mini-SWE-agent

mini-SWE-agent is a lightweight software-engineering agent implemented in ~100 lines of Python. It’s designed to solve real GitHub issues and help with command-line coding tasks while remaining easy to read, extend, and research.

Key design choices keep the scaffold minimal and transparent: the agent issues only bash commands (no tool-calling APIs), executes each action with subprocess.run (no persistent shell state), and maintains a strictly linear message history for straightforward debugging and fine-tuning. The project ships with a simple CLI, a richer terminal UI (mini -v), batch runners for evaluation (e.g., SWE-bench), and an Inspector to browse .traj.json runs.

Models are pluggable through LiteLLM (and a direct Anthropic adapter), so you can use OpenAI, Anthropic, and many other providers—or run local models. Environments include local execution and sandboxed options (Docker, Podman, Singularity/Apptainer, Bubblewrap, and SWe-ReX). Installation is a single pip/pipx/uv command.

The result is a small, hackable agent that’s practical for day-to-day development and robust for reproducible research and benchmarking.