Function executeCodexScenario

scripts/skill-bench.js:678–720

(scenario, codexCmd, tmpDir)

}

function executeCodexScenario(scenario, codexCmd, tmpDir) {
  const outputPath = path.join(tmpDir, '.planning', 'benchmark-results', `${scenario.skill}-${scenario.name}-codex.md`);
  fs.mkdirSync(path.dirname(outputPath), { recursive: true });
  const codexArgs = buildCodexExecArgs({
    projectRoot: tmpDir,
    sandbox: CODEX_SANDBOX,
    model: CODEX_MODEL,
    outputLastMessagePath: outputPath,
    prompt: scenario.input,
    json: true,
  });
  if (CODEX_SERVICE_TIER) {
    codexArgs.splice(1, 0, '-c', `service_tier="${CODEX_SERVICE_TIER}"`);
  }
  try {
    const output = execFileSync(
      codexCmd,
      codexArgs,
      {
        cwd: tmpDir,
        timeout: scenario.timeout,
        encoding: 'utf8',
        env: {
          ...process.env,
          CITADEL_RUNTIME: 'codex',
        },
        maxBuffer: 10 * 1024 * 1024,
      }
    );
    const finalMessage = fs.existsSync(outputPath) ? fs.readFileSync(outputPath, 'utf8') : '';
    return { output: finalMessage || output, rawOutput: output, error: null, timedOut: false, command: ['codex', ...codexArgs].join(' ') };
  } catch (err) {
    if (err.killed) {
      return { output: err.stdout || '', error: 'timed out', timedOut: true, command: ['codex', ...codexArgs].join(' ') };
    }
    return {
      output: err.stdout || '',
      error: err.message || 'unknown error',
      timedOut: false,
      command: ['codex', ...codexArgs].join(' '),
    };
  }
}

// ── Assertion running ─────────────────────────────────────────────────────────