export declare function isFinalUsageFrame(parsed: {
    usage?: unknown;
    choices?: unknown[] | undefined;
}): boolean;
/**
 * Mid-stream silence budget.
 *
 * This is deliberately generous because "no bytes" does **not** mean "dead
 * socket" on an OpenAI-compatible endpoint. Most self-hosted runtimes (vLLM /
 * SGLang and the tool-call parsers layered on top of them) buffer an entire
 * `tool_calls` delta before emitting it, so a model writing a large file goes
 * completely silent on the wire for as long as the generation takes. A 90s
 * budget aborted those healthy streams at `firstToken + 90s`, reported the
 * abort as a network failure, and burned three identical retries that each
 * re-generated the same prefix before one happened to finish inside the window.
 */
export declare const DEFAULT_STREAM_IDLE_TIMEOUT_MS = 240000;
export declare const THINKING_STREAM_IDLE_TIMEOUT_MS = 900000;
export declare const THINKING_STREAM_INITIAL_IDLE_TIMEOUT_MS = 900000;
export declare function streamIdleBudgets(reasoningEnabled: boolean): {
    idleTimeoutMs: number;
    outputIdleTimeoutMs: number;
};
export declare const STREAM_STALL_MARKER = "no model output";
export interface StreamLineReaderOptions {
    signal?: AbortSignal | undefined;
    idleTimeoutMs?: number | undefined;
    maxBytes?: number | undefined;
    onActivity?: (() => void) | undefined;
    outputIdleTimeoutMs?: number | undefined;
    outputProgress?: (() => number) | undefined;
}
export declare function readStreamLines(response: Response, options?: StreamLineReaderOptions): AsyncGenerator<string, void, void>;
export declare function createSseFrameAssembler(options?: {
    maxBufferedBytes?: number;
}): {
    pushLine: (line: string) => string | undefined;
};