maxpool 1.5.27 → 1.5.29

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "maxpool",
-  "version": "1.5.27",
+  "version": "1.5.29",
   "description": "Multi-account Claude Code proxy with adaptive, rate-aware load balancing across Claude accounts",
   "type": "module",
   "main": "src/index.js",

package/src/event-log.js

@@ -99,6 +99,19 @@ export async function rotateIfNeeded() {
  * output (stdout in plain mode; the TUI separately re-points console at its ring
  * buffer, whose _addLog also calls appendEventLog, so TUI-mode lines persist too).
  */
+let stdoutSuppressed = false;
+
+/**
+ * Suppress the console mirror's STDOUT passthrough (the event-log append still
+ * runs). Used during a seamless TUI reload so a worker that doesn't currently own
+ * the terminal — the new worker before its takeover, or the old worker draining
+ * after it handed the TTY off — can't paint plain log lines over the live TUI of
+ * the worker that DOES own it. Per-process (each worker has its own module state),
+ * so it never affects the other worker. Read live inside the mirror, so lifting it
+ * takes effect immediately even for a console wrapper captured earlier.
+ */
+export function setConsoleStdoutSuppressed(v) { stdoutSuppressed = !!v; }
+
 export function installConsoleMirror() {
   if (origConsole) return; // idempotent
   origConsole = { log: console.log.bind(console), error: console.error.bind(console) };
@@ -106,7 +119,7 @@ export function installConsoleMirror() {
     const orig = origConsole[method];
     console[method] = (...args) => {
       try { appendEventLog(args.map(a => (typeof a === 'string' ? a : String(a))).join(' ')); } catch { /* never */ }
-      orig(...args);
+      if (!stdoutSuppressed) orig(...args);
     };
   }
 }
@@ -118,5 +131,5 @@ export const flushEventLog = () => writeChain;
 export function __resetEventLogForTest() {
   if (origConsole) { console.log = origConsole.log; console.error = origConsole.error; origConsole = null; }
   if (rotationTimer) { clearInterval(rotationTimer); rotationTimer = null; }
-  logPath = null; manageRotation = false; writeChain = Promise.resolve(); queued = 0; dropped = 0;
+  logPath = null; manageRotation = false; writeChain = Promise.resolve(); queued = 0; dropped = 0; stdoutSuppressed = false;
 }

package/src/index.js

@@ -3,7 +3,7 @@
 import { spawn, spawnSync } from 'node:child_process';
 import { createInterface } from 'node:readline';
 import { loadOrCreateConfig, loadConfig, saveConfig, atomicConfigUpdate, getConfigPath, loadState, saveState, getStatePath, getLogPath, readGeneration, flushConfigWrites, flushStateWrites } from './config.js';
-import { setEventLogPath, installConsoleMirror } from './event-log.js';
+import { setEventLogPath, installConsoleMirror, setConsoleStdoutSuppressed } from './event-log.js';
 import { SleepGuard } from './sleep-guard.js';
 import { AccountManager } from './account-manager.js';
 import { createProxyServer } from './server.js';
@@ -23,6 +23,10 @@ import {
 const args = process.argv.slice(2);
 const command = args[0];
 const SERVER_RESTART_EXIT_CODE = 75;
+// If a seamless reload doesn't release us (→ take over) within this window, the new
+// worker rolled back — self-heal admission so we don't 503 forever. > the baton
+// readiness timeout (10s) + margin.
+const RELOAD_ROLLBACK_SELFHEAL_MS = Math.max(15_000, Number(process.env.MAXPOOL_RELOAD_SELFHEAL_MS) || 30_000);
 const SERVER_WORKER_ENV = 'MAXPOOL_SERVER_WORKER';
 // Set by the supervisor when it spawns a worker for a seamless reload: that
 // worker boots HEADLESS (plain logs, no writer lease) and waits for the baton.
@@ -48,12 +52,16 @@ function initEventLog(config, { manageRotation = false } = {}) {
   } catch { /* logging must never block startup */ }
 }
 
-// Which reload path an in-process reload request takes. Interactive (a live TUI)
-// reloads via a FULL cold restart so the fresh worker re-renders the TUI — a
-// seamless reload worker boots headless and would strand the user in raw logs.
-// Headless/service keeps the zero-downtime baton (nothing visual to lose).
-function reloadStrategy({ supervised, useTUI }) {
-  return supervised && !useTUI ? 'seamless' : 'cold-restart';
+// Which reload path an in-process reload request takes. ALL supervised reloads —
+// TUI and headless alike — use the zero-downtime baton so the listening socket
+// never closes (no ECONNREFUSED for the ~5 other sessions routing through us). A
+// TUI reload additionally hands the terminal off worker→worker (the new worker
+// re-renders the TUI at takeover; the old worker's terminalHandedOff guard stops
+// it clobbering the shared TTY on exit). Only an UNsupervised process cold-restarts.
+// The user escape hatch MAXPOOL_TUI_COLD_RESTART=1 forces the old cold path (see
+// requestReload) if the TTY handoff ever misbehaves on their terminal.
+function reloadStrategy({ supervised }) {
+  return supervised ? 'seamless' : 'cold-restart';
 }
 
 switch (command) {
@@ -657,20 +665,39 @@ async function serverWorkerCommand() {
 
   const port = config.proxy.port;
   const host = config.proxy.host || '127.0.0.1';
-  // A headless reload worker NEVER drives the TUI (single-owner terminal); it
-  // takes the TUI only on baton takeover. Cold/direct workers use it if on a TTY.
-  const useTUI = process.stdout.isTTY && process.stdin.isTTY && !isReloadWorker;
+  // A reload worker constructs a TUI when on a TTY but only START()s it on baton
+  // takeover (becomePrimary) — single-owner terminal, enforced by the baton order
+  // (old worker's tui.stop() precedes MSG_RELEASED precedes the new worker's
+  // MSG_TAKEOVER→tui.start()). It boots headless-silent until then (stdout muzzled
+  // just below), so its boot logs can't paint over the old worker's live TUI.
+  const useTUI = process.stdout.isTTY && process.stdin.isTTY;
+
+  // A reload worker with a TUI must not write plain logs to the shared terminal
+  // (still owned by the old worker's TUI) while booting toward takeover; suppress
+  // stdout (event-log append still runs). Lifted in becomePrimary at takeover.
+  if (isReloadWorker && useTUI) setConsoleStdoutSuppressed(true);
 
   let tui = null;
   let server = null;
   let syncTimer = null;
   let draining = false;
   let restartController = null;
+  // Set once this worker hands the terminal to a new worker during a seamless TUI
+  // reload; makes this (now non-owner) worker's exit-path restoreTerminal a NO-OP
+  // so it can't flip the shared TTY's raw-mode/alt-screen back and clobber the new
+  // worker's live TUI. Only the current terminal owner (or the supervisor) restores.
+  let terminalHandedOff = false;
+  // Self-heal timer for a seamless reload that rolls back (new worker fails to boot,
+  // we're never released). Cleared on MSG_RELEASE; fires cancelRestart otherwise.
+  let reloadWatchdog = null;
 
   // Best-effort terminal restore on ANY abnormal exit path (uncaughtException,
   // a bare process.exit, a crash) so the user's shell is never left in raw mode
   // or the alt-screen. Idempotent and safe even when no TUI was running.
   const restoreTerminal = () => {
+    // Handed the TTY to a new worker (seamless TUI reload) → the new worker owns
+    // it now; restoring here would clobber its live TUI. No-op.
+    if (terminalHandedOff) return;
     try {
       if (tui?.running) { tui.stop(); return; }
       if (process.stdout.isTTY) process.stdout.write('\x1b[?25h\x1b[?1049l');
@@ -783,14 +810,30 @@ async function serverWorkerCommand() {
   // (not supervised, no IPC), fall back to the abrupt restart.
   const requestReload = () => {
     if (draining) return;
-    // Interactive (live TUI) → full cold restart so the fresh worker re-renders the
-    // TUI; headless/service → zero-downtime seamless baton (nothing visual to lose).
-    // Test-only: force the interactive cold-restart path headless (a pty can't be
-    // allocated in-suite), so the exit-75 respawn is exercised without a real TUI.
-    const forceCold = process.env.MAXPOOL_TEST_FORCE_COLD_RESTART === '1';
-    if (!forceCold && reloadStrategy({ supervised, useTUI }) === 'seamless') {
+    // Supervised → zero-downtime seamless baton (socket never closes; TUI hands the
+    // terminal off worker→worker). Unsupervised → cold restart.
+    // Escape hatches force the cold path: MAXPOOL_TEST_FORCE_COLD_RESTART (in-suite,
+    // no pty) and the USER-facing MAXPOOL_TUI_COLD_RESTART=1 (revert to the old
+    // brief-ECONNREFUSED cold restart if the seamless TTY handoff ever misbehaves).
+    const forceCold = process.env.MAXPOOL_TEST_FORCE_COLD_RESTART === '1'
+      || process.env.MAXPOOL_TUI_COLD_RESTART === '1';
+    if (!forceCold && reloadStrategy({ supervised }) === 'seamless') {
       try {
         process.send({ type: MSG_RELOAD_REQUEST });
+        // Arm the rollback self-heal: restartController already latched
+        // pending/restarting + paused admission (in _restart, before we got here).
+        // On a SUCCESSFUL reload the supervisor sends MSG_RELEASE (→ releaseBaton­
+        // AndDrain clears this) then we exit. On a ROLLBACK we're never released and
+        // never exit — so if this fires, resume serving instead of 503-ing forever.
+        if (reloadWatchdog) clearTimeout(reloadWatchdog);
+        reloadWatchdog = setTimeout(() => {
+          reloadWatchdog = null;
+          if (draining) return; // MSG_RELEASE already arrived → a real reload, not a rollback
+          if (restartController?.cancelRestart()) {
+            console.log('[Maxpool] Reload rolled back (new worker never took over) — resumed serving.');
+          }
+        }, RELOAD_ROLLBACK_SELFHEAL_MS);
+        reloadWatchdog.unref?.();
         return;
       } catch { /* IPC gone — fall through to abrupt restart */ }
     }
@@ -799,6 +842,7 @@ async function serverWorkerCommand() {
 
   restartController = new RestartController({
     pauseAdmission: () => accountManager.setAdmissionPaused(true),
+    resumeAdmission: () => accountManager.setAdmissionPaused(false),
     restartNow: requestReload,
     // Configurable so ops can tune the bounded pre-restart drain, and so the
     // integration test can exercise the force-restart path without a 10s wait.
@@ -954,6 +998,10 @@ async function serverWorkerCommand() {
   // baton (a reload) — freeze the update check so a reload doesn't re-probe npm
   // (only a cold supervisor start self-updates).
   const becomePrimary = async ({ viaTakeover }) => {
+    // We're taking the terminal now (or are the cold-start primary). Lift the
+    // reload-worker boot muzzle so the TUI's first render + any plain-mode fallback
+    // logs reach stdout. (No-op for a worker that never suppressed.)
+    setConsoleStdoutSuppressed(false);
     if (tui) {
       if (tui.start()) {
         console.log(`Listening on ${host}:${port} with ${accounts.length} account(s)`);
@@ -1018,6 +1066,9 @@ async function serverWorkerCommand() {
     const releaseBatonAndDrain = async () => {
       if (draining) { try { process.send({ type: MSG_RELEASED }); } catch { /* ignore */ } return; }
       draining = true;
+      // MSG_RELEASE arrived → the new worker booted OK and this reload is really
+      // proceeding (not a rollback). Disarm the rollback self-heal.
+      if (reloadWatchdog) { clearTimeout(reloadWatchdog); reloadWatchdog = null; }
       if (syncTimer) clearInterval(syncTimer);
       // Stop accepting NEW connections; KEEP in-flight requests alive.
       server.maxpoolBeginDrain?.();
@@ -1041,6 +1092,13 @@ async function serverWorkerCommand() {
       await flushStateWrites();
 
       if (tui?.running) tui.stop(); // restore terminal before the new worker takes it
+      // The terminal is now HANDED OFF: the incoming worker will take the TTY at
+      // MSG_TAKEOVER. Arm the guard UNCONDITIONALLY (even if the TUI was already
+      // stopped) so this worker's later exit(0) restoreTerminal can't clobber the
+      // new worker's TUI. Also muzzle our own remaining drain-time logs (in-flight
+      // requests completing) so they don't paint the new worker's alt-screen.
+      terminalHandedOff = true;
+      setConsoleStdoutSuppressed(true);
       try { process.send({ type: MSG_RELEASED }); } catch { /* ignore */ }
 
       // Drain bounded in-flight on EXISTING access tokens (no refresh needed),
@@ -1569,6 +1627,11 @@ Options:
   --log-to DIR        Log full requests/responses to DIR (server, one file per request)
   --with-key          Include proxy API key in maxpool env output
 
+Env:
+  MAXPOOL_TUI_COLD_RESTART=1   Reload (r) via a full cold restart instead of the
+                               zero-downtime seamless handoff — a fallback if the
+                               terminal ever misbehaves after a reload.
+
 Config: ${getConfigPath()}
 `);
 }

package/src/restart-controller.js

@@ -3,6 +3,7 @@ const NON_UPSTREAM_ROUTES = new Set(['(queued)', '(none available)']);
 export class RestartController {
   constructor({
     pauseAdmission,
+    resumeAdmission = () => {},
     restartNow,
     log = console.log,
     // Bounded drain: wait at most this long for in-flight upstream requests to
@@ -15,6 +16,7 @@ export class RestartController {
     clearTimeoutFn = clearTimeout,
   }) {
     this.pauseAdmission = pauseAdmission;
+    this.resumeAdmission = resumeAdmission;
     this.restartNow = restartNow;
     this.log = log;
     this.drainTimeoutMs = drainTimeoutMs;
@@ -33,6 +35,23 @@ export class RestartController {
     return true;
   }
 
+  /**
+   * Abort a restart that was requested but never completed because the process is
+   * STILL ALIVE — a seamless reload that rolled back (the new worker failed to boot,
+   * so the old worker was never released and never exited). Without this the latched
+   * `pending`/`restarting` flags make requestStarted() return false forever → the
+   * worker 503s EVERY request until a manual restart. Reset the flags + resume
+   * admission so it keeps serving. No-op if no restart is in progress.
+   */
+  cancelRestart() {
+    if (!this.pending && !this.restarting) return false;
+    this.pending = false;
+    this.restarting = false;
+    if (this._drainTimer) { this.clearTimeoutFn(this._drainTimer); this._drainTimer = null; }
+    this.resumeAdmission();
+    return true;
+  }
+
   requestRouted(id, account) {
     if (!this.activeRequests.has(id)) return false;
     if (NON_UPSTREAM_ROUTES.has(account)) this.upstreamRequests.delete(id);

package/src/server.js

@@ -24,6 +24,12 @@ const DEFAULT_RETRY = {
 const UPSTREAM_TTFB_MS = Math.max(5_000, Number(process.env.MAXPOOL_TTFB_MS) || 120_000);        // headers must arrive within this
 const STREAM_IDLE_MS = Math.max(30_000, Number(process.env.MAXPOOL_STREAM_IDLE_MS) || 300_000);  // max gap BETWEEN streamed chunks (reset per chunk)
 const UPSTREAM_BODY_MS = Math.max(30_000, Number(process.env.MAXPOOL_BODY_MS) || 300_000);       // non-streaming body read
+const CLIENT_DRAIN_MS = Math.max(5_000, Number(process.env.MAXPOOL_DRAIN_MS) || 60_000);         // max wait for a backpressured client to drain (half-open client → free the lease)
+// Backstop reaper idle ceiling. Floored WELL above the longest a legit NON-streaming
+// request can go silent (no heartbeat there): TTFB 120s + nonStreamMaxWaitMs 300s +
+// UPSTREAM_BODY_MS 300s ≈ 720s — so a slow-but-alive non-streaming request is never
+// reaped at completion. Streaming holds heartbeat, so they're safe at any ceiling.
+const REQUEST_IDLE_MAX_MS = Math.max(900_000, Number(process.env.MAXPOOL_REQUEST_IDLE_MAX_MS) || 1_200_000);
 
 const DEFAULT_QUEUE = {
   enabled: true,
@@ -167,32 +173,43 @@ export function createProxyServer(accountManager, config, hooks = {}) {
         return;
       }
 
-      // Buffer request body (needed for retry on 429)
-      const bodyChunks = [];
-      for await (const chunk of req) {
-        bodyChunks.push(chunk);
-      }
-      const body = Buffer.concat(bodyChunks);
-      const retryConfig = { ...DEFAULT_RETRY, ...(config.retry || {}) };
-      const queueConfig = { ...DEFAULT_QUEUE, ...(config.queue || {}) };
-      const canRetryBufferedBody = body.length <= retryConfig.maxRetryBufferBytes;
-      const requestInfo = describeRequest(req, body);
-      const maxQueuedBodyBytes = queueConfig.maxQueuedBodyBytes == null
-        ? Infinity
-        : Math.max(0, Number(queueConfig.maxQueuedBodyBytes) || 0);
-      const canQueueBufferedBody = body.length <= maxQueuedBodyBytes;
-      if (!canQueueBufferedBody) {
-        requestInfo.queueBlockedReason = `request body ${body.length} bytes exceeds queue.maxQueuedBodyBytes ${maxQueuedBodyBytes}`;
-      }
-      requestInfo.profile = getMaxpoolProfile(req.headers);
-      requestInfo.sessionKey = headerValue(req.headers, 'x-maxpool-session');
-      if (requestInfo.requiresAnthropicThinkingIntegrity && requestInfo.profile === 'all') {
-        console.log('[Maxpool] Anthropic thinking detected; provider fallback disabled for this session/request');
-      }
-      prepareRuntimeProviders(accountManager, req.headers);
-
+      // The request is now TRACKED (onRequestStart accepted → tui.active +
+      // restartController.activeRequests both hold it). EVERYTHING that can throw —
+      // including body buffering — must run inside the try whose finally fires
+      // onRequestEnd, or a client that dies mid-body-read throws to the OUTER catch
+      // and leaks the active entry AND the restart drain counter (hangs a reload's
+      // drain). requestInfo starts as {} so the error path never dereferences an
+      // undefined (a body-read throw lands before describeRequest runs).
       const ctx = { account: null, status: null };
+      let requestInfo = {};
+      // Backstop reaper (defense-in-depth for any UNKNOWN future hang the specific
+      // TTFB/idle/body/drain guards don't cover). See startIdleRequestReaper.
+      const reaperTimer = startIdleRequestReaper(res, reqId, REQUEST_IDLE_MAX_MS);
       try {
+        // Buffer request body (needed for retry on 429)
+        const bodyChunks = [];
+        for await (const chunk of req) {
+          bodyChunks.push(chunk);
+        }
+        const body = Buffer.concat(bodyChunks);
+        const retryConfig = { ...DEFAULT_RETRY, ...(config.retry || {}) };
+        const queueConfig = { ...DEFAULT_QUEUE, ...(config.queue || {}) };
+        const canRetryBufferedBody = body.length <= retryConfig.maxRetryBufferBytes;
+        requestInfo = describeRequest(req, body);
+        const maxQueuedBodyBytes = queueConfig.maxQueuedBodyBytes == null
+          ? Infinity
+          : Math.max(0, Number(queueConfig.maxQueuedBodyBytes) || 0);
+        const canQueueBufferedBody = body.length <= maxQueuedBodyBytes;
+        if (!canQueueBufferedBody) {
+          requestInfo.queueBlockedReason = `request body ${body.length} bytes exceeds queue.maxQueuedBodyBytes ${maxQueuedBodyBytes}`;
+        }
+        requestInfo.profile = getMaxpoolProfile(req.headers);
+        requestInfo.sessionKey = headerValue(req.headers, 'x-maxpool-session');
+        if (requestInfo.requiresAnthropicThinkingIntegrity && requestInfo.profile === 'all') {
+          console.log('[Maxpool] Anthropic thinking detected; provider fallback disabled for this session/request');
+        }
+        prepareRuntimeProviders(accountManager, req.headers);
+
         await forwardRequest(
           req, res, body, accountManager, upstream, 0, hooks, reqId, ctx, logDir,
           retryConfig, queueConfig, requestInfo, canRetryBufferedBody, canQueueBufferedBody, new Set(),
@@ -200,11 +217,16 @@ export function createProxyServer(accountManager, config, hooks = {}) {
       } catch (err) {
         ctx.status = ctx.status || 502;
         console.error('[Maxpool] Unhandled error:', err);
-        sendErrorResponse(res, requestInfo, 502, {
-          type: 'error',
-          error: { type: 'proxy_error', message: 'Internal proxy error' },
-        });
+        // Only attempt an error body if the socket is still writable — a client
+        // that aborted during body-read leaves res destroyed; writing then throws.
+        if (!res.destroyed && !res.writableEnded && !res.headersSent) {
+          sendErrorResponse(res, requestInfo, 502, {
+            type: 'error',
+            error: { type: 'proxy_error', message: 'Internal proxy error' },
+          });
+        }
       } finally {
+        clearInterval(reaperTimer);
         hooks.onRequestEnd?.(reqId, {
           method: req.method, path: req.url,
           account: ctx.account, status: ctx.status,
@@ -1096,7 +1118,7 @@ function unavailableMessage(accountManager, requestInfo = {}, retryAfter, willRe
   return `All ${n} accounts exhausted. Retry in ${retryAfter}s.`;
 }
 
-export const __serverTest = { unavailableMessage, computeQueueWindowMs, isRetriableUpstreamStatus, headerValue, getMaxpoolProfile, ensureQueueHeartbeat, clearQueueHeartbeat, describeRequest, classifyRateLimit, detectTranscriptOrigin, isAnthropicIncompatBody, streamResponse };
+export const __serverTest = { unavailableMessage, computeQueueWindowMs, isRetriableUpstreamStatus, headerValue, getMaxpoolProfile, ensureQueueHeartbeat, clearQueueHeartbeat, describeRequest, classifyRateLimit, detectTranscriptOrigin, isAnthropicIncompatBody, streamResponse, startIdleRequestReaper };
 
 async function readErrorBody(upstreamRes, limitBytes = 64 * 1024) {
   if (!upstreamRes.body) return '';
@@ -1747,6 +1769,30 @@ function mappedModel(originalModel, account) {
 /**
  * Stream an SSE response to the client, parsing usage data along the way.
  */
+/**
+ * Backstop reaper for a single request: force-abort it if it makes ZERO client-write
+ * progress for idleMs. IDLE-keyed via res.socket.bytesWritten (a legit streaming HOLD
+ * heartbeats → bytesWritten advances → never trips; only a genuinely stuck request
+ * with no writes is reaped). res.destroy() routes into the normal onClose→finally
+ * cleanup so the account lease + onRequestEnd release — never a direct releaseAccount
+ * (which has no idempotency guard). Clock/timer injectable for tests. Returns the
+ * interval handle (caller clearInterval()s it in the request's finally).
+ */
+function startIdleRequestReaper(res, reqId, idleMs, { now = Date.now, setIntervalFn = setInterval } = {}) {
+  let lastBytes = res.socket?.bytesWritten ?? 0;
+  let lastProgressAt = now();
+  const timer = setIntervalFn(() => {
+    const bytes = res.socket?.bytesWritten ?? lastBytes;
+    if (bytes !== lastBytes) { lastBytes = bytes; lastProgressAt = now(); return; }
+    if (now() - lastProgressAt >= idleMs && !res.writableEnded && !res.destroyed) {
+      console.error(`[Maxpool] Request ${reqId} — no write progress for ${Math.round(idleMs / 1000)}s (backstop reaper); force-aborting a stuck request to free its account slot`);
+      res.destroy();
+    }
+  }, Math.min(60_000, idleMs));
+  timer.unref?.();
+  return timer;
+}
+
 async function streamResponse(webStream, res, status, responseHeaders, accountIndex, accountManager, streamLog, requestInfo = {}, idleMs = STREAM_IDLE_MS) {
   const reader = webStream.getReader();
   const decoder = new TextDecoder();
@@ -1822,14 +1868,37 @@ async function streamResponse(webStream, res, status, responseHeaders, accountIn
         parseSSEEvent(event, accountIndex, accountManager, requestInfo);
       }
 
-      // Handle backpressure — also bail out if client disconnects,
-      // because 'drain' will never fire on a destroyed socket
+      // Handle backpressure — bail out if the client disconnects OR goes silently
+      // half-open. A vanished peer (laptop sleep / Wi-Fi drop / lost TCP FIN) leaves
+      // the send buffer full so this branch is entered, but sends no FIN ('close'
+      // never fires) and never ACKs ('drain' never fires) — so a bare await here
+      // hangs FOREVER, pinning the account lease and never firing onRequestEnd (the
+      // phantom-"N active" leak). Bound it: if the client can't drain a small SSE
+      // chunk within CLIENT_DRAIN_MS it's gone → break (same disposition as a
+      // destroyed socket: the finally runs reader.cancel()+res.end(), the lease
+      // frees with success:true, onRequestEnd fires). A legit slow client drains
+      // each episode well within this window; the timer is per-episode, never summed.
       if (!ok) {
-        await new Promise(resolve => {
-          res.once('drain', resolve);
-          res.once('close', resolve);
+        let drainTimer, onDrain, onClose2;
+        const settled = new Promise(resolve => {
+          onDrain = () => resolve(true);
+          onClose2 = () => resolve(true);
+          res.once('drain', onDrain);
+          res.once('close', onClose2);
+          // Bound the wait by the drain cap, but never longer than this stream's
+          // idle bound (keeps it injectable for tests; prod = min(300s, 60s) = 60s).
+          drainTimer = setTimeout(() => resolve(false), Math.min(idleMs, CLIENT_DRAIN_MS));
+          drainTimer.unref?.();
         });
-        if (res.destroyed) break;
+        let drained;
+        try {
+          drained = await settled;
+        } finally {
+          clearTimeout(drainTimer);
+          res.off('drain', onDrain);
+          res.off('close', onClose2);
+        }
+        if (!drained || res.destroyed) break; // client gone/stalled → stop, clean up in finally
       }
     }
 

package/src/tui.js

@@ -1,5 +1,5 @@
 import { createInterface } from 'node:readline';
-import { fetchProfile, loginOAuth } from './oauth.js';
+import { fetchProfile, loginOAuth, tokenFingerprint } from './oauth.js';
 import { appendEventLog } from './event-log.js';
 
 // ── ANSI helpers ─────────────────────────────────────────────
@@ -398,7 +398,12 @@ export class TUI {
       this._confirm(
         'Restart Maxpool?',
         'Pause new requests, drain active work, then start the updated server.',
-        () => { this.stop(); this.onRestart?.(); },
+        // Do NOT stop the TUI here — let the reload path own it: cold restart stops
+        // it in restartWorkerNow, a seamless reload in releaseBatonAndDrain (on
+        // MSG_RELEASE). If the reload ROLLS BACK (new worker fails to boot), neither
+        // fires, so the TUI keeps rendering instead of being stranded in plain-log
+        // mode. (The worker also self-heals admission — see index.js reload watchdog.)
+        () => { this.onRestart?.(); },
       );
     } else if (k === 'a') {
       this.mode = 'accounts';
@@ -725,6 +730,13 @@ export class TUI {
         throw error;
       }
       this._addLog(`Updated account "${name}"`);
+      // Confirm the fresh token actually reached DISK (saveConfig awaited above +
+      // the AM was updated before it, so disk now holds creds.refreshToken). This
+      // is the verifiable proof a re-auth "stuck" — its fp must match the on-disk
+      // token; a later "REJECTED sent fp=<other>" then means an upstream revocation,
+      // not a lost persist. console.log (not _addLog) so it's VISIBLE during the
+      // stopped-TUI login flow, mirroring the sibling "Persisted rotated token" line.
+      console.log(`[Maxpool] Re-authenticated "${name}" — fresh token persisted (fp=${tokenFingerprint(creds.refreshToken)})`);
       return { updated: true, name };
     } else {
       this.config.accounts.push(entry);
@@ -948,12 +960,11 @@ export class TUI {
       lines.push('');
       // Aligned column header (dim + underline so it reads as chrome, not data). It
       // labels the fixed columns that have no inline label; the Ses/Wk/Tok/Req and
-      // Now/15m/1h labels stay inline per row. A short glossary below expands the
-      // abbreviations the header + inline labels can't spell out.
+      // Now/15m/1h labels stay inline per row. It sits DIRECTLY above the rows — the
+      // abbreviation glossary is a FOOTER below the rows (a caption), never between
+      // the header and the data (a left-aligned sentence there reads as a broken,
+      // misaligned second header).
       lines.push(dimUnderline(acctHeader(W)));
-      if (W >= 88) {
-        lines.push(' ' + dim('Ses 5h · Wk 7d · Now in-flight (weight) · 15m/1h served (avg · f fails)'));
-      }
       const showBoth = W >= 70;
       const bw = showBoth
         ? Math.max(5, Math.min(20, Math.floor((W - 56) / 2)))
@@ -962,6 +973,11 @@ export class TUI {
       for (let i = 0; i < this.am.accounts.length; i++) {
         lines.push(this._renderAcct(i, bw, showBoth));
       }
+      // Glossary FOOTER (expands the abbreviations the header + inline labels can't
+      // spell out). Below the rows so it never breaks the header↔column alignment.
+      if (W >= 88) {
+        lines.push(' ' + dim('Ses 5h · Wk 7d · Now in-flight (weight) · 15m/1h served (avg · f fails)'));
+      }
     }
 
     // ── Activity header