/**
 * Tiered token accounting for context virtualization (2.6.0 hardening).
 *
 * The governor must never under-count against a model's REAL context limit.
 * Before this module, the sole safety basis was `chars / 4`, which is a good
 * approximation for English prose but dangerously under-counts content that a
 * real tokenizer expands:
 *
 *   - CJK / non-ASCII text (≈1 token per code point, not 1 per 4 chars)
 *   - emoji / astral-plane code points (≥1 token each, often more)
 *   - minified JSON / dense punctuation (≈1 token per symbol)
 *   - code (symbols and identifiers tokenize finer than prose)
 *
 * Tiering (best to worst):
 *
 *   exact                 -> provider/model tokenizer when available (not yet wired)
 *   tokenizer             -> local deterministic tokenizer (not yet wired)
 *   calibrated            -> bounded uplift derived from provider-reported usage
 *   conservative-estimate -> content-class aware fallback (always biased safe)
 *
 * No exact tokenizer dependency is added: the fallback is deterministic,
 * content-class aware, and biased toward over-estimation.
 */
export type TokenAccountingMode = "exact" | "tokenizer" | "calibrated" | "conservative-estimate";
/**
 * Content class used to pick the conservative divisor. Lower divisor = more
 * conservative (more tokens per character).
 */
export type ContentClass = "prose" | "code" | "json" | "tool-schema" | "tool-result" | "system-prompt" | "log";
export interface TextTokenAnalysis {
    /** ASCII characters (code points < 0x80). */
    asciiChars: number;
    /** ASCII punctuation/symbol characters (not alphanumeric, not whitespace). */
    asciiSymbols: number;
    /** Non-ASCII BMP code points (CJK, accented Latin, etc.). */
    nonAsciiBmpChars: number;
    /** Astral-plane code points (surrogate pairs — emoji, rare scripts). */
    astralChars: number;
    /** Ratio of ASCII symbols to total ASCII (0..1). */
    symbolRatio: number;
}
export interface TokenAccountingOptions {
    /**
     * Bounded conservative uplift multiplier from provider-usage calibration.
     * Must be >= 1; only ever makes estimates larger, never smaller.
     */
    calibratedMultiplier?: number;
}
/** Bound the calibrated uplift so one outlier cannot cripple the budget. */
export declare const MAX_CALIBRATED_MULTIPLIER = 1.5;
/**
 * Analyze a string into the token-relevant character classes. Iterates by code
 * point (not UTF-16 code unit) so surrogate pairs are counted as one astral char.
 */
export declare function analyzeText(text: string): TextTokenAnalysis;
/**
 * Conservative token estimate for a single text payload.
 *
 * Guarantees (bias):
 *   - ASCII prose: ~1 token / 4 chars (matches the historical heuristic).
 *   - Code / JSON / logs: ~1 token / 3 chars or finer.
 *   - Dense punctuation (minified JSON): symbols counted 1:1, remaining /4.
 *   - Non-ASCII BMP (CJK): 1 token per code point.
 *   - Astral (emoji): 2 tokens per code point.
 *
 * These are deliberately upper-biased for the content classes a coding agent
 * actually encounters.
 */
export declare function estimateTextTokens(text: string, contentClass?: ContentClass, options?: TokenAccountingOptions): number;
/** Resolve the effective accounting mode from the current calibration state. */
export declare function resolveAccountingMode(calibratedMultiplier: number | undefined): TokenAccountingMode;
//# sourceMappingURL=token-accounting.d.ts.map