@cup319/mmpl 2.5.0

package/README.md

@@ -0,0 +1,187 @@
+# MMPL MCP
+
+[![npm version](https://img.shields.io/npm/v/mmpl)](https://www.npmjs.com/package/mmpl)
+[![license](https://img.shields.io/npm/l/mmpl)](https://github.com/cnb-t/mmpl/blob/main/LICENSE)
+
+基于 DuckDB 的向量记忆系统，具备完整的记忆生命周期管理，以常驻 HTTP MCP Server 形式运行（v2.5.0+，多 client 共享单进程，避免 DuckDB 文件锁冲突）���
+
+## 核心特性
+
+- **8 个精简工具**：write、search、manage、evolve、compress、palace、symbol、stats
+- **模型自动下载**：首次启动自动从 HuggingFace 下载 ONNX 模型（默认 all-MiniLM-L6-v2，90MB，384维）
+- **跨平台支持**：Windows + Linux，自动适配路径和信号处理
+- **记忆生命周期**：Active → Infrequent → Archived → Permanent（含衰减、强化、SM-2 间隔复习）
+- **混��检索**：BM25（FTS5）+ 向量（HNSW）+ 场景 + 用户画像
+- **L0-L3 分层管道**：原始记录 → 原子事实 → 场景块 → 用户画像
+- **符号压缩**：任务工具调用对 → Mermaid 符号图 + 钻取回放
+- **可选 LLM 分析**：重要性评分、类型分类、实体抽取、情感分析
+
+## 安装
+
+```bash
+npm install @cup319/mmpl
+```
+
+安装后在 `node_modules/mmpl/dist/index.js`。作为 MCP Server 使用时，在配置中指向这个路径即可。
+
+或从 Git 源码构建：
+
+```bash
+git clone https://cnb.cool/cnb-t/mmpl.git
+cd mmpl
+npm install
+npm run build
+```
+
+## 配置
+
+### 部署：常驻 HTTP MCP Server（推荐 — 多 client 共享）
+
+v2.5.0 起，mmpl 默认作为 **Streamable HTTP MCP server** 运行：一个常驻进程独占 DuckDB，被多个 MCP client（ZCode / Claude Desktop / IDE 等）共享。这从根本上避免了 stdio 模式下多 client 各自 spawn 进程、争抢同一个 `.duckdb` 文件导致的 `Connection closed` 错误。
+
+**第一步：启动 server（一次）**
+
+```bash
+node ./node_modules/mmpl/dist/index.js
+# MMPL MCP Server v2.5.0 listening on http://127.0.0.1:7680/mcp
+```
+
+启动时通过环境变量配置 embedding / DB / LLM：
+
+```bash
+EMBEDDING_PROVIDER=onnx \
+DB_PATH=./data/memory.duckdb \
+ONNX_MODEL_PATH=./models/all-MiniLM-L6-v2.onnx \
+ANALYZER_API_URL=https://api.z.ai/api/paas/v4/chat/completions \
+ANALYZER_API_KEY=<your-api-key> \
+ANALYZER_MODEL=glm-4-flash \
+node ./node_modules/mmpl/dist/index.js
+```
+
+**第二步：client 配置（每个 client 一份，只填 URL）**
+
+ZCode (`~/.zcode/cli/config.json`):
+
+```json
+{
+  "mcp": {
+    "servers": {
+      "mmpl": {
+        "type": "http",
+        "url": "http://127.0.0.1:7680/mcp"
+      }
+    }
+  }
+}
+```
+
+Claude Desktop / Claude Code (`claude_desktop_config.json` 或 `~/.claude.json`):
+
+```json
+{
+  "mcpServers": {
+    "mmpl": {
+      "type": "http",
+      "url": "http://127.0.0.1:7680/mcp"
+    }
+  }
+}
+```
+
+**Health check（确认 server 活着）**
+
+```bash
+curl http://127.0.0.1:7680/
+# {"status":"ok","server":"mmpl","sessions":3}
+```
+
+**常驻进程管理（可选）**
+
+HTTP server 启动后不自我守护，靠外部保活：
+
+- **Windows**: `nssm install MemoryPalace node D:\path\to\dist\index.js`，或用 Task Scheduler
+- **Linux/macOS**: systemd unit 或 `pm2 start dist/index.js --name mmpl`
+
+> server 默认只绑 `127.0.0.1`（本机访问）。多 client 共享时**不要**起第二个 server 进程——DuckDB 文件是单进程独占的，第二个进程会启动失败；多个 client 都连同一个 HTTP URL 即可。
+
+### 遗留：stdio 模式
+
+v2.5.0 之前默认走 stdio，每个 client 各 spawn 一个进程。该模式已在 v2.5.0 移除（多 client 场景下 DuckDB 文件锁无法解决）。如确需单 client stdio，可 git checkout v2.4.4 标签。
+
+## 环境变量
+
+**Server 启动时（HTTP 进程的 env）**
+
+| 变量 | 默认值 | 说明 |
+|------|--------|------|
+| `EMBEDDING_PROVIDER` | `onnx` | `onnx` / `api` / `tfidf` |
+| `ONNX_MODEL_PATH` | `./models/all-MiniLM-L6-v2.onnx` | ONNX 模型路径 |
+| `DB_PATH` | `./data/memory.duckdb` | DuckDB 数据库路径 |
+| `ANALYZER_API_URL` | — | LLM API 地址（智谱/OpenAI） |
+| `ANALYZER_API_KEY` | — | LLM API 密钥 |
+| `ANALYZER_MODEL` | `glm-4-flash` | 分析用模型 |
+| `TFIDF_DICT_PATH` | `./data/tfidf-dict.json` | TF-IDF 字典路径（可选，仅 tfidf provider 使用，运行时自动生成） |
+| `MCP_HTTP_HOST` | `127.0.0.1` | HTTP server 绑定地址（`0.0.0.0` 开放外网，需自行加防火墙/鉴权） |
+| `MCP_HTTP_PORT` | `7680` | HTTP server 端口 |
+| `MCP_HTTP_PATH` | `/mcp` | MCP 端点路径 |
+
+## 工具一览
+
+| 工具 | 说明 |
+|------|------|
+| `memory_write` | 写入记忆（encode 编码 + capture 捕获） |
+| `memory_search` | 检索记忆（quick/deep/hybrid/keyword + drill 钻取） |
+| `memory_manage` | 管理生命周期（forget/reinforce/consolidate/reclaim/reembed） |
+| `memory_evolve` | 自动进化（extract 提取 + persona 画像更新） |
+| `memory_compress` | 压缩记忆（compress/decompress/offload） |
+| `memory_palace` | 记忆宫殿（build/explore/visualize/scenes） |
+| `memory_symbol` | 符号记忆（create/search/export） |
+| `memory_stats` | 统计监控（summary/strength/review/pipeline/metrics） |
+
+## ONNX 模型
+
+首次启动时自动从 HuggingFace 下载模型及 tokenizer 文件（~90MB），支持断点续传。
+
+默认使用 `all-MiniLM-L6-v2`（384维，英文优化）。如需多语言支持，设置 `ONNX_MODEL_PATH` 指向 `paraphrase-multilingual-MiniLM-L12-v2`（384维，50+语言，~470MB）。
+
+自动下载失败时，控制台会打印平台适配的手动下载命令。
+
+### 国内网络环境（镜像加速）
+
+HuggingFace 官方源（`huggingface.co`）在国内访问经常超时。MMPL 内置**自动镜像切换**：
+
+- 启动时探测 `huggingface.co`（3 秒超时）
+- 可达 → 用官方源
+- 不可达 → 自动切到国内镜像 `hf-mirror.com`
+
+控制台会打印实际选用的源，例如：
+```
+[ONNX] HuggingFace 官方源不可达 (3000ms 超时)，切换国内镜像 https://hf-mirror.com
+```
+
+**手动指定镜像源**（覆盖自动探测）：设置 `HF_ENDPOINT` 环境变量（HuggingFace 官方约定）：
+
+```bash
+# Windows
+set HF_ENDPOINT=https://hf-mirror.com
+
+# Linux/macOS
+export HF_ENDPOINT=https://hf-mirror.com
+
+# 或者用自定义镜像
+export HF_ENDPOINT=https://your-mirror.example.com
+```
+
+`HF_ENDPOINT` 优先级最高，设置后会跳过自动探测直接使用。
+
+## 开发
+
+```bash
+npm test         # 单元测试（75 用例）
+npm run lint     # 类型检查
+npm run dev      # 开发模式
+```
+
+## 许可证
+
+MIT

package/dist/auto-context-NUQFETSY.js

@@ -0,0 +1,248 @@
+#!/usr/bin/env node
+import {
+  OffloadEngine
+} from "./chunk-JL66S7P7.js";
+import {
+  estimateTokens
+} from "./chunk-CHVFXR24.js";
+import "./chunk-2H7UOFLK.js";
+
+// src/auto-context.ts
+var EMPTY_RESULT = {
+  contextText: "",
+  tokenCount: 0,
+  memoryCount: 0,
+  hasPersona: false,
+  sceneCount: 0,
+  offloadStrategy: "none",
+  pendingL0: 0
+};
+var AutoContextGenerator = class {
+  opts;
+  offloadEngine;
+  constructor(opts) {
+    this.opts = opts;
+    this.offloadEngine = opts.offloadEngine ?? new OffloadEngine({ tokenBudget: opts.tokenBudget ?? 2e3 });
+  }
+  /**
+   * 根据用户查询生成完整的记忆上下文。
+   *
+   * 全程 try-catch，任何子步骤失败都会被吞掉并跳过；整体失败返回空上下文。
+   */
+  async generate(query) {
+    if (!query || !query.trim()) {
+      return { ...EMPTY_RESULT };
+    }
+    try {
+      const topK = this.opts.topK ?? 5;
+      const relevanceThreshold = this.opts.relevanceThreshold ?? 0.5;
+      const scanLimit = this.opts.scanLimit ?? 50;
+      const personaTokenLimit = this.opts.personaTokenLimit ?? 500;
+      const snippetLen = this.opts.memorySnippetLength ?? 120;
+      const entries = [];
+      await this.collectVectorHits(
+        query,
+        entries,
+        relevanceThreshold,
+        scanLimit
+      );
+      await this.collectKeywordHits(query, entries, topK);
+      await this.collectSceneHits(query, entries);
+      const offloadResult = this.offloadEngine.compress(entries);
+      let contextText = "";
+      if (offloadResult.retained.length > 0) {
+        contextText += "## \u76F8\u5173\u8BB0\u5FC6\n";
+        for (const e of offloadResult.retained) {
+          const snippet = e.content.length > snippetLen ? e.content.substring(0, snippetLen) + "\u2026" : e.content;
+          contextText += `- ${snippet}
+`;
+        }
+        if (offloadResult.summary) {
+          contextText += `
+> ${offloadResult.summary}
+`;
+        }
+      }
+      const personaOut = await this.injectPersona(
+        contextText,
+        personaTokenLimit
+      );
+      contextText = personaOut.text;
+      const hasPersona = personaOut.injected;
+      let hasMermaid = false;
+      if (this.opts.mermaidInjector && this.opts.sessionId) {
+        try {
+          const mmdResult = await this.opts.mermaidInjector.injectExisting(this.opts.sessionId);
+          if (mmdResult.injectText) {
+            contextText = mmdResult.injectText + "\n\n" + contextText;
+            hasMermaid = true;
+          }
+        } catch {
+        }
+      }
+      let pendingL0 = 0;
+      try {
+        const raw = await this.opts.storage.getUnconsumedRaw(100);
+        pendingL0 = Array.isArray(raw) ? raw.length : 0;
+      } catch {
+      }
+      const tokenCount = estimateTokens(contextText);
+      return {
+        contextText,
+        tokenCount,
+        memoryCount: offloadResult.retained.filter((e) => e.layer === 1).length,
+        hasPersona,
+        sceneCount: offloadResult.retained.filter((e) => e.layer === 2).length,
+        offloadStrategy: offloadResult.strategy,
+        pendingL0
+      };
+    } catch (err) {
+      console.error(
+        "[auto-context] generate failed:",
+        err instanceof Error ? err.message : err
+      );
+      return { ...EMPTY_RESULT };
+    }
+  }
+  /**
+   * 步骤 1: 向量检索 L1 active memories。
+   * 用 cosine 相似度过滤，命中阈值则加入 entries。
+   */
+  async collectVectorHits(query, entries, relevanceThreshold, scanLimit) {
+    try {
+      if (!this.opts.embeddingService?.available) return;
+      const emb = await this.opts.embeddingService.embed(query);
+      if (!emb || emb.length === 0) return;
+      const memories = await this.opts.storage.getActiveMemories();
+      if (!Array.isArray(memories)) return;
+      for (const m of memories.slice(0, scanLimit)) {
+        let mEmb = m.embedding;
+        if (typeof mEmb === "string") {
+          try {
+            mEmb = JSON.parse(mEmb);
+          } catch {
+            mEmb = null;
+          }
+        }
+        if (!Array.isArray(mEmb) || mEmb.length === 0) continue;
+        const sim = this.cosineSim(emb, mEmb);
+        if (sim >= relevanceThreshold) {
+          entries.push({
+            id: m.id,
+            content: m.content,
+            score: sim,
+            layer: 1,
+            importance: typeof m.importance === "number" ? m.importance : 0.5,
+            strength: typeof m.strength === "number" ? m.strength : 0.5,
+            source: "vector"
+          });
+        }
+      }
+    } catch {
+    }
+  }
+  /**
+   * 步骤 2: BM25 关键词检索补充 L1。
+   * 只加入尚未被向量召回命中的记忆。
+   */
+  async collectKeywordHits(query, entries, topK) {
+    try {
+      const results = await this.opts.storage.searchByKeywords(
+        query,
+        topK,
+        "active"
+      );
+      if (!Array.isArray(results)) return;
+      for (const m of results) {
+        if (entries.find((e) => e.id === m.id)) continue;
+        entries.push({
+          id: m.id,
+          content: m.content,
+          score: 0.7,
+          layer: 1,
+          importance: typeof m.importance === "number" ? m.importance : 0.5,
+          strength: typeof m.strength === "number" ? m.strength : 0.5,
+          source: "bm25"
+        });
+      }
+    } catch {
+    }
+  }
+  /**
+   * 步骤 3: 场景匹配 L2。
+   * 用 label / summary 子串匹配（大小写不敏感）。
+   */
+  async collectSceneHits(query, entries) {
+    try {
+      const scenes = await this.opts.storage.getAllScenes();
+      if (!Array.isArray(scenes)) return;
+      const qLower = query.toLowerCase();
+      for (const s of scenes) {
+        const label = (s.label ?? "").toLowerCase();
+        const summary = (s.summary ?? "").toLowerCase();
+        if (label.includes(qLower) || summary.includes(qLower)) {
+          entries.push({
+            id: s.id,
+            content: `[\u573A\u666F] ${s.label ?? s.id}: ${(s.summary ?? "").substring(
+              0,
+              80
+            )}`,
+            score: 0.6,
+            layer: 2,
+            importance: 0.7,
+            strength: 1,
+            source: "scene"
+          });
+        }
+      }
+    } catch {
+    }
+  }
+  /**
+   * 步骤 6: 注入 L3 用户画像。
+   * 画像内容不参与 offload 压缩，但有自己的 token 上限保护。
+   * 返回更新后的文本和是否注入标志。
+   */
+  async injectPersona(contextText, personaTokenLimit) {
+    try {
+      const persona = await this.opts.storage.getLatestPersona();
+      if (!persona || !persona.content) {
+        return { text: contextText, injected: false };
+      }
+      const personaTokens = estimateTokens(persona.content);
+      if (personaTokens >= personaTokenLimit) {
+        return { text: contextText, injected: false };
+      }
+      const text = contextText + `
+## \u7528\u6237\u753B\u50CF
+${persona.content.substring(0, 300)}
+`;
+      return { text, injected: true };
+    } catch {
+      return { text: contextText, injected: false };
+    }
+  }
+  /**
+   * 余弦相似度（内联实现，不依赖 retrieval.ts）
+   * 不等长或空向量返回 0。
+   */
+  cosineSim(a, b) {
+    if (!a || !b || a.length === 0 || b.length === 0 || a.length !== b.length) {
+      return 0;
+    }
+    let dot = 0;
+    let magA = 0;
+    let magB = 0;
+    for (let i = 0; i < a.length; i++) {
+      dot += a[i] * b[i];
+      magA += a[i] * a[i];
+      magB += b[i] * b[i];
+    }
+    const denom = Math.sqrt(magA) * Math.sqrt(magB);
+    return denom > 0 ? dot / denom : 0;
+  }
+};
+export {
+  AutoContextGenerator
+};
+//# sourceMappingURL=auto-context-NUQFETSY.js.map

package/dist/auto-context-NUQFETSY.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/auto-context.ts"],"sourcesContent":["/**\n * 自动召回上下文生成器 (Auto-Context Generator)\n *\n * 背景:\n *   MCP Server 无法 Hook agent 的生命周期（不像 OpenClaw 这类插件框架可以\n *   拦截 before_prompt_build），因此这里实现一个\"一站式\"工具: agent 在每轮\n *   对话开始时调用一次 generate(query)，即可拿到该轮所需的全部记忆上下文。\n *\n *   这相当于 TencentDB AgentMemory 的 before_prompt_build Hook 的 MCP 版本。\n *\n * 流程:\n *   1. 向量检索 L1 active memories（语义相似）\n *   2. BM25 关键词检索 L1 补充（覆盖向量召回遗漏）\n *   3. 场景匹配 L2（label / summary 子串匹配）\n *   4. L3 用户画像固定注入（不参与压缩）\n *   5. 检查 L0 待处理计数（提示 agent 是否需要 consolidate）\n *   6. 用 OffloadEngine 把 L1+L2 压缩到 token 预算内\n *   7. 格式化为可读 Markdown 文本\n *\n * 全程 try-catch，任何子步骤失败都不会抛出，最坏返回空上下文。\n * 不依赖任何外部库。\n */\n\nimport { OffloadEngine, OffloadEntry } from \"./offload-engine.js\";\nimport { estimateTokens } from \"./token-counter.js\";\n\n/** AutoContextGenerator 配置选项 */\nexport interface AutoContextOptions {\n  /** 存储层（鸭子类型，提供 getActiveMemories / getAllScenes / getLatestPersona / searchByKeywords / getUnconsumedRaw 等） */\n  storage: any;\n  /** Embedding 服务（提供 available / embed） */\n  embeddingService: any;\n  /** 可选的 OffloadEngine，不传则按 tokenBudget 自动创建 */\n  offloadEngine?: OffloadEngine;\n  /** 可选的 MermaidInjector，提供时自动注入任务符号图 */\n  mermaidInjector?: any;\n  /** 可选的 session_id，用于注入已有 Mermaid 会话图 */\n  sessionId?: string;\n  /** 上下文 token 预算，默认 2000 */\n  tokenBudget?: number;\n  /** 每层取 top-k，默认 5 */\n  topK?: number;\n  /** 相关性阈值，默认 0.5 */\n  relevanceThreshold?: number;\n  /** 向量扫描上限（性能保护），默认 50 */\n  scanLimit?: number;\n  /** 画像内容 token 上限，超过则不注入，默认 500 */\n  personaTokenLimit?: number;\n  /** 单条记忆输出截断长度，默认 120 */\n  memorySnippetLength?: number;\n}\n\n/** generate() 返回结果 */\nexport interface AutoContextResult {\n  /** 注入到 prompt 前面的记忆上下文文本 */\n  contextText: string;\n  /** 上下文 token 数 */\n  tokenCount: number;\n  /** 命中的记忆条数（L1 retained） */\n  memoryCount: number;\n  /** 是否包含用户画像 */\n  hasPersona: boolean;\n  /** 命中的场景数（L2 retained） */\n  sceneCount: number;\n  /** 压缩策略名称 */\n  offloadStrategy: string;\n  /** 未消费的 L0 计数 */\n  pendingL0: number;\n}\n\n/** 空结果常量，失败时统一返回 */\nconst EMPTY_RESULT: AutoContextResult = {\n  contextText: \"\",\n  tokenCount: 0,\n  memoryCount: 0,\n  hasPersona: false,\n  sceneCount: 0,\n  offloadStrategy: \"none\",\n  pendingL0: 0,\n};\n\n/**\n * 自动召回上下文生成器\n */\nexport class AutoContextGenerator {\n  private opts: AutoContextOptions;\n  private offloadEngine: OffloadEngine;\n\n  constructor(opts: AutoContextOptions) {\n    this.opts = opts;\n    // 若未提供 offloadEngine，按 tokenBudget 自动创建一个\n    this.offloadEngine =\n      opts.offloadEngine ??\n      new OffloadEngine({ tokenBudget: opts.tokenBudget ?? 2000 });\n  }\n\n  /**\n   * 根据用户查询生成完整的记忆上下文。\n   *\n   * 全程 try-catch，任何子步骤失败都会被吞掉并跳过；整体失败返回空上下文。\n   */\n  async generate(query: string): Promise<AutoContextResult> {\n    // 防御：空 query 直接返回空上下文\n    if (!query || !query.trim()) {\n      return { ...EMPTY_RESULT };\n    }\n\n    try {\n      const topK = this.opts.topK ?? 5;\n      const relevanceThreshold = this.opts.relevanceThreshold ?? 0.5;\n      const scanLimit = this.opts.scanLimit ?? 50;\n      const personaTokenLimit = this.opts.personaTokenLimit ?? 500;\n      const snippetLen = this.opts.memorySnippetLength ?? 120;\n\n      const entries: OffloadEntry[] = [];\n\n      // ===== 1. 向量检索 L1 active memories =====\n      await this.collectVectorHits(\n        query,\n        entries,\n        relevanceThreshold,\n        scanLimit\n      );\n\n      // ===== 2. BM25 关键词检索补充 L1 =====\n      await this.collectKeywordHits(query, entries, topK);\n\n      // ===== 3. 场景匹配 L2 =====\n      await this.collectSceneHits(query, entries);\n\n      // ===== 4. 用 OffloadEngine 压缩到 token 预算 =====\n      const offloadResult = this.offloadEngine.compress(entries);\n\n      // ===== 5. 格式化记忆 + 场景文本 =====\n      let contextText = \"\";\n      if (offloadResult.retained.length > 0) {\n        contextText += \"## 相关记忆\\n\";\n        for (const e of offloadResult.retained) {\n          const snippet =\n            e.content.length > snippetLen\n              ? e.content.substring(0, snippetLen) + \"…\"\n              : e.content;\n          contextText += `- ${snippet}\\n`;\n        }\n        if (offloadResult.summary) {\n          contextText += `\\n> ${offloadResult.summary}\\n`;\n        }\n      }\n\n      // ===== 6. L3 用户画像（固定注入，不参与压缩） =====\n      const personaOut = await this.injectPersona(\n        contextText,\n        personaTokenLimit\n      );\n      contextText = personaOut.text;\n      const hasPersona = personaOut.injected;\n\n      // ===== 6.5 Mermaid 任务符号图注入 =====\n      let hasMermaid = false;\n      if (this.opts.mermaidInjector && this.opts.sessionId) {\n        try {\n          const mmdResult = await this.opts.mermaidInjector.injectExisting(this.opts.sessionId);\n          if (mmdResult.injectText) {\n            contextText = mmdResult.injectText + \"\\n\\n\" + contextText;\n            hasMermaid = true;\n          }\n        } catch {\n          /* ignore mermaid injection failure */\n        }\n      }\n\n      // ===== 7. L0 待处理计数 =====\n      let pendingL0 = 0;\n      try {\n        const raw = await this.opts.storage.getUnconsumedRaw(100);\n        pendingL0 = Array.isArray(raw) ? raw.length : 0;\n      } catch {\n        /* ignore */\n      }\n\n      const tokenCount = estimateTokens(contextText);\n\n      return {\n        contextText,\n        tokenCount,\n        memoryCount: offloadResult.retained.filter((e) => e.layer === 1).length,\n        hasPersona,\n        sceneCount: offloadResult.retained.filter((e) => e.layer === 2).length,\n        offloadStrategy: offloadResult.strategy,\n        pendingL0,\n      };\n    } catch (err) {\n      // 任何意外错误都返回空上下文，绝不抛出\n      console.error(\n        \"[auto-context] generate failed:\",\n        err instanceof Error ? err.message : err\n      );\n      return { ...EMPTY_RESULT };\n    }\n  }\n\n  /**\n   * 步骤 1: 向量检索 L1 active memories。\n   * 用 cosine 相似度过滤，命中阈值则加入 entries。\n   */\n  private async collectVectorHits(\n    query: string,\n    entries: OffloadEntry[],\n    relevanceThreshold: number,\n    scanLimit: number\n  ): Promise<void> {\n    try {\n      if (!this.opts.embeddingService?.available) return;\n      const emb = await this.opts.embeddingService.embed(query);\n      if (!emb || emb.length === 0) return;\n\n      const memories = await this.opts.storage.getActiveMemories();\n      if (!Array.isArray(memories)) return;\n\n      for (const m of memories.slice(0, scanLimit)) {\n        // embedding 可能存成 JSON 字符串\n        let mEmb: any = m.embedding;\n        if (typeof mEmb === \"string\") {\n          try {\n            mEmb = JSON.parse(mEmb);\n          } catch {\n            mEmb = null;\n          }\n        }\n        if (!Array.isArray(mEmb) || mEmb.length === 0) continue;\n\n        const sim = this.cosineSim(emb, mEmb);\n        if (sim >= relevanceThreshold) {\n          entries.push({\n            id: m.id,\n            content: m.content,\n            score: sim,\n            layer: 1,\n            importance: typeof m.importance === \"number\" ? m.importance : 0.5,\n            strength: typeof m.strength === \"number\" ? m.strength : 0.5,\n            source: \"vector\",\n          });\n        }\n      }\n    } catch {\n      /* ignore — 向量检索失败不影响后续步骤 */\n    }\n  }\n\n  /**\n   * 步骤 2: BM25 关键词检索补充 L1。\n   * 只加入尚未被向量召回命中的记忆。\n   */\n  private async collectKeywordHits(\n    query: string,\n    entries: OffloadEntry[],\n    topK: number\n  ): Promise<void> {\n    try {\n      const results = await this.opts.storage.searchByKeywords(\n        query,\n        topK,\n        \"active\"\n      );\n      if (!Array.isArray(results)) return;\n      for (const m of results) {\n        if (entries.find((e) => e.id === m.id)) continue;\n        entries.push({\n          id: m.id,\n          content: m.content,\n          score: 0.7,\n          layer: 1,\n          importance: typeof m.importance === \"number\" ? m.importance : 0.5,\n          strength: typeof m.strength === \"number\" ? m.strength : 0.5,\n          source: \"bm25\",\n        });\n      }\n    } catch {\n      /* ignore */\n    }\n  }\n\n  /**\n   * 步骤 3: 场景匹配 L2。\n   * 用 label / summary 子串匹配（大小写不敏感）。\n   */\n  private async collectSceneHits(\n    query: string,\n    entries: OffloadEntry[]\n  ): Promise<void> {\n    try {\n      const scenes = await this.opts.storage.getAllScenes();\n      if (!Array.isArray(scenes)) return;\n      const qLower = query.toLowerCase();\n      for (const s of scenes) {\n        const label = (s.label ?? \"\").toLowerCase();\n        const summary = (s.summary ?? \"\").toLowerCase();\n        if (label.includes(qLower) || summary.includes(qLower)) {\n          entries.push({\n            id: s.id,\n            content: `[场景] ${s.label ?? s.id}: ${(s.summary ?? \"\").substring(\n              0,\n              80\n            )}`,\n            score: 0.6,\n            layer: 2,\n            importance: 0.7,\n            strength: 1.0,\n            source: \"scene\",\n          });\n        }\n      }\n    } catch {\n      /* ignore */\n    }\n  }\n\n  /**\n   * 步骤 6: 注入 L3 用户画像。\n   * 画像内容不参与 offload 压缩，但有自己的 token 上限保护。\n   * 返回更新后的文本和是否注入标志。\n   */\n  private async injectPersona(\n    contextText: string,\n    personaTokenLimit: number\n  ): Promise<{ text: string; injected: boolean }> {\n    try {\n      const persona = await this.opts.storage.getLatestPersona();\n      if (!persona || !persona.content) {\n        return { text: contextText, injected: false };\n      }\n      const personaTokens = estimateTokens(persona.content);\n      if (personaTokens >= personaTokenLimit) {\n        return { text: contextText, injected: false };\n      }\n      const text =\n        contextText +\n        `\\n## 用户画像\\n${persona.content.substring(0, 300)}\\n`;\n      return { text, injected: true };\n    } catch {\n      return { text: contextText, injected: false };\n    }\n  }\n\n  /**\n   * 余弦相似度（内联实现，不依赖 retrieval.ts）\n   * 不等长或空向量返回 0。\n   */\n  private cosineSim(a: number[], b: number[]): number {\n    if (\n      !a ||\n      !b ||\n      a.length === 0 ||\n      b.length === 0 ||\n      a.length !== b.length\n    ) {\n      return 0;\n    }\n    let dot = 0;\n    let magA = 0;\n    let magB = 0;\n    for (let i = 0; i < a.length; i++) {\n      dot += a[i] * b[i];\n      magA += a[i] * a[i];\n      magB += b[i] * b[i];\n    }\n    const denom = Math.sqrt(magA) * Math.sqrt(magB);\n    return denom > 0 ? dot / denom : 0;\n  }\n}\n"],"mappings":";;;;;;;;;;AAuEA,IAAM,eAAkC;AAAA,EACtC,aAAa;AAAA,EACb,YAAY;AAAA,EACZ,aAAa;AAAA,EACb,YAAY;AAAA,EACZ,YAAY;AAAA,EACZ,iBAAiB;AAAA,EACjB,WAAW;AACb;AAKO,IAAM,uBAAN,MAA2B;AAAA,EACxB;AAAA,EACA;AAAA,EAER,YAAY,MAA0B;AACpC,SAAK,OAAO;AAEZ,SAAK,gBACH,KAAK,iBACL,IAAI,cAAc,EAAE,aAAa,KAAK,eAAe,IAAK,CAAC;AAAA,EAC/D;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,SAAS,OAA2C;AAExD,QAAI,CAAC,SAAS,CAAC,MAAM,KAAK,GAAG;AAC3B,aAAO,EAAE,GAAG,aAAa;AAAA,IAC3B;AAEA,QAAI;AACF,YAAM,OAAO,KAAK,KAAK,QAAQ;AAC/B,YAAM,qBAAqB,KAAK,KAAK,sBAAsB;AAC3D,YAAM,YAAY,KAAK,KAAK,aAAa;AACzC,YAAM,oBAAoB,KAAK,KAAK,qBAAqB;AACzD,YAAM,aAAa,KAAK,KAAK,uBAAuB;AAEpD,YAAM,UAA0B,CAAC;AAGjC,YAAM,KAAK;AAAA,QACT;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AAGA,YAAM,KAAK,mBAAmB,OAAO,SAAS,IAAI;AAGlD,YAAM,KAAK,iBAAiB,OAAO,OAAO;AAG1C,YAAM,gBAAgB,KAAK,cAAc,SAAS,OAAO;AAGzD,UAAI,cAAc;AAClB,UAAI,cAAc,SAAS,SAAS,GAAG;AACrC,uBAAe;AACf,mBAAW,KAAK,cAAc,UAAU;AACtC,gBAAM,UACJ,EAAE,QAAQ,SAAS,aACf,EAAE,QAAQ,UAAU,GAAG,UAAU,IAAI,WACrC,EAAE;AACR,yBAAe,KAAK,OAAO;AAAA;AAAA,QAC7B;AACA,YAAI,cAAc,SAAS;AACzB,yBAAe;AAAA,IAAO,cAAc,OAAO;AAAA;AAAA,QAC7C;AAAA,MACF;AAGA,YAAM,aAAa,MAAM,KAAK;AAAA,QAC5B;AAAA,QACA;AAAA,MACF;AACA,oBAAc,WAAW;AACzB,YAAM,aAAa,WAAW;AAG9B,UAAI,aAAa;AACjB,UAAI,KAAK,KAAK,mBAAmB,KAAK,KAAK,WAAW;AACpD,YAAI;AACF,gBAAM,YAAY,MAAM,KAAK,KAAK,gBAAgB,eAAe,KAAK,KAAK,SAAS;AACpF,cAAI,UAAU,YAAY;AACxB,0BAAc,UAAU,aAAa,SAAS;AAC9C,yBAAa;AAAA,UACf;AAAA,QACF,QAAQ;AAAA,QAER;AAAA,MACF;AAGA,UAAI,YAAY;AAChB,UAAI;AACF,cAAM,MAAM,MAAM,KAAK,KAAK,QAAQ,iBAAiB,GAAG;AACxD,oBAAY,MAAM,QAAQ,GAAG,IAAI,IAAI,SAAS;AAAA,MAChD,QAAQ;AAAA,MAER;AAEA,YAAM,aAAa,eAAe,WAAW;AAE7C,aAAO;AAAA,QACL;AAAA,QACA;AAAA,QACA,aAAa,cAAc,SAAS,OAAO,CAAC,MAAM,EAAE,UAAU,CAAC,EAAE;AAAA,QACjE;AAAA,QACA,YAAY,cAAc,SAAS,OAAO,CAAC,MAAM,EAAE,UAAU,CAAC,EAAE;AAAA,QAChE,iBAAiB,cAAc;AAAA,QAC/B;AAAA,MACF;AAAA,IACF,SAAS,KAAK;AAEZ,cAAQ;AAAA,QACN;AAAA,QACA,eAAe,QAAQ,IAAI,UAAU;AAAA,MACvC;AACA,aAAO,EAAE,GAAG,aAAa;AAAA,IAC3B;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,kBACZ,OACA,SACA,oBACA,WACe;AACf,QAAI;AACF,UAAI,CAAC,KAAK,KAAK,kBAAkB,UAAW;AAC5C,YAAM,MAAM,MAAM,KAAK,KAAK,iBAAiB,MAAM,KAAK;AACxD,UAAI,CAAC,OAAO,IAAI,WAAW,EAAG;AAE9B,YAAM,WAAW,MAAM,KAAK,KAAK,QAAQ,kBAAkB;AAC3D,UAAI,CAAC,MAAM,QAAQ,QAAQ,EAAG;AAE9B,iBAAW,KAAK,SAAS,MAAM,GAAG,SAAS,GAAG;AAE5C,YAAI,OAAY,EAAE;AAClB,YAAI,OAAO,SAAS,UAAU;AAC5B,cAAI;AACF,mBAAO,KAAK,MAAM,IAAI;AAAA,UACxB,QAAQ;AACN,mBAAO;AAAA,UACT;AAAA,QACF;AACA,YAAI,CAAC,MAAM,QAAQ,IAAI,KAAK,KAAK,WAAW,EAAG;AAE/C,cAAM,MAAM,KAAK,UAAU,KAAK,IAAI;AACpC,YAAI,OAAO,oBAAoB;AAC7B,kBAAQ,KAAK;AAAA,YACX,IAAI,EAAE;AAAA,YACN,SAAS,EAAE;AAAA,YACX,OAAO;AAAA,YACP,OAAO;AAAA,YACP,YAAY,OAAO,EAAE,eAAe,WAAW,EAAE,aAAa;AAAA,YAC9D,UAAU,OAAO,EAAE,aAAa,WAAW,EAAE,WAAW;AAAA,YACxD,QAAQ;AAAA,UACV,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,mBACZ,OACA,SACA,MACe;AACf,QAAI;AACF,YAAM,UAAU,MAAM,KAAK,KAAK,QAAQ;AAAA,QACtC;AAAA,QACA;AAAA,QACA;AAAA,MACF;AACA,UAAI,CAAC,MAAM,QAAQ,OAAO,EAAG;AAC7B,iBAAW,KAAK,SAAS;AACvB,YAAI,QAAQ,KAAK,CAAC,MAAM,EAAE,OAAO,EAAE,EAAE,EAAG;AACxC,gBAAQ,KAAK;AAAA,UACX,IAAI,EAAE;AAAA,UACN,SAAS,EAAE;AAAA,UACX,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY,OAAO,EAAE,eAAe,WAAW,EAAE,aAAa;AAAA,UAC9D,UAAU,OAAO,EAAE,aAAa,WAAW,EAAE,WAAW;AAAA,UACxD,QAAQ;AAAA,QACV,CAAC;AAAA,MACH;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,iBACZ,OACA,SACe;AACf,QAAI;AACF,YAAM,SAAS,MAAM,KAAK,KAAK,QAAQ,aAAa;AACpD,UAAI,CAAC,MAAM,QAAQ,MAAM,EAAG;AAC5B,YAAM,SAAS,MAAM,YAAY;AACjC,iBAAW,KAAK,QAAQ;AACtB,cAAM,SAAS,EAAE,SAAS,IAAI,YAAY;AAC1C,cAAM,WAAW,EAAE,WAAW,IAAI,YAAY;AAC9C,YAAI,MAAM,SAAS,MAAM,KAAK,QAAQ,SAAS,MAAM,GAAG;AACtD,kBAAQ,KAAK;AAAA,YACX,IAAI,EAAE;AAAA,YACN,SAAS,kBAAQ,EAAE,SAAS,EAAE,EAAE,MAAM,EAAE,WAAW,IAAI;AAAA,cACrD;AAAA,cACA;AAAA,YACF,CAAC;AAAA,YACD,OAAO;AAAA,YACP,OAAO;AAAA,YACP,YAAY;AAAA,YACZ,UAAU;AAAA,YACV,QAAQ;AAAA,UACV,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAc,cACZ,aACA,mBAC8C;AAC9C,QAAI;AACF,YAAM,UAAU,MAAM,KAAK,KAAK,QAAQ,iBAAiB;AACzD,UAAI,CAAC,WAAW,CAAC,QAAQ,SAAS;AAChC,eAAO,EAAE,MAAM,aAAa,UAAU,MAAM;AAAA,MAC9C;AACA,YAAM,gBAAgB,eAAe,QAAQ,OAAO;AACpD,UAAI,iBAAiB,mBAAmB;AACtC,eAAO,EAAE,MAAM,aAAa,UAAU,MAAM;AAAA,MAC9C;AACA,YAAM,OACJ,cACA;AAAA;AAAA,EAAc,QAAQ,QAAQ,UAAU,GAAG,GAAG,CAAC;AAAA;AACjD,aAAO,EAAE,MAAM,UAAU,KAAK;AAAA,IAChC,QAAQ;AACN,aAAO,EAAE,MAAM,aAAa,UAAU,MAAM;AAAA,IAC9C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,UAAU,GAAa,GAAqB;AAClD,QACE,CAAC,KACD,CAAC,KACD,EAAE,WAAW,KACb,EAAE,WAAW,KACb,EAAE,WAAW,EAAE,QACf;AACA,aAAO;AAAA,IACT;AACA,QAAI,MAAM;AACV,QAAI,OAAO;AACX,QAAI,OAAO;AACX,aAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,aAAO,EAAE,CAAC,IAAI,EAAE,CAAC;AACjB,cAAQ,EAAE,CAAC,IAAI,EAAE,CAAC;AAClB,cAAQ,EAAE,CAAC,IAAI,EAAE,CAAC;AAAA,IACpB;AACA,UAAM,QAAQ,KAAK,KAAK,IAAI,IAAI,KAAK,KAAK,IAAI;AAC9C,WAAO,QAAQ,IAAI,MAAM,QAAQ;AAAA,EACnC;AACF;","names":[]}

package/dist/checkpoint-ABRNCBCA.js

@@ -0,0 +1,68 @@
+#!/usr/bin/env node
+import "./chunk-2H7UOFLK.js";
+
+// src/checkpoint.ts
+import { promises as fs } from "fs";
+import path from "path";
+var CheckpointManager = class {
+  checkpointPath;
+  checkpoint;
+  dirty = false;
+  constructor(dataDir) {
+    this.checkpointPath = path.join(dataDir, ".metadata", "pipeline_checkpoint.json");
+    this.checkpoint = {
+      lastL1ProcessedRawId: null,
+      lastL1Time: 0,
+      lastL2Time: 0,
+      lastL3Time: 0,
+      totalL0Processed: 0,
+      totalL1Extracted: 0,
+      totalL2Built: 0,
+      totalL3Versions: 0,
+      updatedAt: 0
+    };
+  }
+  async init() {
+    try {
+      const raw = await fs.readFile(this.checkpointPath, "utf-8");
+      const parsed = JSON.parse(raw);
+      this.checkpoint = { ...this.checkpoint, ...parsed };
+    } catch {
+    }
+  }
+  recordL1(rawId, extracted) {
+    this.checkpoint.lastL1ProcessedRawId = rawId;
+    this.checkpoint.lastL1Time = Date.now();
+    this.checkpoint.totalL0Processed++;
+    this.checkpoint.totalL1Extracted += extracted;
+    this.dirty = true;
+  }
+  recordL2(sceneCount) {
+    this.checkpoint.lastL2Time = Date.now();
+    this.checkpoint.totalL2Built += sceneCount;
+    this.dirty = true;
+  }
+  recordL3() {
+    this.checkpoint.lastL3Time = Date.now();
+    this.checkpoint.totalL3Versions++;
+    this.dirty = true;
+  }
+  getCheckpoint() {
+    return { ...this.checkpoint };
+  }
+  /** 持久化到磁盘（原子写入：tmp + rename） */
+  async flush() {
+    if (!this.dirty) return;
+    this.checkpoint.updatedAt = Date.now();
+    const dir = path.dirname(this.checkpointPath);
+    await fs.mkdir(dir, { recursive: true });
+    const tmpPath = this.checkpointPath + ".tmp";
+    await fs.writeFile(tmpPath, JSON.stringify(this.checkpoint, null, 2), "utf-8");
+    await fs.rename(tmpPath, this.checkpointPath);
+    this.dirty = false;
+  }
+};
+export {
+  CheckpointManager
+};
+//# sourceMappingURL=checkpoint-ABRNCBCA.js.map

package/dist/checkpoint-ABRNCBCA.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/checkpoint.ts"],"sourcesContent":["import { promises as fs } from \"node:fs\";\nimport path from \"node:path\";\n\n/**\n * 管道检查点 — 追踪记忆处理进度，支持崩溃恢复\n *\n * 灵感: TencentDB Agent Memory 的 checkpoint.ts\n */\n\nexport interface PipelineCheckpoint {\n  /** 最后一次 L1 提取处理的 L0 记录 ID */\n  lastL1ProcessedRawId: string | null;\n  /** 最后一次 L1 提取时间 */\n  lastL1Time: number;\n  /** 最后一次 L2 场景构建时间 */\n  lastL2Time: number;\n  /** 最后一次 L3 画像生成时间 */\n  lastL3Time: number;\n  /** 已处理的 L0 记录总数 */\n  totalL0Processed: number;\n  /** 已提取的 L1 记忆总数 */\n  totalL1Extracted: number;\n  /** 已构建的 L2 场景总数 */\n  totalL2Built: number;\n  /** 已生成的 L3 画像版本数 */\n  totalL3Versions: number;\n  /** 检查点最后更新时间 */\n  updatedAt: number;\n}\n\nexport class CheckpointManager {\n  private checkpointPath: string;\n  private checkpoint: PipelineCheckpoint;\n  private dirty: boolean = false;\n\n  constructor(dataDir: string) {\n    this.checkpointPath = path.join(dataDir, \".metadata\", \"pipeline_checkpoint.json\");\n    this.checkpoint = {\n      lastL1ProcessedRawId: null,\n      lastL1Time: 0, lastL2Time: 0, lastL3Time: 0,\n      totalL0Processed: 0, totalL1Extracted: 0, totalL2Built: 0, totalL3Versions: 0,\n      updatedAt: 0,\n    };\n  }\n\n  async init(): Promise<void> {\n    try {\n      const raw = await fs.readFile(this.checkpointPath, \"utf-8\");\n      const parsed = JSON.parse(raw);\n      this.checkpoint = { ...this.checkpoint, ...parsed };\n    } catch { /* first run */ }\n  }\n\n  recordL1(rawId: string, extracted: number): void {\n    this.checkpoint.lastL1ProcessedRawId = rawId;\n    this.checkpoint.lastL1Time = Date.now();\n    this.checkpoint.totalL0Processed++;\n    this.checkpoint.totalL1Extracted += extracted;\n    this.dirty = true;\n  }\n\n  recordL2(sceneCount: number): void {\n    this.checkpoint.lastL2Time = Date.now();\n    this.checkpoint.totalL2Built += sceneCount;\n    this.dirty = true;\n  }\n\n  recordL3(): void {\n    this.checkpoint.lastL3Time = Date.now();\n    this.checkpoint.totalL3Versions++;\n    this.dirty = true;\n  }\n\n  getCheckpoint(): PipelineCheckpoint { return { ...this.checkpoint }; }\n\n  /** 持久化到磁盘（原子写入：tmp + rename） */\n  async flush(): Promise<void> {\n    if (!this.dirty) return;\n    this.checkpoint.updatedAt = Date.now();\n    const dir = path.dirname(this.checkpointPath);\n    await fs.mkdir(dir, { recursive: true });\n    const tmpPath = this.checkpointPath + \".tmp\";\n    await fs.writeFile(tmpPath, JSON.stringify(this.checkpoint, null, 2), \"utf-8\");\n    await fs.rename(tmpPath, this.checkpointPath);\n    this.dirty = false;\n  }\n}\n"],"mappings":";;;;AAAA,SAAS,YAAY,UAAU;AAC/B,OAAO,UAAU;AA6BV,IAAM,oBAAN,MAAwB;AAAA,EACrB;AAAA,EACA;AAAA,EACA,QAAiB;AAAA,EAEzB,YAAY,SAAiB;AAC3B,SAAK,iBAAiB,KAAK,KAAK,SAAS,aAAa,0BAA0B;AAChF,SAAK,aAAa;AAAA,MAChB,sBAAsB;AAAA,MACtB,YAAY;AAAA,MAAG,YAAY;AAAA,MAAG,YAAY;AAAA,MAC1C,kBAAkB;AAAA,MAAG,kBAAkB;AAAA,MAAG,cAAc;AAAA,MAAG,iBAAiB;AAAA,MAC5E,WAAW;AAAA,IACb;AAAA,EACF;AAAA,EAEA,MAAM,OAAsB;AAC1B,QAAI;AACF,YAAM,MAAM,MAAM,GAAG,SAAS,KAAK,gBAAgB,OAAO;AAC1D,YAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,WAAK,aAAa,EAAE,GAAG,KAAK,YAAY,GAAG,OAAO;AAAA,IACpD,QAAQ;AAAA,IAAkB;AAAA,EAC5B;AAAA,EAEA,SAAS,OAAe,WAAyB;AAC/C,SAAK,WAAW,uBAAuB;AACvC,SAAK,WAAW,aAAa,KAAK,IAAI;AACtC,SAAK,WAAW;AAChB,SAAK,WAAW,oBAAoB;AACpC,SAAK,QAAQ;AAAA,EACf;AAAA,EAEA,SAAS,YAA0B;AACjC,SAAK,WAAW,aAAa,KAAK,IAAI;AACtC,SAAK,WAAW,gBAAgB;AAChC,SAAK,QAAQ;AAAA,EACf;AAAA,EAEA,WAAiB;AACf,SAAK,WAAW,aAAa,KAAK,IAAI;AACtC,SAAK,WAAW;AAChB,SAAK,QAAQ;AAAA,EACf;AAAA,EAEA,gBAAoC;AAAE,WAAO,EAAE,GAAG,KAAK,WAAW;AAAA,EAAG;AAAA;AAAA,EAGrE,MAAM,QAAuB;AAC3B,QAAI,CAAC,KAAK,MAAO;AACjB,SAAK,WAAW,YAAY,KAAK,IAAI;AACrC,UAAM,MAAM,KAAK,QAAQ,KAAK,cAAc;AAC5C,UAAM,GAAG,MAAM,KAAK,EAAE,WAAW,KAAK,CAAC;AACvC,UAAM,UAAU,KAAK,iBAAiB;AACtC,UAAM,GAAG,UAAU,SAAS,KAAK,UAAU,KAAK,YAAY,MAAM,CAAC,GAAG,OAAO;AAC7E,UAAM,GAAG,OAAO,SAAS,KAAK,cAAc;AAC5C,SAAK,QAAQ;AAAA,EACf;AACF;","names":[]}

package/dist/chunk-2H7UOFLK.js

@@ -0,0 +1,11 @@
+#!/usr/bin/env node
+var __defProp = Object.defineProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+
+export {
+  __export
+};
+//# sourceMappingURL=chunk-2H7UOFLK.js.map

package/dist/chunk-2H7UOFLK.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/chunk-CHVFXR24.js

@@ -0,0 +1,19 @@
+#!/usr/bin/env node
+
+// src/token-counter.ts
+function estimateTokens(text) {
+  if (!text) return 0;
+  let count = 0;
+  for (const char of text) {
+    if (/[\u4e00-\u9fff\u3040-\u30ff\uac00-\ud7af]/.test(char)) count += 1;
+    else if (/[a-zA-Z0-9]/.test(char)) count += 0.25;
+    else if (/\s/.test(char)) count += 0.1;
+    else count += 0.5;
+  }
+  return Math.ceil(count);
+}
+
+export {
+  estimateTokens
+};
+//# sourceMappingURL=chunk-CHVFXR24.js.map

package/dist/chunk-CHVFXR24.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/token-counter.ts"],"sourcesContent":["/**\n * 轻量 token 估算器 — 无需 tiktoken 依赖\n *\n * 估算公式:\n *   中文/CJK ~1 token/字\n *   英文字母数字 ~0.25 token/字符 (≈0.75 token/单词)\n *   空格 ~0.1 token\n *   标点符号 ~0.5 token\n *\n * 用于 Context Offload 压缩引擎的 token 预算计算。\n * 纯字符级估算，无外部依赖，开销极低。\n */\n\n/**\n * 估算文本的 token 数\n *\n * @param text 待估算的文本\n * @returns 估算的 token 数（向上取整）\n */\nexport function estimateTokens(text: string): number {\n  if (!text) return 0;\n  let count = 0;\n  for (const char of text) {\n    // CJK 字符（中日韩）: ~1 token each\n    if (/[\\u4e00-\\u9fff\\u3040-\\u30ff\\uac00-\\ud7af]/.test(char)) count += 1;\n    // ASCII 字母数字: ~0.25 token (4 个字符约 1 token)\n    else if (/[a-zA-Z0-9]/.test(char)) count += 0.25;\n    // 空白字符: 0.1 token\n    else if (/\\s/.test(char)) count += 0.1;\n    // 标点及其他符号: 0.5 token\n    else count += 0.5;\n  }\n  return Math.ceil(count);\n}\n\n/**\n * 估算消息数组的总 token 数（含结构开销）\n *\n * @param messages 消息列表，每条含 role 和 content\n * @returns 估算的总 token 数\n */\nexport function estimateMessagesTokens(\n  messages: Array<{ role: string; content: string }>\n): number {\n  let total = 0;\n  for (const msg of messages) {\n    total += estimateTokens(msg.content);\n    total += 4; // role + structural overhead per message\n  }\n  return total + 2; // primer overhead\n}\n"],"mappings":";;;AAmBO,SAAS,eAAe,MAAsB;AACnD,MAAI,CAAC,KAAM,QAAO;AAClB,MAAI,QAAQ;AACZ,aAAW,QAAQ,MAAM;AAEvB,QAAI,4CAA4C,KAAK,IAAI,EAAG,UAAS;AAAA,aAE5D,cAAc,KAAK,IAAI,EAAG,UAAS;AAAA,aAEnC,KAAK,KAAK,IAAI,EAAG,UAAS;AAAA,QAE9B,UAAS;AAAA,EAChB;AACA,SAAO,KAAK,KAAK,KAAK;AACxB;","names":[]}