The rapid adoption of large language models (LLMs) like ChatGPT has introduced new dynamics in software development, particularly within pull request workflows. While prior research has examined the quality of AI-generated code, little is known about how developers actually use these suggestions in real-world collaboration. We analyze 338 pull requests from 255 GitHub repositories containing self-admitted ChatGPT usage, including 645 AI-generated snippets and 3,486 developer-authored patches. We introduce PatchTrack, a tool that classifies whether ChatGPT patches were applied, not applied, or not suggested, enabling fine-grained analysis of AI-assisted decisions. Full adoption of ChatGPT code is rare: the median integration rate was 25%. A qualitative analysis of 89 pull requests with integrated patches revealed recurring patterns of structural integration, selective extraction, and iterative refinement, showing that developers typically treat ChatGPT’s output as a starting point rather than a final implementation.
.
├── LICENSE # The License for the tool - MIT License
├── PatchTrack.py # Main entrypoint of the tool
├── README.md # Readme file to describe how the tool work
├── RQ1_2_3 # Contains all the results for RQ1, RQ2 and RQ3
├── analyzer # Directory for core modules of the tool
│ ├── __init__.py
│ ├── analysis.py # Plotting the classification result
│ ├── classifier.py # Classifies the dataset as PA, PN or NE
│ ├── common.py # Setting n-grams, file types, etc
│ ├── constant.py # All the constant variables used in the tool
│ ├── dataDict.py # Keep track of extracted pr-project-pair information
│ ├── helpers.py # All helper functions e.g. api requests, normalization, etc.
│ ├── main.py # Main PatchTrack class
│ ├── patchLoader.py # Parse and tokenize PR patches. The file should be a diff format
│ ├── sourceLoader.py # Parse and tokenize ChatGPT code snippet
│ └── totals.py # Compute total number of PA, PN or NE classified
├── bin
│ └── os-packages.sh # OS-specific dependencies
├── dataprep
│ ├── __init__.py
│ ├── allPullRequestSharings.zip # DevGPT and Extended json dataset
│ ├── load.py # Functions for loading datasets
│ ├── manual # Procedures on how to generate the extended dataset
│ └── patches.zip # Extracted ChatGPT code snippest and PR patches
├── notebooks # Notebooks containing code for running the experiments
│ ├── __init__.py
│ └── run_experiment.ipynb
├── output # figures and csv files of all the results of running the tool
├── requirements.txt # List of python dependencies required by the tool
├── tests # Testing different component of the tool. This is still a WIP
└── tokens-example.txt # Example file containing GitHub API tokens seperated by a comma (,). This should be renamed to `tokens.txt`
To setup and test PatchTrack tool on your local computer, following the steps below:
The easiest way is using the git clone command:
git clone https://github.com/replication-pack/PatchTrack.gitOperating System: Mac0SX, Linux, WindowsRAM: >= 4 GBStorage: >= 1 GBProcessor: CPU 1.18 GHz or greater
- Git, Python >= 3.10
Let's set python virtual environment;
cd PatchTrack/
python3 -m venv venvActivate the virtual environment
source venv/bin/activatePatchTrack consist of two categories of depencies i.e. (i) OS specific dependencies and (ii) development dependencies. The OS specific dependency is libmagic. To dependencis will be installed automatically when you start the tool.
Datasets are stored in the dataprep directory in zipped files. This will be automatically extracted and placed in the right directory using the step below. Now, let us install the dependencies and load the required datasets.
python PatchTrack.py --initThe above command will install all the required packages, set directories and unzip datasets for the smooth execution of the tool.
Note: PatchTrack has been tested on python >= 3.10
You can also install the OS-specific libraries manually on Ubuntu/Debian or MacOS X, by runing the shell script in the bin directory.
cd bin/
chmod +x os-package.sh
./os-package.shThe above code will automatically detect the OS (Linux or MacOS X) and install the libraries. Note: This is required before installing development specific dependencies.
This is the easiest approach to test the tool and reproduce the classification results presented in the paper. In the notebooks directory, simply run the run_experiment.ipynb file.
If you wish to play with PatchTrack, there are a number of command line arguments for configuring different part of the tool. Some of those option are actively being improved. We are developing a comprehensive documentation of the tool which will be uploaded soon.
Information of the other command line options are provided by using -h or --help
usage: PatchTrack.py [-h] [-i] [-n NUM] [-c NUM] [-v] [-p STR] [-s SOURCE_PATH] [-r]
options:
-h, --help show this help message and exit
-i, --init setup required datasets and directories. This command should be executed at least once
-n NUM, --ngram NUM use n-gram of NUM lines (default: 1)
-c NUM, --context NUM
print NUM lines of context (default: 10)
-v, --verbose enable verbose mode (default: False)
-p STR, --patch_path STR
path to ChatGPT and PR patch files (default: data/patches)
-s SOURCE_PATH, --source_path SOURCE_PATH
path to json files of extracted ChatGPT conversations and code snippets (default: data/extracted)
-r, --restore restore default setting, files and directories
We use GitHub tokens when extracting PR patches. This allows for higher rate limit because of the high number of requests to the GitHub API. Tokens can be set in the tokens.txt file seperated by a comman. The user can add as many tokens as needed. A minimal of 2 tokens can be used to safely execute code and to make sure that the rate limit is not reached for a token.
This repository is MIT licensed. See the LICENSE file for more information.