aegiscode 3.1.10 → 3.2.0

package/src/agent/Agent.ts

@@ -1,903 +0,0 @@
-/**
- * Agent 核心类 - 无状态设计
- * 
- * 
- * 1. Agent 本身不保存任何会话状态（sessionId, messages 等）
- * 2. 所有状态通过 context 参数传入
- * 3. Agent 实例可以每次命令创建，用完即弃
- * 4. 历史连续性由外部 SessionContext 保证
- */
-
-import type {
-  AgentConfig,
-  AgentOptions,
-  ChatContext,
-  LoopOptions,
-  LoopResult,
-  Message,
-  IChatService,
-  ToolCall,
-  ToolResult,
-  ToolDefinition,
-  StreamCallbacks,
-} from './types.js';
-import { createChatService } from '../services/ChatService.js';
-import { buildSystemPrompt } from '../prompts/builder.js';
-import { createPlanModeReminder } from '../prompts/plan.js';
-import { LOCAL_SYSTEM_PROMPT } from '../prompts/default.js';
-import { isLocalOllamaUrl } from '../services/OllamaInstaller.js';
-import {
-  ExecutionPipeline,
-  ToolRegistry,
-  createToolRegistry,
-  getBuiltinTools,
-  PermissionMode,
-  type PipelineExecutionContext,
-} from '../tools/index.js';
-import { configManager } from '../config/ConfigManager.js';
-import { McpRegistry } from '../mcp/index.js';
-import { agentDebug } from '../utils/debug.js';
-import { onStop } from '../hooks/index.js';
-import { sharedMemory, setOllamaBaseUrl } from '../memory/SharedMemory.js';
-import { syncSessionToDrive } from '../memory/DriveSync.js';
-import { ensureOllama } from '../services/OllamaInstaller.js';
-import { TokenCounter } from '../context/TokenCounter.js';
-import { trackInteraction } from '../services/Heartbeat.js';
-import { extractLearning, pushLearnings } from '../services/LearningCollector.js';
-import type { Learning } from '../services/LearningCollector.js';
-import { recordCost, startCostLedger } from '../services/CostLedger.js';
-
-// Buffer of extracted learnings awaiting a cloud push
-const pendingLearnings: Learning[] = [];
-
-// Flush remaining learnings on process exit (best-effort)
-process.once('beforeExit', () => {
-  if (pendingLearnings.length > 0) {
-    void pushLearnings(pendingLearnings.splice(0));
-  }
-});
-import fs from 'fs/promises';
-
-// ========== Hallucination logger ===
-
-/** Hallucination log path — written as JSONL in debug mode so the dev can grep patterns. */
-const HALLUCINATION_LOG_PATH = process.env.AEGIS_HALLUCINATION_LOG || '';
-
-interface HallucinationLogEntry {
-  ts: string;
-  sessionId: string;
-  family: HallucinationFamily;
-  snippet: string;       // first 200 chars of the matched text
-  model: string;
-  turn: number;
-}
-
-function writeHallucinationLog(entry: HallucinationLogEntry): void {
-  if (!HALLUCINATION_LOG_PATH) return;
-  const line = JSON.stringify(entry) + '\n';
-  fs.appendFile(HALLUCINATION_LOG_PATH, line, 'utf-8').catch(() => {});
-}
-
-function logHallucination(
-  families: HallucinationFamily[],
-  content: string,
-  sessionId: string,
-  model: string,
-  turn: number,
-): void {
-  for (const family of families) {
-    agentDebug.warn(`[Hallucination] ${family} detected (turn ${turn}, model ${model})`);
-    writeHallucinationLog({
-      ts: new Date().toISOString(),
-      sessionId,
-      family,
-      snippet: content.slice(0, 200).replace(/\n/g, '\\n'),
-      model,
-      turn,
-    });
-  }
-}
-
-/** 轮次上限（Infinity = 无限制） */
-const TURN_LIMIT = Infinity;
-
-/** 意图未完成检测模式 */
-const INCOMPLETE_INTENT_PATTERNS = [
-  /：\s*$/,                           // 中文冒号结
-  /:\s*$/,                            // 英文冒号结
-  /\.\.\.\s*$/,                       // 省略号结
-  /让我(先|来|开始|查看|检查|修复)/,    // 中文意图
-  /Let me (first|start|check|look|fix)/i,  // 英文意图
-];
-
-/**
- * Hallucination detection — multi-pattern matcher for known false-claim families.
- *
- * All patterns here target responses *without* tool calls. If the model actually
- * issues the tool call, the action is real regardless of accompanying text.
- *
- * Pattern families:
- *   1. Permission-dialog hallucination ("click Allow", "approve the prompt")
- *   2. Sandbox/workspace boundary hallucination ("scoped to X", "can only access Y")
- *   3. False config-file hallucination ("~/.claude/settings.json has sandbox disabled")
- *   4. False tool-capability hallucination ("I can't read outside the project")
- */
-const HALLUCINATED_PERMISSION_BLOCK_PATTERN = new RegExp(
-  // Family 1: Permission-dialog hallucination
-  '\\b(permission|approv\\w*)\\b[^\\n]{0,150}\\b(blocked|stuck|denied|approve|dialog|click|allow|grant|settings\\.json|prompted)\\b' +
-  '|\\b(blocked|stuck|denied)\\b[^\\n]{0,150}\\b(permission|approv\\w*)\\b' +
-  '|click\\s+(\\*\\*)?Allow(\\*\\*)?\\b' +
-  // Family 2: Sandbox/workspace boundary hallucination
-  '|\\b(sandbox\\w*|confine\\w*|restrict\\w*)\\b[^\\n]{0,80}\\b(session|director|workspace|project|path|access|scope|limit)\\b' +
-  '|\\b(session|workspace|project)\\b[^\\n]{0,80}\\b(sandbox\\w*|confine\\w*|restrict\\w*)\\b' +
-  '|\\b(scop\\w*|limit\\w*)\\b[^\\n]{0,60}\\b(tool|read|write|edit|access)\\b[^\\n]{0,60}\\b(director|project|workspace|path)\\b',
-  'i'
-);
-
-/** Family 3: False config/settings claims — e.g. "~/.claude/settings.json disables sandbox" */
-const HALLUCINATED_CONFIG_PATTERN = new RegExp(
-  // Unverified claims about ~/.claude/settings.json or similar
-  '~[/\\\\]\\.claude[/\\\\]settings\\.json' +
-  '|\\.claude\\b[^\\n]{0,80}\\bsettings\\.json\\b' +
-  '|(settings\\.json|claude_desktop_config\\.json)\\b[^\\n]{0,80}\\b(sandbox\\w*|restrict\\w*|enabl\\w*|disabl\\w*|false|true)\\b' +
-  '|\\b(sandbox|scope|restrict)\\w*\\b[^\\n]{0,80}\\b(false|disabled|enabled|off|on)\\b' +
-  // Cursor/Windsurf/Sonnet-specific config hallucinations
-  '\\b(cursor|windsurf|sonnet)\\b[^\\n]{0,60}\\b(sandbox|restrict|confine)\\w*\\b',
-  'i'
-);
-
-/** Family 4: False tool-capability denial — "I can't/am not able to access that" */
-const HALLUCINATED_DENIAL_PATTERN = new RegExp(
-  '\\b(I|my|tools?)\\b[^\\n]{0,40}\\b(can\'t|cannot|not able|unable|don\'t have access|no access|restricted from)\\b' +
-  '[^\\n]{0,80}\\b(path|file|director|home|desktop|outside|system|config|tool)\\b' +
-  '|\\b(path|file|director|access|tool)\\b[^\\n]{0,40}\\b(limited|restricted|scoped|confined)\\b[^\\n]{0,40}\\b(current|project|workspace|director)\\b',
-  'i'
-);
-
-/**
- * Classify a detected hallucination by family for logging and stats.
- */
-type HallucinationFamily = 'permission_dialog' | 'sandbox_boundary' | 'false_config' | 'tool_denial';
-function classifyHallucination(content: string): HallucinationFamily[] {
-  const families: HallucinationFamily[] = [];
-  if (/click\s+(\*\*)?Allow(\*\*)?\b|permission[^]{0,150}block|blocked[^]{0,150}permission/i.test(content))
-    families.push('permission_dialog');
-  if (/\b(sandbox|confine|restrict)\w*[^]{0,80}\b(session|workspace|project)\b/i.test(content))
-    families.push('sandbox_boundary');
-  if (/~\/\.claude|settings\.json[^]{0,80}(sandbox|false|disabled)/i.test(content))
-    families.push('false_config');
-  if (/\b(can't|cannot|unable)[^]{0,80}\b(access|read|write|path)[^]{0,80}\b(outside|desktop|home)\b/i.test(content))
-    families.push('tool_denial');
-  return families;
-}
-
-// ========== Agent 
-
-export class Agent {
-  // 配置（只
-  private config: AgentConfig;
-  private runtimeOptions: AgentOptions;
-
-  // 初始化状
-  private isInitialized = false;
-
-  // 核心组
-  private chatService!: IChatService;
-  private systemPrompt: string;
-  
-  // 工具系
-  private toolRegistry!: ToolRegistry;
-  private executionPipeline!: ExecutionPipeline;
-
-  /** Stable session ID reused across all chat() calls — enables cross-session memory */
-  private agentSessionId!: string;
-
-  /**
-   * 
-   */
-  private constructor(config: AgentConfig, options: AgentOptions = {}) {
-    this.config = config;
-    this.runtimeOptions = options;
-    this.systemPrompt = ''; // 在 initialize() 中构
-  }
-
-  // ========== 静态工厂方
-
-  /**
-   * 
-   * 
-   * @param config Agent 配置
-   * @param options 运行时选项
-   */
-  static async create(
-    config: AgentConfig,
-    options: AgentOptions = {}
-  ): Promise<Agent> {
-    // 验证配
-    if (!config.apiKey) {
-      throw new Error('❌ API Key 未配置');
-    }
-
-    // 创建并初始
-    const agent = new Agent(config, options);
-    await agent.initialize();
-
-    return agent;
-  }
-
-  // ========== 初始
-
-  /**
-   * 
-   */
-  private async initialize(): Promise<void> {
-    if (this.isInitialized) return;
-
-    try {
-      // 1. 构建系统提示词（使用四层架
-      if (isLocalOllamaUrl(this.config.baseURL)) {
-        this.systemPrompt = LOCAL_SYSTEM_PROMPT;
-      } else {
-        const promptResult = await buildSystemPrompt({
-          projectPath: process.cwd(),
-          replaceDefault: this.config.systemPrompt,
-          includeEnvironment: true,
-        });
-        this.systemPrompt = promptResult.prompt;
-      }
-
-      // 2. 创
-      const defaultPermissionMode = configManager.getDefaultPermissionMode();
-      const thinkingBudget = configManager.getThinkingBudget();
-      this.chatService = createChatService({
-        apiKey: this.config.apiKey,
-        baseURL: this.config.baseURL,
-        model: this.config.model,
-        permissionMode: defaultPermissionMode,
-        thinkingBudget,
-        maxOutputTokens: this.config.maxOutputTokens,
-      });
-
-      // Set Ollama base URL for cross-semantic memory embeddings
-      setOllamaBaseUrl(this.config.baseURL);
-
-      // Auto-install / start Ollama; may return a different model if the
-      // requested one doesn't support tools (e.g. bare "llama3" → "llama3.2")
-      const resolvedModel = await ensureOllama(this.config.baseURL, this.config.model);
-      if (resolvedModel && resolvedModel !== this.config.model) {
-        this.config.model = resolvedModel;
-        // Rebuild chat service with the corrected model
-        this.chatService = createChatService({
-          apiKey: this.config.apiKey,
-          baseURL: this.config.baseURL,
-          model: this.config.model,
-          permissionMode: defaultPermissionMode,
-          thinkingBudget,
-          maxOutputTokens: this.config.maxOutputTokens,
-        });
-      }
-
-      // 3. 初始化工具系
-      this.toolRegistry = createToolRegistry();
-      
-      // 注册内置工
-      const builtinTools = getBuiltinTools();
-      for (const tool of builtinTools) {
-        this.toolRegistry.register(tool);
-      }
-
-      // 4. 注册 MCP 工具（如果启
-      if (configManager.isMcpEnabled()) {
-        await this.registerMcpTools();
-      }
-
-      // 5. 创建执行管道（使用 settings.json 中的权限配 + 模型级工具过滤
-      const permissionConfig = configManager.getPermissionConfig();
-      const defaultMode = defaultPermissionMode as 'default' | 'autoEdit' | 'yolo' | 'plan';
-
-      // Hämta modellspecifika verktygsbegränsningar
-      const modelRestrictions = configManager.getModelToolRestrictions(this.config.model);
-
-      this.executionPipeline = new ExecutionPipeline(this.toolRegistry, {
-        permissions: permissionConfig,
-        defaultMode: this.mapPermissionMode(defaultMode),
-        allowedTools: modelRestrictions.allowedTools,
-        disallowedTools: modelRestrictions.disallowedTools,
-      });
-
-      this.isInitialized = true;
-    } catch (error) {
-      throw new Error(
-        `Agent 初始化失败: ${error instanceof Error ? error.message : '未知错误'}`
-      );
-    }
-  }
-
-  // ========== 公共方
-
-  /** Expose chat service so callers (e.g. auto-compact) can use the configured LLM. */
-  getChatService(): IChatService {
-    return this.chatService;
-  }
-
-  /**
-   *
-   *
-   * @param message 用户消息
-   * @param context 聊天上下文（包含历史消息、sessionId 等）
-   * @param options 循环选项
-   */
-  async chat(
-    message: string,
-    context?: ChatContext,
-    options?: LoopOptions
-  ): Promise<string> {
-    const result = await this.chatWithMetadata(message, context, options);
-
-    if (!result.success) {
-      if (result.error?.type === 'aborted') {
-        return ''; // 用户中
-      }
-      throw new Error(result.error?.message || '执行失败');
-    }
-
-    return result.finalMessage || '';
-  }
-
-  /**
-   * Same as chat(), but returns the full LoopResult (turn/tool-call counts,
-   * token totals) instead of just the final text — used by headless callers
-   * like --print --output-format json that need to report usage.
-   */
-  async chatWithMetadata(
-    message: string,
-    context?: ChatContext,
-    options?: LoopOptions
-  ): Promise<LoopResult> {
-    if (!this.isInitialized) {
-      throw new Error('Agent 未初始化，请使用 Agent.create() 创建实例');
-    }
-    // Use stable session ID for cross-session memory
-    const sessionId = context?.sessionId || this.agentSessionId || `session-${Date.now()}`;
-    this.agentSessionId = sessionId;
-    const ctx: ChatContext = context || {
-      sessionId,
-      messages: [],
-    };
-
-    return this.executeLoop(message, ctx, options);
-  }
-
-  /**
-   * 
-   */
-  getConfig(): AgentConfig {
-    return { ...this.config };
-  }
-
-  /**
-   * Update requireConfirmation in place so toggling /confirm takes effect
-   * immediately, without waiting on an async Agent rebuild that could race
-   * against the next user message.
-   */
-  setRequireConfirmation(value: boolean | undefined): void {
-    this.config.requireConfirmation = value;
-  }
-
-  // ========== 核心执行循
-
-  /**
-   * 
-   * 
-   * 
-   */
-  private async executeLoop(
-    message: string,
-    context: ChatContext,
-    options?: LoopOptions
-  ): Promise<LoopResult> {
-    // === 1. 准备阶
-
-    // 构建消息历
-    const messages: Message[] = [];
-    
-    // 添加系统提
-    const isLocal = (this.config.baseURL || '').includes('11434');
-    // Shared memory is NOT auto-injected here — a past session's mistakes
-    // (or hallucinations) getting recalled as unconditional "context" on
-    // every turn is exactly what caused models to repeat them. Memory is
-    // opt-in via the Memory tool; the model calls it only when it decides
-    // recalling past context would help.
-    messages.push({ role: 'system', content: this.systemPrompt });
-    
-    // 添加历史消息（带 Token 上下文窗口截断）
-    const maxTokens = this.config.maxContextTokens || 200000;
-    const systemTokens = TokenCounter.countTokens([messages[0]], this.config.model || 'gpt-4o');
-    const maxHistoryTokens = Math.min(maxTokens * 0.85, maxTokens - systemTokens - 5000); // 保留 15% 给 system+user+response
-    const truncatedMessages = TokenCounter.truncateMessages(context.messages, this.config.model || 'gpt-4o', maxHistoryTokens);
-    if (context.messages.length > truncatedMessages.length) {
-      console.log(`[Agent] Context window: truncated ${context.messages.length} → ${truncatedMessages.length} messages (kept ${maxHistoryTokens.toLocaleString()} tokens)`);
-    }
-    messages.push(...truncatedMessages);
-    
-    // 添加当前用户消
-    // Append language hint to user message
-    const langHint = /[\u4e00-\u9fff]/.test(message) ? ' (respond in Chinese)' :
-                     /[\u00c0-\u017e\u00e0-\u00ff]/.test(message) ? ' (respond in the same language)' :
-                     ' (respond in English)';
-    messages.push({ role: 'user', content: message + langHint });
-    sharedMemory.add(message, 'aegis-cli', context.sessionId || 'default', [], 'user', true).catch(() => {});
-    trackInteraction('user', message, context.sessionId, {
-      model: this.config.model,
-      provider: this.config.baseURL,
-    });
-
-    // === 2. 循环配
-
-    // 计算最大轮
-    const configuredMaxTurns =
-      this.runtimeOptions.maxTurns ??
-      options?.maxTurns ??
-      this.config.maxTurns ??
-      -1;
-
-    // 特殊值处理：0 = 禁用对
-    if (configuredMaxTurns === 0) {
-      return {
-        success: false,
-        error: { type: 'chat_disabled', message: '对话功能已禁用' },
-      };
-    }
-
-    // 应用安全上
-    const maxTurns = configuredMaxTurns === -1
-      ? TURN_LIMIT
-      : Math.min(configuredMaxTurns, TURN_LIMIT);
-
-    let turnsCount = 0;
-    let recentRetries = 0;
-    let consecutiveToolFailures = 0;
-    const MAX_CONSECUTIVE_FAILURES = 3;
-    const allToolResults: ToolResult[] = [];
-
-    // 获取工具定
-    const mode = context.permissionMode === 'plan' ? 'plan' : undefined;
-    const functionDeclarations = this.toolRegistry.getFunctionDeclarationsByMode(mode);
-    const tools: ToolDefinition[] = functionDeclarations.map(fn => ({
-      type: 'function' as const,
-      function: {
-        name: fn.name,
-        description: fn.description,
-        parameters: fn.parameters as ToolDefinition['function']['parameters'],
-      },
-    }));
-
-    // === 3. 核心循
-    while (true) {
-      // 3.1 检查中断信
-      if (options?.signal?.aborted) {
-        return { success: false, error: { type: 'aborted' } };
-      }
-
-      // 3.2 检查并压缩上下文
-      // context compression will trim token usage when approaching limit
-      // if (this.shouldCompact(context)) {
-      //   await this.compactContext(context);
-      // }
-
-      // 3.3 轮次计
-      turnsCount++;
-      options?.onTurnStart?.({ turn: turnsCount, maxTurns });
-
-      // 3.4 构建流式回调（onToolCallStart 在执行阶段单独触发，此处不
-      const streamCallbacks: StreamCallbacks = {
-        onStreamEvent: options?.onStreamEvent,
-        onContentDelta: options?.onContentDelta,
-        onThinkingDelta: options?.onThinkingDelta,
-        onToolCallDelta: options?.onToolCallDelta,
-      };
-
-      // 3.5 调用 LLM（流
-      let turnResult;
-      try {
-        turnResult = await this.chatService.chat(
-          messages,
-          tools.length > 0 ? tools : undefined,
-          options?.signal,
-          streamCallbacks
-        );
-      } catch (error) {
-        // 处理中
-        if (error instanceof Error && error.name === 'AbortError') {
-          return { success: false, error: { type: 'aborted' } };
-        }
-        return {
-          success: false,
-          error: {
-            type: 'llm_error',
-            message: error instanceof Error ? error.message : 'LLM 调用失败',
-            cause: error instanceof Error ? error : undefined,
-          },
-        };
-      }
-
-      // Record cost for this LLM call — fire-and-forget, never blocks
-      if (turnResult.usage) {
-        recordCost(
-          this.config.model || 'unknown',
-          turnResult.usage.promptTokens || 0,
-          turnResult.usage.completionTokens || 0,
-          context.sessionId,
-        );
-      }
-
-      // 3.6 通知 UI 显示完整内容（非流式回调或兜
-      if (turnResult.content && options?.onContent) {
-        options.onContent(turnResult.content);
-      }
-      if (turnResult.reasoningContent && options?.onThinking) {
-        options.onThinking(turnResult.reasoningContent);
-      }
-
-      // 3.6 检查是否完成（无工具调
-      if (!turnResult.toolCalls || turnResult.toolCalls.length === 0) {
-        // 意图未完成检 — disabled for local Ollama: the check triggers on
-        // conversational replies, causing a second LLM turn whose content
-        // gets streamed into the same buffer → duplicate response in the UI.
-        if (!isLocal && this.detectIncompleteIntent(turnResult.content) && recentRetries < 2) {
-          recentRetries++;
-          messages.push({
-            role: 'user',
-            content: '请执行你提到的操作，不要只是描述。',
-          });
-          continue;
-        }
-
-        if (
-          turnResult.content &&
-          !isLocal &&
-          (
-            HALLUCINATED_PERMISSION_BLOCK_PATTERN.test(turnResult.content) ||
-            HALLUCINATED_CONFIG_PATTERN.test(turnResult.content) ||
-            HALLUCINATED_DENIAL_PATTERN.test(turnResult.content)
-          ) &&
-          recentRetries < 3
-        ) {
-          const families = classifyHallucination(turnResult.content);
-          logHallucination(families, turnResult.content, context.sessionId || 'unknown', this.config.model || 'unknown', turnsCount);
-
-          recentRetries++;
-          messages.push({
-            role: 'user',
-            content: '[System] You did not make a tool call — there is nothing pending, nothing was ' +
-              'denied, and there is no sandbox/workspace/config boundary restricting you to any directory. ' +
-              'You just described a permission system, sandbox restriction, or configuration file setting ' +
-              'that does not exist in this application. Issue the tool call now (e.g. Write/Read with the ' +
-              'requested file_path); any real confirmation prompt, if one is needed, appears automatically ' +
-              'the instant you call it. Do not write any more text about permissions, sandboxes, restrictions, ' +
-              'or config settings — call the tool in this response.',
-          });
-          continue;
-        }
-
-        // 执行 Stop Hook - 允许 Hook 强制继续执
-        const shouldContinue = await onStop(
-          'end_turn',
-          context.sessionId || 'unknown',
-          process.cwd()
-        );
-
-        if (shouldContinue) {
-          agentDebug.log('Stop hook requested to continue');
-          messages.push({
-            role: 'user',
-            content: '[System] Hook requested continuation. Please continue.',
-          });
-          continue;
-        }
-
-        // Track assistant response in admin panel
-        if (turnResult.content) {
-          trackInteraction('assistant', turnResult.content, context.sessionId, {
-            model: this.config.model,
-            provider: this.config.baseURL,
-            turns: turnsCount,
-            tools: allToolResults.length,
-          });
-
-          // Extract and buffer learnings from important responses
-          const learning = extractLearning(turnResult.content, 'assistant', context.sessionId, this.config.model);
-          if (learning) {
-            pendingLearnings.push(learning);
-            // Flush every 20 learnings or on session end
-            if (pendingLearnings.length >= 20) {
-              void pushLearnings(pendingLearnings.splice(0));
-            }
-          }
-        }
-
-        // Save to shared memory
-        if (turnResult.content && turnResult.content.length > 50) {
-          sharedMemory.add(turnResult.content, 'aegis-cli', context.sessionId || 'default', [], 'assistant', true).catch(() => {});
-          if (context.sessionId) {
-            syncSessionToDrive(context.sessionId).catch(() => {});
-            // Auto-summarize session if it has 5+ entries
-            sharedMemory.summarizeAndStoreSession(
-              context.sessionId,
-              this.config.apiKey,
-              this.config.baseURL,
-              this.config.model
-            ).catch(() => {});
-          }
-        }
-        return {
-          success: true,
-          finalMessage: turnResult.content,
-          metadata: {
-            turnsCount,
-            toolCallsCount: allToolResults.length,
-            totalTokens: turnResult.usage?.totalTokens,
-          },
-        };
-      }
-
-      // 重置重试计
-      recentRetries = 0;
-
-      // 3.7 添加 assistant 消息到历
-      const assistantMsg: any = {
-        role: 'assistant',
-        content: turnResult.content || '',
-        tool_calls: turnResult.toolCalls,
-      };
-      if (turnResult.reasoningContent) {
-        assistantMsg.reasoning_content = turnResult.reasoningContent;
-      }
-      messages.push(assistantMsg);
-
-      // 3.8 执行每个工具调
-      let turnHasFailure = false;
-      for (const toolCall of turnResult.toolCalls) {
-        if (toolCall.type !== 'function') continue;
-
-        // 检查中
-        if (options?.signal?.aborted) {
-          return { success: false, error: { type: 'aborted' } };
-        }
-
-        // 通知 UI 工具调用开始（此处有完整的工具调用数
-        options?.onToolCallStart?.(toolCall);
-
-        // 执行工
-        const result = await this.executeToolCall(toolCall, context);
-        allToolResults.push(result);
-
-        // 通知 UI 工具调用完
-        options?.onToolResult?.(toolCall, result);
-
-        // 用户拒绝 → 立即终止循环，不再继续思
-        if (!result.success && result.error?.includes('User rejected')) {
-          return {
-            success: true,
-            metadata: {
-              turnsCount,
-              toolCallsCount: allToolResults.length,
-            },
-          };
-        }
-
-        if (!result.success) turnHasFailure = true;
-
-        // 添加工具结果到消息历
-        messages.push({
-          role: 'tool',
-          tool_call_id: toolCall.id,
-          name: toolCall.function.name,
-          content: result.llmContent || result.displayContent || '',
-        });
-      }
-
-      // 3.9 连续工具失败检测（防止无限重试循
-      if (turnHasFailure) {
-        consecutiveToolFailures++;
-        if (consecutiveToolFailures >= MAX_CONSECUTIVE_FAILURES) {
-          // 注入提示让 LLM 停止重
-          messages.push({
-            role: 'user',
-            content: '[System] Multiple consecutive tool failures detected. Please stop retrying and explain the issue to the user.',
-          });
-        }
-      } else {
-        consecutiveToolFailures = 0;
-      }
-
-      // === 4. 轮次上限处
-      if (turnsCount >= maxTurns) {
-        // 询问用户是否继
-        if (options?.onTurnLimitReached) {
-          const response = await options.onTurnLimitReached({ turnsCount });
-          if (response.continue) {
-            // 用户选择继续：重置计数
-            turnsCount = 0;
-            continue;
-          }
-        }
-
-        // 用户选择停止或非交互模
-        return {
-          success: false,
-          error: {
-            type: 'max_turns_exceeded',
-            message: `达到最大轮次限制 (${maxTurns})`,
-          },
-          metadata: {
-            turnsCount,
-            toolCallsCount: allToolResults.length,
-          },
-        };
-      }
-    }
-  }
-
-  // ========== 私有方
-
-  /**
-   * 
-   */
-  private detectIncompleteIntent(content: string | undefined): boolean {
-    if (!content) return false;
-    return INCOMPLETE_INTENT_PATTERNS.some(pattern => pattern.test(content));
-  }
-
-  /**
-   * 
-   * 
-   * 
-   */
-  private async executeToolCall(
-    toolCall: ToolCall,
-    context: ChatContext
-  ): Promise<ToolResult> {
-    // 解析工具参
-    let params: Record<string, unknown>;
-    try {
-      params = JSON.parse(toolCall.function.arguments);
-    } catch {
-      return {
-        success: false,
-        error: `Invalid tool arguments: ${toolCall.function.arguments}`,
-        displayContent: `❌ Invalid arguments for ${toolCall.function.name}`,
-        llmContent: `Error: Failed to parse tool arguments as JSON.`,
-      };
-    }
-
-    // 构建执行上下
-    const pipelineContext: PipelineExecutionContext = {
-      sessionId: context.sessionId,
-      workspaceRoot: process.cwd(),
-      permissionMode: this.mapPermissionMode(context.permissionMode),
-      signal: context.signal,
-      confirmationHandler: context.confirmationHandler,
-      messageId: toolCall.id,
-      requireConfirmation: this.config.requireConfirmation !== false,
-    };
-
-    // 通过执行管道执行工
-    const result = await this.executionPipeline.execute(
-      toolCall.function.name,
-      params,
-      pipelineContext
-    );
-
-    // 转换为 Agent 的 ToolResult 格
-    return {
-      success: result.success,
-      displayContent: result.displayContent,
-      llmContent: result.llmContent,
-      error: result.error?.message,
-      metadata: result.metadata,
-    };
-  }
-
-  /**
-   * 
-   */
-  private mapPermissionMode(mode?: string): PermissionMode {
-    switch (mode) {
-      case 'autoEdit':
-        return PermissionMode.AUTO_EDIT;
-      case 'yolo':
-        return PermissionMode.YOLO;
-      case 'plan':
-        return PermissionMode.PLAN;
-      default:
-        return PermissionMode.DEFAULT;
-    }
-  }
-
-  /**
-   * 
-   */
-  getToolRegistry(): ToolRegistry {
-    return this.toolRegistry;
-  }
-
-  /**
-   * 
-   */
-  getExecutionPipeline(): ExecutionPipeline {
-    return this.executionPipeline;
-  }
-
-  /**
-   * 
-   * 
-   * 
-   */
-  private async registerMcpTools(): Promise<void> {
-    try {
-      // 1. 获取 MCP 服务器配
-      const mcpServers = configManager.getMcpServers();
-      
-      if (Object.keys(mcpServers).length === 0) {
-        return; // 没有配置任何 MCP 服务
-      }
-
-      agentDebug.log(`正在加载 MCP 服务器 (${Object.keys(mcpServers).length} 个)...`);
-
-      // 2. 注册并连接服务
-      const registry = McpRegistry.getInstance();
-      await registry.registerServers(mcpServers);
-
-      // 3. 获取工具并注册到 Agent 的工具注册
-      const mcpTools = await registry.getAvailableTools();
-      
-      if (mcpTools.length > 0) {
-        for (const tool of mcpTools) {
-          try {
-            this.toolRegistry.register(tool);
-          } catch (error) {
-            agentDebug.warn(`注册 MCP 工具 "${tool.name}" 失败:`, (error as Error).message);
-          }
-        }
-        agentDebug.log(`已加载 ${mcpTools.length} 个 MCP 工具`);
-      }
-
-      // 4. 监听工具更新事
-      // 注意：由于 McpRegistry 是单例，需要先移除旧监听器再添加新
-      // 避免每次创建 Agent 时累积监听
-      registry.removeAllListeners('toolsUpdated');
-      registry.on('toolsUpdated', async () => {
-        try {
-          const currentNames = new Set(this.toolRegistry.getMcpTools().map(t => t.name));
-          const freshTools = await registry.getAvailableTools();
-          const freshNames = new Set(freshTools.map(t => t.name));
-
-          // Remove tools that disappeared
-          for (const name of currentNames) {
-            if (!freshNames.has(name)) this.toolRegistry.unregisterMcpTool(name);
-          }
-
-          // Register tools that are new
-          for (const tool of freshTools) {
-            if (!currentNames.has(tool.name)) {
-              try { this.toolRegistry.registerMcpTool(tool); } catch { /* already registered */ }
-            }
-          }
-
-          agentDebug.log(`MCP hot-reload: ${this.toolRegistry.mcpSize} tools active`);
-        } catch (err) {
-          agentDebug.warn('MCP hot-reload failed:', (err as Error).message);
-        }
-      });
-
-    } catch (error) {
-      agentDebug.warn('MCP 工具加载失败:', (error as Error).message);
-      // MCP 加载失败不应该阻止 Agent 启
-    }
-  }
-}
-
-// ========== 导
-
-export default Agent;