metona-ai-desktop/electron/ipc/agent.ts

/**
 * IPC Agent Handlers — Agent 交互域（P2-9 从 handlers.ts 拆分）
 *
 * 职责：
 * 1. agent:sendMessage — 消息发送编排（历史加载/记忆注入/注入检测/runStream/持久化）
 * 2. agent:abortSession — 中断会话（联动 SubAgent）
 * 3. 常驻引擎事件管道 — 流式转发（按会话节流）、状态广播、TRACE 录制（P1-6 补全 9 种事件）、
 *    压缩/死循环/Provider 切换通知
 *
 * P2-10: 事件监听从"每消息 attach/detach"改为常驻管道（按 sessionId 隔离状态），
 * 支持多会话并发流式与多窗口广播。
 */

import { ipcMain } from 'electron';
import type { IPCContext } from './context';
import { broadcast } from './context';
import type {
  MetonaMessage,
  MetonaStreamEvent,
  MetonaError,
  MetonaSystemPrompt,
} from '../harness/types';
import { MetonaErrorCode, MetonaStreamEventType } from '../harness/types';
import { estimateMessagesTokens } from '../harness/utils/token-estimator';
import { DeepSeekAdapter } from '../harness/adapters/deepseek.adapter';
import { OllamaAdapter } from '../harness/adapters/ollama.adapter';
// v0.7.3 P1-1: 用户上下文前置块（动态内容出 system，保 prompt cache 前缀稳定）
import { buildUserContextPrefix, withUserContextPrefix } from '../harness/prompts/user-context';
// v0.7.3 P1-5: 记忆固化触发决策（纯函数）
import { shouldConsolidate } from '../harness/memory/consolidation-policy';
import log from 'electron-log';

/** 构建 "时区名 (UTC±N)" 标签（注入用户上下文前置块；失败回退 UTC） */
function buildTimezoneLabel(): string {
  try {
    const tz = Intl.DateTimeFormat().resolvedOptions().timeZone ?? 'UTC';
    const offset = -new Date().getTimezoneOffset() / 60;
    const offsetStr = offset >= 0 ? `UTC+${offset}` : `UTC${offset}`;
    return `${tz} (${offsetStr})`;
  } catch {
    return 'UTC';
  }
}

/** 单会话的 text_delta 节流状态 */
interface ThrottleState {
  buffer: string;
  lastEventMeta: Pick<
    MetonaStreamEvent,
    'requestId' | 'sessionId' | 'iteration' | 'seq' | 'timestamp' | 'runId'
  > | null;
  flushTimer: ReturnType<typeof setTimeout> | null;
}

/** 单会话的迭代录制状态（TRACE 层） */
interface IterationTrace {
  iteration: number;
  startedAt: number;
  text: string;
  usage?: { input: number; output: number; total: number };
  responded: boolean; // 本轮 llm_response 是否已记录（PARSING 与下一轮 THINKING 去重）
}

export function registerAgentHandlers(ctx: IPCContext): void {
  const {
    agentEngineManager,
    sessionRecorder,
    configService,
    sessionService,
    workspaceService,
    contextBuilder,
    auditService,
    memoryManager,
    promptInjectionDefender,
    outputValidator,
    memoryConsolidator,
    sessionSummaryService,
    orchestrator,
    confirmationHook,
    titleGenerator,
  } = ctx;

  // ===== 常驻事件管道：text_delta 按会话节流（F8） =====
  const throttleStates = new Map<string, ThrottleState>();
  const iterationTraces = new Map<string, IterationTrace>();

  // v0.7.3 P1-5: 会话级记忆固化时间戳（consolidation-policy 频率门控的状态持有方）
  // v0.7.4 P3-5: LRU 化 —— 旧实现每会话一条永不清除，长期运行后无界增长。
  // 容量 200：超过后淘汰最旧的（固化频率窗口 10 分钟，200 个会话远超实际并发上限）。
  const lastConsolidationBySession = new Map<string, number>();
  const MAX_CONSOLIDATION_TRACKED = 200;
  const touchConsolidation = (sessionId: string): number => {
    const now = Date.now();
    lastConsolidationBySession.set(sessionId, now);
    if (lastConsolidationBySession.size > MAX_CONSOLIDATION_TRACKED) {
      // 淘汰最旧（Map 迭代顺序 = 插入顺序）
      const oldestKey = lastConsolidationBySession.keys().next().value as string | undefined;
      if (oldestKey !== undefined) lastConsolidationBySession.delete(oldestKey);
    }
    return now;
  };

  const flushThrottle = (sessionId: string): void => {
    const st = throttleStates.get(sessionId);
    if (!st) return;
    st.flushTimer = null;
    if (!st.buffer || !st.lastEventMeta) {
      st.buffer = '';
      st.lastEventMeta = null;
      return;
    }
    // 构建合并的 text_delta 事件，保留最后一个 delta 的元数据
    const mergedEvent: MetonaStreamEvent = {
      ...st.lastEventMeta,
      type: MetonaStreamEventType.TEXT_DELTA,
      delta: st.buffer,
      seq: st.lastEventMeta.seq,
      timestamp: Date.now(),
    };
    broadcast('agent:streamEvent', mergedEvent);
    st.buffer = '';
    st.lastEventMeta = null;
  };

  const cleanupSessionState = (sessionId: string): void => {
    const st = throttleStates.get(sessionId);
    if (st?.flushTimer) {
      clearTimeout(st.flushTimer);
      flushThrottle(sessionId);
    }
    throttleStates.delete(sessionId);
    iterationTraces.delete(sessionId);
  };

  // ===== 常驻监听：流式事件（节流转发 + TRACE 录制） =====
  agentEngineManager.on('streamEvent', (event: MetonaStreamEvent) => {
    if (!event.sessionId) return;
    const sessionId = event.sessionId;
    const trace = iterationTraces.get(sessionId);

    // F8: text_delta 聚合，其他事件立即转发（先 flush 保证顺序）
    if (event.type === MetonaStreamEventType.TEXT_DELTA && event.delta) {
      const st = throttleStates.get(sessionId) ?? {
        buffer: '',
        lastEventMeta: null,
        flushTimer: null,
      };
      throttleStates.set(sessionId, st);
      if (st.buffer === '') {
        st.lastEventMeta = {
          requestId: event.requestId,
          sessionId: event.sessionId,
          iteration: event.iteration,
          seq: event.seq,
          timestamp: event.timestamp,
          runId: event.runId,
        };
      } else if (st.lastEventMeta) {
        st.lastEventMeta = {
          ...st.lastEventMeta,
          seq: event.seq,
          timestamp: event.timestamp,
          runId: event.runId,
        };
      }
      st.buffer += event.delta;
      if (st.flushTimer === null) {
        st.flushTimer = setTimeout(() => flushThrottle(sessionId), 32);
      }
      // TRACE: 累积本轮 LLM 文本（供 llm_response 记录）
      if (trace) trace.text += event.delta;
      return;
    }

    // 非 text_delta 事件：先 flush 缓冲区，再立即转发（保证事件顺序）
    const st = throttleStates.get(sessionId);
    if (st?.flushTimer) {
      clearTimeout(st.flushTimer);
      flushThrottle(sessionId);
    }

    // TRACE: 工具调用/结果/usage 录制（P1-6 补全）
    switch (event.type) {
      case MetonaStreamEventType.TOOL_CALL_COMPLETE:
        if (event.toolCall) {
          sessionRecorder.recordToolCall({
            sessionId,
            iteration: event.iteration,
            toolName: event.toolCall.name,
            args: event.toolCall.args,
          });
        }
        break;
      case MetonaStreamEventType.TOOL_RESULT:
        if (event.toolResult) {
          const resultPreview =
            typeof event.toolResult.result === 'string'
              ? event.toolResult.result
              : JSON.stringify(event.toolResult.result);
          sessionRecorder.recordToolResult({
            sessionId,
            iteration: event.iteration,
            toolName: event.toolResult.toolName,
            success: event.toolResult.success,
            durationMs: event.toolResult.durationMs ?? 0,
            resultPreview,
            error: event.toolResult.error,
          });
        }
        break;
      case MetonaStreamEventType.USAGE:
        if (trace && event.usage) {
          trace.usage = {
            input: event.usage.inputTokens ?? 0,
            output: event.usage.outputTokens ?? 0,
            total: event.usage.totalTokens ?? 0,
          };
        }
        break;
      case MetonaStreamEventType.DONE:
        // v0.7.4 P1-6: DONE 不再执行 cleanupSessionState —— 引擎的 finish() 顺序是
        // 先发 DONE 再发 TERMINATED stateChange，而 TERMINATED 分支（stateChange 监听）
        // 需要读取 iterationTraces 补记最终迭代的 iteration_end/llm_response。
        // 旧实现在 DONE 时删除 iterationTraces → 最终迭代的 TRACE 完整性缺失。
        // 现仅在 TERMINATED 分支统一收尾（flush throttle buffer + 删 iterationTraces）。
        // 注意：此处仍需转发 DONE 事件到前端（下方 broadcast 统一执行）。
        break;
      default:
        break;
    }

    broadcast('agent:streamEvent', event);
  });

  // ===== 常驻监听：状态变化（广播 + TRACE 迭代录制） =====
  agentEngineManager.on(
    'stateChange',
    (data: {
      previous?: string;
      current?: string;
      state?: string;
      sessionId?: string;
      iteration?: number;
      runId?: string;
    }) => {
      if (data.previous) log.info(`[AGENT] State: ${data.previous} → ${data.current}`);
      broadcast('agent:stateChange', data);

      const sessionId = data.sessionId;
      if (!sessionId || data.iteration == null) return;
      const stateValue = data.state ?? data.current ?? '';
      const trace = iterationTraces.get(sessionId);

      // THINKING 且迭代号变化 → 新迭代开始（关闭上一迭代）
      if (stateValue === 'THINKING' && (!trace || trace.iteration !== data.iteration)) {
        if (trace && !trace.responded) {
          sessionRecorder.recordLLMResponse({
            sessionId,
            iteration: trace.iteration,
            content: trace.text,
            finishReason: 'stop',
            tokenUsage: trace.usage ?? { input: 0, output: 0, total: 0 },
          });
        }
        if (trace) {
          sessionRecorder.recordIterationEnd(sessionId, {
            iteration: trace.iteration,
            durationMs: Date.now() - trace.startedAt,
          });
        }
        iterationTraces.set(sessionId, {
          iteration: data.iteration,
          startedAt: Date.now(),
          text: '',
          responded: false,
        });
        sessionRecorder.recordIterationStart(sessionId, data.iteration);
        const provider = configService.get<string>('llm.provider') ?? '';
        const model = configService.get<string>('llm.model') ?? '';
        sessionRecorder.recordLLMRequest({
          sessionId,
          iteration: data.iteration,
          provider,
          model,
          messageCount: data.iteration + 1,
        });
      }

      // PARSING → 本轮流式结束，记录 llm_response
      if (
        stateValue === 'PARSING' &&
        trace &&
        trace.iteration === data.iteration &&
        !trace.responded
      ) {
        trace.responded = true;
        sessionRecorder.recordLLMResponse({
          sessionId,
          iteration: trace.iteration,
          content: trace.text,
          finishReason: 'stop',
          tokenUsage: trace.usage ?? { input: 0, output: 0, total: 0 },
        });
      }

      // TERMINATED → 补记最终迭代的 iteration_end（正常流程只在下一轮 THINKING 补记，
      // 最终轮无后续迭代，需在此补齐 TRACE 完整性）+ 兜底清理会话管道状态
      if (stateValue === 'TERMINATED') {
        if (trace) {
          sessionRecorder.recordIterationEnd(sessionId, {
            iteration: trace.iteration,
            durationMs: Date.now() - trace.startedAt,
          });
        }
        cleanupSessionState(sessionId);
      }
    },
  );

  // ===== 常驻监听：上下文压缩（toast + streamEvent 通知） =====
  agentEngineManager.on(
    'compressed',
    (data: {
      sessionId?: string;
      iteration?: number;
      originalTokens?: number;
      compressedTokens?: number;
    }) => {
      const savedTokens = Math.max(0, (data.originalTokens ?? 0) - (data.compressedTokens ?? 0));
      // toast 通知用户压缩已发生
      broadcast('toast:show', {
        type: 'info',
        message: `上下文压缩: ${data.originalTokens ?? '?'} → ${data.compressedTokens ?? '?'} tokens（节省 ${savedTokens}）`,
      });
      // 通过 streamEvent 转发，前端 useAgentStream 监听 'compressed' 类型后更新 store
      broadcast('agent:streamEvent', {
        type: 'compressed',
        sessionId: data.sessionId ?? '',
        iteration: data.iteration ?? 0,
        originalTokens: data.originalTokens,
        compressedTokens: data.compressedTokens,
        savedTokens,
        timestamp: Date.now(),
      });
    },
  );

  // ===== 常驻监听：死循环检测（toast 警告） =====
  agentEngineManager.on('deadLoop', (data: { iteration?: number; sessionId?: string }) => {
    log.warn(`[AGENT] Dead loop detected at iteration ${data.iteration ?? '?'}`);
    broadcast('toast:show', {
      type: 'warning',
      message: `检测到死循环（第 ${data.iteration ?? '?'} 轮）：连续3轮重复相同工具调用，已自动终止`,
    });
  });

  // ===== 常驻监听：Provider 故障转移（P1，通知前端 + toast） =====
  agentEngineManager.on(
    'providerSwitched',
    (data: { from?: string; to?: string; reason?: string; sessionId?: string }) => {
      broadcast('agent:providerSwitched', {
        from: data.from,
        to: data.to,
        reason: data.reason ?? 'failover',
        sessionId: data.sessionId ?? '',
      });
      broadcast('toast:show', {
        type: 'warning',
        message: `Provider 故障转移: ${data.from ?? '?'} → ${data.to ?? '?'}（主 Provider 请求失败）`,
      });
    },
  );

  // ===== Agent 消息发送 =====

  ipcMain.handle(
    'agent:sendMessage',
    async (_event, userMessage: MetonaMessage, sessionId: string) => {
      // M-33 修复: 参数校验，防止 undefined/非字符串导致下游异常
      // P1-5 修复: 校验失败时也发 ERROR+DONE 流事件，防止 isStreaming 永久卡死
      const sendErrorEvent = (message: string, sid: string): void => {
        const errorEvent: MetonaStreamEvent = {
          type: MetonaStreamEventType.ERROR,
          requestId: '',
          sessionId: sid,
          iteration: 0,
          seq: 0,
          timestamp: Date.now(),
          error: { code: MetonaErrorCode.UNKNOWN, message, retryable: false },
        };
        broadcast('agent:streamEvent', errorEvent);
        broadcast('agent:streamEvent', { ...errorEvent, type: MetonaStreamEventType.DONE });
      };

      if (!sessionId || typeof sessionId !== 'string') {
        log.warn('[AGENT] sendMessage rejected: invalid sessionId');
        sendErrorEvent('无效的会话 ID', sessionId ?? '');
        return { success: false, error: 'Invalid sessionId' };
      }
      if (
        !userMessage ||
        typeof userMessage !== 'object' ||
        typeof userMessage.content !== 'string'
      ) {
        log.warn('[AGENT] sendMessage rejected: invalid userMessage');
        sendErrorEvent('无效的消息格式', sessionId);
        return { success: false, error: 'Invalid message format' };
      }
      log.info('[AGENT] sendMessage:', sessionId, (userMessage.content ?? '').slice(0, 80));

      // v0.7.4 P1-3: 同会话并发防重入 —— 同一 sessionId 的第二个 invoke 在第一个 run
      // 未结束时到达，会让同一引擎被并行 runStream（工具副作用并发执行 / 内部 run-lock
      // 排队 30s 后强制 abort 旧 run，用户看到"上一次操作未完成"）。此处直接拒绝并
      // 广播明确错误事件，前端 isStreaming 正常收尾。
      if (agentEngineManager.isRunning(sessionId)) {
        const busyMsg = '该会话正在执行任务，请等待完成或先中断后再发送';
        log.warn(`[AGENT] sendMessage rejected: session ${sessionId} is already running`);
        sendErrorEvent(busyMsg, sessionId);
        return { success: false, error: busyMsg };
      }

      // v0.7.4 P1-4: 会话存在性预检 —— sessionId 指向已删除会话时，后续 saveMessage
      // 会因 FK 约束（foreign_keys=ON）抛错，而旧实现该调用在 try 块之外，invoke 直接
      // reject：不发 ERROR/DONE 事件、不 stopRecording，前端 isStreaming 永久卡死。
      // 此处提前校验并走统一的 sendErrorEvent + stopRecording 收尾路径。
      const sessionExists = sessionService.getSession(sessionId) != null;
      if (!sessionExists) {
        const missingMsg = '会话不存在或已被删除，请刷新后重试';
        log.warn(`[AGENT] sendMessage rejected: session ${sessionId} not found`);
        sendErrorEvent(missingMsg, sessionId);
        await sessionRecorder.stopRecording(sessionId, {
          totalIterations: 0,
          totalTokens: 0,
          durationMs: 0,
          terminationReason: 'error',
        });
        return { success: false, error: missingMsg };
      }

      // 发送消息前确保 Adapter 使用最新配置（失败则中止，防止用旧 Provider 的 adapter 发送）
      if (!ctx.reloadAdapter()) {
        const errorMsg =
          'Adapter 加载失败，请检查 LLM 配置（Provider、API Key、Base URL、Model 是否完整）';
        log.error('[AGENT]', errorMsg);
        sendErrorEvent(errorMsg, sessionId);
        await sessionRecorder.stopRecording(sessionId, {
          totalIterations: 0,
          totalTokens: 0,
          durationMs: 0,
          terminationReason: 'error',
        });
        return { success: false, error: errorMsg };
      }

      // TRACE 层：开始录制 / TOOL 层：记录会话开始
      sessionRecorder.startRecording(sessionId);
      auditService.logSessionStart(sessionId);

      // v0.7.4 P1-4: 以下数据准备（保存用户消息/加载历史/构建 System Prompt/记忆检索）
      // 全部移入 try 块 —— 旧实现这些调用在 try 之外，DB 或文件系统抛错时 invoke 直接
      // reject（不发 ERROR/DONE 事件、不 stopRecording，前端 isStreaming 永久卡死）。
      // 统一由下方 catch 收尾：ERROR + DONE 双事件 + 审计 + 录制终止。
      let history: MetonaMessage[] = [];
      let systemPrompt: MetonaSystemPrompt | null = null;
      let userContextPrefix = '';
      let engineUserMessage: MetonaMessage;

      try {
        // 保存用户消息到数据库
        // v0.7.4 回归修复: 透传前端消息 id（ChatMessage.id 由 genMsgId 生成）——
        // 否则 DB 用 msg_<nanoid> 生成不同 id，用户对刚发送消息"仅保存"
        // （updateMessageContent 按 id 匹配）会 0 行更新失败。
        sessionService.saveMessage({
          sessionId,
          role: 'user',
          content: userMessage.content,
          attachments: (userMessage as MetonaMessage & { attachments?: unknown[] }).attachments,
          id: (userMessage as MetonaMessage & { id?: string }).id,
        });

        // P2-11: 分层加载历史——存在滚动摘要时只加载 [摘要 + 近期原文]
        history = sessionSummaryService.buildHistoryMessages(sessionId).slice(0, -1);

        // 从工作空间文件构建 System Prompt
        const workspaceFiles = workspaceService.getFiles();
        systemPrompt = contextBuilder.buildSystemPrompt(workspaceFiles, workspaceService.getPath());

        // v0.3.18 修复: SOUL.md 为空或不存在时降级到默认身份，向前端发 toast 提示用户
        if (contextBuilder.isUsingFallbackRole()) {
          broadcast('toast:show', {
            type: 'info',
            message:
              '未找到 SOUL.md 或内容为空，已使用默认 Metona 身份。可在工作空间根目录创建 SOUL.md 自定义 Agent 人格',
          });
        }

        // 检索与用户消息相关的记忆 + 附件元信息 → 构建用户消息上下文前置块。
        // v0.7.3 P1-1 根治: 记忆注入与附件提示此前追加进 systemPrompt.dynamicReminders，
        // 每条消息都改变 system 字节 → 跨 run 缓存全 miss。现随首条 user 消息注入
        // （LLM 语义等价），system prompt 保持跨 run 字节级稳定。
        try {
          const memories = memoryManager.search(userMessage.content, {
            topK: 5,
            minImportance: 0.3,
          });
          const attachments = (
            userMessage as MetonaMessage & {
              attachments?: Array<{ name: string; type: string; truncated?: boolean }>;
            }
          ).attachments;
          userContextPrefix = buildUserContextPrefix({
            now: Date.now(),
            memories,
            attachments: Array.isArray(attachments) ? attachments : [],
            timezoneLabel: buildTimezoneLabel(),
          });
          if (memories.length > 0) {
            log.debug(`[AGENT] Injected ${memories.length} memories into user context prefix`);
          }
        } catch (err) {
          log.warn('[AGENT] Memory retrieval failed, proceeding without memories:', err);
          // 记忆检索失败时前置块退化为仅含日期时间（附件提示随之丢失可接受——
          // 主进程附件提示是辅助语义，附件内容本体仍在消息中）
          userContextPrefix = buildUserContextPrefix({
            now: Date.now(),
            memories: [],
            attachments: [],
            timezoneLabel: buildTimezoneLabel(),
          });
        }

        // 构建发送给引擎的用户消息副本 —— 前置块只存在于该副本：
        // DB 持久化（上方 saveMessage 已用原始内容）、前端展示、记忆固化、
        // 注入检测均使用原始干净内容，互不污染。
        engineUserMessage = {
          ...userMessage,
          content: withUserContextPrefix(userContextPrefix, userMessage.content),
        };
      } catch (err) {
        log.error('[AGENT] Failed to prepare message context:', err);
        const prepErr = (err as Error).message;
        broadcast('agent:streamEvent', {
          type: MetonaStreamEventType.ERROR,
          requestId: '',
          sessionId,
          iteration: 0,
          seq: 0,
          timestamp: Date.now(),
          error: { code: MetonaErrorCode.UNKNOWN, message: prepErr, retryable: false },
        } satisfies MetonaStreamEvent);
        await sessionRecorder.stopRecording(sessionId, {
          totalIterations: 0,
          totalTokens: 0,
          durationMs: 0,
          terminationReason: 'error',
        });
        return { success: false, error: prepErr };
      }

      // 提示注入检测在 try 内执行（需 systemPrompt 已构建，与引擎运行同域）

      try {
        // 提示注入检测（安全模块）
        // F-8 接通: security.promptInjectionDefense=false 时跳过用户消息检测
        // （工具结果侧的 SecurityScanHook 由 main.ts 按同一配置决定是否挂载）
        // fail-secure: 仅显式 false 才关闭 —— 配置值异常（空串/null/类型错误）时保持防护开启
        const injectionEnabled =
          configService.get<boolean>('security.promptInjectionDefense') !== false;
        if (injectionEnabled) {
          const injectionResult = promptInjectionDefender.detect(userMessage.content);
          if (injectionResult.riskScore >= 7) {
            log.warn('[PromptInjectionDefender] Blocked message:', injectionResult.findings);
            sendErrorEvent(
              `Message blocked by prompt injection defense: ${injectionResult.recommendation}`,
              sessionId,
            );
            await sessionRecorder.stopRecording(sessionId, {
              totalIterations: 0,
              totalTokens: 0,
              durationMs: 0,
              terminationReason: 'error',
            });
            return { success: false, error: 'Message blocked by prompt injection defense' };
          }
          if (injectionResult.riskScore >= 4) {
            log.warn(
              '[PromptInjectionDefender] Suspicious patterns detected:',
              injectionResult.findings,
            );
          }
        }

        // TRACE 层：记录上下文构建
        sessionRecorder.recordContextBuilt(sessionId, {
          tokenCount: estimateMessagesTokens(history),
          usageRatio: 0,
        });

        // 启动 Agent Loop（P2-10: 每会话独立引擎）
        // v0.7.3 P1-1: 传入带上下文前置块的消息副本 —— 原始 userMessage 保持干净
        const engine = agentEngineManager.getEngine(sessionId);
        const output = await engine.runStream(engineUserMessage, sessionId, history, systemPrompt);

        // 输出验证（不阻塞响应，仅记录警告）
        // v0.3.0 修复: 传入 toolResults 和 context，启用事实一致性检查和幻觉检测
        // v0.4.1: warning 及以上级别的 issue 通过 VALIDATION 流事件推送前端展示（此前仅写日志，用户不可感知）
        try {
          const toolResults = output.iterations
            .flatMap((step) => step.toolResults ?? [])
            .map((r) => (typeof r.result === 'string' ? r.result : JSON.stringify(r.result)));
          const context = [...history, { role: 'user', content: userMessage.content }]
            .map((m) => `${m.role}: ${m.content}`)
            .join('\n');
          const validation = await outputValidator.validate(output.finalAnswer, {
            toolResults: toolResults.length > 0 ? toolResults : undefined,
            context,
          });
          if (!validation.valid || validation.issues.length > 0) {
            log.warn('[OutputValidator] Validation issues:', validation.issues);
          }
          log.debug(`[OutputValidator] Score: ${validation.score}, Valid: ${validation.valid}`);

          // v0.4.1: 推送验证结果到前端 — 只推送 warning/error 级（info 级为噪声）
          // 类型谓词收窄 severity，确保与 MetonaValidationPayload.issues 的类型一致
          const visibleIssues = validation.issues
            .filter(
              (i): i is typeof i & { severity: 'warning' | 'error' } =>
                i.severity === 'warning' || i.severity === 'error',
            )
            .slice(0, 5);
          if (visibleIssues.length > 0) {
            broadcast('agent:streamEvent', {
              type: MetonaStreamEventType.VALIDATION,
              requestId: '',
              sessionId,
              iteration: output.iterations.length,
              seq: 0,
              timestamp: Date.now(),
              validation: {
                score: validation.score,
                issues: visibleIssues.map((i) => ({
                  severity: i.severity,
                  type: i.type,
                  message: i.message,
                })),
              },
            } satisfies MetonaStreamEvent);
          }
        } catch (err) {
          log.error('[OutputValidator] Validation failed:', err);
        }

        // 保存每轮迭代的 assistant 消息到数据库（含思考内容和工具调用）
        // 崩溃修复（会话停止根因的持久化侧）: 原 `if (!step.thought) continue;` 把
        // 纯工具调用轮（零文本零思考）整个跳过 — assistant 与 tool 结果都不落库。
        // 后果：① 重启后历史缺失工具上下文（模型"忘记"做过什么，重复调用 →
        // 表现为工具调用不稳定）；② 与 engine 运行时缺陷同源。现改为：
        // 有 thought 或有 toolCalls 的步骤都保存。
        for (const step of output.iterations) {
          if (!step.thought && !(step.toolCalls && step.toolCalls.length > 0)) continue;

          const toolCallsWithResults = step.toolCalls?.map((tc) => {
            const result = step.toolResults?.find((r) => r.toolCallId === tc.id);
            return {
              id: tc.id,
              name: tc.name,
              args: tc.args,
              status: result?.success ? ('success' as const) : ('error' as const),
              result: result?.result,
              durationMs: result?.durationMs,
              error: result?.error,
            };
          });

          // 只有当有内容、思考内容或工具调用时才保存
          if (
            step.thought?.content ||
            step.thought?.reasoningContent ||
            toolCallsWithResults?.length
          ) {
            // C-6 修复: assistant 消息仅有 tool_calls 时 content 必须为 null（而非空字符串）
            const assistantContent =
              toolCallsWithResults?.length && !step.thought?.content
                ? null
                : (step.thought?.content ?? null);
            sessionService.saveMessage({
              sessionId,
              role: 'assistant',
              content: assistantContent,
              reasoningContent: step.thought?.reasoningContent || undefined,
              toolCalls: toolCallsWithResults,
              iteration: step.iteration,
            });
          }

          // v0.3.0 修复: 保存 tool 结果消息到数据库
          // OpenAI 兼容 API 要求 assistant 消息有 tool_calls 时，后续必须有对应的 tool 结果消息
          if (step.toolResults) {
            for (const result of step.toolResults) {
              const resultContent =
                typeof result.result === 'string' ? result.result : JSON.stringify(result.result);
              sessionService.saveMessage({
                sessionId,
                role: 'tool',
                content: result.error ?? resultContent,
                toolResult: result,
                iteration: step.iteration,
              });
            }
          }
        }

        // 更新 Token 统计
        if (output.totalTokenUsage.totalTokens > 0) {
          sessionService.updateTokenUsage(sessionId, output.totalTokenUsage.totalTokens);
        }

        // 更新 MEMORY.md 时间戳
        workspaceService.updateMemoryTimestamp();

        // 会话结束：AI 判断本次对话有哪些重要内容需要持久化到 MEMORY.md
        // v0.7.3 P1-5 节流: 此前每次 run 无条件发起固化 LLM 请求，短寒暄同样触发。
        // 现按 consolidation-policy 决策（总开关 + 内容门控 + 会话级频率窗口）触发；
        // 异步执行不阻塞主流程返回；失败仅记录日志。
        const lastConsolidationAt = lastConsolidationBySession.get(sessionId) ?? 0;
        const hadSuccessfulToolCall = output.iterations.some((step) =>
          (step.toolResults ?? []).some((r) => r.success),
        );
        const decision = shouldConsolidate({
          enabled: configService.get<boolean>('memory.consolidationEnabled') !== false,
          answerChars: output.finalAnswer?.length ?? 0,
          minChars: configService.get<number>('memory.consolidationMinChars') ?? 200,
          hadSuccessfulToolCall,
          lastConsolidationAt,
          now: Date.now(),
          intervalMs: configService.get<number>('memory.consolidationIntervalMs') ?? 600_000,
        });
        if (decision.consolidate) {
          memoryConsolidator
            .consolidate(userMessage.content, output.finalAnswer, output.iterations)
            .then((result) => {
              if (result.appended > 0) {
                touchConsolidation(sessionId);
                log.info(
                  `[AGENT] Memory consolidated: ${result.appended} entries appended to MEMORY.md`,
                );
                broadcast('toast:show', {
                  type: 'info',
                  message: `AI 已将 ${result.appended} 条重要记忆写入 MEMORY.md`,
                });
              }
            })
            .catch((err) => {
              log.warn('[AGENT] Memory consolidation failed:', err);
            });
        } else {
          log.debug(`[AGENT] Memory consolidation skipped (${decision.reason})`);
        }

        // v0.7.3 P4-1: 首个完成的 run 之后生成精炼会话标题（每会话幂等，失败静默）
        if (output.terminationReason === 'completed') {
          titleGenerator
            .maybeGenerateTitle(sessionId, userMessage.content, output.finalAnswer)
            .then((title) => {
              if (title) {
                // 广播重命名结果，前端 Sidebar 实时刷新标题
                broadcast('config:changed', { key: `session.title.${sessionId}`, value: title });
              }
            })
            .catch(() => {
              /* 静默 —— 标题失败已有 debug 日志 */
            });
        }

        // TOOL 层：记录会话结束 / TRACE 层：停止录制
        auditService.logSessionEnd({
          sessionId,
          totalIterations: output.iterations.length,
          totalTokens: output.totalTokenUsage.totalTokens,
          durationMs: output.durationMs,
          terminationReason: output.terminationReason,
        });
        await sessionRecorder.stopRecording(sessionId, {
          totalIterations: output.iterations.length,
          totalTokens: output.totalTokenUsage.totalTokens,
          durationMs: output.durationMs,
          terminationReason: output.terminationReason,
        });

        // P2-11: 会话结束后评估滚动摘要（fire-and-forget，失败仅记录）
        sessionSummaryService.maybeSummarize(sessionId).catch((err) => {
          log.warn('[AGENT] Session summary generation failed:', err);
        });

        log.info(
          `[AGENT] Completed: ${output.terminationReason}, ${output.iterations.length} iterations, ${output.durationMs}ms`,
        );

        return { success: true };
      } catch (error) {
        log.error('[AGENT] Error:', error);

        // TOOL 层：记录错误
        auditService.log({
          sessionId,
          eventType: 'error',
          actor: 'agent',
          target: 'agent_loop',
          details: { error: (error as Error).message },
          outcome: 'error',
        });

        // TRACE 层：停止录制
        await sessionRecorder.stopRecording(sessionId, {
          totalIterations: 0,
          totalTokens: 0,
          durationMs: 0,
          terminationReason: 'error',
        });

        // 发送错误事件到 UI
        const metonaError: MetonaError = {
          code: MetonaErrorCode.UNKNOWN,
          message: (error as Error).message,
          retryable: false,
        };
        broadcast('agent:streamEvent', {
          type: MetonaStreamEventType.ERROR,
          requestId: '',
          sessionId,
          iteration: 0,
          seq: 0,
          timestamp: Date.now(),
          error: metonaError,
        } satisfies MetonaStreamEvent);

        return { success: false, error: (error as Error).message };
      }
    },
  );

  // ===== v0.5.0: DeepSeek 余额查询（复用 DeepSeekAdapter.getBalance，原为死代码） =====
  ipcMain.handle('llm:getBalance', async () => {
    try {
      const adapter = agentEngineManager.getAdapter();
      if (!(adapter instanceof DeepSeekAdapter)) {
        return {
          success: false,
          error: 'Balance query is only supported for the DeepSeek provider',
        };
      }
      const balance = await adapter.getBalance();
      if (!balance) {
        return { success: false, error: '余额查询失败（API Key 无效或网络错误）' };
      }
      return { success: true, data: balance };
    } catch (error) {
      log.warn('[AGENT] getBalance failed:', (error as Error).message);
      return { success: false, error: (error as Error).message };
    }
  });

  // ===== v0.7.2 P3-9: 动态模型列表（六家 adapter 的 listModels 首次获得 IPC 消费者）=====
  // 此前 DeepSeek/OpenAI/Ollama 均实现了实时模型发现（/models、/api/tags），
  // 但 preload/IPC 层无任何通道 —— 设置页模型输入只能靠用户手填。
  // 契约：列表基于"已保存"的 LLM 配置（与引擎实际使用的 adapter 同源）；
  // 配置不完整时显式失败，绝不返回兜底 adapter 的误导性静态列表。
  ipcMain.handle('llm:listModels', async () => {
    try {
      const provider = configService.get<string>('llm.provider') ?? '';
      const model = configService.get<string>('llm.model') ?? '';
      const apiKey = configService.get<string>('llm.apiKey') || '';
      if (!provider || !model) {
        return { success: false, error: 'LLM 未配置（Provider/Model 为空），无法获取模型列表' };
      }
      if (!apiKey && provider !== 'ollama') {
        return { success: false, error: 'API Key 未配置，无法获取模型列表' };
      }
      // 列表基于当前已保存配置 —— 先幂等重载 adapter（配置签名未变时为 no-op）
      if (!ctx.reloadAdapter()) {
        return { success: false, error: 'LLM 配置校验失败，请先在设置中修正配置' };
      }
      const adapter = agentEngineManager.getAdapter();
      if (!adapter.listModels) {
        return { success: false, error: '当前 Provider 不支持模型列表查询' };
      }
      const models = await adapter.listModels();
      return { success: true, data: models };
    } catch (error) {
      log.warn('[AGENT] listModels failed:', (error as Error).message);
      return { success: false, error: (error as Error).message };
    }
  });

  // ===== v0.7.2 P3-10: Ollama 模型下载（adapter.pullModel 首次接线 IPC/UI）=====
  // v0.6.4 P4-1 已实现 pull 的进度回调与外部取消信号，但主进程侧零消费者。
  // 契约：同一时刻仅允许一个下载任务（全局 AbortController）；
  // 进度经 llm:ollamaPullProgress 广播（{ model, status, completed?, total? }），
  // 结束（成功/失败/取消）统一广播 llm:ollamaPullEnded 供前端复位 UI。
  let ollamaPullController: AbortController | null = null;

  ipcMain.handle('llm:ollamaPull', async (_event, modelName: unknown) => {
    if (typeof modelName !== 'string' || !modelName.trim()) {
      return { success: false, error: 'Invalid model name' };
    }
    // 模型名仅进入 JSON body（不经 shell），仍做字符白名单防转义边界
    // 合法形态如 qwen3:8b / llama3.1:70b-instruct-q4_K_M / user/model
    if (!/^[A-Za-z0-9._:/-]+$/.test(modelName.trim())) {
      return { success: false, error: `Invalid model name: ${modelName.slice(0, 60)}` };
    }
    const adapter = agentEngineManager.getAdapter();
    if (!(adapter instanceof OllamaAdapter)) {
      return { success: false, error: '仅 Ollama Provider 支持模型下载' };
    }
    if (ollamaPullController) {
      return { success: false, error: '已有模型下载任务进行中，请先取消' };
    }
    const controller = new AbortController();
    ollamaPullController = controller;
    const trimmed = modelName.trim();
    try {
      await adapter.pullModel(
        trimmed,
        (progress) => {
          broadcast('llm:ollamaPullProgress', { model: trimmed, ...progress });
        },
        controller.signal,
      );
      return { success: true };
    } catch (error) {
      return {
        success: false,
        error: (error as Error).message,
        aborted: controller.signal.aborted,
      };
    } finally {
      ollamaPullController = null;
      broadcast('llm:ollamaPullEnded', { model: trimmed });
    }
  });

  ipcMain.handle('llm:ollamaPullCancel', async () => {
    if (!ollamaPullController) {
      return { success: false, error: '没有进行中的下载任务' };
    }
    ollamaPullController.abort();
    return { success: true };
  });

  // ===== 中断会话 =====

  ipcMain.handle('agent:abortSession', async (_event, sessionId) => {
    log.info('[AGENT] Abort:', sessionId);
    // P2-10: 联动中断该会话派生的所有 SubAgent（消除"会话停了子任务还在跑"）
    // v0.5.1: 记录被中止的 taskId — SubAgent 的 pending 确认以 taskId 为 sessionId，
    // 需一并清理，否则中止后孤儿工具在用户补批残留弹框时会真实执行副作用
    const abortedTaskIds = orchestrator.abortByParent(sessionId);
    agentEngineManager.abort(sessionId);
    // MT-1 修复: 等待当前 run 完全结束再返回，防止用户立即重发时新消息卡在等待中
    await agentEngineManager.waitForAbort(sessionId);
    // v0.3.0 修复: 清理所有等待中的工具确认，避免定时器泄漏和超时 toast 在新会话中弹出
    // v0.5.0: 按会话清理 — 只拒绝被中断会话的 pending，不影响其他并发会话等待中的确认
    confirmationHook.clearPending(sessionId);
    // v0.5.1: 被中止 SubAgent 的 pending 确认一并拒绝（含 SubAgent 递归派生的孙任务）
    // v0.7.3 P2-3: 被中止的 SubAgent 是会话终态 —— 用 forgetSession 连决策记忆一并清理
    // v0.7.4 P3-5: 被中止 SubAgent 的 TRACE 状态（subTraces/subMeta）一并清理 ——
    // 旧实现只覆盖 taskCompleted/taskError 路径，被 abort 的 SubAgent 残留 Map 条目
    for (const taskId of abortedTaskIds) {
      confirmationHook.forgetSession(taskId);
      subTraces.delete(taskId);
      subMeta.delete(taskId);
      // 中止的 SubAgent 录制文件也收尾（TRACE 完整性：标记为中断终止）
      await sessionRecorder.stopRecording(taskId, {
        totalIterations: 0,
        totalTokens: 0,
        durationMs: 0,
        terminationReason: 'user_interrupt',
      });
    }

    // TOOL 层：记录中断
    auditService.log({
      sessionId,
      eventType: 'session_end',
      actor: 'user',
      target: 'session',
      details: { reason: 'user_abort' },
      outcome: 'denied',
    });

    return { success: true };
  });

  // ===== v0.5.0: SubAgent 可观测性 =====
  // 1) 生命周期事件广播给前端（AgentMonitor 的 SubAgent 状态区）
  // 2) SubEngine 流事件录制到独立 TRACE 文件（sessionId = taskId），不污染父会话的流
  //    此前 orchestrator 的 6 个事件全项目零消费者，SubAgent 执行过程对 UI 与 TRACE 完全不可见

  /** SubAgent 元信息（description/depth，供完成/失败事件广播时补全载荷） */
  const subMeta = new Map<string, { description: string; depth: number }>();
  /** SubAgent 迭代录制状态（THINKING 驱动新迭代，text 累积供 llm_response） */
  interface SubTraceState {
    iteration: number;
    startedAt: number;
    text: string;
    usage: { input: number; output: number; total: number };
    responded: boolean;
  }
  const subTraces = new Map<string, SubTraceState>();

  const finishSubTrace = async (
    taskId: string,
    reason: string,
    durationMs: number,
    iterations: number,
  ): Promise<void> => {
    await sessionRecorder.stopRecording(taskId, {
      totalIterations: iterations,
      totalTokens: 0,
      durationMs,
      terminationReason: reason,
    });
    subTraces.delete(taskId);
    subMeta.delete(taskId);
    // v0.7.3 P2-3: SubAgent 终态 —— 决策记忆随任务终结清理（防长期运行泄漏）
    confirmationHook.forgetSession(taskId);
  };

  orchestrator.on(
    'taskDelegated',
    (d: { taskId: string; description: string; parentSessionId: string; depth: number }) => {
      broadcast('subagent:event', {
        taskId: d.taskId,
        parentSessionId: d.parentSessionId,
        description: d.description,
        status: 'delegated',
        depth: d.depth,
      });
      subMeta.set(d.taskId, { description: d.description, depth: d.depth });
      sessionRecorder.startRecording(d.taskId);
    },
  );

  orchestrator.on(
    'taskStarted',
    (d: { taskId: string; description: string; parentSessionId: string; depth: number }) => {
      broadcast('subagent:event', {
        taskId: d.taskId,
        parentSessionId: d.parentSessionId,
        description: d.description,
        status: 'running',
        depth: d.depth,
      });
    },
  );

  orchestrator.on(
    'taskCompleted',
    (r: {
      taskId: string;
      parentSessionId: string;
      success: boolean;
      durationMs: number;
      iterations: number;
    }) => {
      const meta = subMeta.get(r.taskId);
      broadcast('subagent:event', {
        taskId: r.taskId,
        parentSessionId: r.parentSessionId,
        description: meta?.description ?? '',
        status: r.success ? 'completed' : 'error',
        depth: meta?.depth ?? 1,
        durationMs: r.durationMs,
        iterations: r.iterations,
      });
      void finishSubTrace(r.taskId, r.success ? 'completed' : 'error', r.durationMs, r.iterations);
    },
  );

  orchestrator.on(
    'taskError',
    (d: { taskId: string; parentSessionId: string; description: string; error: string }) => {
      const meta = subMeta.get(d.taskId);
      broadcast('subagent:event', {
        taskId: d.taskId,
        parentSessionId: d.parentSessionId,
        description: d.description ?? meta?.description ?? '',
        status: 'error',
        depth: meta?.depth ?? 1,
        error: d.error,
      });
      void finishSubTrace(d.taskId, 'error', 0, 0);
    },
  );

  // SubEngine 状态事件 → TRACE 迭代录制（与主管道相同的事件形状，sessionId = taskId）
  orchestrator.on(
    'subStateChange',
    ({
      taskId,
      data,
    }: {
      taskId: string;
      data: { state?: string; current?: string; iteration?: number };
    }) => {
      const st = subTraces.get(taskId);
      const stateValue = data.state ?? data.current ?? '';

      // THINKING 且迭代号变化 → 新迭代开始（关闭上一迭代）
      if (
        stateValue === 'THINKING' &&
        data.iteration != null &&
        (!st || st.iteration !== data.iteration)
      ) {
        if (st && !st.responded) {
          sessionRecorder.recordLLMResponse({
            sessionId: taskId,
            iteration: st.iteration,
            content: st.text,
            finishReason: 'stop',
            tokenUsage: st.usage,
          });
          sessionRecorder.recordIterationEnd(taskId, {
            iteration: st.iteration,
            durationMs: Date.now() - st.startedAt,
          });
        }
        subTraces.set(taskId, {
          iteration: data.iteration,
          startedAt: Date.now(),
          text: '',
          usage: { input: 0, output: 0, total: 0 },
          responded: false,
        });
        sessionRecorder.recordIterationStart(taskId, data.iteration);
        const provider = configService.get<string>('llm.provider') ?? '';
        const model = configService.get<string>('llm.model') ?? '';
        sessionRecorder.recordLLMRequest({
          sessionId: taskId,
          iteration: data.iteration,
          provider,
          model,
          messageCount: data.iteration + 1,
        });
        return;
      }

      // TERMINATED → 补记最终迭代（正常流程只在下一轮 THINKING 补记，最终轮无后续）
      if (stateValue === 'TERMINATED' && st) {
        if (!st.responded) {
          sessionRecorder.recordLLMResponse({
            sessionId: taskId,
            iteration: st.iteration,
            content: st.text,
            finishReason: 'stop',
            tokenUsage: st.usage,
          });
        }
        sessionRecorder.recordIterationEnd(taskId, {
          iteration: st.iteration,
          durationMs: Date.now() - st.startedAt,
        });
      }
    },
  );

  // SubEngine 流事件 → TRACE 内容录制（文本累积 + 工具调用/结果）
  orchestrator.on(
    'subStreamEvent',
    ({ taskId, event }: { taskId: string; event: MetonaStreamEvent }) => {
      const st = subTraces.get(taskId);
      if (!st) return;

      switch (event.type) {
        case MetonaStreamEventType.TEXT_DELTA:
          if (event.delta) st.text += event.delta;
          break;
        case MetonaStreamEventType.USAGE:
          if (event.usage) {
            st.usage = {
              input: event.usage.inputTokens ?? 0,
              output: event.usage.outputTokens ?? 0,
              total: event.usage.totalTokens ?? 0,
            };
          }
          break;
        case MetonaStreamEventType.TOOL_CALL_COMPLETE:
          if (event.toolCall) {
            sessionRecorder.recordToolCall({
              sessionId: taskId,
              iteration: event.iteration,
              toolName: event.toolCall.name,
              args: event.toolCall.args,
            });
          }
          break;
        case MetonaStreamEventType.TOOL_RESULT:
          if (event.toolResult) {
            const resultPreview =
              typeof event.toolResult.result === 'string'
                ? event.toolResult.result
                : JSON.stringify(event.toolResult.result);
            sessionRecorder.recordToolResult({
              sessionId: taskId,
              iteration: event.iteration,
              toolName: event.toolResult.toolName,
              success: event.toolResult.success,
              durationMs: event.toolResult.durationMs ?? 0,
              resultPreview,
              error: event.toolResult.error,
            });
          }
          break;
        default:
          break;
      }
    },
  );
}