dsh-disk-cleaner 1.2.0 → 1.4.0

package/README.md

@@ -5,7 +5,7 @@
 <!-- badges -->
 | | |
 | --- | --- |
-| **version** | `1.2.0` |
+| **version** | `1.4.0` |
 | **DSH** | `0.2.1-rc.5`（开发与实测版本；用到 `plugins.bundle.config`、`settings.section`） |
 | **license** | [MIT](LICENSE) |
 | **Node** | `>= 22`（仅运行冒烟测试需要） |
@@ -16,7 +16,7 @@
 
 **dsh-disk-cleaner** is a disk-space plugin for [DeepSeek Harness](https://github.com/deepseek-ai) (DSH). It does two different jobs, and keeps them apart on purpose:
 
-1. **Clean** — nineteen known-safe locations (temp folders, package manager stores, browser caches, crash dumps, the Recycle Bin, leftover update installers, per-app caches and logs, superseded plugin versions and applied update packages). Only these are ever deleted.
+1. **Clean** — twenty-eight known-safe locations (temp folders, package manager stores, browser caches, crash dumps, the Recycle Bin, leftover update installers, per-app caches and logs, superseded plugin versions and applied update packages). Only these are ever deleted.
 2. **Assess** — everything the plugin will *not* delete (installed programs, the WSL virtual disk, the Windows component store, chat-app data, your own folders). Measured and listed, never removed.
 
 The assessment ends with a **plan**: the same numbers ordered by what it costs you to act, with the exact Windows command for the steps the plugin cannot perform itself.
@@ -27,9 +27,9 @@ Both the panel and the agent reach the *same* operations: the cleaning logic liv
 
 | Feature | Description |
 | --- | --- |
-| Scan | Measures all nineteen targets concurrently and shows what each occupies |
+| Scan | Measures all twenty-eight targets concurrently and shows what each occupies |
 | Per-drive filter | Chip row of mounted volumes; scanning or cleaning can be limited to one drive, and rows with nothing on that drive disappear |
-| Risk labels | Every target is `safe` or `caution`, with a one-line reason (`in-use` / `rebuild` / `elevation` / `pending-update` / `rollback` / `superseded`) and a permanent legend, so "caution" never means "unspecified" |
+| Risk labels | Every target is `safe` or `caution`, with a one-line reason (`in-use` / `rebuild` / `elevation` / `pending-update` / `rollback` / `superseded` / `history` / `duplicate` / `database-backup`) and a permanent legend, so "caution" never means "unspecified" |
 | Aged deletion | `temp` and `windows-temp` only remove files older than one day; files a running program still holds are skipped |
 | Space assessment | Walks the disk read-only and groups the result into Windows / installed software / your own files, plus the ten largest app-data folders. Any row opens its folder in Explorer — click, right click, or Enter |
 | Honest re-run | Re-running the assessment re-measures; a plain call still answers from the ten-minute cache |
@@ -65,6 +65,15 @@ Every path is built from this table — a caller selects an **id**, never a path
 | `app-temp` | scratch folders apps keep inside their own `%APPDATA%` folder instead of `%TEMP%`: `WorkDaddy\client-tmp`, `PersonalCenter\Temp` | aged, 2 days | caution | the owning app may be running and holding files |
 | `plugin-old-versions` | every version but the newest of each family in the WPS plugin pool `kingsoft\wps\addons\pool\win-i386` | purge | caution | the manager loads the newest; the rest are rollback copies |
 | `update-packages` | update packages already applied: FTNN `.spd` patches whose build number is below the newest `Program Files\FTNN\app\[version]`, and `WorkDaddy\update` installers below the installed version | purge | caution | decided by version comparison, not by age |
+| `session-backups` | `modify_backup` folders under `%USERPROFILE%\.workbuddy\workspace\sessions\*` — a whole-file snapshot of edits per agent session | purge | caution | the session itself lives beside the snapshot and does not read it back |
+| `agent-logs` | date-named log folders under `%USERPROFILE%\.workbuddy\logs` and `\.workbuddy-ai\logs` | aged, 3 days | safe | nothing reads an old log back |
+| `toolchain-snapshots` | duplicate snapshots under `%USERPROFILE%\.dsh\dsh-ops\bash-ipc`: folders whose `.ops-ipc.json` carries the same `patchedSha256` collapse to the newest | purge | caution | a folder without that metadata is never touched |
+| `editor-server-bundles` | every bundle but the newest under `%USERPROFILE%\vscode-remote-wsl\stable` | purge | caution | the editor re-downloads the one it needs |
+| `database-backups` | `zotero.sqlite*.bak` under `%USERPROFILE%\Zotero` | purge | caution | the live `zotero.sqlite` sits beside the copies |
+| `editor-builds` | complete build trees under `%LOCALAPPDATA%\Programs\Microsoft VS Code` that `bin\code.cmd` does not name, plus `old_*` files whose original still sits beside them | purge | caution | no launcher, a launcher naming a missing build, or a lone build means nothing is dropped |
+| `agent-app-caches` | `Cache` / `Code Cache` / `GPUCache` under `%USERPROFILE%\.workbuddy\app` and `\.workbuddy-ai\app` | purge | caution | the sibling `cache` folder holds spilled conversation text and is never a target |
+| `rewind-snapshots` | `session-*` folders under `%USERPROFILE%\.dsh\rewind-snapshots` | aged, 30 days | caution | a recent session stays restorable |
+| `rotated-logs` | `*.old.log` and `*.log.N` files in the `logs` folders of the agent apps | aged, 3 days | safe | nothing reads a rotated file back |
 
 `aged` targets skip anything modified within the retention window; `purge` targets remove every entry. Both keep the target directory itself and leave locked entries alone, counting them as skipped.
 
@@ -139,6 +148,11 @@ These are enforced in `lib/catalog.js` and must not be relaxed:
 - Symlinks and junctions met during a walk are skipped, never followed or deleted.
 - Locked or denied entries are counted as skipped and the walk continues.
 - The target directory itself is kept; only its contents go.
+- Rules that pick "the newest of a group" fail closed: a family with only one member, a folder
+  whose metadata file is missing or whose hash is not a string, and a `vscode-remote-wsl\stable`
+  that holds a single bundle all resolve to *nothing to delete*. The newest member is never returned.
+- `editor-builds` drops a build only when `bin\code.cmd` exists and names a different build that is also on disk; `old_x` goes only while `x` is still beside it.
+- `database-backups` matches `.sqlite*.bak` by name only; the live `zotero.sqlite` is never a candidate.
 - The directory Windows refuses to let an unelevated process write (`SoftwareDistribution\Download`) is labelled `elevation` and excluded from "reclaimable now". `C:\Windows\Temp` is **not** one of them on a default ACL: it grants `BUILTIN\Users` create and delete, so only EBUSY on a live file stops a delete there, which the age limit covers.
 
 ### Install
@@ -175,7 +189,7 @@ The bundle's own `cordis.patch.yml` inserts the plugin row; no profile patch edi
 ### Development
 
 ```bash
-node smoke.mjs                       # 67 assertions, zero dependencies
+node smoke.mjs                       # 68 assertions, zero dependencies
 node --check lib/client.js           # and lib/index.js, lib/catalog.js, lib/audit.js, lib/plan.js
 ```
 
@@ -199,6 +213,9 @@ What the smoke test deliberately does **not** cover: anything that walks a real
 | Dry run deletes nothing | `confirm` omitted, byte counts unchanged |
 | Drive filter rejects unmounted and malformed volumes | `C:extra`, `Z:`, `abc` → 400 |
 | Plan excludes elevation bytes from `reclaimNow` | Live audit against this machine |
+| "Keep only the newest of a group" fails closed | Throwaway fixture tree: a same-hash duplicate collapses to the newest, a folder with no metadata is left alone, a different hash is kept, the live database survives (9/9) |
+| Old editor builds need a launcher that proves which one runs | Throwaway fixture tree: a build the launcher does not name goes, no launcher deletes nothing, a launcher naming a missing build deletes nothing, a lone build stays, `old_x` goes only while `x` exists (7/7) |
+| A dry run reports what a real run frees | The 30-day rewind window reports 263 KB not 283 MB, and scan matches the dry run |
 | Visual layout | Checked by hand in the running Web UI. No browser automation here, so this one is a human check, not a test. |
 
 ### Implementation notes
@@ -223,16 +240,16 @@ What the smoke test deliberately does **not** cover: anything that walks a real
 ```
 ├─ package.json           dsh.bundle.patch + dsh.client { platform: "web" }
 ├─ cordis.patch.yml       - insert:  id: dsh-disk-cleaner
-├─ lib/index.js    (405)  Host: HTTP route + disk_cleaner tool, same run() for both; reveal
-├─ lib/catalog.js  (688)  The 19 targets, measurePath / purgePath / resolveTargets / drives
+├─ lib/index.js    (411)  Host: HTTP route + disk_cleaner tool, same run() for both; reveal
+├─ lib/catalog.js  (969)  The 28 targets, measurePath / purgePath / resolveTargets / drives
 ├─ lib/audit.js    (365)  Read-only assessment, 10-minute cache, areaPaths, exposes no deletion
 ├─ lib/plan.js     (150)  Pure: measurements → ordered steps
-├─ lib/client.js  (1355)  Client half: slots, panel, preferences, zh/en dictionaries, CSS
+├─ lib/client.js  (1379)  Client half: slots, panel, preferences, zh/en dictionaries, CSS
 ├─ locale/{en,zh}.json    Display metadata (meta.title / meta.description)
 ├─ icon.svg               Plugin icon
 ├─ README.md              This file
 ├─ LICENSE                MIT
-└─ smoke.mjs              67 assertions, zero dependencies
+└─ smoke.mjs              68 assertions, zero dependencies
 ```
 
 Client-half slots: `sidebar.panellist` (order 50, disposed in settings-only mode), `main` (keyed by the package name), `settings.section` (order 30, disposed in sidebar-only mode), `plugins.bundle.config` (keyed by the package name).
@@ -244,7 +261,7 @@ Client-half slots: `sidebar.panellist` (order 50, disposed in settings-only mode
 
 **dsh-disk-cleaner** 是 [DeepSeek Harness](https://github.com/deepseek-ai)（DSH）的磁盘空间插件。它做两件不同的事，并且刻意分开：
 
-1. **清理** —— 十九个已知安全的位置（临时目录、包管理器存储、浏览器缓存、崩溃转储、回收站、更新安装包残留、各应用的缓存与日志、被淘汰的插件版本与已应用的更新包）。只有这些会被删除。
+1. **清理** —— 二十八个已知安全的位置（临时目录、包管理器存储、浏览器缓存、崩溃转储、回收站、更新安装包残留、各应用的缓存与日志、被淘汰的插件版本与已应用的更新包）。只有这些会被删除。
 2. **评估** —— 所有本插件**不会**删的东西（已装程序、WSL 虚拟磁盘、Windows 组件库、聊天软件数据、你自己的文件夹）。只测量、只列出，绝不删除。
 
 评估的最后是一份**处理方案**：同一批数字按「你要付出多少代价」排序，本插件做不到的步骤附上确切的 Windows 命令。
@@ -255,9 +272,9 @@ Client-half slots: `sidebar.panellist` (order 50, disposed in settings-only mode
 
 | 功能 | 说明 |
 | --- | --- |
-| 扫描 | 并发测量全部十九个目标，显示各自占用 |
+| 扫描 | 并发测量全部二十八个目标，显示各自占用 |
 | 按盘过滤 | 已挂载卷的胶囊按钮；扫描与清理都可限定单盘，该盘上为空的项会隐藏 |
-| 风险标注 | 每个目标都标 `安全` 或 `注意`，并给一行理由（`in-use` / `rebuild` / `elevation` / `pending-update` / `rollback` / `superseded`），图例常驻，所以「注意」永远不是「说不清」 |
+| 风险标注 | 每个目标都标 `安全` 或 `注意`，并给一行理由（`in-use` / `rebuild` / `elevation` / `pending-update` / `rollback` / `superseded` / `history` / `duplicate` / `database-backup`），图例常驻，所以「注意」永远不是「说不清」 |
 | 保留期删除 | `temp` 与 `windows-temp` 只删一天前的文件；被运行中程序占住的自动跳过 |
 | 空间评估 | 只读遍历磁盘，按 Windows 自身 / 已装软件 / 你自己的文件 分组，另列占用最大的十个应用数据目录。每一行都能打开所在目录——单击、右键或回车 |
 | 重新评估真的重测 | 「重新评估」会重新测量；普通调用仍然走十分钟缓存 |
@@ -293,6 +310,15 @@ Client-half slots: `sidebar.panellist` (order 50, disposed in settings-only mode
 | `app-temp` | 应用自己放在 `%APPDATA%` 里的临时目录（不走 `%TEMP%`）：`WorkDaddy\client-tmp`、`PersonalCenter\Temp` | 保留期 2 天 | 注意 | 应用可能正在运行并占用其中的文件 |
 | `plugin-old-versions` | WPS 插件池 `kingsoft\wps\addons\pool\win-i386` 里，每个插件族除最高版本以外的全部版本 | 全清 | 注意 | 管理器加载最高版本，其余是回滚副本 |
 | `update-packages` | 已应用的更新包：FTNN 的 `.spd` 补丁（build 号低于 `Program Files\FTNN\app\[版本]` 中最新者），以及 `WorkDaddy\update` 里低于已装版本的安装包 | 全清 | 注意 | 按版本号比较判断，不按时间 |
+| `session-backups` | `%USERPROFILE%\.workbuddy\workspace\sessions\*` 下的 `modify_backup` 目录——每个 agent 会话的整文件修改快照 | 全清 | 注意 | 会话本体在快照旁边，不会读回这些文件 |
+| `agent-logs` | `%USERPROFILE%\.workbuddy\logs` 与 `\.workbuddy-ai\logs` 下按日期命名的日志目录 | 保留期 3 天 | 安全 | 旧日志不会被读回 |
+| `toolchain-snapshots` | `%USERPROFILE%\.dsh\dsh-ops\bash-ipc` 下重复的快照：`.ops-ipc.json` 里 `patchedSha256` 相同的目录只留最新一份 | 全清 | 注意 | 没有该元数据的目录永不触碰 |
+| `editor-server-bundles` | `%USERPROFILE%\vscode-remote-wsl\stable` 下除最新以外的服务包 | 全清 | 注意 | 编辑器需要时会重新下载 |
+| `database-backups` | `%USERPROFILE%\Zotero` 下的 `zotero.sqlite*.bak` | 全清 | 注意 | 正在用的 `zotero.sqlite` 就在副本旁边 |
+| `editor-builds` | `%LOCALAPPDATA%\Programs\Microsoft VS Code` 下 `bin\code.cmd` 未指向的完整构建目录，以及原件仍在旁边的 `old_*` 文件 | 全清 | 注意 | 没有启动器、启动器指向的构建不在磁盘上、或只有一份构建时，一律不删 |
+| `agent-app-caches` | `%USERPROFILE%\.workbuddy\app` 与 `\.workbuddy-ai\app` 下的 `Cache` / `Code Cache` / `GPUCache` | 全清 | 注意 | 同级的 `cache` 目录存的是对话溢写内容，从不作为目标 |
+| `rewind-snapshots` | `%USERPROFILE%\.dsh\rewind-snapshots` 下的 `session-*` 目录 | 保留期 30 天 | 注意 | 最近的会话仍然可以回滚 |
+| `rotated-logs` | 各 agent 应用 `logs` 目录下的 `*.old.log` 与 `*.log.N` 文件 | 保留期 3 天 | 安全 | 轮转掉的日志不会被读回 |
 
 保留期方式会跳过保留期内修改过的文件；全清方式清掉全部条目。两者都保留目标目录本身，并把锁定或无权限的条目跳过计数。
 
@@ -308,8 +334,8 @@ Client-half slots: `sidebar.panellist` (order 50, disposed in settings-only mode
 | `reclaim-caution` | 注意目标共 ≥200 MB（不含需提权的） | 是，一个按钮 |
 | `wsl-compact` | WSL 映像 ≥1 GB | 否 —— `wsl --shutdown` + `diskpart` 的 `compact vdisk` |
 | `winsxs-cleanup` | WinSxS ≥1 GB | 否 —— `Dism.exe /Online /Cleanup-Image /StartComponentCleanup` |
-| `elevation-clean` | 需提权目标 ≥200 MB | 否 —— 以管理员身份运行本插件 |
 | `hibernate-off` | `system-files` ≥200 MB | 否 —— `powercfg.exe /hibernate off` |
+| `elevation-clean` | 需提权目标 ≥200 MB | 否 —— 以管理员身份运行本插件 |
 | `app-review` | 单个应用数据目录 ≥1 GB，取前 3 个且未被其它步骤覆盖 | 否 —— 走卸载，绝不删目录 |
 | `move-folder` | 个人文件夹 ≥2 GB，且另有盘有空间 | 否 —— 属性 → 位置 |
 | `pressure` | 系统盘可用 < 12% | 否 —— 作为上述步骤的背景 |
@@ -367,6 +393,10 @@ Client-half slots: `sidebar.panellist` (order 50, disposed in settings-only mode
 - 遍历时遇到符号链接与 junction 一律跳过，不跟随也不删除。
 - 锁定或无权限的条目计为跳过，遍历继续。
 - 保留目标目录本身，只清内容。
+- 「只留最新一份」这类规则一律选保守的一侧：族内只有一个成员、元数据文件缺失或哈希不是字符串、
+  `vscode-remote-wsl\stable` 下只有一个服务包，结果都是*没有可删项*。最新那一份永远不会被返回。
+- `editor-builds` 只有在 `bin\code.cmd` 存在且指向另一份确实在磁盘上的构建时才删；`old_x` 也只有在 `x` 仍在旁边时才删。
+- `database-backups` 只按名字匹配 `.sqlite*.bak`，正在用的 `zotero.sqlite` 永不成为候选。
 - 那个拒绝普通权限写入的 Windows 目录标为 `elevation`，并排除在「现在能回收」之外。
 
 ### 安装
@@ -407,7 +437,7 @@ git clone https://github.com/stallenlu/dsh-disk-cleaner
 ### 开发
 
 ```bash
-node smoke.mjs                       # 67 条断言，零依赖
+node smoke.mjs                       # 68 条断言，零依赖
 node --check lib/client.js           # 以及 lib/index.js、lib/catalog.js、lib/audit.js、lib/plan.js
 ```
 
@@ -431,6 +461,9 @@ node --check lib/client.js           # 以及 lib/index.js、lib/catalog.js、li
 | 试运行不删任何东西 | 不传 `confirm`，字节数不变 |
 | 盘过滤拒绝未挂载与畸形卷名 | `C:extra`、`Z:`、`abc` → 400 |
 | 方案不把提权目标的字节算进「现在能回收」 | 在本机跑真实 audit |
+| 「只留最新一份」一律选保守一侧 | 一次性 fixture 目录：同哈希的重复项收敛到最新、无元数据的目录不动、不同哈希保留、正在用的数据库完好（9/9） |
+| 旧版编辑器构建要有启动器作证才删 | 一次性 fixture 目录：启动器未指向的构建会删、没有启动器一律不删、启动器指向的构建不在磁盘上也不删、只有一份构建时不删、`old_x` 只在 `x` 仍在时删（7/7） |
+| 预演报告的字节数等于真删能释放的字节数 | 30 天的回滚窗口报 263 KB 而不是 283 MB，且扫描结果与预演一致 |
 | 视觉排版 | 在运行中的 Web UI 里手工确认过。本环境没有浏览器自动化，所以这条是人眼检查，不是测试。 |
 
 ### 实现要点
@@ -455,16 +488,16 @@ node --check lib/client.js           # 以及 lib/index.js、lib/catalog.js、li
 ```
 ├─ package.json           dsh.bundle.patch + dsh.client { platform: "web" }
 ├─ cordis.patch.yml       - insert:  id: dsh-disk-cleaner
-├─ lib/index.js    (405)  Host：HTTP 路由 + disk_cleaner 工具，共用同一个 run()，含 reveal
-├─ lib/catalog.js  (688)  19 个目标，measurePath / purgePath / resolveTargets / drives
+├─ lib/index.js    (411)  Host：HTTP 路由 + disk_cleaner 工具，共用同一个 run()，含 reveal
+├─ lib/catalog.js  (969)  28 个目标，measurePath / purgePath / resolveTargets / drives
 ├─ lib/audit.js    (365)  只读评估，10 分钟缓存，带 areaPaths，不暴露任何删除
 ├─ lib/plan.js     (150)  纯函数：测量结果 → 有序步骤
-├─ lib/client.js  (1355)  Client 半面：slot、面板、偏好、zh/en 字典、CSS
+├─ lib/client.js  (1379)  Client 半面：slot、面板、偏好、zh/en 字典、CSS
 ├─ locale/{en,zh}.json    展示元数据（meta.title / meta.description）
 ├─ icon.svg               插件图标
 ├─ README.md              本文件
 ├─ LICENSE                MIT
-└─ smoke.mjs              67 条断言，零依赖
+└─ smoke.mjs              68 条断言，零依赖
 ```
 
 Client 半面注册的 slot：`sidebar.panellist`（order 50，仅设置页模式下拆除）、`main`（以包名为 key）、`settings.section`（order 30，仅侧边栏模式下拆除）、`plugins.bundle.config`（以包名为 key）。

package/lib/catalog.js

@@ -274,14 +274,209 @@ const supersededUpdatePackages = () => [
   ...pooledOldVersions(path.join(env('APPDATA') ?? '', 'WorkDaddy', 'update'), /^(.+)-([0-9][0-9._]*)(\.[A-Za-z0-9]+)?$/),
 ];
 
+/** The user profile, which roots every target built from a dot-directory in the home folder. */
+const home = () => env('USERPROFILE') ?? '';
+
+/**
+ * Per-session backup dumps an editor's agent keeps beside a workspace, such as WorkBuddy's
+ * `sessions/<id>/modify_backup`. Each file is a whole-file snapshot ten-odd days old; the
+ * session's live state is elsewhere in the same folder, so the dumps are pure history.
+ */
+const sessionBackups = () => {
+  const sessions = path.join(home(), '.workbuddy', 'workspace', 'sessions');
+  const out = [];
+  for (const entry of listDir(sessions)) {
+    if (!entry.isDirectory() || entry.isSymbolicLink()) continue;
+    const backup = path.join(sessions, entry.name, 'modify_backup');
+    if (isDir(backup)) out.push(backup);
+  }
+  return out;
+};
+
+/** `2026-11-04` and `20261104` both name one day; anything else is not a log archive. */
+const LOG_DAY = /^(\d{4})-?(\d{2})-?(\d{2})$/;
+
+/**
+ * Day-named log archives an app writes and never prunes. The age limit on the target does
+ * the pruning per file, so this only has to name the folders that look like a date.
+ */
+const datedLogDirs = (roots) => {
+  const out = [];
+  for (const root of roots) {
+    for (const entry of listDir(root)) {
+      if (!entry.isDirectory() || entry.isSymbolicLink()) continue;
+      if (LOG_DAY.exec(entry.name) === null) continue;
+      out.push(path.join(root, entry.name));
+    }
+  }
+  return out;
+};
+
+/**
+ * Sibling directories that repeat each other, decided by the content hash the owner writes
+ * into its own metadata file. A snapshot keeper leaves the previous extraction behind after
+ * an upgrade; two of them here were byte-identical. The newest stays, and a sibling with no
+ * metadata proves nothing, so it is never touched.
+ */
+const repeatedSiblings = (parent, metaName, key) => {
+  if (!isDir(parent)) return [];
+  const groups = new Map();
+  for (const entry of listDir(parent)) {
+    if (!entry.isDirectory() || entry.isSymbolicLink()) continue;
+    const abs = path.join(parent, entry.name);
+    let stamp;
+    try {
+      stamp = JSON.parse(fs.readFileSync(path.join(abs, metaName), 'utf8'))[key];
+    } catch {
+      continue;
+    }
+    if (typeof stamp !== 'string' || stamp === '') continue;
+    if (!groups.has(stamp)) groups.set(stamp, []);
+    let mtimeMs = 0;
+    try {
+      mtimeMs = fs.statSync(abs).mtimeMs;
+    } catch {
+      /* unreadable sorts last, so a broken copy is the one dropped */
+    }
+    groups.get(stamp).push({ abs, mtimeMs });
+  }
+  const out = [];
+  for (const list of groups.values()) {
+    if (list.length < 2) continue;
+    list.sort((a, b) => b.mtimeMs - a.mtimeMs);
+    for (const older of list.slice(1)) out.push(older.abs);
+  }
+  return out;
+};
+
+/** Every child directory but the most recently written one; a lone child is left alone. */
+const keepNewestChild = (parent) => {
+  if (!isDir(parent)) return [];
+  const stamped = listDir(parent)
+    .filter((entry) => entry.isDirectory() && !entry.isSymbolicLink())
+    .map((entry) => {
+      const abs = path.join(parent, entry.name);
+      let mtimeMs = 0;
+      try {
+        mtimeMs = fs.statSync(abs).mtimeMs;
+      } catch {
+        /* unreadable sorts last */
+      }
+      return { abs, mtimeMs };
+    })
+    .sort((a, b) => b.mtimeMs - a.mtimeMs);
+  if (stamped.length < 2) return [];
+  return stamped.slice(1).map((item) => item.abs);
+};
+
+/**
+ * An editor that ships a whole runtime per commit leaves the previous one behind after every
+ * update: VS Code kept two complete 764 MB trees plus `old_Code.exe`. Unlike a version holder,
+ * the tree names are commit hashes, so the newest mtime proves nothing — the launcher does.
+ * `bin\code.cmd` names the live hash, and a build is only dropped when that launcher exists,
+ * parses, and points at a different build that is also present. No launcher, no deletion.
+ *
+ * The renamed `old_*` copies are dead weight only while the file they duplicate still sits
+ * beside them; a lone `old_x` may be the only copy left, so it stays.
+ */
+const vscodeSuperseded = () => {
+  const appDir = path.join(env('LOCALAPPDATA') ?? '', 'Programs', 'Microsoft VS Code');
+  if (!isDir(appDir)) return [];
+  const out = [];
+  let live = '';
+  try {
+    const launcher = fs.readFileSync(path.join(appDir, 'bin', 'code.cmd'), 'utf8');
+    const match = /[\\/]([0-9a-f]{10,})[\\/]/.exec(launcher);
+    if (match !== null) live = match[1];
+  } catch {
+    /* no launcher means no evidence about which build runs, so nothing is dropped */
+  }
+  if (live !== '') {
+    const builds = listDir(appDir)
+      .filter((entry) => entry.isDirectory() && !entry.isSymbolicLink())
+      .map((entry) => entry.name)
+      .filter((name) => /^[0-9a-f]{10,}$/.test(name));
+    if (builds.length >= 2 && builds.includes(live)) {
+      for (const name of builds) if (name !== live) out.push(path.join(appDir, name));
+    }
+  }
+  for (const dir of [appDir, path.join(appDir, 'bin')]) {
+    for (const entry of listDir(dir)) {
+      if (!entry.isFile() || entry.isSymbolicLink()) continue;
+      const match = /^old_(.+)$/.exec(entry.name);
+      if (match === null) continue;
+      if (!isFile(path.join(dir, match[1]))) continue;
+      out.push(path.join(dir, entry.name));
+    }
+  }
+  return out;
+};
+
+/**
+ * The Chromium caches an Electron-based agent app writes under its own app folder. The root
+ * is spelled out per app because the sibling `cache` next to it (`~/.workbuddy/cache`) holds
+ * spilled conversation text rather than anything regenerable — it stays.
+ */
+const agentAppCaches = () => {
+  const out = [];
+  for (const name of ['.workbuddy', '.workbuddy-ai']) {
+    const app = path.join(home(), name, 'app');
+    out.push(...subdirs(path.join(app, 'session'), ['Cache', 'Code Cache', 'GPUCache', 'DawnCache', 'ShaderCache']));
+    out.push(...subdirs(app, ['cache', 'CodeCache']));
+  }
+  return out;
+};
+
+/**
+ * Rewind snapshots this harness writes one folder per session, never pruning. The folders are
+ * named by session id, so age is the only signal; the target's 30-day floor keeps a recent
+ * session restorable.
+ */
+const rewindSnapshots = () => {
+  const root = path.join(home(), '.dsh', 'rewind-snapshots');
+  return listDir(root)
+    .filter((entry) => entry.isDirectory() && !entry.isSymbolicLink())
+    .map((entry) => path.join(root, entry.name));
+};
+
+/**
+ * A log file rotated out of the way — `daemon.old.log`, `main.log.1` — as opposed to the
+ * day-named folders the `agent-logs` target covers. Nothing reads a rotated file at runtime.
+ */
+const rotatedLogs = () => {
+  const out = [];
+  for (const name of ['.workbuddy', '.workbuddy-ai', '.qoder', '.zcode', '.codebuddy']) {
+    const logs = path.join(home(), name, 'logs');
+    for (const entry of listDir(logs)) {
+      if (!entry.isFile() || entry.isSymbolicLink()) continue;
+      if (!/\.(old\.log|log\.[0-9]+)$/i.test(entry.name)) continue;
+      out.push(path.join(logs, entry.name));
+    }
+  }
+  return out;
+};
+
+/** The one-generation database copies Zotero writes beside the library it is using. */
+const zoteroBackups = () => {
+  const data = path.join(home(), 'Zotero');
+  return listDir(data)
+    .filter((entry) => entry.isFile() && !entry.isSymbolicLink())
+    .filter((entry) => /\.sqlite[0-9.]*\.bak$/i.test(entry.name))
+    .map((entry) => path.join(data, entry.name));
+};
+
 /**
  * `purge` targets drop every entry under the path. `aged` targets keep anything modified
  * within `minAgeDays`, because those files may belong to a process that is still running.
  *
  * `note` says why a target carries the `caution` risk, in one stable code the UI
  * translates: 'in-use' (a running program may hold these files), 'rebuild' (the owner
- * regenerates them, so the first use afterwards is slower), or 'elevation' (Windows
- * refuses an unelevated delete). Safe targets have none.
+ * regenerates them, so the first use afterwards is slower), 'elevation' (Windows
+ * refuses an unelevated delete), 'pending-update' (a download that may not be applied
+ * yet), 'rollback' (an older version kept beside the live one), 'superseded' (an update
+ * package the installed version already contains), 'history' (a snapshot nothing reads
+ * back), 'duplicate' (a byte-identical copy of a live directory) or 'database-backup' (a
+ * copy of a database file made before an upgrade). Safe targets have none.
  */
 const CATALOG = [
   {
@@ -437,6 +632,79 @@ const CATALOG = [
     roots: ['APPDATA'],
     paths: () => supersededUpdatePackages(),
   },
+  {
+    id: 'session-backups',
+    kind: 'purge',
+    risk: 'caution',
+    note: 'history',
+    roots: ['USERPROFILE'],
+    paths: () => sessionBackups(),
+  },
+  {
+    id: 'agent-logs',
+    kind: 'aged',
+    minAgeDays: 3,
+    risk: 'safe',
+    roots: ['USERPROFILE'],
+    paths: () => datedLogDirs([path.join(home(), '.workbuddy', 'logs'), path.join(home(), '.workbuddy-ai', 'logs')]),
+  },
+  {
+    id: 'toolchain-snapshots',
+    kind: 'purge',
+    risk: 'caution',
+    note: 'duplicate',
+    roots: ['USERPROFILE'],
+    paths: () => repeatedSiblings(path.join(home(), '.dsh', 'dsh-ops', 'bash-ipc'), '.ops-ipc.json', 'patchedSha256'),
+  },
+  {
+    id: 'editor-server-bundles',
+    kind: 'purge',
+    risk: 'caution',
+    note: 'superseded',
+    roots: ['USERPROFILE'],
+    paths: () => keepNewestChild(path.join(home(), 'vscode-remote-wsl', 'stable')),
+  },
+  {
+    id: 'database-backups',
+    kind: 'purge',
+    risk: 'caution',
+    note: 'database-backup',
+    roots: ['USERPROFILE'],
+    paths: () => zoteroBackups(),
+  },
+  {
+    id: 'editor-builds',
+    kind: 'purge',
+    risk: 'caution',
+    note: 'superseded',
+    roots: ['LOCALAPPDATA'],
+    paths: () => vscodeSuperseded(),
+  },
+  {
+    id: 'agent-app-caches',
+    kind: 'purge',
+    risk: 'caution',
+    note: 'rebuild',
+    roots: ['USERPROFILE'],
+    paths: () => agentAppCaches(),
+  },
+  {
+    id: 'rewind-snapshots',
+    kind: 'aged',
+    minAgeDays: 30,
+    risk: 'caution',
+    note: 'history',
+    roots: ['USERPROFILE'],
+    paths: () => rewindSnapshots(),
+  },
+  {
+    id: 'rotated-logs',
+    kind: 'aged',
+    minAgeDays: 3,
+    risk: 'safe',
+    roots: ['USERPROFILE'],
+    paths: () => rotatedLogs(),
+  },
 ];
 
 const byId = new Map(CATALOG.map((entry) => [entry.id, entry]));
@@ -506,10 +774,16 @@ export function resolveTargets(only, drive) {
   });
 }
 
-export async function measurePath(abs) {
+/**
+ * Total the files under `abs`. With a `cutoff`, only files older than it count, so a dry run
+ * reports what a real run would actually delete instead of the whole folder.
+ */
+export async function measurePath(abs, cutoff) {
   if (isFile(abs)) {
     try {
-      return { bytes: fs.statSync(abs).size, files: 1 };
+      const stat = fs.statSync(abs);
+      if (cutoff !== undefined && stat.mtimeMs > cutoff) return { bytes: 0, files: 0 };
+      return { bytes: stat.size, files: 1 };
     } catch {
       return { bytes: 0, files: 0 };
     }
@@ -536,7 +810,9 @@ export async function measurePath(abs) {
         }
         if (!entry.isFile()) continue;
         try {
-          bytes += (await fsp.stat(path.join(dir, entry.name))).size;
+          const stat = await fsp.stat(path.join(dir, entry.name));
+          if (cutoff !== undefined && stat.mtimeMs > cutoff) continue;
+          bytes += stat.size;
           files += 1;
         } catch {
           /* a file that vanished or is locked contributes nothing */
@@ -549,15 +825,20 @@ export async function measurePath(abs) {
   return { bytes, files };
 }
 
-/** Measure every resolved target concurrently, keeping the host event loop free. */
+/**
+ * Measure every resolved target concurrently, keeping the host event loop free. Each target's
+ * own age limit is applied, so the numbers answer "what would a clean free right now" rather
+ * than "how much is in the folder" — the two differ for every `aged` target.
+ */
 export async function scan(only, drive) {
   const targets = resolveTargets(only, drive);
   const measured = await Promise.all(
     targets.map(async (target) => {
+      const cutoff = target.minAgeDays > 0 ? Date.now() - target.minAgeDays * DAY : undefined;
       let bytes = 0;
       let files = 0;
       for (const abs of target.paths) {
-        const size = await measurePath(abs);
+        const size = await measurePath(abs, cutoff);
         bytes += size.bytes;
         files += size.files;
       }
@@ -628,7 +909,7 @@ export async function clean(ids, options = {}) {
       let bytes = 0;
       let files = 0;
       for (const abs of target.paths) {
-        const size = await measurePath(abs);
+        const size = await measurePath(abs, cutoff);
         bytes += size.bytes;
         files += size.files;
       }

package/lib/client.js

@@ -41,6 +41,9 @@ window.__ModuleLoader__.load({
       'why_pending-update': '只清 {n} 天前的下载，避免删掉还没装的更新',
       why_rollback: '每个插件/程序只留版本号最高的一份，其余是回滚副本',
       why_superseded: '版本号不高于已装版本的更新包，装着的那份已经生效',
+      why_history: '每个会话的整文件备份，删了不影响会话本身',
+      why_duplicate: '内容完全相同的重复副本，保留最新那份',
+      'why_database-backup': '升级前的数据库副本，正在用的数据库还在旁边',
       pathCount: '{n} 处',
       driveLabel: '清理范围',
       allDrives: '全部磁盘',
@@ -66,6 +69,15 @@ window.__ModuleLoader__.load({
       'target_app-temp': '各应用临时目录',
       'target_plugin-old-versions': '插件旧版本',
       'target_update-packages': '已应用的更新包',
+      'target_session-backups': '会话备份快照',
+      'target_agent-logs': '按天的日志归档',
+      'target_toolchain-snapshots': '重复的工具链快照',
+      'target_editor-server-bundles': '多余的编辑器服务包',
+      'target_database-backups': '数据库备份副本',
+      'target_editor-builds': '编辑器的旧版构建',
+      'target_agent-app-caches': 'AI 应用的浏览器缓存',
+      'target_rewind-snapshots': '过期的回滚快照',
+      'target_rotated-logs': '轮转掉的日志',
       auditTitle: '空间评估（本插件不会删除）',
       auditHint: '统计缓存之外真正占空间的东西：已装软件、WSL 虚拟磁盘、Windows 组件库，以及你自己的文件。这些一律不由本插件删除，只列出供你判断。',
       auditRun: '开始评估',
@@ -185,6 +197,9 @@ window.__ModuleLoader__.load({
       'why_pending-update': 'only downloads older than {n} day(s), so a pending update survives',
       why_rollback: 'only the highest version of each family stays; the rest are rollback copies',
       why_superseded: 'the update package version is not above what is installed, so it is already applied',
+      why_history: 'a whole-file backup per session; removing it does not touch the session',
+      why_duplicate: 'an identical copy; the newest one is kept',
+      'why_database-backup': 'a copy made before an upgrade; the live database is beside it',
       pathCount: '{n} location(s)',
       driveLabel: 'Limit to',
       allDrives: 'All drives',
@@ -210,6 +225,15 @@ window.__ModuleLoader__.load({
       'target_app-temp': 'App scratch folders',
       'target_plugin-old-versions': 'Old plugin versions',
       'target_update-packages': 'Applied update packages',
+      'target_session-backups': 'Session backup snapshots',
+      'target_agent-logs': 'Dated log archives',
+      'target_toolchain-snapshots': 'Duplicate toolchain snapshots',
+      'target_editor-server-bundles': 'Extra editor server bundles',
+      'target_database-backups': 'Database backup copies',
+      'target_editor-builds': 'Old editor builds',
+      'target_agent-app-caches': 'Browser caches of the AI apps',
+      'target_rewind-snapshots': 'Expired rewind snapshots',
+      'target_rotated-logs': 'Rotated-out log files',
       auditTitle: 'Space assessment (this plugin deletes none of it)',
       auditHint: 'What actually takes up room outside caches: installed programs, WSL disk images, the Windows component store, and your own files. None of it is removed here; it is listed so you can decide.',
       auditRun: 'Run assessment',

package/lib/index.js

@@ -197,6 +197,10 @@ const NOTE_MEANING = {
   elevation: 'Windows refuses an unelevated delete here; run the plugin elevated to reach these bytes',
   rollback: 'an older version the app keeps for rollback; what it runs now is the newest of that family',
   superseded: 'an update package whose version is at or below what is already installed',
+  'pending-update': 'a downloaded update inside the age limit, so it may not be applied yet',
+  history: 'a backup snapshot of file edits that nothing reads back; the live state sits beside it',
+  duplicate: 'a byte-identical copy of a directory that is still present',
+  'database-backup': 'a copy of a database file taken before an upgrade; the live database is beside it',
 };
 
 const formatBytes = (bytes) => {
@@ -274,7 +278,11 @@ export function apply(ctx) {
         description:
           'Inspect and reclaim disk space on this Windows machine from known-safe cache locations only: ' +
           'temp directories, package manager caches (npm, pnpm, pip, uv), crash dumps, browser and editor ' +
-          'caches, and the Windows internet cache. Call with action "overview" to list drives and target ids, ' +
+          'caches, the Windows internet cache, per-app caches, logs and scratch folders under %APPDATA%, ' +
+          'superseded plugin versions and applied update packages, the backup snapshots, dated logs, ' +
+          'rotated-out log files, duplicate toolchain copies and database backups under the user profile, ' +
+          'editor builds their launcher no longer names, and the Chromium caches of the agent apps. ' +
+          'Call with action "overview" to list drives and target ids, ' +
           '"scan" to measure them, and "clean" to delete. clean without confirm: true is a dry run that only ' +
           'reports what it would free, so the user can approve the amount before anything is removed. ' +
           'Action "audit" instead assesses the space that CANNOT be cleaned automatically — installed programs, ' +

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-disk-cleaner",
-  "version": "1.2.0",
+  "version": "1.4.0",
   "type": "module",
   "description": "Disk Cleaner for DSH: scan known-safe cache locations, see what they occupy, and reclaim the space from a panel or from the agent.",
   "keywords": [

package/smoke.mjs

@@ -262,13 +262,19 @@ check('tool description warns audit is read-only', /read-only: it never removes
 check('every effect is labelled', applied.every((label) => typeof label === 'string' && label.startsWith('dsh-disk-cleaner:')));
 
 const ids = [...catalog.matchAll(/^\s{4}id: '([a-z-]+)',$/gm)].map((m) => m[1]);
-check('catalog has the documented 19 targets', ids.length === 19, `got ${ids.length}`);
+check('catalog has the documented 28 targets', ids.length === 28, `got ${ids.length}`);
 check('every catalog id has a client label', ids.every((id) => zhKeys.includes(`target_${id}`) && enKeys.includes(`target_${id}`)));
-check('every note code has an explanation', ['why_safe', 'why_in-use', 'why_rebuild', 'why_elevation', 'why_pending-update', 'why_rollback', 'why_superseded'].every((k) => zhKeys.includes(k) && enKeys.includes(k)));
+check('every note code has an explanation', ['why_safe', 'why_in-use', 'why_rebuild', 'why_elevation', 'why_pending-update', 'why_rollback', 'why_superseded', 'why_history', 'why_duplicate', 'why_database-backup'].every((k) => zhKeys.includes(k) && enKeys.includes(k)));
 const areaIds = [...auditSrc.matchAll(/^\s{6}id: '([a-z-]+)',$/gm)].map((m) => m[1]);
 check('every assessment area has a translated label and reason', areaIds.length > 0 && areaIds.every((id) => zhKeys.includes(`area_${id}`) && zhKeys.includes(`area_${id}_why`) && enKeys.includes(`area_${id}`) && enKeys.includes(`area_${id}_why`)), `areas: ${areaIds.join(',')}`);
 check('a reveal request names an area, never a path', index.includes("action === 'reveal'") && auditSrc.includes('export function areaPaths') && !/revealPath[\s\S]{0,200}payload\.path/.test(index));
 check('no caller may pass a raw path', !/parameters[\s\S]{0,400}paths/.test(index) && catalog.includes('Callers select catalog ids'));
+check(
+  'a dry run reports what a real run would delete, not the folder size',
+  catalog.includes('export async function measurePath(abs, cutoff)') &&
+    (catalog.match(/await measurePath\(abs, cutoff\)/g) ?? []).length === 2,
+  `${(catalog.match(/await measurePath\(abs, cutoff\)/g) ?? []).length} cutoff-aware call sites`,
+);
 
 const codes = [...new Set([...plan.matchAll(/code: '([a-z-]+)'/g)].map((m) => m[1]))];
 check('plan emits the documented 9 step codes', codes.length === 9, `got ${codes.length}`);