Moqui AI Policy

Community
1

I think it’s time to consider what to do about AI PRs and contributions.

We’ve gotten several PRs recently that have been almost entirely AI that come with good features, but come with some AI slop:

I’m thinking we just reject these outright and request for specific changes if they want any code to go upstream.

Any thoughts?

2

I guess we can silently ignore things we don’t like but it would be better to give reasons for rejection.

3

I cannot think of any AI functionality that must touch the framework directly. It seems AI can be abstracted away in higher layers. Even then use cases must be clear to avoid buzzword features that get abandoned quickly.

Real critical work right now might be upgrading, modernizing, lots of widget system work, java groovy upgrades and so on. We have a lot of critical foundational work that might be more pressing at the moment. I’m pretty sure AI is relatively an easier problem to solve once we sort all this out

4

I tried the System Improvements by hunterino · Pull Request #674 · moqui/moqui-framework · GitHub and, aside from editing some configuration files, Moqui started without any problems. I’m still in the early stages of testing, but everything seems to be running fine. Virtual threads aren’t active yet.

5

Platform improvement pairs well with async processing and that’s going to be gold for AI. It seems to me like the future of AI in business will mostly consist of weaving LLM calls with classical business logic, still encapsulated in transactions. We’re set up very well for that work.

6

120 files in 674! Seems like it needs to get more focused in its direction. This is like someone typed “upgrade everything and document it” as a directive to Claude. There’s an MCP server (with no screen support! :-D), a new health check servlet and lots of new documentation files (ie. executive summary). All things that would benefit from being decomposed into separate digestible commits. I don’t think our AI strategy can be a sequence of “fix everything” requests and remain testable/coherent.

I’m totally riding the vibe code train myself but I got group feedback (and think it makes good sense) to implement these new AI tools as components rather than modifications to core. There is no reason MCP can’t be done as a separate component. It should be, in fact, because there are obviously lots of opinions.