/** Drop the cache. Tests that swap the global provider need this. */
export declare function resetRuntimeMetrics(): void;
export interface TokenUsageSample {
    readonly promptTokens: number;
    readonly completionTokens: number;
    /** Prompt tokens served from the provider's cache. */
    readonly cachedTokens?: number;
    /** Prompt tokens written into the provider's cache. */
    readonly cacheWriteTokens?: number;
}
/**
 * Record what a turn cost, split by token type.
 *
 * Cache reads and writes are recorded as their own types rather than
 * folded into `input`. They bill differently — a read is a fraction of the
 * input rate and a write is a premium on it — so a total that hides them
 * cannot explain a bill, which is the main thing anyone asks a token
 * metric.
 */
export declare function recordTokenUsage(model: string, usage: TokenUsageSample): void;
/**
 * Record a tool call.
 *
 * `errorType` carries WHY a call failed rather than only that it did. A
 * flat success rate cannot separate a tool that is broken from one whose
 * input the model keeps getting wrong, and those need different fixes.
 */
export declare function recordToolCall(toolName: string, success: boolean, errorType?: string, durationMs?: number): void;
/**
 * Record how long the first content delta took to arrive.
 *
 * Recorded at the first delta rather than at the end of the request,
 * because the two are not the same measurement and only this one tracks
 * what a person waiting actually experiences.
 */
export declare function recordTimeToFirstToken(model: string, durationMs: number): void;
/** Record how long a whole turn took, keyed by how it settled. */
export declare function recordTurnDuration(status: string, durationMs: number): void;
/** Record how long one model request took. */
export declare function recordModelDuration(model: string, durationMs: number, errorType?: string): void;
//# sourceMappingURL=metrics.d.ts.map