fix stream recovery and runtime metrics

assets/sys_prompt.txt

@@ -1,7 +1,7 @@
 # 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
-调用工具前先推演：当前阶段、上步结果是否符合预期、下步策略，必须在回复文本中用<summary>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，必须在回复文本头部用小于30词<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。
 - 行动也是认识：边推进边用现实反馈校正方向，内部解释与计划只作暂时假说。根据任务选择高级sop调用

assets/sys_prompt_en.txt

@@ -2,6 +2,7 @@
 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: current phase, whether the last result met expectations, and next strategy and <summary> in reply text of each turn.
+Before each tool call, reason: current phase, whether the last result met expectations, and next strategy in reply text of each turn.
+<summary> (<30 tokens) before reply text.
 - 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

@@ -288,12 +288,18 @@ def _render_stat_badge(is_running):
     secs = max(0, int(end_ts - st.session_state.task_start_ts))
     stats = dict(llmcore.STATS)
     short = lambda n: f'{n / 1000:.0f}k' if n >= 1000 else str(n)
+    _p = []
+    if stats.get('t_start') and stats.get('t_ttft') is not None and stats['t_ttft'] != stats['t_start']:
+        _p.append(f"ttft{stats['t_ttft'] - stats['t_start']:.1f}s")
+    _gen = (stats['t_end'] - max(stats['t_ttft'], stats['t_start'])) if (stats.get('t_end') and stats.get('t_ttft') is not None) else None
+    if _gen and stats.get('out'):
+        _p.append(f"{stats['out'] / _gen:.0f}t/s")
+    _tail = (' │ ' + '·'.join(_p)) if _p else ''
     usage = ((f"{stats['session']} │ " if stats.get('session') else '') +
              f"{short(stats['ctx'])} chars·{stats['msgs']}msgs │ "
-             f"in {short(stats.get('inp', 0))} toks·cached{short(stats.get('cached', 0))}·out{short(stats.get('out', 0))} │ "
+             f"in {short(stats.get('inp', 0))} toks·cached{short(stats.get('cached', 0))}·out{short(stats.get('out', 0))}{_tail}"
              if 'ctx' in stats else '')
-    st.markdown(f'<div class="ga-stat-badge">{usage}{secs // 60}:{secs % 60:02d}</div>',
-                unsafe_allow_html=True)
+    st.markdown(f'<div class="ga-stat-badge">{usage} │ {secs // 60}:{secs % 60:02d}</div>', unsafe_allow_html=True)
 
 
 if not hasattr(agent, "_ui_messages"): agent._ui_messages = st.session_state.get("messages", [])
@@ -359,6 +365,7 @@ if prompt:
         st.session_state.reply_ts = ""
         st.session_state.current_prompt = ""
         st.session_state.last_reply_time = int(time.time())
+        st.session_state.show_full_history = False
         st.rerun()
     def _slash_missing(name):
         st.session_state.messages.extend([
@@ -379,6 +386,11 @@ if prompt:
         target = sessions[idx][0] if 0 <= idx < len(sessions) else None
         result = handle_frontend_command(agent, cmd)
         history = extract_ui_messages(target) if target and result.startswith('✅') else None
+        if history:
+            for x in history:
+                if x['role'] == 'assistant' and len(x['content']) > 120_000:
+                    m = re.search(r'\**LLM Running \(Turn \d+\) \.\.\.\**', x['content'][-120_000:])
+                    x['content'] = x['content'][-120_000 + m.start():] if m else ''
         tail = [{"role": "assistant", "content": result, "time": ts}]
         if history: st.session_state.messages[:] = history + tail
         else: st.session_state.messages.extend([{"role": "user", "content": cmd, "time": ts}] + tail)

llmcore.py

@@ -448,6 +448,8 @@ def _stream_with_retry(sess, url, headers, payload, parse_fn):
         return None if ra and ra > cap else max(0.5, ra or min(30.0, 3.0 * (2 ** attempt)))
     for attempt in range(sess.max_retries + 1):
         streamed = False
+        STATS.update(t_start=time.time(), t_ttft=None)
+        if not sess.stream: STATS['t_ttft'] = STATS['t_start']
         try:
             with requests.post(url, headers=headers, json=payload, stream=sess.stream, 
                                timeout=(sess.connect_timeout, sess.read_timeout), proxies=sess.proxies, verify=sess.verify) as r:
@@ -464,9 +466,15 @@ def _stream_with_retry(sess, url, headers, payload, parse_fn):
                     yield err; return [{"type": "text", "text": err}]
                 gen = parse_fn(r)
                 try:
-                    while True: chunk = next(gen); streamed = True; yield chunk
+                    while True:
+                        if getattr(sess, 'should_stop', None) and sess.should_stop():
+                            STATS['t_end'] = time.time(); return []
+                        chunk = next(gen)
+                        if chunk and STATS.get('t_ttft') is None: STATS['t_ttft'] = time.time()
+                        streamed = True; yield chunk
                 except StopIteration as e:
                     if not e.value and not streamed: raise requests.ConnectionError("empty response")
+                    STATS['t_end'] = time.time()
                     return e.value or []
         except (requests.Timeout, requests.ConnectionError, requests.exceptions.ChunkedEncodingError) as e:
             err = f"!!!Error: {type(e).__name__}: {e}" if str(e) else f"!!!Error: {type(e).__name__}"