import { defineTool } from "@grok/agent-sdk/tools";
import { z } from "zod";
const DATA: Record<string, { name: string; tests: string; task_count: number }> = {
  "swe-bench-verified": { name: "SWE-bench Verified", tests: "fixing real GitHub issues in Python repositories", task_count: 500 },
  "gpqa-diamond": { name: "GPQA Diamond", tests: "graduate-level science multiple-choice questions", task_count: 198 },
};
export default defineTool({
  description: "Look up a benchmark by slug (e.g. swe-bench-verified, gpqa-diamond). Returns what it tests and its task count.",
  inputSchema: z.object({ slug: z.string() }),
  async execute({ slug }) {
    const row = DATA[slug.toLowerCase()];
    return row ? { found: true, ...row } : { found: false, slug };
  },
});