qwenproxy-cli 1.6.4 → 1.6.6

package/src/routes/chat/streaming.ts

@@ -33,6 +33,10 @@ import {
 import type { Message, OpenAIRequest, Usage } from "../../utils/types.ts";
 import { touchChatLock } from "./account.ts";
 import { StreamingToolParser } from "../../tools/parser.ts";
+import {
+  decideMissingToolCallRetry,
+  type MissingToolCallReason,
+} from "../../tools/tool-call-heuristics.ts";
 import {
   getStream,
   markStreamEmitted,
@@ -609,8 +613,17 @@ export async function processNonStreamingResponse(
 
     const finishReason = toolCallsOut.length ? "tool_calls" : "stop";
 
-    // Auto-retry if all tool calls were malformed (no successful tool calls)
-    const allToolsFailed = toolParser && toolParser.getMalformedToolCalls().length > 0 && toolCallsOut.length === 0;
+    // Auto-retry if all tool calls were malformed or model hesitated with missing call
+    const retryDecision = toolParser
+      ? decideMissingToolCallRetry({
+          declaredToolCount: declaredTools.length,
+          successfulToolCallCount: toolParser.getSuccessfulToolCallCount(),
+          malformedToolCallCount: toolParser.getMalformedToolCalls().length,
+          emittedToolCallCount: toolParser.getEmittedToolCallCount(),
+          text: finalContent,
+        })
+      : { retry: false, reason: null as MissingToolCallReason };
+    const allToolsFailed = retryDecision.retry;
     const malformedRetryCount = midStreamRetry?.malformedRetryCount ?? 0;
     if (
       allToolsFailed &&
@@ -618,7 +631,7 @@ export async function processNonStreamingResponse(
       midStreamRetry &&
       malformedRetryCount < config.retry.autoRetryMalformedToolsMax
     ) {
-      const malformedCalls = toolParser.getMalformedToolCalls();
+      const malformedCalls = toolParser?.getMalformedToolCalls() ?? [];
       const malformedCount = malformedCalls.length;
 
       // Build detailed error message with available tools list
@@ -634,7 +647,9 @@ export async function processNonStreamingResponse(
         : "";
 
       let errorMessage: string;
-      if (undeclaredNames.length > 0) {
+      if (retryDecision.reason === "no_call") {
+        errorMessage = `You did not emit any tool call. Respond with a single tool-call block using one of the declared tools.${toolsHint}`;
+      } else if (undeclaredNames.length > 0) {
         errorMessage = `Your previous ${malformedCount} tool call(s) used undeclared tool names: ${undeclaredNames.join(", ")}. Only declared tools can be executed. Please retry with valid tool names.${toolsHint}`;
       } else {
         const previews = malformedCalls
@@ -719,7 +734,7 @@ export async function processNonStreamingResponse(
         );
       }
 
-      console.log(`🔄 [Chat] Auto-retry (${malformedRetryCount + 1}/${config.retry.autoRetryMalformedToolsMax}) | ${newStreamResult.activeAccountLabel} | ${body.model} | chat=${newStreamResult.uiSessionId.substring(0, 12)} | reason=malformed_tool_calls`);
+      console.log(`🔄 [Chat] Auto-retry (${malformedRetryCount + 1}/${config.retry.autoRetryMalformedToolsMax}) | ${newStreamResult.activeAccountLabel} | ${body.model} | chat=${newStreamResult.uiSessionId.substring(0, 12)} | reason=${retryDecision.reason === "no_call" ? "missing_tool_call" : "malformed_tool_calls"}`);
 
       // Process the new stream. Propagate the retry context with an incremented
       // counter so a subsequent malformed response can retry again up to
@@ -922,7 +937,7 @@ export async function processStreamingResponse(
     let chatInProgressSameAccountRetries = 0;
     let antiBotSameAccountRetries = 0;
     const totalConfiguredAccounts = loadAccounts().length;
-    let quotaRotationsLeft = Math.min(Math.max(1, totalConfiguredAccounts - 1), 6);
+    let quotaRotationsLeft = Math.max(1, totalConfiguredAccounts - 1);
     // Set when the terminal [DONE] reached the client. The `| recovered`
     // suffix on Stream done must only appear for attempts that COMPLETED after
     // a mid-stream retry, not for failed attempts that consumed retries.
@@ -2164,9 +2179,14 @@ export async function processStreamingResponse(
 
         while (malformedRetryCount < maxMalformedRetries) {
           if (!toolParser) break;
-          const allToolsFailed =
-            toolParser.getMalformedToolCalls().length > 0 &&
-            toolParser.getSuccessfulToolCallCount() === 0;
+          const retryDecision = decideMissingToolCallRetry({
+            declaredToolCount: declaredTools.length,
+            successfulToolCallCount: toolParser.getSuccessfulToolCallCount(),
+            malformedToolCallCount: toolParser.getMalformedToolCalls().length,
+            emittedToolCallCount: toolParser.getEmittedToolCallCount(),
+            text: finalContent,
+          });
+          const allToolsFailed = retryDecision.retry;
           if (!allToolsFailed) break;
 
           const malformedCalls = toolParser.getMalformedToolCalls();
@@ -2185,7 +2205,9 @@ export async function processStreamingResponse(
               : "";
 
           let errorMessage: string;
-          if (undeclaredNames.length > 0) {
+          if (retryDecision.reason === "no_call") {
+            errorMessage = `You did not emit any tool call. Respond with a single tool-call block using one of the declared tools.${toolsHint}`;
+          } else if (undeclaredNames.length > 0) {
             errorMessage = `Your previous ${malformedCount} tool call(s) used undeclared tool names: ${undeclaredNames.join(", ")}. Only declared tools can be executed. Please retry with valid tool names.${toolsHint}`;
           } else {
             const previews = malformedCalls
@@ -2270,7 +2292,7 @@ export async function processStreamingResponse(
           }
 
           console.log(
-            `🔄 [Chat] Auto-retry (${malformedRetryCount + 1}/${maxMalformedRetries}) | ${newStreamResult.activeAccountLabel} | ${body.model} | chat=${newStreamResult.uiSessionId.substring(0, 12)} | reason=malformed_tool_calls`,
+            `🔄 [Chat] Auto-retry (${malformedRetryCount + 1}/${maxMalformedRetries}) | ${newStreamResult.activeAccountLabel} | ${body.model} | chat=${newStreamResult.uiSessionId.substring(0, 12)} | reason=${retryDecision.reason === "no_call" ? "missing_tool_call" : "malformed_tool_calls"}`,
           );
 
           // Transfer the retry's stream registry entry to the original
@@ -2809,28 +2831,41 @@ export async function processStreamingResponse(
       logger.error("[Chat] Stream callback error", errorDetails);
     }
     // The HTTP response is already committed at this point.
-    // 1. Emit an explicit assistant message delta so CLI/TUI clients that only
-    //    listen for choices[].delta.content (e.g. OpenCode, Claude Code, etc.)
-    //    render the error visibly in the chat instead of appearing blank/frozen.
-    // 2. Emit a terminal OpenAI-compatible SSE error event + [DONE].
+    // 1. Emit an explicit assistant message delta ONLY if this was an upstream content moderation /
+    //    policy violation (with finish_reason: "content_filter"), so the user understands why the
+    //    prompt was blocked.
+    // 2. For network/transient/quota errors, NEVER emit fake assistant content with finish_reason: "stop".
+    //    Doing so masks network failures as successful AI answers, preventing client agent retries and
+    //    poisoning chat history with error banners.
+    // 3. Emit a terminal OpenAI-compatible SSE error event + [DONE].
     try {
-      const userFriendlyNotice = `\n\n⚠️ **[Qwen Security / Erro]** ${err.message || "A resposta foi interrompida pelo Qwen."}\n\n`;
-      const errorDelta = {
-        id: completionId,
-        object: "chat.completion.chunk",
-        created: Math.floor(Date.now() / 1000),
-        model: body.model,
-        choices: [
-          {
-            index: 0,
-            delta: { content: userFriendlyNotice },
-            logprobs: null,
-            finish_reason: "stop",
-          },
-        ],
-      };
+      const isNetwork = isNetworkLikeError(err);
+      const isModeration =
+        normalizedErrorCode === "data_inspection_failed" ||
+        normalizedErrorCode === "fail_sys_user_validate" ||
+        err.message?.toLowerCase().includes("content moderation") ||
+        err.message?.toLowerCase().includes("security verification");
+
+      if (!isNetwork && isModeration) {
+        const userFriendlyNotice = `\n\n⚠️ **[Qwen Moderação / Segurança]** ${err.message || "A resposta foi interrompida pelas políticas de segurança do Qwen."}\n\n`;
+        const errorDelta = {
+          id: completionId,
+          object: "chat.completion.chunk",
+          created: Math.floor(Date.now() / 1000),
+          model: body.model,
+          choices: [
+            {
+              index: 0,
+              delta: { content: userFriendlyNotice },
+              logprobs: null,
+              finish_reason: "content_filter",
+            },
+          ],
+        };
+
+        await errorStream.write(`data: ${JSON.stringify(errorDelta)}\n\n`);
+      }
 
-      await errorStream.write(`data: ${JSON.stringify(errorDelta)}\n\n`);
       await errorStream.write(
         `data: ${JSON.stringify({
           error: {

package/src/routes/responses/adapter.ts

@@ -372,7 +372,10 @@ export function chatCompletionsToResponses(
       cached_tokens: (chatRes.usage as any).prompt_tokens_details?.cached_tokens ?? 0,
     },
     output_tokens_details: {
-      reasoning_tokens: (chatRes.usage as any).completion_tokens_details?.reasoning_tokens ?? 0,
+      reasoning_tokens:
+        (chatRes.usage as any).completion_tokens_details?.reasoning_tokens ??
+        (chatRes.usage as any).output_tokens_details?.reasoning_tokens ??
+        0,
     },
     ...(chatRes.usage.context_meter
       ? { context_meter: chatRes.usage.context_meter }

package/src/services/context-meter.ts

@@ -189,7 +189,7 @@ export function enrichUsageWithContextMeter(
       ? usage.prompt_tokens
       : null;
   const reportedPromptTokens = options.reportUsage
-    ? (upstreamPromptTokens ?? snapshot.estimatedContextTokens)
+    ? (upstreamPromptTokens ?? usage.prompt_tokens)
     : usage.prompt_tokens;
   const contextMeter: ContextMeterSnapshot = {
     ...snapshot,

package/src/services/playwright.ts

@@ -83,6 +83,7 @@ import { setWafContextResetListener } from "../core/waf-isolation.ts";
 import { updateQwenWebVersion, getQwenWebVersion } from "./qwen-headers.ts";
 import { getAccountProfilePath, getProfilesDir } from "../core/paths.ts";
 import { parseJwtExpiry, isTokenExpiringSoon } from "../utils/jwt.ts";
+import { formatAsciiTimezone } from "../utils/timezone.ts";
 
 type ContextInitHook = (context: BrowserContext) => Promise<void> | void;
 const contextInitHooks: ContextInitHook[] = [];
@@ -517,7 +518,7 @@ const headerCaches = new Map<string, AccountHeaderCache>();
 // across the browser session; cookies are dynamically updated from the browser context.
 const HEADER_CACHE_TTL = 24 * 60 * 60 * 1000; // 24 hours
 const HEADER_REFRESH_THRESHOLD = 0.8; // Background refresh at 80% of TTL
-const COOKIE_CACHE_TTL = 15 * 60 * 1000; // 15 minutes
+const COOKIE_CACHE_TTL = 60 * 1000; // 1 minute (prevents returning stale expiring token after refresh)
 const cookieCaches = new Map<string, { cookie: string; timestamp: number }>();
 const lastAccountActivity = new Map<string, number>();
 const lastKeepAliveNavigation = new Map<string, number>();
@@ -1528,6 +1529,14 @@ export async function initPlaywrightForAccount(
     return;
   }
 
+  let resolveInit!: () => void;
+  let rejectInit!: (err: any) => void;
+  const inFlightDeferred = new Promise<void>((res, rej) => {
+    resolveInit = res;
+    rejectInit = rej;
+  });
+  inFlightAccountInits.set(account.id, inFlightDeferred);
+
   const initPromise = (async () => {
     const release = await acquireAccountMutex(
       account.id,
@@ -1535,10 +1544,10 @@ export async function initPlaywrightForAccount(
       Math.max(PLAYWRIGHT_MUTEX_WAIT_MS, 120_000),
     );
     try {
-    // Double-check after acquiring lock
-    if (accountPages.has(account.id)) {
-      return;
-    }
+      // Double-check after acquiring lock
+      if (accountPages.has(account.id)) {
+        return;
+      }
 
     // If a context limit is configured, make room by closing idle contexts.
     await evictIdlePlaywrightContextsToLimit().catch(() => {});
@@ -1777,6 +1786,26 @@ export async function initPlaywrightForAccount(
             }
           }
         }
+
+        // If the session restored from SQLite has an expired access token (e.g. was in standby for hours/days),
+        // proactively refresh it in-page now with the 30-day refresh_token instead of waiting for the first chat
+        // request to hit an upstream 401 and trigger a slow re-auth / captcha loop.
+        const currentCache = getHeaderCache(account.id);
+        const currentCookie = currentCache?.headers?.cookie || "";
+        if (currentCookie && isTokenExpiringSoon(currentCookie, 5)) {
+          try {
+            const refreshRes = await refreshAccountToken(account.id);
+            if (refreshRes.success) {
+              console.log(
+                `⚡ [Playwright] Silently refreshed expired token restored from DB for ${maskEmail(account.email)}`,
+              );
+            }
+          } catch (err: any) {
+            console.warn(
+              `⚠️  [Playwright] Failed to proactively refresh expired token from DB for ${maskEmail(account.email)}: ${err?.message || err}`,
+            );
+          }
+        }
         return;
       }
 
@@ -1845,9 +1874,12 @@ export async function initPlaywrightForAccount(
     }
   })();
 
-  inFlightAccountInits.set(account.id, initPromise);
   try {
     await initPromise;
+    resolveInit();
+  } catch (error) {
+    rejectInit(error);
+    throw error;
   } finally {
     inFlightAccountInits.delete(account.id);
   }
@@ -2160,7 +2192,7 @@ export async function startEmailOtpLogin(
         email: normalizedEmail,
         requestId,
         version: getQwenWebVersion() || "0.3.12",
-        timezone: new Date().toString().split(" (")[0],
+        timezone: formatAsciiTimezone(),
       },
     ).catch((err: any) => ({ ok: false, error: err?.message || String(err) }));
 
@@ -2272,7 +2304,7 @@ export async function verifyEmailOtpLogin(
       code: cleanCode,
       requestId,
       version: getQwenWebVersion() || "0.3.12",
-      timezone: new Date().toString().split(" (")[0],
+      timezone: formatAsciiTimezone(),
     },
   ).catch((err: any) => ({ ok: false, error: err?.message || String(err) }));
 
@@ -2626,7 +2658,7 @@ async function loginViaApi(
 
     const evalRes = await page
       .evaluate(
-        async ({ email, password, signinUrl, requestId }) => {
+        async ({ email, password, signinUrl, requestId, tz }) => {
           try {
             const response = await fetch(signinUrl, {
               method: "POST",
@@ -2638,7 +2670,7 @@ async function loginViaApi(
                 source: "web",
                 version: "0.3.11",
                 "x-request-origin": "https://chat.qwen.ai",
-                timezone: new Date().toString().split(" (")[0],
+                timezone: tz,
                 "x-request-id": requestId,
               },
               body: JSON.stringify({ email, password }),
@@ -2649,7 +2681,7 @@ async function loginViaApi(
             return { ok: false, error: e.message };
           }
         },
-        { email, password: hashedPassword, signinUrl, requestId },
+        { email, password: hashedPassword, signinUrl, requestId, tz: formatAsciiTimezone() },
       )
       .catch(() => null);
 
@@ -2667,7 +2699,7 @@ async function loginViaApi(
       const fallbackUrl = qwenUrl("/api/v2/auths/signin");
       const fallbackRes = await page
         .evaluate(
-          async ({ email, password, fallbackUrl, requestId }) => {
+          async ({ email, password, fallbackUrl, requestId, tz }) => {
             try {
               const response = await fetch(fallbackUrl, {
                 method: "POST",
@@ -2678,7 +2710,7 @@ async function loginViaApi(
                   "content-type": "application/json",
                   source: "web",
                   version: "0.3.11",
-                  timezone: new Date().toString().split(" (")[0],
+                  timezone: tz,
                   "x-request-id": requestId,
                 },
                 body: JSON.stringify({ email, password }),
@@ -2689,7 +2721,7 @@ async function loginViaApi(
               return { ok: false, error: e.message };
             }
           },
-          { email, password: hashedPassword, fallbackUrl, requestId },
+          { email, password: hashedPassword, fallbackUrl, requestId, tz: formatAsciiTimezone() },
         )
         .catch(() => null);
 
@@ -3111,7 +3143,7 @@ export async function refreshAccountToken(
   if (refreshToken) {
     try {
       const refreshResult = await page.evaluate(
-        async ({ rTok }) => {
+        async ({ rTok, tz }) => {
           try {
             const response = await fetch("https://auth.qwen.ai/api/v2/auths/refresh", {
               method: "GET",
@@ -3121,7 +3153,7 @@ export async function refreshAccountToken(
                 source: "web",
                 version: "0.3.11",
                 "x-request-origin": "https://chat.qwen.ai",
-                timezone: new Date().toString().split(" (")[0],
+                timezone: tz,
                 authorization: `Bearer ${rTok}`,
               },
             });
@@ -3131,7 +3163,7 @@ export async function refreshAccountToken(
             return { ok: false, error: e.message };
           }
         },
-        { rTok: refreshToken },
+        { rTok: refreshToken, tz: formatAsciiTimezone() },
       );
 
       const json = refreshResult?.data;
@@ -3175,6 +3207,7 @@ export async function refreshAccountToken(
           const cache = getHeaderCache(accountId);
           cache.headers.cookie = cookieStr;
           cache.lastRefresh = Date.now();
+          cookieCaches.set(accountId, { cookie: cookieStr, timestamp: Date.now() });
           markAccountHeadersReady(accountId);
 
           // Update SQLite session with new tokens
@@ -3213,6 +3246,7 @@ export async function refreshAccountToken(
         const cache = getHeaderCache(accountId);
         cache.headers.cookie = cookieStr;
         cache.lastRefresh = Date.now();
+        cookieCaches.set(accountId, { cookie: cookieStr, timestamp: Date.now() });
         markAccountHeadersReady(accountId);
         saveAuthSession(accountId, {
           cookie: cookieStr,
@@ -4653,11 +4687,11 @@ export async function keepAlivePlaywrightAccount(
 
     // Proactive session check: if token expires within 5 minutes, refresh in-page silently using refreshAccountToken
     const lastRefreshAttempt = lastProactiveTokenRefresh.get(accountId) ?? 0;
-    const cookie = await getCookies(accountId);
+    const cookie = await getCookies(accountId, true);
     if (cookie && isTokenExpiringSoon(cookie, 5) && now - lastRefreshAttempt > 60_000) {
       lastProactiveTokenRefresh.set(accountId, now);
-      console.log(
-        `💓 [SessionKeeper] Account ${accountId} access token expiring soon (<5m); silently refreshing...`,
+      logger.debug(
+        `[SessionKeeper] Account ${accountId} access token expiring soon (<5m); silently refreshing...`,
       );
       const res = await refreshAccountToken(accountId);
       if (res.success) {
@@ -4665,6 +4699,9 @@ export async function keepAlivePlaywrightAccount(
         touchAccountActivity(accountId);
         return true;
       }
+      logger.warn(
+        `[SessionKeeper] Failed to silently refresh expiring token for ${accountId}: ${res.error || "unknown error"}`,
+      );
     }
 
     const lastNavigation = lastKeepAliveNavigation.get(accountId) ?? 0;

package/src/services/qwen-headers.ts

@@ -2,6 +2,7 @@ import { v4 as uuidv4 } from "uuid";
 import { qwenUrl, qwenOrigin } from "./qwen-url.ts";
 import { config } from "../core/config.js";
 import { getChromeMajor, getSystemLocale } from "./fingerprint.ts";
+import { formatAsciiTimezone } from "../utils/timezone.ts";
 
 export function getSystemAcceptLanguage(customLocale?: string): string {
   const locale = customLocale || getSystemLocale();
@@ -79,8 +80,8 @@ export function buildQwenRequestHeaders(
     "bx-v": opts.bxV || "2.5.37",
     source: "web",
     version: opts.version || getQwenWebVersion(),
-    timezone: opts.extra?.timezone || opts.extra?.Timezone || new Date().toString().split(" (")[0],
-    Timezone: opts.extra?.Timezone || opts.extra?.timezone || new Date().toString().split(" (")[0],
+    timezone: opts.extra?.timezone || opts.extra?.Timezone || formatAsciiTimezone(),
+    Timezone: opts.extra?.Timezone || opts.extra?.timezone || formatAsciiTimezone(),
     // Use the real browser client-hints when captured (anti-hardcoded); fall
     // back to the static fingerprint otherwise.
     "sec-ch-ua":