thincoder 0.7.8 → 0.8.0

package/src/memory/docs.mjs

@@ -0,0 +1,262 @@
+/**
+ * memory/docs.mjs — 文档索引同步、检索、agent 工具生成
+ */
+
+import { readFile, readdir, stat } from "node:fs/promises"
+import { join } from "node:path"
+import { embed, cosine, toBlob, fromBlob } from "../embedding.mjs"
+import { commitAndPush } from "../git/gitmem.mjs"
+import { DOC_EXTS, SKIP_DIRS } from "./schema.mjs"
+import { buildFtsQuery, put, search, putMarkdown } from "./core.mjs"
+import { _upsertDocFile, yieldTick } from "./code-index.mjs"
+import { markIndexedCommit } from "./code-sync.mjs"
+
+/**
+ * 同步文档索引：扫描 dir 下所有 .md/.mdc/.txt/.rst/.adoc → 分块 → upsert 到 doc_chunks。
+ * 按 mtime 增量。
+ */
+export async function docSync(memory, dir, { onProgress } = {}) {
+  const files = []
+  async function walk(d) {
+    let entries
+    try { entries = await readdir(d, { withFileTypes: true }) } catch { return }
+    for (const e of entries) {
+      if (e.isDirectory()) {
+        if (SKIP_DIRS.has(e.name) || e.name.startsWith(".")) continue
+        await walk(join(d, e.name))
+      } else if (e.isFile()) {
+        const ext = e.name.slice(e.name.lastIndexOf(".")).toLowerCase()
+        if (DOC_EXTS.has(ext)) files.push(join(d, e.name))
+      }
+    }
+  }
+  await walk(dir)
+
+  const indexed = new Map(
+    memory.db.prepare(`SELECT path, mtime_ms FROM doc_chunks WHERE origin = ?`).all(dir).map((r) => [r.path, r.mtime_ms])
+  )
+  const seen = new Set()
+
+  onProgress?.({ phase: "scan", total: files.length })
+
+  let updated = 0, removed = 0, skipped = 0, failed = 0
+  const errors = []
+  for (let i = 0; i < files.length; i++) {
+    const abs = files[i]
+    const rel = abs.slice(dir.length + 1).replaceAll("\\", "/")
+    seen.add(rel)
+
+    let mtimeMs
+    try { mtimeMs = Math.floor((await stat(abs)).mtimeMs) } catch { continue }
+    if (indexed.get(rel) === mtimeMs) {
+      skipped++
+      continue
+    }
+
+    try {
+      const text = await readFile(abs, "utf8")
+      const lines = text.split("\n")
+      _upsertDocFile(memory, dir, rel, lines, mtimeMs)
+      updated++
+    } catch (e) {
+      failed++
+      if (errors.length < 5) errors.push(`${rel}: ${e.message}`)
+    }
+    await yieldTick()
+
+    if (onProgress && i % 10 === 0) {
+      onProgress({ phase: "index", current: i + 1, total: files.length, updated, removed, skipped, failed })
+    }
+  }
+
+  for (const stale of indexed.keys()) {
+    if (!seen.has(stale)) {
+      memory.db.prepare(`DELETE FROM doc_chunks WHERE origin = ? AND path = ?`).run(dir, stale)
+      removed++
+    }
+  }
+
+  onProgress?.({ phase: "done", total: files.length, updated, removed, skipped, failed })
+  markIndexedCommit(memory, dir)
+  return { updated, removed, skipped, failed, errors, total: files.length }
+}
+
+/**
+ * 文档检索：FTS5(BM25) + 可选向量余弦，RRF 合并。
+ * 无 embedder 时退化为纯 FTS；ftsQuery 为空且有 embedder 时退化为纯向量。
+ */
+export async function docSearch(memory, query, { limit = 5 } = {}) {
+  const ftsQuery = buildFtsQuery(query)
+  if (!ftsQuery && !memory.embedder) return []
+
+  const ftsOriginFilter = memory.codeOrigin ? `AND d.origin = ?` : ""
+  const vecOriginFilter = memory.codeOrigin ? `AND origin = ?` : ""
+  const originParams = memory.codeOrigin ? [memory.codeOrigin] : []
+
+  const ftsList = ftsQuery ? memory.db.prepare(`
+    SELECT d.rowid, d.path, d.language, d.heading, d.content, d.line_start, d.line_end, bm25(doc_chunks_fts) AS rank
+    FROM doc_chunks_fts JOIN doc_chunks d ON d.rowid = doc_chunks_fts.rowid
+    WHERE doc_chunks_fts MATCH ? ${ftsOriginFilter}
+    ORDER BY rank LIMIT ?
+  `).all(ftsQuery, ...originParams, Math.max(limit * 4, 20)) : []
+
+  if (!memory.embedder) return ftsList.slice(0, limit)
+
+  try { await ensureDocEmbeddings(memory) } catch (e) {
+    console.error(`[docs] embedding ensure failed, falling back to FTS-only: ${e.message}`)
+    return ftsList.slice(0, limit)
+  }
+  let qvec
+  try { [qvec] = await embed(memory.embedder, [query]) } catch (e) {
+    console.error(`[docs] query embedding failed, falling back to FTS-only: ${e.message}`)
+    return ftsList.slice(0, limit)
+  }
+  const rows = memory.db.prepare(`SELECT rowid, embedding FROM doc_chunks WHERE embedding IS NOT NULL ${vecOriginFilter}`).all(...originParams)
+  const vecList = rows
+    .map((r) => ({ rowid: r.rowid, score: cosine(qvec, fromBlob(r.embedding)) }))
+    .sort((a, b) => b.score - a.score)
+    .slice(0, Math.max(limit * 4, 20))
+
+  const K = 60
+  const scores = new Map()
+  ftsList.forEach((r, i) => scores.set(r.rowid, (scores.get(r.rowid) ?? 0) + 1 / (K + i + 1)))
+  vecList.forEach((r, i) => scores.set(r.rowid, (scores.get(r.rowid) ?? 0) + 1 / (K + i + 1)))
+
+  const fetchChunk = memory.db.prepare(`
+    SELECT path, language, heading, content, line_start, line_end FROM doc_chunks WHERE rowid = ?
+  `)
+  return [...scores.entries()]
+    .sort((a, b) => b[1] - a[1])
+    .slice(0, limit)
+    .map(([rowid]) => fetchChunk.get(rowid))
+    .filter(Boolean)
+}
+
+/** 惰性补算 doc_chunks 缺失的向量 */
+export async function ensureDocEmbeddings(memory) {
+  if (!memory.embedder) return
+  const modelKey = memory.embedder.model
+  const stored = memory.db.prepare(`SELECT value FROM meta WHERE key = 'doc_embedding_model'`).get()?.value
+  if (stored !== modelKey) {
+    memory.db.prepare(`UPDATE doc_chunks SET embedding = NULL`).run()
+    memory.db.prepare(`INSERT INTO meta (key, value) VALUES ('doc_embedding_model', ?)
+      ON CONFLICT (key) DO UPDATE SET value = excluded.value`).run(modelKey)
+  }
+
+  const pending = memory.db.prepare(`SELECT rowid, path, heading, content FROM doc_chunks WHERE embedding IS NULL LIMIT 64`).all()
+  if (pending.length === 0) return
+
+  const texts = pending.map((r) => `${r.heading || r.path}\n${r.content.slice(0, 2000)}`)
+  const vecs = await embed(memory.embedder, texts)
+
+  const update = memory.db.prepare(`UPDATE doc_chunks SET embedding = ? WHERE rowid = ?`)
+  pending.forEach((r, i) => update.run(toBlob(vecs[i]), r.rowid))
+}
+
+/** 生成 doc_search 工具（只读）。 */
+export function docSearchTool(memory) {
+  return {
+    name: "doc_search",
+    description:
+      "Search the project's documentation (README, design docs, guides, markdown files) for relevant information. Use this to find design decisions, coding conventions, architecture docs, or project rules. Prefer this over code_search when you need to understand the project's intended design rather than existing implementation.",
+    parameters: {
+      type: "object",
+      properties: {
+        query: { type: "string", description: "Natural language search query" },
+        limit: { type: "number", description: "Max results (default 5)" },
+      },
+      required: ["query"],
+    },
+    readonly: true,
+    async execute(args) {
+      const results = await docSearch(memory, args.query, { limit: args.limit ?? 5 })
+      if (results.length === 0) return "(no matching documentation)"
+      return results.map((r) =>
+        `${r.path}${r.heading ? ` > ${r.heading}` : ""} (L${r.line_start}-L${r.line_end}):\n${r.content.slice(0, 2000)}`
+      ).join("\n\n---\n\n")
+    },
+  }
+}
+
+// ---------------------------------------------------------------- agent 工具
+
+/**
+ * 生成记忆相关的两个 agent 工具（遵循 tools.mjs 的工具形状）。
+ * memory_put 是有副作用工具（需权限确认），memory_search 只读。
+ * opts: { cwd, projectDir, author, team: { dir, name } | null }
+ */
+export function memoryTools(memory, opts = {}) {
+  const projectDir = opts.projectDir ? join(opts.cwd ?? process.cwd(), opts.projectDir) : null
+  return [
+    {
+      name: "memory_put",
+      description:
+        "Save a piece of knowledge to long-term memory. Use when you learn something worth remembering across sessions: a project convention, a debugging insight, an architecture decision. Types: rule (coding standards), knowledge (project facts), decision (architecture decisions), pattern (debugging/workflow patterns). Scopes: personal (default, private to you), project (shared via this repo's .thincoder/memory/), team (org-wide team repo, if configured).",
+      parameters: {
+        type: "object",
+        properties: {
+          type: { type: "string", enum: ["rule", "knowledge", "decision", "pattern"] },
+          title: { type: "string", description: "Short title" },
+          content: { type: "string", description: "Full content to remember" },
+          tags: { type: "string", description: "Space-separated tags" },
+          scope: { type: "string", enum: ["personal", "project", "team"], description: "Where to save (default personal)" },
+        },
+        required: ["type", "title", "content"],
+      },
+      readonly: false,
+      async execute(args) {
+        const scope = args.scope ?? "personal"
+        if (scope === "personal") {
+          const id = await put(memory, args)
+          return `Saved to personal memory (id=${id}): [${args.type}] ${args.title}`
+        }
+        if (scope === "project") {
+          if (!projectDir) throw new Error("project scope unavailable: no project directory configured")
+          const filename = await putMarkdown(memory, {
+            layer: "project",
+            dir: projectDir,
+            type: args.type,
+            title: args.title,
+            content: args.content,
+            tags: (args.tags ?? "").split(/\s+/).filter(Boolean),
+            author: opts.author ?? "unknown",
+          })
+          return `Saved to project memory (${filename}): [${args.type}] ${args.title}\nNote: file written to the repo; commit it yourself when ready.`
+        }
+        if (!opts.team?.dir) {
+          throw new Error("team scope not configured: set memory.team in ~/.thincoder/config.json")
+        }
+        const filename = await putMarkdown(memory, {
+          layer: "team",
+          dir: opts.team.dir,
+          type: args.type,
+          title: args.title,
+          content: args.content,
+          tags: (args.tags ?? "").split(/\s+/).filter(Boolean),
+          author: opts.author ?? "unknown",
+        })
+        await commitAndPush(opts.team.dir, filename, `memory: [${args.type}] ${args.title}`)
+        return `Saved to team memory and pushed (${filename}): [${args.type}] ${args.title}`
+      },
+    },
+    {
+      name: "memory_search",
+      description:
+        "Search long-term memory across all layers (personal/project/team) for relevant knowledge saved in previous sessions. Use the same language as the memories being searched.",
+      parameters: {
+        type: "object",
+        properties: {
+          query: { type: "string", description: "Natural language search query" },
+          limit: { type: "number", description: "Max results (default 5)" },
+        },
+        required: ["query"],
+      },
+      readonly: true,
+      async execute(args) {
+        const results = await search(memory, args.query, { limit: args.limit ?? 5 })
+        if (results.length === 0) return "(no matching memories)"
+        return results.map((r) => `[${r.layer}][${r.type}] ${r.title}\n${r.content}`).join("\n\n")
+      },
+    },
+  ]
+}

package/src/memory/schema.mjs

@@ -0,0 +1,426 @@
+/**
+ * memory/schema.mjs — 数据库 schema 定义、迁移、CJK 分词
+ * v1：node:sqlite + FTS5 单机实现，零依赖。
+ *
+ * 中文检索方案：FTS5 unicode61 分词 + CJK 逐字加空格（写入和查询两侧同样处理）。
+ * 效果：中文按字索引，"分号" 这类双字词也能命中；ASCII 仍按整词。
+ */
+
+import { DatabaseSync } from "node:sqlite"
+import { mkdirSync } from "node:fs"
+import { dirname } from "node:path"
+
+export const VALID_TYPES = new Set(["rule", "knowledge", "decision", "pattern"])
+export const SCHEMA_VERSION = 9
+
+// 代码索引：源码文件扩展名
+export const CODE_EXTS = new Set([".mjs", ".js", ".ts", ".tsx", ".jsx", ".py", ".rs", ".go", ".java", ".c", ".h", ".cpp", ".hpp", ".rb", ".swift", ".kt", ".sh", ".bash", ".sql", ".yaml", ".yml", ".toml", ".json", ".css", ".html", ".vue", ".svelte"])
+// 文档索引：markdown / 纯文本（分开索引，便于 LLM 区分"设计规范"和"现存代码"）
+export const DOC_EXTS = new Set([".md", ".mdc", ".txt", ".rst", ".adoc"])
+// 总是跳过的目录名
+export const SKIP_DIRS = new Set(["node_modules", ".git", "dist", "build", ".turbo", "coverage", "__pycache__", ".venv", "venv", "target", ".next", ".nuxt", ".svelte-kit"])
+// 大文件阈值（行数）：超过此行数按符号分块，否则整文件入索引
+export const BIG_FILE_LINES = 2000
+
+/**
+ * CJK 逐字加空格：让 unicode61 把每个汉字/日韩字当独立 token。
+ * 写入和查询必须使用同一处理，检索才能对上。
+ */
+export function segmentCJK(text) {
+  return text.replace(
+    /[぀-ヿ㐀-䶿一-鿿豈-﫿가-힯]+/g,
+    (run) => [...run].join(" "),
+  )
+}
+
+/**
+ * 打开/初始化记忆库。dbPath 不存在会自动创建。
+ * 返回的 memory 对象即接口，后续函数的第一个参数都是它。
+ */
+export function createMemory({ dbPath }) {
+  mkdirSync(dirname(dbPath), { recursive: true })
+  const db = new DatabaseSync(dbPath)
+  // WAL 读写不互锁（TUI 检索和后台索引可并发）；busy_timeout 防多进程同库直接 SQLITE_BUSY
+  db.exec(`PRAGMA journal_mode = WAL`)
+  db.exec(`PRAGMA busy_timeout = 3000`)
+
+  db.exec(`
+    CREATE TABLE IF NOT EXISTS entries (
+      id INTEGER PRIMARY KEY AUTOINCREMENT,
+      type TEXT NOT NULL CHECK(type IN ('rule','knowledge','decision','pattern')),
+      title TEXT NOT NULL,
+      content TEXT NOT NULL,
+      tags TEXT NOT NULL DEFAULT '',
+      seg_title TEXT NOT NULL DEFAULT '',
+      seg_content TEXT NOT NULL DEFAULT '',
+      seg_tags TEXT NOT NULL DEFAULT '',
+      created_at INTEGER NOT NULL,
+      updated_at INTEGER NOT NULL
+    )
+  `)
+
+  migrate(db)
+  return { db }
+}
+
+/** 按 user_version 逐步迁移。整体单事务——任何一步失败都回滚，不留半成品 schema */
+export function migrate(db) {
+  const { user_version: version } = db.prepare(`PRAGMA user_version`).get()
+  if (version >= SCHEMA_VERSION) return
+
+  db.exec("BEGIN IMMEDIATE")
+  try {
+  if (version < 2) {
+    // v1(trigram) 或空库 → v2(unicode61 + CJK 逐字)：重建 FTS 和触发器
+    db.exec(`
+      DROP TRIGGER IF EXISTS entries_ai;
+      DROP TRIGGER IF EXISTS entries_ad;
+      DROP TRIGGER IF EXISTS entries_au;
+      DROP TABLE IF EXISTS entries_fts;
+    `)
+    // 老库（v1）没有 seg 列，补上
+    const columns = db.prepare(`PRAGMA table_info(entries)`).all().map((c) => c.name)
+    for (const col of ["seg_title", "seg_content", "seg_tags"]) {
+      if (!columns.includes(col)) {
+        db.exec(`ALTER TABLE entries ADD COLUMN ${col} TEXT NOT NULL DEFAULT ''`)
+      }
+    }
+    // 回填 seg 列（JS 侧分字，SQL 做不了）
+    const rows = db.prepare(`SELECT id, title, content, tags FROM entries`).all()
+    const update = db.prepare(`UPDATE entries SET seg_title = ?, seg_content = ?, seg_tags = ? WHERE id = ?`)
+    for (const r of rows) {
+      update.run(segmentCJK(r.title), segmentCJK(r.content), segmentCJK(r.tags), r.id)
+    }
+
+    db.exec(`
+      CREATE VIRTUAL TABLE entries_fts USING fts5(
+        seg_title, seg_content, seg_tags,
+        content='entries', content_rowid='id',
+        tokenize='unicode61'
+      )
+    `)
+    db.exec(`
+      CREATE TRIGGER entries_ai AFTER INSERT ON entries BEGIN
+        INSERT INTO entries_fts(rowid, seg_title, seg_content, seg_tags)
+        VALUES (new.id, new.seg_title, new.seg_content, new.seg_tags);
+      END;
+      CREATE TRIGGER entries_ad AFTER DELETE ON entries BEGIN
+        INSERT INTO entries_fts(entries_fts, rowid, seg_title, seg_content, seg_tags)
+        VALUES ('delete', old.id, old.seg_title, old.seg_content, old.seg_tags);
+      END;
+      CREATE TRIGGER entries_au AFTER UPDATE ON entries BEGIN
+        INSERT INTO entries_fts(entries_fts, rowid, seg_title, seg_content, seg_tags)
+        VALUES ('delete', old.id, old.seg_title, old.seg_content, old.seg_tags);
+        INSERT INTO entries_fts(rowid, seg_title, seg_content, seg_tags)
+        VALUES (new.id, new.seg_title, new.seg_content, new.seg_tags);
+      END;
+    `)
+    db.exec(`INSERT INTO entries_fts(entries_fts) VALUES('rebuild')`)
+    db.exec(`PRAGMA user_version = 2`)
+  }
+
+  if (version < 3) {
+    // v3：markdown 层（project/team）的 files 表 + FTS + 触发器
+    db.exec(`
+      CREATE TABLE IF NOT EXISTS files (
+        layer TEXT NOT NULL CHECK(layer IN ('project','team')),
+        path TEXT NOT NULL,
+        type TEXT NOT NULL CHECK(type IN ('rule','knowledge','decision','pattern')),
+        title TEXT NOT NULL,
+        content TEXT NOT NULL,
+        tags TEXT NOT NULL DEFAULT '',
+        author TEXT NOT NULL DEFAULT '',
+        embedding BLOB,
+        mtime_ms INTEGER NOT NULL DEFAULT 0,
+        seg_title TEXT NOT NULL DEFAULT '',
+        seg_content TEXT NOT NULL DEFAULT '',
+        seg_tags TEXT NOT NULL DEFAULT '',
+        updated_at INTEGER NOT NULL,
+        PRIMARY KEY (layer, path)
+      )
+    `)
+    db.exec(`
+      CREATE VIRTUAL TABLE IF NOT EXISTS files_fts USING fts5(
+        seg_title, seg_content, seg_tags,
+        content='files', content_rowid='rowid',
+        tokenize='unicode61'
+      )
+    `)
+    db.exec(`
+      CREATE TRIGGER files_ai AFTER INSERT ON files BEGIN
+        INSERT INTO files_fts(rowid, seg_title, seg_content, seg_tags)
+        VALUES (new.rowid, new.seg_title, new.seg_content, new.seg_tags);
+      END;
+      CREATE TRIGGER files_ad AFTER DELETE ON files BEGIN
+        INSERT INTO files_fts(files_fts, rowid, seg_title, seg_content, seg_tags)
+        VALUES ('delete', old.rowid, old.seg_title, old.seg_content, old.seg_tags);
+      END;
+      CREATE TRIGGER files_au AFTER UPDATE ON files BEGIN
+        INSERT INTO files_fts(files_fts, rowid, seg_title, seg_content, seg_tags)
+        VALUES ('delete', old.rowid, old.seg_title, old.seg_content, old.seg_tags);
+        INSERT INTO files_fts(rowid, seg_title, seg_content, seg_tags)
+        VALUES (new.rowid, new.seg_title, new.seg_content, new.seg_tags);
+      END;
+    `)
+    db.exec(`PRAGMA user_version = 3`)
+  }
+
+  if (version < 4) {
+    // v4：personal 的 entries 表加向量列（files 表在 v3 已带）；meta 表存 embedding 模型名
+    const columns = db.prepare(`PRAGMA table_info(entries)`).all().map((c) => c.name)
+    if (!columns.includes("embedding")) {
+      db.exec(`ALTER TABLE entries ADD COLUMN embedding BLOB`)
+    }
+    db.exec(`CREATE TABLE IF NOT EXISTS meta (key TEXT PRIMARY KEY, value TEXT)`)
+    db.exec(`PRAGMA user_version = 4`)
+  }
+
+  if (version < 5) {
+    // v5：files 表加 origin 列（项目绝对路径），防止跨项目记忆串台
+    const columns = db.prepare(`PRAGMA table_info(files)`).all().map((c) => c.name)
+    if (!columns.includes("origin")) {
+      db.exec(`ALTER TABLE files ADD COLUMN origin TEXT NOT NULL DEFAULT ''`)
+    }
+    db.exec(`PRAGMA user_version = 5`)
+  }
+
+  if (version < 6) {
+    // v6：代码索引——code_chunks 表 + FTS5（与 files 表同样模式）
+    db.exec(`
+      CREATE TABLE IF NOT EXISTS code_chunks (
+        path TEXT NOT NULL,
+        language TEXT NOT NULL,
+        chunk_type TEXT NOT NULL CHECK(chunk_type IN ('file','symbol')),
+        symbol_name TEXT NOT NULL DEFAULT '',
+        content TEXT NOT NULL,
+        line_start INTEGER NOT NULL DEFAULT 0,
+        line_end INTEGER NOT NULL DEFAULT 0,
+        mtime_ms INTEGER NOT NULL DEFAULT 0,
+        embedding BLOB,
+        seg_content TEXT NOT NULL DEFAULT '',
+        PRIMARY KEY (path, line_start)
+      )
+    `)
+    db.exec(`
+      CREATE VIRTUAL TABLE IF NOT EXISTS code_chunks_fts USING fts5(
+        path, symbol_name, seg_content,
+        content='code_chunks', content_rowid='rowid',
+        tokenize='unicode61'
+      )
+    `)
+    db.exec(`
+      CREATE TRIGGER code_chunks_ai AFTER INSERT ON code_chunks BEGIN
+        INSERT INTO code_chunks_fts(rowid, path, symbol_name, seg_content)
+        VALUES (new.rowid, new.path, new.symbol_name, new.seg_content);
+      END;
+      CREATE TRIGGER code_chunks_ad AFTER DELETE ON code_chunks BEGIN
+        INSERT INTO code_chunks_fts(code_chunks_fts, rowid, path, symbol_name, seg_content)
+        VALUES ('delete', old.rowid, old.path, old.symbol_name, old.seg_content);
+      END;
+      CREATE TRIGGER code_chunks_au AFTER UPDATE ON code_chunks BEGIN
+        INSERT INTO code_chunks_fts(code_chunks_fts, rowid, path, symbol_name, seg_content)
+        VALUES ('delete', old.rowid, old.path, old.symbol_name, old.seg_content);
+        INSERT INTO code_chunks_fts(rowid, path, symbol_name, seg_content)
+        VALUES (new.rowid, new.path, new.symbol_name, new.seg_content);
+      END;
+    `)
+    db.exec(`PRAGMA user_version = 6`)
+  }
+
+  if (version < 7) {
+    // v7：文档索引——doc_chunks 表 + FTS5（与 code_chunks 同模式），markdown 按 ## 标题分块
+    db.exec(`
+      CREATE TABLE IF NOT EXISTS doc_chunks (
+        path TEXT NOT NULL,
+        language TEXT NOT NULL DEFAULT 'markdown',
+        heading TEXT NOT NULL DEFAULT '',
+        content TEXT NOT NULL,
+        line_start INTEGER NOT NULL DEFAULT 0,
+        line_end INTEGER NOT NULL DEFAULT 0,
+        mtime_ms INTEGER NOT NULL DEFAULT 0,
+        embedding BLOB,
+        seg_content TEXT NOT NULL DEFAULT '',
+        PRIMARY KEY (path, line_start)
+      )
+    `)
+    db.exec(`
+      CREATE VIRTUAL TABLE IF NOT EXISTS doc_chunks_fts USING fts5(
+        path, heading, seg_content,
+        content='doc_chunks', content_rowid='rowid',
+        tokenize='unicode61'
+      )
+    `)
+    db.exec(`
+      CREATE TRIGGER doc_chunks_ai AFTER INSERT ON doc_chunks BEGIN
+        INSERT INTO doc_chunks_fts(rowid, path, heading, seg_content)
+        VALUES (new.rowid, new.path, new.heading, new.seg_content);
+      END;
+      CREATE TRIGGER doc_chunks_ad AFTER DELETE ON doc_chunks BEGIN
+        INSERT INTO doc_chunks_fts(doc_chunks_fts, rowid, path, heading, seg_content)
+        VALUES ('delete', old.rowid, old.path, old.heading, old.seg_content);
+      END;
+      CREATE TRIGGER doc_chunks_au AFTER UPDATE ON doc_chunks BEGIN
+        INSERT INTO doc_chunks_fts(doc_chunks_fts, rowid, path, heading, seg_content)
+        VALUES ('delete', old.rowid, old.path, old.heading, old.seg_content);
+        INSERT INTO doc_chunks_fts(rowid, path, heading, seg_content)
+        VALUES (new.rowid, new.path, new.heading, new.seg_content);
+      END;
+    `)
+    db.exec(`PRAGMA user_version = 7`)
+  }
+
+  if (version < 8) {
+    // v8：code_chunks/doc_chunks 加 origin 列（项目根目录绝对路径），主键改为 (origin, path, line_start)。
+    // SQLite 不能 ALTER 主键 → 直接删表重建（下次 codeSync/docSync 自动重索引）。
+    db.exec(`
+      DROP TRIGGER IF EXISTS code_chunks_ai;
+      DROP TRIGGER IF EXISTS code_chunks_ad;
+      DROP TRIGGER IF EXISTS code_chunks_au;
+      DROP TABLE IF EXISTS code_chunks_fts;
+      DROP TABLE IF EXISTS code_chunks;
+      DROP TRIGGER IF EXISTS doc_chunks_ai;
+      DROP TRIGGER IF EXISTS doc_chunks_ad;
+      DROP TRIGGER IF EXISTS doc_chunks_au;
+      DROP TABLE IF EXISTS doc_chunks_fts;
+      DROP TABLE IF EXISTS doc_chunks;
+    `)
+    db.exec(`
+      CREATE TABLE code_chunks (
+        origin TEXT NOT NULL DEFAULT '',
+        path TEXT NOT NULL,
+        language TEXT NOT NULL,
+        chunk_type TEXT NOT NULL CHECK(chunk_type IN ('file','symbol')),
+        symbol_name TEXT NOT NULL DEFAULT '',
+        content TEXT NOT NULL,
+        line_start INTEGER NOT NULL DEFAULT 0,
+        line_end INTEGER NOT NULL DEFAULT 0,
+        mtime_ms INTEGER NOT NULL DEFAULT 0,
+        embedding BLOB,
+        seg_content TEXT NOT NULL DEFAULT '',
+        PRIMARY KEY (origin, path, line_start)
+      )
+    `)
+    db.exec(`
+      CREATE VIRTUAL TABLE code_chunks_fts USING fts5(
+        path, symbol_name, seg_content,
+        content='code_chunks', content_rowid='rowid',
+        tokenize='unicode61'
+      )
+    `)
+    db.exec(`
+      CREATE TRIGGER code_chunks_ai AFTER INSERT ON code_chunks BEGIN
+        INSERT INTO code_chunks_fts(rowid, path, symbol_name, seg_content)
+        VALUES (new.rowid, new.path, new.symbol_name, new.seg_content);
+      END;
+      CREATE TRIGGER code_chunks_ad AFTER DELETE ON code_chunks BEGIN
+        INSERT INTO code_chunks_fts(code_chunks_fts, rowid, path, symbol_name, seg_content)
+        VALUES ('delete', old.rowid, old.path, old.symbol_name, old.seg_content);
+      END;
+      CREATE TRIGGER code_chunks_au AFTER UPDATE ON code_chunks BEGIN
+        INSERT INTO code_chunks_fts(code_chunks_fts, rowid, path, symbol_name, seg_content)
+        VALUES ('delete', old.rowid, old.path, old.symbol_name, old.seg_content);
+        INSERT INTO code_chunks_fts(rowid, path, symbol_name, seg_content)
+        VALUES (new.rowid, new.path, new.symbol_name, new.seg_content);
+      END;
+    `)
+    db.exec(`
+      CREATE TABLE doc_chunks (
+        origin TEXT NOT NULL DEFAULT '',
+        path TEXT NOT NULL,
+        language TEXT NOT NULL DEFAULT 'markdown',
+        heading TEXT NOT NULL DEFAULT '',
+        content TEXT NOT NULL,
+        line_start INTEGER NOT NULL DEFAULT 0,
+        line_end INTEGER NOT NULL DEFAULT 0,
+        mtime_ms INTEGER NOT NULL DEFAULT 0,
+        embedding BLOB,
+        seg_content TEXT NOT NULL DEFAULT '',
+        PRIMARY KEY (origin, path, line_start)
+      )
+    `)
+    db.exec(`
+      CREATE VIRTUAL TABLE doc_chunks_fts USING fts5(
+        path, heading, seg_content,
+        content='doc_chunks', content_rowid='rowid',
+        tokenize='unicode61'
+      )
+    `)
+    db.exec(`
+      CREATE TRIGGER doc_chunks_ai AFTER INSERT ON doc_chunks BEGIN
+        INSERT INTO doc_chunks_fts(rowid, path, heading, seg_content)
+        VALUES (new.rowid, new.path, new.heading, new.seg_content);
+      END;
+      CREATE TRIGGER doc_chunks_ad AFTER DELETE ON doc_chunks BEGIN
+        INSERT INTO doc_chunks_fts(doc_chunks_fts, rowid, path, heading, seg_content)
+        VALUES ('delete', old.rowid, old.path, old.heading, old.seg_content);
+      END;
+      CREATE TRIGGER doc_chunks_au AFTER UPDATE ON doc_chunks BEGIN
+        INSERT INTO doc_chunks_fts(doc_chunks_fts, rowid, path, heading, seg_content)
+        VALUES ('delete', old.rowid, old.path, old.heading, old.seg_content);
+        INSERT INTO doc_chunks_fts(rowid, path, heading, seg_content)
+        VALUES (new.rowid, new.path, new.heading, new.seg_content);
+      END;
+    `)
+    db.exec(`PRAGMA user_version = 8`)
+  }
+
+  if (version < 9) {
+    // v9: files 表 PK 加 origin，防跨项目 project 层记忆覆盖
+    // SQLite 不能 ALTER 主键 → 删表重建（下次 syncDir 自动重索引）
+    db.exec(`
+      DROP TRIGGER IF EXISTS files_ai;
+      DROP TRIGGER IF EXISTS files_ad;
+      DROP TRIGGER IF EXISTS files_au;
+      DROP TABLE IF EXISTS files_fts;
+      DROP TABLE IF EXISTS files;
+    `)
+    db.exec(`
+      CREATE TABLE files (
+        layer TEXT NOT NULL CHECK(layer IN ('project','team')),
+        origin TEXT NOT NULL DEFAULT '',
+        path TEXT NOT NULL,
+        type TEXT NOT NULL CHECK(type IN ('rule','knowledge','decision','pattern')),
+        title TEXT NOT NULL,
+        content TEXT NOT NULL,
+        tags TEXT NOT NULL DEFAULT '',
+        author TEXT NOT NULL DEFAULT '',
+        embedding BLOB,
+        mtime_ms INTEGER NOT NULL DEFAULT 0,
+        seg_title TEXT NOT NULL DEFAULT '',
+        seg_content TEXT NOT NULL DEFAULT '',
+        seg_tags TEXT NOT NULL DEFAULT '',
+        updated_at INTEGER NOT NULL,
+        PRIMARY KEY (layer, origin, path)
+      )
+    `)
+    db.exec(`
+      CREATE VIRTUAL TABLE files_fts USING fts5(
+        seg_title, seg_content, seg_tags,
+        content='files', content_rowid='rowid',
+        tokenize='unicode61'
+      )
+    `)
+    db.exec(`
+      CREATE TRIGGER files_ai AFTER INSERT ON files BEGIN
+        INSERT INTO files_fts(rowid, seg_title, seg_content, seg_tags)
+        VALUES (new.rowid, new.seg_title, new.seg_content, new.seg_tags);
+      END;
+      CREATE TRIGGER files_ad AFTER DELETE ON files BEGIN
+        INSERT INTO files_fts(files_fts, rowid, seg_title, seg_content, seg_tags)
+        VALUES ('delete', old.rowid, old.seg_title, old.seg_content, old.seg_tags);
+      END;
+      CREATE TRIGGER files_au AFTER UPDATE ON files BEGIN
+        INSERT INTO files_fts(files_fts, rowid, seg_title, seg_content, seg_tags)
+        VALUES ('delete', old.rowid, old.seg_title, old.seg_content, old.seg_tags);
+        INSERT INTO files_fts(rowid, seg_title, seg_content, seg_tags)
+        VALUES (new.rowid, new.seg_title, new.seg_content, new.seg_tags);
+      END;
+    `)
+    db.exec(`PRAGMA user_version = 9`)
+  }
+  db.exec("COMMIT")
+  } catch (err) {
+    db.exec("ROLLBACK")
+    throw err
+  }
+}