Hi there 
,
sorry for the clickbait. I actually don’t believe adding AI slop to the Codebase of Django is a great idea, BUT…
I review are sparse commodity in the community and I am forever grateful for everyone who reviews my PRs and helps me move patches forward. In the last couple of weeks and month, I have begun to love AI review more and more. Especially on GitHub with the ability to set them up a default for each nur PR. I find the feedback as an author mostly helpful to iron out all blunders prior to a human review. As a review, I love the AI summaries Copilot creates. I helps me to quickly switch context to a different PR.
I’d propose to enable Copilot reviews on GitHub to show our core team some love and present them this lightly better PRs.
Obviously I am only one voice. I am curious to hear other experiences on this topic.
Cheerio,
Joe
Please avoid sensationalism and edit the title to what you’re actually proposing.
I’ve had some mixed success with AI reviews. Recently, I did see ChatGPT codex spot a subtle issue that humans missed, which I was impressed with. Most of the time, though, I find it spews out unnecessary words, redescribing code changes in prose, which I find just adds noise to the discussion thread.
It’s possible right now to use the Copilot review feature in your personal fork. For example: Fixed #36649 -- Added |= operator support to MultiValueDict and QueryDict. by adamchainz · Pull Request #6 · adamchainz/django · GitHub . (Here the review ended up being one of those “redescribe changes in prose” comments.)
I think that might be the best path, for now, to test how well the feature can work for Django. It seems likely that before we consider enabling the feature for everyone, we’d want a custom instructions file that tweaks the agent to provide a useful review, including Django’s existing contribution guidelines.
1 Like
Hi, Adam.
Yes, you’re right, my bad. I updated the title.
Agreed, it certainly needs a bit of instruction. Which you can provide via agent files.
I am currently working for a company I don’t know if I can name, but their coding agent rhymes with Kevin. I’ve seen how well an agent can perform when pointed towards the right resources in the repo. I’d be happy to share some insights.
But this isn’t about generating code but reviewing it. I agree; if a PR is perfect, a summary isn’t much help. But I noticed that GitHub’s Copilot is getting increasingly better at providing edge case and adversarial insights. Which are easily overlooked while writing code.
I am not proposing a default review. Nonetheless, it might be worth giving it a try here and there.
Best,
Joe
Sure. I think the next step here would be to try it out on some PRs in a personal fork, and see if it’s possible to write an agent file to make the review useful and Django-aligned.
I’d also be interested in what the fellows or other contributors think.
I have proposed an AI team for Django to consider the feasibility of those types of tooling changes for the project, amongst many other things.
Personally I’ve spent lots of time experimenting with AI code review on other open source projects. I find it promising but there are lots of false positives and rough edges. Copilot in particular creates a lot of noise since it’s so integrated within the main PR interface that everyone uses.
When I get the chance I’ll try out those tools on my backlog of Django PR reviews and report back!
Agreed, the false positives can be annoying, but I appreciate the 10% that’s valuable. Given that reviews are a sparse resource, that’s certainly better than no feedback 
1 Like