import type { AIProviderName } from "../constants/enums.js";
import type { OpenAICompatBuildBodyArgs, OpenAICompatStreamLifecycleListeners } from "../types/index.js";
import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
/**
 * LiteLLM Provider — direct HTTP, no AI SDK. Talks to a LiteLLM proxy
 * server (or any deployment that speaks OpenAI chat-completions + the
 * `/v1/models` and `/v1/embeddings` endpoints).
 *
 * All request/stream/tool-loop orchestration lives in
 * `OpenAIChatCompletionsProvider`. This class adds LiteLLM-specific
 * behaviour: OTel span wrap with cost (`onStreamStart`), Gemini 2.5
 * maxTokens skip (`adjustBuildBodyOptions`), ModelAccessDeniedError on
 * 403, 10-minute model cache (`getAvailableModels`), `LITELLM_FALLBACK_MODELS`
 * env-driven fallback list, and native `/v1/embeddings`.
 */
export declare class LiteLLMProvider extends OpenAIChatCompletionsProvider {
    private static modelsCache;
    private static modelsCacheTime;
    private static readonly MODELS_CACHE_DURATION;
    constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: {
        apiKey?: string;
        baseURL?: string;
    });
    protected getProviderName(): AIProviderName;
    protected getDefaultModel(): string;
    protected getFallbackModelName(): string;
    protected getFallbackModels(): string[];
    /**
     * Gemini 2.5 models on LiteLLM have a known compatibility issue with
     * `max_tokens` — strip it before the wire body is built. Applies to
     * both streaming and non-streaming paths.
     */
    protected adjustBuildBodyOptions(modelId: string, opts: OpenAICompatBuildBodyArgs["options"]): OpenAICompatBuildBodyArgs["options"];
    /**
     * Wrap the stream in an OTel span to capture provider-level latency,
     * token usage, finish reason, and cost. Matches the pre-migration
     * behaviour where streamText was wrapped in `neurolink.provider.streamText`.
     */
    protected onStreamStart(modelId: string): OpenAICompatStreamLifecycleListeners | undefined;
    formatProviderError(error: unknown): Error;
    /**
     * Get available models from LiteLLM proxy `/v1/models` endpoint.
     * Caches results for 10 minutes; falls back to env-driven list or a
     * minimal safe default if the API fetch fails.
     */
    getAvailableModels(): Promise<string[]>;
    private fetchModelsFromAPI;
    /**
     * Generate an embedding for a single text input via native /v1/embeddings.
     */
    embed(text: string, modelName?: string): Promise<number[]>;
    /**
     * Generate embeddings for multiple text inputs via native /v1/embeddings.
     */
    embedMany(texts: string[], modelName?: string): Promise<number[][]>;
    private callEmbeddings;
}