import { OpenAI, Raw } from '@vscode/prompt-tsx';
import { TelemetryData } from '../../telemetry/common/telemetryData';
import { ThinkingData, ThinkingDataInMessage } from '../../thinking/common/thinking';
import { ICopilotReference, RequestId } from './fetch';
/**
 * How the logprobs field looks in the OpenAI API chunks.
 */
export interface APILogprobs {
    text_offset: number[];
    token_logprobs: number[];
    top_logprobs?: {
        [key: string]: number;
    }[];
    tokens: string[];
}
/**
 * Usage statistics for the completion request.
 */
export interface APIUsage {
    /**
     * Number of tokens in the prompt.
     */
    prompt_tokens: number;
    /**
     * Number of tokens in the generated completion.
     */
    completion_tokens: number;
    /**
     * Total number of tokens used in the request (prompt + completion).
     */
    total_tokens: number;
    /**
     * Breakdown of tokens used in the prompt.
     */
    prompt_tokens_details?: {
        cached_tokens: number;
        cache_creation_input_tokens?: number;
        /**
         * Anthropic-specific: per-TTL breakdown of cache-creation (write) input
         * tokens. Mirrors Anthropic's `usage.cache_creation` object verbatim.
         * Only populated for Anthropic Messages API responses where the server
         * reports the split; absent for all other providers and for older
         * Anthropic responses that don't include the breakdown.
         */
        anthropic_cache_creation?: {
            /** Cache-creation tokens written with the 1h (extended) TTL — billed at 2x base input rate. */
            ephemeral_1h_input_tokens?: number;
            /** Cache-creation tokens written with the default 5m TTL — billed at 1.25x base input rate. */
            ephemeral_5m_input_tokens?: number;
        };
    };
    /**
     * Breakdown of tokens used in a completion.
     *
     * @remark it's an optional field because Copilot Proxy returns this information but not CAPI as of 18 Jun 2025
     */
    completion_tokens_details?: {
        /**
         * Tokens generated by the model for reasoning.
         */
        reasoning_tokens: number;
        /**
         * When using Predicted Outputs, the number of tokens in the prediction that appeared in the completion.
         */
        accepted_prediction_tokens: number;
        /**
         * When using Predicted Outputs, the number of tokens in the prediction that did not appear in the completion.
         * However, like reasoning tokens, these tokens are still counted in the total completion tokens for purposes of billing,
         * output, and context window limits.
         */
        rejected_prediction_tokens: number;
    };
    /**
     * Copilot billing usage for this request, returned by CAPI.
     */
    copilot_usage?: {
        total_nano_aiu: number;
    };
}
export declare function isApiUsage(obj: unknown): obj is APIUsage;
/**
 * Converts a nano-AIU value from copilot_usage to a credit number.
 * Returns `undefined` when the value is missing or negative.
 */
export declare function nanoAiuToCredits(nanoAiu: number | undefined): number | undefined;
export interface APIJsonData {
    text: string;
    tokens: readonly string[];
    logprobs?: APILogprobs;
}
export interface APIErrorResponse {
    code: number;
    message: string;
    metadata?: Record<string, any>;
}
export declare const openAIContextManagementCompactionType = "compaction";
export declare const modelsWithoutResponsesContextManagement: Set<string>;
export interface OpenAIContextManagement {
    type: typeof openAIContextManagementCompactionType;
    compact_threshold: number;
}
export interface OpenAIContextManagementResponse {
    encrypted_content: string;
    type: typeof openAIContextManagementCompactionType;
    id: string;
}
export declare enum ChatRole {
    System = "system",
    User = "user",
    Assistant = "assistant",
    Function = "function",
    Tool = "tool"
}
export type CAPIChatMessage = OpenAI.ChatMessage & {
    /**
     * CAPI references used in this message.
     */
    copilot_references?: ICopilotReference[];
    /**
     * CAPI confirmations used in this message.
     */
    copilot_confirmations?: {
        state: string;
        confirmation: any;
    }[];
    copilot_cache_control?: {
        'type': 'ephemeral';
    };
} & ThinkingDataInMessage;
export declare function getCAPITextPart(content: string | OpenAI.ChatCompletionContentPart[] | OpenAI.ChatCompletionContentPart): string;
export type RawMessageConversionCallback = (message: CAPIChatMessage, thinkingData?: ThinkingData) => void;
/**
 * Converts a raw TSX chat message to CAPI's format.
 *
 * **Extra:** the raw message can have `copilot_references` and
 * `copilot_confirmations` properties, which are copied to the CAPI message.
 */
export declare function rawMessageToCAPI(message: Raw.ChatMessage, callback?: RawMessageConversionCallback): CAPIChatMessage;
export declare function rawMessageToCAPI(message: Raw.ChatMessage[], callback?: RawMessageConversionCallback): CAPIChatMessage[];
export declare enum FinishedCompletionReason {
    /**
     * Reason generated by the server. See https://platform.openai.com/docs/guides/gpt/chat-completions-api
     */
    Stop = "stop",
    /**
     * Reason generated by the server. See https://platform.openai.com/docs/guides/gpt/chat-completions-api
     */
    Length = "length",
    /**
     * Reason generated by the server. See https://platform.openai.com/docs/guides/gpt/chat-completions-api
     */
    FunctionCall = "function_call",
    /**
     * Reason generated by the server. See https://platform.openai.com/docs/guides/gpt/chat-completions-api
     */
    ToolCalls = "tool_calls",
    /**
     * Reason generated by the server. See https://platform.openai.com/docs/guides/gpt/chat-completions-api
     */
    ContentFilter = "content_filter",
    /**
     * Reason generated by the server. The model itself declined, as opposed to {@link ContentFilter}
     * where a separate system blocked the response.
     */
    Refusal = "refusal",
    /**
     * Reason generated by the server (CAPI). Happens when the stream cannot be completed and the server must terminate the response.
     */
    ServerError = "error",
    /**
     * Reason generated by the client when the finish callback asked for processing to stop.
     */
    ClientTrimmed = "client-trimmed",
    /**
     * Reason generated by the client when we never received a finish_reason for this particular completion (indicates a server-side bug)
     */
    ClientIterationDone = "Iteration Done",
    /**
     * Reason generated by the client when we never received a finish_reason for this particular completion (indicates a server-side bug)
     */
    ClientDone = "DONE"
}
export interface IToolCall {
    index: number;
    id?: string;
    function?: {
        name: string;
        arguments: string;
    };
}
/**
 * Contains the possible reasons a response can be filtered
 */
export declare enum FilterReason {
    /**
     * Content deemed to be hateful
     */
    Hate = "hate",
    /**
     * Content deemed to cause self harm
     */
    SelfHarm = "self_harm",
    /**
     * Content deemed to be sexual in nature
     */
    Sexual = "sexual",
    /**
     * Content deemed to be violent in nature
     */
    Violence = "violence",
    /**
     * Content contains copyrighted material
     */
    Copyright = "snippy",
    /**
     * The prompt was filtered, the reason was not provided
     */
    Prompt = "prompt"
}
export interface ChatCompletion {
    message: Raw.ChatMessage;
    choiceIndex: number;
    requestId: RequestId;
    tokens: readonly string[];
    usage: APIUsage | undefined;
    model: string;
    blockFinished: boolean;
    finishReason: FinishedCompletionReason;
    filterReason?: FilterReason;
    telemetryData: TelemetryData;
    error?: APIErrorResponse;
}
export interface ChoiceLogProbs {
    content: ChoiceLogProbsContent[];
}
export interface TokenLogProb {
    bytes: number[];
    token: string;
    logprob: number;
}
export interface ChoiceLogProbsContent extends TokenLogProb {
    top_logprobs: TokenLogProb[];
}
//# sourceMappingURL=openai.d.ts.map