import { OpenAI } from 'openai';
import type { LLMCompletionMessage, LLMCompletionResult, LLMInstanceOptions, LLMModel, LLMModelUsage, LLMSendMessageOptions, LLMStructuredResult } from '../../../LLMService.typedefs';
import { LLMProviders } from '../../../LLMService.typedefs';
import { type LLMLoggerInterface } from '../../../utilities/logger';
import { type LLMReporterInterface } from '../../../utilities/reporter';
import { OpenAICompatibleCompletionService } from '../../../providers/OpenAICompatible/OpenAICompatibleCompletion.service';
import { type LLMSchemaInterface } from '../../../utilities/schema';
export declare class LLMAPICompletionService<Reporter extends LLMReporterInterface<any> | undefined> extends OpenAICompatibleCompletionService<LLMProviders.LLMAPI, Reporter> {
    constructor(logger: LLMLoggerInterface | undefined, reporter: Reporter, options?: LLMInstanceOptions[LLMProviders.LLMAPI]);
    get instance(): OpenAI;
    private combineInstructions;
    protected buildChatMessages(options: {
        messages: LLMCompletionMessage[];
        model: LLMModel<LLMProviders.LLMAPI>;
    }): OpenAI.ChatCompletionMessageParam[];
    protected sendTextMessage(options: LLMSendMessageOptions<LLMProviders.LLMAPI, Reporter>): Promise<LLMCompletionResult>;
    protected sendStructuredMessage(options: LLMSendMessageOptions<LLMProviders.LLMAPI, Reporter> & {
        schema: LLMSchemaInterface;
    }): Promise<LLMStructuredResult>;
    /**
     * A route handed audio it never fetched still answers, fluently and from
     * nothing, so the reported audio token count is the only thing separating a
     * real read from a fabrication.
     */
    protected assertCompletionAccepted(options: {
        request: LLMSendMessageOptions<LLMProviders.LLMAPI, Reporter>;
        usage: LLMModelUsage;
    }): void;
    /**
     * Audio tokens are counted inside `prompt_tokens`, so the audio half is moved
     * out of the text half rather than added on top: the gateway prices text and
     * audio input separately and would otherwise bill the recording twice.
     */
    protected extractUsageFromCompletion(completion: OpenAI.ChatCompletion): {
        inputTextTokens: number;
        inputAudioTokens: number;
        inputCachedTextTokens: number;
        outputTextTokens: number;
        outputReasoningTokens: number;
    };
    /**
     * Builds the provider-native message array and the response format before the
     * request is attempted, so the success and the error path trace the same
     * transcript rather than one of them falling back to a raw
     * `{ message, history }` object. It runs outside the request's `try` on
     * purpose: combining messages and rendering the schema is local work, and its
     * failure is a configuration error rather than a provider one.
     */
    private prepareStructuredRequest;
}