BugCraft: Automated Bug Reproduction and Action Synthesis

BugCraft is a powerful tool designed to streamline bug reproduction by automatically generating detailed, actionable steps to replicate software issues. It leverages a combination of step synthesis and action modeling to produce comprehensive reproduction workflows.

BugCraft-Bench

The bugcraft_bench folder contains the final 86 bug reports that can be reliably reproduced, serving as our benchmark.

Prerequisites

• Windows OS: This tool uses pywin32 for window manipulation, making it Windows-specific.
• Capable Hardware: Ensure your PC meets the requirements to run OmniParser, particularly regarding GPU and memory.
• Python 3.8+: Required for running the application.

Setup Instructions

1. Clone Repository

   git clone https://github.com/erayyap/bugcraft
   cd bugcraft

2. Install OmniParser Weights

   • From https://huggingface.co/microsoft/OmniParser, download:

     • icon_caption_blip2
     • icon_caption_florence

   • Place them in the respective directories:

     ./action_model/OmniParser/weights/
     

3. Create Virtual Environment

   python -m venv venv
   venv\Scripts\activate
   pip install -r requirements.txt

4. Configure Environment Create a .env file in the root directory of the project with the following variables:

   OPENAI_API_KEY=TO_BE_FILLED  # Required
   TAVILY_API_KEY=ONLY_FILL_IF_YOU_USE_SEARCH  # Optional for web search
   
   # Path configurations (You must configure these!)
   WIKI_DIRECTORY=C:\\path\\to\\wiki\\output_pages
   FLORENCE_PATH=C:/path/to/icon_caption_florence
   ICON_MODEL_PATH=C:\\path\\to\\model_v1_5.pt
   
   # Model configurations
   STEP_SYNTH_MODEL_NAME="gpt-4o"
   ACTION_MODEL_NAME="gpt-4o"

5. Prepare Bug Reports Place all bug reports you want to process in the bug_reports folder, located in the root directory of the project.

Usage

Important: Run the following commands in a terminal with administrator privileges. This is crucial for BugCraft to interact with application windows correctly and perform necessary actions.

Full Pipeline (Step Synthesizer + Action Model):

python main.py

Step Synthesizer Only:

python main.py --only-step

Configuration Options

Step Synthesizer

Feature	Description
USE_WIKI	Enables the use of an internal knowledge base (Requires wiki setup).
USE_SEARCH	Enables web search capabilities using Tavily.
USE_MOB_CHECKER	Enables a separate call to an LLM to verify mob interactions.
USE_ALTERNATE_SOLUTIONS	Enables a separate call to an LLM to check for alternative solutions in S2R steps.
USE_REASONING_TRAJECTORY	Instead of directly including wiki/search pages in the context, utilizes reasoning trajectories (Refer to the paper for details).
USE_FINAL_CLUSTERING	Performs a final LLM call to integrate information from images/pictures into the S2R.

Action Model

Feature	Description
MAKE_FULLSCREEN	Maximizes the Minecraft window to fullscreen. This can slightly enhance the traversal capabilities of GPT-4o but increases OmniParser's execution time and raises the cost of OpenAI calls due to larger image inputs.
SEPARATE_THOUGHT	Instead of generating thought and action in separate LLM calls, generates them concurrently.
USE_CORRECTION	Before executing an action, sends another LLM call to verify its correctness. This also helps to prevent the agent from getting stuck in loops.

Important Notes

1. Administrator Privileges: Ensure you execute the commands with administrator privileges to allow BugCraft to function correctly.
2. Path Configuration: Double-check that all paths in the .env file use the correct Windows-style backslashes (\).
3. API Keys:
   • An OpenAI API key (specifically for gpt-4o) is mandatory for the default configuration.
   • A Tavily API key is only required if you enable USE_SEARCH=True.

For optimal performance, we recommend:

• 16GB+ RAM
• NVIDIA GPU with 8GB+ VRAM

If you want the dataset, our run infos and all the configuration pre-done, you can reach all of that as a packaged zip in: https://figshare.com/s/70b5d141e69ff822cd40