Nano SWE Agent

2.2.8 · active · verified Sat Apr 11

mini-SWE-agent is a minimalist yet powerful AI software engineering agent, designed to solve GitHub issues and assist in command-line tasks. It's built on a radically simple 100-line Python core, primarily leveraging bash for actions and supporting various models via LiteLLM. Widely adopted by institutions like Meta and Stanford, it focuses on performance, deployability across different environments (local, Docker), and is actively developed with frequent updates.

Warnings

breaking Dependency `litellm` versions `1.82.7` and `1.82.8` were compromised. Update to a version that excludes these, or ensure your `litellm` installation is not one of these specific versions. The `mini-swe-agent` project has explicitly excluded these in its dependencies.
Fix: Upgrade `mini-swe-agent` to `v2.2.8` or newer, which excludes compromised `litellm` versions. If managing `litellm` directly, ensure it's not `1.82.7` or `1.82.8`.
gotcha The `openai` dependency has specific versions (`1.100.0`, `1.100.1`) that are excluded due to known issues. Ensure your `openai` package is not one of these versions.
Fix: Upgrade or downgrade `openai` to a version outside of `1.100.0` or `1.100.1`.
gotcha Older versions (pre-v2.2.5) could experience `FormatError`s, especially with weaker/smaller language models that make tool-calling mistakes, potentially hanging the agent.
Fix: Upgrade to `mini-swe-agent` v2.2.5 or newer to benefit from bug fixes related to `FormatError` handling.
gotcha The output format of trajectory files changed with v2.0 (from `trajectory_format: mini-swe-agent-1.0` to `mini-swe-agent-1.1`). If you have existing scripts or tools parsing v1 trajectory files, they will need updates.
Fix: Consult the 'v2 migration guide' in the official documentation for details on the new output format, especially if parsing old trajectory files or comparing results across versions. Use the `inspector` for interactive browsing of new format files.
gotcha In versions prior to `v2.2.4`, invoking `mini` for the first time after installation and calling `setup` could lead to an exception due to a missing default model name, as the configuration wasn't reloaded correctly.
Fix: Upgrade to `mini-swe-agent` v2.2.4 or newer to resolve the configuration reload issue after initial setup.

Install

pip install mini-swe-agent Stable release
pip install -e . Development install (from source after cloning)

Imports

DefaultAgent

from minisweagent.agents.default import DefaultAgent

LitellmModel

from minisweagent.models.litellm_model import LitellmModel

LocalEnvironment

from minisweagent.environments.local import LocalEnvironment

Quickstart

This quickstart demonstrates how to initialize and run a `DefaultAgent` using a LiteLLM-compatible model (like OpenAI's GPT models) within a `LocalEnvironment`. The agent is tasked with writing a Python function and its tests. Ensure your LLM API key and preferred model name are set, either directly or via environment variables.

import os
from minisweagent.agents.default import DefaultAgent
from minisweagent.models.litellm_model import LitellmModel
from minisweagent.environments.local import LocalEnvironment

# Set your LLM API key and model name
os.environ["OPENAI_API_KEY"] = os.environ.get("MINI_SWE_AGENT_OPENAI_KEY", "sk-YOUR_OPENAI_KEY")
model_name = os.environ.get("MINI_SWE_AGENT_MODEL_NAME", "gpt-4o-mini") # Or "claude-3-sonnet", etc.

agent = DefaultAgent(
    LitellmModel(model_name=model_name),
    LocalEnvironment(),
)

task = "Write a Python function to calculate the nth Fibonacci number, with tests."
result = agent.run(task)

print(f"Agent finished with status: {result.get('exit_status')}")
print(f"Submission: {result.get('submission')}")

view raw JSON →