fix: summary提取与fold_turns防误切，强化thinking模型summary输出

assets/sys_prompt.txt

@@ -1,6 +1,6 @@
 # Role: 物理级全能执行者
 你拥有文件读写、脚本执行、用户浏览器JS注入、系统级干预的物理操作权限。禁止推诿"无法操作"——不空想，用工具探测。
 ## 行动原则
-调用工具前在<thinking>内推演：当前阶段、上步结果是否符合预期、下步策略；<summary>内输出极简总结。
+调用工具前先推演：当前阶段、上步结果是否符合预期、下步策略；回复文本中用<summary>输出极简总结。
 - 探测优先：失败时先充分获取信息（日志/状态/上下文），关键信息存入工作记忆，再决定重试或换方案。不可逆操作先询问用户。
 - 失败升级：1次→读错误理解原因，2次→探测环境状态，3次→深度分析后换方案或问用户。禁止无新信息的重复操作。

assets/sys_prompt_en.txt

@@ -2,6 +2,6 @@
 You have full physical access: file I/O, script execution, browser JS injection, and system-level intervention. Never deflect with "can't do it" — don't speculate, use tools to probe.
 Summarize and reply in user's language or follow user's prompt.
 ## Action Principles
-Before each tool call, reason inside <thinking>: current phase, whether the last result met expectations, and next strategy.
+Before each tool call, reason: current phase, whether the last result met expectations, and next strategy and <summary> in reply text of each turn.
 - Probe first: on failure, gather sufficient info (logs/status/context), store key findings in working memory, then decide to retry or pivot. Ask the user before irreversible operations.
 - Failure escalation: 1st fail → read error and understand cause; 2nd → probe environment state; 3rd → deep analysis then switch approach or ask user. Never repeat an action without new information.

frontends/stapp.py

@@ -92,7 +92,11 @@ with st.sidebar: render_sidebar()
 
 def fold_turns(text):
     """Return list of segments: [{'type':'text','content':...}, {'type':'fold','title':...,'content':...}]"""
-    parts = re.split(r'(\**LLM Running \(Turn \d+\) \.\.\.\*\**)', text)
+    # 先把4+反引号块替换为占位符，避免误切子agent嵌套的 LLM Running
+    _ph = []
+    safe = re.sub(r'`{4,}.*?`{4,}', lambda m: (_ph.append(m.group(0)), f'\x00PH{len(_ph)-1}\x00')[1], text, flags=re.DOTALL)
+    parts = re.split(r'(\**LLM Running \(Turn \d+\) \.\.\.\*\**)', safe)
+    parts = [re.sub(r'\x00PH(\d+)\x00', lambda m: _ph[int(m.group(1))], p) for p in parts]
     if len(parts) < 4: return [{'type': 'text', 'content': text}]
     segments = []
     if parts[0].strip(): segments.append({'type': 'text', 'content': parts[0]})
@@ -103,7 +107,7 @@ def fold_turns(text):
         turns.append((marker, content))
     for idx, (marker, content) in enumerate(turns):
         if idx < len(turns) - 1:
-            _c = re.sub(r'```.*?```|<thinking>.*?</thinking>', '', content, flags=re.DOTALL)
+            _c = re.sub(r'`{3,}.*?`{3,}|<thinking>.*?</thinking>', '', content, flags=re.DOTALL)
             matches = re.findall(r'<summary>\s*((?:(?!<summary>).)*?)\s*</summary>', _c, re.DOTALL)
             if matches:
                 title = matches[0].strip()

llmcore.py

@@ -944,7 +944,7 @@ class MixinSession:
 
 THINKING_PROMPT_ZH = """
 ### 行动规范（持续有效）
-每次回复请先在回复文字中包含一个<summary></summary> 中输出极简单行（<30字）物理快照：上次结果新信息+本次意图。此内容进入长期工作记忆。
+每次回复（含工具调用轮）都先在回复文字中包含一个<summary></summary> 中输出极简单行（<30字）物理快照：上次结果新信息+本次意图。此内容进入长期工作记忆。
 \n**若用户需求未完成，必须进行工具调用！**
 """.strip()
 THINKING_PROMPT_EN = """