@livekit/agents-plugin-openai 1.2.1 → 1.2.3

package/dist/responses/llm.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/responses/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport type { APIConnectOptions } from '@livekit/agents';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  APITimeoutError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  llm,\n  log,\n  toError,\n} from '@livekit/agents';\nimport OpenAI from 'openai';\nimport type { ChatModels } from '../models.js';\nimport { WSLLM } from '../ws/llm.js';\n\nexport interface LLMOptions {\n  model: string | ChatModels;\n  apiKey?: string;\n  baseURL?: string;\n  client?: OpenAI;\n  temperature?: number;\n  parallelToolCalls?: boolean;\n  toolChoice?: llm.ToolChoice;\n  store?: boolean;\n  metadata?: Record<string, string>;\n  strictToolSchema?: boolean;\n\n  /**\n   * Whether to use the WebSocket API.\n   * @default true\n   */\n  useWebSocket?: boolean;\n}\n\ntype HttpLLMOptions = Omit<LLMOptions, 'useWebSocket'>;\n\nconst defaultLLMOptions: LLMOptions = {\n  model: 'gpt-4.1',\n  apiKey: process.env.OPENAI_API_KEY,\n  strictToolSchema: true,\n  useWebSocket: true,\n};\n\nclass ResponsesHttpLLM extends llm.LLM {\n  #client: OpenAI;\n  #opts: HttpLLMOptions;\n\n  constructor(opts: Partial<HttpLLMOptions> = defaultLLMOptions) {\n    super();\n\n    this.#opts = { ...defaultLLMOptions, ...opts };\n    if (this.#opts.apiKey === undefined && this.#opts.client === undefined) {\n      throw new Error('OpenAI API key is required, whether as an argument or as $OPENAI_API_KEY');\n    }\n\n    this.#client =\n      this.#opts.client ||\n      new OpenAI({\n        baseURL: this.#opts.baseURL,\n        apiKey: this.#opts.apiKey,\n      });\n  }\n\n  override label(): string {\n    return 'openai.responses.LLM';\n  }\n\n  override get model(): string {\n    return this.#opts.model;\n  }\n\n  override chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    extraKwargs?: Record<string, unknown>;\n  }): ResponsesHttpLLMStream {\n    const modelOptions: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined ? parallelToolCalls : this.#opts.parallelToolCalls;\n\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      modelOptions.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice =\n      toolChoice !== undefined ? toolChoice : (this.#opts.toolChoice as llm.ToolChoice | undefined);\n\n    if (toolChoice) {\n      modelOptions.tool_choice = toolChoice;\n    }\n\n    if (this.#opts.temperature !== undefined) {\n      modelOptions.temperature = this.#opts.temperature;\n    }\n\n    if (this.#opts.store !== undefined) {\n      modelOptions.store = this.#opts.store;\n    }\n\n    if (this.#opts.metadata) {\n      modelOptions.metadata = this.#opts.metadata;\n    }\n\n    return new ResponsesHttpLLMStream(this, {\n      model: this.#opts.model,\n      client: this.#client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema: this.#opts.strictToolSchema ?? true,\n    });\n  }\n}\n\nclass ResponsesHttpLLMStream extends llm.LLMStream {\n  private model: string | ChatModels;\n  private client: OpenAI;\n  private modelOptions: Record<string, unknown>;\n  private strictToolSchema: boolean;\n  private responseId: string;\n\n  constructor(\n    llm: ResponsesHttpLLM,\n    {\n      model,\n      client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema,\n    }: {\n      model: string | ChatModels;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      connOptions: APIConnectOptions;\n      modelOptions: Record<string, unknown>;\n      strictToolSchema: boolean;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.model = model;\n    this.client = client;\n    this.modelOptions = modelOptions;\n    this.strictToolSchema = strictToolSchema;\n    this.responseId = '';\n  }\n\n  protected async run(): Promise<void> {\n    let retryable = true;\n\n    try {\n      const messages = (await this.chatCtx.toProviderFormat(\n        'openai.responses',\n      )) as OpenAI.Responses.ResponseInputItem[];\n\n      const tools = this.toolCtx\n        ? Object.entries(this.toolCtx).map(([name, func]) => {\n            const oaiParams = {\n              type: 'function' as const,\n              name: name,\n              description: func.description,\n              parameters: llm.toJsonSchema(\n                func.parameters,\n                true,\n                this.strictToolSchema,\n              ) as unknown as OpenAI.Responses.FunctionTool['parameters'],\n            } as OpenAI.Responses.FunctionTool;\n\n            if (this.strictToolSchema) {\n              oaiParams.strict = true;\n            }\n\n            return oaiParams;\n          })\n        : undefined;\n\n      const requestOptions: Record<string, unknown> = { ...this.modelOptions };\n      if (!tools) {\n        delete requestOptions.tool_choice;\n      }\n\n      const stream = await this.client.responses.create(\n        {\n          model: this.model,\n          input: messages,\n          tools: tools,\n          stream: true,\n          ...requestOptions,\n        },\n        {\n          timeout: this.connOptions.timeoutMs,\n        },\n      );\n\n      for await (const event of stream) {\n        retryable = false;\n        let chunk: llm.ChatChunk | undefined;\n\n        switch (event.type) {\n          case 'error':\n            this.handleError(event);\n            break;\n          case 'response.created':\n            this.handleResponseCreated(event);\n            break;\n          case 'response.output_item.done':\n            chunk = this.handleResponseOutputItemDone(event);\n            break;\n          case 'response.output_text.delta':\n            chunk = this.handleResponseOutputTextDelta(event);\n            break;\n          case 'response.completed':\n            chunk = this.handleResponseCompleted(event);\n            break;\n        }\n\n        if (chunk) {\n          this.queue.put(chunk);\n        }\n      }\n    } catch (error) {\n      if (\n        error instanceof APIStatusError ||\n        error instanceof APITimeoutError ||\n        error instanceof APIConnectionError\n      ) {\n        throw error;\n      } else if (error instanceof OpenAI.APIConnectionTimeoutError) {\n        throw new APITimeoutError({ options: { retryable } });\n      } else if (error instanceof OpenAI.APIError) {\n        throw new APIStatusError({\n          message: error.message,\n          options: {\n            statusCode: error.status,\n            body: error.error,\n            requestId: error.requestID,\n            retryable,\n          },\n        });\n      } else {\n        throw new APIConnectionError({\n          message: toError(error).message,\n          options: { retryable },\n        });\n      }\n    }\n  }\n\n  private handleError(event: OpenAI.Responses.ResponseErrorEvent): void {\n    throw new APIStatusError({\n      message: event.message,\n      options: {\n        statusCode: -1,\n        retryable: false,\n      },\n    });\n  }\n\n  private handleResponseCreated(event: OpenAI.Responses.ResponseCreatedEvent): void {\n    this.responseId = event.response.id;\n  }\n\n  private handleResponseOutputItemDone(\n    event: OpenAI.Responses.ResponseOutputItemDoneEvent,\n  ): llm.ChatChunk | undefined {\n    let chunk: llm.ChatChunk | undefined;\n\n    if (event.item.type === 'function_call') {\n      chunk = {\n        id: this.responseId,\n        delta: {\n          role: 'assistant',\n          content: undefined,\n          toolCalls: [\n            llm.FunctionCall.create({\n              callId: event.item.call_id || '',\n              name: event.item.name,\n              args: event.item.arguments,\n            }),\n          ],\n        },\n      };\n    }\n    return chunk;\n  }\n\n  private handleResponseOutputTextDelta(\n    event: OpenAI.Responses.ResponseTextDeltaEvent,\n  ): llm.ChatChunk {\n    return {\n      id: this.responseId,\n      delta: {\n        role: 'assistant',\n        content: event.delta,\n      },\n    };\n  }\n\n  private handleResponseCompleted(\n    event: OpenAI.Responses.ResponseCompletedEvent,\n  ): llm.ChatChunk | undefined {\n    if (event.response.usage) {\n      return {\n        id: this.responseId,\n        usage: {\n          completionTokens: event.response.usage.output_tokens,\n          promptTokens: event.response.usage.input_tokens,\n          promptCachedTokens: event.response.usage.input_tokens_details.cached_tokens,\n          totalTokens: event.response.usage.total_tokens,\n        },\n      };\n    }\n    return undefined;\n  }\n}\n\nexport class LLM extends llm.LLM {\n  #opts: LLMOptions;\n  #llm: llm.LLM;\n  #logger = log();\n\n  /**\n   * Create a new instance of OpenAI Responses LLM.\n   *\n   * @remarks\n   * `apiKey` must be set to your OpenAI API key, either using the argument or by setting the\n   * `OPENAI_API_KEY` environment variable.\n   */\n  constructor(opts: Partial<LLMOptions> = defaultLLMOptions) {\n    super();\n\n    this.#opts = { ...defaultLLMOptions, ...opts };\n    const { useWebSocket, client, ...baseOpts } = this.#opts;\n\n    if (useWebSocket) {\n      if (client !== undefined) {\n        this.#logger.warn(\n          'WebSocket mode does not support custom client; provided client will be ignored',\n        );\n      }\n      this.#llm = new WSLLM(baseOpts);\n    } else {\n      this.#llm = new ResponsesHttpLLM({ ...baseOpts, client });\n    }\n\n    // Forward events from the inner delegate so consumers listening on this\n    // wrapper instance (e.g. AgentActivity) receive them.\n    this.#llm.on('metrics_collected', (metrics) => this.emit('metrics_collected', metrics));\n    this.#llm.on('error', (error) => this.emit('error', error));\n  }\n\n  override label(): string {\n    return this.#llm.label();\n  }\n\n  override get model(): string {\n    return this.#llm.model;\n  }\n\n  override prewarm(): void {\n    this.#llm.prewarm();\n  }\n\n  // Ref: python livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/responses/llm.py - 229-233 lines\n  override async aclose(): Promise<void> {\n    await this.#llm.aclose();\n  }\n\n  async close(): Promise<void> {\n    await this.aclose();\n  }\n\n  override chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    extraKwargs?: Record<string, unknown>;\n  }): llm.LLMStream {\n    return this.#llm.chat({\n      chatCtx,\n      toolCtx,\n      connOptions,\n      parallelToolCalls,\n      toolChoice,\n      extraKwargs,\n    });\n  }\n}\n"],"mappings":"AAIA;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AACP,OAAO,YAAY;AAEnB,SAAS,aAAa;AAuBtB,MAAM,oBAAgC;AAAA,EACpC,OAAO;AAAA,EACP,QAAQ,QAAQ,IAAI;AAAA,EACpB,kBAAkB;AAAA,EAClB,cAAc;AAChB;AAEA,MAAM,yBAAyB,IAAI,IAAI;AAAA,EACrC;AAAA,EACA;AAAA,EAEA,YAAY,OAAgC,mBAAmB;AAC7D,UAAM;AAEN,SAAK,QAAQ,EAAE,GAAG,mBAAmB,GAAG,KAAK;AAC7C,QAAI,KAAK,MAAM,WAAW,UAAa,KAAK,MAAM,WAAW,QAAW;AACtE,YAAM,IAAI,MAAM,0EAA0E;AAAA,IAC5F;AAEA,SAAK,UACH,KAAK,MAAM,UACX,IAAI,OAAO;AAAA,MACT,SAAS,KAAK,MAAM;AAAA,MACpB,QAAQ,KAAK,MAAM;AAAA,IACrB,CAAC;AAAA,EACL;AAAA,EAES,QAAgB;AACvB,WAAO;AAAA,EACT;AAAA,EAEA,IAAa,QAAgB;AAC3B,WAAO,KAAK,MAAM;AAAA,EACpB;AAAA,EAES,KAAK;AAAA,IACZ;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAO2B;AACzB,UAAM,eAAwC,EAAE,GAAI,eAAe,CAAC,EAAG;AAEvE,wBACE,sBAAsB,SAAY,oBAAoB,KAAK,MAAM;AAEnE,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,mBAAa,sBAAsB;AAAA,IACrC;AAEA,iBACE,eAAe,SAAY,aAAc,KAAK,MAAM;AAEtD,QAAI,YAAY;AACd,mBAAa,cAAc;AAAA,IAC7B;AAEA,QAAI,KAAK,MAAM,gBAAgB,QAAW;AACxC,mBAAa,cAAc,KAAK,MAAM;AAAA,IACxC;AAEA,QAAI,KAAK,MAAM,UAAU,QAAW;AAClC,mBAAa,QAAQ,KAAK,MAAM;AAAA,IAClC;AAEA,QAAI,KAAK,MAAM,UAAU;AACvB,mBAAa,WAAW,KAAK,MAAM;AAAA,IACrC;AAEA,WAAO,IAAI,uBAAuB,MAAM;AAAA,MACtC,OAAO,KAAK,MAAM;AAAA,MAClB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,KAAK,MAAM,oBAAoB;AAAA,IACnD,CAAC;AAAA,EACH;AACF;AAEA,MAAM,+BAA+B,IAAI,UAAU;AAAA,EACzC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YACEA,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GASA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,QAAQ;AACb,SAAK,SAAS;AACd,SAAK,eAAe;AACpB,SAAK,mBAAmB;AACxB,SAAK,aAAa;AAAA,EACpB;AAAA,EAEA,MAAgB,MAAqB;AACnC,QAAI,YAAY;AAEhB,QAAI;AACF,YAAM,WAAY,MAAM,KAAK,QAAQ;AAAA,QACnC;AAAA,MACF;AAEA,YAAM,QAAQ,KAAK,UACf,OAAO,QAAQ,KAAK,OAAO,EAAE,IAAI,CAAC,CAAC,MAAM,IAAI,MAAM;AACjD,cAAM,YAAY;AAAA,UAChB,MAAM;AAAA,UACN;AAAA,UACA,aAAa,KAAK;AAAA,UAClB,YAAY,IAAI;AAAA,YACd,KAAK;AAAA,YACL;AAAA,YACA,KAAK;AAAA,UACP;AAAA,QACF;AAEA,YAAI,KAAK,kBAAkB;AACzB,oBAAU,SAAS;AAAA,QACrB;AAEA,eAAO;AAAA,MACT,CAAC,IACD;AAEJ,YAAM,iBAA0C,EAAE,GAAG,KAAK,aAAa;AACvE,UAAI,CAAC,OAAO;AACV,eAAO,eAAe;AAAA,MACxB;AAEA,YAAM,SAAS,MAAM,KAAK,OAAO,UAAU;AAAA,QACzC;AAAA,UACE,OAAO,KAAK;AAAA,UACZ,OAAO;AAAA,UACP;AAAA,UACA,QAAQ;AAAA,UACR,GAAG;AAAA,QACL;AAAA,QACA;AAAA,UACE,SAAS,KAAK,YAAY;AAAA,QAC5B;AAAA,MACF;AAEA,uBAAiB,SAAS,QAAQ;AAChC,oBAAY;AACZ,YAAI;AAEJ,gBAAQ,MAAM,MAAM;AAAA,UAClB,KAAK;AACH,iBAAK,YAAY,KAAK;AACtB;AAAA,UACF,KAAK;AACH,iBAAK,sBAAsB,KAAK;AAChC;AAAA,UACF,KAAK;AACH,oBAAQ,KAAK,6BAA6B,KAAK;AAC/C;AAAA,UACF,KAAK;AACH,oBAAQ,KAAK,8BAA8B,KAAK;AAChD;AAAA,UACF,KAAK;AACH,oBAAQ,KAAK,wBAAwB,KAAK;AAC1C;AAAA,QACJ;AAEA,YAAI,OAAO;AACT,eAAK,MAAM,IAAI,KAAK;AAAA,QACtB;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UACE,iBAAiB,kBACjB,iBAAiB,mBACjB,iBAAiB,oBACjB;AACA,cAAM;AAAA,MACR,WAAW,iBAAiB,OAAO,2BAA2B;AAC5D,cAAM,IAAI,gBAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,OAAO,UAAU;AAC3C,cAAM,IAAI,eAAe;AAAA,UACvB,SAAS,MAAM;AAAA,UACf,SAAS;AAAA,YACP,YAAY,MAAM;AAAA,YAClB,MAAM,MAAM;AAAA,YACZ,WAAW,MAAM;AAAA,YACjB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,OAAO;AACL,cAAM,IAAI,mBAAmB;AAAA,UAC3B,SAAS,QAAQ,KAAK,EAAE;AAAA,UACxB,SAAS,EAAE,UAAU;AAAA,QACvB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,YAAY,OAAkD;AACpE,UAAM,IAAI,eAAe;AAAA,MACvB,SAAS,MAAM;AAAA,MACf,SAAS;AAAA,QACP,YAAY;AAAA,QACZ,WAAW;AAAA,MACb;AAAA,IACF,CAAC;AAAA,EACH;AAAA,EAEQ,sBAAsB,OAAoD;AAChF,SAAK,aAAa,MAAM,SAAS;AAAA,EACnC;AAAA,EAEQ,6BACN,OAC2B;AAC3B,QAAI;AAEJ,QAAI,MAAM,KAAK,SAAS,iBAAiB;AACvC,cAAQ;AAAA,QACN,IAAI,KAAK;AAAA,QACT,OAAO;AAAA,UACL,MAAM;AAAA,UACN,SAAS;AAAA,UACT,WAAW;AAAA,YACT,IAAI,aAAa,OAAO;AAAA,cACtB,QAAQ,MAAM,KAAK,WAAW;AAAA,cAC9B,MAAM,MAAM,KAAK;AAAA,cACjB,MAAM,MAAM,KAAK;AAAA,YACnB,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA,EAEQ,8BACN,OACe;AACf,WAAO;AAAA,MACL,IAAI,KAAK;AAAA,MACT,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM;AAAA,MACjB;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,wBACN,OAC2B;AAC3B,QAAI,MAAM,SAAS,OAAO;AACxB,aAAO;AAAA,QACL,IAAI,KAAK;AAAA,QACT,OAAO;AAAA,UACL,kBAAkB,MAAM,SAAS,MAAM;AAAA,UACvC,cAAc,MAAM,SAAS,MAAM;AAAA,UACnC,oBAAoB,MAAM,SAAS,MAAM,qBAAqB;AAAA,UAC9D,aAAa,MAAM,SAAS,MAAM;AAAA,QACpC;AAAA,MACF;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACF;AAEO,MAAM,YAAY,IAAI,IAAI;AAAA,EAC/B;AAAA,EACA;AAAA,EACA,UAAU,IAAI;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASd,YAAY,OAA4B,mBAAmB;AACzD,UAAM;AAEN,SAAK,QAAQ,EAAE,GAAG,mBAAmB,GAAG,KAAK;AAC7C,UAAM,EAAE,cAAc,QAAQ,GAAG,SAAS,IAAI,KAAK;AAEnD,QAAI,cAAc;AAChB,UAAI,WAAW,QAAW;AACxB,aAAK,QAAQ;AAAA,UACX;AAAA,QACF;AAAA,MACF;AACA,WAAK,OAAO,IAAI,MAAM,QAAQ;AAAA,IAChC,OAAO;AACL,WAAK,OAAO,IAAI,iBAAiB,EAAE,GAAG,UAAU,OAAO,CAAC;AAAA,IAC1D;AAIA,SAAK,KAAK,GAAG,qBAAqB,CAAC,YAAY,KAAK,KAAK,qBAAqB,OAAO,CAAC;AACtF,SAAK,KAAK,GAAG,SAAS,CAAC,UAAU,KAAK,KAAK,SAAS,KAAK,CAAC;AAAA,EAC5D;AAAA,EAES,QAAgB;AACvB,WAAO,KAAK,KAAK,MAAM;AAAA,EACzB;AAAA,EAEA,IAAa,QAAgB;AAC3B,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAES,UAAgB;AACvB,SAAK,KAAK,QAAQ;AAAA,EACpB;AAAA;AAAA,EAGA,MAAe,SAAwB;AACrC,UAAM,KAAK,KAAK,OAAO;AAAA,EACzB;AAAA,EAEA,MAAM,QAAuB;AAC3B,UAAM,KAAK,OAAO;AAAA,EACpB;AAAA,EAES,KAAK;AAAA,IACZ;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAOkB;AAChB,WAAO,KAAK,KAAK,KAAK;AAAA,MACpB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AACF;","names":["llm"]}
+{"version":3,"sources":["../../src/responses/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport type { APIConnectOptions } from '@livekit/agents';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  APITimeoutError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  llm,\n  log,\n  toError,\n} from '@livekit/agents';\nimport OpenAI from 'openai';\nimport type { ChatModels } from '../models.js';\nimport { WSLLM } from '../ws/llm.js';\n\nexport interface LLMOptions {\n  model: string | ChatModels;\n  apiKey?: string;\n  baseURL?: string;\n  client?: OpenAI;\n  temperature?: number;\n  parallelToolCalls?: boolean;\n  toolChoice?: llm.ToolChoice;\n  store?: boolean;\n  metadata?: Record<string, string>;\n  strictToolSchema?: boolean;\n\n  /**\n   * Whether to use the WebSocket API.\n   * @default true\n   */\n  useWebSocket?: boolean;\n}\n\ntype HttpLLMOptions = Omit<LLMOptions, 'useWebSocket'>;\n\nconst defaultLLMOptions: LLMOptions = {\n  model: 'gpt-4.1',\n  apiKey: process.env.OPENAI_API_KEY,\n  strictToolSchema: true,\n  useWebSocket: true,\n};\n\nclass ResponsesHttpLLM extends llm.LLM {\n  #client: OpenAI;\n  #opts: HttpLLMOptions;\n\n  constructor(opts: Partial<HttpLLMOptions> = defaultLLMOptions) {\n    super();\n\n    this.#opts = { ...defaultLLMOptions, ...opts };\n    if (this.#opts.apiKey === undefined && this.#opts.client === undefined) {\n      throw new Error('OpenAI API key is required, whether as an argument or as $OPENAI_API_KEY');\n    }\n\n    this.#client =\n      this.#opts.client ||\n      new OpenAI({\n        baseURL: this.#opts.baseURL,\n        apiKey: this.#opts.apiKey,\n      });\n  }\n\n  override label(): string {\n    return 'openai.responses.LLM';\n  }\n\n  override get model(): string {\n    return this.#opts.model;\n  }\n\n  override chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    extraKwargs?: Record<string, unknown>;\n  }): ResponsesHttpLLMStream {\n    const modelOptions: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined ? parallelToolCalls : this.#opts.parallelToolCalls;\n\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      modelOptions.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice =\n      toolChoice !== undefined ? toolChoice : (this.#opts.toolChoice as llm.ToolChoice | undefined);\n\n    if (toolChoice) {\n      modelOptions.tool_choice = toolChoice;\n    }\n\n    if (this.#opts.temperature !== undefined) {\n      modelOptions.temperature = this.#opts.temperature;\n    }\n\n    if (this.#opts.store !== undefined) {\n      modelOptions.store = this.#opts.store;\n    }\n\n    if (this.#opts.metadata) {\n      modelOptions.metadata = this.#opts.metadata;\n    }\n\n    return new ResponsesHttpLLMStream(this, {\n      model: this.#opts.model,\n      client: this.#client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema: this.#opts.strictToolSchema ?? true,\n    });\n  }\n}\n\nclass ResponsesHttpLLMStream extends llm.LLMStream {\n  private model: string | ChatModels;\n  private client: OpenAI;\n  private modelOptions: Record<string, unknown>;\n  private strictToolSchema: boolean;\n  private responseId: string;\n\n  constructor(\n    llm: ResponsesHttpLLM,\n    {\n      model,\n      client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema,\n    }: {\n      model: string | ChatModels;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      connOptions: APIConnectOptions;\n      modelOptions: Record<string, unknown>;\n      strictToolSchema: boolean;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.model = model;\n    this.client = client;\n    this.modelOptions = modelOptions;\n    this.strictToolSchema = strictToolSchema;\n    this.responseId = '';\n  }\n\n  protected async run(): Promise<void> {\n    let retryable = true;\n\n    try {\n      const messages = (await this.chatCtx.toProviderFormat(\n        'openai.responses',\n      )) as OpenAI.Responses.ResponseInputItem[];\n\n      const tools = this.toolCtx\n        ? Object.entries(this.toolCtx).map(([name, func]) => {\n            const oaiParams = {\n              type: 'function' as const,\n              name: name,\n              description: func.description,\n              parameters: llm.toJsonSchema(\n                func.parameters,\n                true,\n                this.strictToolSchema,\n              ) as unknown as OpenAI.Responses.FunctionTool['parameters'],\n            } as OpenAI.Responses.FunctionTool;\n\n            if (this.strictToolSchema) {\n              oaiParams.strict = true;\n            }\n\n            return oaiParams;\n          })\n        : undefined;\n\n      const requestOptions: Record<string, unknown> = { ...this.modelOptions };\n      if (!tools) {\n        delete requestOptions.tool_choice;\n      }\n\n      const stream = await this.client.responses.create(\n        {\n          model: this.model,\n          input: messages,\n          tools: tools,\n          stream: true,\n          ...requestOptions,\n        },\n        {\n          timeout: this.connOptions.timeoutMs,\n        },\n      );\n\n      for await (const event of stream) {\n        retryable = false;\n        let chunk: llm.ChatChunk | undefined;\n\n        switch (event.type) {\n          case 'error':\n            this.handleError(event);\n            break;\n          case 'response.created':\n            this.handleResponseCreated(event);\n            break;\n          case 'response.output_item.done':\n            chunk = this.handleResponseOutputItemDone(event);\n            break;\n          case 'response.output_text.delta':\n            chunk = this.handleResponseOutputTextDelta(event);\n            break;\n          case 'response.completed':\n            chunk = this.handleResponseCompleted(event);\n            break;\n          case 'response.failed':\n            this.handleResponseFailed(event);\n            break;\n        }\n\n        if (chunk) {\n          this.queue.put(chunk);\n        }\n      }\n    } catch (error) {\n      if (\n        error instanceof APIStatusError ||\n        error instanceof APITimeoutError ||\n        error instanceof APIConnectionError\n      ) {\n        throw error;\n      } else if (error instanceof OpenAI.APIConnectionTimeoutError) {\n        throw new APITimeoutError({ options: { retryable } });\n      } else if (error instanceof OpenAI.APIError) {\n        throw new APIStatusError({\n          message: error.message,\n          options: {\n            statusCode: error.status,\n            body: error.error,\n            requestId: error.requestID,\n            retryable,\n          },\n        });\n      } else {\n        throw new APIConnectionError({\n          message: toError(error).message,\n          options: { retryable },\n        });\n      }\n    }\n  }\n\n  private handleError(event: OpenAI.Responses.ResponseErrorEvent): void {\n    throw new APIStatusError({\n      message: event.message,\n      options: {\n        statusCode: -1,\n        retryable: false,\n      },\n    });\n  }\n\n  private handleResponseFailed(event: OpenAI.Responses.ResponseFailedEvent): void {\n    throw new APIStatusError({\n      message: event.response.error?.message ?? 'Response failed',\n      options: { statusCode: -1, retryable: false },\n    });\n  }\n\n  private handleResponseCreated(event: OpenAI.Responses.ResponseCreatedEvent): void {\n    this.responseId = event.response.id;\n  }\n\n  private handleResponseOutputItemDone(\n    event: OpenAI.Responses.ResponseOutputItemDoneEvent,\n  ): llm.ChatChunk | undefined {\n    let chunk: llm.ChatChunk | undefined;\n\n    if (event.item.type === 'function_call') {\n      chunk = {\n        id: this.responseId,\n        delta: {\n          role: 'assistant',\n          content: undefined,\n          toolCalls: [\n            llm.FunctionCall.create({\n              callId: event.item.call_id || '',\n              name: event.item.name,\n              args: event.item.arguments,\n            }),\n          ],\n        },\n      };\n    }\n    return chunk;\n  }\n\n  private handleResponseOutputTextDelta(\n    event: OpenAI.Responses.ResponseTextDeltaEvent,\n  ): llm.ChatChunk {\n    return {\n      id: this.responseId,\n      delta: {\n        role: 'assistant',\n        content: event.delta,\n      },\n    };\n  }\n\n  private handleResponseCompleted(\n    event: OpenAI.Responses.ResponseCompletedEvent,\n  ): llm.ChatChunk | undefined {\n    if (event.response.usage) {\n      return {\n        id: this.responseId,\n        usage: {\n          completionTokens: event.response.usage.output_tokens,\n          promptTokens: event.response.usage.input_tokens,\n          promptCachedTokens: event.response.usage.input_tokens_details.cached_tokens,\n          totalTokens: event.response.usage.total_tokens,\n        },\n      };\n    }\n    return undefined;\n  }\n}\n\nexport class LLM extends llm.LLM {\n  #opts: LLMOptions;\n  #llm: llm.LLM;\n  #logger = log();\n\n  /**\n   * Create a new instance of OpenAI Responses LLM.\n   *\n   * @remarks\n   * `apiKey` must be set to your OpenAI API key, either using the argument or by setting the\n   * `OPENAI_API_KEY` environment variable.\n   */\n  constructor(opts: Partial<LLMOptions> = defaultLLMOptions) {\n    super();\n\n    this.#opts = { ...defaultLLMOptions, ...opts };\n    const { useWebSocket, client, ...baseOpts } = this.#opts;\n\n    if (useWebSocket) {\n      if (client !== undefined) {\n        this.#logger.warn(\n          'WebSocket mode does not support custom client; provided client will be ignored',\n        );\n      }\n      this.#llm = new WSLLM(baseOpts);\n    } else {\n      this.#llm = new ResponsesHttpLLM({ ...baseOpts, client });\n    }\n\n    // Forward events from the inner delegate so consumers listening on this\n    // wrapper instance (e.g. AgentActivity) receive them.\n    this.#llm.on('metrics_collected', (metrics) => this.emit('metrics_collected', metrics));\n    this.#llm.on('error', (error) => this.emit('error', error));\n  }\n\n  override label(): string {\n    return this.#llm.label();\n  }\n\n  override get model(): string {\n    return this.#llm.model;\n  }\n\n  override prewarm(): void {\n    this.#llm.prewarm();\n  }\n\n  // Ref: python livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/responses/llm.py - 229-233 lines\n  override async aclose(): Promise<void> {\n    await this.#llm.aclose();\n  }\n\n  async close(): Promise<void> {\n    await this.aclose();\n  }\n\n  override chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    extraKwargs?: Record<string, unknown>;\n  }): llm.LLMStream {\n    return this.#llm.chat({\n      chatCtx,\n      toolCtx,\n      connOptions,\n      parallelToolCalls,\n      toolChoice,\n      extraKwargs,\n    });\n  }\n}\n"],"mappings":"AAIA;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AACP,OAAO,YAAY;AAEnB,SAAS,aAAa;AAuBtB,MAAM,oBAAgC;AAAA,EACpC,OAAO;AAAA,EACP,QAAQ,QAAQ,IAAI;AAAA,EACpB,kBAAkB;AAAA,EAClB,cAAc;AAChB;AAEA,MAAM,yBAAyB,IAAI,IAAI;AAAA,EACrC;AAAA,EACA;AAAA,EAEA,YAAY,OAAgC,mBAAmB;AAC7D,UAAM;AAEN,SAAK,QAAQ,EAAE,GAAG,mBAAmB,GAAG,KAAK;AAC7C,QAAI,KAAK,MAAM,WAAW,UAAa,KAAK,MAAM,WAAW,QAAW;AACtE,YAAM,IAAI,MAAM,0EAA0E;AAAA,IAC5F;AAEA,SAAK,UACH,KAAK,MAAM,UACX,IAAI,OAAO;AAAA,MACT,SAAS,KAAK,MAAM;AAAA,MACpB,QAAQ,KAAK,MAAM;AAAA,IACrB,CAAC;AAAA,EACL;AAAA,EAES,QAAgB;AACvB,WAAO;AAAA,EACT;AAAA,EAEA,IAAa,QAAgB;AAC3B,WAAO,KAAK,MAAM;AAAA,EACpB;AAAA,EAES,KAAK;AAAA,IACZ;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAO2B;AACzB,UAAM,eAAwC,EAAE,GAAI,eAAe,CAAC,EAAG;AAEvE,wBACE,sBAAsB,SAAY,oBAAoB,KAAK,MAAM;AAEnE,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,mBAAa,sBAAsB;AAAA,IACrC;AAEA,iBACE,eAAe,SAAY,aAAc,KAAK,MAAM;AAEtD,QAAI,YAAY;AACd,mBAAa,cAAc;AAAA,IAC7B;AAEA,QAAI,KAAK,MAAM,gBAAgB,QAAW;AACxC,mBAAa,cAAc,KAAK,MAAM;AAAA,IACxC;AAEA,QAAI,KAAK,MAAM,UAAU,QAAW;AAClC,mBAAa,QAAQ,KAAK,MAAM;AAAA,IAClC;AAEA,QAAI,KAAK,MAAM,UAAU;AACvB,mBAAa,WAAW,KAAK,MAAM;AAAA,IACrC;AAEA,WAAO,IAAI,uBAAuB,MAAM;AAAA,MACtC,OAAO,KAAK,MAAM;AAAA,MAClB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,KAAK,MAAM,oBAAoB;AAAA,IACnD,CAAC;AAAA,EACH;AACF;AAEA,MAAM,+BAA+B,IAAI,UAAU;AAAA,EACzC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YACEA,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GASA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,QAAQ;AACb,SAAK,SAAS;AACd,SAAK,eAAe;AACpB,SAAK,mBAAmB;AACxB,SAAK,aAAa;AAAA,EACpB;AAAA,EAEA,MAAgB,MAAqB;AACnC,QAAI,YAAY;AAEhB,QAAI;AACF,YAAM,WAAY,MAAM,KAAK,QAAQ;AAAA,QACnC;AAAA,MACF;AAEA,YAAM,QAAQ,KAAK,UACf,OAAO,QAAQ,KAAK,OAAO,EAAE,IAAI,CAAC,CAAC,MAAM,IAAI,MAAM;AACjD,cAAM,YAAY;AAAA,UAChB,MAAM;AAAA,UACN;AAAA,UACA,aAAa,KAAK;AAAA,UAClB,YAAY,IAAI;AAAA,YACd,KAAK;AAAA,YACL;AAAA,YACA,KAAK;AAAA,UACP;AAAA,QACF;AAEA,YAAI,KAAK,kBAAkB;AACzB,oBAAU,SAAS;AAAA,QACrB;AAEA,eAAO;AAAA,MACT,CAAC,IACD;AAEJ,YAAM,iBAA0C,EAAE,GAAG,KAAK,aAAa;AACvE,UAAI,CAAC,OAAO;AACV,eAAO,eAAe;AAAA,MACxB;AAEA,YAAM,SAAS,MAAM,KAAK,OAAO,UAAU;AAAA,QACzC;AAAA,UACE,OAAO,KAAK;AAAA,UACZ,OAAO;AAAA,UACP;AAAA,UACA,QAAQ;AAAA,UACR,GAAG;AAAA,QACL;AAAA,QACA;AAAA,UACE,SAAS,KAAK,YAAY;AAAA,QAC5B;AAAA,MACF;AAEA,uBAAiB,SAAS,QAAQ;AAChC,oBAAY;AACZ,YAAI;AAEJ,gBAAQ,MAAM,MAAM;AAAA,UAClB,KAAK;AACH,iBAAK,YAAY,KAAK;AACtB;AAAA,UACF,KAAK;AACH,iBAAK,sBAAsB,KAAK;AAChC;AAAA,UACF,KAAK;AACH,oBAAQ,KAAK,6BAA6B,KAAK;AAC/C;AAAA,UACF,KAAK;AACH,oBAAQ,KAAK,8BAA8B,KAAK;AAChD;AAAA,UACF,KAAK;AACH,oBAAQ,KAAK,wBAAwB,KAAK;AAC1C;AAAA,UACF,KAAK;AACH,iBAAK,qBAAqB,KAAK;AAC/B;AAAA,QACJ;AAEA,YAAI,OAAO;AACT,eAAK,MAAM,IAAI,KAAK;AAAA,QACtB;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UACE,iBAAiB,kBACjB,iBAAiB,mBACjB,iBAAiB,oBACjB;AACA,cAAM;AAAA,MACR,WAAW,iBAAiB,OAAO,2BAA2B;AAC5D,cAAM,IAAI,gBAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,OAAO,UAAU;AAC3C,cAAM,IAAI,eAAe;AAAA,UACvB,SAAS,MAAM;AAAA,UACf,SAAS;AAAA,YACP,YAAY,MAAM;AAAA,YAClB,MAAM,MAAM;AAAA,YACZ,WAAW,MAAM;AAAA,YACjB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,OAAO;AACL,cAAM,IAAI,mBAAmB;AAAA,UAC3B,SAAS,QAAQ,KAAK,EAAE;AAAA,UACxB,SAAS,EAAE,UAAU;AAAA,QACvB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,YAAY,OAAkD;AACpE,UAAM,IAAI,eAAe;AAAA,MACvB,SAAS,MAAM;AAAA,MACf,SAAS;AAAA,QACP,YAAY;AAAA,QACZ,WAAW;AAAA,MACb;AAAA,IACF,CAAC;AAAA,EACH;AAAA,EAEQ,qBAAqB,OAAmD;AArRlF;AAsRI,UAAM,IAAI,eAAe;AAAA,MACvB,WAAS,WAAM,SAAS,UAAf,mBAAsB,YAAW;AAAA,MAC1C,SAAS,EAAE,YAAY,IAAI,WAAW,MAAM;AAAA,IAC9C,CAAC;AAAA,EACH;AAAA,EAEQ,sBAAsB,OAAoD;AAChF,SAAK,aAAa,MAAM,SAAS;AAAA,EACnC;AAAA,EAEQ,6BACN,OAC2B;AAC3B,QAAI;AAEJ,QAAI,MAAM,KAAK,SAAS,iBAAiB;AACvC,cAAQ;AAAA,QACN,IAAI,KAAK;AAAA,QACT,OAAO;AAAA,UACL,MAAM;AAAA,UACN,SAAS;AAAA,UACT,WAAW;AAAA,YACT,IAAI,aAAa,OAAO;AAAA,cACtB,QAAQ,MAAM,KAAK,WAAW;AAAA,cAC9B,MAAM,MAAM,KAAK;AAAA,cACjB,MAAM,MAAM,KAAK;AAAA,YACnB,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA,EAEQ,8BACN,OACe;AACf,WAAO;AAAA,MACL,IAAI,KAAK;AAAA,MACT,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM;AAAA,MACjB;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,wBACN,OAC2B;AAC3B,QAAI,MAAM,SAAS,OAAO;AACxB,aAAO;AAAA,QACL,IAAI,KAAK;AAAA,QACT,OAAO;AAAA,UACL,kBAAkB,MAAM,SAAS,MAAM;AAAA,UACvC,cAAc,MAAM,SAAS,MAAM;AAAA,UACnC,oBAAoB,MAAM,SAAS,MAAM,qBAAqB;AAAA,UAC9D,aAAa,MAAM,SAAS,MAAM;AAAA,QACpC;AAAA,MACF;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACF;AAEO,MAAM,YAAY,IAAI,IAAI;AAAA,EAC/B;AAAA,EACA;AAAA,EACA,UAAU,IAAI;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASd,YAAY,OAA4B,mBAAmB;AACzD,UAAM;AAEN,SAAK,QAAQ,EAAE,GAAG,mBAAmB,GAAG,KAAK;AAC7C,UAAM,EAAE,cAAc,QAAQ,GAAG,SAAS,IAAI,KAAK;AAEnD,QAAI,cAAc;AAChB,UAAI,WAAW,QAAW;AACxB,aAAK,QAAQ;AAAA,UACX;AAAA,QACF;AAAA,MACF;AACA,WAAK,OAAO,IAAI,MAAM,QAAQ;AAAA,IAChC,OAAO;AACL,WAAK,OAAO,IAAI,iBAAiB,EAAE,GAAG,UAAU,OAAO,CAAC;AAAA,IAC1D;AAIA,SAAK,KAAK,GAAG,qBAAqB,CAAC,YAAY,KAAK,KAAK,qBAAqB,OAAO,CAAC;AACtF,SAAK,KAAK,GAAG,SAAS,CAAC,UAAU,KAAK,KAAK,SAAS,KAAK,CAAC;AAAA,EAC5D;AAAA,EAES,QAAgB;AACvB,WAAO,KAAK,KAAK,MAAM;AAAA,EACzB;AAAA,EAEA,IAAa,QAAgB;AAC3B,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAES,UAAgB;AACvB,SAAK,KAAK,QAAQ;AAAA,EACpB;AAAA;AAAA,EAGA,MAAe,SAAwB;AACrC,UAAM,KAAK,KAAK,OAAO;AAAA,EACzB;AAAA,EAEA,MAAM,QAAuB;AAC3B,UAAM,KAAK,OAAO;AAAA,EACpB;AAAA,EAES,KAAK;AAAA,IACZ;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAOkB;AAChB,WAAO,KAAK,KAAK,KAAK;AAAA,MACpB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AACF;","names":["llm"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@livekit/agents-plugin-openai",
-  "version": "1.2.1",
+  "version": "1.2.3",
   "description": "OpenAI plugin for LiveKit Node Agents",
   "main": "dist/index.js",
   "require": "dist/index.cjs",
@@ -30,9 +30,9 @@
     "@types/ws": "^8.5.10",
     "tsup": "^8.3.5",
     "typescript": "^5.0.0",
-    "@livekit/agents-plugin-silero": "1.2.1",
-    "@livekit/agents": "1.2.1",
-    "@livekit/agents-plugins-test": "1.2.1"
+    "@livekit/agents": "1.2.3",
+    "@livekit/agents-plugin-silero": "1.2.3",
+    "@livekit/agents-plugins-test": "1.2.3"
   },
   "dependencies": {
     "@livekit/mutex": "^1.1.1",
@@ -42,7 +42,7 @@
   "peerDependencies": {
     "@livekit/rtc-node": "^0.13.24",
     "zod": "^3.25.76 || ^4.1.8",
-    "@livekit/agents": "1.2.1"
+    "@livekit/agents": "1.2.3"
   },
   "scripts": {
     "build": "tsup --onSuccess \"pnpm build:types\"",

package/src/llm.ts

@@ -521,7 +521,7 @@ export class LLM extends llm.LLM {
     return new LLMStream(this as unknown as inference.LLM, {
       model: this.#opts.model,
       providerFmt: this.#providerFmt,
-      client: this.#client,
+      client: this.#client as any,
       chatCtx,
       toolCtx,
       connOptions,

package/src/realtime/realtime_model.ts

@@ -419,6 +419,12 @@ export class RealtimeSession extends llm.RealtimeSession {
     this.sendEvent(this.createSessionUpdateEvent());
   }
 
+  private hasInflightResponse(): boolean {
+    return (
+      this.currentGeneration !== undefined || Object.keys(this.responseCreatedFutures).length > 0
+    );
+  }
+
   sendEvent(command: api_proto.ClientEvent): void {
     this.messageChannel.put(command);
   }
@@ -624,6 +630,13 @@ export class RealtimeSession extends llm.RealtimeSession {
   }
 
   updateOptions({ toolChoice }: { toolChoice?: llm.ToolChoice }): void {
+    const currentToolChoice = toOaiToolChoice(this.oaiRealtimeModel._options.toolChoice);
+    const nextToolChoice = toOaiToolChoice(toolChoice);
+    if (currentToolChoice === nextToolChoice) {
+      this.oaiRealtimeModel._options.toolChoice = toolChoice;
+      return;
+    }
+
     const options: api_proto.SessionUpdateEvent['session'] = {
       type: 'realtime',
     };
@@ -677,6 +690,10 @@ export class RealtimeSession extends llm.RealtimeSession {
   }
 
   async interrupt(): Promise<void> {
+    if (!this.hasInflightResponse()) {
+      return;
+    }
+
     this.sendEvent({
       type: 'response.cancel',
     } as api_proto.ResponseCancelEvent);
@@ -831,6 +848,13 @@ export class RealtimeSession extends llm.RealtimeSession {
       }
       this.itemDeleteFutures = {};
 
+      for (const handle of Object.values(this.responseCreatedFutures)) {
+        if (!handle.doneFut.done) {
+          handle.doneFut.reject(new Error('Session reconnected'));
+        }
+      }
+      this.responseCreatedFutures = {};
+
       // Clear audio-capable item tracking - restored items are text-only on the server
       this.audioCapableItemIds.clear();
 
@@ -1236,13 +1260,24 @@ export class RealtimeSession extends llm.RealtimeSession {
       throw new Error('item.id is not set');
     }
 
+    const serverEventType = event.type as string;
+    const incomingItem = openAIItemToLivekitItem(event.item);
+    const existingItem = this.remoteChatCtx.get(event.item.id);
+    const pendingCreateFuture = this.itemCreateFutures[event.item.id];
+    if (existingItem && serverEventType === 'conversation.item.added' && !pendingCreateFuture) {
+      // The server may emit a later input-audio-backed view of a user item whose
+      // transcribed text variant we already inserted locally under the same ID.
+      // Treat that as idempotent instead of surfacing a duplicate-ID error.
+      return;
+    }
+
     try {
-      this.remoteChatCtx.insert(event.previous_item_id, openAIItemToLivekitItem(event.item));
+      this.remoteChatCtx.insert(event.previous_item_id, incomingItem);
     } catch (error) {
       this.#logger.error({ error, itemId: event.item.id }, 'failed to insert conversation item');
     }
 
-    const fut = this.itemCreateFutures[event.item.id];
+    const fut = pendingCreateFuture;
     if (fut) {
       fut.resolve();
       delete this.itemCreateFutures[event.item.id];

package/src/responses/llm.ts

@@ -228,6 +228,9 @@ class ResponsesHttpLLMStream extends llm.LLMStream {
           case 'response.completed':
             chunk = this.handleResponseCompleted(event);
             break;
+          case 'response.failed':
+            this.handleResponseFailed(event);
+            break;
         }
 
         if (chunk) {
@@ -272,6 +275,13 @@ class ResponsesHttpLLMStream extends llm.LLMStream {
     });
   }
 
+  private handleResponseFailed(event: OpenAI.Responses.ResponseFailedEvent): void {
+    throw new APIStatusError({
+      message: event.response.error?.message ?? 'Response failed',
+      options: { statusCode: -1, retryable: false },
+    });
+  }
+
   private handleResponseCreated(event: OpenAI.Responses.ResponseCreatedEvent): void {
     this.responseId = event.response.id;
   }