@librechat/agents 3.1.80 → 3.1.81

package/dist/cjs/llm/vertexai/index.cjs

@@ -50,34 +50,33 @@ function repairStreamUsageMetadata(current, generationInfoUsage) {
  * - The signature for a functionCall part is an empty string
  *
  * This function correlates each "model" content block in the formatted request
- * back to its originating AI message, then re-attaches non-empty signatures
- * that the library failed to apply.
+ * back to its originating AI message by *position*, then re-attaches non-empty
+ * signatures that the library failed to apply. AI messages without signatures
+ * still consume their slot — filtering them out shifted later messages onto
+ * the wrong content block and dropped real signatures on the floor.
  */
 function fixThoughtSignatures(contents, input) {
-    // Collect AI messages that have signatures, in order
-    const aiMessages = input.filter((msg) => messages.isAIMessage(msg) &&
-        Array.isArray(msg.additional_kwargs?.signatures) &&
-        msg.additional_kwargs.signatures.length > 0);
-    // Collect "model" content blocks from the formatted request, in order
+    // All AI messages, in order — non-signature ones still consume positional
+    // slots so later messages line up with their model content blocks.
+    const aiMessages = input.filter(messages.isAIMessage);
     const modelContents = contents.filter((c) => c.role === 'model');
-    // They should correspond 1:1 in order (both derived from the same input sequence)
     const count = Math.min(aiMessages.length, modelContents.length);
     for (let i = 0; i < count; i++) {
-        const msg = aiMessages[i];
+        const signatures = aiMessages[i].additional_kwargs
+            ?.signatures;
+        if (!Array.isArray(signatures) || signatures.length === 0)
+            continue;
         const content = modelContents[i];
-        const signatures = msg.additional_kwargs?.signatures;
-        // Collect non-empty signatures that aren't already attached to any part
         const attachedSignatures = new Set(content.parts
             .map((p) => p.thoughtSignature)
             .filter((s) => s != null && s !== ''));
-        const availableSignatures = signatures?.filter((s) => s != null && s !== '' && !attachedSignatures.has(s));
-        // Assign available signatures to functionCall parts missing one, in order
+        const availableSignatures = signatures.filter((s) => s != null && s !== '' && !attachedSignatures.has(s));
         let sigIdx = 0;
         for (const part of content.parts) {
             if ('functionCall' in part &&
                 (part.thoughtSignature == null || part.thoughtSignature === '') &&
-                sigIdx < (availableSignatures?.length ?? 0)) {
-                part.thoughtSignature = availableSignatures?.[sigIdx];
+                sigIdx < availableSignatures.length) {
+                part.thoughtSignature = availableSignatures[sigIdx];
                 sigIdx++;
             }
         }
@@ -457,5 +456,6 @@ class ChatVertexAI extends googleGauth.ChatGoogle {
 }
 
 exports.ChatVertexAI = ChatVertexAI;
+exports.fixThoughtSignatures = fixThoughtSignatures;
 exports.repairStreamUsageMetadata = repairStreamUsageMetadata;
 //# sourceMappingURL=index.cjs.map

package/dist/cjs/llm/vertexai/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","sources":["../../../../src/llm/vertexai/index.ts"],"sourcesContent":["import { ChatGoogle } from '@langchain/google-gauth';\nimport { ChatConnection } from '@langchain/google-common';\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 { AIMessageChunk, isAIMessage } 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, then re-attaches non-empty signatures\n * that the library failed to apply.\n */\nfunction fixThoughtSignatures(\n  contents: GeminiContent[],\n  input: BaseMessage[]\n): void {\n  // Collect AI messages that have signatures, in order\n  const aiMessages = input.filter(\n    (msg) =>\n      isAIMessage(msg) &&\n      Array.isArray((msg.additional_kwargs as AdditionalKwargs)?.signatures) &&\n      (msg.additional_kwargs.signatures as string[]).length > 0\n  );\n\n  // Collect \"model\" content blocks from the formatted request, in order\n  const modelContents = contents.filter((c) => c.role === 'model');\n\n  // They should correspond 1:1 in order (both derived from the same input sequence)\n  const count = Math.min(aiMessages.length, modelContents.length);\n  for (let i = 0; i < count; i++) {\n    const msg = aiMessages[i];\n    const content = modelContents[i];\n    const signatures = (msg.additional_kwargs as AdditionalKwargs)?.signatures;\n\n    // Collect non-empty signatures that aren't already attached to any part\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 != null && s !== '' && !attachedSignatures.has(s)\n    );\n\n    // Assign available signatures to functionCall parts missing one, in order\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 ?? 0)\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"],"names":["isAIMessage","ChatConnection","ChatGoogle","messages","AIMessageChunk"],"mappings":";;;;;;AAcA;;;;;;;;;;;;;;;;;;;;;;;AAuBG;AACG,SAAU,yBAAyB,CACvC,OAAkC,EAClC,mBAA8C,EAAA;AAE9C,IAAA,IAAI,CAAC,OAAO;AAAE,QAAA,OAAO,OAAO;AAC5B,IAAA,IAAI,CAAC,mBAAmB;AAAE,QAAA,OAAO,OAAO;AACxC,IAAA,IAAI,mBAAmB,CAAC,YAAY,KAAK,OAAO,CAAC,YAAY;AAAE,QAAA,OAAO,OAAO;AAC7E,IAAA,IAAI,mBAAmB,CAAC,aAAa,IAAI,OAAO,CAAC,aAAa;AAC5D,QAAA,OAAO,OAAO;AAChB,IAAA,OAAO,mBAAmB;AAC5B;AAQA;;;;;;;;;;;;;AAaG;AACH,SAAS,oBAAoB,CAC3B,QAAyB,EACzB,KAAoB,EAAA;;AAGpB,IAAA,MAAM,UAAU,GAAG,KAAK,CAAC,MAAM,CAC7B,CAAC,GAAG,KACFA,oBAAW,CAAC,GAAG,CAAC;QAChB,KAAK,CAAC,OAAO,CAAE,GAAG,CAAC,iBAAsC,EAAE,UAAU,CAAC;QACrE,GAAG,CAAC,iBAAiB,CAAC,UAAuB,CAAC,MAAM,GAAG,CAAC,CAC5D;;AAGD,IAAA,MAAM,aAAa,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,KAAK,OAAO,CAAC;;AAGhE,IAAA,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC;AAC/D,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,EAAE,CAAC,EAAE,EAAE;AAC9B,QAAA,MAAM,GAAG,GAAG,UAAU,CAAC,CAAC,CAAC;AACzB,QAAA,MAAM,OAAO,GAAG,aAAa,CAAC,CAAC,CAAC;AAChC,QAAA,MAAM,UAAU,GAAI,GAAG,CAAC,iBAAsC,EAAE,UAAU;;AAG1E,QAAA,MAAM,kBAAkB,GAAG,IAAI,GAAG,CAChC,OAAO,CAAC;aACL,GAAG,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,gBAAgB;AAC7B,aAAA,MAAM,CAAC,CAAC,CAAC,KAAkB,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC,CACrD;QACD,MAAM,mBAAmB,GAAG,UAAU,EAAE,MAAM,CAC5C,CAAC,CAAC,KAAK,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,kBAAkB,CAAC,GAAG,CAAC,CAAC,CAAC,CAC3D;;QAGD,IAAI,MAAM,GAAG,CAAC;AACd,QAAA,KAAK,MAAM,IAAI,IAAI,OAAO,CAAC,KAAK,EAAE;YAChC,IACE,cAAc,IAAI,IAAI;iBACrB,IAAI,CAAC,gBAAgB,IAAI,IAAI,IAAI,IAAI,CAAC,gBAAgB,KAAK,EAAE,CAAC;gBAC/D,MAAM,IAAI,mBAAmB,EAAE,MAAM,IAAI,CAAC,CAAC,EAC3C;gBACA,IAAI,CAAC,gBAAgB,GAAG,mBAAmB,GAAG,MAAM,CAAC;AACrD,gBAAA,MAAM,EAAE;YACV;QACF;IACF;AACF;AAEA,MAAM,oBAAqB,SAAQC,2BAAqC,CAAA;AACtE,IAAA,cAAc;AAEd,IAAA,MAAM,UAAU,CACd,KAAoB,EACpB,UAAsC,EAAA;AAEtC,QAAA,MAAM,aAAa,IAAI,MAAM,KAAK,CAAC,UAAU,CAC3C,KAAK,EACL,UAAU,CACX,CAAkB;QACnB,IAAI,aAAa,CAAC,gBAAgB,EAAE,cAAc,EAAE,cAAc,KAAK,EAAE,EAAE;;YAEzE,IACE,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,eAAe,KAAK,KAAK,EACvE;gBACA,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,eAAe,GAAG,IAAI;YACtE;AACA,YAAA,OAAO,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,cAAc;QACrE;QACA,IAAI,IAAI,CAAC,cAAc,EAAE,aAAa,IAAI,IAAI,EAAE;AAC9C,YAAA,aAAa,CAAC,gBAAgB,KAAK,EAAE;;;AAGrC,YAAA,MAAM,EAAE,cAAc,EAAE,CAAC,EAAE,GAAG,sBAAsB,EAAE,GACnD,aAAa,CAAC,gBAAgB,CAAC,cAElB,IAAI,EAAE;AAEpB,YAAA,aAAa,CAAC,gBACf,CAAC,cAAc,GAAG;AACjB,gBAAA,GAAG,sBAAsB;AACzB,gBAAA,aAAa,EAAE,IAAI,CAAC,cAAc,CAAC,aAAa;gBAChD,IAAI,IAAI,CAAC,cAAc,CAAC,eAAe,IAAI,IAAI,IAAI;AACjD,oBAAA,eAAe,EAAE,IAAI,CAAC,cAAc,CAAC,eAAe;iBACrD,CAAC;aACH;QACH;AACA,QAAA,IAAI,aAAa,CAAC,QAAQ,EAAE;AAC1B,YAAA,oBAAoB,CAAC,aAAa,CAAC,QAAQ,EAAE,KAAK,CAAC;;AAEnD,YAAA,KAAK,MAAM,OAAO,IAAI,aAAa,CAAC,QAAQ,EAAE;AAC5C,gBAAA,IAAI,OAAO,CAAC,IAAI,KAAK,UAAU,EAAE;AAC9B,oBAAA,OAA4B,CAAC,IAAI,GAAG,MAAM;gBAC7C;YACF;QACF;AACA,QAAA,OAAO,aAAa;IACtB;AACD;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyRG;AACG,MAAO,YAAa,SAAQC,sBAAU,CAAA;IAC1C,YAAY,GAAG,CAAC,WAAW,EAAE,aAAa,EAAE,UAAU,CAAC;IACvD,qBAAqB,GAAG,KAAK;AAC7B,IAAA,cAAc;AAEd,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,mBAAmB;IAC5B;IAIA,WAAA,CACE,aAA8C,EAC9C,MAA6C,EAAA;AAE7C,QAAA,MAAM,MAAM,GACV,OAAO,aAAa,KAAK;AACvB,cAAE,EAAE,IAAI,MAAM,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,aAAa;cACzC,aAAa;QACnB,MAAM,qBAAqB,GAAG,MAAM,EAAE,cAAc,KAAK,EAAE;AAC3D,QAAA,KAAK,CAAC;AACJ,YAAA,GAAG,MAAM;AACT,YAAA,YAAY,EAAE,KAAK;AACpB,SAAA,CAAC;AACF,QAAA,IAAI,CAAC,qBAAqB,GAAG,qBAAqB;AAClD,QAAA,IAAI,CAAC,cAAc,GAAG,MAAM,EAAE,cAAc;IAC9C;AACA,IAAA,gBAAgB,CACd,OAA+C,EAAA;QAE/C,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;AAC9C,QAAA,IAAI,IAAI,CAAC,qBAAqB,EAAE;AAC9B,YAAA,MAAM,CAAC,kBAAkB,GAAG,EAAE;QAChC;AACA,QAAA,OAAO,MAAM;IACf;IACA,OAAO,qBAAqB,CAC1BC,UAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;AAErC,QAAA,IAAI,aAAwC;AAC5C,QAAA,WAAW,MAAM,KAAK,IAAI,KAAK,CAAC,qBAAqB,CACnDA,UAAQ,EACR,OAAO,EACP,UAAU,CACX,EAAE;AACD,YAAA,MAAM,QAAQ,GACZ,KAAK,CAAC,cACP,EAAE,cAAc;YACjB,IAAI,QAAQ,EAAE;gBACZ,aAAa,GAAG,QAAQ;YAC1B;AACA,YAAA,IAAI,KAAK,CAAC,OAAO,YAAYC,uBAAc,EAAE;AAC3C,gBAAA,MAAM,QAAQ,GAAG,yBAAyB,CACxC,KAAK,CAAC,OAAO,CAAC,cAAc,EAC5B,aAAa,CACd;gBACD,IAAI,QAAQ,KAAK,KAAK,CAAC,OAAO,CAAC,cAAc,EAAE;AAC7C,oBAAA,KAAK,CAAC,OAAO,CAAC,cAAc,GAAG,QAAQ;gBACzC;YACF;AACA,YAAA,MAAM,KAAK;QACb;IACF;IACA,eAAe,CACb,MAAyC,EACzC,MAA8B,EAAA;;;QAI9B,MAAM,cAAc,GAAG,MAAM,EAAE,cAAc,IAAI,IAAI,CAAC,cAAc;QAEpE,MAAM,UAAU,GAAG,IAAI,oBAAoB,CACzC,EAAE,GAAG,MAAM,EAAE,GAAG,IAAI,EAAE,EACtB,IAAI,CAAC,MAAM,EACX,MAAM,EACN,KAAK,CACN;AACD,QAAA,UAAU,CAAC,cAAc,GAAG,cAAc;AAC1C,QAAA,IAAI,CAAC,UAAU,GAAG,UAAU;QAE5B,MAAM,kBAAkB,GAAG,IAAI,oBAAoB,CACjD,EAAE,GAAG,MAAM,EAAE,GAAG,IAAI,EAAE,EACtB,IAAI,CAAC,MAAM,EACX,MAAM,EACN,IAAI,CACL;AACD,QAAA,kBAAkB,CAAC,cAAc,GAAG,cAAc;AAClD,QAAA,IAAI,CAAC,kBAAkB,GAAG,kBAAkB;IAC9C;AACD;;;;;"}
+{"version":3,"file":"index.cjs","sources":["../../../../src/llm/vertexai/index.ts"],"sourcesContent":["import { ChatGoogle } from '@langchain/google-gauth';\nimport { ChatConnection } from '@langchain/google-common';\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 { AIMessageChunk, isAIMessage } 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"],"names":["isAIMessage","ChatConnection","ChatGoogle","messages","AIMessageChunk"],"mappings":";;;;;;AAcA;;;;;;;;;;;;;;;;;;;;;;;AAuBG;AACG,SAAU,yBAAyB,CACvC,OAAkC,EAClC,mBAA8C,EAAA;AAE9C,IAAA,IAAI,CAAC,OAAO;AAAE,QAAA,OAAO,OAAO;AAC5B,IAAA,IAAI,CAAC,mBAAmB;AAAE,QAAA,OAAO,OAAO;AACxC,IAAA,IAAI,mBAAmB,CAAC,YAAY,KAAK,OAAO,CAAC,YAAY;AAAE,QAAA,OAAO,OAAO;AAC7E,IAAA,IAAI,mBAAmB,CAAC,aAAa,IAAI,OAAO,CAAC,aAAa;AAC5D,QAAA,OAAO,OAAO;AAChB,IAAA,OAAO,mBAAmB;AAC5B;AAQA;;;;;;;;;;;;;;;AAeG;AACG,SAAU,oBAAoB,CAClC,QAAyB,EACzB,KAAoB,EAAA;;;IAIpB,MAAM,UAAU,GAAG,KAAK,CAAC,MAAM,CAACA,oBAAW,CAAC;AAC5C,IAAA,MAAM,aAAa,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,KAAK,OAAO,CAAC;AAEhE,IAAA,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC;AAC/D,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,EAAE,CAAC,EAAE,EAAE;AAC9B,QAAA,MAAM,UAAU,GAAI,UAAU,CAAC,CAAC,CAAC,CAAC;AAChC,cAAE,UAAU;AACd,QAAA,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;YAAE;AAE3D,QAAA,MAAM,OAAO,GAAG,aAAa,CAAC,CAAC,CAAC;AAChC,QAAA,MAAM,kBAAkB,GAAG,IAAI,GAAG,CAChC,OAAO,CAAC;aACL,GAAG,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,gBAAgB;AAC7B,aAAA,MAAM,CAAC,CAAC,CAAC,KAAkB,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC,CACrD;QACD,MAAM,mBAAmB,GAAG,UAAU,CAAC,MAAM,CAC3C,CAAC,CAAC,KAAkB,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,kBAAkB,CAAC,GAAG,CAAC,CAAC,CAAC,CACxE;QAED,IAAI,MAAM,GAAG,CAAC;AACd,QAAA,KAAK,MAAM,IAAI,IAAI,OAAO,CAAC,KAAK,EAAE;YAChC,IACE,cAAc,IAAI,IAAI;iBACrB,IAAI,CAAC,gBAAgB,IAAI,IAAI,IAAI,IAAI,CAAC,gBAAgB,KAAK,EAAE,CAAC;AAC/D,gBAAA,MAAM,GAAG,mBAAmB,CAAC,MAAM,EACnC;AACA,gBAAA,IAAI,CAAC,gBAAgB,GAAG,mBAAmB,CAAC,MAAM,CAAC;AACnD,gBAAA,MAAM,EAAE;YACV;QACF;IACF;AACF;AAEA,MAAM,oBAAqB,SAAQC,2BAAqC,CAAA;AACtE,IAAA,cAAc;AAEd,IAAA,MAAM,UAAU,CACd,KAAoB,EACpB,UAAsC,EAAA;AAEtC,QAAA,MAAM,aAAa,IAAI,MAAM,KAAK,CAAC,UAAU,CAC3C,KAAK,EACL,UAAU,CACX,CAAkB;QACnB,IAAI,aAAa,CAAC,gBAAgB,EAAE,cAAc,EAAE,cAAc,KAAK,EAAE,EAAE;;YAEzE,IACE,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,eAAe,KAAK,KAAK,EACvE;gBACA,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,eAAe,GAAG,IAAI;YACtE;AACA,YAAA,OAAO,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,cAAc;QACrE;QACA,IAAI,IAAI,CAAC,cAAc,EAAE,aAAa,IAAI,IAAI,EAAE;AAC9C,YAAA,aAAa,CAAC,gBAAgB,KAAK,EAAE;;;AAGrC,YAAA,MAAM,EAAE,cAAc,EAAE,CAAC,EAAE,GAAG,sBAAsB,EAAE,GACnD,aAAa,CAAC,gBAAgB,CAAC,cAElB,IAAI,EAAE;AAEpB,YAAA,aAAa,CAAC,gBACf,CAAC,cAAc,GAAG;AACjB,gBAAA,GAAG,sBAAsB;AACzB,gBAAA,aAAa,EAAE,IAAI,CAAC,cAAc,CAAC,aAAa;gBAChD,IAAI,IAAI,CAAC,cAAc,CAAC,eAAe,IAAI,IAAI,IAAI;AACjD,oBAAA,eAAe,EAAE,IAAI,CAAC,cAAc,CAAC,eAAe;iBACrD,CAAC;aACH;QACH;AACA,QAAA,IAAI,aAAa,CAAC,QAAQ,EAAE;AAC1B,YAAA,oBAAoB,CAAC,aAAa,CAAC,QAAQ,EAAE,KAAK,CAAC;;AAEnD,YAAA,KAAK,MAAM,OAAO,IAAI,aAAa,CAAC,QAAQ,EAAE;AAC5C,gBAAA,IAAI,OAAO,CAAC,IAAI,KAAK,UAAU,EAAE;AAC9B,oBAAA,OAA4B,CAAC,IAAI,GAAG,MAAM;gBAC7C;YACF;QACF;AACA,QAAA,OAAO,aAAa;IACtB;AACD;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyRG;AACG,MAAO,YAAa,SAAQC,sBAAU,CAAA;IAC1C,YAAY,GAAG,CAAC,WAAW,EAAE,aAAa,EAAE,UAAU,CAAC;IACvD,qBAAqB,GAAG,KAAK;AAC7B,IAAA,cAAc;AAEd,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,mBAAmB;IAC5B;IAIA,WAAA,CACE,aAA8C,EAC9C,MAA6C,EAAA;AAE7C,QAAA,MAAM,MAAM,GACV,OAAO,aAAa,KAAK;AACvB,cAAE,EAAE,IAAI,MAAM,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,aAAa;cACzC,aAAa;QACnB,MAAM,qBAAqB,GAAG,MAAM,EAAE,cAAc,KAAK,EAAE;AAC3D,QAAA,KAAK,CAAC;AACJ,YAAA,GAAG,MAAM;AACT,YAAA,YAAY,EAAE,KAAK;AACpB,SAAA,CAAC;AACF,QAAA,IAAI,CAAC,qBAAqB,GAAG,qBAAqB;AAClD,QAAA,IAAI,CAAC,cAAc,GAAG,MAAM,EAAE,cAAc;IAC9C;AACA,IAAA,gBAAgB,CACd,OAA+C,EAAA;QAE/C,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;AAC9C,QAAA,IAAI,IAAI,CAAC,qBAAqB,EAAE;AAC9B,YAAA,MAAM,CAAC,kBAAkB,GAAG,EAAE;QAChC;AACA,QAAA,OAAO,MAAM;IACf;IACA,OAAO,qBAAqB,CAC1BC,UAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;AAErC,QAAA,IAAI,aAAwC;AAC5C,QAAA,WAAW,MAAM,KAAK,IAAI,KAAK,CAAC,qBAAqB,CACnDA,UAAQ,EACR,OAAO,EACP,UAAU,CACX,EAAE;AACD,YAAA,MAAM,QAAQ,GACZ,KAAK,CAAC,cACP,EAAE,cAAc;YACjB,IAAI,QAAQ,EAAE;gBACZ,aAAa,GAAG,QAAQ;YAC1B;AACA,YAAA,IAAI,KAAK,CAAC,OAAO,YAAYC,uBAAc,EAAE;AAC3C,gBAAA,MAAM,QAAQ,GAAG,yBAAyB,CACxC,KAAK,CAAC,OAAO,CAAC,cAAc,EAC5B,aAAa,CACd;gBACD,IAAI,QAAQ,KAAK,KAAK,CAAC,OAAO,CAAC,cAAc,EAAE;AAC7C,oBAAA,KAAK,CAAC,OAAO,CAAC,cAAc,GAAG,QAAQ;gBACzC;YACF;AACA,YAAA,MAAM,KAAK;QACb;IACF;IACA,eAAe,CACb,MAAyC,EACzC,MAA8B,EAAA;;;QAI9B,MAAM,cAAc,GAAG,MAAM,EAAE,cAAc,IAAI,IAAI,CAAC,cAAc;QAEpE,MAAM,UAAU,GAAG,IAAI,oBAAoB,CACzC,EAAE,GAAG,MAAM,EAAE,GAAG,IAAI,EAAE,EACtB,IAAI,CAAC,MAAM,EACX,MAAM,EACN,KAAK,CACN;AACD,QAAA,UAAU,CAAC,cAAc,GAAG,cAAc;AAC1C,QAAA,IAAI,CAAC,UAAU,GAAG,UAAU;QAE5B,MAAM,kBAAkB,GAAG,IAAI,oBAAoB,CACjD,EAAE,GAAG,MAAM,EAAE,GAAG,IAAI,EAAE,EACtB,IAAI,CAAC,MAAM,EACX,MAAM,EACN,IAAI,CACL;AACD,QAAA,kBAAkB,CAAC,cAAc,GAAG,cAAc;AAClD,QAAA,IAAI,CAAC,kBAAkB,GAAG,kBAAkB;IAC9C;AACD;;;;;;"}

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

@@ -48,34 +48,33 @@ function repairStreamUsageMetadata(current, generationInfoUsage) {
  * - The signature for a functionCall part is an empty string
  *
  * This function correlates each "model" content block in the formatted request
- * back to its originating AI message, then re-attaches non-empty signatures
- * that the library failed to apply.
+ * back to its originating AI message by *position*, then re-attaches non-empty
+ * signatures that the library failed to apply. AI messages without signatures
+ * still consume their slot — filtering them out shifted later messages onto
+ * the wrong content block and dropped real signatures on the floor.
  */
 function fixThoughtSignatures(contents, input) {
-    // Collect AI messages that have signatures, in order
-    const aiMessages = input.filter((msg) => isAIMessage(msg) &&
-        Array.isArray(msg.additional_kwargs?.signatures) &&
-        msg.additional_kwargs.signatures.length > 0);
-    // Collect "model" content blocks from the formatted request, in order
+    // All AI messages, in order — non-signature ones still consume positional
+    // slots so later messages line up with their model content blocks.
+    const aiMessages = input.filter(isAIMessage);
     const modelContents = contents.filter((c) => c.role === 'model');
-    // They should correspond 1:1 in order (both derived from the same input sequence)
     const count = Math.min(aiMessages.length, modelContents.length);
     for (let i = 0; i < count; i++) {
-        const msg = aiMessages[i];
+        const signatures = aiMessages[i].additional_kwargs
+            ?.signatures;
+        if (!Array.isArray(signatures) || signatures.length === 0)
+            continue;
         const content = modelContents[i];
-        const signatures = msg.additional_kwargs?.signatures;
-        // Collect non-empty signatures that aren't already attached to any part
         const attachedSignatures = new Set(content.parts
             .map((p) => p.thoughtSignature)
             .filter((s) => s != null && s !== ''));
-        const availableSignatures = signatures?.filter((s) => s != null && s !== '' && !attachedSignatures.has(s));
-        // Assign available signatures to functionCall parts missing one, in order
+        const availableSignatures = signatures.filter((s) => s != null && s !== '' && !attachedSignatures.has(s));
         let sigIdx = 0;
         for (const part of content.parts) {
             if ('functionCall' in part &&
                 (part.thoughtSignature == null || part.thoughtSignature === '') &&
-                sigIdx < (availableSignatures?.length ?? 0)) {
-                part.thoughtSignature = availableSignatures?.[sigIdx];
+                sigIdx < availableSignatures.length) {
+                part.thoughtSignature = availableSignatures[sigIdx];
                 sigIdx++;
             }
         }
@@ -454,5 +453,5 @@ class ChatVertexAI extends ChatGoogle {
     }
 }
 
-export { ChatVertexAI, repairStreamUsageMetadata };
+export { ChatVertexAI, fixThoughtSignatures, repairStreamUsageMetadata };
 //# sourceMappingURL=index.mjs.map

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

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","sources":["../../../../src/llm/vertexai/index.ts"],"sourcesContent":["import { ChatGoogle } from '@langchain/google-gauth';\nimport { ChatConnection } from '@langchain/google-common';\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 { AIMessageChunk, isAIMessage } 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, then re-attaches non-empty signatures\n * that the library failed to apply.\n */\nfunction fixThoughtSignatures(\n  contents: GeminiContent[],\n  input: BaseMessage[]\n): void {\n  // Collect AI messages that have signatures, in order\n  const aiMessages = input.filter(\n    (msg) =>\n      isAIMessage(msg) &&\n      Array.isArray((msg.additional_kwargs as AdditionalKwargs)?.signatures) &&\n      (msg.additional_kwargs.signatures as string[]).length > 0\n  );\n\n  // Collect \"model\" content blocks from the formatted request, in order\n  const modelContents = contents.filter((c) => c.role === 'model');\n\n  // They should correspond 1:1 in order (both derived from the same input sequence)\n  const count = Math.min(aiMessages.length, modelContents.length);\n  for (let i = 0; i < count; i++) {\n    const msg = aiMessages[i];\n    const content = modelContents[i];\n    const signatures = (msg.additional_kwargs as AdditionalKwargs)?.signatures;\n\n    // Collect non-empty signatures that aren't already attached to any part\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 != null && s !== '' && !attachedSignatures.has(s)\n    );\n\n    // Assign available signatures to functionCall parts missing one, in order\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 ?? 0)\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"],"names":[],"mappings":";;;;AAcA;;;;;;;;;;;;;;;;;;;;;;;AAuBG;AACG,SAAU,yBAAyB,CACvC,OAAkC,EAClC,mBAA8C,EAAA;AAE9C,IAAA,IAAI,CAAC,OAAO;AAAE,QAAA,OAAO,OAAO;AAC5B,IAAA,IAAI,CAAC,mBAAmB;AAAE,QAAA,OAAO,OAAO;AACxC,IAAA,IAAI,mBAAmB,CAAC,YAAY,KAAK,OAAO,CAAC,YAAY;AAAE,QAAA,OAAO,OAAO;AAC7E,IAAA,IAAI,mBAAmB,CAAC,aAAa,IAAI,OAAO,CAAC,aAAa;AAC5D,QAAA,OAAO,OAAO;AAChB,IAAA,OAAO,mBAAmB;AAC5B;AAQA;;;;;;;;;;;;;AAaG;AACH,SAAS,oBAAoB,CAC3B,QAAyB,EACzB,KAAoB,EAAA;;AAGpB,IAAA,MAAM,UAAU,GAAG,KAAK,CAAC,MAAM,CAC7B,CAAC,GAAG,KACF,WAAW,CAAC,GAAG,CAAC;QAChB,KAAK,CAAC,OAAO,CAAE,GAAG,CAAC,iBAAsC,EAAE,UAAU,CAAC;QACrE,GAAG,CAAC,iBAAiB,CAAC,UAAuB,CAAC,MAAM,GAAG,CAAC,CAC5D;;AAGD,IAAA,MAAM,aAAa,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,KAAK,OAAO,CAAC;;AAGhE,IAAA,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC;AAC/D,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,EAAE,CAAC,EAAE,EAAE;AAC9B,QAAA,MAAM,GAAG,GAAG,UAAU,CAAC,CAAC,CAAC;AACzB,QAAA,MAAM,OAAO,GAAG,aAAa,CAAC,CAAC,CAAC;AAChC,QAAA,MAAM,UAAU,GAAI,GAAG,CAAC,iBAAsC,EAAE,UAAU;;AAG1E,QAAA,MAAM,kBAAkB,GAAG,IAAI,GAAG,CAChC,OAAO,CAAC;aACL,GAAG,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,gBAAgB;AAC7B,aAAA,MAAM,CAAC,CAAC,CAAC,KAAkB,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC,CACrD;QACD,MAAM,mBAAmB,GAAG,UAAU,EAAE,MAAM,CAC5C,CAAC,CAAC,KAAK,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,kBAAkB,CAAC,GAAG,CAAC,CAAC,CAAC,CAC3D;;QAGD,IAAI,MAAM,GAAG,CAAC;AACd,QAAA,KAAK,MAAM,IAAI,IAAI,OAAO,CAAC,KAAK,EAAE;YAChC,IACE,cAAc,IAAI,IAAI;iBACrB,IAAI,CAAC,gBAAgB,IAAI,IAAI,IAAI,IAAI,CAAC,gBAAgB,KAAK,EAAE,CAAC;gBAC/D,MAAM,IAAI,mBAAmB,EAAE,MAAM,IAAI,CAAC,CAAC,EAC3C;gBACA,IAAI,CAAC,gBAAgB,GAAG,mBAAmB,GAAG,MAAM,CAAC;AACrD,gBAAA,MAAM,EAAE;YACV;QACF;IACF;AACF;AAEA,MAAM,oBAAqB,SAAQ,cAAqC,CAAA;AACtE,IAAA,cAAc;AAEd,IAAA,MAAM,UAAU,CACd,KAAoB,EACpB,UAAsC,EAAA;AAEtC,QAAA,MAAM,aAAa,IAAI,MAAM,KAAK,CAAC,UAAU,CAC3C,KAAK,EACL,UAAU,CACX,CAAkB;QACnB,IAAI,aAAa,CAAC,gBAAgB,EAAE,cAAc,EAAE,cAAc,KAAK,EAAE,EAAE;;YAEzE,IACE,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,eAAe,KAAK,KAAK,EACvE;gBACA,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,eAAe,GAAG,IAAI;YACtE;AACA,YAAA,OAAO,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,cAAc;QACrE;QACA,IAAI,IAAI,CAAC,cAAc,EAAE,aAAa,IAAI,IAAI,EAAE;AAC9C,YAAA,aAAa,CAAC,gBAAgB,KAAK,EAAE;;;AAGrC,YAAA,MAAM,EAAE,cAAc,EAAE,CAAC,EAAE,GAAG,sBAAsB,EAAE,GACnD,aAAa,CAAC,gBAAgB,CAAC,cAElB,IAAI,EAAE;AAEpB,YAAA,aAAa,CAAC,gBACf,CAAC,cAAc,GAAG;AACjB,gBAAA,GAAG,sBAAsB;AACzB,gBAAA,aAAa,EAAE,IAAI,CAAC,cAAc,CAAC,aAAa;gBAChD,IAAI,IAAI,CAAC,cAAc,CAAC,eAAe,IAAI,IAAI,IAAI;AACjD,oBAAA,eAAe,EAAE,IAAI,CAAC,cAAc,CAAC,eAAe;iBACrD,CAAC;aACH;QACH;AACA,QAAA,IAAI,aAAa,CAAC,QAAQ,EAAE;AAC1B,YAAA,oBAAoB,CAAC,aAAa,CAAC,QAAQ,EAAE,KAAK,CAAC;;AAEnD,YAAA,KAAK,MAAM,OAAO,IAAI,aAAa,CAAC,QAAQ,EAAE;AAC5C,gBAAA,IAAI,OAAO,CAAC,IAAI,KAAK,UAAU,EAAE;AAC9B,oBAAA,OAA4B,CAAC,IAAI,GAAG,MAAM;gBAC7C;YACF;QACF;AACA,QAAA,OAAO,aAAa;IACtB;AACD;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyRG;AACG,MAAO,YAAa,SAAQ,UAAU,CAAA;IAC1C,YAAY,GAAG,CAAC,WAAW,EAAE,aAAa,EAAE,UAAU,CAAC;IACvD,qBAAqB,GAAG,KAAK;AAC7B,IAAA,cAAc;AAEd,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,mBAAmB;IAC5B;IAIA,WAAA,CACE,aAA8C,EAC9C,MAA6C,EAAA;AAE7C,QAAA,MAAM,MAAM,GACV,OAAO,aAAa,KAAK;AACvB,cAAE,EAAE,IAAI,MAAM,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,aAAa;cACzC,aAAa;QACnB,MAAM,qBAAqB,GAAG,MAAM,EAAE,cAAc,KAAK,EAAE;AAC3D,QAAA,KAAK,CAAC;AACJ,YAAA,GAAG,MAAM;AACT,YAAA,YAAY,EAAE,KAAK;AACpB,SAAA,CAAC;AACF,QAAA,IAAI,CAAC,qBAAqB,GAAG,qBAAqB;AAClD,QAAA,IAAI,CAAC,cAAc,GAAG,MAAM,EAAE,cAAc;IAC9C;AACA,IAAA,gBAAgB,CACd,OAA+C,EAAA;QAE/C,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;AAC9C,QAAA,IAAI,IAAI,CAAC,qBAAqB,EAAE;AAC9B,YAAA,MAAM,CAAC,kBAAkB,GAAG,EAAE;QAChC;AACA,QAAA,OAAO,MAAM;IACf;IACA,OAAO,qBAAqB,CAC1B,QAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;AAErC,QAAA,IAAI,aAAwC;AAC5C,QAAA,WAAW,MAAM,KAAK,IAAI,KAAK,CAAC,qBAAqB,CACnD,QAAQ,EACR,OAAO,EACP,UAAU,CACX,EAAE;AACD,YAAA,MAAM,QAAQ,GACZ,KAAK,CAAC,cACP,EAAE,cAAc;YACjB,IAAI,QAAQ,EAAE;gBACZ,aAAa,GAAG,QAAQ;YAC1B;AACA,YAAA,IAAI,KAAK,CAAC,OAAO,YAAY,cAAc,EAAE;AAC3C,gBAAA,MAAM,QAAQ,GAAG,yBAAyB,CACxC,KAAK,CAAC,OAAO,CAAC,cAAc,EAC5B,aAAa,CACd;gBACD,IAAI,QAAQ,KAAK,KAAK,CAAC,OAAO,CAAC,cAAc,EAAE;AAC7C,oBAAA,KAAK,CAAC,OAAO,CAAC,cAAc,GAAG,QAAQ;gBACzC;YACF;AACA,YAAA,MAAM,KAAK;QACb;IACF;IACA,eAAe,CACb,MAAyC,EACzC,MAA8B,EAAA;;;QAI9B,MAAM,cAAc,GAAG,MAAM,EAAE,cAAc,IAAI,IAAI,CAAC,cAAc;QAEpE,MAAM,UAAU,GAAG,IAAI,oBAAoB,CACzC,EAAE,GAAG,MAAM,EAAE,GAAG,IAAI,EAAE,EACtB,IAAI,CAAC,MAAM,EACX,MAAM,EACN,KAAK,CACN;AACD,QAAA,UAAU,CAAC,cAAc,GAAG,cAAc;AAC1C,QAAA,IAAI,CAAC,UAAU,GAAG,UAAU;QAE5B,MAAM,kBAAkB,GAAG,IAAI,oBAAoB,CACjD,EAAE,GAAG,MAAM,EAAE,GAAG,IAAI,EAAE,EACtB,IAAI,CAAC,MAAM,EACX,MAAM,EACN,IAAI,CACL;AACD,QAAA,kBAAkB,CAAC,cAAc,GAAG,cAAc;AAClD,QAAA,IAAI,CAAC,kBAAkB,GAAG,kBAAkB;IAC9C;AACD;;;;"}
+{"version":3,"file":"index.mjs","sources":["../../../../src/llm/vertexai/index.ts"],"sourcesContent":["import { ChatGoogle } from '@langchain/google-gauth';\nimport { ChatConnection } from '@langchain/google-common';\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 { AIMessageChunk, isAIMessage } 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"],"names":[],"mappings":";;;;AAcA;;;;;;;;;;;;;;;;;;;;;;;AAuBG;AACG,SAAU,yBAAyB,CACvC,OAAkC,EAClC,mBAA8C,EAAA;AAE9C,IAAA,IAAI,CAAC,OAAO;AAAE,QAAA,OAAO,OAAO;AAC5B,IAAA,IAAI,CAAC,mBAAmB;AAAE,QAAA,OAAO,OAAO;AACxC,IAAA,IAAI,mBAAmB,CAAC,YAAY,KAAK,OAAO,CAAC,YAAY;AAAE,QAAA,OAAO,OAAO;AAC7E,IAAA,IAAI,mBAAmB,CAAC,aAAa,IAAI,OAAO,CAAC,aAAa;AAC5D,QAAA,OAAO,OAAO;AAChB,IAAA,OAAO,mBAAmB;AAC5B;AAQA;;;;;;;;;;;;;;;AAeG;AACG,SAAU,oBAAoB,CAClC,QAAyB,EACzB,KAAoB,EAAA;;;IAIpB,MAAM,UAAU,GAAG,KAAK,CAAC,MAAM,CAAC,WAAW,CAAC;AAC5C,IAAA,MAAM,aAAa,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,KAAK,OAAO,CAAC;AAEhE,IAAA,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC;AAC/D,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,EAAE,CAAC,EAAE,EAAE;AAC9B,QAAA,MAAM,UAAU,GAAI,UAAU,CAAC,CAAC,CAAC,CAAC;AAChC,cAAE,UAAU;AACd,QAAA,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;YAAE;AAE3D,QAAA,MAAM,OAAO,GAAG,aAAa,CAAC,CAAC,CAAC;AAChC,QAAA,MAAM,kBAAkB,GAAG,IAAI,GAAG,CAChC,OAAO,CAAC;aACL,GAAG,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,gBAAgB;AAC7B,aAAA,MAAM,CAAC,CAAC,CAAC,KAAkB,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC,CACrD;QACD,MAAM,mBAAmB,GAAG,UAAU,CAAC,MAAM,CAC3C,CAAC,CAAC,KAAkB,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,kBAAkB,CAAC,GAAG,CAAC,CAAC,CAAC,CACxE;QAED,IAAI,MAAM,GAAG,CAAC;AACd,QAAA,KAAK,MAAM,IAAI,IAAI,OAAO,CAAC,KAAK,EAAE;YAChC,IACE,cAAc,IAAI,IAAI;iBACrB,IAAI,CAAC,gBAAgB,IAAI,IAAI,IAAI,IAAI,CAAC,gBAAgB,KAAK,EAAE,CAAC;AAC/D,gBAAA,MAAM,GAAG,mBAAmB,CAAC,MAAM,EACnC;AACA,gBAAA,IAAI,CAAC,gBAAgB,GAAG,mBAAmB,CAAC,MAAM,CAAC;AACnD,gBAAA,MAAM,EAAE;YACV;QACF;IACF;AACF;AAEA,MAAM,oBAAqB,SAAQ,cAAqC,CAAA;AACtE,IAAA,cAAc;AAEd,IAAA,MAAM,UAAU,CACd,KAAoB,EACpB,UAAsC,EAAA;AAEtC,QAAA,MAAM,aAAa,IAAI,MAAM,KAAK,CAAC,UAAU,CAC3C,KAAK,EACL,UAAU,CACX,CAAkB;QACnB,IAAI,aAAa,CAAC,gBAAgB,EAAE,cAAc,EAAE,cAAc,KAAK,EAAE,EAAE;;YAEzE,IACE,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,eAAe,KAAK,KAAK,EACvE;gBACA,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,eAAe,GAAG,IAAI;YACtE;AACA,YAAA,OAAO,aAAa,CAAC,gBAAgB,CAAC,cAAc,CAAC,cAAc;QACrE;QACA,IAAI,IAAI,CAAC,cAAc,EAAE,aAAa,IAAI,IAAI,EAAE;AAC9C,YAAA,aAAa,CAAC,gBAAgB,KAAK,EAAE;;;AAGrC,YAAA,MAAM,EAAE,cAAc,EAAE,CAAC,EAAE,GAAG,sBAAsB,EAAE,GACnD,aAAa,CAAC,gBAAgB,CAAC,cAElB,IAAI,EAAE;AAEpB,YAAA,aAAa,CAAC,gBACf,CAAC,cAAc,GAAG;AACjB,gBAAA,GAAG,sBAAsB;AACzB,gBAAA,aAAa,EAAE,IAAI,CAAC,cAAc,CAAC,aAAa;gBAChD,IAAI,IAAI,CAAC,cAAc,CAAC,eAAe,IAAI,IAAI,IAAI;AACjD,oBAAA,eAAe,EAAE,IAAI,CAAC,cAAc,CAAC,eAAe;iBACrD,CAAC;aACH;QACH;AACA,QAAA,IAAI,aAAa,CAAC,QAAQ,EAAE;AAC1B,YAAA,oBAAoB,CAAC,aAAa,CAAC,QAAQ,EAAE,KAAK,CAAC;;AAEnD,YAAA,KAAK,MAAM,OAAO,IAAI,aAAa,CAAC,QAAQ,EAAE;AAC5C,gBAAA,IAAI,OAAO,CAAC,IAAI,KAAK,UAAU,EAAE;AAC9B,oBAAA,OAA4B,CAAC,IAAI,GAAG,MAAM;gBAC7C;YACF;QACF;AACA,QAAA,OAAO,aAAa;IACtB;AACD;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyRG;AACG,MAAO,YAAa,SAAQ,UAAU,CAAA;IAC1C,YAAY,GAAG,CAAC,WAAW,EAAE,aAAa,EAAE,UAAU,CAAC;IACvD,qBAAqB,GAAG,KAAK;AAC7B,IAAA,cAAc;AAEd,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,mBAAmB;IAC5B;IAIA,WAAA,CACE,aAA8C,EAC9C,MAA6C,EAAA;AAE7C,QAAA,MAAM,MAAM,GACV,OAAO,aAAa,KAAK;AACvB,cAAE,EAAE,IAAI,MAAM,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,aAAa;cACzC,aAAa;QACnB,MAAM,qBAAqB,GAAG,MAAM,EAAE,cAAc,KAAK,EAAE;AAC3D,QAAA,KAAK,CAAC;AACJ,YAAA,GAAG,MAAM;AACT,YAAA,YAAY,EAAE,KAAK;AACpB,SAAA,CAAC;AACF,QAAA,IAAI,CAAC,qBAAqB,GAAG,qBAAqB;AAClD,QAAA,IAAI,CAAC,cAAc,GAAG,MAAM,EAAE,cAAc;IAC9C;AACA,IAAA,gBAAgB,CACd,OAA+C,EAAA;QAE/C,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;AAC9C,QAAA,IAAI,IAAI,CAAC,qBAAqB,EAAE;AAC9B,YAAA,MAAM,CAAC,kBAAkB,GAAG,EAAE;QAChC;AACA,QAAA,OAAO,MAAM;IACf;IACA,OAAO,qBAAqB,CAC1B,QAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;AAErC,QAAA,IAAI,aAAwC;AAC5C,QAAA,WAAW,MAAM,KAAK,IAAI,KAAK,CAAC,qBAAqB,CACnD,QAAQ,EACR,OAAO,EACP,UAAU,CACX,EAAE;AACD,YAAA,MAAM,QAAQ,GACZ,KAAK,CAAC,cACP,EAAE,cAAc;YACjB,IAAI,QAAQ,EAAE;gBACZ,aAAa,GAAG,QAAQ;YAC1B;AACA,YAAA,IAAI,KAAK,CAAC,OAAO,YAAY,cAAc,EAAE;AAC3C,gBAAA,MAAM,QAAQ,GAAG,yBAAyB,CACxC,KAAK,CAAC,OAAO,CAAC,cAAc,EAC5B,aAAa,CACd;gBACD,IAAI,QAAQ,KAAK,KAAK,CAAC,OAAO,CAAC,cAAc,EAAE;AAC7C,oBAAA,KAAK,CAAC,OAAO,CAAC,cAAc,GAAG,QAAQ;gBACzC;YACF;AACA,YAAA,MAAM,KAAK;QACb;IACF;IACA,eAAe,CACb,MAAyC,EACzC,MAA8B,EAAA;;;QAI9B,MAAM,cAAc,GAAG,MAAM,EAAE,cAAc,IAAI,IAAI,CAAC,cAAc;QAEpE,MAAM,UAAU,GAAG,IAAI,oBAAoB,CACzC,EAAE,GAAG,MAAM,EAAE,GAAG,IAAI,EAAE,EACtB,IAAI,CAAC,MAAM,EACX,MAAM,EACN,KAAK,CACN;AACD,QAAA,UAAU,CAAC,cAAc,GAAG,cAAc;AAC1C,QAAA,IAAI,CAAC,UAAU,GAAG,UAAU;QAE5B,MAAM,kBAAkB,GAAG,IAAI,oBAAoB,CACjD,EAAE,GAAG,MAAM,EAAE,GAAG,IAAI,EAAE,EACtB,IAAI,CAAC,MAAM,EACX,MAAM,EACN,IAAI,CACL;AACD,QAAA,kBAAkB,CAAC,cAAc,GAAG,cAAc;AAClD,QAAA,IAAI,CAAC,kBAAkB,GAAG,kBAAkB;IAC9C;AACD;;;;"}

package/dist/types/llm/vertexai/index.d.ts

@@ -1,5 +1,5 @@
 import { ChatGoogle } from '@langchain/google-gauth';
-import type { GoogleAIModelRequestParams, GoogleAbstractedClient } from '@langchain/google-common';
+import type { GeminiContent, GoogleAIModelRequestParams, GoogleAbstractedClient } from '@langchain/google-common';
 import type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';
 import type { BaseMessage, UsageMetadata } from '@langchain/core/messages';
 import type { ChatGenerationChunk } from '@langchain/core/outputs';
@@ -29,6 +29,23 @@ import type { GoogleThinkingConfig, VertexAIClientOptions } from '@/types';
  * overriding `_convertToUsageMetadata`.
  */
 export declare function repairStreamUsageMetadata(current: UsageMetadata | undefined, generationInfoUsage: UsageMetadata | undefined): UsageMetadata | undefined;
+/**
+ * Fixes thought signatures on functionCall parts in the formatted Gemini request.
+ *
+ * `@langchain/google-common` stores signatures as a flat array in
+ * `additional_kwargs.signatures` (one per response part) and re-attaches them
+ * by index only when `signatures.length === parts.length`. This fails when:
+ * - The API omits a signature (length mismatch)
+ * - Streaming chunks merge with different part counts
+ * - The signature for a functionCall part is an empty string
+ *
+ * This function correlates each "model" content block in the formatted request
+ * back to its originating AI message by *position*, then re-attaches non-empty
+ * signatures that the library failed to apply. AI messages without signatures
+ * still consume their slot — filtering them out shifted later messages onto
+ * the wrong content block and dropped real signatures on the floor.
+ */
+export declare function fixThoughtSignatures(contents: GeminiContent[], input: BaseMessage[]): void;
 /**
  * Integration with Google Vertex AI chat models.
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "3.1.80",
+  "version": "3.1.81",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",

package/src/llm/vertexai/fixThoughtSignatures.test.ts

@@ -0,0 +1,154 @@
+import { expect, test, describe } from '@jest/globals';
+import type { GeminiContent } from '@langchain/google-common';
+import { AIMessage, HumanMessage, ToolMessage } from '@langchain/core/messages';
+import { fixThoughtSignatures } from './index';
+
+const SIG_A = 'AY89a1/sigA==';
+const SIG_B = 'AY89a1/sigB==';
+
+const buildContents = (
+  blocks: Array<['user' | 'model' | 'function', GeminiContent['parts']]>
+): GeminiContent[] =>
+  blocks.map(([role, parts]) => ({ role, parts }) as GeminiContent);
+
+describe('fixThoughtSignatures', () => {
+  test('attaches signature to functionCall part when prior turn is a plain-text AI message (issue LibreChat#13006-followup)', () => {
+    // Reproduces the live failure from the issue: a Gemini 3 conversation
+    // where turn 1 was plain text ("Hello!") and turn 2 emitted a tool call
+    // with a thought signature. The plain-text AI message has no signatures,
+    // so the old position-by-filter code matched the toolcall AIMessage with
+    // the WRONG model content.
+    const helloAi = new AIMessage('Hello! How can I help you today?');
+    const toolcallAi = new AIMessage({
+      content: '',
+      tool_calls: [
+        { name: 'bash_tool', args: { command: 'echo hi' }, id: 'tc1' },
+      ],
+      additional_kwargs: { signatures: [SIG_A, ''] },
+    });
+    const input = [
+      new HumanMessage('hi there'),
+      helloAi,
+      new HumanMessage('run something'),
+      toolcallAi,
+      new ToolMessage({ content: 'ok', tool_call_id: 'tc1' }),
+    ];
+
+    const contents = buildContents([
+      ['user', [{ text: 'hi there' }]],
+      ['model', [{ text: 'Hello! How can I help you today?' }]],
+      ['user', [{ text: 'run something' }]],
+      [
+        'model',
+        [{ functionCall: { name: 'bash_tool', args: { command: 'echo hi' } } }],
+      ],
+      [
+        'user',
+        [
+          {
+            functionResponse: {
+              name: 'bash_tool',
+              response: { content: 'ok' },
+            },
+          },
+        ],
+      ],
+    ]);
+
+    fixThoughtSignatures(contents, input);
+
+    expect(contents[1].parts[0].thoughtSignature).toBeUndefined();
+    expect(contents[3].parts[0]).toMatchObject({
+      functionCall: { name: 'bash_tool' },
+      thoughtSignature: SIG_A,
+    });
+  });
+
+  test('attaches signatures across multiple tool-call turns by position', () => {
+    const turn1 = new AIMessage({
+      content: '',
+      tool_calls: [{ name: 'a', args: {}, id: 't1' }],
+      additional_kwargs: { signatures: [SIG_A, ''] },
+    });
+    const turn2 = new AIMessage({
+      content: '',
+      tool_calls: [{ name: 'b', args: {}, id: 't2' }],
+      additional_kwargs: { signatures: [SIG_B, ''] },
+    });
+
+    const input = [
+      new HumanMessage('q1'),
+      turn1,
+      new ToolMessage({ content: '1', tool_call_id: 't1' }),
+      new HumanMessage('q2'),
+      turn2,
+      new ToolMessage({ content: '2', tool_call_id: 't2' }),
+    ];
+    const contents = buildContents([
+      ['user', [{ text: 'q1' }]],
+      ['model', [{ functionCall: { name: 'a', args: {} } }]],
+      ['user', [{ functionResponse: { name: 'a', response: {} } }]],
+      ['user', [{ text: 'q2' }]],
+      ['model', [{ functionCall: { name: 'b', args: {} } }]],
+      ['user', [{ functionResponse: { name: 'b', response: {} } }]],
+    ]);
+
+    fixThoughtSignatures(contents, input);
+
+    expect(contents[1].parts[0].thoughtSignature).toBe(SIG_A);
+    expect(contents[4].parts[0].thoughtSignature).toBe(SIG_B);
+  });
+
+  test('does not overwrite signatures already attached by the library', () => {
+    const ai = new AIMessage({
+      content: '',
+      tool_calls: [{ name: 'a', args: {}, id: 't1' }],
+      additional_kwargs: { signatures: [SIG_A] },
+    });
+    const input = [new HumanMessage('q'), ai];
+    const contents = buildContents([
+      ['user', [{ text: 'q' }]],
+      [
+        'model',
+        [{ functionCall: { name: 'a', args: {} }, thoughtSignature: SIG_B }],
+      ],
+    ]);
+
+    fixThoughtSignatures(contents, input);
+
+    expect(contents[1].parts[0].thoughtSignature).toBe(SIG_B);
+  });
+
+  test('no-op when AI message has no signatures', () => {
+    const ai = new AIMessage({
+      content: '',
+      tool_calls: [{ name: 'a', args: {}, id: 't1' }],
+    });
+    const input = [new HumanMessage('q'), ai];
+    const contents = buildContents([
+      ['user', [{ text: 'q' }]],
+      ['model', [{ functionCall: { name: 'a', args: {} } }]],
+    ]);
+
+    fixThoughtSignatures(contents, input);
+
+    expect(contents[1].parts[0].thoughtSignature).toBeUndefined();
+  });
+
+  test('skips empty-string signatures', () => {
+    const ai = new AIMessage({
+      content: '',
+      tool_calls: [{ name: 'a', args: {}, id: 't1' }],
+      additional_kwargs: { signatures: ['', '', ''] },
+    });
+    const input = [new HumanMessage('q'), ai];
+    const contents = buildContents([
+      ['user', [{ text: 'q' }]],
+      ['model', [{ functionCall: { name: 'a', args: {} } }]],
+    ]);
+
+    fixThoughtSignatures(contents, input);
+
+    expect(contents[1].parts[0].thoughtSignature).toBeUndefined();
+  });
+});

package/src/llm/vertexai/index.ts

@@ -65,50 +65,44 @@ type AdditionalKwargs =
  * - The signature for a functionCall part is an empty string
  *
  * This function correlates each "model" content block in the formatted request
- * back to its originating AI message, then re-attaches non-empty signatures
- * that the library failed to apply.
+ * back to its originating AI message by *position*, then re-attaches non-empty
+ * signatures that the library failed to apply. AI messages without signatures
+ * still consume their slot — filtering them out shifted later messages onto
+ * the wrong content block and dropped real signatures on the floor.
  */
-function fixThoughtSignatures(
+export function fixThoughtSignatures(
   contents: GeminiContent[],
   input: BaseMessage[]
 ): void {
-  // Collect AI messages that have signatures, in order
-  const aiMessages = input.filter(
-    (msg) =>
-      isAIMessage(msg) &&
-      Array.isArray((msg.additional_kwargs as AdditionalKwargs)?.signatures) &&
-      (msg.additional_kwargs.signatures as string[]).length > 0
-  );
-
-  // Collect "model" content blocks from the formatted request, in order
+  // All AI messages, in order — non-signature ones still consume positional
+  // slots so later messages line up with their model content blocks.
+  const aiMessages = input.filter(isAIMessage);
   const modelContents = contents.filter((c) => c.role === 'model');
 
-  // They should correspond 1:1 in order (both derived from the same input sequence)
   const count = Math.min(aiMessages.length, modelContents.length);
   for (let i = 0; i < count; i++) {
-    const msg = aiMessages[i];
-    const content = modelContents[i];
-    const signatures = (msg.additional_kwargs as AdditionalKwargs)?.signatures;
+    const signatures = (aiMessages[i].additional_kwargs as AdditionalKwargs)
+      ?.signatures;
+    if (!Array.isArray(signatures) || signatures.length === 0) continue;
 
-    // Collect non-empty signatures that aren't already attached to any part
+    const content = modelContents[i];
     const attachedSignatures = new Set(
       content.parts
         .map((p) => p.thoughtSignature)
         .filter((s): s is string => s != null && s !== '')
     );
-    const availableSignatures = signatures?.filter(
-      (s) => s != null && s !== '' && !attachedSignatures.has(s)
+    const availableSignatures = signatures.filter(
+      (s): s is string => s != null && s !== '' && !attachedSignatures.has(s)
     );
 
-    // Assign available signatures to functionCall parts missing one, in order
     let sigIdx = 0;
     for (const part of content.parts) {
       if (
         'functionCall' in part &&
         (part.thoughtSignature == null || part.thoughtSignature === '') &&
-        sigIdx < (availableSignatures?.length ?? 0)
+        sigIdx < availableSignatures.length
       ) {
-        part.thoughtSignature = availableSignatures?.[sigIdx];
+        part.thoughtSignature = availableSignatures[sigIdx];
         sigIdx++;
       }
     }