@huyooo/ai-chat-core 0.2.45 → 0.3.2

package/src/orchestrator/context-summarizer.ts

@@ -0,0 +1,123 @@
+/**
+ * Context Summarizer — 专用上下文总结模块
+ *
+ * 当对话历史超过当前模型的 context window 时，通过外部注入的 summarize 回调执行 AI 总结。
+ * 回调由宿主提供（如 SuperX 走 ai-server 的 toolAI），ai-chat-core 不关心具体调用方式。
+ */
+
+import type { ProtocolMessage } from '../protocols';
+import {
+  buildSummarizeRequest,
+  applySummary,
+  estimateTotalTokens,
+  estimateStringTokens,
+  buildSingleMessageSummarizeRequest,
+  type CompactConfig,
+} from './context-compressor';
+import { createModuleLogger } from '../logger';
+
+const logger = createModuleLogger('ContextSummarizer');
+
+// ==================== 类型 ====================
+
+/** 外部注入的 AI 总结函数，model 由 orchestrator 根据场景传入 */
+export type SummarizeFn = (systemPrompt: string, userPrompt: string, model: string) => Promise<string>;
+
+export interface SummarizeResult {
+  messages: ProtocolMessage[];
+  success: boolean;
+  /** AI 生成的原始摘要文本，前端持久化时需加 [对话历史摘要] 前缀 */
+  summaryContent?: string;
+}
+
+// ==================== 核心函数 ====================
+
+/**
+ * 使用 AI 总结对话历史
+ *
+ * 1. 用 buildSummarizeRequest 分离「待总结的中间历史」和「保留的最近消息」
+ * 2. 调用外部注入的 summarizeFn 执行总结（走 ai-server 等）
+ * 3. 用 applySummary 组装压缩后的消息列表
+ */
+export async function summarizeHistory(
+  summarizeFn: SummarizeFn,
+  messages: ProtocolMessage[],
+  config: CompactConfig,
+  model: string,
+): Promise<SummarizeResult> {
+  const { summarizeMessages, keepMessages } = buildSummarizeRequest(messages, config);
+
+  if (summarizeMessages.length === 0) {
+    logger.info('中间历史太短，跳过 AI 总结');
+    return { messages, success: true };
+  }
+
+  const systemPrompt = summarizeMessages[0].content;
+  const userPrompt = summarizeMessages[1].content;
+
+  logger.info({ model, tokens: estimateTotalTokens(summarizeMessages), messageCount: messages.length }, `开始 AI 总结 (${model}): ~${estimateTotalTokens(summarizeMessages)} tokens, ${messages.length} 条消息`);
+
+  try {
+    const summaryContent = await summarizeFn(systemPrompt, userPrompt, model);
+
+    if (!summaryContent.trim()) {
+      logger.warn('AI 总结返回空内容');
+      return { messages, success: false };
+    }
+
+    const compressed = applySummary(messages, summaryContent, keepMessages);
+    logger.info({ originalCount: messages.length, compressedCount: compressed.length, tokens: estimateTotalTokens(compressed) }, `AI 总结完成: ${messages.length} → ${compressed.length} 条消息, ~${estimateTotalTokens(compressed)} tokens`);
+    return { messages: compressed, success: true, summaryContent };
+  } catch (err) {
+    logger.error({ err }, 'AI 总结异常');
+    return { messages, success: false };
+  }
+}
+
+// ==================== 单条消息内容压缩 ====================
+
+export interface ContentCompressResult {
+  /** 压缩后的摘要，替换原消息 content 发给目标模型 */
+  summary: string;
+  success: boolean;
+  /** 原文估算 token 数 */
+  originalTokens: number;
+  /** 摘要估算 token 数 */
+  compressedTokens: number;
+}
+
+/**
+ * 使用 AI 压缩单条超长用户消息
+ *
+ * 直接调用 qwen-long（10M 窗口）总结用户贴的超长文档，
+ * 返回摘要文本，由 orchestrator 替换消息 content 发给目标模型。
+ */
+export async function compressSingleMessage(
+  summarizeFn: SummarizeFn,
+  message: ProtocolMessage,
+  model: string,
+): Promise<ContentCompressResult> {
+  const { summarizeMessages } = buildSingleMessageSummarizeRequest(message);
+  const systemPrompt = summarizeMessages[0].content;
+  const userPrompt = summarizeMessages[1].content;
+  const originalTokens = estimateStringTokens(message.content);
+
+  logger.info({ model, originalTokens }, `开始单条消息压缩 (${model}): ~${originalTokens} tokens`);
+
+  try {
+    const summary = await summarizeFn(systemPrompt, userPrompt, model);
+
+    if (!summary.trim()) {
+      logger.warn('单条消息压缩返回空内容');
+      return { summary: '', success: false, originalTokens, compressedTokens: 0 };
+    }
+
+    const compressedTokens = estimateStringTokens(summary);
+    const suffixed = `${summary}\n\n（原文约 ${originalTokens} tokens，已由 AI 压缩提取要点）`;
+    logger.info({ originalTokens, compressedTokens }, `单条消息压缩完成: ~${originalTokens} → ~${compressedTokens} tokens`);
+    return { summary: suffixed, success: true, originalTokens, compressedTokens };
+  } catch (err) {
+    logger.error({ err }, '单条消息压缩异常');
+    return { summary: '', success: false, originalTokens, compressedTokens: 0 };
+  }
+}

package/src/orchestrator/index.ts

@@ -0,0 +1,20 @@
+/**
+ * Orchestrator 模块导出（barrel）
+ *
+ * - ChatOrchestrator：RawEvent → ChatEvent 循环、工具调用与压缩编排
+ * - context-summarizer：超长历史外部总结入口
+ * - types：编排配置与钩子类型
+ */
+
+export { ChatOrchestrator, createOrchestrator } from './orchestrator';
+export { summarizeHistory } from './context-summarizer';
+export type { SummarizeFn, SummarizeResult } from './context-summarizer';
+
+export type {
+  ToolExecutionHooks,
+  OrchestratorToolExecutor,
+  GetAutoRunConfigCallback,
+  OrchestratorConfig,
+  OrchestratorContext,
+  OrchestratorOptions,
+} from './types';