swesh | SWE-agent shell

Live-SWE-agent

mini-swe-agent

ℹ️ About | ✨ Features | 🚀 Setup | 💻 Usage | 📊 Comparison | 📜 Attribution

ℹ️ About

swesh is an interactive CLI evolution of Live-SWE-agent and mini-swe-agent.

It combines the "100 lines of Python" philosophy of mini with the self-evolving capabilities of live, while adding robust features for daily development.

Why swesh?

1. Daemonless Execution: Run secure, isolated environments using Podman without needing root access or a Docker daemon.
2. Robust Local AI: Optimized for open-source models (e.g., Mistral, OpenRouter) with built-in message cleaning and fault-tolerant cost tracking.
3. Interactive Shell: Enhanced generic agent shell that supports ! commands for direct shell escape interactions.
4. Batch & Interactive: Supports both single-task interactive runs and full SWE-bench evaluations out of the box.

⚠️ Security Warning: swesh executes bash commands with your user privileges. Always review commands before approving in confirm mode. For untrusted tasks, use --sandbox mode. See docs/security.md for details.

---

✨ Features

• Interactive REPL w/ Shell Escape:
  • Inputs starting with ! execute directly via subprocess and re-prompt the agent.
  • Full history preservation for debugging.
• Native Podman Support:
  • Explicit --environment-class podman support.
  • Works with rootless containers and user sockets.
• SWE-bench Verified:
  • Fully compatible with the official evaluation harness.
  • Includes specific fixes (timeouts, API compatibility) for running benchmarks on consumer hardware/APIs.
• Generic Model Support:
  • Works with any OpenAI-compatible API (Mistral, Anthropic, vLLM, etc.).
  • RobustLitellmModel handles API quirks automatically.

---

🚀 Setup

Python version: Requires Python 3.14 (this repo currently targets 3.14-only; other versions are unsupported).

swesh offers flexible installation options similar to mini-swe-agent.

1. Installation

Option 1: Local Install (Recommended)

# Install with uv (fastest)
uv sync

Option 2: Developer Setup

git clone https://github.com/elusznik/swesh.git
cd swesh
./sync_remotes.sh   # Adds upstream remotes (live-swe-agent, mini-swe-agent)
uv sync

Note: swesh tracks two upstream projects. Run ./sync_remotes.sh periodically to pull latest changes from both.

2. Configure Podman (Recommended)

swesh is built for Podman. Ensure you have the socket enabled for compatibility:

# Install podman
sudo apt install podman

# Enable the socket
systemctl --user enable --now podman.socket

# Export the host variable (add to ~/.bashrc)
export DOCKER_HOST=unix:///run/user/$(id -u)/podman/podman.sock

3. Set API Keys

Configure your .env file (~/.config/mini-swe-agent/.env):

MISTRAL_API_KEY=...
OPENAI_API_KEY=...
ANTHROPIC_API_KEY=...
# Check https://docs.litellm.ai/docs/providers for other providers

See Global Configuration for more options.

---

💻 Usage

1. Interactive Mode

Run swesh (or mini) to start an interactive session.

swesh --model mistral/codestral-latest --task "Fix the bug in src/utils.py"

Interactive Commands:

• !ls -la: Run a shell command directly.
• exit: valid exit command.

2. Visual Mode

swesh defaults to an opencode-style Textual TUI (continuous transcript + bottom input bar).

To use the legacy pager-style Textual UI (the same as mini -v), run:

swesh -v

To override the opencode UI styling, set MSWEA_OPENCODE_STYLE_PATH to a .tcss file.

3. Live-SWE-agent Self-Evolution

Enable the self-evolving features by using the Live-SWE-agent config:

mini --config config/livesweagent.yaml --model mistral/codestral-latest --task "Refactor the memory module"

4. SWE-bench Evaluation

Run full benchmarks on the SWE-bench Verified dataset using Podman:

uv run src/minisweagent/run/extra/swebench.py \
  --config config/livesweagent_swebench.yaml \
  --subset verified \
  --split test \
  --model mistral/devstral-2512 \
  --slice 0:1 \
  --environment-class podman \
  --workers 1

See docs/guides/running_swebench.md for a full guide.

---

📊 Comparison

Live-SWE-agent (the basis for swesh) is the current state-of-the-art on SWE-bench Verified.

Claude Opus 4.5 + Live-SWE-agent scores 79.2%, leading all open-source scaffolds. Gemini 3 Pro + Live-SWE-agent scores 77.4%.

swesh brings these SOTA capabilities to your local command line with improved developer ergonomics.

---

📜 Attribution

swesh is built by Nikodem Eluszkiewicz, heavily based on the incredible work of the Live-SWE-agent and mini-swe-agent teams.

Citing swesh

If you use this project, please cite:

@misc{swesh,
  author = {Eluszkiewicz, Nikodem},
  title = {swesh: Software Engineering Shell},
  year = {2025},
  howpublished = {\url{https://github.com/elusznik/swesh}},
  note = {GitHub repository}
}

Original Works

Please also credit the original projects:

Live-SWE-agent:

@article{livesweagent,
  author    = {Xia, Chunqiu Steven and Wang, Zhe and Yang, Yan and Wei, Yuxiang and Zhang, Lingming},
  title     = {Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly?},
  year      = {2025},
  journal   = {arXiv preprint},
}

SWE-agent:

@inproceedings{yang2024sweagent,
  title={{SWE}-agent: Agent-Computer Interfaces Enable Automated Software Engineering},
  author={John Yang and Carlos E Jimenez and Alexander Wettig and Kilian Lieret and Shunyu Yao and Karthik R Narasimhan and Ofir Press},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024},
  url={https://arxiv.org/abs/2405.15793}
}

Acknowledgements

• mini-swe-agent
• Live-SWE-agent
• SWE-bench