import type { CompletionRequest, Message } from './messages.js';
/**
 * How a long conversation is cut to fit: keep the last messages or the last tokens, optionally with
 * a summary of what was cut, or `auto` to choose from the limits configured and whether summaries
 * are on.
 */
export type ContextWindowStrategy = 'last-messages' | 'last-tokens' | 'last-messages-with-summary' | 'last-tokens-with-summary' | 'auto';
/**
 * Who writes the summary of cut messages: a local extractive summary, the model through the client,
 * or your own function.
 */
export type ContextSummaryMode = 'local' | 'provider' | 'custom';
/** What a custom summarizer receives. */
export interface ContextSummaryInput {
    /** The request being trimmed. */
    request: CompletionRequest;
    /** The messages to summarize. */
    messages: Message[];
    /** Those messages as one text, ready to put in a prompt. */
    serializedMessages: string;
    /** Longest the summary may be, in tokens. */
    maxTokens: number;
    /** Model to summarize with. */
    model?: string;
    /** Instructions for the summary. */
    instruction: string;
    /** Sampling temperature for the summary. */
    temperature?: number;
}
/** Writes a summary of the messages cut from a conversation. */
export type ContextSummarizer = (input: ContextSummaryInput) => string | Promise<string>;
/** How cut messages are summarized. */
export interface ContextSummaryConfig {
    /** Turns summaries on. */
    enabled?: boolean;
    /** Who writes the summary. Defaults to `local`. */
    mode?: ContextSummaryMode;
    /** Model used in `provider` mode. */
    model?: string;
    /** Longest a summary may be, in tokens. Defaults to 512. */
    maxTokens?: number;
    /** Sampling temperature in `provider` mode. */
    temperature?: number;
    /** Instructions given to the summarizer. */
    instruction?: string;
    /** Heading the summary is inserted under. Defaults to `Earlier conversation summary`. */
    label?: string;
    /** Role of the inserted summary message. Defaults to `system`. */
    insertAsRole?: 'system' | 'user';
    /** Your own summarizer, used instead of either built-in mode. */
    summarizer?: ContextSummarizer;
    /**
     * Falls back to a local summary when `provider` mode fails. Defaults to true; `false` fails the
     * request instead.
     */
    fallbackToLocal?: boolean;
}
/** Keeps long conversations within a model's context window. */
export interface ContextWindowConfig {
    /** Turns trimming on. */
    enabled?: boolean;
    /** How the conversation is cut. */
    strategy?: ContextWindowStrategy;
    /** Conversation messages kept, system messages aside. Defaults to 20. */
    lastMessages?: number;
    /** Longest the request may be, in estimated tokens. */
    maxInputTokens?: number;
    /**
     * Tokens held back for the summary when trimming by tokens. Defaults to the summary's
     * `maxTokens`.
     */
    summaryReserveTokens?: number;
    /** Keeps system messages whatever else is cut. Defaults to true. */
    preserveSystemMessages?: boolean;
    /** How cut messages are summarized. */
    summary?: ContextSummaryConfig;
}
/** What trimming did to a request. */
export interface ContextWindowUsage {
    /** The strategy actually applied, which `auto` resolves to one of the others. */
    strategy: ContextWindowStrategy;
    /** Estimated tokens before trimming. */
    beforeTokens: number;
    /** Estimated tokens after trimming. */
    afterTokens: number;
    /** Tokens saved. */
    savedTokens: number;
    /** Messages before trimming. */
    originalMessages: number;
    /** Messages after trimming, the summary included. */
    finalMessages: number;
    /** Messages kept unchanged. */
    keptMessages: number;
    /** Messages folded into a summary. */
    summarizedMessages: number;
    /** Messages removed without a summary. */
    droppedMessages: number;
    /** Estimated tokens in the summary. */
    summaryTokens: number;
    /** Summaries inserted. */
    summariesCreated: number;
    /** Who wrote the summary. */
    summaryMode?: ContextSummaryMode;
    /** Model that wrote the summary, in `provider` mode. */
    summaryModel?: string;
}
/** A trimmed value with what trimming did to it. */
export interface ContextWindowResult<T> {
    /** The trimmed value. */
    value: T;
    /** Token and message counts before and after. */
    usage: ContextWindowUsage;
    /** Techniques applied, for response metadata. */
    techniquesApplied: string[];
    /** Anything the caller should know, such as a summary that fell back to the local mode. */
    warnings: string[];
}