SlothSort 以離線 AI 輔助分類及整理檔案
SlothSort 🦥 — An Offline AI File Organizer That Rewards Laziness
D9292S 建立 SlothSort,一款透過 Ollama 在本機運行 llama3.2 的 CLI 工具,可按檔案副檔名及文字內容分類、整理檔名並移入分類資料夾。工具預設提供 dry-run 預覽,並支援 --watch、以 SHA-256 偵測重複檔案的 --dedupe,以及 slothsort undo 回復操作。完整程式碼、文件及自動化測試套件已在 GitHub 開源。
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
What I Built
Most developers have a Downloads folder that looks like a digital landfill: dozens of document(1).pdf, Screenshot 2026-10-21.png, and camera uploads like IMG_20241019_142356.jpg.
Organizing them manually is tedious, and simple extension-moving scripts fail because they can't understand what a file actually contains.
In honor of International Sloth Day and this week's theme — Slow and Steady: build something with open-source AI that rewards laziness — I built SlothSort.
SlothSort is a lightweight CLI that uses a local open-weight model (llama3.2) via Ollama to scan messy folders, understand file context, clean up names into readable kebab-case, and sort everything into category folders.
Core Features
- 🧠 Contextual Sorting: Classifies files using both extensions and text header previews (peeking at the first few lines of text/CSV/code).
- 🏷️ Clean Renaming: Standardizes messy timestamps and downloads (
Screenshot 2026-10-21 000814.png→screenshot-2026-10-21-000814.png). - 🦥 True Sloth Mode (
--watch): Runs quietly in the background, tidying up new downloads automatically while you take a nap. - 📑 Deduplication (
--dedupe): Uses fast SHA-256 matching to isolate duplicate downloads (invoice (1).pdf) into aDuplicates/folder. - 🛡️ Defensive by Default: Safe dry-run previews by default, collision protection (
-1,-2), and instant 1-command rollback (slothsort undo).
Demo
Here is a live terminal session running against a messy directory:
$ slothsort ~/Downloads --dedupe
┌──────────────── 🦥 slothsort v1.0.0 ────────────────┐
│ 📂 /Users/deepa/Downloads │
│ ⚙️ DRY-RUN Model: llama3.2 • Offline Inference │
└─────────────────────────────────────────────────────┘
Taxonomy Breakdown
┌────────────────────┬──────────┬────────────┬────────────┬────────────────┐
│ Category │ Files │ Size │ Renamed │ Inference │
├────────────────────┼──────────┼────────────┼────────────┼────────────────┤
│ 📦 Archives │ 3 │ 45.0 MB │ 1 │ 1/3 local │
│ 🎵 Audio │ 2 │ 18.5 MB │ 2 │ rules │
│ 💻 Code │ 5 │ 0.2 MB │ 2 │ rules │
│ 📊 Data │ 3 │ 1.1 MB │ 3 │ rules │
│ 🎨 Design │ 2 │ 24.0 MB │ 2 │ rules │
│ 📄 Documents │ 7 │ 2.8 MB │ 6 │ 1/7 local │
│ 📑 Duplicates │ 2 │ 1.4 MB │ 0 │ dedupe │
│ 🖼️ Images │ 4 │ 6.9 MB │ 4 │ rules │
│ ⚙️ Installers │ 2 │ 157.9 MB │ 2 │ rules │
│ 🎬 Videos │ 1 │ 240.0 MB │ 1 │ 1/1 local │
├────────────────────┼──────────┼────────────┼────────────┼────────────────┤
│ Total │ 31 │ 497.8 MB │ 23 │ — │
└────────────────────┴──────────┴────────────┴────────────┴────────────────┘
┌─ Proposed Directory Hierarchy ──────────────────────────────────────────────┐
│ 📁 Downloads/ │
│ ├── 📄 Documents/ (7 files) │
│ │ ├── analytics_october_metrics.csv → analytics-october-metrics.csv • │
│ │ │ Financial data report │
│ │ ├── budget_forecast_q4.xlsx → budget-forecast-q4.xlsx • standardized │
│ │ │ kebab-case │
│ │ ├── final_report_FINAL_v3_USETHISONE.docx → │
│ │ │ final-report-final-v3-usethisone.docx • standardized kebab-case │
│ │ └── meeting_notes_oct.txt → meeting-notes-oct.txt • standardized │
│ │ kebab-case │
│ ├── 📑 Duplicates/ (2 files) │
│ │ ├── document(2).pdf → Duplicates/document(2).pdf • duplicate of │
│ │ │ document(1).pdf │
│ │ └── IMG_20241020_083211.jpg → Duplicates/IMG_20241020_083211.jpg • │
│ │ duplicate of IMG_20241019_142356.jpg │
│ ├── 🖼️ Images/ (4 files) │
│ │ ├── Screenshot 2026-10-21 000814.png → │
│ │ │ screenshot-2026-10-21-000814.png • standardized kebab-case │
│ │ └── wallpaper_4k_mountain.jpg → wallpaper-4k-mountain.jpg • │
│ │ standardized kebab-case │
│ └── ⚙️ Installers/ (2 files) │
│ ├── ABDownloadManager_1.11.1_windows_x64.exe → │
│ │ abdownloadmanager-1.11.1-windows-x64.exe • standardized kebab-case │
│ └── VSCodeUserSetup-x64-1.92.0.exe → vscodeusersetup-x64-1.92.0.exe • │
│ standardized kebab-case │
└─────────────────────────────────────────────────────────────────────────────┘
┌─ Next Step ─────────────────────────────────────────────────────────────────┐
│ Dry-run preview complete. No files were modified on disk. │
│ │
│ To apply these changes, run: │
│ slothsort "~/Downloads" --execute --dedupe │
└─────────────────────────────────────────────────────────────────────────────┘
When you pass --execute, the moves happen atomically. If you ever change your mind:
$ slothsort undo ~/Downloads
┌─ Rollback Complete ─────────────────────────────────────────────────────────┐
│ ✓ Successfully restored 31 file(s) to original locations. │
└─────────────────────────────────────────────────────────────────────────────┘
Code
The complete codebase, documentation, and automated test suite are open-sourced on GitHub:
👉 GitHub Repository: D9292S/slothsort
slothsort/
├── pyproject.toml # Modern Hatchling package configuration
├── README.md # Full documentation and command reference
├── LICENSE # MIT License
├── .slothsort.toml.example # Configuration template for user overrides
├── slothsort/
│ ├── __init__.py # Version metadata
│ ├── config.py # Taxonomy mapping, model configurations, defaults
│ ├── scanner.py # Filtering, snippet extraction, SHA-256 duplicate hashing
│ ├── categorizer.py # Ollama JSON inference + Algorithmic Hallucination Guard
│ ├── organizer.py # Collision resolution (-1, -2), moves, atomic undo manifest
│ └── cli.py # Rich terminal UI, diff tree, watch mode, commands
├── demo/
│ └── simulate_mess.py # Demo script to generate 31 realistic messy files
└── tests/
├── test_scanner.py # Filtering, snippets, duplicate detection tests
├── test_safety.py # Hallucination guard & name preservation tests
├── test_parser.py # Resilient JSON parsing tests (lists, dicts, markdown)
└── test_organizer.py # Collision resolution & end-to-end rollback tests
How I Built It
SlothSort is built with Python 3.10+, Ollama, Meta Llama 3.2 (3B), Click, Rich, and Pydantic.
Technical Highlights
- Local Inference: Files are batched and sent to Ollama with strict JSON schemas enforced by Pydantic.
- Content Sniffing: For ambiguous text/code/data files under 2 MB, the scanner peeks at the first 300 characters to provide the model with grounding context.
-
The Hallucination Guard (
is_safe_rename): During testing, I caught a 3B model edge case: when given a batch containing multiple installers, Llama 3.2 occasionally cross-pollinated names (trying to renameABDownloadManager.exetobitdefender.exe). I wrote an algorithmic safety guard that verifies suggested names never borrow tokens from other files in the batch, falling back to safe formatting if cross-contamination is detected. -
Safety & Testing: Includes an atomic rollback log (
.slothsort-undo.json) and 12 unit tests inpytestverifying parsing, collision handling, and hallucination rejection.
Why Does Open Innovation Matter?
For personal file management, local open-source AI is fundamentally superior to closed cloud APIs:
- 🔒 Absolute Privacy: Your Downloads folder contains tax drafts, bank statements, and personal records. With local models, zero bytes leave your machine.
- 💸 Zero Recurring Cost: Organizing thousands of files costs \$0.00 in API tokens.
- ✈️ 100% Offline: Operates in airplane mode, on trains, or without internet.
- 🎛️ Model Portability: Swap models freely based on your hardware (
--model llama3.2,--model phi3, or--model mistral). - 🛡️ Resilient Fallback: If Ollama is offline or files exceed 500 MB, SlothSort gracefully falls back to deterministic rules. It never crashes.
My Agent Session
I built SlothSort through an iterative agent pair-programming session with an AI coding assistant. The agent helped scaffold the initial Click/Rich CLI architecture, iterate on small-model prompt engineering with Ollama, and write the 12-test pytest suite that verified the algorithmic hallucination guard.
The Lazy Payoff
Manually sorting an 800-file Downloads folder usually takes 45 minutes of tedious drag-and-drop.
With SlothSort, it takes 5 seconds with one command — or zero effort if you leave --watch running in the background.
來源:DEV Community · dev.to