/**
 * How many tokens the model's context window holds.
 *
 * The loop has never known this, which is why every context rail it had was a
 * number a host guessed at: `contextTokenBudget` is an absolute figure with no
 * default and no way to reach it from `createVendo`, so in practice nothing
 * bounded the prompt but the provider's own 400. A window table is the smallest
 * thing that turns that into a decision the loop can make for itself.
 *
 * Matched by SUBSTRING, longest first, because a model id does not arrive as a
 * bare name. It arrives prefixed by whichever gateway routed it and suffixed by a
 * snapshot date — `us.anthropic.claude-sonnet-4-6-20260101` — so a table of exact
 * keys would miss every real id and silently run the whole shipment on the
 * default.
 *
 * No tokenizer, no network lookup, no per-provider metadata fetch: a table in the
 * repo is wrong slowly and visibly, which is the failure mode a host can fix with
 * the override below.
 */
import type { LanguageModel } from "ai";
/** The window assumed for a model this table does not name. Deliberately the
 *  smallest window still in wide use: under-guessing costs one early compaction,
 *  over-guessing costs a 400 in the middle of somebody's turn. */
export declare const DEFAULT_CONTEXT_WINDOW_TOKENS = 128000;
/**
 * Substring → window, longest match wins.
 *
 * Family entries carry the family's standard window and member entries carry the
 * exceptions, so a new dated snapshot of a known family is right on the day it
 * ships rather than on the day someone remembers this file. Every figure is the
 * window available on a PLAIN request: Anthropic's 1M window is behind a beta
 * header we do not send, so claiming it here would trade one early compaction for
 * a request the provider rejects.
 */
export declare const MODEL_CONTEXT_WINDOWS: readonly (readonly [match: string, tokens: number])[];
/** Record what the provider reported for `model`, from the turn that called it. */
export declare function rememberResolvedModelId(model: LanguageModel, reported: string | undefined): void;
/** What the provider reported for `model`, if a call has reported yet. The
 *  metering path reads it too: usage is priced on the model that served the
 *  tokens, and a lazy seat's own id is a family name, not a model. */
export declare function resolvedModelId(model: LanguageModel): string | undefined;
/**
 * THE one new public knob of this shipment.
 *
 * `override` is the BYO escape and it wins outright, table hit or not: a host on
 * a model this repo has never heard of, or on a seat whose entry has gone stale,
 * needs a way to be right that does not involve waiting for a release. It has to
 * be a positive WHOLE number of tokens to be a window at all, and this is the
 * only place either door is checked: nothing in the stack parses a harness's
 * options schema, so the per-turn knob arrives exactly as unvalidated as the
 * deployment one — which is why the rule lives HERE, at the function both doors
 * reach, and not in a declaration beside one of them. Both ends of the range
 * fail the same way, silently and in opposite
 * directions: a zero puts the trigger at zero, so every turn pays for a
 * summarizer pass and then sheds the conversation to its last message; an
 * infinity puts the trigger past every estimate there is, so compaction never
 * fires again and the provider's 400 is the only rail left. A fraction is the
 * zero in disguise — `triggerTokens` floors the window times the ratio, so any
 * window under ~1.24 tokens clears `> 0` and still trips at zero.
 * `Number.isInteger` is false for `NaN` and both infinities too, so it is the
 * whole rule in one call.
 */
export declare function contextWindowTokens(model: LanguageModel, override?: number): number;
//# sourceMappingURL=model-windows.d.ts.map