Name: AutoCodeRover
Uploaded: 2025-02-24T05:30:58.340Z
Duration: 2 min 33 s
Description: AI-powered platform automating code maintenance tasks like debugging and refactoring.

autocoderover

AI-powered platform automating code maintenance tasks like debugging and refactoring.

AutoCodeRover

Hello everyone. I'm Hyen, run a PhD student from National University of Singapore. I'm happy to introduce AutoCodeRover, our new technique for autonomous program improvement. This work is done together with Von J J Fan and a big Joy Field three.

So why have we made AutoCodeRover? In recent years, software engineering has become more and more automated. For example, we can do test generation automatically with symbolic execution. We can do automated program repair, and we can do code automatic code generation with.

Various LLMs in this case one would wonder can software maintenance be automated and autonomous as well. With AutoCodeRover our answer is yes. So what does AutoCodeRover do? In a word, AutoCodeRover takes in a GitHub issue and outputs a patch that resolves the issue. It can do both back fixes and feature additions. 

AutoCodeRover works in two stages. Given an issue statement, AutoCodeRover would first Rover in the code base and collect relevant context to find out the Buy locations. In the second stage.

AutoCodeRover tries to generate patches at the buggy locations to actually resolve the issue. Let us get into more details about the two stages. 

In the first stage, AutoCodeRover has a set of code search APIs. AutoCodeRover would interact with an LM agent to decide what APIs to call and with what arguments. The decided API calls would then be executed locally by AutoCodeRover to retrieve the relevant code snippets. The code snippets would be returned to the agent for further analysis. The code search.

Can go on for multiple runs until the agent decides that enough context has been collected to pinpoint the buggy location. 

Let us see a real world example from the jungle project. So here is the issue statement talking about a missing piece of error information in the code. AutoCodeRover would provide a set of six search APIs such as search class, search method in file, and search method in class. AutoCodeRover would prompt the agent to analyze the issue and select necessary API cause. 

So what does the agent?

Select in this case the agent selects two API calls for searching classes and another two API calls for searching for method. So these four API calls would be executed locally by AutoCodeRover. For example, for this class we have retrieved all its member variables as well as all its member functions. This context would be an analyzed further by the agent. 

Here is the analysis. As we can see, this context has already provided the agent with some insights into the root cause and the resolution of the issue we have.

Another round of code search to do before we have enough context to decide on the buggy location. So here is the buggy method which is the two python method in the method Choice field class. With this py location we would then go into the second stage of the workflow. 

In the second stage AutoCodeRover would first retrieve the code at the Buy location. Then prompt another allow M agent to write a patch. If the patch it if it turns out that the patch cannot be applied to the program autocode Rover would prompt.

The agent to write another patch. Finally, we would get a patch that can be applied to the program. Back to this example, we would first retrieve the code at this buggy two py two python method. 

Then we would prompt the agent to write a patch in such a format. So here is the agent return patch. It turns out that the patch can be applied to the program. 

And here is the patch that we have extract extracted. It turns out that this patch can be applied to the program and it can correctly resolve the issue. So that's the end of the workflow. This is just one of many issues that can be solved by our AutoCodeRover. For more information, you can go to our project homepage. You are also welcome to contact us by email to have any discussion. Thank you for watching.

Note, that there may be slight differences in text and video. Watch video for the original content.

AutoCodeRover Video