classifier: Add fallback model for transient LLM failures

The 26b model occasionally hangs or fails outright on the local Ollama
box, which used to drop every such update into 'manual' for no good
reason. Retry the same prompt with a smaller model before giving up,
and surface both errors when even the fallback fails.

OLLAMA_FALLBACK_MODEL is new config, see .env.sample.

.env.sample

@@ -23,3 +23,4 @@ UPTIME_KUMA_PASSWORD=your-password
 # Ollama for AI-driven auto-update classification
 OLLAMA_BASE_URL=http://192.168.1.20:11434
 OLLAMA_MODEL=gemma4:26b
+OLLAMA_FALLBACK_MODEL=gemma4:12b

update_monitor/classifier.py

@@ -29,9 +29,11 @@ class Classifier:
         self,
         base_url: str = "",
         model: str = "",
+        fallback_model: str = "",
     ) -> None:
         self.base_url = base_url or os.getenv("OLLAMA_BASE_URL", "http://192.168.1.20:11434")
         self.model = model or os.getenv("OLLAMA_MODEL", "gemma4:26b")
+        self.fallback_model = fallback_model or os.getenv("OLLAMA_FALLBACK_MODEL", "gemma4:12b")
 
     def classify(
         self,
@@ -54,9 +56,23 @@ class Classifier:
         context = context_file.read_text(encoding="utf-8").strip() if context_file.is_file() else None
         messages = self._build_messages(service_dir.name, examples, release_notes, current, latest, context)
 
+        try:
+            return self._classify_with_model(self.model, messages, timeout)
+        except Exception as primary_exc:
+            if not self.fallback_model or self.fallback_model == self.model:
+                return {"action": "manual", "reason": f"Classification error: {primary_exc}"}
+            try:
+                return self._classify_with_model(self.fallback_model, messages, timeout)
+            except Exception as fallback_exc:
+                return {
+                    "action": "manual",
+                    "reason": f"Classification error: {primary_exc}; fallback ({self.fallback_model}) also failed: {fallback_exc}",
+                }
+
+    def _classify_with_model(self, model: str, messages: List[Dict[str, str]], timeout: float) -> Dict[str, str]:
         try:
             payload = json.dumps({
-                "model": self.model,
+                "model": model,
                 "messages": messages,
                 "stream": False,
                 "options": {"temperature": 0},
@@ -79,7 +95,7 @@ class Classifier:
                 raise ValueError(f"Unexpected action value: {action!r}")
             return {"action": action, "reason": result.get("reason", "")}
         except Exception as exc:
-            return {"action": "manual", "reason": f"Classification error: {exc}"}
+            raise RuntimeError(f"model {model}: {exc}") from exc
 
     def _build_messages(
         self,