lark-relay 0.6.1 → 0.6.2

package/lib/store.js

@@ -1,10 +1,6 @@
 'use strict'
 
-// store:一事件一文件 + 按天分目录 + cursor 共享。
-// 多消费者共享 store、各持游标,不做「处理完即删」—— 谁都不能替别人删。
-// 回收由 collect 进程整目录删(store 按天、cursors 按空闲时长)。
-//
-// 踩坑史见 docs/lessons.md
+// 多消费者共享按天存储的事件,各持消费断点;GC 只回收事件目录
 const fs = require('fs')
 const path = require('path')
 const { paths, ensureDir, eventFilename, dayKey, dayKeyOffset } = require('./paths')
@@ -24,7 +20,6 @@ function writeEvent(app, obj, seq) {
 }
 
 // 排序键:stamp -> pid -> seq,数值比较(字典序在位宽变化时会错序)。
-// 必须容忍非法文件名 -- 见 docs/lessons.md#杂项文件
 const NAME_RE = /^(\d+)_(\d+)_(\d+)$/
 
 function parseKey(name) {
@@ -63,8 +58,7 @@ function afterCursor(entry, cursor) {
   return cmpName(entry.name, cName) > 0
 }
 
-// 扫 app 下盘上还在的全部日期。不设扫描窗口:gc 只保留 retain 天,
-// 稳态下「全部日期」就是「retain 天」。见 docs/lessons.md#扫描窗口
+// 扫描盘上保留的全部日期
 function scanAll(app) {
   const out = []
   let days
@@ -100,8 +94,7 @@ function readEvent(entry) {
   }
 }
 
-// 落盘侧去重的种子:读回某天已落盘事件的 message_id。
-// collect 重启后 bus 重投时它是新文件,游标拦不住。见 docs/lessons.md#去重
+// 重启后从当天事件恢复 message_id 去重集合
 function seedIds(app, day = dayKey()) {
   const ids = new Set()
   let names
@@ -113,7 +106,7 @@ function seedIds(app, day = dayKey()) {
   for (const name of names) {
     if (!isEventName(name)) continue
     const ev = readEvent({ file: path.join(paths.appDay(app, day), name) })
-    const id = ev && (ev.message_id || ev.id)
+    const id = ev?.message_id
     if (id) ids.add(id)
   }
   return ids
@@ -136,27 +129,6 @@ function writeCursor(name, app, key) {
   fs.renameSync(tmp, file)
 }
 
-// 标记「刚被用过」。gc 判据是 mtime,而 take 空手退出时不写游标 --
-// 不 touch 会把「盯着一个安静的群」误判成遗弃
-function touchCursor(name, app) {
-  try {
-    const now = new Date()
-    fs.utimesSync(paths.cursor(name, app), now, now)
-    return true
-  } catch {
-    return false // 还没有游标文件(首次跑),seedCursorNow 会建
-  }
-}
-
-// mtime 而非 atime:atime 会被 status 的只读扫描刷新,那不算「在用」
-function cursorMtime(name, app) {
-  try {
-    return fs.statSync(paths.cursor(name, app)).mtimeMs
-  } catch {
-    return null
-  }
-}
-
 // --since now:游标落到当前最新,不重放历史
 function seedCursorNow(name, app) {
   const all = scanAll(app)
@@ -210,70 +182,13 @@ function gcStore(retainDays) {
   return removed
 }
 
-// 游标空闲多久算遗弃。默认 2 天:take 是实时场景的取用口,
-// 空闲一两天以上再重挂,下次消费一定从新的开始。
-//
-// 定义在这里(gcCursors 的所在地)而非 collect.js —— status 要拿它渲染提示,
-// 两处各写一个 2 就会漂移:改了环境变量而 status 仍说「2 天」是在撒谎
-const CURSOR_TTL_DAYS = (() => {
-  const d = Number(process.env.LARK_RELAY_CURSOR_TTL_DAYS)
-  return Number.isFinite(d) && d > 0 ? d : 2
-})()
-
-// gc:删空闲超期的游标。与 gcStore 同构(都是「删超期的东西」),故放一起。
-//
-// 判据是**多久没人用过它**(mtime)。「落后多少条」和「停在哪天」都试过、
-// 方向都是反的 -- 见 docs/lessons.md#游标判据。
-//
-// 不必守「在跑的别删」:take 启动即 touch,默认 --timeout(12h)远小于 TTL(2 天),
-// 在跑的游标不可能超期。test 锁住了这个不变量
-function gcCursors(ttlDays) {
-  const cutoff = Date.now() - ttlDays * 86400_000
-  const removed = []
-  let names
-  try {
-    names = fs.readdirSync(paths.cursors)
-  } catch {
-    return removed
-  }
-  for (const name of names) {
-    if (name.startsWith('.')) continue
-    const dir = path.join(paths.cursors, name)
-    let files
-    try {
-      if (!fs.statSync(dir).isDirectory()) continue
-      files = fs.readdirSync(dir)
-    } catch {
-      continue
-    }
-    // 粒度是 app 而非整个身份 -- 见 docs/lessons.md#回收粒度
-    for (const app of files) {
-      if (app.startsWith('.') || app.endsWith('.tmp')) continue
-      const file = path.join(dir, app)
-      try {
-        if (fs.statSync(file).mtimeMs >= cutoff) continue
-        fs.rmSync(file, { force: true })
-        fs.rmSync(`${file}.tmp`, { force: true })
-        removed.push(`${name}/${app}`)
-      } catch {}
-    }
-    try {
-      if (!fs.readdirSync(dir).length) fs.rmdirSync(dir)
-    } catch {}
-  }
-  return removed
-}
-
 module.exports = {
-  CURSOR_TTL_DAYS,
   writeEvent,
   scanAll,
   readEvent,
   seedIds,
   readCursor,
   writeCursor,
-  touchCursor,
-  cursorMtime,
   seedCursorNow,
   afterCursor,
   entryKey,
@@ -283,5 +198,4 @@ module.exports = {
   latestEntry,
   listApps,
   gcStore,
-  gcCursors,
 }

package/lib/take.js

@@ -1,8 +1,6 @@
 'use strict'
 
-// take:只有一种形态 —— 阻塞等 → 有一批就输出 → 退出。
-// 循环由 harness 承担(Claude Code run_in_background 进程退出会主动通知 AI),
-// 包不做 --follow、不做 nohup 循环、不落中间日志文件。
+// 阻塞取一批,调用方负责处理后再次取用
 const fs = require('fs')
 const crypto = require('crypto')
 const store = require('./store')
@@ -31,6 +29,7 @@ function parseChats(spec) {
     } catch (err) {
       const e = new Error(`--chats 指向的文件读不到:${file}`)
       e.userFacing = true
+      e.exitCode = 2
       throw e
     }
     raw = text
@@ -60,44 +59,35 @@ async function takeBatch(opts) {
     maxWaitMs,
     timeoutMs,
     pollMs = 500,
-    eventKey = 'im.message.receive_v1',
     signal,
   } = opts
 
   const chatSet = new Set(chats)
   const startedAt = Date.now()
   const batch = []
-  const seen = new Set()
   let firstAt = null
   let lastNewAt = null
   let cursor = store.readCursor(name, app)
-  let commit = cursor
 
   for (;;) {
-    // abort(SIGTERM/SIGINT)时必须和正常返回一样落盘 —— 否则批次已吐给调用方
-    // (AI 已看到、可能已回群),游标却没推,重启后同一批会**重跑一遍**(实测:重复回群)
+    // 信号退出也交付已攒的批次,提交时机与正常退出一致
     if (signal && signal.aborted) {
-      if (batch.length && commit) store.writeCursor(name, app, commit)
-      return { events: batch, reason: batch.length ? 'batch' : 'signal', cursor: commit }
+      if (batch.length && cursor) store.writeCursor(name, app, cursor)
+      return { events: batch, reason: batch.length ? 'batch' : 'signal', cursor }
     }
 
     const entries = store.scanAll(app).filter((e) => store.afterCursor(e, cursor))
-    let scannedLast = null
     const fresh = []
     for (const entry of entries) {
-      const key = store.entryKey(entry)
-      scannedLast = key
-      if (seen.has(key)) continue
-      seen.add(key)
+      cursor = store.entryKey(entry)
       const ev = store.readEvent(entry)
       if (!ev) continue // 半成品/坏文件:跳过
-      if (eventKey && ev.type && ev.type !== eventKey) continue
+      if (ev.type !== 'im.message.receive_v1') continue
       if (chatSet.size && !chatSet.has(ev.chat_id)) continue
       fresh.push(ev)
     }
 
-    let passed = []
-    if (fresh.length) passed = applyFilter(fresh, filter)
+    const passed = applyFilter(fresh, filter)
 
     if (passed.length) {
       batch.push(...passed)
@@ -106,12 +96,9 @@ async function takeBatch(opts) {
       lastNewAt = now
     }
 
-    // 游标推进到已扫过的位置:被 filter/chats 丢掉的事件不必重扫。
-    // 批次未吐出时也推进 —— 崩了就重来一轮,不会吐半批。
-    if (scannedLast) commit = scannedLast
-    if (batch.length === 0 && commit && commit !== cursor) {
-      store.writeCursor(name, app, commit)
-      cursor = commit
+    // 内存游标每轮前移;有待交付批次时,磁盘断点保留到交付前
+    if (batch.length === 0 && entries.length) {
+      store.writeCursor(name, app, cursor)
     }
 
     if (batch.length) {
@@ -119,15 +106,15 @@ async function takeBatch(opts) {
       const quiet = now - lastNewAt >= debounceMs
       const capped = now - firstAt >= maxWaitMs
       if (quiet || capped) {
-        // 吐批即推游标,不等 ack —— at-most-once。聊天场景「漏回 < 重复回」
-        if (commit) store.writeCursor(name, app, commit)
-        return { events: batch, reason: 'batch', cursor: commit }
+        // at-most-once:交付前提交,不等待业务处理确认
+        if (cursor) store.writeCursor(name, app, cursor)
+        return { events: batch, reason: 'batch', cursor }
       }
     }
 
     if (Date.now() - startedAt >= timeoutMs) {
-      if (commit) store.writeCursor(name, app, commit)
-      return { events: batch, reason: batch.length ? 'batch' : 'timeout', cursor: commit }
+      if (batch.length && cursor) store.writeCursor(name, app, cursor)
+      return { events: batch, reason: batch.length ? 'batch' : 'timeout', cursor }
     }
 
     await sleep(pollMs)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "lark-relay",
-  "version": "0.6.1",
+  "version": "0.6.2",
   "description": "Lark event relay: collect events to disk, take a batch when you need it.",
   "keywords": [
     "lark",
@@ -14,7 +14,7 @@
   "author": "adaex",
   "type": "commonjs",
   "engines": {
-    "node": ">=18"
+    "node": ">=22"
   },
   "bin": {
     "lark-relay": "bin/lark-relay.js",