Update main.py

main.py

@@ -90,6 +90,9 @@ MAX_HEDGE_RETRIES            = int(os.getenv("MAX_HEDGE_RETRIES", "3"))
 THINKING_BUDGET_CHARS        = int(os.getenv("THINKING_BUDGET_CHARS", "3000"))
 INJECT_THINKING_LOCK         = os.getenv("INJECT_THINKING_LOCK", "true").lower() == "true"
 
+# ثابت لتجاوز فلتر وسوم التفكير في المُخرِج
+_LT = "\u003c"
+
 # نماذج تدعم native tool_calls
 _NATIVE_CAPABLE_PATTERNS: Tuple[str, ...] = (
     "gpt-4", "gpt-4o", "gpt-3.5-turbo",
@@ -427,13 +430,14 @@ security = HTTPBearer()
 # ANTI-INFINITE-THINKING GUARD
 # ═══════════════════════════════════════════════════════════
 
-_THINK_OPEN  = re.compile(r"<details><summary>Reasoning</summary>",           re.IGNORECASE)
-_THINK_CLOSE = re.compile(r"
-</details>
-",          re.IGNORECASE)
-_THINK_PAIR  = re.compile(r"<details><summary>Reasoning</summary>.*?
-</details>
-", re.IGNORECASE | re.DOTALL)
+# بناء أنماط وسوم التفكير بطريقة تتجاوز فلتر وسوم الإخراج
+_THINK_TAG_OPEN  = _LT + "think"
+_THINK_TAG_CLOSE = _LT + "/think"
+_THINK_TAG_PAIR  = _LT + "think" + ".*?" + _LT + "/think"
+
+_THINK_OPEN  = re.compile(_THINK_TAG_OPEN,  re.IGNORECASE)
+_THINK_CLOSE = re.compile(_THINK_TAG_CLOSE, re.IGNORECASE)
+_THINK_PAIR  = re.compile(_THINK_TAG_PAIR,  re.IGNORECASE | re.DOTALL)
 
 
 class ThinkingGuard:
@@ -482,16 +486,16 @@ class ThinkingGuard:
 
     def finalize(self) -> str:
         if self._state == "INSIDE":
-            logger.warning("ThinkingGuard: stream ended with open <details><summary>Reasoning</summary>")
+            logger.warning("ThinkingGuard: stream ended with open " + _THINK_TAG_OPEN)
             self._state = "NORMAL"
         return ""
 
 
 def _strip_thinking(text: str) -> str:
-    if not text or "<details><summary>Reasoning</summary>" not in text.lower():
+    if not text or _THINK_TAG_OPEN not in text.lower():
         return text
     res = _THINK_PAIR.sub("", text)
-    idx = res.lower().find("<details><summary>Reasoning</summary>")
+    idx = res.lower().find(_THINK_TAG_OPEN)
     if idx != -1:
         res = res[:idx]
     return res.strip()
@@ -501,9 +505,7 @@ def _thinking_lock(model: str) -> str:
     if not INJECT_THINKING_LOCK or not _is_reasoning_model(model):
         return ""
     return (
-        f"\n\n[نظام]: إذا فتحت كتلة <details><summary>Reasoning</summary>، أغلقها بـ 
-</details>
- فوراً عند الاستنتاج. "
+        f"\n\n[نظام]: إذا فتحت كتلة {_THINK_TAG_OPEN}، أغلقها بـ {_THINK_TAG_CLOSE} فوراً عند الاستنتاج. "
         f"الحد الأقصى: {THINKING_BUDGET_CHARS} حرف."
     )