jinzd-ai-cli 0.4.89 → 0.4.91

package/README.md

@@ -7,7 +7,7 @@
 [![npm version](https://img.shields.io/npm/v/jinzd-ai-cli)](https://www.npmjs.com/package/jinzd-ai-cli)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 [![Node.js](https://img.shields.io/badge/node-%3E%3D20-brightgreen)](https://nodejs.org)
-[![Tests](https://img.shields.io/badge/tests-529%20passing-brightgreen)]()
+[![Tests](https://img.shields.io/badge/tests-647%20passing-brightgreen)]()
 [![GitHub Release](https://img.shields.io/github/v/release/jinzhengdong/ai-cli)](https://github.com/jinzhengdong/ai-cli/releases)
 [![CI](https://github.com/jinzhengdong/ai-cli/actions/workflows/ci.yml/badge.svg)](https://github.com/jinzhengdong/ai-cli/actions/workflows/ci.yml)
 
@@ -31,6 +31,10 @@
 - **Conversation Branching** *(v0.4.74+)* — `/branch list/new/switch/delete/rename` inside the REPL, plus a 🌿 "fork here" button on every replay step — explore alternate directions without losing the original thread
 - **Symbol Index** *(v0.4.76+)* — persistent tree-sitter index for TS/JS/TSX/Python powers three new AI tools: `find_symbol`, `get_outline`, `find_references`. Orders of magnitude faster than grep for definition lookups; background refresh on REPL startup, `/index status|rebuild|clear` to manage
 - **Semantic Code Search** *(v0.4.77+)* — `search_code` tool finds code by meaning, not name. Local sentence embeddings (multilingual MiniLM, 117 MB one-time download) score symbols by cosine similarity against natural-language queries in English or Chinese ("where are users authenticated", "哪里做了速率限制"). No API key, runs on CPU. Manage with `/index semantic-rebuild|semantic-clear`
+- **MCP Server Mode** *(v0.4.84+)* — `aicli mcp-serve` reverses ai-cli into an MCP server (JSON-RPC 2.0 over stdio), exposing its 26 built-in tools (incl. `find_symbol` / `search_code` / `run_tests`) to Claude Desktop / Cursor / any MCP client. Opt-in destructive-tool allow, `--tools` whitelist, `--cwd` override
+- **Session Sensitive-Data Redaction** *(v0.4.88+)* — unified redactor scrubs `password=` / `api_key` / bearer tokens / OpenAI-style keys from every message **before it hits disk**. Query text is redacted too, so secrets never reach embeddings or logs. `/security status` + `/security scan` to audit
+- **Human-like Long-Term Memory** *(v0.4.89+, B4)* — semantic index over every past chat session + `recall_memory` AI tool + `/memory rebuild|refresh|status|recall` commands. AI is prompted to auto-recall when it sees "last time" / "之前" / ambiguous references. Reuses the same MiniLM embedder as semantic code search
+- **Web UI Memory Panel** *(v0.4.90+, B4)* — new 🧠 Memory sidebar tab with semantic search across past chats; each hit has **➕ Inject** (quotes the snippet into the chat input as a markdown blockquote so you can review/edit before sending — no silent context injection) and **↗ Load** (jumps to source session). Bulk "Inject top 3" for recall bundles
 - **Streaming Tool Use** — Real-time streaming of AI reasoning and tool calls as they happen
 - **Sub-Agents** — Delegate complex subtasks to isolated child agents with independent tool loops
 - **Extended Thinking** — Claude deep reasoning mode with `/think` toggle
@@ -41,7 +45,7 @@
 - **PWA Support** — Install Web UI as a desktop/mobile app, accessible over LAN
 - **Hierarchical Context** — 3-layer context files (global / project / subdirectory) auto-injected
 - **Headless Mode** — `ai-cli -p "prompt"` for CI/CD pipelines and scripting
-- **40 REPL Commands** — Session management, checkpointing, code review, security review, rewind, scaffolding, cross-session history search, and more
+- **43 REPL Commands** — Session management, checkpointing, code review, security review/scan, rewind, scaffolding, cross-session history search, chat-memory recall, and more
 - **GitHub Actions CI/CD** — Automated testing on Node 20/22 + npm publish on release tags
 - **Cross-Platform** — Windows, macOS, Linux
 

package/README.zh-CN.md

@@ -7,7 +7,7 @@
 [![npm version](https://img.shields.io/npm/v/jinzd-ai-cli)](https://www.npmjs.com/package/jinzd-ai-cli)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 [![Node.js](https://img.shields.io/badge/node-%3E%3D20-brightgreen)](https://nodejs.org)
-[![Tests](https://img.shields.io/badge/tests-529%20passing-brightgreen)]()
+[![Tests](https://img.shields.io/badge/tests-647%20passing-brightgreen)]()
 [![GitHub Release](https://img.shields.io/github/v/release/jinzhengdong/ai-cli)](https://github.com/jinzhengdong/ai-cli/releases)
 [![CI](https://github.com/jinzhengdong/ai-cli/actions/workflows/ci.yml/badge.svg)](https://github.com/jinzhengdong/ai-cli/actions/workflows/ci.yml)
 
@@ -23,6 +23,10 @@
 - **对话分支**（v0.4.74+）— REPL 内 `/branch list/new/switch/delete/rename/diff/cherry-pick`，Web UI 回放面板每条消息旁 🌿 "fork here" 按钮，任意位置开辟新分支探索不同方向，原对话保持不变；v0.4.81 起所有子命令支持 id / title / 唯一前缀
 - **符号索引**（v0.4.76+，C1）— tree-sitter 持久化索引（TS/JS/TSX/JSX/Python）+ 3 个只读 AI 工具（`find_symbol` / `get_outline` / `find_references`）+ `/index status/rebuild/clear`；启动后台增量刷新，`write_file` 后自动 upsert
 - **语义代码搜索**（v0.4.77+，C2）— `search_code` AI 工具 + 本地 sentence embedding（`paraphrase-multilingual-MiniLM-L12-v2`，117 MB 一次性下载，384 维，CPU 运行）；支持中英文自然语言查询代码；`/index semantic-rebuild/semantic-clear` 管理
+- **MCP Server 模式**（v0.4.84+，E1）— `aicli mcp-serve` 把 aicli 反转为 MCP 服务器（JSON-RPC 2.0 over stdio），把 26 个内置工具（含 `find_symbol` / `search_code` / `run_tests`）输出给 Claude Desktop / Cursor / 任何 MCP 客户端；支持 `--tools` 白名单、`--cwd` 覆盖、`--allow-destructive`
+- **Session 敏感数据脱敏**（v0.4.88+）— 统一 redactor 在 session **落盘前**自动替换 `password=` / `api_key` / bearer token / OpenAI key 为 `[REDACTED:*]`；查询文本也走脱敏，secret 不会进 embedding 或日志；`/security status` + `/security scan` 审计
+- **类人长期记忆**（v0.4.89+，B4）— 聊天记录语义索引跨 session 可召回 + `recall_memory` AI 工具 + `/memory rebuild/refresh/status/recall/index-clear`；AI 看到"上次"/"之前"/指代不明时自动回忆。复用 C2 的 MiniLM embedder
+- **Web UI Memory 面板**（v0.4.90+，B4）— sidebar 新增 🧠 Memory 标签页，跨 session 语义搜索；每条 hit 带 **➕ Inject**（把片段作为 markdown 引用块塞进聊天输入框，用户可在上面继续打字编辑——不是静默注入上下文）和 **↗ Load**（跳转对应 session）；顶部"➕ Inject top 3"一键批量
 - **流式工具调用** — 实时流式展示 AI 推理过程和工具调用
 - **子代理系统** — 将复杂子任务委派给独立子代理执行
 - **深度推理** — Claude Extended Thinking，`/think` 一键切换
@@ -33,7 +37,7 @@
 - **PWA 支持** — Web UI 可安装为桌面/移动应用，支持局域网访问
 - **三层级上下文** — 全局 / 项目 / 子目录上下文文件自动注入
 - **无头模式** — `aicli -p "提示词"` 用于 CI/CD 管道和脚本
-- **40 个 REPL 命令** — 会话管理、检查点、代码审查、安全审查、对话回退、脚手架等
+- **43 个 REPL 命令** — 会话管理、检查点、代码审查、安全审查/扫描、对话回退、脚手架、聊天记忆召回等
 - **GitHub Actions CI/CD** — Node 20/22 自动测试 + Release tag 自动发布 npm
 - **跨平台** — Windows、macOS、Linux
 

package/dist/{batch-UMQYXVKG.js → batch-KZQHKPE5.js}

@@ -1,9 +1,9 @@
 #!/usr/bin/env node
 import {
   ConfigManager
-} from "./chunk-3O3U3L5W.js";
+} from "./chunk-US6MQO6W.js";
 import "./chunk-2ZD3YTVM.js";
-import "./chunk-E7YC4GWV.js";
+import "./chunk-AKCXRW2Q.js";
 
 // src/cli/batch.ts
 import Anthropic from "@anthropic-ai/sdk";

package/dist/chat-index-QKFH7ZP6.js

@@ -0,0 +1,17 @@
+import {
+  buildChatIndex,
+  chunkSession,
+  clearChatIndex,
+  getChatIndexStatus,
+  loadChatIndex,
+  searchChatMemory
+} from "./chunk-5S3PIG5O.js";
+import "./chunk-JV5N65KN.js";
+export {
+  buildChatIndex,
+  chunkSession,
+  clearChatIndex,
+  getChatIndexStatus,
+  loadChatIndex,
+  searchChatMemory
+};

package/dist/chunk-5S3PIG5O.js

@@ -0,0 +1,453 @@
+import {
+  EMBEDDING_DIM,
+  embed,
+  embedOne
+} from "./chunk-JV5N65KN.js";
+
+// src/memory/chat-index.ts
+import fs from "fs";
+import path from "path";
+import os from "os";
+import crypto from "crypto";
+
+// src/security/redactor.ts
+var DEFAULT_PATTERNS = [
+  // password: xxx / password = xxx / password="xxx"
+  // Covers YAML / JSON / shell-ish / env-file forms.
+  { kind: "password", regex: /\b(password|passwd|pwd)\s*[:=]\s*["']?([^\s"',;{}]{4,200})["']?/gi },
+  // PGPASSWORD=xxx (explicit bash env-var form, separate rule because no quotes usually)
+  { kind: "pgpassword-env", regex: /\b(PGPASSWORD)=([^\s"']{4,200})/g },
+  // JDBC/PG/MySQL/Mongo connection strings with inline credentials
+  // postgresql://user:pass@host/db → redact pass
+  { kind: "db-uri-password", regex: /(\b(?:postgres(?:ql)?|mysql|mongodb(?:\+srv)?|redis|amqp|mssql):\/\/[^:\s]+:)([^@\s]+)(@)/gi },
+  // Anthropic API keys
+  { kind: "anthropic-key", regex: /(sk-ant-[a-zA-Z0-9_-]{90,})/g },
+  // OpenAI / generic sk- keys — requires length ≥32 to avoid eating short identifiers
+  { kind: "openai-key", regex: /(sk-(?:proj-)?[a-zA-Z0-9_-]{32,})/g },
+  // GitHub personal access tokens
+  { kind: "github-pat", regex: /\b(ghp_[a-zA-Z0-9]{36})\b/g },
+  { kind: "github-oauth", regex: /\b(gho_[a-zA-Z0-9]{36})\b/g },
+  { kind: "github-install", regex: /\b(ghs_[a-zA-Z0-9]{36})\b/g },
+  // Slack tokens
+  { kind: "slack-bot", regex: /\b(xoxb-\d+-\d+-[a-zA-Z0-9]+)\b/g },
+  { kind: "slack-user", regex: /\b(xoxp-\d+-\d+-\d+-[a-zA-Z0-9]+)\b/g },
+  // AWS access key IDs (AKIA...) and secret access keys are context-dependent;
+  // we only catch the ID because secret key alone is indistinguishable from random base64.
+  { kind: "aws-access-key-id", regex: /\b(AKIA[0-9A-Z]{16})\b/g },
+  // Google API keys
+  { kind: "google-api-key", regex: /\b(AIza[0-9A-Za-z_-]{35})\b/g },
+  // Generic "api_key": "..." / "apiKey": "..." / api-key=xxx
+  { kind: "api-key", regex: /\b(api[_-]?key)\s*[:=]\s*["']?([a-zA-Z0-9_\-.]{16,200})["']?/gi },
+  // Generic token: xxx (only when value looks token-shaped; avoids eating human prose)
+  { kind: "token", regex: /\b(token|access[_-]?token|bearer[_-]?token)\s*[:=]\s*["']?([a-zA-Z0-9_\-.]{20,300})["']?/gi },
+  // Bearer <token> in Authorization headers
+  { kind: "bearer", regex: /\b(Authorization:\s*Bearer\s+)([a-zA-Z0-9_\-.=]{20,500})/g },
+  // Private key PEM blocks — catch the header+footer together
+  { kind: "private-key", regex: /-----BEGIN [A-Z ]*PRIVATE KEY-----[\s\S]*?-----END [A-Z ]*PRIVATE KEY-----/g }
+];
+function render(placeholder, kind) {
+  return placeholder.replace("{kind}", kind);
+}
+function redactString(input, options) {
+  if (!options.enabled || !input) return { redacted: input, hits: [] };
+  const placeholder = options.placeholder ?? "[REDACTED:{kind}]";
+  const patterns = [
+    ...options.patterns ?? DEFAULT_PATTERNS,
+    ...(options.customRegexes ?? []).flatMap((src, i) => {
+      try {
+        const flags = src.match(/^\/.*\/([gimsuy]*)$/)?.[1] ?? "";
+        const body = src.replace(/^\/(.*)\/[gimsuy]*$/, "$1");
+        const regex = new RegExp(body, flags.includes("g") ? flags : flags + "g");
+        return [{ kind: `custom-${i}`, regex }];
+      } catch {
+        return [];
+      }
+    })
+  ];
+  let redacted = input;
+  const hits = [];
+  for (const { kind, regex } of patterns) {
+    const rx = new RegExp(regex.source, regex.flags);
+    redacted = redacted.replace(rx, (...args) => {
+      const match = args[0];
+      const probe = new RegExp(rx.source).exec(match);
+      const captureCount = probe ? probe.length - 1 : 0;
+      const g1 = captureCount >= 1 ? args[1] : void 0;
+      const g2 = captureCount >= 2 ? args[2] : void 0;
+      const offset = args[1 + captureCount];
+      if (captureCount >= 2 && typeof g2 === "string") {
+        hits.push({ kind, start: offset + (g1?.length ?? 0), length: g2.length, secret: g2 });
+        return `${g1}${render(placeholder, kind)}`;
+      }
+      hits.push({ kind, start: offset, length: match.length, secret: g1 ?? match });
+      return render(placeholder, kind);
+    });
+  }
+  return { redacted, hits };
+}
+function redactJson(value, options) {
+  if (!options.enabled) return { value, hits: [] };
+  const allHits = [];
+  function walk(v) {
+    if (typeof v === "string") {
+      const r = redactString(v, options);
+      allHits.push(...r.hits);
+      return r.redacted;
+    }
+    if (Array.isArray(v)) return v.map(walk);
+    if (v && typeof v === "object") {
+      const out = {};
+      for (const [k, vv] of Object.entries(v)) {
+        out[k] = walk(vv);
+      }
+      return out;
+    }
+    return v;
+  }
+  const redacted = walk(value);
+  return { value: redacted, hits: allHits };
+}
+
+// src/memory/chat-index.ts
+var MEMORY_DIR_NAME = "memory-index";
+var CHUNKS_FILE = "chunks.json";
+var VECTORS_FILE = "vectors.vec";
+var VEC_MAGIC = 1094929750;
+var VEC_VERSION = 1;
+var VEC_HEADER_BYTES = 16;
+function memoryIndexDir() {
+  return path.join(os.homedir(), ".aicli", MEMORY_DIR_NAME);
+}
+function chunksPath() {
+  return path.join(memoryIndexDir(), CHUNKS_FILE);
+}
+function vectorsPath() {
+  return path.join(memoryIndexDir(), VECTORS_FILE);
+}
+function historyDir() {
+  return path.join(os.homedir(), ".aicli", "history");
+}
+var MAX_CHUNK_CHARS = 1200;
+var MIN_CHUNK_CHARS = 40;
+function extractMessageText(msg) {
+  if (typeof msg.content === "string") return msg.content;
+  if (Array.isArray(msg.content)) {
+    return msg.content.filter((p) => p && p.type === "text" && typeof p.text === "string").map((p) => p.text).join("\n");
+  }
+  return "";
+}
+function chunkSession(session) {
+  const chunks = [];
+  let pending = null;
+  const flush = () => {
+    if (!pending) return;
+    const rawText = pending.parts.join("\n").trim();
+    if (rawText.length < MIN_CHUNK_CHARS) {
+      pending = null;
+      return;
+    }
+    const { redacted } = redactString(rawText, { enabled: true });
+    const id = crypto.createHash("sha1").update(`${session.id}|${pending.start}|${pending.end}|${redacted.length}`).digest("hex").slice(0, 16);
+    chunks.push({
+      id,
+      sessionId: session.id,
+      sessionTitle: session.title,
+      provider: session.provider,
+      model: session.model,
+      startMessageIdx: pending.start,
+      endMessageIdx: pending.end,
+      text: redacted,
+      timestamp: pending.latestTs,
+      roles: pending.roles
+    });
+    pending = null;
+  };
+  for (let i = 0; i < session.messages.length; i++) {
+    const m = session.messages[i];
+    if (m.role !== "user" && m.role !== "assistant" && m.role !== "system") continue;
+    const text = extractMessageText(m).trim();
+    if (!text) continue;
+    const ts = m.timestamp ?? (/* @__PURE__ */ new Date()).toISOString();
+    const prefix = m.role === "user" ? "[USER] " : m.role === "assistant" ? "[AI] " : "[SYS] ";
+    const part = `${prefix}${text}`;
+    if (!pending) {
+      pending = { start: i, end: i, parts: [part], roles: [m.role], latestTs: ts };
+      continue;
+    }
+    const projected = pending.parts.reduce((n, p) => n + p.length + 1, 0) + part.length;
+    if (projected > MAX_CHUNK_CHARS) {
+      flush();
+      pending = { start: i, end: i, parts: [part], roles: [m.role], latestTs: ts };
+    } else {
+      pending.parts.push(part);
+      pending.end = i;
+      pending.roles.push(m.role);
+      pending.latestTs = ts;
+    }
+  }
+  flush();
+  return chunks;
+}
+function writeVectorsFile(chunks, vectors) {
+  if (chunks.length * EMBEDDING_DIM !== vectors.length) {
+    throw new Error(
+      `writeVectorsFile: length mismatch \u2014 ${chunks.length} chunks vs ${vectors.length / EMBEDDING_DIM} vectors`
+    );
+  }
+  const dir = memoryIndexDir();
+  fs.mkdirSync(dir, { recursive: true });
+  const totalBytes = VEC_HEADER_BYTES + vectors.byteLength;
+  const buf = Buffer.alloc(totalBytes);
+  buf.writeUInt32LE(VEC_MAGIC, 0);
+  buf.writeUInt32LE(VEC_VERSION, 4);
+  buf.writeUInt32LE(chunks.length, 8);
+  buf.writeUInt32LE(EMBEDDING_DIM, 12);
+  Buffer.from(vectors.buffer, vectors.byteOffset, vectors.byteLength).copy(buf, VEC_HEADER_BYTES);
+  const target = vectorsPath();
+  const tmp = `${target}.tmp`;
+  fs.writeFileSync(tmp, buf);
+  fs.renameSync(tmp, target);
+}
+function readVectorsFile(expectedCount) {
+  const p = vectorsPath();
+  if (!fs.existsSync(p)) return null;
+  let buf;
+  try {
+    buf = fs.readFileSync(p);
+  } catch {
+    return null;
+  }
+  if (buf.length < VEC_HEADER_BYTES) return null;
+  const magic = buf.readUInt32LE(0);
+  const version = buf.readUInt32LE(4);
+  const count = buf.readUInt32LE(8);
+  const dim = buf.readUInt32LE(12);
+  if (magic !== VEC_MAGIC || version !== VEC_VERSION || dim !== EMBEDDING_DIM) return null;
+  if (count !== expectedCount) return null;
+  const expected = VEC_HEADER_BYTES + count * dim * 4;
+  if (buf.length !== expected) return null;
+  return new Float32Array(
+    buf.buffer.slice(buf.byteOffset + VEC_HEADER_BYTES, buf.byteOffset + expected)
+  );
+}
+function writeIndexFile(idx) {
+  const dir = memoryIndexDir();
+  fs.mkdirSync(dir, { recursive: true });
+  const target = chunksPath();
+  const tmp = `${target}.tmp`;
+  fs.writeFileSync(tmp, JSON.stringify(idx, null, 2), "utf-8");
+  fs.renameSync(tmp, target);
+}
+function readIndexFile() {
+  const p = chunksPath();
+  if (!fs.existsSync(p)) return null;
+  try {
+    const raw = fs.readFileSync(p, "utf-8");
+    const data = JSON.parse(raw);
+    if (data.version !== 1) return null;
+    return data;
+  } catch {
+    return null;
+  }
+}
+function loadChatIndex() {
+  const idx = readIndexFile();
+  if (!idx) return null;
+  const vectors = readVectorsFile(idx.chunks.length);
+  if (!vectors) return null;
+  return { idx, vectors };
+}
+function clearChatIndex() {
+  try {
+    if (fs.existsSync(chunksPath())) fs.unlinkSync(chunksPath());
+  } catch {
+  }
+  try {
+    if (fs.existsSync(vectorsPath())) fs.unlinkSync(vectorsPath());
+  } catch {
+  }
+}
+function listSessionFiles() {
+  const dir = historyDir();
+  if (!fs.existsSync(dir)) return [];
+  const out = [];
+  for (const name of fs.readdirSync(dir)) {
+    if (!name.endsWith(".json")) continue;
+    const id = name.replace(/\.json$/, "");
+    const p = path.join(dir, name);
+    try {
+      const st = fs.statSync(p);
+      out.push({ id, path: p, mtime: st.mtimeMs });
+    } catch {
+    }
+  }
+  return out;
+}
+function readSession(p) {
+  try {
+    const data = JSON.parse(fs.readFileSync(p, "utf-8"));
+    if (!data.id || !Array.isArray(data.messages)) return null;
+    return data;
+  } catch {
+    return null;
+  }
+}
+async function buildChatIndex(options = {}) {
+  const t0 = Date.now();
+  const onProgress = options.onProgress ?? (() => {
+  });
+  onProgress({ stage: "scanning" });
+  const files = listSessionFiles();
+  const existing = options.full ? null : loadChatIndex();
+  const prevMtimes = existing?.idx.sessionMtimes ?? {};
+  const prevChunksBySession = /* @__PURE__ */ new Map();
+  const prevVectorsByChunkId = /* @__PURE__ */ new Map();
+  if (existing) {
+    for (let i = 0; i < existing.idx.chunks.length; i++) {
+      const c = existing.idx.chunks[i];
+      const arr = prevChunksBySession.get(c.sessionId) ?? [];
+      arr.push(c);
+      prevChunksBySession.set(c.sessionId, arr);
+      prevVectorsByChunkId.set(
+        c.id,
+        existing.vectors.slice(i * EMBEDDING_DIM, (i + 1) * EMBEDDING_DIM)
+      );
+    }
+  }
+  onProgress({ stage: "chunking" });
+  const stats = {
+    sessionsScanned: files.length,
+    sessionsIndexed: 0,
+    sessionsSkipped: 0,
+    chunksTotal: 0,
+    chunksAdded: 0,
+    chunksRemoved: 0,
+    durationMs: 0
+  };
+  const newMtimes = {};
+  const finalChunks = [];
+  const finalVectors = [];
+  const toEmbed = [];
+  for (const f of files) {
+    newMtimes[f.id] = f.mtime;
+    const prevMtime = prevMtimes[f.id];
+    if (prevMtime === f.mtime && prevChunksBySession.has(f.id)) {
+      stats.sessionsSkipped++;
+      const cached = prevChunksBySession.get(f.id);
+      for (const c of cached) {
+        const v = prevVectorsByChunkId.get(c.id);
+        if (!v) continue;
+        finalChunks.push(c);
+        finalVectors.push(v);
+      }
+      continue;
+    }
+    const sess = readSession(f.path);
+    if (!sess) continue;
+    stats.sessionsIndexed++;
+    const chunks = chunkSession(sess);
+    for (const c of chunks) {
+      finalChunks.push(c);
+      toEmbed.push(c);
+      stats.chunksAdded++;
+    }
+  }
+  if (existing) {
+    for (const prevId of Object.keys(prevMtimes)) {
+      if (!(prevId in newMtimes)) {
+        const removed = prevChunksBySession.get(prevId) ?? [];
+        stats.chunksRemoved += removed.length;
+      }
+    }
+  }
+  stats.chunksTotal = finalChunks.length;
+  const BATCH = 16;
+  onProgress({ stage: "embedding", processed: 0, total: toEmbed.length });
+  const newVectorsByChunkId = /* @__PURE__ */ new Map();
+  for (let i = 0; i < toEmbed.length; i += BATCH) {
+    const batch = toEmbed.slice(i, i + BATCH);
+    const vecs = await embed(batch.map((c) => c.text));
+    for (let j = 0; j < batch.length; j++) {
+      newVectorsByChunkId.set(batch[j].id, vecs[j]);
+    }
+    onProgress({ stage: "embedding", processed: Math.min(i + BATCH, toEmbed.length), total: toEmbed.length });
+  }
+  const flat = new Float32Array(finalChunks.length * EMBEDDING_DIM);
+  for (let i = 0; i < finalChunks.length; i++) {
+    const c = finalChunks[i];
+    const v = newVectorsByChunkId.get(c.id) ?? prevVectorsByChunkId.get(c.id);
+    if (!v || v.length !== EMBEDDING_DIM) {
+      continue;
+    }
+    flat.set(v, i * EMBEDDING_DIM);
+  }
+  onProgress({ stage: "saving" });
+  const idx = {
+    version: 1,
+    built: (/* @__PURE__ */ new Date()).toISOString(),
+    model: "Xenova/paraphrase-multilingual-MiniLM-L12-v2",
+    sessionMtimes: newMtimes,
+    chunks: finalChunks
+  };
+  writeIndexFile(idx);
+  writeVectorsFile(finalChunks, flat);
+  stats.durationMs = Date.now() - t0;
+  onProgress({ stage: "done" });
+  return stats;
+}
+async function searchChatMemory(query, options = {}) {
+  const topK = options.topK ?? 5;
+  const minScore = options.minScore ?? 0.25;
+  const loaded = loadChatIndex();
+  if (!loaded || loaded.idx.chunks.length === 0) return [];
+  const { idx, vectors } = loaded;
+  const { redacted } = redactString(query, { enabled: true });
+  const qvec = await embedOne(redacted);
+  const candidates = [];
+  for (let i = 0; i < idx.chunks.length; i++) {
+    const c = idx.chunks[i];
+    if (options.sessionId && c.sessionId !== options.sessionId) continue;
+    if (options.excludeSessionId && c.sessionId === options.excludeSessionId) continue;
+    let score = 0;
+    const base = i * EMBEDDING_DIM;
+    for (let d = 0; d < EMBEDDING_DIM; d++) {
+      score += vectors[base + d] * qvec[d];
+    }
+    if (score < minScore) continue;
+    candidates.push({ chunk: c, score });
+  }
+  candidates.sort((a, b) => b.score - a.score);
+  return candidates.slice(0, topK);
+}
+function getChatIndexStatus() {
+  const status = {
+    exists: false,
+    chunks: 0,
+    sessions: 0,
+    vecFileSizeBytes: 0,
+    chunksFileSizeBytes: 0
+  };
+  try {
+    if (fs.existsSync(vectorsPath())) status.vecFileSizeBytes = fs.statSync(vectorsPath()).size;
+    if (fs.existsSync(chunksPath())) status.chunksFileSizeBytes = fs.statSync(chunksPath()).size;
+  } catch {
+  }
+  const idx = readIndexFile();
+  if (!idx) return status;
+  status.exists = true;
+  status.chunks = idx.chunks.length;
+  status.sessions = Object.keys(idx.sessionMtimes).length;
+  status.built = idx.built;
+  status.model = idx.model;
+  return status;
+}
+
+export {
+  redactJson,
+  chunkSession,
+  loadChatIndex,
+  clearChatIndex,
+  buildChatIndex,
+  searchChatMemory,
+  getChatIndexStatus
+};

package/dist/{chunk-E7YC4GWV.js → chunk-AKCXRW2Q.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 
 // src/core/constants.ts
-var VERSION = "0.4.89";
+var VERSION = "0.4.91";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{chunk-TKYNTXKB.js → chunk-IGNYJUZU.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   TEST_TIMEOUT
-} from "./chunk-E7YC4GWV.js";
+} from "./chunk-AKCXRW2Q.js";
 
 // src/tools/builtin/run-tests.ts
 import { execSync } from "child_process";

package/dist/chunk-JV5N65KN.js

@@ -0,0 +1,50 @@
+// src/symbols/embedder.ts
+import path from "path";
+import os from "os";
+import fs from "fs";
+var EMBEDDING_MODEL_ID = "Xenova/paraphrase-multilingual-MiniLM-L12-v2";
+var EMBEDDING_DIM = 384;
+var pipelinePromise = null;
+function cacheDir() {
+  return path.join(os.homedir(), ".aicli", "models");
+}
+async function getEmbedder() {
+  if (pipelinePromise) return pipelinePromise;
+  pipelinePromise = (async () => {
+    const mod = await import("@huggingface/transformers");
+    const dir = cacheDir();
+    fs.mkdirSync(dir, { recursive: true });
+    mod.env.cacheDir = dir;
+    mod.env.allowRemoteModels = true;
+    mod.env.allowLocalModels = true;
+    const pipe = await mod.pipeline("feature-extraction", EMBEDDING_MODEL_ID, {
+      // Keep the ONNX session in float32; int8 quantization exists but the
+      // quality drop on short code identifiers is noticeable.
+      dtype: "fp32"
+    });
+    return pipe;
+  })();
+  return pipelinePromise;
+}
+async function embed(texts) {
+  if (texts.length === 0) return [];
+  const pipe = await getEmbedder();
+  const out = await pipe(texts, { pooling: "mean", normalize: true });
+  const batch = texts.length;
+  const dim = EMBEDDING_DIM;
+  const rows = new Array(batch);
+  for (let i = 0; i < batch; i++) {
+    rows[i] = new Float32Array(out.data.buffer, out.data.byteOffset + i * dim * 4, dim).slice();
+  }
+  return rows;
+}
+async function embedOne(text) {
+  const [vec] = await embed([text]);
+  return vec;
+}
+
+export {
+  EMBEDDING_DIM,
+  embed,
+  embedOne
+};

package/dist/{chunk-GTKJUEBS.js → chunk-MO7MWNWC.js}

@@ -2,13 +2,15 @@ import {
   loadIndex
 } from "./chunk-BJAT4GNC.js";
 import {
-  EMBEDDING_DIM,
-  embed,
-  embedOne,
   loadVectorStore,
   saveVectorStore,
   searchVectorStore
-} from "./chunk-XMA222FQ.js";
+} from "./chunk-PASCDYMH.js";
+import {
+  EMBEDDING_DIM,
+  embed,
+  embedOne
+} from "./chunk-JV5N65KN.js";
 
 // src/symbols/semantic.ts
 function pathTokens(absFile, root) {

package/dist/{chunk-FVRLRIKC.js → chunk-MYQANQ6F.js}

@@ -2,7 +2,7 @@
 import {
   schemaToJsonSchema,
   truncateForPersist
-} from "./chunk-ABPT6XCI.js";
+} from "./chunk-Z2V6EYPQ.js";
 import {
   AuthError,
   ProviderError,
@@ -21,7 +21,7 @@ import {
   MCP_PROTOCOL_VERSION,
   MCP_TOOL_PREFIX,
   VERSION
-} from "./chunk-E7YC4GWV.js";
+} from "./chunk-AKCXRW2Q.js";
 
 // src/providers/claude.ts
 import Anthropic from "@anthropic-ai/sdk";