diff --git a/agent.py b/agent.py
index 0c1260b..ed5047f 100644
--- a/agent.py
+++ b/agent.py
@@ -242,7 +242,7 @@ class TaskRunner(threading.Thread):
                         [{'role': 'system',
                           'content': VISION_SYSTEM_PROMPT if self.vision else SYSTEM_PROMPT},
                          {'role': 'user', 'content': user_content}],
-                        cfg=self.llm_cfg, temperature=0.2, max_tokens=500)
+                        cfg=self.llm_cfg, max_tokens=1024)
                 except LLMError as e:
                     self._log(f'LLM 错误: {e}')
                     self._finish(browser, 'error', f'LLM 决策失败: {e}', steps)
@@ -403,7 +403,7 @@ class TaskRunner(threading.Thread):
                   'content': VISION_SYSTEM_PROMPT if self.vision else SYSTEM_PROMPT},
                  {'role': 'user',
                   'content': self._build_user_content(retry_msg, browser, self.current.get('step', 0) + 99)}],
-                cfg=self.llm_cfg, temperature=0.2, max_tokens=400)
+                cfg=self.llm_cfg, max_tokens=800)
             new_action = retry_dec.get('action', '')
             if new_action == 'fail':
                 return 'error', f'自愈放弃: {retry_dec.get("summary", last_err)}', {}
diff --git a/llm.py b/llm.py
index d1bf82d..250282e 100644
--- a/llm.py
+++ b/llm.py
@@ -89,11 +89,25 @@ def chat(messages, cfg=None, temperature=None, max_tokens=None, timeout=None):
         raise LLMError(f'LLM 调用失败: {e}')
 
     try:
-        return data['choices'][0]['message']['content']
+        msg = data['choices'][0]['message']
+        content = msg.get('content') or ''
+        if not content.strip():
+            # thinking 模型（如 qwen3/deepseek-r1 系）：最终答案可能为空，
+            # 思考过程放在 reasoning_content 里（常包含最终 JSON），用它兜底
+            reasoning = msg.get('reasoning_content') or ''
+            if reasoning.strip():
+                return reasoning
+        return content
     except (KeyError, IndexError, TypeError):
         raise LLMError(f'LLM 响应异常: {str(data)[:300]}')
 
 
+def _trim_content(text, limit=300):
+    """截断过长的模型输出（避免超长思考内容反复进入上下文）"""
+    text = (text or '').strip()
+    return text if len(text) <= limit else text[:limit] + '...(已截断)'
+
+
 def chat_json(messages, cfg=None, temperature=None, max_tokens=None, retries=2):
     """调用 LLM 并强制解析 JSON，失败重试"""
     last_err = None
@@ -110,8 +124,8 @@ def chat_json(messages, cfg=None, temperature=None, max_tokens=None, retries=2):
             last_err = str(e)
         if i < retries:
             messages = messages + [
-                {'role': 'assistant', 'content': last_content or ''},
+                {'role': 'assistant', 'content': _trim_content(last_content)},
                 {'role': 'user',
-                 'content': f'刚才的输出不是合法 JSON，请只输出严格的 JSON 对象。错误: {last_err}'},
+                 'content': f'刚才的输出不是合法 JSON，请只输出严格的 JSON 对象，不要输出思考过程和代码块。错误: {last_err}'},
             ]
     raise LLMError(f'LLM JSON 解析失败: {last_err}')