disk-clean

English | 简体中文

Windows disk cleanup & analysis CLI — scan, classify, suggest, safely organize (with automatic shortcut fixing) and clean your disks. Zero runtime dependencies, works offline.

⚠️ Safety first: every destructive command is dry-run by default. Nothing is moved or deleted unless you pass --yes. Moves are logged and rollback-able.

🌐 Language notice (please read before reporting a bug). The CLI prints Chinese console output, and the built-in suggestion titles are Chinese. --lang en|zh only switches the structural labels of the report — the Markdown heading, the column/field names and the "Generated / Elapsed / Status" line. It does not translate the CLI's own messages, the suggestion titles, or the category names. Treat this README as the English documentation for a Chinese-language CLI rather than as a description of an English UI; the wording above used to imply the latter. A real translation pass is tracked separately and is not in this release.

What's new in v0.7.2

Useful on a real disk. Classification used to answer "Other - 76.6%" on a real 588 GB drive full of game assets and model weights: the largest category was "unclassified". It is now 11.8%, and the report lists the biggest extensions that still have no rule, so the gap points at itself instead of hiding inside a total.

• Classification rebuilt from real occupancy: game assets (.pak alone: 296.88 GB), AI model weights, build artifacts, mobile packages - plus directory context for names like .bin that carry no meaning on their own.
• The GUI no longer asks for admin on every launch. It never needed it for scan/clean/organize/dedup; the three features that do need it now say so explicitly.
• Errors point somewhere: clean with no type lists the options, junk suggests junk-temp, scna suggests scan. When nothing is close, it stays quiet.
• disk-clean doctor for first runs; report --html for a shareable single-file report; scan C:\ D:\ now confirms the range before traversing minutes of disk.

See docs/RELEASE_NOTES-v0.7.2.md.

What's new in v0.7.1

Trustworthy data. A patch release — no new capabilities, two batches of fixes.

• Tests no longer touch your real data. Twelve of the fifteen suites (as they stood then) read and wrote your actual ~/.disk-clean/: scan reports were overwritten with temp-tree results, the audit log was polluted, and test files were sent to your recycle bin. Fully isolated now, with two guards — a static one (a new suite that forgets to isolate fails the run) and a content-hash snapshot of the real directory taken before and after.
• Reports have a provenance and a history. Generation time and scan range now appear in the CLI, in MCP responses and in the GUI, with a ">24 h — rescan recommended" warning. Every scan is also archived (gzip) to ~/.disk-clean/reports/, keeping the newest N via retention.reports — a config key that existed for several versions with nothing reading it. List them with disk-clean report --history.
• disk-clean build-info prints the version, commit and whether the tree was clean when this binary was built. The release checksums.txt carries the same commit, so a downloaded exe can be checked against the source revision instead of trusted by filename.
• Suggestion lines fixed. They printed [junk-temp] junk-temp because the CLI read label on an object whose field is title. The sample output documented in this README had been correct all along; the code was wrong.
• Two dead config keys removed: they were never read by the engine, yet the MCP tool description promised them as working features. (Which keys, and why: see CHANGELOG — this file documents what the tool can do, so it does not list capabilities that were deleted.)

See docs/RELEASE_NOTES-v0.7.1.md for details.

What's new in v0.6.0

Security hardening — upgrade recommended for all v0.5.x users. An independent review of v0.5.0 found one P0 and several safety defects; all were reproduced before being fixed.

• Hardlink merges no longer accept sampled matches (P0, irreversible data loss). Files over 32MB are compared by head-64KB + tail-64KB only and flagged approx:true — but CLI, MCP and GUI all merged those groups anyway. Reproduced: two 33MB files with identical head/tail and different middles became a hardlink, and the second file's content was gone, with the .dsk-dup-bak backup already deleted. hardlinkGroup() now refuses approx groups outright; exact groups still merge normally, and MCP points the model at disk_clean duplicates (recycle bin, recoverable) instead.
• The safety gate is now genuinely single-source. README claimed the protected-path list lived only in lib/guard.js; four more copies were alive (each with 7 segments instead of 16, missing windows.old, boot, efi, recovery and 5 more), and the copy in organize.js also lacked the trailing match — so C:\Users\me\windows was refused by clean and allowed by organize.
• Hardlink merging consults the gate, and a scan root that is itself protected (C:\Windows\System32\drivers) is now refused instead of walked.
• Organize destinations can no longer traverse: C:\整理区\..\boot\x used to pass a prefix regex and resolve to C:\boot\x.
• Partial cleans no longer report success. If 3 of 5 items are locked or denied, the tool now returns partial:true and the MCP layer raises isError, so a model cannot announce "cleanup complete".
• dedup-map.json has one schema (MCP wrote entries, CLI/GUI wrote merged — so files merged by the AI could not be rolled back elsewhere); recycle-bin matching uses canonKey on both paths (GUI could not list its own short-name cleanups); rollback maps are written atomically; malformed URL escapes return 400 instead of throwing.
• Release integrity: a stale checksums.txt from CI's own build used to ship alongside the local exe — v0.5.0 shipped sha256=72ce9f21… for an exe hashing 3cbdc188…. The publish script now recomputes, uploads and downloads back to verify.

New test/safety-gates.js guards every item above — including an end-to-end proof with real >32MB files. The authoritative suite list is test/all.js (never restate a suite count here: test/docs-consistency.js fails the build when a number in this file disagrees with the code).

What's new in v0.5.0

• AI integration moved to MCP: a built-in MCP server (disk-clean mcp, stdio, zero dependencies) exposes 12 disk tools to any MCP client — DeepSeek Harness, Claude Desktop, Cursor. The previous DSH-only plugin form (~3400 lines of DSH-specific code that could never render its panel in the shipped build) is gone.
• 12 MCP tools: disk_scan / disk_report / disk_drives / disk_clean / disk_organize / disk_dedup / disk_health / disk_quota / disk_mftscan / disk_audit / disk_recycle / disk_config, each carrying MCP annotations (readOnlyHint / destructiveHint). Destructive tools are dry-run by default and require an explicit confirm:true.
• One safety gate, one source: new lib/guard.js (protected paths / OneDrive / scan-root membership) replaces the copies in clean.js, organize.js and serve.js. Protected segments grew to 16 (adds Windows.old, $Windows.~BT, Recovery, PerfLogs, MSOCache, Config.Msi, Boot, EFI).
• One version source: new lib/version.js. The version used to live in 6 places; now one is writable and the rest derive from it, bump-version.js self-verifies, and test/version-consistency.js guards against drift.
• Real defects fixed: a 0-item clean no longer reports success; emptying the recycle bin reports the true item count and flags it irreversible; hardlink rollback is now an atomic sequence (no whole-file reads, no delete-then-write); static file serving uses path.relative for containment; query-string tokens are rejected by default; temp detection is exact-segment (temporary-report is no longer mistaken for a temp dir).

What's new in v0.4.1

• OneDrive cloud-safe: \OneDrive\ files are counted but excluded from dedup hash and destructive suggestions; report/UI shows "OneDrive cloud files excluded (avoids silent download/delete)".
• Bulk low-risk clean: Clean Center → "One-click clean (low-risk)" previews junk-temp + empty-dirs together, single confirm; medium/high-risk types still require per-card confirm.
• Health trend: /api/health-check returns trend sparklines (last 20) with 60s history throttle + 30s cache; UI shows Temp/Wear/Error trends.
• Recycle restore (audit-matched): GET /api/recycle/list / POST /api/recycle/restore + Clean Center "Recycle restore" (scope A).
• Engine-core single source: lib/engine-core.js is the sole editable core; lib/engine.js is a thin wrapper, plugin shims hybrid-require it.

Why disk-clean?

Install

Option A — single EXE (recommended)

Download disk-clean-win-x64.exe from Releases — no Node.js required.

.\disk-clean-win-x64.exe scan D:\

Option B — via Node.js (>= 18.15)

npm install -g disk-clean    # or: git clone + npm link
disk-clean scan D:\

The npm channel was stuck on 0.4.1 (which predates both the MCP server and the v0.6.0 security fixes) from 2026-08-25 until 0.6.0 was published on 2026-09-11. If you installed before then, upgrade: npm install -g disk-clean@latest.

Option C — MCP server (AI-driven, works with any client)

disk-clean mcp starts a standard MCP (Model Context Protocol) server that exposes 12 disk tools over stdio. Zero dependencies — no SDK, no build step.

// DeepSeek Harness / Claude Desktop / Cursor — same config everywhere
{
  "mcpServers": {
    "disk-clean": { "command": "npx", "args": ["-y", "disk-clean", "mcp"] }
  }
}

Then just ask:

"D: is nearly full — scan it, tell me what's safe to remove, then clean up temp files and empty folders."

The typical call chain is disk_drives (capacity) → disk_scan (build a report) → disk_clean (dry-run preview, then confirm:true). Destructive tools only preview unless explicitly confirmed; system paths, OneDrive folders and anything outside the scanned roots are refused outright. For local development use node bin/disk-clean-mcp.js directly.

- name: '@deepseek-ai/dsh-mcp-client'
  config:
    serverName: disk-clean
    transport: stdio
    command: node
    args: ['D:\\deepseekHerness\\disk-clean-cli\\bin\\disk-clean-mcp.js']

Option D — Native GUI (WebView2 window)

Download disk-clean-setup-<ver>.exe from Releases — a native desktop window (WinForms + WebView2) with a bilingual (zh/en) dashboard: one-click scan, suggestions, cleanup, plus advanced tabs (organize / health / dedup / quota / MFT / schedule / config / audit). The installer detects and bootstraps .NET 8 Desktop Runtime and WebView2 when missing. See docs/GUI-PLAN.md and README.zh-CN.md.

Quick start

# 1. Scan a drive (report: JSON + Markdown)
disk-clean scan C:\ D:\

# 2. Read the report (terminal summary + Markdown render)
disk-clean report

# 3. Generate an organize plan (loose dirs → <drive>:\整理区\<category>\)
disk-clean organize plan

# ...also include program/game dirs (⚠ moves will rewrite shortcuts)
disk-clean organize plan --include-program

# 4. Preview, then actually move (dry-run by default; --yes to run)
disk-clean organize apply
disk-clean organize apply --yes

# 5. Undo the last batch (moves + shortcuts restored)
disk-clean organize rollback --yes

# 6. Clean junk (temp / empty dirs / duplicates / recycle bin)
disk-clean clean empty-dirs              # preview
disk-clean clean empty-dirs --yes        # execute (moves to recycle bin)

# 7. Fix broken shortcuts manually
disk-clean fix-shortcuts pairs.json

# 8. View the audit log
disk-clean audit

# 9. Fast MFT scan (needs admin, ~8x faster than traversal)
disk-clean mftscan D:

# 10. Full-disk duplicate detection (excludes system/program dirs)
disk-clean dedup D:\

# 11. Merge duplicates into hardlinks to free space (rollback-able)
disk-clean dedup D:\ --hardlink --yes
disk-clean dedup rollback

# 12. Per-user quota analysis (needs admin)
disk-clean quota C:

# 13. SMART / SSD health check
disk-clean health

# 14. Create a system restore point before destructive ops
disk-clean organize apply --yes --restore-point

# 15. English report
disk-clean scan D:\ --lang en

# 16. Start the MCP server (for AI clients; stdio, long-running)
disk-clean mcp

Commands

Safety model

• Dry-run by default — every destructive command prints what it would do; pass --yes to actually run.
• Recycle bin — junk/empty/duplicate items are moved to the recycle bin, not permanently deleted.
• Rollback — directory moves append to organize-map.json; organize rollback restores the last batch (including shortcuts).
• Protected paths — 16 protected segments are always refused: \windows\, \windows.old\, \program files\, \program files (x86)\, \programdata\, \winsxs\, \system volume information\, \$recycle.bin\, \$windows.~bt\, \$windows.~ws\, \recovery\, \perflogs\, \msocache\, \config.msi\, \boot\, \efi\. The list exists in exactly one place (lib/guard.js), and test/docs-consistency.js fails if this enumeration and that list stop matching.
• OneDrive cloud sync — paths containing a \OneDrive\ segment are refused for destructive operations too (deleting syncs to the cloud; hashing triggers silent placeholder downloads).
• Audit log — every action is appended to ~/.disk-clean/audit.jsonl (time / type / paths / result / real item counts).
• Exit codes — 0 ok · 1 user cancel/args · 2 runtime error · 3 scan cancelled.

State files live in ~/.disk-clean/:

audit.jsonl            # audit log
report.json / .md      # latest report
reports/               # report history (one gzip archive per scan, newest N kept)
organize-map.json      # rollback mapping
organize-plan.json     # last plan

Example output

▶ 正在扫描: D:\
✔ 扫描完成  (43.1s)
  总大小   : 630.2 GB   文件: 1381931   目录: 196460
  报告     : C:\Users\Administrator\.disk-clean\report.json
  Markdown : C:\Users\Administrator\.disk-clean\report.md

── 智能建议 ──
  [organize-folders] 目录整理建议 — 13 项 (164.0 GB)
  [stale-large] 清理长期未使用的大文件 — 14 项 (15.4 GB)
  [duplicates] 重复文件 — 46 组 (可释放约 110 MB)
  [recycle-bin] 清空回收站 — 67 MB
  [empty-dirs] 删除空文件夹 — 4490 项

▶ MFT 直读扫描: D:
✔ 扫描完成  (5.5s)   ← ~8x faster than traversal
  D:  总大小: 532 GB  文件: 1362274  目录: 196450

▶ 全盘重复检测: D:\
✔ 扫描完成  (35.1s)  重复组: 1503  可释放: 17.3 GB

▶ 配额分析: C:
  administrator   1.1 TB  (74.9%)
      ├ Desktop  77.8 GB
      ├ AppData  875 GB

See docs/demo-report.md for a full Markdown report sample.

Development

npm run check     # syntax check all modules
npm test          # full suite (19 suites, incl. MCP protocol + rules integrity + safety gates + docs consistency)
npm run mcp       # start an MCP server locally
powershell -File scripts\build-sea.ps1   # build exe + sha256 (adds the build fingerprint)

• Engine: lib/engine-core.js (the only editable core) → lib/engine.js (thin wrapper) — zero-dependency Node (native fs), PowerShell used only for COM shortcut fixing.
• Safety gate: lib/guard.js (single source). Version: lib/version.js (single source).
• MCP: lib/mcp/server.js (protocol core, zero-dependency) + lib/mcp/tools.js (12 tools, all thin wrappers over lib/*).
• The CLI and the MCP server are both thin shells — there is no second copy of the business logic.

Reusable process

See docs/RELEASE-PLAYBOOK.md — the step-by-step SOP for building the CLI/GUI, publishing to GitHub and npm, and shipping Release assets (reuse for new projects). GUI specifics live in the gui-development skill; release specifics in release-sop.

Roadmap

See ROADMAP.md — 12 phases: config rules, scheduled scans, MFT fast scan, SMART health, full-disk dedup, per-user quota, restore points, i18n.

License

MIT