AI Agent to analyze Nightly Composes

Project Discussion
,
1

Hi All,

I have been working with the AI/ML SIG on building a library of AI Skills. I see these skills as a great way to make contributing to Fedora more accessible. A first concrete use case and proof of concept I worked on is an AI Agent designed to analyze Rawhide nightly compose logs, and find the root cause of failures.

It is a great practical example on how AI and skills can be integrated into Fedora’s workflow and infrastructure. I wanted to share the proof of concept that has been running autonomously since last Friday.

Current Solution

The Agent reads the daily rawhide compose tracker issue and analyzes the failure. Once done it adds a comment with its findings.

This is at the early stages and mainly aimed at showcasing how AI Agents could be useful for Fedora as well as gathering feedback. Long term it could graduate from Communishift and be deployed in Fedora’s Infra OpenShift cluster.

How you can participate:

  • Do you want to help with composes failures? Take a look at the Agent analysis in the compose-tracker-issues and see if there is something you can help with.
  • If you enjoy working with LLMs, we would love your help expanding the Skill Library.
  • Come chat with the team on Matrix at #ai-ml:fedoraproject.org or share your thoughts right here on Fedora Discussion using the ai-ml-sig tag.
3 Likes
2

Thanks for that, by the way! I appreciate you driving energy into this. :rocket:

If you get time to share more, I would be curious to know how you used this part. I assume the Gitea/Forgejo APIs are more or less identical as far as endpoints go?

I appreciate that you got to both create and refine this in a practical, open way. I wonder how we might make a more strategic ask for contribution. Like, are there a specific group or types of packagers or QA folks who may also want to use the skill at some point, and would it work well for them or not?

It would be a useful exercise to determine what the better-worn path should be for experimental AI use projects moving from CommuniShift to a production setup should look like. I’d like to think through that!

3

Yes, IIUC Forgejo is a fork of Gitea and there is API compatibility up to a certain version, see Forgejo numbering scheme | Forgejo – Beyond coding. We forge.

In practice, I would expect that most basic needs will be compatible.

Here, the Agent is the main user of the skill, but anyone should be able to run it locally. I think what would be interesting as next steps, and opportunity to contribute is to have a skill to analyze the compose tracker issues and try to fix them.

Yes, definitely! Sounds useful to have some guideline, security best practices etc …