qwenproxy-cli 1.2.1 → 1.2.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "qwenproxy-cli",
-  "version": "1.2.1",
+  "version": "1.2.2",
   "description": "High-performance OpenAI & Anthropic compatible API gateway for Qwen with multi-account rotation, interactive TUI, and resilient tool calling.",
   "main": "src/index.ts",
   "bin": {

package/src/routes/anthropic/index.ts

@@ -105,6 +105,9 @@ app.post("/v1/messages", async (c) => {
           "Content-Type": "application/json",
           Authorization: `Bearer ${process.env.API_KEY || config.apiKey || ""}`,
           "x-qwenproxy-route": "Anthropic",
+          ...(c.req.header("x-qwenproxy-chat-mode")
+            ? { "x-qwenproxy-chat-mode": c.req.header("x-qwenproxy-chat-mode")! }
+            : {}),
         },
         body: JSON.stringify({
           ...openaiRequest,

package/src/routes/chat/context.ts

@@ -87,16 +87,17 @@ export async function buildFinalContext(
     ? false
     : useThreadNative && (hasExplicitConversationKey || !isNewSession); // has assistant messages = continuation of existing chat
 
-  // Compute sessionId: only generate a persistent session ID when we have
-  // an explicit conversation key. Otherwise, generate an ephemeral ID for
-  // logging/metrics only (not used for thread reuse). Temp mode never persists
-  // a thread, so it has no session id.
+  // Compute sessionId: deterministic session ID derived from the conversation
+  // key (or "implicit-thread") + system instructions + first user message.
+  // Including completeInstructions in ALL modes (not just explicit keys)
+  // prevents cross-project session collisions when different projects share
+  // the same first user message but have different system prompts / tools.
   const sessionId = isStateless
     ? null
     : (conversationKey || useThreadNative)
       ? deriveSessionId(
           messages,
-          conversationKey ? completeInstructions : "",
+          completeInstructions,
           conversationKey ?? "implicit-thread",
         )
       : null;

package/src/routes/responses/index.ts

@@ -136,6 +136,9 @@ app.post("/v1/responses", async (c) => {
                   "Content-Type": "application/json",
                   Authorization: `Bearer ${process.env.API_KEY || config.apiKey || ""}`,
                   "x-qwenproxy-route": "Responses",
+                  ...(c.req.header("x-qwenproxy-chat-mode")
+                    ? { "x-qwenproxy-chat-mode": c.req.header("x-qwenproxy-chat-mode")! }
+                    : {}),
                 },
                 body: JSON.stringify({
                   ...chatRequest,
@@ -300,6 +303,9 @@ app.post("/v1/responses", async (c) => {
             "Content-Type": "application/json",
             Authorization: `Bearer ${process.env.API_KEY || config.apiKey || ""}`,
             "x-qwenproxy-route": "Responses",
+            ...(c.req.header("x-qwenproxy-chat-mode")
+              ? { "x-qwenproxy-chat-mode": c.req.header("x-qwenproxy-chat-mode")! }
+              : {}),
           },
           body: JSON.stringify(chatRequest),
         },

package/src/services/playwright.ts

@@ -3353,6 +3353,10 @@ export function isPlaywrightInitialized(accountId: string): boolean {
   return accountPages.has(accountId);
 }
 
+export function isPlaywrightInitializing(accountId: string): boolean {
+  return inFlightAccountInits.has(accountId);
+}
+
 export function isAccountRecentlyActive(
   accountId: string,
   maxIdleMs = 300_000,

package/src/services/qwen.ts

@@ -27,7 +27,7 @@ import {
   syncModelMetadata,
 } from "../core/model-registry.ts";
 import { type Page, type BrowserContext } from "patchright";
-import { withAccountPage, assertAntiBotHeaders, onBrowserContextCreated } from "./playwright.ts";
+import { withAccountPage, assertAntiBotHeaders, onBrowserContextCreated, isPlaywrightInitializing } from "./playwright.ts";
 import { recoverBaxiaCaptcha } from "./captcha-coordinator.ts";
 import { startBaxiaCaptchaWatcher } from "./captcha-solver.ts";
 import { isAccountBusy } from "../core/account-concurrency.ts";
@@ -1767,6 +1767,14 @@ export async function disableNativeTools(accountId?: string): Promise<void> {
   ) {
     return;
   }
+  // Defer if the account's Playwright browser is still initializing: the
+  // settings POST requires the page mutex, and init legitimately holds it for
+  // 20-30s (navigation + bx SDK + header capture). Attempting to acquire the
+  // mutex now would time out (5s default in withQwenBrowserPage) and log a
+  // spurious WARN. The startup flow calls us again after init finishes.
+  if (accountId && isPlaywrightInitializing(accountId)) {
+    return;
+  }
   disablingNativeToolsInProgress.add(cacheKey);
 
   try {
@@ -2827,7 +2835,7 @@ async function createQwenStreamInternal(
             auto_thinking: mode === "auto",
             thinking_mode: thinkingMode,
             ...(thinkingEnabled ? { thinking_format: "summary" } : {}),
-            auto_search: true,
+            auto_search: false,
           };
         })(),
         extra: {

package/src/tools/parser.ts

@@ -1703,6 +1703,12 @@ export class StreamingToolParser {
       result.text += literalBlock;
     }
 
+    // Count undeclared/malformed tool calls toward the per-turn cap.
+    // Without this, the model can generate unlimited undeclared tool calls
+    // (e.g. Qwen-native WebSearch/WebFetch) that bypass the cap entirely,
+    // causing infinite generation until TOTAL_REQUEST_TIMEOUT (10 min).
+    this.emittedToolCallCount++;
+
     this.advanceMarkdownState(literalBlock);
     this.pendingLeadIn = "";
   }