@librechat/agents 3.3.12 → 3.4.0

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

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","names":[],"sources":["../../../../src/llm/openrouter/index.ts"],"sourcesContent":["import type {\n  ChatOpenAICallOptions,\n  OpenAIChatInput,\n  OpenAIClient,\n} from '@langchain/openai';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { SeenScalarMetadata } from '@/llm/openai/streamMetadata';\nimport type { PromptCacheTtl } from '@/messages/cache';\nimport { dropRepeatedScalarMetadata } from '@/llm/openai/streamMetadata';\nimport { ChatOpenAI, emitStreamChunkCallback } from '@/llm/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  useResponsesApi?: boolean;\n  promptCache?: boolean;\n  /**\n   * Prompt-cache breakpoint TTL. Defaults to `'1h'` (extended cache) when\n   * `promptCache` is enabled; set `'5m'` for the legacy 5-minute behavior.\n   * OpenRouter forwards this to Claude upstreams (Anthropic / Bedrock / Vertex),\n   * which downgrade to 5m where the extended TTL isn't supported.\n   */\n  promptCacheTtl?: PromptCacheTtl;\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  cache_control?: { type: 'ephemeral'; ttl?: '1h' };\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    delete fieldsWithoutPromptCache.promptCacheTtl;\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      preserveToolCacheControl: true,\n      responsesPromptCache: _fields.promptCache,\n      responsesPromptCacheTtl: _fields.promptCacheTtl,\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  // OpenRouter widens OpenAI reasoning with extra effort levels ('xhigh' | 'none' | 'minimal').\n  override invocationParams(\n    options?: this['ParsedCallOptions'],\n    extra?: InvocationParamsExtra\n  ): OpenRouterInvocationParams {\n    type MutableParams = Omit<\n      OpenAIClient.Chat.ChatCompletionCreateParams,\n      'messages'\n    > & {\n      reasoning_effort?: string;\n      reasoning?: OpenRouterReasoning;\n      cache_control?: { type: 'ephemeral'; ttl?: '1h' };\n    };\n\n    const optionsWithDefaults = this._combineCallOptions(options);\n    const useResponsesApi = this._useResponsesApi(options);\n    const params = (\n      useResponsesApi\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    // Consume the raw (un-deduped) stream so `finish_reason` survives as the\n    // reasoning-flush signal below; de-duplicate after that, before emitting.\n    const seenScalarMetadata: SeenScalarMetadata = new Map();\n\n    for await (const generationChunk of super._streamRawResponseChunks(\n      messages,\n      options,\n      undefined\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      } else {\n        delete generationChunk.message.additional_kwargs.reasoning_details;\n      }\n\n      dropRepeatedScalarMetadata(generationChunk, seenScalarMetadata);\n      await emitStreamChunkCallback(generationChunk, runManager);\n      yield generationChunk;\n    }\n  }\n}\n"],"mappings":";;;AAiFA,SAAS,sBACP,OACwC;CACxC,OACE,OAAO,UAAU,YACjB,UAAU,QACV,UAAU,SACV,MAAM,SAAS;AAEnB;AAEA,SAAS,2BACP,OAC6C;CAC7C,OACE,OAAO,UAAU,YACjB,UAAU,QACV,UAAU,SACV,MAAM,SAAS;AAEnB;AAEA,SAAS,oBAAoB,OAA6C;CACxE,IAAI,CAAC,MAAM,QAAQ,KAAK,GACtB,OAAO,CAAC;CAEV,OAAO,MAAM,QACV,WACC,sBAAsB,MAAM,KAAK,2BAA2B,MAAM,CACtE;AACF;AAEA,IAAa,iBAAb,cAAoC,WAAW;CAC7C;;CAEA;CAEA,YAAY,SAA8B;EACxC,MAAM,2BAAgD,EAAE,GAAG,QAAQ;EACnE,OAAO,yBAAyB;EAChC,OAAO,yBAAyB;EAEhC,MAAM,EACJ,mBACA,WAAW,qBACX,cAAc,CAAC,GACf,GAAG,WACD;EAGJ,MAAM,EAAE,WAAW,aAAa,GAAG,oBAAoB;EAGvD,MAAM,kBACJ,eAAe,QAAQ,uBAAuB,OAC1C;GACA,GAAG;GACH,GAAG;EACL,IACE,KAAA;EACN,MAAM,mBACJ,oBACC,sBAAsB,OAAO,EAAE,SAAS,KAAK,IAAI,KAAA;EACpD,MAAM,oBACJ,oBAAoB,OAChB,kBACA;GAAE,GAAG;GAAiB,WAAW;EAAiB;EAExD,MAAM;GACJ,GAAG;GACH,aAAa;GACb,yBAAyB;GACzB,kCAAkC;GAClC,0BAA0B;GAC1B,sBAAsB,QAAQ;GAC9B,yBAAyB,QAAQ;EACnC,CAAC;EAGD,IAAI,mBAAmB,MACrB,KAAK,sBAAsB;EAG7B,KAAK,mBAAmB;CAC1B;CACA,OAAO,UAAiC;EACtC,OAAO;CACT;CAGA,iBACE,SACA,OAC4B;EAU5B,MAAM,sBAAsB,KAAK,oBAAoB,OAAO;EAE5D,MAAM,SADkB,KAAK,iBAAiB,OAE9B,IACV,KAAK,UAAU,iBAAiB,mBAAmB,IACnD,KAAK,YAAY,iBAAiB,qBAAqB,KAAK;EAKlE,OAAO,OAAO;EAGd,MAAM,YAAY,KAAK,yBAAyB,mBAAmB;EACnE,IAAI,aAAa,MACf,OAAO,YAAY;OAEnB,OAAO,OAAO;EAGhB,OAAO;CACT;CAEA,yBACE,SACiC;EACjC,IAAI;EAGJ,IAAI,KAAK,uBAAuB,MAC9B,YAAY,EAAE,GAAG,KAAK,oBAAoB;EAI5C,MAAM,cAAc,KAAK,mBAAmB,OAAO;EACnD,IAAI,aAAa,UAAU,MACzB,YAAY;GACV,GAAG;GACH,QAAQ,YAAY;EACtB;EAIF,MAAM,gBAAiB,SACnB;EACJ,IAAI,iBAAiB,MACnB,YAAY;GAAE,GAAG;GAAW,GAAG;EAAc;EAI/C,IAAI,aAAa,QAAQ,KAAK,qBAAqB,MACjD,YAAY,EAAE,SAAS,KAAK;EAG9B,OAAO;CACT;CAEA,OAAgB,sBACd,UACA,SACA,YACqC;EACrC,MAAM,uCAAuB,IAAI,IAG/B;EACF,MAAM,yCAAyB,IAAI,IAGjC;EAGF,MAAM,qCAAyC,IAAI,IAAI;EAEvD,WAAW,MAAM,mBAAmB,MAAM,yBACxC,UACA,SACA,KAAA,CACF,GAAG;GACD,IAAI,uBAAuB;GAC3B,MAAM,mBAAmB,oBACvB,gBAAgB,QAAQ,kBAAkB,iBAC5C;GAEA,KAAK,MAAM,UAAU,kBAAkB;IACrC,IAAI,OAAO,SAAS,kBAAkB;KACpC,wBAAwB,OAAO,QAAQ;KACvC,MAAM,QAAQ,OAAO,SAAS;KAC9B,MAAM,WAAW,qBAAqB,IAAI,KAAK;KAC/C,IAAI,YAAY,MAAM;MACpB,SAAS,OAAO,GAAG,SAAS,QAAQ,KAAK,OAAO,QAAQ;MACxD;KACF;KACA,qBAAqB,IAAI,OAAO;MAC9B,GAAG;MACH,MAAM,OAAO,QAAQ;KACvB,CAAC;KACD;IACF;IACA,IAAI,OAAO,MAAM,MACf,uBAAuB,IAAI,OAAO,IAAI,EAAE,GAAG,OAAO,CAAC;GAEvD;GAEA,IACE,qBAAqB,SAAS,KAC9B,gBAAgB,QAAQ,kBAAkB,aAAa,MAEvD,gBAAgB,QAAQ,kBAAkB,YACxC;GAGJ,IAAI,gBAAgB,gBAAgB,iBAAiB,MAAM;IACzD,MAAM,wBAAwB,CAC5B,GAAG,qBAAqB,OAAO,GAC/B,GAAG,uBAAuB,OAAO,CACnC;IACA,IAAI,sBAAsB,SAAS,GACjC,gBAAgB,QAAQ,kBAAkB,oBACxC;SAEF,OAAO,gBAAgB,QAAQ,kBAAkB;GAErD,OACE,OAAO,gBAAgB,QAAQ,kBAAkB;GAGnD,2BAA2B,iBAAiB,kBAAkB;GAC9D,MAAM,wBAAwB,iBAAiB,UAAU;GACzD,MAAM;EACR;CACF;AACF"}
+{"version":3,"file":"index.mjs","names":[],"sources":["../../../../src/llm/openrouter/index.ts"],"sourcesContent":["import type {\n  ChatOpenAICallOptions,\n  OpenAIChatInput,\n  OpenAIClient,\n} from '@langchain/openai';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { SeenScalarMetadata } from '@/llm/openai/streamMetadata';\nimport type { PromptCacheTtl } from '@/messages/cache';\nimport { dropRepeatedScalarMetadata } from '@/llm/openai/streamMetadata';\nimport { ChatOpenAI, emitStreamChunkCallback } from '@/llm/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  useResponsesApi?: boolean;\n  promptCache?: boolean;\n  /**\n   * Prompt-cache breakpoint TTL. Defaults to `'1h'` (extended cache) when\n   * `promptCache` is enabled; set `'5m'` for the legacy 5-minute behavior.\n   * OpenRouter forwards this to Claude upstreams (Anthropic / Bedrock / Vertex),\n   * which downgrade to 5m where the extended TTL isn't supported.\n   */\n  promptCacheTtl?: PromptCacheTtl;\n}\n\nexport type ChatOpenRouterInput = Partial<\n  ChatOpenRouterCallOptions & OpenAIChatInput\n> & {\n  /** Minimum delay in ms between visible streamed deltas (default 25; 0 disables). */\n  _lc_stream_delay?: number;\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  cache_control?: { type: 'ephemeral'; ttl?: '1h' };\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    delete fieldsWithoutPromptCache.promptCacheTtl;\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      preserveToolCacheControl: true,\n      responsesPromptCache: _fields.promptCache,\n      responsesPromptCacheTtl: _fields.promptCacheTtl,\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  // OpenRouter widens OpenAI reasoning with extra effort levels ('xhigh' | 'none' | 'minimal').\n  override invocationParams(\n    options?: this['ParsedCallOptions'],\n    extra?: InvocationParamsExtra\n  ): OpenRouterInvocationParams {\n    type MutableParams = Omit<\n      OpenAIClient.Chat.ChatCompletionCreateParams,\n      'messages'\n    > & {\n      reasoning_effort?: string;\n      reasoning?: OpenRouterReasoning;\n      cache_control?: { type: 'ephemeral'; ttl?: '1h' };\n    };\n\n    const optionsWithDefaults = this._combineCallOptions(options);\n    const useResponsesApi = this._useResponsesApi(options);\n    const params = (\n      useResponsesApi\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    // Consume the raw (un-deduped) stream so `finish_reason` survives as the\n    // reasoning-flush signal below; de-duplicate after that, before emitting.\n    const seenScalarMetadata: SeenScalarMetadata = new Map();\n\n    for await (const generationChunk of super._streamRawResponseChunks(\n      messages,\n      options,\n      undefined\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      } else {\n        delete generationChunk.message.additional_kwargs.reasoning_details;\n      }\n\n      dropRepeatedScalarMetadata(generationChunk, seenScalarMetadata);\n      await emitStreamChunkCallback(generationChunk, runManager);\n      yield generationChunk;\n    }\n  }\n}\n"],"mappings":";;;AAoFA,SAAS,sBACP,OACwC;CACxC,OACE,OAAO,UAAU,YACjB,UAAU,QACV,UAAU,SACV,MAAM,SAAS;AAEnB;AAEA,SAAS,2BACP,OAC6C;CAC7C,OACE,OAAO,UAAU,YACjB,UAAU,QACV,UAAU,SACV,MAAM,SAAS;AAEnB;AAEA,SAAS,oBAAoB,OAA6C;CACxE,IAAI,CAAC,MAAM,QAAQ,KAAK,GACtB,OAAO,CAAC;CAEV,OAAO,MAAM,QACV,WACC,sBAAsB,MAAM,KAAK,2BAA2B,MAAM,CACtE;AACF;AAEA,IAAa,iBAAb,cAAoC,WAAW;CAC7C;;CAEA;CAEA,YAAY,SAA8B;EACxC,MAAM,2BAAgD,EAAE,GAAG,QAAQ;EACnE,OAAO,yBAAyB;EAChC,OAAO,yBAAyB;EAEhC,MAAM,EACJ,mBACA,WAAW,qBACX,cAAc,CAAC,GACf,GAAG,WACD;EAGJ,MAAM,EAAE,WAAW,aAAa,GAAG,oBAAoB;EAGvD,MAAM,kBACJ,eAAe,QAAQ,uBAAuB,OAC1C;GACA,GAAG;GACH,GAAG;EACL,IACE,KAAA;EACN,MAAM,mBACJ,oBACC,sBAAsB,OAAO,EAAE,SAAS,KAAK,IAAI,KAAA;EACpD,MAAM,oBACJ,oBAAoB,OAChB,kBACA;GAAE,GAAG;GAAiB,WAAW;EAAiB;EAExD,MAAM;GACJ,GAAG;GACH,aAAa;GACb,yBAAyB;GACzB,kCAAkC;GAClC,0BAA0B;GAC1B,sBAAsB,QAAQ;GAC9B,yBAAyB,QAAQ;EACnC,CAAC;EAGD,IAAI,mBAAmB,MACrB,KAAK,sBAAsB;EAG7B,KAAK,mBAAmB;CAC1B;CACA,OAAO,UAAiC;EACtC,OAAO;CACT;CAGA,iBACE,SACA,OAC4B;EAU5B,MAAM,sBAAsB,KAAK,oBAAoB,OAAO;EAE5D,MAAM,SADkB,KAAK,iBAAiB,OAE9B,IACV,KAAK,UAAU,iBAAiB,mBAAmB,IACnD,KAAK,YAAY,iBAAiB,qBAAqB,KAAK;EAKlE,OAAO,OAAO;EAGd,MAAM,YAAY,KAAK,yBAAyB,mBAAmB;EACnE,IAAI,aAAa,MACf,OAAO,YAAY;OAEnB,OAAO,OAAO;EAGhB,OAAO;CACT;CAEA,yBACE,SACiC;EACjC,IAAI;EAGJ,IAAI,KAAK,uBAAuB,MAC9B,YAAY,EAAE,GAAG,KAAK,oBAAoB;EAI5C,MAAM,cAAc,KAAK,mBAAmB,OAAO;EACnD,IAAI,aAAa,UAAU,MACzB,YAAY;GACV,GAAG;GACH,QAAQ,YAAY;EACtB;EAIF,MAAM,gBAAiB,SACnB;EACJ,IAAI,iBAAiB,MACnB,YAAY;GAAE,GAAG;GAAW,GAAG;EAAc;EAI/C,IAAI,aAAa,QAAQ,KAAK,qBAAqB,MACjD,YAAY,EAAE,SAAS,KAAK;EAG9B,OAAO;CACT;CAEA,OAAgB,sBACd,UACA,SACA,YACqC;EACrC,MAAM,uCAAuB,IAAI,IAG/B;EACF,MAAM,yCAAyB,IAAI,IAGjC;EAGF,MAAM,qCAAyC,IAAI,IAAI;EAEvD,WAAW,MAAM,mBAAmB,MAAM,yBACxC,UACA,SACA,KAAA,CACF,GAAG;GACD,IAAI,uBAAuB;GAC3B,MAAM,mBAAmB,oBACvB,gBAAgB,QAAQ,kBAAkB,iBAC5C;GAEA,KAAK,MAAM,UAAU,kBAAkB;IACrC,IAAI,OAAO,SAAS,kBAAkB;KACpC,wBAAwB,OAAO,QAAQ;KACvC,MAAM,QAAQ,OAAO,SAAS;KAC9B,MAAM,WAAW,qBAAqB,IAAI,KAAK;KAC/C,IAAI,YAAY,MAAM;MACpB,SAAS,OAAO,GAAG,SAAS,QAAQ,KAAK,OAAO,QAAQ;MACxD;KACF;KACA,qBAAqB,IAAI,OAAO;MAC9B,GAAG;MACH,MAAM,OAAO,QAAQ;KACvB,CAAC;KACD;IACF;IACA,IAAI,OAAO,MAAM,MACf,uBAAuB,IAAI,OAAO,IAAI,EAAE,GAAG,OAAO,CAAC;GAEvD;GAEA,IACE,qBAAqB,SAAS,KAC9B,gBAAgB,QAAQ,kBAAkB,aAAa,MAEvD,gBAAgB,QAAQ,kBAAkB,YACxC;GAGJ,IAAI,gBAAgB,gBAAgB,iBAAiB,MAAM;IACzD,MAAM,wBAAwB,CAC5B,GAAG,qBAAqB,OAAO,GAC/B,GAAG,uBAAuB,OAAO,CACnC;IACA,IAAI,sBAAsB,SAAS,GACjC,gBAAgB,QAAQ,kBAAkB,oBACxC;SAEF,OAAO,gBAAgB,QAAQ,kBAAkB;GAErD,OACE,OAAO,gBAAgB,QAAQ,kBAAkB;GAGnD,2BAA2B,iBAAiB,kBAAkB;GAC9D,MAAM,wBAAwB,iBAAiB,UAAU;GACzD,MAAM;EACR;CACF;AACF"}

package/dist/esm/llm/providers.mjs

@@ -1,12 +1,12 @@
 import "../common/enum.mjs";
 import "../common/index.mjs";
+import { CustomChatMistralAI } from "./mistral/index.mjs";
 import { AzureChatOpenAI, ChatDeepSeek, ChatMoonshot, ChatOpenAI, ChatXAI } from "./openai/index.mjs";
 import { CustomChatGoogleGenerativeAI } from "./google/index.mjs";
 import { CustomChatBedrockConverse } from "./bedrock/index.mjs";
 import { CustomAnthropic } from "./anthropic/index.mjs";
 import { ChatOpenRouter } from "./openrouter/index.mjs";
 import { ChatVertexAI } from "./vertexai/index.mjs";
-import { ChatMistralAI } from "@langchain/mistralai";
 //#region src/llm/providers.ts
 const llmProviders = {
 	["xai"]: ChatXAI,
@@ -14,8 +14,8 @@ const llmProviders = {
 	["azureOpenAI"]: AzureChatOpenAI,
 	["vertexai"]: ChatVertexAI,
 	["deepseek"]: ChatDeepSeek,
-	["mistralai"]: ChatMistralAI,
-	["mistral"]: ChatMistralAI,
+	["mistralai"]: CustomChatMistralAI,
+	["mistral"]: CustomChatMistralAI,
 	["anthropic"]: CustomAnthropic,
 	["openrouter"]: ChatOpenRouter,
 	["bedrock"]: CustomChatBedrockConverse,

package/dist/esm/llm/providers.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"providers.mjs","names":[],"sources":["../../../src/llm/providers.ts"],"sourcesContent":["// src/llm/providers.ts\nimport { ChatMistralAI } from '@langchain/mistralai';\nimport type {\n  ChatModelConstructorMap,\n  ProviderOptionsMap,\n  ChatModelMap,\n} from '@/types';\nimport {\n  AzureChatOpenAI,\n  ChatDeepSeek,\n  ChatMoonshot,\n  ChatOpenAI,\n  ChatXAI,\n} from '@/llm/openai';\nimport { CustomChatGoogleGenerativeAI } from '@/llm/google';\nimport { CustomChatBedrockConverse } from '@/llm/bedrock';\nimport { CustomAnthropic } from '@/llm/anthropic';\nimport { ChatOpenRouter } from '@/llm/openrouter';\nimport { ChatVertexAI } from '@/llm/vertexai';\nimport { Providers } from '@/common';\n\nexport const llmProviders: Partial<ChatModelConstructorMap> = {\n  [Providers.XAI]: ChatXAI,\n  [Providers.OPENAI]: ChatOpenAI,\n  [Providers.AZURE]: AzureChatOpenAI,\n  [Providers.VERTEXAI]: ChatVertexAI,\n  [Providers.DEEPSEEK]: ChatDeepSeek,\n  [Providers.MISTRALAI]: ChatMistralAI,\n  [Providers.MISTRAL]: ChatMistralAI,\n  [Providers.ANTHROPIC]: CustomAnthropic,\n  [Providers.OPENROUTER]: ChatOpenRouter,\n  [Providers.BEDROCK]: CustomChatBedrockConverse,\n  // [Providers.ANTHROPIC]: ChatAnthropic,\n  [Providers.GOOGLE]: CustomChatGoogleGenerativeAI,\n  [Providers.MOONSHOT]: ChatMoonshot,\n};\n\nexport const manualToolStreamProviders = new Set<Providers | string>([\n  Providers.ANTHROPIC,\n  Providers.BEDROCK,\n]);\n\nexport const getChatModelClass = <P extends Providers>(\n  provider: P\n): new (config: ProviderOptionsMap[P]) => ChatModelMap[P] => {\n  const ChatModelClass = llmProviders[provider];\n  if (!ChatModelClass) {\n    throw new Error(`Unsupported LLM provider: ${provider}`);\n  }\n\n  return ChatModelClass;\n};\n"],"mappings":";;;;;;;;;;AAqBA,MAAa,eAAiD;UAC3C;aACG;kBACD;eACG;eACA;gBACC;cACF;gBACE;iBACC;cACH;aAED;eACE;AACxB;AAEA,MAAa,4BAA4B,IAAI,IAAwB,CAAA,aAAA,SAGrE,CAAC;AAED,MAAa,qBACX,aAC2D;CAC3D,MAAM,iBAAiB,aAAa;CACpC,IAAI,CAAC,gBACH,MAAM,IAAI,MAAM,6BAA6B,UAAU;CAGzD,OAAO;AACT"}
+{"version":3,"file":"providers.mjs","names":[],"sources":["../../../src/llm/providers.ts"],"sourcesContent":["// src/llm/providers.ts\nimport { CustomChatMistralAI } from '@/llm/mistral';\nimport type {\n  ChatModelConstructorMap,\n  ProviderOptionsMap,\n  ChatModelMap,\n} from '@/types';\nimport {\n  AzureChatOpenAI,\n  ChatDeepSeek,\n  ChatMoonshot,\n  ChatOpenAI,\n  ChatXAI,\n} from '@/llm/openai';\nimport { CustomChatGoogleGenerativeAI } from '@/llm/google';\nimport { CustomChatBedrockConverse } from '@/llm/bedrock';\nimport { CustomAnthropic } from '@/llm/anthropic';\nimport { ChatOpenRouter } from '@/llm/openrouter';\nimport { ChatVertexAI } from '@/llm/vertexai';\nimport { Providers } from '@/common';\n\nexport const llmProviders: Partial<ChatModelConstructorMap> = {\n  [Providers.XAI]: ChatXAI,\n  [Providers.OPENAI]: ChatOpenAI,\n  [Providers.AZURE]: AzureChatOpenAI,\n  [Providers.VERTEXAI]: ChatVertexAI,\n  [Providers.DEEPSEEK]: ChatDeepSeek,\n  [Providers.MISTRALAI]: CustomChatMistralAI,\n  [Providers.MISTRAL]: CustomChatMistralAI,\n  [Providers.ANTHROPIC]: CustomAnthropic,\n  [Providers.OPENROUTER]: ChatOpenRouter,\n  [Providers.BEDROCK]: CustomChatBedrockConverse,\n  // [Providers.ANTHROPIC]: ChatAnthropic,\n  [Providers.GOOGLE]: CustomChatGoogleGenerativeAI,\n  [Providers.MOONSHOT]: ChatMoonshot,\n};\n\nexport const manualToolStreamProviders = new Set<Providers | string>([\n  Providers.ANTHROPIC,\n  Providers.BEDROCK,\n]);\n\nexport const getChatModelClass = <P extends Providers>(\n  provider: P\n): new (config: ProviderOptionsMap[P]) => ChatModelMap[P] => {\n  const ChatModelClass = llmProviders[provider];\n  if (!ChatModelClass) {\n    throw new Error(`Unsupported LLM provider: ${provider}`);\n  }\n\n  return ChatModelClass;\n};\n"],"mappings":";;;;;;;;;;AAqBA,MAAa,eAAiD;UAC3C;aACG;kBACD;eACG;eACA;gBACC;cACF;gBACE;iBACC;cACH;aAED;eACE;AACxB;AAEA,MAAa,4BAA4B,IAAI,IAAwB,CAAA,aAAA,SAGrE,CAAC;AAED,MAAa,qBACX,aAC2D;CAC3D,MAAM,iBAAiB,aAAa;CACpC,IAAI,CAAC,gBACH,MAAM,IAAI,MAAM,6BAA6B,UAAU;CAGzD,OAAO;AACT"}

package/dist/esm/llm/stream/chunkAdapters.mjs

@@ -0,0 +1,195 @@
+import { smoothStream } from "./smoother.mjs";
+import { AIMessageChunk } from "@langchain/core/messages";
+import { ChatGenerationChunk } from "@langchain/core/outputs";
+//#region src/llm/stream/chunkAdapters.ts
+/**
+* Rebuilds a generation chunk carrying one piece of a split plain-text chunk.
+* Unsplit pieces return the original chunk untouched, so disabled smoothing is
+* byte-identical to no smoothing. Usage metadata, additional kwargs and
+* response metadata survive only on the first piece: the aggregator merges
+* dicts by concatenating string fields, so replicating them across pieces
+* would duplicate reasoning text and scalar metadata once per piece.
+*/
+function cloneGenerationChunkPiece(chunk, piece) {
+	if (piece.isFirst && piece.isLast) return chunk;
+	const message = chunk.message;
+	return new ChatGenerationChunk({
+		text: piece.text,
+		generationInfo: piece.isFirst ? chunk.generationInfo : void 0,
+		message: new AIMessageChunk(Object.assign({}, message, {
+			content: piece.text,
+			usage_metadata: piece.isFirst ? message.usage_metadata : void 0,
+			additional_kwargs: piece.isFirst ? message.additional_kwargs : {},
+			response_metadata: piece.isFirst ? message.response_metadata : {}
+		}))
+	});
+}
+const SPLITTABLE_TEXT_PART_KEYS = new Set([
+	"type",
+	"text",
+	"index"
+]);
+/**
+* Returns the sole plain text part of a single-element complex content array
+* (the shape google-common emits for text deltas), or null when the part
+* carries anything beyond `type`/`text`/`index` (thought signatures, media,
+* function calls) and must not be sliced. The part must carry a numeric
+* `index`: it is the aggregator's list-merge key, and without it each split
+* piece would append as a separate content part instead of merging back
+* into one — index-less parts pace whole instead.
+*/
+function getSplittableTextPart(message, chunkText) {
+	const content = message.content;
+	if (!Array.isArray(content) || content.length !== 1) return null;
+	const part = content[0];
+	if (part == null || typeof part !== "object") return null;
+	const record = part;
+	if (typeof record.text !== "string" || record.text !== chunkText) return null;
+	if (record.type != null && record.type !== "text") return null;
+	if (typeof record.index !== "number") return null;
+	if (!Object.keys(record).every((key) => SPLITTABLE_TEXT_PART_KEYS.has(key))) return null;
+	return record;
+}
+function clonePartGenerationChunkPiece(chunk, part, piece) {
+	if (piece.isFirst && piece.isLast) return chunk;
+	const message = chunk.message;
+	return new ChatGenerationChunk({
+		text: piece.text,
+		generationInfo: piece.isFirst ? chunk.generationInfo : void 0,
+		message: new AIMessageChunk(Object.assign({}, message, {
+			content: [Object.assign({}, part, { text: piece.text })],
+			usage_metadata: piece.isFirst ? message.usage_metadata : void 0,
+			additional_kwargs: piece.isFirst ? message.additional_kwargs : {},
+			response_metadata: piece.isFirst ? message.response_metadata : {}
+		}))
+	});
+}
+/**
+* Provider-agnostic classification of a `ChatGenerationChunk` for the
+* smoothing engine:
+* - splittable: plain visible text — string content equal to `chunk.text`, or
+*   a single plain text part (google-common's delta shape) — with no
+*   logprobs / finish_reason; sliced adaptively at the pacing cadence.
+* - atomic: any other text-bearing chunk (complex content arrays, logprobs,
+*   finish_reason, provider-specific reasoning payloads surfaced via
+*   `getAtomicText`) — paced as one piece, never split.
+* - passthrough: tool-call deltas, usage-only and metadata chunks — strict
+*   FIFO, zero delay.
+*/
+/**
+* google-common stamps `logprobs: { content: [] }` on every chunk, so only
+* logprobs that actually carry data may block splitting here. This is looser
+* than the OpenAI-family adapter in `llm/openai/index.ts`, which treats ANY
+* logprobs as atomic (its providers only attach logprobs when requested, and
+* the DeepSeek suite pins that contract) — do not unify the two predicates.
+*/
+function hasMeaningfulLogprobs(generationInfo) {
+	const logprobs = generationInfo?.logprobs;
+	if (logprobs == null) return false;
+	const content = logprobs.content;
+	if (Array.isArray(content)) return content.length > 0;
+	return true;
+}
+/** google-common reports the terminal reason camelCase (`finishReason`). */
+function hasFinishReason(generationInfo) {
+	return generationInfo?.finish_reason != null || generationInfo?.finishReason != null;
+}
+/**
+* Chunks that pair visible text with reasoning payloads in
+* `additional_kwargs` (Gemini thought summaries, reasoning_content deltas,
+* OpenRouter reasoning_details) must pace whole: split pieces would each
+* carry the same kwargs and downstream merging — the aggregator's dict merge
+* or OpenRouter's reasoning_details accumulation — duplicates them once per
+* piece.
+*/
+/**
+* Mixed text/tool-call deltas must never split: cloned pieces would each
+* carry the same tool_call_chunks and downstream accumulation would corrupt
+* the assembled tool arguments.
+*/
+function hasToolCallChunks(message) {
+	return (message.tool_call_chunks?.length ?? 0) > 0;
+}
+function hasReasoningKwargs(message) {
+	const kwargs = message.additional_kwargs;
+	if (typeof kwargs.reasoning_content === "string" && kwargs.reasoning_content !== "") return true;
+	if (Array.isArray(kwargs.reasoning_details) && kwargs.reasoning_details.length > 0) return true;
+	return kwargs.reasoning != null;
+}
+/**
+* Extracts the visible reasoning text a kwargs-borne delta contributes, so
+* reasoning-only chunks (Gemini thoughts, DeepSeek reasoning_content,
+* OpenAI reasoning summaries, OpenRouter reasoning_details) pace atomically
+* at the cadence instead of passing through unsmoothed.
+*/
+function getReasoningKwargsText(message) {
+	const kwargs = message.additional_kwargs;
+	if (typeof kwargs.reasoning_content === "string" && kwargs.reasoning_content !== "") return kwargs.reasoning_content;
+	const reasoning = kwargs.reasoning;
+	if (typeof reasoning === "string" && reasoning !== "") return reasoning;
+	if (reasoning != null && typeof reasoning === "object") {
+		const summaryText = reasoning.summary?.[0]?.text;
+		if (typeof summaryText === "string" && summaryText !== "") return summaryText;
+	}
+	const details = kwargs.reasoning_details;
+	if (Array.isArray(details)) {
+		let text = "";
+		for (const detail of details) if (detail != null && typeof detail === "object") {
+			const detailText = detail.text;
+			if (typeof detailText === "string") text += detailText;
+		}
+		if (text !== "") return text;
+	}
+	return "";
+}
+function toGenerationSmoothItem(chunk, getAtomicText) {
+	const { message } = chunk;
+	const isMessageChunk = message instanceof AIMessageChunk;
+	const cleanGenerationInfo = !hasMeaningfulLogprobs(chunk.generationInfo) && !hasFinishReason(chunk.generationInfo);
+	if (Boolean(chunk.text) && isMessageChunk && typeof message.content === "string" && message.content === chunk.text && cleanGenerationInfo && !hasReasoningKwargs(message) && !hasToolCallChunks(message)) return {
+		text: chunk.text,
+		smooth: true,
+		emit: (piece) => cloneGenerationChunkPiece(chunk, piece)
+	};
+	const splittablePart = Boolean(chunk.text) && isMessageChunk && cleanGenerationInfo && !hasReasoningKwargs(message) && !hasToolCallChunks(message) ? getSplittableTextPart(message, chunk.text) : null;
+	if (splittablePart != null) return {
+		text: chunk.text,
+		smooth: true,
+		emit: (piece) => clonePartGenerationChunkPiece(chunk, splittablePart, piece)
+	};
+	const pacedText = chunk.text || (isMessageChunk && getAtomicText != null ? getAtomicText(message) : "");
+	if (pacedText !== "") return {
+		text: pacedText,
+		smooth: true,
+		atomic: true,
+		emit: () => chunk
+	};
+	return {
+		text: "",
+		smooth: false,
+		emit: () => chunk
+	};
+}
+/**
+* Wraps a provider's raw chunk stream with adaptive smoothing and per-piece
+* `handleLLMNewToken` dispatch. The raw stream must NOT dispatch runManager
+* callbacks itself — callback-echo consumers would otherwise observe the
+* unsmoothed deltas.
+*/
+async function* smoothGenerationChunks({ chunks, delayMs, signal, runManager }) {
+	const source = (async function* () {
+		for await (const chunk of chunks) yield toGenerationSmoothItem(chunk, getReasoningKwargsText);
+	})();
+	for await (const outputChunk of smoothStream({
+		source,
+		delayMs,
+		signal
+	})) {
+		yield outputChunk;
+		await runManager?.handleLLMNewToken(outputChunk.text || "", void 0, void 0, void 0, void 0, { chunk: outputChunk });
+	}
+}
+//#endregion
+export { getReasoningKwargsText, hasReasoningKwargs, hasToolCallChunks, smoothGenerationChunks };
+
+//# sourceMappingURL=chunkAdapters.mjs.map

package/dist/esm/llm/stream/chunkAdapters.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"chunkAdapters.mjs","names":[],"sources":["../../../../src/llm/stream/chunkAdapters.ts"],"sourcesContent":["import { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { SmoothItem, SmoothPiece } from '@/llm/stream/smoother';\nimport { smoothStream } from '@/llm/stream/smoother';\n\n/**\n * Rebuilds a generation chunk carrying one piece of a split plain-text chunk.\n * Unsplit pieces return the original chunk untouched, so disabled smoothing is\n * byte-identical to no smoothing. Usage metadata, additional kwargs and\n * response metadata survive only on the first piece: the aggregator merges\n * dicts by concatenating string fields, so replicating them across pieces\n * would duplicate reasoning text and scalar metadata once per piece.\n */\nexport function cloneGenerationChunkPiece(\n  chunk: ChatGenerationChunk,\n  piece: SmoothPiece\n): ChatGenerationChunk {\n  if (piece.isFirst && piece.isLast) {\n    return chunk;\n  }\n  const message = chunk.message as AIMessageChunk;\n  return new ChatGenerationChunk({\n    text: piece.text,\n    generationInfo: piece.isFirst ? chunk.generationInfo : undefined,\n    message: new AIMessageChunk(\n      Object.assign({}, message, {\n        content: piece.text,\n        usage_metadata: piece.isFirst ? message.usage_metadata : undefined,\n        additional_kwargs: piece.isFirst ? message.additional_kwargs : {},\n        response_metadata: piece.isFirst ? message.response_metadata : {},\n      })\n    ),\n  });\n}\n\nconst SPLITTABLE_TEXT_PART_KEYS: ReadonlySet<string> = new Set([\n  'type',\n  'text',\n  'index',\n]);\n\n/**\n * Returns the sole plain text part of a single-element complex content array\n * (the shape google-common emits for text deltas), or null when the part\n * carries anything beyond `type`/`text`/`index` (thought signatures, media,\n * function calls) and must not be sliced. The part must carry a numeric\n * `index`: it is the aggregator's list-merge key, and without it each split\n * piece would append as a separate content part instead of merging back\n * into one — index-less parts pace whole instead.\n */\nfunction getSplittableTextPart(\n  message: AIMessageChunk,\n  chunkText: string\n): Record<string, unknown> | null {\n  const content = message.content;\n  if (!Array.isArray(content) || content.length !== 1) {\n    return null;\n  }\n  const part: unknown = content[0];\n  if (part == null || typeof part !== 'object') {\n    return null;\n  }\n  const record = part as Record<string, unknown>;\n  if (typeof record.text !== 'string' || record.text !== chunkText) {\n    return null;\n  }\n  if (record.type != null && record.type !== 'text') {\n    return null;\n  }\n  if (typeof record.index !== 'number') {\n    return null;\n  }\n  if (!Object.keys(record).every((key) => SPLITTABLE_TEXT_PART_KEYS.has(key))) {\n    return null;\n  }\n  return record;\n}\n\nfunction clonePartGenerationChunkPiece(\n  chunk: ChatGenerationChunk,\n  part: Record<string, unknown>,\n  piece: SmoothPiece\n): ChatGenerationChunk {\n  if (piece.isFirst && piece.isLast) {\n    return chunk;\n  }\n  const message = chunk.message as AIMessageChunk;\n  return new ChatGenerationChunk({\n    text: piece.text,\n    generationInfo: piece.isFirst ? chunk.generationInfo : undefined,\n    message: new AIMessageChunk(\n      Object.assign({}, message, {\n        content: [Object.assign({}, part, { text: piece.text })],\n        usage_metadata: piece.isFirst ? message.usage_metadata : undefined,\n        additional_kwargs: piece.isFirst ? message.additional_kwargs : {},\n        response_metadata: piece.isFirst ? message.response_metadata : {},\n      })\n    ),\n  });\n}\n\n/**\n * Provider-agnostic classification of a `ChatGenerationChunk` for the\n * smoothing engine:\n * - splittable: plain visible text — string content equal to `chunk.text`, or\n *   a single plain text part (google-common's delta shape) — with no\n *   logprobs / finish_reason; sliced adaptively at the pacing cadence.\n * - atomic: any other text-bearing chunk (complex content arrays, logprobs,\n *   finish_reason, provider-specific reasoning payloads surfaced via\n *   `getAtomicText`) — paced as one piece, never split.\n * - passthrough: tool-call deltas, usage-only and metadata chunks — strict\n *   FIFO, zero delay.\n */\n/**\n * google-common stamps `logprobs: { content: [] }` on every chunk, so only\n * logprobs that actually carry data may block splitting here. This is looser\n * than the OpenAI-family adapter in `llm/openai/index.ts`, which treats ANY\n * logprobs as atomic (its providers only attach logprobs when requested, and\n * the DeepSeek suite pins that contract) — do not unify the two predicates.\n */\nfunction hasMeaningfulLogprobs(\n  generationInfo: ChatGenerationChunk['generationInfo']\n): boolean {\n  const logprobs = generationInfo?.logprobs;\n  if (logprobs == null) {\n    return false;\n  }\n  const content = (logprobs as { content?: unknown }).content;\n  if (Array.isArray(content)) {\n    return content.length > 0;\n  }\n  return true;\n}\n\n/** google-common reports the terminal reason camelCase (`finishReason`). */\nfunction hasFinishReason(\n  generationInfo: ChatGenerationChunk['generationInfo']\n): boolean {\n  return (\n    generationInfo?.finish_reason != null || generationInfo?.finishReason != null\n  );\n}\n\n/**\n * Chunks that pair visible text with reasoning payloads in\n * `additional_kwargs` (Gemini thought summaries, reasoning_content deltas,\n * OpenRouter reasoning_details) must pace whole: split pieces would each\n * carry the same kwargs and downstream merging — the aggregator's dict merge\n * or OpenRouter's reasoning_details accumulation — duplicates them once per\n * piece.\n */\n/**\n * Mixed text/tool-call deltas must never split: cloned pieces would each\n * carry the same tool_call_chunks and downstream accumulation would corrupt\n * the assembled tool arguments.\n */\nexport function hasToolCallChunks(message: AIMessageChunk): boolean {\n  return (message.tool_call_chunks?.length ?? 0) > 0;\n}\n\nexport function hasReasoningKwargs(message: AIMessageChunk): boolean {\n  const kwargs = message.additional_kwargs;\n  if (\n    typeof kwargs.reasoning_content === 'string' &&\n    kwargs.reasoning_content !== ''\n  ) {\n    return true;\n  }\n  if (\n    Array.isArray(kwargs.reasoning_details) &&\n    kwargs.reasoning_details.length > 0\n  ) {\n    return true;\n  }\n  return kwargs.reasoning != null;\n}\n\n/**\n * Extracts the visible reasoning text a kwargs-borne delta contributes, so\n * reasoning-only chunks (Gemini thoughts, DeepSeek reasoning_content,\n * OpenAI reasoning summaries, OpenRouter reasoning_details) pace atomically\n * at the cadence instead of passing through unsmoothed.\n */\nexport function getReasoningKwargsText(message: AIMessageChunk): string {\n  const kwargs = message.additional_kwargs;\n  if (\n    typeof kwargs.reasoning_content === 'string' &&\n    kwargs.reasoning_content !== ''\n  ) {\n    return kwargs.reasoning_content;\n  }\n  const reasoning = kwargs.reasoning;\n  if (typeof reasoning === 'string' && reasoning !== '') {\n    return reasoning;\n  }\n  if (reasoning != null && typeof reasoning === 'object') {\n    const summaryText = (\n      reasoning as { summary?: { text?: unknown }[] }\n    ).summary?.[0]?.text;\n    if (typeof summaryText === 'string' && summaryText !== '') {\n      return summaryText;\n    }\n  }\n  const details = kwargs.reasoning_details;\n  if (Array.isArray(details)) {\n    let text = '';\n    for (const detail of details) {\n      if (detail != null && typeof detail === 'object') {\n        const detailText = (detail as { text?: unknown }).text;\n        if (typeof detailText === 'string') {\n          text += detailText;\n        }\n      }\n    }\n    if (text !== '') {\n      return text;\n    }\n  }\n  return '';\n}\n\nexport function toGenerationSmoothItem(\n  chunk: ChatGenerationChunk,\n  getAtomicText?: (message: AIMessageChunk) => string\n): SmoothItem<ChatGenerationChunk> {\n  const { message } = chunk;\n  const isMessageChunk = message instanceof AIMessageChunk;\n  const cleanGenerationInfo =\n    !hasMeaningfulLogprobs(chunk.generationInfo) &&\n    !hasFinishReason(chunk.generationInfo);\n  const splittable =\n    Boolean(chunk.text) &&\n    isMessageChunk &&\n    typeof message.content === 'string' &&\n    message.content === chunk.text &&\n    cleanGenerationInfo &&\n    !hasReasoningKwargs(message) &&\n    !hasToolCallChunks(message);\n\n  if (splittable) {\n    return {\n      text: chunk.text,\n      smooth: true,\n      emit: (piece) => cloneGenerationChunkPiece(chunk, piece),\n    };\n  }\n\n  const splittablePart =\n    Boolean(chunk.text) &&\n    isMessageChunk &&\n    cleanGenerationInfo &&\n    !hasReasoningKwargs(message) &&\n    !hasToolCallChunks(message)\n      ? getSplittableTextPart(message, chunk.text)\n      : null;\n  if (splittablePart != null) {\n    return {\n      text: chunk.text,\n      smooth: true,\n      emit: (piece) =>\n        clonePartGenerationChunkPiece(chunk, splittablePart, piece),\n    };\n  }\n\n  const pacedText =\n    chunk.text ||\n    (isMessageChunk && getAtomicText != null ? getAtomicText(message) : '');\n  if (pacedText !== '') {\n    return {\n      text: pacedText,\n      smooth: true,\n      atomic: true,\n      emit: () => chunk,\n    };\n  }\n\n  return { text: '', smooth: false, emit: () => chunk };\n}\n\n/**\n * Wraps a provider's raw chunk stream with adaptive smoothing and per-piece\n * `handleLLMNewToken` dispatch. The raw stream must NOT dispatch runManager\n * callbacks itself — callback-echo consumers would otherwise observe the\n * unsmoothed deltas.\n */\nexport async function* smoothGenerationChunks({\n  chunks,\n  delayMs,\n  signal,\n  runManager,\n}: {\n  chunks: AsyncGenerator<ChatGenerationChunk>;\n  delayMs: number;\n  signal?: AbortSignal;\n  runManager?: CallbackManagerForLLMRun;\n}): AsyncGenerator<ChatGenerationChunk> {\n  const source = (async function* (): AsyncGenerator<\n    SmoothItem<ChatGenerationChunk>\n    > {\n    for await (const chunk of chunks) {\n      yield toGenerationSmoothItem(chunk, getReasoningKwargsText);\n    }\n  })();\n\n  for await (const outputChunk of smoothStream({ source, delayMs, signal })) {\n    yield outputChunk;\n    await runManager?.handleLLMNewToken(\n      outputChunk.text || '',\n      undefined,\n      undefined,\n      undefined,\n      undefined,\n      { chunk: outputChunk }\n    );\n  }\n}\n"],"mappings":";;;;;;;;;;;;AAcA,SAAgB,0BACd,OACA,OACqB;CACrB,IAAI,MAAM,WAAW,MAAM,QACzB,OAAO;CAET,MAAM,UAAU,MAAM;CACtB,OAAO,IAAI,oBAAoB;EAC7B,MAAM,MAAM;EACZ,gBAAgB,MAAM,UAAU,MAAM,iBAAiB,KAAA;EACvD,SAAS,IAAI,eACX,OAAO,OAAO,CAAC,GAAG,SAAS;GACzB,SAAS,MAAM;GACf,gBAAgB,MAAM,UAAU,QAAQ,iBAAiB,KAAA;GACzD,mBAAmB,MAAM,UAAU,QAAQ,oBAAoB,CAAC;GAChE,mBAAmB,MAAM,UAAU,QAAQ,oBAAoB,CAAC;EAClE,CAAC,CACH;CACF,CAAC;AACH;AAEA,MAAM,4BAAiD,IAAI,IAAI;CAC7D;CACA;CACA;AACF,CAAC;;;;;;;;;;AAWD,SAAS,sBACP,SACA,WACgC;CAChC,MAAM,UAAU,QAAQ;CACxB,IAAI,CAAC,MAAM,QAAQ,OAAO,KAAK,QAAQ,WAAW,GAChD,OAAO;CAET,MAAM,OAAgB,QAAQ;CAC9B,IAAI,QAAQ,QAAQ,OAAO,SAAS,UAClC,OAAO;CAET,MAAM,SAAS;CACf,IAAI,OAAO,OAAO,SAAS,YAAY,OAAO,SAAS,WACrD,OAAO;CAET,IAAI,OAAO,QAAQ,QAAQ,OAAO,SAAS,QACzC,OAAO;CAET,IAAI,OAAO,OAAO,UAAU,UAC1B,OAAO;CAET,IAAI,CAAC,OAAO,KAAK,MAAM,CAAC,CAAC,OAAO,QAAQ,0BAA0B,IAAI,GAAG,CAAC,GACxE,OAAO;CAET,OAAO;AACT;AAEA,SAAS,8BACP,OACA,MACA,OACqB;CACrB,IAAI,MAAM,WAAW,MAAM,QACzB,OAAO;CAET,MAAM,UAAU,MAAM;CACtB,OAAO,IAAI,oBAAoB;EAC7B,MAAM,MAAM;EACZ,gBAAgB,MAAM,UAAU,MAAM,iBAAiB,KAAA;EACvD,SAAS,IAAI,eACX,OAAO,OAAO,CAAC,GAAG,SAAS;GACzB,SAAS,CAAC,OAAO,OAAO,CAAC,GAAG,MAAM,EAAE,MAAM,MAAM,KAAK,CAAC,CAAC;GACvD,gBAAgB,MAAM,UAAU,QAAQ,iBAAiB,KAAA;GACzD,mBAAmB,MAAM,UAAU,QAAQ,oBAAoB,CAAC;GAChE,mBAAmB,MAAM,UAAU,QAAQ,oBAAoB,CAAC;EAClE,CAAC,CACH;CACF,CAAC;AACH;;;;;;;;;;;;;;;;;;;;AAqBA,SAAS,sBACP,gBACS;CACT,MAAM,WAAW,gBAAgB;CACjC,IAAI,YAAY,MACd,OAAO;CAET,MAAM,UAAW,SAAmC;CACpD,IAAI,MAAM,QAAQ,OAAO,GACvB,OAAO,QAAQ,SAAS;CAE1B,OAAO;AACT;;AAGA,SAAS,gBACP,gBACS;CACT,OACE,gBAAgB,iBAAiB,QAAQ,gBAAgB,gBAAgB;AAE7E;;;;;;;;;;;;;;AAeA,SAAgB,kBAAkB,SAAkC;CAClE,QAAQ,QAAQ,kBAAkB,UAAU,KAAK;AACnD;AAEA,SAAgB,mBAAmB,SAAkC;CACnE,MAAM,SAAS,QAAQ;CACvB,IACE,OAAO,OAAO,sBAAsB,YACpC,OAAO,sBAAsB,IAE7B,OAAO;CAET,IACE,MAAM,QAAQ,OAAO,iBAAiB,KACtC,OAAO,kBAAkB,SAAS,GAElC,OAAO;CAET,OAAO,OAAO,aAAa;AAC7B;;;;;;;AAQA,SAAgB,uBAAuB,SAAiC;CACtE,MAAM,SAAS,QAAQ;CACvB,IACE,OAAO,OAAO,sBAAsB,YACpC,OAAO,sBAAsB,IAE7B,OAAO,OAAO;CAEhB,MAAM,YAAY,OAAO;CACzB,IAAI,OAAO,cAAc,YAAY,cAAc,IACjD,OAAO;CAET,IAAI,aAAa,QAAQ,OAAO,cAAc,UAAU;EACtD,MAAM,cACJ,UACA,UAAU,EAAE,EAAE;EAChB,IAAI,OAAO,gBAAgB,YAAY,gBAAgB,IACrD,OAAO;CAEX;CACA,MAAM,UAAU,OAAO;CACvB,IAAI,MAAM,QAAQ,OAAO,GAAG;EAC1B,IAAI,OAAO;EACX,KAAK,MAAM,UAAU,SACnB,IAAI,UAAU,QAAQ,OAAO,WAAW,UAAU;GAChD,MAAM,aAAc,OAA8B;GAClD,IAAI,OAAO,eAAe,UACxB,QAAQ;EAEZ;EAEF,IAAI,SAAS,IACX,OAAO;CAEX;CACA,OAAO;AACT;AAEA,SAAgB,uBACd,OACA,eACiC;CACjC,MAAM,EAAE,YAAY;CACpB,MAAM,iBAAiB,mBAAmB;CAC1C,MAAM,sBACJ,CAAC,sBAAsB,MAAM,cAAc,KAC3C,CAAC,gBAAgB,MAAM,cAAc;CAUvC,IARE,QAAQ,MAAM,IAAI,KAClB,kBACA,OAAO,QAAQ,YAAY,YAC3B,QAAQ,YAAY,MAAM,QAC1B,uBACA,CAAC,mBAAmB,OAAO,KAC3B,CAAC,kBAAkB,OAAO,GAG1B,OAAO;EACL,MAAM,MAAM;EACZ,QAAQ;EACR,OAAO,UAAU,0BAA0B,OAAO,KAAK;CACzD;CAGF,MAAM,iBACJ,QAAQ,MAAM,IAAI,KAClB,kBACA,uBACA,CAAC,mBAAmB,OAAO,KAC3B,CAAC,kBAAkB,OAAO,IACtB,sBAAsB,SAAS,MAAM,IAAI,IACzC;CACN,IAAI,kBAAkB,MACpB,OAAO;EACL,MAAM,MAAM;EACZ,QAAQ;EACR,OAAO,UACL,8BAA8B,OAAO,gBAAgB,KAAK;CAC9D;CAGF,MAAM,YACJ,MAAM,SACL,kBAAkB,iBAAiB,OAAO,cAAc,OAAO,IAAI;CACtE,IAAI,cAAc,IAChB,OAAO;EACL,MAAM;EACN,QAAQ;EACR,QAAQ;EACR,YAAY;CACd;CAGF,OAAO;EAAE,MAAM;EAAI,QAAQ;EAAO,YAAY;CAAM;AACtD;;;;;;;AAQA,gBAAuB,uBAAuB,EAC5C,QACA,SACA,QACA,cAMsC;CACtC,MAAM,UAAU,mBAEZ;EACF,WAAW,MAAM,SAAS,QACxB,MAAM,uBAAuB,OAAO,sBAAsB;CAE9D,EAAA,CAAG;CAEH,WAAW,MAAM,eAAe,aAAa;EAAE;EAAQ;EAAS;CAAO,CAAC,GAAG;EACzE,MAAM;EACN,MAAM,YAAY,kBAChB,YAAY,QAAQ,IACpB,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,EAAE,OAAO,YAAY,CACvB;CACF;AACF"}