qwenproxy-cli 1.6.1 → 1.6.3

package/src/utils/overload-detector.ts

@@ -0,0 +1,55 @@
+/**
+ * Upstream Qwen overload response detector.
+ *
+ * When Qwen Web servers are overloaded, the model sometimes generates
+ * an in-band plain text apology (HTTP 200 SSE) instead of a structured error:
+ *   - "Estamos com alta demanda no momento. Tente novamente mais tarde."
+ *   - "We are experiencing high demand right now."
+ *   - "The server is busy, please try again later."
+ *
+ * Detecting these in-band messages before emitting output to the client allows
+ * QwenProxy to quarantine the overloaded account and auto-retry on another
+ * account transparently, so coding agents (Claude Code, OpenCode, Codex)
+ * never receive plain text error apologies as assistant responses.
+ */
+
+const OVERLOAD_PATTERNS = [
+  /high\s+demand/i,
+  /problem\s+connecting/i,
+  /try\s+again\s+later/i,
+  /temporarily\s+unavailable/i,
+  /service\s+(is\s+)?overloaded/i,
+  /too\s+many\s+requests/i,
+  /server\s+is\s+busy/i,
+  /experiencing\s+high\s+(traffic|volume|load)/i,
+  /unable\s+to\s+connect/i,
+  /connection\s+(problem|error|issue)/i,
+  /alta\s+demanda/i,
+  /problema\s+de\s+conex/i,
+  /tente\s+novamente\s+mais\s+tarde/i,
+  /servidor\s+ocupado/i,
+  /indispon[íi]vel\s+no\s+momento/i,
+  /servi[çc]o\s+est[áa]\s+com\s+alta\s+demanda/i,
+];
+
+export const OVERLOAD_COOLDOWN_MS = 5 * 60 * 1000; // 5 minutes
+
+export function isOverloadMessage(content: string | null | undefined): boolean {
+  if (!content) return false;
+  const text = content.trim();
+
+  // Overload apologies are always short (typically 20-150 characters)
+  if (!text || text.length > 500) return false;
+
+  // Never match legitimate code blocks, markdown structures or tool calls
+  if (
+    text.includes("```") ||
+    text.includes("<tool_call") ||
+    text.includes("<qpx_call") ||
+    text.includes("</")
+  ) {
+    return false;
+  }
+
+  return OVERLOAD_PATTERNS.some((pattern) => pattern.test(text));
+}