/**
 * Type-safe MLLM (Multimodal Large Language Model) vendor classes.
 *
 * MLLM vendors handle real-time audio end-to-end, bypassing the standard
 * ASR → LLM → TTS pipeline. Calling `agent.withMllm(vendor)` automatically
 * sets `mllm.enable: true`.
 */
import type { MllmConfig, MllmTurnDetectionConfig } from "../types.js";
import { BaseCNMLLM, BaseMLLM, type BaseMllmOptions } from "./base.js";
export declare const GeminiLiveModels: {
    /** Public ID for the low-latency voice model. */
    readonly Live38: "models/gemini-3.8-live";
    /** Public ID for the reasoning voice model. */
    readonly Live38ExtendedThinking: "models/gemini-3.8-live-extended-thinking";
};
/** The model Gemini Live sends by default. */
export declare const GEMINI_MLLM_DEFAULT_MODEL: string;
/** Default Gemini Developer API endpoint for Gemini 3.8 models. */
export declare const GEMINI_MLLM_URL = "https://generativelanguage.googleapis.com";
/** @deprecated Use {@link GEMINI_MLLM_URL}. */
export declare const GEMINI_PREVIEW_MLLM_URL: string;
/** A Gemini Live model. Known models autocomplete while future model IDs remain accepted. */
export type GeminiLiveModel = (typeof GeminiLiveModels)[keyof typeof GeminiLiveModels] | (string & {});
/** Reasoning budgets accepted by Gemini Extended Thinking. */
export declare const GeminiThinkingLevels: readonly ["low", "medium", "high"];
export type GeminiThinkingLevel = (typeof GeminiThinkingLevels)[number];
/** Known Gemini output voices; future voice names remain accepted. */
export type GeminiLiveVoice = "Puck" | "Charon" | "Kore" | "Fenrir" | "Aoede" | "Leda" | "Orus" | "Zephyr" | (string & {});
/** @deprecated Use {@link GeminiLiveVoice}. */
export type GeminiPreviewVoice = GeminiLiveVoice;
export declare const OPENAI_GPT_LIVE_VENDOR: "openai_gpt_live";
export declare function isOpenAIGPTLiveConfig(config: unknown): boolean;
/** GPT Live v3 options. */
export interface OpenAIGPTLiveOptions extends BaseMllmOptions {
    apiKey: string;
    /** Full WebSocket URL, used verbatim except OpenAI's legacy /v1/live route. */
    url?: string;
    /** @deprecated Use prompt. Serialized as prompt; an explicit prompt wins. */
    instructions?: string;
    greeting?: string;
    failureMessage?: string;
    inputModalities?: string[];
    outputModalities?: string[];
    messages?: Record<string, unknown>[];
    /** Additional provider fields; explicit options take precedence. */
    params?: Record<string, unknown>;
    /** @deprecated Unsupported in v3; setting this raises an error. */
    inputAudioTranscription?: Record<string, unknown>;
    /** @deprecated Ignored with a warning; v3 performs endpointing internally. */
    turnDetection?: MllmTurnDetectionConfig;
    /** Defaults to gpt-live-1. */
    model?: string;
    /** Output voice; provider default marin. Custom voice objects require PR #1522; use params after rollout. */
    voice?: string;
    /** Session instructions. */
    prompt?: string;
    /** Host when url is omitted; default wss://api.openai.com. */
    baseUrl?: string;
    /** WebSocket path; default /v1/live/sessions. */
    path?: string;
    /** Optional OpenAI-Alpha selector for preview contracts. Omitted by default. */
    alphaSelector?: string;
    /** Extra provider request headers as a JSON string; protocol headers win. */
    headers?: string;
    /** Assistant silence boundary in ms; provider default 600. Zero disables inference. */
    outputIdleEndMs?: number;
    /** Caller silence boundary in ms; provider default 1500. */
    inputIdleEndMs?: number;
    /** Speech amplitude threshold on the 16-bit scale; provider default 50. */
    outputSilencePeak?: number;
    /** Graph PCM sample rate; provider default 24000. */
    outputSampleRate?: number;
    /** Initial audio cushion; provider default 0. Negative disables pacing. */
    outputBufferMs?: number;
    /** Mic append batching in ms. Join default 0; extension class default 100. */
    inputBatchMs?: number;
    /** Advertise graph tools; provider default false. Does not control delegate built-ins. */
    toolEnabled?: boolean;
    /** Tool delegation mode; provider default responses. Fixed for the session. */
    delegation?: "client" | "responses";
    /** Tool delegate model; provider default gpt-5.6-sol. */
    responsesModel?: string;
    /** Interrupt playback on caller speech; provider default false. */
    interruptOnUserTurn?: boolean;
    /** Unmodelled v3 session fields. Cannot override model, delegation, audio, instructions or input. */
    sessionParams?: Record<string, unknown>;
}
export declare class OpenAIGPTLive extends BaseMLLM {
    private readonly options;
    constructor(options: OpenAIGPTLiveOptions);
    toConfig(): MllmConfig;
}
/**
 * Constructor options for OpenAI Realtime API.
 */
export interface OpenAIRealtimeOptions extends BaseMllmOptions {
    /** OpenAI API key */
    apiKey: string;
    /** Model name (e.g., 'gpt-4o-realtime-preview') */
    model?: string;
    /** Voice identifier for audio output */
    voice?: string;
    /** System instructions that define agent behavior */
    instructions?: string;
    /** Audio transcription settings */
    inputAudioTranscription?: Record<string, unknown>;
    /** WebSocket URL. Defaults to `wss://api.openai.com/v1/realtime` when omitted or empty. */
    url?: string;
    /** Agent greeting message */
    greetingMessage?: string;
    /** Input modalities (e.g., ['audio'], ['audio', 'text']) */
    inputModalities?: string[];
    /** Output modalities (e.g., ['text', 'audio']) */
    outputModalities?: string[];
    /** Conversation messages for short-term memory */
    messages?: Record<string, unknown>[];
    /** Additional MLLM parameters */
    params?: Record<string, unknown>;
    /** MLLM turn detection configuration. Overrides top-level turn_detection. */
    turnDetection?: MllmTurnDetectionConfig;
    /** Message played on failure */
    failureMessage?: string;
}
/**
 * OpenAI Realtime API MLLM vendor.
 *
 * @example
 * ```typescript
 * const client = new AgoraClient({ area: Area.US, appId: '...', appCertificate: '...' });
 * const agent = new Agent({ client })
 *   .withMllm(new OpenAIRealtime({
 *     apiKey: process.env.OPENAI_API_KEY,
 *     greetingMessage: 'Hello! How can I help you?',
 *   }));
 * ```
 */
export declare class OpenAIRealtime extends BaseMLLM {
    private readonly options;
    constructor(options: OpenAIRealtimeOptions);
    toConfig(): MllmConfig;
}
/** Parameters accepted by Azure OpenAI Realtime. */
export interface AzureOpenAIRealtimeParams {
    /** System instructions that define agent behavior */
    instructions?: string;
    /** Model or deployment model identifier */
    model?: string;
    /** Voice identifier for audio output */
    voice?: string;
}
/** Constructor options for Azure OpenAI Realtime API. */
export interface AzureOpenAIRealtimeOptions extends BaseMllmOptions {
    /** Azure OpenAI API key */
    apiKey: string;
    /** Azure OpenAI Realtime WebSocket URL, including deployment routing when required */
    url: string;
    /** Model or deployment model identifier */
    model?: string;
    /** Voice identifier for audio output */
    voice?: string;
    /** System instructions that define agent behavior */
    instructions?: string;
    /** Number of conversation history messages cached by the MLLM */
    maxHistory?: number;
    /** Agent greeting message */
    greetingMessage?: string;
    /** Output modalities (e.g., ['text', 'audio']) */
    outputModalities?: string[];
    /** Conversation messages for short-term memory */
    messages?: Record<string, unknown>[];
    /** Azure Realtime model parameters */
    params?: AzureOpenAIRealtimeParams;
    /** Required MLLM turn detection configuration. Overrides top-level turn_detection. */
    turnDetection: MllmTurnDetectionConfig;
}
/**
 * Azure OpenAI Realtime MLLM vendor for global deployments.
 *
 * @example
 * ```typescript
 * const agent = new Agent({ client }).withMllm(new AzureOpenAIRealtime({
 *   apiKey: process.env.AZURE_OPENAI_API_KEY,
 *   url: 'wss://example.openai.azure.com/openai/realtime',
 *   model: 'gpt-4o-realtime-preview',
 *   maxHistory: 32,
 *   turnDetection: { mode: 'server_vad' },
 * }));
 * ```
 */
export declare class AzureOpenAIRealtime extends BaseMLLM {
    private readonly options;
    constructor(options: AzureOpenAIRealtimeOptions);
    toConfig(): MllmConfig;
}
/**
 * Constructor options for Google Gemini Live (direct API, non-Vertex AI).
 */
export interface GeminiLiveOptions extends BaseMllmOptions {
    /** Google API key */
    apiKey: string;
    /** Model name (e.g., 'gemini-live-2.5-flash') */
    model?: GeminiLiveModel;
    /** Sent only for models/gemini-3.8-live-extended-thinking. */
    thinkingLevel?: GeminiThinkingLevel;
    /** Languages for Gemini 3.8, sent as params.language_codes. */
    languageCodes?: readonly string[];
    /** Endpoint override; Gemini 3.8 defaults to the Developer API host. */
    url?: string;
    /** System instructions for the model */
    instructions?: string;
    /** Voice name (e.g., 'Aoede', 'Charon') */
    voice?: GeminiLiveVoice;
    affectiveDialog?: boolean;
    proactiveAudio?: boolean;
    transcribeAgent?: boolean;
    transcribeUser?: boolean;
    httpOptions?: Record<string, unknown>;
    /** Agent greeting message */
    greetingMessage?: string;
    /** Input modalities (e.g., ['audio'], ['audio', 'text']) */
    inputModalities?: string[];
    /** Output modalities (e.g., ['text', 'audio']) */
    outputModalities?: string[];
    /** Conversation messages for short-term memory */
    messages?: Record<string, unknown>[];
    /** Additional MLLM parameters passed directly to the model */
    additionalParams?: Record<string, unknown>;
    /** MLLM turn detection configuration. Overrides top-level turn_detection. */
    turnDetection?: MllmTurnDetectionConfig;
    /** Message played on failure */
    failureMessage?: string;
}
/**
 * Google Gemini Live MLLM vendor (direct API, non-Vertex AI).
 *
 * Uses a Google API key. For Vertex AI / ADC credentials use {@link VertexAI} instead.
 *
 * @example
 * ```typescript
 * const client = new AgoraClient({ area: Area.US, appId: '...', appCertificate: '...' });
 * const agent = new Agent({ client })
 *   .withMllm(new GeminiLive({
 *     apiKey: process.env.GOOGLE_API_KEY,
 *     model: 'gemini-live-2.5-flash',
 *     greetingMessage: 'Hello! Gemini is listening.',
 *   }));
 * ```
 */
export declare class GeminiLive extends BaseMLLM {
    private readonly options;
    constructor(options: GeminiLiveOptions);
    toConfig(): MllmConfig;
}
/**
 * Constructor options for Google Gemini Live (Vertex AI).
 */
export interface VertexAIOptions extends BaseMllmOptions {
    /** Model name (e.g., 'gemini-live-2.5-flash-preview-native-audio-09-2025') */
    model: string;
    /** WebSocket URL for real-time communication */
    url?: string;
    /** Google Cloud project ID */
    projectId: string;
    /** Google Cloud location/region */
    location: string;
    /** Application Default Credentials JSON string */
    adcCredentialsString: string;
    /** System instructions for the model */
    instructions?: string;
    /** Voice name (e.g., 'Aoede', 'Charon') */
    voice?: string;
    affectiveDialog?: boolean;
    proactiveAudio?: boolean;
    transcribeAgent?: boolean;
    transcribeUser?: boolean;
    httpOptions?: Record<string, unknown>;
    /** Agent greeting message */
    greetingMessage?: string;
    /** Input modalities (e.g., ['audio'], ['audio', 'text']) */
    inputModalities?: string[];
    /** Output modalities (e.g., ['text', 'audio']) */
    outputModalities?: string[];
    /** Conversation messages for short-term memory */
    messages?: Record<string, unknown>[];
    /** Additional MLLM parameters */
    additionalParams?: Record<string, unknown>;
    /** MLLM turn detection configuration. Overrides top-level turn_detection. */
    turnDetection?: MllmTurnDetectionConfig;
    /** Message played on failure */
    failureMessage?: string;
}
/**
 * Google Gemini Live (Vertex AI) MLLM vendor.
 *
 * @example
 * ```typescript
 * const client = new AgoraClient({ area: Area.US, appId: '...', appCertificate: '...' });
 * const agent = new Agent({ client })
 *   .withMllm(new VertexAI({
 *     model: 'gemini-live-2.5-flash-preview-native-audio-09-2025',
 *     projectId: process.env.GOOGLE_PROJECT_ID,
 *     location: 'us-central1',
 *     adcCredentialsString: process.env.GOOGLE_ADC_CREDENTIALS,
 *     instructions: 'You are a helpful voice assistant.',
 *     voice: 'Aoede',
 *     greetingMessage: 'Hello! Gemini is listening.',
 *   }));
 * ```
 */
export declare class VertexAI extends BaseMLLM {
    private readonly options;
    constructor(options: VertexAIOptions);
    toConfig(): MllmConfig;
}
/**
 * Constructor options for xAI Grok Realtime API.
 */
export interface XaiGrokOptions extends BaseMllmOptions {
    /** xAI API key */
    apiKey: string;
    /** WebSocket URL for real-time communication (defaults to xAI Realtime API) */
    url?: string;
    /** Voice identifier (e.g., 'eve', 'rex') */
    voice?: string;
    /** Language code (e.g., 'en') */
    language?: string;
    /** Audio sample rate in Hz (e.g., 24000) */
    sampleRate?: number;
    /** Agent greeting message */
    greetingMessage?: string;
    /** Message played on failure */
    failureMessage?: string;
    /** Input modalities (e.g., ['audio'], ['audio', 'text']) */
    inputModalities?: string[];
    /** Output modalities (e.g., ['audio'], ['text', 'audio']) */
    outputModalities?: string[];
    /** Conversation messages for short-term memory */
    messages?: Record<string, unknown>[];
    /** Additional MLLM parameters passed directly to xAI */
    params?: Record<string, unknown>;
    /** MLLM turn detection configuration. Overrides top-level turn_detection. */
    turnDetection?: MllmTurnDetectionConfig;
}
/**
 * xAI Grok MLLM vendor (`mllm.vendor`: `"xai"`).
 *
 * Uses the xAI Realtime API WebSocket URL by default. Do not name future xAI ASR/TTS
 * wrappers `XaiRealtime`; use `XaiSTT` / `XaiTTS` when those pipelines are added.
 *
 * @example
 * ```typescript
 * const client = new AgoraClient({ area: Area.US, appId: '...', appCertificate: '...' });
 * const agent = new Agent({ client })
 *   .withMllm(new XaiGrok({
 *     apiKey: process.env.XAI_API_KEY,
 *     voice: 'eve',
 *     language: 'en',
 *     sampleRate: 24000,
 *     greetingMessage: 'Hello, how can I help?',
 *   }));
 * ```
 */
export declare class XaiGrok extends BaseMLLM {
    private readonly options;
    constructor(options: XaiGrokOptions);
    toConfig(): MllmConfig;
}
/** Constructor options for Alibaba Cloud Qwen Omni Realtime. */
export interface QwenOmniOptions extends BaseMllmOptions {
    /** Alibaba Cloud DashScope API key */
    apiKey: string;
    /** Qwen Omni model identifier */
    model: string;
    /** Qwen Omni Realtime WebSocket URL */
    url: string;
    /** Voice identifier for audio output */
    voice?: string;
    /** System instructions that define agent behavior */
    instructions?: string;
    /** Agent greeting message */
    greetingMessage?: string;
    /** Message played on failure */
    failureMessage?: string;
    /** Input modalities (e.g., ['audio'], ['audio', 'text']) */
    inputModalities?: string[];
    /** Output modalities (e.g., ['text', 'audio']) */
    outputModalities?: string[];
    /** Conversation messages for short-term memory */
    messages?: Record<string, unknown>[];
    /** Additional Qwen Omni parameters */
    params?: Record<string, unknown>;
    /** MLLM turn detection configuration. Overrides top-level turn_detection. */
    turnDetection?: MllmTurnDetectionConfig;
}
/**
 * Alibaba Cloud Qwen Omni Realtime MLLM vendor for Chinese mainland deployments.
 *
 * @example
 * ```typescript
 * const agent = new Agent({ client }).withMllm(new QwenOmni({
 *   apiKey: process.env.DASHSCOPE_API_KEY,
 *   model: 'qwen3.5-omni-plus-realtime',
 *   url: 'wss://dashscope.aliyuncs.com/api-ws/v1/realtime',
 *   greetingMessage: '你好，有什么可以帮你？',
 * }));
 * ```
 */
export declare class QwenOmni extends BaseCNMLLM {
    private readonly options;
    constructor(options: QwenOmniOptions);
    toConfig(): MllmConfig;
}