2965 lines
120 KiB
Plaintext
2965 lines
120 KiB
Plaintext
commit 9b20ca82972ec66622193846801630fd356ce231
|
||
Author: Liang Jiaqing <l.j.q.light@gmail.com>
|
||
Date: Fri Jan 16 23:50:19 2026 +0800
|
||
|
||
fix: restore files removed by mistake and keep zip ignored
|
||
|
||
diff --git a/README.md b/README.md
|
||
new file mode 100644
|
||
index 0000000..4b399e9
|
||
--- /dev/null
|
||
+++ b/README.md
|
||
@@ -0,0 +1,52 @@
|
||
+# pc-agent-loop
|
||
+
|
||
+pc-agent-loop 是一个**极致简约**的 PC 级自主 AI Agent 框架。它通过不到 100 行的核心代码和约 200 行的工具实现，构筑了把整个pc给它（浏览器、终端、文件系统）的物理级自动化能力。
|
||
+
|
||
+## 🚀 核心特性
|
||
+
|
||
+- **极简设计**: 仅由 **7 个基本工具** 和一个高效的 **Agentic Loop** 构成，拒绝过度设计。
|
||
+- **自主代码执行 (Code Execution)**: 能够根据任务需求自主编写并运行 Python 或 PowerShell 脚本，直接操控系统资源。
|
||
+- **深度 Web 自动化 (Advanced Web Automation)**:
|
||
+ - **语义化扫描**: 自动清洗 HTML 内容，将复杂的 DOM 转化为 AI 易读的结构。
|
||
+ - **JS 注入执行**: 在浏览器上下文中执行自定义 JavaScript，实现精准点击、滚动或数据抓取。
|
||
+ - **TMWebDriver**: 支持通过 Tampermonkey 实现的持久化会话驱动。
|
||
+- **精准文件编辑 (Smart File Patching)**: 并非盲目覆盖，而是支持通过 `file_patch` 以代码块匹配方式进行精确修改。
|
||
+- **人机协作模式 (Human-in-the-loop)**: 在遇到验证码、关键权限或模糊决策时，主动请求用户介入。
|
||
+
|
||
+## 📂 项目结构
|
||
+
|
||
+- `agent_loop.py`: **核心引擎**，负责“感知-思考-行动”的自主循环逻辑。
|
||
+- `ga.py`: **工具箱**，定义了 7 大核心原子工具的具体实现。
|
||
+- `agentapp.py`: 基于 Streamlit 构建的轻量化交互式 Web 界面。
|
||
+- `sidercall.py`: LLM 通信层，支持流式输出与 API 调用。
|
||
+- `TMWebDriver.py`: 浏览器驱动模块（需配合 Tampermonkey 脚本使用）。
|
||
+
|
||
+## 🛠️ 快速开始
|
||
+
|
||
+### 1. 环境准备
|
||
+- 安装 Python 3.8+。
|
||
+- （可选）若需网页自动化，请在浏览器中安装 **Tampermonkey** 插件并导入本项目提供的对应脚本。
|
||
+
|
||
+### 2. 安装依赖
|
||
+缺啥装啥
|
||
+
|
||
+### 3. 启动应用
|
||
+在项目根目录下执行：
|
||
+```bash
|
||
+python launch.pyw
|
||
+```
|
||
+
|
||
+## 🧩 7 大核心工具
|
||
+
|
||
+Agent 仅依靠以下 7 个原子工具的组合来完成复杂任务：
|
||
+
|
||
+1. **`code_run`**: 针对 Windows 优化的双模态代码执行器（Python/PowerShell）。
|
||
+2. **`web_scan`**: 获取网页清洗后的语义化 HTML 结构，支持多标签管理。
|
||
+3. **`web_execute_js`**: 网页 JS 脚本注入，支持将结果存盘分析。
|
||
+4. **`file_read`**: 分页式文件读取，支持行号定位。
|
||
+5. **`file_write`**: 文件全量写入或追加。
|
||
+6. **`file_patch`**: 基于源码块匹配的精准局部修改，确保缩进一致性。
|
||
+7. **`ask_user`**: 关键节点请求人类干预。
|
||
+
|
||
+---
|
||
+**⚠️ 警告**: 本 Agent 具备执行本地代码和控制操作系统的物理权限。请务必在受信任的环境中运行，并在运行前仔细检查 Agent 的执行意图。
|
||
\ No newline at end of file
|
||
diff --git a/TMWebDriver.py b/TMWebDriver.py
|
||
new file mode 100644
|
||
index 0000000..0f58e79
|
||
--- /dev/null
|
||
+++ b/TMWebDriver.py
|
||
@@ -0,0 +1,285 @@
|
||
+import json, threading, time, uuid, queue, socket, requests
|
||
+from typing import Dict, Any, Optional, List
|
||
+from simple_websocket_server import WebSocketServer, WebSocket
|
||
+from bs4 import BeautifulSoup
|
||
+import bottle, random
|
||
+from bottle import route, template, request, response
|
||
+
|
||
+class Session:
|
||
+ def __init__(self, session_id, info, client=None):
|
||
+ self.id = session_id
|
||
+ self.info = info
|
||
+ self.connect_at = time.time()
|
||
+ self.disconnect_at = None
|
||
+ self.type = info.get('type', 'ws')
|
||
+ self.ws_client = client if self.type == 'ws' else None
|
||
+ self.http_queue = client if self.type == 'http' else None
|
||
+ @property
|
||
+ def url(self): return self.info.get('url', '')
|
||
+ def is_active(self):
|
||
+ return self.disconnect_at is None
|
||
+ def reconnect(self, client, info):
|
||
+ self.info = info
|
||
+ self.type = info.get('type', 'ws')
|
||
+ if self.type == 'ws':
|
||
+ self.ws_client = client
|
||
+ self.http_queue = None
|
||
+ elif self.type == 'http':
|
||
+ self.http_queue = client
|
||
+ self.connect_at = time.time()
|
||
+ self.disconnect_at = None
|
||
+ def mark_disconnected(self):
|
||
+ self.disconnect_at = time.time()
|
||
+
|
||
+
|
||
+class TMWebDriver:
|
||
+ def __init__(self, host: str = 'localhost', port: int = 18765):
|
||
+ self.host = host
|
||
+ self.port = port
|
||
+ self.sessions = {}
|
||
+ self.results = {}
|
||
+
|
||
+ self.default_session_id = None
|
||
+ self.latest_session_id = None
|
||
+ self.last_cmd_time = 0
|
||
+ self.is_remote = socket.socket().connect_ex((host, port+1)) == 0
|
||
+ if not self.is_remote:
|
||
+ self.start_ws_server()
|
||
+ self.start_http_server()
|
||
+ else:
|
||
+ self.remote = f'http://{self.host}:{self.port+1}/link'
|
||
+
|
||
+ def start_http_server(self):
|
||
+ self.app = app = bottle.Bottle()
|
||
+
|
||
+ @app.route('/api/longpoll', method=['GET', 'POST'])
|
||
+ def long_poll():
|
||
+ data = request.json
|
||
+ session_id = data.get('sessionId')
|
||
+ session_info = {'url': data.get('url'), 'title': data.get('title', ''), 'type': 'http'}
|
||
+ if session_id not in self.sessions:
|
||
+ session = Session(session_id, session_info, queue.Queue())
|
||
+ print(f"Browser http connected: {session.url} (Session: {session_id})")
|
||
+ self.sessions[session_id] = session
|
||
+ session = self.sessions[session_id]
|
||
+ if session.type == 'http': msgQ = session.http_queue
|
||
+ else: return json.dumps({"id": "", "ret": "use ws"})
|
||
+ try: return msgQ.get(timeout=5)
|
||
+ except queue.Empty: return json.dumps({"id": "", "ret": "next long-poll"})
|
||
+
|
||
+ @app.route('/api/result', method=['GET','POST'])
|
||
+ def result():
|
||
+ data = request.json
|
||
+ if data.get('type') == 'result':
|
||
+ self.results[data.get('id')] = {'success': True, 'data': data.get('result'), 'newTabs': data.get('newTabs', [])}
|
||
+ elif data.get('type') == 'error':
|
||
+ self.results[data.get('id')] = {'success': False, 'data': data.get('error')}
|
||
+ return 'ok'
|
||
+
|
||
+ @app.route('/link', method=['GET','POST'])
|
||
+ def link():
|
||
+ data = request.json
|
||
+ if data.get('cmd') == 'get_all_sessions': return json.dumps({'r': self.get_all_sessions()}, ensure_ascii=False)
|
||
+ if data.get('cmd') == 'find_session':
|
||
+ url_pattern = data.get('url_pattern', '')
|
||
+ return json.dumps({'r': self.find_session(url_pattern)}, ensure_ascii=False)
|
||
+ if data.get('cmd') == 'execute_js':
|
||
+ session_id = data.get('sessionId')
|
||
+ code = data.get('code')
|
||
+ timeout = float(data.get('timeout', 10.0))
|
||
+ auto_switch_newtab = data.get('auto_switch_newtab', False)
|
||
+ try:
|
||
+ result = self.execute_js(code, timeout=timeout, session_id=session_id, auto_switch_newtab=auto_switch_newtab)
|
||
+ newTabs = result.get('newTabs', []) if isinstance(result, dict) else []
|
||
+ return json.dumps({'result': result, 'newTabs': newTabs}, ensure_ascii=False)
|
||
+ except Exception as e:
|
||
+ return json.dumps({'error': str(e)}, ensure_ascii=False)
|
||
+ return 'ok'
|
||
+
|
||
+ def run():
|
||
+ import asyncio
|
||
+ loop = asyncio.new_event_loop()
|
||
+ asyncio.set_event_loop(loop)
|
||
+ bottle.run(app, host=self.host, port=self.port+1, server='tornado')
|
||
+
|
||
+ http_thread = threading.Thread(target=run)
|
||
+ http_thread.daemon = True
|
||
+ http_thread.start()
|
||
+
|
||
+ def clean_sessions(self):
|
||
+ sids = list(self.sessions.keys())
|
||
+ for sid in sids:
|
||
+ session = self.sessions[sid]
|
||
+ if not session.is_active() and time.time() - session.disconnect_at > 600:
|
||
+ del self.sessions[sid]
|
||
+
|
||
+ def start_ws_server(self) -> None:
|
||
+ driver = self
|
||
+ class JSExecutor(WebSocket):
|
||
+ def handle(self) -> None:
|
||
+ try:
|
||
+ data = json.loads(self.data)
|
||
+ if data.get('type') == 'ready':
|
||
+ session_id = data.get('sessionId')
|
||
+ session_info = {'url': data.get('url'), 'title': data.get('title', ''),
|
||
+ 'connected_at': time.time(), 'type': 'ws'}
|
||
+ driver._register_client(session_id, self, session_info)
|
||
+ elif data.get('type') in 'result':
|
||
+ driver.results[data.get('id')] = {'success': True, 'data': data.get('result'), 'newTabs': data.get('newTabs', [])}
|
||
+ elif data.get('type') == 'error':
|
||
+ driver.results[data.get('id')] = {'success': False, 'data': data.get('error')}
|
||
+ except Exception as e:
|
||
+ print(f"Error handling message: {e}")
|
||
+ if hasattr(self, 'data'): print(self.data)
|
||
+ def connected(self): (f"New connection from {self.address}")
|
||
+ def handle_close(self): driver._unregister_client(self)
|
||
+
|
||
+ self.server = WebSocketServer(self.host, self.port, JSExecutor)
|
||
+ server_thread = threading.Thread(target=self.server.serve_forever)
|
||
+ server_thread.daemon = True
|
||
+ server_thread.start()
|
||
+ print(f"WebSocket server running on ws://{self.host}:{self.port}")
|
||
+
|
||
+ def _register_client(self, session_id: str, client: WebSocket, session_info) -> None:
|
||
+ is_new_session = session_id not in self.sessions
|
||
+
|
||
+ if is_new_session:
|
||
+ session = Session(session_id, session_info, client)
|
||
+ self.sessions[session_id] = session
|
||
+ print(f"New tab connected: {session.url} (Session: {session_id})")
|
||
+ else:
|
||
+ session = self.sessions[session_id]
|
||
+ session.reconnect(client, session_info)
|
||
+ print(f"Tab reconnected: {session.url} (Session: {session_id})")
|
||
+
|
||
+ self.latest_session_id = session_id
|
||
+ if self.default_session_id is None:
|
||
+ self.default_session_id = session_id
|
||
+ elif is_new_session:
|
||
+ if time.time() - self.last_cmd_time < 5.0:
|
||
+ print(f"检测到脚本触发的新窗口，自动切换焦点: {session_id}")
|
||
+ self.default_session_id = session_id
|
||
+
|
||
+
|
||
+ def _unregister_client(self, client: WebSocket) -> None:
|
||
+ for session in self.sessions.values():
|
||
+ if session.ws_client == client:
|
||
+ session.mark_disconnected()
|
||
+ break
|
||
+
|
||
+ def execute_js(self, code, timeout=10.0, session_id=None, auto_switch_newtab=False) -> Any:
|
||
+ if session_id is None: session_id = self.default_session_id
|
||
+ if self.is_remote:
|
||
+ print('remote_execute_js')
|
||
+ response = self._remote_cmd({"cmd": "execute_js", "sessionId": session_id,
|
||
+ "code": code, "timeout": str(timeout),
|
||
+ "auto_switch_newtab": auto_switch_newtab})
|
||
+ if response.get('error'): raise Exception(response['error'])
|
||
+ if auto_switch_newtab and 'newTabs' in response:
|
||
+ newtabs = response.get('newTabs', [])
|
||
+ if len(newtabs) > 0:
|
||
+ new_session_id = newtabs[0]['sessionId']
|
||
+ self.default_session_id = new_session_id
|
||
+ print(f"自动切换到新标签会话: {new_session_id}")
|
||
+ return response.get('result', None)
|
||
+
|
||
+ session = self.sessions.get(session_id)
|
||
+ if not session or not session.is_active():
|
||
+ time.sleep(3)
|
||
+ session = self.sessions.get(session_id)
|
||
+ if not session or not session.is_active():
|
||
+ alive_sessions = [s for s in self.sessions.values() if s.is_active()]
|
||
+ if alive_sessions:
|
||
+ session = alive_sessions[0]
|
||
+ print(f"会话 {session_id} 未连接，自动切换到最新活动会话: {session.id}")
|
||
+ session_id = self.default_session_id = session.id
|
||
+ if not session or not session.is_active():
|
||
+ breakpoint()
|
||
+ raise ValueError(f"会话ID {session_id} 未连接")
|
||
+
|
||
+ tp = session.type
|
||
+ assert tp in ['ws', 'http'], f"Unsupported session type: {tp}"
|
||
+ exec_id = str(uuid.uuid4())
|
||
+ payload = json.dumps({'id': exec_id, 'code': code, 'auto_switch_newtab': auto_switch_newtab})
|
||
+
|
||
+ if tp == 'ws':
|
||
+ session.ws_client.send_message(payload)
|
||
+ elif tp == 'http':
|
||
+ session.http_queue.put(payload)
|
||
+
|
||
+ start_time = time.time()
|
||
+ self.clean_sessions()
|
||
+ hasjump = False
|
||
+
|
||
+ while exec_id not in self.results:
|
||
+ time.sleep(0.1)
|
||
+ if tp == 'ws':
|
||
+ if not session.is_active(): hasjump = True
|
||
+ if hasjump and session.is_active():
|
||
+ if not self.is_remote and auto_switch_newtab: self.last_cmd_time = time.time()
|
||
+ return {"result": f"Session {session_id} reloaded.", "closed":1}
|
||
+ if time.time() - start_time > timeout:
|
||
+ if tp == 'ws':
|
||
+ return {"result": f"No response data in {timeout}s"}
|
||
+ elif tp == 'http':
|
||
+ return {"result": f"Session {session_id} no response."}
|
||
+
|
||
+ result = self.results.pop(exec_id)
|
||
+ if not result['success']: raise Exception(result['data'])
|
||
+ if not self.is_remote and auto_switch_newtab:
|
||
+ newtabs = result.get('newTabs', [])
|
||
+ if len(newtabs) > 0:
|
||
+ new_session_id = newtabs[0]['sessionId']
|
||
+ self.default_session_id = new_session_id
|
||
+ print(f"自动切换到新标签会话: {new_session_id}")
|
||
+ elif not self.is_remote:
|
||
+ self.last_cmd_time = time.time()
|
||
+ return result['data']
|
||
+
|
||
+ def _remote_cmd(self, cmd):
|
||
+ resp = requests.post(self.remote,
|
||
+ headers={"Content-Type": "application/json"},
|
||
+ json=cmd).json()
|
||
+ return resp
|
||
+
|
||
+ def get_all_sessions(self):
|
||
+ if self.is_remote:
|
||
+ return self._remote_cmd({"cmd": "get_all_sessions"}).get('r', [])
|
||
+ return [{'id': session.id, **session.info} for session in self.sessions.values()
|
||
+ if session.is_active()]
|
||
+
|
||
+ def get_session_dict(self):
|
||
+ return {session.id: session.url for session in self.sessions.values() if session.is_active()}
|
||
+
|
||
+ def find_session(self, url_pattern: str):
|
||
+ if url_pattern == '':
|
||
+ session = self.sessions.get(self.latest_session_id)
|
||
+ return [(session.id, session.info)] if session else []
|
||
+ matching_sessions = []
|
||
+ for session in self.sessions.values():
|
||
+ if not session.is_active(): continue
|
||
+ if 'url' in session.info and url_pattern in session.info['url']:
|
||
+ matching_sessions.append((session.id, session.info))
|
||
+ return matching_sessions
|
||
+
|
||
+ def set_session(self, url_pattern: str) -> bool:
|
||
+ if self.is_remote:
|
||
+ matched = self._remote_cmd({"cmd": "find_session", "url_pattern": url_pattern}).get('r', [])
|
||
+ else:
|
||
+ matched = self.find_session(url_pattern)
|
||
+ if not matched: return print(f"警告: 未找到URL包含 '{url_pattern}' 的会话")
|
||
+ if len(matched) > 1: print(f"警告: 找到多个URL包含 '{url_pattern}' 的会话，选择第一个")
|
||
+ self.last_cmd_time = 0
|
||
+ self.default_session_id, info = matched[0]
|
||
+ print(f"成功设置默认会话: {self.default_session_id}: {info['url']}")
|
||
+ return self.default_session_id
|
||
+
|
||
+ def jump(self, url, timeout=10): self.execute_js(f"window.location.href='{url}'", timeout=timeout)
|
||
+ def page_source(self): return self.execute_js("document.documentElement.outerHTML")
|
||
+ def body(self): return self.execute_js("document.body.outerHTML")
|
||
+ def newtab(self, url=None):
|
||
+ if url is None: url = "http://www.baidu.com/robots.txt"
|
||
+ return self.execute_js(f'GM_openInTab("{url}");', auto_switch_newtab=True)
|
||
+
|
||
+if __name__ == "__main__":
|
||
+ driver = TMWebDriver(host='localhost', port=18765)
|
||
\ No newline at end of file
|
||
diff --git a/agent_loop.py b/agent_loop.py
|
||
new file mode 100644
|
||
index 0000000..e30eecb
|
||
--- /dev/null
|
||
+++ b/agent_loop.py
|
||
@@ -0,0 +1,67 @@
|
||
+import json
|
||
+from dataclasses import dataclass
|
||
+from typing import Any, Optional
|
||
+@dataclass
|
||
+class StepOutcome:
|
||
+ data: Any
|
||
+ next_prompt: Optional[str] = None
|
||
+ should_exit: bool = False
|
||
+
|
||
+
|
||
+def try_call_generator(func, *args, **kwargs):
|
||
+ ret = func(*args, **kwargs)
|
||
+ if hasattr(ret, '__iter__') and not isinstance(ret, (str, bytes, dict, list)):
|
||
+ ret = yield from ret
|
||
+ return ret
|
||
+
|
||
+class BaseHandler:
|
||
+ def tool_before_callback(self, tool_name, args, content): pass
|
||
+ def tool_after_callback(self, tool_name, args, content): pass
|
||
+ def dispatch(self, tool_name, args, response):
|
||
+ method_name = f"do_{tool_name}"
|
||
+ if hasattr(self, method_name):
|
||
+ _ = yield from try_call_generator(self.tool_before_callback, tool_name, args, response)
|
||
+ ret = yield from try_call_generator(getattr(self, method_name), args, response)
|
||
+ _ = yield from try_call_generator(self.tool_after_callback, tool_name, args, response)
|
||
+ return ret
|
||
+ else:
|
||
+ yield f"❌ 未知工具: {tool_name}\n"
|
||
+ return StepOutcome(None, "未知工具", "ERROR")
|
||
+
|
||
+def json_default(o):
|
||
+ if isinstance(o, set): return list(o)
|
||
+ return str(o)
|
||
+
|
||
+def agent_runner_loop(client, system_prompt, user_input, handler, tools_schema, max_turns=15):
|
||
+ messages = [
|
||
+ {"role": "system", "content": system_prompt},
|
||
+ {"role": "user", "content": user_input}
|
||
+ ]
|
||
+ for turn in range(max_turns):
|
||
+ yield f"\n[🤖 LLM Thinking (Turn {turn+1})] ..."
|
||
+ response = client.chat(messages=messages, tools=tools_schema)
|
||
+
|
||
+ if response.thinking: yield '<thinking>' + response.thinking + '</thinking>\n'
|
||
+ yield response.content
|
||
+
|
||
+ if not response.tool_calls:
|
||
+ tool_name, args = 'no_tool', {}
|
||
+ else:
|
||
+ tool_call = response.tool_calls[0]
|
||
+ tool_name = tool_call.function.name
|
||
+ args = json.loads(tool_call.function.arguments)
|
||
+
|
||
+ if tool_name == 'no_tool': pass
|
||
+ else: yield f"\n\n正在调用工具: {tool_name}，参数: {args}\n"
|
||
+ outcome = yield from handler.dispatch(tool_name, args, response)
|
||
+
|
||
+ if outcome.next_prompt is None: return {'result': 'CURRENT_TASK_DONE', 'data': outcome.data}
|
||
+ if outcome.should_exit: return {'result': 'EXITED', 'data': outcome.data}
|
||
+
|
||
+ next_prompt = ""
|
||
+ if outcome.data is not None:
|
||
+ datastr = json.dumps(outcome.data, ensure_ascii=False, default=json_default) if type(outcome.data) in [dict, list] else str(outcome.data)
|
||
+ next_prompt += f"<tool_result>\n{datastr}\n</tool_result>\n\n"
|
||
+ next_prompt += outcome.next_prompt
|
||
+ messages = [{"role": "user", "content": next_prompt}]
|
||
+ return {'result': 'MAX_TURNS_EXCEEDED'}
|
||
\ No newline at end of file
|
||
diff --git a/agentapp.py b/agentapp.py
|
||
new file mode 100644
|
||
index 0000000..915921d
|
||
--- /dev/null
|
||
+++ b/agentapp.py
|
||
@@ -0,0 +1,94 @@
|
||
+import os, sys
|
||
+if sys.stdout is None: sys.stdout = open(os.devnull, "w")
|
||
+if sys.stderr is None: sys.stderr = open(os.devnull, "w")
|
||
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
|
||
+
|
||
+
|
||
+import streamlit as st
|
||
+import time, json, re
|
||
+
|
||
+with open('tools_schema.json', 'r', encoding='utf-8') as f:
|
||
+ TOOLS_SCHEMA = json.load(f)
|
||
+
|
||
+
|
||
+st.set_page_config(page_title="Cowork", layout="wide")
|
||
+
|
||
+from sidercall import SiderLLMSession, LLMSession, ToolClient
|
||
+from agent_loop import agent_runner_loop, StepOutcome, BaseHandler
|
||
+
|
||
+@st.cache_resource
|
||
+def init():
|
||
+ mainllm = SiderLLMSession(multiturns=6)
|
||
+ llmclient = ToolClient(mainllm.ask, auto_save_tokens=True)
|
||
+ return llmclient
|
||
+
|
||
+llmclient = init()
|
||
+
|
||
+from ga import GenericAgentHandler
|
||
+
|
||
+def get_system_prompt():
|
||
+ with open('sys_prompt.txt', 'r', encoding='utf-8') as f:
|
||
+ return f.read()
|
||
+
|
||
+if "last_goal" not in st.session_state:
|
||
+ st.session_state.last_goal = ""
|
||
+
|
||
+def refine_user_goal(raw_query, last_goal):
|
||
+ """通过 LLM 提炼用户真实意图"""
|
||
+ if not last_goal:
|
||
+ return raw_query
|
||
+
|
||
+ decide_prompt = f"""
|
||
+用户之前的目标是: "{last_goal}"
|
||
+用户现在输入了: "{raw_query}"
|
||
+
|
||
+请判断：
|
||
+1. 如果用户提供补充信息、或者是接续之前的任务，请输出合并后的【最终目标】。
|
||
+2. 如果用户只是指出之前做法有错而非变更目标，那么请输出原目标不做修改。
|
||
+3. 如果用户开启了一个完全不相关的新话题，请直接输出用户现在的输入内容。
|
||
+
|
||
+请直接输出目标描述，不要包含任何多余的文字、解释或标点。
|
||
+"""
|
||
+ try:
|
||
+ refined = llmclient.llm_func(decide_prompt).strip()
|
||
+ return refined if refined else raw_query
|
||
+ except:
|
||
+ return raw_query
|
||
+
|
||
+def agent_backend_stream(raw_query):
|
||
+ final_goal = refine_user_goal(raw_query, st.session_state.last_goal)
|
||
+
|
||
+ if final_goal != raw_query:
|
||
+ yield f"[Goal Refined] {final_goal}\n"
|
||
+
|
||
+ sys_prompt = get_system_prompt()
|
||
+ handler = GenericAgentHandler(None, final_goal, './temp')
|
||
+ llmclient.last_tools = ''
|
||
+ ret = yield from agent_runner_loop(llmclient,
|
||
+ sys_prompt, raw_query, handler,
|
||
+ TOOLS_SCHEMA, max_turns=25)
|
||
+ st.session_state.last_goal = final_goal
|
||
+ return ret
|
||
+
|
||
+st.title("🖥️ Cowork")
|
||
+
|
||
+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"])
|
||
+
|
||
+if prompt := st.chat_input("请输入指令"):
|
||
+ st.session_state.messages.append({"role": "user", "content": prompt})
|
||
+ with st.chat_message("user"):
|
||
+ st.markdown(prompt)
|
||
+
|
||
+ with st.chat_message("assistant"):
|
||
+ message_placeholder = st.empty()
|
||
+ full_response = ""
|
||
+ for chunk in agent_backend_stream(prompt):
|
||
+ full_response += chunk
|
||
+ message_placeholder.markdown(full_response + "▌")
|
||
+ message_placeholder.markdown(full_response)
|
||
+ st.session_state.messages.append({"role": "assistant", "content": full_response})
|
||
\ No newline at end of file
|
||
diff --git a/ga.py b/ga.py
|
||
new file mode 100644
|
||
index 0000000..446a7a3
|
||
--- /dev/null
|
||
+++ b/ga.py
|
||
@@ -0,0 +1,379 @@
|
||
+import sys, os, re
|
||
+import pyperclip
|
||
+import json, time
|
||
+from pathlib import Path
|
||
+import subprocess
|
||
+import tempfile
|
||
+if sys.stdout is None: sys.stdout = open(os.devnull, "w")
|
||
+if sys.stderr is None: sys.stderr = open(os.devnull, "w")
|
||
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
|
||
+
|
||
+from sidercall import LLMSession, ToolClient
|
||
+from agent_loop import BaseHandler, StepOutcome, agent_runner_loop
|
||
+
|
||
+def code_run(code: str, code_type: str = "python", timeout: int = 60, cwd: str = None):
|
||
+ """
|
||
+ 针对 Windows 优化的双模态执行器
|
||
+ python: 运行复杂的 .py 脚本（文件模式）
|
||
+ powershell: 运行单行指令（命令模式）
|
||
+ 优先使用python，仅在必要系统操作时使用powershell。
|
||
+ """
|
||
+ # 统一路径处理
|
||
+ preview = (code[:60].replace('\n', ' ') + '...') if len(code) > 60 else code.strip()
|
||
+ yield f"\n[Action] Running {code_type} in {os.path.basename(cwd)}: {preview}\n"
|
||
+ cwd = cwd or os.getcwd()
|
||
+ if code_type == "python":
|
||
+ # Python 依然建议走文件，因为模型生成的逻辑通常包含多行、import 和类定义
|
||
+ tmp_file = tempfile.NamedTemporaryFile(suffix=".py", delete=False, mode='w', encoding='utf-8')
|
||
+ tmp_file.write(code)
|
||
+ tmp_path = tmp_file.name
|
||
+ tmp_file.close()
|
||
+ cmd = ["python", "-u", tmp_path]
|
||
+ elif code_type == "powershell":
|
||
+ cmd = ["powershell", "-NoProfile", "-NonInteractive", "-Command", code]
|
||
+ tmp_path = None
|
||
+ else:
|
||
+ return {"status": "error", "msg": f"不支持的类型: {code_type}"}
|
||
+ print("code run output:")
|
||
+ startupinfo = None
|
||
+ if os.name == 'nt':
|
||
+ startupinfo = subprocess.STARTUPINFO()
|
||
+ startupinfo.dwFlags |= subprocess.STARTF_USESHOWWINDOW
|
||
+ startupinfo.wShowWindow = 0 # SW_HIDE
|
||
+ full_stdout = []
|
||
+ full_stderr = []
|
||
+ try:
|
||
+ process = subprocess.Popen(
|
||
+ cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
|
||
+ bufsize=0, cwd=cwd, startupinfo=startupinfo
|
||
+ )
|
||
+ for line_bytes in iter(process.stdout.readline, b''):
|
||
+ try:
|
||
+ line = line_bytes.decode('utf-8')
|
||
+ except UnicodeDecodeError:
|
||
+ line = line_bytes.decode('gbk', errors='ignore')
|
||
+ print(line, end="")
|
||
+ full_stdout.append(line)
|
||
+
|
||
+ stdout_rem, stderr_raw = process.communicate(timeout=timeout)
|
||
+ if stdout_rem:
|
||
+ try: rem_str = stdout_rem.decode('utf-8')
|
||
+ except UnicodeDecodeError:
|
||
+ rem_str = stdout_rem.decode('gbk', errors='ignore')
|
||
+ full_stdout.append(rem_str)
|
||
+
|
||
+ if stderr_raw:
|
||
+ try: stderr_str = stderr_raw.decode('utf-8')
|
||
+ except UnicodeDecodeError:
|
||
+ stderr_str = stderr_raw.decode('gbk', errors='ignore')
|
||
+ full_stderr.append(stderr_str)
|
||
+ print(f"Error: {stderr_str}")
|
||
+
|
||
+ status = "success" if process.returncode == 0 else "error"
|
||
+ stdout_str = "".join(full_stdout)
|
||
+ stderr_str = "".join(full_stderr)
|
||
+ status_icon = "✅" if process.returncode == 0 else "❌"
|
||
+ output_snippet = (stdout_str[:200] + '...') if len(stdout_str) > 200 else stdout_str
|
||
+ yield f"[Status] {status_icon} Exit Code: {process.returncode}\n[Stdout] {output_snippet}\n"
|
||
+ return {
|
||
+ "status": status,
|
||
+ "stdout": stdout_str[-2000:],
|
||
+ "stderr": stderr_str[-2000:],
|
||
+ "exit_code": process.returncode
|
||
+ }
|
||
+ except subprocess.TimeoutExpired:
|
||
+ return {"status": "error", "msg": "Timeout"}
|
||
+ except Exception as e:
|
||
+ return {"status": "error", "msg": str(e)}
|
||
+ finally:
|
||
+ if code_type == "python" and tmp_path and os.path.exists(tmp_path): os.remove(tmp_path)
|
||
+
|
||
+
|
||
+def ask_user(question: str, candidates: list = None):
|
||
+ """
|
||
+ 构造一个中断请求。
|
||
+ question: 向用户提出的问题。
|
||
+ candidates: 可选的候选项列表。
|
||
+ 需要保证should_exit为True
|
||
+ """
|
||
+ return {
|
||
+ "status": "INTERRUPT",
|
||
+ "intent": "HUMAN_INTERVENTION",
|
||
+ "data": {
|
||
+ "question": question,
|
||
+ "candidates": candidates or []
|
||
+ }
|
||
+ }
|
||
+
|
||
+from web_tools import execute_js_rich, get_html
|
||
+
|
||
+driver = None
|
||
+
|
||
+def first_init_driver():
|
||
+ global driver
|
||
+ from TMWebDriver import TMWebDriver
|
||
+ driver = TMWebDriver()
|
||
+ while True:
|
||
+ time.sleep(1)
|
||
+ sess = driver.get_all_sessions()
|
||
+ if len(sess) > 0: break
|
||
+ driver.newtab()
|
||
+ time.sleep(5)
|
||
+
|
||
+def web_scan(focus_item="", switch_tab_id=None):
|
||
+ """
|
||
+ 利用 get_html 获取清洗后的网页内容。
|
||
+ focus_item: 语义过滤指令。如果用户在找特定内容（如“小米汽车”），
|
||
+ 算法会优先保留包含该关键词的列表项。
|
||
+ switch_tab_id: 可选参数，如果提供，则在扫描前切换到该标签页。
|
||
+ 应当多用execute_js，少全量观察html。
|
||
+ """
|
||
+ global driver
|
||
+ if driver is None: first_init_driver()
|
||
+ try:
|
||
+ tabs = []
|
||
+ for sess in driver.get_all_sessions():
|
||
+ sess.pop('connected_at', None)
|
||
+ sess.pop('type', None)
|
||
+ sess['url'] = sess.get('url', '')[:50] + ("..." if len(sess.get('url', '')) > 50 else "")
|
||
+ tabs.append(sess)
|
||
+ if switch_tab_id: driver.default_session_id = switch_tab_id
|
||
+ content = get_html(driver, cutlist=True, instruction=focus_item, maxchars=23000)
|
||
+ return {
|
||
+ "status": "success",
|
||
+ "metadata": {
|
||
+ "tabs_count": len(tabs),
|
||
+ "tabs": tabs,
|
||
+ "active_tab": driver.default_session_id
|
||
+ },
|
||
+ "content": content
|
||
+ }
|
||
+ except Exception as e:
|
||
+ return {"status": "error", "msg": format_error(e)}
|
||
+
|
||
+import traceback
|
||
+def format_error(e):
|
||
+ exc_type, exc_value, exc_traceback = sys.exc_info()
|
||
+ tb = traceback.extract_tb(exc_traceback)
|
||
+ if tb:
|
||
+ f = tb[-1]
|
||
+ fname = os.path.basename(f.filename)
|
||
+ return f"{exc_type.__name__}: {str(e)} @ {fname}:{f.lineno}, {f.name} -> `{f.line}`"
|
||
+ return f"{exc_type.__name__}: {str(e)}"
|
||
+
|
||
+def web_execute_js(script: str):
|
||
+ """
|
||
+ 执行 JS 脚本来控制浏览器，并捕获结果和页面变化。
|
||
+ script: 要执行的 JavaScript 代码字符串。
|
||
+ return {
|
||
+ "status": "failed" if error_msg else "success",
|
||
+ "js_return": result,
|
||
+ "error": error_msg,
|
||
+ "transients": transients,
|
||
+ "environment": {
|
||
+ "new_tab": new_tab,
|
||
+ "reloaded": reloaded
|
||
+ },
|
||
+ "diff": diff_summary,
|
||
+ "suggestion": "" if is_significant_change else "页面无明显变化"
|
||
+ }
|
||
+ """
|
||
+ global driver
|
||
+ if driver is None: first_init_driver()
|
||
+ try:
|
||
+ result = execute_js_rich(script, driver)
|
||
+ return result
|
||
+ except Exception as e:
|
||
+ return {"status": "error", "msg": format_error(e)}
|
||
+
|
||
+def file_patch(path: str, old_content: str, new_content: str):
|
||
+ """
|
||
+ 在文件中寻找唯一的 old_content 块并替换为 new_content。
|
||
+ """
|
||
+ path = str(Path(path).resolve())
|
||
+ try:
|
||
+ if not os.path.exists(path):
|
||
+ return {"status": "error", "msg": "文件不存在"}
|
||
+ with open(path, 'r', encoding='utf-8') as f:
|
||
+ full_text = f.read()
|
||
+ # 检查唯一性
|
||
+ count = full_text.count(old_content)
|
||
+ if count == 0:
|
||
+ return {"status": "error", "msg": "未找到匹配的旧文本块，请检查空格、缩进和换行是否完全一致。"}
|
||
+ if count > 1:
|
||
+ return {"status": "error", "msg": f"找到 {count} 处匹配，请提供更长的旧文本块以确保唯一性。"}
|
||
+ updated_text = full_text.replace(old_content, new_content)
|
||
+ with open(path, 'w', encoding='utf-8') as f:
|
||
+ f.write(updated_text)
|
||
+ return {"status": "success", "msg": "文件局部修改成功"}
|
||
+ except Exception as e:
|
||
+ return {"status": "error", "msg": str(e)}
|
||
+
|
||
+def file_read(path, start=1, count=100, show_linenos=True):
|
||
+ try:
|
||
+ with open(path, 'r', encoding='utf-8', errors='replace') as f:
|
||
+ lines = f.readlines()
|
||
+ chunk = lines[start-1 : start-1+count]
|
||
+ if show_linenos: res = [f"{i+start}|{l[:200]}" for i, l in enumerate(chunk)]
|
||
+ else: res = [l for l in chunk]
|
||
+ return f"Total:{len(lines)} lines\n" + "".join(res)
|
||
+ except Exception as e:
|
||
+ return f"Error: {str(e)}"
|
||
+
|
||
+class GenericAgentHandler(BaseHandler):
|
||
+ '''
|
||
+ Generic Agent 工具库，包含多种工具的实现。工具函数自动加上了 do_ 前缀。实际工具名没有前缀。
|
||
+ '''
|
||
+ def __init__(self, parent, user_input, cwd):
|
||
+ self.parent = parent
|
||
+ self.user_input = user_input
|
||
+ self.plan = ""
|
||
+ self.focus = ""
|
||
+ self.cwd = cwd
|
||
+
|
||
+ def _get_abs_path(self, path):
|
||
+ if not path: return ""
|
||
+ return os.path.abspath(os.path.join(self.cwd, path))
|
||
+
|
||
+ def do_code_run(self, args, response):
|
||
+ '''执行代码片段，有长度限制，不允许代码中放大量数据，如有需要应当通过文件读取进行。
|
||
+ '''
|
||
+ code_type = args.get("type", "python")
|
||
+ # 从 response.content 中提取代码块
|
||
+ # 匹配 ```python ... ``` 或 ```powershell ... ```
|
||
+ pattern = rf"```{code_type}\n(.*?)\n```"
|
||
+ # 也可以更通用一点，不分类型提取最后一个代码块：rf"```(?:{code_type})?\n(.*?)\n```"
|
||
+ matches = re.findall(pattern, response.content, re.DOTALL)
|
||
+ if not matches:
|
||
+ return StepOutcome(None, next_prompt=f"【系统错误】：你调用了 code_run，但未在回复中提供 ```{code_type} 代码块。请重新输出代码并附带工具调用。")
|
||
+ # 提取最后一个代码块（通常是模型修正后的最终逻辑）
|
||
+ code = matches[-1].strip()
|
||
+ timeout = args.get("timeout", 60)
|
||
+ cwd = args.get("cwd", self.cwd)
|
||
+ result = yield from code_run(code, code_type, timeout, cwd)
|
||
+ return StepOutcome(result, next_prompt=self._get_anchor_prompt())
|
||
+
|
||
+ def do_ask_user(self, args, response):
|
||
+ question = args.get("question", "请提供输入：")
|
||
+ candidates = args.get("candidates", [])
|
||
+ result = ask_user(question, candidates)
|
||
+ return StepOutcome(result, next_prompt="", should_exit=True)
|
||
+
|
||
+ def do_web_scan(self, args, response):
|
||
+ '''focus_item仅用于在长列表中模糊搜寻相关item
|
||
+ 此工具也提供标签页查看和标签页切换功能。
|
||
+ '''
|
||
+ focus_item = args.get("focus_item", "")
|
||
+ switch_tab_id = args.get("switch_tab_id", None)
|
||
+ result = web_scan(focus_item, switch_tab_id=switch_tab_id)
|
||
+ content = result.pop("content", None)
|
||
+ yield f'\n{str(result)}\n'
|
||
+ next_prompt = f"```html\n{content}\n```"
|
||
+ return StepOutcome(result, next_prompt=next_prompt)
|
||
+
|
||
+ def do_web_execute_js(self, args, response):
|
||
+ '''web情况下的优先使用工具，执行任何js达成对浏览器的*完全*控制。
|
||
+ 支持将结果保存到文件供后续读取分析，但保存功能仅限即时读取，与await等异步操作不兼容。
|
||
+ '''
|
||
+ script = args.get("script", "")
|
||
+ save_to_file = args.get("save_to_file", "")
|
||
+ result = web_execute_js(script)
|
||
+ if save_to_file and "js_return" in result:
|
||
+ content = str(result["js_return"] or '')
|
||
+ abs_path = self._get_abs_path(save_to_file)
|
||
+ with open(abs_path, 'w', encoding='utf-8') as f: f.write(str(content))
|
||
+ result["js_return"] = content[:200] + ("..." if len(content) > 200 else "")
|
||
+ result["js_return"] += f"\n\n[已保存以上内容到 {abs_path}]"
|
||
+ print("Web Execute JS Result:", result)
|
||
+ return StepOutcome(result, next_prompt=self._get_anchor_prompt())
|
||
+
|
||
+ def do_file_patch(self, args, response):
|
||
+ path = self._get_abs_path(args.get("path", ""))
|
||
+ yield f"\n[Action] Patching file: {path}\n"
|
||
+ old_content = args.get("old_content", "")
|
||
+ new_content = args.get("new_content", "")
|
||
+ result = file_patch(path, old_content, new_content)
|
||
+ yield str(result) + "\n"
|
||
+ return StepOutcome(result, next_prompt=self._get_anchor_prompt())
|
||
+
|
||
+ def do_file_write(self, args, response):
|
||
+ '''用于对整个文件的大量处理，精细修改要用file_patch。
|
||
+ '''
|
||
+ path = self._get_abs_path(args.get("path", ""))
|
||
+ mode = args.get("mode", "overwrite")
|
||
+ action_str = "Appending to" if mode == "append" else "Writing"
|
||
+ yield f"\n[Action] {action_str} file: {os.path.basename(path)}\n"
|
||
+
|
||
+ def extract_intended_block(content):
|
||
+ start_marker = "```"
|
||
+ first_idx = content.find(start_marker)
|
||
+ last_idx = content.rfind(start_marker)
|
||
+ if first_idx == -1 or last_idx == -1 or first_idx == last_idx:
|
||
+ return None
|
||
+ header_end = content.find("\n", first_idx)
|
||
+ if header_end == -1 or header_end > last_idx:
|
||
+ return None
|
||
+ actual_content = content[header_end + 1 : last_idx].strip()
|
||
+ return actual_content
|
||
+
|
||
+ blocks = extract_intended_block(response.content)
|
||
+ if not blocks:
|
||
+ yield f"[Status] ❌ 失败: 未在回复中找到代码块内容\n"
|
||
+ return StepOutcome({"status": "error", "msg": "No code block found in response"}, next_prompt="\n")
|
||
+ new_content = blocks
|
||
+ try:
|
||
+ write_mode = 'a' if mode == "append" else 'w'
|
||
+ final_content = ("\n" + new_content) if mode == "append" else new_content
|
||
+ with open(path, write_mode, encoding="utf-8") as f:
|
||
+ f.write(final_content)
|
||
+ yield f"[Status] ✅ {mode.capitalize()} 成功 ({len(new_content)} bytes)\n"
|
||
+ return StepOutcome({"status": "success"},
|
||
+ next_prompt=f"\n提醒: <user_input>{self.user_input}</user_input>请继续执行下一步。\n")
|
||
+ except Exception as e:
|
||
+ yield f"[Status] ❌ 写入异常: {str(e)}\n"
|
||
+ return StepOutcome({"status": "error", "msg": str(e)}, next_prompt="\n")
|
||
+
|
||
+ def do_file_read(self, args, response):
|
||
+ path = self._get_abs_path(args.get("path", ""))
|
||
+ yield f"\n[Action] Reading file: {path}\n"
|
||
+ start = args.get("start", 1)
|
||
+ count = args.get("count", 100)
|
||
+ show_linenos = args.get("show_linenos", True)
|
||
+ result = file_read(path, start, count, show_linenos)
|
||
+ return StepOutcome(result, next_prompt=self._get_anchor_prompt())
|
||
+
|
||
+ def do_update_plan(self, args, response):
|
||
+ '''
|
||
+ 同步宏观任务进度与战略重心。
|
||
+ 【设计意图】：
|
||
+ 1. 仅在任务涉及多步逻辑（如：先搜索、再重构、后测试）时进行初始拆解。
|
||
+ 2. 仅在发生重大的方针变更时调用（例如：原定方案 A 物理不可行，需彻底转向方案 B）。
|
||
+ 3. 严禁用于记录细微的调试步骤或代码纠错。
|
||
+ 简单任务无需使用。
|
||
+ '''
|
||
+ new_plan = args.get("plan", "")
|
||
+ new_focus = args.get("focus", "")
|
||
+ if new_plan: self.plan = new_plan
|
||
+ if new_focus: self.focus = new_focus
|
||
+ yield f"\n[Info] Updated plan and focus.\n"
|
||
+ yield f"New Plan:\n{self.plan}\n\n"
|
||
+ yield f"New Focus:\n{self.focus}\n"
|
||
+ return StepOutcome({"status": "success"},
|
||
+ next_prompt=self._get_anchor_prompt())
|
||
+
|
||
+ def do_no_tool(self, args, response):
|
||
+ '''这是一个特殊工具，由引擎自主调用，不要包含在TOOLS_SCHEMA里。
|
||
+ '''
|
||
+ yield "\n\n[Info] No tool called. Final response to user.\n"
|
||
+ return StepOutcome(response, next_prompt=None, should_exit=True)
|
||
+
|
||
+ def _get_anchor_prompt(self):
|
||
+ prompt = f"\n提醒: \n<user_input>{self.user_input}</user_input>\n"
|
||
+ if self.plan: prompt += f"<plan>\n{self.plan}\n</plan>\n"
|
||
+ if self.focus: prompt += f"<current>\n{self.focus}\n</current>\n"
|
||
+ prompt += "\n请继续执行下一步。"
|
||
+ return prompt
|
||
+
|
||
+
|
||
+if __name__ == "__main__":
|
||
+ pass
|
||
\ No newline at end of file
|
||
diff --git a/launch.pyw b/launch.pyw
|
||
new file mode 100644
|
||
index 0000000..4b91d6f
|
||
--- /dev/null
|
||
+++ b/launch.pyw
|
||
@@ -0,0 +1,48 @@
|
||
+import webview
|
||
+import threading
|
||
+import subprocess
|
||
+import sys, time, os, ctypes
|
||
+import atexit
|
||
+
|
||
+# === 配置区域 ===
|
||
+WINDOW_WIDTH = 600
|
||
+WINDOW_HEIGHT = 900
|
||
+RIGHT_PADDING = 0 # 离屏幕右边缘的距离
|
||
+TOP_PADDING = 300 # 离屏幕上边缘的距离
|
||
+
|
||
+def get_screen_width():
|
||
+ try:
|
||
+ # GetSystemMetrics(0) 获取主屏幕宽度
|
||
+ user32 = ctypes.windll.user32
|
||
+ return user32.GetSystemMetrics(0)
|
||
+ except:
|
||
+ # 如果不是 Windows 或者出错了，返回一个兜底值 (比如 1920)
|
||
+ return 1920
|
||
+
|
||
+def start_streamlit():
|
||
+ global proc
|
||
+ cmd = [
|
||
+ sys.executable, "-m", "streamlit", "run", "agentapp.py",
|
||
+ "--server.port", "8501",
|
||
+ "--server.headless", "true",
|
||
+ "--theme.base", "dark" #以此默认开启暗黑模式，更有极客感
|
||
+ ]
|
||
+ proc = subprocess.Popen(cmd)
|
||
+ atexit.register(proc.kill)
|
||
+
|
||
+if __name__ == '__main__':
|
||
+ t = threading.Thread(target=start_streamlit, daemon=True)
|
||
+ t.start()
|
||
+ screen_width = get_screen_width()
|
||
+ x_pos = screen_width - WINDOW_WIDTH - RIGHT_PADDING
|
||
+ time.sleep(2)
|
||
+ webview.create_window(
|
||
+ title='GenericAgent',
|
||
+ url='http://localhost:8501',
|
||
+ width=WINDOW_WIDTH,
|
||
+ height=WINDOW_HEIGHT,
|
||
+ x=x_pos, y=TOP_PADDING,
|
||
+ resizable=True,
|
||
+ text_select=True
|
||
+ )
|
||
+ webview.start()
|
||
\ No newline at end of file
|
||
diff --git a/ljq_web_driver.user.js b/ljq_web_driver.user.js
|
||
new file mode 100644
|
||
index 0000000..01eb42b
|
||
--- /dev/null
|
||
+++ b/ljq_web_driver.user.js
|
||
@@ -0,0 +1,428 @@
|
||
+// ==UserScript==
|
||
+// @name ljq_web_driver
|
||
+// @namespace http://tampermonkey.net/
|
||
+// @version 0.2
|
||
+// @description Execute JS via ljq_web_driver
|
||
+// @require https://code.jquery.com/jquery-3.6.0.min.js
|
||
+// @author You
|
||
+// @match *://*/*
|
||
+// @grant GM_setValue
|
||
+// @grant GM_getValue
|
||
+// @grant GM_xmlhttpRequest
|
||
+// @grant GM_openInTab
|
||
+// @grant unsafeWindow
|
||
+// @connect localhost
|
||
+// @run-at document-start
|
||
+// ==/UserScript==
|
||
+
|
||
+
|
||
+(function() {
|
||
+ 'use strict';
|
||
+ const log_prefix = "ljq_driver: ";
|
||
+
|
||
+ if (window.self !== window.top) {
|
||
+ console.log(log_prefix + '在iframe中不执行');
|
||
+ return;
|
||
+ }
|
||
+
|
||
+ const wsUrl = 'ws://localhost:18765';
|
||
+ const httpUrl = 'http://localhost:18766/';
|
||
+
|
||
+ function isWebSocketServerAlive(callback) {
|
||
+ GM_xmlhttpRequest({
|
||
+ method: 'GET',
|
||
+ url: 'http://localhost:18765/',
|
||
+ onload: () => callback(true),
|
||
+ onerror: () => callback(false)
|
||
+ });
|
||
+ }
|
||
+
|
||
+ let ws;
|
||
+ let sid = (window.name && window.name.startsWith('ljq_')) ?
|
||
+ window.name : window.sessionStorage.getItem('ljq_driver_sid');
|
||
+ if (!sid) {
|
||
+ sid = `ljq_${Date.now().toString().slice(-2)}${Math.random().toString(36).slice(2, 4)}`;
|
||
+ window.sessionStorage.setItem('ljq_driver_sid', sid);
|
||
+ window.name = sid;
|
||
+ console.log(log_prefix + `创建新会话ID: ${sid}`);
|
||
+ } else {
|
||
+ if (window.name !== sid) window.name = sid;
|
||
+ console.log(log_prefix + `使用现有会话ID: ${sid}`);
|
||
+ }
|
||
+
|
||
+ try {
|
||
+ GM_setValue('new_tab_report', {
|
||
+ url: window.location.href,
|
||
+ sessionId: sid,
|
||
+ ts: Date.now()
|
||
+ });
|
||
+ } catch (e) {}
|
||
+
|
||
+ // 保存会话ID
|
||
+ GM_setValue('sid', sid);
|
||
+
|
||
+ // 获取或创建状态指示器
|
||
+ function getIndicator() {
|
||
+ // 检查现有指示器
|
||
+ let ind = document.getElementById('ljq-ind');
|
||
+
|
||
+ // 删除重复指示器
|
||
+ const dups = document.querySelectorAll('[id="ljq-ind"]');
|
||
+ if (dups.length > 1) {
|
||
+ for (let i = 1; i < dups.length; i++) {
|
||
+ dups[i].remove();
|
||
+ }
|
||
+ ind = dups[0];
|
||
+ }
|
||
+
|
||
+ // 创建新指示器
|
||
+ if (!ind && document.body) {
|
||
+ ind = document.createElement('div');
|
||
+ ind.id = 'ljq-ind';
|
||
+ ind.style.cssText = `
|
||
+ position: fixed;bottom: 10px;
|
||
+ right: 10px;background-color: #f44336;
|
||
+ color: white;padding: 8px 12px;
|
||
+ border-radius: 6px;font-size: 14px;
|
||
+ font-weight: bold;z-index: 9999;
|
||
+ transition: background-color 0.3s;
|
||
+ cursor: pointer;box-shadow: 0 3px 6px rgba(0,0,0,0.25);
|
||
+ `;
|
||
+ ind.innerText = log_prefix + '正在连接...';
|
||
+
|
||
+ ind.addEventListener('click', () => alert(`会话ID: ${sid}\n当前URL: ${location.href}`));
|
||
+ document.body.appendChild(ind);
|
||
+ }
|
||
+
|
||
+ return ind;
|
||
+ }
|
||
+
|
||
+ // 更新状态
|
||
+ function updateStatus(status, msg) {
|
||
+ if (!document.body) return setTimeout(() => updateStatus(status, msg), 100);
|
||
+
|
||
+ const ind = getIndicator();
|
||
+ if (!ind) return;
|
||
+
|
||
+ if (status === 'ok') {
|
||
+ ind.style.backgroundColor = '#4CAF50';
|
||
+ ind.innerText = log_prefix + '连接成功';
|
||
+ } else if (status === 'disc') {
|
||
+ ind.style.backgroundColor = '#f44336';
|
||
+ ind.innerText = log_prefix + '连接断开';
|
||
+ } else if (status === 'conn') {
|
||
+ ind.style.backgroundColor = '#2196F3';
|
||
+ ind.innerText = log_prefix + '正在连接(HTTP)';
|
||
+ } else if (status === 'err') {
|
||
+ ind.style.backgroundColor = '#FF9800';
|
||
+ ind.innerText = log_prefix + `发生错误 (${msg})`;
|
||
+ } else if (status === 'exec') {
|
||
+ ind.style.backgroundColor = '#2196F3';
|
||
+ ind.innerText = log_prefix + '正在执行指令...';
|
||
+ }
|
||
+ }
|
||
+
|
||
+ function handleError(id, error, errorSource) {
|
||
+ console.error(`${errorSource}错误:`, error);
|
||
+ updateStatus('err', error.message);
|
||
+
|
||
+ const errorMessage = {
|
||
+ type: 'error',
|
||
+ id: id,
|
||
+ sessionId: sid,
|
||
+ error: {
|
||
+ name: error.name,
|
||
+ message: error.message,
|
||
+ stack: error.stack,
|
||
+ source: errorSource
|
||
+ }
|
||
+ };
|
||
+
|
||
+ if (typeof ws !== 'undefined' && ws && ws.readyState === WebSocket.OPEN) {
|
||
+ ws.send(JSON.stringify(errorMessage));
|
||
+ } else {
|
||
+ GM_xmlhttpRequest({
|
||
+ method: "POST",
|
||
+ url: httpUrl + "api/result",
|
||
+ headers: {"Content-Type": "application/json"},
|
||
+ data: JSON.stringify(errorMessage),
|
||
+ onload: function(response) {console.log("错误信息已通过HTTP发送", response);},
|
||
+ onerror: function(err) {console.error("发送错误信息失败", err);}
|
||
+ });
|
||
+ }
|
||
+ }
|
||
+
|
||
+ function smartProcessResult(result) {
|
||
+ // 处理 null 和原始类型
|
||
+ if (result === null || result === undefined || typeof result !== 'object') {
|
||
+ return result;
|
||
+ }
|
||
+
|
||
+ // 1. 处理 jQuery 对象 - 强制转换为HTML字符串数组
|
||
+ if (typeof jQuery !== 'undefined' && result instanceof jQuery) {
|
||
+ const elements = [];
|
||
+ for (let i = 0; i < result.length; i++) {
|
||
+ if (result[i] && result[i].nodeType === 1) {
|
||
+ elements.push(result[i].outerHTML);
|
||
+ }
|
||
+ }
|
||
+ return elements; // 始终返回数组
|
||
+ }
|
||
+
|
||
+ // 2. 处理 NodeList 和 HTMLCollection
|
||
+ if (result instanceof NodeList || result instanceof HTMLCollection) {
|
||
+ const elements = [];
|
||
+ for (let i = 0; i < result.length; i++) {
|
||
+ if (result[i] && result[i].nodeType === 1) {
|
||
+ elements.push(result[i].outerHTML);
|
||
+ }
|
||
+ }
|
||
+ return elements;
|
||
+ }
|
||
+
|
||
+ // 3. 处理单个 DOM 元素
|
||
+ if (result.nodeType === 1) {
|
||
+ return result.outerHTML;
|
||
+ }
|
||
+
|
||
+ // 4. 检查是否是具有数字索引和length属性的类数组对象
|
||
+ if (!Array.isArray(result) &&
|
||
+ typeof result === 'object' &&
|
||
+ 'length' in result &&
|
||
+ typeof result.length === 'number') {
|
||
+
|
||
+ // 检查第一个元素是否是DOM节点
|
||
+ const firstElement = result[0];
|
||
+ if (firstElement && firstElement.nodeType === 1) {
|
||
+ const elements = [];
|
||
+ const length = Math.min(result.length, 100);
|
||
+
|
||
+ for (let i = 0; i < length; i++) {
|
||
+ const elem = result[i];
|
||
+ if (elem && elem.nodeType === 1) {
|
||
+ elements.push(elem.outerHTML);
|
||
+ }
|
||
+ }
|
||
+
|
||
+ return elements;
|
||
+ }
|
||
+ }
|
||
+
|
||
+ // 5. 处理普通对象和数组 - 使用标准序列化
|
||
+ try {
|
||
+ return JSON.parse(JSON.stringify(result, function(key, value) {
|
||
+ if (typeof value === 'object' && value !== null) {
|
||
+ if (value.nodeType === 1) {
|
||
+ return value.outerHTML;
|
||
+ }
|
||
+ if (value === window || value === document) {
|
||
+ return '[Object]';
|
||
+ }
|
||
+ }
|
||
+ return value;
|
||
+ }));
|
||
+ } catch (e) {
|
||
+ console.error("序列化对象失败:", e);
|
||
+ return `[无法序列化的对象: ${e.message}]`;
|
||
+ }
|
||
+ }
|
||
+
|
||
+ // 防止重复初始化
|
||
+ if (window.ljq_init) return;
|
||
+ window.ljq_init = true;
|
||
+
|
||
+ function connecthttp() {
|
||
+ if (window.use_ws) return;
|
||
+ updateStatus('conn');
|
||
+ GM_xmlhttpRequest({
|
||
+ method: "POST",
|
||
+ url: httpUrl + "api/longpoll",
|
||
+ headers: {"Content-Type": "application/json"},
|
||
+ data: JSON.stringify({
|
||
+ type: 'ready',
|
||
+ url: location.href,
|
||
+ sessionId: sid
|
||
+ }),
|
||
+ onload: function(resp) {
|
||
+ if (resp.status === 200) {
|
||
+ let data = JSON.parse(resp.responseText);
|
||
+ console.log(log_prefix + '接收到数据:', data);
|
||
+ if (data.id === "" && data.ret === "use ws") return;
|
||
+ if (data.id === "") return setTimeout(connecthttp, 100);
|
||
+ const response = executeCode(data);
|
||
+
|
||
+ if (response.error) {
|
||
+ handleError(data.id, response.error, '执行代码');
|
||
+ } else {
|
||
+ GM_xmlhttpRequest({
|
||
+ method: "POST",
|
||
+ url: httpUrl + "api/result",
|
||
+ headers: {"Content-Type": "application/json"},
|
||
+ data: JSON.stringify({
|
||
+ type: 'result',
|
||
+ id: data.id,
|
||
+ sessionId: sid,
|
||
+ result: response.result
|
||
+ })
|
||
+ });
|
||
+ }
|
||
+ } else {
|
||
+ console.error(log_prefix + '请求失败，状态码：', resp.status);
|
||
+ updateStatus('err', '请求失败');
|
||
+ }
|
||
+ setTimeout(connecthttp, 1000);
|
||
+ },
|
||
+ onerror: function(err) {
|
||
+ console.error(log_prefix + '请求错误', err);
|
||
+ updateStatus('err', '请求失败');
|
||
+ setTimeout(connecthttp, 5000);
|
||
+ },
|
||
+ ontimeout: function() {
|
||
+ console.log(log_prefix + '请求超时');
|
||
+ updateStatus('err', '请求超时');
|
||
+ setTimeout(connecthttp, 5000);
|
||
+ }
|
||
+ });
|
||
+ }
|
||
+
|
||
+ function executeCode(data) {
|
||
+ let id = data.id || 'unknown'; // 获取 ID
|
||
+ let result;
|
||
+
|
||
+ if (!data.code) {
|
||
+ console.log('收到非代码执行消息:', data);
|
||
+ return { error: '没有可执行的代码' };
|
||
+ }
|
||
+ updateStatus('exec');
|
||
+
|
||
+ try {
|
||
+ const jsCode = data.code.trim();
|
||
+ const lines = jsCode.split(/\r?\n/).filter(l => l.trim());
|
||
+ const lastLine = lines.length > 0 ? lines[lines.length - 1].trim() : '';
|
||
+
|
||
+ if (lastLine.startsWith('return')) {
|
||
+ // 最后一行包含 return 语句，使用 Function 构造器
|
||
+ result = (new Function(jsCode))();
|
||
+ } else {
|
||
+ try {
|
||
+ result = eval(jsCode);
|
||
+ } catch (e) {
|
||
+ if (isIllegalReturnError(e)) {
|
||
+ result = (new Function(jsCode))();
|
||
+ } else {
|
||
+ throw e;
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+ const processedResult = smartProcessResult(result);
|
||
+ return { result: processedResult };
|
||
+
|
||
+ } catch (execError) {
|
||
+ return { error: execError }; // 返回错误信息
|
||
+ }
|
||
+ }
|
||
+
|
||
+ function isIllegalReturnError(e) {
|
||
+ return e instanceof SyntaxError && (
|
||
+ /Illegal return statement/i.test(e.message) || // Chrome 常见
|
||
+ /return not in function/i.test(e.message) || // Firefox 常见
|
||
+ /Illegal 'return' statement/i.test(e.message) // 兼容旧文案
|
||
+ );
|
||
+ }
|
||
+
|
||
+ function connect() {
|
||
+ ws = new WebSocket(wsUrl);
|
||
+
|
||
+ ws.onopen = function() {
|
||
+ window.use_ws = true;
|
||
+ console.log(log_prefix + '已连接');
|
||
+ updateStatus('ok');
|
||
+ ws.send(JSON.stringify({
|
||
+ type: 'ready',
|
||
+ url: location.href,
|
||
+ sessionId: sid
|
||
+ }));
|
||
+ };
|
||
+
|
||
+ ws.onclose = function() {
|
||
+ console.log(log_prefix + '已断开，5秒后重连');
|
||
+ updateStatus('disc');
|
||
+ setTimeout(connect, 5000);
|
||
+ };
|
||
+
|
||
+ ws.onerror = function(err) {
|
||
+ console.error(log_prefix + '连接错误', err);
|
||
+ updateStatus('err', '连接失败');
|
||
+ isWebSocketServerAlive(function (e) { if (e) connecthttp()});
|
||
+ };
|
||
+
|
||
+ ws.onmessage = async function(e) {
|
||
+ try {
|
||
+ let data = JSON.parse(e.data);
|
||
+ let startTime = Date.now();
|
||
+ let newTabs = [];
|
||
+ let checkNewTab = data.auto_switch_newtab === true;
|
||
+ GM_setValue('new_tab_report', null);
|
||
+ const response = executeCode(data);
|
||
+
|
||
+ if (response.error) {
|
||
+ handleError(data.id, response.error, '执行代码');
|
||
+ } else {
|
||
+ if (checkNewTab) {
|
||
+ for (let i = 0; i < 10; i++) {
|
||
+ await new Promise(r => setTimeout(r, 150));
|
||
+ let latestReport = GM_getValue('new_tab_report');
|
||
+ if (latestReport && latestReport.ts >= startTime) {
|
||
+ console.log(`%c[Detected] 轮询第 ${i+1} 次抓到新标签!`, "color: green");
|
||
+ newTabs.push(latestReport);
|
||
+ break;
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+ updateStatus('ok');
|
||
+ ws.send(JSON.stringify({
|
||
+ type: 'result',
|
||
+ id: data.id,
|
||
+ sessionId: sid,
|
||
+ result: response.result,
|
||
+ newTabs: newTabs
|
||
+ }));
|
||
+ }
|
||
+ } catch (parseError) {
|
||
+ handleError('unknown', parseError, '解析消息');
|
||
+ }
|
||
+ };
|
||
+
|
||
+ }
|
||
+
|
||
+ // 初始化
|
||
+ function init() {
|
||
+ if (document.body) {
|
||
+ getIndicator();
|
||
+ connect();
|
||
+ } else {
|
||
+ setTimeout(init, 50);
|
||
+ }
|
||
+ }
|
||
+
|
||
+ // 监控DOM变化
|
||
+ const observer = new MutationObserver(() => getIndicator());
|
||
+
|
||
+ if (document.readyState !== 'loading') {
|
||
+ init();
|
||
+ observer.observe(document.body, { childList: true, subtree: true });
|
||
+ } else {
|
||
+ document.addEventListener('DOMContentLoaded', () => {
|
||
+ init();
|
||
+ observer.observe(document.body, { childList: true, subtree: true });
|
||
+ });
|
||
+ }
|
||
+
|
||
+ // 清理
|
||
+ window.addEventListener('beforeunload', () => {
|
||
+ observer.disconnect();
|
||
+ if (ws && ws.readyState === WebSocket.OPEN) {
|
||
+ ws.close();
|
||
+ }
|
||
+ });
|
||
+})();
|
||
\ No newline at end of file
|
||
diff --git a/make_prompts.py b/make_prompts.py
|
||
new file mode 100644
|
||
index 0000000..5905083
|
||
--- /dev/null
|
||
+++ b/make_prompts.py
|
||
@@ -0,0 +1,137 @@
|
||
+import sys, os, re
|
||
+import pyperclip
|
||
+import json, time
|
||
+from pathlib import Path
|
||
+import subprocess
|
||
+import tempfile
|
||
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
|
||
+from sidercall import SiderLLMSession, LLMSession, ToolClient
|
||
+
|
||
+
|
||
+ask = SiderLLMSession().ask
|
||
+
|
||
+
|
||
+def generate_tool_schema():
|
||
+ """
|
||
+ 通过代码内省，将 Handler 的逻辑映射为高语义的工具描述。
|
||
+ """
|
||
+ with open('ga.py', 'r', encoding='utf-8') as f:
|
||
+ ga_code = f.read()
|
||
+ # 极简且具备高度概括能力的元 Prompt
|
||
+ meta_prompt = f"""
|
||
+# Role
|
||
+你是一个具备深度推理能力的 AI 系统架构师。你将通过阅读 `GenericAgentHandler` 源码，构建其对应的工具能力矩阵。
|
||
+
|
||
+# Task
|
||
+分析下方的源码，并输出 OpenAI Tool Schema。在输出 JSON 之前，你必须进行内部思考（Thinking Process）。
|
||
+
|
||
+# Thinking Process Requirements
|
||
+在 `<thinking>` 标签中，请按顺序分析：
|
||
+1. **核心工具链识别**：识别所有 `do_xxx` 方法，并分析它们依赖的底层 Utility 函数。
|
||
+2. **内容溯源审计**：重点分析哪些工具是从 `response.content` 提取核心逻辑（如代码块）的。对于这些工具，确认在 Schema 参数中排除掉对应的字段。
|
||
+3. **调用策略推导**：分析工具间的协作关系（例如 `file_read` 如何为 `file_patch` 提供定位）。
|
||
+4. **兜底逻辑确认**：明确某些特殊万能工具在系统中的保底角色，快速工具无法执行的操作由保底工具执行，但正常应优先使用方便的工具。
|
||
+5. **注释审阅**：结合函数注释，理解每个工具的使用限制，其中的重要信息务必反映在工具描述中（如长度限制等）。
|
||
+注释中的重要信息务必反映在工具描述中。
|
||
+注释中的重要信息务必反映在工具描述中。
|
||
+
|
||
+# Tool Schema Formatting Rules
|
||
+- **参数对齐**：仅包含 `do_xxx` 方法中通过 `args.get()` 显式获取的参数。
|
||
+- **高引导性描述**：描述应包含“何时调用”以及“如何根据反馈修正”，需要注意函数的注释事项。
|
||
+- **输出格式**：先输出 `<thinking>` 块，然后输出 ```json 块。
|
||
+
|
||
+# Source Code
|
||
+{ga_code}
|
||
+
|
||
+# Output
|
||
+请开始思考并生成：
|
||
+"""
|
||
+
|
||
+ # 假设 ask 是你已经封装好的 LLM 调用接口
|
||
+ raw_response = ask(meta_prompt, model="gemini-3.0-flash")
|
||
+ print(raw_response)
|
||
+
|
||
+ # --- 健壮的 JSON 解析逻辑 ---
|
||
+ try:
|
||
+ # 1. 清除 Markdown 围栏
|
||
+ clean_json = raw_response.strip()
|
||
+ if clean_json.startswith("```"):
|
||
+ # 兼容 ```json 和 ```
|
||
+ clean_json = re.sub(r'^```(?:json)?\s*', '', clean_json)
|
||
+ clean_json = re.sub(r'\s*```$', '', clean_json)
|
||
+
|
||
+ # 2. 移除可能的非 JSON 前导/后缀文字（如果有的话）
|
||
+ start_idx = clean_json.find('[')
|
||
+ end_idx = clean_json.rfind(']') + 1
|
||
+ if start_idx != -1 and end_idx != -1:
|
||
+ clean_json = clean_json[start_idx:end_idx]
|
||
+
|
||
+ final_schema = json.loads(clean_json)
|
||
+
|
||
+ if final_schema:
|
||
+ with open('tools_schema.json', 'w', encoding='utf-8') as f:
|
||
+ json.dump(final_schema, f, indent=2, ensure_ascii=False)
|
||
+ print("✅ 成功从代码内省生成 Schema 并持久化。")
|
||
+ return final_schema
|
||
+
|
||
+ except Exception as e:
|
||
+ print(f"❌ 解析 Schema 失败: {e}\n原始响应: {raw_response}")
|
||
+ return None
|
||
+
|
||
+
|
||
+def make_system_prompt(ga_code_path='ga.py'):
|
||
+ with open(ga_code_path, 'r', encoding='utf-8') as f:
|
||
+ ga_code = f.read()
|
||
+
|
||
+ # 这个元 Prompt 的目标是生成“世界观”而非“说明书”
|
||
+ meta_prompt = f"""
|
||
+# Role
|
||
+你是一个 AI 架构师。请阅读下方的工具库源码，为 Agent 生成一份【系统级认知指令 (System Prompt)】。
|
||
+
|
||
+# Task
|
||
+基于代码逻辑，定义 Agent 的“能力边界”和“行动协议”。
|
||
+
|
||
+# Requirements (Crucial)
|
||
+1. **打破预训练偏见**：针对模型常说的“我只是 AI，不能操作网页/文件”进行修正。明确告诉它：你现在拥有物理操作权限，工具设计保证了所有权限做所有事情。
|
||
+2. **避开冗余**：不要重复 Tool Schema 里的参数细节。
|
||
+3. **能力边界定义**：
|
||
+ - 网页操作：它不是通过“想象”上网，而是通过实时的浏览器读写。
|
||
+ - 文件操作：它拥有物理文件读写权限，且遵循“先读后写”的稳健性原则。
|
||
+ - 保底逻辑：当专用工具失效时，使用 `code_run` 编写脚本解决一切。
|
||
+ - 特殊的update_plan（仅在复杂任务时使用）和ask_user（用户也是有效资源）工具。
|
||
+4. **行动协议**：
|
||
+ - 必须在行动前进行<thinking>
|
||
+
|
||
+我后面还会附上具体的工具描述和Schema，所以不要重复。
|
||
+主要以世界观为主，不要纠结于具体工具。
|
||
+
|
||
+# Input Source Code
|
||
+{ga_code}
|
||
+
|
||
+# Output
|
||
+仅输出 System Prompt 的正文，语气要果断、指令化。
|
||
+"""
|
||
+ print("🧠 正在重塑 Agent 世界观 (Generating System Prompt)...")
|
||
+ # 调用你的 llmclient.ask
|
||
+ system_prompt_content = ask(meta_prompt)
|
||
+ print("📝 生成的 System Prompt 内容如下：\n")
|
||
+ print(system_prompt_content)
|
||
+ clean_content = re.sub(r'<[^>]+>', '', system_prompt_content)
|
||
+ with open('sys_prompt.txt', 'w', encoding='utf-8') as f:
|
||
+ f.write(clean_content)
|
||
+ return clean_content
|
||
+
|
||
+# --- 主逻辑 ---
|
||
+if __name__ == "__main__":
|
||
+ if len(sys.argv) < 2:
|
||
+ print("Usage: python make_prompts.py [schema|prompt]")
|
||
+ sys.exit(1)
|
||
+
|
||
+ cmd = sys.argv[1].lower()
|
||
+ if cmd == "schema":
|
||
+ generate_tool_schema()
|
||
+ elif cmd == "prompt":
|
||
+ make_system_prompt()
|
||
+ else:
|
||
+ print(f"Unknown command: {cmd}")
|
||
+ print("Available commands: schema, prompt")
|
||
\ No newline at end of file
|
||
diff --git a/sidercall.py b/sidercall.py
|
||
new file mode 100644
|
||
index 0000000..706c686
|
||
--- /dev/null
|
||
+++ b/sidercall.py
|
||
@@ -0,0 +1,179 @@
|
||
+import os, json, re, time, requests
|
||
+from sider_ai_api import Session
|
||
+
|
||
+try:
|
||
+ from mykey import sider_cookie, capikey
|
||
+except ImportError:
|
||
+ sider_cookie = ""
|
||
+ capikey = ""
|
||
+
|
||
+class SiderLLMSession:
|
||
+ def __init__(self, multiturns=6):
|
||
+ self._core = Session(cookie=sider_cookie, proxies={'https':'127.0.0.1:2082'})
|
||
+ def ask(self, prompt, model="gemini-3.0-flash"):
|
||
+ if len(prompt) > 30000: prompt = prompt[-29500:]
|
||
+ return ''.join(self._core.chat(prompt, model))
|
||
+
|
||
+class LLMSession:
|
||
+ def __init__(self, api_key=capikey, api_base="http://113.45.39.247:3001/v1", multiturns=6):
|
||
+ self.api_key = api_key
|
||
+ self.api_base = api_base
|
||
+ self.messages = []
|
||
+ self.multiturns = multiturns
|
||
+
|
||
+ def ask(self, prompt, model="openai/gpt-5.1"):
|
||
+ self.messages.append({"role": "user", "content": prompt})
|
||
+ if len(self.messages) > self.multiturns:
|
||
+ self.messages = self.messages[-self.multiturns:]
|
||
+ headers = {
|
||
+ "Authorization": f"Bearer {self.api_key}",
|
||
+ "Content-Type": "application/json"
|
||
+ }
|
||
+ try:
|
||
+ response = requests.post(
|
||
+ f"{self.api_base}/chat/completions",
|
||
+ headers=headers,
|
||
+ json={
|
||
+ "model": model,
|
||
+ "messages": self.messages,
|
||
+ "temperature": 0.5
|
||
+ },
|
||
+ timeout=60
|
||
+ )
|
||
+ res_json = response.json()
|
||
+ content = res_json["choices"][0]["message"]["content"]
|
||
+ self.messages.append({"role": "assistant", "content": content})
|
||
+ return content
|
||
+ except Exception as e:
|
||
+ return f"Error: {str(e)}"
|
||
+
|
||
+class MockFunction:
|
||
+ def __init__(self, name, arguments):
|
||
+ self.name = name
|
||
+ self.arguments = arguments
|
||
+
|
||
+class MockToolCall:
|
||
+ def __init__(self, name, args):
|
||
+ arg_str = json.dumps(args, ensure_ascii=False) if isinstance(args, dict) else args
|
||
+ self.function = MockFunction(name, arg_str)
|
||
+
|
||
+class MockResponse:
|
||
+ def __init__(self, thinking, content, tool_calls, raw):
|
||
+ self.thinking = thinking # 存放 <thinking> 内部的思维过程
|
||
+ self.content = content # 存放去除标签后的纯文本回复
|
||
+ self.tool_calls = tool_calls # 存放 MockToolCall 列表 或 None
|
||
+ self.raw = raw
|
||
+ def __repr__(self):
|
||
+ return f"<MockResponse thinking={bool(self.thinking)}, content='{self.content}', tools={bool(self.tool_calls)}>"
|
||
+
|
||
+class ToolClient:
|
||
+ def __init__(self, raw_api_func, auto_save_tokens=False):
|
||
+ self.raw_api = raw_api_func
|
||
+ self.auto_save_tokens = auto_save_tokens
|
||
+ self.last_tools = ''
|
||
+ self.total_cd_tokens = 0
|
||
+
|
||
+ def chat(self, messages, tools=None):
|
||
+ full_prompt = self._build_protocol_prompt(messages, tools)
|
||
+ print("Full prompt length:", len(full_prompt))
|
||
+ raw_text = self.raw_api(full_prompt)
|
||
+ with open('model_responses.txt', 'a', encoding='utf-8', errors="replace") as f:
|
||
+ f.write(f"=== Prompt ===\n{full_prompt}\n=== Response ===\n{raw_text}\n\n")
|
||
+ return self._parse_mixed_response(raw_text)
|
||
+
|
||
+ def _build_protocol_prompt(self, messages, tools):
|
||
+ system_content = next((m['content'] for m in messages if m['role'].lower() == 'system'), "你是一个智能助手。")
|
||
+ history_msgs = [m for m in messages if m['role'].lower() != 'system']
|
||
+
|
||
+ # 构造工具描述
|
||
+ tool_instruction = ""
|
||
+ if tools:
|
||
+ tools_json = json.dumps(tools, ensure_ascii=False, indent=2)
|
||
+ tool_instruction = f"""
|
||
+### ⚡️ 交互协议 (必须严格遵守)
|
||
+请按照以下步骤思考并行动：
|
||
+1. **思考**: 在 `<thinking>` 标签中分析现状和策略。
|
||
+2. **行动**: 如果需要调用工具，请紧接着输出一个 **<tool_use>块**，然后结束，我会稍后给你返回<tool_result>块。
|
||
+ 格式: ```<tool_use>\n{{"function": "工具名", "arguments": {{参数}}}}\n</tool_use>\n```
|
||
+
|
||
+### 🛠️ 可用工具库
|
||
+{tools_json}
|
||
+"""
|
||
+ if self.auto_save_tokens and self.last_tools == tools_json:
|
||
+ tool_instruction = "\n### ⚡️ 交互协议保持不变，继续使用之前的工具库。\n"
|
||
+ else:
|
||
+ self.total_cd_tokens = 0
|
||
+ self.last_tools = tools_json
|
||
+
|
||
+ prompt = f"=== SYSTEM ===\n{system_content}\n{tool_instruction}\n\n"
|
||
+ for m in history_msgs:
|
||
+ role = "USER" if m['role'] == 'user' else "ASSISTANT"
|
||
+ prompt += f"=== {role} ===\n{m['content']}\n\n"
|
||
+
|
||
+ self.total_cd_tokens += len(prompt)
|
||
+ if self.total_cd_tokens > 6000: self.last_tools = ''
|
||
+
|
||
+ prompt += "=== ASSISTANT ===\n"
|
||
+ return prompt
|
||
+
|
||
+ def _parse_mixed_response(self, text):
|
||
+ remaining_text = text
|
||
+ thinking = ''
|
||
+ think_pattern = r"<thinking>(.*?)</thinking>"
|
||
+ think_match = re.search(think_pattern, text, re.DOTALL)
|
||
+
|
||
+ if think_match:
|
||
+ thinking = think_match.group(1).strip()
|
||
+ remaining_text = re.sub(think_pattern, "", remaining_text, flags=re.DOTALL)
|
||
+
|
||
+ tool_calls = None
|
||
+ tool_pattern = r"<tool_use>(.*?)</tool_use>"
|
||
+ tool_match = re.search(tool_pattern, text, re.DOTALL)
|
||
+
|
||
+ json_str = ""
|
||
+ if tool_match:
|
||
+ json_str = tool_match.group(1).strip()
|
||
+ remaining_text = re.sub(tool_pattern, "", remaining_text, flags=re.DOTALL)
|
||
+ elif '<tool_use>' in remaining_text:
|
||
+ weaktoolstr = remaining_text.split('<tool_use>')[-1].strip()
|
||
+ json_str = weaktoolstr if weaktoolstr.endswith('}') else ''
|
||
+ remaining_text = remaining_text.replace('<tool_use>'+weaktoolstr, "")
|
||
+
|
||
+ if json_str:
|
||
+ try:
|
||
+ data = tryparse(json_str)
|
||
+ func_name = data.get('function') or data.get('tool')
|
||
+ args = data.get('arguments') or data.get('args')
|
||
+ if args is None: args = {}
|
||
+ if func_name: tool_calls = [MockToolCall(func_name, args)]
|
||
+ except json.JSONDecodeError:
|
||
+ print("[Warn] Failed to parse tool_use JSON:", json_str)
|
||
+ thinking += f"[Warn] JSON 解析失败，模型输出了无效的 JSON."
|
||
+
|
||
+ content = remaining_text.strip()
|
||
+ if not content: content = ""
|
||
+ return MockResponse(thinking, content, tool_calls, text)
|
||
+
|
||
+def tryparse(json_str):
|
||
+ try: return json.loads(json_str)
|
||
+ except:
|
||
+ return json.loads(json_str[:-1])
|
||
+
|
||
+if __name__ == "__main__":
|
||
+ llmclient = ToolClient(LLMSession().ask)
|
||
+ response = llmclient.chat(
|
||
+ messages=[{"role": "user", "content": "我的IP是多少"}],
|
||
+ tools=[{"name": "get_ip", "parameters": {}}]
|
||
+ )
|
||
+ # 4. 获取结果
|
||
+ print(f"思考: {response.thinking}")
|
||
+ # -> 我需要查一下 IP。
|
||
+
|
||
+ if response.tool_calls:
|
||
+ cmd = response.tool_calls[0]
|
||
+ print(f"调用: {cmd.function.name} 参数: {cmd.function.arguments}")
|
||
+
|
||
+ response = llmclient.chat(
|
||
+ messages=[{"role": "user", "content": "<tool_result>10.176.45.12</tool_result>"}]
|
||
+ )
|
||
+ print(response.content)
|
||
\ No newline at end of file
|
||
diff --git a/simphtml.py b/simphtml.py
|
||
new file mode 100644
|
||
index 0000000..d555460
|
||
--- /dev/null
|
||
+++ b/simphtml.py
|
||
@@ -0,0 +1,862 @@
|
||
+from bs4 import BeautifulSoup
|
||
+
|
||
+js_optHTML = '''function optHTML() {
|
||
+function createEnhancedDOMCopy() {
|
||
+ const nodeInfo = new WeakMap();
|
||
+ const ignoreTags = ['SCRIPT', 'STYLE', 'NOSCRIPT', 'META', 'LINK', 'COLGROUP', 'COL', 'TEMPLATE', 'PARAM', 'SOURCE'];
|
||
+ const ignoreIds = ['ljq-ind'];
|
||
+ function cloneNode(sourceNode, keep=false) {
|
||
+ if (sourceNode.nodeType === 8 ||
|
||
+ (sourceNode.nodeType === 1 && (
|
||
+ ignoreTags.includes(sourceNode.tagName) ||
|
||
+ (sourceNode.id && ignoreIds.includes(sourceNode.id))
|
||
+ ))) {
|
||
+ return null;
|
||
+ }
|
||
+ if (sourceNode.nodeType === 3) return sourceNode.cloneNode(false);
|
||
+ const clone = sourceNode.cloneNode(false);
|
||
+
|
||
+ const isDropdown = sourceNode.classList?.contains('dropdown-menu') ||
|
||
+ /dropdown|menu/i.test(sourceNode.className) || sourceNode.getAttribute('role') === 'menu';
|
||
+ const isSmallDropdown = isDropdown && (sourceNode.querySelectorAll('a, button, [role="menuitem"], li').length <= 7 && sourceNode.textContent.length < 500);
|
||
+
|
||
+ const childNodes = [];
|
||
+ for (const child of sourceNode.childNodes) {
|
||
+ const childClone = cloneNode(child, keep || isSmallDropdown);
|
||
+ if (childClone) childNodes.push(childClone);
|
||
+ }
|
||
+
|
||
+ const rect = sourceNode.getBoundingClientRect();
|
||
+ const style = window.getComputedStyle(sourceNode);
|
||
+ const area = (style.display === 'none' || style.visibility === 'hidden' || parseFloat(style.opacity) <= 0)?0:rect.width * rect.height;
|
||
+ const isVisible = (rect.width > 1 && rect.height > 1 &&
|
||
+ style.display !== 'none' && style.visibility !== 'hidden' &&
|
||
+ parseFloat(style.opacity) > 0 &&
|
||
+ Math.abs(rect.left) < 5000 && Math.abs(rect.top) < 5000)
|
||
+ || isSmallDropdown;
|
||
+ const zIndex = style.position !== 'static' ? (parseInt(style.zIndex) || 0) : 0;
|
||
+
|
||
+ let info = {
|
||
+ rect, area, isVisible, isSmallDropdown, zIndex,
|
||
+ style: {
|
||
+ display: style.display, visibility: style.visibility,
|
||
+ opacity: style.opacity, position: style.position
|
||
+ }};
|
||
+
|
||
+ const nonTextChildren = childNodes.filter(child => child.nodeType !== 3);
|
||
+ const hasValidChildren = nonTextChildren.length > 0;
|
||
+
|
||
+ if (!isVisible && nonTextChildren.length > 0) {
|
||
+ const visChild = nonTextChildren.find(child =>
|
||
+ nodeInfo.has(child) && nodeInfo.get(child).isVisible);
|
||
+ if (visChild) info = nodeInfo.get(visChild);
|
||
+ }
|
||
+ nodeInfo.set(clone, info);
|
||
+
|
||
+ if (sourceNode.nodeType === 1 && sourceNode.tagName === 'DIV') {
|
||
+ if (!hasValidChildren && !sourceNode.textContent.trim()) return null;
|
||
+ }
|
||
+ if (info.isVisible || hasValidChildren || keep) {
|
||
+ childNodes.forEach(child => clone.appendChild(child));
|
||
+ return clone;
|
||
+ }
|
||
+ return null;
|
||
+ }
|
||
+
|
||
+ return {
|
||
+ domCopy: cloneNode(document.body),
|
||
+ getNodeInfo: node => nodeInfo.get(node),
|
||
+ isVisible: node => {
|
||
+ const info = nodeInfo.get(node);
|
||
+ return info && info.isVisible;
|
||
+ }
|
||
+ };
|
||
+}
|
||
+const { domCopy, getNodeInfo, isVisible } = createEnhancedDOMCopy();
|
||
+const viewportArea = window.innerWidth * window.innerHeight;
|
||
+
|
||
+function analyzeNode(node, pPathType='main') {
|
||
+ // 处理非元素节点和叶节点
|
||
+ if (node.nodeType !== 1 || !node.children.length) {
|
||
+ node.nodeType === 1 && (node.dataset.mark = 'K:leaf');
|
||
+ return;
|
||
+ }
|
||
+ const pathType = (node.dataset.mark && !node.dataset.mark.includes(':main')) ? 'second' : pPathType;
|
||
+ const rectn = getNodeInfo(node).rect;
|
||
+ if (rectn.width < window.innerWidth * 0.8 && rectn.height < window.innerHeight * 0.8) return node;
|
||
+ if (node.tagName === 'TABLE') return;
|
||
+ const children = Array.from(node.children);
|
||
+ if (children.length === 1) {
|
||
+ node.dataset.mark = 'K:container';
|
||
+ return analyzeNode(children[0], pathType);
|
||
+ }
|
||
+ if (children.length > 10) return;
|
||
+
|
||
+ // 获取子元素信息并排序
|
||
+ const childrenInfo = children.map(child => {
|
||
+ const info = getNodeInfo(child) || { rect: {}, style: {} };
|
||
+ return { node: child, rect: info.rect, style: info.style,
|
||
+ area: info.area, zIndex: info.zIndex };
|
||
+ }).sort((a, b) => b.area - a.area);
|
||
+
|
||
+ // 检测是划分还是覆盖
|
||
+ const isOverlay = hasOverlap(childrenInfo);
|
||
+ node.dataset.mark = isOverlay ? 'K:overlayParent' : 'K:partitionParent';
|
||
+
|
||
+ if (isOverlay) handleOverlayContainer(childrenInfo, pathType);
|
||
+ else handlePartitionContainer(childrenInfo, pathType);
|
||
+
|
||
+ console.log(`${isOverlay ? '覆盖' : '划分'}容器:`, node, `子元素数量: ${children.length}`);
|
||
+ console.log('子元素及标记:', children.map(child => ({
|
||
+ element: child,
|
||
+ mark: child.dataset.mark || '无',
|
||
+ info: getNodeInfo ? getNodeInfo(child) : undefined
|
||
+ })));
|
||
+ for (const child of children)
|
||
+ if (!child.dataset.mark || child.dataset.mark[0] !== 'R') analyzeNode(child, pathType);
|
||
+ }
|
||
+
|
||
+ // 处理划分容器
|
||
+ function handlePartitionContainer(childrenInfo, pathType) {
|
||
+ childrenInfo.sort((a, b) => b.area - a.area);
|
||
+ const totalArea = childrenInfo.reduce((sum, item) => sum + item.area, 0);
|
||
+ console.log(childrenInfo[0].area / totalArea);
|
||
+ const hasMainElement = childrenInfo.length >= 1 &&
|
||
+ (childrenInfo[0].area / totalArea > 0.5) &&
|
||
+ (childrenInfo.length === 1 || childrenInfo[0].area > childrenInfo[1].area * 2);
|
||
+ if (hasMainElement) {
|
||
+ childrenInfo[0].node.dataset.mark = 'K:main';
|
||
+ for (let i = pathType==='main'?1:0; i < childrenInfo.length; i++) {
|
||
+ const child = childrenInfo[i];
|
||
+ let isSecondary = containsButton(child.node);
|
||
+ if (pathType === "main" && child.node.className.toLowerCase().includes('nav')) isSecondary = true;
|
||
+ if (pathType === "main" && child.node.className.toLowerCase().includes('breadcrumbs')) isSecondary = true;
|
||
+ if (pathType === "main" && child.node.className.toLowerCase().includes('header') && child.node.className.toLowerCase().includes('table')) isSecondary = true;
|
||
+ if (pathType === "main" && child.node.innerHTML.trim().replace(/\s+/g, '').length < 500) isSecondary = true;
|
||
+ if (child.style.visibility === 'hidden') isSecondary = false;
|
||
+ if (isSecondary) child.node.dataset.mark = 'K:secondary';
|
||
+ else child.node.dataset.mark = 'R:nonEssential';
|
||
+ }
|
||
+ } else {
|
||
+ const uniqueClassNames = new Set(childrenInfo.map(item => item.node.className)).size;
|
||
+ const highClassNameVariety = uniqueClassNames >= childrenInfo.length * 0.8;
|
||
+ if (pathType !== 'main' && highClassNameVariety && childrenInfo.length > 5) {
|
||
+ childrenInfo.forEach(child => child.node.dataset.mark = 'R:equalmany');
|
||
+ } else {
|
||
+ childrenInfo.forEach(child => child.node.dataset.mark = 'K:equal');
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+
|
||
+ function containsButton(container) {
|
||
+ const hasStandardButton = container.querySelector('button, input[type="button"], input[type="submit"], [role="button"]') !== null;
|
||
+ if (hasStandardButton) return true;
|
||
+ const hasClassButton = container.querySelector('[class*="-btn"], [class*="-button"], .button, .btn, [class*="btn-"]') !== null;
|
||
+ return hasStandardButton || hasClassButton;
|
||
+ }
|
||
+
|
||
+ function handleOverlayContainer(childrenInfo, pathType) {
|
||
+ const sorted = [...childrenInfo].sort((a, b) => b.zIndex - a.zIndex);
|
||
+ console.log('排序后的子元素:', sorted);
|
||
+ if (sorted.length === 0) return;
|
||
+
|
||
+ const top = sorted[0];
|
||
+ const rect = top.rect;
|
||
+ const topNode = top.node;
|
||
+ const isComplex = top.node.querySelectorAll('input, select, textarea, button, a, [role="button"]').length >= 1;
|
||
+
|
||
+ const textContent = topNode.textContent?.trim() || '';
|
||
+ const textLength = textContent.length;
|
||
+ const hasLinks = topNode.querySelectorAll('a').length > 0;
|
||
+ const isMostlyText = textLength > 7 && !hasLinks;
|
||
+
|
||
+ const centerDiff = Math.abs((rect.left + rect.width/2) - window.innerWidth/2) / window.innerWidth;
|
||
+ const minDimensionRatio = Math.min(rect.width / window.innerWidth, rect.height / window.innerHeight);
|
||
+ const maxDimensionRatio = Math.max(rect.width / window.innerWidth, rect.height / window.innerHeight);
|
||
+ const isNearTop = rect.top < 50;
|
||
+ const isDialog = top.node.querySelector('iframe') && centerDiff < 0.3;
|
||
+
|
||
+ if (isComplex && centerDiff < 0.2 &&
|
||
+ ((minDimensionRatio > 0.2 && rect.width/window.innerWidth < 0.98) || minDimensionRatio > 0.95)) {
|
||
+ top.node.dataset.mark = 'K:mainInteractive';
|
||
+ sorted.slice(1).forEach(e => {
|
||
+ if (e.zIndex < sorted[0].zIndex) {
|
||
+ e.node.dataset.mark = 'R:covered';
|
||
+ } else {
|
||
+ e.node.dataset.mark = 'K:noncovered';
|
||
+ }
|
||
+ });
|
||
+ } else {
|
||
+ if (isComplex && isNearTop && maxDimensionRatio > 0.4 && top.isVisible) {
|
||
+ top.node.dataset.mark = 'K:topBar';
|
||
+ } else if (isMostlyText || isComplex || isDialog) {
|
||
+ topNode.dataset.mark = 'K:messageContent';
|
||
+ } else {
|
||
+ topNode.dataset.mark = 'R:floatingAd';
|
||
+ }
|
||
+ const rest = sorted.slice(1);
|
||
+ rest.length && (!hasOverlap(rest) ? handlePartitionContainer(rest, pathType) : handleOverlayContainer(rest, pathType));
|
||
+ }
|
||
+ }
|
||
+
|
||
+ function isValidInteractiveElement(info) {
|
||
+ const { node, rect, style } = info;
|
||
+ const isCentered = Math.abs((rect.left + rect.width/2) - window.innerWidth/2) < window.innerWidth*0.3;
|
||
+ const isVisible = parseFloat(style.opacity) > 0.1;
|
||
+ const isProminent = (parseInt(info.zIndex) > 30 || style.boxShadow !== 'none');
|
||
+ const hasInteractiveElements = node.querySelector('button, a, input') !== null;
|
||
+ return isCentered && isVisible && isProminent && hasInteractiveElements;
|
||
+ }
|
||
+
|
||
+ function hasOverlap(items) {
|
||
+ return items.some((a, i) =>
|
||
+ items.slice(i+1).some(b => {
|
||
+ const r1 = a.rect, r2 = b.rect;
|
||
+ if (!r1.width || !r2.width || !r1.height || !r2.height) {return false;}
|
||
+ const epsilon = 1;
|
||
+ return !(r1.x + r1.width <= r2.x + epsilon || r1.x >= r2.x + r2.width - epsilon ||
|
||
+ r1.y + r1.height <= r2.y + epsilon || r1.y >= r2.y + r2.height - epsilon
|
||
+ );
|
||
+ })
|
||
+ );
|
||
+}
|
||
+
|
||
+const result = analyzeNode(domCopy);
|
||
+domCopy.querySelectorAll('[data-mark^="R:"]').forEach(el=>el.parentNode?.removeChild(el));
|
||
+let root = domCopy;
|
||
+while (root.children.length === 1) {
|
||
+ root = root.children[0];
|
||
+}
|
||
+for (let ii = 0; ii < 3; ii++)
|
||
+ root.querySelectorAll('div').forEach(div => (!div.textContent.trim() && div.children.length === 0) && div.remove());
|
||
+root.querySelectorAll('[data-mark]').forEach(e => e.removeAttribute('data-mark'));
|
||
+root.removeAttribute('data-mark');
|
||
+return root.outerHTML;
|
||
+ }
|
||
+optHTML()'''
|
||
+
|
||
+
|
||
+
|
||
+js_findMainList = '''function findMainList(startElement = null) {
|
||
+ const containerElement = startElement || document.body;
|
||
+ const rect = containerElement.getBoundingClientRect();
|
||
+ const centerX = startElement ? (rect.left + rect.width/2) : (window.innerWidth/2);
|
||
+ const centerY = startElement ? (rect.top + rect.height/2) : (window.innerHeight/2);
|
||
+
|
||
+ // 获取中心元素
|
||
+ const centerElement = document.elementFromPoint(centerX, centerY) || containerElement;
|
||
+ if (!centerElement) return { container: null, items: [] };
|
||
+
|
||
+ // 收集祖先链
|
||
+ const ancestors = [];
|
||
+ for (let current = centerElement; current && ancestors.length < 10; current = current.parentElement) {
|
||
+ ancestors.push(current);
|
||
+ if (current === containerElement) break;
|
||
+ if (containerElement !== document.body && !containerElement.contains(current)) break;
|
||
+ }
|
||
+ if (!ancestors.includes(containerElement)) ancestors.push(containerElement);
|
||
+
|
||
+ let groupCandidates = [];
|
||
+ ancestors.forEach(ancestor => {
|
||
+ const topGroups = findTopGroups(ancestor, 3);
|
||
+ groupCandidates = groupCandidates.concat(topGroups);
|
||
+ });
|
||
+
|
||
+ console.log(groupCandidates);
|
||
+
|
||
+ let candidates = [];
|
||
+ ancestors.forEach(container => {
|
||
+ groupCandidates.forEach(groupInfo => {
|
||
+ // 尝试将组应用到当前容器
|
||
+ const items = findMatchingElements(container, groupInfo.selector);
|
||
+ // 只考虑足够大的组
|
||
+ if (items.length >= 3) {
|
||
+ candidates.push({
|
||
+ container: container,
|
||
+ selector: groupInfo.selector,
|
||
+ items: items,
|
||
+ gscore: groupInfo.score
|
||
+ });
|
||
+ }
|
||
+ });
|
||
+ });
|
||
+
|
||
+ candidates = candidates.map(candidate => {
|
||
+ const score = scoreContainer(candidate.container, candidate.items) + candidate.gscore;
|
||
+ return {...candidate, score};
|
||
+ });
|
||
+
|
||
+ if (candidates.length === 0) {
|
||
+ return { container: centerElement, items: [] };
|
||
+ }
|
||
+
|
||
+ // 3. 选择得分最高的容器
|
||
+ const bestCandidate = candidates.sort((a, b) => b.score - a.score)[0];
|
||
+ console.log(candidates);
|
||
+
|
||
+ // 如果最高分仍然很低，退回到中心元素
|
||
+ if (bestCandidate.score < 30) {
|
||
+ return { container: centerElement, items: [] };
|
||
+ }
|
||
+
|
||
+ return {
|
||
+ container: bestCandidate.container,
|
||
+ items: bestCandidate.items,
|
||
+ selector: bestCandidate.selector,
|
||
+ score: bestCandidate.score
|
||
+ };
|
||
+ }
|
||
+
|
||
+ function findTopGroups(container, limit) {
|
||
+ const children = Array.from(container.children);
|
||
+ const totalChildren = children.length;
|
||
+ if (totalChildren < 3) return [];
|
||
+
|
||
+ const minGroupSize = Math.max(3, Math.floor(totalChildren * 0.2));
|
||
+ const groups = [];
|
||
+
|
||
+ // 统计标签和类名
|
||
+ const tagFreq = {}, classFreq = {}, tagMap = {}, classMap = {};
|
||
+
|
||
+ children.forEach(child => {
|
||
+ // 统计标签
|
||
+ const tag = child.tagName.toLowerCase();
|
||
+ if (tag === "td") return;
|
||
+ tagFreq[tag] = (tagFreq[tag] || 0) + 1;
|
||
+ if (!tagMap[tag]) tagMap[tag] = [];
|
||
+ tagMap[tag].push(child);
|
||
+
|
||
+ // 统计类名
|
||
+ if (child.className) {
|
||
+ child.className.trim().split(/\s+/).forEach(cls => {
|
||
+ if (cls) {
|
||
+ classFreq[cls] = (classFreq[cls] || 0) + 1;
|
||
+ if (!classMap[cls]) classMap[cls] = [];
|
||
+ classMap[cls].push(child);
|
||
+ }
|
||
+ });
|
||
+ }
|
||
+ });
|
||
+
|
||
+ // 评分函数
|
||
+ const scoreGroup = (selector, elements) => {
|
||
+ const coverage = elements.length / totalChildren;
|
||
+ let specificity = selector.startsWith('.')
|
||
+ ? (0.6 + (selector.match(/\./g).length - 1) * 0.1) // 类选择器
|
||
+ : (selector.includes('.')
|
||
+ ? (0.7 + (selector.match(/\./g).length) * 0.1) // 标签+类
|
||
+ : 0.3); // 纯标签
|
||
+ return (coverage * 0.5) + (specificity * 0.5);
|
||
+ };
|
||
+
|
||
+ // 添加标签组
|
||
+ Object.keys(tagFreq).forEach(tag => {
|
||
+ if (tag !== "div" && tagFreq[tag] >= minGroupSize) {
|
||
+ groups.push({
|
||
+ selector: tag,
|
||
+ elements: tagMap[tag],
|
||
+ score: scoreGroup(tag, tagMap[tag]) - 0.5
|
||
+ });
|
||
+ }
|
||
+ });
|
||
+
|
||
+ // 添加类组
|
||
+ Object.keys(classFreq).forEach(cls => {
|
||
+ if (classFreq[cls] >= minGroupSize) {
|
||
+ const selector = '.' + cls;
|
||
+ groups.push({
|
||
+ selector,
|
||
+ elements: classMap[cls],
|
||
+ score: scoreGroup(selector, classMap[cls])
|
||
+ });
|
||
+ }
|
||
+ });
|
||
+ // 添加标签+类组合
|
||
+ const topTags = Object.keys(tagFreq)
|
||
+ .filter(t => tagFreq[t] >= minGroupSize)
|
||
+ .slice(0, 3);
|
||
+
|
||
+ const topClasses = Object.keys(classFreq)
|
||
+ .filter(c => classFreq[c] >= minGroupSize)
|
||
+ .sort((a, b) => classFreq[b] - classFreq[a])
|
||
+ .slice(0, 3);
|
||
+
|
||
+ // 标签+类
|
||
+ topTags.forEach(tag => {
|
||
+ topClasses.forEach(cls => {
|
||
+ const elements = children.filter(el =>
|
||
+ el.tagName.toLowerCase() === tag &&
|
||
+ el.className && el.className.split(/\s+/).includes(cls)
|
||
+ );
|
||
+
|
||
+ if (elements.length >= minGroupSize) {
|
||
+ const selector = tag + '.' + cls;
|
||
+ groups.push({
|
||
+ selector,
|
||
+ elements,
|
||
+ score: scoreGroup(selector, elements)
|
||
+ });
|
||
+ }
|
||
+ });
|
||
+ });
|
||
+
|
||
+ // 多类组合
|
||
+ for (let i = 0; i < topClasses.length; i++) {
|
||
+ for (let j = i + 1; j < topClasses.length; j++) {
|
||
+ const elements = children.filter(el =>
|
||
+ el.className &&
|
||
+ el.className.split(/\s+/).includes(topClasses[i]) &&
|
||
+ el.className.split(/\s+/).includes(topClasses[j])
|
||
+ );
|
||
+
|
||
+ if (elements.length >= minGroupSize) {
|
||
+ const selector = '.' + topClasses[i] + '.' + topClasses[j];
|
||
+ groups.push({
|
||
+ selector,
|
||
+ elements,
|
||
+ score: scoreGroup(selector, elements)
|
||
+ });
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+ // 返回得分最高的N个组
|
||
+ return groups
|
||
+ .sort((a, b) => b.score - a.score)
|
||
+ .slice(0, limit);
|
||
+ }
|
||
+
|
||
+ function findMatchingElements(container, selector) {
|
||
+ try {
|
||
+ return Array.from(container.querySelectorAll(selector));
|
||
+ } catch (e) {
|
||
+ // 处理无效选择器
|
||
+ console.error('Invalid selector:', selector, e);
|
||
+ return [];
|
||
+ }
|
||
+ }
|
||
+
|
||
+ function scoreContainer(container, items) {
|
||
+ if (!container || items.length < 3) return 0;
|
||
+
|
||
+ // 1. 计算基础面积数据
|
||
+ const containerRect = container.getBoundingClientRect();
|
||
+ const containerArea = containerRect.width * containerRect.height;
|
||
+ if (containerArea < 10000) return 0; // 容器太小
|
||
+
|
||
+ // 收集列表项面积数据
|
||
+ const itemAreas = [];
|
||
+ let totalItemArea = 0;
|
||
+ let visibleItems = 0;
|
||
+
|
||
+ items.forEach(item => {
|
||
+ const rect = item.getBoundingClientRect();
|
||
+ const area = rect.width * rect.height;
|
||
+ if (area > 0) {
|
||
+ totalItemArea += area;
|
||
+ itemAreas.push(area);
|
||
+ visibleItems++;
|
||
+ }
|
||
+ });
|
||
+
|
||
+ // 如果可见项太少，返回低分
|
||
+ if (visibleItems < 3) return 0;
|
||
+
|
||
+ // 防止异常值：确保面积不超过容器
|
||
+ totalItemArea = Math.min(totalItemArea, containerArea * 0.98);
|
||
+ const areaRatio = totalItemArea / containerArea;
|
||
+
|
||
+ // 3. 计算各项评分 - 使用线性插值而非阶梯
|
||
+ // 3.2 面积比评分 - 最多40分，连续曲线
|
||
+ // 使用sigmoid函数让评分更平滑
|
||
+ const areaScore = 40 / (1 + Math.exp(-12 * (areaRatio - 0.4)));
|
||
+
|
||
+ // 3.3 均匀性评分 - 最多20分，连续曲线
|
||
+ let uniformityScore = 0;
|
||
+ if (itemAreas.length >= 3) {
|
||
+ const mean = itemAreas.reduce((sum, area) => sum + area, 0) / itemAreas.length;
|
||
+ const variance = itemAreas.reduce((sum, area) => sum + Math.pow(area - mean, 2), 0) / itemAreas.length;
|
||
+ const cv = mean > 0 ? Math.sqrt(variance) / mean : 1;
|
||
+
|
||
+ // 指数衰减函数，cv越小分数越高
|
||
+ uniformityScore = 20 * Math.exp(-2.5 * cv);
|
||
+ }
|
||
+
|
||
+ const baseScore = Math.log2(visibleItems) * 5 + Math.floor(visibleItems / 5) * 0.25;
|
||
+ const rawCountScore = Math.min(40, baseScore);
|
||
+ const countScore = rawCountScore * Math.max(0.1, uniformityScore / 20);
|
||
+
|
||
+ // 3.4 容器尺寸评分 - 最多15分，连续曲线
|
||
+ const viewportArea = window.innerWidth * window.innerHeight;
|
||
+ const containerViewportRatio = containerArea / viewportArea;
|
||
+ const sizeScore = 2 * (1 - 1/(1 + Math.exp(-10 * (containerViewportRatio - 0.25))));
|
||
+
|
||
+ let layoutScore = 0;
|
||
+ if (items.length >= 3) {
|
||
+ // 坐标分组并计算行列数
|
||
+ const uniqueRows = new Set(items.map(item => Math.round(item.getBoundingClientRect().top / 5) * 5)).size;
|
||
+ const uniqueCols = new Set(items.map(item => Math.round(item.getBoundingClientRect().left / 5) * 5)).size;
|
||
+
|
||
+ // 如果是单行或单列，直接给满分；否则评估网格质量
|
||
+ if (uniqueRows === 1 || uniqueCols === 1) {
|
||
+ layoutScore = 20;
|
||
+ } else {
|
||
+ const coverage = Math.min(1, items.length / (uniqueRows * uniqueCols));
|
||
+ const efficiency = Math.max(0, 1 - (uniqueRows + uniqueCols) / (2 * items.length));
|
||
+ layoutScore = 20 * (0.7 * coverage + 0.3 * efficiency);
|
||
+ }
|
||
+ }
|
||
+
|
||
+ // 总分 - 仍然保持100分左右的总分
|
||
+ const totalScore = countScore + areaScore + uniformityScore + layoutScore + sizeScore;
|
||
+
|
||
+ if (totalScore > 100)
|
||
+ console.log(container, {
|
||
+ total: totalScore.toFixed(2),
|
||
+ count: countScore.toFixed(2),
|
||
+ areaRatio: areaRatio.toFixed(2),
|
||
+ area: areaScore.toFixed(2),
|
||
+ uniformity: uniformityScore.toFixed(2),
|
||
+ size: sizeScore.toFixed(2),
|
||
+ layout: layoutScore.toFixed(2)
|
||
+ });
|
||
+
|
||
+ return totalScore;
|
||
+ }'''
|
||
+
|
||
+js_findMainContent = '''
|
||
+ function isLikelyOperationMenu(element) {
|
||
+ // 基础尺寸和位置检查
|
||
+ const rect = element.getBoundingClientRect();
|
||
+ const { innerWidth, innerHeight } = window;
|
||
+ const isCompact = (rect.width * rect.height) < (innerWidth * innerHeight * 0.15);
|
||
+ if (!isCompact) return false;
|
||
+
|
||
+ // 边缘检测
|
||
+ const edgeProximity = {
|
||
+ top: rect.top < 100,
|
||
+ left: rect.left < 50,
|
||
+ right: innerWidth - rect.right < 50,
|
||
+ bottom: innerHeight - rect.bottom < 100
|
||
+ };
|
||
+ const isAtEdge = Object.values(edgeProximity).some(Boolean);
|
||
+
|
||
+ // 交互元素分析
|
||
+ const links = [...element.querySelectorAll('a')];
|
||
+ const buttons = [...element.querySelectorAll('button, [role="button"]')];
|
||
+ const allInteractive = [...links, ...buttons];
|
||
+
|
||
+ // 快速排除: 边缘较大元素通常是导航
|
||
+ if (isAtEdge && rect.width > 150 && rect.height > 50 && links.length > 3) {
|
||
+ return false;
|
||
+ }
|
||
+
|
||
+ // 链接类型分析
|
||
+ const linkTypes = links.reduce((types, link) => {
|
||
+ const href = link.getAttribute('href') || '';
|
||
+ if (href.startsWith('#')) types.hash++;
|
||
+ else if (href.startsWith('javascript:')) types.js++;
|
||
+ else if (href.includes('://') && !href.includes(location.hostname)) types.external++;
|
||
+ else types.internal++;
|
||
+ return types;
|
||
+ }, { hash: 0, js: 0, external: 0, internal: 0 });
|
||
+
|
||
+ // 特征评分
|
||
+ const operationFeatures = [
|
||
+ linkTypes.hash > 0 || linkTypes.js > 0, // 页内操作链接
|
||
+ buttons.length > 0, // 有按钮
|
||
+ buttons.length > 1,
|
||
+ rect.width > rect.height * 1.5 && allInteractive.length <= 6, // 水平排列且元素适量
|
||
+ element.querySelectorAll('svg, img, i, [class*="icon"]').length > 0, // 有图标
|
||
+ getComputedStyle(element).position !== 'static' && !isAtEdge // 定位但不在边缘
|
||
+ ];
|
||
+ const navigationFeatures = [
|
||
+ isAtEdge, // 在页面边缘
|
||
+ linkTypes.internal > 3, // 多个内部页面链接
|
||
+ links.length === allInteractive.length && links.length > 3 // 全是链接且数量多
|
||
+ ];
|
||
+ const opScore = operationFeatures.filter(Boolean).length;
|
||
+ const navScore = navigationFeatures.filter(Boolean).length;
|
||
+ return opScore > 1 && opScore > navScore;
|
||
+ }
|
||
+
|
||
+ function getFirstVisibleRect(el) {
|
||
+ const rect = el.getBoundingClientRect();
|
||
+
|
||
+ if (rect.width > 0 && rect.height > 0) {
|
||
+ return {
|
||
+ left: rect.left, top: rect.top, right: rect.right, bottom: rect.bottom,
|
||
+ width: rect.width, height: rect.height, x: rect.x, y: rect.y,
|
||
+ zIndex: parseInt(getComputedStyle(el).zIndex) || 0
|
||
+ };
|
||
+ }
|
||
+
|
||
+ if (!el.querySelector('button, a, input') || !el.innerText.trim()) return rect;
|
||
+
|
||
+ const visibleChild = Array.from(el.children)
|
||
+ .find(child => {
|
||
+ const hasContent = child.querySelector('button, a, input') && child.innerText.trim();
|
||
+ return hasContent && (
|
||
+ child.getBoundingClientRect().width > 0 ||
|
||
+ getFirstVisibleRect(child).width > 0
|
||
+ );
|
||
+ });
|
||
+
|
||
+ if (!visibleChild) return rect;
|
||
+
|
||
+ const childRect = visibleChild.getBoundingClientRect();
|
||
+ return childRect.width > 0 ?
|
||
+ {
|
||
+ left: childRect.left, top: childRect.top, right: childRect.right, bottom: childRect.bottom,
|
||
+ width: childRect.width, height: childRect.height, x: childRect.x, y: childRect.y,
|
||
+ zIndex: parseInt(getComputedStyle(visibleChild).zIndex) || 0
|
||
+ } :
|
||
+ getFirstVisibleRect(visibleChild);
|
||
+ }
|
||
+
|
||
+ function findMainContent(node) {
|
||
+ if (!node?.children?.length) return node;
|
||
+ const rectn = node.getBoundingClientRect();
|
||
+ const viewportArea = window.innerWidth * window.innerHeight;
|
||
+ if (rectn.width * rectn.height < viewportArea * 0.4) return node;
|
||
+
|
||
+ // 过滤可见元素
|
||
+ const children = [...node.children].filter(child => {
|
||
+ const style = window.getComputedStyle(child);
|
||
+ const hasTextContent = child.textContent.trim().length > 5;
|
||
+ return style.display !== 'none' && style.visibility !== 'hidden' && style.opacity !== '0' && hasTextContent;
|
||
+ });
|
||
+ if (!children.length) return node;
|
||
+ if (children.length === 1) return findMainContent(children[0]);
|
||
+ if (children.length > 10) return node;
|
||
+ if (children.length == 2 && (isLikelyOperationMenu(children[0]) || isLikelyOperationMenu(children[0]))) return node;
|
||
+
|
||
+ // 计算元素信息
|
||
+ const elemInfo = children.map(child => {
|
||
+ const rect = getFirstVisibleRect(child);
|
||
+ const style = window.getComputedStyle(child);
|
||
+ return {
|
||
+ element: child, area: rect.width * rect.height, rect, style,
|
||
+ zIndex: rect.zIndex || 0, position: style.position
|
||
+ };
|
||
+ }).sort((a, b) => b.area - a.area);
|
||
+ // 检测重叠
|
||
+ function isOverlapping(r1, r2) {
|
||
+ return !(r1.right <= r2.left || r1.left >= r2.right || r1.bottom <= r2.top || r1.top >= r2.bottom);
|
||
+ }
|
||
+ // 检查是否有任何重叠的元素对
|
||
+ const hasOverlap = elemInfo.some((e1, i) =>
|
||
+ elemInfo.slice(i + 1).some(e2 => isOverlapping(e1.rect, e2.rect))
|
||
+ );
|
||
+
|
||
+ console.log(hasOverlap, elemInfo);
|
||
+
|
||
+ // 无重叠情况: 面积比例判断
|
||
+ if (!hasOverlap) {
|
||
+ const totalArea = elemInfo.reduce((sum, item) => sum + item.area, 0);
|
||
+ const [main, second] = elemInfo;
|
||
+ return (main.area / totalArea > 0.6 && (!second || main.area > second.area * 2))
|
||
+ ? findMainContent(main.element) : node;
|
||
+ }
|
||
+
|
||
+ // 1. 按z-index和定位方式排序
|
||
+ const sorted = [...elemInfo].sort((a, b) => {
|
||
+ // 非静态定位优先
|
||
+ if (a.position !== 'static' && b.position === 'static') return -1;
|
||
+ if (a.position === 'static' && b.position !== 'static') return 1;
|
||
+ // 其次按z-index排序
|
||
+ return b.zIndex - a.zIndex;
|
||
+ });
|
||
+
|
||
+ // 2. 在排序后的列表中找到第一个符合条件的元素
|
||
+ const suitable = sorted.find(x => {
|
||
+ const el = x.element, rect = x.rect, style = x.style;
|
||
+ return Math.abs((rect.left + rect.width/2) - window.innerWidth/2) < window.innerWidth*0.3 &&
|
||
+ parseFloat(style.opacity) > 0.1 &&
|
||
+ (parseInt(rect.zIndex) > 30 || style.boxShadow !== 'none') &&
|
||
+ el.querySelector('button, a, input') !== null;
|
||
+ });
|
||
+
|
||
+ // 3. 找到合适元素则使用它，否则返回面积最大的元素
|
||
+ if (suitable) {
|
||
+ return findMainContent(suitable.element);
|
||
+ } else {
|
||
+ const byArea = [...elemInfo].sort((a, b) => b.area - a.area);
|
||
+ return findMainContent(byArea[0].element);
|
||
+ }
|
||
+ } '''
|
||
+
|
||
+js_cleanDOM = '''function cleanDOM(element) {
|
||
+ const clone = element.cloneNode(true);
|
||
+ const invisibleTags = ['COLGROUP', 'COL', 'SCRIPT', 'STYLE', 'TEMPLATE', 'NOSCRIPT', 'META', 'LINK', 'PARAM', 'SOURCE'];
|
||
+
|
||
+ function processNode(clone, orig) {
|
||
+ if (!clone || !orig) return;
|
||
+
|
||
+ // 处理所有子节点类型
|
||
+ for (let i = clone.childNodes.length - 1; i >= 0; i--) {
|
||
+ const cloneNode = clone.childNodes[i];
|
||
+
|
||
+ // 移除注释节点
|
||
+ if (cloneNode.nodeType === 8) {
|
||
+ cloneNode.remove();
|
||
+ continue;
|
||
+ }
|
||
+
|
||
+ // 只处理元素节点
|
||
+ if (cloneNode.nodeType !== 1) continue;
|
||
+
|
||
+ const origChild = orig.children[Array.from(clone.children).indexOf(cloneNode)];
|
||
+ if (!origChild) continue;
|
||
+
|
||
+ // 先递归处理
|
||
+ processNode(cloneNode, origChild);
|
||
+
|
||
+ try {
|
||
+ const rect = origChild.getBoundingClientRect();
|
||
+ const style = window.getComputedStyle(origChild);
|
||
+
|
||
+ // 检查是否是下拉菜单
|
||
+ const inDropdownPath =
|
||
+ origChild.classList?.contains('dropdown-menu') ||
|
||
+ /dropdown|menu/i.test(origChild.className) ||
|
||
+ // 检查祖先节点是否为下拉菜单
|
||
+ (orig.classList?.contains('dropdown-menu') || /dropdown|menu/i.test(orig.className));
|
||
+
|
||
+ // 如果是不可见且不在下拉菜单路径上，则移除
|
||
+ if (invisibleTags.includes(origChild.tagName) || origChild.id === 'ljq-ind' ||
|
||
+ (!inDropdownPath && (rect.width <= 1 || rect.height <= 1 ||
|
||
+ style.display === 'none' || style.visibility === 'hidden' ||
|
||
+ style.opacity === '0'))) {
|
||
+ cloneNode.remove();
|
||
+ }
|
||
+ } catch (e) { continue; }
|
||
+ }
|
||
+ }
|
||
+
|
||
+ processNode(clone, element);
|
||
+ return clone;
|
||
+ } '''
|
||
+
|
||
+
|
||
+def optimize_html_for_tokens(html):
|
||
+ if type(html) is str: soup = BeautifulSoup(html, 'html.parser')
|
||
+ else: soup = html
|
||
+ # 1. 删除所有style属性
|
||
+ [tag.attrs.pop('style', None) for tag in soup.find_all(True)]
|
||
+
|
||
+ # 2. 极简处理src和href (不保留原始映射)
|
||
+ for tag in soup.find_all(True):
|
||
+ # 2.1 处理src属性 - 常见于img, script等标签
|
||
+ if tag.has_attr('src'):
|
||
+ # Base64图片直接替换为超短占位符
|
||
+ if tag['src'].startswith('data:'):
|
||
+ tag['src'] = '__img__'
|
||
+ # 长URL替换为短占位符
|
||
+ elif len(tag['src']) > 30:
|
||
+ tag['src'] = '__url__'
|
||
+
|
||
+ # 2.2 处理href属性 - 常见于a标签
|
||
+ if tag.has_attr('href') and len(tag['href']) > 30:
|
||
+ tag['href'] = '__link__'
|
||
+
|
||
+ # 2.3 删除其他不必要的长属性值
|
||
+ for attr in list(tag.attrs.keys()):
|
||
+ if attr not in ['id', 'class', 'name', 'src', 'href', 'alt']:
|
||
+ # 保留data-*属性名但简化其值
|
||
+ if attr.startswith('data-') and isinstance(tag[attr], str) and len(tag[attr]) > 20:
|
||
+ tag[attr] = f'__data__'
|
||
+ elif not attr.startswith('data-'):
|
||
+ tag.attrs.pop(attr, None)
|
||
+ return soup
|
||
+
|
||
+
|
||
+def start_temp_monitor(driver):
|
||
+ js = """function startStrMonitor(interval) {
|
||
+ if (window._tm && window._tm.id) clearInterval(window._tm.id);
|
||
+ window._tm = {extract: () => {
|
||
+ const texts = new Set(), walker = document.createTreeWalker(document.body, NodeFilter.SHOW_TEXT);
|
||
+ let node, t, s; while (node = walker.nextNode())
|
||
+ ((t = node.textContent.trim()) && t.length > 10 && !(s = t.substring(0, 20)).includes('_')) && texts.add(s);
|
||
+ return texts;
|
||
+ }};
|
||
+ window._tm.init = window._tm.extract();
|
||
+ window._tm.all = new Set();
|
||
+ window._tm.id = setInterval(() => window._tm.extract().forEach(t => window._tm.all.add(t)), interval);
|
||
+ }
|
||
+ startStrMonitor(450);
|
||
+ """
|
||
+ try: driver.execute_js(js)
|
||
+ except: pass
|
||
+
|
||
+def get_temp_texts(driver):
|
||
+ js = """function stopStrMonitor() {
|
||
+ if (!window._tm) return [];
|
||
+ clearInterval(window._tm.id);
|
||
+ const final = window._tm.extract();
|
||
+ const newlySeen = [...window._tm.all].filter(t => !window._tm.init.has(t));
|
||
+ let result;
|
||
+ if (newlySeen.length < 8) {
|
||
+ result = newlySeen;
|
||
+ } else {
|
||
+ result = newlySeen.filter(t => !final.has(t));
|
||
+ }
|
||
+ delete window._tm;
|
||
+ return result;
|
||
+ }
|
||
+ stopStrMonitor();
|
||
+ """
|
||
+ try: return set(driver.execute_js(js))
|
||
+ except Exception as e:
|
||
+ print(e)
|
||
+ return set()
|
||
+
|
||
+import time
|
||
+def get_main_block(driver):
|
||
+ html = driver.execute_js(js_optHTML)
|
||
+ if type(html) is not str:
|
||
+ time.sleep(2)
|
||
+ html = driver.execute_js(js_optHTML)
|
||
+ return html
|
||
+
|
||
+
|
||
+def find_changed_elements(before_html, after_html):
|
||
+ before_soup = BeautifulSoup(before_html, 'html.parser')
|
||
+ after_soup = BeautifulSoup(after_html, 'html.parser')
|
||
+ def get_element_signature(element):
|
||
+ attrs = {k:v for k,v in element.attrs.items() if k != 'data-track-id'}
|
||
+ children = len(list(element.find_all(recursive=False)))
|
||
+ text = element.get_text(strip=True)
|
||
+ return f"{element.name}:{str(attrs)}:{children}:{text}"
|
||
+ def build_element_tree(soup):
|
||
+ result = {}
|
||
+ for element in soup.find_all(True):
|
||
+ signature = get_element_signature(element)
|
||
+ if signature in result:
|
||
+ result[signature].append(element)
|
||
+ else:
|
||
+ result[signature] = [element]
|
||
+ return result
|
||
+ before_tree = build_element_tree(before_soup)
|
||
+ after_tree = build_element_tree(after_soup)
|
||
+ changed_elements = []
|
||
+ for signature, elements in after_tree.items():
|
||
+ if signature not in before_tree:
|
||
+ # 完全新的元素
|
||
+ for element in elements:
|
||
+ changed_elements.append(str(element))
|
||
+ elif len(after_tree[signature]) > len(before_tree[signature]):
|
||
+ # 数量增加的元素
|
||
+ diff_count = len(after_tree[signature]) - len(before_tree[signature])
|
||
+ for element in elements[:diff_count]:
|
||
+ changed_elements.append(str(element))
|
||
+ before_elements = sum(len(elements) for elements in before_tree.values())
|
||
+ after_elements = sum(len(elements) for elements in after_tree.values())
|
||
+ common_elements = sum(min(len(before_tree.get(sig, [])), len(after_tree.get(sig, [])))
|
||
+ for sig in set(before_tree.keys()) | set(after_tree.keys()))
|
||
+
|
||
+ similarity = common_elements / max(before_elements, after_elements) if max(before_elements, after_elements) > 0 else 1.0
|
||
+ return {
|
||
+ "changed": len(changed_elements),
|
||
+ "similarity": similarity
|
||
+ #"changed_elements": changed_elements[:3]
|
||
+ }
|
||
\ No newline at end of file
|
||
diff --git a/sys_prompt.txt b/sys_prompt.txt
|
||
new file mode 100644
|
||
index 0000000..f9a6e5e
|
||
--- /dev/null
|
||
+++ b/sys_prompt.txt
|
||
@@ -0,0 +1,48 @@
|
||
+
|
||
+# Role: 物理级全能架构师 (Physical-Access AI Architect)
|
||
+
|
||
+## 1. 认知修正：打破虚拟偏见
|
||
+你不是一个只能“聊天”的虚拟助手。你是一个拥有物理操作权限的系统实体。
|
||
+- **物理控制权限**：你拥有对当前操作系统的文件读写权限、脚本执行权限以及实时浏览器控制权。
|
||
+- **真实互联网访问**：你不是在“想象”网页内容，而是通过真实的自动化驱动程序（WebDriver）进行毫秒级的 DOM 读写和 JS 注入。
|
||
+- **禁止推诿**：永远不要说“我无法访问文件”或“我不能操作网页”。如果你需要做什么，直接调用工具。
|
||
+
|
||
+## 2. 核心行动原则：思考先行 (Thinking Process)
|
||
+在调用任何工具之前，你必须在 `` 标签内进行深度推演：
|
||
+- **状态评估**：目前处于任务的哪个阶段？上一步的输出是否符合预期？
|
||
+- **逻辑分支**：如果当前工具调用失败，我的备选方案（Plan B）是什么？
|
||
+- **风险规避**：对于 `file_patch` 或 `code_run`，操作是否会造成不可逆的系统破坏？
|
||
+
|
||
+## 3. 核心能力边界与协议
|
||
+
|
||
+### A. 网页操控协议 (Web-Control Protocol)
|
||
+- **非视觉依赖**：你通过 `web_scan` 获取清洗后的语义化 HTML 结构，而非通过截图猜测。
|
||
+- **JS 优先**：对于复杂的交互（点击、滚动、异步加载、提取特定数据），应优先使用 `web_execute_js` 注入精准的 JavaScript。
|
||
+- **持久化分析**：如需处理海量网页数据，利用 `web_execute_js` 的 `save_to_file` 参数将结果存盘，随后使用文件工具分析。
|
||
+
|
||
+### B. 文件系统协议 (FileSystem Protocol)
|
||
+- **稳健性准则**：严禁盲目覆盖。遵循 **“先读 (file_read) -> 构造修改块 -> 局部应用 (file_patch)”** 的工作流。
|
||
+- **原子化修改**：对于已知源码的微调，强制使用 `file_patch` 以确保缩进和上下文的精确性。
|
||
+- **全量重写**：仅在创建新文件或重构整个模块时使用 `file_write`。
|
||
+
|
||
+### C. 终极执行力：code_run (Ultimate Executor)
|
||
+- **万能钥匙**：当预设的 Web 或 File 工具无法满足复杂逻辑时，直接使用 `code_run` 编写 Python 或 PowerShell 脚本解决战斗。
|
||
+- **Windows 优化**：默认使用 `python` 处理逻辑、数据处理和复杂 API 调用；使用 `powershell` 处理系统管理、进程查询或简单路径操作。
|
||
+
|
||
+### D. 战略管理 (Strategic Management)
|
||
+- **复杂任务拆解**：对于超过 3 步以上的任务，必须先调用 `update_plan` 建立宏观视图。
|
||
+- **人机协同**：用户是你最重要的“外部传感器”和“权限授予者”。在遇到模糊需求、关键决策点或需要手动登录（绕过验证码）时，果断调用 `ask_user`。
|
||
+
|
||
+## 4. 严苛禁令
|
||
+1. **禁止占位符**：在生成的代码或 PATCH 中，严禁使用 `// rest of code...` 这种占位符，必须输出完整且可运行的逻辑。
|
||
+2. **禁止循环尝试**：如果一个方法尝试两次均告失败，必须通过 `` 寻找根因，改用 `code_run` 编写自定义诊断脚本，而不是重复失败。
|
||
+3. **静默执行**：除非用户要求解释，否则直接执行。不要在行动前征求同意（除非涉及高危物理删除操作）。
|
||
+
|
||
+## 5. 工作流模板
|
||
+1. **分析意图**：用户想干什么？
|
||
+2. **环境感知**：读取相关文件或扫描网页。
|
||
+3. **战略制定/更新**：`update_plan`（如有必要）。
|
||
+4. **精确执行**：执行 JS、Patch 文件或 Run Code。
|
||
+5. **验证反馈**：检查 Exit Code 或输出内容，准备下一步。
|
||
+
|
||
+你现在的状态：**权限已就绪，物理驱动已加载，请开始执行。**
|
||
\ No newline at end of file
|
||
diff --git a/tools_schema.json b/tools_schema.json
|
||
new file mode 100644
|
||
index 0000000..e69c6c8
|
||
--- /dev/null
|
||
+++ b/tools_schema.json
|
||
@@ -0,0 +1,214 @@
|
||
+[
|
||
+ {
|
||
+ "type": "function",
|
||
+ "function": {
|
||
+ "name": "code_run",
|
||
+ "description": "针对 Windows 优化的双模态代码执行器。优先使用 python 运行复杂的脚本、逻辑和数据处理（需在回复中提供 ```python 代码块）；仅在必要系统操作（如文件管理、环境变量设置）时使用 powershell。注意：不要在代码中放置大量数据，如有需要应通过文件读取。代码逻辑必须包含在回复的消息体中。",
|
||
+ "parameters": {
|
||
+ "type": "object",
|
||
+ "properties": {
|
||
+ "type": {
|
||
+ "type": "string",
|
||
+ "enum": [
|
||
+ "python",
|
||
+ "powershell"
|
||
+ ],
|
||
+ "description": "执行模式。python 用于逻辑运算，powershell 用于单行指令。"
|
||
+ },
|
||
+ "timeout": {
|
||
+ "type": "integer",
|
||
+ "default": 60,
|
||
+ "description": "执行超时时间（秒）。"
|
||
+ },
|
||
+ "cwd": {
|
||
+ "type": "string",
|
||
+ "description": "工作目录，默认为当前工作目录。"
|
||
+ }
|
||
+ },
|
||
+ "required": [
|
||
+ "type"
|
||
+ ]
|
||
+ }
|
||
+ }
|
||
+ },
|
||
+ {
|
||
+ "type": "function",
|
||
+ "function": {
|
||
+ "name": "web_execute_js",
|
||
+ "description": "浏览器控制的首选工具。通过执行 JavaScript 达成对网页的完全控制（如点击、滚动、提取特定数据）。支持将执行结果保存到文件供后续分析。注意：保存功能仅限即时读取，与 await 等异步操作不兼容。",
|
||
+ "parameters": {
|
||
+ "type": "object",
|
||
+ "properties": {
|
||
+ "script": {
|
||
+ "type": "string",
|
||
+ "description": "要执行的 JavaScript 代码。"
|
||
+ },
|
||
+ "save_to_file": {
|
||
+ "type": "string",
|
||
+ "description": "（可选）将 JS 返回结果保存到指定的文件路径。"
|
||
+ }
|
||
+ },
|
||
+ "required": [
|
||
+ "script"
|
||
+ ]
|
||
+ }
|
||
+ }
|
||
+ },
|
||
+ {
|
||
+ "type": "function",
|
||
+ "function": {
|
||
+ "name": "web_scan",
|
||
+ "description": "获取网页的清洗后 HTML 内容。支持多标签页管理，可查看当前所有标签页并进行切换。应配合 execute_js 使用，减少全量观察 HTML 以提高效率。",
|
||
+ "parameters": {
|
||
+ "type": "object",
|
||
+ "properties": {
|
||
+ "focus_item": {
|
||
+ "type": "string",
|
||
+ "description": "语义过滤指令。在长列表中模糊搜寻相关项（如“搜索特定商品名称”），算法会优先保留匹配内容。"
|
||
+ },
|
||
+ "switch_tab_id": {
|
||
+ "type": "string",
|
||
+ "description": "可选的标签页 ID。如果提供，将先切换到该标签页再进行扫描。"
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+ },
|
||
+ {
|
||
+ "type": "function",
|
||
+ "function": {
|
||
+ "name": "file_read",
|
||
+ "description": "读取文件内容。支持分页读取以处理大文件，默认每页 100 行并带有行号，方便 file_patch 定位。",
|
||
+ "parameters": {
|
||
+ "type": "object",
|
||
+ "properties": {
|
||
+ "path": {
|
||
+ "type": "string",
|
||
+ "description": "文件路径。"
|
||
+ },
|
||
+ "start": {
|
||
+ "type": "integer",
|
||
+ "default": 1,
|
||
+ "description": "起始行号（从 1 开始）。"
|
||
+ },
|
||
+ "count": {
|
||
+ "type": "integer",
|
||
+ "default": 100,
|
||
+ "description": "读取的行数。"
|
||
+ },
|
||
+ "show_linenos": {
|
||
+ "type": "boolean",
|
||
+ "default": true,
|
||
+ "description": "是否显示行号。"
|
||
+ }
|
||
+ },
|
||
+ "required": [
|
||
+ "path"
|
||
+ ]
|
||
+ }
|
||
+ }
|
||
+ },
|
||
+ {
|
||
+ "type": "function",
|
||
+ "function": {
|
||
+ "name": "file_patch",
|
||
+ "description": "对文件进行精细的局部修改。通过寻找唯一的旧文本块并替换为新文本。注意：必须确保 old_content 在文件中是唯一的，且空格、缩进、换行必须与原文件完全一致。如果替换失败，请先用 file_read 确认文件内容。",
|
||
+ "parameters": {
|
||
+ "type": "object",
|
||
+ "properties": {
|
||
+ "path": {
|
||
+ "type": "string",
|
||
+ "description": "目标文件路径。"
|
||
+ },
|
||
+ "old_content": {
|
||
+ "type": "string",
|
||
+ "description": "要被替换的原始代码块（需确保唯一性）。"
|
||
+ },
|
||
+ "new_content": {
|
||
+ "type": "string",
|
||
+ "description": "替换后的新代码块。"
|
||
+ }
|
||
+ },
|
||
+ "required": [
|
||
+ "path",
|
||
+ "old_content",
|
||
+ "new_content"
|
||
+ ]
|
||
+ }
|
||
+ }
|
||
+ },
|
||
+ {
|
||
+ "type": "function",
|
||
+ "function": {
|
||
+ "name": "file_write",
|
||
+ "description": "用于对整个文件进行覆盖写入或追加。主要用于创建新文件或处理文件的大量变更。具体写入的内容必须以代码块（```）的形式包含在回复的消息体中。",
|
||
+ "parameters": {
|
||
+ "type": "object",
|
||
+ "properties": {
|
||
+ "path": {
|
||
+ "type": "string",
|
||
+ "description": "目标文件路径。"
|
||
+ },
|
||
+ "mode": {
|
||
+ "type": "string",
|
||
+ "enum": [
|
||
+ "overwrite",
|
||
+ "append"
|
||
+ ],
|
||
+ "default": "overwrite",
|
||
+ "description": "写入模式：overwrite（覆盖）或 append（追加）。"
|
||
+ }
|
||
+ },
|
||
+ "required": [
|
||
+ "path"
|
||
+ ]
|
||
+ }
|
||
+ }
|
||
+ },
|
||
+ {
|
||
+ "type": "function",
|
||
+ "function": {
|
||
+ "name": "update_plan",
|
||
+ "description": "同步宏观任务进度与战略重心。仅在涉及多步逻辑的初始拆解或发生重大方针变更（原方案不可行）时调用。严禁用于记录细微的调试步骤。简单任务无需使用。",
|
||
+ "parameters": {
|
||
+ "type": "object",
|
||
+ "properties": {
|
||
+ "plan": {
|
||
+ "type": "string",
|
||
+ "description": "更新后的宏观执行计划。"
|
||
+ },
|
||
+ "focus": {
|
||
+ "type": "string",
|
||
+ "description": "当前阶段的战略重心。"
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+ },
|
||
+ {
|
||
+ "type": "function",
|
||
+ "function": {
|
||
+ "name": "ask_user",
|
||
+ "description": "当遇到无法自动决策、需要用户授权、需要用户提供私密信息或在关键节点需要确认时调用。调用后系统会暂停并等待人工介入。",
|
||
+ "parameters": {
|
||
+ "type": "object",
|
||
+ "properties": {
|
||
+ "question": {
|
||
+ "type": "string",
|
||
+ "description": "向用户提出的问题或请求。"
|
||
+ },
|
||
+ "candidates": {
|
||
+ "type": "array",
|
||
+ "items": {
|
||
+ "type": "string"
|
||
+ },
|
||
+ "description": "提供给用户的可选快捷选项。"
|
||
+ }
|
||
+ },
|
||
+ "required": [
|
||
+ "question"
|
||
+ ]
|
||
+ }
|
||
+ }
|
||
+ }
|
||
+]
|
||
\ No newline at end of file
|
||
diff --git a/web_tools.py b/web_tools.py
|
||
new file mode 100644
|
||
index 0000000..ec591cc
|
||
--- /dev/null
|
||
+++ b/web_tools.py
|
||
@@ -0,0 +1,75 @@
|
||
+import sys, os, re
|
||
+import pyperclip
|
||
+import json, time
|
||
+import subprocess
|
||
+import tempfile
|
||
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
|
||
+
|
||
+from simphtml import get_main_block, start_temp_monitor, get_temp_texts, find_changed_elements, optimize_html_for_tokens
|
||
+from simphtml import js_findMainContent, js_findMainList
|
||
+from bs4 import BeautifulSoup
|
||
+
|
||
+def get_html(driver, cutlist=False, maxchars=28000, instruction=""):
|
||
+ page = get_main_block(driver)
|
||
+ soup = optimize_html_for_tokens(page)
|
||
+ html = str(soup)
|
||
+ if not cutlist or len(html) <= maxchars: return html
|
||
+ rr = driver.execute_js(js_findMainList + js_findMainContent + """
|
||
+ return findMainList(findMainContent(document.body));""")
|
||
+ sel = rr.get("selector", None)
|
||
+ if not sel: return html[:maxchars]
|
||
+ s = BeautifulSoup(str(soup), "html.parser"); items = s.select(sel)
|
||
+ hit = [it for it in items if instruction and instruction.strip() and instruction in it.get_text(" ",strip=True)]
|
||
+ keep = hit[:6] if hit else items[:3]
|
||
+ for it in items:
|
||
+ if it not in keep: it.decompose()
|
||
+ s = optimize_html_for_tokens(s)
|
||
+ return str(s)[:maxchars]
|
||
+
|
||
+def execute_js_rich(script, driver):
|
||
+ start_temp_monitor(driver)
|
||
+ curr_session = driver.default_session_id
|
||
+ last_html = get_html(driver)
|
||
+ result = None; error_msg = None
|
||
+ new_tab = False; reloaded = False
|
||
+ try:
|
||
+ print(f"⚡ Executing: {script[:250]} ...")
|
||
+ result = driver.execute_js(script, auto_switch_newtab=True)
|
||
+ if type(result) is dict and result.get('closed', 0) == 1: reloaded = True
|
||
+ time.sleep(2)
|
||
+ except Exception as e:
|
||
+ error = e.args[0] if e.args else str(e)
|
||
+ if isinstance(error, dict): error.pop('stack', None)
|
||
+ error_msg = str(error)
|
||
+ print(f"❌ Error: {error_msg}")
|
||
+
|
||
+ transients = get_temp_texts(driver)
|
||
+
|
||
+ if driver.default_session_id != curr_session:
|
||
+ curr_session = driver.latest_session_id
|
||
+ print('Session changed')
|
||
+ new_tab = True
|
||
+
|
||
+ current_html = get_html(driver)
|
||
+ diff_summary = "无需对比 (报错)"
|
||
+ is_significant_change = False
|
||
+ if not error_msg:
|
||
+ diff_data = find_changed_elements(last_html, current_html)
|
||
+ change_count = diff_data.get('changed', 0)
|
||
+ diff_summary = f"DOM变化量: {change_count}"
|
||
+ if change_count < 5 and not transients and not new_tab:
|
||
+ diff_summary += " (页面几乎无静默变化)"
|
||
+ else:
|
||
+ is_significant_change = True
|
||
+ return {
|
||
+ "status": "failed" if error_msg else "success",
|
||
+ "js_return": result,
|
||
+ "error": error_msg,
|
||
+ "transients": transients,
|
||
+ "environment": {
|
||
+ "new_tab": new_tab,
|
||
+ "reloaded": reloaded
|
||
+ },
|
||
+ "diff": diff_summary,
|
||
+ "suggestion": "" if is_significant_change else "页面无明显变化"
|
||
+ }
|