mirror of
https://github.com/jxxghp/MoviePilot.git
synced 2026-04-01 01:41:59 +08:00
407 lines
14 KiB
Python
407 lines
14 KiB
Python
"""
|
||
活动日志中间件 - 自动记录 Agent 每次交互的操作摘要。
|
||
|
||
按日期存储在 CONFIG_PATH/agent/activity/YYYY-MM-DD.md 中，
|
||
每次 Agent 执行完毕后自动调用 LLM 对本轮对话生成简洁的活动摘要，
|
||
并在每次 Agent 启动时加载近几天的活动日志注入系统提示词。
|
||
"""
|
||
|
||
import re
|
||
from collections.abc import Awaitable, Callable
|
||
from datetime import datetime, timedelta
|
||
from typing import Annotated, Any, NotRequired, TypedDict
|
||
|
||
from anyio import Path as AsyncPath
|
||
from langchain.agents.middleware.types import (
|
||
AgentMiddleware,
|
||
AgentState,
|
||
ContextT,
|
||
ModelRequest,
|
||
ModelResponse,
|
||
PrivateStateAttr, # noqa
|
||
ResponseT,
|
||
)
|
||
from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
|
||
from langgraph.runtime import Runtime
|
||
|
||
from app.agent.middleware.utils import append_to_system_message
|
||
from app.log import logger
|
||
|
||
# 活动日志保留天数
|
||
DEFAULT_RETENTION_DAYS = 7
|
||
|
||
# 注入系统提示词时加载的天数
|
||
PROMPT_LOAD_DAYS = 3
|
||
|
||
# 每日日志文件最大大小 (256KB)
|
||
MAX_LOG_FILE_SIZE = 256 * 1024
|
||
|
||
# 提取本轮对话上下文的最大字符数（避免过长的对话消耗太多 token）
|
||
MAX_CONTEXT_FOR_SUMMARY = 4000
|
||
|
||
# LLM 总结的提示词
|
||
SUMMARY_PROMPT = """请根据以下 AI 助手与用户的对话记录，生成一条简洁的活动摘要（中文，一句话，不超过80字）。
|
||
摘要应包含：用户的需求是什么、助手做了什么、结果如何。
|
||
只输出摘要内容，不要加任何前缀、标点序号或解释。
|
||
|
||
对话记录：
|
||
{conversation}"""
|
||
|
||
|
||
class ActivityLogState(AgentState):
|
||
"""ActivityLogMiddleware 的状态模型。"""
|
||
|
||
activity_log_contents: NotRequired[Annotated[dict[str, str], PrivateStateAttr]]
|
||
"""将日期字符串映射到日志内容的字典。标记为私有，不包含在最终代理状态中。"""
|
||
|
||
|
||
class ActivityLogStateUpdate(TypedDict):
|
||
"""ActivityLogMiddleware 的状态更新。"""
|
||
|
||
activity_log_contents: dict[str, str]
|
||
|
||
|
||
def _extract_last_round(messages: list) -> list | None:
|
||
"""从完整消息列表中提取最后一轮交互。
|
||
|
||
从最后一条 HumanMessage 到消息末尾即为本轮交互。
|
||
|
||
参数：
|
||
messages: Agent 执行后的完整消息列表。
|
||
|
||
返回：
|
||
本轮交互的消息子列表，如果无有效交互则返回 None。
|
||
"""
|
||
if not messages:
|
||
return None
|
||
|
||
# 找到最后一条用户消息的索引
|
||
last_human_idx = None
|
||
for i in range(len(messages) - 1, -1, -1):
|
||
if isinstance(messages[i], HumanMessage) and messages[i].content:
|
||
last_human_idx = i
|
||
break
|
||
|
||
if last_human_idx is None:
|
||
return None
|
||
|
||
round_messages = messages[last_human_idx:]
|
||
|
||
# 检查是否为系统心跳消息
|
||
user_msg = round_messages[0]
|
||
user_content = (
|
||
user_msg.content if isinstance(user_msg.content, str) else str(user_msg.content)
|
||
)
|
||
if user_content.strip().startswith("[System Heartbeat]"):
|
||
return None
|
||
|
||
return round_messages
|
||
|
||
|
||
def _format_conversation_for_summary(round_messages: list) -> str:
|
||
"""将本轮对话消息格式化为文本，供 LLM 总结。
|
||
|
||
参数：
|
||
round_messages: 本轮交互的消息列表。
|
||
|
||
返回：
|
||
格式化后的对话文本。
|
||
"""
|
||
lines = []
|
||
total_len = 0
|
||
|
||
for msg in round_messages:
|
||
if isinstance(msg, HumanMessage):
|
||
content = msg.content if isinstance(msg.content, str) else str(msg.content)
|
||
line = f"用户: {content}"
|
||
elif isinstance(msg, AIMessage):
|
||
if hasattr(msg, "tool_calls") and msg.tool_calls:
|
||
tool_names = [
|
||
tc["name"]
|
||
for tc in msg.tool_calls
|
||
if isinstance(tc, dict) and "name" in tc
|
||
]
|
||
line = f"助手调用工具: {', '.join(tool_names)}"
|
||
elif msg.content:
|
||
content = (
|
||
msg.content if isinstance(msg.content, str) else str(msg.content)
|
||
)
|
||
line = f"助手: {content}"
|
||
else:
|
||
continue
|
||
elif isinstance(msg, ToolMessage):
|
||
content = msg.content if isinstance(msg.content, str) else str(msg.content)
|
||
# 工具返回可能很长，截断
|
||
if len(content) > 200:
|
||
content = content[:200] + "..."
|
||
line = f"工具返回: {content}"
|
||
else:
|
||
continue
|
||
|
||
# 控制总长度
|
||
if total_len + len(line) > MAX_CONTEXT_FOR_SUMMARY:
|
||
lines.append("...(后续对话省略)")
|
||
break
|
||
lines.append(line)
|
||
total_len += len(line)
|
||
|
||
return "\n".join(lines)
|
||
|
||
|
||
async def _summarize_with_llm(conversation_text: str) -> str | None:
|
||
"""调用 LLM 对对话文本生成活动摘要。
|
||
|
||
参数：
|
||
conversation_text: 格式化后的对话文本。
|
||
|
||
返回：
|
||
LLM 生成的摘要字符串，失败时返回 None。
|
||
"""
|
||
try:
|
||
from app.helper.llm import LLMHelper
|
||
|
||
llm = LLMHelper.get_llm(streaming=False)
|
||
prompt = SUMMARY_PROMPT.format(conversation=conversation_text)
|
||
response = await llm.ainvoke(prompt)
|
||
summary = response.content.strip()
|
||
# 清理模型可能输出的前缀（如 "摘要：" "总结："）
|
||
summary = re.sub(r"^(摘要|总结|活动记录)[：:]\s*", "", summary)
|
||
return summary if summary else None
|
||
except Exception as e:
|
||
logger.debug("LLM summarization failed: %s", e)
|
||
return None
|
||
|
||
|
||
ACTIVITY_LOG_SYSTEM_PROMPT = """<activity_log>
|
||
{activity_log}
|
||
</activity_log>
|
||
|
||
<activity_log_guidelines>
|
||
The above <activity_log> contains a record of your recent interactions with the user, automatically maintained by the system.
|
||
|
||
**How to use this information:**
|
||
- Reference past activities when relevant to provide continuity (e.g., "之前帮你订阅了《XXX》，现在有更新了")
|
||
- Use activity history to understand ongoing tasks and user patterns
|
||
- When the user asks "你之前帮我做了什么" or similar questions, refer to this log
|
||
- Activity logs are automatically recorded after each interaction - you do NOT need to manually update them
|
||
|
||
**What is automatically logged:**
|
||
- Each user interaction: what was asked, which tools were used, and the outcome
|
||
- Timestamps for all activities
|
||
- The log is organized by date for easy reference
|
||
|
||
**Important:**
|
||
- Activity logs are READ-ONLY from your perspective - the system manages them automatically
|
||
- Do not attempt to edit or write to activity log files
|
||
- For long-term preferences and knowledge, continue to use MEMORY.md
|
||
- Activity logs are retained for {retention_days} days and then automatically cleaned up
|
||
</activity_log_guidelines>
|
||
"""
|
||
|
||
|
||
class ActivityLogMiddleware(AgentMiddleware[ActivityLogState, ContextT, ResponseT]): # noqa
|
||
"""自动记录和加载 Agent 活动日志的中间件。
|
||
|
||
- abefore_agent: 加载近几天的活动日志
|
||
- awrap_model_call: 将活动日志注入系统提示词
|
||
- aafter_agent: 从本次对话中提取摘要并追加到当日日志文件
|
||
|
||
参数：
|
||
activity_dir: 活动日志存储目录路径。
|
||
retention_days: 日志保留天数（默认 7 天）。
|
||
prompt_load_days: 注入系统提示词时加载的天数（默认 3 天）。
|
||
"""
|
||
|
||
state_schema = ActivityLogState
|
||
|
||
def __init__(
|
||
self,
|
||
*,
|
||
activity_dir: str,
|
||
retention_days: int = DEFAULT_RETENTION_DAYS,
|
||
prompt_load_days: int = PROMPT_LOAD_DAYS,
|
||
) -> None:
|
||
self.activity_dir = activity_dir
|
||
self.retention_days = retention_days
|
||
self.prompt_load_days = prompt_load_days
|
||
|
||
def _get_log_path(self, date_str: str) -> AsyncPath:
|
||
"""获取指定日期的日志文件路径。"""
|
||
return AsyncPath(self.activity_dir) / f"{date_str}.md"
|
||
|
||
def _format_activity_log(self, contents: dict[str, str]) -> str:
|
||
"""格式化活动日志用于系统提示词注入。"""
|
||
if not contents:
|
||
return ACTIVITY_LOG_SYSTEM_PROMPT.format(
|
||
activity_log="(暂无活动记录)",
|
||
retention_days=self.retention_days,
|
||
)
|
||
|
||
# 按日期排序（最近的在前）
|
||
sorted_dates = sorted(contents.keys(), reverse=True)
|
||
sections = []
|
||
for date_str in sorted_dates:
|
||
content = contents[date_str].strip()
|
||
if content:
|
||
sections.append(f"### {date_str}\n{content}")
|
||
|
||
if not sections:
|
||
return ACTIVITY_LOG_SYSTEM_PROMPT.format(
|
||
activity_log="(暂无活动记录)",
|
||
retention_days=self.retention_days,
|
||
)
|
||
|
||
log_body = "\n\n".join(sections)
|
||
return ACTIVITY_LOG_SYSTEM_PROMPT.format(
|
||
activity_log=log_body,
|
||
retention_days=self.retention_days,
|
||
)
|
||
|
||
async def _load_recent_logs(self) -> dict[str, str]:
|
||
"""加载近几天的活动日志。"""
|
||
contents: dict[str, str] = {}
|
||
today = datetime.now().date()
|
||
|
||
for i in range(self.prompt_load_days):
|
||
date = today - timedelta(days=i)
|
||
date_str = date.strftime("%Y-%m-%d")
|
||
log_path = self._get_log_path(date_str)
|
||
|
||
if await log_path.exists():
|
||
try:
|
||
content = await log_path.read_text(encoding="utf-8")
|
||
contents[date_str] = content
|
||
logger.debug("Loaded activity log for %s", date_str)
|
||
except Exception as e:
|
||
logger.warning("Failed to load activity log %s: %s", date_str, e)
|
||
|
||
return contents
|
||
|
||
async def _append_activity(self, summary: str) -> None:
|
||
"""将一条活动记录追加到当日日志文件。"""
|
||
today_str = datetime.now().strftime("%Y-%m-%d")
|
||
now_str = datetime.now().strftime("%H:%M")
|
||
log_path = self._get_log_path(today_str)
|
||
|
||
# 确保目录存在
|
||
dir_path = AsyncPath(self.activity_dir)
|
||
if not await dir_path.exists():
|
||
await dir_path.mkdir(parents=True, exist_ok=True)
|
||
|
||
# 检查文件大小
|
||
if await log_path.exists():
|
||
stat = await log_path.stat()
|
||
if stat.st_size >= MAX_LOG_FILE_SIZE:
|
||
logger.warning(
|
||
"Activity log %s exceeds size limit (%d bytes), skipping append",
|
||
today_str,
|
||
stat.st_size,
|
||
)
|
||
return
|
||
|
||
# 追加记录
|
||
entry = f"- **{now_str}** {summary}\n"
|
||
try:
|
||
if await log_path.exists():
|
||
existing = await log_path.read_text(encoding="utf-8")
|
||
await log_path.write_text(existing + entry, encoding="utf-8")
|
||
else:
|
||
header = f"# {today_str} 活动日志\n\n"
|
||
await log_path.write_text(header + entry, encoding="utf-8")
|
||
logger.debug("Activity logged: %s", summary[:80])
|
||
except Exception as e:
|
||
logger.warning("Failed to append activity log: %s", e)
|
||
|
||
async def _cleanup_old_logs(self) -> None:
|
||
"""清理超过保留天数的旧日志文件。"""
|
||
dir_path = AsyncPath(self.activity_dir)
|
||
if not await dir_path.exists():
|
||
return
|
||
|
||
cutoff_date = datetime.now().date() - timedelta(days=self.retention_days)
|
||
date_pattern = re.compile(r"^(\d{4}-\d{2}-\d{2})\.md$")
|
||
|
||
try:
|
||
async for path in dir_path.iterdir():
|
||
if not await path.is_file():
|
||
continue
|
||
match = date_pattern.match(path.name)
|
||
if not match:
|
||
continue
|
||
try:
|
||
file_date = datetime.strptime(match.group(1), "%Y-%m-%d").date()
|
||
if file_date < cutoff_date:
|
||
await path.unlink()
|
||
logger.debug("Cleaned up old activity log: %s", path.name)
|
||
except ValueError:
|
||
continue
|
||
except Exception as e:
|
||
logger.warning("Failed to cleanup old activity logs: %s", e)
|
||
|
||
async def abefore_agent(
|
||
self, state: ActivityLogState, runtime: Runtime
|
||
) -> ActivityLogStateUpdate | None:
|
||
"""在 Agent 执行前加载近期活动日志。"""
|
||
# 如果已经加载则跳过
|
||
if "activity_log_contents" in state:
|
||
return None
|
||
|
||
contents = await self._load_recent_logs()
|
||
|
||
# 趁机清理旧日志（低频操作，不影响性能）
|
||
await self._cleanup_old_logs()
|
||
|
||
return ActivityLogStateUpdate(activity_log_contents=contents)
|
||
|
||
def modify_request(self, request: ModelRequest[ContextT]) -> ModelRequest[ContextT]:
|
||
"""将活动日志注入系统消息。"""
|
||
contents = request.state.get("activity_log_contents", {})
|
||
activity_log_prompt = self._format_activity_log(contents)
|
||
|
||
new_system_message = append_to_system_message(
|
||
request.system_message, activity_log_prompt
|
||
)
|
||
return request.override(system_message=new_system_message)
|
||
|
||
async def awrap_model_call(
|
||
self,
|
||
request: ModelRequest[ContextT],
|
||
handler: Callable[
|
||
[ModelRequest[ContextT]], Awaitable[ModelResponse[ResponseT]]
|
||
],
|
||
) -> ModelResponse[ResponseT]:
|
||
"""异步包装模型调用，注入活动日志到系统提示词。"""
|
||
modified_request = self.modify_request(request)
|
||
return await handler(modified_request)
|
||
|
||
async def aafter_agent(
|
||
self, state: ActivityLogState, runtime: Runtime
|
||
) -> dict[str, Any] | None:
|
||
"""Agent 执行完毕后，调用 LLM 对本轮对话生成摘要并追加到当日活动日志。"""
|
||
try:
|
||
messages = state.get("messages", [])
|
||
if not messages:
|
||
return None
|
||
|
||
# 提取本轮交互
|
||
round_messages = _extract_last_round(messages)
|
||
if not round_messages:
|
||
return None
|
||
|
||
# 格式化对话文本
|
||
conversation_text = _format_conversation_for_summary(round_messages)
|
||
if not conversation_text:
|
||
return None
|
||
|
||
# 调用 LLM 生成摘要
|
||
summary = await _summarize_with_llm(conversation_text)
|
||
if summary:
|
||
await self._append_activity(summary)
|
||
except Exception as e:
|
||
logger.warning("Failed to record activity: %s", e)
|
||
|
||
return None
|
||
|
||
|
||
__all__ = ["ActivityLogMiddleware"]
|