Bug Agent

An autonomous agent to help maintainers research issues/bugs. All changes are local, the agent is intentionally designed to have a human in the loop.

Pipeline:

1. GitHub Issues API — fetch open issues from a repo.
2. Issue Summarizer — convert raw issue JSON into a structured object.

• LM advisor — use SmolAgents with a lightweight open-source model (default: HuggingFaceTB/SmolLM-1.7B) to analyze the issue and provide guidance. Current code always prefers CodeAgent with tools and stops on CodeAgent failure instead of falling back to direct model.generate().

Installation

Install dependencies:

pip install -r requirements.txt

Quick Start

The default model HuggingFaceTB/SmolLM-1.7B requires ~4GB RAM. For best results, run on Google Colab (free tier has 16GB) or a machine with sufficient memory.

(Optional) Change the model used by the advisor. Default is HuggingFaceTB/SmolLM-1.7B:

export MODEL_NAME=HuggingFaceTB/SmolLM-1.7B
# Windows PowerShell
$env:MODEL_NAME = 'HuggingFaceTB/SmolLM-1.7B'

Run the agent (defaults to the latest open Transformers issue):

python -m src.main

• To process a specific issue:

python -m src.main --issue 44593

Example Labels (from Transformers)

• #44593, #44910 → check PR
• #44485 → references to vllm and sglang
• #44829 → need GPU
• #44869 → existing issue/PR
• #44912 → start by recreating issue
• #44995 → PR #44950 will fix this