/**
 * Observe provider-reported usage and relay-estimated model output without
 * changing the response path. In particular, this module never changes (or
 * buffers) a response chunk: it only keeps bounded parsing state alongside
 * the stream.
 */
/**
 * Protocols the USAGE observer actually parses. Deliberately narrower than
 * `StreamCommitProtocol`: no call site observes a native OpenAI Responses body
 * (Responses front-door traffic is translated to Anthropic before it is proxied),
 * so there is no `"openai-responses"` member to re-add if one ever appears.
 */
export type UsageProtocol = "anthropic-messages" | "openai-chat";
export interface UsageAccumulator {
    /** Provider-reported prompt/input tokens. */
    inputTokens: number | undefined;
    /** Provider-reported completion/output tokens. */
    outputTokens: number | undefined;
    /** OpenAI's explicitly reported cached prompt/input tokens. */
    cachedInputTokens: number | undefined;
    /** Anthropic reports cache writes and cache reads as distinct facts. */
    cacheCreationInputTokens: number | undefined;
    cacheReadInputTokens: number | undefined;
    /** Compatibility alias consumed by existing model telemetry. */
    completionTokens: number | undefined;
    /** chars/4 estimate over model-authored content, never response framing. */
    estimatedOutputTokens: number | undefined;
}
export declare function createUsageAccumulator(): UsageAccumulator;
/**
 * Return a response whose body is byte-for-byte the input body while usage is
 * observed synchronously as each chunk passes through.
 */
export declare function observeUsage(response: Response, protocol: UsageProtocol, accumulator: UsageAccumulator, options?: {
    streamed?: boolean;
}): Response;