Solve GitHub issues instantly with AI | Hacker News Reader

Solve GitHub issues instantly with AI

mendable.ai

3 pointsericciarla3 comments

Hey HN,

It’s Eric, Nick, Garrett, and Caleb from the SideGuide team.

We’re excited to announce Mendable, a GitHub action that automatically suggests fixes for your GitHub issues with GPT-3. You can test it out right now with any Github Issue on our website!

Open-source maintainers hold the internet on their shoulders. But maintenance can be a thankless job. Issues pile up; the longer you wait to respond, the less happy users will be. We hope Mendable will ease the burden on open-source and help maintainers spend more time on meaningful work.

We’ve been exploring tools to help with developer experience for a while. With SideGuide, we built a community of 50k developers on our Instagram and watched the Discord server grow to thousands of developers. That’s when we noticed how difficult it was to answer all the questions and engage with our community.

So, we built Mendable. In its current form, Mendable is a Github action that connects to the GPT-3 API via our server. It senses when new issues are opened, then picks a prompt based on the tag the user places in the issue. Our immediate next steps will be to add support for questions on Discord and Slack, and build a tool that allows you to review responses before they post (via discord, slack, or email). Once OpenAI releases fine-tuning support for text-davinci-003, we will explore training our model to improve the quality of the responses.

Right now, Mendable is just an MVP but we have some useful features. Some are available now, with others coming soon:

- Instant Issue Solver

- Different Tag, Different Solution

If you’ve made it this far, thank you . We’re excited to hear the HN community's feedback and suggestions. We have a long way to go, but we hope Mendable will help make maintenance as easy as possible!

Best, E, C, G, N from SideGuide

> We don't look at your repository's code

Are the suggestions liable to be in any random language?

Also, it looks like your recommended workflow comments on all issues. But not all issues are fixable problems. How does this bot respond to non-bugs – things like suggestions, questions, and the like? It'd be cool to have more screenshots depicting this.

ChatGPT is also quite easy to trick into spouting abusive responses. Have you red-teamed getting Mendable to output mean stuff?

Hi, Nick here from SideGuide. Thanks for your questions and suggestions. 1. As of right now it seems like 90% of the responses can infer the language pretty well from the issue context. Although we are working hard to improve this. 2. We use the issue's labels to identify the issue type and choose an appropriate prompt depending on what type it is. Thanks for the suggestion, will add more screenshots about that asap. 3. We're currently only using GPT-3 (davinci-003) on production. But we are currently testing ChatGPT too, but we don't think it is ready yet due to it is down time/rate limiting and lack of fine-tuning capabilities. Because this is still an early MVP, we haven't gotten around to filtering abusive responses just yet.

Reply on news.ycombinator.com