feat: fold_turns UI折叠 + 放宽输出限制 + 压缩history/key_info标签 + context_win调大

frontends/stapp.py

@@ -66,6 +66,23 @@ def render_sidebar():
         st.caption("🔴 自主行动已停止")
 with st.sidebar: render_sidebar()
 
+def fold_turns(text):
+    parts = re.split(r'(\**LLM Running \(Turn \d+\) \.\.\.*\**)', text)
+    if len(parts) < 4: return text
+    result = parts[0]
+    turns = []
+    for i in range(1, len(parts), 2):
+        marker = parts[i].strip('*')
+        content = parts[i+1] if i+1 < len(parts) else ''
+        turns.append((marker, content))
+    for idx, (marker, content) in enumerate(turns):
+        if idx < len(turns) - 1:
+            m = re.search(r'<summary>\s*(.*?)\|*</summary>', content, re.DOTALL)
+            title = m.group(1).strip() if m else marker
+            result += f'<details><summary>{title}</summary>\n\n{content}\n</details>\n\n'
+        else:
+            result += marker + content
+    return result
 
 def agent_backend_stream(prompt):
     display_queue = agent.put_task(prompt, source="user")
@@ -84,7 +101,11 @@ def agent_backend_stream(prompt):
 
 if "messages" not in st.session_state: st.session_state.messages = []
 for msg in st.session_state.messages:
-    with st.chat_message(msg["role"]): st.markdown(msg["content"], unsafe_allow_html=False)
+    with st.chat_message(msg["role"]):
+        if msg["role"] == "assistant":
+            st.markdown(fold_turns(msg["content"]), unsafe_allow_html=True)
+        else:
+            st.markdown(msg["content"], unsafe_allow_html=False)
 
 # IME composition fix (macOS only) - prevents Enter from submitting during CJK input
 if os.name != 'nt':
@@ -99,8 +120,8 @@ if prompt := st.chat_input("请输入指令"):
         message_placeholder = st.empty()
         response = ''
         for response in agent_backend_stream(prompt):
-            message_placeholder.markdown(response + "▌", unsafe_allow_html=False)
-        message_placeholder.markdown(response, unsafe_allow_html=False)
+            message_placeholder.markdown(fold_turns(response) + "...", unsafe_allow_html=True)
+        message_placeholder.markdown(fold_turns(response), unsafe_allow_html=True)
     st.session_state.messages.append({"role": "assistant", "content": response})
     st.session_state.last_reply_time = int(time.time())
 

ga.py

@@ -73,12 +73,12 @@ def code_run(code, code_type="python", timeout=60, cwd=None, code_cwd=None, stop
         status = "success" if exit_code == 0 else "error"
         status_icon = "✅" if exit_code == 0 else "❌"
         if exit_code is None: status_icon = "⏳" 
-        output_snippet = smart_format(stdout_str, max_str_len=600, omit_str='\n[omitted long output]\n')
+        output_snippet = smart_format(stdout_str, max_str_len=600, omit_str='\n\n[omitted long output]\n\n')
         yield f"[Status] {status_icon} Exit Code: {exit_code}\n[Stdout]\n{output_snippet}\n"
         if process.stdout: threading.Thread(target=process.stdout.close, daemon=True).start()
         return {
             "status": status,
-            "stdout": smart_format(stdout_str, max_str_len=8000, omit_str='\n[omitted long output]\n'),
+            "stdout": smart_format(stdout_str, max_str_len=10000, omit_str='\n\n[omitted long output]\n\n'),
             "exit_code": exit_code
         }
     except Exception as e:
@@ -137,7 +137,7 @@ def web_scan(tabs_only=False, switch_tab_id=None, text_only=False):
                 "active_tab": driver.default_session_id
             }
         }
-        if not tabs_only: result["content"] = get_html(driver, cutlist=True, maxchars=28000, text_only=text_only)
+        if not tabs_only: result["content"] = get_html(driver, cutlist=True, maxchars=38000, text_only=text_only)
         return result
     except Exception as e:
         return {"status": "error", "msg": format_error(e)}
@@ -354,7 +354,7 @@ class GenericAgentHandler(BaseHandler):
         except: pass
         yield f"JS 执行结果:\n{smart_format(result)}\n"
         next_prompt = self._get_anchor_prompt(skip=args.get('_index', 0) > 0)
-        return StepOutcome(smart_format(result, max_str_len=5000), next_prompt=next_prompt)
+        return StepOutcome(smart_format(result, max_str_len=8000), next_prompt=next_prompt)
     
     def do_file_patch(self, args, response):
         path = self._get_abs_path(args.get("path", ""))

llmcore.py

@@ -15,19 +15,18 @@ proxy = mykeys.get("proxy", 'http://127.0.0.1:2082')
 proxies = {"http": proxy, "https": proxy} if proxy else None
 
 def compress_history_tags(messages, keep_recent=10, max_len=800):
-    """Compress <thinking>/<tool_use>/<tool_result> tags in older messages to save tokens.
-    Supports both prompt-style (ClaudeSession/LLMSession) and content-style (NativeClaudeSession) messages."""
+    """Compress <thinking>/<tool_use>/<tool_result> tags in older messages to save tokens."""
     compress_history_tags._cd = getattr(compress_history_tags, '_cd', 0) + 1
     if compress_history_tags._cd % 5 != 0: return messages
     _before = sum(len(json.dumps(m, ensure_ascii=False)) for m in messages)
     _pats = {tag: re.compile(rf'(<{tag}>)([\s\S]*?)(</{tag}>)') for tag in ('thinking', 'think', 'tool_use', 'tool_result')}
+    _hist_pat = re.compile(r'<(history|key_info)>[\s\S]*?</\1>')
     def _trunc(text):
+        text = _hist_pat.sub(lambda m: f'<{m.group(1)}>[...]</{m.group(1)}>', text)
         for pat in _pats.values(): text = pat.sub(lambda m: m.group(1) + m.group(2)[:max_len] + '...' + m.group(3) if len(m.group(2)) > max_len else m.group(0), text)
         return text
     for i, msg in enumerate(messages):
         if i >= len(messages) - keep_recent: break
-        if 'prompt' in msg: msg['prompt'] = _trunc(msg['prompt'])
-        elif 'content' in msg and 'prompt' not in msg:
         c = msg['content']
         if isinstance(c, str): msg['content'] = _trunc(c)
         elif isinstance(c, list):
@@ -40,7 +39,7 @@ def compress_history_tags(messages, keep_recent=10, max_len=800):
 def _sanitize_leading_user_msg(msg):
     """把 user 消息里的 tool_result 块改写成纯文本，避免孤立引用。
     history 统一使用 Claude content-block 格式：content 是 list of blocks。"""
-    msg = dict(msg)  # 浅拷贝外层 dict；content 在 L56 整体替换而非原地修改，故原对象的 content 不受影响
+    msg = dict(msg)  # 浅拷贝外层 dict
     content = msg.get('content')
     if not isinstance(content, list): return msg
     texts = []
@@ -411,7 +410,7 @@ class BaseSession:
         self.api_key = cfg['apikey']
         self.api_base = cfg['apibase'].rstrip('/')
         self.default_model = cfg.get('model', '')
-        self.context_win = cfg.get('context_win', 20000)
+        self.context_win = cfg.get('context_win', 24000)
         self.history = []
         self.lock = threading.Lock()
         self.system = ""
@@ -481,7 +480,7 @@ class LLMSession(BaseSession):
 class NativeClaudeSession(BaseSession):
     def __init__(self, cfg):
         super().__init__(cfg)
-        self.context_win = cfg.get("context_win", 24000)
+        self.context_win = cfg.get("context_win", 28000)
         self.no_system_prompt = cfg.get("no_system_prompt", False)
 
     def raw_ask(self, messages, tools=None, system=None, model=None, temperature=0.5, max_tokens=6144):