Four jobs on your laptop: ask, write a test, fix a bug, add
one small function. Only touches the folder you point at. Daily work
is py-harness plus Ollama llama3.1:8b. Not everyday-ready.
Site: yauhenbichel.github.io/py-harness.
python3 -m venv .venv && source .venv/bin/activate
pip install py-harness-cliThe command is py-harness. The PyPI name is py-harness-cli because
py-harness collides with another package. Do not pip install py-harness or pip install pyharness.
Sample project — clone, then cd demo/orders. Do not brief this
repository.
git clone https://github.com/YauhenBichel/py-harness.git
cd py-harness/demo/orders
py-harness brief
py-harness ask "what does compute_total return?"
py-harness run "find the NameError and fix it"Activate .venv in every new terminal. If the shell says
command not found: py-harness, it is not active. Daily work needs
ollama pull llama3.1:8b. From a clone,
python3 scripts/run/install.py is the editable install.
ask never writes. run writes, then runs the tests. The NameError
sample is built into the tool (no model). Another folder:
py-harness ask ~/app "what does add return?".
Start · Commands · Contributing · Security
Replay: asciinema play docs/media/pip-demo.cast.
A longer session (8B ask): docs/media/live-demo.gif.
Full log: Live demo.
VS Code — py-harness editors vscode, then Tasks: Run Task.
Replay: asciinema play docs/media/vscode-demo.cast.
VS Code.
Cursor — py-harness editors cursor --allow-writes, then chat or Tasks.
Replay: asciinema play docs/media/cursor-demo.cast.
Cursor.
One laptop. 29 Aug–5 Sep 2026. Not everyday-ready.
| What I tried | Result |
|---|---|
| 0.5B as daily work | 0 / 4 vibe, 0 / 2 parse |
Four Start commands on demo/orders |
0 / 4, then 4 / 4 after the harness |
| Same bench, code must run | 8B 6–9 / 9; 7B coder 7 / 9; 30B timeout |
| Same-night daily (write tests, clamp, sum × 3) | 8B 9 / 9, 7B coder 7 / 9. Keep the 8B |
| Extra 7B–8B on disk | 8192-warmed SWE daily clamp: first generate 180s. Not a score |
| A real repository (4,580 files) | reading works; writing 1 / 12 |
The long table: Experiments. Every score: Results. Which tags timed out: Hub models.
| If you want | Go here |
|---|---|
| What each folder is | Folders |
| Tests | PYTHONPATH=src python -m unittest discover -s tests -q |
| 0.5B style prior | YauhenBichel/python-vibe-0.5b |
| Train / serve / API | Commands |
| A first issue | good first issue |
Vulnerabilities: a public GitHub issue. Do not paste live keys.
Thank you to everyone who has helped py-harness.
Filled from GitHub commits (bots omitted). Contributor graph · good first issue · Action