Reason Machines examples

Run Reason on DeepSWE coding tasks in Docker.

Setup

On macOS or Linux, install Docker and uv, and start Docker. You also need Git and curl. Run these commands in the same terminal:

git clone https://github.com/reason-machines/examples.git
cd examples
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python \
  https://github.com/reason-machines/examples/releases/download/v0.3.9/reason_machines_pier.tar.gz

mkdir -p deep-swe
curl -LsSf https://github.com/datacurve-ai/deep-swe/archive/0b9fabbb63b9104d678fe965e1632f2dd9eaa2ea.tar.gz \
  | tar -xz --strip-components=1 -C deep-swe

This installs Pier. Before each production run, the adapter resolves the current stable Reason CLI release, verifies its published checksum, and installs it inside Docker automatically.

Run

Get a Reason API key and enable openai/gpt-5.6-luna in your workspace's Models settings. Replace the placeholder with your key, then run:

export REASON_API_KEY="YOUR_REASON_API_KEY"
DOCKER_DEFAULT_PLATFORM=linux/amd64 .venv/bin/pier run \
  --path deep-swe/tasks/csstree-shorthand-expansion-compression \
  --agent-import-path reasonmachines_pier:ReasonAgent \
  --model openai/gpt-5.6-luna \
  --jobs-dir outputs/pier

This runs one task in Docker, including on Apple Silicon. Model/API charges may apply.

Allow up to 3 hours for the agent, plus 30 minutes for verification. It can finish sooner. The first run also downloads several GB of Docker images. Keep your terminal and Docker running, and your computer awake.

Results and logs are saved in outputs/pier/. A completed run can still fail tests.

More tasks and API-key permissions.