diff --git a/README.md b/README.md index a3f2272..5892354 100644 --- a/README.md +++ b/README.md @@ -123,6 +123,7 @@ _✨ 支持多API预设、MCP协议、内置工具、联网搜索、视觉模型 | LLMCHAT__DEFAULT_PRESET | 否 | off | 默认使用的预设名称，配置为off则为关闭 | | LLMCHAT__RANDOM_TRIGGER_PROB | 否 | 0.05 | 默认随机触发概率 [0, 1] | | LLMCHAT__DEFAULT_PROMPT | 否 | 你的回答应该尽量简洁、幽默、可以使用一些语气词、颜文字。你应该拒绝回答任何政治相关的问题。 | 默认提示词 | +| LLMCHAT__MCP_SERVER_CWD | 否 | 无 | command类型MCP服务器全局工作目录（cwd） | | LLMCHAT__BLACKLIST_USER_IDS | 否 | [] | 黑名单用户ID列表，机器人将不会处理黑名单用户的消息 | | LLMCHAT__IGNORE_PREFIXES | 否 | [] | 需要忽略的消息前缀列表，匹配到这些前缀的消息不会处理 | | LLMCHAT__MCP_SERVERS | 否 | {} | MCP服务器配置，具体见下表 | @@ -165,7 +166,7 @@ LLMCHAT__MCP_SERVERS同样为一个dict，key为服务器名称，value配置的 | 配置项 | 必填 | 默认值 | 说明 | |:-----:|:----:|:----:|:----:| | command | stdio服务器必填 | 无 | stdio服务器MCP命令 | -| arg | 否 | [] | stdio服务器MCP命令参数 | +| args | 否 | [] | stdio服务器MCP命令参数 | | env | 否 | {} | stdio服务器环境变量 | | url | sse服务器必填 | 无 | sse服务器地址 | | headers | 否 | {} | sse模式下http请求头，用于认证或其他设置 | diff --git a/nonebot_plugin_llmchat/__init__.py b/nonebot_plugin_llmchat/__init__.py index cbca006..af6f315 100755 --- a/nonebot_plugin_llmchat/__init__.py +++ b/nonebot_plugin_llmchat/__init__.py @@ -349,261 +349,266 @@ async def process_messages(context_id: int, is_group: bool = True): logger.info( f"开始处理{chat_type}消息 {context_type}：{context_id} 当前队列长度：{state.queue.qsize()}" ) - while not state.queue.empty(): - event = await state.queue.get() - if is_group: - logger.debug(f"从队列获取消息 群号：{context_id} 消息ID：{event.message_id}") - group_id = context_id - else: - logger.debug(f"从队列获取消息 用户：{context_id} 消息ID：{event.message_id}") - group_id = None - past_events_snapshot = [] - mcp_client = MCPClient.get_instance(plugin_config.mcp_servers) - try: - # 构建系统提示，分成多行以满足行长限制 - chat_type = "群聊" if is_group else "私聊" - bot_names = "、".join(list(driver.config.nickname)) - default_prompt = (state.group_prompt) or plugin_config.default_prompt - - system_lines = [ - f"我想要你帮我在{chat_type}中闲聊，大家一般叫你{bot_names}。", - "我将会在后面的信息中告诉你每条信息的发送者和发送时间，你可以直接称呼发送者为他对应的昵称。", - "你的回复需要遵守以下几点规则：", - "- 你可以使用多条消息回复，每两条消息之间使用分隔，前后不需要包含额外的换行和空格。", - "- 除外，消息中不应该包含其他类似的标记。", - "- 不要使用markdown或者html，聊天软件不支持解析，换行请用换行符。", - "- 你应该以普通人的方式发送消息，每条消息字数要尽量少一些，应该倾向于使用更多条的消息回复。", - "- 代码则不需要分段，用单独的一条消息发送。", - "- 请使用发送者的昵称称呼发送者，你可以礼貌地问候发送者，但只需要在" - "第一次回答这位发送者的问题时问候他。", - "- 你有引用某条消息的能力，使用[CQ:reply,id=（消息id）]来引用。", - "- 如果有多条消息，你应该优先回复提到你的，一段时间之前的就不要回复了，也可以直接选择不回复。", - "- 如果你选择完全不回复，你只需要直接输出一个。", - "- 如果你需要思考的话，你应该尽量少思考，以节省时间。", - ] - + try: + while not state.queue.empty(): + event = await state.queue.get() if is_group: - system_lines += [ - "- 你有at群成员的能力，只需要在某条消息中插入[CQ:at,qq=（QQ号）]，" - "也就是CQ码。at发送者是非必要的，你可以根据你自己的想法at某个人。", + logger.debug(f"从队列获取消息 群号：{context_id} 消息ID：{event.message_id}") + group_id = context_id + else: + logger.debug(f"从队列获取消息 用户：{context_id} 消息ID：{event.message_id}") + group_id = None + past_events_snapshot = [] + mcp_client = MCPClient.get_instance( + plugin_config.mcp_servers, + plugin_config.mcp_server_cwd, + ) + try: + # 构建系统提示，分成多行以满足行长限制 + chat_type = "群聊" if is_group else "私聊" + bot_names = "、".join(list(driver.config.nickname)) + default_prompt = (state.group_prompt) or plugin_config.default_prompt + + system_lines = [ + f"我想要你帮我在{chat_type}中闲聊，大家一般叫你{bot_names}。", + "我将会在后面的信息中告诉你每条信息的发送者和发送时间，你可以直接称呼发送者为他对应的昵称。", + "你的回复需要遵守以下几点规则：", + "- 你可以使用多条消息回复，每两条消息之间使用分隔，前后不需要包含额外的换行和空格。", + "- 除外，消息中不应该包含其他类似的标记。", + "- 不要使用markdown或者html，聊天软件不支持解析，换行请用换行符。", + "- 你应该以普通人的方式发送消息，每条消息字数要尽量少一些，应该倾向于使用更多条的消息回复。", + "- 代码则不需要分段，用单独的一条消息发送。", + "- 请使用发送者的昵称称呼发送者，你可以礼貌地问候发送者，但只需要在" + "第一次回答这位发送者的问题时问候他。", + "- 你有引用某条消息的能力，使用[CQ:reply,id=（消息id）]来引用。", + "- 如果有多条消息，你应该优先回复提到你的，一段时间之前的就不要回复了，也可以直接选择不回复。", + "- 如果你选择完全不回复，你只需要直接输出一个。", + "- 如果你需要思考的话，你应该尽量少思考，以节省时间。", ] - system_lines += [ - "下面是关于你性格的设定，如果设定中提到让你扮演某个人，或者设定中有提到名字，则优先使用设定中的名字。", - default_prompt, - ] + if is_group: + system_lines += [ + "- 你有at群成员的能力，只需要在某条消息中插入[CQ:at,qq=（QQ号）]，" + "也就是CQ码。at发送者是非必要的，你可以根据你自己的想法at某个人。", + ] - systemPrompt = "\n".join(system_lines) - if preset.support_mcp: - systemPrompt += "\n你也可以使用一些工具，下面是关于这些工具的额外说明：\n" - for mcp_name, mcp_config in plugin_config.mcp_servers.items(): - if mcp_config.additional_prompt: - systemPrompt += f"{mcp_name}：{mcp_config.additional_prompt}" - systemPrompt += "\n" + system_lines += [ + "下面是关于你性格的设定，如果设定中提到让你扮演某个人，或者设定中有提到名字，则优先使用设定中的名字。", + default_prompt, + ] - logger.debug(f"构建系统提示词：\n{systemPrompt}") + systemPrompt = "\n".join(system_lines) + if preset.support_mcp: + systemPrompt += "\n你也可以使用一些工具，下面是关于这些工具的额外说明：\n" + for mcp_name, mcp_config in plugin_config.mcp_servers.items(): + if mcp_config.additional_prompt: + systemPrompt += f"{mcp_name}：{mcp_config.additional_prompt}" + systemPrompt += "\n" - messages: list[ChatCompletionMessageParam] = [ - {"role": "system", "content": systemPrompt} - ] + logger.debug(f"构建系统提示词：\n{systemPrompt}") - while len(state.history) > 0 and state.history[0]["role"] != "user": - state.history.popleft() + messages: list[ChatCompletionMessageParam] = [ + {"role": "system", "content": systemPrompt} + ] - messages += list(state.history)[-plugin_config.history_size * 2 :] + while len(state.history) > 0 and state.history[0]["role"] != "user": + state.history.popleft() - # 没有未处理的消息说明已经被处理了，跳过 - if state.past_events.__len__() < 1: - break + messages += list(state.history)[-plugin_config.history_size * 2 :] - content: list[ChatCompletionContentPartParam] = [] + # 没有未处理的消息说明已经被处理了，跳过 + if state.past_events.__len__() < 1: + break - # 将机器人错过的消息推送给LLM - past_events_snapshot = list(state.past_events) - state.past_events.clear() - for ev in past_events_snapshot: - text_content = format_message(ev) - content.append({"type": "text", "text": text_content}) + content: list[ChatCompletionContentPartParam] = [] - # 将消息中的图片转成 base64 - if preset.support_image: - base64_images = await process_images(ev) - for base64_image in base64_images: - content.append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}}) + # 将机器人错过的消息推送给LLM + past_events_snapshot = list(state.past_events) + state.past_events.clear() + for ev in past_events_snapshot: + text_content = format_message(ev) + content.append({"type": "text", "text": text_content}) - new_messages: list[ChatCompletionMessageParam] = [ - {"role": "user", "content": content} - ] + # 将消息中的图片转成 base64 + if preset.support_image: + base64_images = await process_images(ev) + for base64_image in base64_images: + content.append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}}) - logger.debug( - f"发送API请求 模型：{preset.model_name} 历史消息数：{len(messages)}" - ) + new_messages: list[ChatCompletionMessageParam] = [ + {"role": "user", "content": content} + ] - client_config = { - "model": preset.model_name, - "max_tokens": preset.max_tokens, - "temperature": preset.temperature, - "timeout": 60, - "extra_body": preset.extra_body, - } + logger.debug( + f"发送API请求 模型：{preset.model_name} 历史消息数：{len(messages)}" + ) - if preset.support_mcp: - available_tools = await mcp_client.get_available_tools(is_group) - client_config["tools"] = available_tools - - response = await client.chat.completions.create( - **client_config, - messages=messages + new_messages, - ) - - if response.usage is not None: - logger.debug(f"收到API响应 使用token数：{response.usage.total_tokens}") - - message = response.choices[0].message - - # 处理响应并处理工具调用 - while preset.support_mcp and message and message.tool_calls: - llm_reply: ChatCompletionMessageParam = { - "role": "assistant", - "content": message.content, - "tool_calls": [tool_call.model_dump() for tool_call in message.tool_calls] + client_config = { + "model": preset.model_name, + "max_tokens": preset.max_tokens, + "temperature": preset.temperature, + "timeout": 60, + "extra_body": preset.extra_body, } - if preset.request_with_reasoning_content: - llm_reply["reasoning_content"] = message.reasoning_content# pyright: ignore[reportGeneralTypeIssues] + if preset.support_mcp: + available_tools = await mcp_client.get_available_tools(is_group) + client_config["tools"] = available_tools - # 发送LLM调用工具时的回复，一般没有 - if message.content: - await send_split_messages(handler, message.content) - - # 处理每个工具调用 - new_messages.append(llm_reply) - - for tool_call in message.tool_calls: - logger.debug(f"处理工具调用：{tool_call.function.name} 参数：{tool_call.function.arguments}") - - tool_name = tool_call.function.name - try: - tool_args = json.loads(tool_call.function.arguments) - except (json.JSONDecodeError, TypeError, ValueError) as e: - error_message = ( - f"工具调用参数格式错误，无法解析 {tool_name} 的 arguments: {e!s}. " - f"原始参数: {tool_call.function.arguments}" - ) - logger.warning(error_message) - new_messages.append({ - "role": "tool", - "tool_call_id": tool_call.id, - "content": error_message, - }) - continue - - # 发送工具调用提示 - await handler.send(Message(f"正在使用{mcp_client.get_friendly_name(tool_name)}")) - - if is_group: - result = await mcp_client.call_tool( - tool_name, - tool_args, - group_id=event.group_id, - bot_id=str(event.self_id) - ) - else: - result = await mcp_client.call_tool( - tool_name, - tool_args, - bot_id=str(event.self_id) - ) - - new_messages.append({ - "role": "tool", - "tool_call_id": tool_call.id, - "content": str(result) - }) - - # 将工具调用的结果交给 LLM response = await client.chat.completions.create( **client_config, messages=messages + new_messages, ) + if response.usage is not None: + logger.debug(f"收到API响应 使用token数：{response.usage.total_tokens}") + message = response.choices[0].message - # 安全检查：确保 message 不为 None - if not message: - logger.error("API 响应中的 message 为 None") - await handler.send(Message("服务暂时不可用，请稍后再试")) - return + # 处理响应并处理工具调用 + while preset.support_mcp and message and message.tool_calls: + llm_reply: ChatCompletionMessageParam = { + "role": "assistant", + "content": message.content, + "tool_calls": [tool_call.model_dump() for tool_call in message.tool_calls] + } - reply, matched_reasoning_content = pop_reasoning_content( - message.content - ) - reasoning_content: str | None = ( - getattr(message, "reasoning_content", None) - or matched_reasoning_content - ) + if preset.request_with_reasoning_content: + llm_reply["reasoning_content"] = message.reasoning_content # pyright: ignore[reportGeneralTypeIssues] - llm_reply: ChatCompletionMessageParam = { - "role": "assistant", - "content": reply, - } + # 发送LLM调用工具时的回复，一般没有 + if message.content: + await send_split_messages(handler, message.content) - reply_images = getattr(message, "images", None) + # 处理每个工具调用 + new_messages.append(llm_reply) - if reply_images: - # openai的sdk里的assistant消息暂时没有images字段，需要单独处理 - llm_reply["images"] = reply_images # pyright: ignore[reportGeneralTypeIssues] + for tool_call in message.tool_calls: + logger.debug(f"处理工具调用：{tool_call.function.name} 参数：{tool_call.function.arguments}") - if preset.request_with_reasoning_content: - llm_reply["reasoning_content"] = reasoning_content# pyright: ignore[reportGeneralTypeIssues] + tool_name = tool_call.function.name + try: + tool_args = json.loads(tool_call.function.arguments) + except (json.JSONDecodeError, TypeError, ValueError) as e: + error_message = ( + f"工具调用参数格式错误，无法解析 {tool_name} 的 arguments: {e!s}. " + f"原始参数: {tool_call.function.arguments}" + ) + logger.warning(error_message) + new_messages.append({ + "role": "tool", + "tool_call_id": tool_call.id, + "content": error_message, + }) + continue - new_messages.append(llm_reply) + # 发送工具调用提示 + await handler.send(Message(f"正在使用{mcp_client.get_friendly_name(tool_name)}")) - # 请求成功后再保存历史记录，保证user和assistant穿插，防止R1模型报错 - for message in new_messages: - state.history.append(message) + if is_group: + result = await mcp_client.call_tool( + tool_name, + tool_args, + group_id=event.group_id, + bot_id=str(event.self_id) + ) + else: + result = await mcp_client.call_tool( + tool_name, + tool_args, + bot_id=str(event.self_id) + ) - if state.output_reasoning_content and reasoning_content: - try: - bot = get_bot(str(event.self_id)) - if is_group: - await bot.send_group_forward_msg( - group_id=group_id, - messages=build_reasoning_forward_nodes( - bot.self_id, reasoning_content - ), - ) - else: - await bot.send_private_forward_msg( - user_id=context_id, - messages=build_reasoning_forward_nodes( - bot.self_id, reasoning_content - ), - ) - except Exception as e: - logger.error(f"合并转发消息发送失败：\n{e!s}\n") + new_messages.append({ + "role": "tool", + "tool_call_id": tool_call.id, + "content": str(result) + }) - assert reply is not None - await send_split_messages(handler, reply) + # 将工具调用的结果交给 LLM + response = await client.chat.completions.create( + **client_config, + messages=messages + new_messages, + ) - if reply_images: - logger.debug(f"API响应 图片数：{len(reply_images)}") - for i, image in enumerate(reply_images, start=1): - logger.debug(f"正在发送第{i}张图片") - image_base64 = image["image_url"]["url"].removeprefix("data:image/png;base64,") - image_msg = MessageSegment.image(base64.b64decode(image_base64)) - await handler.send(image_msg) + message = response.choices[0].message - except Exception as e: - logger.opt(exception=e).error(f"API请求失败 {'群号' if is_group else '用户'}：{context_id}") - # 如果在处理过程中出现异常，恢复未处理的消息到state中 - state.past_events.extendleft(reversed(past_events_snapshot)) - await handler.send(Message(f"服务暂时不可用，请稍后再试\n{e!s}")) - finally: - state.processing = False - state.queue.task_done() - # 不再需要每次都清理MCPClient，因为它现在是单例 - # await mcp_client.cleanup() + # 安全检查：确保 message 不为 None + if not message: + logger.error("API 响应中的 message 为 None") + await handler.send(Message("服务暂时不可用，请稍后再试")) + return + + reply, matched_reasoning_content = pop_reasoning_content( + message.content + ) + reasoning_content: str | None = ( + getattr(message, "reasoning_content", None) + or matched_reasoning_content + ) + + llm_reply: ChatCompletionMessageParam = { + "role": "assistant", + "content": reply, + } + + reply_images = getattr(message, "images", None) + + if reply_images: + # openai的sdk里的assistant消息暂时没有images字段，需要单独处理 + llm_reply["images"] = reply_images # pyright: ignore[reportGeneralTypeIssues] + + if preset.request_with_reasoning_content: + llm_reply["reasoning_content"] = reasoning_content # pyright: ignore[reportGeneralTypeIssues] + + new_messages.append(llm_reply) + + # 请求成功后再保存历史记录，保证user和assistant穿插，防止R1模型报错 + for message in new_messages: + state.history.append(message) + + if state.output_reasoning_content and reasoning_content: + try: + bot = get_bot(str(event.self_id)) + if is_group: + await bot.send_group_forward_msg( + group_id=group_id, + messages=build_reasoning_forward_nodes( + bot.self_id, reasoning_content + ), + ) + else: + await bot.send_private_forward_msg( + user_id=context_id, + messages=build_reasoning_forward_nodes( + bot.self_id, reasoning_content + ), + ) + except Exception as e: + logger.error(f"合并转发消息发送失败：\n{e!s}\n") + + assert reply is not None + await send_split_messages(handler, reply) + + if reply_images: + logger.debug(f"API响应 图片数：{len(reply_images)}") + for i, image in enumerate(reply_images, start=1): + logger.debug(f"正在发送第{i}张图片") + image_base64 = image["image_url"]["url"].removeprefix("data:image/png;base64,") + image_msg = MessageSegment.image(base64.b64decode(image_base64)) + await handler.send(image_msg) + + except Exception as e: + logger.opt(exception=e).error(f"API请求失败 {'群号' if is_group else '用户'}：{context_id}") + # 如果在处理过程中出现异常，恢复未处理的消息到state中 + state.past_events.extendleft(reversed(past_events_snapshot)) + await handler.send(Message(f"服务暂时不可用，请稍后再试\n{e!s}")) + finally: + state.queue.task_done() + # 不再需要每次都清理MCPClient，因为它现在是单例 + # await mcp_client.cleanup() + finally: + state.processing = False # 预设切换命令 diff --git a/nonebot_plugin_llmchat/config.py b/nonebot_plugin_llmchat/config.py index 06bc55d..90fafcb 100755 --- a/nonebot_plugin_llmchat/config.py +++ b/nonebot_plugin_llmchat/config.py @@ -48,6 +48,10 @@ class ScopedConfig(BaseModel): "你的回答应该尽量简洁、幽默、可以使用一些语气词、颜文字。你应该拒绝回答任何政治相关的问题。", description="默认提示词", ) + mcp_server_cwd: str | None = Field( + None, + description="command类型MCP服务器的全局工作目录（cwd）" + ) mcp_servers: dict[str, MCPServerConfig] = Field({}, description="MCP服务器配置") blacklist_user_ids: set[int] = Field(set(), description="黑名单用户ID列表") ignore_prefixes: list[str] = Field( diff --git a/nonebot_plugin_llmchat/mcpclient.py b/nonebot_plugin_llmchat/mcpclient.py index 5dc6b6c..022b30d 100644 --- a/nonebot_plugin_llmchat/mcpclient.py +++ b/nonebot_plugin_llmchat/mcpclient.py @@ -18,12 +18,20 @@ class MCPClient: _SESSION_TTL_SECONDS = 600 _SESSION_CLEANUP_INTERVAL_SECONDS = 60 - def __new__(cls, server_config: dict[str, MCPServerConfig] | None = None): + def __new__( + cls, + server_config: dict[str, MCPServerConfig] | None = None, + default_command_cwd: str | None = None, + ): if cls._instance is None: cls._instance = super().__new__(cls) return cls._instance - def __init__(self, server_config: dict[str, MCPServerConfig] | None = None): + def __init__( + self, + server_config: dict[str, MCPServerConfig] | None = None, + default_command_cwd: str | None = None, + ): if self._initialized: return @@ -32,6 +40,7 @@ class MCPClient: logger.info(f"正在初始化MCPClient单例，共有{len(server_config)}个服务器配置") self.server_config = server_config + self.default_command_cwd = default_command_cwd self.sessions = {} self.exit_stack = AsyncExitStack() self._session_exit_stacks: dict[str, AsyncExitStack] = {} @@ -47,12 +56,16 @@ class MCPClient: logger.debug("MCPClient单例初始化成功") @classmethod - def get_instance(cls, server_config: dict[str, MCPServerConfig] | None = None): + def get_instance( + cls, + server_config: dict[str, MCPServerConfig] | None = None, + default_command_cwd: str | None = None, + ): """获取MCPClient实例""" if cls._instance is None: if server_config is None: raise ValueError("server_config must be provided for first initialization") - cls._instance = cls(server_config) + cls._instance = cls(server_config, default_command_cwd) return cls._instance @classmethod @@ -79,8 +92,15 @@ class MCPClient: sse_client(url=config.url, headers=config.headers) ) elif config.command: + stdio_params: dict[str, Any] = { + "command": config.command, + "args": config.args or [], + "env": config.env or {}, + } + if self.default_command_cwd: + stdio_params["cwd"] = self.default_command_cwd transport = await session_stack.enter_async_context( - cast(Any, stdio_client(StdioServerParameters(**config.model_dump()))) + cast(Any, stdio_client(StdioServerParameters(**stdio_params))) ) else: raise ValueError("Server config must have either url or command")