Losita/mai-bot
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat：新增sub_agent，暂时用于emoji选择；修改部分配置位置

Browse Source
This commit is contained in:
SengokuCola
2026-04-03 21:57:35 +08:00
parent ce580d1f8b
commit 07597bc1d7
9 changed files with 300 additions and 88 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
src/chat/emoji_system/maisaka_tool.py
View File

@@ -1,5 +1,6 @@
"""Maisaka 表情工具内置能力。"""
from collections.abc import Awaitable, Callable
from dataclasses import dataclass, field
from typing import Any, Optional, Sequence
@@ -17,6 +18,11 @@ from .emoji_manager import _serialize_emoji_for_hook, emoji_manager, emoji_manag
logger = get_logger("emoji_maisaka_tool")
EmojiSelector = Callable[
[str, str, Sequence[str] | None, int],
Awaitable[tuple[MaiEmoji | None, str]],
]
@dataclass(slots=True)
class MaisakaEmojiSendResult:
@@ -198,13 +204,14 @@ async def send_emoji_for_maisaka(
requested_emotion: str = "",
reasoning: str = "",
context_texts: Sequence[str] | None = None,
emoji_selector: EmojiSelector | None = None,
) -> MaisakaEmojiSendResult:
"""为 Maisaka 选择并发送一个表情。"""
normalized_requested_emotion = requested_emotion.strip()
normalized_reasoning = reasoning.strip()
normalized_context_texts = _normalize_context_texts(context_texts)
sample_size = 30
sample_size = 20
before_select_result = await _get_runtime_manager().invoke_hook(
"emoji.maisaka.before_select",
@@ -232,12 +239,20 @@ async def send_emoji_for_maisaka(
normalized_context_texts = _normalize_context_texts(before_select_kwargs.get("context_texts"))
sample_size = _coerce_positive_int(before_select_kwargs.get("sample_size"), sample_size)
selected_emoji, matched_emotion = await select_emoji_for_maisaka(
requested_emotion=normalized_requested_emotion,
reasoning=normalized_reasoning,
context_texts=normalized_context_texts,
sample_size=sample_size,
)
if emoji_selector is None:
selected_emoji, matched_emotion = await select_emoji_for_maisaka(
requested_emotion=normalized_requested_emotion,
reasoning=normalized_reasoning,
context_texts=normalized_context_texts,
sample_size=sample_size,
)
else:
selected_emoji, matched_emotion = await emoji_selector(
normalized_requested_emotion,
normalized_reasoning,
normalized_context_texts,
sample_size,
)
after_select_result = await _get_runtime_manager().invoke_hook(
"emoji.maisaka.after_select",
stream_id=stream_id,

2
src/chat/replyer/maisaka_generator.py
View File

@@ -151,7 +151,7 @@ class MaisakaReplyGenerator:
content = self._normalize_content(content_body)
if not content:
continue
visible_speaker = speaker_name or global_config.maisaka.user_name.strip() or "User"
visible_speaker = speaker_name or global_config.maisaka.cli_user_name.strip() or "User"
parts.append(f"{timestamp} {visible_speaker}: {content}")
continue

2
src/chat/replyer/maisaka_generator_multi.py
View File

@@ -162,7 +162,7 @@ class MaisakaReplyGenerator:
def _build_history_messages(self, chat_history: List[LLMContextMessage]) -> List[Message]:
"""将 replyer 上下文拆成多条 LLM 消息。"""
bot_nickname = global_config.bot.nickname.strip() or "Bot"
default_user_name = global_config.maisaka.user_name.strip() or "User"
default_user_name = global_config.maisaka.cli_user_name.strip() or "User"
messages: List[Message] = []
for message in chat_history:

2
src/cli/maisaka_cli.py
View File

@@ -67,7 +67,7 @@ class BufferCLI:
timestamp=timestamp,
platform=BufferCLI._CLI_PLATFORM,
)
user_name = global_config.maisaka.user_name.strip() or "用户"
user_name = global_config.maisaka.cli_user_name.strip() or "用户"
message.message_info = MessageInfo(
user_info=UserInfo(
user_id=BufferCLI._CLI_USER_ID,

73
src/config/official_configs.py
View File

@@ -282,7 +282,24 @@ class ChatConfig(ConfigBase):
"x-icon": "list",
},
)
"""_wrap_为指定聊天添加额外的 prompt 配置列表"""
direct_image_input: bool = Field(
default=True,
json_schema_extra={
"x-widget": "switch",
"x-icon": "image",
},
)
"""是否直接输入图片"""
replyer_generator_type: Literal["legacy", "multi"] = Field(
default="legacy",
json_schema_extra={
"x-widget": "select",
"x-icon": "git-branch",
},
)
"""Maisaka replyer 生成器类型：legacy（旧版单 prompt）/ multi（多消息版）"""
enable_talk_value_rules: bool = Field(
default=True,
@@ -964,6 +981,14 @@ class DebugConfig(ConfigBase):
"x-icon": "brain",
},
)
show_maisaka_thinking: bool = Field(
default=True,
json_schema_extra={
"x-widget": "switch",
"x-icon": "brain",
},
)
"""是否显示回复器推理"""
show_jargon_prompt: bool = Field(
@@ -1427,16 +1452,7 @@ class MaiSakaConfig(ConfigBase):
},
)
"""启用知识库模块"""
show_thinking: bool = Field(
default=True,
json_schema_extra={
"x-widget": "switch",
"x-icon": "brain",
},
)
"""是否显示MaiSaka思考过程"""
user_name: str = Field(
cli_user_name: str = Field(
default="用户",
json_schema_extra={
"x-widget": "input",
@@ -1445,33 +1461,6 @@ class MaiSakaConfig(ConfigBase):
)
"""MaiSaka 使用的用户名称"""
direct_image_input: bool = Field(
default=True,
json_schema_extra={
"x-widget": "switch",
"x-icon": "image",
},
)
"""是否直接输入图片"""
merge_user_messages: bool = Field(
default=True,
json_schema_extra={
"x-widget": "switch",
"x-icon": "merge",
},
)
"""是否将新接收的用户发言合并为单个用户消息"""
replyer_generator_type: Literal["legacy", "multi"] = Field(
default="legacy",
json_schema_extra={
"x-widget": "select",
"x-icon": "git-branch",
},
)
"""Maisaka replyer 生成器类型：legacy（旧版单 prompt）/ multi（多消息版）"""
max_internal_rounds: int = Field(
default=6,
ge=1,
@@ -1511,14 +1500,14 @@ class MaiSakaConfig(ConfigBase):
)
"""工具筛选阶段最多保留的非内置工具数量"""
terminal_image_display_mode: Literal["legacy", "path_link"] = Field(
default="legacy",
show_image_path: bool = Field(
default=True,
json_schema_extra={
"x-widget": "select",
"x-widget": "switch",
"x-icon": "image",
},
)
"""图片展示模式：legacy（仅显示元信息）/ path_link（可点击本地路径）"""
"""是否显示图片本地路径"""
class MCPAuthorizationConfig(ConfigBase):

153
src/maisaka/builtin_tool/send_emoji.py
View File

@@ -1,16 +1,40 @@
"""send_emoji 内置工具。"""
from datetime import datetime
from random import sample
from secrets import token_hex
from typing import Any, Dict, Optional
import asyncio
from pydantic import BaseModel, Field as PydanticField
from src.chat.emoji_system.emoji_manager import emoji_manager
from src.chat.emoji_system.maisaka_tool import send_emoji_for_maisaka
from src.common.data_models.message_component_data_model import ImageComponent, MessageSequence, TextComponent
from src.common.data_models.image_data_model import MaiEmoji
from src.common.logger import get_logger
from src.core.tooling import ToolExecutionContext, ToolExecutionResult, ToolInvocation, ToolSpec
from src.maisaka.context_messages import LLMContextMessage
from src.llm_models.payload_content.resp_format import RespFormat, RespFormatType
from src.maisaka.context_messages import LLMContextMessage, ReferenceMessage, ReferenceMessageType, SessionBackedMessage
from .context import BuiltinToolRuntimeContext
logger = get_logger("maisaka_builtin_send_emoji")
_EMOJI_SUB_AGENT_CONTEXT_LIMIT = 12
_EMOJI_SUB_AGENT_MAX_TOKENS = 240
_EMOJI_SUB_AGENT_SAMPLE_SIZE = 20
_EMOJI_SUCCESS_MESSAGE = "???????"
class EmojiSelectionResult(BaseModel):
"""表情包子代理的结构化选择结果。"""
emoji_id: str = PydanticField(default="", description="选中的候选表情包 ID。")
matched_emotion: str = PydanticField(default="", description="本次命中的情绪标签，可为空。")
reason: str = PydanticField(default="", description="简短选择理由。")
def get_tool_spec() -> ToolSpec:
"""获取 send_emoji 工具声明。"""
@@ -33,6 +57,105 @@ def get_tool_spec() -> ToolSpec:
)
async def _build_emoji_candidate_message(emoji: MaiEmoji, candidate_id: str) -> SessionBackedMessage:
"""构建供子代理挑选的图片候选消息。"""
image_bytes = await asyncio.to_thread(emoji.full_path.read_bytes)
raw_message = MessageSequence(
[
TextComponent(f"ID: {candidate_id}"),
ImageComponent(binary_hash=str(emoji.file_hash or ""), binary_data=image_bytes),
]
)
return SessionBackedMessage(
raw_message=raw_message,
visible_text=f"ID: {candidate_id}",
timestamp=datetime.now(),
source_kind="emoji_candidate",
)
async def _select_emoji_with_sub_agent(
tool_ctx: BuiltinToolRuntimeContext,
requested_emotion: str,
reasoning: str,
context_texts: list[str],
sample_size: int,
) -> tuple[MaiEmoji | None, str]:
"""通过临时子代理从候选表情包中选出一个结果。"""
available_emojis = list(emoji_manager.emojis)
if not available_emojis:
return None, ""
effective_sample_size = min(max(sample_size, 1), _EMOJI_SUB_AGENT_SAMPLE_SIZE, len(available_emojis))
sampled_emojis = sample(available_emojis, effective_sample_size)
candidate_map: dict[str, MaiEmoji] = {}
candidate_messages: list[LLMContextMessage] = []
for emoji in sampled_emojis:
candidate_id = token_hex(4)
while candidate_id in candidate_map:
candidate_id = token_hex(4)
candidate_map[candidate_id] = emoji
candidate_messages.append(await _build_emoji_candidate_message(emoji, candidate_id))
context_text = "\n".join(context_texts[-5:]) if context_texts else "（暂无额外上下文）"
system_prompt = (
"你是 Maisaka 的临时表情包选择子代理。\n"
"你会收到一段群聊上下文，以及若干条候选表情包消息。每条候选消息里都有一个临时 ID。\n"
"你的任务是根据上下文、当前语气和发送意图，从候选里选出最合适的一个表情包。\n"
"必须只从候选消息中选择，不能编造新的 ID。\n"
"如果提供了 requested_emotion，请优先考虑与其接近的候选；如果没有完全匹配，则选择最符合上下文语气的候选。\n"
"你必须返回一个 JSON 对象（json object），不要输出任何 JSON 之外的内容。\n"
'返回格式固定为：{"emoji_id":"候选ID","matched_emotion":"情绪标签","reason":"简短理由"}'
)
prompt_message = ReferenceMessage(
content=(
f"[选择任务]\n"
f"requested_emotion: {requested_emotion or '未指定'}\n"
f"reasoning: {reasoning or '辅助表达当前语气和情绪'}\n"
f"recent_context:\n{context_text}\n"
'请只输出 JSON。'
),
timestamp=datetime.now(),
reference_type=ReferenceMessageType.TOOL_HINT,
remaining_uses_value=1,
display_prefix="[表情包选择任务]",
)
response = await tool_ctx.runtime.run_sub_agent(
context_message_limit=_EMOJI_SUB_AGENT_CONTEXT_LIMIT,
system_prompt=system_prompt,
extra_messages=[prompt_message, *candidate_messages],
max_tokens=_EMOJI_SUB_AGENT_MAX_TOKENS,
response_format=RespFormat(
format_type=RespFormatType.JSON_SCHEMA,
schema=EmojiSelectionResult,
),
)
try:
selection = EmojiSelectionResult.model_validate_json(response.content or "")
except Exception as exc:
logger.warning(f"{tool_ctx.runtime.log_prefix} 表情包子代理结果解析失败，将回退到候选首项: {exc}")
fallback_emoji = sampled_emojis[0] if sampled_emojis else None
return fallback_emoji, requested_emotion
selected_emoji = candidate_map.get(selection.emoji_id.strip())
if selected_emoji is None:
logger.warning(
f"{tool_ctx.runtime.log_prefix} 表情包子代理返回了无效 ID: {selection.emoji_id!r}，将回退到候选首项"
)
fallback_emoji = sampled_emojis[0] if sampled_emojis else None
return fallback_emoji, requested_emotion
matched_emotion = selection.matched_emotion.strip()
if not matched_emotion:
matched_emotion = requested_emotion.strip()
return selected_emoji, matched_emotion
async def handle_tool(
tool_ctx: BuiltinToolRuntimeContext,
invocation: ToolInvocation,
@@ -64,6 +187,13 @@ async def handle_tool(
requested_emotion=emotion,
reasoning=tool_ctx.engine.last_reasoning_content,
context_texts=context_texts,
emoji_selector=lambda requested_emotion, reasoning, context_texts, sample_size: _select_emoji_with_sub_agent(
tool_ctx,
requested_emotion,
reasoning,
list(context_texts or []),
sample_size,
),
)
except Exception as exc:
logger.exception(f"{tool_ctx.runtime.log_prefix} 发送表情包时发生异常: {exc}")
@@ -74,28 +204,29 @@ async def handle_tool(
structured_content=structured_result,
)
structured_result["description"] = send_result.description
structured_result["emotion"] = list(send_result.emotions)
structured_result["matched_emotion"] = send_result.matched_emotion
structured_result["message"] = send_result.message
if send_result.success:
structured_result["message"] = _EMOJI_SUCCESS_MESSAGE
logger.info(
f"{tool_ctx.runtime.log_prefix} 表情包发送成功 "
f"描述={send_result.description!r} 情绪标签={send_result.emotions} "
f"请求情绪={emotion!r} 命中情绪={send_result.matched_emotion!r}"
f"{tool_ctx.runtime.log_prefix} ??????? "
f"??={send_result.description!r} ????={send_result.emotions} "
f"????={emotion!r} ????={send_result.matched_emotion!r}"
)
tool_ctx.append_sent_emoji_to_chat_history(
emoji_base64=send_result.emoji_base64,
success_message=send_result.message,
success_message=_EMOJI_SUCCESS_MESSAGE,
)
structured_result["success"] = True
return tool_ctx.build_success_result(
invocation.tool_name,
send_result.message,
_EMOJI_SUCCESS_MESSAGE,
structured_content=structured_result,
)
structured_result["description"] = send_result.description
structured_result["emotion"] = list(send_result.emotions)
structured_result["matched_emotion"] = send_result.matched_emotion
structured_result["message"] = send_result.message
logger.warning(
f"{tool_ctx.runtime.log_prefix} 表情包发送失败 "
f"请求情绪={emotion!r} 错误信息={send_result.message}"

69
src/maisaka/chat_loop_service.py
View File

@@ -210,7 +210,7 @@ class MaisakaChatLoopService:
self._extra_tools: List[ToolOption] = []
self._interrupt_flag: asyncio.Event | None = None
self._tool_registry: ToolRegistry | None = None
self._prompts_loaded = False
self._prompts_loaded = chat_system_prompt is not None
self._prompt_load_lock = asyncio.Lock()
self._personality_prompt = self._build_personality_prompt()
if chat_system_prompt is None:
@@ -392,7 +392,12 @@ class MaisakaChatLoopService:
"""设置当前 planner 请求使用的中断标记。"""
self._interrupt_flag = interrupt_flag
def _build_request_messages(self, selected_history: List[LLMContextMessage]) -> List[Message]:
def _build_request_messages(
self,
selected_history: List[LLMContextMessage],
*,
system_prompt: Optional[str] = None,
) -> List[Message]:
"""构造发给大模型的消息列表。
Args:
@@ -404,7 +409,7 @@ class MaisakaChatLoopService:
messages: List[Message] = []
system_msg = MessageBuilder().set_role(RoleType.System)
system_msg.add_text_content(self._chat_system_prompt)
system_msg.add_text_content(system_prompt if system_prompt is not None else self._chat_system_prompt)
messages.append(system_msg.build())
for msg in selected_history:
@@ -691,7 +696,13 @@ class MaisakaChatLoopService:
return extract_category_ids_from_result(generation_result.response or "")
async def chat_loop_step(self, chat_history: List[LLMContextMessage]) -> ChatResponse:
async def chat_loop_step(
self,
chat_history: List[LLMContextMessage],
*,
response_format: RespFormat | None = None,
tool_definitions: Sequence[ToolDefinitionInput] | None = None,
) -> ChatResponse:
"""执行一轮 Maisaka 规划器请求。
Args:
@@ -701,8 +712,9 @@ class MaisakaChatLoopService:
ChatResponse: 本轮规划器返回结果。
"""
await self.ensure_chat_prompt_loaded()
selected_history, selection_reason = self._select_llm_context_messages(chat_history)
if not self._prompts_loaded:
await self.ensure_chat_prompt_loaded()
selected_history, selection_reason = self.select_llm_context_messages(chat_history)
built_messages = self._build_request_messages(selected_history)
def message_factory(_client: BaseClient) -> List[Message]:
@@ -719,7 +731,9 @@ class MaisakaChatLoopService:
return built_messages
all_tools: List[ToolDefinitionInput]
if self._tool_registry is not None:
if tool_definitions is not None:
all_tools = list(tool_definitions)
elif self._tool_registry is not None:
tool_specs = await self._tool_registry.list_tools()
filtered_tool_specs = await self._filter_tool_specs_for_planner(selected_history, tool_specs)
all_tools = [tool_spec.to_llm_definition() for tool_spec in filtered_tool_specs]
@@ -748,10 +762,10 @@ class MaisakaChatLoopService:
ordered_panels = PromptCLIVisualizer.build_prompt_panels(
built_messages,
image_display_mode=global_config.maisaka.terminal_image_display_mode,
image_display_mode="path_link" if global_config.maisaka.show_image_path else "legacy",
)
if global_config.maisaka.show_thinking and ordered_panels:
if global_config.debug.show_maisaka_thinking and ordered_panels:
console.print(
Panel(
Group(*ordered_panels),
@@ -776,6 +790,7 @@ class MaisakaChatLoopService:
tool_options=all_tools if all_tools else None,
temperature=self._temperature,
max_tokens=self._max_tokens,
response_format=response_format,
interrupt_flag=self._interrupt_flag,
),
)
@@ -837,6 +852,40 @@ class MaisakaChatLoopService:
total_tokens=total_tokens,
)
@staticmethod
def select_llm_context_messages(
chat_history: List[LLMContextMessage],
*,
max_context_size: Optional[int] = None,
) -> tuple[List[LLMContextMessage], str]:
"""??????? LLM ???????"""
effective_context_size = max(1, int(max_context_size or global_config.chat.max_context_size))
selected_indices: List[int] = []
counted_message_count = 0
for index in range(len(chat_history) - 1, -1, -1):
message = chat_history[index]
if message.to_llm_message() is None:
continue
selected_indices.append(index)
if message.count_in_context:
counted_message_count += 1
if counted_message_count >= effective_context_size:
break
if not selected_indices:
return [], f"???????? {effective_context_size} ? user/assistant??? 0 ??"
selected_indices.reverse()
selected_history = [chat_history[index] for index in selected_indices]
selected_history = MaisakaChatLoopService._drop_leading_orphan_tool_results(selected_history)
return (
selected_history,
f"???????? {effective_context_size} ? user/assistant??????????? {len(selected_history)} ?",
)
@staticmethod
def _select_llm_context_messages(chat_history: List[LLMContextMessage]) -> tuple[List[LLMContextMessage], str]:
"""选择真正发送给 LLM 的上下文消息。
@@ -905,4 +954,4 @@ class MaisakaChatLoopService:
if first_valid_index == 0:
return selected_history
return selected_history[first_valid_index:]
return selected_history[first_valid_index:]

14
src/maisaka/reasoning_engine.py
View File

@@ -266,7 +266,7 @@ class MaisakaReasoningEngine:
source_sequence = message.raw_message
planner_components = clone_message_sequence(source_sequence).components
if global_config.maisaka.direct_image_input:
if global_config.chat.direct_image_input:
await self._hydrate_visual_components(planner_components)
if planner_components and isinstance(planner_components[0], TextComponent):
planner_components[0].text = planner_prefix + planner_components[0].text
@@ -610,16 +610,8 @@ class MaisakaReasoningEngine:
return f"你尝试回复消息 {target_message_id or 'unknown'}，但失败了：{error_text}"
if invocation.tool_name == "send_emoji":
description = str(structured_content.get("description") or "").strip()
emotion_list = structured_content.get("emotion")
if isinstance(emotion_list, list):
emotion_text = "、".join(str(item).strip() for item in emotion_list if str(item).strip())
else:
emotion_text = ""
if result.success and description:
if emotion_text:
return f"你发送了表情包：{description}（情绪：{emotion_text}）"
return f"你发送了表情包：{description}"
if result.success:
return "你发送了表情包。"
return f"你尝试发送表情包，但失败了：{self._truncate_tool_record_text(result.error_message or history_content, 120)}"
if invocation.tool_name == "wait":

44
src/maisaka/runtime.py
View File

@@ -1,6 +1,6 @@
"""Maisaka 非 CLI 运行时。"""
from typing import Any, Literal, Optional
from typing import Any, Literal, Optional, Sequence
import asyncio
import time
@@ -20,12 +20,14 @@ from src.core.tooling import ToolRegistry
from src.know_u.knowledge import KnowledgeLearner
from src.learners.expression_learner import ExpressionLearner
from src.learners.jargon_miner import JargonMiner
from src.llm_models.payload_content.resp_format import RespFormat
from src.llm_models.payload_content.tool_option import ToolDefinitionInput
from src.mcp_module import MCPManager
from src.mcp_module.host_llm_bridge import MCPHostLLMBridge
from src.mcp_module.provider import MCPToolProvider
from src.plugin_runtime.tool_provider import PluginToolProvider
from .chat_loop_service import MaisakaChatLoopService
from .chat_loop_service import ChatResponse, MaisakaChatLoopService
from .context_messages import LLMContextMessage
from .reasoning_engine import MaisakaReasoningEngine
from .tool_provider import MaisakaBuiltinToolProvider
@@ -197,6 +199,40 @@ class MaisakaHeartFlowChatting:
self._tool_registry.register_provider(PluginToolProvider())
self._chat_loop_service.set_tool_registry(self._tool_registry)
async def run_sub_agent(
self,
*,
context_message_limit: int,
system_prompt: str,
extra_messages: Optional[Sequence[LLMContextMessage]] = None,
max_tokens: int = 512,
response_format: RespFormat | None = None,
temperature: float = 0.2,
tool_definitions: Optional[Sequence[ToolDefinitionInput]] = None,
) -> ChatResponse:
"""运行一个复制上下文的临时子代理，并在完成后立即销毁。"""
selected_history, _ = MaisakaChatLoopService.select_llm_context_messages(
self._chat_history,
max_context_size=context_message_limit,
)
sub_agent_history = list(selected_history)
if extra_messages:
sub_agent_history.extend(list(extra_messages))
sub_agent = MaisakaChatLoopService(
chat_system_prompt=system_prompt,
session_id=self.session_id,
is_group_chat=self.chat_stream.is_group_session,
temperature=temperature,
max_tokens=max_tokens,
)
return await sub_agent.chat_loop_step(
sub_agent_history,
response_format=response_format,
tool_definitions=[] if tool_definitions is None else tool_definitions,
)
async def _main_loop(self) -> None:
try:
while self._running:
@@ -421,7 +457,7 @@ class MaisakaHeartFlowChatting:
if self.chat_stream.user_id:
return UserInfo(
user_id=self.chat_stream.user_id,
user_nickname=global_config.maisaka.user_name.strip() or "用户",
user_nickname=global_config.maisaka.cli_user_name.strip() or "用户",
user_cardname=None,
)
return UserInfo(user_id="maisaka_user", user_nickname="用户", user_cardname=None)
@@ -455,7 +491,7 @@ class MaisakaHeartFlowChatting:
tool_results: Optional[list[str]] = None,
) -> None:
"""在终端展示当前聊天流的上下文占用、规划结果与工具摘要。"""
if not global_config.maisaka.show_thinking:
if not global_config.debug.show_maisaka_thinking:
return
session_name = chat_manager.get_session_name(self.session_id) or self.session_id