fix: use Prompture's strip_think_tags and clean_json_text instead of hand-rolled regexes

chat() and chat_json() now delegate think-tag stripping and JSON
cleanup to Prompture's built-in utilities (strip_think_tags,
clean_json_text).  Manual regexes are kept only in the OpenAI
fallback path.  Adds LM Studio integration test script.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/app/utils/llm_client.py

@@ -16,6 +16,7 @@ from ..config import Config
 try:
     from prompture.agents import Conversation
     from prompture.infra.provider_env import ProviderEnvironment
+    from prompture.extraction.tools import strip_think_tags, clean_json_text
     _HAS_PROMPTURE = True
 except ImportError:
     _HAS_PROMPTURE = False
@@ -121,12 +122,11 @@ class LLMClient:
         """
         if _HAS_PROMPTURE:
             content = self._chat_prompture(messages, temperature, max_tokens)
+            return strip_think_tags(content)
         else:
             content = self._chat_openai(messages, temperature, max_tokens, response_format)
-
-        # 部分模型（如MiniMax M2.5）会在content中包含<think>思考内容，需要移除
-        content = re.sub(r'<think>[\s\S]*?</think>', '', content).strip()
-        return content
+            # Fallback: strip think tags with regex when Prompture is not available
+            return re.sub(r'<think>[\s\S]*?</think>', '', content).strip()
 
     def chat_json(
         self,
@@ -147,17 +147,18 @@ class LLMClient:
         """
         if _HAS_PROMPTURE:
             response = self._chat_prompture(messages, temperature, max_tokens)
+            # Prompture's clean_json_text strips think tags + markdown fences
+            cleaned = clean_json_text(response)
         else:
             response = self._chat_openai(
                 messages, temperature, max_tokens,
                 response_format={"type": "json_object"},
             )
-
-        # 清理markdown代码块标记
-        cleaned = response.strip()
-        cleaned = re.sub(r'^```(?:json)?\s*\n?', '', cleaned, flags=re.IGNORECASE)
-        cleaned = re.sub(r'\n?```\s*$', '', cleaned)
-        cleaned = cleaned.strip()
+            # Fallback cleaning when Prompture is not available
+            cleaned = re.sub(r'<think>[\s\S]*?</think>', '', response).strip()
+            cleaned = re.sub(r'^```(?:json)?\s*\n?', '', cleaned, flags=re.IGNORECASE)
+            cleaned = re.sub(r'\n?```\s*$', '', cleaned)
+            cleaned = cleaned.strip()
 
         try:
             return json.loads(cleaned)

backend/scripts/test_lmstudio.py

@@ -0,0 +1,68 @@
+"""
+Quick test: MiroFish LLMClient → LM Studio via Prompture
+"""
+import sys, os
+
+# Add backend to path so we can import app modules
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+# Override env vars for LM Studio before Config loads
+os.environ["LLM_MODEL_NAME"] = "lmstudio/deepseek/deepseek-r1-0528-qwen3-8b"
+os.environ["LLM_BASE_URL"] = "http://localhost:1234/v1"
+os.environ["LLM_API_KEY"] = "lm-studio"
+# Provide a dummy ZEP key so Config.validate() won't complain
+os.environ.setdefault("ZEP_API_KEY", "dummy")
+
+from app.utils.llm_client import LLMClient
+
+def test_basic_chat():
+    print("=== Test 1: Basic chat ===")
+    client = LLMClient()
+    from app.utils.llm_client import _HAS_PROMPTURE
+    print(f"  Backend: Prompture={_HAS_PROMPTURE}")
+    print(f"  Model:   {client.model}")
+    response = client.chat([
+        {"role": "system", "content": "You are a helpful assistant. Reply in one sentence."},
+        {"role": "user", "content": "What is social media simulation?"},
+    ], temperature=0.5, max_tokens=256)
+    print(f"  Response: {response[:300]}")
+    print()
+
+def test_json_chat():
+    print("=== Test 2: JSON response ===")
+    client = LLMClient()
+    result = client.chat_json([
+        {"role": "system", "content": "You are a JSON-only assistant. Always respond with valid JSON."},
+        {"role": "user", "content": 'Return a JSON object with keys "platform" and "agents" (an integer). Example: {"platform":"twitter","agents":5}'},
+    ], temperature=0.2, max_tokens=256)
+    print(f"  Parsed JSON: {result}")
+    print(f"  Type: {type(result)}")
+    print()
+
+def test_multi_turn():
+    print("=== Test 3: Multi-turn conversation ===")
+    client = LLMClient()
+    r1 = client.chat([
+        {"role": "user", "content": "My name is MiroFish. Remember it."},
+    ], max_tokens=128)
+    print(f"  Turn 1: {r1[:200]}")
+
+    r2 = client.chat([
+        {"role": "user", "content": "My name is MiroFish. Remember it."},
+        {"role": "assistant", "content": r1},
+        {"role": "user", "content": "What is my name?"},
+    ], max_tokens=128)
+    print(f"  Turn 2: {r2[:200]}")
+    print()
+
+if __name__ == "__main__":
+    print(f"Prompture installed: True")
+    print(f"LM Studio endpoint: http://localhost:1234/v1\n")
+    try:
+        test_basic_chat()
+        test_json_chat()
+        test_multi_turn()
+        print("All tests passed!")
+    except Exception as e:
+        print(f"ERROR: {e}")
+        import traceback; traceback.print_exc()