistantIndex] = newMessage;
     if (newRemainingCount > 0) {
         result.context = context.reverse();
         return result;
@@ -182,17 +209,28 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
     const firstMessage = newContext[newContext.length - 1];
     const firstMessageType = newContext[newContext.length - 1].getType();
     if (firstMessageType === 'tool') {
-        startType = 'ai';
+        startType = ['ai', 'human'];
     }
-    if (startType != null && startType && newContext.length > 0) {
-        const requiredTypeIndex = newContext.findIndex(msg => msg.getType() === startType);
+    if (startType != null && startType.length > 0 && newContext.length > 0) {
+        let requiredTypeIndex = -1;
+        let totalTokens = 0;
+        for (let i = newContext.length - 1; i >= 0; i--) {
+            const currentType = newContext[i]?.getType() ?? '';
+            if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {
+                requiredTypeIndex = i + 1;
+                break;
+            }
+            const originalIndex = originalLength - 1 - i;
+            totalTokens += indexTokenCountMap[originalIndex] ?? 0;
+        }
         if (requiredTypeIndex > 0) {
-            newContext = newContext.slice(requiredTypeIndex);
+            currentTokenCount -= totalTokens;
+            newContext = newContext.slice(0, requiredTypeIndex);
         }
     }
     if (firstMessageType === 'ai') {
-        const content = addThinkingBlock(firstMessage, thinkingBlock);
-        newContext[newContext.length - 1].content = content;
+        const newMessage = addThinkingBlock(firstMessage, thinkingBlock);
+        newContext[newContext.length - 1] = newMessage;
     }
     else {
         newContext.push(thinkingMessage);
@@ -212,6 +250,7 @@ function createPruneMessages(factoryParams) {
     let lastTurnStartIndex = factoryParams.startIndex;
     let lastCutOffIndex = 0;
     let totalTokens = (Object.values(indexTokenCountMap)).reduce((a, b) => a + b, 0);
+    let runThinkingStartIndex = -1;
     return function pruneMessages(params) {
         let currentUsage;
         if (params.usageMetadata && (checkValidNumber(params.usageMetadata.input_tokens)
@@ -268,7 +307,7 @@ function createPruneMessages(factoryParams) {
         if (totalTokens <= factoryParams.maxTokens) {
             return { context: params.messages, indexTokenCountMap };
         }
-        const { context } = getMessagesWithinTokenLimit({
+        const { context, thinkingStartIndex } = getMessagesWithinTokenLimit({
             maxContextTokens: factoryParams.maxTokens,
             messages: params.messages,
             indexTokenCountMap,
@@ -276,8 +315,11 @@ function createPruneMessages(factoryParams) {
             thinkingEnabled: factoryParams.thinkingEnabled,
             tokenCounter: factoryParams.tokenCounter,
             reasoningType: factoryParams.provider === Providers.BEDROCK ? ContentTypes.REASONING_CONTENT : ContentTypes.THINKING,
+            thinkingStartIndex: factoryParams.thinkingEnabled === true ? runThinkingStartIndex : undefined,
         });
-        lastCutOffIndex = Math.max(params.messages.length - context.length, 0);
+        runThinkingStartIndex = thinkingStartIndex ?? -1;
+        /** The index is the first value of `context`, index relative to `params.messages` */
+        lastCutOffIndex = Math.max(params.messages.length - (context.length - (context[0]?.getType() === 'system' ? 1 : 0)), 0);
         return { context, indexTokenCountMap };
     };
 }

package/dist/esm/messages/prune.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"prune.mjs","sources":["../../../src/messages/prune.ts"],"sourcesContent":["import { AIMessage, BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { ThinkingContentText, MessageContentComplex, ReasoningContentText } from '@/types/stream';\nimport type { TokenCounter } from '@/types/run';\nimport { ContentTypes, Providers } from '@/common';\n\nexport type PruneMessagesFactoryParams = {\n  provider?: Providers;\n  maxTokens: number;\n  startIndex: number;\n  tokenCounter: TokenCounter;\n  indexTokenCountMap: Record<string, number>;\n  thinkingEnabled?: boolean;\n};\nexport type PruneMessagesParams = {\n  messages: BaseMessage[];\n  usageMetadata?: Partial<UsageMetadata>;\n  startType?: ReturnType<BaseMessage['getType']>;\n}\n\nfunction isIndexInContext(arrayA: unknown[], arrayB: unknown[], targetIndex: number): boolean {\n  const startingIndexInA = arrayA.length - arrayB.length;\n  return targetIndex >= startingIndexInA;\n}\n\nfunction addThinkingBlock(message: AIMessage, thinkingBlock: ThinkingContentText | ReasoningContentText): MessageContentComplex[] {\n  const content: MessageContentComplex[] = Array.isArray(message.content)\n    ? message.content as MessageContentComplex[]\n    : [{\n      type: ContentTypes.TEXT,\n      text: message.content,\n    }];\n  content.unshift(thinkingBlock);\n  return content;\n}\n\n/**\n * Calculates the total tokens from a single usage object\n *\n * @param usage The usage metadata object containing token information\n * @returns An object containing the total input and output tokens\n */\nexport function calculateTotalTokens(usage: Partial<UsageMetadata>): UsageMetadata {\n  const baseInputTokens = Number(usage.input_tokens) || 0;\n  const cacheCreation = Number(usage.input_token_details?.cache_creation) || 0;\n  const cacheRead = Number(usage.input_token_details?.cache_read) || 0;\n\n  const totalInputTokens = baseInputTokens + cacheCreation + cacheRead;\n  const totalOutputTokens = Number(usage.output_tokens) || 0;\n\n  return {\n    input_tokens: totalInputTokens,\n    output_tokens: totalOutputTokens,\n    total_tokens: totalInputTokens + totalOutputTokens\n  };\n}\n\n/**\n * Processes an array of messages and returns a context of messages that fit within a specified token limit.\n * It iterates over the messages from newest to oldest, adding them to the context until the token limit is reached.\n *\n * @param options Configuration options for processing messages\n * @returns Object containing the message context, remaining tokens, messages not included, and summary index\n */\nexport function getMessagesWithinTokenLimit({\n  messages: _messages,\n  maxContextTokens,\n  indexTokenCountMap,\n  startType: _startType,\n  thinkingEnabled,\n  tokenCounter,\n  reasoningType = ContentTypes.THINKING,\n}: {\n  messages: BaseMessage[];\n  maxContextTokens: number;\n  indexTokenCountMap: Record<string, number | undefined>;\n  tokenCounter: TokenCounter;\n  startType?: string;\n  thinkingEnabled?: boolean;\n  reasoningType?: ContentTypes.THINKING | ContentTypes.REASONING_CONTENT;\n}): {\n  context: BaseMessage[];\n  remainingContextTokens: number;\n  messagesToRefine: BaseMessage[];\n} {\n  // Every reply is primed with <|start|>assistant<|message|>, so we\n  // start with 3 tokens for the label after all messages have been counted.\n  let currentTokenCount = 3;\n  const instructions = _messages[0]?.getType() === 'system' ? _messages[0] : undefined;\n  const instructionsTokenCount = instructions != null ? indexTokenCountMap[0] ?? 0 : 0;\n  const initialContextTokens = maxContextTokens - instructionsTokenCount;\n  let remainingContextTokens = initialContextTokens;\n  let startType = _startType;\n  const originalLength = _messages.length;\n  const messages = [..._messages];\n  /**\n   * IMPORTANT: this context array gets reversed at the end, since the latest messages get pushed first.\n   *\n   * This may be confusing to read, but it is done to ensure the context is in the correct order for the model.\n   * */\n  let context: Array<BaseMessage | undefined> = [];\n\n  let thinkingStartIndex = -1;\n  let thinkingEndIndex = -1;\n  let thinkingBlock: ThinkingContentText | ReasoningContentText | undefined;\n  const endIndex = instructions != null ? 1 : 0;\n  const prunedMemory: BaseMessage[] = [];\n\n  if (currentTokenCount < remainingContextTokens) {\n    let currentIndex = messages.length;\n    while (messages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > endIndex) {\n      currentIndex--;\n      if (messages.length === 1 && instructions) {\n        break;\n      }\n      const poppedMessage = messages.pop();\n      if (!poppedMessage) continue;\n      const messageType = poppedMessage.getType();\n      if (thinkingEnabled === true && thinkingEndIndex === -1 && (currentIndex === (originalLength - 1)) && (messageType === 'ai' || messageType === 'tool')) {\n        thinkingEndIndex = currentIndex;\n      }\n      if (thinkingEndIndex > -1 && !thinkingBlock  && thinkingStartIndex < 0 && messageType === 'ai' && Array.isArray(poppedMessage.content)) {\n        thinkingBlock = (poppedMessage.content.find((content) => content.type === reasoningType)) as ThinkingContentText | undefined;\n        thinkingStartIndex = thinkingBlock != null ? currentIndex : -1;\n      }\n      /** False start, the latest message was not part of a multi-assistant/tool sequence of messages */\n      if (\n        thinkingEndIndex > -1\n        && currentIndex === (thinkingEndIndex - 1)\n        && (messageType !== 'ai' && messageType !== 'tool')\n      ) {\n        thinkingEndIndex = -1;\n      }\n\n      const tokenCount = indexTokenCountMap[currentIndex] ?? 0;\n\n      if (prunedMemory.length === 0 && ((currentTokenCount + tokenCount) <= remainingContextTokens)) {\n        context.push(poppedMessage);\n        currentTokenCount += tokenCount;\n      } else {\n        prunedMemory.push(poppedMessage);\n        if (thinkingEndIndex > -1) {\n          continue;\n        }\n        break;\n      }\n    }\n\n    if (thinkingEndIndex > -1 && context[context.length - 1]?.getType() === 'tool') {\n      startType = 'ai';\n    }\n\n    if (startType != null && startType && context.length > 0) {\n      const requiredTypeIndex = context.findIndex(msg => msg?.getType() === startType);\n\n      if (requiredTypeIndex > 0) {\n        context = context.slice(requiredTypeIndex);\n      }\n    }\n  }\n\n  if (instructions && originalLength > 0) {\n    context.push(_messages[0] as BaseMessage);\n    messages.shift();\n  }\n\n  remainingContextTokens -= currentTokenCount;\n  const result = {\n    remainingContextTokens,\n    context: [] as BaseMessage[],\n    messagesToRefine: prunedMemory,\n  };\n\n  if (prunedMemory.length === 0 || thinkingEndIndex < 0 || (thinkingStartIndex > -1 && isIndexInContext(_messages, context, thinkingStartIndex))) {\n    // we reverse at this step to ensure the context is in the correct order for the model, and we need to work backwards\n    result.context = context.reverse() as BaseMessage[];\n    return result;\n  }\n\n  if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no \"AI\" messages with thinking blocks.');\n  }\n\n  if (!thinkingBlock) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no thinking block found.');\n  }\n\n  // Since we have a thinking sequence, we need to find the last assistant message\n  // in the latest AI/tool sequence to add the thinking block that falls outside of the current context\n  // Latest messages are ordered first.\n  let assistantIndex = -1;\n  for (let i = 0; i < context.length; i++) {\n    const currentMessage = context[i];\n    const type = currentMessage?.getType();\n    if (type === 'ai') {\n      assistantIndex = i;\n    }\n    if (assistantIndex > -1 && (type === 'human' || type === 'system')) {\n      break;\n    }\n  }\n\n  if (assistantIndex === -1) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no \"AI\" messages to append thinking blocks to.');\n  }\n\n  thinkingStartIndex = originalLength - 1 - assistantIndex;\n  const thinkingTokenCount = tokenCounter(new AIMessage({ content: [thinkingBlock] }));\n  const newRemainingCount = remainingContextTokens - thinkingTokenCount;\n  const content: MessageContentComplex[] = addThinkingBlock(context[assistantIndex] as AIMessage, thinkingBlock);\n  (context[assistantIndex] as AIMessage).content = content;\n  if (newRemainingCount > 0) {\n    result.context = context.reverse() as BaseMessage[];\n    return result;\n  }\n\n  const thinkingMessage: AIMessage = context[assistantIndex] as AIMessage;\n  // now we need to an additional round of pruning but making the thinking block fit\n  const newThinkingMessageTokenCount = (indexTokenCountMap[thinkingStartIndex] ?? 0) + thinkingTokenCount;\n  remainingContextTokens = initialContextTokens - newThinkingMessageTokenCount;\n  currentTokenCount = 3;\n  let newContext: BaseMessage[] = [];\n  const secondRoundMessages = [..._messages];\n  let currentIndex = secondRoundMessages.length;\n  while (secondRoundMessages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > thinkingStartIndex) {\n    currentIndex--;\n    const poppedMessage = secondRoundMessages.pop();\n    if (!poppedMessage) continue;\n    const tokenCount = indexTokenCountMap[currentIndex] ?? 0;\n    if ((currentTokenCount + tokenCount) <= remainingContextTokens) {\n      newContext.push(poppedMessage);\n      currentTokenCount += tokenCount;\n    } else {\n      messages.push(poppedMessage);\n      break;\n    }\n  }\n\n  const firstMessage: AIMessage = newContext[newContext.length - 1];\n  const firstMessageType = newContext[newContext.length - 1].getType();\n  if (firstMessageType === 'tool') {\n    startType = 'ai';\n  }\n\n  if (startType != null && startType && newContext.length > 0) {\n    const requiredTypeIndex = newContext.findIndex(msg => msg.getType() === startType);\n    if (requiredTypeIndex > 0) {\n      newContext = newContext.slice(requiredTypeIndex);\n    }\n  }\n\n  if (firstMessageType === 'ai') {\n    const content = addThinkingBlock(firstMessage, thinkingBlock);\n    newContext[newContext.length - 1].content = content;\n  } else {\n    newContext.push(thinkingMessage);\n  }\n\n  if (instructions && originalLength > 0) {\n    newContext.push(_messages[0] as BaseMessage);\n    secondRoundMessages.shift();\n  }\n\n  result.context = newContext.reverse();\n  return result;\n}\n\nexport function checkValidNumber(value: unknown): value is number {\n  return typeof value === 'number' && !isNaN(value) && value > 0;\n}\n\nexport function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {\n  const indexTokenCountMap = { ...factoryParams.indexTokenCountMap };\n  let lastTurnStartIndex = factoryParams.startIndex;\n  let lastCutOffIndex = 0;\n  let totalTokens = (Object.values(indexTokenCountMap)).reduce((a, b) => a + b, 0);\n  return function pruneMessages(params: PruneMessagesParams): {\n    context: BaseMessage[];\n    indexTokenCountMap: Record<string, number>;\n  } {\n    let currentUsage: UsageMetadata | undefined;\n    if (params.usageMetadata && (\n      checkValidNumber(params.usageMetadata.input_tokens)\n      || (\n        checkValidNumber(params.usageMetadata.input_token_details)\n        && (\n          checkValidNumber(params.usageMetadata.input_token_details.cache_creation)\n          || checkValidNumber(params.usageMetadata.input_token_details.cache_read)\n        )\n      )\n    ) && checkValidNumber(params.usageMetadata.output_tokens)) {\n      currentUsage = calculateTotalTokens(params.usageMetadata);\n      totalTokens = currentUsage.total_tokens;\n    }\n\n    for (let i = lastTurnStartIndex; i < params.messages.length; i++) {\n      const message = params.messages[i];\n      // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition\n      if (i === lastTurnStartIndex && indexTokenCountMap[i] === undefined && currentUsage) {\n        indexTokenCountMap[i] = currentUsage.output_tokens;\n      // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition\n      } else if (indexTokenCountMap[i] === undefined) {\n        indexTokenCountMap[i] = factoryParams.tokenCounter(message);\n        totalTokens += indexTokenCountMap[i];\n      }\n    }\n\n    // If `currentUsage` is defined, we need to distribute the current total tokens to our `indexTokenCountMap`,\n    // We must distribute it in a weighted manner, so that the total token count is equal to `currentUsage.total_tokens`,\n    // relative the manually counted tokens in `indexTokenCountMap`.\n    // EDGE CASE: when the resulting context gets pruned, we should not distribute the usage for messages that are not in the context.\n    if (currentUsage) {\n      // Calculate the sum of tokens only for indices at or after lastCutOffIndex\n      const totalIndexTokens = Object.entries(indexTokenCountMap).reduce((sum, [key, value]) => {\n        // Convert string key to number and check if it's >= lastCutOffIndex\n        const numericKey = Number(key);\n        if (numericKey === 0 && params.messages[0].getType() === 'system') {\n          return sum + value;\n        }\n        return numericKey >= lastCutOffIndex ? sum + value : sum;\n      }, 0);\n\n      // Calculate ratio based only on messages that remain in the context\n      const ratio = currentUsage.total_tokens / totalIndexTokens;\n      const isRatioSafe = ratio >= 1/3 && ratio <= 2.5;\n\n      // Apply the ratio adjustment only to messages at or after lastCutOffIndex, and only if the ratio is safe\n      if (isRatioSafe) {\n        for (const key in indexTokenCountMap) {\n          const numericKey = Number(key);\n          if (numericKey === 0 && params.messages[0].getType() === 'system') {\n            indexTokenCountMap[key] = Math.round(indexTokenCountMap[key] * ratio);\n          } else if (numericKey >= lastCutOffIndex) {\n            // Only adjust token counts for messages still in the context\n            indexTokenCountMap[key] = Math.round(indexTokenCountMap[key] * ratio);\n          }\n        }\n      }\n    }\n\n    lastTurnStartIndex = params.messages.length;\n    if (totalTokens <= factoryParams.maxTokens) {\n      return { context: params.messages, indexTokenCountMap };\n    }\n\n    const { context } = getMessagesWithinTokenLimit({\n      maxContextTokens: factoryParams.maxTokens,\n      messages: params.messages,\n      indexTokenCountMap,\n      startType: params.startType,\n      thinkingEnabled: factoryParams.thinkingEnabled,\n      tokenCounter: factoryParams.tokenCounter,\n      reasoningType: factoryParams.provider === Providers.BEDROCK ? ContentTypes.REASONING_CONTENT : ContentTypes.THINKING,\n    });\n    lastCutOffIndex = Math.max(params.messages.length - context.length, 0);\n\n    return { context, indexTokenCountMap };\n  };\n}\n"],"names":[],"mappings":";;;AAmBA,SAAS,gBAAgB,CAAC,MAAiB,EAAE,MAAiB,EAAE,WAAmB,EAAA;IACjF,MAAM,gBAAgB,GAAG,MAAM,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM;IACtD,OAAO,WAAW,IAAI,gBAAgB;AACxC;AAEA,SAAS,gBAAgB,CAAC,OAAkB,EAAE,aAAyD,EAAA;IACrG,MAAM,OAAO,GAA4B,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO;UAClE,OAAO,CAAC;AACV,UAAE,CAAC;gBACD,IAAI,EAAE,YAAY,CAAC,IAAI;gBACvB,IAAI,EAAE,OAAO,CAAC,OAAO;AACtB,aAAA,CAAC;AACJ,IAAA,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;AAC9B,IAAA,OAAO,OAAO;AAChB;AAEA;;;;;AAKG;AACG,SAAU,oBAAoB,CAAC,KAA6B,EAAA;IAChE,MAAM,eAAe,GAAG,MAAM,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC;AACvD,IAAA,MAAM,aAAa,GAAG,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,cAAc,CAAC,IAAI,CAAC;AAC5E,IAAA,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,UAAU,CAAC,IAAI,CAAC;AAEpE,IAAA,MAAM,gBAAgB,GAAG,eAAe,GAAG,aAAa,GAAG,SAAS;IACpE,MAAM,iBAAiB,GAAG,MAAM,CAAC,KAAK,CAAC,aAAa,CAAC,IAAI,CAAC;IAE1D,OAAO;AACL,QAAA,YAAY,EAAE,gBAAgB;AAC9B,QAAA,aAAa,EAAE,iBAAiB;QAChC,YAAY,EAAE,gBAAgB,GAAG;KAClC;AACH;AAEA;;;;;;AAMG;AACG,SAAU,2BAA2B,CAAC,EAC1C,QAAQ,EAAE,SAAS,EACnB,gBAAgB,EAChB,kBAAkB,EAClB,SAAS,EAAE,UAAU,EACrB,eAAe,EACf,YAAY,EACZ,aAAa,GAAG,YAAY,CAAC,QAAQ,GAStC,EAAA;;;IAOC,IAAI,iBAAiB,GAAG,CAAC;IACzB,MAAM,YAAY,GAAG,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,QAAQ,GAAG,SAAS,CAAC,CAAC,CAAC,GAAG,SAAS;AACpF,IAAA,MAAM,sBAAsB,GAAG,YAAY,IAAI,IAAI,GAAG,kBAAkB,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC;AACpF,IAAA,MAAM,oBAAoB,GAAG,gBAAgB,GAAG,sBAAsB;IACtE,IAAI,sBAAsB,GAAG,oBAAoB;IACjD,IAAI,SAAS,GAAG,UAAU;AAC1B,IAAA,MAAM,cAAc,GAAG,SAAS,CAAC,MAAM;AACvC,IAAA,MAAM,QAAQ,GAAG,CAAC,GAAG,SAAS,CAAC;AAC/B;;;;AAIK;IACL,IAAI,OAAO,GAAmC,EAAE;AAEhD,IAAA,IAAI,kBAAkB,GAAG,EAAE;AAC3B,IAAA,IAAI,gBAAgB,GAAG,EAAE;AACzB,IAAA,IAAI,aAAqE;AACzE,IAAA,MAAM,QAAQ,GAAG,YAAY,IAAI,IAAI,GAAG,CAAC,GAAG,CAAC;IAC7C,MAAM,YAAY,GAAkB,EAAE;AAEtC,IAAA,IAAI,iBAAiB,GAAG,sBAAsB,EAAE;AAC9C,QAAA,IAAI,YAAY,GAAG,QAAQ,CAAC,MAAM;AAClC,QAAA,OAAO,QAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,GAAG,sBAAsB,IAAI,YAAY,GAAG,QAAQ,EAAE;AACnG,YAAA,YAAY,EAAE;YACd,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,IAAI,YAAY,EAAE;gBACzC;;AAEF,YAAA,MAAM,aAAa,GAAG,QAAQ,CAAC,GAAG,EAAE;AACpC,YAAA,IAAI,CAAC,aAAa;gBAAE;AACpB,YAAA,MAAM,WAAW,GAAG,aAAa,CAAC,OAAO,EAAE;AAC3C,YAAA,IAAI,eAAe,KAAK,IAAI,IAAI,gBAAgB,KAAK,EAAE,KAAK,YAAY,MAAM,cAAc,GAAG,CAAC,CAAC,CAAC,KAAK,WAAW,KAAK,IAAI,IAAI,WAAW,KAAK,MAAM,CAAC,EAAE;gBACtJ,gBAAgB,GAAG,YAAY;;YAEjC,IAAI,gBAAgB,GAAG,EAAE,IAAI,CAAC,aAAa,IAAK,kBAAkB,GAAG,CAAC,IAAI,WAAW,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,aAAa,CAAC,OAAO,CAAC,EAAE;gBACtI,aAAa,IAAI,aAAa,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,OAAO,KAAK,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAoC;AAC5H,gBAAA,kBAAkB,GAAG,aAAa,IAAI,IAAI,GAAG,YAAY,GAAG,EAAE;;;YAGhE,IACE,gBAAgB,GAAG;AAChB,mBAAA,YAAY,MAAM,gBAAgB,GAAG,CAAC;oBACrC,WAAW,KAAK,IAAI,IAAI,WAAW,KAAK,MAAM,CAAC,EACnD;gBACA,gBAAgB,GAAG,EAAE;;YAGvB,MAAM,UAAU,GAAG,kBAAkB,CAAC,YAAY,CAAC,IAAI,CAAC;AAExD,YAAA,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC,KAAK,CAAC,iBAAiB,GAAG,UAAU,KAAK,sBAAsB,CAAC,EAAE;AAC7F,gBAAA,OAAO,CAAC,IAAI,CAAC,aAAa,CAAC;gBAC3B,iBAAiB,IAAI,UAAU;;iBAC1B;AACL,gBAAA,YAAY,CAAC,IAAI,CAAC,aAAa,CAAC;AAChC,gBAAA,IAAI,gBAAgB,GAAG,EAAE,EAAE;oBACzB;;gBAEF;;;AAIJ,QAAA,IAAI,gBAAgB,GAAG,EAAE,IAAI,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,MAAM,EAAE;YAC9E,SAAS,GAAG,IAAI;;AAGlB,QAAA,IAAI,SAAS,IAAI,IAAI,IAAI,SAAS,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE;AACxD,YAAA,MAAM,iBAAiB,GAAG,OAAO,CAAC,SAAS,CAAC,GAAG,IAAI,GAAG,EAAE,OAAO,EAAE,KAAK,SAAS,CAAC;AAEhF,YAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;AACzB,gBAAA,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,iBAAiB,CAAC;;;;AAKhD,IAAA,IAAI,YAAY,IAAI,cAAc,GAAG,CAAC,EAAE;QACtC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAgB,CAAC;QACzC,QAAQ,CAAC,KAAK,EAAE;;IAGlB,sBAAsB,IAAI,iBAAiB;AAC3C,IAAA,MAAM,MAAM,GAAG;QACb,sBAAsB;AACtB,QAAA,OAAO,EAAE,EAAmB;AAC5B,QAAA,gBAAgB,EAAE,YAAY;KAC/B;IAED,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC,IAAI,gBAAgB,GAAG,CAAC,KAAK,kBAAkB,GAAG,EAAE,IAAI,gBAAgB,CAAC,SAAS,EAAE,OAAO,EAAE,kBAAkB,CAAC,CAAC,EAAE;;AAE9I,QAAA,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,EAAmB;AACnD,QAAA,OAAO,MAAM;;IAGf,IAAI,gBAAgB,GAAG,EAAE,IAAI,kBAAkB,GAAG,CAAC,EAAE;AACnD,QAAA,MAAM,IAAI,KAAK,CAAC,mGAAmG,CAAC;;IAGtH,IAAI,CAAC,aAAa,EAAE;AAClB,QAAA,MAAM,IAAI,KAAK,CAAC,qFAAqF,CAAC;;;;;AAMxG,IAAA,IAAI,cAAc,GAAG,EAAE;AACvB,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;AACvC,QAAA,MAAM,cAAc,GAAG,OAAO,CAAC,CAAC,CAAC;AACjC,QAAA,MAAM,IAAI,GAAG,cAAc,EAAE,OAAO,EAAE;AACtC,QAAA,IAAI,IAAI,KAAK,IAAI,EAAE;YACjB,cAAc,GAAG,CAAC;;AAEpB,QAAA,IAAI,cAAc,GAAG,EAAE,KAAK,IAAI,KAAK,OAAO,IAAI,IAAI,KAAK,QAAQ,CAAC,EAAE;YAClE;;;AAIJ,IAAA,IAAI,cAAc,KAAK,EAAE,EAAE;AACzB,QAAA,MAAM,IAAI,KAAK,CAAC,2GAA2G,CAAC;;AAG9H,IAAA,kBAAkB,GAAG,cAAc,GAAG,CAAC,GAAG,cAAc;AACxD,IAAA,MAAM,kBAAkB,GAAG,YAAY,CAAC,IAAI,SAAS,CAAC,EAAE,OAAO,EAAE,CAAC,aAAa,CAAC,EAAE,CAAC,CAAC;AACpF,IAAA,MAAM,iBAAiB,GAAG,sBAAsB,GAAG,kBAAkB;IACrE,MAAM,OAAO,GAA4B,gBAAgB,CAAC,OAAO,CAAC,cAAc,CAAc,EAAE,aAAa,CAAC;AAC7G,IAAA,OAAO,CAAC,cAAc,CAAe,CAAC,OAAO,GAAG,OAAO;AACxD,IAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;AACzB,QAAA,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,EAAmB;AACnD,QAAA,OAAO,MAAM;;AAGf,IAAA,MAAM,eAAe,GAAc,OAAO,CAAC,cAAc,CAAc;;AAEvE,IAAA,MAAM,4BAA4B,GAAG,CAAC,kBAAkB,CAAC,kBAAkB,CAAC,IAAI,CAAC,IAAI,kBAAkB;AACvG,IAAA,sBAAsB,GAAG,oBAAoB,GAAG,4BAA4B;IAC5E,iBAAiB,GAAG,CAAC;IACrB,IAAI,UAAU,GAAkB,EAAE;AAClC,IAAA,MAAM,mBAAmB,GAAG,CAAC,GAAG,SAAS,CAAC;AAC1C,IAAA,IAAI,YAAY,GAAG,mBAAmB,CAAC,MAAM;AAC7C,IAAA,OAAO,mBAAmB,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,GAAG,sBAAsB,IAAI,YAAY,GAAG,kBAAkB,EAAE;AACxH,QAAA,YAAY,EAAE;AACd,QAAA,MAAM,aAAa,GAAG,mBAAmB,CAAC,GAAG,EAAE;AAC/C,QAAA,IAAI,CAAC,aAAa;YAAE;QACpB,MAAM,UAAU,GAAG,kBAAkB,CAAC,YAAY,CAAC,IAAI,CAAC;QACxD,IAAI,CAAC,iBAAiB,GAAG,UAAU,KAAK,sBAAsB,EAAE;AAC9D,YAAA,UAAU,CAAC,IAAI,CAAC,aAAa,CAAC;YAC9B,iBAAiB,IAAI,UAAU;;aAC1B;AACL,YAAA,QAAQ,CAAC,IAAI,CAAC,aAAa,CAAC;YAC5B;;;IAIJ,MAAM,YAAY,GAAc,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;AACjE,IAAA,MAAM,gBAAgB,GAAG,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,OAAO,EAAE;AACpE,IAAA,IAAI,gBAAgB,KAAK,MAAM,EAAE;QAC/B,SAAS,GAAG,IAAI;;AAGlB,IAAA,IAAI,SAAS,IAAI,IAAI,IAAI,SAAS,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;AAC3D,QAAA,MAAM,iBAAiB,GAAG,UAAU,CAAC,SAAS,CAAC,GAAG,IAAI,GAAG,CAAC,OAAO,EAAE,KAAK,SAAS,CAAC;AAClF,QAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;AACzB,YAAA,UAAU,GAAG,UAAU,CAAC,KAAK,CAAC,iBAAiB,CAAC;;;AAIpD,IAAA,IAAI,gBAAgB,KAAK,IAAI,EAAE;QAC7B,MAAM,OAAO,GAAG,gBAAgB,CAAC,YAAY,EAAE,aAAa,CAAC;QAC7D,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,OAAO,GAAG,OAAO;;SAC9C;AACL,QAAA,UAAU,CAAC,IAAI,CAAC,eAAe,CAAC;;AAGlC,IAAA,IAAI,YAAY,IAAI,cAAc,GAAG,CAAC,EAAE;QACtC,UAAU,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAgB,CAAC;QAC5C,mBAAmB,CAAC,KAAK,EAAE;;AAG7B,IAAA,MAAM,CAAC,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE;AACrC,IAAA,OAAO,MAAM;AACf;AAEM,SAAU,gBAAgB,CAAC,KAAc,EAAA;AAC7C,IAAA,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC;AAChE;AAEM,SAAU,mBAAmB,CAAC,aAAyC,EAAA;IAC3E,MAAM,kBAAkB,GAAG,EAAE,GAAG,aAAa,CAAC,kBAAkB,EAAE;AAClE,IAAA,IAAI,kBAAkB,GAAG,aAAa,CAAC,UAAU;IACjD,IAAI,eAAe,GAAG,CAAC;IACvB,IAAI,WAAW,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,kBAAkB,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;IAChF,OAAO,SAAS,aAAa,CAAC,MAA2B,EAAA;AAIvD,QAAA,IAAI,YAAuC;AAC3C,QAAA,IAAI,MAAM,CAAC,aAAa,KACtB,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,YAAY;AAC/C,gBACD,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB;oBAEvD,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB,CAAC,cAAc;uBACrE,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB,CAAC,UAAU,CAAC,CACzE,CACF,CACF,IAAI,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,aAAa,CAAC,EAAE;AACzD,YAAA,YAAY,GAAG,oBAAoB,CAAC,MAAM,CAAC,aAAa,CAAC;AACzD,YAAA,WAAW,GAAG,YAAY,CAAC,YAAY;;AAGzC,QAAA,KAAK,IAAI,CAAC,GAAG,kBAAkB,EAAE,CAAC,GAAG,MAAM,CAAC,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;YAChE,MAAM,OAAO,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC;;AAElC,YAAA,IAAI,CAAC,KAAK,kBAAkB,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,SAAS,IAAI,YAAY,EAAE;AACnF,gBAAA,kBAAkB,CAAC,CAAC,CAAC,GAAG,YAAY,CAAC,aAAa;;;AAE7C,iBAAA,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,SAAS,EAAE;gBAC9C,kBAAkB,CAAC,CAAC,CAAC,GAAG,aAAa,CAAC,YAAY,CAAC,OAAO,CAAC;AAC3D,gBAAA,WAAW,IAAI,kBAAkB,CAAC,CAAC,CAAC;;;;;;;QAQxC,IAAI,YAAY,EAAE;;YAEhB,MAAM,gBAAgB,GAAG,MAAM,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,KAAK,CAAC,KAAI;;AAEvF,gBAAA,MAAM,UAAU,GAAG,MAAM,CAAC,GAAG,CAAC;AAC9B,gBAAA,IAAI,UAAU,KAAK,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,KAAK,QAAQ,EAAE;oBACjE,OAAO,GAAG,GAAG,KAAK;;AAEpB,gBAAA,OAAO,UAAU,IAAI,eAAe,GAAG,GAAG,GAAG,KAAK,GAAG,GAAG;aACzD,EAAE,CAAC,CAAC;;AAGL,YAAA,MAAM,KAAK,GAAG,YAAY,CAAC,YAAY,GAAG,gBAAgB;YAC1D,MAAM,WAAW,GAAG,KAAK,IAAI,CAAC,GAAC,CAAC,IAAI,KAAK,IAAI,GAAG;;YAGhD,IAAI,WAAW,EAAE;AACf,gBAAA,KAAK,MAAM,GAAG,IAAI,kBAAkB,EAAE;AACpC,oBAAA,MAAM,UAAU,GAAG,MAAM,CAAC,GAAG,CAAC;AAC9B,oBAAA,IAAI,UAAU,KAAK,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,KAAK,QAAQ,EAAE;AACjE,wBAAA,kBAAkB,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;;AAChE,yBAAA,IAAI,UAAU,IAAI,eAAe,EAAE;;AAExC,wBAAA,kBAAkB,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;;;;;AAM7E,QAAA,kBAAkB,GAAG,MAAM,CAAC,QAAQ,CAAC,MAAM;AAC3C,QAAA,IAAI,WAAW,IAAI,aAAa,CAAC,SAAS,EAAE;YAC1C,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,QAAQ,EAAE,kBAAkB,EAAE;;AAGzD,QAAA,MAAM,EAAE,OAAO,EAAE,GAAG,2BAA2B,CAAC;YAC9C,gBAAgB,EAAE,aAAa,CAAC,SAAS;YACzC,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,kBAAkB;YAClB,SAAS,EAAE,MAAM,CAAC,SAAS;YAC3B,eAAe,EAAE,aAAa,CAAC,eAAe;YAC9C,YAAY,EAAE,aAAa,CAAC,YAAY;AACxC,YAAA,aAAa,EAAE,aAAa,CAAC,QAAQ,KAAK,SAAS,CAAC,OAAO,GAAG,YAAY,CAAC,iBAAiB,GAAG,YAAY,CAAC,QAAQ;AACrH,SAAA,CAAC;AACF,QAAA,eAAe,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC;AAEtE,QAAA,OAAO,EAAE,OAAO,EAAE,kBAAkB,EAAE;AACxC,KAAC;AACH;;;;"}
+{"version":3,"file":"prune.mjs","sources":["../../../src/messages/prune.ts"],"sourcesContent":["import { AIMessage, BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { ThinkingContentText, MessageContentComplex, ReasoningContentText } from '@/types/stream';\nimport type { TokenCounter } from '@/types/run';\nimport { ContentTypes, Providers } from '@/common';\n\nexport type PruneMessagesFactoryParams = {\n  provider?: Providers;\n  maxTokens: number;\n  startIndex: number;\n  tokenCounter: TokenCounter;\n  indexTokenCountMap: Record<string, number>;\n  thinkingEnabled?: boolean;\n};\nexport type PruneMessagesParams = {\n  messages: BaseMessage[];\n  usageMetadata?: Partial<UsageMetadata>;\n  startType?: ReturnType<BaseMessage['getType']>;\n}\n\nfunction isIndexInContext(arrayA: unknown[], arrayB: unknown[], targetIndex: number): boolean {\n  const startingIndexInA = arrayA.length - arrayB.length;\n  return targetIndex >= startingIndexInA;\n}\n\nfunction addThinkingBlock(message: AIMessage, thinkingBlock: ThinkingContentText | ReasoningContentText): AIMessage {\n  const content: MessageContentComplex[] = Array.isArray(message.content)\n    ? message.content as MessageContentComplex[]\n    : [{\n      type: ContentTypes.TEXT,\n      text: message.content,\n    }];\n  /** Edge case, the message already has the thinking block */\n  if (content[0].type === thinkingBlock.type) {\n    return message;\n  }\n  content.unshift(thinkingBlock);\n  return new AIMessage({\n    ...message,\n    content\n  });\n}\n\n/**\n * Calculates the total tokens from a single usage object\n *\n * @param usage The usage metadata object containing token information\n * @returns An object containing the total input and output tokens\n */\nexport function calculateTotalTokens(usage: Partial<UsageMetadata>): UsageMetadata {\n  const baseInputTokens = Number(usage.input_tokens) || 0;\n  const cacheCreation = Number(usage.input_token_details?.cache_creation) || 0;\n  const cacheRead = Number(usage.input_token_details?.cache_read) || 0;\n\n  const totalInputTokens = baseInputTokens + cacheCreation + cacheRead;\n  const totalOutputTokens = Number(usage.output_tokens) || 0;\n\n  return {\n    input_tokens: totalInputTokens,\n    output_tokens: totalOutputTokens,\n    total_tokens: totalInputTokens + totalOutputTokens\n  };\n}\n\nexport type PruningResult = {\n  context: BaseMessage[];\n  remainingContextTokens: number;\n  messagesToRefine: BaseMessage[];\n  thinkingStartIndex?: number;\n};\n\n/**\n * Processes an array of messages and returns a context of messages that fit within a specified token limit.\n * It iterates over the messages from newest to oldest, adding them to the context until the token limit is reached.\n *\n * @param options Configuration options for processing messages\n * @returns Object containing the message context, remaining tokens, messages not included, and summary index\n */\nexport function getMessagesWithinTokenLimit({\n  messages: _messages,\n  maxContextTokens,\n  indexTokenCountMap,\n  startType: _startType,\n  thinkingEnabled,\n  tokenCounter,\n  thinkingStartIndex: _thinkingStartIndex = -1,\n  reasoningType = ContentTypes.THINKING,\n}: {\n  messages: BaseMessage[];\n  maxContextTokens: number;\n  indexTokenCountMap: Record<string, number | undefined>;\n  startType?: string | string[];\n  thinkingEnabled?: boolean;\n  tokenCounter: TokenCounter;\n  thinkingStartIndex?: number;\n  reasoningType?: ContentTypes.THINKING | ContentTypes.REASONING_CONTENT;\n}): PruningResult {\n  // Every reply is primed with <|start|>assistant<|message|>, so we\n  // start with 3 tokens for the label after all messages have been counted.\n  let currentTokenCount = 3;\n  const instructions = _messages[0]?.getType() === 'system' ? _messages[0] : undefined;\n  const instructionsTokenCount = instructions != null ? indexTokenCountMap[0] ?? 0 : 0;\n  const initialContextTokens = maxContextTokens - instructionsTokenCount;\n  let remainingContextTokens = initialContextTokens;\n  let startType = _startType;\n  const originalLength = _messages.length;\n  const messages = [..._messages];\n  /**\n   * IMPORTANT: this context array gets reversed at the end, since the latest messages get pushed first.\n   *\n   * This may be confusing to read, but it is done to ensure the context is in the correct order for the model.\n   * */\n  let context: Array<BaseMessage | undefined> = [];\n\n  let thinkingStartIndex = _thinkingStartIndex;\n  let thinkingEndIndex = -1;\n  let thinkingBlock: ThinkingContentText | ReasoningContentText | undefined;\n  const endIndex = instructions != null ? 1 : 0;\n  const prunedMemory: BaseMessage[] = [];\n\n  if (_thinkingStartIndex > -1) {\n    const thinkingMessageContent = messages[_thinkingStartIndex]?.content;\n    if (Array.isArray(thinkingMessageContent)) {\n      thinkingBlock = thinkingMessageContent.find((content) => content.type === reasoningType) as ThinkingContentText | undefined;\n    }\n  }\n\n  if (currentTokenCount < remainingContextTokens) {\n    let currentIndex = messages.length;\n    while (messages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > endIndex) {\n      currentIndex--;\n      if (messages.length === 1 && instructions) {\n        break;\n      }\n      const poppedMessage = messages.pop();\n      if (!poppedMessage) continue;\n      const messageType = poppedMessage.getType();\n      if (thinkingEnabled === true && thinkingEndIndex === -1 && (currentIndex === (originalLength - 1)) && (messageType === 'ai' || messageType === 'tool')) {\n        thinkingEndIndex = currentIndex;\n      }\n      if (thinkingEndIndex > -1 && !thinkingBlock  && thinkingStartIndex < 0 && messageType === 'ai' && Array.isArray(poppedMessage.content)) {\n        thinkingBlock = (poppedMessage.content.find((content) => content.type === reasoningType)) as ThinkingContentText | undefined;\n        thinkingStartIndex = thinkingBlock != null ? currentIndex : -1;\n      }\n      /** False start, the latest message was not part of a multi-assistant/tool sequence of messages */\n      if (\n        thinkingEndIndex > -1\n        && currentIndex === (thinkingEndIndex - 1)\n        && (messageType !== 'ai' && messageType !== 'tool')\n      ) {\n        thinkingEndIndex = -1;\n      }\n\n      const tokenCount = indexTokenCountMap[currentIndex] ?? 0;\n\n      if (prunedMemory.length === 0 && ((currentTokenCount + tokenCount) <= remainingContextTokens)) {\n        context.push(poppedMessage);\n        currentTokenCount += tokenCount;\n      } else {\n        prunedMemory.push(poppedMessage);\n        if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {\n          continue;\n        }\n        break;\n      }\n    }\n\n    if (context[context.length - 1]?.getType() === 'tool') {\n      startType = ['ai', 'human'];\n    }\n\n    if (startType != null && startType.length > 0 && context.length > 0) {\n      let requiredTypeIndex = -1;\n\n      let totalTokens = 0;\n      for (let i = context.length - 1; i >= 0; i--) {\n        const currentType = context[i]?.getType() ?? '';\n        if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {\n          requiredTypeIndex = i + 1;\n          break;\n        }\n        const originalIndex = originalLength - 1 - i;\n        totalTokens += indexTokenCountMap[originalIndex] ?? 0;\n      }\n\n      if (requiredTypeIndex > 0) {\n        currentTokenCount -= totalTokens;\n        context = context.slice(0, requiredTypeIndex);\n      }\n    }\n  }\n\n  if (instructions && originalLength > 0) {\n    context.push(_messages[0] as BaseMessage);\n    messages.shift();\n  }\n\n  remainingContextTokens -= currentTokenCount;\n  const result: PruningResult = {\n    remainingContextTokens,\n    context: [] as BaseMessage[],\n    messagesToRefine: prunedMemory,\n  };\n\n  if (thinkingStartIndex > -1) {\n    result.thinkingStartIndex = thinkingStartIndex;\n  }\n\n  if (prunedMemory.length === 0 || thinkingEndIndex < 0 || (thinkingStartIndex > -1 && isIndexInContext(_messages, context, thinkingStartIndex))) {\n    // we reverse at this step to ensure the context is in the correct order for the model, and we need to work backwards\n    result.context = context.reverse() as BaseMessage[];\n    return result;\n  }\n\n  if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no \"AI\" messages with thinking blocks.');\n  }\n\n  if (!thinkingBlock) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no thinking block found.');\n  }\n\n  // Since we have a thinking sequence, we need to find the last assistant message\n  // in the latest AI/tool sequence to add the thinking block that falls outside of the current context\n  // Latest messages are ordered first.\n  let assistantIndex = -1;\n  for (let i = 0; i < context.length; i++) {\n    const currentMessage = context[i];\n    const type = currentMessage?.getType();\n    if (type === 'ai') {\n      assistantIndex = i;\n    }\n    if (assistantIndex > -1 && (type === 'human' || type === 'system')) {\n      break;\n    }\n  }\n\n  if (assistantIndex === -1) {\n    throw new Error('The payload is malformed. There is a thinking sequence but no \"AI\" messages to append thinking blocks to.');\n  }\n\n  thinkingStartIndex = originalLength - 1 - assistantIndex;\n  const thinkingTokenCount = tokenCounter(new AIMessage({ content: [thinkingBlock] }));\n  const newRemainingCount = remainingContextTokens - thinkingTokenCount;\n  const newMessage = addThinkingBlock(context[assistantIndex] as AIMessage, thinkingBlock);\n  context[assistantIndex] = newMessage;\n  if (newRemainingCount > 0) {\n    result.context = context.reverse() as BaseMessage[];\n    return result;\n  }\n\n  const thinkingMessage: AIMessage = context[assistantIndex] as AIMessage;\n  // now we need to an additional round of pruning but making the thinking block fit\n  const newThinkingMessageTokenCount = (indexTokenCountMap[thinkingStartIndex] ?? 0) + thinkingTokenCount;\n  remainingContextTokens = initialContextTokens - newThinkingMessageTokenCount;\n  currentTokenCount = 3;\n  let newContext: BaseMessage[] = [];\n  const secondRoundMessages = [..._messages];\n  let currentIndex = secondRoundMessages.length;\n  while (secondRoundMessages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > thinkingStartIndex) {\n    currentIndex--;\n    const poppedMessage = secondRoundMessages.pop();\n    if (!poppedMessage) continue;\n    const tokenCount = indexTokenCountMap[currentIndex] ?? 0;\n    if ((currentTokenCount + tokenCount) <= remainingContextTokens) {\n      newContext.push(poppedMessage);\n      currentTokenCount += tokenCount;\n    } else {\n      messages.push(poppedMessage);\n      break;\n    }\n  }\n\n  const firstMessage: AIMessage = newContext[newContext.length - 1];\n  const firstMessageType = newContext[newContext.length - 1].getType();\n  if (firstMessageType === 'tool') {\n    startType = ['ai', 'human'];\n  }\n\n  if (startType != null && startType.length > 0 && newContext.length > 0) {\n    let requiredTypeIndex = -1;\n\n    let totalTokens = 0;\n    for (let i = newContext.length - 1; i >= 0; i--) {\n      const currentType = newContext[i]?.getType() ?? '';\n      if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {\n        requiredTypeIndex = i + 1;\n        break;\n      }\n      const originalIndex = originalLength - 1 - i;\n      totalTokens += indexTokenCountMap[originalIndex] ?? 0;\n    }\n\n    if (requiredTypeIndex > 0) {\n      currentTokenCount -= totalTokens;\n      newContext = newContext.slice(0, requiredTypeIndex);\n    }\n  }\n\n  if (firstMessageType === 'ai') {\n    const newMessage = addThinkingBlock(firstMessage, thinkingBlock);\n    newContext[newContext.length - 1] = newMessage;\n  } else {\n    newContext.push(thinkingMessage);\n  }\n\n  if (instructions && originalLength > 0) {\n    newContext.push(_messages[0] as BaseMessage);\n    secondRoundMessages.shift();\n  }\n\n  result.context = newContext.reverse();\n  return result;\n}\n\nexport function checkValidNumber(value: unknown): value is number {\n  return typeof value === 'number' && !isNaN(value) && value > 0;\n}\n\nexport function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {\n  const indexTokenCountMap = { ...factoryParams.indexTokenCountMap };\n  let lastTurnStartIndex = factoryParams.startIndex;\n  let lastCutOffIndex = 0;\n  let totalTokens = (Object.values(indexTokenCountMap)).reduce((a, b) => a + b, 0);\n  let runThinkingStartIndex = -1;\n  return function pruneMessages(params: PruneMessagesParams): {\n    context: BaseMessage[];\n    indexTokenCountMap: Record<string, number>;\n  } {\n    let currentUsage: UsageMetadata | undefined;\n    if (params.usageMetadata && (\n      checkValidNumber(params.usageMetadata.input_tokens)\n      || (\n        checkValidNumber(params.usageMetadata.input_token_details)\n        && (\n          checkValidNumber(params.usageMetadata.input_token_details.cache_creation)\n          || checkValidNumber(params.usageMetadata.input_token_details.cache_read)\n        )\n      )\n    ) && checkValidNumber(params.usageMetadata.output_tokens)) {\n      currentUsage = calculateTotalTokens(params.usageMetadata);\n      totalTokens = currentUsage.total_tokens;\n    }\n\n    for (let i = lastTurnStartIndex; i < params.messages.length; i++) {\n      const message = params.messages[i];\n      // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition\n      if (i === lastTurnStartIndex && indexTokenCountMap[i] === undefined && currentUsage) {\n        indexTokenCountMap[i] = currentUsage.output_tokens;\n      // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition\n      } else if (indexTokenCountMap[i] === undefined) {\n        indexTokenCountMap[i] = factoryParams.tokenCounter(message);\n        totalTokens += indexTokenCountMap[i];\n      }\n    }\n\n    // If `currentUsage` is defined, we need to distribute the current total tokens to our `indexTokenCountMap`,\n    // We must distribute it in a weighted manner, so that the total token count is equal to `currentUsage.total_tokens`,\n    // relative the manually counted tokens in `indexTokenCountMap`.\n    // EDGE CASE: when the resulting context gets pruned, we should not distribute the usage for messages that are not in the context.\n    if (currentUsage) {\n      // Calculate the sum of tokens only for indices at or after lastCutOffIndex\n      const totalIndexTokens = Object.entries(indexTokenCountMap).reduce((sum, [key, value]) => {\n        // Convert string key to number and check if it's >= lastCutOffIndex\n        const numericKey = Number(key);\n        if (numericKey === 0 && params.messages[0].getType() === 'system') {\n          return sum + value;\n        }\n        return numericKey >= lastCutOffIndex ? sum + value : sum;\n      }, 0);\n\n      // Calculate ratio based only on messages that remain in the context\n      const ratio = currentUsage.total_tokens / totalIndexTokens;\n      const isRatioSafe = ratio >= 1/3 && ratio <= 2.5;\n\n      // Apply the ratio adjustment only to messages at or after lastCutOffIndex, and only if the ratio is safe\n      if (isRatioSafe) {\n        for (const key in indexTokenCountMap) {\n          const numericKey = Number(key);\n          if (numericKey === 0 && params.messages[0].getType() === 'system') {\n            indexTokenCountMap[key] = Math.round(indexTokenCountMap[key] * ratio);\n          } else if (numericKey >= lastCutOffIndex) {\n            // Only adjust token counts for messages still in the context\n            indexTokenCountMap[key] = Math.round(indexTokenCountMap[key] * ratio);\n          }\n        }\n      }\n    }\n\n    lastTurnStartIndex = params.messages.length;\n    if (totalTokens <= factoryParams.maxTokens) {\n      return { context: params.messages, indexTokenCountMap };\n    }\n\n    const { context, thinkingStartIndex } = getMessagesWithinTokenLimit({\n      maxContextTokens: factoryParams.maxTokens,\n      messages: params.messages,\n      indexTokenCountMap,\n      startType: params.startType,\n      thinkingEnabled: factoryParams.thinkingEnabled,\n      tokenCounter: factoryParams.tokenCounter,\n      reasoningType: factoryParams.provider === Providers.BEDROCK ? ContentTypes.REASONING_CONTENT : ContentTypes.THINKING,\n      thinkingStartIndex: factoryParams.thinkingEnabled === true ? runThinkingStartIndex : undefined,\n    });\n    runThinkingStartIndex = thinkingStartIndex ?? -1;\n    /** The index is the first value of `context`, index relative to `params.messages` */\n    lastCutOffIndex = Math.max(params.messages.length - (context.length - (context[0]?.getType() === 'system' ? 1 : 0)), 0);\n\n    return { context, indexTokenCountMap };\n  };\n}\n"],"names":[],"mappings":";;;AAmBA,SAAS,gBAAgB,CAAC,MAAiB,EAAE,MAAiB,EAAE,WAAmB,EAAA;IACjF,MAAM,gBAAgB,GAAG,MAAM,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM;IACtD,OAAO,WAAW,IAAI,gBAAgB;AACxC;AAEA,SAAS,gBAAgB,CAAC,OAAkB,EAAE,aAAyD,EAAA;IACrG,MAAM,OAAO,GAA4B,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO;UAClE,OAAO,CAAC;AACV,UAAE,CAAC;gBACD,IAAI,EAAE,YAAY,CAAC,IAAI;gBACvB,IAAI,EAAE,OAAO,CAAC,OAAO;AACtB,aAAA,CAAC;;IAEJ,IAAI,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,KAAK,aAAa,CAAC,IAAI,EAAE;AAC1C,QAAA,OAAO,OAAO;;AAEhB,IAAA,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;IAC9B,OAAO,IAAI,SAAS,CAAC;AACnB,QAAA,GAAG,OAAO;QACV;AACD,KAAA,CAAC;AACJ;AAEA;;;;;AAKG;AACG,SAAU,oBAAoB,CAAC,KAA6B,EAAA;IAChE,MAAM,eAAe,GAAG,MAAM,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC;AACvD,IAAA,MAAM,aAAa,GAAG,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,cAAc,CAAC,IAAI,CAAC;AAC5E,IAAA,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,UAAU,CAAC,IAAI,CAAC;AAEpE,IAAA,MAAM,gBAAgB,GAAG,eAAe,GAAG,aAAa,GAAG,SAAS;IACpE,MAAM,iBAAiB,GAAG,MAAM,CAAC,KAAK,CAAC,aAAa,CAAC,IAAI,CAAC;IAE1D,OAAO;AACL,QAAA,YAAY,EAAE,gBAAgB;AAC9B,QAAA,aAAa,EAAE,iBAAiB;QAChC,YAAY,EAAE,gBAAgB,GAAG;KAClC;AACH;AASA;;;;;;AAMG;AACa,SAAA,2BAA2B,CAAC,EAC1C,QAAQ,EAAE,SAAS,EACnB,gBAAgB,EAChB,kBAAkB,EAClB,SAAS,EAAE,UAAU,EACrB,eAAe,EACf,YAAY,EACZ,kBAAkB,EAAE,mBAAmB,GAAG,EAAE,EAC5C,aAAa,GAAG,YAAY,CAAC,QAAQ,GAUtC,EAAA;;;IAGC,IAAI,iBAAiB,GAAG,CAAC;IACzB,MAAM,YAAY,GAAG,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,QAAQ,GAAG,SAAS,CAAC,CAAC,CAAC,GAAG,SAAS;AACpF,IAAA,MAAM,sBAAsB,GAAG,YAAY,IAAI,IAAI,GAAG,kBAAkB,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC;AACpF,IAAA,MAAM,oBAAoB,GAAG,gBAAgB,GAAG,sBAAsB;IACtE,IAAI,sBAAsB,GAAG,oBAAoB;IACjD,IAAI,SAAS,GAAG,UAAU;AAC1B,IAAA,MAAM,cAAc,GAAG,SAAS,CAAC,MAAM;AACvC,IAAA,MAAM,QAAQ,GAAG,CAAC,GAAG,SAAS,CAAC;AAC/B;;;;AAIK;IACL,IAAI,OAAO,GAAmC,EAAE;IAEhD,IAAI,kBAAkB,GAAG,mBAAmB;AAC5C,IAAA,IAAI,gBAAgB,GAAG,EAAE;AACzB,IAAA,IAAI,aAAqE;AACzE,IAAA,MAAM,QAAQ,GAAG,YAAY,IAAI,IAAI,GAAG,CAAC,GAAG,CAAC;IAC7C,MAAM,YAAY,GAAkB,EAAE;AAEtC,IAAA,IAAI,mBAAmB,GAAG,EAAE,EAAE;QAC5B,MAAM,sBAAsB,GAAG,QAAQ,CAAC,mBAAmB,CAAC,EAAE,OAAO;AACrE,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,sBAAsB,CAAC,EAAE;AACzC,YAAA,aAAa,GAAG,sBAAsB,CAAC,IAAI,CAAC,CAAC,OAAO,KAAK,OAAO,CAAC,IAAI,KAAK,aAAa,CAAoC;;;AAI/H,IAAA,IAAI,iBAAiB,GAAG,sBAAsB,EAAE;AAC9C,QAAA,IAAI,YAAY,GAAG,QAAQ,CAAC,MAAM;AAClC,QAAA,OAAO,QAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,GAAG,sBAAsB,IAAI,YAAY,GAAG,QAAQ,EAAE;AACnG,YAAA,YAAY,EAAE;YACd,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,IAAI,YAAY,EAAE;gBACzC;;AAEF,YAAA,MAAM,aAAa,GAAG,QAAQ,CAAC,GAAG,EAAE;AACpC,YAAA,IAAI,CAAC,aAAa;gBAAE;AACpB,YAAA,MAAM,WAAW,GAAG,aAAa,CAAC,OAAO,EAAE;AAC3C,YAAA,IAAI,eAAe,KAAK,IAAI,IAAI,gBAAgB,KAAK,EAAE,KAAK,YAAY,MAAM,cAAc,GAAG,CAAC,CAAC,CAAC,KAAK,WAAW,KAAK,IAAI,IAAI,WAAW,KAAK,MAAM,CAAC,EAAE;gBACtJ,gBAAgB,GAAG,YAAY;;YAEjC,IAAI,gBAAgB,GAAG,EAAE,IAAI,CAAC,aAAa,IAAK,kBAAkB,GAAG,CAAC,IAAI,WAAW,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,aAAa,CAAC,OAAO,CAAC,EAAE;gBACtI,aAAa,IAAI,aAAa,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,OAAO,KAAK,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAoC;AAC5H,gBAAA,kBAAkB,GAAG,aAAa,IAAI,IAAI,GAAG,YAAY,GAAG,EAAE;;;YAGhE,IACE,gBAAgB,GAAG;AAChB,mBAAA,YAAY,MAAM,gBAAgB,GAAG,CAAC;oBACrC,WAAW,KAAK,IAAI,IAAI,WAAW,KAAK,MAAM,CAAC,EACnD;gBACA,gBAAgB,GAAG,EAAE;;YAGvB,MAAM,UAAU,GAAG,kBAAkB,CAAC,YAAY,CAAC,IAAI,CAAC;AAExD,YAAA,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC,KAAK,CAAC,iBAAiB,GAAG,UAAU,KAAK,sBAAsB,CAAC,EAAE;AAC7F,gBAAA,OAAO,CAAC,IAAI,CAAC,aAAa,CAAC;gBAC3B,iBAAiB,IAAI,UAAU;;iBAC1B;AACL,gBAAA,YAAY,CAAC,IAAI,CAAC,aAAa,CAAC;gBAChC,IAAI,gBAAgB,GAAG,EAAE,IAAI,kBAAkB,GAAG,CAAC,EAAE;oBACnD;;gBAEF;;;AAIJ,QAAA,IAAI,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,MAAM,EAAE;AACrD,YAAA,SAAS,GAAG,CAAC,IAAI,EAAE,OAAO,CAAC;;AAG7B,QAAA,IAAI,SAAS,IAAI,IAAI,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE;AACnE,YAAA,IAAI,iBAAiB,GAAG,EAAE;YAE1B,IAAI,WAAW,GAAG,CAAC;AACnB,YAAA,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE;gBAC5C,MAAM,WAAW,GAAG,OAAO,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE;gBAC/C,IAAI,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,GAAG,SAAS,CAAC,QAAQ,CAAC,WAAW,CAAC,GAAG,WAAW,KAAK,SAAS,EAAE;AAC1F,oBAAA,iBAAiB,GAAG,CAAC,GAAG,CAAC;oBACzB;;AAEF,gBAAA,MAAM,aAAa,GAAG,cAAc,GAAG,CAAC,GAAG,CAAC;AAC5C,gBAAA,WAAW,IAAI,kBAAkB,CAAC,aAAa,CAAC,IAAI,CAAC;;AAGvD,YAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;gBACzB,iBAAiB,IAAI,WAAW;gBAChC,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,iBAAiB,CAAC;;;;AAKnD,IAAA,IAAI,YAAY,IAAI,cAAc,GAAG,CAAC,EAAE;QACtC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAgB,CAAC;QACzC,QAAQ,CAAC,KAAK,EAAE;;IAGlB,sBAAsB,IAAI,iBAAiB;AAC3C,IAAA,MAAM,MAAM,GAAkB;QAC5B,sBAAsB;AACtB,QAAA,OAAO,EAAE,EAAmB;AAC5B,QAAA,gBAAgB,EAAE,YAAY;KAC/B;AAED,IAAA,IAAI,kBAAkB,GAAG,EAAE,EAAE;AAC3B,QAAA,MAAM,CAAC,kBAAkB,GAAG,kBAAkB;;IAGhD,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC,IAAI,gBAAgB,GAAG,CAAC,KAAK,kBAAkB,GAAG,EAAE,IAAI,gBAAgB,CAAC,SAAS,EAAE,OAAO,EAAE,kBAAkB,CAAC,CAAC,EAAE;;AAE9I,QAAA,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,EAAmB;AACnD,QAAA,OAAO,MAAM;;IAGf,IAAI,gBAAgB,GAAG,EAAE,IAAI,kBAAkB,GAAG,CAAC,EAAE;AACnD,QAAA,MAAM,IAAI,KAAK,CAAC,mGAAmG,CAAC;;IAGtH,IAAI,CAAC,aAAa,EAAE;AAClB,QAAA,MAAM,IAAI,KAAK,CAAC,qFAAqF,CAAC;;;;;AAMxG,IAAA,IAAI,cAAc,GAAG,EAAE;AACvB,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;AACvC,QAAA,MAAM,cAAc,GAAG,OAAO,CAAC,CAAC,CAAC;AACjC,QAAA,MAAM,IAAI,GAAG,cAAc,EAAE,OAAO,EAAE;AACtC,QAAA,IAAI,IAAI,KAAK,IAAI,EAAE;YACjB,cAAc,GAAG,CAAC;;AAEpB,QAAA,IAAI,cAAc,GAAG,EAAE,KAAK,IAAI,KAAK,OAAO,IAAI,IAAI,KAAK,QAAQ,CAAC,EAAE;YAClE;;;AAIJ,IAAA,IAAI,cAAc,KAAK,EAAE,EAAE;AACzB,QAAA,MAAM,IAAI,KAAK,CAAC,2GAA2G,CAAC;;AAG9H,IAAA,kBAAkB,GAAG,cAAc,GAAG,CAAC,GAAG,cAAc;AACxD,IAAA,MAAM,kBAAkB,GAAG,YAAY,CAAC,IAAI,SAAS,CAAC,EAAE,OAAO,EAAE,CAAC,aAAa,CAAC,EAAE,CAAC,CAAC;AACpF,IAAA,MAAM,iBAAiB,GAAG,sBAAsB,GAAG,kBAAkB;IACrE,MAAM,UAAU,GAAG,gBAAgB,CAAC,OAAO,CAAC,cAAc,CAAc,EAAE,aAAa,CAAC;AACxF,IAAA,OAAO,CAAC,cAAc,CAAC,GAAG,UAAU;AACpC,IAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;AACzB,QAAA,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,EAAmB;AACnD,QAAA,OAAO,MAAM;;AAGf,IAAA,MAAM,eAAe,GAAc,OAAO,CAAC,cAAc,CAAc;;AAEvE,IAAA,MAAM,4BAA4B,GAAG,CAAC,kBAAkB,CAAC,kBAAkB,CAAC,IAAI,CAAC,IAAI,kBAAkB;AACvG,IAAA,sBAAsB,GAAG,oBAAoB,GAAG,4BAA4B;IAC5E,iBAAiB,GAAG,CAAC;IACrB,IAAI,UAAU,GAAkB,EAAE;AAClC,IAAA,MAAM,mBAAmB,GAAG,CAAC,GAAG,SAAS,CAAC;AAC1C,IAAA,IAAI,YAAY,GAAG,mBAAmB,CAAC,MAAM;AAC7C,IAAA,OAAO,mBAAmB,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,GAAG,sBAAsB,IAAI,YAAY,GAAG,kBAAkB,EAAE;AACxH,QAAA,YAAY,EAAE;AACd,QAAA,MAAM,aAAa,GAAG,mBAAmB,CAAC,GAAG,EAAE;AAC/C,QAAA,IAAI,CAAC,aAAa;YAAE;QACpB,MAAM,UAAU,GAAG,kBAAkB,CAAC,YAAY,CAAC,IAAI,CAAC;QACxD,IAAI,CAAC,iBAAiB,GAAG,UAAU,KAAK,sBAAsB,EAAE;AAC9D,YAAA,UAAU,CAAC,IAAI,CAAC,aAAa,CAAC;YAC9B,iBAAiB,IAAI,UAAU;;aAC1B;AACL,YAAA,QAAQ,CAAC,IAAI,CAAC,aAAa,CAAC;YAC5B;;;IAIJ,MAAM,YAAY,GAAc,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;AACjE,IAAA,MAAM,gBAAgB,GAAG,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,OAAO,EAAE;AACpE,IAAA,IAAI,gBAAgB,KAAK,MAAM,EAAE;AAC/B,QAAA,SAAS,GAAG,CAAC,IAAI,EAAE,OAAO,CAAC;;AAG7B,IAAA,IAAI,SAAS,IAAI,IAAI,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;AACtE,QAAA,IAAI,iBAAiB,GAAG,EAAE;QAE1B,IAAI,WAAW,GAAG,CAAC;AACnB,QAAA,KAAK,IAAI,CAAC,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE;YAC/C,MAAM,WAAW,GAAG,UAAU,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE;YAClD,IAAI,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,GAAG,SAAS,CAAC,QAAQ,CAAC,WAAW,CAAC,GAAG,WAAW,KAAK,SAAS,EAAE;AAC1F,gBAAA,iBAAiB,GAAG,CAAC,GAAG,CAAC;gBACzB;;AAEF,YAAA,MAAM,aAAa,GAAG,cAAc,GAAG,CAAC,GAAG,CAAC;AAC5C,YAAA,WAAW,IAAI,kBAAkB,CAAC,aAAa,CAAC,IAAI,CAAC;;AAGvD,QAAA,IAAI,iBAAiB,GAAG,CAAC,EAAE;YACzB,iBAAiB,IAAI,WAAW;YAChC,UAAU,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,iBAAiB,CAAC;;;AAIvD,IAAA,IAAI,gBAAgB,KAAK,IAAI,EAAE;QAC7B,MAAM,UAAU,GAAG,gBAAgB,CAAC,YAAY,EAAE,aAAa,CAAC;QAChE,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,GAAG,UAAU;;SACzC;AACL,QAAA,UAAU,CAAC,IAAI,CAAC,eAAe,CAAC;;AAGlC,IAAA,IAAI,YAAY,IAAI,cAAc,GAAG,CAAC,EAAE;QACtC,UAAU,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAgB,CAAC;QAC5C,mBAAmB,CAAC,KAAK,EAAE;;AAG7B,IAAA,MAAM,CAAC,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE;AACrC,IAAA,OAAO,MAAM;AACf;AAEM,SAAU,gBAAgB,CAAC,KAAc,EAAA;AAC7C,IAAA,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC;AAChE;AAEM,SAAU,mBAAmB,CAAC,aAAyC,EAAA;IAC3E,MAAM,kBAAkB,GAAG,EAAE,GAAG,aAAa,CAAC,kBAAkB,EAAE;AAClE,IAAA,IAAI,kBAAkB,GAAG,aAAa,CAAC,UAAU;IACjD,IAAI,eAAe,GAAG,CAAC;IACvB,IAAI,WAAW,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,kBAAkB,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;AAChF,IAAA,IAAI,qBAAqB,GAAG,EAAE;IAC9B,OAAO,SAAS,aAAa,CAAC,MAA2B,EAAA;AAIvD,QAAA,IAAI,YAAuC;AAC3C,QAAA,IAAI,MAAM,CAAC,aAAa,KACtB,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,YAAY;AAC/C,gBACD,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB;oBAEvD,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB,CAAC,cAAc;uBACrE,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,mBAAmB,CAAC,UAAU,CAAC,CACzE,CACF,CACF,IAAI,gBAAgB,CAAC,MAAM,CAAC,aAAa,CAAC,aAAa,CAAC,EAAE;AACzD,YAAA,YAAY,GAAG,oBAAoB,CAAC,MAAM,CAAC,aAAa,CAAC;AACzD,YAAA,WAAW,GAAG,YAAY,CAAC,YAAY;;AAGzC,QAAA,KAAK,IAAI,CAAC,GAAG,kBAAkB,EAAE,CAAC,GAAG,MAAM,CAAC,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;YAChE,MAAM,OAAO,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC;;AAElC,YAAA,IAAI,CAAC,KAAK,kBAAkB,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,SAAS,IAAI,YAAY,EAAE;AACnF,gBAAA,kBAAkB,CAAC,CAAC,CAAC,GAAG,YAAY,CAAC,aAAa;;;AAE7C,iBAAA,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,SAAS,EAAE;gBAC9C,kBAAkB,CAAC,CAAC,CAAC,GAAG,aAAa,CAAC,YAAY,CAAC,OAAO,CAAC;AAC3D,gBAAA,WAAW,IAAI,kBAAkB,CAAC,CAAC,CAAC;;;;;;;QAQxC,IAAI,YAAY,EAAE;;YAEhB,MAAM,gBAAgB,GAAG,MAAM,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,KAAK,CAAC,KAAI;;AAEvF,gBAAA,MAAM,UAAU,GAAG,MAAM,CAAC,GAAG,CAAC;AAC9B,gBAAA,IAAI,UAAU,KAAK,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,KAAK,QAAQ,EAAE;oBACjE,OAAO,GAAG,GAAG,KAAK;;AAEpB,gBAAA,OAAO,UAAU,IAAI,eAAe,GAAG,GAAG,GAAG,KAAK,GAAG,GAAG;aACzD,EAAE,CAAC,CAAC;;AAGL,YAAA,MAAM,KAAK,GAAG,YAAY,CAAC,YAAY,GAAG,gBAAgB;YAC1D,MAAM,WAAW,GAAG,KAAK,IAAI,CAAC,GAAC,CAAC,IAAI,KAAK,IAAI,GAAG;;YAGhD,IAAI,WAAW,EAAE;AACf,gBAAA,KAAK,MAAM,GAAG,IAAI,kBAAkB,EAAE;AACpC,oBAAA,MAAM,UAAU,GAAG,MAAM,CAAC,GAAG,CAAC;AAC9B,oBAAA,IAAI,UAAU,KAAK,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,KAAK,QAAQ,EAAE;AACjE,wBAAA,kBAAkB,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;;AAChE,yBAAA,IAAI,UAAU,IAAI,eAAe,EAAE;;AAExC,wBAAA,kBAAkB,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;;;;;AAM7E,QAAA,kBAAkB,GAAG,MAAM,CAAC,QAAQ,CAAC,MAAM;AAC3C,QAAA,IAAI,WAAW,IAAI,aAAa,CAAC,SAAS,EAAE;YAC1C,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,QAAQ,EAAE,kBAAkB,EAAE;;AAGzD,QAAA,MAAM,EAAE,OAAO,EAAE,kBAAkB,EAAE,GAAG,2BAA2B,CAAC;YAClE,gBAAgB,EAAE,aAAa,CAAC,SAAS;YACzC,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,kBAAkB;YAClB,SAAS,EAAE,MAAM,CAAC,SAAS;YAC3B,eAAe,EAAE,aAAa,CAAC,eAAe;YAC9C,YAAY,EAAE,aAAa,CAAC,YAAY;AACxC,YAAA,aAAa,EAAE,aAAa,CAAC,QAAQ,KAAK,SAAS,CAAC,OAAO,GAAG,YAAY,CAAC,iBAAiB,GAAG,YAAY,CAAC,QAAQ;AACpH,YAAA,kBAAkB,EAAE,aAAa,CAAC,eAAe,KAAK,IAAI,GAAG,qBAAqB,GAAG,SAAS;AAC/F,SAAA,CAAC;AACF,QAAA,qBAAqB,GAAG,kBAAkB,IAAI,EAAE;;AAEhD,QAAA,eAAe,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,IAAI,OAAO,CAAC,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,QAAQ,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AAEvH,QAAA,OAAO,EAAE,OAAO,EAAE,kBAAkB,EAAE;AACxC,KAAC;AACH;;;;"}

package/dist/types/messages/prune.d.ts

@@ -21,6 +21,12 @@ export type PruneMessagesParams = {
  * @returns An object containing the total input and output tokens
  */
 export declare function calculateTotalTokens(usage: Partial<UsageMetadata>): UsageMetadata;
+export type PruningResult = {
+    context: BaseMessage[];
+    remainingContextTokens: number;
+    messagesToRefine: BaseMessage[];
+    thinkingStartIndex?: number;
+};
 /**
  * Processes an array of messages and returns a context of messages that fit within a specified token limit.
  * It iterates over the messages from newest to oldest, adding them to the context until the token limit is reached.
@@ -28,19 +34,16 @@ export declare function calculateTotalTokens(usage: Partial<UsageMetadata>): Usa
  * @param options Configuration options for processing messages
  * @returns Object containing the message context, remaining tokens, messages not included, and summary index
  */
-export declare function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, indexTokenCountMap, startType: _startType, thinkingEnabled, tokenCounter, reasoningType, }: {
+export declare function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, indexTokenCountMap, startType: _startType, thinkingEnabled, tokenCounter, thinkingStartIndex: _thinkingStartIndex, reasoningType, }: {
     messages: BaseMessage[];
     maxContextTokens: number;
     indexTokenCountMap: Record<string, number | undefined>;
-    tokenCounter: TokenCounter;
-    startType?: string;
+    startType?: string | string[];
     thinkingEnabled?: boolean;
+    tokenCounter: TokenCounter;
+    thinkingStartIndex?: number;
     reasoningType?: ContentTypes.THINKING | ContentTypes.REASONING_CONTENT;
-}): {
-    context: BaseMessage[];
-    remainingContextTokens: number;
-    messagesToRefine: BaseMessage[];
-};
+}): PruningResult;
 export declare function checkValidNumber(value: unknown): value is number;
 export declare function createPruneMessages(factoryParams: PruneMessagesFactoryParams): (params: PruneMessagesParams) => {
     context: BaseMessage[];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "2.3.8",
+  "version": "2.3.91",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",

package/src/messages/prune.ts

@@ -22,15 +22,22 @@ function isIndexInContext(arrayA: unknown[], arrayB: unknown[], targetIndex: num
   return targetIndex >= startingIndexInA;
 }
 
-function addThinkingBlock(message: AIMessage, thinkingBlock: ThinkingContentText | ReasoningContentText): MessageContentComplex[] {
+function addThinkingBlock(message: AIMessage, thinkingBlock: ThinkingContentText | ReasoningContentText): AIMessage {
   const content: MessageContentComplex[] = Array.isArray(message.content)
     ? message.content as MessageContentComplex[]
     : [{
       type: ContentTypes.TEXT,
       text: message.content,
     }];
+  /** Edge case, the message already has the thinking block */
+  if (content[0].type === thinkingBlock.type) {
+    return message;
+  }
   content.unshift(thinkingBlock);
-  return content;
+  return new AIMessage({
+    ...message,
+    content
+  });
 }
 
 /**
@@ -54,6 +61,13 @@ export function calculateTotalTokens(usage: Partial<UsageMetadata>): UsageMetada
   };
 }
 
+export type PruningResult = {
+  context: BaseMessage[];
+  remainingContextTokens: number;
+  messagesToRefine: BaseMessage[];
+  thinkingStartIndex?: number;
+};
+
 /**
  * Processes an array of messages and returns a context of messages that fit within a specified token limit.
  * It iterates over the messages from newest to oldest, adding them to the context until the token limit is reached.
@@ -68,20 +82,18 @@ export function getMessagesWithinTokenLimit({
   startType: _startType,
   thinkingEnabled,
   tokenCounter,
+  thinkingStartIndex: _thinkingStartIndex = -1,
   reasoningType = ContentTypes.THINKING,
 }: {
   messages: BaseMessage[];
   maxContextTokens: number;
   indexTokenCountMap: Record<string, number | undefined>;
-  tokenCounter: TokenCounter;
-  startType?: string;
+  startType?: string | string[];
   thinkingEnabled?: boolean;
+  tokenCounter: TokenCounter;
+  thinkingStartIndex?: number;
   reasoningType?: ContentTypes.THINKING | ContentTypes.REASONING_CONTENT;
-}): {
-  context: BaseMessage[];
-  remainingContextTokens: number;
-  messagesToRefine: BaseMessage[];
-} {
+}): PruningResult {
   // Every reply is primed with <|start|>assistant<|message|>, so we
   // start with 3 tokens for the label after all messages have been counted.
   let currentTokenCount = 3;
@@ -99,12 +111,19 @@ export function getMessagesWithinTokenLimit({
    * */
   let context: Array<BaseMessage | undefined> = [];
 
-  let thinkingStartIndex = -1;
+  let thinkingStartIndex = _thinkingStartIndex;
   let thinkingEndIndex = -1;
   let thinkingBlock: ThinkingContentText | ReasoningContentText | undefined;
   const endIndex = instructions != null ? 1 : 0;
   const prunedMemory: BaseMessage[] = [];
 
+  if (_thinkingStartIndex > -1) {
+    const thinkingMessageContent = messages[_thinkingStartIndex]?.content;
+    if (Array.isArray(thinkingMessageContent)) {
+      thinkingBlock = thinkingMessageContent.find((content) => content.type === reasoningType) as ThinkingContentText | undefined;
+    }
+  }
+
   if (currentTokenCount < remainingContextTokens) {
     let currentIndex = messages.length;
     while (messages.length > 0 && currentTokenCount < remainingContextTokens && currentIndex > endIndex) {
@@ -138,22 +157,34 @@ export function getMessagesWithinTokenLimit({
         currentTokenCount += tokenCount;
       } else {
         prunedMemory.push(poppedMessage);
-        if (thinkingEndIndex > -1) {
+        if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {
           continue;
         }
         break;
       }
     }
 
-    if (thinkingEndIndex > -1 && context[context.length - 1]?.getType() === 'tool') {
-      startType = 'ai';
+    if (context[context.length - 1]?.getType() === 'tool') {
+      startType = ['ai', 'human'];
     }
 
-    if (startType != null && startType && context.length > 0) {
-      const requiredTypeIndex = context.findIndex(msg => msg?.getType() === startType);
+    if (startType != null && startType.length > 0 && context.length > 0) {
+      let requiredTypeIndex = -1;
+
+      let totalTokens = 0;
+      for (let i = context.length - 1; i >= 0; i--) {
+        const currentType = context[i]?.getType() ?? '';
+        if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {
+          requiredTypeIndex = i + 1;
+          break;
+        }
+        const originalIndex = originalLength - 1 - i;
+        totalTokens += indexTokenCountMap[originalIndex] ?? 0;
+      }
 
       if (requiredTypeIndex > 0) {
-        context = context.slice(requiredTypeIndex);
+        currentTokenCount -= totalTokens;
+        context = context.slice(0, requiredTypeIndex);
       }
     }
   }
@@ -164,12 +195,16 @@ export function getMessagesWithinTokenLimit({
   }
 
   remainingContextTokens -= currentTokenCount;
-  const result = {
+  const result: PruningResult = {
     remainingContextTokens,
     context: [] as BaseMessage[],
     messagesToRefine: prunedMemory,
   };
 
+  if (thinkingStartIndex > -1) {
+    result.thinkingStartIndex = thinkingStartIndex;
+  }
+
   if (prunedMemory.length === 0 || thinkingEndIndex < 0 || (thinkingStartIndex > -1 && isIndexInContext(_messages, context, thinkingStartIndex))) {
     // we reverse at this step to ensure the context is in the correct order for the model, and we need to work backwards
     result.context = context.reverse() as BaseMessage[];
@@ -206,8 +241,8 @@ export function getMessagesWithinTokenLimit({
   thinkingStartIndex = originalLength - 1 - assistantIndex;
   const thinkingTokenCount = tokenCounter(new AIMessage({ content: [thinkingBlock] }));
   const newRemainingCount = remainingContextTokens - thinkingTokenCount;
-  const content: MessageContentComplex[] = addThinkingBlock(context[assistantIndex] as AIMessage, thinkingBlock);
-  (context[assistantIndex] as AIMessage).content = content;
+  const newMessage = addThinkingBlock(context[assistantIndex] as AIMessage, thinkingBlock);
+  context[assistantIndex] = newMessage;
   if (newRemainingCount > 0) {
     result.context = context.reverse() as BaseMessage[];
     return result;
@@ -238,19 +273,32 @@ export function getMessagesWithinTokenLimit({
   const firstMessage: AIMessage = newContext[newContext.length - 1];
   const firstMessageType = newContext[newContext.length - 1].getType();
   if (firstMessageType === 'tool') {
-    startType = 'ai';
+    startType = ['ai', 'human'];
   }
 
-  if (startType != null && startType && newContext.length > 0) {
-    const requiredTypeIndex = newContext.findIndex(msg => msg.getType() === startType);
+  if (startType != null && startType.length > 0 && newContext.length > 0) {
+    let requiredTypeIndex = -1;
+
+    let totalTokens = 0;
+    for (let i = newContext.length - 1; i >= 0; i--) {
+      const currentType = newContext[i]?.getType() ?? '';
+      if (Array.isArray(startType) ? startType.includes(currentType) : currentType === startType) {
+        requiredTypeIndex = i + 1;
+        break;
+      }
+      const originalIndex = originalLength - 1 - i;
+      totalTokens += indexTokenCountMap[originalIndex] ?? 0;
+    }
+
     if (requiredTypeIndex > 0) {
-      newContext = newContext.slice(requiredTypeIndex);
+      currentTokenCount -= totalTokens;
+      newContext = newContext.slice(0, requiredTypeIndex);
     }
   }
 
   if (firstMessageType === 'ai') {
-    const content = addThinkingBlock(firstMessage, thinkingBlock);
-    newContext[newContext.length - 1].content = content;
+    const newMessage = addThinkingBlock(firstMessage, thinkingBlock);
+    newContext[newContext.length - 1] = newMessage;
   } else {
     newContext.push(thinkingMessage);
   }
@@ -273,6 +321,7 @@ export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
   let lastTurnStartIndex = factoryParams.startIndex;
   let lastCutOffIndex = 0;
   let totalTokens = (Object.values(indexTokenCountMap)).reduce((a, b) => a + b, 0);
+  let runThinkingStartIndex = -1;
   return function pruneMessages(params: PruneMessagesParams): {
     context: BaseMessage[];
     indexTokenCountMap: Record<string, number>;
@@ -342,7 +391,7 @@ export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
       return { context: params.messages, indexTokenCountMap };
     }
 
-    const { context } = getMessagesWithinTokenLimit({
+    const { context, thinkingStartIndex } = getMessagesWithinTokenLimit({
       maxContextTokens: factoryParams.maxTokens,
       messages: params.messages,
       indexTokenCountMap,
@@ -350,8 +399,11 @@ export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
       thinkingEnabled: factoryParams.thinkingEnabled,
       tokenCounter: factoryParams.tokenCounter,
       reasoningType: factoryParams.provider === Providers.BEDROCK ? ContentTypes.REASONING_CONTENT : ContentTypes.THINKING,
+      thinkingStartIndex: factoryParams.thinkingEnabled === true ? runThinkingStartIndex : undefined,
     });
-    lastCutOffIndex = Math.max(params.messages.length - context.length, 0);
+    runThinkingStartIndex = thinkingStartIndex ?? -1;
+    /** The index is the first value of `context`, index relative to `params.messages` */
+    lastCutOffIndex = Math.max(params.messages.length - (context.length - (context[0]?.getType() === 'system' ? 1 : 0)), 0);
 
     return { context, indexTokenCountMap };
   };

package/src/specs/prune.test.ts

@@ -1,7 +1,7 @@
 // src/specs/prune.test.ts
 import { config } from 'dotenv';
 config();
-import { HumanMessage, AIMessage, SystemMessage, BaseMessage } from '@langchain/core/messages';
+import { HumanMessage, AIMessage, SystemMessage, BaseMessage, ToolMessage } from '@langchain/core/messages';
 import type { RunnableConfig } from '@langchain/core/runnables';
 import type { UsageMetadata } from '@langchain/core/messages';
 import type * as t from '@/types';
@@ -512,6 +512,187 @@ describe('Prune Messages Tests', () => {
     });
   });
 
+  describe('Tool Message Handling', () => {
+    it('should ensure context does not start with a tool message by finding an AI message', () => {
+      const tokenCounter = createTestTokenCounter();
+      const messages = [
+        new SystemMessage('System instruction'),
+        new AIMessage('AI message 1'),
+        new ToolMessage({ content: 'Tool result 1', tool_call_id: 'tool1' }),
+        new AIMessage('AI message 2'),
+        new ToolMessage({ content: 'Tool result 2', tool_call_id: 'tool2' })
+      ];
+
+      const indexTokenCountMap = {
+        0: 17, // System instruction
+        1: 12, // AI message 1
+        2: 13, // Tool result 1
+        3: 12, // AI message 2
+        4: 13  // Tool result 2
+      };
+
+      // Create a pruneMessages function with a token limit that will only include the last few messages
+      const pruneMessages = createPruneMessages({
+        maxTokens: 58, // Only enough for system + last 3 messages + 3, but should not include a parent-less tool message
+        startIndex: 0,
+        tokenCounter,
+        indexTokenCountMap: { ...indexTokenCountMap }
+      });
+
+      const result = pruneMessages({ messages });
+
+      // The context should include the system message, AI message 2, and Tool result 2
+      // It should NOT start with Tool result 2 alone
+      expect(result.context.length).toBe(3);
+      expect(result.context[0]).toBe(messages[0]); // System message
+      expect(result.context[1]).toBe(messages[3]); // AI message 2
+      expect(result.context[2]).toBe(messages[4]); // Tool result 2
+    });
+
+    it('should ensure context does not start with a tool message by finding a human message', () => {
+      const tokenCounter = createTestTokenCounter();
+      const messages = [
+        new SystemMessage('System instruction'),
+        new HumanMessage('Human message 1'),
+        new AIMessage('AI message 1'),
+        new ToolMessage({ content: 'Tool result 1', tool_call_id: 'tool1' }),
+        new HumanMessage('Human message 2'),
+        new ToolMessage({ content: 'Tool result 2', tool_call_id: 'tool2' })
+      ];
+
+      const indexTokenCountMap = {
+        0: 17, // System instruction
+        1: 15, // Human message 1
+        2: 12, // AI message 1
+        3: 13, // Tool result 1
+        4: 15, // Human message 2
+        5: 13  // Tool result 2
+      };
+
+      // Create a pruneMessages function with a token limit that will only include the last few messages
+      const pruneMessages = createPruneMessages({
+        maxTokens: 48, // Only enough for system + last 2 messages
+        startIndex: 0,
+        tokenCounter,
+        indexTokenCountMap: { ...indexTokenCountMap }
+      });
+
+      const result = pruneMessages({ messages });
+
+      // The context should include the system message, Human message 2, and Tool result 2
+      // It should NOT start with Tool result 2 alone
+      expect(result.context.length).toBe(3);
+      expect(result.context[0]).toBe(messages[0]); // System message
+      expect(result.context[1]).toBe(messages[4]); // Human message 2
+      expect(result.context[2]).toBe(messages[5]); // Tool result 2
+    });
+
+    it('should handle the case where a tool message is followed by an AI message', () => {
+      const tokenCounter = createTestTokenCounter();
+      const messages = [
+        new SystemMessage('System instruction'),
+        new HumanMessage('Human message'),
+        new AIMessage('AI message with tool use'),
+        new ToolMessage({ content: 'Tool result', tool_call_id: 'tool1' }),
+        new AIMessage('AI message after tool')
+      ];
+
+      const indexTokenCountMap = {
+        0: 17, // System instruction
+        1: 13, // Human message
+        2: 22, // AI message with tool use
+        3: 11, // Tool result
+        4: 19  // AI message after tool
+      };
+
+      const pruneMessages = createPruneMessages({
+        maxTokens: 50,
+        startIndex: 0,
+        tokenCounter,
+        indexTokenCountMap: { ...indexTokenCountMap }
+      });
+
+      const result = pruneMessages({ messages });
+
+      expect(result.context.length).toBe(2);
+      expect(result.context[0]).toBe(messages[0]); // System message
+      expect(result.context[1]).toBe(messages[4]); // AI message after tool
+    });
+
+    it('should handle the case where a tool message is followed by a human message', () => {
+      const tokenCounter = createTestTokenCounter();
+      const messages = [
+        new SystemMessage('System instruction'),
+        new HumanMessage('Human message 1'),
+        new AIMessage('AI message with tool use'),
+        new ToolMessage({ content: 'Tool result', tool_call_id: 'tool1' }),
+        new HumanMessage('Human message 2')
+      ];
+
+      const indexTokenCountMap = {
+        0: 17, // System instruction
+        1: 15, // Human message 1
+        2: 22, // AI message with tool use
+        3: 11, // Tool result
+        4: 15  // Human message 2
+      };
+
+      const pruneMessages = createPruneMessages({
+        maxTokens: 46,
+        startIndex: 0,
+        tokenCounter,
+        indexTokenCountMap: { ...indexTokenCountMap }
+      });
+
+      const result = pruneMessages({ messages });
+
+      expect(result.context.length).toBe(2);
+      expect(result.context[0]).toBe(messages[0]); // System message
+      expect(result.context[1]).toBe(messages[4]); // Human message 2
+    });
+
+    it('should handle complex sequence with multiple tool messages', () => {
+      const tokenCounter = createTestTokenCounter();
+      const messages = [
+        new SystemMessage('System instruction'),
+        new HumanMessage('Human message 1'),
+        new AIMessage('AI message 1 with tool use'),
+        new ToolMessage({ content: 'Tool result 1', tool_call_id: 'tool1' }),
+        new AIMessage('AI message 2 with tool use'),
+        new ToolMessage({ content: 'Tool result 2', tool_call_id: 'tool2' }),
+        new AIMessage('AI message 3 with tool use'),
+        new ToolMessage({ content: 'Tool result 3', tool_call_id: 'tool3' })
+      ];
+
+      const indexTokenCountMap = {
+        0: 17, // System instruction
+        1: 15, // Human message 1
+        2: 26, // AI message 1 with tool use
+        3: 13, // Tool result 1
+        4: 26, // AI message 2 with tool use
+        5: 13, // Tool result 2
+        6: 26, // AI message 3 with tool use
+        7: 13  // Tool result 3
+      };
+
+      const pruneMessages = createPruneMessages({
+        maxTokens: 111,
+        startIndex: 0,
+        tokenCounter,
+        indexTokenCountMap: { ...indexTokenCountMap }
+      });
+
+      const result = pruneMessages({ messages });
+
+      expect(result.context.length).toBe(5);
+      expect(result.context[0]).toBe(messages[0]); // System message
+      expect(result.context[1]).toBe(messages[4]); // AI message 2 with tool use
+      expect(result.context[2]).toBe(messages[5]); // Tool result 2
+      expect(result.context[3]).toBe(messages[6]); // AI message 3 with tool use
+      expect(result.context[4]).toBe(messages[7]); // Tool result 3
+    });
+  });
+
   describe('Integration with Run', () => {
     it('should initialize Run with custom token counter and process messages', async () => {
       const provider = Providers.OPENAI;

package/src/specs/token-distribution-edge-case.test.ts

@@ -235,9 +235,9 @@ describe('Token Distribution Edge Case Tests', () => {
     });
 
     // Add two more messages
+    messages.push(new HumanMessage('Message 4'));
     const extendedMessages = [
       ...messages,
-      new HumanMessage('Message 4'),
       new AIMessage('Response 4')
     ];
 
@@ -257,6 +257,7 @@ describe('Token Distribution Edge Case Tests', () => {
     // The context should include the system message and some of the latest messages
     expect(thirdResult.context.length).toBeGreaterThan(0);
     expect(thirdResult.context[0].content).toBe('System instruction');
+    expect(thirdResult.context[1].content).toBe('Response 4');
 
     // Find which messages are in the final context
     const contextMessageIndices = thirdResult.context.map(msg => {
@@ -282,14 +283,12 @@ describe('Token Distribution Edge Case Tests', () => {
     // Verify that messages not in the context have their original token counts or previously adjusted values
     for (let i = 0; i < extendedMessages.length; i++) {
       if (!contextMessageIndices.includes(i)) {
-        // This message is not in the context, so its token count should not have been adjusted in the last operation
         const expectedValue = i < messages.length
           ? (secondResult.indexTokenCountMap[i] || indexTokenCountMap[i])
-          : (indexTokenCountMap as Record<string, number | undefined>)[i] ?? indexTokenCountMap[i - 1];
+          : (indexTokenCountMap as Record<string, number | undefined>)[i] ?? 0;
 
-        // For defined values, we can check that they're close to what we expect
         const difference = Math.abs((thirdResult.indexTokenCountMap[i] || 0) - expectedValue);
-        expect(difference).toBeLessThan(20); // Allow for some implementation differences
+        expect(difference).toBe(0);
       }
     }
   });