maxpool 1.24.16 → 1.24.18

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "maxpool",
-  "version": "1.24.16",
+  "version": "1.24.18",
   "description": "Multi-account Claude Code proxy with adaptive, rate-aware load balancing across Claude accounts",
   "type": "module",
   "main": "src/index.js",

package/src/account-manager.js

@@ -4375,8 +4375,12 @@ export class AccountManager {
     const configTokens = new Set(
       entries.map(e => e.token).filter(Boolean),
     );
+    // Same rule for a provider KIND config already covers: a header-derived fallback
+    // of that kind is a duplicate login even when its token differs (a pre-rotation key).
+    const configKinds = new Set(entries.filter(e => e.token).map(e => e.provider).filter(Boolean));
     for (const a of [...this.accounts]) {
-      if (!a.configSourced && a.type === 'provider' && a.credential && configTokens.has(a.credential)) {
+      if (!a.configSourced && a.type === 'provider' && a.credential
+          && (configTokens.has(a.credential) || (a.runtime && configKinds.has(a.provider)))) {
         const idx = this.accounts.indexOf(a);
         if (idx >= 0 && this.accounts[idx].inFlight === 0) this.removeAccount(idx);
       }

package/src/server.js

@@ -163,7 +163,14 @@ export function createProxyServer(accountManager, config, hooks = {}) {
   // upstream-bound request headers elsewhere.
   const workerStamp = String(process.pid);
 
-  const server = http.createServer(async (req, res) => {
+  // REQUEST-ARRIVAL TIMEOUTS (2026-10-11 outage). Node's defaults (300s request
+  // timeout, 30s checking interval) are what fired live — but they must be in the
+  // constructor to be settable at all (property assignment after listen is a no-op
+  // in Node 20, measured), so they are read here, env-tunable, never mutated later.
+  const server = http.createServer({
+    requestTimeout: Math.max(1_000, Number(process.env.MAXPOOL_REQUEST_TIMEOUT_MS) || 300_000),
+    connectionsCheckingInterval: Math.max(100, Number(process.env.MAXPOOL_CONNECTIONS_CHECK_MS) || 30_000),
+  }, async (req, res) => {
     try {
       // Disable Nagle on the client socket: flush each SSE event immediately instead
       // of coalescing small writes into bursts. Smooths streaming cadence toward the
@@ -343,6 +350,28 @@ export function createProxyServer(accountManager, config, hooks = {}) {
   // this releasing worker.
   server.maxpoolBeginDrain = () => { draining = true; };
 
+  // STALLED-UPLOAD ANSWER (2026-10-11 outage). When a request body never finishes
+  // arriving, Node's requestTimeout fires and its default handler writes a BODY-LESS
+  // `408 Request Timeout`, which Claude Code surfaces as the opaque "API Error: 408
+  // status code (no body)" (1,711 of them overnight, while the Mac's network stack
+  // was starved). Answer with the standard Anthropic error envelope instead, naming
+  // the actual cause, so the user sees a real message and the CLI treats it like any
+  // other retryable API error. Every other client error keeps Node's own handling.
+  server.on('clientError', (err, socket) => {
+    if (err?.code === 'ERR_HTTP_REQUEST_TIMEOUT' && socket.writable && !socket._httpMessage?.headersSent) {
+      const body = JSON.stringify({
+        type: 'error',
+        error: {
+          type: 'request_timeout',
+          message: 'Maxpool did not receive the full request in time (the upload stalled — usually a network problem on this machine). Retrying should work once the network recovers.',
+        },
+      });
+      socket.end(`HTTP/1.1 408 Request Timeout\r\nContent-Type: application/json\r\nContent-Length: ${Buffer.byteLength(body)}\r\nConnection: close\r\n\r\n${body}`);
+      return;
+    }
+    socket.destroy(err);
+  });
+
   return server;
 }
 
@@ -3015,9 +3044,29 @@ function prepareRuntimeProviders(accountManager, headers) {
       .filter(a => a.configSourced && a.credential)
       .map(a => a.credential),
   );
+  // THE PROXY OWNS THE KEYS (2026-10-10). When config already manages accounts of a
+  // provider kind, a session-supplied key for that kind is ignored outright. Dedup by
+  // token alone was not enough: a session launched before a key rotation keeps sending
+  // the OLD key for its whole life, and it surfaced as a duplicate "glm-fallback"
+  // account for the same z.ai login (1,578 requests on the old key in 2 days).
+  // A kind counts as proxy-managed only while at least one of its config accounts
+  // HOLDS a key: if every one is unresolved (the secret-unresolved outages of
+  // 2026-10-05/08) the session key still stands in as the safety net.
+  const configKinds = new Set(
+    (accountManager.accounts || []).filter(a => a.configSourced && a.credential).map(a => a.provider),
+  );
+  // Self-clean: a fallback created before the config keys loaded (or restored from
+  // state) is dropped as soon as it is idle — no restart, no waiting for old sessions.
+  // (removeAccount itself refuses an account with a request in flight.)
+  for (const a of [...(accountManager.accounts || [])]) {
+    if (a.runtime && !a.configSourced && a.type === 'provider' && configKinds.has(a.provider)
+        && accountManager.removeAccount(accountManager.accounts.indexOf(a))) {
+      console.log(`[Maxpool] Dropped session-supplied "${a.name}" — ${a.provider} keys are managed by the pool config`);
+    }
+  }
 
   const zaiToken = headerValue(headers, 'x-maxpool-zai-token');
-  if (zaiToken && !configTokens.has(zaiToken)) {
+  if (zaiToken && !configKinds.has('zai') && !configTokens.has(zaiToken)) {
     const opus = headerValue(headers, 'x-maxpool-zai-opus-model') || headerValue(headers, 'x-maxpool-zai-model') || 'glm-5.3';
     const sonnet = headerValue(headers, 'x-maxpool-zai-sonnet-model') || headerValue(headers, 'x-maxpool-zai-model') || opus;
     const haiku = headerValue(headers, 'x-maxpool-zai-haiku-model') || 'glm-5.3';
@@ -3036,7 +3085,7 @@ function prepareRuntimeProviders(accountManager, headers) {
   }
 
   const kimiToken = headerValue(headers, 'x-maxpool-kimi-token');
-  if (kimiToken && !configTokens.has(kimiToken)) {
+  if (kimiToken && !configKinds.has('kimi') && !configTokens.has(kimiToken)) {
     // Fallback only — `cc all` always sends x-maxpool-kimi-model from the llm_config SSOT,
     // so this is what a bare/older client gets. Kept current deliberately: it read
     // 'kimi-k2.7' while the fleet had moved to k3.
@@ -3652,3 +3701,6 @@ function computeRetryAfter(accountManager, requestInfo = {}) {
   const ms = accountManager.nextRetryForRequest?.(requestInfo, new Set())?.retryAfterMs ?? Infinity;
   return ms === Infinity ? 60 : Math.max(1, Math.ceil(ms / 1000));
 }
+
+// Test seam: prepareRuntimeProviders with a plain headers object.
+export const __test = { prepareRuntimeProviders };