Function main

scripts/bench/agent-control-benchmark.mjs:46–98

()

  });

async function main() {
  assertBinary(simdeckBin, "Run npm run build:cli before benchmarking.");

  const versions = {
    simdeck: readVersion(simdeckBin, ["--version"]),
    "agent-device": readVersion("agent-device", ["--version"]),
    argent: readVersion("argent", ["--version"]),
  };

  const udid = options.udid || selectBootedIphoneUdid();
  if (!udid) {
    throw new Error(
      "No booted iPhone simulator found. Boot one first or pass --udid <UDID>.",
    );
  }

  const metadata = {
    generatedAt: new Date().toISOString(),
    udid,
    reps: options.reps,
    versions,
    notes: [
      "Cold rows stop tool services where practical before measuring the next first command.",
      "Setup/reset work is excluded from action timings and uses xcrun/simdeck to normalize Settings state.",
      "Argent has no selector tap/wait primitive in this comparison, so comparable navigation uses AX-derived coordinates.",
    ],
  };

  console.log(`Benchmarking simulator ${udid}`);
  console.log(
    `Versions: simdeck ${versions.simdeck}, agent-device ${versions["agent-device"]}, argent ${versions.argent}`,
  );

  await benchmarkSimDeck(udid);
  await benchmarkAgentDevice(udid);
  await benchmarkArgent(udid);

  const report = buildReport(metadata, rows);
  const outDir =
    options.outDir ||
    path.join(tempRoot, new Date().toISOString().replace(/[:.]/g, "-"));
  mkdirSync(outDir, { recursive: true });
  const jsonPath = path.join(outDir, "agent-control-benchmark.json");
  const markdownPath = path.join(outDir, "agent-control-benchmark.md");
  writeFileSync(jsonPath, JSON.stringify({ metadata, rows }, null, 2));
  writeFileSync(markdownPath, report);

  console.log("");
  console.log(report);
  console.log("");
  console.log(`Wrote ${jsonPath}`);
  console.log(`Wrote ${markdownPath}`);
}

async function benchmarkSimDeck(udid) {
  section("simdeck");