Merge pull request #4 from echoVic/feat/thinking-block-reasoning

Feat/thinking block reasoning

Author: echoVic

.blade/settings.local.json

@@ -1,7 +1,9 @@
 {
   "permissions": {
     "allow": [
-      "TestTool"
+      "TestTool",
+      "Bash(git status)",
+      "Bash(git diff *)"
     ],
     "ask": [],
     "deny": []

src/agent/Agent.ts

@@ -185,7 +185,7 @@ export class Agent {
         baseUrl: modelConfig.baseUrl,
         temperature: modelConfig.temperature ?? this.config.temperature,
         maxContextTokens: modelConfig.maxContextTokens ?? this.config.maxContextTokens, // 上下文窗口（压缩判断）
-        maxOutputTokens: this.config.maxOutputTokens, // 输出限制（API max_tokens）
+        maxOutputTokens: modelConfig.maxOutputTokens ?? this.config.maxOutputTokens, // 输出限制（API max_tokens）
         timeout: this.config.timeout,
       });
 
@@ -278,8 +278,8 @@ export class Agent {
           : await this.runLoop(enhancedMessage, context, loopOptions);
 
       if (!result.success) {
-        // 如果是用户中止，返回空字符串（不抛出异常）
-        if (result.error?.type === 'aborted') {
+        // 如果是用户中止或用户拒绝，返回空字符串（不抛出异常）
+        if (result.error?.type === 'aborted' || result.metadata?.shouldExitLoop) {
           return ''; // 返回空字符串，让调用方自行处理
         }
         // 其他错误则抛出异常
@@ -530,7 +530,6 @@ IMPORTANT: Execute according to the approved plan above. Follow the steps exactl
         // checkAndCompactInLoop 返回是否发生了压缩
         // 🆕 传入上一轮 LLM 返回的真实 prompt tokens（比估算更准确）
         const didCompact = await this.checkAndCompactInLoop(
-          messages,
           context,
           turnsCount,
           lastPromptTokens, // 首轮为 undefined，使用估算；后续轮次使用真实值
@@ -669,6 +668,11 @@ IMPORTANT: Execute according to the approved plan above. Follow the steps exactl
         logger.debug('当前权限模式:', context.permissionMode);
         logger.debug('================================\n');
 
+        // 🆕 如果 LLM 返回了 thinking 内容（DeepSeek R1 等），通知 UI
+        if (turnResult.reasoningContent && options?.onThinking) {
+          options.onThinking(turnResult.reasoningContent);
+        }
+
         // 🆕 如果 LLM 返回了 content，立即显示
         if (turnResult.content && turnResult.content.trim() && options?.onContent) {
           options.onContent(turnResult.content);
@@ -858,7 +862,7 @@ IMPORTANT: Execute according to the approved plan above. Follow the steps exactl
             }
             logger.debug('==================================\n');
 
-            // 🆕 检查是否应该退出循环（ExitPlanMode 返回时设置此标记）
+            // 🆕 检查是否应该退出循环（ExitPlanMode 或用户拒绝时设置此标记）
             if (result.metadata?.shouldExitLoop) {
               logger.debug('🚪 检测到退出循环标记，结束 Agent 循环');
 
@@ -1361,15 +1365,13 @@ IMPORTANT: Execute according to the approved plan above. Follow the steps exactl
    * 在 Agent 循环中检查并执行压缩
    * 仅使用 LLM 返回的真实 usage.promptTokens 进行判断（不再估算）
    *
-   * @param messages - 实际发送给 LLM 的消息数组
    * @param context - 聊天上下文
    * @param currentTurn - 当前轮次
    * @param actualPromptTokens - LLM 返回的真实 prompt tokens（必须，来自上一轮响应）
    * @param onCompacting - 压缩状态回调
    * @returns 是否发生了压缩
    */
   private async checkAndCompactInLoop(
-    messages: Message[],
     context: ChatContext,
     currentTurn: number,
     actualPromptTokens?: number,

src/agent/types.ts

@@ -169,7 +169,7 @@ export interface LoopResult {
     configuredMaxTurns?: number;
     actualMaxTurns?: number;
     hitSafetyLimit?: boolean;
-    shouldExitLoop?: boolean; // ExitPlanMode 设置此标记以退出循环
+    shouldExitLoop?: boolean; // ExitPlanMode 或用户拒绝时设置此标记以退出循环
     targetMode?: PermissionMode; // Plan 模式批准后的目标权限模式
     planContent?: string; // Plan 模式批准后的方案内容
   };

src/config/types.ts

@@ -59,8 +59,13 @@ export interface ModelConfig {
   // 可选：模型特定参数
   temperature?: number;
   maxContextTokens?: number; // 上下文窗口大小
+  maxOutputTokens?: number; // 输出 token 限制
   topP?: number;
   topK?: number;
+
+  // Thinking 模型配置（如 DeepSeek R1）
+  supportsThinking?: boolean; // 手动覆盖自动检测结果
+  thinkingBudget?: number; // 思考 token 预算（可选）
 }
 
 export interface BladeConfig {

src/services/ChatServiceInterface.ts

@@ -41,11 +41,13 @@ export interface ChatConfig {
  */
 export interface ChatResponse {
   content: string;
+  reasoningContent?: string; // Thinking 模型的推理过程（如 DeepSeek R1）
   toolCalls?: ChatCompletionMessageToolCall[];
   usage?: {
     promptTokens: number;
     completionTokens: number;
     totalTokens: number;
+    reasoningTokens?: number; // Thinking 模型消耗的推理 tokens
   };
 }
 
@@ -54,6 +56,7 @@ export interface ChatResponse {
  */
 export interface StreamChunk {
   content?: string;
+  reasoningContent?: string; // Thinking 模型的推理过程片段
   // biome-ignore lint/suspicious/noExplicitAny: 不同 provider 的 tool call 类型不同
   toolCalls?: any[];
   finishReason?: string;

src/services/OpenAIChatService.ts

@@ -193,13 +193,26 @@ export class OpenAIChatService implements IChatService {
         (tc): tc is ChatCompletionMessageToolCall => tc.type === 'function'
       );
 
+      // 提取 reasoning_content（DeepSeek R1 等 thinking 模型的扩展字段）
+      const extendedMessage = choice.message as typeof choice.message & {
+        reasoning_content?: string;
+      };
+      const reasoningContent = extendedMessage.reasoning_content || undefined;
+
+      // 提取 reasoning_tokens（thinking 模型的扩展 usage 字段）
+      const extendedUsage = completion.usage as typeof completion.usage & {
+        reasoning_tokens?: number;
+      };
+
       const response = {
         content: choice.message.content || '',
+        reasoningContent,
         toolCalls: toolCalls,
         usage: {
           promptTokens: completion.usage?.prompt_tokens || 0,
           completionTokens: completion.usage?.completion_tokens || 0,
           totalTokens: completion.usage?.total_tokens || 0,
+          reasoningTokens: extendedUsage?.reasoning_tokens,
         },
       };
 
@@ -309,6 +322,7 @@ export class OpenAIChatService implements IChatService {
 
       let chunkCount = 0;
       let totalContent = '';
+      let totalReasoningContent = '';
       let toolCallsReceived = false;
 
       for await (const chunk of stream) {
@@ -326,10 +340,19 @@ export class OpenAIChatService implements IChatService {
           continue;
         }
 
+        // 提取 reasoning_content（DeepSeek R1 等 thinking 模型的扩展字段）
+        const extendedDelta = delta as typeof delta & {
+          reasoning_content?: string;
+        };
+
         if (delta.content) {
           totalContent += delta.content;
         }
 
+        if (extendedDelta.reasoning_content) {
+          totalReasoningContent += extendedDelta.reasoning_content;
+        }
+
         if (delta.tool_calls && !toolCallsReceived) {
           toolCallsReceived = true;
           _logger.debug('🔧 [ChatService] Tool calls detected in stream');
@@ -341,13 +364,15 @@ export class OpenAIChatService implements IChatService {
           _logger.debug('📊 [ChatService] Stream summary:', {
             totalChunks: chunkCount,
             totalContentLength: totalContent.length,
+            totalReasoningContentLength: totalReasoningContent.length,
             hadToolCalls: toolCallsReceived,
             duration: Date.now() - startTime + 'ms',
           });
         }
 
         yield {
           content: delta.content || undefined,
+          reasoningContent: extendedDelta.reasoning_content || undefined,
           toolCalls: delta.tool_calls,
           finishReason: finishReason || undefined,
         };

src/services/VersionChecker.ts

@@ -18,8 +18,8 @@ const CACHE_DIR = path.join(
 );
 const CACHE_FILE = path.join(CACHE_DIR, 'version-cache.json');
 
-// 缓存有效期：24 小时
-const CACHE_TTL = 24 * 60 * 60 * 1000;
+// 缓存有效期：1 小时
+const CACHE_TTL = 1 * 60 * 60 * 1000;
 
 // npm registry URL
 const NPM_REGISTRY_URL = `https://registry.npmmirror.com/${PACKAGE_NAME}/latest`;
@@ -224,13 +224,58 @@ export function formatUpdateMessage(result: VersionCheckResult): string | null {
 }
 
 /**
- * 启动时版本检查（后台执行，不阻塞）
+ * 执行自动升级（后台进程，不阻塞主进程）
+ * @returns 升级提示消息
+ */
+async function performUpgrade(
+  currentVersion: string,
+  latestVersion: string
+): Promise<string> {
+  const { spawn } = await import('child_process');
+
+  try {
+    const updateCommand = `npm install -g blade-code@${latestVersion} --registry https://registry.npmjs.org`;
+
+    // 使用 spawn + detached + unref 在后台运行升级
+    // 这样主进程退出后，升级进程会继续运行完成安装
+    const updateProcess = spawn(updateCommand, {
+      stdio: 'ignore',
+      shell: true,
+      detached: true,
+    });
+    updateProcess.unref();
+
+    return `⬆️ 正在后台升级 ${currentVersion} → ${latestVersion}，下次启动生效`;
+  } catch {
+    return (
+      `\x1b[33m⚠️  Update available: ${currentVersion} → ${latestVersion}\x1b[0m\n` +
+      `   Run \x1b[36mnpm install -g ${PACKAGE_NAME}@latest\x1b[0m to update`
+    );
+  }
+}
+
+/**
+ * 启动时版本检查并自动升级
  *
- * @returns Promise<string | null> 更新提示消息，如果没有更新则返回 null
+ * @param autoUpgrade - 是否自动升级（默认 true）
+ * @returns Promise<string | null> 提示消息，如果没有更新则返回 null
  */
-export async function checkVersionOnStartup(): Promise<string | null> {
+export async function checkVersionOnStartup(
+  autoUpgrade = true
+): Promise<string | null> {
   try {
     const result = await checkVersion();
+
+    if (!result.hasUpdate || !result.latestVersion) {
+      return null;
+    }
+
+    // 自动升级
+    if (autoUpgrade) {
+      return await performUpgrade(result.currentVersion, result.latestVersion);
+    }
+
+    // 仅显示提示
     return formatUpdateMessage(result);
   } catch {
     return null;

src/store/selectors/index.ts

@@ -342,3 +342,23 @@ export const useIsModal = (modal: ActiveModal) =>
  */
 export const useIsBusy = () =>
   useBladeStore((state) => state.session.isThinking || state.command.isProcessing);
+
+// ==================== Thinking 模式选择器 ====================
+
+/**
+ * 获取 Thinking 模式是否启用
+ */
+export const useThinkingModeEnabled = () =>
+  useBladeStore((state) => state.app.thinkingModeEnabled);
+
+/**
+ * 获取当前 Thinking 内容（流式接收中）
+ */
+export const useCurrentThinkingContent = () =>
+  useBladeStore((state) => state.session.currentThinkingContent);
+
+/**
+ * 获取 Thinking 内容是否展开
+ */
+export const useThinkingExpanded = () =>
+  useBladeStore((state) => state.session.thinkingExpanded);

src/store/slices/appSlice.ts

@@ -32,6 +32,7 @@ const initialAppState: AppState = {
   modelEditorTarget: null,
   todos: [],
   awaitingSecondCtrlC: false,
+  thinkingModeEnabled: false, // Thinking 模式默认关闭
 };
 
 /**
@@ -137,5 +138,25 @@ export const createAppSlice: StateCreator<BladeStore, [], [], AppSlice> = (set)
         app: { ...state.app, awaitingSecondCtrlC: awaiting },
       }));
     },
+
+    // ==================== Thinking 模式相关 actions ====================
+
+    /**
+     * 设置 Thinking 模式开关状态
+     */
+    setThinkingModeEnabled: (enabled: boolean) => {
+      set((state) => ({
+        app: { ...state.app, thinkingModeEnabled: enabled },
+      }));
+    },
+
+    /**
+     * 切换 Thinking 模式开关
+     */
+    toggleThinkingMode: () => {
+      set((state) => ({
+        app: { ...state.app, thinkingModeEnabled: !state.app.thinkingModeEnabled },
+      }));
+    },
   },
 });