Cache last-working LLM to avoid retrying unavailable models (faster subsequent rewrites)

ai_ext.py

@@ -73,6 +73,8 @@ QWEN_TEXT_MODELS = [m.strip() for m in os.getenv(
     "QWEN_TEXT_MODELS",
     "Qwen/Qwen2.5-72B-Instruct,meta-llama/Llama-3.3-70B-Instruct,Qwen/Qwen2.5-7B-Instruct"
 ).split(",") if m.strip()]
+_WORKING_MODEL_TEXT = None  # cached last-working text model
+_WORKING_MODEL_VL = None    # cached last-working vision model
 DATA_DIR = "/data" if os.path.isdir("/data") else "/app/data"
 SHORTS_DIR = os.path.join(DATA_DIR, "ai_shorts")
 HEADERS = {
@@ -544,6 +546,11 @@ async def qwen_generate(prompt: str, image_url: Optional[str] = None, max_tokens
     else:
         # Text-only summary -> FAST text models first, VL only as last resort.
         candidate = QWEN_TEXT_MODELS + [QWEN_VL_MODEL]
+    # Use the last-known-working model first to avoid wasting time on unavailable models.
+    global _WORKING_MODEL_TEXT, _WORKING_MODEL_VL
+    cached_ok = _WORKING_MODEL_VL if has_images else _WORKING_MODEL_TEXT
+    if cached_ok and cached_ok in candidate:
+        candidate = [cached_ok] + [m for m in candidate if m != cached_ok]
     for m in candidate:
         if m and m not in models:
             models.append(m)
@@ -580,6 +587,8 @@ async def qwen_generate(prompt: str, image_url: Optional[str] = None, max_tokens
             txt = (resp.choices[0].message.content or "").strip()
             if txt:
                 LAST_QWEN_ERROR = ""
+                if has_images: _WORKING_MODEL_VL = model
+                else: _WORKING_MODEL_TEXT = model
                 return txt
         except Exception as e:
             errors.append(f"{model}: {type(e).__name__}: {str(e)[:220]}")