release: v8.0.4 — 会话 SQLite 存储 + 去加载动画 + 展示截断 + 工具调用解析加固
- 会话数据改用 SQLite（chat_history/conversations.db，WAL）：sessions / snapshots 两表， 首次启动自动迁移旧 JSON，/chat 与 /resume 改走 SQLite，memory 搜索兼容数据库 - 去掉流式显示（_typeprint 改整段打印）与常驻加载动画（移除 show_animation 线程及逐帧输出） - 展示给用户的消息过长自动截断（仅展示层，底层 messages 完整保存） - 工具调用解析加固：_loads_json 四级兜底修复；新增 _extract_last_action_json 成对错括号扫描； 续轮也下发 Function Calling 工具，减少长文本格式错误浪费 token - 版本号统一更新至 8.0.4（README/about.txt/constants/TUI banner/LSP·MCP clientInfo） - 12 项 session 单测 + 8 项 JSON 提取边界用例全部通过 Co-Authored-By: WorkBuddy <workbuddy@tencent.com>
This commit is contained in:
shuiyu67andWorkBuddy committed
1 parent
66de8eea13
commit
749fb3f91c
15 files changed
+510
-249
No files matched your search
+145
-139
@@ -375,16 +375,44 @@ multi 操作支持一次修改多处：
|
||||
return "\n".join(prompts)
|
||||
|
||||
def _loads_json(self, s):
|
||||
"""json.loads，但容忍未成对出现的反斜杠（Windows 路径 C:\\Users 常见）。
|
||||
"""json.loads，带多级容错修复（模型输出的 JSON 常常不规整）。
|
||||
|
||||
纯 JSON 模式下模型常把路径写成单反斜杠，直接 json.loads 会报非法转义。
|
||||
这里先按标准解析；失败时把「未成对的反斜杠」转义后再试一次。
|
||||
尝试顺序：
|
||||
1) 标准解析
|
||||
2) 未成对的反斜杠转义（Windows 路径 C:\\Users 常见）
|
||||
3) 字符串内的未转义换行/制表符（多行 args 常见）
|
||||
4) 去掉对象/数组里多余的尾随逗号（模型高频错误）
|
||||
"""
|
||||
try:
|
||||
return json.loads(s)
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
# 2) 反斜杠修复
|
||||
try:
|
||||
fixed = re.sub(r'(?<!\\)\\(?!\\\\)', r'\\\\', s)
|
||||
return json.loads(fixed)
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
# 3) 字符串内未转义换行/制表符修复
|
||||
try:
|
||||
repaired = re.sub(
|
||||
r'"([^"]*(?:\n[^"]*)*)"',
|
||||
lambda m: '"' + (m.group(1)
|
||||
.replace('\\', '\\\\')
|
||||
.replace('\n', '\\n')
|
||||
.replace('\r', '\\r')
|
||||
.replace('\t', '\\t')
|
||||
.replace('"', '\\"')) + '"',
|
||||
s, flags=re.DOTALL)
|
||||
return json.loads(repaired)
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
# 4) 尾随逗号修复
|
||||
try:
|
||||
trimmed = re.sub(r',\s*([}\]])', r'\1', s)
|
||||
return json.loads(trimmed)
|
||||
except json.JSONDecodeError:
|
||||
raise
|
||||
|
||||
def _parse_mixed_response(self, response):
|
||||
"""
|
||||
@@ -465,69 +493,87 @@ multi 操作支持一次修改多处：
|
||||
else:
|
||||
return text_content, json_objects
|
||||
|
||||
# 方法5: 查找内联的JSON对象
|
||||
json_start_markers = [
|
||||
'{"action": "use_tool"', '{"action": "continue"',
|
||||
'{"action":"use_tool"', '{"action":"continue"',
|
||||
'{ "action": "use_tool"', '{ "action": "continue"',
|
||||
'{ "action":"use_tool"', '{ "action":"continue"'
|
||||
]
|
||||
for marker in json_start_markers:
|
||||
json_start = response.find(marker)
|
||||
if json_start != -1:
|
||||
brace_count = 0
|
||||
in_json = False
|
||||
for i in range(json_start, len(response)):
|
||||
char = response[i]
|
||||
if char == '{':
|
||||
if not in_json:
|
||||
in_json = True
|
||||
brace_count += 1
|
||||
elif char == '}':
|
||||
brace_count -= 1
|
||||
if in_json and brace_count == 0:
|
||||
json_str = response[json_start:i+1]
|
||||
try:
|
||||
json_data = self._loads_json(json_str)
|
||||
if isinstance(json_data, dict) and json_data.get('action') in ['use_tool', 'continue']:
|
||||
text_content = response[:json_start].strip()
|
||||
return text_content, json_data
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
break
|
||||
# 方法5: 稳健抽取最后一个「控制指令」JSON（兼容长文本后格式不规整的情况）。
|
||||
# 结构化 FC 已是主路径；这里是纯 JSON 兜底：不再依赖脆弱的 marker 匹配，
|
||||
# 而是扫描所有平衡括号的 JSON 对象，取最后一个含控制键的，最大限度避免
|
||||
# 「长篇文本 + 格式错乱的工具调用」被解析失败、白白浪费 token。
|
||||
text_before, json_data = self._extract_last_action_json(response)
|
||||
if json_data is not None:
|
||||
return text_before, json_data
|
||||
|
||||
# 方法6: 通用JSON对象查找
|
||||
potential_jsons = re.findall(r'\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}', response, re.DOTALL)
|
||||
for json_str in potential_jsons:
|
||||
return response, None
|
||||
|
||||
def _extract_last_action_json(self, response):
|
||||
"""在可能夹杂长篇正文的响应里，稳健抽取最后一个控制指令 JSON。
|
||||
|
||||
控制指令指含以下任一字段的 JSON 对象：
|
||||
action / continue / need_continue / think_more / message
|
||||
返回 (text_before, json_data)；找不到返回 (response, None)。
|
||||
"""
|
||||
# 整段就是 JSON 数组（多个工具调用）时直接返回，保留并发执行能力
|
||||
stripped = response.strip()
|
||||
if stripped.startswith('['):
|
||||
try:
|
||||
json_data = self._loads_json(json_str)
|
||||
if isinstance(json_data, dict) and json_data.get('action') in ['use_tool', 'continue']:
|
||||
json_start = response.find(json_str)
|
||||
text_content = response[:json_start].strip()
|
||||
return text_content, json_data
|
||||
arr = self._loads_json(stripped)
|
||||
if isinstance(arr, list) and arr:
|
||||
return "", arr
|
||||
except json.JSONDecodeError:
|
||||
continue
|
||||
pass
|
||||
|
||||
# 方法7: 批量JSON
|
||||
CONTROL_KEYS = {"action", "continue", "need_continue", "think_more", "message"}
|
||||
candidates = [] # (start_index, parsed_dict)
|
||||
n = len(response)
|
||||
i = 0
|
||||
while i < n:
|
||||
if response[i] == '{':
|
||||
# 从 i 起做括号/字符串感知的平衡解析，取出一个完整 JSON 对象
|
||||
depth = 0
|
||||
in_str = False
|
||||
esc = False
|
||||
end = -1
|
||||
for j in range(i, n):
|
||||
c = response[j]
|
||||
if in_str:
|
||||
if esc:
|
||||
esc = False
|
||||
elif c == '\\':
|
||||
esc = True
|
||||
elif c == '"':
|
||||
in_str = False
|
||||
continue
|
||||
if c == '"':
|
||||
in_str = True
|
||||
elif c == '{':
|
||||
depth += 1
|
||||
elif c == '}':
|
||||
depth -= 1
|
||||
if depth == 0:
|
||||
end = j
|
||||
break
|
||||
if end != -1:
|
||||
frag = response[i:end + 1]
|
||||
try:
|
||||
obj = self._loads_json(frag)
|
||||
except json.JSONDecodeError:
|
||||
obj = None
|
||||
if isinstance(obj, dict) and (set(obj.keys()) & CONTROL_KEYS):
|
||||
candidates.append((i, obj))
|
||||
i = end + 1
|
||||
continue
|
||||
i += 1
|
||||
|
||||
if candidates:
|
||||
start, obj = candidates[-1]
|
||||
return response[:start].strip(), obj
|
||||
|
||||
# 退路：整段若为 JSON 数组（如 [{"action":"use_tool"}, ...]）
|
||||
try:
|
||||
cleaned = response.strip()
|
||||
for action in ['use_tool', 'continue']:
|
||||
last_brace_pos = cleaned.rfind(f'{{"action": "{action}"')
|
||||
if last_brace_pos == -1:
|
||||
last_brace_pos = cleaned.rfind(f'{{"action":"{action}"')
|
||||
if last_brace_pos != -1:
|
||||
json_part = cleaned[last_brace_pos:]
|
||||
if json_part.startswith('['):
|
||||
parsed = self._loads_json(json_part)
|
||||
if isinstance(parsed, list):
|
||||
text_content = cleaned[:last_brace_pos].strip()
|
||||
return text_content, parsed
|
||||
else:
|
||||
parsed = self._loads_json(json_part)
|
||||
if isinstance(parsed, dict):
|
||||
text_content = cleaned[:last_brace_pos].strip()
|
||||
return text_content, parsed
|
||||
except:
|
||||
stripped = response.strip()
|
||||
if stripped.startswith('['):
|
||||
arr = self._loads_json(stripped)
|
||||
if isinstance(arr, list) and arr:
|
||||
return "", arr
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
|
||||
return response, None
|
||||
@@ -922,10 +968,9 @@ multi 操作支持一次修改多处：
|
||||
return None
|
||||
except Exception:
|
||||
return None
|
||||
# 工具结果续轮：模型只需决定「继续/收尾」，不需要 FC schema，省下每轮重发的 token。
|
||||
# 若模型仍想调工具，走纯 JSON action（无需 schema 即可），行为等价、成本更低。
|
||||
if current_input and str(current_input).lstrip().startswith("工具执行结果"):
|
||||
return None
|
||||
# 工具结果续轮也下发 FC 工具：让模型在续轮里继续用结构化 function calling
|
||||
# 调用工具，避免把工具调用写成散文 JSON（长文本后格式错乱会浪费 token）。
|
||||
# 仅对带 _fc_supported() 的引擎下发 tools（Manual 等不支持的引擎不传 tools）。
|
||||
try:
|
||||
from .fc_tools import mcp_to_openai_tools, prune_tools
|
||||
all_tools = mcp_to_openai_tools(self.liugin_manager)
|
||||
@@ -956,90 +1001,55 @@ multi 操作支持一次修改多处：
|
||||
response = self._call_engine_with_tools(current_input, system_prompt, fc_tools)
|
||||
return response
|
||||
|
||||
# CLI 模式：带动画和跨平台 ESC 取消
|
||||
animation_running = threading.Event()
|
||||
animation_running.set()
|
||||
# CLI 模式：去掉流式显示与逐帧加载动画，改为「一次性等待 + 完整输出」。
|
||||
# 仍用线程跑生成以支持 ESC 取消，但不再喷任何中间动画/情话帧。
|
||||
esc_pressed = threading.Event()
|
||||
response_ready = threading.Event()
|
||||
response_result = [None] # 用列表存储，方便线程内修改
|
||||
|
||||
# 是否真实交互终端：非 TTY（管道 / 日志镜像 / 聊天前端捕获 stdout）时，
|
||||
# \r 无法原地覆盖，逐帧动画会污染输出流，必须禁用。
|
||||
# 是否真实交互终端：非 TTY（管道 / 日志镜像 / 聊天前端捕获 stdout）时
|
||||
# 不打印任何等待提示，避免污染捕获流。
|
||||
is_tty = sys.stdout.isatty()
|
||||
|
||||
love_sentences = []
|
||||
love_file_path = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), LOVE_FILE_PATH)
|
||||
if os.path.exists(love_file_path):
|
||||
try:
|
||||
with open(love_file_path, 'r', encoding='utf-8') as f:
|
||||
love_sentences = [line.strip() for line in f.readlines()
|
||||
if line.strip()
|
||||
and not line.strip().lower().startswith(('http://', 'https://'))]
|
||||
except Exception:
|
||||
pass
|
||||
if not love_sentences:
|
||||
love_sentences = ["AI正在思考中...", "请稍等片刻...", "正在处理您的请求..."]
|
||||
if is_tty:
|
||||
# 仅一行静态提示，非动画
|
||||
print(f"{Fore.MAGENTA}💭 思考中… (按 ESC 取消){Style.RESET_ALL}", end="", flush=True)
|
||||
|
||||
# ── 跨平台 ESC 检测 ──
|
||||
def check_for_esc():
|
||||
"""跨平台 ESC 键检测"""
|
||||
import platform
|
||||
if platform.system() == 'Windows':
|
||||
try:
|
||||
import msvcrt
|
||||
while animation_running.is_set():
|
||||
while not response_ready.is_set() and not esc_pressed.is_set():
|
||||
if platform.system() == 'Windows':
|
||||
try:
|
||||
import msvcrt
|
||||
if msvcrt.kbhit():
|
||||
key = msvcrt.getch()
|
||||
if ord(key) == 27: # ESC
|
||||
esc_pressed.set()
|
||||
break
|
||||
time.sleep(0.05)
|
||||
except ImportError:
|
||||
pass
|
||||
else:
|
||||
# Linux / macOS — 用 select 做非阻塞读取
|
||||
import select
|
||||
import sys
|
||||
try:
|
||||
# 保存原始终端设置
|
||||
import tty, termios
|
||||
fd = sys.stdin.fileno()
|
||||
old_settings = termios.tcgetattr(fd)
|
||||
except ImportError:
|
||||
break
|
||||
else:
|
||||
# Linux / macOS — 用 select 做非阻塞读取
|
||||
import select
|
||||
try:
|
||||
tty.setcbreak(fd) # cbreak 模式：无需回车即可读取
|
||||
while animation_running.is_set():
|
||||
if select.select([sys.stdin], [], [], 0.05)[0]:
|
||||
ch = sys.stdin.read(1)
|
||||
import sys as _sys
|
||||
import tty, termios
|
||||
fd = _sys.stdin.fileno()
|
||||
old_settings = termios.tcgetattr(fd)
|
||||
try:
|
||||
tty.setcbreak(fd) # cbreak 模式：无需回车即可读取
|
||||
if select.select([_sys.stdin], [], [], 0.05)[0]:
|
||||
ch = _sys.stdin.read(1)
|
||||
if ord(ch) == 27: # ESC
|
||||
esc_pressed.set()
|
||||
break
|
||||
finally:
|
||||
termios.tcsetattr(fd, termios.TCSADRAIN, old_settings)
|
||||
except Exception:
|
||||
# 降级：无法检测 ESC，只能等 AI 完成
|
||||
pass
|
||||
|
||||
# ── 动画线程 ──
|
||||
def show_animation():
|
||||
# 非 TTY：不跑逐帧 \r 动画（避免残帧进入捕获流），但保留一句随机情话保人味
|
||||
if not is_tty:
|
||||
try:
|
||||
print(f"💭 {random.choice(love_sentences)}", flush=True)
|
||||
except Exception:
|
||||
pass
|
||||
return
|
||||
last_change_time = time.time()
|
||||
current_sentence = random.choice(love_sentences)
|
||||
while animation_running.is_set():
|
||||
current_time = time.time()
|
||||
if current_time - last_change_time >= 5:
|
||||
current_sentence = random.choice(love_sentences)
|
||||
last_change_time = current_time
|
||||
animation_chars = "|/-\\"
|
||||
char_idx = int((current_time * 10) % len(animation_chars))
|
||||
display_sentence = current_sentence[:47] + "..." if len(current_sentence) > 50 else current_sentence
|
||||
print(f"\r{Fore.MAGENTA}{display_sentence} {animation_chars[char_idx]}{Style.RESET_ALL}", end="", flush=True)
|
||||
time.sleep(0.1)
|
||||
finally:
|
||||
termios.tcsetattr(fd, termios.TCSADRAIN, old_settings)
|
||||
except Exception:
|
||||
# 降级：无法检测 ESC，只能等 AI 完成
|
||||
break
|
||||
|
||||
# ── AI 响应线程 ──
|
||||
def generate_in_thread():
|
||||
@@ -1056,9 +1066,9 @@ multi 操作支持一次修改多处：
|
||||
finally:
|
||||
response_ready.set()
|
||||
|
||||
# 启动三个线程
|
||||
# 启动线程（生成 + ESC 检测）
|
||||
threads = []
|
||||
for target in (show_animation, check_for_esc, generate_in_thread):
|
||||
for target in (check_for_esc, generate_in_thread):
|
||||
t = threading.Thread(target=target, daemon=True)
|
||||
t.start()
|
||||
threads.append(t)
|
||||
@@ -1070,19 +1080,16 @@ multi 操作支持一次修改多处：
|
||||
|
||||
# ESC 被按下 → 取消
|
||||
if esc_pressed.is_set():
|
||||
# 清除动画行
|
||||
if is_tty:
|
||||
print("\r" + " " * 80 + "\r", end="", flush=True)
|
||||
print(f"\n{Fore.YELLOW}⚡ AI 请求已取消 (ESC){Style.RESET_ALL}")
|
||||
return "[AI请求已取消]"
|
||||
|
||||
# 响应正常返回
|
||||
# 响应正常返回：清除等待行，不做逐字流式打印。
|
||||
# 完整结果由 process_conversation 统一经 _display_response 透出（含展示层截断）。
|
||||
response = response_result[0]
|
||||
if response:
|
||||
# 清除动画行后再输出
|
||||
if is_tty:
|
||||
print("\r" + " " * 80 + "\r", end="", flush=True)
|
||||
self._typeprint(response, Fore.CYAN)
|
||||
if is_tty:
|
||||
print("\r" + " " * 80 + "\r", end="", flush=True)
|
||||
|
||||
if response and self.shared_conversation_history is not None:
|
||||
self.shared_conversation_history.append({
|
||||
@@ -1093,7 +1100,6 @@ multi 操作支持一次修改多处：
|
||||
return response
|
||||
|
||||
finally:
|
||||
animation_running.clear()
|
||||
for t in threads:
|
||||
t.join(timeout=1)
|
||||
|
||||
|
||||
Reference in new issue
Block a user