/**
 * darwin eval <agent> --tasks <file.json>
 *
 * Offline eval: run stored prompt versions (or all of them) over a frozen
 * task set, score each output with the built-in critic, and print/write a
 * per-task comparison table with arm deltas against the baseline.
 *
 *   darwin eval writer --tasks tasks.json                # active version vs v1
 *   darwin eval writer --tasks tasks.json --versions v1,v3
 *   darwin eval writer --tasks tasks.json --all-versions --runs 3
 *   darwin eval writer --tasks tasks.json --dry          # wiring check, 0 LLM calls
 *
 * Task-set format (JSON): `[{"id": "t1", "type": "tech", "task": "..."}, …]`
 * or `{"tasks": [...]}` — the same shape as `benchmark/seed-tasks.json`.
 */
interface EvalFlags {
    agentName: string;
    tasksPath?: string;
    /** Explicit version labels (`--versions v1,v3`). */
    versions?: string[];
    allVersions: boolean;
    runsPerCell: number;
    json: boolean;
    dry: boolean;
}
export declare function parseEvalArgs(args: string[]): EvalFlags;
export declare function evalCommand(args: string[]): Promise<void>;
export {};
//# sourceMappingURL=eval.d.ts.map