Build an automated OSS issue fixer

[mhauru]

This is a pet idea of mine: Using something like AutoGPT, build an LLM-based tool that can identify easily fixable issues in open source repos, clone the repoe and fix them, have a human from the team check over the work, and then send a PR. Could also be a force multiplier. A feasibility study should be done first, i.e. is this at all realistic. My first attempts at using AutoGPT were disappointing, but I suspect a bespoke tool for our needs would do better.

[llewelld]

Lovely idea; this could be a potential Hackweek project as well, with the feasibility study as a starting point.

[JimMadge]

Should be careful not to upset people here.

I believe a number of projects reject automatic contributions (LLM, static analysis) outright and may block people/organisations that propose them.

There is also a risk of creating GPL (or other license) violations. This is part of the reason for this type of contribution being rejected. Also, it there could be reputational damage for the pro-open-source group to violate FOSS license terms.

[mhauru]

Fully agree on proceeding carefully, @JimMadge. The first step, and it's a massive step, would be to build the tool that can make the code changes. That by itself could be extremely useful, even if people only use it on their own projects. Anything beyond that that looks like automation would be a follow-up conversation.

I do think that in the long term having public benefit LLMs improve the quality of our shared open source ecosystem and infrastructure is something humanity should do.

[mhauru]

In yesterday's tech talk Angie Grishina, an enrichment student, talked about her PhD research which was highly related to this. Should reach out to her if we ever get to planning this one.