/**
* Rate-limit header parser + optional client-side token bucket.
*
* - `parseRateLimitHeaders(headers)` reads OpenAI-style and IETF-draft
*   `x-ratelimit-*` headers; returns null when none recognized.
* - `withRateLimit(misina, opts)` adds an in-process limiter that gates
*   each request before dispatch, learns the real budget from response
*   headers, and backs off on 429.
*/
import type { MisinaPlugin } from "../types.mjs";
export interface RateLimitBucket {
	limit: number | undefined;
	remaining: number | undefined;
	resetAt: Date | undefined;
}
export interface RateLimitInfo {
	requests: RateLimitBucket | undefined;
	tokens: RateLimitBucket | undefined;
}
export declare function parseRateLimitHeaders(headers: Headers): RateLimitInfo | null;
export interface RateLimitOptions {
	/** Initial requests-per-minute budget. Default: Infinity (unlimited). */
	rpm?: number;
	/** Initial tokens-per-minute budget. Default: Infinity (unlimited). */
	tpm?: number;
	/**
	* Estimate token cost for a request before dispatch. Used to acquire
	* tokens from the TPM bucket. Return 0 to skip the TPM gate for a
	* request. Default: () => 0 (no TPM accounting).
	*/
	estimateTokens?: (request: Request) => number;
	/**
	* Override the source of "now" — useful in tests with fake timers.
	* Default: `() => Date.now()`.
	*/
	now?: () => number;
}
/**
* Token-bucket rate limiter wired into a `Misina` instance via hooks.
* Two buckets — requests-per-minute and tokens-per-minute — refill
* linearly. `beforeRequest` waits until both buckets cover the cost.
* `onComplete` reads `x-ratelimit-remaining-*` and adjusts the bucket
* to the real budget the server reports. 429 responses drain both
* buckets aggressively so subsequent calls back off.
*/
export declare function rateLimit(options?: RateLimitOptions): MisinaPlugin;