Files
maestro/scripts/bench-run.ts
T
oss-sync 57702b4f34
CI / build-and-test (push) Canceled after 0s
sync: update from private repo (2336949a)
2026-09-17 11:29:31 +00:00

159 lines
6.1 KiB
TypeScript

#!/usr/bin/env -S npx tsx
/**
* CLI entry for the benchmark runner.
*
* npm run bench # all tasks
* npm run bench -- --task=composite-mini-report
* npm run bench -- --server=http://127.0.0.1:9876
* BENCH_JUDGE=off npm run bench # skip LLM judge (axis D not graded, excluded from the total)
*
* Prereq: scripts/server.sh start (or `npm start`) running on the configured port.
*/
import * as fs from 'fs';
import * as path from 'path';
import { parse as parseYaml } from 'yaml';
import { startFixtureServer } from '../src/bench/fixture-server.js';
import { runOneTask } from '../src/bench/runner.js';
import { exitCodeForResults, runAndGradeTask } from '../src/bench/grade-run.js';
import { loadJudgeConfigFromProvider } from '../src/bench/judge.js';
import { formatAxisCell, formatTotalCell, writeRunSummary } from '../src/bench/summary.js';
import type { BenchResult, BenchTask } from '../src/bench/types.js';
import { loadConfig } from '../src/config.js';
import { logger } from '../src/logger.js';
interface Args {
taskFilter?: string;
serverUrl: string;
benchRoot: string;
resultDir: string;
}
function parseArgs(): Args {
const argv = process.argv.slice(2);
let taskFilter: string | undefined;
let serverUrl = process.env['BENCH_SERVER_URL'] ?? 'http://127.0.0.1:9876';
for (const a of argv) {
if (a.startsWith('--task=')) taskFilter = a.slice('--task='.length);
else if (a.startsWith('--server=')) serverUrl = a.slice('--server='.length);
}
const benchRoot = path.resolve(process.cwd(), 'bench');
const runId = new Date().toISOString().replace(/[:.]/g, '-');
const resultDir = path.join(benchRoot, 'results', runId);
return { taskFilter, serverUrl, benchRoot, resultDir };
}
function loadTasks(benchRoot: string, filter?: string): BenchTask[] {
const taskDir = path.join(benchRoot, 'tasks');
if (!fs.existsSync(taskDir)) return [];
const files = fs.readdirSync(taskDir).filter((f) => f.endsWith('.yaml') || f.endsWith('.yml'));
const tasks: BenchTask[] = [];
for (const f of files) {
const parsed = parseYaml(fs.readFileSync(path.join(taskDir, f), 'utf-8')) as BenchTask;
if (filter && parsed.id !== filter) continue;
tasks.push(parsed);
}
return tasks;
}
function copyWorkspaceArtifacts(rawWorkspace: string, dest: string): void {
if (!rawWorkspace || !fs.existsSync(rawWorkspace)) return;
fs.mkdirSync(dest, { recursive: true });
function copyDir(src: string, target: string): void {
fs.mkdirSync(target, { recursive: true });
for (const entry of fs.readdirSync(src, { withFileTypes: true })) {
const s = path.join(src, entry.name);
const t = path.join(target, entry.name);
if (entry.isDirectory()) copyDir(s, t);
else if (entry.isFile()) {
try {
fs.copyFileSync(s, t);
} catch (err) {
logger.warn(`[bench] copy failed ${s} -> ${t}: ${(err as Error).message}`);
}
}
}
}
for (const sub of ['logs', 'output']) {
const src = path.join(rawWorkspace, sub);
if (fs.existsSync(src)) copyDir(src, path.join(dest, sub));
}
}
async function main(): Promise<void> {
const args = parseArgs();
const config = loadConfig();
const judgeConfig = loadJudgeConfigFromProvider(config.provider);
const tasks = loadTasks(args.benchRoot, args.taskFilter);
if (tasks.length === 0) {
logger.error(`[bench] no tasks found (filter=${args.taskFilter ?? 'none'})`);
process.exit(1);
}
fs.mkdirSync(args.resultDir, { recursive: true });
logger.info(`[bench] run dir: ${args.resultDir}`);
logger.info(`[bench] server: ${args.serverUrl}`);
logger.info(`[bench] judge: ${judgeConfig ? `${judgeConfig.endpoint} model=${judgeConfig.model}` : 'off'}`);
const fixtureServer = await startFixtureServer(path.join(args.benchRoot, 'fixtures', 'web'));
const results: BenchResult[] = [];
try {
for (const task of tasks) {
logger.info(`[bench] === task: ${task.id} ===`);
const taskResultDir = path.join(args.resultDir, task.id);
fs.mkdirSync(taskResultDir, { recursive: true });
const result = await runAndGradeTask(task, {
runTask: (t) =>
runOneTask(t, {
serverUrl: args.serverUrl,
benchRoot: args.benchRoot,
webPort: fixtureServer.port,
}),
judgeConfig,
});
if (result.outcome === 'failed') {
logger.error(`[bench] runner failed for ${task.id}: ${result.error}`);
} else if (result.raw) {
copyWorkspaceArtifacts(result.raw.workspacePath, path.join(taskResultDir, 'workspace'));
}
fs.writeFileSync(path.join(taskResultDir, 'result.json'), JSON.stringify(result, null, 2), 'utf-8');
results.push(result);
}
} finally {
await fixtureServer.close();
}
const summaryPath = writeRunSummary(args.resultDir, results);
logger.info(`[bench] wrote ${summaryPath}`);
// Concise stdout summary so CI / humans see totals without opening the file.
// Axes that were not graded print their status, never a percentage, and the
// scored weight travels with the total so two runs are comparable on purpose.
for (const r of results) {
if (r.outcome === 'failed') {
// eslint-disable-next-line no-console
console.log(`${r.taskId}: FAILED (not scored) error=${r.error ?? 'unknown'}`);
continue;
}
const weight = r.scoring ? `${r.scoring.scoredWeight}/${r.scoring.maxWeight}` : '?';
// eslint-disable-next-line no-console
console.log(
`${r.taskId}: total=${formatTotalCell(r)}/100 (scored weight ${weight}) status=${r.raw?.status ?? '-'} ` +
`tools=${formatAxisCell(r.axes?.tools)} checklist=${formatAxisCell(r.axes?.checklist)} ` +
`instr=${formatAxisCell(r.axes?.instructions)} reasoning=${formatAxisCell(r.axes?.reasoning)}`,
);
}
const failedCount = results.filter((r) => r.outcome === 'failed').length;
if (failedCount > 0) {
// eslint-disable-next-line no-console
console.log(`${failedCount} of ${results.length} task(s) failed to run and were not scored.`);
}
process.exitCode = exitCodeForResults(results);
}
main().catch((err) => {
// eslint-disable-next-line no-console
console.error(err);
process.exit(1);
});