feat: /compress 上下文压缩 + Skill 渐进式加载

2026-04-24 12:35:49 +08:00
parent fe85251738
commit 8e53a411d6
5 changed files with 220 additions and 43 deletions
@@ -1,10 +1,11 @@
 /**
- * Context Manager - 智能上下文管理 (v4.0)
- * 三层策略：滑动窗口 + 消息摘要压缩 + 记忆注入
+ * Context Manager - 智能上下文管理 (v5.0)
+ * 三层策略：滑动窗口 + LLM 摘要压缩 + 记忆注入
+ * 支持自动压缩与手动 /compress 触发
 */

-import type { OllamaMessage } from '../types.js';
-import { logInfo, logWarn } from './log-service.js';
+import type { OllamaMessage, OllamaStreamChunk } from '../types.js';
+import { logInfo, logWarn, logSuccess, logError } from './log-service.js';

 /** 粗略估算 token 数 */
 export function estimateTokens(text: string): number {
@@ -19,6 +20,13 @@ export function estimateTokens(text: string): number {
  return Math.ceil(chineseChars / 1.5 + otherChars / 4);
 }

+/** 自动压缩阈值：当消息 token 占 context window 比例超过此值时触发自动压缩 */
+export const AUTO_COMPRESS_THRESHOLD = 0.5;
+
+/** 压缩后保留首尾消息数 */
+const COMPRESS_KEEP_HEAD = 2;
+const COMPRESS_KEEP_TAIL = 2;
+
 export interface ContextBuildOptions {
  /** 滑动窗口大小（最近 N 条消息完整保留） */
  windowSize?: number;
@@ -41,10 +49,10 @@ const DEFAULT_OPTIONS: Required<ContextBuildOptions> = {
 };

 /**
- * 构建发送给模型的 messages
+ * 构建发送给模型的 messages（同步，滑动窗口）
 * 三层策略：
 *   a. 滑动窗口：最近 N 条消息完整保留
- *   b. 更早的消息：每 N 条压缩为一段摘要
+ *   b. 更早的消息：每 N 条压缩为一段摘要（快速文本截取）
 *   c. 系统 prompt 注入记忆上下文
 */
 export function buildContext(
@@ -64,9 +72,9 @@ export function buildContext(
  }

  // 提取已有的 system 消息
-  const existingSystem = allMessages.find(m => m.role === 'system');
-  if (existingSystem) {
-    systemContent += existingSystem.content;
+  const existingSystem = allMessages.filter(m => m.role === 'system');
+  for (const sys of existingSystem) {
+    systemContent += sys.content + '\n';
  }

  if (systemContent.trim()) {
@@ -77,7 +85,6 @@ export function buildContext(
  const nonSystemMessages = allMessages.filter(m => m.role !== 'system');

  if (nonSystemMessages.length <= opts.windowSize) {
-    // 消息不多，全部保留
    result.push(...nonSystemMessages);
    return result;
  }
@@ -85,11 +92,21 @@ export function buildContext(
  // 滑动窗口：最近 N 条
  const recentMessages = nonSystemMessages.slice(-opts.windowSize);

-  // 更早的消息：压缩为摘要
+  // 更早的消息：已压缩标记的保留原样，未压缩的做快速摘要
  const olderMessages = nonSystemMessages.slice(0, -opts.windowSize);
-  const summaries = summarizeOlderMessages(olderMessages, opts.summaryBatchSize);
+  const compressedMsgs = olderMessages.filter(m => m.compressed);
+  const uncompressedMsgs = olderMessages.filter(m => !m.compressed);

-  result.push(...summaries, ...recentMessages);
+  // 已压缩的消息直接保留
+  result.push(...compressedMsgs);
+
+  // 未压缩的消息做快速摘要
+  if (uncompressedMsgs.length > 0) {
+    const summaries = summarizeOlderMessages(uncompressedMsgs, opts.summaryBatchSize);
+    result.push(...summaries);
+  }
+
+  result.push(...recentMessages);

  // Token 估算和裁剪
  const trimmed = trimByTokenLimit(result, opts.maxTokens);
@@ -101,7 +118,132 @@ export function buildContext(
 }

 /**
- * 将较早的消息每 batchSize 条压缩为一段摘要
+ * 判断是否需要自动压缩
+ * 当总 token 数超过 context window 的 AUTO_COMPRESS_THRESHOLD 比例时返回 true
+ */
+export function shouldAutoCompress(messages: OllamaMessage[], numCtx: number): boolean {
+  const totalTokens = estimateTokens(messages.map(m => m.content || '').join(''));
+  const threshold = numCtx * AUTO_COMPRESS_THRESHOLD;
+  return totalTokens > threshold;
+}
+
+/**
+ * LLM 摘要压缩：调用模型对中间消息生成摘要
+ * 保留首尾各 keepHead/keepTail 条消息，中间用 LLM 摘要替换
+ *
+ * @returns 压缩后的消息列表（包含 compressed 标记的摘要消息）
+ */
+export async function compressWithLLM(
+  messages: OllamaMessage[],
+  api: { chatStream: (params: Record<string, unknown>, onChunk: (chunk: OllamaStreamChunk) => void, ac?: AbortController) => Promise<void> },
+  model: string,
+  options: {
+    keepHead?: number;
+    keepTail?: number;
+    maxSummaryTokens?: number;
+    abortController?: AbortController;
+  } = {}
+): Promise<OllamaMessage[]> {
+  const keepHead = options.keepHead ?? COMPRESS_KEEP_HEAD;
+  const keepTail = options.keepTail ?? COMPRESS_KEEP_TAIL;
+  const maxSummaryTokens = options.maxSummaryTokens ?? 500;
+
+  // 分离 system 和非 system 消息
+  const systemMsgs = messages.filter(m => m.role === 'system');
+  const nonSystemMsgs = messages.filter(m => m.role !== 'system');
+
+  if (nonSystemMsgs.length <= keepHead + keepTail + 2) {
+    logInfo('上下文压缩: 消息太少，跳过压缩');
+    return messages;
+  }
+
+  const head = nonSystemMsgs.slice(0, keepHead);
+  const tail = nonSystemMsgs.slice(-keepTail);
+  const middle = nonSystemMsgs.slice(keepHead, nonSystemMsgs.length - keepTail);
+
+  // 过滤掉已经压缩过的消息（避免重复压缩）
+  const uncompressedMiddle = middle.filter(m => !m.compressed);
+  if (uncompressedMiddle.length === 0) {
+    logInfo('上下文压缩: 中间消息已全部压缩，跳过');
+    return messages;
+  }
+
+  // 构建对话文本
+  const conversationText = uncompressedMiddle.map(m => {
+    const role = m.role === 'user' ? '用户' : 'AI';
+    let content = m.content || '';
+    if (content.length > 500) content = content.slice(0, 500) + '...';
+    if (m.tool_calls?.length) {
+      const toolNames = m.tool_calls.map(t => t.function.name).join(', ');
+      content += ` [工具调用: ${toolNames}]`;
+    }
+    return `${role}: ${content}`;
+  }).join('\n');
+
+  logInfo(`上下文压缩: 开始 LLM 摘要，${uncompressedMiddle.length} 条消息待压缩`);
+
+  let summary = '';
+  try {
+    await api.chatStream(
+      {
+        model,
+        messages: [{
+          role: 'user',
+          content: `请将以下对话摘要为一段简洁的上下文总结。保留关键信息：讨论的主题、得出的结论、用户的偏好、未完成的任务、关键工具调用结果。用中文输出，不超过 ${maxSummaryTokens} 字。\n\n对话记录：\n${conversationText}`
+        }],
+        stream: true,
+        think: false,
+        options: { num_ctx: 8192, temperature: 0.3 }
+      },
+      (chunk: OllamaStreamChunk) => {
+        if (chunk.message?.content) {
+          summary += chunk.message.content;
+        }
+      },
+      options.abortController
+    );
+  } catch (err) {
+    if ((err as Error).name === 'AbortError') {
+      logWarn('上下文压缩: LLM 调用被中止');
+      return messages;
+    }
+    logError('上下文压缩: LLM 调用失败', (err as Error).message);
+    return messages;
+  }
+
+  if (!summary.trim()) {
+    logWarn('上下文压缩: 模型未返回摘要内容');
+    return messages;
+  }
+
+  // 构建压缩后的摘要消息
+  const summaryMsg: OllamaMessage = {
+    role: 'system',
+    content: `📋 以下是对之前对话的摘要（已压缩 ${uncompressedMiddle.length} 条消息）：\n\n${summary}`,
+    compressed: true
+  };
+
+  // 保留已压缩的中间消息 + 新摘要
+  const alreadyCompressed = middle.filter(m => m.compressed);
+
+  const result: OllamaMessage[] = [
+    ...systemMsgs,
+    ...head,
+    ...alreadyCompressed,
+    summaryMsg,
+    ...tail
+  ];
+
+  const beforeTokens = estimateTokens(messages.map(m => m.content || '').join(''));
+  const afterTokens = estimateTokens(result.map(m => m.content || '').join(''));
+
+  logSuccess(`上下文压缩完成: ${messages.length} 条 → ${result.length} 条, tokens: ${beforeTokens} → ${afterTokens}`);
+
+  return result;
+}
+
+/**
+ * 将较早的消息每 batchSize 条压缩为一段摘要（快速文本截取，不调用 LLM）
 */
 function summarizeOlderMessages(messages: OllamaMessage[], batchSize: number): OllamaMessage[] {
  const summaries: OllamaMessage[] = [];
@@ -111,7 +253,8 @@ function summarizeOlderMessages(messages: OllamaMessage[], batchSize: number): O
    const summary = createQuickSummary(batch);
    summaries.push({
      role: 'system',
-      content: `【更早的对话摘要（第 ${Math.floor(i / batchSize) + 1} 部分）】\n${summary}`
+      content: `【更早的对话摘要（第 ${Math.floor(i / batchSize) + 1} 部分）】\n${summary}`,
+      compressed: true
    });
  }

@@ -127,12 +270,10 @@ function createQuickSummary(messages: OllamaMessage[]): string {
  for (const msg of messages) {
    const role = msg.role === 'user' ? '用户' : 'AI';
    const content = msg.content || '';
-    // 取前 100 字符作为摘要
    const preview = content.length > 100 ? content.slice(0, 100) + '...' : content;
    if (preview.trim()) {
      parts.push(`${role}: ${preview}`);
    }
-    // 如果有工具调用，记录工具名
    if (msg.tool_calls?.length) {
      const toolNames = msg.tool_calls.map(t => t.function.name).join(', ');
      parts.push(`  [工具: ${toolNames}]`);
@@ -149,7 +290,6 @@ function trimByTokenLimit(messages: OllamaMessage[], maxTokens: number): OllamaM
  let totalTokens = 0;
  const result: OllamaMessage[] = [];

-  // 从后往前添加（优先保留最近的消息）
  for (let i = messages.length - 1; i >= 0; i--) {
    const msg = messages[i];
    const msgTokens = estimateTokens(msg.content || '') +
@@ -164,7 +304,7 @@ function trimByTokenLimit(messages: OllamaMessage[], maxTokens: number): OllamaM
    }

    if (totalTokens + msgTokens > maxTokens && result.length > 2) {
-      break; // 已到限制，保留至少 system + 1 条消息
+      break;
    }

    totalTokens += msgTokens;