metona-ai-desktop/electron/services/session-summary.service.ts

/**
 * Session Summary Service — 会话摘要分层上下文（P2-11）
 *
 * 解决"历史消息全量加载"问题：sendMessage 原实现从 DB 加载全部历史消息进 LLM
 * 上下文，超长会话（数百条消息）token 成本线性膨胀，只能依赖运行时压缩兜底。
 *
 * 分层策略（DB 层持久化，跨 run 生效）：
 * 1. 会话消息数超过阈值（50 条）后，将较早消息（保留尾部 20 条原文）交由 LLM
 *    生成滚动摘要，持久化到 session_summaries 表（含 summarized_until_rowid 游标）
 * 2. 下次 sendMessage 只加载：[摘要消息] + [rowid > 游标的近期原文消息]
 * 3. 摘要随新消息增量滚动更新（旧摘要作为上下文参与新一轮总结）
 *
 * 与 engine 运行时压缩（compressMessages）的关系：
 * 运行时压缩处理"单次 run 内"的上下文膨胀；本服务处理"跨 run"的 DB 历史分层，
 * 两者互补，运行时压缩触发频率将显著下降。
 *
 * @see electron/services/database.service.ts — session_summaries 表结构
 */

import { nanoid } from 'nanoid';
import type Database from 'better-sqlite3';
import log from 'electron-log';
import type { IMetonaProviderAdapter, MetonaMessage, MetonaRequest } from '../harness/types';
import type { SessionService } from './session.service';

/** 触发摘要的最小消息总数（低于此值保持全量加载） */
const MIN_MESSAGES_TO_SUMMARIZE = 50;
/** 摘要后保留的尾部原文条数 */
const TAIL_KEEP = 20;
/** 每次摘要需新增的最小未总结消息数（避免每条消息都触发 LLM 摘要） */
const MIN_NEW_TO_SUMMARIZE = 15;
/** LLM 摘要调用超时 */
const SUMMARY_TIMEOUT_MS = 30_000;
/** 传给 LLM 的单条消息内容截断 */
const PER_MESSAGE_TRUNCATE = 600;
/** 传给 LLM 的总字符上限 */
const MAX_DIGEST_CHARS = 24_000;
/** v0.5.4: 多轮图片记忆 — 历史上下文注入的最大图片数（从最新向前收集，防 token 爆炸） */
const MAX_HISTORY_IMAGES = 10;

export class SessionSummaryService {
  constructor(
    private getDB: () => Database.Database,
    private sessionService: SessionService,
    private adapterGetter: () => IMetonaProviderAdapter,
  ) {}

  /**
   * 构建分层历史消息（sendMessage 的加载入口）
   *
   * @returns MetonaMessage 数组：存在摘要时为 [摘要消息, 近期原文...]，否则全量原文
   */
  buildHistoryMessages(sessionId: string): MetonaMessage[] {
    const existing = this.getSummary(sessionId);
    const tail = this.sessionService.getMessages(sessionId, {
      afterRowid: existing?.summarizedUntilRowid ?? 0,
    });

    const messages: MetonaMessage[] = tail
      .filter((m) => m.role !== 'system')
      .map((m) => ({
        role: m.role as MetonaMessage['role'],
        content: m.content,
        reasoningContent: m.reasoningContent,
        toolCalls: m.toolCalls as MetonaMessage['toolCalls'],
        toolResult: m.toolResult as MetonaMessage['toolResult'],
        // v0.5.4: 保留 attachments 供 restoreHistoryImages 恢复图片（多轮图片记忆）
        attachments: (m as { attachments?: unknown[] }).attachments,
        timestamp: m.timestamp,
        iteration: m.iteration,
      }));

    // v0.5.4: 多轮图片记忆 — 从持久化的 attachments（压缩 base64 preview）恢复 images，
    // 历史轮次的图片重新注入 LLM 上下文（此前仅发送当轮可见，跨轮即"失忆"）。
    // 数量上限防 token 爆炸：只取最近 MAX_HISTORY_IMAGES 张（从最新消息向前收集）。
    // 摘要区间（summarizedUntilRowid 之前）的图片无法恢复 — 符合滚动摘要的语义。
    this.restoreHistoryImages(messages);

    if (existing && messages.length > 0) {
      // 摘要以 assistant 角色注入（与 engine 运行时压缩的注入策略一致）
      const summaryMessage: MetonaMessage = {
        role: 'assistant',
        content: `[Context Summary] The following is a rolling summary of earlier conversation history:\n\n${existing.summary}`,
        timestamp: 0,
      };
      return [summaryMessage, ...messages];
    }
    return messages;
  }

  /**
   * v0.5.4: 恢复历史消息的 images（多轮图片记忆）
   *
   * attachments 中 type=image 的 preview（1024px JPEG 压缩 base64）在消息
   * 持久化时已保存（与编辑重发的恢复逻辑同源）。此处将其映射回
   * MetonaMessage.images，让历史轮次图片随上下文回传 LLM。
   *
   * 上限策略：从最新消息向前收集，最多 MAX_HISTORY_IMAGES 张 —
   * 每张 1024px 图约数百至千余 token，无上限的长会话会迅速吃满上下文。
   */
  private restoreHistoryImages(messages: MetonaMessage[]): void {
    let remaining = MAX_HISTORY_IMAGES;
    for (let i = messages.length - 1; i >= 0 && remaining > 0; i--) {
      const raw = messages[i] as MetonaMessage & {
        attachments?: Array<{ type?: string; preview?: string }>;
      };
      const imageAttachments = (raw.attachments ?? []).filter(
        (a) => a.type === 'image' && typeof a.preview === 'string' && a.preview.length > 0,
      );
      if (imageAttachments.length === 0) continue;

      const take = Math.min(imageAttachments.length, remaining);
      // 优先保留靠后的图片（时间更近）
      const picked = imageAttachments.slice(-take);
      messages[i].images = picked.map((a) => ({
        url: a.preview as string,
        detail: 'auto' as const,
      }));
      remaining -= take;
    }
  }

  /**
   * 会话结束后评估并生成滚动摘要（fire-and-forget 调用，失败仅记录日志）
   */
  async maybeSummarize(sessionId: string): Promise<void> {
    const rows = this.sessionService.getMessages(sessionId);
    if (rows.length < MIN_MESSAGES_TO_SUMMARIZE) return;

    const existing = this.getSummary(sessionId);
    const lastSummarized = existing?.summarizedUntilRowid ?? 0;
    const unsummarized = rows.filter((r) => (r.rowId ?? 0) > lastSummarized);

    // 未总结增量不足（保留尾部 TAIL_KEEP 后仍需 ≥ MIN_NEW_TO_SUMMARIZE 条）
    if (unsummarized.length <= TAIL_KEEP + MIN_NEW_TO_SUMMARIZE) return;

    const toSummarize = unsummarized.slice(0, unsummarized.length - TAIL_KEEP);
    if (toSummarize.length < MIN_NEW_TO_SUMMARIZE) return;

    const summary = await this.summarizeViaLLM(existing?.summary ?? '', toSummarize);
    if (!summary) return;

    const untilRowid = toSummarize[toSummarize.length - 1].rowId ?? 0;
    this.saveSummary(sessionId, summary, untilRowid);
    log.info(
      `[SessionSummary] session ${sessionId}: summarized ${toSummarize.length} messages (up to rowid ${untilRowid}), kept ${TAIL_KEEP} recent`,
    );
  }

  // ===== 摘要表 CRUD =====

  getSummary(sessionId: string): { summary: string; summarizedUntilRowid: number } | null {
    const db = this.getDB();
    const row = db
      .prepare('SELECT summary, summarized_until_rowid FROM session_summaries WHERE session_id = ?')
      .get(sessionId) as { summary: string; summarized_until_rowid: number } | undefined;
    return row ? { summary: row.summary, summarizedUntilRowid: row.summarized_until_rowid } : null;
  }

  saveSummary(sessionId: string, summary: string, untilRowid: number): void {
    const db = this.getDB();
    db.prepare(
      `
      INSERT INTO session_summaries (session_id, summary, summarized_until_rowid, updated_at)
      VALUES (?, ?, ?, ?)
      ON CONFLICT(session_id) DO UPDATE SET
        summary = excluded.summary,
        summarized_until_rowid = excluded.summarized_until_rowid,
        updated_at = excluded.updated_at
    `,
    ).run(sessionId, summary, untilRowid, Date.now());
  }

  // ===== LLM 摘要 =====

  /**
   * 调用 LLM 生成滚动摘要（复用主 Provider adapter）
   *
   * @param priorSummary 既有摘要（滚动总结上下文，可为空）
   * @param messages 本轮待总结的消息
   */
  private async summarizeViaLLM(
    priorSummary: string,
    messages: Array<{ role: string; content: string | null }>,
  ): Promise<string | null> {
    let total = 0;
    const transcript = messages
      .map((m) => {
        const content = (m.content ?? '').slice(0, PER_MESSAGE_TRUNCATE);
        if (total < MAX_DIGEST_CHARS) {
          total += content.length;
          return `[${m.role.toUpperCase()}] ${content}`;
        }
        return null;
      })
      .filter((s): s is string => s !== null)
      .join('\n\n');

    const request: MetonaRequest = {
      meta: {
        sessionId: 'session-summary',
        iteration: 0,
        requestId: `ss_${nanoid(12)}`,
        timestamp: Date.now(),
        agentVersion: '1.0.0',
      },
      systemPrompt: {
        roleDefinition: 'You are a conversation summarizer for an AI agent application.',
        outputConstraints:
          'Produce a rolling summary of the conversation history below. ' +
          'If a prior summary exists, merge it with the new content into one updated summary. ' +
          'Preserve key facts, decisions, tool outcomes, file paths, and open questions needed for future reasoning. ' +
          'Output in the same language as the conversation. Maximum 400 words. Output ONLY the summary text.',
        safetyGuidelines:
          'Do not include sensitive data like passwords or API keys in the summary.',
      },
      messages: [
        {
          role: 'user',
          content: `${priorSummary ? `## Prior summary (merge and update):\n\n${priorSummary}\n\n` : ''}## New messages to summarize:\n\n${transcript}`,
          timestamp: Date.now(),
        },
      ],
      params: {
        maxTokens: 2048,
        temperature: 0.0,
        stream: false,
        thinkingEnabled: false,
        thinkingEffort: 'low',
      },
    };

    try {
      const timeoutPromise = new Promise<never>((_, reject) => {
        setTimeout(() => reject(new Error('summary timeout')), SUMMARY_TIMEOUT_MS);
      });
      const response = await Promise.race([this.adapterGetter().send(request), timeoutPromise]);
      const summary = response.content.trim();
      return summary || null;
    } catch (err) {
      log.warn('[SessionSummary] LLM summarization failed:', (err as Error).message);
      return null;
    }
  }
}