fix: LLM thinking 출력 대응 — max_tokens 증가 + JSON 추출 강화

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

scripts/pkm_utils.py

@@ -111,22 +111,29 @@ def llm_generate(prompt: str, model: str = "mlx-community/Qwen3.5-35B-A3B-4bit",
     import requests
     messages = []
     if json_mode:
-        messages.append({"role": "system", "content": "You must respond ONLY with valid JSON. No thinking, no explanation, no markdown."})
-    # Qwen3.5: /nothink 접미사로 thinking 출력 억제
-    messages.append({"role": "user", "content": prompt + " /nothink"})
+        messages.append({"role": "system", "content": "IMPORTANT: Output ONLY valid JSON. No thinking process, no explanation, no markdown fences. Start your response with { and end with }."})
+    messages.append({"role": "user", "content": prompt})
     resp = requests.post(f"{host}/v1/chat/completions", json={
         "model": model,
         "messages": messages,
-        "temperature": 0.3,
-        "max_tokens": 1024,
-    }, timeout=120)
+        "temperature": 0.1 if json_mode else 0.3,
+        "max_tokens": 2048,
+    }, timeout=180)
     resp.raise_for_status()
     content = resp.json()["choices"][0]["message"]["content"]
-    # JSON 블록 추출 (```json ... ``` 감싸기 대응)
+    # thinking 블록 제거 (Qwen3.5 thinking 모델 대응)
+    if "<think>" in content and "</think>" in content:
+        content = content.split("</think>")[-1].strip()
+    # JSON 블록 추출
     if "```json" in content:
         content = content.split("```json")[1].split("```")[0].strip()
     elif "```" in content:
         content = content.split("```")[1].split("```")[0].strip()
+    # { 로 시작하는 JSON 추출
+    import re
+    json_match = re.search(r'\{[\s\S]*\}', content)
+    if json_match:
+        content = json_match.group(0)
     return content