调整相关逻辑

main/xiaozhi-server/core/connection.py

@@ -42,7 +42,7 @@ from core.providers.tts.dto.dto import ContentType, TTSMessageDTO, SentenceType
 from config.logger import setup_logging, build_module_string, create_connection_logger
 from config.manage_api_client import DeviceNotFoundException, DeviceBindException, generate_and_save_chat_title
 from core.utils.prompt_manager import PromptManager
-from core.utils.voiceprint_provider import VoiceprintProvider
+from core.utils.voiceprint_provider import VoiceprintProvider, UNKNOWN_SPEAKER_NAME
 from core.utils.util import get_system_error_response
 from core.utils import textUtils
 
@@ -159,7 +159,6 @@ class ConnectionHandler:
         self.asr_audio = []  # 存储PCM帧列表，供VAD和ASR共享
         self.asr_audio_queue = queue.Queue()
         self.current_speaker = None  # 当前说话人展示名
-        self.current_speaker_id = None  # 声纹识别得到的 speaker_id
         self.introduced_speakers = set()  # 已"首次引入"的说话人，控制只在首轮带名字
         self.system_introduced_speakers = set()  # 已在 system 注入过身份的说话人，控制 system 身份只首轮出现
 
@@ -1053,7 +1052,8 @@ class ConnectionHandler:
         # 更新系统prompt至上下文
         self.dialogue.update_system_message(self.prompt)
 
-    def chat(self, query, depth=0):
+    def chat(self, query, depth=0, speaker=None):
+        # speaker: 本轮声纹身份快照，随工具调用下传
         # 保存当前任务的sentence_id到局部变量，避免被新任务覆盖
         current_sentence_id = None
 
@@ -1123,7 +1123,7 @@ class ConnectionHandler:
             # 避免每轮在 system 重复出现名字诱导模型反复称呼
             speaker_for_system = None
             cs = (self.current_speaker or "").strip()
-            if cs and cs != "未知说话人" and cs not in self.system_introduced_speakers:
+            if cs and cs != UNKNOWN_SPEAKER_NAME and cs not in self.system_introduced_speakers:
                 self.system_introduced_speakers.add(cs)
                 speaker_for_system = cs
 
@@ -1338,7 +1338,7 @@ class ConnectionHandler:
 
                     future = asyncio.run_coroutine_threadsafe(
                         self.func_handler.handle_llm_function_call(
-                            self, tool_call_data
+                            self, tool_call_data, speaker=speaker
                         ),
                         self.loop,
                     )
@@ -1501,7 +1501,7 @@ class ConnectionHandler:
                         )
                     )
 
-            self.chat(None, depth=depth + 1)
+            self.chat(None, depth=depth + 1, speaker=speaker)
 
     def _report_worker(self):
         """聊天记录上报工作线程"""

main/xiaozhi-server/core/handle/intentHandler.py

@@ -17,7 +17,7 @@ from core.providers.tts.dto.dto import TTSMessageDTO, SentenceType
 TAG = __name__
 
 
-async def handle_user_intent(conn: "ConnectionHandler", text):
+async def handle_user_intent(conn: "ConnectionHandler", text, speaker=None):
     # 预处理输入文本，处理可能的JSON格式
     try:
         if text.strip().startswith("{") and text.strip().endswith("}"):
@@ -172,7 +172,7 @@ async def process_intent_result(
                 try:
                     result = asyncio.run_coroutine_threadsafe(
                         conn.func_handler.handle_llm_function_call(
-                            conn, function_call_data
+                            conn, function_call_data, speaker=speaker
                         ),
                         conn.loop,
                     ).result(timeout=tool_call_timeout)

main/xiaozhi-server/core/handle/receiveAudioHandle.py

@@ -40,7 +40,8 @@ async def resume_vad_detection(conn: "ConnectionHandler"):
     conn.just_woken_up = False
 
 
-async def startToChat(conn: "ConnectionHandler", text):
+async def startToChat(conn: "ConnectionHandler", text, speaker=None):
+    """speaker: 本轮声纹身份快照，随调用链传递。"""
     # 检查输入是否是JSON格式（包含说话人信息）
     speaker_name = None
     actual_text = text
@@ -69,13 +70,8 @@ async def startToChat(conn: "ConnectionHandler", text):
         # 如果解析失败，继续使用原始文本
         pass
 
-    # 保存说话人信息到连接对象。speaker_id 由声纹识别写入，这里只在没有说话人时清掉，
-    # 避免上一轮身份落到没有声纹的文本消息上。
-    if speaker_name:
-        conn.current_speaker = speaker_name
-    else:
-        conn.current_speaker = None
-        conn.current_speaker_id = None
+    # 说话人名供提示词注入；声纹身份走 speaker 快照，不挂连接
+    conn.current_speaker = speaker_name
 
     if conn.need_bind:
         await check_bind_device(conn)
@@ -94,7 +90,7 @@ async def startToChat(conn: "ConnectionHandler", text):
         await handleAbortMessage(conn)
 
     # 首先进行意图分析，使用实际文本内容
-    intent_handled = await handle_user_intent(conn, actual_text)
+    intent_handled = await handle_user_intent(conn, actual_text, speaker=speaker)
 
     if intent_handled:
         # 如果意图已被处理，不再进行聊天
@@ -106,7 +102,7 @@ async def startToChat(conn: "ConnectionHandler", text):
     # 准备开始新会话
     conn.client_abort = False
 
-    conn.executor.submit(conn.chat, actual_text)
+    conn.executor.submit(conn.chat, actual_text, depth=0, speaker=speaker)
 
 
 async def no_voice_close_connect(conn: "ConnectionHandler", have_voice):

main/xiaozhi-server/core/providers/asr/base.py

@@ -17,6 +17,7 @@ from core.providers.asr.dto.dto import InterfaceType
 from core.handle.receiveAudioHandle import startToChat
 from core.handle.reportHandle import enqueue_asr_report
 from core.utils.util import remove_punctuation_and_length
+from core.utils.voiceprint_provider import SpeakerIdentification
 from core.handle.receiveAudioHandle import handleAudioMessage
 from typing import Optional, Tuple, List, NamedTuple, TYPE_CHECKING
 
@@ -119,13 +120,11 @@ class ASRProviderBase(ABC):
             else:
                 raw_text, _ = asr_result
 
-            speaker_id = None
             speaker_name = ""
             if isinstance(voiceprint_result, Exception):
                 logger.bind(tag=TAG).error(f"声纹识别失败: {voiceprint_result}")
             elif voiceprint_result is not None:
                 speaker_name = voiceprint_result.name or ""
-                speaker_id = voiceprint_result.speaker_id
 
             # 判断 ASR 结果类型
             if isinstance(raw_text, dict):
@@ -166,15 +165,16 @@ class ASRProviderBase(ABC):
             self.stop_ws_connection()
 
             if text_len > 0:
-                # 声纹 id 不进入发给 LLM 的文本，留在连接上供本轮 MCP 接入点调用读取。
-                if isinstance(speaker_id, str) and speaker_id.strip():
-                    conn.current_speaker_id = speaker_id.strip()
-                else:
-                    conn.current_speaker_id = None
+                # 本轮声纹身份快照，随调用链供 MCP 接入点工具使用
+                speaker = (
+                    voiceprint_result
+                    if isinstance(voiceprint_result, SpeakerIdentification)
+                    else None
+                )
                 audio_snapshot = asr_audio_task.copy()
                 enqueue_asr_report(conn, enhanced_text, audio_snapshot)
                 # 使用自定义模块进行上报
-                await startToChat(conn, enhanced_text)
+                await startToChat(conn, enhanced_text, speaker=speaker)
         except Exception as e:
             logger.bind(tag=TAG).error(f"处理语音停止失败: {e}")
             import traceback

main/xiaozhi-server/core/providers/tools/base/tool_executor.py

@@ -11,9 +11,9 @@ class ToolExecutor(ABC):
 
     @abstractmethod
     async def execute(
-        self, conn, tool_name: str, arguments: Dict[str, Any]
+        self, conn, tool_name: str, arguments: Dict[str, Any], speaker=None
     ) -> ActionResponse:
-        """执行工具调用"""
+        """执行工具调用。speaker: 可选的说话人身份快照。"""
         pass
 
     @abstractmethod

main/xiaozhi-server/core/providers/tools/device_iot/iot_executor.py

@@ -15,7 +15,7 @@ class DeviceIoTExecutor(ToolExecutor):
         self.iot_tools: Dict[str, ToolDefinition] = {}
 
     async def execute(
-        self, conn, tool_name: str, arguments: Dict[str, Any]
+        self, conn, tool_name: str, arguments: Dict[str, Any], speaker=None
     ) -> ActionResponse:
         """执行设备端IoT工具"""
         if not self.has_tool(tool_name):

main/xiaozhi-server/core/providers/tools/device_mcp/mcp_executor.py

@@ -16,7 +16,7 @@ class DeviceMCPExecutor(ToolExecutor):
         self.conn = conn
 
     async def execute(
-        self, conn: "ConnectionHandler", tool_name: str, arguments: Dict[str, Any]
+        self, conn: "ConnectionHandler", tool_name: str, arguments: Dict[str, Any], speaker=None
     ) -> ActionResponse:
         """执行设备端MCP工具"""
         if not hasattr(conn, "mcp_client") or not conn.mcp_client:

main/xiaozhi-server/core/providers/tools/mcp_endpoint/mcp_endpoint_executor.py

@@ -1,8 +1,9 @@
 """MCP接入点工具执行器"""
 
-from typing import Dict, Any
+from typing import Dict, Any, Optional
 from ..base import ToolType, ToolDefinition, ToolExecutor
 from plugins_func.register import Action, ActionResponse
+from core.utils.voiceprint_provider import SpeakerIdentification
 from .mcp_endpoint_handler import call_mcp_endpoint_tool
 from .mcp_speaker import should_forward_speaker, speaker_call_meta
 
@@ -14,9 +15,9 @@ class MCPEndpointExecutor(ToolExecutor):
         self.conn = conn
 
     async def execute(
-        self, conn, tool_name: str, arguments: Dict[str, Any]
+        self, conn, tool_name: str, arguments: Dict[str, Any], speaker: Optional[SpeakerIdentification] = None
     ) -> ActionResponse:
-        """执行MCP接入点工具"""
+        """执行MCP接入点工具。speaker: 本轮说话人快照。"""
         if not hasattr(conn, "mcp_endpoint_client") or not conn.mcp_endpoint_client:
             return ActionResponse(
                 action=Action.ERROR,
@@ -34,11 +35,10 @@ class MCPEndpointExecutor(ToolExecutor):
             import json
 
             meta = None
-            if should_forward_speaker(getattr(conn, "config", None)):
-                meta = speaker_call_meta(
-                    getattr(conn, "current_speaker_id", None),
-                    getattr(conn, "current_speaker", None),
-                )
+            if speaker is not None and should_forward_speaker(
+                getattr(conn, "config", None)
+            ):
+                meta = speaker_call_meta(speaker.speaker_id, speaker.name)
 
             args_str = json.dumps(arguments) if arguments else "{}"
 

main/xiaozhi-server/core/providers/tools/mcp_endpoint/mcp_speaker.py

@@ -13,7 +13,10 @@ def should_forward_speaker(config: Optional[dict]) -> bool:
     """mcp_forward_speaker 缺省为开启。显式 false 时不附带身份。"""
     if not isinstance(config, dict):
         return True
-    return bool(config.get("mcp_forward_speaker", True))
+    value = config.get("mcp_forward_speaker", True)
+    if value is None:
+        return True
+    return str(value).strip().lower() in ("true", "1", "yes")
 
 
 def speaker_call_meta(speaker_id: Any, speaker_name: Any) -> Optional[Dict[str, str]]:

main/xiaozhi-server/core/providers/tools/server_mcp/mcp_executor.py

@@ -22,7 +22,7 @@ class ServerMCPExecutor(ToolExecutor):
             await self.mcp_manager.initialize_servers()
 
     async def execute(
-        self, conn, tool_name: str, arguments: Dict[str, Any]
+        self, conn, tool_name: str, arguments: Dict[str, Any], speaker=None
     ) -> ActionResponse:
         """执行服务端MCP工具"""
         if not self._initialized or not self.mcp_manager:

main/xiaozhi-server/core/providers/tools/server_plugins/plugin_executor.py

@@ -17,7 +17,7 @@ class ServerPluginExecutor(ToolExecutor):
         self.config = conn.config
 
     async def execute(
-        self, conn: "ConnectionHandler", tool_name: str, arguments: Dict[str, Any]
+        self, conn: "ConnectionHandler", tool_name: str, arguments: Dict[str, Any], speaker=None
     ) -> ActionResponse:
         """执行服务端插件工具"""
         func_item = all_function_registry.get(tool_name)

main/xiaozhi-server/core/providers/tools/unified_tool_handler.py

@@ -137,16 +137,16 @@ class UnifiedToolHandler:
         return self.tool_manager.has_tool(tool_name)
 
     async def handle_llm_function_call(
-        self, conn, function_call_data: Dict[str, Any]
+        self, conn, function_call_data: Dict[str, Any], speaker=None
     ) -> Optional[ActionResponse]:
-        """处理LLM函数调用"""
+        """处理LLM函数调用。speaker: 本轮说话人快照。"""
         try:
             # 处理多函数调用
             if "function_calls" in function_call_data:
                 responses = []
                 for call in function_call_data["function_calls"]:
                     result = await self.tool_manager.execute_tool(
-                        call["name"], call.get("arguments", {})
+                        call["name"], call.get("arguments", {}), speaker=speaker
                     )
                     responses.append(result)
                 return self._combine_responses(responses)
@@ -175,7 +175,9 @@ class UnifiedToolHandler:
                 self.logger.warning(f"发送工具调用显示消息失败: {e}")
 
             # 执行工具调用
-            result = await self.tool_manager.execute_tool(function_name, arguments)
+            result = await self.tool_manager.execute_tool(
+                function_name, arguments, speaker=speaker
+            )
             return result
 
         except Exception as e:

main/xiaozhi-server/core/providers/tools/unified_tool_manager.py

@@ -71,9 +71,9 @@ class ToolManager:
         return tool_def.tool_type if tool_def else None
 
     async def execute_tool(
-        self, tool_name: str, arguments: Dict[str, Any]
+        self, tool_name: str, arguments: Dict[str, Any], speaker=None
     ) -> ActionResponse:
-        """执行工具调用"""
+        """执行工具调用。speaker: 本轮说话人快照。"""
         try:
             # 查找工具类型
             tool_type = self.get_tool_type(tool_name)
@@ -93,7 +93,7 @@ class ToolManager:
 
             # 执行工具
             self.logger.info(f"执行工具: {tool_name}，参数: {arguments}")
-            result = await executor.execute(self.conn, tool_name, arguments)
+            result = await executor.execute(self.conn, tool_name, arguments, speaker=speaker)
             self.logger.debug(f"工具执行结果: {result}")
             return result
 

main/xiaozhi-server/core/utils/dialogue.py

@@ -3,6 +3,8 @@ import re
 from typing import List, Dict
 from datetime import datetime
 
+from core.utils.voiceprint_provider import UNKNOWN_SPEAKER_NAME
+
 
 class Message:
     def __init__(
@@ -125,7 +127,7 @@ class Dialogue:
                 current_speaker_name = (current_speaker or "").strip()
                 # 仅在本轮注入了有效身份时才输出 speakers_info，避免列表里的名字每轮
                 # 重复出现诱导模型反复称呼；后续轮不再注入身份，靠对话历史首轮保留
-                if current_speaker_name and current_speaker_name != "未知说话人":
+                if current_speaker_name and current_speaker_name != UNKNOWN_SPEAKER_NAME:
                     speakers = voiceprint_config.get("speakers", [])
                     speakers_info = "\n<speakers_info>"
                     speakers_info += f"\n当前说话人：{current_speaker_name}"

main/xiaozhi-server/core/utils/voiceprint_provider.py

@@ -23,7 +23,10 @@ class SpeakerIdentification(NamedTuple):
 
 
 def resolve_speaker_identification(
-    speaker_id, score, speaker_map, similarity_threshold
+    speaker_id: Optional[str],
+    score: float,
+    speaker_map: Dict[str, Dict[str, str]],
+    similarity_threshold: float,
 ) -> SpeakerIdentification:
     """把声纹接口返回的 speaker_id / score 解析成展示名，并保留命中的 id。"""
     if score < similarity_threshold:

main/xiaozhi-server/tests/core/providers/tools/test_mcp_endpoint_speaker.py

@@ -8,7 +8,7 @@ from core.providers.tools.mcp_endpoint.mcp_speaker import (
     should_forward_speaker,
     speaker_call_meta,
 )
-from core.utils.voiceprint_provider import UNKNOWN_SPEAKER_NAME
+from core.utils.voiceprint_provider import UNKNOWN_SPEAKER_NAME, SpeakerIdentification
 
 
 def test_meta_includes_id_and_name():
@@ -29,6 +29,14 @@ def test_forwarding_defaults_on_and_can_be_disabled():
     assert should_forward_speaker({}) is True
     assert should_forward_speaker({"mcp_forward_speaker": True}) is True
     assert should_forward_speaker({"mcp_forward_speaker": False}) is False
+    # 字符串 false 可关闭；空值视为缺省开启
+    assert should_forward_speaker({"mcp_forward_speaker": "false"}) is False
+    assert should_forward_speaker({"mcp_forward_speaker": "False"}) is False
+    assert should_forward_speaker({"mcp_forward_speaker": "0"}) is False
+    assert should_forward_speaker({"mcp_forward_speaker": "true"}) is True
+    assert should_forward_speaker({"mcp_forward_speaker": None}) is True
+    # 空串按关闭处理，与 manager-api 的 Boolean.parseBoolean("") 一致
+    assert should_forward_speaker({"mcp_forward_speaker": ""}) is False
 
 
 def test_payload_adds_meta_without_touching_arguments():
@@ -57,11 +65,9 @@ class _Client:
 
 
 class _Conn:
-    def __init__(self, config=None, speaker_id="test1", speaker_name="张三"):
+    def __init__(self, config=None):
         self.mcp_endpoint_client = _Client()
         self.config = {} if config is None else config
-        self.current_speaker_id = speaker_id
-        self.current_speaker = speaker_name
 
 
 @pytest.mark.asyncio
@@ -80,7 +86,10 @@ async def test_executor_forwards_speaker_in_meta_only(monkeypatch):
     arguments = {"area": "客厅"}
     conn = _Conn()
     result = await executor_mod.MCPEndpointExecutor(conn).execute(
-        conn, "HassTurnOn", arguments
+        conn,
+        "HassTurnOn",
+        arguments,
+        speaker=SpeakerIdentification("test1", "张三"),
     )
 
     assert json.loads(captured["args"]) == {"area": "客厅"}
@@ -101,7 +110,12 @@ async def test_executor_skips_meta_when_disabled(monkeypatch):
 
     monkeypatch.setattr(executor_mod, "call_mcp_endpoint_tool", fake_call)
     conn = _Conn(config={"mcp_forward_speaker": False})
-    await executor_mod.MCPEndpointExecutor(conn).execute(conn, "HassTurnOn", {"area": "客厅"})
+    await executor_mod.MCPEndpointExecutor(conn).execute(
+        conn,
+        "HassTurnOn",
+        {"area": "客厅"},
+        speaker=SpeakerIdentification("test1", "张三"),
+    )
     assert captured["meta"] is None
 
 
@@ -116,6 +130,33 @@ async def test_executor_omits_meta_for_unknown_speaker(monkeypatch):
         return "done"
 
     monkeypatch.setattr(executor_mod, "call_mcp_endpoint_tool", fake_call)
-    conn = _Conn(speaker_id=None, speaker_name=UNKNOWN_SPEAKER_NAME)
-    await executor_mod.MCPEndpointExecutor(conn).execute(conn, "HassTurnOn", {})
+    conn = _Conn()
+    await executor_mod.MCPEndpointExecutor(conn).execute(
+        conn,
+        "HassTurnOn",
+        {},
+        speaker=SpeakerIdentification(None, UNKNOWN_SPEAKER_NAME),
+    )
     assert captured["meta"] is None
+
+
+@pytest.mark.asyncio
+async def test_executor_reads_snapshot_not_conn_state(monkeypatch):
+    """meta 取自快照，不受连接态影响。"""
+    from core.providers.tools.mcp_endpoint import mcp_endpoint_executor as executor_mod
+
+    captured = {}
+
+    async def fake_call(client, tool_name, args_str="{}", timeout=30, meta=None):
+        captured["meta"] = meta
+        return "done"
+
+    monkeypatch.setattr(executor_mod, "call_mcp_endpoint_tool", fake_call)
+    conn = _Conn()
+    await executor_mod.MCPEndpointExecutor(conn).execute(
+        conn,
+        "HassTurnOn",
+        {},
+        speaker=SpeakerIdentification("test1", "张三"),
+    )
+    assert captured["meta"] == {"speakerId": "test1", "speakerName": "张三"}