/**
 * Rolling context management for a talk session.
 *
 * A realtime conversation has no compaction step: everything the model has heard
 * stays in its context and is re-billed on every turn. So the gateway keeps its
 * own estimate and drops the oldest turns once the estimate passes a budget.
 */
import type { TalkTurnRecord } from "./types.js";
/** Where truncation starts. Roughly a long conversation, well under any
 *  provider's window — the budget exists to bound cost, not to avoid an error. */
export declare const TALK_CONTEXT_BUDGET_TOKENS = 6000;
/** A single turn this large is a research request wearing a conversation's
 *  clothes; the orb offers a text session instead of answering it out loud. */
export declare const TALK_HANDOFF_TURN_TOKENS = 1500;
/** Four characters per token. A documented estimate, not the provider's count —
 *  the real numbers arrive with each turn's usage and are what gets billed. */
export declare function estimateTokens(text: string): number;
export declare function contextTokens(turns: readonly TalkTurnRecord[]): number;
export interface TruncationResult {
    turns: TalkTurnRecord[];
    /** How many turns this call dropped. */
    dropped: number;
}
/**
 * Drop oldest turns until the estimate fits the budget. The newest turn is
 * always kept, even when it exceeds the budget on its own: a session with
 * nothing in context cannot answer the thing the user just said. That case is
 * what `handoffSuggested` is for.
 */
export declare function truncateTurns(turns: readonly TalkTurnRecord[], budgetTokens: number): TruncationResult;
/** Whether this turn is heavy enough to be worth moving to a text session. */
export declare function handoffSuggested(turn: TalkTurnRecord): boolean;
//# sourceMappingURL=context.d.ts.map