REQUIREMENTS

Architecture

Item	Requirement
CPU architecture	x86_64 / amd64 (`docker-env/Dockerfile` and language toolchains target linux-amd64)
ARM / Apple Silicon	Not officially packaged
GPU	Not required

Use case	Recommendation
Inspecting Zenodo `results.zip` / `results.xlsx` and running analysis scripts (`ablation.py`, `cost.py`, `run_test_comparison.sh`)	≥ 4 CPU cores, ≥ 8 GB RAM, ≥ 20 GB free disk
Building and running the Docker image (`scripts/start_docker.sh`)	≥ 8 CPU cores, ≥ 16 GB RAM, ≥ 40 GB free disk
Re-running all paper experiments (118 projects)	Multi-day wall-clock time and non-trivial LLM cost

Docker (Engine 24+ recommended) or a Podman-compatible Docker CLI — primary way to run ReCodeAgent (see README.md § Getting Started)
Network access during docker build to fetch language toolchains
Linux host recommended inside/alongside Docker; macOS or Windows with WSL2 may work for host-side scripts
Python 3.10+ (3.12 recommended) for analysis scripts in src/analysis/
unzip (or equivalent) to unpack Zenodo results.zip
Optional host packages for analysis plots / RQ2 similarity: matplotlib, numpy, scipy, seaborn, pandas (and scikit-learn / sentence-transformers if embedding similarity is enabled)

Running agents requires an LLM backend configured by the user (keys are not included in this artifact):

AWS Bedrock + Claude (paper setup): aws configure with access key, secret, and region; Claude model enabled in that region (awscli)
OpenRouter or other providers: set the env block in .claude/settings.local.json (see README.md § Credentials)

Published experiment outputs are also available on Zenodo without re-running agents: https://doi.org/10.5281/zenodo.19214481.