gardel/Mofox-Core
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat：统一normal和focus的动作调整,emoji统一可选随机激活或llm激活

Browse Source
This commit is contained in:
SengokuCola
2025-07-06 18:36:14 +08:00
parent 6c117742a9
commit 498d72384f
20 changed files with 217 additions and 748 deletions
Download Patch File Download Diff File Expand all files Collapse all files

323
src/chat/planner_actions/action_manager.py Normal file
View File

@@ -0,0 +1,323 @@
from typing import Dict, List, Optional, Type, Any
from src.plugin_system.base.base_action import BaseAction
from src.chat.message_receive.chat_stream import ChatStream
from src.common.logger import get_logger
from src.plugin_system.core.component_registry import component_registry
from src.plugin_system.base.component_types import ComponentType
logger = get_logger("action_manager")
# 定义动作信息类型
ActionInfo = Dict[str, Any]
class ActionManager:
"""
动作管理器，用于管理各种类型的动作
现在统一使用新插件系统，简化了原有的新旧兼容逻辑。
"""
# 类常量
DEFAULT_RANDOM_PROBABILITY = 0.3
DEFAULT_MODE = "all"
DEFAULT_ACTIVATION_TYPE = "always"
def __init__(self):
"""初始化动作管理器"""
# 所有注册的动作集合
self._registered_actions: Dict[str, ActionInfo] = {}
# 当前正在使用的动作集合，默认加载默认动作
self._using_actions: Dict[str, ActionInfo] = {}
# 默认动作集，仅作为快照，用于恢复默认
self._default_actions: Dict[str, ActionInfo] = {}
# 加载插件动作
self._load_plugin_actions()
# 初始化时将默认动作加载到使用中的动作
self._using_actions = self._default_actions.copy()
def _load_plugin_actions(self) -> None:
"""
加载所有插件系统中的动作
"""
try:
# 从新插件系统获取Action组件
self._load_plugin_system_actions()
logger.debug("从插件系统加载Action组件成功")
except Exception as e:
logger.error(f"加载插件动作失败: {e}")
def _load_plugin_system_actions(self) -> None:
"""从插件系统的component_registry加载Action组件"""
try:
from src.plugin_system.core.component_registry import component_registry
from src.plugin_system.base.component_types import ComponentType
# 获取所有Action组件
action_components = component_registry.get_components_by_type(ComponentType.ACTION)
for action_name, action_info in action_components.items():
if action_name in self._registered_actions:
logger.debug(f"Action组件 {action_name} 已存在，跳过")
continue
# 将插件系统的ActionInfo转换为ActionManager格式
converted_action_info = {
"description": action_info.description,
"parameters": getattr(action_info, "action_parameters", {}),
"require": getattr(action_info, "action_require", []),
"associated_types": getattr(action_info, "associated_types", []),
"enable_plugin": action_info.enabled,
# 激活类型相关
"focus_activation_type": action_info.focus_activation_type.value,
"normal_activation_type": action_info.normal_activation_type.value,
"random_activation_probability": action_info.random_activation_probability,
"llm_judge_prompt": action_info.llm_judge_prompt,
"activation_keywords": action_info.activation_keywords,
"keyword_case_sensitive": action_info.keyword_case_sensitive,
# 模式和并行设置
"mode_enable": action_info.mode_enable.value,
"parallel_action": action_info.parallel_action,
# 插件信息
"_plugin_name": getattr(action_info, "plugin_name", ""),
}
self._registered_actions[action_name] = converted_action_info
# 如果启用，也添加到默认动作集
if action_info.enabled:
self._default_actions[action_name] = converted_action_info
logger.debug(
f"从插件系统加载Action组件: {action_name} (插件: {getattr(action_info, 'plugin_name', 'unknown')})"
)
logger.info(f"从插件系统加载了 {len(action_components)} 个Action组件")
except Exception as e:
logger.error(f"从插件系统加载Action组件失败: {e}")
import traceback
logger.error(traceback.format_exc())
def create_action(
self,
action_name: str,
action_data: dict,
reasoning: str,
cycle_timers: dict,
thinking_id: str,
chat_stream: ChatStream,
log_prefix: str,
shutting_down: bool = False,
) -> Optional[BaseAction]:
"""
创建动作处理器实例
Args:
action_name: 动作名称
action_data: 动作数据
reasoning: 执行理由
cycle_timers: 计时器字典
thinking_id: 思考ID
chat_stream: 聊天流
log_prefix: 日志前缀
shutting_down: 是否正在关闭
Returns:
Optional[BaseAction]: 创建的动作处理器实例，如果动作名称未注册则返回None
"""
try:
# 获取组件类 - 明确指定查询Action类型
component_class = component_registry.get_component_class(action_name, ComponentType.ACTION)
if not component_class:
logger.warning(f"{log_prefix} 未找到Action组件: {action_name}")
return None
# 获取组件信息
component_info = component_registry.get_component_info(action_name, ComponentType.ACTION)
if not component_info:
logger.warning(f"{log_prefix} 未找到Action组件信息: {action_name}")
return None
# 获取插件配置
plugin_config = component_registry.get_plugin_config(component_info.plugin_name)
# 创建动作实例
instance = component_class(
action_data=action_data,
reasoning=reasoning,
cycle_timers=cycle_timers,
thinking_id=thinking_id,
chat_stream=chat_stream,
log_prefix=log_prefix,
shutting_down=shutting_down,
plugin_config=plugin_config,
)
logger.debug(f"创建Action实例成功: {action_name}")
return instance
except Exception as e:
logger.error(f"创建Action实例失败 {action_name}: {e}")
import traceback
logger.error(traceback.format_exc())
return None
def get_registered_actions(self) -> Dict[str, ActionInfo]:
"""获取所有已注册的动作集"""
return self._registered_actions.copy()
def get_default_actions(self) -> Dict[str, ActionInfo]:
"""获取默认动作集"""
return self._default_actions.copy()
def get_using_actions(self) -> Dict[str, ActionInfo]:
"""获取当前正在使用的动作集合"""
return self._using_actions.copy()
def get_using_actions_for_mode(self, mode: str) -> Dict[str, ActionInfo]:
"""
根据聊天模式获取可用的动作集合
Args:
mode: 聊天模式 ("focus", "normal", "all")
Returns:
Dict[str, ActionInfo]: 在指定模式下可用的动作集合
"""
filtered_actions = {}
for action_name, action_info in self._using_actions.items():
action_mode = action_info.get("mode_enable", "all")
# 检查动作是否在当前模式下启用
if action_mode == "all" or action_mode == mode:
filtered_actions[action_name] = action_info
logger.debug(f"动作 {action_name} 在模式 {mode} 下可用 (mode_enable: {action_mode})")
logger.debug(f"模式 {mode} 下可用动作: {list(filtered_actions.keys())}")
return filtered_actions
def add_action_to_using(self, action_name: str) -> bool:
"""
添加已注册的动作到当前使用的动作集
Args:
action_name: 动作名称
Returns:
bool: 添加是否成功
"""
if action_name not in self._registered_actions:
logger.warning(f"添加失败: 动作 {action_name} 未注册")
return False
if action_name in self._using_actions:
logger.info(f"动作 {action_name} 已经在使用中")
return True
self._using_actions[action_name] = self._registered_actions[action_name]
logger.info(f"添加动作 {action_name} 到使用集")
return True
def remove_action_from_using(self, action_name: str) -> bool:
"""
从当前使用的动作集中移除指定动作
Args:
action_name: 动作名称
Returns:
bool: 移除是否成功
"""
if action_name not in self._using_actions:
logger.warning(f"移除失败: 动作 {action_name} 不在当前使用的动作集中")
return False
del self._using_actions[action_name]
logger.debug(f"已从使用集中移除动作 {action_name}")
return True
def add_action(self, action_name: str, description: str, parameters: Dict = None, require: List = None) -> bool:
"""
添加新的动作到注册集
Args:
action_name: 动作名称
description: 动作描述
parameters: 动作参数定义，默认为空字典
require: 动作依赖项，默认为空列表
Returns:
bool: 添加是否成功
"""
if action_name in self._registered_actions:
return False
if parameters is None:
parameters = {}
if require is None:
require = []
action_info = {"description": description, "parameters": parameters, "require": require}
self._registered_actions[action_name] = action_info
return True
def remove_action(self, action_name: str) -> bool:
"""从注册集移除指定动作"""
if action_name not in self._registered_actions:
return False
del self._registered_actions[action_name]
# 如果在使用集中也存在，一并移除
if action_name in self._using_actions:
del self._using_actions[action_name]
return True
def temporarily_remove_actions(self, actions_to_remove: List[str]) -> None:
"""临时移除使用集中的指定动作"""
for name in actions_to_remove:
self._using_actions.pop(name, None)
def restore_actions(self) -> None:
"""恢复到默认动作集"""
logger.debug(
f"恢复动作集: 从 {list(self._using_actions.keys())} 恢复到默认动作集 {list(self._default_actions.keys())}"
)
self._using_actions = self._default_actions.copy()
def add_system_action_if_needed(self, action_name: str) -> bool:
"""
根据需要添加系统动作到使用集
Args:
action_name: 动作名称
Returns:
bool: 是否成功添加
"""
if action_name in self._registered_actions and action_name not in self._using_actions:
self._using_actions[action_name] = self._registered_actions[action_name]
logger.info(f"临时添加系统动作到使用集: {action_name}")
return True
return False
def get_action(self, action_name: str) -> Optional[Type[BaseAction]]:
"""
获取指定动作的处理器类
Args:
action_name: 动作名称
Returns:
Optional[Type[BaseAction]]: 动作处理器类，如果不存在则返回None
"""
from src.plugin_system.core.component_registry import component_registry
return component_registry.get_component_class(action_name)

548
src/chat/planner_actions/action_modifier.py Normal file
View File

@@ -0,0 +1,548 @@
from typing import List, Optional, Any, Dict
from src.chat.heart_flow.observation.observation import Observation
from src.common.logger import get_logger
from src.chat.heart_flow.observation.hfcloop_observation import HFCloopObservation
from src.chat.heart_flow.observation.chatting_observation import ChattingObservation
from src.chat.message_receive.chat_stream import get_chat_manager
from src.config.config import global_config
from src.llm_models.utils_model import LLMRequest
import random
import asyncio
import hashlib
import time
from src.chat.planner_actions.action_manager import ActionManager
from src.chat.utils.chat_message_builder import get_raw_msg_before_timestamp_with_chat, build_readable_messages
logger = get_logger("action_manager")
class ActionModifier:
"""动作处理器
用于处理Observation对象和根据激活类型处理actions。
集成了原有的modify_actions功能和新的激活类型处理功能。
支持并行判定和智能缓存优化。
"""
def __init__(self, action_manager: ActionManager, chat_id: str):
"""初始化动作处理器"""
self.chat_id = chat_id
self.chat_stream = get_chat_manager().get_stream(self.chat_id)
self.log_prefix = f"[{get_chat_manager().get_stream_name(self.chat_id) or self.chat_id}]"
self.action_manager = action_manager
# 用于LLM判定的小模型
self.llm_judge = LLMRequest(
model=global_config.model.utils_small,
request_type="action.judge",
)
# 缓存相关属性
self._llm_judge_cache = {} # 缓存LLM判定结果
self._cache_expiry_time = 30 # 缓存过期时间（秒）
self._last_context_hash = None # 上次上下文的哈希值
async def modify_actions(
self,
mode: str = "focus",
observations: Optional[List[Observation]] = None,
message_content: str = "",
):
"""
动作修改流程，整合传统观察处理和新的激活类型判定
这个方法处理完整的动作管理流程：
1. 基于观察的传统动作修改（循环历史分析、类型匹配等）
2. 基于激活类型的智能动作判定，最终确定可用动作集
处理后，ActionManager 将包含最终的可用动作集，供规划器直接使用
"""
logger.debug(f"{self.log_prefix}开始完整动作修改流程")
removals_s1 = []
removals_s2 = []
self.action_manager.restore_actions()
all_actions = self.action_manager.get_using_actions_for_mode(mode)
message_list_before_now_half = get_raw_msg_before_timestamp_with_chat(
chat_id=self.chat_stream.stream_id,
timestamp=time.time(),
limit=int(global_config.chat.max_context_size * 0.5),
)
chat_content = build_readable_messages(
message_list_before_now_half,
replace_bot_name=True,
merge_messages=False,
timestamp_mode="relative",
read_mark=0.0,
show_actions=True,
)
if message_content:
chat_content = chat_content + "\n" + f"现在，最新的消息是：{message_content}"
# === 第一阶段：传统观察处理 ===
if observations:
for obs in observations:
if isinstance(obs, HFCloopObservation):
# 获取适用于FOCUS模式的动作
removals_from_loop = await self.analyze_loop_actions(obs)
if removals_from_loop:
removals_s1.extend(removals_from_loop)
# 检查动作的关联类型
chat_context = self.chat_stream.context
type_mismatched_actions = self._check_action_associated_types(all_actions, chat_context)
if type_mismatched_actions:
removals_s1.extend(type_mismatched_actions)
# 应用第一阶段的移除
for action_name, reason in removals_s1:
self.action_manager.remove_action_from_using(action_name)
logger.debug(f"{self.log_prefix}阶段一移除动作: {action_name}，原因: {reason}")
# === 第二阶段：激活类型判定 ===
if chat_content is not None:
logger.debug(f"{self.log_prefix}开始激活类型判定阶段")
# 获取当前使用的动作集（经过第一阶段处理）
current_using_actions = self.action_manager.get_using_actions_for_mode(mode)
# 获取因激活类型判定而需要移除的动作
removals_s2 = await self._get_deactivated_actions_by_type(
current_using_actions,
mode,
chat_content,
)
# 应用第二阶段的移除
for action_name, reason in removals_s2:
self.action_manager.remove_action_from_using(action_name)
logger.debug(f"{self.log_prefix}阶段二移除动作: {action_name}，原因: {reason}")
# === 统一日志记录 ===
all_removals = removals_s1 + removals_s2
if all_removals:
removals_summary = " | ".join([f"{name}({reason})" for name, reason in all_removals])
logger.info(
f"{self.log_prefix}{mode}模式动作修改流程结束，最终可用动作: {list(self.action_manager.get_using_actions_for_mode(mode).keys())}||移除记录: {removals_summary}"
)
def _check_action_associated_types(self, all_actions, chat_context):
type_mismatched_actions = []
for action_name, data in all_actions.items():
if data.get("associated_types"):
if not chat_context.check_types(data["associated_types"]):
associated_types_str = ", ".join(data["associated_types"])
reason = f"适配器不支持（需要: {associated_types_str}）"
type_mismatched_actions.append((action_name, reason))
logger.debug(
f"{self.log_prefix}决定移除动作: {action_name}，原因: {reason}"
)
return type_mismatched_actions
async def _get_deactivated_actions_by_type(
self,
actions_with_info: Dict[str, Any],
mode: str = "focus",
chat_content: str = "",
) -> List[tuple[str, str]]:
"""
根据激活类型过滤，返回需要停用的动作列表及原因
Args:
actions_with_info: 带完整信息的动作字典
chat_content: 聊天内容
Returns:
List[Tuple[str, str]]: 需要停用的 (action_name, reason) 元组列表
"""
deactivated_actions = []
# 分类处理不同激活类型的actions
llm_judge_actions = {}
actions_to_check = list(actions_with_info.items())
random.shuffle(actions_to_check)
for action_name, action_info in actions_to_check:
activation_type = f"{mode}_activation_type"
activation_type = action_info.get(activation_type, "always")
if activation_type == "always":
continue # 总是激活，无需处理
elif activation_type == "random":
probability = action_info.get("random_activation_probability", ActionManager.DEFAULT_RANDOM_PROBABILITY)
if not (random.random() < probability):
reason = f"RANDOM类型未触发（概率{probability}）"
deactivated_actions.append((action_name, reason))
logger.debug(f"{self.log_prefix}未激活动作: {action_name}，原因: {reason}")
elif activation_type == "keyword":
if not self._check_keyword_activation(action_name, action_info, chat_content):
keywords = action_info.get("activation_keywords", [])
reason = f"关键词未匹配（关键词: {keywords}）"
deactivated_actions.append((action_name, reason))
logger.debug(f"{self.log_prefix}未激活动作: {action_name}，原因: {reason}")
elif activation_type == "llm_judge":
llm_judge_actions[action_name] = action_info
else:
logger.warning(f"{self.log_prefix}未知的激活类型: {activation_type}，跳过处理")
# 并行处理LLM_JUDGE类型
if llm_judge_actions:
llm_results = await self._process_llm_judge_actions_parallel(
llm_judge_actions,
chat_content,
)
for action_name, should_activate in llm_results.items():
if not should_activate:
reason = "LLM判定未激活"
deactivated_actions.append((action_name, reason))
logger.debug(f"{self.log_prefix}未激活动作: {action_name}，原因: {reason}")
return deactivated_actions
async def process_actions_for_planner(
self, observed_messages_str: str = "", chat_context: Optional[str] = None, extra_context: Optional[str] = None
) -> Dict[str, Any]:
"""
[已废弃] 此方法现在已被整合到 modify_actions() 中
为了保持向后兼容性而保留，但建议直接使用 ActionManager.get_using_actions()
规划器应该直接从 ActionManager 获取最终的可用动作集，而不是调用此方法
新的架构：
1. 主循环调用 modify_actions() 处理完整的动作管理流程
2. 规划器直接使用 ActionManager.get_using_actions() 获取最终动作集
"""
logger.warning(
f"{self.log_prefix}process_actions_for_planner() 已废弃，建议规划器直接使用 ActionManager.get_using_actions()"
)
# 为了向后兼容，仍然返回当前使用的动作集
current_using_actions = self.action_manager.get_using_actions()
all_registered_actions = self.action_manager.get_registered_actions()
# 构建完整的动作信息
result = {}
for action_name in current_using_actions.keys():
if action_name in all_registered_actions:
result[action_name] = all_registered_actions[action_name]
return result
def _generate_context_hash(self, chat_content: str) -> str:
"""生成上下文的哈希值用于缓存"""
context_content = f"{chat_content}"
return hashlib.md5(context_content.encode("utf-8")).hexdigest()
async def _process_llm_judge_actions_parallel(
self,
llm_judge_actions: Dict[str, Any],
chat_content: str = "",
) -> Dict[str, bool]:
"""
并行处理LLM判定actions，支持智能缓存
Args:
llm_judge_actions: 需要LLM判定的actions
chat_content: 聊天内容
Returns:
Dict[str, bool]: action名称到激活结果的映射
"""
# 生成当前上下文的哈希值
current_context_hash = self._generate_context_hash(chat_content)
current_time = time.time()
results = {}
tasks_to_run = {}
# 检查缓存
for action_name, action_info in llm_judge_actions.items():
cache_key = f"{action_name}_{current_context_hash}"
# 检查是否有有效的缓存
if (
cache_key in self._llm_judge_cache
and current_time - self._llm_judge_cache[cache_key]["timestamp"] < self._cache_expiry_time
):
results[action_name] = self._llm_judge_cache[cache_key]["result"]
logger.debug(
f"{self.log_prefix}使用缓存结果 {action_name}: {'激活' if results[action_name] else '未激活'}"
)
else:
# 需要进行LLM判定
tasks_to_run[action_name] = action_info
# 如果有需要运行的任务，并行执行
if tasks_to_run:
logger.debug(f"{self.log_prefix}并行执行LLM判定，任务数: {len(tasks_to_run)}")
# 创建并行任务
tasks = []
task_names = []
for action_name, action_info in tasks_to_run.items():
task = self._llm_judge_action(
action_name,
action_info,
chat_content,
)
tasks.append(task)
task_names.append(action_name)
# 并行执行所有任务
try:
task_results = await asyncio.gather(*tasks, return_exceptions=True)
# 处理结果并更新缓存
for _, (action_name, result) in enumerate(zip(task_names, task_results)):
if isinstance(result, Exception):
logger.error(f"{self.log_prefix}LLM判定action {action_name} 时出错: {result}")
results[action_name] = False
else:
results[action_name] = result
# 更新缓存
cache_key = f"{action_name}_{current_context_hash}"
self._llm_judge_cache[cache_key] = {"result": result, "timestamp": current_time}
logger.debug(f"{self.log_prefix}并行LLM判定完成，耗时: {time.time() - current_time:.2f}s")
except Exception as e:
logger.error(f"{self.log_prefix}并行LLM判定失败: {e}")
# 如果并行执行失败，为所有任务返回False
for action_name in tasks_to_run.keys():
results[action_name] = False
# 清理过期缓存
self._cleanup_expired_cache(current_time)
return results
def _cleanup_expired_cache(self, current_time: float):
"""清理过期的缓存条目"""
expired_keys = []
for cache_key, cache_data in self._llm_judge_cache.items():
if current_time - cache_data["timestamp"] > self._cache_expiry_time:
expired_keys.append(cache_key)
for key in expired_keys:
del self._llm_judge_cache[key]
if expired_keys:
logger.debug(f"{self.log_prefix}清理了 {len(expired_keys)} 个过期缓存条目")
async def _llm_judge_action(
self,
action_name: str,
action_info: Dict[str, Any],
chat_content: str = "",
) -> bool:
"""
使用LLM判定是否应该激活某个action
Args:
action_name: 动作名称
action_info: 动作信息
observed_messages_str: 观察到的聊天消息
chat_context: 聊天上下文
extra_context: 额外上下文
Returns:
bool: 是否应该激活此action
"""
try:
# 构建判定提示词
action_description = action_info.get("description", "")
action_require = action_info.get("require", [])
custom_prompt = action_info.get("llm_judge_prompt", "")
# 构建基础判定提示词
base_prompt = f"""
你需要判断在当前聊天情况下，是否应该激活名为"{action_name}"的动作。
动作描述：{action_description}
动作使用场景：
"""
for req in action_require:
base_prompt += f"- {req}\n"
if custom_prompt:
base_prompt += f"\n额外判定条件：\n{custom_prompt}\n"
if chat_content:
base_prompt += f"\n当前聊天记录：\n{chat_content}\n"
base_prompt += """
请根据以上信息判断是否应该激活这个动作。
只需要回答"是"或"否"，不要有其他内容。
"""
# 调用LLM进行判定
response, _ = await self.llm_judge.generate_response_async(prompt=base_prompt)
# 解析响应
response = response.strip().lower()
# print(base_prompt)
# print(f"LLM判定动作 {action_name}：响应='{response}'")
should_activate = "是" in response or "yes" in response or "true" in response
logger.debug(
f"{self.log_prefix}LLM判定动作 {action_name}：响应='{response}'，结果={'激活' if should_activate else '不激活'}"
)
return should_activate
except Exception as e:
logger.error(f"{self.log_prefix}LLM判定动作 {action_name} 时出错: {e}")
# 出错时默认不激活
return False
def _check_keyword_activation(
self,
action_name: str,
action_info: Dict[str, Any],
chat_content: str = "",
) -> bool:
"""
检查是否匹配关键词触发条件
Args:
action_name: 动作名称
action_info: 动作信息
observed_messages_str: 观察到的聊天消息
chat_context: 聊天上下文
extra_context: 额外上下文
Returns:
bool: 是否应该激活此action
"""
activation_keywords = action_info.get("activation_keywords", [])
case_sensitive = action_info.get("keyword_case_sensitive", False)
if not activation_keywords:
logger.warning(f"{self.log_prefix}动作 {action_name} 设置为关键词触发但未配置关键词")
return False
# 构建检索文本
search_text = ""
if chat_content:
search_text += chat_content
# if chat_context:
# search_text += f" {chat_context}"
# if extra_context:
# search_text += f" {extra_context}"
# 如果不区分大小写，转换为小写
if not case_sensitive:
search_text = search_text.lower()
# 检查每个关键词
matched_keywords = []
for keyword in activation_keywords:
check_keyword = keyword if case_sensitive else keyword.lower()
if check_keyword in search_text:
matched_keywords.append(keyword)
if matched_keywords:
logger.debug(f"{self.log_prefix}动作 {action_name} 匹配到关键词: {matched_keywords}")
return True
else:
logger.debug(f"{self.log_prefix}动作 {action_name} 未匹配到任何关键词: {activation_keywords}")
return False
async def analyze_loop_actions(self, obs: HFCloopObservation) -> List[tuple[str, str]]:
"""分析最近的循环内容并决定动作的移除
Returns:
List[Tuple[str, str]]: 包含要删除的动作及原因的元组列表
[("action3", "some reason")]
"""
removals = []
# 获取最近10次循环
recent_cycles = obs.history_loop[-10:] if len(obs.history_loop) > 10 else obs.history_loop
if not recent_cycles:
return removals
reply_sequence = [] # 记录最近的动作序列
for cycle in recent_cycles:
action_result = cycle.loop_plan_info.get("action_result", {})
action_type = action_result.get("action_type", "unknown")
reply_sequence.append(action_type == "reply")
# 计算连续回复的相关阈值
max_reply_num = int(global_config.focus_chat.consecutive_replies * 3.2)
sec_thres_reply_num = int(global_config.focus_chat.consecutive_replies * 2)
one_thres_reply_num = int(global_config.focus_chat.consecutive_replies * 1.5)
# 获取最近max_reply_num次的reply状态
if len(reply_sequence) >= max_reply_num:
last_max_reply_num = reply_sequence[-max_reply_num:]
else:
last_max_reply_num = reply_sequence[:]
# 详细打印阈值和序列信息，便于调试
logger.info(
f"连续回复阈值: max={max_reply_num}, sec={sec_thres_reply_num}, one={one_thres_reply_num}，"
f"最近reply序列: {last_max_reply_num}"
)
# print(f"consecutive_replies: {consecutive_replies}")
# 根据最近的reply情况决定是否移除reply动作
if len(last_max_reply_num) >= max_reply_num and all(last_max_reply_num):
# 如果最近max_reply_num次都是reply，直接移除
reason = f"连续回复过多（最近{len(last_max_reply_num)}次全是reply，超过阈值{max_reply_num}）"
removals.append(("reply", reason))
# reply_count = len(last_max_reply_num) - no_reply_count
elif len(last_max_reply_num) >= sec_thres_reply_num and all(last_max_reply_num[-sec_thres_reply_num:]):
# 如果最近sec_thres_reply_num次都是reply，40%概率移除
removal_probability = 0.4 / global_config.focus_chat.consecutive_replies
if random.random() < removal_probability:
reason = f"连续回复较多（最近{sec_thres_reply_num}次全是reply，{removal_probability:.2f}概率移除，触发移除）"
removals.append(("reply", reason))
elif len(last_max_reply_num) >= one_thres_reply_num and all(last_max_reply_num[-one_thres_reply_num:]):
# 如果最近one_thres_reply_num次都是reply，20%概率移除
removal_probability = 0.2 / global_config.focus_chat.consecutive_replies
if random.random() < removal_probability:
reason = f"连续回复检测（最近{one_thres_reply_num}次全是reply，{removal_probability:.2f}概率移除，触发移除）"
removals.append(("reply", reason))
else:
logger.debug(f"{self.log_prefix}连续回复检测：无需移除reply动作，最近回复模式正常")
return removals
def get_available_actions_count(self) -> int:
"""获取当前可用动作数量（排除默认的no_action）"""
current_actions = self.action_manager.get_using_actions_for_mode("normal")
# 排除no_action（如果存在）
filtered_actions = {k: v for k, v in current_actions.items() if k != "no_action"}
return len(filtered_actions)
def should_skip_planning(self) -> bool:
"""判断是否应该跳过规划过程"""
available_count = self.get_available_actions_count()
if available_count == 0:
logger.debug(f"{self.log_prefix} 没有可用动作，跳过规划")
return True
return False

369
src/chat/planner_actions/planner_focus.py Normal file
View File

@@ -0,0 +1,369 @@
import json # <--- 确保导入 json
import traceback
from typing import List, Dict, Any, Optional
from rich.traceback import install
from src.llm_models.utils_model import LLMRequest
from src.config.config import global_config
from src.chat.focus_chat.info.info_base import InfoBase
from src.chat.focus_chat.info.obs_info import ObsInfo
from src.chat.focus_chat.info.action_info import ActionInfo
from src.common.logger import get_logger
from src.chat.utils.prompt_builder import Prompt, global_prompt_manager
from src.chat.planner_actions.action_manager import ActionManager
from json_repair import repair_json
from src.chat.heart_flow.utils_chat import get_chat_type_and_target_info
from datetime import datetime
logger = get_logger("planner")
install(extra_lines=3)
def init_prompt():
Prompt(
"""
{time_block}
{indentify_block}
你现在需要根据聊天内容，选择的合适的action来参与聊天。
{chat_context_description}，以下是具体的聊天内容：
{chat_content_block}
{moderation_prompt}
现在请你根据聊天内容选择合适的action:
{action_options_text}
请根据动作示例，以严格的 JSON 格式输出，且仅包含 JSON 内容：
""",
"simple_planner_prompt",
)
Prompt(
"""
{time_block}
{indentify_block}
你现在需要根据聊天内容，选择的合适的action来参与聊天。
{chat_context_description}，以下是具体的聊天内容：
{chat_content_block}
{moderation_prompt}
现在请你选择合适的action:
{action_options_text}
请根据动作示例，以严格的 JSON 格式输出，且仅包含 JSON 内容：
""",
"simple_planner_prompt_private",
)
Prompt(
"""
动作：{action_name}
动作描述：{action_description}
{action_require}
{{
"action": "{action_name}",{action_parameters}
}}
""",
"action_prompt",
)
class ActionPlanner:
def __init__(self, log_prefix: str, action_manager: ActionManager):
self.log_prefix = log_prefix
self.action_manager = action_manager
# LLM规划器配置
self.planner_llm = LLMRequest(
model=global_config.model.planner,
request_type="focus.planner", # 用于动作规划
)
self.utils_llm = LLMRequest(
model=global_config.model.utils_small,
request_type="focus.planner", # 用于动作规划
)
async def plan(
self, all_plan_info: List[InfoBase],loop_start_time: float
) -> Dict[str, Any]:
"""
规划器 (Planner): 使用LLM根据上下文决定做出什么动作。
参数:
all_plan_info: 所有计划信息
running_memorys: 回忆信息
loop_start_time: 循环开始时间
"""
action = "no_reply" # 默认动作
reasoning = "规划器初始化默认"
action_data = {}
try:
# 获取观察信息
extra_info: list[str] = []
extra_info = []
observed_messages = []
observed_messages_str = ""
chat_type = "group"
is_group_chat = True
chat_id = None # 添加chat_id变量
for info in all_plan_info:
if isinstance(info, ObsInfo):
observed_messages = info.get_talking_message()
observed_messages_str = info.get_talking_message_str_truncate_short()
chat_type = info.get_chat_type()
is_group_chat = chat_type == "group"
# 从ObsInfo中获取chat_id
chat_id = info.get_chat_id()
else:
extra_info.append(info.get_processed_info())
# 获取聊天类型和目标信息
chat_target_info = None
if chat_id:
try:
# 重新获取更准确的聊天信息
is_group_chat_updated, chat_target_info = get_chat_type_and_target_info(chat_id)
# 如果获取成功，更新is_group_chat
if is_group_chat_updated is not None:
is_group_chat = is_group_chat_updated
logger.debug(
f"{self.log_prefix}获取到聊天信息 - 群聊: {is_group_chat}, 目标信息: {chat_target_info}"
)
except Exception as e:
logger.warning(f"{self.log_prefix}获取聊天目标信息失败: {e}")
chat_target_info = None
# 获取经过modify_actions处理后的最终可用动作集
# 注意：动作的激活判定现在在主循环的modify_actions中完成
# 使用Focus模式过滤动作
current_available_actions_dict = self.action_manager.get_using_actions_for_mode("focus")
# 获取完整的动作信息
all_registered_actions = self.action_manager.get_registered_actions()
current_available_actions = {}
for action_name in current_available_actions_dict.keys():
if action_name in all_registered_actions:
current_available_actions[action_name] = all_registered_actions[action_name]
else:
logger.warning(f"{self.log_prefix}使用中的动作 {action_name} 未在已注册动作中找到")
# 如果没有可用动作或只有no_reply动作，直接返回no_reply
if not current_available_actions or (
len(current_available_actions) == 1 and "no_reply" in current_available_actions
):
action = "no_reply"
reasoning = "没有可用的动作" if not current_available_actions else "只有no_reply动作可用，跳过规划"
logger.info(f"{self.log_prefix}{reasoning}")
self.action_manager.restore_actions()
logger.debug(
f"{self.log_prefix}[focus]沉默后恢复到默认动作集, 当前可用: {list(self.action_manager.get_using_actions().keys())}"
)
return {
"action_result": {"action_type": action, "action_data": action_data, "reasoning": reasoning},
"observed_messages": observed_messages,
}
# --- 构建提示词 (调用修改后的 PromptBuilder 方法) ---
prompt = await self.build_planner_prompt(
is_group_chat=is_group_chat, # <-- Pass HFC state
chat_target_info=chat_target_info, # <-- 传递获取到的聊天目标信息
observed_messages_str=observed_messages_str, # <-- Pass local variable
current_available_actions=current_available_actions, # <-- Pass determined actions
)
# --- 调用 LLM (普通文本生成) ---
llm_content = None
try:
prompt = f"{prompt}"
llm_content, (reasoning_content, _) = await self.planner_llm.generate_response_async(prompt=prompt)
logger.info(f"{self.log_prefix}规划器原始提示词: {prompt}")
logger.info(f"{self.log_prefix}规划器原始响应: {llm_content}")
if reasoning_content:
logger.info(f"{self.log_prefix}规划器推理: {reasoning_content}")
except Exception as req_e:
logger.error(f"{self.log_prefix}LLM 请求执行失败: {req_e}")
reasoning = f"LLM 请求失败，你的模型出现问题: {req_e}"
action = "no_reply"
if llm_content:
try:
fixed_json_string = repair_json(llm_content)
if isinstance(fixed_json_string, str):
try:
parsed_json = json.loads(fixed_json_string)
except json.JSONDecodeError as decode_error:
logger.error(f"JSON解析错误: {str(decode_error)}")
parsed_json = {}
else:
# 如果repair_json直接返回了字典对象，直接使用
parsed_json = fixed_json_string
# 处理repair_json可能返回列表的情况
if isinstance(parsed_json, list):
if parsed_json:
# 取列表中最后一个元素（通常是最完整的）
parsed_json = parsed_json[-1]
logger.warning(f"{self.log_prefix}LLM返回了多个JSON对象，使用最后一个: {parsed_json}")
else:
parsed_json = {}
# 确保parsed_json是字典
if not isinstance(parsed_json, dict):
logger.error(f"{self.log_prefix}解析后的JSON不是字典类型: {type(parsed_json)}")
parsed_json = {}
# 提取决策，提供默认值
extracted_action = parsed_json.get("action", "no_reply")
extracted_reasoning = ""
# 将所有其他属性添加到action_data
action_data = {}
for key, value in parsed_json.items():
if key not in ["action", "reasoning"]:
action_data[key] = value
action_data["loop_start_time"] = loop_start_time
# 对于reply动作不需要额外处理，因为相关字段已经在上面的循环中添加到action_data
if extracted_action not in current_available_actions:
logger.warning(
f"{self.log_prefix}LLM 返回了当前不可用或无效的动作: '{extracted_action}' (可用: {list(current_available_actions.keys())})，将强制使用 'no_reply'"
)
action = "no_reply"
reasoning = f"LLM 返回了当前不可用的动作 '{extracted_action}' (可用: {list(current_available_actions.keys())})。原始理由: {extracted_reasoning}"
else:
# 动作有效且可用
action = extracted_action
reasoning = extracted_reasoning
except Exception as json_e:
logger.warning(f"{self.log_prefix}解析LLM响应JSON失败 {json_e}. LLM原始输出: '{llm_content}'")
traceback.print_exc()
reasoning = f"解析LLM响应JSON失败: {json_e}. 将使用默认动作 'no_reply'."
action = "no_reply"
except Exception as outer_e:
logger.error(f"{self.log_prefix}Planner 处理过程中发生意外错误，规划失败，将执行 no_reply: {outer_e}")
traceback.print_exc()
action = "no_reply"
reasoning = f"Planner 内部处理错误: {outer_e}"
# 恢复到默认动作集
self.action_manager.restore_actions()
logger.debug(
f"{self.log_prefix}规划后恢复到默认动作集, 当前可用: {list(self.action_manager.get_using_actions().keys())}"
)
action_result = {"action_type": action, "action_data": action_data, "reasoning": reasoning}
plan_result = {
"action_result": action_result,
"observed_messages": observed_messages,
"action_prompt": prompt,
}
return plan_result
async def build_planner_prompt(
self,
is_group_chat: bool, # Now passed as argument
chat_target_info: Optional[dict], # Now passed as argument
observed_messages_str: str,
current_available_actions: Dict[str, ActionInfo],
) -> str:
"""构建 Planner LLM 的提示词 (获取模板并填充数据)"""
try:
chat_context_description = "你现在正在一个群聊中"
chat_target_name = None # Only relevant for private
if not is_group_chat and chat_target_info:
chat_target_name = (
chat_target_info.get("person_name") or chat_target_info.get("user_nickname") or "对方"
)
chat_context_description = f"你正在和 {chat_target_name} 私聊"
chat_content_block = ""
if observed_messages_str:
chat_content_block = f"\n{observed_messages_str}"
else:
chat_content_block = "你还未开始聊天"
action_options_block = ""
# 根据聊天类型选择不同的动作prompt模板
action_template_name = "action_prompt_private" if not is_group_chat else "action_prompt"
for using_actions_name, using_actions_info in current_available_actions.items():
using_action_prompt = await global_prompt_manager.get_prompt_async(action_template_name)
if using_actions_info["parameters"]:
param_text = "\n"
for param_name, param_description in using_actions_info["parameters"].items():
param_text += f' "{param_name}":"{param_description}"\n'
param_text = param_text.rstrip("\n")
else:
param_text = ""
require_text = ""
for require_item in using_actions_info["require"]:
require_text += f"- {require_item}\n"
require_text = require_text.rstrip("\n")
# 根据模板类型决定是否包含description参数
if action_template_name == "action_prompt_private":
# 私聊模板不包含description参数
using_action_prompt = using_action_prompt.format(
action_name=using_actions_name,
action_parameters=param_text,
action_require=require_text,
)
else:
# 群聊模板包含description参数
using_action_prompt = using_action_prompt.format(
action_name=using_actions_name,
action_description=using_actions_info["description"],
action_parameters=param_text,
action_require=require_text,
)
action_options_block += using_action_prompt
# moderation_prompt_block = "请不要输出违法违规内容，不要输出色情，暴力，政治相关内容，如有敏感内容，请规避。"
moderation_prompt_block = ""
# 获取当前时间
time_block = f"当前时间：{datetime.now().strftime('%Y-%m-%d %H:%M:%S')}"
bot_name = global_config.bot.nickname
if global_config.bot.alias_names:
bot_nickname = f",也有人叫你{','.join(global_config.bot.alias_names)}"
else:
bot_nickname = ""
bot_core_personality = global_config.personality.personality_core
indentify_block = f"你的名字是{bot_name}{bot_nickname}，你{bot_core_personality}："
# 根据聊天类型选择不同的prompt模板
template_name = "simple_planner_prompt_private" if not is_group_chat else "simple_planner_prompt"
planner_prompt_template = await global_prompt_manager.get_prompt_async(template_name)
prompt = planner_prompt_template.format(
time_block=time_block,
chat_context_description=chat_context_description,
chat_content_block=chat_content_block,
action_options_text=action_options_block,
moderation_prompt=moderation_prompt_block,
indentify_block=indentify_block,
)
return prompt
except Exception as e:
logger.error(f"构建 Planner 提示词时出错: {e}")
logger.error(traceback.format_exc())
return "构建 Planner Prompt 时出错"
init_prompt()

306
src/chat/planner_actions/planner_normal.py Normal file
View File

@@ -0,0 +1,306 @@
import json
from typing import Dict, Any
from rich.traceback import install
from src.llm_models.utils_model import LLMRequest
from src.config.config import global_config
from src.common.logger import get_logger
from src.chat.utils.prompt_builder import Prompt, global_prompt_manager
from src.individuality.individuality import get_individuality
from src.chat.planner_actions.action_manager import ActionManager
from src.chat.message_receive.message import MessageThinking
from json_repair import repair_json
from src.chat.utils.chat_message_builder import build_readable_messages, get_raw_msg_before_timestamp_with_chat
import time
import traceback
logger = get_logger("normal_chat_planner")
install(extra_lines=3)
def init_prompt():
Prompt(
"""
你的自我认知是：
{self_info_block}
请记住你的性格，身份和特点。
你是群内的一员，你现在正在参与群内的闲聊，以下是群内的聊天内容：
{chat_context}
基于以上聊天上下文和用户的最新消息，选择最合适的action。
注意，除了下面动作选项之外，你在聊天中不能做其他任何事情，这是你能力的边界，现在请你选择合适的action:
{action_options_text}
重要说明：
- "no_action" 表示只进行普通聊天回复，不执行任何额外动作
- 其他action表示在普通回复的基础上，执行相应的额外动作
你必须从上面列出的可用action中选择一个，并说明原因。
{moderation_prompt}
请以动作的输出要求，以严格的 JSON 格式输出，且仅包含 JSON 内容。不要有任何其他文字或解释：
""",
"normal_chat_planner_prompt",
)
Prompt(
"""
动作：{action_name}
动作描述：{action_description}
{action_require}
{{
"action": "{action_name}",{action_parameters}
}}
""",
"normal_chat_action_prompt",
)
class NormalChatPlanner:
def __init__(self, log_prefix: str, action_manager: ActionManager):
self.log_prefix = log_prefix
# LLM规划器配置
self.planner_llm = LLMRequest(
model=global_config.model.planner,
request_type="normal.planner", # 用于normal_chat动作规划
)
self.action_manager = action_manager
async def plan(self, message: MessageThinking) -> Dict[str, Any]:
"""
Normal Chat 规划器: 使用LLM根据上下文决定做出什么动作。
参数:
message: 思考消息对象
sender_name: 发送者名称
"""
action = "no_action" # 默认动作改为no_action
reasoning = "规划器初始化默认"
action_data = {}
try:
# 设置默认值
nickname_str = ""
for nicknames in global_config.bot.alias_names:
nickname_str += f"{nicknames},"
name_block = f"你的名字是{global_config.bot.nickname},你的昵称有{nickname_str}，有人也会用这些昵称称呼你。"
personality_block = get_individuality().get_personality_prompt(x_person=2, level=2)
identity_block = get_individuality().get_identity_prompt(x_person=2, level=2)
self_info = name_block + personality_block + identity_block
# 获取当前可用的动作，使用Normal模式过滤
current_available_actions = self.action_manager.get_using_actions_for_mode("normal")
# 注意：动作的激活判定现在在 normal_chat_action_modifier 中完成
# 这里直接使用经过 action_modifier 处理后的最终动作集
# 符合职责分离原则：ActionModifier负责动作管理，Planner专注于决策
# 如果没有可用动作，直接返回no_action
if not current_available_actions:
logger.debug(f"{self.log_prefix}规划器: 没有可用动作，返回no_action")
return {
"action_result": {
"action_type": action,
"action_data": action_data,
"reasoning": reasoning,
"is_parallel": True,
},
"chat_context": "",
"action_prompt": "",
}
# 构建normal_chat的上下文 (使用与normal_chat相同的prompt构建方法)
message_list_before_now = get_raw_msg_before_timestamp_with_chat(
chat_id=message.chat_stream.stream_id,
timestamp=time.time(),
limit=global_config.chat.max_context_size,
)
chat_context = build_readable_messages(
message_list_before_now,
replace_bot_name=True,
merge_messages=False,
timestamp_mode="relative",
read_mark=0.0,
show_actions=True,
)
# 构建planner的prompt
prompt = await self.build_planner_prompt(
self_info_block=self_info,
chat_context=chat_context,
current_available_actions=current_available_actions,
)
if not prompt:
logger.warning(f"{self.log_prefix}规划器: 构建提示词失败")
return {
"action_result": {
"action_type": action,
"action_data": action_data,
"reasoning": reasoning,
"is_parallel": False,
},
"chat_context": chat_context,
"action_prompt": "",
}
# 使用LLM生成动作决策
try:
content, (reasoning_content, model_name) = await self.planner_llm.generate_response_async(prompt)
logger.info(f"{self.log_prefix}规划器原始提示词: {prompt}")
logger.info(f"{self.log_prefix}规划器原始响应: {content}")
if reasoning_content:
logger.info(f"{self.log_prefix}规划器推理: {reasoning_content}")
# 解析JSON响应
try:
# 尝试修复JSON
fixed_json = repair_json(content)
action_result = json.loads(fixed_json)
action = action_result.get("action", "no_action")
reasoning = action_result.get("reasoning", "未提供原因")
# 提取其他参数作为action_data
action_data = {k: v for k, v in action_result.items() if k not in ["action", "reasoning"]}
# 验证动作是否在可用动作列表中，或者是特殊动作
if action not in current_available_actions:
logger.warning(f"{self.log_prefix}规划器选择了不可用的动作: {action}, 回退到no_action")
action = "no_action"
reasoning = f"选择的动作{action}不在可用列表中，回退到no_action"
action_data = {}
except json.JSONDecodeError as e:
logger.warning(f"{self.log_prefix}规划器JSON解析失败: {e}, 内容: {content}")
action = "no_action"
reasoning = "JSON解析失败，使用默认动作"
action_data = {}
except Exception as e:
logger.error(f"{self.log_prefix}规划器LLM调用失败: {e}")
action = "no_action"
reasoning = "LLM调用失败，使用默认动作"
action_data = {}
except Exception as outer_e:
logger.error(f"{self.log_prefix}规划器异常: {outer_e}")
# 设置异常时的默认值
current_available_actions = {}
chat_context = "无法获取聊天上下文"
prompt = ""
action = "no_action"
reasoning = "规划器出现异常，使用默认动作"
action_data = {}
# 检查动作是否支持并行执行
is_parallel = False
if action in current_available_actions:
action_info = current_available_actions[action]
is_parallel = action_info.get("parallel_action", False)
logger.debug(
f"{self.log_prefix}规划器决策动作:{action}, 动作信息: '{action_data}', 理由: {reasoning}, 并行执行: {is_parallel}"
)
# 恢复到默认动作集
self.action_manager.restore_actions()
logger.debug(
f"{self.log_prefix}规划后恢复到默认动作集, 当前可用: {list(self.action_manager.get_using_actions().keys())}"
)
# 构建 action 记录
action_record = {
"action_type": action,
"action_data": action_data,
"reasoning": reasoning,
"timestamp": time.time(),
"model_name": model_name if "model_name" in locals() else None,
}
action_result = {
"action_type": action,
"action_data": action_data,
"reasoning": reasoning,
"is_parallel": is_parallel,
"action_record": json.dumps(action_record, ensure_ascii=False),
}
plan_result = {
"action_result": action_result,
"chat_context": chat_context,
"action_prompt": prompt,
}
return plan_result
async def build_planner_prompt(
self,
self_info_block: str,
chat_context: str,
current_available_actions: Dict[str, Any],
) -> str:
"""构建 Normal Chat Planner LLM 的提示词"""
try:
# 构建动作选项文本
action_options_text = ""
for action_name, action_info in current_available_actions.items():
action_description = action_info.get("description", "")
action_parameters = action_info.get("parameters", {})
action_require = action_info.get("require", [])
if action_parameters:
param_text = "\n"
# print(action_parameters)
for param_name, param_description in action_parameters.items():
param_text += f' "{param_name}":"{param_description}"\n'
param_text = param_text.rstrip("\n")
else:
param_text = ""
require_text = ""
for require_item in action_require:
require_text += f"- {require_item}\n"
require_text = require_text.rstrip("\n")
# 构建单个动作的提示
action_prompt = await global_prompt_manager.format_prompt(
"normal_chat_action_prompt",
action_name=action_name,
action_description=action_description,
action_parameters=param_text,
action_require=require_text,
)
action_options_text += action_prompt + "\n\n"
# 审核提示
moderation_prompt = "请确保你的回复符合平台规则，避免不当内容。"
# 使用模板构建最终提示词
prompt = await global_prompt_manager.format_prompt(
"normal_chat_planner_prompt",
self_info_block=self_info_block,
action_options_text=action_options_text,
moderation_prompt=moderation_prompt,
chat_context=chat_context,
)
return prompt
except Exception as e:
logger.error(f"{self.log_prefix}构建Planner提示词失败: {e}")
traceback.print_exc()
return ""
init_prompt()