This commit is contained in:
HuangHai
2026-01-28 16:34:52 +08:00
parent d2fc0cd7c7
commit 4d8a4801b8
24 changed files with 202 additions and 535 deletions

View File

@@ -1,333 +0,0 @@
# coding=utf-8
import os
import sys
import logging
import asyncio
import hashlib
import json
import numpy as np
import cv2
# 添加项目根目录到 sys.path
project_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if project_root not in sys.path:
sys.path.append(project_root)
from WeiXin import WxUtil
from WeiXin.WxUtil import perform_input_action
from Util.LlmUtil import get_llm_response
from Util import Win32Patch
# 配置日志
log_dir = WxUtil.LOG_DIR
if not os.path.exists(log_dir):
os.makedirs(log_dir)
log_file_path = os.path.join(log_dir, "T2_ChatMonitor.log")
# 设置 logger
logger = logging.getLogger("T2_ChatMonitor")
logger.setLevel(logging.INFO)
if logger.hasHandlers():
logger.handlers.clear()
file_handler = logging.FileHandler(log_file_path, encoding='utf-8', mode='w')
file_handler.setFormatter(logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s'))
logger.addHandler(file_handler)
stream_handler = logging.StreamHandler()
stream_handler.setFormatter(logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s'))
logger.addHandler(stream_handler)
logger.propagate = False
logger.info(f"🚀 日志文件路径: {os.path.abspath(log_file_path)}")
# 同时将 WxUtil 的日志也输出到同一个文件
wx_logger = logging.getLogger("WxUtil")
wx_logger.propagate = False # 防止日志向上传递导致重复 (因为 WxUtil 中调用了 basicConfig)
if not any(isinstance(h, logging.FileHandler) and os.path.abspath(h.baseFilename) == os.path.abspath(log_file_path) for h in wx_logger.handlers):
wx_logger.addHandler(file_handler)
wx_logger.addHandler(stream_handler) # 确保 WxUtil 也输出到控制台
class ChatMonitorBot:
"""
大张老师自动巡课系统 (CV版)
"""
def __init__(self):
self.device = None
self.screenshot_path = os.path.join(WxUtil.OUTPUT_DIR, "T2_ChatMonitor_live_shot.jpg")
self.debug_view_path = os.path.join(WxUtil.OUTPUT_DIR, "T2_ChatMonitor_debug_view.jpg")
self.dialogue_log = []
self.input_pos = None
self.last_screen_hash = None
self.last_processed_msg_hash = None
# [User Requested] 移除持久化存储,只在内存中记录,重启即忘
self.processed_hashes = set()
self.check_interval = 3 # 检查频率 (秒)
self.persona = (
"你是一名1999年毕业、拥有27年一线教学经验的小学高级女教师,名叫‘大张老师’。你目前在‘长春市少惠林作文素养培养中心’工作。"
"你不仅是一位作文教学专家,更是一位心思细腻、能与家长共情的教育智者。"
"你的回复风格应该是:温柔、知性、亲切,就像一位邻家大姐姐在聊天。"
"【严格约束】:\n"
"1. 绝对禁止发散!绝对禁止幻觉!\n"
"2. 知道什么就说什么,不要乱讲话,不要自己编造内容!\n"
"3. 仅针对家长明确表达的内容进行回复。\n"
"4. 严禁使用列表格式。严禁使用‘首先、其次’等逻辑词。\n"
"5. 回复必须简练,字数严格控制在 50 字以内!\n"
"6. 对方问什么就答什么。例如问‘学校叫什么’,就只回答‘少惠林’,不要回复地址和电话!\n"
"如果涉及到校区信息,必须且只能使用以下真实数据:\n"
"- 单位/学校名称:长春市少惠林作文素养培养中心(简称:少惠林)\n"
"- 地址:南环城路与临河街交汇,TOUCH12街3楼325号\n"
"- 联系人:小张老师(电话:18686619970)\n"
"- 每学期开学招收小学三年级至六年级,初中七年级的学生入学,其它年段不招生。\n"
)
def _record_processed_hash(self, msg_hash):
"""记录已处理的消息哈希 (仅内存)"""
self.processed_hashes.add(msg_hash)
# 仅保留最近 100 条记录,防止无限增长
if len(self.processed_hashes) > 100:
# 简单丢弃旧的(转列表切片再转回集合)
temp = list(self.processed_hashes)[-100:]
self.processed_hashes = set(temp)
async def get_reply(self, last_message_text, context_text=""):
prompt = (
f"【教师人设】:{self.persona}\n\n"
f"【上下文对话内容】:\n{context_text}\n\n"
f"【最后一条待回复消息】:\n{last_message_text}\n\n"
"【任务要求】:\n"
"请作为大张老师回复家长。**必须且只能针对最后一条消息进行回复!**\n"
"参考上下文对话内容,确保回复逻辑连贯。\n"
"严禁发散,严禁编造家长没说过的情况。如果不清楚家长的意图,就温柔询问。\n"
"字数严格控制在 50 字以内。直接输出回复正文。"
)
full_response = ""
async for chunk in get_llm_response(prompt, stream=False):
full_response += chunk
return full_response.strip().strip('"').strip('“').strip('”')
def step_1_prepare_env(self):
"""步骤1: 环境准备"""
logger.info("--- [Step 1] 环境准备 ---")
WxUtil.setup_script_environment()
return True
def step_2_connect_device(self):
"""步骤2: 连接设备"""
logger.info("--- [Step 2] 连接设备 ---")
self.device = WxUtil.connect_device()
if not self.device:
logger.error("❌ 设备连接失败,请检查手机是否连接且开启了调试模式")
return False
return True
def get_image_hash(self, file_path):
"""计算图片的 MD5 哈希值 (忽略顶部 100 像素的状态栏)"""
if not os.path.exists(file_path):
return None
try:
# 使用 OpenCV 读取图片
img = cv2.imread(file_path)
if img is None:
# 如果读取失败,回退到文件哈希
with open(file_path, "rb") as f:
return hashlib.md5(f.read()).hexdigest()
# 裁剪掉顶部 150 像素 (状态栏/时间)
h, w = img.shape[:2]
if h > 150:
cropped_img = img[150:h, 0:w]
else:
cropped_img = img
# 计算裁剪后数据的哈希
return hashlib.md5(cropped_img.tobytes()).hexdigest()
except Exception as e:
logger.error(f"计算哈希出错: {e}, 回退到文件哈希")
with open(file_path, "rb") as f:
return hashlib.md5(f.read()).hexdigest()
def get_stable_message_hash(self, msg):
"""
计算消息的稳定哈希值(忽略坐标等易变字段)
仅包含: sender, content, time_display, type
"""
if not msg:
return ""
stable_data = {
"sender": msg.get("sender", ""),
"content": msg.get("content") or "", # 确保 None 转为空字符串
"time_display": msg.get("time_display", ""),
"type": msg.get("type", "")
}
# 序列化并计算哈希
msg_str = json.dumps(stable_data, sort_keys=True, ensure_ascii=False)
return hashlib.md5(msg_str.encode('utf-8')).hexdigest()
async def run(self):
"""
主运行循环
"""
logger.info("🚀 正在启动 T2_ChatMonitor (Auto-Reply)...")
# 定义 JSON 序列化辅助函数
def numpy_serializer(obj):
if isinstance(obj, np.integer):
return int(obj)
if isinstance(obj, np.floating):
return float(obj)
if isinstance(obj, np.ndarray):
return obj.tolist()
raise TypeError(f"Type {type(obj)} not serializable")
# 1. 环境准备
if not self.step_1_prepare_env(): return
if not self.step_2_connect_device(): return
# [User Requested] 移除首屏概念,直接进入监控循环
# 以前说过什么都不管了,只关注最后一条
logger.info("🚀 启动完成,直接进入实时监控阶段...")
# 3. 进入循环阶段
while True:
try:
# A. 截图并计算哈希
self.device.screenshot(self.screenshot_path)
current_screen_hash = self.get_image_hash(self.screenshot_path)
# B. 如果屏幕无变化,则跳过识别
if current_screen_hash == self.last_screen_hash:
await asyncio.sleep(self.check_interval)
continue
self.last_screen_hash = current_screen_hash
logger.info("📸 屏幕发生变化,正在分析...")
# C. 分析最新图片
dialogue_log, input_pos = await WxUtil.analyze_chat_image(
self.screenshot_path,
self.debug_view_path,
device=self.device,
process_strategy="UNREAD" # 监控阶段:只处理带红点的新语音
)
if not dialogue_log:
logger.info("😴 未识别到有效消息")
await asyncio.sleep(self.check_interval)
continue
logger.info(f"📊 当前识别到 {len(dialogue_log)} 条消息,最后一条: {dialogue_log[-1]}")
# 更新当前对话日志(可用于上下文参考)
self.dialogue_log = dialogue_log
self.input_pos = input_pos
# D. 只关注最后一条消息
last_msg = dialogue_log[-1]
# 计算稳定哈希(忽略坐标变化)
current_msg_hash = self.get_stable_message_hash(last_msg)
# E. 判断是否需要回复 (对方发送且非重复消息)
sender = last_msg.get('sender', '')
# Check if hash is already processed (in-memory only)
is_processed = current_msg_hash in self.processed_hashes
# Log only if it changed from last *in-memory* check to avoid spam
if is_processed and current_msg_hash != self.last_processed_msg_hash:
# logger.info(f"🚫 [监控] 消息哈希已存在于历史记录中,跳过回复 (Hash: {current_msg_hash})")
self.last_processed_msg_hash = current_msg_hash
if not is_processed and current_msg_hash != self.last_processed_msg_hash:
if sender != "我":
event_shot = WxUtil.get_next_debug_path("event_new_msg")
self.device.screenshot(event_shot)
logger.info(f"💡 [监控] 发现新消息: {last_msg},保存现场截图: {event_shot}")
# 获取上下文文本 (格式化为 Sender: Content)
context_text = "\n".join([f"{m.get('time_display', '') + ' ' if m.get('time_display') else ''}{m.get('sender')}: {m.get('content')}" for m in dialogue_log[:-1]])
last_content = last_msg.get('content') or ""
# 兜底逻辑:如果最后一条是语音且内容为空(可能因无红点未被 UNREAD 策略处理),尝试强制转换
if last_msg.get('type') == 'voice' and not last_content.strip():
logger.info("⚠️ [监控] 最后一条语音消息未获取到内容(可能已读无红点),尝试强制转换...")
# 强制使用 LAST 策略重试
dialogue_log_retry, _ = await WxUtil.analyze_chat_image(
self.screenshot_path,
self.debug_view_path,
device=self.device,
process_strategy="LAST"
)
if dialogue_log_retry:
# 更新引用
self.dialogue_log = dialogue_log_retry
dialogue_log = dialogue_log_retry
last_msg = dialogue_log[-1]
last_content = last_msg.get('content') or ""
logger.info(f"🔄 [重试] 强制转换后内容: {last_content}")
# 重新构建哈希
current_msg_hash = self.get_stable_message_hash(last_msg)
# 再次检查是否已处理 (因为内容变了,哈希变了)
if current_msg_hash in self.processed_hashes:
logger.info(f"🚫 [重试] 转换后发现该消息已处理,跳过。")
self.last_processed_msg_hash = current_msg_hash
# 跳过本次循环的剩余部分
continue
# 生成回复
reply = await self.get_reply(last_content, context_text)
if reply:
logger.info(f"🤖 [监控] LLM 建议回复: {reply}")
if self.input_pos:
logger.info(f"⚡ [监控] 执行自动回复...")
# input_pos 是 ((x,y), box) 格式,取第一个元素坐标点
target_pos = self.input_pos[0] if isinstance(self.input_pos, (list, tuple)) and len(self.input_pos) == 2 and isinstance(self.input_pos[0], (list, tuple)) else self.input_pos
# 简单兼容处理:如果 input_pos[0] 是 tuple/list 且 input_pos[1] 是 None/box,则取 input_pos[0]
if isinstance(self.input_pos, (list, tuple)) and len(self.input_pos) == 2 and isinstance(self.input_pos[0], (list, tuple)):
target_pos = self.input_pos[0]
perform_input_action(self.device, target_pos, reply)
# 发送后截图留存
reply_sent_shot = WxUtil.get_next_debug_path("event_reply_sent")
self.device.screenshot(reply_sent_shot)
logger.info(f"✅ [监控] 回复已发送,保存发送后截图: {reply_sent_shot}")
self._record_processed_hash(current_msg_hash)
self.last_processed_msg_hash = current_msg_hash
else:
logger.error("❌ 未找到输入框位置,无法发送回复")
else:
logger.info("⚪ [监控] LLM 认为无需回复")
self._record_processed_hash(current_msg_hash)
self.last_processed_msg_hash = current_msg_hash
else:
# 是我发的消息,更新哈希,不再处理
self.last_processed_msg_hash = current_msg_hash
await asyncio.sleep(self.check_interval)
except Exception as e:
logger.error(f"Error in monitoring loop: {e}", exc_info=True)
await asyncio.sleep(self.check_interval)
async def run_main():
"""
运行自动巡课机器人
"""
bot = ChatMonitorBot()
await bot.run()
if __name__ == "__main__":
# 应用 Win32 补丁
Win32Patch.patch()
asyncio.run(run_main())

View File

@@ -40,9 +40,16 @@ def parse_wechat_time(time_str):
支持: "10:03", "昨天 10:03", "星期三 10:03", "2025年1月1日 10:03"
"""
try:
clean_str = time_str.strip()
# 0. 预处理:过滤纯数字(防止电话号码被误识别为时间)
# 微信时间戳通常包含中文或冒号,单纯的数字串(如 "18686619970")不是有效时间
if re.match(r'^\d+$', clean_str):
logger.warning(f"忽略疑似电话号码/纯数字的时间字符串: '{clean_str}'")
return ""
now = datetime.now()
today = now.date()
clean_str = time_str.strip()
# 1. HH:mm (当天)
# 注意:有时候 OCR 会把冒号识别成其他字符,这里假设是标准的 HH:mm
@@ -130,10 +137,11 @@ def parse_wechat_time(time_str):
dt = datetime.combine(today, datetime.min.time().replace(hour=h, minute=m))
return dt.strftime("%Y-%m-%d %H:%M")
return clean_str # 解析失败,返回原串
# 解析失败,返回空字符串,避免将无关文本(如电话号码)误认为时间注入到上下文中
return ""
except Exception as e:
logger.warning(f"时间解析失败 '{time_str}': {e}")
return time_str
return ""
def get_next_debug_path(desc="step"):
"""获取下一个顺序命名的调试图片路径 (debug_N_desc.jpg)"""
@@ -305,6 +313,11 @@ def _detect_bubble_color(img, bbox):
# 且亮度不能太低 (太低可能是黑色背景或深色物体)
if g > 150:
return "white"
# 特殊补丁:如果 B, G, R 都很接近且在 130 左右,可能是微信的背景灰 (通常用于时间戳或系统消息)
if 110 < r < 160 and 110 < g < 160 and 110 < b < 160:
if abs(r - g) < 15 and abs(g - b) < 15:
return "system_gray"
return "unknown"
@@ -359,7 +372,7 @@ def _scan_chat_messages(image_path):
# 微信菜单关键字(用于排除干扰)
MENU_KEYWORDS = ["听筒播放", "收藏", "背景播放", "删除", "多选", "取消转文字", "转文字", "引用", "提醒"]
# 忽略的系统消息内容
IGNORE_CONTENT = ["撤回了一条消息", "打招呼的消息", "拍了拍", "你撤回了一条消息", "引用", "Clear Text", "Switch IME", "Done"]
IGNORE_CONTENT = ["撤回了一条消息", "打招呼的消息", "拍了拍", "你撤回了一条消息", "引用", "Clear Text", "Switch IME", "Done", "按住说话", "发送"]
# 5. 整合所有消息
messages = []
@@ -367,9 +380,9 @@ def _scan_chat_messages(image_path):
# 绘制过滤区域边界 (可视化)
cv2.line(debug_img, (0, 150), (w, 150), (255, 0, 255), 2) # 顶部线
cv2.line(debug_img, (0, h - 100), (w, h - 100), (255, 0, 255), 2) # 底部线
cv2.line(debug_img, (0, h - 60), (w, h - 60), (255, 0, 255), 2) # 底部线 (放宽到底部 60px)
cv2.putText(debug_img, "TOP_FILTER", (10, 140), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 0, 255), 1)
cv2.putText(debug_img, "BOTTOM_FILTER", (10, h - 110), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 0, 255), 1)
cv2.putText(debug_img, "BOTTOM_FILTER", (10, h - 70), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 0, 255), 1)
claimed_ocr_indices = set()
@@ -379,7 +392,7 @@ def _scan_chat_messages(image_path):
cv2.circle(debug_img, (ax, ay), 10, (255, 255, 0), -1)
# 过滤掉顶部和底部的非聊天区域
if ay < 150 or ay > h - 100:
if ay < 150 or ay > h - 60:
logger.info(f"忽略区域外语音图标: ({ax}, {ay})")
cv2.rectangle(debug_img, (ax-35, ay-35), (ax+35, ay+35), (128, 128, 128), 1)
cv2.putText(debug_img, "FILTERED", (ax - 40, ay - 45), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (128, 128, 128), 1)
@@ -416,25 +429,32 @@ def _scan_chat_messages(image_path):
if voice_is_left:
# 语音在左 (对方): 文本必须也是左对齐
# - min_x 必须靠左 (< 300)
# - min_x 必须靠左 (< 450) [Fix] 放宽阈值,防止长文本或缩进文本被过滤
# - max_x 不能太靠右 (> w - 150),否则可能是"我"的消息
if min_x > 300 or max_x > w - 150:
if min_x > 450 or max_x > w - 150:
logger.debug(f"忽略文本 '{text[:10]}' (Left Voice): min_x={min_x}, max_x={max_x} 不满足左对齐条件")
continue
else:
# 语音在右 (我): 文本必须也是右对齐
# - max_x 必须靠右 (> w - 300)
# - min_x 不能太靠左 (< 100)
if max_x < w - 300 or min_x < 100:
logger.debug(f"忽略文本 '{text[:10]}' (Right Voice): min_x={min_x}, max_x={max_x} 不满足右对齐条件")
continue
if -50 < c_y - ay < 800 and abs(c_x - ax) < 500:
# [Fix] 放宽 X 轴判定范围 (500 -> 600) 以适应更宽的文本
if -50 < c_y - ay < 800 and abs(c_x - ax) < 600:
# 检查中间是否有其他语音图标
has_intermediate_audio = False
for other_ax, other_ay in audio_matches:
# 只有当中间的语音图标在 [150, h-60] 的有效聊天区域内时,才视为阻断
if ay + 20 < other_ay < c_y - 10:
has_intermediate_audio = True
logger.info(f"语音({ax},{ay}) 被中间语音图标({other_ax},{other_ay}) 阻断,无法关联文本 '{text[:10]}...'")
break
if 150 <= other_ay <= h - 60:
has_intermediate_audio = True
logger.info(f"语音({ax},{ay}) 被中间有效区域内的语音图标({other_ax},{other_ay}) 阻断,无法关联文本 '{text[:10]}...'")
break
else:
logger.info(f"语音({ax},{ay}) 忽略非聊天区域(Y={other_ay})的语音图标阻断")
if has_intermediate_audio:
continue
@@ -532,7 +552,7 @@ def _scan_chat_messages(image_path):
c_x = int((bbox[0][0] + bbox[2][0]) / 2)
c_y = int((bbox[0][1] + bbox[2][1]) / 2)
if c_y < 150 or c_y > h - 100:
if c_y < 150 or c_y > h - 60:
continue
# 判定发送者 (增强版: 几何 + 颜色)
@@ -544,6 +564,9 @@ def _scan_chat_messages(image_path):
sender = "我"
elif sender_color == "white":
sender = "对方"
elif sender_color == "system_gray":
# 灰底文字通常是时间戳或系统消息,由系统发送,几何上居中
sender = "system"
# 2. 几何特征强制修正 (Double Check)
# 假设头像+边距约占 15% 宽度
@@ -561,12 +584,15 @@ def _scan_chat_messages(image_path):
# 规则 B: 如果这一行极其靠左 (小于 35% 宽度),且不靠右,那肯定是"对方"
# 扩大判定范围,防止因为 OCR 稍微缩进导致判定失效
# 注意:如果颜色明确为"我"(绿色),则跳过此规则,因为"我"的长消息也可能靠左
elif min_x < w * 0.35 and max_x < w * 0.85:
elif min_x < w * 0.35 and max_x < w * 0.75: # 修正:max_x 阈值从 0.85 降低到 0.75
if sender == "我":
logger.info(f"Geometry says '对方' (min_x={min_x} < {w*0.35}) but Color is '我' (Green). Trusting Color.")
elif sender == "system":
# 即使颜色是系统灰,但如果位置极其靠左,也可能是对方的某种特殊气泡
pass
else:
sender = "对方"
# 规则 C: 如果颜色是 unknown,且不在极端位置,使用中心点兜底
if sender == "unknown":
c_x = int((min_x + max_x) / 2)
@@ -584,6 +610,9 @@ def _scan_chat_messages(image_path):
# [Fix] 如果颜色明确是绿色,说明是"我"的左对齐文本(长文换行),不应被几何规则强制改为"对方"
if sender == "我" and sender_color == "green":
logger.info(f"Geometry says '对方' (center={c_x} < {w*0.45}) but Color is 'green'. Keeping '我'.")
elif sender == "system":
# 系统消息允许居中或偏左
pass
else:
if sender == "我":
logger.warning(f"Sender detected as '我' by color but center is left ({c_x} < {w*0.45}). Correcting to '对方'.")
@@ -592,7 +621,10 @@ def _scan_chat_messages(image_path):
elif c_x > w * 0.55:
if sender == "对方":
logger.warning(f"Sender detected as '对方' by color but center is right ({c_x} > {w*0.55}). Correcting to '我'.")
sender = "我"
elif sender == "system":
pass
else:
sender = "我"
else:
logger.info(f"Message in middle zone ({w*0.45} < {c_x} < {w*0.55}), trusting color detection: {sender}")
else:
@@ -665,10 +697,12 @@ def _scan_chat_messages(image_path):
return final_messages_with_time, debug_img, chat_title
async def analyze_chat_image(image_path, output_path, device=None, target_name="对方", process_strategy="ALL"):
async def analyze_chat_image(image_path, output_path, device=None, target_name="对方", process_strategy="ALL", restore_processed_voice=True):
"""
全面采用 CV + OCR 识别微信聊天截图中的最后一条消息
:param process_strategy: 语音处理策略 (ALL/UNREAD/LAST)
:param restore_processed_voice: 是否在转文字后还原(隐藏文字)。默认为 True。
设为 False 可防止最后一条消息在无内容时陷入"转文字->还原->空内容"的死循环。
注意:此函数现在包含一个循环,如果发现需要转文字的语音,会逐个处理并重新截图。
"""
try:
@@ -799,21 +833,39 @@ async def analyze_chat_image(image_path, output_path, device=None, target_name="
# 1. 截图 (但不立即 OCR,而是丢给异步任务)
peek_shot = get_next_debug_path("step_peek_content")
d.screenshot(peek_shot)
logger.info("已截图,启动异步OCR任务以提取内容...")
logger.info(f"已截图 {peek_shot},启动异步OCR任务以提取内容...")
async def _async_ocr_task(img_path, target_y):
"""内部异步任务:在线程池中运行 OCR"""
loop = asyncio.get_running_loop()
# 在默认执行器(线程池)中运行耗时的 _scan_chat_messages
msgs, _, _ = await loop.run_in_executor(None, _scan_chat_messages, img_path)
found = None
for pm in msgs:
if pm['type'] == 'voice' and pm.get('is_converted'):
if abs(pm['y'] - target_y) < 50:
found = pm.get('content')
break
return target_y, found
try:
loop = asyncio.get_running_loop()
# 在默认执行器(线程池)中运行耗时的 _scan_chat_messages
logger.info(f"🚀 [Async OCR] 开始分析截图 {os.path.basename(img_path)} (目标 Y={target_y})")
msgs, _, _ = await loop.run_in_executor(None, _scan_chat_messages, img_path)
found = None
# 收集所有可能是该语音消息转换出的文本
all_found_texts = []
for pm in msgs:
if pm['type'] == 'voice' and pm.get('is_converted'):
# 容差稍微放大,因为转文字展开后 Y 坐标会变
if abs(pm['y'] - target_y) < 150: # 进一步放宽容差
content = pm.get('content', '').strip()
if content:
all_found_texts.append((pm['y'], content))
if all_found_texts:
# 按 Y 轴排序,确保多行文本顺序正确
all_found_texts.sort(key=lambda x: x[0])
found = " ".join([t[1] for t in all_found_texts])
logger.info(f"✨ [Async OCR] 在 Y={target_y} 附近找到转换文字: {found}")
if not found:
logger.warning(f"⚠️ [Async OCR] 未能在 Y={target_y} 附近找到已转换文字")
return target_y, found
except Exception as e:
logger.error(f"❌ [Async OCR] 任务执行失败: {e}")
return target_y, None
# 创建并保存任务
task = asyncio.create_task(_async_ocr_task(peek_shot, vy))
@@ -822,44 +874,61 @@ async def analyze_chat_image(image_path, output_path, device=None, target_name="
analyze_chat_image._ocr_tasks = []
analyze_chat_image._ocr_tasks.append(task)
# 2. 立即还原状态 (取消转文字)
# 2. 还原状态 (取消转文字)
# 注意:由于 OCR 还没出结果,我们无法精确定位展开后的文字位置
# 但通常点击原语音气泡位置 (vx, vy) 也能触发菜单
logger.info("准备还原状态 (取消转文字)...")
d.long_click(vx, vy, 1.0) # 盲点原坐标
logger.info("正在快速寻找'隐藏文字'按钮...")
cancel_template = os.path.join(TEMPLATE_DIR, "cancel_zhuan_wen_zi.jpg")
cancel_btn = None
poll_start = time.time()
while time.time() - poll_start < 3.0:
restore_menu_shot = get_next_debug_path("step_restore_poll")
d.screenshot(restore_menu_shot)
cancel_btn = find_template_match(restore_menu_shot, cancel_template, threshold=0.7)
if restore_processed_voice:
logger.info("准备还原状态 (取消转文字)...")
d.long_click(vx, vy, 1.0) # 盲点原坐标
logger.info("正在快速寻找'隐藏文字'按钮...")
cancel_template = os.path.join(TEMPLATE_DIR, "cancel_zhuan_wen_zi.jpg")
cancel_btn = None
poll_start = time.time()
while time.time() - poll_start < 3.0:
restore_menu_shot = get_next_debug_path("step_restore_poll")
d.screenshot(restore_menu_shot)
cancel_btn = find_template_match(restore_menu_shot, cancel_template, threshold=0.7)
if cancel_btn:
break
time.sleep(0.2)
if cancel_btn:
cx, cy = int(cancel_btn[0]), int(cancel_btn[1])
logger.info(f"✅ 找到'隐藏文字'按钮: ({cx}, {cy}),点击还原...")
safe_device_click(d, cx, cy)
convert_closed_count += 1
time.sleep(2.0) # 等待收起动画
else:
logger.warning("❌ 未找到'隐藏文字'按钮,无法还原状态!(后续可能导致重复处理)")
# 3. 准备下一次循环
if len(target_voices) == 1:
logger.info("✅ 当前屏幕所有目标语音已处理完毕,无需再次全屏扫描。")
final_messages = messages # 使用本轮初始扫描的消息列表
break
time.sleep(0.2)
if cancel_btn:
cx, cy = int(cancel_btn[0]), int(cancel_btn[1])
logger.info(f"✅ 找到'隐藏文字'按钮: ({cx}, {cy}),点击还原...")
safe_device_click(d, cx, cy)
convert_closed_count += 1
time.sleep(2.0) # 等待收起动画
# 重新截图,因为界面可能微调,或者只是恢复了
next_screenshot = get_next_debug_path("step_restored")
d.screenshot(next_screenshot)
current_image_path = next_screenshot
current_output_path = get_next_debug_path("flag_restored")
continue
else:
logger.warning("❌ 未找到'隐藏文字'按钮,无法还原状态!(后续可能导致重复处理)")
# 3. 准备下一次循环
# 重新截图,因为界面可能微调,或者只是恢复了
next_screenshot = get_next_debug_path("step_restored")
d.screenshot(next_screenshot)
current_image_path = next_screenshot
current_output_path = get_next_debug_path("flag_restored")
continue
logger.info("⏩ [配置] 跳过还原状态步骤 (保持文字展开)。")
# 即使不还原,我们也不建议继续处理下一条,因为界面已经大幅变动(展开了文字)。
# 除非我们重新截图并重新定位。
# 但在这里,如果 restore_processed_voice=False,通常意味着我们只关心最后一条(LAST策略),或者我们接受界面变动。
# 为了安全起见,如果不还原,我们最好终止循环(假设只处理这一条,或者下一轮主循环再处理其他的)
# 否则后续的 target_voices 坐标全都不准了。
logger.info("🛑 因不还原状态,终止本轮多语音处理循环,等待下一次主监控循环。")
final_messages = messages # 这里的 messages 其实是展开前的,但没关系,我们的内容通过 captured_voice_contents 注入
break
else:
logger.warning("❌ 未找到'转文字'按钮,可能是已转换或误判")
@@ -887,12 +956,13 @@ async def analyze_chat_image(image_path, output_path, device=None, target_name="
if captured_voice_contents:
logger.info(f"正在注入 {len(captured_voice_contents)} 条已还原的语音内容...")
for m in final_messages:
if m['type'] == 'voice' and not m.get('content'):
if m['type'] == 'voice' and (not m.get('content') or m.get('content').strip() == ""):
for py, content in captured_voice_contents.items():
if abs(m['y'] - py) < 30: # 匹配原始 Y 坐标
# 注入时的容差也要放大,因为 final_messages 的 Y 可能和点击时的 vy 略有差异
if abs(m['y'] - py) < 100:
m['content'] = content
m['is_converted'] = True # 标记为逻辑上已转换
logger.info(f" -> 注入内容: {content[:10]}...")
logger.info(f" -> 注入内容到 Y={m['y']} (原 py={py}): {content[:20]}...")
break
# 构造返回值
@@ -931,15 +1001,16 @@ async def analyze_chat_image(image_path, output_path, device=None, target_name="
# 尝试注入异步获取的语音内容
if msg['type'] == 'voice':
# 模糊匹配 Y 坐标 (增大容差到 100,应对界面滚动)
for y_key, content in captured_voice_contents.items():
if abs(msg['y'] - y_key) < 100:
msg['is_converted'] = True
msg['content'] = content
logger.info(f"注入语音内容到最终消息列表: {content}")
break
# 优先检查 content 是否为空或为 placeholder
if not msg.get('content') or msg.get('content').strip() == "":
for y_key, content in captured_voice_contents.items():
if abs(msg['y'] - y_key) < 100:
msg['is_converted'] = True
msg['content'] = content
logger.info(f"✅ [注入] 成功将异步语音内容 '{content}' 注入到 Y={msg['y']} 的消息中")
break
# 无论是否有内容,都加入 dialogue_log
# 如果是语音且没内容,T2 会有兜底逻辑去处理
if msg['type'] == 'text':
if msg.get('content'): # 文本消息没内容通常是识别错误,可以丢弃
dialogue_log.append(msg)