android2harmony 0.1.3 → 0.1.5

package/tools/autotest/engine/self-test-runner.ts

@@ -1,1024 +0,0 @@
-/**
- * Self-Test Runner — launches HarmonyOS on-device self-tests via the Node.js
- * `@autotest/agent` engine (batch_runner), then reports status from output files.
- *
- *   run     — prepare env (npm install + hdc install), spawn batch detached, return PID.
- *   status  — report state from output files (summary.json / task_results.jsonl) + native liveness.
- *   kill    — terminate the background batch.
- *
- * STDOUT carries strict JSON parsed by the caller; all logging goes to stderr /
- * the log file. JSON shapes + exit codes are a frozen contract (see agents/self-tester.md).
- */
-import crypto from 'node:crypto';
-import { execFileSync, spawn, spawnSync } from 'node:child_process';
-import fs from 'node:fs';
-import path from 'node:path';
-import { fileURLToPath } from 'node:url';
-import { Command } from 'commander';
-import { createLogger, type Logger } from '../runner/logger.ts';
-import { killProcessTree, pidAlive } from '../runner/process-utils.ts';
-import {
-  type CheckResult,
-  checkHap,
-  checkInputs,
-  checkPre,
-} from '../validate.ts';
-
-const HAP_INSTALL_WAIT = 3; // post-install settle, in seconds (reduced from the original 10s).
-const INSTALL_TIMEOUT = 120;
-const DEFAULT_CATEGORY = 'self_test';
-
-let logger: Logger = createLogger(null, 'INFO');
-
-function timestampForFile(d: Date): string {
-  const pad = (n: number) => String(n).padStart(2, '0');
-  return (
-    `${d.getFullYear()}${pad(d.getMonth() + 1)}${pad(d.getDate())}_` +
-    `${pad(d.getHours())}${pad(d.getMinutes())}${pad(d.getSeconds())}`
-  );
-}
-
-/** Synchronous sleep (PowerShell on Windows, `sleep` on POSIX). Best-effort. */
-function sleep(sec: number): void {
-  if (sec <= 0) return;
-  const ms = sec * 1000;
-  try {
-    if (process.platform === 'win32') {
-      spawnSync('powershell.exe', ['-NoProfile', '-Command', `Start-Sleep -Seconds ${sec}`], {
-        stdio: 'ignore',
-        timeout: ms + 5000,
-        windowsHide: true,
-      });
-    } else {
-      spawnSync('sleep', [String(sec)], { stdio: 'ignore', timeout: ms + 5000 });
-    }
-  } catch {
-    // best effort
-  }
-}
-
-function setupLogging(outputDir: string): void {
-  const logPath = path.join(outputDir, `self_test_${timestampForFile(new Date())}.log`);
-  fs.mkdirSync(path.dirname(logPath), { recursive: true });
-  logger = createLogger(fs.createWriteStream(logPath, { encoding: 'utf-8' }), 'DEBUG');
-  logger.info(`日志文件: ${logPath}`);
-}
-
-// ---- external tool discovery ----------------------------------------------
-
-/** Find `name` on PATH via `where` (Windows) or `which` (POSIX). Null on failure. */
-function findOnPath(name: string): string | null {
-  const finder = process.platform === 'win32' ? 'where' : 'which';
-  try {
-    const out = execFileSync(finder, [name], {
-      encoding: 'utf-8',
-      windowsHide: true,
-      timeout: 5000,
-    });
-    const first = out.split(/\r?\n/).find((l) => l.trim());
-    return first ? first.trim() : null;
-  } catch {
-    return null;
-  }
-}
-
-function getHdc(): string {
-  const found = findOnPath('hdc');
-  if (!found) {
-    logger.error('hdc 未找到');
-    process.exit(1);
-  }
-  return found;
-}
-
-/**
- * Absolute path to the compiled `batch-launcher.js` — the detached background
- * process that drives `new AutoTestAgent({...})` programmatically. Sits next to
- * this module under `dist/test-tools/autotest/` (and `src/...` in dev mode via
- * `HOMETRANS_AUTOTEST_DIR`).
- */
-function batchLauncherPath(): string {
-  const here = path.dirname(fileURLToPath(import.meta.url));
-  const p = path.join(here, 'batch-launcher.ts');
-  if (!fs.existsSync(p)) {
-    logger.error(
-      `未找到 batch-launcher: ${p}\n` +
-        '请先在 HomeTrans 根目录执行 `npm run build` 重新构建 dist/。',
-    );
-    process.exit(1);
-  }
-  return p;
-}
-
-/**
- * Model params injected via stdin by the a2h plugin tool (which resolved them
- * via `deps.resolveModelParams`). `provider` is not surfaced by that contract,
- * so it defaults to `openai` (OpenAI-compatible — works with any baseURL the
- * `@ai-sdk/openai` client can target). No `~/.hometrans/config.json` involved.
- */
-interface ModelConfig {
-  apiKey: string;
-  modelName: string;
-  baseURL: string;
-}
-
-/** Optional device/agent overrides passed through from CLI args. */
-interface DeviceOpts {
-  deviceSn?: string;
-  ip?: string;
-  port?: string;
-  agentMode?: string;
-  maxSteps?: string;
-  promptVersion?: string;
-}
-
-/**
- * Read `{ apiKey, modelName, baseURL }` JSON from stdin (fd 0). The plugin
- * tool writes the blob and closes the write end; `readFileSync(0)` returns
- * once EOF is seen. Throws on missing/invalid/empty apiKey.
- */
-function readModelFromStdin(): ModelConfig {
-  let raw = '';
-  try {
-    raw = fs.readFileSync(0, 'utf-8');
-  } catch (e) {
-    throw new Error(`--model-stdin: failed to read stdin: ${(e as Error).message}`);
-  }
-  let parsed: unknown;
-  try {
-    parsed = JSON.parse(raw);
-  } catch (e) {
-    throw new Error(`--model-stdin: stdin is not valid JSON: ${(e as Error).message}`);
-  }
-  const m = (parsed ?? {}) as Partial<ModelConfig>;
-  if (!m.apiKey) {
-    throw new Error('--model-stdin: stdin JSON missing apiKey');
-  }
-  return { apiKey: m.apiKey, modelName: m.modelName ?? '', baseURL: m.baseURL ?? '' };
-}
-
-/**
- * Bridge the stdin-injected model params + CLI device/agent options into the
- * `AUTOTEST_*` env vars the batch launcher reads — these become the
- * `new AutoTestAgent({ apiKey, modelName, … })` constructor args, so the
- * key/model/device flow straight into the agent constructor with no
- * config.yaml file and no `~/.hometrans/config.json`.
- */
-function agentEnvFromModel(model: ModelConfig, dev: DeviceOpts, base: NodeJS.ProcessEnv): NodeJS.ProcessEnv {
-  const env: NodeJS.ProcessEnv = { ...base };
-  env.AUTOTEST_API_KEY = model.apiKey;
-  if (model.modelName) env.AUTOTEST_MODEL = model.modelName;
-  if (model.baseURL) env.AUTOTEST_BASE_URL = model.baseURL;
-  // resolveModelParams returns no provider; OpenAI-compatible default.
-  env.AUTOTEST_PROVIDER = 'openai';
-  if (dev.deviceSn) env.AUTOTEST_DEVICE_SN = dev.deviceSn;
-  if (dev.ip) env.AUTOTEST_IP = dev.ip;
-  if (dev.port) env.AUTOTEST_PORT = dev.port;
-  env.AUTOTEST_MODE = dev.agentMode || 'single';
-  if (dev.maxSteps) env.AUTOTEST_MAX_STEPS = dev.maxSteps;
-  if (dev.promptVersion) env.AUTOTEST_PROMPT_VERSION = dev.promptVersion;
-  return env;
-}
-
-// ---- sync subprocess runner ------------------------------------------------
-
-interface RunCmdResult {
-  returncode: number;
-  stdout: string;
-  stderr: string;
-}
-
-/**
- * Synchronous subprocess runner with logging. On non-zero exit logs
- * stdout/stderr; if `check` (default true), exits with the failing code.
- */
-function runCmd(
-  cmd: string[],
-  opts: { cwd?: string; check?: boolean; timeout?: number; env?: NodeJS.ProcessEnv } = {},
-): RunCmdResult {
-  const { cwd, check = true, timeout, env } = opts;
-  const cmdStr = cmd.join(' ');
-  logger.info(`>>> ${cmdStr}`);
-  const result = spawnSync(cmd[0], cmd.slice(1), {
-    cwd,
-    env,
-    encoding: 'utf-8',
-    timeout: timeout ? timeout * 1000 : undefined,
-    maxBuffer: 1024 * 1024 * 1024,
-    windowsHide: true,
-    stdio: ['ignore', 'pipe', 'pipe'],
-  });
-  if (result.error) {
-    logger.error(`命令异常: ${result.error.message}`);
-    throw result.error;
-  }
-  const status = result.status ?? 0;
-  const stdout = result.stdout ?? '';
-  const stderr = result.stderr ?? '';
-  if (status !== 0) {
-    logger.warning(`退出码=${status}: ${cmdStr}`);
-    if (stdout) logger.warning(`  stdout: ${stdout.trim()}`);
-    if (stderr) logger.warning(`  stderr: ${stderr.trim()}`);
-    if (check) process.exit(status);
-  } else {
-    logger.debug(`成功 (rc=0): ${cmdStr}`);
-  }
-  return { returncode: status, stdout, stderr };
-}
-
-// ---- testcases JSONL normalizer -------------------------------------------
-
-/**
- * Normalize `--testcases` into a JSONL path. A JSON array (legacy
- * `testcases.json`) is converted to JSONL at `<taskSubdir>/testcases.jsonl`,
- * filling missing `uuid` (md5 of case_name or `case_<idx>`, first 8 hex) and
- * `spec` (""). Already-JSONL input is validated line-by-line and passed through.
- */
-export function prepareTestcasesJsonl(inputPath: string, taskSubdir: string): string {
-  if (!fs.existsSync(inputPath)) {
-    logger.error(`测试用例文件不存在: ${inputPath}`);
-    process.exit(1);
-  }
-  let text = fs.readFileSync(inputPath, 'utf-8');
-  if (text.charCodeAt(0) === 0xfeff) text = text.slice(1); // strip BOM
-  const stripped = text.trimStart();
-  if (!stripped) {
-    logger.error(`测试用例文件为空: ${inputPath}`);
-    process.exit(1);
-  }
-
-  if (stripped[0] === '[') {
-    let arr: unknown;
-    try {
-      arr = JSON.parse(text);
-    } catch (e) {
-      logger.error(`测试用例 JSON 数组解析失败 ${inputPath}: ${e instanceof Error ? e.message : e}`);
-      process.exit(1);
-    }
-    if (!Array.isArray(arr)) {
-      logger.error(`测试用例文件根不是数组: ${inputPath}`);
-      process.exit(1);
-    }
-    const outPath = path.join(taskSubdir, 'testcases.jsonl');
-    const lines = arr.map((entry, idx) => {
-      const i = idx + 1;
-      if (!entry || typeof entry !== 'object' || Array.isArray(entry)) {
-        logger.error(`用例 #${i} 不是 JSON 对象: ${JSON.stringify(entry)}`);
-        process.exit(1);
-      }
-      const c = entry as Record<string, unknown>;
-      if (!c.uuid) {
-        const seed = typeof c.case_name === 'string' && c.case_name ? c.case_name : `case_${i}`;
-        c.uuid = crypto.createHash('md5').update(seed, 'utf-8').digest('hex').slice(0, 8);
-      }
-      if (!('spec' in c)) c.spec = '';
-      return JSON.stringify(c);
-    });
-    fs.writeFileSync(outPath, lines.join('\n') + '\n', 'utf-8');
-    logger.info(`已将 JSON 数组转换为 JSONL: ${outPath}（${arr.length} 条用例）`);
-    return outPath;
-  }
-
-  const lines = text.split(/\r?\n/).filter((l) => l.trim());
-  lines.forEach((line, i) => {
-    try {
-      JSON.parse(line);
-    } catch (e) {
-      logger.error(`JSONL 第 ${i + 1} 行解析失败 ${inputPath}: ${e instanceof Error ? e.message : e}`);
-      process.exit(1);
-    }
-  });
-  logger.info(`测试用例文件已为 JSONL，原样透传: ${inputPath}（${lines.length} 条用例）`);
-  return inputPath;
-}
-
-// ---- HAP install -----------------------------------------------------------
-
-/**
- * Install every `.hap`/`.hsp` from a comma-separated `--hap` arg (each entry a
- * file or a directory) in one `hdc install -r` transaction. Real apps ship an
- * entry HAP plus HSPs, so at least one `.hap` is required. Success is judged by
- * hdc's stdout heuristic — hdc can return 0 on failure.
- */
-function installHaps(hapArg: string, hdc: string): void {
-  const entries = hapArg
-    .split(',')
-    .map((s) => s.trim())
-    .filter((s) => s);
-  if (entries.length === 0) {
-    logger.error('hap-path 为空');
-    process.exit(1);
-  }
-  const pkgs: string[] = [];
-  for (const e of entries) {
-    if (!fs.existsSync(e)) {
-      logger.error(`hap 路径不存在: ${e}`);
-      process.exit(1);
-    }
-    const stat = fs.statSync(e);
-    const lower = e.toLowerCase();
-    if (stat.isDirectory()) {
-      const found = fs
-        .readdirSync(e)
-        .filter((f) => f.toLowerCase().endsWith('.hap') || f.toLowerCase().endsWith('.hsp'))
-        .sort()
-        .map((f) => path.join(e, f));
-      if (found.length === 0) {
-        logger.error(`目录内无 .hap/.hsp 包: ${e}`);
-        process.exit(1);
-      }
-      pkgs.push(...found);
-    } else if (stat.isFile() && (lower.endsWith('.hap') || lower.endsWith('.hsp'))) {
-      pkgs.push(e);
-    } else {
-      logger.error(`既不是目录也不是 .hap/.hsp 文件: ${e}`);
-      process.exit(1);
-    }
-  }
-  // Resolve to native absolute paths (backslashes on Windows) AND dedup in one
-  // pass. Resolving is required, not cosmetic: hdc on Windows mis-parses a
-  // forward-slash *drive-absolute* path (e.g. `C:/a/b.hap`) as relative and
-  // prepends its own cwd (`C:\cwd\C:/a/b.hap` → invalid), silently skipping the
-  // install after Step 2a already uninstalled the app. `path.resolve` yields the
-  // native `C:\a\b.hap` form hdc accepts. On POSIX it's a harmless no-op for
-  // already-absolute inputs. Callers assemble --hap from multiple sources
-  // (package-set dir, comma-list, build-fixer's hsp/ copy) that can overlap, so
-  // dedup on the same resolved key.
-  const seen = new Set<string>();
-  const deduped: string[] = [];
-  for (const p of pkgs) {
-    const resolved = path.resolve(p);
-    if (seen.has(resolved)) continue;
-    seen.add(resolved);
-    deduped.push(resolved);
-  }
-  if (!deduped.some((p) => p.toLowerCase().endsWith('.hap'))) {
-    logger.error(`缺少 entry HAP（至少需要一个 .hap）: ${hapArg}`);
-    process.exit(1);
-  }
-
-  // hdc install -r accepts multiple package paths in one transaction.
-  logger.info(`安装 ${deduped.length} 个包: ${deduped.map((p) => path.basename(p)).join(', ')}`);
-  const result = runCmd([hdc, 'install', '-r', ...deduped], {
-    timeout: INSTALL_TIMEOUT,
-    check: false,
-  });
-  const combined = `${(result.stdout || '').trim()}\n${(result.stderr || '').trim()}`;
-  const combinedLower = combined.toLowerCase();
-  // msg:error is case-sensitive; the rest match on lowercased output.
-  const hasError =
-    combined.includes('msg:error') ||
-    combinedLower.includes('failed to install') ||
-    (combined.includes('code:') && combinedLower.includes('error:'));
-  const hasSuccess = combinedLower.includes('successfully');
-  if (result.returncode === 0 && !hasError && hasSuccess) {
-    logger.info('安装成功');
-    sleep(HAP_INSTALL_WAIT);
-  } else if (result.returncode === 0 && !hasError) {
-    logger.warning(`安装结果不确定（rc=0 但无 success 消息）: ${combined}`);
-    sleep(HAP_INSTALL_WAIT);
-  } else {
-    logger.error(`安装失败: rc=${result.returncode}, output=${combined}`);
-    process.exit(1);
-  }
-}
-
-// ---- run subcommand --------------------------------------------------------
-
-export interface RunArgs {
-  testcases: string;
-  hap: string;
-  bundleName: string;
-  taskDir: string;
-  outputDir?: string;
-  category: string;
-  modelStdin?: boolean;
-  deviceSn?: string;
-  ip?: string;
-  port?: string;
-  agentMode?: string;
-  maxSteps?: string;
-  promptVersion?: string;
-  timeout?: number | 'auto';
-}
-
-// Per-case outer poll budget when `--timeout auto` is used: mirrors the
-// batch-launcher hard cap CASE_TIMEOUT_MS=10min + 20% margin = 720s/case. The
-// CLI already parses `--testcases`, so it derives the total itself instead of
-// making the caller compute `caseCount * 720` in a (shell-specific) wrapper.
-const PER_CASE_TIMEOUT_SEC = 720;
-
-export interface SpawnBatchOpts {
-  cmd: string[];
-  cwd: string;
-  env: NodeJS.ProcessEnv;
-  stdoutLog: string;
-  pidFile: string;
-  taskDir: string;
-  taskSubdir: string;
-}
-
-/**
- * Spawn the batch-launcher detached and write pidFile. Resolves with the PID
- * on the spawn event, rejects with an Error on the spawn-error event (ENOENT,
- * EACCES, …). Caller is responsible for emitting JSON (RUNNING or FAILED) —
- * this function is silent on stdout so the same code path can serve both the
- * backward-compat "emit RUNNING immediately" mode and the `--timeout`
- * "suppress RUNNING, poll until terminal" mode. `spawnFn` is injectable for tests.
- */
-export function spawnBatch(opts: SpawnBatchOpts, spawnFn: typeof spawn = spawn): Promise<number> {
-  return new Promise((resolve, reject) => {
-    const { cmd, cwd, env, stdoutLog, pidFile, taskDir, taskSubdir } = opts;
-    const logFd = fs.openSync(stdoutLog, 'w');
-    let settled = false;
-    const fail = (error: string): void => {
-      if (settled) return;
-      settled = true;
-      try { fs.closeSync(logFd); } catch { /* ignore */ }
-      try { fs.rmSync(pidFile, { force: true }); } catch { /* ignore */ }
-      logger.error(`后台 batch 进程启动失败: ${error}`);
-      reject(new Error(error));
-    };
-    const proc = spawnFn(cmd[0], cmd.slice(1), {
-      cwd,
-      stdio: ['ignore', logFd, logFd],
-      env,
-      detached: true,
-      windowsHide: true,
-    });
-    // spawn() reports failure asynchronously via the 'error' event and success
-    // via the 'spawn' event. proc.pid is set synchronously even for some
-    // failures, so emitting RUNNING off the pid could hand the caller a
-    // live-looking but dead PID. Gate both outcomes on the events (with a
-    // `settled` latch) so exactly one of resolve/reject fires.
-    proc.on('error', (e: Error) => fail(e.message));
-    proc.on('spawn', () => {
-      if (settled) return;
-      settled = true;
-      fs.writeFileSync(pidFile, String(proc.pid), 'utf-8');
-      proc.unref();
-      logger.info(`后台进程 PID=${proc.pid}`);
-      resolve(proc.pid);
-    });
-  });
-}
-
-/**
- * Poll `probeStatus` every 60 seconds until a terminal status is reached
- * (COMPLETED / CRASHED / NOT_STARTED) or `timeoutSec` elapses. On timeout,
- * auto-kill the batch tree via `killProcessTree` and return a `TIMEOUT`
- * result with exit code 5 (new). The first probe happens immediately (no
- * leading sleep) so a fast-finished batch is detected without a 60s delay.
- *
- * Returns `{ json, exitCode }` — caller emits `JSON.stringify(json)` to stdout
- * and `process.exit(exitCode)`.
- *
- * Exit codes: COMPLETED=0, CRASHED=3, NOT_STARTED=4, TIMEOUT=5.
- */
-export function pollUntilTerminal(
-  taskDir: string,
-  outputDir: string,
-  timeoutSec: number,
-): { json: Record<string, unknown>; exitCode: number } {
-  const start = Date.now();
-
-  while (true) {
-    const r = probeStatus(taskDir, outputDir);
-
-    if (r.status === 'COMPLETED') {
-      return {
-        json: {
-          status: 'COMPLETED',
-          pid: r.pid,
-          task_subdir: r.taskSubdir,
-          cases_done: r.casesDone,
-          pass_count: r.passCount ?? 0,
-          fail_count: r.failCount ?? 0,
-          unknown_count: r.unknownCount ?? 0,
-          pass_rate: r.passRate ?? 0,
-          log_tail: r.logTail,
-        },
-        exitCode: 0,
-      };
-    }
-
-    if (r.status === 'CRASHED') {
-      return {
-        json: {
-          status: 'CRASHED',
-          pid: r.pid,
-          task_subdir: r.taskSubdir,
-          cases_done: r.casesDone,
-          last_case: r.lastCase,
-          log_tail: r.logTail,
-        },
-        exitCode: 3,
-      };
-    }
-
-    if (r.status === 'NOT_STARTED') {
-      return {
-        json: { status: 'NOT_STARTED' },
-        exitCode: 4,
-      };
-    }
-
-    // r.status === 'RUNNING' — check timeout before sleeping
-    const elapsed = (Date.now() - start) / 1000;
-    if (elapsed >= timeoutSec) {
-      // Auto-kill the batch tree, then return TIMEOUT.
-      if (r.pid != null) {
-        try {
-          killProcessTree(r.pid);
-          sleep(2);
-        } catch {
-          // best effort
-        }
-      }
-      return {
-        json: {
-          status: 'TIMEOUT',
-          pid: r.pid,
-          task_subdir: r.taskSubdir,
-          cases_done: r.casesDone,
-          last_case: r.lastCase,
-          log_tail: r.logTail,
-        },
-        exitCode: 5,
-      };
-    }
-
-    // Still RUNNING and within budget — sleep 60s then probe again.
-    sleep(60);
-  }
-}
-
-/** `run` subcommand: prepare environment, then spawn the batch-launcher detached. */
-export async function cmdRun(args: RunArgs): Promise<void> {
-  let taskDir: string;
-  let outputDir: string;
-  let taskSubdir: string;
-  let pidFile: string;
-  let stdoutLog: string;
-  let batchEnv: NodeJS.ProcessEnv;
-  let cmd: string[];
-  let effectiveTimeout: number | undefined;
-  try {
-    taskDir = path.resolve(args.taskDir);
-    fs.mkdirSync(taskDir, { recursive: true });
-    outputDir = args.outputDir ? path.resolve(args.outputDir) : taskDir;
-    fs.mkdirSync(outputDir, { recursive: true });
-
-    setupLogging(outputDir);
-    logger.info(`Node: ${process.execPath}`);
-    logger.info(
-      `参数: testcases=${args.testcases}, hap=${args.hap}, bundle=${args.bundleName}, ` +
-        `task_dir=${taskDir}, category=${args.category}, model_stdin=${args.modelStdin ? 'yes' : 'no'}` +
-        (args.deviceSn ? `, device_sn=${args.deviceSn}` : '') +
-        (args.agentMode ? `, agent_mode=${args.agentMode}` : '') +
-        (args.timeout === 'auto' ? `, timeout=auto` : args.timeout ? `, timeout=${args.timeout}s` : ''),
-    );
-
-    // Per-run task subdir; status subcommand globs `task_*` to find the latest.
-    taskSubdir = path.join(taskDir, `task_${timestampForFile(new Date())}`);
-    fs.mkdirSync(taskSubdir, { recursive: true });
-
-    pidFile = path.join(taskDir, 'batch.pid');
-    stdoutLog = path.join(taskDir, 'batch_stdout.log');
-
-    // Kill any previous run's batch still recorded alive in the pid file.
-    if (fs.existsSync(pidFile)) {
-      try {
-        const oldPid = parseInt(fs.readFileSync(pidFile, 'utf-8').trim(), 10);
-        if (Number.isFinite(oldPid) && pidAlive(oldPid)) {
-          logger.warning(`杀掉上次残留进程 PID=${oldPid}`);
-          killProcessTree(oldPid);
-          sleep(2);
-        }
-      } catch {
-        // best effort
-      }
-    }
-
-    // a2h: model params come from stdin (--model-stdin, piped by the plugin
-    // tool which resolved them via deps.resolveModelParams). No
-    // ~/.hometrans/config.json and no HOMETRANS_MODEL_* env fallback.
-    if (!args.modelStdin) {
-      throw new Error(
-        'a2h selftest run requires --model-stdin: the plugin tool pipes ' +
-        '{apiKey, modelName, baseURL} via stdin. No config-file fallback.',
-      );
-    }
-    const model = readModelFromStdin();
-    logger.info('已从 stdin 读取模型参数（apiKey/model/baseURL 将通过 new AutoTestAgent 构造器传入）');
-    const devOpts: DeviceOpts = {
-      deviceSn: args.deviceSn,
-      ip: args.ip,
-      port: args.port,
-      agentMode: args.agentMode,
-      maxSteps: args.maxSteps,
-      promptVersion: args.promptVersion,
-    };
-
-    const hdc = getHdc();
-
-    logger.info(`=== Step 2a: uninstall ${args.bundleName} ===`);
-    runCmd([hdc, 'uninstall', args.bundleName], { timeout: INSTALL_TIMEOUT, check: false });
-
-    logger.info(`=== Step 2b: install ${args.hap} ===`);
-    installHaps(args.hap, hdc);
-
-    const testcasesJsonl = prepareTestcasesJsonl(path.resolve(args.testcases), taskSubdir);
-
-    // Resolve `--timeout auto` now that the testcases are parsed: total budget =
-    // caseCount × 720s (min 1 case). A numeric --timeout passes through; omitted
-    // stays undefined → backward-compat immediate-RUNNING mode below.
-    if (args.timeout === 'auto') {
-      const caseCount = fs.readFileSync(testcasesJsonl, 'utf-8').split(/\r?\n/).filter((l) => l.trim()).length;
-      effectiveTimeout = Math.max(caseCount, 1) * PER_CASE_TIMEOUT_SEC;
-      logger.info(`--timeout auto: ${caseCount} 条用例 × ${PER_CASE_TIMEOUT_SEC}s = ${effectiveTimeout}s`);
-    } else {
-      effectiveTimeout = args.timeout;
-    }
-
-    // Node engine: `node <batch-launcher.js>` constructs `new AutoTestAgent(...)`
-    // from AUTOTEST_* env and runs the testcases batch in-process. cwd=taskSubdir
-    // keeps the per-case report dirs and task_results.jsonl inside the subdir.
-    cmd = [
-      process.execPath, batchLauncherPath(),
-      '--task-dir', taskSubdir,
-      '--testcases', testcasesJsonl,
-      '--category', args.category,
-    ];
-    batchEnv = agentEnvFromModel(model, devOpts, process.env);
-  } catch (e) {
-    const error = e instanceof Error ? e.message : String(e);
-    logger.error(`异常:\n${e instanceof Error ? e.stack ?? error : String(e)}`);
-    console.log(JSON.stringify({ status: 'FAILED', error }));
-    process.exit(1);
-  }
-
-  logger.info('=== Step 3: 后台启动 AutoTest (Node @autotest/agent, new AutoTestAgent) ===');
-
-  // No --timeout: backward-compat mode — emit RUNNING JSON immediately on successful spawn.
-  try {
-    // ponytail: the non-null assertions are safe because the prep try/catch above
-    // either assigned all these locals or process.exit(1) — no path reaches here
-    // with an unassigned variable. If the catch ever stops exiting, drop the `!`.
-    const pid = await spawnBatch({ cmd: cmd!, cwd: taskSubdir!, env: batchEnv!, stdoutLog, pidFile, taskDir: taskDir!, taskSubdir: taskSubdir! });
-    if (effectiveTimeout) {
-      const r = pollUntilTerminal(taskDir!, outputDir!, effectiveTimeout);
-      console.log(JSON.stringify(r.json));
-      process.exit(r.exitCode);
-    } else {
-      console.log(
-        JSON.stringify({
-          status: 'RUNNING',
-          pid,
-          pid_file: pidFile,
-          stdout_log: stdoutLog,
-          task_dir: taskDir,
-          task_subdir: taskSubdir,
-        }),
-      );
-    }
-  } catch (e) {
-    const error = e instanceof Error ? e.message : String(e);
-    console.log(JSON.stringify({ status: 'FAILED', error }));
-    process.exit(1);
-  }
-}
-
-export interface ProbeResult {
-  status: 'NOT_STARTED' | 'COMPLETED' | 'RUNNING' | 'CRASHED';
-  pid: number | null;
-  taskSubdir: string | null;
-  casesDone: number;
-  lastCase: string;
-  logTail: string;
-  // COMPLETED-only (undefined for other statuses):
-  passCount?: number;
-  failCount?: number;
-  unknownCount?: number;
-  passRate?: string | number;
-}
-
-/**
- * Read batch state from on-disk artifacts (no JSON emission, no exit).
- * Same logic as `cmdStatus` but returns a structured `ProbeResult` so callers
- * (`cmdStatus` for one-shot, `pollUntilTerminal` for the `run --timeout` loop)
- * can share one source of truth.
- */
-export function probeStatus(taskDir: string, outputDir: string): ProbeResult {
-  const pidFile = path.join(taskDir, 'batch.pid');
-  if (!fs.existsSync(pidFile)) {
-    return {
-      status: 'NOT_STARTED',
-      pid: null,
-      taskSubdir: null,
-      casesDone: 0,
-      lastCase: '',
-      logTail: '',
-    };
-  }
-
-  const pid = parseInt(fs.readFileSync(pidFile, 'utf-8').trim(), 10);
-  const taskSubdir = findTaskSubdir(taskDir);
-
-  let casesDone = 0;
-  let lastCase = '';
-  const jsonlFile = taskSubdir ? path.join(taskSubdir, 'task_results.jsonl') : null;
-  if (jsonlFile && fs.existsSync(jsonlFile)) {
-    const text = fs.readFileSync(jsonlFile, 'utf-8').trim();
-    const lines = text ? text.split(/\r?\n/) : [];
-    casesDone = lines.length;
-    if (lines.length) {
-      try {
-        const last = JSON.parse(lines[lines.length - 1]) as { case_name?: string };
-        if (typeof last.case_name === 'string') lastCase = last.case_name;
-      } catch {
-        // ignore
-      }
-    }
-  }
-
-  let logTail = '';
-  const logFile = findLatestLog(outputDir);
-  if (logFile && fs.existsSync(logFile)) {
-    const text = fs.readFileSync(logFile, 'utf-8').trim();
-    logTail = (text ? text.split(/\r?\n/) : []).slice(-5).join('\n');
-  }
-
-  const summaryFile = taskSubdir ? path.join(taskSubdir, 'summary.json') : null;
-  if (summaryFile && fs.existsSync(summaryFile)) {
-    let summary: Record<string, unknown> = {};
-    try {
-      summary = JSON.parse(fs.readFileSync(summaryFile, 'utf-8'));
-    } catch {
-      // ignore
-    }
-    const num = (v: unknown, fallback: number) => (typeof v === 'number' ? v : fallback);
-    // ponytail: cast `pass_rate` (unknown from JSON) to the documented schema
-    // (string like "30.00%" or absent). `num(...)` would zero a string rate, so
-    // we keep the `?? 0` fallback verbatim from the original cmdStatus.
-    return {
-      status: 'COMPLETED',
-      pid,
-      taskSubdir,
-      casesDone: num(summary.total_cases, casesDone),
-      lastCase,
-      passCount: num(summary.pass_count, num(summary.success_count, 0)),
-      failCount: num(summary.fail_count, 0),
-      unknownCount: num(summary.unknown_count, 0),
-      passRate: (summary.pass_rate as string | number | undefined) ?? 0,
-      logTail,
-    };
-  }
-
-  const alive = pidAlive(pid);
-  return {
-    status: alive ? 'RUNNING' : 'CRASHED',
-    pid,
-    taskSubdir,
-    casesDone,
-    lastCase,
-    logTail,
-  };
-}
-
-// ---- status subcommand -----------------------------------------------------
-
-export interface StatusArgs {
-  taskDir: string;
-  outputDir?: string;
-}
-
-/** Latest `task_*` subdir in `taskDir` (name-sorted desc), or null. */
-export function findTaskSubdir(taskDir: string): string | null {
-  if (!fs.existsSync(taskDir)) return null;
-  let entries: fs.Dirent[];
-  try {
-    entries = fs.readdirSync(taskDir, { withFileTypes: true });
-  } catch {
-    return null;
-  }
-  const candidates = entries
-    .filter((e) => e.isDirectory() && e.name.startsWith('task_'))
-    .map((e) => e.name)
-    .sort()
-    .reverse();
-  return candidates.length ? path.join(taskDir, candidates[0]) : null;
-}
-
-/** Latest `self_test_*.log` (desc), falling back to `batch_stdout.log`, or null. */
-export function findLatestLog(outputDir: string): string | null {
-  if (!fs.existsSync(outputDir)) return null;
-  let entries: fs.Dirent[];
-  try {
-    entries = fs.readdirSync(outputDir, { withFileTypes: true });
-  } catch {
-    return null;
-  }
-  const selfTest = entries
-    .filter((e) => e.isFile() && e.name.startsWith('self_test_') && e.name.endsWith('.log'))
-    .map((e) => e.name)
-    .sort()
-    .reverse();
-  const batchLog = entries
-    .filter((e) => e.isFile() && e.name === 'batch_stdout.log')
-    .map((e) => e.name);
-  const ordered = [...selfTest, ...batchLog];
-  return ordered.length ? path.join(outputDir, ordered[0]) : null;
-}
-
-/**
- * `status` subcommand: report state from output files.
- *   summary.json present            → COMPLETED (exit 0)
- *   else pid alive                  → RUNNING   (exit 2)
- *   else                            → CRASHED   (exit 3)
- * No pid file at all → NOT_STARTED (exit 4).
- */
-export function cmdStatus(args: StatusArgs): void {
-  const taskDir = path.resolve(args.taskDir);
-  const outputDir = args.outputDir ? path.resolve(args.outputDir) : taskDir;
-
-  const r = probeStatus(taskDir, outputDir);
-
-  if (r.status === 'NOT_STARTED') {
-    console.log(JSON.stringify({ status: 'NOT_STARTED' }));
-    process.exit(4);
-  }
-
-  if (r.status === 'COMPLETED') {
-    console.log(
-      JSON.stringify({
-        status: 'COMPLETED',
-        pid: r.pid,
-        task_subdir: r.taskSubdir,
-        cases_done: r.casesDone,
-        pass_count: r.passCount ?? 0,
-        fail_count: r.failCount ?? 0,
-        unknown_count: r.unknownCount ?? 0,
-        pass_rate: r.passRate ?? 0,
-        log_tail: r.logTail,
-      }),
-    );
-    process.exit(0);
-  }
-
-  // RUNNING or CRASHED
-  console.log(
-    JSON.stringify({
-      status: r.status,
-      pid: r.pid,
-      cases_done: r.casesDone,
-      last_case: r.lastCase,
-      task_subdir: r.taskSubdir,
-      log_tail: r.logTail,
-    }),
-  );
-  process.exit(r.status === 'RUNNING' ? 2 : 3);
-}
-
-// ---- kill subcommand -------------------------------------------------------
-
-export interface KillArgs {
-  taskDir: string;
-}
-
-/** `kill` subcommand: terminate the batch-launcher background process by pid. */
-export function cmdKill(args: KillArgs): void {
-  const taskDir = path.resolve(args.taskDir);
-  const pidFile = path.join(taskDir, 'batch.pid');
-  if (!fs.existsSync(pidFile)) {
-    console.log(JSON.stringify({ status: 'NO_PID_FILE', killed: false }));
-    return;
-  }
-  let pid: number;
-  try {
-    pid = parseInt(fs.readFileSync(pidFile, 'utf-8').trim(), 10);
-  } catch {
-    console.log(JSON.stringify({ status: 'NO_PID_FILE', killed: false }));
-    return;
-  }
-  if (!pidAlive(pid)) {
-    console.log(JSON.stringify({ status: 'ALREADY_DEAD', pid, killed: false }));
-    return;
-  }
-  killProcessTree(pid);
-  console.log(JSON.stringify({ status: 'KILLED', pid, killed: true }));
-}
-
-// ---- CLI entry -------------------------------------------------------------
-
-/** Wire up `run` / `status` / `kill` subcommands via commander. */
-export function main(): void {
-  const program = new Command();
-  program.name('self-test-runner').description('自测执行器 - 统一的进程管理和测试执行');
-
-  program
-    .command('run')
-    .description('准备环境并执行用例')
-    .requiredOption('--testcases <path>', '测试用例文件，JSON 数组或 JSONL')
-    .requiredOption('--hap <path>', '签名包：单个 .hap/.hsp 文件、目录，或逗号分隔的多个文件/目录')
-    .requiredOption('--bundle-name <name>', '包名')
-    .option('--task-dir <dir>', '任务目录（内部会创建 task_<timestamp> 子目录）', 'task')
-    .option('--output-dir <dir>', '日志目录')
-    .option('--category <name>', '测试分类名称，传给 batch-launcher', DEFAULT_CATEGORY)
-    .option('--model-stdin', '从 stdin 读取 {apiKey, modelName, baseURL}（a2h 插件工具注入；无 config 回退）')
-    .option('--device-sn <sn>', '设备序列号（传给 AutoTestAgent.deviceSn）')
-    .option('--ip <ip>', '设备 IP（传给 AutoTestAgent.ip）')
-    .option('--port <port>', '设备端口（传给 AutoTestAgent.port）')
-    .option('--agent-mode <mode>', 'agent 模式（默认 single）', 'single')
-    .option('--max-steps <n>', '单用例最大步数')
-    .option('--prompt-version <ver>', 'prompt 版本（v1.0/v2.0/v3.0，默认 v2.0）', 'v2.0')
-    .option('--timeout <sec>', '阻塞直到终态或超时（秒），或 `auto` = 按用例数 × 720s 自动推导。不传则立即返回 RUNNING JSON。超时后自动 kill batch 并返回 TIMEOUT JSON（exit 5）', v => {
-      if (v === 'auto') return 'auto' as const;
-      const n = Number(v);
-      if (!Number.isInteger(n) || n <= 0) {
-        throw new Error(`--timeout 必须是正整数或 'auto'，收到: ${v}`);
-      }
-      return n;
-    })
-    .action(async (opts: RunArgs) => {
-      await cmdRun({
-        testcases: opts.testcases,
-        hap: opts.hap,
-        bundleName: opts.bundleName,
-        taskDir: opts.taskDir,
-        outputDir: opts.outputDir,
-        category: opts.category,
-        modelStdin: opts.modelStdin,
-        deviceSn: opts.deviceSn,
-        ip: opts.ip,
-        port: opts.port,
-        agentMode: opts.agentMode,
-        maxSteps: opts.maxSteps,
-        promptVersion: opts.promptVersion,
-        timeout: opts.timeout,
-      });
-    });
-
-  program
-    .command('status')
-    .description('查询执行状态')
-    .requiredOption('--task-dir <dir>')
-    .option('--output-dir <dir>')
-    .action((opts: StatusArgs) => {
-      cmdStatus({ taskDir: opts.taskDir, outputDir: opts.outputDir });
-    });
-
-  program
-    .command('kill')
-    .description('终止进程')
-    .requiredOption('--task-dir <dir>')
-    .action((opts: KillArgs) => {
-      cmdKill({ taskDir: opts.taskDir });
-    });
-
-  // ---- Validation gates (replace the self-tester.md `uv run python -c` lines) --
-  // Each prints its message to stdout on pass, or stderr + exit(1) on fail —
-  // matching the prompt's `… && echo OK` expectation. These are NOT part of the
-  // run/status/kill stdout-JSON contract; they are plain pass/fail checks.
-  const gate = (fn: () => CheckResult): void => {
-    const r = fn();
-    if (r.ok) {
-      console.log(r.message);
-    } else {
-      console.error(r.message);
-      process.exit(1);
-    }
-  };
-
-  program
-    .command('check-hap')
-    .description('校验 hap_path：各项存在且至少含一个入口 .hap')
-    .requiredOption('--hap <path>', '签名包：文件、目录，或逗号分隔的多个文件/目录')
-    .action((opts: { hap: string }) => gate(() => checkHap(opts.hap)));
-
-  program
-    .command('check-inputs')
-    .description('校验 testcases.json（非空数组）与 app-metadata.json（必需字段）可解析')
-    .requiredOption('--testcases <path>')
-    .requiredOption('--metadata <path>')
-    .action((opts: { testcases: string; metadata: string }) =>
-      gate(() => checkInputs(opts.testcases, opts.metadata)),
-    );
-
-  program
-    .command('check-pre')
-    .description('校验 testcases.json 中 [PRE] 用例从索引 0 起连续')
-    .requiredOption('--testcases <path>')
-    .action((opts: { testcases: string }) => gate(() => checkPre(opts.testcases)));
-
-  program.parseAsync(process.argv).catch((err) => {
-    console.error(err instanceof Error ? err.message : err);
-    process.exit(1);
-  });
-}
-
-const __filename = fileURLToPath(import.meta.url);
-const isMainEntry = !!process.argv[1] && path.resolve(process.argv[1]) === __filename;
-if (isMainEntry) {
-  main();
-}