@librechat/agents 3.1.80 → 3.1.82

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

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","sources":["../../../../src/llm/openrouter/index.ts"],"sourcesContent":["import { ChatOpenAI } from '@/llm/openai';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type {\n  ChatOpenAICallOptions,\n  OpenAIChatInput,\n  OpenAIClient,\n} from '@langchain/openai';\n\nexport type OpenRouterReasoningEffort =\n  | 'xhigh'\n  | 'high'\n  | 'medium'\n  | 'low'\n  | 'minimal'\n  | 'none';\n\nexport interface OpenRouterReasoning {\n  effort?: OpenRouterReasoningEffort;\n  max_tokens?: number;\n  exclude?: boolean;\n  enabled?: boolean;\n}\n\nexport interface ChatOpenRouterCallOptions\n  extends Omit<ChatOpenAICallOptions, 'reasoning'> {\n  /** @deprecated Use `reasoning` object instead */\n  include_reasoning?: boolean;\n  reasoning?: OpenRouterReasoning;\n  modelKwargs?: OpenAIChatInput['modelKwargs'];\n}\n\nexport type ChatOpenRouterInput = Partial<\n  ChatOpenRouterCallOptions & OpenAIChatInput\n>;\n\n/** invocationParams return type extended with OpenRouter reasoning */\nexport type OpenRouterInvocationParams = Omit<\n  OpenAIClient.Chat.ChatCompletionCreateParams,\n  'messages'\n> & {\n  reasoning?: OpenRouterReasoning;\n};\n\ntype InvocationParamsExtra = {\n  streaming?: boolean;\n};\n\ninterface OpenRouterReasoningTextDetail {\n  type: 'reasoning.text';\n  text?: string;\n  format?: string;\n  index?: number;\n}\n\ninterface OpenRouterReasoningEncryptedDetail {\n  type: 'reasoning.encrypted';\n  id?: string;\n  data?: string;\n  format?: string;\n  index?: number;\n}\n\ntype OpenRouterReasoningDetail =\n  | OpenRouterReasoningTextDetail\n  | OpenRouterReasoningEncryptedDetail;\n\nfunction isReasoningTextDetail(\n  value: unknown\n): value is OpenRouterReasoningTextDetail {\n  return (\n    typeof value === 'object' &&\n    value !== null &&\n    'type' in value &&\n    value.type === 'reasoning.text'\n  );\n}\n\nfunction isReasoningEncryptedDetail(\n  value: unknown\n): value is OpenRouterReasoningEncryptedDetail {\n  return (\n    typeof value === 'object' &&\n    value !== null &&\n    'type' in value &&\n    value.type === 'reasoning.encrypted'\n  );\n}\n\nfunction getReasoningDetails(value: unknown): OpenRouterReasoningDetail[] {\n  if (!Array.isArray(value)) {\n    return [];\n  }\n  return value.filter(\n    (detail): detail is OpenRouterReasoningDetail =>\n      isReasoningTextDetail(detail) || isReasoningEncryptedDetail(detail)\n  );\n}\n\nexport class ChatOpenRouter extends ChatOpenAI {\n  private openRouterReasoning?: OpenRouterReasoning;\n  /** @deprecated Use `reasoning` object instead */\n  private includeReasoning?: boolean;\n\n  constructor(_fields: ChatOpenRouterInput) {\n    const {\n      include_reasoning,\n      reasoning: openRouterReasoning,\n      modelKwargs = {},\n      ...fields\n    } = _fields;\n\n    // Extract reasoning from modelKwargs if provided there (e.g., from LLMConfig)\n    const { reasoning: mkReasoning, ...restModelKwargs } = modelKwargs as {\n      reasoning?: OpenRouterReasoning;\n    } & Record<string, unknown>;\n\n    super({\n      ...fields,\n      modelKwargs: restModelKwargs,\n      includeReasoningDetails: true,\n      convertReasoningDetailsToContent: true,\n    });\n\n    // Merge reasoning config: modelKwargs.reasoning < constructor reasoning\n    if (mkReasoning != null || openRouterReasoning != null) {\n      this.openRouterReasoning = {\n        ...mkReasoning,\n        ...openRouterReasoning,\n      };\n    }\n\n    this.includeReasoning = include_reasoning;\n  }\n  static lc_name(): 'LibreChatOpenRouter' {\n    return 'LibreChatOpenRouter';\n  }\n\n  // @ts-expect-error - OpenRouter reasoning extends OpenAI Reasoning with additional\n  // effort levels ('xhigh' | 'none' | 'minimal') not in ReasoningEffort.\n  // The parent's generic conditional return type cannot be widened in an override.\n  override invocationParams(\n    options?: this['ParsedCallOptions'],\n    extra?: InvocationParamsExtra\n  ): OpenRouterInvocationParams {\n    type MutableParams = Omit<\n      OpenAIClient.Chat.ChatCompletionCreateParams,\n      'messages'\n    > & { reasoning_effort?: string; reasoning?: OpenRouterReasoning };\n\n    const optionsWithDefaults = this._combineCallOptions(options);\n    const params = (\n      this._useResponsesApi(options)\n        ? this.responses.invocationParams(optionsWithDefaults)\n        : this.completions.invocationParams(optionsWithDefaults, extra)\n    ) as MutableParams;\n\n    // Remove the OpenAI-native reasoning_effort that the parent sets;\n    // OpenRouter uses a `reasoning` object instead\n    delete params.reasoning_effort;\n\n    // Build the OpenRouter reasoning config\n    const reasoning = this.buildOpenRouterReasoning(optionsWithDefaults);\n    if (reasoning != null) {\n      params.reasoning = reasoning;\n    } else {\n      delete params.reasoning;\n    }\n\n    return params;\n  }\n\n  private buildOpenRouterReasoning(\n    options?: this['ParsedCallOptions']\n  ): OpenRouterReasoning | undefined {\n    let reasoning: OpenRouterReasoning | undefined;\n\n    // 1. Instance-level reasoning config (from constructor)\n    if (this.openRouterReasoning != null) {\n      reasoning = { ...this.openRouterReasoning };\n    }\n\n    // 2. LangChain-style reasoning params (from parent's `this.reasoning`)\n    const lcReasoning = this.getReasoningParams(options);\n    if (lcReasoning?.effort != null) {\n      reasoning = {\n        ...reasoning,\n        effort: lcReasoning.effort as OpenRouterReasoningEffort,\n      };\n    }\n\n    // 3. Call-level reasoning override\n    const callReasoning = (options as ChatOpenRouterCallOptions | undefined)\n      ?.reasoning;\n    if (callReasoning != null) {\n      reasoning = { ...reasoning, ...callReasoning };\n    }\n\n    // 4. Legacy include_reasoning backward compatibility\n    if (reasoning == null && this.includeReasoning === true) {\n      reasoning = { enabled: true };\n    }\n\n    return reasoning;\n  }\n\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const reasoningTextByIndex = new Map<\n      number,\n      OpenRouterReasoningTextDetail\n    >();\n    const reasoningEncryptedById = new Map<\n      string,\n      OpenRouterReasoningEncryptedDetail\n    >();\n\n    for await (const generationChunk of super._streamResponseChunks(\n      messages,\n      options,\n      runManager\n    )) {\n      let currentReasoningText = '';\n      const reasoningDetails = getReasoningDetails(\n        generationChunk.message.additional_kwargs.reasoning_details\n      );\n\n      for (const detail of reasoningDetails) {\n        if (detail.type === 'reasoning.text') {\n          currentReasoningText += detail.text ?? '';\n          const index = detail.index ?? 0;\n          const existing = reasoningTextByIndex.get(index);\n          if (existing != null) {\n            existing.text = `${existing.text ?? ''}${detail.text ?? ''}`;\n            continue;\n          }\n          reasoningTextByIndex.set(index, {\n            ...detail,\n            text: detail.text ?? '',\n          });\n          continue;\n        }\n        if (detail.id != null) {\n          reasoningEncryptedById.set(detail.id, { ...detail });\n        }\n      }\n\n      if (\n        currentReasoningText.length > 0 &&\n        generationChunk.message.additional_kwargs.reasoning == null\n      ) {\n        generationChunk.message.additional_kwargs.reasoning =\n          currentReasoningText;\n      }\n\n      if (generationChunk.generationInfo?.finish_reason != null) {\n        const finalReasoningDetails = [\n          ...reasoningTextByIndex.values(),\n          ...reasoningEncryptedById.values(),\n        ];\n        if (finalReasoningDetails.length > 0) {\n          generationChunk.message.additional_kwargs.reasoning_details =\n            finalReasoningDetails;\n        } else {\n          delete generationChunk.message.additional_kwargs.reasoning_details;\n        }\n        yield generationChunk;\n        continue;\n      }\n\n      delete generationChunk.message.additional_kwargs.reasoning_details;\n      yield generationChunk;\n    }\n  }\n}\n"],"names":[],"mappings":";;AAoEA,SAAS,qBAAqB,CAC5B,KAAc,EAAA;AAEd,IAAA,QACE,OAAO,KAAK,KAAK,QAAQ;AACzB,QAAA,KAAK,KAAK,IAAI;AACd,QAAA,MAAM,IAAI,KAAK;AACf,QAAA,KAAK,CAAC,IAAI,KAAK,gBAAgB;AAEnC;AAEA,SAAS,0BAA0B,CACjC,KAAc,EAAA;AAEd,IAAA,QACE,OAAO,KAAK,KAAK,QAAQ;AACzB,QAAA,KAAK,KAAK,IAAI;AACd,QAAA,MAAM,IAAI,KAAK;AACf,QAAA,KAAK,CAAC,IAAI,KAAK,qBAAqB;AAExC;AAEA,SAAS,mBAAmB,CAAC,KAAc,EAAA;IACzC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE;AACzB,QAAA,OAAO,EAAE;IACX;AACA,IAAA,OAAO,KAAK,CAAC,MAAM,CACjB,CAAC,MAAM,KACL,qBAAqB,CAAC,MAAM,CAAC,IAAI,0BAA0B,CAAC,MAAM,CAAC,CACtE;AACH;AAEM,MAAO,cAAe,SAAQ,UAAU,CAAA;AACpC,IAAA,mBAAmB;;AAEnB,IAAA,gBAAgB;AAExB,IAAA,WAAA,CAAY,OAA4B,EAAA;AACtC,QAAA,MAAM,EACJ,iBAAiB,EACjB,SAAS,EAAE,mBAAmB,EAC9B,WAAW,GAAG,EAAE,EAChB,GAAG,MAAM,EACV,GAAG,OAAO;;QAGX,MAAM,EAAE,SAAS,EAAE,WAAW,EAAE,GAAG,eAAe,EAAE,GAAG,WAE5B;AAE3B,QAAA,KAAK,CAAC;AACJ,YAAA,GAAG,MAAM;AACT,YAAA,WAAW,EAAE,eAAe;AAC5B,YAAA,uBAAuB,EAAE,IAAI;AAC7B,YAAA,gCAAgC,EAAE,IAAI;AACvC,SAAA,CAAC;;QAGF,IAAI,WAAW,IAAI,IAAI,IAAI,mBAAmB,IAAI,IAAI,EAAE;YACtD,IAAI,CAAC,mBAAmB,GAAG;AACzB,gBAAA,GAAG,WAAW;AACd,gBAAA,GAAG,mBAAmB;aACvB;QACH;AAEA,QAAA,IAAI,CAAC,gBAAgB,GAAG,iBAAiB;IAC3C;AACA,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,qBAAqB;IAC9B;;;;IAKS,gBAAgB,CACvB,OAAmC,EACnC,KAA6B,EAAA;QAO7B,MAAM,mBAAmB,GAAG,IAAI,CAAC,mBAAmB,CAAC,OAAO,CAAC;QAC7D,MAAM,MAAM,IACV,IAAI,CAAC,gBAAgB,CAAC,OAAO;cACzB,IAAI,CAAC,SAAS,CAAC,gBAAgB,CAAC,mBAAmB;AACrD,cAAE,IAAI,CAAC,WAAW,CAAC,gBAAgB,CAAC,mBAAmB,EAAE,KAAK,CAAC,CACjD;;;QAIlB,OAAO,MAAM,CAAC,gBAAgB;;QAG9B,MAAM,SAAS,GAAG,IAAI,CAAC,wBAAwB,CAAC,mBAAmB,CAAC;AACpE,QAAA,IAAI,SAAS,IAAI,IAAI,EAAE;AACrB,YAAA,MAAM,CAAC,SAAS,GAAG,SAAS;QAC9B;aAAO;YACL,OAAO,MAAM,CAAC,SAAS;QACzB;AAEA,QAAA,OAAO,MAAM;IACf;AAEQ,IAAA,wBAAwB,CAC9B,OAAmC,EAAA;AAEnC,QAAA,IAAI,SAA0C;;AAG9C,QAAA,IAAI,IAAI,CAAC,mBAAmB,IAAI,IAAI,EAAE;AACpC,YAAA,SAAS,GAAG,EAAE,GAAG,IAAI,CAAC,mBAAmB,EAAE;QAC7C;;QAGA,MAAM,WAAW,GAAG,IAAI,CAAC,kBAAkB,CAAC,OAAO,CAAC;AACpD,QAAA,IAAI,WAAW,EAAE,MAAM,IAAI,IAAI,EAAE;AAC/B,YAAA,SAAS,GAAG;AACV,gBAAA,GAAG,SAAS;gBACZ,MAAM,EAAE,WAAW,CAAC,MAAmC;aACxD;QACH;;QAGA,MAAM,aAAa,GAAI;AACrB,cAAE,SAAS;AACb,QAAA,IAAI,aAAa,IAAI,IAAI,EAAE;YACzB,SAAS,GAAG,EAAE,GAAG,SAAS,EAAE,GAAG,aAAa,EAAE;QAChD;;QAGA,IAAI,SAAS,IAAI,IAAI,IAAI,IAAI,CAAC,gBAAgB,KAAK,IAAI,EAAE;AACvD,YAAA,SAAS,GAAG,EAAE,OAAO,EAAE,IAAI,EAAE;QAC/B;AAEA,QAAA,OAAO,SAAS;IAClB;IAES,OAAO,qBAAqB,CACnC,QAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;AAErC,QAAA,MAAM,oBAAoB,GAAG,IAAI,GAAG,EAGjC;AACH,QAAA,MAAM,sBAAsB,GAAG,IAAI,GAAG,EAGnC;AAEH,QAAA,WAAW,MAAM,eAAe,IAAI,KAAK,CAAC,qBAAqB,CAC7D,QAAQ,EACR,OAAO,EACP,UAAU,CACX,EAAE;YACD,IAAI,oBAAoB,GAAG,EAAE;AAC7B,YAAA,MAAM,gBAAgB,GAAG,mBAAmB,CAC1C,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,iBAAiB,CAC5D;AAED,YAAA,KAAK,MAAM,MAAM,IAAI,gBAAgB,EAAE;AACrC,gBAAA,IAAI,MAAM,CAAC,IAAI,KAAK,gBAAgB,EAAE;AACpC,oBAAA,oBAAoB,IAAI,MAAM,CAAC,IAAI,IAAI,EAAE;AACzC,oBAAA,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,IAAI,CAAC;oBAC/B,MAAM,QAAQ,GAAG,oBAAoB,CAAC,GAAG,CAAC,KAAK,CAAC;AAChD,oBAAA,IAAI,QAAQ,IAAI,IAAI,EAAE;AACpB,wBAAA,QAAQ,CAAC,IAAI,GAAG,CAAA,EAAG,QAAQ,CAAC,IAAI,IAAI,EAAE,CAAA,EAAG,MAAM,CAAC,IAAI,IAAI,EAAE,EAAE;wBAC5D;oBACF;AACA,oBAAA,oBAAoB,CAAC,GAAG,CAAC,KAAK,EAAE;AAC9B,wBAAA,GAAG,MAAM;AACT,wBAAA,IAAI,EAAE,MAAM,CAAC,IAAI,IAAI,EAAE;AACxB,qBAAA,CAAC;oBACF;gBACF;AACA,gBAAA,IAAI,MAAM,CAAC,EAAE,IAAI,IAAI,EAAE;AACrB,oBAAA,sBAAsB,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,GAAG,MAAM,EAAE,CAAC;gBACtD;YACF;AAEA,YAAA,IACE,oBAAoB,CAAC,MAAM,GAAG,CAAC;gBAC/B,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,SAAS,IAAI,IAAI,EAC3D;AACA,gBAAA,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,SAAS;AACjD,oBAAA,oBAAoB;YACxB;YAEA,IAAI,eAAe,CAAC,cAAc,EAAE,aAAa,IAAI,IAAI,EAAE;AACzD,gBAAA,MAAM,qBAAqB,GAAG;oBAC5B,GAAG,oBAAoB,CAAC,MAAM,EAAE;oBAChC,GAAG,sBAAsB,CAAC,MAAM,EAAE;iBACnC;AACD,gBAAA,IAAI,qBAAqB,CAAC,MAAM,GAAG,CAAC,EAAE;AACpC,oBAAA,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,iBAAiB;AACzD,wBAAA,qBAAqB;gBACzB;qBAAO;AACL,oBAAA,OAAO,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,iBAAiB;gBACpE;AACA,gBAAA,MAAM,eAAe;gBACrB;YACF;AAEA,YAAA,OAAO,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,iBAAiB;AAClE,YAAA,MAAM,eAAe;QACvB;IACF;AACD;;;;"}
+{"version":3,"file":"index.mjs","sources":["../../../../src/llm/openrouter/index.ts"],"sourcesContent":["import { ChatOpenAI } from '@/llm/openai';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type {\n  ChatOpenAICallOptions,\n  OpenAIChatInput,\n  OpenAIClient,\n} from '@langchain/openai';\n\nexport type OpenRouterReasoningEffort =\n  | 'xhigh'\n  | 'high'\n  | 'medium'\n  | 'low'\n  | 'minimal'\n  | 'none';\n\nexport interface OpenRouterReasoning {\n  effort?: OpenRouterReasoningEffort;\n  max_tokens?: number;\n  exclude?: boolean;\n  enabled?: boolean;\n}\n\nexport interface ChatOpenRouterCallOptions\n  extends Omit<ChatOpenAICallOptions, 'reasoning'> {\n  /** @deprecated Use `reasoning` object instead */\n  include_reasoning?: boolean;\n  reasoning?: OpenRouterReasoning;\n  modelKwargs?: OpenAIChatInput['modelKwargs'];\n  promptCache?: boolean;\n}\n\nexport type ChatOpenRouterInput = Partial<\n  ChatOpenRouterCallOptions & OpenAIChatInput\n>;\n\n/** invocationParams return type extended with OpenRouter reasoning */\nexport type OpenRouterInvocationParams = Omit<\n  OpenAIClient.Chat.ChatCompletionCreateParams,\n  'messages'\n> & {\n  reasoning?: OpenRouterReasoning;\n};\n\ntype InvocationParamsExtra = {\n  streaming?: boolean;\n};\n\ninterface OpenRouterReasoningTextDetail {\n  type: 'reasoning.text';\n  text?: string;\n  format?: string;\n  index?: number;\n}\n\ninterface OpenRouterReasoningEncryptedDetail {\n  type: 'reasoning.encrypted';\n  id?: string;\n  data?: string;\n  format?: string;\n  index?: number;\n}\n\ntype OpenRouterReasoningDetail =\n  | OpenRouterReasoningTextDetail\n  | OpenRouterReasoningEncryptedDetail;\n\nfunction isReasoningTextDetail(\n  value: unknown\n): value is OpenRouterReasoningTextDetail {\n  return (\n    typeof value === 'object' &&\n    value !== null &&\n    'type' in value &&\n    value.type === 'reasoning.text'\n  );\n}\n\nfunction isReasoningEncryptedDetail(\n  value: unknown\n): value is OpenRouterReasoningEncryptedDetail {\n  return (\n    typeof value === 'object' &&\n    value !== null &&\n    'type' in value &&\n    value.type === 'reasoning.encrypted'\n  );\n}\n\nfunction getReasoningDetails(value: unknown): OpenRouterReasoningDetail[] {\n  if (!Array.isArray(value)) {\n    return [];\n  }\n  return value.filter(\n    (detail): detail is OpenRouterReasoningDetail =>\n      isReasoningTextDetail(detail) || isReasoningEncryptedDetail(detail)\n  );\n}\n\nexport class ChatOpenRouter extends ChatOpenAI {\n  private openRouterReasoning?: OpenRouterReasoning;\n  /** @deprecated Use `reasoning` object instead */\n  private includeReasoning?: boolean;\n\n  constructor(_fields: ChatOpenRouterInput) {\n    const fieldsWithoutPromptCache: ChatOpenRouterInput = { ..._fields };\n    delete fieldsWithoutPromptCache.promptCache;\n\n    const {\n      include_reasoning,\n      reasoning: openRouterReasoning,\n      modelKwargs = {},\n      ...fields\n    } = fieldsWithoutPromptCache;\n\n    // Extract reasoning from modelKwargs if provided there (e.g., from LLMConfig)\n    const { reasoning: mkReasoning, ...restModelKwargs } = modelKwargs as {\n      reasoning?: OpenRouterReasoning;\n    } & Record<string, unknown>;\n    const mergedReasoning =\n      mkReasoning != null || openRouterReasoning != null\n        ? {\n          ...mkReasoning,\n          ...openRouterReasoning,\n        }\n        : undefined;\n    const runtimeReasoning =\n      mergedReasoning ??\n      (include_reasoning === true ? { enabled: true } : undefined);\n    const parentModelKwargs =\n      runtimeReasoning == null\n        ? restModelKwargs\n        : { ...restModelKwargs, reasoning: runtimeReasoning };\n\n    super({\n      ...fields,\n      modelKwargs: parentModelKwargs,\n      includeReasoningDetails: true,\n      convertReasoningDetailsToContent: true,\n    });\n\n    // Merge reasoning config: modelKwargs.reasoning < constructor reasoning\n    if (mergedReasoning != null) {\n      this.openRouterReasoning = mergedReasoning;\n    }\n\n    this.includeReasoning = include_reasoning;\n  }\n  static lc_name(): 'LibreChatOpenRouter' {\n    return 'LibreChatOpenRouter';\n  }\n\n  // @ts-expect-error - OpenRouter reasoning extends OpenAI Reasoning with additional\n  // effort levels ('xhigh' | 'none' | 'minimal') not in ReasoningEffort.\n  // The parent's generic conditional return type cannot be widened in an override.\n  override invocationParams(\n    options?: this['ParsedCallOptions'],\n    extra?: InvocationParamsExtra\n  ): OpenRouterInvocationParams {\n    type MutableParams = Omit<\n      OpenAIClient.Chat.ChatCompletionCreateParams,\n      'messages'\n    > & { reasoning_effort?: string; reasoning?: OpenRouterReasoning };\n\n    const optionsWithDefaults = this._combineCallOptions(options);\n    const params = (\n      this._useResponsesApi(options)\n        ? this.responses.invocationParams(optionsWithDefaults)\n        : this.completions.invocationParams(optionsWithDefaults, extra)\n    ) as MutableParams;\n\n    // Remove the OpenAI-native reasoning_effort that the parent sets;\n    // OpenRouter uses a `reasoning` object instead\n    delete params.reasoning_effort;\n\n    // Build the OpenRouter reasoning config\n    const reasoning = this.buildOpenRouterReasoning(optionsWithDefaults);\n    if (reasoning != null) {\n      params.reasoning = reasoning;\n    } else {\n      delete params.reasoning;\n    }\n\n    return params;\n  }\n\n  private buildOpenRouterReasoning(\n    options?: this['ParsedCallOptions']\n  ): OpenRouterReasoning | undefined {\n    let reasoning: OpenRouterReasoning | undefined;\n\n    // 1. Instance-level reasoning config (from constructor)\n    if (this.openRouterReasoning != null) {\n      reasoning = { ...this.openRouterReasoning };\n    }\n\n    // 2. LangChain-style reasoning params (from parent's `this.reasoning`)\n    const lcReasoning = this.getReasoningParams(options);\n    if (lcReasoning?.effort != null) {\n      reasoning = {\n        ...reasoning,\n        effort: lcReasoning.effort as OpenRouterReasoningEffort,\n      };\n    }\n\n    // 3. Call-level reasoning override\n    const callReasoning = (options as ChatOpenRouterCallOptions | undefined)\n      ?.reasoning;\n    if (callReasoning != null) {\n      reasoning = { ...reasoning, ...callReasoning };\n    }\n\n    // 4. Legacy include_reasoning backward compatibility\n    if (reasoning == null && this.includeReasoning === true) {\n      reasoning = { enabled: true };\n    }\n\n    return reasoning;\n  }\n\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const reasoningTextByIndex = new Map<\n      number,\n      OpenRouterReasoningTextDetail\n    >();\n    const reasoningEncryptedById = new Map<\n      string,\n      OpenRouterReasoningEncryptedDetail\n    >();\n\n    for await (const generationChunk of super._streamResponseChunks(\n      messages,\n      options,\n      runManager\n    )) {\n      let currentReasoningText = '';\n      const reasoningDetails = getReasoningDetails(\n        generationChunk.message.additional_kwargs.reasoning_details\n      );\n\n      for (const detail of reasoningDetails) {\n        if (detail.type === 'reasoning.text') {\n          currentReasoningText += detail.text ?? '';\n          const index = detail.index ?? 0;\n          const existing = reasoningTextByIndex.get(index);\n          if (existing != null) {\n            existing.text = `${existing.text ?? ''}${detail.text ?? ''}`;\n            continue;\n          }\n          reasoningTextByIndex.set(index, {\n            ...detail,\n            text: detail.text ?? '',\n          });\n          continue;\n        }\n        if (detail.id != null) {\n          reasoningEncryptedById.set(detail.id, { ...detail });\n        }\n      }\n\n      if (\n        currentReasoningText.length > 0 &&\n        generationChunk.message.additional_kwargs.reasoning == null\n      ) {\n        generationChunk.message.additional_kwargs.reasoning =\n          currentReasoningText;\n      }\n\n      if (generationChunk.generationInfo?.finish_reason != null) {\n        const finalReasoningDetails = [\n          ...reasoningTextByIndex.values(),\n          ...reasoningEncryptedById.values(),\n        ];\n        if (finalReasoningDetails.length > 0) {\n          generationChunk.message.additional_kwargs.reasoning_details =\n            finalReasoningDetails;\n        } else {\n          delete generationChunk.message.additional_kwargs.reasoning_details;\n        }\n        yield generationChunk;\n        continue;\n      }\n\n      delete generationChunk.message.additional_kwargs.reasoning_details;\n      yield generationChunk;\n    }\n  }\n}\n"],"names":[],"mappings":";;AAqEA,SAAS,qBAAqB,CAC5B,KAAc,EAAA;AAEd,IAAA,QACE,OAAO,KAAK,KAAK,QAAQ;AACzB,QAAA,KAAK,KAAK,IAAI;AACd,QAAA,MAAM,IAAI,KAAK;AACf,QAAA,KAAK,CAAC,IAAI,KAAK,gBAAgB;AAEnC;AAEA,SAAS,0BAA0B,CACjC,KAAc,EAAA;AAEd,IAAA,QACE,OAAO,KAAK,KAAK,QAAQ;AACzB,QAAA,KAAK,KAAK,IAAI;AACd,QAAA,MAAM,IAAI,KAAK;AACf,QAAA,KAAK,CAAC,IAAI,KAAK,qBAAqB;AAExC;AAEA,SAAS,mBAAmB,CAAC,KAAc,EAAA;IACzC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE;AACzB,QAAA,OAAO,EAAE;IACX;AACA,IAAA,OAAO,KAAK,CAAC,MAAM,CACjB,CAAC,MAAM,KACL,qBAAqB,CAAC,MAAM,CAAC,IAAI,0BAA0B,CAAC,MAAM,CAAC,CACtE;AACH;AAEM,MAAO,cAAe,SAAQ,UAAU,CAAA;AACpC,IAAA,mBAAmB;;AAEnB,IAAA,gBAAgB;AAExB,IAAA,WAAA,CAAY,OAA4B,EAAA;AACtC,QAAA,MAAM,wBAAwB,GAAwB,EAAE,GAAG,OAAO,EAAE;QACpE,OAAO,wBAAwB,CAAC,WAAW;AAE3C,QAAA,MAAM,EACJ,iBAAiB,EACjB,SAAS,EAAE,mBAAmB,EAC9B,WAAW,GAAG,EAAE,EAChB,GAAG,MAAM,EACV,GAAG,wBAAwB;;QAG5B,MAAM,EAAE,SAAS,EAAE,WAAW,EAAE,GAAG,eAAe,EAAE,GAAG,WAE5B;QAC3B,MAAM,eAAe,GACnB,WAAW,IAAI,IAAI,IAAI,mBAAmB,IAAI;AAC5C,cAAE;AACA,gBAAA,GAAG,WAAW;AACd,gBAAA,GAAG,mBAAmB;AACvB;cACC,SAAS;QACf,MAAM,gBAAgB,GACpB,eAAe;AACf,aAAC,iBAAiB,KAAK,IAAI,GAAG,EAAE,OAAO,EAAE,IAAI,EAAE,GAAG,SAAS,CAAC;AAC9D,QAAA,MAAM,iBAAiB,GACrB,gBAAgB,IAAI;AAClB,cAAE;cACA,EAAE,GAAG,eAAe,EAAE,SAAS,EAAE,gBAAgB,EAAE;AAEzD,QAAA,KAAK,CAAC;AACJ,YAAA,GAAG,MAAM;AACT,YAAA,WAAW,EAAE,iBAAiB;AAC9B,YAAA,uBAAuB,EAAE,IAAI;AAC7B,YAAA,gCAAgC,EAAE,IAAI;AACvC,SAAA,CAAC;;AAGF,QAAA,IAAI,eAAe,IAAI,IAAI,EAAE;AAC3B,YAAA,IAAI,CAAC,mBAAmB,GAAG,eAAe;QAC5C;AAEA,QAAA,IAAI,CAAC,gBAAgB,GAAG,iBAAiB;IAC3C;AACA,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,qBAAqB;IAC9B;;;;IAKS,gBAAgB,CACvB,OAAmC,EACnC,KAA6B,EAAA;QAO7B,MAAM,mBAAmB,GAAG,IAAI,CAAC,mBAAmB,CAAC,OAAO,CAAC;QAC7D,MAAM,MAAM,IACV,IAAI,CAAC,gBAAgB,CAAC,OAAO;cACzB,IAAI,CAAC,SAAS,CAAC,gBAAgB,CAAC,mBAAmB;AACrD,cAAE,IAAI,CAAC,WAAW,CAAC,gBAAgB,CAAC,mBAAmB,EAAE,KAAK,CAAC,CACjD;;;QAIlB,OAAO,MAAM,CAAC,gBAAgB;;QAG9B,MAAM,SAAS,GAAG,IAAI,CAAC,wBAAwB,CAAC,mBAAmB,CAAC;AACpE,QAAA,IAAI,SAAS,IAAI,IAAI,EAAE;AACrB,YAAA,MAAM,CAAC,SAAS,GAAG,SAAS;QAC9B;aAAO;YACL,OAAO,MAAM,CAAC,SAAS;QACzB;AAEA,QAAA,OAAO,MAAM;IACf;AAEQ,IAAA,wBAAwB,CAC9B,OAAmC,EAAA;AAEnC,QAAA,IAAI,SAA0C;;AAG9C,QAAA,IAAI,IAAI,CAAC,mBAAmB,IAAI,IAAI,EAAE;AACpC,YAAA,SAAS,GAAG,EAAE,GAAG,IAAI,CAAC,mBAAmB,EAAE;QAC7C;;QAGA,MAAM,WAAW,GAAG,IAAI,CAAC,kBAAkB,CAAC,OAAO,CAAC;AACpD,QAAA,IAAI,WAAW,EAAE,MAAM,IAAI,IAAI,EAAE;AAC/B,YAAA,SAAS,GAAG;AACV,gBAAA,GAAG,SAAS;gBACZ,MAAM,EAAE,WAAW,CAAC,MAAmC;aACxD;QACH;;QAGA,MAAM,aAAa,GAAI;AACrB,cAAE,SAAS;AACb,QAAA,IAAI,aAAa,IAAI,IAAI,EAAE;YACzB,SAAS,GAAG,EAAE,GAAG,SAAS,EAAE,GAAG,aAAa,EAAE;QAChD;;QAGA,IAAI,SAAS,IAAI,IAAI,IAAI,IAAI,CAAC,gBAAgB,KAAK,IAAI,EAAE;AACvD,YAAA,SAAS,GAAG,EAAE,OAAO,EAAE,IAAI,EAAE;QAC/B;AAEA,QAAA,OAAO,SAAS;IAClB;IAES,OAAO,qBAAqB,CACnC,QAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;AAErC,QAAA,MAAM,oBAAoB,GAAG,IAAI,GAAG,EAGjC;AACH,QAAA,MAAM,sBAAsB,GAAG,IAAI,GAAG,EAGnC;AAEH,QAAA,WAAW,MAAM,eAAe,IAAI,KAAK,CAAC,qBAAqB,CAC7D,QAAQ,EACR,OAAO,EACP,UAAU,CACX,EAAE;YACD,IAAI,oBAAoB,GAAG,EAAE;AAC7B,YAAA,MAAM,gBAAgB,GAAG,mBAAmB,CAC1C,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,iBAAiB,CAC5D;AAED,YAAA,KAAK,MAAM,MAAM,IAAI,gBAAgB,EAAE;AACrC,gBAAA,IAAI,MAAM,CAAC,IAAI,KAAK,gBAAgB,EAAE;AACpC,oBAAA,oBAAoB,IAAI,MAAM,CAAC,IAAI,IAAI,EAAE;AACzC,oBAAA,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,IAAI,CAAC;oBAC/B,MAAM,QAAQ,GAAG,oBAAoB,CAAC,GAAG,CAAC,KAAK,CAAC;AAChD,oBAAA,IAAI,QAAQ,IAAI,IAAI,EAAE;AACpB,wBAAA,QAAQ,CAAC,IAAI,GAAG,CAAA,EAAG,QAAQ,CAAC,IAAI,IAAI,EAAE,CAAA,EAAG,MAAM,CAAC,IAAI,IAAI,EAAE,EAAE;wBAC5D;oBACF;AACA,oBAAA,oBAAoB,CAAC,GAAG,CAAC,KAAK,EAAE;AAC9B,wBAAA,GAAG,MAAM;AACT,wBAAA,IAAI,EAAE,MAAM,CAAC,IAAI,IAAI,EAAE;AACxB,qBAAA,CAAC;oBACF;gBACF;AACA,gBAAA,IAAI,MAAM,CAAC,EAAE,IAAI,IAAI,EAAE;AACrB,oBAAA,sBAAsB,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,GAAG,MAAM,EAAE,CAAC;gBACtD;YACF;AAEA,YAAA,IACE,oBAAoB,CAAC,MAAM,GAAG,CAAC;gBAC/B,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,SAAS,IAAI,IAAI,EAC3D;AACA,gBAAA,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,SAAS;AACjD,oBAAA,oBAAoB;YACxB;YAEA,IAAI,eAAe,CAAC,cAAc,EAAE,aAAa,IAAI,IAAI,EAAE;AACzD,gBAAA,MAAM,qBAAqB,GAAG;oBAC5B,GAAG,oBAAoB,CAAC,MAAM,EAAE;oBAChC,GAAG,sBAAsB,CAAC,MAAM,EAAE;iBACnC;AACD,gBAAA,IAAI,qBAAqB,CAAC,MAAM,GAAG,CAAC,EAAE;AACpC,oBAAA,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,iBAAiB;AACzD,wBAAA,qBAAqB;gBACzB;qBAAO;AACL,oBAAA,OAAO,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,iBAAiB;gBACpE;AACA,gBAAA,MAAM,eAAe;gBACrB;YACF;AAEA,YAAA,OAAO,eAAe,CAAC,OAAO,CAAC,iBAAiB,CAAC,iBAAiB;AAClE,YAAA,MAAM,eAAe;QACvB;IACF;AACD;;;;"}

package/dist/esm/llm/openrouter/toolCache.mjs

@@ -0,0 +1,53 @@
+import { _convertToOpenAITool } from '../openai/index.mjs';
+
+const CACHE_CONTROL = { type: 'ephemeral' };
+function getToolName(tool) {
+    const candidate = tool;
+    if (typeof candidate.name === 'string') {
+        return candidate.name;
+    }
+    if (typeof candidate.function?.name === 'string') {
+        return candidate.function.name;
+    }
+    return undefined;
+}
+function hasDeferredMarker(tool) {
+    return tool.defer_loading === true;
+}
+function toOpenRouterTool(tool) {
+    const converted = _convertToOpenAITool(tool);
+    if (hasDeferredMarker(tool)) {
+        return { ...converted, defer_loading: true };
+    }
+    return converted;
+}
+function markCacheControl(tool) {
+    return {
+        ...tool,
+        cache_control: CACHE_CONTROL,
+    };
+}
+function partitionAndMarkOpenRouterToolCache(tools, isDeferred) {
+    if (tools == null || tools.length === 0) {
+        return tools;
+    }
+    const staticTools = [];
+    const deferredTools = [];
+    for (const tool of tools) {
+        const converted = toOpenRouterTool(tool);
+        const name = getToolName(converted) ?? getToolName(tool);
+        if (name != null && isDeferred(name)) {
+            deferredTools.push(converted);
+            continue;
+        }
+        staticTools.push(converted);
+    }
+    if (staticTools.length === 0) {
+        return [...deferredTools];
+    }
+    staticTools[staticTools.length - 1] = markCacheControl(staticTools[staticTools.length - 1]);
+    return [...staticTools, ...deferredTools];
+}
+
+export { partitionAndMarkOpenRouterToolCache };
+//# sourceMappingURL=toolCache.mjs.map

package/dist/esm/llm/openrouter/toolCache.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"toolCache.mjs","sources":["../../../../src/llm/openrouter/toolCache.ts"],"sourcesContent":["import type { BindToolsInput } from '@langchain/core/language_models/chat_models';\nimport type { OpenAIClient } from '@langchain/openai';\nimport type { GraphTools } from '@/types';\nimport { _convertToOpenAITool } from '@/llm/openai';\n\nconst CACHE_CONTROL = { type: 'ephemeral' as const };\n\ntype OpenRouterToolWithCacheControl = OpenAIClient.ChatCompletionTool & {\n  cache_control?: typeof CACHE_CONTROL;\n  defer_loading?: boolean;\n};\n\ntype ToolNameCandidate = {\n  name?: unknown;\n  function?: {\n    name?: unknown;\n  };\n  defer_loading?: unknown;\n};\n\nfunction getToolName(tool: unknown): string | undefined {\n  const candidate = tool as ToolNameCandidate;\n  if (typeof candidate.name === 'string') {\n    return candidate.name;\n  }\n  if (typeof candidate.function?.name === 'string') {\n    return candidate.function.name;\n  }\n  return undefined;\n}\n\nfunction hasDeferredMarker(tool: unknown): boolean {\n  return (tool as ToolNameCandidate).defer_loading === true;\n}\n\nfunction toOpenRouterTool(tool: unknown): OpenRouterToolWithCacheControl {\n  const converted = _convertToOpenAITool(\n    tool as BindToolsInput\n  ) as OpenRouterToolWithCacheControl;\n\n  if (hasDeferredMarker(tool)) {\n    return { ...converted, defer_loading: true };\n  }\n\n  return converted;\n}\n\nfunction markCacheControl(\n  tool: OpenRouterToolWithCacheControl\n): OpenRouterToolWithCacheControl {\n  return {\n    ...tool,\n    cache_control: CACHE_CONTROL,\n  };\n}\n\nexport function partitionAndMarkOpenRouterToolCache(\n  tools: GraphTools | undefined,\n  isDeferred: (toolName: string) => boolean\n): GraphTools | undefined {\n  if (tools == null || tools.length === 0) {\n    return tools;\n  }\n\n  const staticTools: OpenRouterToolWithCacheControl[] = [];\n  const deferredTools: OpenRouterToolWithCacheControl[] = [];\n\n  for (const tool of tools as readonly unknown[]) {\n    const converted = toOpenRouterTool(tool);\n    const name = getToolName(converted) ?? getToolName(tool);\n\n    if (name != null && isDeferred(name)) {\n      deferredTools.push(converted);\n      continue;\n    }\n\n    staticTools.push(converted);\n  }\n\n  if (staticTools.length === 0) {\n    return [...deferredTools] as GraphTools;\n  }\n\n  staticTools[staticTools.length - 1] = markCacheControl(\n    staticTools[staticTools.length - 1]\n  );\n\n  return [...staticTools, ...deferredTools] as GraphTools;\n}\n"],"names":[],"mappings":";;AAKA,MAAM,aAAa,GAAG,EAAE,IAAI,EAAE,WAAoB,EAAE;AAepD,SAAS,WAAW,CAAC,IAAa,EAAA;IAChC,MAAM,SAAS,GAAG,IAAyB;AAC3C,IAAA,IAAI,OAAO,SAAS,CAAC,IAAI,KAAK,QAAQ,EAAE;QACtC,OAAO,SAAS,CAAC,IAAI;IACvB;IACA,IAAI,OAAO,SAAS,CAAC,QAAQ,EAAE,IAAI,KAAK,QAAQ,EAAE;AAChD,QAAA,OAAO,SAAS,CAAC,QAAQ,CAAC,IAAI;IAChC;AACA,IAAA,OAAO,SAAS;AAClB;AAEA,SAAS,iBAAiB,CAAC,IAAa,EAAA;AACtC,IAAA,OAAQ,IAA0B,CAAC,aAAa,KAAK,IAAI;AAC3D;AAEA,SAAS,gBAAgB,CAAC,IAAa,EAAA;AACrC,IAAA,MAAM,SAAS,GAAG,oBAAoB,CACpC,IAAsB,CACW;AAEnC,IAAA,IAAI,iBAAiB,CAAC,IAAI,CAAC,EAAE;QAC3B,OAAO,EAAE,GAAG,SAAS,EAAE,aAAa,EAAE,IAAI,EAAE;IAC9C;AAEA,IAAA,OAAO,SAAS;AAClB;AAEA,SAAS,gBAAgB,CACvB,IAAoC,EAAA;IAEpC,OAAO;AACL,QAAA,GAAG,IAAI;AACP,QAAA,aAAa,EAAE,aAAa;KAC7B;AACH;AAEM,SAAU,mCAAmC,CACjD,KAA6B,EAC7B,UAAyC,EAAA;IAEzC,IAAI,KAAK,IAAI,IAAI,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE;AACvC,QAAA,OAAO,KAAK;IACd;IAEA,MAAM,WAAW,GAAqC,EAAE;IACxD,MAAM,aAAa,GAAqC,EAAE;AAE1D,IAAA,KAAK,MAAM,IAAI,IAAI,KAA2B,EAAE;AAC9C,QAAA,MAAM,SAAS,GAAG,gBAAgB,CAAC,IAAI,CAAC;QACxC,MAAM,IAAI,GAAG,WAAW,CAAC,SAAS,CAAC,IAAI,WAAW,CAAC,IAAI,CAAC;QAExD,IAAI,IAAI,IAAI,IAAI,IAAI,UAAU,CAAC,IAAI,CAAC,EAAE;AACpC,YAAA,aAAa,CAAC,IAAI,CAAC,SAAS,CAAC;YAC7B;QACF;AAEA,QAAA,WAAW,CAAC,IAAI,CAAC,SAAS,CAAC;IAC7B;AAEA,IAAA,IAAI,WAAW,CAAC,MAAM,KAAK,CAAC,EAAE;AAC5B,QAAA,OAAO,CAAC,GAAG,aAAa,CAAe;IACzC;AAEA,IAAA,WAAW,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,CAAC,GAAG,gBAAgB,CACpD,WAAW,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,CAAC,CACpC;AAED,IAAA,OAAO,CAAC,GAAG,WAAW,EAAE,GAAG,aAAa,CAAe;AACzD;;;;"}

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/esm/tools/ToolNode.mjs

@@ -32,6 +32,9 @@ import './local/attachments.mjs';
 function isSend(value) {
     return value instanceof Send;
 }
+function isHandoffToolName(name) {
+    return name.startsWith(Constants.LC_TRANSFER_TO_);
+}
 /**
  * Format a fail-closed diagnostic for malformed approval-decision
  * fields. Hosts deserialize resume payloads from untyped JSON, so
@@ -394,6 +397,58 @@ class ToolNode extends RunnableCallable {
     getFileCheckpointer() {
         return this.fileCheckpointer;
     }
+    *getRegisteredHandoffNames() {
+        if (this.directToolNames != null) {
+            for (const toolName of this.directToolNames) {
+                yield toolName;
+            }
+        }
+        for (const toolName of this.toolMap.keys()) {
+            if (this.directToolNames?.has(toolName) === true) {
+                continue;
+            }
+            yield toolName;
+        }
+    }
+    hasRegisteredHandoffTool() {
+        for (const toolName of this.getRegisteredHandoffNames()) {
+            if (isHandoffToolName(toolName)) {
+                return true;
+            }
+        }
+        return false;
+    }
+    getHandoffToolNameSuggestion(callName) {
+        if (!isHandoffToolName(callName)) {
+            return undefined;
+        }
+        let suggestion;
+        for (const toolName of this.getRegisteredHandoffNames()) {
+            if (!isHandoffToolName(toolName) ||
+                toolName.length >= callName.length ||
+                !callName.startsWith(toolName)) {
+                continue;
+            }
+            if (suggestion == null || toolName.length > suggestion.length) {
+                suggestion = toolName;
+            }
+        }
+        return suggestion;
+    }
+    shouldHandleUnknownHandoffLocally(callName, hasRegisteredHandoffTool) {
+        if (!isHandoffToolName(callName) || this.toolMap.has(callName)) {
+            return false;
+        }
+        return hasRegisteredHandoffTool ?? this.hasRegisteredHandoffTool();
+    }
+    getUnknownToolErrorMessage(callName) {
+        const suggestion = this.getHandoffToolNameSuggestion(callName);
+        if (suggestion == null) {
+            return `Tool "${callName}" not found.`;
+        }
+        return (`Tool "${callName}" not found. Did you mean "${suggestion}"? ` +
+            'Handoff tool names must match exactly.');
+    }
     /**
      * Flush the per-Run direct-path turn cache. Called by the Graph at
      * end-of-Run via `clearHeavyState`. The map intentionally survives
@@ -479,7 +534,7 @@ class ToolNode extends RunnableCallable {
         const runId = batchScopeId ?? config.configurable?.run_id;
         try {
             if (tool === undefined) {
-                throw new Error(`Tool "${call.name}" not found.`);
+                throw new Error(this.getUnknownToolErrorMessage(call.name));
             }
             /**
              * `usageCount` is the per-tool-name invocation index that
@@ -2121,8 +2176,9 @@ class ToolNode extends RunnableCallable {
         const turn = this.toolOutputRegistry?.nextTurn(batchScopeId) ?? 0;
         let outputs;
         if (this.isSendInput(input)) {
-            const isDirectTool = this.directToolNames?.has(input.lg_tool_call.name);
-            if (this.eventDrivenMode && isDirectTool !== true) {
+            const isLocalTool = this.directToolNames?.has(input.lg_tool_call.name) === true ||
+                this.shouldHandleUnknownHandoffLocally(input.lg_tool_call.name);
+            if (this.eventDrivenMode && !isLocalTool) {
                 return this.executeViaEvent([input.lg_tool_call], config, input, {
                     batchIndices: [0],
                     turn,
@@ -2205,26 +2261,28 @@ class ToolNode extends RunnableCallable {
                         false));
             }) ?? [];
             if (this.eventDrivenMode && filteredCalls.length > 0) {
-                const filteredIndices = filteredCalls.map((_, idx) => idx);
-                if (!this.directToolNames || this.directToolNames.size === 0) {
-                    return this.executeViaEvent(filteredCalls, config, input, {
-                        batchIndices: filteredIndices,
-                        turn,
-                        batchScopeId,
-                    });
-                }
+                const directToolNames = this.directToolNames;
+                const hasRegisteredHandoffTool = this.hasRegisteredHandoffTool();
                 const directEntries = [];
                 const eventEntries = [];
                 for (let i = 0; i < filteredCalls.length; i++) {
                     const call = filteredCalls[i];
                     const entry = { call, batchIndex: i };
-                    if (this.directToolNames.has(call.name)) {
+                    if (directToolNames?.has(call.name) === true ||
+                        this.shouldHandleUnknownHandoffLocally(call.name, hasRegisteredHandoffTool)) {
                         directEntries.push(entry);
                     }
                     else {
                         eventEntries.push(entry);
                     }
                 }
+                if (directEntries.length === 0) {
+                    return this.executeViaEvent(filteredCalls, config, input, {
+                        batchIndices: eventEntries.map((entry) => entry.batchIndex),
+                        turn,
+                        batchScopeId,
+                    });
+                }
                 const directCalls = directEntries.map((e) => e.call);
                 const directIndices = directEntries.map((e) => e.batchIndex);
                 const eventCalls = eventEntries.map((e) => e.call);