import type { Model } from "../../internal/llm.js";
import { type ModelPricing } from "../pricing.js";
/** How a provider counts cached tokens in its reported `input`. */
export type CacheFamily = "input-includes-cached" | "input-excludes-cached";
/**
 * Resolve the prompt-cache accounting family for a model. **Auto-inferred from `model.api`** by
 * default (Anthropic/Bedrock report `input` EXCLUDING cached tokens; OpenAI/vLLM/Google INCLUDE them),
 * but a deployment can **override** it via `model.params.promptCacheFamily` when `model.api` doesn't
 * match the brain that actually serves the model — e.g. a routing brain that dispatches by `provider`
 * while `api` is left at a default. Accepts the explicit values or the aliases `"anthropic"`/`"openai"`.
 * (design/18: a route-by-provider setup with `api` hardcoded misjudged the family → 1090% hit rate.)
 */
export declare function cacheFamilyOf(model: Model): CacheFamily;
/** Total prompt tokens presented this turn, normalized so `cachedTokens` is always ⊆ `promptTokens`. */
export declare function promptTokensOf(family: CacheFamily, u: {
    input: number;
    cacheRead: number;
    cacheWrite: number;
}): number;
/**
 * Normalize one turn's reported usage into its `{ totalInputTokens, costMicroUsd }` — the authoritative
 * cost re-priced in core from injected pricing (so cached tokens aren't charged at full price). Shared
 * by the run loop's per-turn accounting and post-task memory consolidation (design/41) so a pricing
 * change lives in ONE place. We emit only 5-min-TTL cache writes, so `cacheWriteTokensLong` is 0.
 */
export declare function usageCostMicroUsd(family: CacheFamily, u: {
    input?: number;
    output?: number;
    cacheRead?: number;
    cacheWrite?: number;
}, pricing: ModelPricing): {
    totalInputTokens: number;
    costMicroUsd: number;
};
//# sourceMappingURL=usage-accounting.d.ts.map