thincoder 0.7.0 → 0.7.1

package/src/memory.mjs

@@ -19,7 +19,7 @@ import { embed, cosine, toBlob, fromBlob } from "./embedding.mjs"
 import { commitAndPush } from "./gitmem.mjs"
 
 const VALID_TYPES = new Set(["rule", "knowledge", "decision", "pattern"])
-const SCHEMA_VERSION = 7
+const SCHEMA_VERSION = 8
 
 // 代码索引：源码文件扩展名
 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"])
@@ -37,6 +37,9 @@ const BIG_FILE_LINES = 2000
 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 (
@@ -57,11 +60,13 @@ export function createMemory({ dbPath }) {
   return { db }
 }
 
-/** 按 user_version 逐步迁移 */
+/** 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 */
 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(`
@@ -260,6 +265,107 @@ function migrate(db) {
     `)
     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)。
+    // 旧主键不含 origin，多项目共用一个记忆库时同相对路径互相覆盖、codeSync(B) 会把 A 的块当 stale 清掉。
+    // SQLite 不能 ALTER 主键，且索引是易失品 → 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`)
+  }
+  db.exec("COMMIT")
+  } catch (err) {
+    db.exec("ROLLBACK")
+    throw err
+  }
 }
 
 /**
@@ -326,7 +432,10 @@ export async function search(memory, query, { limit = 5 } = {}) {
   return [...scores.entries()]
     .sort((a, b) => b[1] - a[1])
     .slice(0, limit)
-    .map(([id, score]) => ({ ...fetchEntry(memory, id), rrf: score }))
+    .map(([id, score]) => {
+      const entry = fetchEntry(memory, id)
+      return entry ? { ...entry, rrf: score } : null // fetchEntry 为 null 时不能展开（会漏出 { rrf } 空壳）
+    })
     .filter(Boolean)
 }
 
@@ -670,18 +779,18 @@ function yieldTick() {
   return new Promise((r) => setTimeout(r, 0))
 }
 
-function _upsertCodeFile(memory, rel, lines, lang, mtimeMs) {
+function _upsertCodeFile(memory, origin, rel, lines, lang, mtimeMs) {
   const chunks = chunkCode(lines, rel)
   memory.db.exec("BEGIN")
   try {
-    memory.db.prepare(`DELETE FROM code_chunks WHERE path = ?`).run(rel)
+    memory.db.prepare(`DELETE FROM code_chunks WHERE origin = ? AND path = ?`).run(origin, rel)
     const insert = memory.db.prepare(`
-      INSERT INTO code_chunks (path, language, chunk_type, symbol_name, content, line_start, line_end, mtime_ms, seg_content)
-      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+      INSERT INTO code_chunks (origin, path, language, chunk_type, symbol_name, content, line_start, line_end, mtime_ms, seg_content)
+      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
     `)
     for (const c of chunks) {
       const isFile = c.name === rel
-      insert.run(rel, lang, isFile ? "file" : "symbol", isFile ? "" : c.name.slice(rel.length + 1), c.content, c.line_start, c.line_end, mtimeMs, segmentCJK(c.content))
+      insert.run(origin, rel, lang, isFile ? "file" : "symbol", isFile ? "" : c.name.slice(rel.length + 1), c.content, c.line_start, c.line_end, mtimeMs, segmentCJK(c.content))
     }
     memory.db.exec("COMMIT")
   } catch (e) {
@@ -690,18 +799,18 @@ function _upsertCodeFile(memory, rel, lines, lang, mtimeMs) {
   }
 }
 
-function _upsertDocFile(memory, rel, lines, mtimeMs) {
+function _upsertDocFile(memory, origin, rel, lines, mtimeMs) {
   const chunks = chunkMarkdown(lines, rel)
   const lang = rel.endsWith(".rst") ? "rst" : rel.endsWith(".adoc") ? "asciidoc" : rel.endsWith(".txt") ? "text" : "markdown"
   memory.db.exec("BEGIN")
   try {
-    memory.db.prepare(`DELETE FROM doc_chunks WHERE path = ?`).run(rel)
+    memory.db.prepare(`DELETE FROM doc_chunks WHERE origin = ? AND path = ?`).run(origin, rel)
     const insert = memory.db.prepare(`
-      INSERT INTO doc_chunks (path, language, heading, content, line_start, line_end, mtime_ms, seg_content)
-      VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+      INSERT INTO doc_chunks (origin, path, language, heading, content, line_start, line_end, mtime_ms, seg_content)
+      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
     `)
     for (const c of chunks) {
-      insert.run(rel, lang, c.heading, c.content, c.line_start, c.line_end, mtimeMs, segmentCJK(c.content))
+      insert.run(origin, rel, lang, c.heading, c.content, c.line_start, c.line_end, mtimeMs, segmentCJK(c.content))
     }
     memory.db.exec("COMMIT")
   } catch (e) {
@@ -733,15 +842,16 @@ export async function codeSync(memory, dir, { onProgress } = {}) {
   }
   await walk(dir)
 
-  // 取已索引文件的 mtime 快照
+  // 取已索引文件的 mtime 快照（只看本 origin，别的项目的块不归这里管）
   const indexed = new Map(
-    memory.db.prepare(`SELECT path, mtime_ms FROM code_chunks`).all().map((r) => [r.path, r.mtime_ms])
+    memory.db.prepare(`SELECT path, mtime_ms FROM code_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
+  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("\\", "/")
@@ -754,72 +864,81 @@ export async function codeSync(memory, dir, { onProgress } = {}) {
       continue
     }
 
-    // 读文件并分块
-    let text
-    try { text = await readFile(abs, "utf8") } catch { continue }
-    const lines = text.split("\n")
-    const lang = detectLanguage(abs)
-    _upsertCodeFile(memory, rel, lines, lang, mtimeMs)
-    updated++
+    // 单文件失败不拖垮整轮同步：记下错误继续，调用方看 failed/errors
+    try {
+      const text = await readFile(abs, "utf8")
+      const lines = text.split("\n")
+      const lang = detectLanguage(abs)
+      _upsertCodeFile(memory, dir, rel, lines, lang, 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 })
+      onProgress({ phase: "index", current: i + 1, total: files.length, updated, removed, skipped, failed })
     }
   }
 
-  // 清理磁盘上已消失的文件块
+  // 清理磁盘上已消失的文件块（仅本 origin）
   for (const stale of indexed.keys()) {
     if (!seen.has(stale)) {
-      memory.db.prepare(`DELETE FROM code_chunks WHERE path = ?`).run(stale)
+      memory.db.prepare(`DELETE FROM code_chunks WHERE origin = ? AND path = ?`).run(dir, stale)
       removed++
     }
   }
 
-  onProgress?.({ phase: "done", total: files.length, updated, removed, skipped })
-  return { updated, removed, skipped, total: files.length }
+  onProgress?.({ phase: "done", total: files.length, updated, removed, skipped, failed })
+  return { updated, removed, skipped, failed, errors, total: files.length }
 }
 
 /**
  * 代码检索：FTS5(BM25) + 可选向量余弦，RRF 合并。
- * 无 embedder 时退化为纯 FTS。
- * 返回 [{ path, language, symbol_name, content, line_start, line_end, rank }]
+ * 无 embedder 时退化为纯 FTS；ftsQuery 为空（纯标点查询）且有 embedder 时退化为纯向量。
+ * 返回 [{ path, language, symbol_name, content, line_start, line_end }]
  */
 export async function codeSearch(memory, query, { limit = 5 } = {}) {
   const ftsQuery = buildFtsQuery(query)
-  if (!ftsQuery) return []
+  if (!ftsQuery && !memory.embedder) return []
+
+  // codeOrigin 设置时只检索本项目（与 files 表的 projectOrigin 过滤同模式）；未设置时不过滤
+  const ftsOriginFilter = memory.codeOrigin ? `AND c.origin = ?` : ""
+  const vecOriginFilter = memory.codeOrigin ? `AND origin = ?` : ""
+  const originParams = memory.codeOrigin ? [memory.codeOrigin] : []
 
-  const ftsList = memory.db.prepare(`
-    SELECT c.path, c.language, c.symbol_name, c.content, c.line_start, c.line_end, bm25(code_chunks_fts) AS rank
+  const ftsList = ftsQuery ? memory.db.prepare(`
+    SELECT c.rowid, c.path, c.language, c.symbol_name, c.content, c.line_start, c.line_end, bm25(code_chunks_fts) AS rank
     FROM code_chunks_fts JOIN code_chunks c ON c.rowid = code_chunks_fts.rowid
-    WHERE code_chunks_fts MATCH ?
+    WHERE code_chunks_fts MATCH ? ${ftsOriginFilter}
     ORDER BY rank LIMIT ?
-  `).all(ftsQuery, Math.max(limit * 4, 20))
+  `).all(ftsQuery, ...originParams, Math.max(limit * 4, 20)) : []
 
   if (!memory.embedder) return ftsList.slice(0, limit)
 
   // 向量通道
   await ensureEmbeddings(memory)
   const [qvec] = await embed(memory.embedder, [query])
-  const rows = memory.db.prepare(`SELECT path, line_start, embedding FROM code_chunks WHERE embedding IS NOT NULL`).all()
+  const rows = memory.db.prepare(`SELECT rowid, embedding FROM code_chunks WHERE embedding IS NOT NULL ${vecOriginFilter}`).all(...originParams)
   const vecList = rows
-    .map((r) => ({ key: `${r.path}:${r.line_start}`, score: cosine(qvec, fromBlob(r.embedding)) }))
+    .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))
 
-  // RRF 合并
+  // RRF 合并（按 rowid 对齐两个通道；解析结果回表取，纯向量命中的块也能浮现）
   const K = 60
   const scores = new Map()
-  ftsList.forEach((r, i) => scores.set(`${r.path}:${r.line_start}`, (scores.get(`${r.path}:${r.line_start}`) ?? 0) + 1 / (K + i + 1)))
-  vecList.forEach((r, i) => scores.set(r.key, (scores.get(r.key) ?? 0) + 1 / (K + i + 1)))
+  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, symbol_name, content, line_start, line_end FROM code_chunks WHERE rowid = ?
+  `)
   return [...scores.entries()]
     .sort((a, b) => b[1] - a[1])
     .slice(0, limit)
-    .map(([key]) => {
-      const [path, lineStr] = key.split(/:(?=\d+$)/)
-      return ftsList.find((r) => r.path === path && String(r.line_start) === lineStr)
-    })
+    .map(([rowid]) => fetchChunk.get(rowid))
     .filter(Boolean)
 }
 
@@ -878,13 +997,18 @@ export function codeSearchTool(memory) {
 export async function reindexFile(memory, cwd, absPath) {
   const ext = absPath.slice(absPath.lastIndexOf(".")).toLowerCase()
   const rel = relative(cwd, absPath).replaceAll("\\", "/")
-  if (rel.startsWith("..")) return // 越界路径拒索引
+  // 越界路径拒索引：必须是 ".." 或 "../..." 才算越界——文件名以 .. 开头（如 ..foo.js）不算
+  if (rel === ".." || rel.startsWith("../")) return
+  // 与 walk 策略一致：跳过隐藏目录和 SKIP_DIRS 里的文件（node_modules/.git/dist…），
+  // 否则这些文件虽然全量扫描时被跳过，单文件增量更新却会漏进来
+  const dirs = rel.split("/").slice(0, -1)
+  if (dirs.some((d) => SKIP_DIRS.has(d) || d.startsWith("."))) return
 
   let text
   try { text = await readFile(absPath, "utf8") } catch {
     // 文件已删：清理索引
-    if (CODE_EXTS.has(ext)) memory.db.prepare(`DELETE FROM code_chunks WHERE path = ?`).run(rel)
-    else if (DOC_EXTS.has(ext)) memory.db.prepare(`DELETE FROM doc_chunks WHERE path = ?`).run(rel)
+    if (CODE_EXTS.has(ext)) memory.db.prepare(`DELETE FROM code_chunks WHERE origin = ? AND path = ?`).run(cwd, rel)
+    else if (DOC_EXTS.has(ext)) memory.db.prepare(`DELETE FROM doc_chunks WHERE origin = ? AND path = ?`).run(cwd, rel)
     return
   }
   const lines = text.split("\n")
@@ -893,11 +1017,11 @@ export async function reindexFile(memory, cwd, absPath) {
     const lang = detectLanguage(absPath)
     let mtimeMs = 0
     try { mtimeMs = Math.floor((await stat(absPath)).mtimeMs) } catch { /* 新文件 */ }
-    _upsertCodeFile(memory, rel, lines, lang, mtimeMs)
+    _upsertCodeFile(memory, cwd, rel, lines, lang, mtimeMs)
   } else if (DOC_EXTS.has(ext)) {
     let mtimeMs = 0
     try { mtimeMs = Math.floor((await stat(absPath)).mtimeMs) } catch { /* 新文件 */ }
-    _upsertDocFile(memory, rel, lines, mtimeMs)
+    _upsertDocFile(memory, cwd, rel, lines, mtimeMs)
   }
 }
 
@@ -951,13 +1075,14 @@ export async function docSync(memory, dir, { onProgress } = {}) {
   await walk(dir)
 
   const indexed = new Map(
-    memory.db.prepare(`SELECT path, mtime_ms FROM doc_chunks`).all().map((r) => [r.path, r.mtime_ms])
+    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
+  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("\\", "/")
@@ -970,65 +1095,80 @@ export async function docSync(memory, dir, { onProgress } = {}) {
       continue
     }
 
-    let text
-    try { text = await readFile(abs, "utf8") } catch { continue }
-    const lines = text.split("\n")
-    _upsertDocFile(memory, rel, lines, mtimeMs)
-    updated++
+    // 单文件失败不拖垮整轮同步（与 codeSync 同策略）
+    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 })
+      onProgress({ phase: "index", current: i + 1, total: files.length, updated, removed, skipped, failed })
     }
   }
 
-  // 清理磁盘上消失的文件
+  // 清理磁盘上消失的文件（仅本 origin）
   for (const stale of indexed.keys()) {
     if (!seen.has(stale)) {
-      memory.db.prepare(`DELETE FROM doc_chunks WHERE path = ?`).run(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 })
-  return { updated, removed, skipped, total: files.length }
+  onProgress?.({ phase: "done", total: files.length, updated, removed, skipped, failed })
+  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（纯标点查询）且有 embedder 时退化为纯向量。
+ * 返回 [{ path, language, heading, content, line_start, line_end }]
  */
 export async function docSearch(memory, query, { limit = 5 } = {}) {
   const ftsQuery = buildFtsQuery(query)
-  if (!ftsQuery) return []
+  if (!ftsQuery && !memory.embedder) return []
 
-  const ftsList = memory.db.prepare(`
-    SELECT d.path, d.language, d.heading, d.content, d.line_start, d.line_end, bm25(doc_chunks_fts) AS rank
+  // codeOrigin 设置时只检索本项目（与 codeSearch 同模式）；未设置时不过滤
+  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 ?
+    WHERE doc_chunks_fts MATCH ? ${ftsOriginFilter}
     ORDER BY rank LIMIT ?
-  `).all(ftsQuery, Math.max(limit * 4, 20))
+  `).all(ftsQuery, ...originParams, Math.max(limit * 4, 20)) : []
 
   if (!memory.embedder) return ftsList.slice(0, limit)
 
+  // 向量通道
+  await ensureDocEmbeddings(memory)
   const [qvec] = await embed(memory.embedder, [query])
-  const rows = memory.db.prepare(`SELECT path, line_start, embedding FROM doc_chunks WHERE embedding IS NOT NULL`).all()
+  const rows = memory.db.prepare(`SELECT rowid, embedding FROM doc_chunks WHERE embedding IS NOT NULL ${vecOriginFilter}`).all(...originParams)
   const vecList = rows
-    .map((r) => ({ key: `${r.path}:${r.line_start}`, score: cosine(qvec, fromBlob(r.embedding)) }))
+    .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))
 
+  // RRF 合并（按 rowid 对齐两个通道；解析结果回表取，纯向量命中的块也能浮现）
   const K = 60
   const scores = new Map()
-  ftsList.forEach((r, i) => scores.set(`${r.path}:${r.line_start}`, (scores.get(`${r.path}:${r.line_start}`) ?? 0) + 1 / (K + i + 1)))
-  vecList.forEach((r, i) => scores.set(r.key, (scores.get(r.key) ?? 0) + 1 / (K + i + 1)))
+  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(([key]) => {
-      const [path, lineStr] = key.split(/:(?=\d+$)/)
-      return ftsList.find((r) => r.path === path && String(r.line_start) === lineStr)
-    })
+    .map(([rowid]) => fetchChunk.get(rowid))
     .filter(Boolean)
 }
 

package/src/provider.mjs

@@ -106,21 +106,25 @@ export async function chat(provider, { messages, tools, onToken, onReasoning, si
     result.content += continued.content
     result.reasoning += continued.reasoning ?? ""
     for (const tc of continued.toolCalls ?? []) {
-    if (tc.index == null) { result.toolCalls = continued.toolCalls; break }
-    const s = result.toolCalls[tc.index] ??= { id: "", name: "", arguments: "" }
-    if (tc.id) s.id = tc.id
-    s.name += tc.name ?? ""
-    s.arguments += tc.arguments ?? ""
-  }
+      if (tc.index == null) { result.toolCalls = continued.toolCalls; break }
+      const s = result.toolCalls[tc.index] ??= { id: "", name: "", arguments: "" }
+      if (tc.id) s.id = tc.id
+      s.name += tc.name ?? ""
+      s.arguments += tc.arguments ?? ""
+    }
     result.finishReason = continued.finishReason
     if (continued.usage) {
-    result.usage = {
-      prompt_tokens: (result.usage?.prompt_tokens??0) + (continued.usage.prompt_tokens??0),
-      completion_tokens: (result.usage?.completion_tokens??0) + (continued.usage.completion_tokens??0),
-      total_tokens: (result.usage?.total_tokens??0) + (continued.usage.total_tokens??0),
+      const sum = (k) => (result.usage?.[k] ?? 0) + (continued.usage[k] ?? 0)
+      result.usage = {
+        prompt_tokens: sum("prompt_tokens"),
+        completion_tokens: sum("completion_tokens"),
+        total_tokens: sum("total_tokens"),
+        // 缓存命中/未命中也要累计（DeepSeek 计费与状态栏展示依赖这两个字段）
+        prompt_cache_hit_tokens: sum("prompt_cache_hit_tokens"),
+        prompt_cache_miss_tokens: sum("prompt_cache_miss_tokens"),
+      }
     }
   }
-  }
   return result
 }
 

package/src/repomap.mjs

@@ -74,17 +74,15 @@ function parsePyOutline(lines) {
   for (const line of lines) {
     const fromRe = line.match(/^from\s+(\S+)\s+import\s+(.+)/)
     if (fromRe) {
-      const mod = fromRe[1]
-      if (!mod.startsWith(".")) continue // 只本地
-      imports.push(normalizeExt(mod.replace(/^\.+/, "")))
+      const rel = pyRelPath(fromRe[1])
+      if (rel) imports.push(rel)
       continue
     }
     const impRe = line.match(/^import\s+(.+)/)
     if (impRe) {
       for (const mod of impRe[1].split(",")) {
-        const m = mod.trim().split(/\s+/)[0]
-        if (!m.startsWith(".")) continue
-        imports.push(normalizeExt(m.replace(/^\.+/, "")))
+        const rel = pyRelPath(mod.trim().split(/\s+/)[0])
+        if (rel) imports.push(rel)
       }
       continue
     }
@@ -94,6 +92,19 @@ function parsePyOutline(lines) {
   return { imports: [...new Set(imports)], symbols: [...new Set(symbols)] }
 }
 
+/**
+ * Python 相对导入 → 相对文件路径：
+ * 前导 n 个点表示上溯 n-1 层（"."=当前包），模块点号转路径分隔符。
+ * 非相对导入（不以 . 开头）或纯包导入（"from . import x"）返回 null。
+ */
+function pyRelPath(mod) {
+  if (!mod?.startsWith(".")) return null
+  const dots = mod.match(/^\.+/)[0].length
+  const rest = mod.slice(dots).replaceAll(".", "/")
+  if (!rest) return null
+  return normalizeExt("../".repeat(dots - 1) + rest)
+}
+
 function normalizeExt(p) {
   return p.replace(/\.(m?js|jsx|tsx?)$/i, "")
 }

package/src/session.mjs

@@ -45,8 +45,8 @@ function saveManifest(cwd, m) {
   writeSessionFile(manifestPath(cwd), m)
 }
 
-/** 归档当前会话到空闲槽位——满了踢最老 */
-export function archiveCurrent(cwd) {
+/** 归档当前会话到空闲槽位——满了踢最老；exclude 指定一个不许被踢的槽位（switchToSlot 的目标槽） */
+export function archiveCurrent(cwd, { exclude } = {}) {
   const src = sessionPath(cwd)
   if (!existsSync(src)) return
   const m = loadManifest(cwd)
@@ -57,11 +57,13 @@ export function archiveCurrent(cwd) {
     slot = 1
     while (m.slots[slot]) slot++
   } else {
-    slot = Number(entries.sort((a, b) => a[1] - b[1])[0][0])
+    const candidates = entries.filter(([n]) => Number(n) !== exclude)
+    slot = Number((candidates.length ? candidates : entries).sort((a, b) => a[1] - b[1])[0][0])
   }
 
   const dst = slotPath(cwd, slot)
-  writeFileSync(dst, readFileSync(src, "utf8"), "utf8") // 复制（rename 会丢当前）
+  // 复制（rename 会丢当前）；走原子写，防中途崩溃留下截断的 JSON 丢归档
+  writeSessionFile(dst, JSON.parse(readFileSync(src, "utf8")))
   m.slots[slot] = Date.now()
   delete m.slots._currentName
   saveManifest(cwd, m)
@@ -82,7 +84,8 @@ export function switchToSlot(cwd, slot) {
   if (!m.slots[slot]) return null
 
   // 归档当前（内部写 manifest；之后我们的 m 已过期，需重读）
-  archiveCurrent(cwd)
+  // 满槽时排除目标槽：否则最老槽=目标槽，归档会把目标覆盖掉再复制回来，目标会话永久丢失
+  archiveCurrent(cwd, { exclude: slot })
 
   // 槽位文件 → 当前（copy+unlink，不用 rename：Windows rename 目标已存在会抛 EPERM）
   const src = slotPath(cwd, slot)

package/src/skills.mjs

@@ -28,13 +28,17 @@ export async function loadSkills(cwd) {
     try {
       const s = await stat(p)
       if (!s.isFile()) continue
-      // 提取描述（前 400 字符里第一段非空、非标题行）
+      // 提取描述（前 400 字符里第一段非空、非标题行）；文件带 frontmatter 时整块跳过，
+      // 否则会把 frontmatter 字段行（如 "name: x"）误当描述
       const head = await readFile(p, "utf8")
       const body = head.slice(0, 400).split("\n")
       let desc = ""
+      let inFrontmatter = false
       for (const line of body) {
         const t = line.trim()
-        if (t && !t.startsWith("#") && !t.startsWith("---")) {
+        if (t === "---") { inFrontmatter = !inFrontmatter; continue }
+        if (inFrontmatter) continue
+        if (t && !t.startsWith("#")) {
           desc = t.slice(0, 120)
           break
         }

package/src/tools/apply_patch.md

@@ -0,0 +1,11 @@
+Apply a unified diff to one or more files, atomically: if any hunk fails to apply, nothing is written.
+
+Parameters:
+- patch (required): Unified diff text. One `--- a/path` / `+++ b/path` header pair per file, then `@@ -old,count +new,count @@` hunks. Use `--- /dev/null` to create a new file.
+
+Notes:
+- Use this for multi-file changes (e.g. rename an interface + update all callers) — one call, all-or-nothing
+- Hunks are located by their context/removed lines, not line numbers — but the context must match the file EXACTLY. Read the files first and generate the patch from actual content
+- If a hunk's context matches multiple locations it is rejected — add more surrounding context lines
+- Deleting files is not supported — use the delete tool
+- For single-file small edits, edit is simpler; for full rewrites, write is simpler

package/src/tools/checkpoint.md

@@ -0,0 +1,11 @@
+List, create, and restore workspace snapshots (checkpoints). Git repositories only.
+
+Parameters:
+- action (required): "list" | "create" | "rewind"
+- id: snapshot id (required for rewind)
+
+Notes:
+- A checkpoint is AUTO-CREATED before every user task. If uncommitted work was destroyed (by you, a git command, or a failed refactor), use action=list then action=rewind with the latest id to recover it
+- A checkpoint captures all uncommitted state: tracked-file changes (as a diff) plus copies of untracked files
+- Rewind first snapshots the current state, so rewinding is itself reversible
+- Create one manually before risky bulk operations