@mortiseai/stem 0.0.16 → 0.0.17

package/bin/stem-supervisor-lib.mjs

@@ -1,9 +1,9 @@
 // supervisor 逻辑库(2026-08-12 OOM 治理)— bin/stem.mjs 的看护实现。
 // 独立成库的唯一理由:入口脚本 import 即执行,smoke 无法直接测;这里的函数
 // 全部纯逻辑/可注入,scripts/smoke-supervisor.ts 逐项锁语义。
-import { existsSync, mkdirSync, mkdtempSync, readFileSync, readdirSync, rmSync, statSync, writeFileSync } from 'node:fs'
+import { appendFileSync, existsSync, mkdirSync, mkdtempSync, readFileSync, readdirSync, rmSync, statSync, writeFileSync } from 'node:fs'
 import { dirname, join } from 'node:path'
-import { tmpdir } from 'node:os'
+import { homedir, tmpdir } from 'node:os'
 import { spawn, spawnSync } from 'node:child_process'
 
 // 配置类错误的专用退出码(stem-entry-cli 侧同值):见 78 不重启。
@@ -99,7 +99,7 @@ export function injectRestartPrompt(queuePath, isZh, crashInfo) {
  *  与 queue sidecar 同目录,`<sessionId>.crash.json`;crashes 数组保最近 20 条,
  *  下次排查直接看这里而不用翻 supervisor stderr。写失败静默(标记是诊断增强,
  *  不能反过来阻断重启)。 */
-export function writeCrashMarker(queuePath, sessionId, crashInfo, nowMs) {
+export function writeCrashMarker(queuePath, sessionId, crashInfo, nowMs, extras = {}) {
   if (typeof queuePath !== 'string' || !queuePath) return false
   const markerPath = queuePath.endsWith('.queue.json')
     ? queuePath.slice(0, -'.queue.json'.length) + '.crash.json'
@@ -109,11 +109,17 @@ export function writeCrashMarker(queuePath, sessionId, crashInfo, nowMs) {
     const parsed = JSON.parse(readFileSync(markerPath, 'utf8'))
     if (parsed && parsed.v === 1 && Array.isArray(parsed.crashes)) crashes = parsed.crashes
   } catch { /* 无文件/坏文件 → 全新标记 */ }
-  crashes.push({
+  const entry = {
     atMs: nowMs,
     exitCode: crashInfo && typeof crashInfo.code === 'number' ? crashInfo.code : null,
     signal: crashInfo && typeof crashInfo.signal === 'string' ? crashInfo.signal : null,
-  })
+  }
+  // 附加取证字段(2026-08-17 第五次崩溃:原生 SIGABRT 的 stderr 转储全部丢失,
+  // 只有 {exitCode, signal} 无法定因)— 全部可选、additive,v 保持 1,旧读方无感。
+  for (const [k, v] of Object.entries(extras)) {
+    if (v !== undefined && v !== null && v !== '') entry[k] = v
+  }
+  crashes.push(entry)
   if (crashes.length > 20) crashes = crashes.slice(-20)
   try {
     mkdirSync(dirname(markerPath), { recursive: true })
@@ -161,12 +167,81 @@ export function resetTerminal() {
   try { process.stdin.setRawMode?.(false) } catch { /* 非 TTY / 已复位 */ }
 }
 
+/** stderr 尾部环形缓冲(2026-08-17 第五次崩溃:V8 fatal error 的原生转储直写
+ *  fd 2 打上 TUI 后丢失,`FATAL ERROR:` 首行是定因的唯一证据)。Buffer 数组累积,
+ *  超限从头丢弃;read() 返回尾部 utf8 文本。 */
+export function createStderrTail(maxBytes = 64 * 1024) {
+  const chunks = []
+  let total = 0
+  return {
+    push(chunk) {
+      const buf = Buffer.isBuffer(chunk) ? chunk : Buffer.from(String(chunk))
+      chunks.push(buf)
+      total += buf.length
+      while (total > maxBytes && chunks.length > 1) total -= chunks.shift().length
+    },
+    read() {
+      const text = Buffer.concat(chunks).toString('utf8')
+      return text.length > maxBytes ? text.slice(-maxBytes) : text
+    },
+  }
+}
+
+/** 崩溃时刻之后生成的 macOS 原生崩溃报告(ReportCrash 对 SIGABRT/SIGSEGV 必产出,
+ *  含完整原生栈与 abort 原因)。只找不搬:报告数百 KB 且含系统信息。 */
+export function findDiagnosticReport(sinceMs, dir) {
+  let entries
+  try { entries = readdirSync(dir).filter(f => f.endsWith('.ips')) } catch { return null }
+  let best = null
+  for (const f of entries) {
+    try {
+      const t = statSync(join(dir, f)).mtimeMs
+      if (t >= sinceMs && (!best || t > best.t)) best = { f, t }
+    } catch { /* 读取中/权限失败跳过 */ }
+  }
+  return best ? join(dir, best.f) : null
+}
+
+/** 把 .ips 报告路径回填到 crash.json 最新一条崩溃条目(报告由 ReportCrash 异步
+ *  生成,writeCrashMarker 时通常还不存在,退避 sleep 后二次回填)。失败静默。 */
+export function attachDiagnosticReport(queuePath, sessionId, reportPath) {
+  if (typeof queuePath !== 'string' || !queuePath || !reportPath) return false
+  const markerPath = queuePath.endsWith('.queue.json')
+    ? queuePath.slice(0, -'.queue.json'.length) + '.crash.json'
+    : join(dirname(queuePath), `${sessionId || 'session'}.crash.json`)
+  try {
+    const parsed = JSON.parse(readFileSync(markerPath, 'utf8'))
+    if (!parsed || parsed.v !== 1 || !Array.isArray(parsed.crashes) || parsed.crashes.length === 0) return false
+    parsed.crashes[parsed.crashes.length - 1].diagnosticReport = reportPath
+    writeFileSync(markerPath, JSON.stringify(parsed, null, 2) + '\n')
+    return true
+  } catch { return false }
+}
+
 function defaultSpawnChild(args, env) {
-  const child = spawn(process.execPath, args, { stdio: 'inherit', env })
+  // stderr 走 pipe 而非 inherit:原生崩溃转储(V8 fatal / abort)绕过 JS 层拦截
+  // 直写 fd 2,inherit 时直接打上 TUI 且父进程零捕获(2026-08-17 事故)。pipe 后
+  // 实时透传回终端(行为不变)+ 环形缓冲留尾部作取证。TUI 渲染门只看 stdout/stdin
+  // 的 isTTY,stderr 非 TTY 的影响仅限着色(exitCleanup 的 hint 着色已改跟 stdout)。
+  const tail = createStderrTail()
+  const child = spawn(process.execPath, args, { stdio: ['inherit', 'inherit', 'pipe'], env })
+  child.stderr?.on('data', chunk => {
+    tail.push(chunk)
+    try { process.stderr.write(chunk) } catch { /* 终端已关闭等,透传失败无害 */ }
+  })
   return {
     pid: child.pid,
+    stderrTail: () => tail.read(),
     exited: new Promise(resolve => {
-      child.on('exit', (code, signal) => resolve({ code, signal }))
+      // 'close' 优先:exit 触发时 stderr pipe 可能还有未读完的崩溃转储,close 保证
+      // flush 完、stderrTail 不缺尾巴。但 close 会被继承了 stderr fd 的孙进程
+      // (MCP server 等)拖住 — 子进程死了管道还开着,close 迟迟不来(实测孙进程
+      // 活多久拖多久)→ 崩溃检测被无限延迟。所以 exit 后给 2s flush 宽限兜底,
+      // 两者先到先赢(resolve 幂等)。
+      child.on('close', (code, signal) => resolve({ code, signal }))
+      child.on('exit', (code, signal) => {
+        setTimeout(() => resolve({ code, signal }), 2000)
+      })
       child.on('error', () => resolve({ code: 1, signal: null }))
     }),
   }
@@ -188,6 +263,10 @@ export async function runSupervisor(distEntry, argv, options = {}) {
     now = Date.now,
     cwd = process.cwd(),
     announceFile = join(mkdtempSync(join(tmpdir(), 'stem-supervisor-')), 'session.json'),
+    // supervisor 自身诊断的持久化落点(2026-08-17 事故:崩溃信息只写终端,滚屏即丢)。
+    supervisorLogPath = join(cwd, 'logs', 'supervisor.log'),
+    // macOS ReportCrash 报告目录(smoke 可注入假目录)。
+    diagnosticReportsDir = join(homedir(), 'Library', 'Logs', 'DiagnosticReports'),
     resetTerminalFn = resetTerminal,
     onCrashCleanupPid = pid => {
       // Windows 下 best-effort 清理子进程树残留(MCP npx/uvx 僵尸);根进程已死时
@@ -199,6 +278,23 @@ export async function runSupervisor(distEntry, argv, options = {}) {
     installSignalHandlers = true,
   } = options
 
+  // log() 只写终端;logLine() 同时 append 到 supervisor.log(ISO 时间戳)—
+  // 崩溃路径一律用 logLine,子进程日志闸门(STEM_DEBUG)关着时这里仍有据可查。
+  const logFileOnly = text => {
+    if (!supervisorLogPath) return
+    try {
+      mkdirSync(dirname(supervisorLogPath), { recursive: true })
+      appendFileSync(supervisorLogPath, `${new Date(now()).toISOString()} ${text}\n`)
+    } catch { /* 诊断落盘失败不阻断看护 */ }
+  }
+  const logLine = text => {
+    log(text)
+    logFileOnly(text)
+  }
+  // stem 版本一次性读取(dist/cli.mjs → ../package.json),best-effort。
+  let stemVersion = null
+  try { stemVersion = JSON.parse(readFileSync(join(dirname(distEntry), '..', 'package.json'), 'utf8')).version ?? null } catch { /* smoke 假路径等 */ }
+
   let { baseArgs, resumeId } = stripResumeFlag(argv)
   let crashTimes = []
   let childAlive = false
@@ -241,8 +337,18 @@ export async function runSupervisor(distEntry, argv, options = {}) {
     crashTimes = crashTimes.filter(x => t - x < CRASH_WINDOW_MS)
     crashTimes.push(t)
     resetTerminalFn()
+    // stderr 尾部取证:pipe 捕获的原生崩溃转储(FATAL ERROR / abort 栈)。
+    // crash.json 里截 8KB;supervisor.log 无条件记一份 — 即便子进程死在首次
+    // announce 前(crash.json 写不了),这里也有据。
+    const stderrTail = typeof child.stderrTail === 'function' ? child.stderrTail() : ''
+    const stderrTail8k = stderrTail.length > 8 * 1024 ? stderrTail.slice(-8 * 1024) : stderrTail
+    if (stderrTail8k.trim()) {
+      logLine(isZh
+        ? `stem: 崩溃 stderr 尾部(code=${code ?? '-'}, signal=${signal ?? '-'}):\n${stderrTail8k}`
+        : `stem: crash stderr tail (code=${code ?? '-'}, signal=${signal ?? '-'}):\n${stderrTail8k}`)
+    }
     if (crashTimes.length >= CRASH_GIVE_UP_COUNT) {
-      log(isZh
+      logLine(isZh
         ? `stem: 60 秒内连续崩溃 ${crashTimes.length} 次(code=${code ?? '-'}, signal=${signal ?? '-'}),停止自动重启。`
         : `stem: crashed ${crashTimes.length} times within 60s (code=${code ?? '-'}, signal=${signal ?? '-'}); giving up on auto-restart.`)
       return typeof code === 'number' && code !== 0 ? code : 1
@@ -262,12 +368,25 @@ export async function runSupervisor(distEntry, argv, options = {}) {
       }
       for (const p of Array.isArray(announce.jsonlPaths) ? announce.jsonlPaths : []) {
         if (typeof p === 'string' && existsSync(p) && repairJsonlTail(p)) {
-          log(isZh ? `stem: 已修补崩溃截断的会话记录 ${p}` : `stem: repaired crash-truncated session log ${p}`)
+          logLine(isZh ? `stem: 已修补崩溃截断的会话记录 ${p}` : `stem: repaired crash-truncated session log ${p}`)
           break
         }
       }
       if (typeof announce.queuePath === 'string') {
-        writeCrashMarker(announce.queuePath, announce.sessionId, { code, signal }, t)
+        // .tasks.json / .taskstate.json 的 mtime = 子进程最后一次任务落盘时刻,
+        // 用于回答"崩溃前它空闲了多久"(本次事故:空闲 8.5 分钟零遥测)。
+        const sidecarMtime = suffix => {
+          try { return statSync(announce.queuePath.replace(/\.queue\.json$/, suffix)).mtimeMs } catch { return undefined }
+        }
+        writeCrashMarker(announce.queuePath, announce.sessionId, { code, signal }, t, {
+          stderrTail: stderrTail8k,
+          args,
+          nodeVersion: process.version,
+          stemVersion,
+          announceUpdatedAtMs: typeof announce.updatedAtMs === 'number' ? announce.updatedAtMs : undefined,
+          tasksSnapshotMtimeMs: sidecarMtime('.tasks.json'),
+          taskStateMtimeMs: sidecarMtime('.taskstate.json'),
+        })
         injectRestartPrompt(announce.queuePath, isZh, { code, signal })
       }
     }
@@ -275,9 +394,25 @@ export async function runSupervisor(distEntry, argv, options = {}) {
     onCrashCleanupPid(child.pid)
 
     const delayMs = Math.min(1000 * 2 ** (crashTimes.length - 1), RESTART_BACKOFF_MAX_MS)
-    log(isZh
+    logLine(isZh
       ? `stem: 进程异常退出(code=${code ?? '-'}, signal=${signal ?? '-'}),${Math.round(delayMs / 1000)}s 后自动恢复会话${resumeId ? ` ${resumeId}` : ''}…(STEM_NO_SUPERVISOR=1 可关闭看护)`
       : `stem: process crashed (code=${code ?? '-'}, signal=${signal ?? '-'}); resuming session${resumeId ? ` ${resumeId}` : ''} in ${Math.round(delayMs / 1000)}s… (set STEM_NO_SUPERVISOR=1 to disable)`)
     await sleep(delayMs)
+
+    // macOS 信号死亡 → ReportCrash 异步生成 .ips 报告(实测崩溃后 ~5s 才落盘,
+    // 退避窗口等不到):崩溃后 8s 延迟扫描,找到就回填 crash.json 最新条目 +
+    // 落 supervisor.log。只写文件不写终端(此时新子进程 TUI 已接管屏幕);
+    // unref 不阻退出;只记路径不搬内容。
+    if (process.platform === 'darwin' && signal) {
+      const queuePathForAttach = announce && typeof announce.queuePath === 'string' ? announce.queuePath : null
+      const sessionForAttach = announce ? announce.sessionId : null
+      const scanTimer = setTimeout(() => {
+        const report = findDiagnosticReport(t, diagnosticReportsDir)
+        if (!report) return
+        logFileOnly(isZh ? `stem: 原生崩溃报告 ${report}` : `stem: native crash report ${report}`)
+        if (queuePathForAttach) attachDiagnosticReport(queuePathForAttach, sessionForAttach, report)
+      }, 8000)
+      scanTimer.unref?.()
+    }
   }
 }