@a9i5k4/dsh-auto-memory 2.2.2 → 2.2.4

package/lib/subagent-gc.js

@@ -0,0 +1,239 @@
+/** 子代理痕迹回收(pre 线,2026-09-08)。
+ *
+ * 背景(实测 2026-09-08):DSH 会为每个子代理创建持久化会话
+ * `~/.dsh/sessions/<工作区>/<裸 uuid>/session.jsonl.zstd`,header 形如
+ * `{origin:'subagent', parentSession, delegationDepth}`;一次性委派结束后文件仍留在磁盘。
+ * 本插件的 auto-memory-* 子代理(自动沉淀 / 时段总结 / 问候 / 蒸馏 / 关键词)高频产生:
+ * 实测全机 686 个子代理会话(302MB)中 638 个(93%)来自本插件,数量上千后拖慢会话列表
+ * 与投影缓存加载,并可能触发子代理目录诊断报错。
+ *
+ * 本模块只做两件事,边界清晰、可单测:
+ *   1) 扫描:识别「本插件产生的、已结束的一次性子代理会话」
+ *      (origin==='subagent' && descriptor.label 以 auto-memory- 开头 && descriptor.mode==='one-shot');
+ *   2) 回收:把会话目录 + 投影缓存文件**移动**到备份目录(不做不可逆删除,可整体回滚)。
+ *
+ * 安全边界:
+ *   - 绝不触碰普通会话(目录名以 session- 开头 / header.origin !== 'subagent');
+ *   - continuable 子代理(可续聊)一律跳过;
+ *   - keepMs 之内新产生的会话跳过(避免与运行中的子代理竞争);
+ *   - 任一步失败只记录,不抛错(调用方 fail-soft)。
+ */
+
+import { readdir, stat, rename, mkdir, access } from 'node:fs/promises'
+import { readFile } from 'node:fs/promises'
+import { zstdDecompressSync } from 'node:zlib'
+import path from 'node:path'
+
+/** 本插件 spawn 的子代理 label 前缀(runSubagent 的 label 参数)。 */
+export const PLUGIN_LABEL_PREFIX = 'auto-memory-'
+
+/** 默认保留时长:3 天(兜底 GC 用;任务结束即删路径不受此限制)。 */
+export const DEFAULT_KEEP_MS = 3 * 24 * 60 * 60 * 1000
+
+/**
+ * 按块头定位多帧 zstd 文件的每个帧区间。
+ * 会话日志是逐事件追加的独立 zstd 帧,单次 zstdDecompressSync 只解第一帧。
+ * @param {Buffer} buf - 原始文件内容。
+ * @returns {{start:number,end:number}[]} 帧区间。
+ */
+export function scanZstdFrames(buf) {
+  const frames = []
+  let off = 0
+  while (off + 4 <= buf.length && buf.readUInt32LE(off) === 0xfd2fb528) {
+    const start = off
+    const fhd = buf.readUInt8(off + 4)
+    const fcsFlag = fhd >> 6
+    const single = (fhd >> 5) & 1
+    const dictFlag = fhd & 3
+    const cksum = (fhd >> 2) & 1
+    let p = off + 5
+    if (!single) p += 1
+    if (dictFlag === 1) p += 1
+    else if (dictFlag === 2) p += 2
+    else if (dictFlag === 3) p += 4
+    if (fcsFlag === 0) { if (single) p += 1 }
+    else if (fcsFlag === 1) p += 2
+    else if (fcsFlag === 2) p += 4
+    else p += 8
+    off = p
+    let broken = false
+    for (;;) {
+      if (off + 3 > buf.length) { broken = true; break }
+      const bh = buf.readUIntLE(off, 3)
+      off += 3
+      const last = bh & 1
+      const btype = (bh >> 1) & 3
+      let bsize = bh >>> 3
+      if (btype === 1) bsize = 1
+      else if (btype === 3) { broken = true; break }
+      off += bsize
+      if (last) break
+    }
+    if (cksum) off += 4
+    if (broken && off > buf.length) break
+    frames.push({ start, end: Math.min(off, buf.length) })
+    if (off >= buf.length) break
+  }
+  return frames
+}
+
+/**
+ * 多帧 zstd 全量解压为文本(单帧文件同样适用)。
+ * @param {Buffer} buf - 原始文件内容。
+ * @returns {string} 解压文本(坏帧跳过)。
+ */
+export function decodeZstdFrames(buf) {
+  const parts = []
+  for (const fr of scanZstdFrames(buf)) {
+    try { parts.push(zstdDecompressSync(buf.subarray(fr.start, fr.end))) } catch (e) { /* 坏帧跳过 */ }
+  }
+  return Buffer.concat(parts).toString('utf8')
+}
+
+/**
+ * 从会话日志文本里取首帧 header 与 subagent descriptor。
+ * @param {string} text - 解压后的 JSONL 文本。
+ * @returns {{header:object|null,descriptor:{label?:string,mode?:string}|null,eventCount:number}}
+ */
+export function parseSessionHead(text) {
+  let header = null
+  let descriptor = null
+  let eventCount = 0
+  for (const ln of String(text).split('\n')) {
+    if (!ln) continue
+    let ev = null
+    try { ev = JSON.parse(ln) } catch (e) { continue }
+    eventCount++
+    if (ev.type === 'session' && header === null) header = ev
+    else if (ev.type === 'subagent/descriptor' && descriptor === null && ev.data && typeof ev.data === 'object') {
+      descriptor = { label: ev.data.label, mode: ev.data.mode }
+    }
+    if (header !== null && descriptor !== null) break
+  }
+  return { header, descriptor, eventCount }
+}
+
+/**
+ * 判定一个会话目录是否属于「本插件产生的一次性子代理痕迹」。
+ * @param {{header:object|null,descriptor:object|null}} head - parseSessionHead 的结果。
+ * @param {{labelPrefix?:string,keepMs?:number,now?:number,mtimeMs?:number}} opts - 判据。
+ * @returns {{hit:boolean,reason:string,label:string,mode:string}}
+ */
+export function isPluginOneShotSubagent(head, opts = {}) {
+  const labelPrefix = opts.labelPrefix || PLUGIN_LABEL_PREFIX
+  const keepMs = Number.isFinite(opts.keepMs) ? Number(opts.keepMs) : DEFAULT_KEEP_MS
+  const now = Number.isFinite(opts.now) ? Number(opts.now) : Date.now()
+  const header = head && head.header
+  const descriptor = head && head.descriptor
+  const label = String((descriptor && descriptor.label) || '')
+  const mode = String((descriptor && descriptor.mode) || '')
+  if (!header) return { hit: false, reason: 'no-header', label, mode }
+  if (header.origin !== 'subagent') return { hit: false, reason: 'not-subagent', label, mode }
+  if (header.parentSession === undefined || header.parentSession === null || header.parentSession === '') {
+    return { hit: false, reason: 'no-parent', label, mode }
+  }
+  if (label === '' || !label.startsWith(labelPrefix)) return { hit: false, reason: 'foreign-label', label, mode }
+  if (mode === 'continuable') return { hit: false, reason: 'continuable', label, mode }
+  if (mode !== 'one-shot') return { hit: false, reason: 'unknown-mode', label, mode }
+  const mt = Number(opts.mtimeMs || 0)
+  if (keepMs > 0 && mt > 0 && now - mt < keepMs) return { hit: false, reason: 'too-recent', label, mode }
+  return { hit: true, reason: 'hit', label, mode }
+}
+
+/**
+ * 扫描会话根目录,列出可回收的本插件子代理会话。
+ * @param {{sessionsRoot:string,labelPrefix?:string,keepMs?:number,now?:number,maxEntries?:number}} opts - 扫描参数。
+ * @returns {Promise<{candidates:object[],scanned:number,subagents:number,bytes:number,reasons:Record<string,number>}>}
+ */
+export async function scanPluginSubagentSessions(opts = {}) {
+  const sessionsRoot = opts.sessionsRoot
+  if (!sessionsRoot) throw new Error('scanPluginSubagentSessions: sessionsRoot required')
+  const labelPrefix = opts.labelPrefix || PLUGIN_LABEL_PREFIX
+  const keepMs = Number.isFinite(opts.keepMs) ? Number(opts.keepMs) : DEFAULT_KEEP_MS
+  const now = Number.isFinite(opts.now) ? Number(opts.now) : Date.now()
+  const maxEntries = Number(opts.maxEntries || 20000)
+  const out = { candidates: [], scanned: 0, subagents: 0, bytes: 0, reasons: {} }
+  const bump = (r) => { out.reasons[r] = (out.reasons[r] || 0) + 1 }
+  let wsDirs = []
+  try { wsDirs = await readdir(sessionsRoot, { withFileTypes: true }) } catch (e) { return out }
+  for (const ws of wsDirs) {
+    if (!ws.isDirectory()) continue
+    const wsDir = ws.name
+    let entries = []
+    try { entries = await readdir(path.join(sessionsRoot, wsDir), { withFileTypes: true }) } catch (e) { continue }
+    for (const ent of entries) {
+      if (!ent.isDirectory()) continue
+      if (out.scanned >= maxEntries) break
+      out.scanned++
+      const sid = ent.name
+      const dir = path.join(sessionsRoot, wsDir, sid)
+      const file = path.join(dir, 'session.jsonl.zstd')
+      let st = null
+      try { st = await stat(file) } catch (e) { bump('no-file'); continue }
+      let text = ''
+      try { text = decodeZstdFrames(await readFile(file)) } catch (e) { bump('decode-error'); continue }
+      const head = parseSessionHead(text)
+      if (!head.header) { bump('no-header'); continue }
+      if (head.header.origin === 'subagent') out.subagents++
+      const verdict = isPluginOneShotSubagent(head, { labelPrefix, keepMs, now, mtimeMs: st.mtimeMs })
+      bump(verdict.reason)
+      if (!verdict.hit) continue
+      out.candidates.push({
+        wsDir,
+        sid,
+        dir,
+        file,
+        label: verdict.label,
+        mode: verdict.mode,
+        parentSession: String(head.header.parentSession || ''),
+        delegationDepth: Number(head.header.delegationDepth || 0),
+        sizeBytes: st.size,
+        mtimeMs: st.mtimeMs,
+      })
+      out.bytes += st.size
+    }
+  }
+  out.candidates.sort((a, b) => a.mtimeMs - b.mtimeMs)
+  return out
+}
+
+/** 目标路径是否已存在。 */
+async function exists(p) {
+  try { await access(p); return true } catch (e) { return false }
+}
+
+/**
+ * 把候选会话移动到备份目录(可回滚),并顺带移走对应的投影缓存文件。
+ * 不做任何删除;同名冲突时跳过并记录。
+ * @param {{candidates:object[],backupRoot:string,projcacheRoot?:string,apply?:boolean}} opts - 回收参数。
+ * @returns {Promise<{moved:object[],skipped:object[],failed:object[],bytes:number,applied:boolean}>}
+ */
+export async function recycleSessions(opts = {}) {
+  const apply = opts.apply === true
+  const backupRoot = opts.backupRoot
+  const projcacheRoot = opts.projcacheRoot || ''
+  const out = { moved: [], skipped: [], failed: [], bytes: 0, applied: apply }
+  if (!backupRoot) throw new Error('recycleSessions: backupRoot required')
+  for (const c of opts.candidates || []) {
+    const destDir = path.join(backupRoot, c.wsDir || '_', c.sid)
+    try {
+      if (await exists(destDir)) { out.skipped.push({ sid: c.sid, reason: 'backup-exists' }); continue }
+      if (!apply) { out.moved.push({ sid: c.sid, label: c.label, sizeBytes: c.sizeBytes, to: destDir, dryRun: true }); out.bytes += c.sizeBytes; continue }
+      await mkdir(path.dirname(destDir), { recursive: true })
+      await rename(c.dir, destDir)
+      out.moved.push({ sid: c.sid, label: c.label, sizeBytes: c.sizeBytes, to: destDir })
+      out.bytes += c.sizeBytes
+      if (projcacheRoot) {
+        const cacheFile = path.join(projcacheRoot, c.sid)
+        if (await exists(cacheFile)) {
+          const cacheDest = path.join(backupRoot, '_projcache', c.sid)
+          await mkdir(path.dirname(cacheDest), { recursive: true })
+          await rename(cacheFile, cacheDest).catch(() => {})
+        }
+      }
+    } catch (e) {
+      out.failed.push({ sid: c.sid, error: e && e.message ? e.message : String(e) })
+    }
+  }
+  return out
+}

package/lib/water-window.js

@@ -0,0 +1,88 @@
+/** 上下文窗口解析(2.2.4,2026-09-08)。
+ *
+ * 背景:水位 = 当前上下文占用 / 模型窗口。占用取官方 tokenMeter(与聊天框 context ring 同源),
+ * 窗口此前有三条来源:手动覆盖 > 官方 request/context 的 contextWindow > settings.yaml 解析 > fallback。
+ * 实测发现两条路同时失效,导致窗口恒为 fallback 131072(真实模型 1M),水位被算成 150%+:
+ *   1) 官方 request/context 只在会话早期追加(本会话 2999 个事件里仅 2 条),而旧代码只扫最近 256 条 → 永远扫不到;
+ *   2) settings.yaml 的 llm-deepseek 段是 flow 风格 YAML(`{ models: [ { id: x, contextWindow: 1000000, ... } ] }`),
+ *      旧解析器只认 block 风格(`- id: x` + 行尾纯数字),整段解析不出来。
+ *
+ * 本模块把「settings.yaml → 模型窗口映射」与「按 provider/model 选窗口」抽成纯函数,便于 smoke 覆盖。
+ */
+
+/**
+ * 从 settings.yaml 文本解析 `模型 id → contextWindow` 映射。
+ * 同时支持 block 风格(llm-pi-ai 那种 `- id: x` / `contextWindow: N`)与 flow 风格(llm-deepseek 那种 `id: x,` / `contextWindow: N,`)。
+ * @param {string} text - settings.yaml 原文。
+ * @returns {{byRoute:Record<string,number>,byModel:Record<string,number>}} byRoute 键为 provider/id,byModel 键为 id。
+ */
+export function parseModelWindowsPre(text) {
+  const byRoute = {}
+  const byModel = {}
+  if (!text || typeof text !== 'string') return { byRoute, byModel }
+  const lines = String(text).split(/\r?\n/)
+  let currentProvider = ''
+  let currentId = null
+  for (const line of lines) {
+    // provider 段(block 风格,2 空格缩进的裸键,如 `opencode-go:`);flow 风格段名同样以冒号结尾但缩进更深
+    const pm = line.match(/^ {2}([A-Za-z0-9_.-]+):\s*$/)
+    if (pm) { currentProvider = pm[1]; currentId = null; continue }
+    // 模型 id:兼容 block(`- id: x`)与 flow(`id: x,`)两种写法,允许行尾逗号
+    const im = line.match(/^\s*(?:-\s*)?id:\s*([A-Za-z0-9._:-]+?)\s*,?\s*$/)
+    if (im) { currentId = im[1]; if (byModel[currentId] === undefined) byModel[currentId] = 0; continue }
+    // 窗口:允许行尾逗号(flow 风格)
+    const cm = line.match(/^\s*contextWindow:\s*([0-9]+)\s*,?\s*$/)
+    if (cm && currentId) {
+      const value = Number(cm[1])
+      const route = currentProvider + '/' + currentId
+      if (byRoute[route] === undefined) byRoute[route] = value
+      if (!byModel[currentId]) byModel[currentId] = value
+    }
+  }
+  return { byRoute, byModel }
+}
+
+/**
+ * 按 provider/model 选窗口:先精确匹配 provider/model,再按 model 名全局兜底。
+ * @param {{byRoute:Record<string,number>,byModel:Record<string,number>}} windows - parseModelWindowsPre 结果。
+ * @param {string} provider - 路由 provider(可空)。
+ * @param {string} model - 模型 id。
+ * @returns {number} 窗口 token 数;未命中返回 0。
+ */
+export function pickWindowPre(windows, provider, model) {
+  if (!windows || !model) return 0
+  const byRoute = windows.byRoute || {}
+  const byModel = windows.byModel || {}
+  const exact = byRoute[(provider || '') + '/' + model]
+  if (Number.isFinite(exact) && exact > 0) return exact
+  const direct = byModel[model]
+  if (Number.isFinite(direct) && direct > 0) return direct
+  // 后缀兜底:路由 id 可能带 provider 前缀或版本后缀差异
+  for (const route of Object.keys(byRoute)) {
+    if (route.endsWith('/' + model) && byRoute[route] > 0) return byRoute[route]
+  }
+  for (const id of Object.keys(byModel)) {
+    if (id.endsWith('-' + model) && byModel[id] > 0) return byModel[id]
+  }
+  return 0
+}
+
+/**
+ * 从会话事件里取最近一条 request/context 的 contextWindow(路由权威值)。
+ * 不限制扫描窗口:该事件只在请求头部变化时追加,长会话里往往只存在于开头。
+ * @param {object[]} events - 会话事件数组。
+ * @param {number} maxScan - 最多向前扫多少条(默认全部;设小值可限制开销)。
+ * @returns {number} contextWindow;未找到返回 0。
+ */
+export function findOfficialContextWindowPre(events, maxScan = 0) {
+  if (!Array.isArray(events) || !events.length) return 0
+  const floor = maxScan > 0 ? Math.max(0, events.length - maxScan) : 0
+  for (let i = events.length - 1; i >= floor; i--) {
+    const ev = events[i]
+    if (ev && ev.type === 'request/context') {
+      const cw = Number(ev && ev.data && ev.data.contextWindow)
+      return Number.isFinite(cw) && cw > 0 ? cw : 0
+    }
+  }
+  return 0
+}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@a9i5k4/dsh-auto-memory",
   "description": "Proactive associative memory for DSH: zero-prompt recall injected before the model speaks, three-layer auto-consolidation, skill crystallization, and Astra-style context management - handoff ledgers, PLAN whiteboard, water-level sensing. Local-first, model-agnostic, zero deps. 主动联想记忆+Astra 式上下文管理:自动唤回/自动沉淀/技能固化/交接账本与白板跨窗口续命/水位感知。",
-  "version": "2.2.2",
+  "version": "2.2.4",
   "type": "module",
   "main": "lib/index.js",
   "exports": {
@@ -12,6 +12,7 @@
   "files": [
     "lib",
     "python",
+    "docs",
     "cordis.patch.yml",
     "!python/bench",
     "!python/__pycache__"