/**
 * Context Budget System for Kuralle.
 *
 * Manages token allocation across the multiple content sources that compose
 * each LLM request. Ensures no single section can monopolize the context window.
 */
/**
 * ContextBudgetConfig defines the token allocation strategy for the LLM
 * context window. It governs how the finite context window is partitioned
 * across the multiple content sources that compose each LLM request.
 */
export interface ContextBudgetConfig {
    /** Total context window size of the target model, in tokens. Default: 128,000. */
    modelContextWindow: number;
    /** Tokens reserved for the LLM's response generation. Default: 4,096. */
    responseReserve: number;
    /** Maximum tokens allocated to auto-retrieve (RAG) context injection. Default: 4,000. */
    maxAutoRetrieveTokens: number;
    /** Maximum tokens allocated to the working memory section. Default: 2,000. */
    maxWorkingMemoryTokens: number;
    /** Maximum tokens allocated to the extraction snapshot section. Default: 2,000. */
    maxExtractionTokens: number;
    /** Maximum tokens allocated to cross-session long-term memory. Default: 2,000. */
    maxLongTermMemoryTokens: number;
    /**
     * Maximum tokens allocated to the base system prompt.
     * 0 means no limit — the base prompt is measured, not capped.
     * Default: 0.
     */
    maxBasePromptTokens: number;
}
export declare const DEFAULT_CONTEXT_BUDGET: ContextBudgetConfig;
/** Preset for realtime / voice agents — smaller window and allocations for low-latency turns. */
export declare const VOICE_CONTEXT_BUDGET: ContextBudgetConfig;
/**
 * Token estimation function. Uses a rough 4:1 character-to-token ratio.
 * Consistent with ContextManager.ts.
 */
export declare function estimateTokenCount(text: string): number;
/**
 * Computes the remaining token budget available for message history after
 * all system prompt sections have been allocated.
 *
 * Floors at 1000 tokens to ensure the LLM always sees at least minimal
 * conversation history.
 */
export declare function computeMessageHistoryBudget(config: ContextBudgetConfig, measuredBasePromptTokens: number, measuredPolicyInjectionTokens: number): number;
/**
 * Truncates a text string to fit within a token budget.
 * Truncation is at the nearest sentence boundary when possible.
 */
export declare function truncateToTokenBudget(text: string, maxTokens: number): string;
/**
 * Budget-aware working memory formatter.
 * Iterates entries in insertion order, dropping entries that exceed the budget.
 * Entries are atomic — they are either fully included or fully dropped.
 */
export declare function formatMemoryWithBudget(memory: Record<string, unknown>, maxTokens: number, allowlist?: string[]): string;
/**
 * Tracks the last pre-flight input token estimate and compares it to provider-reported actuals.
 */
export declare class ContextBudget {
    private readonly config;
    private lastEstimate;
    constructor(config: ContextBudgetConfig);
    /** Call after assembling the request (e.g. from `onBeforeModelCall` estimated tokens). */
    recordPreFlightEstimate(estimatedInputTokens: number): void;
    /**
     * Compares the last pre-flight estimate to actual input tokens from the provider.
     * Logs a warning when absolute drift percentage exceeds 20%.
     */
    validateActual(actualInputTokens: number): {
        estimated: number;
        actual: number;
        drift: number;
        driftPct: number;
    };
    get modelContextWindow(): number;
}