@librechat/agents 3.2.33 → 3.2.35

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

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","names":[],"sources":["../../../../src/llm/vertexai/index.ts"],"sourcesContent":["import { ChatGoogle } from '@langchain/google-gauth';\nimport { ChatConnection } from '@langchain/google-common';\nimport { AIMessageChunk, isAIMessage } from '@langchain/core/messages';\nimport type {\n  GeminiContent,\n  GeminiRequest,\n  GoogleAIModelRequestParams,\n  GoogleAbstractedClient,\n} from '@langchain/google-common';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { GoogleThinkingConfig, VertexAIClientOptions } from '@/types';\n\n/**\n * `@langchain/google-common`'s `_streamResponseChunks` emits usage on TWO\n * different paths within the same stream:\n *\n *   - Streaming chunks set `chunk.generationInfo.usage_metadata` via\n *     `responseToUsageMetadata`, which correctly sums\n *     `candidatesTokenCount + thoughtsTokenCount` and includes\n *     `output_token_details.reasoning`.\n *   - The trailing fallback chunk (emitted after the API stream exhausts)\n *     attaches its own `chunk.message.usage_metadata` built inline as\n *     `output_tokens = candidatesTokenCount` only — dropping\n *     `thoughtsTokenCount` and `output_token_details` entirely.\n *\n * After `AIMessageChunk.concat`, only `message.usage_metadata` survives —\n * which is the buggy fallback value. This breaks the documented\n * `total_tokens === input_tokens + output_tokens` invariant and silently\n * undercharges thinking models for reasoning tokens.\n *\n * The repair: track the last `generationInfo.usage_metadata` we see, and\n * when the fallback chunk arrives with its buggy `message.usage_metadata`,\n * replace it with the tracked good value. `CustomChatGoogleGenerativeAI`\n * solves the same problem for the Google API path differently — by\n * overriding `_convertToUsageMetadata`.\n */\nexport function repairStreamUsageMetadata(\n  current: UsageMetadata | undefined,\n  generationInfoUsage: UsageMetadata | undefined\n): UsageMetadata | undefined {\n  if (!current) return current;\n  if (!generationInfoUsage) return current;\n  if (generationInfoUsage.total_tokens !== current.total_tokens) return current;\n  if (generationInfoUsage.output_tokens <= current.output_tokens)\n    return current;\n  return generationInfoUsage;\n}\n\ntype AdditionalKwargs =\n  | undefined\n  | (BaseMessage['additional_kwargs'] & {\n      signatures?: Array<string | undefined>;\n    });\n\n/**\n * Fixes thought signatures on functionCall parts in the formatted Gemini request.\n *\n * `@langchain/google-common` stores signatures as a flat array in\n * `additional_kwargs.signatures` (one per response part) and re-attaches them\n * by index only when `signatures.length === parts.length`. This fails when:\n * - The API omits a signature (length mismatch)\n * - Streaming chunks merge with different part counts\n * - The signature for a functionCall part is an empty string\n *\n * This function correlates each \"model\" content block in the formatted request\n * back to its originating AI message by *position*, then re-attaches non-empty\n * signatures that the library failed to apply. AI messages without signatures\n * still consume their slot — filtering them out shifted later messages onto\n * the wrong content block and dropped real signatures on the floor.\n */\nexport function fixThoughtSignatures(\n  contents: GeminiContent[],\n  input: BaseMessage[]\n): void {\n  // All AI messages, in order — non-signature ones still consume positional\n  // slots so later messages line up with their model content blocks.\n  const aiMessages = input.filter(isAIMessage);\n  const modelContents = contents.filter((c) => c.role === 'model');\n\n  const count = Math.min(aiMessages.length, modelContents.length);\n  for (let i = 0; i < count; i++) {\n    const signatures = (aiMessages[i].additional_kwargs as AdditionalKwargs)\n      ?.signatures;\n    if (!Array.isArray(signatures) || signatures.length === 0) continue;\n\n    const content = modelContents[i];\n    const attachedSignatures = new Set(\n      content.parts\n        .map((p) => p.thoughtSignature)\n        .filter((s): s is string => s != null && s !== '')\n    );\n    const availableSignatures = signatures.filter(\n      (s): s is string => s != null && s !== '' && !attachedSignatures.has(s)\n    );\n\n    let sigIdx = 0;\n    for (const part of content.parts) {\n      if (\n        'functionCall' in part &&\n        (part.thoughtSignature == null || part.thoughtSignature === '') &&\n        sigIdx < availableSignatures.length\n      ) {\n        part.thoughtSignature = availableSignatures[sigIdx];\n        sigIdx++;\n      }\n    }\n  }\n}\n\nclass CustomChatConnection extends ChatConnection<VertexAIClientOptions> {\n  thinkingConfig?: GoogleThinkingConfig;\n\n  async formatData(\n    input: BaseMessage[],\n    parameters: GoogleAIModelRequestParams\n  ): Promise<unknown> {\n    const formattedData = (await super.formatData(\n      input,\n      parameters\n    )) as GeminiRequest;\n    if (formattedData.generationConfig?.thinkingConfig?.thinkingBudget === -1) {\n      // -1 means \"let the model decide\" - delete the property so the API doesn't receive an invalid value\n      if (\n        formattedData.generationConfig.thinkingConfig.includeThoughts === false\n      ) {\n        formattedData.generationConfig.thinkingConfig.includeThoughts = true;\n      }\n      delete formattedData.generationConfig.thinkingConfig.thinkingBudget;\n    }\n    if (this.thinkingConfig?.thinkingLevel != null) {\n      formattedData.generationConfig ??= {};\n      // thinkingLevel and thinkingBudget cannot coexist — the API rejects the request.\n      // Remove thinkingBudget when thinkingLevel is set.\n      const { thinkingBudget: _, ...existingThinkingConfig } =\n        (formattedData.generationConfig.thinkingConfig as\n          | Record<string, unknown>\n          | undefined) ?? {};\n      (\n        formattedData.generationConfig as Record<string, unknown>\n      ).thinkingConfig = {\n        ...existingThinkingConfig,\n        thinkingLevel: this.thinkingConfig.thinkingLevel,\n        ...(this.thinkingConfig.includeThoughts != null && {\n          includeThoughts: this.thinkingConfig.includeThoughts,\n        }),\n      };\n    }\n    if (formattedData.contents) {\n      fixThoughtSignatures(formattedData.contents, input);\n      // gemini-3.1+ models reject role=\"function\"; convert to role=\"user\"\n      for (const content of formattedData.contents) {\n        if (content.role === 'function') {\n          (content as { role: string }).role = 'user';\n        }\n      }\n    }\n    return formattedData;\n  }\n}\n\n/**\n * Integration with Google Vertex AI chat models.\n *\n * Setup:\n * Install `@langchain/google-vertexai` and set your stringified\n * Vertex AI credentials as an environment variable named `GOOGLE_APPLICATION_CREDENTIALS`.\n *\n * ```bash\n * npm install @langchain/google-vertexai\n * export GOOGLE_APPLICATION_CREDENTIALS=\"path/to/credentials\"\n * ```\n *\n * ## [Constructor args](https://api.js.langchain.com/classes/_langchain_google_vertexai.index.ChatVertexAI.html#constructor.new_ChatVertexAI)\n *\n * ## [Runtime args](https://api.js.langchain.com/interfaces/langchain_google_common_types.GoogleAIBaseLanguageModelCallOptions.html)\n *\n * Runtime args can be passed as the second argument to any of the base runnable methods `.invoke`. `.stream`, `.batch`, etc.\n * They can also be passed via `.withConfig`, or the second arg in `.bindTools`, like shown in the examples below:\n *\n * ```typescript\n * // When calling `.withConfig`, call options should be passed via the first argument\n * const llmWithArgsBound = llm.withConfig({\n *   stop: [\"\\n\"],\n *   tools: [...],\n * });\n *\n * // When calling `.bindTools`, call options should be passed via the second argument\n * const llmWithTools = llm.bindTools(\n *   [...],\n *   {\n *     tool_choice: \"auto\",\n *   }\n * );\n * ```\n *\n * ## Examples\n *\n * <details open>\n * <summary><strong>Instantiate</strong></summary>\n *\n * ```typescript\n * import { ChatVertexAI } from '@langchain/google-vertexai';\n *\n * const llm = new ChatVertexAI({\n *   model: \"gemini-1.5-pro\",\n *   temperature: 0,\n *   // other params...\n * });\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Invoking</strong></summary>\n *\n * ```typescript\n * const input = `Translate \"I love programming\" into French.`;\n *\n * // Models also accept a list of chat messages or a formatted prompt\n * const result = await llm.invoke(input);\n * console.log(result);\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"J'adore programmer\\\" \\n\\nHere's why this is the best translation:\\n\\n* **J'adore** means \\\"I love\\\" and conveys a strong passion.\\n* **Programmer** is the French verb for \\\"to program.\\\"\\n\\nThis translation is natural and idiomatic in French. \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 63,\n *     \"total_tokens\": 72\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Streaming Chunks</strong></summary>\n *\n * ```typescript\n * for await (const chunk of await llm.stream(input)) {\n *   console.log(chunk);\n * }\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"J'adore programmer\\\" \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {\n *     \"finishReason\": \"stop\"\n *   },\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 8,\n *     \"total_tokens\": 17\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Aggregate Streamed Chunks</strong></summary>\n *\n * ```typescript\n * import { AIMessageChunk } from '@langchain/core/messages';\n * import { concat } from '@langchain/core/utils/stream';\n *\n * const stream = await llm.stream(input);\n * let full: AIMessageChunk | undefined;\n * for await (const chunk of stream) {\n *   full = !full ? chunk : concat(full, chunk);\n * }\n * console.log(full);\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"J'adore programmer\\\" \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {\n *     \"finishReason\": \"stop\"\n *   },\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 8,\n *     \"total_tokens\": 17\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Bind tools</strong></summary>\n *\n * ```typescript\n * import { z } from 'zod';\n *\n * const GetWeather = {\n *   name: \"GetWeather\",\n *   description: \"Get the current weather in a given location\",\n *   schema: z.object({\n *     location: z.string().describe(\"The city and state, e.g. San Francisco, CA\")\n *   }),\n * }\n *\n * const GetPopulation = {\n *   name: \"GetPopulation\",\n *   description: \"Get the current population in a given location\",\n *   schema: z.object({\n *     location: z.string().describe(\"The city and state, e.g. San Francisco, CA\")\n *   }),\n * }\n *\n * const llmWithTools = llm.bindTools([GetWeather, GetPopulation]);\n * const aiMsg = await llmWithTools.invoke(\n *   \"Which city is hotter today and which is bigger: LA or NY?\"\n * );\n * console.log(aiMsg.tool_calls);\n * ```\n *\n * ```txt\n * [\n *   {\n *     name: 'GetPopulation',\n *     args: { location: 'New York City, NY' },\n *     id: '33c1c1f47e2f492799c77d2800a43912',\n *     type: 'tool_call'\n *   }\n * ]\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Structured Output</strong></summary>\n *\n * ```typescript\n * import { z } from 'zod';\n *\n * const Joke = z.object({\n *   setup: z.string().describe(\"The setup of the joke\"),\n *   punchline: z.string().describe(\"The punchline to the joke\"),\n *   rating: z.number().optional().describe(\"How funny the joke is, from 1 to 10\")\n * }).describe('Joke to tell user.');\n *\n * const structuredLlm = llm.withStructuredOutput(Joke, { name: \"Joke\" });\n * const jokeResult = await structuredLlm.invoke(\"Tell me a joke about cats\");\n * console.log(jokeResult);\n * ```\n *\n * ```txt\n * {\n *   setup: 'What do you call a cat that loves to bowl?',\n *   punchline: 'An alley cat!'\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Usage Metadata</strong></summary>\n *\n * ```typescript\n * const aiMsgForMetadata = await llm.invoke(input);\n * console.log(aiMsgForMetadata.usage_metadata);\n * ```\n *\n * ```txt\n * { input_tokens: 9, output_tokens: 8, total_tokens: 17 }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Stream Usage Metadata</strong></summary>\n *\n * ```typescript\n * const streamForMetadata = await llm.stream(\n *   input,\n *   {\n *     streamUsage: true\n *   }\n * );\n * let fullForMetadata: AIMessageChunk | undefined;\n * for await (const chunk of streamForMetadata) {\n *   fullForMetadata = !fullForMetadata ? chunk : concat(fullForMetadata, chunk);\n * }\n * console.log(fullForMetadata?.usage_metadata);\n * ```\n *\n * ```txt\n * { input_tokens: 9, output_tokens: 8, total_tokens: 17 }\n * ```\n * </details>\n *\n * <br />\n */\nexport class ChatVertexAI extends ChatGoogle {\n  lc_namespace = ['langchain', 'chat_models', 'vertexai'];\n  dynamicThinkingBudget = false;\n  thinkingConfig?: GoogleThinkingConfig;\n\n  static lc_name(): 'LibreChatVertexAI' {\n    return 'LibreChatVertexAI';\n  }\n\n  constructor(model: string, fields?: Omit<VertexAIClientOptions, 'model'>);\n  constructor(fields?: VertexAIClientOptions);\n  constructor(\n    modelOrFields?: string | VertexAIClientOptions,\n    params?: Omit<VertexAIClientOptions, 'model'>\n  ) {\n    const fields =\n      typeof modelOrFields === 'string'\n        ? { ...(params ?? {}), model: modelOrFields }\n        : modelOrFields;\n    const dynamicThinkingBudget = fields?.thinkingBudget === -1;\n    super({\n      ...fields,\n      platformType: 'gcp',\n    });\n    this.dynamicThinkingBudget = dynamicThinkingBudget;\n    this.thinkingConfig = fields?.thinkingConfig;\n  }\n  invocationParams(\n    options?: this['ParsedCallOptions'] | undefined\n  ): GoogleAIModelRequestParams {\n    const params = super.invocationParams(options);\n    if (this.dynamicThinkingBudget) {\n      params.maxReasoningTokens = -1;\n    }\n    return params;\n  }\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    let lastGoodUsage: UsageMetadata | undefined;\n    for await (const chunk of super._streamResponseChunks(\n      messages,\n      options,\n      runManager\n    )) {\n      const genUsage = (\n        chunk.generationInfo as { usage_metadata?: UsageMetadata } | undefined\n      )?.usage_metadata;\n      if (genUsage) {\n        lastGoodUsage = genUsage;\n      }\n      if (chunk.message instanceof AIMessageChunk) {\n        const repaired = repairStreamUsageMetadata(\n          chunk.message.usage_metadata,\n          lastGoodUsage\n        );\n        if (repaired !== chunk.message.usage_metadata) {\n          chunk.message.usage_metadata = repaired;\n        }\n      }\n      yield chunk;\n    }\n  }\n  buildConnection(\n    fields: VertexAIClientOptions | undefined,\n    client: GoogleAbstractedClient\n  ): void {\n    // Note: buildConnection is called from super() BEFORE this.thinkingConfig is set,\n    // so we must read thinkingConfig from `fields` directly.\n    const thinkingConfig = fields?.thinkingConfig ?? this.thinkingConfig;\n\n    const connection = new CustomChatConnection(\n      { ...fields, ...this },\n      this.caller,\n      client,\n      false\n    );\n    connection.thinkingConfig = thinkingConfig;\n    this.connection = connection;\n\n    const streamedConnection = new CustomChatConnection(\n      { ...fields, ...this },\n      this.caller,\n      client,\n      true\n    );\n    streamedConnection.thinkingConfig = thinkingConfig;\n    this.streamedConnection = streamedConnection;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsCA,SAAgB,0BACd,SACA,qBAC2B;CAC3B,IAAI,CAAC,SAAS,OAAO;CACrB,IAAI,CAAC,qBAAqB,OAAO;CACjC,IAAI,oBAAoB,iBAAiB,QAAQ,cAAc,OAAO;CACtE,IAAI,oBAAoB,iBAAiB,QAAQ,eAC/C,OAAO;CACT,OAAO;AACT;;;;;;;;;;;;;;;;;AAwBA,SAAgB,qBACd,UACA,OACM;CAGN,MAAM,aAAa,MAAM,OAAO,WAAW;CAC3C,MAAM,gBAAgB,SAAS,QAAQ,MAAM,EAAE,SAAS,OAAO;CAE/D,MAAM,QAAQ,KAAK,IAAI,WAAW,QAAQ,cAAc,MAAM;CAC9D,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;EAC9B,MAAM,aAAc,WAAW,EAAE,CAAC,mBAC9B;EACJ,IAAI,CAAC,MAAM,QAAQ,UAAU,KAAK,WAAW,WAAW,GAAG;EAE3D,MAAM,UAAU,cAAc;EAC9B,MAAM,qBAAqB,IAAI,IAC7B,QAAQ,MACL,KAAK,MAAM,EAAE,gBAAgB,CAAC,CAC9B,QAAQ,MAAmB,KAAK,QAAQ,MAAM,EAAE,CACrD;EACA,MAAM,sBAAsB,WAAW,QACpC,MAAmB,KAAK,QAAQ,MAAM,MAAM,CAAC,mBAAmB,IAAI,CAAC,CACxE;EAEA,IAAI,SAAS;EACb,KAAK,MAAM,QAAQ,QAAQ,OACzB,IACE,kBAAkB,SACjB,KAAK,oBAAoB,QAAQ,KAAK,qBAAqB,OAC5D,SAAS,oBAAoB,QAC7B;GACA,KAAK,mBAAmB,oBAAoB;GAC5C;EACF;CAEJ;AACF;AAEA,IAAM,uBAAN,cAAmC,eAAsC;CACvE;CAEA,MAAM,WACJ,OACA,YACkB;EAClB,MAAM,gBAAiB,MAAM,MAAM,WACjC,OACA,UACF;EACA,IAAI,cAAc,kBAAkB,gBAAgB,mBAAmB,IAAI;GAEzE,IACE,cAAc,iBAAiB,eAAe,oBAAoB,OAElE,cAAc,iBAAiB,eAAe,kBAAkB;GAElE,OAAO,cAAc,iBAAiB,eAAe;EACvD;EACA,IAAI,KAAK,gBAAgB,iBAAiB,MAAM;GAC9C,cAAc,qBAAqB,CAAC;GAGpC,MAAM,EAAE,gBAAgB,GAAG,GAAG,2BAC3B,cAAc,iBAAiB,kBAEd,CAAC;GACrB,cACgB,iBACd,iBAAiB;IACjB,GAAG;IACH,eAAe,KAAK,eAAe;IACnC,GAAI,KAAK,eAAe,mBAAmB,QAAQ,EACjD,iBAAiB,KAAK,eAAe,gBACvC;GACF;EACF;EACA,IAAI,cAAc,UAAU;GAC1B,qBAAqB,cAAc,UAAU,KAAK;GAElD,KAAK,MAAM,WAAW,cAAc,UAClC,IAAI,QAAQ,SAAS,YACnB,QAA8B,OAAO;EAG3C;EACA,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4RA,IAAa,eAAb,cAAkC,WAAW;CAC3C,eAAe;EAAC;EAAa;EAAe;CAAU;CACtD,wBAAwB;CACxB;CAEA,OAAO,UAA+B;EACpC,OAAO;CACT;CAIA,YACE,eACA,QACA;EACA,MAAM,SACJ,OAAO,kBAAkB,WACrB;GAAE,GAAI,UAAU,CAAC;GAAI,OAAO;EAAc,IAC1C;EACN,MAAM,wBAAwB,QAAQ,mBAAmB;EACzD,MAAM;GACJ,GAAG;GACH,cAAc;EAChB,CAAC;EACD,KAAK,wBAAwB;EAC7B,KAAK,iBAAiB,QAAQ;CAChC;CACA,iBACE,SAC4B;EAC5B,MAAM,SAAS,MAAM,iBAAiB,OAAO;EAC7C,IAAI,KAAK,uBACP,OAAO,qBAAqB;EAE9B,OAAO;CACT;CACA,OAAO,sBACL,UACA,SACA,YACqC;EACrC,IAAI;EACJ,WAAW,MAAM,SAAS,MAAM,sBAC9B,UACA,SACA,UACF,GAAG;GACD,MAAM,WACJ,MAAM,gBACL;GACH,IAAI,UACF,gBAAgB;GAElB,IAAI,MAAM,mBAAmB,gBAAgB;IAC3C,MAAM,WAAW,0BACf,MAAM,QAAQ,gBACd,aACF;IACA,IAAI,aAAa,MAAM,QAAQ,gBAC7B,MAAM,QAAQ,iBAAiB;GAEnC;GACA,MAAM;EACR;CACF;CACA,gBACE,QACA,QACM;EAGN,MAAM,iBAAiB,QAAQ,kBAAkB,KAAK;EAEtD,MAAM,aAAa,IAAI,qBACrB;GAAE,GAAG;GAAQ,GAAG;EAAK,GACrB,KAAK,QACL,QACA,KACF;EACA,WAAW,iBAAiB;EAC5B,KAAK,aAAa;EAElB,MAAM,qBAAqB,IAAI,qBAC7B;GAAE,GAAG;GAAQ,GAAG;EAAK,GACrB,KAAK,QACL,QACA,IACF;EACA,mBAAmB,iBAAiB;EACpC,KAAK,qBAAqB;CAC5B;AACF"}
+{"version":3,"file":"index.mjs","names":[],"sources":["../../../../src/llm/vertexai/index.ts"],"sourcesContent":["import { ChatGoogle } from '@langchain/google-gauth';\nimport { ChatConnection } from '@langchain/google-common';\nimport { AIMessageChunk, isAIMessage } from '@langchain/core/messages';\nimport type {\n  GeminiContent,\n  GeminiRequest,\n  GoogleAIModelRequestParams,\n  GoogleAbstractedClient,\n} from '@langchain/google-common';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { GoogleThinkingConfig, VertexAIClientOptions } from '@/types';\nimport {\n  STREAMED_TOOL_CALL_SEAL_METADATA_KEY,\n  STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY,\n  GOOGLE_STREAMED_TOOL_CALL_ADAPTER,\n} from '@/tools/streamedToolCallSeals';\n\n/**\n * `@langchain/google-common`'s `_streamResponseChunks` emits usage on TWO\n * different paths within the same stream:\n *\n *   - Streaming chunks set `chunk.generationInfo.usage_metadata` via\n *     `responseToUsageMetadata`, which correctly sums\n *     `candidatesTokenCount + thoughtsTokenCount` and includes\n *     `output_token_details.reasoning`.\n *   - The trailing fallback chunk (emitted after the API stream exhausts)\n *     attaches its own `chunk.message.usage_metadata` built inline as\n *     `output_tokens = candidatesTokenCount` only — dropping\n *     `thoughtsTokenCount` and `output_token_details` entirely.\n *\n * After `AIMessageChunk.concat`, only `message.usage_metadata` survives —\n * which is the buggy fallback value. This breaks the documented\n * `total_tokens === input_tokens + output_tokens` invariant and silently\n * undercharges thinking models for reasoning tokens.\n *\n * The repair: track the last `generationInfo.usage_metadata` we see, and\n * when the fallback chunk arrives with its buggy `message.usage_metadata`,\n * replace it with the tracked good value. `CustomChatGoogleGenerativeAI`\n * solves the same problem for the Google API path differently — by\n * overriding `_convertToUsageMetadata`.\n */\nexport function repairStreamUsageMetadata(\n  current: UsageMetadata | undefined,\n  generationInfoUsage: UsageMetadata | undefined\n): UsageMetadata | undefined {\n  if (!current) return current;\n  if (!generationInfoUsage) return current;\n  if (generationInfoUsage.total_tokens !== current.total_tokens) return current;\n  if (generationInfoUsage.output_tokens <= current.output_tokens)\n    return current;\n  return generationInfoUsage;\n}\n\n/**\n * The Gemini API delivers function calls as complete objects — never as\n * partial arg deltas. `@langchain/google-common` pre-parses each streamed\n * functionCall part into `tool_calls` (invalid args land in\n * `invalid_tool_calls` instead), so a chunk whose tool-call chunks all parsed\n * cleanly is sealed on arrival for eager tool execution. Anything that fails\n * the parse check is left unstamped and falls back to the lazy path.\n */\nexport function sealCompleteStreamedToolCalls(message: AIMessageChunk): void {\n  const chunkCount = message.tool_call_chunks?.length ?? 0;\n  if (\n    chunkCount === 0 ||\n    (message.invalid_tool_calls?.length ?? 0) > 0 ||\n    (message.tool_calls?.length ?? 0) !== chunkCount\n  ) {\n    return;\n  }\n  message.response_metadata = {\n    ...message.response_metadata,\n    [STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY]:\n      GOOGLE_STREAMED_TOOL_CALL_ADAPTER,\n    [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'all' },\n  };\n}\n\ntype AdditionalKwargs =\n  | undefined\n  | (BaseMessage['additional_kwargs'] & {\n      signatures?: Array<string | undefined>;\n    });\n\n/**\n * Fixes thought signatures on functionCall parts in the formatted Gemini request.\n *\n * `@langchain/google-common` stores signatures as a flat array in\n * `additional_kwargs.signatures` (one per response part) and re-attaches them\n * by index only when `signatures.length === parts.length`. This fails when:\n * - The API omits a signature (length mismatch)\n * - Streaming chunks merge with different part counts\n * - The signature for a functionCall part is an empty string\n *\n * This function correlates each \"model\" content block in the formatted request\n * back to its originating AI message by *position*, then re-attaches non-empty\n * signatures that the library failed to apply. AI messages without signatures\n * still consume their slot — filtering them out shifted later messages onto\n * the wrong content block and dropped real signatures on the floor.\n */\nexport function fixThoughtSignatures(\n  contents: GeminiContent[],\n  input: BaseMessage[]\n): void {\n  // All AI messages, in order — non-signature ones still consume positional\n  // slots so later messages line up with their model content blocks.\n  const aiMessages = input.filter(isAIMessage);\n  const modelContents = contents.filter((c) => c.role === 'model');\n\n  const count = Math.min(aiMessages.length, modelContents.length);\n  for (let i = 0; i < count; i++) {\n    const signatures = (aiMessages[i].additional_kwargs as AdditionalKwargs)\n      ?.signatures;\n    if (!Array.isArray(signatures) || signatures.length === 0) continue;\n\n    const content = modelContents[i];\n    const attachedSignatures = new Set(\n      content.parts\n        .map((p) => p.thoughtSignature)\n        .filter((s): s is string => s != null && s !== '')\n    );\n    const availableSignatures = signatures.filter(\n      (s): s is string => s != null && s !== '' && !attachedSignatures.has(s)\n    );\n\n    let sigIdx = 0;\n    for (const part of content.parts) {\n      if (\n        'functionCall' in part &&\n        (part.thoughtSignature == null || part.thoughtSignature === '') &&\n        sigIdx < availableSignatures.length\n      ) {\n        part.thoughtSignature = availableSignatures[sigIdx];\n        sigIdx++;\n      }\n    }\n  }\n}\n\nclass CustomChatConnection extends ChatConnection<VertexAIClientOptions> {\n  thinkingConfig?: GoogleThinkingConfig;\n\n  async formatData(\n    input: BaseMessage[],\n    parameters: GoogleAIModelRequestParams\n  ): Promise<unknown> {\n    const formattedData = (await super.formatData(\n      input,\n      parameters\n    )) as GeminiRequest;\n    if (formattedData.generationConfig?.thinkingConfig?.thinkingBudget === -1) {\n      // -1 means \"let the model decide\" - delete the property so the API doesn't receive an invalid value\n      if (\n        formattedData.generationConfig.thinkingConfig.includeThoughts === false\n      ) {\n        formattedData.generationConfig.thinkingConfig.includeThoughts = true;\n      }\n      delete formattedData.generationConfig.thinkingConfig.thinkingBudget;\n    }\n    if (this.thinkingConfig?.thinkingLevel != null) {\n      formattedData.generationConfig ??= {};\n      // thinkingLevel and thinkingBudget cannot coexist — the API rejects the request.\n      // Remove thinkingBudget when thinkingLevel is set.\n      const { thinkingBudget: _, ...existingThinkingConfig } =\n        (formattedData.generationConfig.thinkingConfig as\n          | Record<string, unknown>\n          | undefined) ?? {};\n      (\n        formattedData.generationConfig as Record<string, unknown>\n      ).thinkingConfig = {\n        ...existingThinkingConfig,\n        thinkingLevel: this.thinkingConfig.thinkingLevel,\n        ...(this.thinkingConfig.includeThoughts != null && {\n          includeThoughts: this.thinkingConfig.includeThoughts,\n        }),\n      };\n    }\n    if (formattedData.contents) {\n      fixThoughtSignatures(formattedData.contents, input);\n      // gemini-3.1+ models reject role=\"function\"; convert to role=\"user\"\n      for (const content of formattedData.contents) {\n        if (content.role === 'function') {\n          (content as { role: string }).role = 'user';\n        }\n      }\n    }\n    return formattedData;\n  }\n}\n\n/**\n * Integration with Google Vertex AI chat models.\n *\n * Setup:\n * Install `@langchain/google-vertexai` and set your stringified\n * Vertex AI credentials as an environment variable named `GOOGLE_APPLICATION_CREDENTIALS`.\n *\n * ```bash\n * npm install @langchain/google-vertexai\n * export GOOGLE_APPLICATION_CREDENTIALS=\"path/to/credentials\"\n * ```\n *\n * ## [Constructor args](https://api.js.langchain.com/classes/_langchain_google_vertexai.index.ChatVertexAI.html#constructor.new_ChatVertexAI)\n *\n * ## [Runtime args](https://api.js.langchain.com/interfaces/langchain_google_common_types.GoogleAIBaseLanguageModelCallOptions.html)\n *\n * Runtime args can be passed as the second argument to any of the base runnable methods `.invoke`. `.stream`, `.batch`, etc.\n * They can also be passed via `.withConfig`, or the second arg in `.bindTools`, like shown in the examples below:\n *\n * ```typescript\n * // When calling `.withConfig`, call options should be passed via the first argument\n * const llmWithArgsBound = llm.withConfig({\n *   stop: [\"\\n\"],\n *   tools: [...],\n * });\n *\n * // When calling `.bindTools`, call options should be passed via the second argument\n * const llmWithTools = llm.bindTools(\n *   [...],\n *   {\n *     tool_choice: \"auto\",\n *   }\n * );\n * ```\n *\n * ## Examples\n *\n * <details open>\n * <summary><strong>Instantiate</strong></summary>\n *\n * ```typescript\n * import { ChatVertexAI } from '@langchain/google-vertexai';\n *\n * const llm = new ChatVertexAI({\n *   model: \"gemini-1.5-pro\",\n *   temperature: 0,\n *   // other params...\n * });\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Invoking</strong></summary>\n *\n * ```typescript\n * const input = `Translate \"I love programming\" into French.`;\n *\n * // Models also accept a list of chat messages or a formatted prompt\n * const result = await llm.invoke(input);\n * console.log(result);\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"J'adore programmer\\\" \\n\\nHere's why this is the best translation:\\n\\n* **J'adore** means \\\"I love\\\" and conveys a strong passion.\\n* **Programmer** is the French verb for \\\"to program.\\\"\\n\\nThis translation is natural and idiomatic in French. \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 63,\n *     \"total_tokens\": 72\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Streaming Chunks</strong></summary>\n *\n * ```typescript\n * for await (const chunk of await llm.stream(input)) {\n *   console.log(chunk);\n * }\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"J'adore programmer\\\" \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {\n *     \"finishReason\": \"stop\"\n *   },\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 8,\n *     \"total_tokens\": 17\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Aggregate Streamed Chunks</strong></summary>\n *\n * ```typescript\n * import { AIMessageChunk } from '@langchain/core/messages';\n * import { concat } from '@langchain/core/utils/stream';\n *\n * const stream = await llm.stream(input);\n * let full: AIMessageChunk | undefined;\n * for await (const chunk of stream) {\n *   full = !full ? chunk : concat(full, chunk);\n * }\n * console.log(full);\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"J'adore programmer\\\" \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {\n *     \"finishReason\": \"stop\"\n *   },\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 8,\n *     \"total_tokens\": 17\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Bind tools</strong></summary>\n *\n * ```typescript\n * import { z } from 'zod';\n *\n * const GetWeather = {\n *   name: \"GetWeather\",\n *   description: \"Get the current weather in a given location\",\n *   schema: z.object({\n *     location: z.string().describe(\"The city and state, e.g. San Francisco, CA\")\n *   }),\n * }\n *\n * const GetPopulation = {\n *   name: \"GetPopulation\",\n *   description: \"Get the current population in a given location\",\n *   schema: z.object({\n *     location: z.string().describe(\"The city and state, e.g. San Francisco, CA\")\n *   }),\n * }\n *\n * const llmWithTools = llm.bindTools([GetWeather, GetPopulation]);\n * const aiMsg = await llmWithTools.invoke(\n *   \"Which city is hotter today and which is bigger: LA or NY?\"\n * );\n * console.log(aiMsg.tool_calls);\n * ```\n *\n * ```txt\n * [\n *   {\n *     name: 'GetPopulation',\n *     args: { location: 'New York City, NY' },\n *     id: '33c1c1f47e2f492799c77d2800a43912',\n *     type: 'tool_call'\n *   }\n * ]\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Structured Output</strong></summary>\n *\n * ```typescript\n * import { z } from 'zod';\n *\n * const Joke = z.object({\n *   setup: z.string().describe(\"The setup of the joke\"),\n *   punchline: z.string().describe(\"The punchline to the joke\"),\n *   rating: z.number().optional().describe(\"How funny the joke is, from 1 to 10\")\n * }).describe('Joke to tell user.');\n *\n * const structuredLlm = llm.withStructuredOutput(Joke, { name: \"Joke\" });\n * const jokeResult = await structuredLlm.invoke(\"Tell me a joke about cats\");\n * console.log(jokeResult);\n * ```\n *\n * ```txt\n * {\n *   setup: 'What do you call a cat that loves to bowl?',\n *   punchline: 'An alley cat!'\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Usage Metadata</strong></summary>\n *\n * ```typescript\n * const aiMsgForMetadata = await llm.invoke(input);\n * console.log(aiMsgForMetadata.usage_metadata);\n * ```\n *\n * ```txt\n * { input_tokens: 9, output_tokens: 8, total_tokens: 17 }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Stream Usage Metadata</strong></summary>\n *\n * ```typescript\n * const streamForMetadata = await llm.stream(\n *   input,\n *   {\n *     streamUsage: true\n *   }\n * );\n * let fullForMetadata: AIMessageChunk | undefined;\n * for await (const chunk of streamForMetadata) {\n *   fullForMetadata = !fullForMetadata ? chunk : concat(fullForMetadata, chunk);\n * }\n * console.log(fullForMetadata?.usage_metadata);\n * ```\n *\n * ```txt\n * { input_tokens: 9, output_tokens: 8, total_tokens: 17 }\n * ```\n * </details>\n *\n * <br />\n */\nexport class ChatVertexAI extends ChatGoogle {\n  lc_namespace = ['langchain', 'chat_models', 'vertexai'];\n  dynamicThinkingBudget = false;\n  thinkingConfig?: GoogleThinkingConfig;\n\n  static lc_name(): 'LibreChatVertexAI' {\n    return 'LibreChatVertexAI';\n  }\n\n  constructor(model: string, fields?: Omit<VertexAIClientOptions, 'model'>);\n  constructor(fields?: VertexAIClientOptions);\n  constructor(\n    modelOrFields?: string | VertexAIClientOptions,\n    params?: Omit<VertexAIClientOptions, 'model'>\n  ) {\n    const fields =\n      typeof modelOrFields === 'string'\n        ? { ...(params ?? {}), model: modelOrFields }\n        : modelOrFields;\n    const dynamicThinkingBudget = fields?.thinkingBudget === -1;\n    super({\n      ...fields,\n      platformType: 'gcp',\n    });\n    this.dynamicThinkingBudget = dynamicThinkingBudget;\n    this.thinkingConfig = fields?.thinkingConfig;\n  }\n  invocationParams(\n    options?: this['ParsedCallOptions'] | undefined\n  ): GoogleAIModelRequestParams {\n    const params = super.invocationParams(options);\n    if (this.dynamicThinkingBudget) {\n      params.maxReasoningTokens = -1;\n    }\n    return params;\n  }\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    let lastGoodUsage: UsageMetadata | undefined;\n    for await (const chunk of super._streamResponseChunks(\n      messages,\n      options,\n      runManager\n    )) {\n      const genUsage = (\n        chunk.generationInfo as { usage_metadata?: UsageMetadata } | undefined\n      )?.usage_metadata;\n      if (genUsage) {\n        lastGoodUsage = genUsage;\n      }\n      if (chunk.message instanceof AIMessageChunk) {\n        const repaired = repairStreamUsageMetadata(\n          chunk.message.usage_metadata,\n          lastGoodUsage\n        );\n        if (repaired !== chunk.message.usage_metadata) {\n          chunk.message.usage_metadata = repaired;\n        }\n        sealCompleteStreamedToolCalls(chunk.message);\n      }\n      yield chunk;\n    }\n  }\n  buildConnection(\n    fields: VertexAIClientOptions | undefined,\n    client: GoogleAbstractedClient\n  ): void {\n    // Note: buildConnection is called from super() BEFORE this.thinkingConfig is set,\n    // so we must read thinkingConfig from `fields` directly.\n    const thinkingConfig = fields?.thinkingConfig ?? this.thinkingConfig;\n\n    const connection = new CustomChatConnection(\n      { ...fields, ...this },\n      this.caller,\n      client,\n      false\n    );\n    connection.thinkingConfig = thinkingConfig;\n    this.connection = connection;\n\n    const streamedConnection = new CustomChatConnection(\n      { ...fields, ...this },\n      this.caller,\n      client,\n      true\n    );\n    streamedConnection.thinkingConfig = thinkingConfig;\n    this.streamedConnection = streamedConnection;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2CA,SAAgB,0BACd,SACA,qBAC2B;CAC3B,IAAI,CAAC,SAAS,OAAO;CACrB,IAAI,CAAC,qBAAqB,OAAO;CACjC,IAAI,oBAAoB,iBAAiB,QAAQ,cAAc,OAAO;CACtE,IAAI,oBAAoB,iBAAiB,QAAQ,eAC/C,OAAO;CACT,OAAO;AACT;;;;;;;;;AAUA,SAAgB,8BAA8B,SAA+B;CAC3E,MAAM,aAAa,QAAQ,kBAAkB,UAAU;CACvD,IACE,eAAe,MACd,QAAQ,oBAAoB,UAAU,KAAK,MAC3C,QAAQ,YAAY,UAAU,OAAO,YAEtC;CAEF,QAAQ,oBAAoB;EAC1B,GAAG,QAAQ;GACV,0CACC;GACD,uCAAuC,EAAE,MAAM,MAAM;CACxD;AACF;;;;;;;;;;;;;;;;;AAwBA,SAAgB,qBACd,UACA,OACM;CAGN,MAAM,aAAa,MAAM,OAAO,WAAW;CAC3C,MAAM,gBAAgB,SAAS,QAAQ,MAAM,EAAE,SAAS,OAAO;CAE/D,MAAM,QAAQ,KAAK,IAAI,WAAW,QAAQ,cAAc,MAAM;CAC9D,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;EAC9B,MAAM,aAAc,WAAW,EAAE,CAAC,mBAC9B;EACJ,IAAI,CAAC,MAAM,QAAQ,UAAU,KAAK,WAAW,WAAW,GAAG;EAE3D,MAAM,UAAU,cAAc;EAC9B,MAAM,qBAAqB,IAAI,IAC7B,QAAQ,MACL,KAAK,MAAM,EAAE,gBAAgB,CAAC,CAC9B,QAAQ,MAAmB,KAAK,QAAQ,MAAM,EAAE,CACrD;EACA,MAAM,sBAAsB,WAAW,QACpC,MAAmB,KAAK,QAAQ,MAAM,MAAM,CAAC,mBAAmB,IAAI,CAAC,CACxE;EAEA,IAAI,SAAS;EACb,KAAK,MAAM,QAAQ,QAAQ,OACzB,IACE,kBAAkB,SACjB,KAAK,oBAAoB,QAAQ,KAAK,qBAAqB,OAC5D,SAAS,oBAAoB,QAC7B;GACA,KAAK,mBAAmB,oBAAoB;GAC5C;EACF;CAEJ;AACF;AAEA,IAAM,uBAAN,cAAmC,eAAsC;CACvE;CAEA,MAAM,WACJ,OACA,YACkB;EAClB,MAAM,gBAAiB,MAAM,MAAM,WACjC,OACA,UACF;EACA,IAAI,cAAc,kBAAkB,gBAAgB,mBAAmB,IAAI;GAEzE,IACE,cAAc,iBAAiB,eAAe,oBAAoB,OAElE,cAAc,iBAAiB,eAAe,kBAAkB;GAElE,OAAO,cAAc,iBAAiB,eAAe;EACvD;EACA,IAAI,KAAK,gBAAgB,iBAAiB,MAAM;GAC9C,cAAc,qBAAqB,CAAC;GAGpC,MAAM,EAAE,gBAAgB,GAAG,GAAG,2BAC3B,cAAc,iBAAiB,kBAEd,CAAC;GACrB,cACgB,iBACd,iBAAiB;IACjB,GAAG;IACH,eAAe,KAAK,eAAe;IACnC,GAAI,KAAK,eAAe,mBAAmB,QAAQ,EACjD,iBAAiB,KAAK,eAAe,gBACvC;GACF;EACF;EACA,IAAI,cAAc,UAAU;GAC1B,qBAAqB,cAAc,UAAU,KAAK;GAElD,KAAK,MAAM,WAAW,cAAc,UAClC,IAAI,QAAQ,SAAS,YACnB,QAA8B,OAAO;EAG3C;EACA,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4RA,IAAa,eAAb,cAAkC,WAAW;CAC3C,eAAe;EAAC;EAAa;EAAe;CAAU;CACtD,wBAAwB;CACxB;CAEA,OAAO,UAA+B;EACpC,OAAO;CACT;CAIA,YACE,eACA,QACA;EACA,MAAM,SACJ,OAAO,kBAAkB,WACrB;GAAE,GAAI,UAAU,CAAC;GAAI,OAAO;EAAc,IAC1C;EACN,MAAM,wBAAwB,QAAQ,mBAAmB;EACzD,MAAM;GACJ,GAAG;GACH,cAAc;EAChB,CAAC;EACD,KAAK,wBAAwB;EAC7B,KAAK,iBAAiB,QAAQ;CAChC;CACA,iBACE,SAC4B;EAC5B,MAAM,SAAS,MAAM,iBAAiB,OAAO;EAC7C,IAAI,KAAK,uBACP,OAAO,qBAAqB;EAE9B,OAAO;CACT;CACA,OAAO,sBACL,UACA,SACA,YACqC;EACrC,IAAI;EACJ,WAAW,MAAM,SAAS,MAAM,sBAC9B,UACA,SACA,UACF,GAAG;GACD,MAAM,WACJ,MAAM,gBACL;GACH,IAAI,UACF,gBAAgB;GAElB,IAAI,MAAM,mBAAmB,gBAAgB;IAC3C,MAAM,WAAW,0BACf,MAAM,QAAQ,gBACd,aACF;IACA,IAAI,aAAa,MAAM,QAAQ,gBAC7B,MAAM,QAAQ,iBAAiB;IAEjC,8BAA8B,MAAM,OAAO;GAC7C;GACA,MAAM;EACR;CACF;CACA,gBACE,QACA,QACM;EAGN,MAAM,iBAAiB,QAAQ,kBAAkB,KAAK;EAEtD,MAAM,aAAa,IAAI,qBACrB;GAAE,GAAG;GAAQ,GAAG;EAAK,GACrB,KAAK,QACL,QACA,KACF;EACA,WAAW,iBAAiB;EAC5B,KAAK,aAAa;EAElB,MAAM,qBAAqB,IAAI,qBAC7B;GAAE,GAAG;GAAQ,GAAG;EAAK,GACrB,KAAK,QACL,QACA,IACF;EACA,mBAAmB,iBAAiB;EACpC,KAAK,qBAAqB;CAC5B;AACF"}

package/dist/esm/main.mjs

@@ -2,7 +2,7 @@ import { isPresent, unescapeObject } from "./utils/misc.mjs";
 import { ANTHROPIC_TOOL_TOKEN_MULTIPLIER, DEFAULT_TOOL_TOKEN_MULTIPLIER } from "./common/constants.mjs";
 import { CODE_EXECUTION_TOOLS, Callback, CommonEvents, Constants, ContentTypes, EnvVar, GraphEvents, GraphNodeActions, GraphNodeKeys, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, Providers, StepTypes, TitleMethod, ToolCallTypes } from "./common/enum.mjs";
 import "./common/index.mjs";
-import { TokenEncoderManager, createTokenCounter, encodingForModel, estimateAnthropicImageTokens, estimateOpenAIImageTokens, extractImageDimensions, getTokenCountForMessage } from "./utils/tokens.mjs";
+import { TokenEncoderManager, apportionTokenCounts, createTokenCounter, encodingForModel, estimateAnthropicImageTokens, estimateOpenAIImageTokens, extractImageDimensions, getTokenCountForMessage } from "./utils/tokens.mjs";
 import { convertMessagesToContent, findLastIndex, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, getConverseOverrideMessage, modifyDeltaProperties } from "./messages/core.mjs";
 import { getMessageId } from "./messages/ids.mjs";
 import { HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, truncateToolInput, truncateToolResultContent } from "./utils/truncation.mjs";
@@ -12,7 +12,7 @@ import { DEFAULT_RESERVE_RATIO, ORIGINAL_CONTENT_MAX_CHARS, calculateTotalTokens
 import { ensureThinkingBlockInMessages, formatAgentMessages, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, labelContentByAgent, shiftIndexTokenCountMap, withMessageRole } from "./messages/format.mjs";
 import { addBedrockCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, stripAnthropicCacheControl, stripBedrockCacheControl } from "./messages/cache.mjs";
 import { makeIsDeferred, partitionAndMarkAnthropicToolCache } from "./messages/anthropicToolCache.mjs";
-import { formatContentStrings } from "./messages/content.mjs";
+import { formatContentStrings, isLegacyConvertible } from "./messages/content.mjs";
 import { extractToolDiscoveries, hasToolSearchInCurrentTurn } from "./messages/tools.mjs";
 import { REMOVE_ALL_MESSAGES, createRemoveAllMessage, messagesStateReducer } from "./messages/reducer.mjs";
 import { splitAtRecencyBoundary } from "./messages/recency.mjs";
@@ -96,4 +96,4 @@ import { Runnable, RunnableLambda, RunnableSequence } from "./langchain/runnable
 import { DynamicStructuredTool, StructuredTool, Tool, tool } from "./langchain/tools.mjs";
 import "./langchain/index.mjs";
 import { BaseCheckpointSaver, Command, INTERRUPT, MemorySaver, interrupt, isInterrupted } from "@langchain/langgraph";
-export { AIMessage, AIMessageChunk, ANTHROPIC_TOOL_TOKEN_MULTIPLIER, AgentSession, BASH_SHELL_GUIDANCE, BaseCheckpointSaver, BaseMessage, BaseMessageChunk, BashExecutionToolDefinition, BashExecutionToolDescription, BashExecutionToolName, BashExecutionToolSchema, BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, BashToolOutputReferencesGuide, CLOUDFLARE_BASH_CODING_TOOL_NAMES, CLOUDFLARE_CODING_TOOL_NAMES, CODE_ARTIFACT_PATH_GUIDANCE, CODE_EXECUTION_TOOLS, Calculator, CalculatorSchema, CalculatorToolDefinition, CalculatorToolDescription, CalculatorToolName, Callback, ChatModelStreamHandler, ChatOpenRouter, CloudflareBashExecutionToolDescription, CloudflareCodeExecutionToolDescription, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, Command, CommonEvents, CompileCheckToolName, Constants, ContentTypes, CustomOpenAIClient, DATE_RANGE, DEFAULT_CONTEXT_PRUNING_SETTINGS, DEFAULT_COUNTRY_DESCRIPTION, DEFAULT_HOOK_TIMEOUT_MS, DEFAULT_QUERY_DESCRIPTION, DEFAULT_RESERVE_RATIO, DEFAULT_TOOL_TOKEN_MULTIPLIER, DynamicStructuredTool, EnvVar, FAILED_EXECUTION_FILE_REMINDER, FakeChatModel, Graph, GraphEvents, GraphNodeActions, GraphNodeKeys, HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, HOOK_EVENTS, HandlerRegistry, HookRegistry, HumanMessage, INTERRUPT, JsonlSessionStore, LLMStreamHandler, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, LOCAL_SPAWN_TIMEOUT_MS, LocalBashExecutionToolDescription, LocalCodeExecutionToolDescription, LocalEditFileToolName, LocalEditFileToolSchema, LocalFileCheckpointerImpl, LocalGlobSearchToolName, LocalGlobSearchToolSchema, LocalGrepSearchToolName, LocalGrepSearchToolSchema, LocalListDirectoryToolName, LocalListDirectoryToolSchema, LocalReadFileToolSchema, LocalWriteFileToolName, LocalWriteFileToolSchema, MAX_CACHE_SIZE, MAX_PATTERN_LENGTH, MemorySaver, ModelEndHandler, MultiAgentGraph, ORIGINAL_CONTENT_MAX_CHARS, ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, PromptTemplate, Providers, REMOVE_ALL_MESSAGES, ReadFileToolDefinition, ReadFileToolDescription, ReadFileToolName, ReadFileToolSchema, Run, Runnable, RunnableCallable, RunnableLambda, RunnableSequence, SEPARATORS, SessionManager, SkillToolDefinition, SkillToolDescription, SkillToolName, SkillToolSchema, SplitStreamHandler, StandardGraph, StepTypes, StructuredTool, SubagentExecutor, SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, SystemMessage, TMP_SCRATCH_OUTPUT_REMINDER, TestChatStreamHandler, TestLLMStreamHandler, TitleMethod, TokenEncoderManager, Tool, ToolCallTypes, ToolEndHandler, ToolMessage, ToolNode, ToolSearchToolDefinition, ToolSearchToolDescription, ToolSearchToolName, ToolSearchToolSchema, WebSearchToolDefinition, WebSearchToolDescription, WebSearchToolName, WebSearchToolSchema, _createBashProgramForTests, _resetLocalEngineWarningsForTests, _resetRipgrepCacheForTests, _resetSyntaxCheckProbeCacheForTests, _resetUnrecognizedTriggerWarnings, addBedrockCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, appendCodeSessionFileSummary, appendFailedExecutionFileReminder, appendTmpScratchReminder, applyContextPruning, applyEdit, applyPreToolUseHooksForBridge, askUserQuestion, attemptInvoke, bashAstFindingsToErrors, buildBashExecutionToolDescription, buildChildInputs, buildCodeApiHttpErrorMessage, buildSandboxRuntimeConfig, buildSubagentToolParams, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, calculateTotalTokens, checkValidNumber, classifyAttachment, composeEventHandlers, convertMessagesToContent, countNestedGroups, countrySchema, createAgentSession, createBashExecutionTool, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, createCloudflareBashExecutionTool, createCloudflareBashProgrammaticToolCallingTool, createCloudflareBridgeRuntime, createCloudflareCodeExecutionTool, createCloudflareCodingToolBundle, createCloudflareCodingTools, createCloudflareExecutionTool, createCloudflareLocalExecutionConfig, createCloudflareProgrammaticToolCallingTool, createCloudflareWorkspaceFS, createCodeExecutionTool, createCompileCheckTool, createCompileCheckToolDefinition, createContentAggregator, createFakeStreamingLLM, createHandlers, createLocalBashExecutionTool, createLocalBashProgrammaticToolCallingTool, createLocalCodeExecutionTool, createLocalCodingToolBundle, createLocalCodingToolDefinitions, createLocalCodingToolRegistry, createLocalCodingTools, createLocalEditFileTool, createLocalFileCheckpointer, createLocalGlobSearchTool, createLocalGrepSearchTool, createLocalListDirectoryTool, createLocalProgrammaticToolCallingTool, createLocalReadFileTool, createLocalWriteFileTool, createMetadataAggregator, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, createPruneMessages, createRemoveAllMessage, createRunHandlers, createSchemaOnlyTool, createSchemaOnlyTools, createSearchTool, createSubagentToolDefinition, createTokenCounter, createToolPolicyHook, createToolSearch, createWorkspacePolicyHook, dateSchema, decodeFile, defaultOmitOptions, deserializeMessage, emptyOutputMessage, encodeFile, encodingForModel, enforceOriginalContentCap, ensureThinkingBlockInMessages, escapeRegexSpecialChars, estimateAnthropicImageTokens, estimateOpenAIImageTokens, executeCloudflareBash, executeCloudflareCode, executeHooks, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, executeParallelSearches, executeTools, extractErrorMessage, extractImageDimensions, extractMcpServerName, extractTextFromContent, extractToolDiscoveries, extractUsedBashToolNames, extractUsedToolNames, fetchSessionFiles, filterBashToolsByUsage, filterSubagentResult, filterToolsByUsage, findLastIndex, formatAgentMessages, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, formatCloudflareOutput, formatCompletedResponse, formatContentStrings, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, formatServerListing, formatSkillCatalog, getAvailableMcpServers, getBaseToolName, getBufferString, getChatModelClass, getChunkContent, getCloudflareWorkspaceRoot, getCodeBaseURL, getConverseOverrideMessage, getDeferredToolsListing, getLocalCwd, getLocalSessionId, getMaxOutputTokensKey, getMessageId, getMessagesWithinTokenLimit, getReadRoots, getSpawn, getTokenCountForMessage, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, handleServerToolResult, handleToolCallChunks, handleToolCalls, hasNestedQuantifier, hasNestedQuantifiers, hasToolSearchInCurrentTurn, imageAttachmentContent, imagesSchema, initializeModel, interrupt, isAIMessage, isAnthropicLike, isBaseMessage, isContextOverflowError, isDangerousPattern, isFromAnyMcpServer, isFromMcpServer, isGoogleLike, isInterrupted, isLikelyContextOverflowError, isOpenAILike, isPresent, isThinkingEnabled, isToolMessage, isZodSchema, joinKeys, labelContentByAgent, locateEdit, makeIsDeferred, makeRequest, maskConsumedToolResults, matchesQuery, messagesStateReducer, modifyDeltaProperties, newsSchema, normalizeBashToolResultsForReplay, normalizeServerFilter, normalizeToBashIdentifier, normalizeToPythonIdentifier, partitionAndMarkAnthropicToolCache, performLocalSearch, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, querySchema, repairOrphanedToolMessages, resetIfNotEmpty, resolveCloudflareSandbox, resolveCodeApiAuthHeaders, resolveContextPruningSettings, resolveLocalExecutionConfig, resolveLocalExecutionTools, resolveLocalToolRegistry, resolveLocalToolsForBinding, resolveSubagentConfigs, resolveWorkspacePath, resolveWorkspacePathSafe, runBashAstChecks, runPostEditSyntaxCheck, sanitizeOrphanToolBlocks, sanitizeRegex, serializeMessage, shellQuote, shiftIndexTokenCountMap, shouldTriggerSummarization, sleep, spawnLocalProcess, splitAtRecencyBoundary, stripAnthropicCacheControl, stripBedrockCacheControl, stripCodeSessionFileSummary, summarizeEvent, toJsonSchema, tool, toolResultTypes, toolsCondition, truncateLocalOutput, truncateToolInput, truncateToolResultContent, tryFallbackProviders, unescapeObject, unwrapToolResponse, validateBashCommand, validateCloudflareBashCommand, videosSchema, withMessageRole };
+export { AIMessage, AIMessageChunk, ANTHROPIC_TOOL_TOKEN_MULTIPLIER, AgentSession, BASH_SHELL_GUIDANCE, BaseCheckpointSaver, BaseMessage, BaseMessageChunk, BashExecutionToolDefinition, BashExecutionToolDescription, BashExecutionToolName, BashExecutionToolSchema, BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, BashToolOutputReferencesGuide, CLOUDFLARE_BASH_CODING_TOOL_NAMES, CLOUDFLARE_CODING_TOOL_NAMES, CODE_ARTIFACT_PATH_GUIDANCE, CODE_EXECUTION_TOOLS, Calculator, CalculatorSchema, CalculatorToolDefinition, CalculatorToolDescription, CalculatorToolName, Callback, ChatModelStreamHandler, ChatOpenRouter, CloudflareBashExecutionToolDescription, CloudflareCodeExecutionToolDescription, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, Command, CommonEvents, CompileCheckToolName, Constants, ContentTypes, CustomOpenAIClient, DATE_RANGE, DEFAULT_CONTEXT_PRUNING_SETTINGS, DEFAULT_COUNTRY_DESCRIPTION, DEFAULT_HOOK_TIMEOUT_MS, DEFAULT_QUERY_DESCRIPTION, DEFAULT_RESERVE_RATIO, DEFAULT_TOOL_TOKEN_MULTIPLIER, DynamicStructuredTool, EnvVar, FAILED_EXECUTION_FILE_REMINDER, FakeChatModel, Graph, GraphEvents, GraphNodeActions, GraphNodeKeys, HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, HOOK_EVENTS, HandlerRegistry, HookRegistry, HumanMessage, INTERRUPT, JsonlSessionStore, LLMStreamHandler, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, LOCAL_SPAWN_TIMEOUT_MS, LocalBashExecutionToolDescription, LocalCodeExecutionToolDescription, LocalEditFileToolName, LocalEditFileToolSchema, LocalFileCheckpointerImpl, LocalGlobSearchToolName, LocalGlobSearchToolSchema, LocalGrepSearchToolName, LocalGrepSearchToolSchema, LocalListDirectoryToolName, LocalListDirectoryToolSchema, LocalReadFileToolSchema, LocalWriteFileToolName, LocalWriteFileToolSchema, MAX_CACHE_SIZE, MAX_PATTERN_LENGTH, MemorySaver, ModelEndHandler, MultiAgentGraph, ORIGINAL_CONTENT_MAX_CHARS, ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, PromptTemplate, Providers, REMOVE_ALL_MESSAGES, ReadFileToolDefinition, ReadFileToolDescription, ReadFileToolName, ReadFileToolSchema, Run, Runnable, RunnableCallable, RunnableLambda, RunnableSequence, SEPARATORS, SessionManager, SkillToolDefinition, SkillToolDescription, SkillToolName, SkillToolSchema, SplitStreamHandler, StandardGraph, StepTypes, StructuredTool, SubagentExecutor, SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, SystemMessage, TMP_SCRATCH_OUTPUT_REMINDER, TestChatStreamHandler, TestLLMStreamHandler, TitleMethod, TokenEncoderManager, Tool, ToolCallTypes, ToolEndHandler, ToolMessage, ToolNode, ToolSearchToolDefinition, ToolSearchToolDescription, ToolSearchToolName, ToolSearchToolSchema, WebSearchToolDefinition, WebSearchToolDescription, WebSearchToolName, WebSearchToolSchema, _createBashProgramForTests, _resetLocalEngineWarningsForTests, _resetRipgrepCacheForTests, _resetSyntaxCheckProbeCacheForTests, _resetUnrecognizedTriggerWarnings, addBedrockCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, appendCodeSessionFileSummary, appendFailedExecutionFileReminder, appendTmpScratchReminder, applyContextPruning, applyEdit, applyPreToolUseHooksForBridge, apportionTokenCounts, askUserQuestion, attemptInvoke, bashAstFindingsToErrors, buildBashExecutionToolDescription, buildChildInputs, buildCodeApiHttpErrorMessage, buildSandboxRuntimeConfig, buildSubagentToolParams, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, calculateTotalTokens, checkValidNumber, classifyAttachment, composeEventHandlers, convertMessagesToContent, countNestedGroups, countrySchema, createAgentSession, createBashExecutionTool, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, createCloudflareBashExecutionTool, createCloudflareBashProgrammaticToolCallingTool, createCloudflareBridgeRuntime, createCloudflareCodeExecutionTool, createCloudflareCodingToolBundle, createCloudflareCodingTools, createCloudflareExecutionTool, createCloudflareLocalExecutionConfig, createCloudflareProgrammaticToolCallingTool, createCloudflareWorkspaceFS, createCodeExecutionTool, createCompileCheckTool, createCompileCheckToolDefinition, createContentAggregator, createFakeStreamingLLM, createHandlers, createLocalBashExecutionTool, createLocalBashProgrammaticToolCallingTool, createLocalCodeExecutionTool, createLocalCodingToolBundle, createLocalCodingToolDefinitions, createLocalCodingToolRegistry, createLocalCodingTools, createLocalEditFileTool, createLocalFileCheckpointer, createLocalGlobSearchTool, createLocalGrepSearchTool, createLocalListDirectoryTool, createLocalProgrammaticToolCallingTool, createLocalReadFileTool, createLocalWriteFileTool, createMetadataAggregator, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, createPruneMessages, createRemoveAllMessage, createRunHandlers, createSchemaOnlyTool, createSchemaOnlyTools, createSearchTool, createSubagentToolDefinition, createTokenCounter, createToolPolicyHook, createToolSearch, createWorkspacePolicyHook, dateSchema, decodeFile, defaultOmitOptions, deserializeMessage, emptyOutputMessage, encodeFile, encodingForModel, enforceOriginalContentCap, ensureThinkingBlockInMessages, escapeRegexSpecialChars, estimateAnthropicImageTokens, estimateOpenAIImageTokens, executeCloudflareBash, executeCloudflareCode, executeHooks, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, executeParallelSearches, executeTools, extractErrorMessage, extractImageDimensions, extractMcpServerName, extractTextFromContent, extractToolDiscoveries, extractUsedBashToolNames, extractUsedToolNames, fetchSessionFiles, filterBashToolsByUsage, filterSubagentResult, filterToolsByUsage, findLastIndex, formatAgentMessages, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, formatCloudflareOutput, formatCompletedResponse, formatContentStrings, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, formatServerListing, formatSkillCatalog, getAvailableMcpServers, getBaseToolName, getBufferString, getChatModelClass, getChunkContent, getCloudflareWorkspaceRoot, getCodeBaseURL, getConverseOverrideMessage, getDeferredToolsListing, getLocalCwd, getLocalSessionId, getMaxOutputTokensKey, getMessageId, getMessagesWithinTokenLimit, getReadRoots, getSpawn, getTokenCountForMessage, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, handleServerToolResult, handleToolCallChunks, handleToolCalls, hasNestedQuantifier, hasNestedQuantifiers, hasToolSearchInCurrentTurn, imageAttachmentContent, imagesSchema, initializeModel, interrupt, isAIMessage, isAnthropicLike, isBaseMessage, isContextOverflowError, isDangerousPattern, isFromAnyMcpServer, isFromMcpServer, isGoogleLike, isInterrupted, isLegacyConvertible, isLikelyContextOverflowError, isOpenAILike, isPresent, isThinkingEnabled, isToolMessage, isZodSchema, joinKeys, labelContentByAgent, locateEdit, makeIsDeferred, makeRequest, maskConsumedToolResults, matchesQuery, messagesStateReducer, modifyDeltaProperties, newsSchema, normalizeBashToolResultsForReplay, normalizeServerFilter, normalizeToBashIdentifier, normalizeToPythonIdentifier, partitionAndMarkAnthropicToolCache, performLocalSearch, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, querySchema, repairOrphanedToolMessages, resetIfNotEmpty, resolveCloudflareSandbox, resolveCodeApiAuthHeaders, resolveContextPruningSettings, resolveLocalExecutionConfig, resolveLocalExecutionTools, resolveLocalToolRegistry, resolveLocalToolsForBinding, resolveSubagentConfigs, resolveWorkspacePath, resolveWorkspacePathSafe, runBashAstChecks, runPostEditSyntaxCheck, sanitizeOrphanToolBlocks, sanitizeRegex, serializeMessage, shellQuote, shiftIndexTokenCountMap, shouldTriggerSummarization, sleep, spawnLocalProcess, splitAtRecencyBoundary, stripAnthropicCacheControl, stripBedrockCacheControl, stripCodeSessionFileSummary, summarizeEvent, toJsonSchema, tool, toolResultTypes, toolsCondition, truncateLocalOutput, truncateToolInput, truncateToolResultContent, tryFallbackProviders, unescapeObject, unwrapToolResponse, validateBashCommand, validateCloudflareBashCommand, videosSchema, withMessageRole };

package/dist/esm/messages/content.mjs

@@ -2,6 +2,16 @@ import "../common/enum.mjs";
 import "../common/index.mjs";
 //#region src/messages/content.ts
 /**
+* Whether {@link formatContentStrings} will flatten this message's content:
+* a human/ai/system message whose content is an array of text-only blocks.
+*/
+const isLegacyConvertible = (message) => {
+	const messageType = message.getType();
+	if (!(messageType === "human" || messageType === "ai" || messageType === "system")) return false;
+	if (!Array.isArray(message.content)) return false;
+	return message.content.every((block) => block.type === "text");
+};
+/**
 * Formats an array of messages for LangChain, making sure all content fields are strings
 * @param {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} payload - The array of messages to format.
 * @returns {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} - The array of formatted LangChain messages, including ToolMessages for tool calls.
@@ -9,20 +19,7 @@ import "../common/index.mjs";
 const formatContentStrings = (payload) => {
 	const result = [];
 	for (const message of payload) {
-		const messageType = message.getType();
-		if (!(messageType === "human" || messageType === "ai" || messageType === "system")) {
-			result.push(message);
-			continue;
-		}
-		if (typeof message.content === "string") {
-			result.push(message);
-			continue;
-		}
-		if (!Array.isArray(message.content)) {
-			result.push(message);
-			continue;
-		}
-		if (!message.content.every((block) => block.type === "text")) {
+		if (!isLegacyConvertible(message)) {
 			result.push(message);
 			continue;
 		}
@@ -35,6 +32,6 @@ const formatContentStrings = (payload) => {
 	return result;
 };
 //#endregion
-export { formatContentStrings };
+export { formatContentStrings, isLegacyConvertible };
 
 //# sourceMappingURL=content.mjs.map

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

@@ -1 +1 @@
-{"version":3,"file":"content.mjs","names":[],"sources":["../../../src/messages/content.ts"],"sourcesContent":["import type { BaseMessage } from '@langchain/core/messages';\nimport { ContentTypes } from '@/common';\n\n/**\n * Formats an array of messages for LangChain, making sure all content fields are strings\n * @param {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} payload - The array of messages to format.\n * @returns {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} - The array of formatted LangChain messages, including ToolMessages for tool calls.\n */\nexport const formatContentStrings = (\n  payload: Array<BaseMessage>\n): Array<BaseMessage> => {\n  // Create a new array to store the processed messages\n  const result: Array<BaseMessage> = [];\n\n  for (const message of payload) {\n    const messageType = message.getType();\n    const isValidMessage =\n      messageType === 'human' ||\n      messageType === 'ai' ||\n      messageType === 'system';\n\n    if (!isValidMessage) {\n      result.push(message);\n      continue;\n    }\n\n    // If content is already a string, add as-is\n    if (typeof message.content === 'string') {\n      result.push(message);\n      continue;\n    }\n\n    // If content is not an array, add as-is\n    if (!Array.isArray(message.content)) {\n      result.push(message);\n      continue;\n    }\n\n    // Check if all content blocks are text type\n    const allTextBlocks = message.content.every(\n      (block) => block.type === ContentTypes.TEXT\n    );\n\n    // Only convert to string if all blocks are text type\n    if (!allTextBlocks) {\n      result.push(message);\n      continue;\n    }\n\n    // Reduce text types to a single string\n    const content = message.content.reduce((acc, curr) => {\n      if (curr.type === ContentTypes.TEXT) {\n        return `${acc}${curr[ContentTypes.TEXT] || ''}\\n`;\n      }\n      return acc;\n    }, '');\n\n    message.content = content.trim();\n    result.push(message);\n  }\n\n  return result;\n};\n"],"mappings":";;;;;;;;AAQA,MAAa,wBACX,YACuB;CAEvB,MAAM,SAA6B,CAAC;CAEpC,KAAK,MAAM,WAAW,SAAS;EAC7B,MAAM,cAAc,QAAQ,QAAQ;EAMpC,IAAI,EAJF,gBAAgB,WAChB,gBAAgB,QAChB,gBAAgB,WAEG;GACnB,OAAO,KAAK,OAAO;GACnB;EACF;EAGA,IAAI,OAAO,QAAQ,YAAY,UAAU;GACvC,OAAO,KAAK,OAAO;GACnB;EACF;EAGA,IAAI,CAAC,MAAM,QAAQ,QAAQ,OAAO,GAAG;GACnC,OAAO,KAAK,OAAO;GACnB;EACF;EAQA,IAAI,CALkB,QAAQ,QAAQ,OACnC,UAAU,MAAM,SAAA,MAIF,GAAG;GAClB,OAAO,KAAK,OAAO;GACnB;EACF;EAUA,QAAQ,UAPQ,QAAQ,QAAQ,QAAQ,KAAK,SAAS;GACpD,IAAI,KAAK,SAAA,QACP,OAAO,GAAG,MAAM,KAAA,WAA2B,GAAG;GAEhD,OAAO;EACT,GAAG,EAEqB,CAAC,CAAC,KAAK;EAC/B,OAAO,KAAK,OAAO;CACrB;CAEA,OAAO;AACT"}
+{"version":3,"file":"content.mjs","names":[],"sources":["../../../src/messages/content.ts"],"sourcesContent":["import type {\n  BaseMessage,\n  MessageContentComplex,\n} from '@langchain/core/messages';\nimport { ContentTypes } from '@/common';\n\n/**\n * Whether {@link formatContentStrings} will flatten this message's content:\n * a human/ai/system message whose content is an array of text-only blocks.\n */\nexport const isLegacyConvertible = (message: BaseMessage): boolean => {\n  const messageType = message.getType();\n  const isValidMessage =\n    messageType === 'human' || messageType === 'ai' || messageType === 'system';\n  if (!isValidMessage) {\n    return false;\n  }\n  if (!Array.isArray(message.content)) {\n    return false;\n  }\n  return message.content.every((block) => block.type === ContentTypes.TEXT);\n};\n\n/**\n * Formats an array of messages for LangChain, making sure all content fields are strings\n * @param {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} payload - The array of messages to format.\n * @returns {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} - The array of formatted LangChain messages, including ToolMessages for tool calls.\n */\nexport const formatContentStrings = (\n  payload: Array<BaseMessage>\n): Array<BaseMessage> => {\n  // Create a new array to store the processed messages\n  const result: Array<BaseMessage> = [];\n\n  for (const message of payload) {\n    if (!isLegacyConvertible(message)) {\n      result.push(message);\n      continue;\n    }\n\n    // Reduce text types to a single string\n    const blocks = message.content as MessageContentComplex[];\n    const content = blocks.reduce((acc, curr) => {\n      if (curr.type === ContentTypes.TEXT) {\n        return `${acc}${curr[ContentTypes.TEXT] || ''}\\n`;\n      }\n      return acc;\n    }, '');\n\n    message.content = content.trim();\n    result.push(message);\n  }\n\n  return result;\n};\n"],"mappings":";;;;;;;AAUA,MAAa,uBAAuB,YAAkC;CACpE,MAAM,cAAc,QAAQ,QAAQ;CAGpC,IAAI,EADF,gBAAgB,WAAW,gBAAgB,QAAQ,gBAAgB,WAEnE,OAAO;CAET,IAAI,CAAC,MAAM,QAAQ,QAAQ,OAAO,GAChC,OAAO;CAET,OAAO,QAAQ,QAAQ,OAAO,UAAU,MAAM,SAAA,MAA0B;AAC1E;;;;;;AAOA,MAAa,wBACX,YACuB;CAEvB,MAAM,SAA6B,CAAC;CAEpC,KAAK,MAAM,WAAW,SAAS;EAC7B,IAAI,CAAC,oBAAoB,OAAO,GAAG;GACjC,OAAO,KAAK,OAAO;GACnB;EACF;EAWA,QAAQ,UARO,QAAQ,QACA,QAAQ,KAAK,SAAS;GAC3C,IAAI,KAAK,SAAA,QACP,OAAO,GAAG,MAAM,KAAA,WAA2B,GAAG;GAEhD,OAAO;EACT,GAAG,EAEqB,CAAC,CAAC,KAAK;EAC/B,OAAO,KAAK,OAAO;CACrB;CAEA,OAAO;AACT"}

package/dist/esm/messages/prune.mjs

@@ -685,15 +685,24 @@ function createPruneMessages(factoryParams) {
 	let originalToolContentSize = 0;
 	const contextPruningSettings = resolveContextPruningSettings(factoryParams.contextPruningConfig);
 	return function pruneMessages(params) {
-		if (params.messages.length === 0) return {
-			context: [],
-			indexTokenCountMap,
-			messagesToRefine: [],
-			prePruneContextTokens: 0,
-			remainingContextTokens: factoryParams.maxTokens,
-			calibrationRatio,
-			resolvedInstructionOverhead: bestInstructionOverhead
-		};
+		if (params.messages.length === 0) {
+			/** Post-compaction calls still invoke the model — report the same
+			*  reserve-adjusted budget fields as the populated paths */
+			const emptyInstructionTokens = factoryParams.getInstructionTokens?.() ?? 0;
+			const emptyReserveRatio = factoryParams.reserveRatio ?? .05;
+			const emptyBudget = factoryParams.maxTokens - (emptyReserveRatio > 0 && emptyReserveRatio < 1 ? Math.round(factoryParams.maxTokens * emptyReserveRatio) : 0);
+			return {
+				context: [],
+				indexTokenCountMap,
+				messagesToRefine: [],
+				prePruneContextTokens: 0,
+				remainingContextTokens: Math.max(0, emptyBudget - emptyInstructionTokens),
+				calibrationRatio,
+				resolvedInstructionOverhead: bestInstructionOverhead,
+				contextBudget: emptyBudget,
+				effectiveInstructionTokens: emptyInstructionTokens
+			};
+		}
 		if (factoryParams.provider === "openAI" && factoryParams.thinkingEnabled === true) for (let i = lastTurnStartIndex; i < params.messages.length; i++) {
 			const m = params.messages[i];
 			if (m.getType() === "ai" && typeof m.additional_kwargs.reasoning_content === "string" && Array.isArray(m.additional_kwargs.provider_specific_fields?.thinking_blocks) && m.tool_calls && (m.tool_calls?.length ?? 0) > 0) {
@@ -799,7 +808,9 @@ function createPruneMessages(factoryParams) {
 				remainingContextTokens: 0,
 				contextPressure: pruningBudget > 0 ? calibratedTotalTokens / pruningBudget : 0,
 				calibrationRatio,
-				resolvedInstructionOverhead: bestInstructionOverhead
+				resolvedInstructionOverhead: bestInstructionOverhead,
+				contextBudget: pruningBudget,
+				effectiveInstructionTokens: currentInstructionTokens
 			};
 		}
 		totalTokens = sumTokenCounts(indexTokenCountMap, params.messages.length);
@@ -892,7 +903,9 @@ function createPruneMessages(factoryParams) {
 			contextPressure,
 			originalToolContent: originalToolContent.size > 0 ? originalToolContent : void 0,
 			calibrationRatio,
-			resolvedInstructionOverhead: bestInstructionOverhead
+			resolvedInstructionOverhead: bestInstructionOverhead,
+			contextBudget: pruningBudget,
+			effectiveInstructionTokens: currentInstructionTokens
 		};
 		const rawSpaceBudget = calibrationRatio > 0 ? Math.round(pruningBudget / calibrationRatio) : pruningBudget;
 		const rawSpaceInstructionTokens = calibrationRatio > 0 ? Math.round(currentInstructionTokens / calibrationRatio) : currentInstructionTokens;
@@ -1085,7 +1098,10 @@ function createPruneMessages(factoryParams) {
 				for (const [key, value] of Object.entries(preEmergencyTokenCounts)) indexTokenCountMap[key] = value;
 			}
 		}
-		const remainingContextTokens = Math.max(0, Math.min(pruningBudget, initialRemainingContextTokens + reclaimedTokens));
+		/** Scale raw-space remaining back to calibrated/provider units so it is
+		*  directly comparable with pruningBudget and prePruneContextTokens */
+		const rawRemaining = Math.max(0, initialRemainingContextTokens + reclaimedTokens);
+		const remainingContextTokens = Math.max(0, Math.min(pruningBudget, calibrationRatio > 0 ? Math.round(rawRemaining * calibrationRatio) : rawRemaining));
 		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);
@@ -1098,7 +1114,9 @@ function createPruneMessages(factoryParams) {
 			contextPressure,
 			originalToolContent: originalToolContent.size > 0 ? originalToolContent : void 0,
 			calibrationRatio,
-			resolvedInstructionOverhead: bestInstructionOverhead
+			resolvedInstructionOverhead: bestInstructionOverhead,
+			contextBudget: pruningBudget,
+			effectiveInstructionTokens: currentInstructionTokens
 		};
 	};
 }