SlothSort: I built a tiny agent harness from scratch, then taught it to tidy my Downloads folder so I never have to

Hitesh Chouhan — Sun, 25 Oct 2026 08:42:11 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 3 — theme: **Slow and Steady.

What I built

Two things, really:

1. A small agent harness, written from scratch. About 350 lines of standard-library Python, zero dependencies. It runs a model↔tools loop over a local open-weight model: the model never touches the filesystem — it can only reply with JSON tool calls (list_files, classify_file, finish_plan, apply_plan, undo_last_run). The harness validates every call, executes it, feeds the result back, retries malformed replies with feedback, and writes a full session trace (JSONL) of every model decision, tool call, refusal and fallback. If the model fumbles a step, the harness says so in the trace — nothing is silently faked.

2. SlothSort, the harness's job: a file tidier that rewards laziness. Point it at a messy folder. The agent lists the files, classifies each one into a category with a cleaner filename and a one-line reason, and shows you a plan. It is the "Slow and Steady" deal: the sloth does the filing, slowly and carefully; you keep your energy.

python demo/make_mess.py                 # a folder with 12 sad files
python -m slothsort demo/messy           # dry run: plan only
python -m slothsort demo/messy --apply   # confirmed? then it moves
python -m slothsort undo demo/messy      # regret? fully reversible

The brain is Gemma 3 (gemma3:1b), Google's open-weight model, served locally by Ollama. The harness is model-agnostic — Gemma is a component, not a landlord (more on that below).

How it saves effort (the lazy payoff)

I never sort a Downloads folder by hand again. The agent does the boring classification pass — including renaming invoice_final_FINAL_v3.pdf into something a human would choose.
It is careful precisely because I am lazy: dry-run by default, an interactive confirmation before anything moves, never overwrite (collisions get a -1 suffix), and an undo log that reverses the last run exactly. A lazy tool you can't trust is just more work.
It skips hidden files and its own logs. Even its mistakes are auditable through the trace.

Demo

Repo: https://github.com/hiteshchouhan22/slothsort
Live demo (GitHub Actions, one click, real model): the repo ships a demo.yml workflow that installs Ollama, pulls gemma3:1b, runs the agent against the sample messy folder — plan, apply, undo — and uploads the full transcript and session trace as artifacts. No staging, no mocked model.
A local run transcript and instructions are also in the repo under demo/ (https://github.com/hiteshchouhan22/slothsort).

Why open innovation matters here

This project only makes sense with open-weight models:

Privacy by architecture. Filenames are personal — tax returns, ID scans, salary letters. With a local open model, they never leave the machine. A cloud API would make the whole tool a bad idea.
Offline and free forever. No API key, no per-token bill, no rate limits. It runs on a plane.
Swappable models. Because Gemma is open weight and served behind a plain local API, I can swap in qwen2.5:0.5b or llama3.2:1b with one environment variable and the harness doesn't change a line. Open weights turn "the model" into a replaceable part — which is exactly what made a from-scratch harness worth building: the loop, the tools, the safety gates and the trace are mine; the brain is interchangeable.
Small models are enough — if the harness is honest. A 1B model will not reliably one-shot a 40-file JSON plan. But guided one file at a time, with validation and retries, it does the job. That gap between "demo-ware" and "works on my machine" is where open tooling lets you experiment for free until it clicks.

Prize categories

I'm entering:

Best Use of Gemma — Gemma 3 (gemma3:1b, open weight, via Ollama) is SlothSort's default and demoed model: it makes every classification decision inside the harness's tool loop, and the full session trace of its decisions ships with the demo.

Stack

Python 3 (standard library only), Ollama, Gemma 3. Tests: 17, using scripted fake model clients — plan parsing, collision handling, dry-run safety, harness apply-gate refusal, and apply+undo round-trips, all green without needing a model installed.

— Hitesh Chouhan