import type { ProviderId, ToolResult } from "../types.js";
export interface ImageToolRunOptions {
    signal?: AbortSignal | undefined;
    onOutput?: ((chunk: string, stream: "stdout" | "stderr") => void) | undefined;
    /** Active route — decides vision support and the per-image size budget. */
    llmProvider?: ProviderId | undefined;
    llmModel?: string | undefined;
}
export declare function imageOcr(args: Record<string, unknown>, options?: ImageToolRunOptions): Promise<ToolResult>;
/**
 * Put real image bytes in front of a vision-capable model.
 *
 * The agent could previously only reach an image through `image.ocr`, so a
 * model that had just captured a screenshot to verify its own work had no way
 * to actually look at it — it got a Tesseract transcript of whatever text
 * happened to be legible, which is useless for "does this render correctly"
 * and actively misleading for photos, 3D scenes and charts.
 *
 * Tool results are text-only on every provider wire, so this tool returns the
 * prepared bytes on {@link ToolResult.images} and the agent replays them as a
 * follow-up user turn — the same path a user attachment takes, which every
 * multimodal adapter (OpenAI, Anthropic, Gemini, Ollama) already serializes.
 */
export declare function imageView(args: Record<string, unknown>, options?: ImageToolRunOptions): Promise<ToolResult>;