feat: LLM 调用日志 + ModelSelector 优化 + devtools.bat 编码修复

- 新增 call_log.go: 全局环形缓冲区记录每次 LLM 调用(模型/Token/耗时/错误)
- OpenAIProvider.doChat/ChatStreamWithTools 自动记录调用日志
- ai-core 暴露 GET /api/v1/llm-calls 端点, DevTools 代理 + UI 面板
- ModelSelector.envProvider 改为单例缓存, 避免重复创建 HTTP Client
- 新增 PurposeToolCalling 适配器, 后台思考工具调用走专用路由
- envFallback 超时 120s→180s, 显式设置 MaxRetries
- devtools.bat 全英文, 解决 Windows CMD GBK 编码乱码问题

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/ai-core/cmd/main.go

@@ -67,7 +67,8 @@ func main() {
 		APIKey:        cfg.LLMAPIKey,
 		Model:         cfg.LLMModel,
 		FallbackModel: cfg.LLMFallbackModel,
-		Timeout:       120 * time.Second,
+		MaxRetries:    3,
+		Timeout:       180 * time.Second,
 	}
 
 	// 创建 ModelSelector (优先使用 models.json，回退到 .env)
@@ -82,10 +83,12 @@ func main() {
 	thinkerAdapter := llm.NewAdapter(provider)
 	provider, _ = modelSelector.Select(context.Background(), llm.PurposeMemoryExtraction)
 	memoryAdapter := llm.NewAdapter(provider)
+	provider, _ = modelSelector.Select(context.Background(), llm.PurposeToolCalling)
+	toolAdapter := llm.NewAdapter(provider)
 
 	if configLoader != nil && configLoader.HasConfig() {
-		log.Printf("LLM适配器已就绪: models.json 驱动 (chat=%s, intent=%s, think=%s, memory=%s)",
-			chatAdapter.ModelName(), intentAdapter.ModelName(), thinkerAdapter.ModelName(), memoryAdapter.ModelName())
+		log.Printf("LLM适配器已就绪: models.json 驱动 (chat=%s, intent=%s, think=%s, memory=%s, tool=%s)",
+			chatAdapter.ModelName(), intentAdapter.ModelName(), thinkerAdapter.ModelName(), memoryAdapter.ModelName(), toolAdapter.ModelName())
 	} else {
 		log.Printf("LLM适配器已就绪: .env 驱动 (模型=%s)", chatAdapter.ModelName())
 	}
@@ -193,6 +196,7 @@ func main() {
 		personaLoader,
 		memRetriever,
 		thinkerAdapter,
+		toolAdapter,
 		iotClient,
 		memStore,
 		toolRegistry,
@@ -311,6 +315,22 @@ func main() {
 		w.Write([]byte(`{"status":"ok","service":"ai-core","model":"` + chatAdapter.ModelName() + `"}`))
 	})
 
+	// LLM 调用日志（调试用）
+	mux.HandleFunc("/api/v1/llm-calls", func(w http.ResponseWriter, r *http.Request) {
+		limit := 50
+		if n, err := fmt.Sscanf(r.URL.Query().Get("limit"), "%d", &limit); n != 1 || err != nil || limit <= 0 {
+			limit = 50
+		}
+		if limit > 500 {
+			limit = 500
+		}
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]interface{}{
+			"calls": llm.GetCalls(limit),
+			"total": len(llm.GetCalls(0)),
+		})
+	})
+
 	// 启动HTTP服务
 	srv := &http.Server{
 		Addr:    ":" + cfg.Port,

backend/ai-core/internal/background/thinker.go

@@ -43,6 +43,7 @@ type Thinker struct {
 	personaLoader *persona.Loader
 	memRetriever  *memory.Retriever
 	llmAdapter    *llm.Adapter
+	toolAdapter   *llm.Adapter // 工具调用专用适配器
 	iotClient     *tools.IoTClient
 
 	// 记忆管理
@@ -224,6 +225,7 @@ func NewThinker(
 	personaLoader *persona.Loader,
 	memRetriever *memory.Retriever,
 	llmAdapter *llm.Adapter,
+	toolAdapter *llm.Adapter,
 	iotClient *tools.IoTClient,
 	memoryStore *memory.Store,
 	toolRegistry *tools.Registry,
@@ -237,6 +239,7 @@ func NewThinker(
 		personaLoader:       personaLoader,
 		memRetriever:        memRetriever,
 		llmAdapter:          llmAdapter,
+		toolAdapter:         toolAdapter,
 		iotClient:           iotClient,
 		thinkInterval:       cfg.ThinkInterval,
 		silenceTimeout:      cfg.SilenceTimeout,
@@ -598,7 +601,7 @@ func (t *Thinker) performThink(triggerReason string) {
 	var toolCallRecords []map[string]interface{}
 
 	for round := 0; round <= maxToolRounds; round++ {
-		resp, err := t.llmAdapter.ChatWithTools(ctx, messages, openAITools)
+		resp, err := t.toolAdapter.ChatWithTools(ctx, messages, openAITools)
 		if err != nil {
 			log.Printf("[后台思考] LLM调用失败 (round=%d): %v", round, err)
 			return

backend/ai-core/internal/llm/call_log.go

@@ -0,0 +1,74 @@
+package llm
+
+import (
+	"sync"
+	"time"
+)
+
+// CallRecord records a single LLM API call.
+type CallRecord struct {
+	Time             time.Time     `json:"time"`
+	Model            string        `json:"model"`
+	Duration         time.Duration `json:"duration_ms"`
+	PromptTokens     int           `json:"prompt_tokens"`
+	CompletionTokens int           `json:"completion_tokens"`
+	TotalTokens      int           `json:"total_tokens"`
+	Success          bool          `json:"success"`
+	Error            string        `json:"error,omitempty"`
+}
+
+// CallLogger is a thread-safe ring buffer for LLM call records.
+type CallLogger struct {
+	mu       sync.RWMutex
+	records  []CallRecord
+	capacity int
+	head     int
+	size     int
+}
+
+var globalCallLogger = &CallLogger{capacity: 500}
+
+// LogCall records an LLM call. Safe for concurrent use.
+func LogCall(r CallRecord) {
+	globalCallLogger.log(r)
+}
+
+// GetCalls returns recent call records, newest first.
+func GetCalls(limit int) []CallRecord {
+	return globalCallLogger.get(limit)
+}
+
+func (cl *CallLogger) log(r CallRecord) {
+	cl.mu.Lock()
+	defer cl.mu.Unlock()
+
+	if cl.records == nil {
+		cl.records = make([]CallRecord, cl.capacity)
+	}
+
+	r.Time = time.Now()
+	cl.records[cl.head] = r
+	cl.head = (cl.head + 1) % cl.capacity
+	if cl.size < cl.capacity {
+		cl.size++
+	}
+}
+
+func (cl *CallLogger) get(limit int) []CallRecord {
+	cl.mu.RLock()
+	defer cl.mu.RUnlock()
+
+	if limit <= 0 || limit > cl.size {
+		limit = cl.size
+	}
+
+	result := make([]CallRecord, limit)
+	for i := 0; i < limit; i++ {
+		idx := (cl.head - 1 - i) % cl.capacity
+		if idx < 0 {
+			idx += cl.capacity
+		}
+		result[i] = cl.records[idx]
+	}
+	return result
+}

backend/ai-core/internal/llm/openai.go

@@ -139,14 +139,38 @@ func (p *OpenAIProvider) ChatStreamWithTools(ctx context.Context, messages []mod
 	go func() {
 		defer close(ch)
 
+		startTime := time.Now()
+		modelName := p.config.Model
+		var streamErr error
+		var finalUsage *model.Usage
+
+		defer func() {
+			r := CallRecord{
+				Model:    modelName,
+				Duration: time.Since(startTime),
+				Success:  streamErr == nil,
+			}
+			if streamErr != nil {
+				r.Error = streamErr.Error()
+			}
+			if finalUsage != nil {
+				r.PromptTokens = finalUsage.PromptTokens
+				r.CompletionTokens = finalUsage.CompletionTokens
+				r.TotalTokens = finalUsage.TotalTokens
+			}
+			LogCall(r)
+		}()
+
 		resp, err := p.doChatStream(ctx, messages, p.config.Model, tools)
 		if err != nil {
 			// Fallback
 			if p.config.FallbackModel != "" {
 				logger.Printf("[LLM] 流式调用主模型失败，降级: %v", err)
+				modelName = p.config.FallbackModel
 				resp, err = p.doChatStream(ctx, messages, p.config.FallbackModel, tools)
 			}
 			if err != nil {
+				streamErr = err
 				ch <- StreamChunk{Error: err, Done: true}
 				return
 			}
@@ -184,20 +208,22 @@ func (p *OpenAIProvider) ChatStreamWithTools(ctx context.Context, messages []mod
 					ch <- StreamChunk{Content: deltaStr}
 				}
 				if streamResp.Choices[0].FinishReason != "" {
-					usage := &model.Usage{}
 					if streamResp.Usage != nil {
-						usage.PromptTokens = streamResp.Usage.PromptTokens
-						usage.CompletionTokens = streamResp.Usage.CompletionTokens
-						usage.TotalTokens = streamResp.Usage.TotalTokens
+						finalUsage = &model.Usage{
+							PromptTokens:     streamResp.Usage.PromptTokens,
+							CompletionTokens: streamResp.Usage.CompletionTokens,
+							TotalTokens:      streamResp.Usage.TotalTokens,
+						}
 					}
-					ch <- StreamChunk{Done: true, Usage: usage}
+					ch <- StreamChunk{Done: true, Usage: finalUsage}
 					return
 				}
 			}
 		}
 
 		if err := scanner.Err(); err != nil {
-			ch <- StreamChunk{Error: fmt.Errorf("读取流式响应失败: %w", err), Done: true}
+			streamErr = fmt.Errorf("读取流式响应失败: %w", err)
+			ch <- StreamChunk{Error: streamErr, Done: true}
 			return
 		}
 
@@ -222,7 +248,25 @@ type openAIStreamChoice struct {
 }
 
 // doChat 执行同步对话请求
-func (p *OpenAIProvider) doChat(ctx context.Context, messages []model.LLMMessage, modelName string, stream bool, tools []OpenAITool) (*model.LLMResponse, error) {
+func (p *OpenAIProvider) doChat(ctx context.Context, messages []model.LLMMessage, modelName string, stream bool, tools []OpenAITool) (llmResp *model.LLMResponse, err error) {
+	startTime := time.Now()
+	defer func() {
+		r := CallRecord{
+			Model:    modelName,
+			Duration: time.Since(startTime),
+			Success:  err == nil,
+		}
+		if err != nil {
+			r.Error = err.Error()
+		}
+		if llmResp != nil {
+			r.PromptTokens = llmResp.Usage.PromptTokens
+			r.CompletionTokens = llmResp.Usage.CompletionTokens
+			r.TotalTokens = llmResp.Usage.TotalTokens
+		}
+		LogCall(r)
+	}()
+
 	// 转换消息格式
 	oaiMessages := make([]openAIMessage, len(messages))
 	for i, msg := range messages {
@@ -304,7 +348,7 @@ func (p *OpenAIProvider) doChat(ctx context.Context, messages []model.LLMMessage
 
 	// 检查是否有工具调用
 	choice := oaiResp.Choices[0]
-	llmResp := &model.LLMResponse{
+	llmResp = &model.LLMResponse{
 		Content:          contentString(choice.Message.Content),
 		FinishReason:     choice.FinishReason,
 		ReasoningContent: choice.Message.ReasoningContent,

backend/ai-core/internal/llm/selector.go

@@ -25,10 +25,11 @@ var ErrModelNotRequired = fmt.Errorf("model not required, caller should degrade
 
 // ModelSelector routes requests to the best available LLMProvider based on purpose.
 type ModelSelector struct {
-	loader *config.Loader
-	envCfg OpenAIConfig
-	mu     sync.RWMutex
-	cache  map[string]LLMProvider
+	loader    *config.Loader
+	envCfg    OpenAIConfig
+	mu        sync.RWMutex
+	cache     map[string]LLMProvider
+	cachedEnv LLMProvider // cached env fallback, created once
 }
 
 // NewModelSelector creates a ModelSelector. If loader is nil or has no config,
@@ -81,7 +82,12 @@ func (s *ModelSelector) DefaultAdapter() *Adapter {
 }
 
 func (s *ModelSelector) envProvider() LLMProvider {
-	return NewOpenAIProvider(s.envCfg)
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.cachedEnv == nil {
+		s.cachedEnv = NewOpenAIProvider(s.envCfg)
+	}
+	return s.cachedEnv
 }
 
 func (s *ModelSelector) getOrCreateProvider(modelID string, cfg *config.ModelsConfigData) (LLMProvider, error) {

devtools.bat

@@ -2,7 +2,7 @@
 setlocal enabledelayedexpansion
 :: ========================================
 :: Cyrene DevTools CLI (Windows CMD)
-:: 用法: devtools.bat [命令] [选项]
+:: Usage: devtools.bat [command] [options]
 :: ========================================
 
 set "SCRIPT_DIR=%~dp0"
@@ -13,7 +13,7 @@ set "LOG_DIR=%DEVTOOLS_DIR%\logs"
 set "LOG_FILE=%LOG_DIR%\sh.log"
 set "DB_COMPOSE_FILE=%ROOT%docker-compose.dev.db.yml"
 
-:: 解析第一个参数作为命令
+:: Parse first argument as command
 set CMD=%1
 if "%CMD%"=="" set CMD=start
 if "%CMD%"=="help" goto :help

devtools/public/index.html

@@ -696,6 +696,9 @@ input[type="range"] { accent-color: var(--accent); padding: 0; }
     <button class="nav-item" data-panel="modelConfig">
       <span class="nav-icon">🤖</span><span class="nav-label">模型配置</span>
     </button>
+    <button class="nav-item" data-panel="llmCalls">
+      <span class="nav-icon">📊</span><span class="nav-label">LLM 调用</span>
+    </button>
   </nav>
   <div class="sidebar-footer">
     <span id="ws-dot" class="disconnected"></span>
@@ -737,6 +740,7 @@ input[type="range"] { accent-color: var(--accent); padding: 0; }
     <!-- 客户端管理 -->
     <div class="panel" id="panel-clients"></div>
     <div class="panel" id="panel-modelConfig"></div>
+    <div class="panel" id="panel-llmCalls"></div>
   </div>
 </div>
 
@@ -801,7 +805,8 @@ const STATE = {
 	  modelConfigModels: [],
 	  modelConfigRouting: [],
 	  fetchedModels: [],
-	  expandedThinkingLogs: {},
+	  llmCallsData: [],
+  expandedThinkingLogs: {},
 };
 
 // ========== WebSocket ==========
@@ -1015,6 +1020,7 @@ function switchPanel(name) {
     chatPlatforms: '💬 第三方聊天配置与消息日志',
     clients: '📱 客户端管理',
 	    modelConfig: '🤖 模型配置管理',
+    llmCalls: '📊 LLM 调用日志',
   };
   document.getElementById('panel-title').textContent = titles[name] || name;
 
@@ -1041,6 +1047,7 @@ function switchPanel(name) {
     case 'chatPlatforms': renderChatPlatformsPanel(); stopSessionsAutoRefresh(); stopDashboardAutoRefresh(); stopDbAutoRefresh(); stopIoTRefresh(); stopTimelineAutoRefresh(); startChatAutoRefresh(); break;
     case 'clients': renderClientsPanel(); stopSessionsAutoRefresh(); stopDashboardAutoRefresh(); stopDbAutoRefresh(); stopIoTRefresh(); stopTimelineAutoRefresh(); break;
 	    case 'modelConfig': renderModelConfigPanel(); stopSessionsAutoRefresh(); stopDashboardAutoRefresh(); stopDbAutoRefresh(); stopIoTRefresh(); stopTimelineAutoRefresh(); break;
+	    case 'llmCalls': renderLlmCallsPanel(); stopSessionsAutoRefresh(); stopDashboardAutoRefresh(); stopDbAutoRefresh(); stopIoTRefresh(); stopTimelineAutoRefresh(); break;
   }
 }
 
@@ -4463,6 +4470,63 @@ async function editClientNote(clientID) {
   }
 }
 
+// ========== LLM Calls Panel ==========
+async function renderLlmCallsPanel() {
+  var panel = document.getElementById('panel-llmCalls');
+  panel.innerHTML = '<div class="loading">🔄 加载中...</div>';
+
+  try {
+    var resp = await api('/api/llm-calls?limit=100');
+    var calls = resp.calls || [];
+    STATE.llmCallsData = calls;
+
+    if (calls.length === 0) {
+      panel.innerHTML = '<div class="empty-state">暂无 LLM 调用记录<br><small>发送一条消息后刷新查看</small></div>';
+      return;
+    }
+
+    var totalTokens = calls.reduce(function(s, c) { return s + (c.total_tokens || 0); }, 0);
+    var successCount = calls.filter(function(c) { return c.success; }).length;
+
+    var html = '<div class="stats-row" style="margin-bottom:16px">' +
+      '<div class="stat-card"><div class="stat-value">' + calls.length + '</div><div class="stat-label">总调用</div></div>' +
+      '<div class="stat-card"><div class="stat-value">' + successCount + '/' + calls.length + '</div><div class="stat-label">成功</div></div>' +
+      '<div class="stat-card"><div class="stat-value">' + formatTokens(totalTokens) + '</div><div class="stat-label">总 Token</div></div>' +
+      '</div>';
+
+    html += '<div class="table-wrap"><table class="data-table">' +
+      '<thead><tr>' +
+      '<th>时间</th><th>模型</th><th>耗时</th><th>Prompt</th><th>Completion</th><th>Total</th><th>状态</th>' +
+      '</tr></thead><tbody>';
+
+    calls.forEach(function(c) {
+      var statusClass = c.success ? 'status-ok' : 'status-err';
+      var statusText = c.success ? '✓' : '✗ ' + escHtml(c.error || '');
+      var durMs = (c.duration_ms || 0) / 1000000;
+      html += '<tr>' +
+        '<td style="font-size:11px;white-space:nowrap">' + formatTime(c.time) + '</td>' +
+        '<td style="font-size:12px;font-family:monospace">' + escHtml(c.model) + '</td>' +
+        '<td style="font-size:11px">' + (durMs > 0 ? (durMs / 1000).toFixed(2) + 's' : '-') + '</td>' +
+        '<td style="font-size:11px">' + (c.prompt_tokens || 0).toLocaleString() + '</td>' +
+        '<td style="font-size:11px">' + (c.completion_tokens || 0).toLocaleString() + '</td>' +
+        '<td style="font-size:11px;font-weight:600">' + (c.total_tokens || 0).toLocaleString() + '</td>' +
+        '<td><span class="' + statusClass + '">' + statusText + '</span></td>' +
+        '</tr>';
+    });
+
+    html += '</tbody></table></div>';
+    panel.innerHTML = html;
+  } catch (err) {
+    panel.innerHTML = '<div class="error-state">加载失败: ' + escHtml(err.message) + '<br><small>确认 AI-Core 服务已启动</small></div>';
+  }
+}
+
+function formatTokens(n) {
+  if (n >= 1000000) return (n / 1000000).toFixed(1) + 'M';
+  if (n >= 1000) return (n / 1000).toFixed(1) + 'K';
+  return String(n);
+}
+
 </script>
 <script src="iot-panel.js"></script>
 <script>
@@ -4471,7 +4535,7 @@ async function editClientNote(clientID) {
 // Listen for browser back/forward navigation.
 window.addEventListener('hashchange', function() {
   var hash = location.hash.replace('#', '');
-  var validPanels = ['dashboard', 'memory', 'sessions', 'services', 'iot', 'performance', 'database', 'toolCalls', 'stt', 'thinking', 'timeline', 'chatPlatforms', 'clients', 'modelConfig'];
+  var validPanels = ['dashboard', 'memory', 'sessions', 'services', 'iot', 'performance', 'database', 'toolCalls', 'stt', 'thinking', 'timeline', 'chatPlatforms', 'clients', 'modelConfig', 'llmCalls'];
   if (hash && validPanels.indexOf(hash) >= 0 && hash !== STATE.activePanel) {
     switchPanel(hash);
   }
@@ -4482,7 +4546,7 @@ refreshStatus();
 
 // Restore last panel from URL hash, or default to dashboard.
 var initHash = location.hash.replace('#', '');
-var validPanels = ['dashboard', 'memory', 'sessions', 'services', 'iot', 'performance', 'database', 'toolCalls', 'stt', 'thinking', 'timeline', 'chatPlatforms', 'clients', 'modelConfig'];
+var validPanels = ['dashboard', 'memory', 'sessions', 'services', 'iot', 'performance', 'database', 'toolCalls', 'stt', 'thinking', 'timeline', 'chatPlatforms', 'clients', 'modelConfig', 'llmCalls'];
 if (initHash && validPanels.indexOf(initHash) >= 0) {
   switchPanel(initHash);
 } else {

devtools/src/index.js

@@ -19,6 +19,7 @@ import { processManager } from './process-manager.js';
 import { performanceMonitor } from './performance.js';
 import { SERVICES, DEVTOOLS_PORT, LOGS_DIR, logFile, GATEWAY_URL, TOOL_ENGINE_URL, ADMIN_USERNAME, ADMIN_PASSWORD } from './config.js';
 
+const AI_CORE_URL = process.env.AI_CORE_URL || 'http://localhost:8081';
 const MEMORY_SERVICE_URL = process.env.MEMORY_SERVICE_URL || 'http://localhost:8091';
 const VOICE_SERVICE_URL = process.env.VOICE_SERVICE_URL || 'http://localhost:8093';
 const PLATFORM_BRIDGE_URL = process.env.PLATFORM_BRIDGE_URL || 'http://localhost:8095';
@@ -168,6 +169,31 @@ async function proxyToGateway(path, opts = {}) {
   }
 }
 
+/**
+ * 代理请求到 AI-Core
+ */
+async function proxyToAICore(path, opts = {}) {
+  const url = `${AI_CORE_URL}${path}`;
+  try {
+    const resp = await fetch(url, {
+      ...opts,
+      headers: { 'Content-Type': 'application/json', ...opts.headers },
+      signal: AbortSignal.timeout(10000),
+    });
+    const body = await resp.json().catch(() => null);
+    return { status: resp.status, body };
+  } catch (err) {
+    return {
+      status: 502,
+      body: {
+        error: `AI-Core 不可达: ${err.message}`,
+        errorType: 'ai_core_unreachable',
+        hint: 'AI-Core 服务未启动，请先在「服务管理」面板中启动 AI-Core',
+      },
+    };
+  }
+}
+
 // ========== REST API 路由 ==========
 
 // ---- 健康检查 ----
@@ -1003,6 +1029,13 @@ app.get('/api/voice/health', async (_req, res) => {
   res.status(result.status).json(result.body);
 });
 
+// GET /api/llm-calls — LLM 调用日志 (代理到 AI-Core)
+app.get('/api/llm-calls', async (req, res) => {
+  const limit = parseInt(req.query.limit) || 50;
+  const result = await proxyToAICore(`/api/v1/llm-calls?limit=${Math.min(limit, 500)}`);
+  res.status(result.status).json(result.body);
+});
+
 /**
  * 代理请求到 Memory-Service
  * @param {string} path - Memory-Service API 路径