Merge remote-tracking branch 'upstream/main' into merge-upstream-main-20260710

assets/supergrok_proxy.py

@@ -6,7 +6,7 @@ SuperGrok Local Proxy - CLI
   python assets/supergrok_proxy.py login
   python assets/supergrok_proxy.py serve --port 15433
   python assets/supergrok_proxy.py models
-  python assets/supergrok_proxy.py test --model grok-4.3
+  python assets/supergrok_proxy.py test --model grok-4.5
 
 GenericAgent/mykey 配置示例：
   native_oai_config_supergrok_proxy = {

frontends/conductor.py

@@ -84,6 +84,20 @@ def _apply_desktop_model(agent: "GenericAgent") -> None:
         _fallback_usable_model(agent)
 
 
+def _select_llm(agent: "GenericAgent", llm: Any) -> bool:
+    if llm is None or str(llm).strip() == "": return False
+    q = str(llm).strip()
+    if isinstance(llm, int) or q.isdigit(): agent.next_llm(int(q)); return True
+    q = q.lower(); agent.load_llm_sessions()
+    for i, c in enumerate(agent.llmclients):
+        vals = []
+        for fn in (lambda: agent.get_llm_name(c), lambda: agent.get_llm_name(c, model=True), lambda: c.backend.name, lambda: c.backend.model):
+            try: vals.append(str(fn()).lower())
+            except Exception: pass
+        if any(q in v for v in vals): agent.next_llm(i); return True
+    raise ValueError(f"llm not found: {llm}")
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     # 服务启动（事件循环已就绪）：捕获 loop 供工作线程跨线程推 WS 广播，并起主agent
@@ -104,6 +118,7 @@ class ChatIn(BaseModel):
 
 class StartSubagentIn(BaseModel):
     prompt: str
+    llm: Any = None
 
 class ApprovalIn(BaseModel):
     prompt: str
@@ -112,6 +127,7 @@ class ApprovalIn(BaseModel):
 class SubagentActionIn(BaseModel):
     action: str = "intervene"  # intervene | abort | kill
     msg: str = ""
+    llm: Any = None
 
 @dataclass
 class SubAgentState:
@@ -257,13 +273,13 @@ class SubagentPool:
                 s.agent.task_queue.put("EXIT")  
                 with self.lock: self.subagents.pop(sid, None)  
             if to_abort: push_cards()
-    def start_subagent(self, prompt: str) -> dict:
+    def start_subagent(self, prompt: str, llm: Any = None) -> dict:
         sid = short_id()
         agent = GenericAgent()
         agent.inc_out = True
         agent.verbose = False
         agent.no_print = True
-        _apply_desktop_model(agent)
+        if not _select_llm(agent, llm): _apply_desktop_model(agent)
         th = start_agent_runner(agent, f"subagent-{sid}")
         state = SubAgentState(id=sid, agent=agent, prompt=prompt, status="running", thread=th)
         with self.lock: self.subagents[sid] = state
@@ -275,10 +291,11 @@ class SubagentPool:
         threading.Thread(target=monitor_display_queue, args=(sid, dq, True), name=f"monitor-{sid}", daemon=True).start()
         push_cards()
         return {"id": sid, "status": "running"}
-    def input_subagent(self, sid: str, msg: str) -> dict:
+    def input_subagent(self, sid: str, msg: str, llm: Any = None) -> dict:
         with self.lock: s = self.subagents.get(sid)
         if not s: return {"error": "subagent not found", "id": sid}
         if s.status == "running": return {"error": "subagent is still running, cannot input/reply. Start a new subagent instead.", "id": sid}
+        _select_llm(s.agent, llm)
         s.prompt = msg
         s.reply = ""
         s.status = "running"
@@ -299,10 +316,10 @@ READMES = {
 Conductor API\tBase: http://{HOST}:{PORT}
 
 POST /chat\tbody: {{"msg": "..."}}\t给用户发消息
-POST /subagent\tbody: {{"prompt": "..."}}\t启动新subagent，返回 {{"id": "xxx"}}
+POST /subagent\tbody: {{"prompt": "..."}}\t启动新subagent，返回 {{"id": "xxx"}}；指定模型加参数llm(数字/名称)
 POST /approval\tbody: {{"prompt": "...", "source": "..."}}\t推一条待批任务到前端(后端不存)，用户同意则直接派发为subagent
 POST /subagent/{{id}}\tbody: {{"action": "keyinfo", "msg": "..."}}\t注入key_info（agent下轮可见）
-POST /subagent/{{id}}\tbody: {{"action": "input", "msg": "..."}}\t开新一轮任务（agent停下后追加）
+POST /subagent/{{id}}\tbody: {{"action": "input", "msg": "..."}}\t开新一轮任务；指定模型加参数llm(数字/名称)
 POST /subagent/{{id}}\tbody: {{"action": "stop"}}\t中断执行但保留（可继续input/reply）
 GET /chat?last=N\t返回最近N条对话（默认20）
 GET /subagent\t返回 {{"items": [...]}}\t查看所有subagent状态
@@ -506,7 +523,7 @@ INSTR_DISPATCHED = "Task received. I'll handle THIS TASK from here. You MUST to
 
 @app.post("/subagent")
 def api_start_subagent(body: StartSubagentIn):
-    result = pool.start_subagent(body.prompt)
+    result = pool.start_subagent(body.prompt, body.llm)
     result["instruction"] = INSTR_DISPATCHED
     return result
 
@@ -520,7 +537,7 @@ def api_subagent_action(sid: str, body: SubagentActionIn):
         result["instruction"] = "Received. I'll incorporate this. You MUST to do other task or end your reply."
         return result
     if action in ("input", "reply", "append", "message", "msg"):
-        result = pool.input_subagent(sid, body.msg)
+        result = pool.input_subagent(sid, body.msg, body.llm)
         result["instruction"] = INSTR_DISPATCHED
         return result
     if action in ("abort", "stop"):
@@ -533,10 +550,11 @@ def api_subagent_action(sid: str, body: SubagentActionIn):
 
 @app.get("/chat")
 def api_get_chat(last: int = 20):
+    items = [m.copy() for m in chat_messages[-last:]]
     for m in chat_messages:
         if m.get("role") == "user" and not m.get("read"): m["read"] = True
     schedule_broadcast({"type": "chat_read"})
-    return {"items": chat_messages[-last:]}
+    return {"items": items}
 
 @app.post("/chat")
 def api_chat(body: ChatIn):

frontends/stapp.py

@@ -88,6 +88,7 @@ def get_controller():
         threading.Thread(target=ag.run, daemon=True).start()
         while True:
             b['ev'].wait(); b['ev'].clear()
+            if ag.llm_no != agent.llm_no: ag.next_llm(agent.llm_no)
             dq = ag.put_task(build_prompt(b['obj']), source="controller")
             while 'done' not in (it := dq.get()): pass
             ms = re.findall(r'<next_prompt>(.*?)</next_prompt>', it['done'], re.S)
@@ -278,7 +279,8 @@ def render_main_stream(prompt=None):
             b = get_controller()
             b['obj'] = st.session_state.get('loop_prompt_input', ''); b['ready'] = False; b['ev'].set()
 
-if "messages" not in st.session_state: st.session_state.messages = []
+if not hasattr(agent, "_ui_messages"): agent._ui_messages = st.session_state.get("messages", [])
+if "messages" not in st.session_state: st.session_state.messages = agent._ui_messages
 for msg in st.session_state.messages:
     with st.chat_message(msg["role"]):
         # 用 slot=st.empty() + with slot.container(): ... 的外壳，DOM 路径和流式渲染完全一致，跨 rerun 对齐

frontends/worldline.py

@@ -1081,10 +1081,9 @@ def render_native_history(history, ts=None):
     `_parse_native_history(_pairs(render_native_history(h))) == h`.
 
     Byte-for-byte matches llmcore's `_write_llm_log` for the tool backend:
-    Prompt = `json.dumps(merged, ensure_ascii=False, indent=2)`; Response =
-    `str(content_blocks)` which for a list equals `repr(...)` (survives
-    True/False/None that JSON would emit as true/false/null and break
-    `ast.literal_eval`). Only complete (user, assistant) pairs are emitted."""
+    Prompt/Response each put one content block per line, with their wrappers
+    attached to the first/last lines. Response keeps Python repr so
+    True/False/None survive `ast.literal_eval`. Only complete pairs are emitted."""
     ts = ts or datetime.now().strftime('%Y-%m-%d %H:%M:%S')
     out, i, n = [], 0, len(history)
     while i < n:
@@ -1092,8 +1091,13 @@ def render_native_history(history, ts=None):
         if isinstance(u, dict) and u.get('role') == 'user' and i + 1 < n:
             a = history[i + 1]
             if isinstance(a, dict) and a.get('role') == 'assistant':
-                out.append(f"=== Prompt === {ts}\n{json.dumps(u, ensure_ascii=False, indent=2)}\n\n")
-                out.append(f"=== Response === {ts}\n{a.get('content')!r}\n\n")
+                prompt_blocks = u.get('content', [])
+                prompt_raw = '{"role": "user", "content": [\n' + ",\n".join(
+                    json.dumps(b, ensure_ascii=False) for b in prompt_blocks) + "]}"
+                out.append(f"=== Prompt === {ts}\n{prompt_raw}\n\n")
+                blocks = a.get('content', [])
+                raw = "[" + ",\n".join(repr(b) for b in blocks) + "]"
+                out.append(f"=== Response === {ts}\n{raw}\n\n")
                 i += 2
                 continue
         i += 1

ga.py

@@ -400,6 +400,7 @@ class GenericAgentHandler(BaseHandler):
                 with open(path, 'a' if mode == "append" else 'w', encoding="utf-8") as f: f.write(new_content)
             yield f"[Status] ✅ {mode.capitalize()} 成功 ({len(new_content)} bytes)\n"
             next_prompt = self._get_anchor_prompt(skip=args.get('_index', 0) > 0)
+            if len(new_content) > 5000: next_prompt = "\n[SYSTEM TIPS] WRITE TOO LONG! MUST RECHECK HALLUCINATIONS! SMALL WRITES OR PATCHES NEXT TIME!"
             return StepOutcome({"status": "success", 'writed_bytes': len(new_content)}, next_prompt=next_prompt)
         except Exception as e:
             yield f"[Status] ❌ 写入异常: {str(e)}\n"
@@ -427,6 +428,7 @@ class GenericAgentHandler(BaseHandler):
     
     def export_history(self, fn):
         with open(fn, 'w', encoding='utf-8') as f: json.dump(self.parent.llmclient.backend.history, f, ensure_ascii=False)
+    def enter_project_mode(self, name): self.parent._ga_project_mode_name = name
     def _in_plan_mode(self): return self.working.get('in_plan_mode')
     def _exit_plan_mode(self): self.working.pop('in_plan_mode', None)
     def enter_plan_mode(self, plan_path): 
@@ -539,9 +541,9 @@ class GenericAgentHandler(BaseHandler):
     def _get_anchor_prompt(self, skip=False):
         if skip: return "\n"
         h = self.history_info; W = 30
-        earlier = f'<earlier_context>\n{self._fold_earlier(h[:-W])}\n</earlier_context>\n' if len(h) > W else ""
-        h_str = "\n".join(h[-W:])
-        prompt = f"\n### [WORKING MEMORY]\n{earlier}<history>\n{h_str}\n</history>"
+        earlier = f'<earlier_context>\n{self._fold_earlier(h[:-W])}\n</earlier_context>\n' if len(h) > W and self.current_turn % 4 == 1 else ""
+        history = f'<history>\n{"\n".join(h[-W:])}\n</history>' if self.current_turn % 2 == 1 else ""
+        prompt = f"\n### [WORKING MEMORY]\n{earlier}{history}"
         prompt += f"\nCurrent turn: {self.current_turn}\n"
         if self.working.get('key_info'): prompt += f"\n<key_info>{self.working.get('key_info')}</key_info>"
         if self.working.get('related_sop'): prompt += f"\n有不清晰的地方请再次读取{self.working.get('related_sop')}"

llmcore.py

@@ -117,6 +117,10 @@ def auto_make_url(base, path):
     return f"{b}/{p}" if re.search(r'/v\d+(/|$)', b) else f"{b}/v1/{p}"
 
 def _parse_claude_json(data):
+    if data.get("stop_reason") == "refusal":
+        err = "[Error: Claude refusal]"
+        yield err
+        return [{"type": "text", "text": err}]
     content_blocks = data.get("content", [])
     _record_usage(data.get("usage", {}), "messages")
     for b in content_blocks:
@@ -182,6 +186,7 @@ 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 !!!]"
+        elif stop_reason == "refusal": warn = "\n\n[Error: Claude refusal]"
     if current_block:
         if current_block["type"] == "tool_use":
             try: current_block["input"] = json.loads(tool_json_buf) if tool_json_buf else {}
@@ -552,6 +557,7 @@ class BaseSession:
         self.service_tier = _enum('service_tier', {'auto', 'default', 'priority', 'flex'})
         self.thinking_type = _enum('thinking_type', {'adaptive', 'enabled', 'disabled'})
         self.thinking_budget_tokens = cfg.get('thinking_budget_tokens')
+        self.omit_thinking = cfg.get('omit_thinking', False)  # Exclude thinking from session history
         mode = str(cfg.get('api_mode', 'chat_completions')).strip().lower().replace('-', '_')
         self.api_mode = 'responses' if mode in ('responses', 'response') else 'chat_completions'
         self.temperature = cfg.get('temperature', 1)
@@ -730,7 +736,9 @@ class NativeClaudeSession(BaseSession):
         except StopIteration as e: content_blocks = e.value or []
         if content_blocks and (_injected := _ensure_text_block(content_blocks)): yield _injected
         if content_blocks and not (len(content_blocks) == 1 and content_blocks[0].get("text", "").startswith("!!!Error:")):
-            self.history.append({"role": "assistant", "content": content_blocks})
+            history_blocks = content_blocks
+            if self.omit_thinking: history_blocks = [b for b in content_blocks if b.get("type") != "thinking"]
+            self.history.append({"role": "assistant", "content": history_blocks})
         text_parts = [b["text"] for b in content_blocks if b.get("type") == "text"]
         content = "\n".join(text_parts).strip()
         tool_calls = [MockToolCall(b["name"], b.get("input", {}), id=b.get("id", "")) for b in content_blocks if b.get("type") == "tool_use"]
@@ -743,7 +751,8 @@ class NativeClaudeSession(BaseSession):
             if think_match:
                 thinking = think_match.group(1).strip()
                 content = re.sub(think_pattern, "", content, flags=re.DOTALL)
-        return MockResponse(thinking, content, tool_calls, str(content_blocks))
+        raw = "[" + ",\n".join(repr(b) for b in content_blocks) + "]"
+        return MockResponse(thinking, content, tool_calls, raw)
 
 class NativeOAISession(NativeClaudeSession):
     native_ua = "codex_exec/0.139.0 (Windows 10.0.26200; x86_64) unknown (codex_exec; 0.139.0)"
@@ -1055,7 +1064,8 @@ class NativeToolClient:
         final_content = tool_result_blocks + filtered_content
         if not final_content: final_content = [{"type": "text", "text": "."}]
         merged = {"role": "user", "content": final_content}
-        _write_llm_log('Prompt', json.dumps(merged, ensure_ascii=False, indent=2), self.log_path)
+        prompt_raw = '{"role": "user", "content": [\n' + ",\n".join(json.dumps(b, ensure_ascii=False) for b in final_content) + "]}"
+        _write_llm_log('Prompt', prompt_raw, self.log_path)
         gen = self.backend.ask(merged)
         try:
             while True: 

memory/L4_raw_sessions/compress_session.py

@@ -156,6 +156,7 @@ def batch_process(src, l4_dir=None, dry_run=True):
     l4_dir = os.path.normpath(l4_dir or L4_DIR)
     raw_files = sorted(src) if isinstance(src, (list, tuple)) else \
                 sorted(glob.glob(os.path.join(src, 'model_responses_*.txt')))
+    raw_files = sorted(raw_files, key=os.path.getmtime)[:-10]  # always retain 10 newest raw files
     if not raw_files:
         print("No raw files found"); return {'processed': 0, 'skipped': 0, 'errors': 0, 'new_sessions': 0}
 

memory/project_mode_sop.md

@@ -7,14 +7,14 @@ Project Mode = 跨会话保持项目认知的工作模式
 
 ## 进入
 
-锚 = `./.active_project.<宿主pid>`，只对当前 GA 进程有效：GA 关闭即自动失活；多开 GA 各自激活不同项目，互不干扰。（下文路径一律以 cwd 为基准，cwd 即 GA 的 temp 目录；禁写 `temp/xxx` 前缀，会嵌套出 temp/temp）
+激活态保存在当前 Agent 实例；各会话互不干扰，关闭 GA 自动失效。（下文路径以 cwd 为基准，禁写 `temp/xxx` 前缀）
 
 - 用户只说「进入项目模式」未指明项目：列出 `./projects/` 下各项目（名字 + memory 行数 + 最后修改时间），ask_user 让用户选定后再继续
 - 用户明确说「进入/切换到 <项目名> 项目」：视为已确认，直接执行：
 
 1. 建目录 `./projects/<项目名>/`，无则创建 `project_memory.md`（空文件即可）
-2. 写文件锚，必须用 code_run（锚文件名含宿主 pid，ppid 即 GA 宿主进程）：
-   `import os; open(f'./.active_project.{os.getppid()}', 'w').write('<项目名>')`
+2. 用 `code_run` 的 `inline_eval=true` 绑定当前 Agent（`handler` 已注入，无需 import）：
+   `handler.enter_project_mode('<项目名>')`
 3. 回读 `project_memory.md` 全文，向用户复述项目现状
 
 ## 期间纪律
@@ -25,6 +25,5 @@ Project Mode = 跨会话保持项目认知的工作模式
 
 ## 离开
 
-用户表示「离开项目模式」时：删除 `./.active_project.<宿主pid>`（仅关闭激活态，项目目录与 `project_memory.md` 原样保留）
-切换到另一项目时无需先离开，直接按「进入」覆盖文件锚即可
-GA 关闭不需任何操作：锚随进程消亡自动失效，残留文件由插件下次启动时清扫
+明确要求离开时，用 `code_run`（`inline_eval=true`）执行 `handler.enter_project_mode(None)`。
+切换项目直接换项目名

plugins/project_mode.py

@@ -6,11 +6,7 @@
 由模型按 L1 中的指针与线索（行数/大小）自行判断是否用 file 工具读取。
 利用 messages 是 list 引用的事实，直接 mutate 即反映到真正发给 LLM 的内容。
 
-激活态载体 = 文件锚 temp/.active_project.<宿主pid>（存当前项目名）。PID 键控：
-  - 锚只对写它的那个 GA 进程有效 → 多开 GA 各自激活不同项目，互不可见
-  - GA 关闭即自动失活（重启后 pid 变，旧锚作废）；重新激活需经用户确认（SOP 指示）
-  - 进入：agent 读 project_mode_sop，经用户确认后写锚（code_run 中 os.getppid() 即宿主 pid）
-  - 退出：删除该文件。插件加载时清扫旧版无后缀锚与自己 pid 的前世残留（不碰他进程的锚）
+激活态保存在当前 GenericAgent 实例的 _ga_project_mode_name；各会话互不干扰。
 
 目录约定：
   temp/projects/<项目名>/project_memory.md   单文件全文注入的项目记忆
@@ -21,43 +17,13 @@ import plugins.hooks as hooks
 
 _PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 _TEMP = os.path.join(_PROJECT_ROOT, 'temp')
-_ANCHOR = os.path.join(_TEMP, f'.active_project.{os.getpid()}')
-
-
-def _cleanup_stale_anchors():
-    """清扫无主锚：仅删两类能纯靠文件名判定的——旧版无后缀的、自己 pid 的（刚启动
-    不可能写过锚，必是 pid 复用的前世残留）。他进程的锚一律不碰：故意不做存活探测，
-    os.kill(pid, 0) 在 Windows 上语义是终止进程而非探活，任何平台分支都不值得为
-    删除无害残留文件引入杀进程风险（激活判定只认自己 pid 的锚，残留不影响行为）。"""
-    import glob
-    for path in glob.glob(os.path.join(_TEMP, '.active_project*')):
-        pid = path.rsplit('.', 1)[-1]
-        if path != _ANCHOR and pid.isdigit():
-            continue  # 他进程的锚（含已死进程的），不碰
-        try: os.remove(path)
-        except OSError: pass
-
-
-_cleanup_stale_anchors()
 
 
 def _active_project(ctx=None):
-    """返回当前激活的项目名；未激活返回 None。
-
-    兼容策略:
-      - 新 TUI 多会话可在当前 GenericAgent 实例上设置 _ga_project_mode_name。
-        只要该属性存在,就以它为准;值为 None/空串表示该 agent 普通模式。
-      - 其它 UI / 旧 SOP 不设置该属性,继续读取 pid 键控文件锚。
-    异常不在此捕获——hooks.trigger 统一捕获并打印，保持可观测。"""
-    parent = None
-    if isinstance(ctx, dict):
-        handler = ctx.get('handler')
-        parent = getattr(handler, 'parent', None)
-    if parent is not None and hasattr(parent, '_ga_project_mode_name'):
-        return getattr(parent, '_ga_project_mode_name', None) or None
-    if not os.path.isfile(_ANCHOR):
-        return None
-    return open(_ANCHOR, encoding='utf-8').read().strip() or None
+    """返回当前 Agent 实例激活的项目名；未激活返回 None。"""
+    handler = ctx.get('handler') if isinstance(ctx, dict) else None
+    parent = getattr(handler, 'parent', None)
+    return getattr(parent, '_ga_project_mode_name', None) or None
 
 
 def _project_dir(name):