hotreload mykeys, fix SSE residual block, unify context_win, cleanup

agentmain.py

@@ -1,16 +1,12 @@
 import os, sys, threading, queue, time, json, re, random, locale
 os.environ.setdefault('GA_LANG', 'zh' if any(k in (locale.getlocale()[0] or '').lower() for k in ('zh', 'chinese')) else 'en')
-try:
-    import readline
-except Exception:
-    readline = None
 if sys.stdout is None: sys.stdout = open(os.devnull, "w")
 elif hasattr(sys.stdout, 'reconfigure'): sys.stdout.reconfigure(errors='replace')
 if sys.stderr is None: sys.stderr = open(os.devnull, "w")
 elif hasattr(sys.stderr, 'reconfigure'): sys.stderr.reconfigure(errors='replace')
 sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
 
-from llmcore import LLMSession, ToolClient, ClaudeSession, MixinSession, NativeToolClient, NativeClaudeSession, NativeOAISession
+from llmcore import reload_mykeys, LLMSession, ToolClient, ClaudeSession, MixinSession, NativeToolClient, NativeClaudeSession, NativeOAISession
 from agent_loop import agent_runner_loop
 from ga import GenericAgentHandler, smart_format, get_global_memory, format_error, consume_file
 
@@ -47,7 +43,20 @@ class GeneraticAgent:
     def __init__(self):
         script_dir = os.path.dirname(os.path.abspath(__file__))
         os.makedirs(os.path.join(script_dir, 'temp'), exist_ok=True)
-        from llmcore import mykeys
+        self.lock = threading.Lock()
+        self.task_dir = None
+        self.history = []
+        self.task_queue = queue.Queue() 
+        self.is_running = False; self.stop_sig = False
+        self.llm_no = 0;  self.inc_out = False
+        self.handler = None; self.verbose = True
+        self.load_llm_sessions()
+
+    def load_llm_sessions(self):
+        mykeys, changed = reload_mykeys()
+        if not changed and hasattr(self, 'llmclients'): return
+        try: oldhistory = self.llmclient.backend.history
+        except: oldhistory = None
         llm_sessions = []
         for k, cfg in mykeys.items():
             if not any(x in k for x in ['api', 'config', 'cookie']): continue
@@ -66,16 +75,11 @@ class GeneraticAgent:
                     else: llm_sessions[i] = ToolClient(mixin)
                 except Exception as e: print(f'[WARN] Failed to init MixinSession with cfg {s["mixin_cfg"]}: {e}')
         self.llmclients = llm_sessions
-        self.lock = threading.Lock()
-        self.task_dir = None
-        self.history = []
-        self.task_queue = queue.Queue() 
-        self.is_running = False; self.stop_sig = False
-        self.llm_no = 0;  self.inc_out = False
-        self.handler = None; self.verbose = True
-        self.llmclient = self.llmclients[self.llm_no]
-
+        self.llmclient = self.llmclients[self.llm_no%len(self.llmclients)]
+        if oldhistory: self.llmclient.backend.history = oldhistory
+    
     def next_llm(self, n=-1):
+        self.load_llm_sessions()
         self.llm_no = ((self.llm_no + 1) if n < 0 else n) % len(self.llmclients)
         lastc = self.llmclient
         self.llmclient = self.llmclients[self.llm_no]
@@ -85,7 +89,9 @@ class GeneraticAgent:
         name = self.get_llm_name(model=True)
         if 'glm' in name or 'minimax' in name or 'kimi' in name: load_tool_schema('_cn')
         else: load_tool_schema()
-    def list_llms(self): return [(i, self.get_llm_name(b), i == self.llm_no) for i, b in enumerate(self.llmclients)]
+    def list_llms(self): 
+        self.load_llm_sessions()
+        return [(i, self.get_llm_name(b), i == self.llm_no) for i, b in enumerate(self.llmclients)]
     def get_llm_name(self, b=None, model=False):
         b = self.llmclient if b is None else b
         if isinstance(b, dict): return 'BADCONFIG_MIXIN'
@@ -139,12 +145,8 @@ class GeneraticAgent:
                 handler.working['passed_sessions'] = ps = self.handler.working.get('passed_sessions', 0) + 1
                 if ps > 0: handler.working['key_info'] += f'\n[SYSTEM] 此为 {ps} 个对话前设置的key_info，若已在新任务，先更新或清除工作记忆。\n'
             self.handler = handler
-            user_input = raw_query
-            if source == 'feishu' and len(self.history) > 1:   # 如果有历史记录且来自飞书，注入到首轮 user_input 中（支持/restore恢复上下文）
-                user_input = handler._get_anchor_prompt() + f"\n\n### 用户当前消息\n{raw_query}"
-            #if 'gpt' in self.get_llm_name(model=True): handler._done_hooks.append('请确定任务是否完成，如果完成请给出信息完整的简报回答，如未完成需要继续工具调用直到完成任务，确实需要问用户应使用ask_user工具')
             # although new handler, the **full** history is in llmclient, so it is full history!
-            gen = agent_runner_loop(self.llmclient, sys_prompt, user_input, 
+            gen = agent_runner_loop(self.llmclient, sys_prompt, raw_query, 
                                 handler, TOOLS_SCHEMA, max_turns=70, verbose=self.verbose)
             try:
                 full_resp = ""; last_pos = 0
@@ -251,6 +253,8 @@ if __name__ == '__main__':
                 except Exception as e: print(f'[Reflect] on_done error: {e}')
             if getattr(mod, 'ONCE', False): print('[Reflect] ONCE=True, exiting.'); break
     else:
+        try: import readline
+        except Exception: pass
         agent.inc_out = True
         while True:
             q = input('> ').strip()

assets/code_run_header.py

@@ -24,7 +24,4 @@ def _pinit(self, *a, **k):
     if os.name == 'nt': k['creationflags'] = (k.get('creationflags') or 0) | 0x08000000
     _Pi(self, *a, **k)
 subprocess.Popen.__init__ = _pinit
-if hasattr(os, 'startfile'):
-    def _nosf(*a, **k): raise RuntimeError("startfile disabled in sandbox")
-    os.startfile = _nosf
 sys.excepthook = lambda t, v, tb: (sys.__excepthook__(t, v, tb), print(f"\n[Agent Hint]: NO GUESSING! You MUST probe first. If missing common package, pip.")) if issubclass(t, (ImportError, AttributeError)) else sys.__excepthook__(t, v, tb)

frontends/stapp.py

@@ -36,13 +36,16 @@ if 'autonomous_enabled' not in st.session_state: st.session_state.autonomous_ena
 
 @st.fragment
 def render_sidebar():
+    llm_options = agent.list_llms()
     current_idx = agent.llm_no
-    st.caption(f"LLM Core: {current_idx}: {agent.get_llm_name()}", help="点击切换备用链路")
+    llm_labels = {idx: f"{idx}: {(name or '').strip()}" for idx, name, _ in llm_options}
+    st.caption(f"LLM Core: {llm_labels.get(current_idx, str(current_idx))}", help="下拉切换备用链路")
+    selected_idx = st.selectbox("备用链路", [idx for idx, _, _ in llm_options], index=next((i for i, (idx, _, _) in enumerate(llm_options) if idx == current_idx), 0), format_func=llm_labels.get, label_visibility="collapsed", key="sidebar_llm_select")
+    if selected_idx != current_idx:
+        agent.next_llm(selected_idx); st.rerun(scope="fragment")
     last_reply_time = st.session_state.get('last_reply_time', 0)
     if last_reply_time > 0:
         st.caption(f"空闲时间：{int(time.time()) - last_reply_time}秒", help="当超过30分钟未收到回复时，系统会自动任务")
-    if st.button("切换备用链路"):
-        agent.next_llm(); st.rerun(scope="fragment")
     if st.button("强行停止任务"):
         agent.abort(); st.toast("已发送停止信号"); st.rerun()
     if st.button("重新注入工具"):

ga.py

@@ -445,7 +445,7 @@ class GenericAgentHandler(BaseHandler):
         if not response or (not content.strip() and not thinking.strip()):
             yield "[Warn] LLM returned an empty response. Retrying...\n"
             return StepOutcome({}, next_prompt="[System] Blank response, regenerate and tooluse")
-        if len(content) > 100 and ('未收到完整响应 !!!]' in content[-100:] or '!!!Error: [SSL:' in content[-100:]):
+        if len(content) > 50 and ('未收到完整响应 !!!]' in content[-100:] or '!!!Error: [SSL:' in content[-100:]):
             return StepOutcome({}, next_prompt="[System] Incomplete response. Regenerate and tooluse.")
         if 'max_tokens !!!]' in content[-100:]:
             return StepOutcome({}, next_prompt="[System] max_tokens limit reached. Use multi small steps to do it.")
@@ -490,10 +490,11 @@ class GenericAgentHandler(BaseHandler):
         '''Agent觉得当前任务完成后有重要信息需要记忆时调用此工具。'''
         prompt = '''### [总结提炼经验] 既然你觉得当前任务有重要信息需要记忆，请提取最近一次任务中【事实验证成功且长期有效】的环境事实、用户偏好、重要步骤，更新记忆。
 本工具是标记开启结算过程，若已在更新记忆过程或没有值得记忆的点，忽略本次调用。
-**提取行动验证成功的信息**：
+**如果没有经验证的，未来能用上的信息，忽略本次调用！**
+**只能提取行动验证成功的信息**：
 - **环境事实**（路径/凭证/配置）→ `file_patch` 更新 L2，同步 L1
 - **复杂任务经验**（关键坑点/前置条件/重要步骤）→ L3 精简 SOP（只记你被坑得多次重试的核心要点）
-**禁止**：临时变量、具体推理过程、未验证信息、通用常识、你可以轻松复现的细节。
+**禁止**：临时变量、具体推理过程、未验证信息、通用常识、你可以轻松复现的细节、只是做了但没有验证的信息
 **操作**：严格遵循提供的L0的记忆更新SOP。先 `file_read` 看现有 → 判断类型 → 最小化更新 → 无新内容跳过，保证对记忆库最小局部修改。\n
 ''' + get_global_memory()
         yield "[Info] Start distilling good memory for long-term storage.\n"

llmcore.py

@@ -1,26 +1,33 @@
-import os, json, re, time, requests, sys, threading, urllib3, base64, mimetypes, uuid
+import os, json, re, time, requests, sys, threading, urllib3, base64, importlib, uuid
 from datetime import datetime
 urllib3.disable_warnings(urllib3.exceptions.InsecureRequestWarning)
 _RESP_CACHE_KEY = str(uuid.uuid4())
 
 def _load_mykeys():
+    global _mykey_path
     try:
-        import mykey; return {k: v for k, v in vars(mykey).items() if not k.startswith('_')}
+        import mykey; importlib.reload(mykey); _mykey_path = mykey.__file__
+        return {k: v for k, v in vars(mykey).items() if not k.startswith('_')}
     except ImportError: pass
-    p = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'mykey.json')
+    _mykey_path = p = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'mykey.json')
     if not os.path.exists(p): raise Exception('[ERROR] mykey.py or mykey.json not found, please create one from mykey_template.')
     with open(p, encoding='utf-8') as f: return json.load(f)
 
+_mykey_path = _mykey_mtime = None
+def reload_mykeys():
+    global _mykey_mtime
+    mt = os.stat(_mykey_path).st_mtime_ns if _mykey_path else -1
+    if mt == _mykey_mtime: return globals().get('mykeys', {}), False
+    mk = _load_mykeys(); _mykey_mtime = os.stat(_mykey_path).st_mtime_ns
+    print(f'[Info] Load mykeys from {_mykey_path}')
+    globals().update(mykeys=mk)
+    if mk.get('langfuse_config'):
+        try: from plugins import langfuse_tracing
+        except Exception: pass
+    return mk, True
+
 def __getattr__(name):  # once guard in PEP 562
-    if name in ('mykeys', 'proxies'):  
-        mk = _load_mykeys()
-        proxy = mk.get("proxy", 'http://127.0.0.1:2082')
-        px = {"http": proxy, "https": proxy} if proxy else None
-        globals().update(mykeys=mk, proxies=px)
-        if mk.get('langfuse_config'):
-            try: from plugins import langfuse_tracing
-            except Exception: pass
-        return globals()[name]
+    if name == 'mykeys': return reload_mykeys()[0]
     raise AttributeError(f"module 'llmcore' has no attribute {name}")
 
 def compress_history_tags(messages, keep_recent=10, max_len=800, force=False):
@@ -158,6 +165,11 @@ def _parse_claude_sse(resp_lines):
     if not warn:
         if not got_message_stop and not stop_reason: warn = "\n\n[!!! 流异常中断，未收到完整响应 !!!]"
         elif stop_reason == "max_tokens": warn = "\n\n[!!! Response truncated: max_tokens !!!]"
+    if current_block:
+        if current_block["type"] == "tool_use":
+            try: current_block["input"] = json.loads(tool_json_buf) if tool_json_buf else {}
+            except: current_block["input"] = {"_raw": tool_json_buf}
+        content_blocks.append(current_block); current_block = None
     if warn:
         print(f"[WARN] {warn.strip()}")
         content_blocks.append({"type": "text", "text": warn}); yield warn
@@ -488,7 +500,7 @@ class BaseSession:
         self.api_key = cfg['apikey']
         self.api_base = cfg['apibase'].rstrip('/')
         self.model = cfg.get('model', '')
-        self.context_win = cfg.get('context_win', 24000)
+        self.context_win = cfg.get('context_win', 28000)
         self.history = []
         self.lock = threading.Lock()
         self.system = ""
@@ -598,7 +610,6 @@ def _fix_messages(messages):
 class NativeClaudeSession(BaseSession):
     def __init__(self, cfg):
         super().__init__(cfg)
-        self.context_win = cfg.get("context_win", 28000)
         self.fake_cc_system_prompt = cfg.get("fake_cc_system_prompt", False)
         self.user_agent = cfg.get("user_agent", "claude-cli/2.1.113 (external, cli)")
         self._session_id = str(uuid.uuid4())