fix: strip markdown code fences from LLM responses before json.loads

- glm-5.2 returns JSON wrapped in ```json ... ``` code fences
- Added fence stripping in services.py, query_understanding.py, jobs.py

app/plugins/builtins/ai_proactive/jobs.py

@@ -225,6 +225,12 @@ async def deep_analysis(
             content = response.choices[0].message.content
             if not content:
                 return
+            # Strip markdown code fences if present
+            content = content.strip()
+            if content.startswith("```"):
+                content = content.split("\n", 1)[-1] if "\n" in content else content[3:]
+            if content.endswith("```"):
+                content = content[:-3].strip()
             suggestion_data = json.loads(content)
             if not suggestion_data.get("title") or not suggestion_data.get("content"):
                 return

app/plugins/builtins/ai_proactive/services.py

@@ -435,6 +435,12 @@ async def generate_suggestion(
         content = response.choices[0].message.content
         if not content:
             return None
+        # Strip markdown code fences if present (e.g. ```json ... ```)
+        content = content.strip()
+        if content.startswith("```"):
+            content = content.split("\n", 1)[-1] if "\n" in content else content[3:]
+        if content.endswith("```"):
+            content = content[:-3].strip()
         result = json.loads(content)
         # Validate required fields
         if not result.get("title") or not result.get("content"):