Index multiple commits within multiple branches of a repository

[kirusfg]

I dug through the issues and code and I am pretty sure indexing multiple branches and multiple commits of a repository is not supported yet, and it's a dealbreaker for me. When setting up a repository, Tabby could ask the user the branch to use, or, ideally, a number of them.

The closest thing related to this I found is #3122, but I think it is only talking about the commit message from the last commit. I guess a lot of thought would have to be put into how to provide context to the model in a multi-branch, multi-commit setup...

Is this feature planned? What would be the biggest design concerns of this feature?

[wsxiaoys]

Thank you for providing your feedback. Could you please also elaborate on the specific use case detailing how the ability to index from multiple branches could be exceptionally beneficial ("dealbreaker") for your scenario?

[kirusfg]

I stumbled upon this because of a very trivial thing: the active branch master is an old, no longer maintained branch in all of our repositories that we just didn't bother to delete. We have dev, main, and a bunch of feat/new-feature and fix/bug branches. At the very least, choosing what branch to index would have been nice. But again, the main branch that only has changes from the last release might be missing features, fixes, and refactors from dev and other branches.

We often use separate branches to work on features in parallel. It would be great if one could ask Tabby what major changes have been made in another branch (this requires the model to look at commits after branch was created) to know how to continue development of their own features so that the merge is not as painful.

I also can see multi-commit indexing to be beneficial. It's probably not necessary to process all commits, just N last commits. The autocomplete model then could incorporate this additional context to make better suggestions. One example is Supermaven - they claim to be using code diffs to train the model. I don't know if providing chunks from older commits would work with openly available models that were not specifically trained on such data, though. But if it works, the chat model would now have additional context to more accurately describe what has been done in a branch, e.g. if it's feat/new-feature, it would describe the feature better.

Otherwise, use-cases of Tabby are somewhat limited for me. I can see big enterprise company employees using it to quickly get insight into parts of the codebase they are not familiar with, or using it for autocomplete in their company's style, but nothing other than that.

[kirusfg]

After some searching, I haven't found projects claiming to be able to do all this. GitHub Copilot in VS Code does re-index or switch index used (or whatever it does) when a branch is switched. At least that would have been convenient - being able to switch the branch of the codebase provided as the context to the chat and autocomplete models.