Hacker News new | ask | show | jobs
by progbits 496 days ago
Ah this again?

Commit message is not just a quick summary of what, it's also a historical record of why. Can't generate the latter from the diff.

6 comments

Yep. You should stop and answer “why does this commit exist?” since that is the question you will be asking yourself when you discover it with `git blame` or `git bisect`. It amazes me how many devs are just willing to slap a -m on `git commit` and just say “fix bug” or “pr feedback”.
Thank you for the response, My target is people that are bad at writing commits, barely or don't even write commits. Also from what you arguing I've worked on alot of projects including opensource and the majority just provide a quick summary so this would be of great use. Here are some commits the bot wrote

feat: Enhance Auto-Commit Bot with new features and improvements - Reworked the description to provide a more detailed overview of the tool's capabilities. - Added badges for license and Python version. - Updated the installation instructions to include environment variable setup for the Google Gemini API key. - Implemented logging for all actions to facilitate debugging and provide a comprehensive record of events. - Included a usage example to demonstrate the workflow of the tool. - Improved the commit message generation using the Google Gemini API for increased accuracy and context.

And

feat: integrate Gemini API for commit message generation This commit introduces integration with the Gemini API to enhance the automated commit message generation process.

The following changes were made:

- Added a `CommitMessageGenerator` class to generate commit messages using the Gemini API. - Modified the `ChangeDetector` to handle the generation and staging of commit messages. - Updated the CLI to accept an API key for Gemini API access. - Added error handling for missing API keys. - Implemented tests for the `CommitMessageGenerator`.

Yes, when spelunking why something has changed having good rationality would be a godsend... however let's face reality:

1. Majority of commit messages are low quality and would benefit significantly from a good summary of what was done.

2. Margin of commit messages is often too small for documenting the rationale - this job is better left for tickets.

> 2. Margin of commit messages is often too small for documenting the rationale - this job is better left for tickets.

The commit message lives with the code. The number of times in my career that a company has migrated, changed, consolidated, or otherwise made all those links in commit messages obsolete, well I don't quite yet need two hands. But I see a lot of dead ends to context in code bases.

Nothing stopping someone from deciding to migrate VCSs (or even just repos) and all of a sudden end up in the place that 15 years of codebase history is squashed down to a single new starting commit. :(
Don’t forget that most modern version control systems are distributed. The same can not be said for ticketing systems.
If the content of the commit is too small to be meaningful, the commit is too small. Either wait until you’ve done more before committing, or rebase -i before pushing your branch and merge up any piecemeal commits into better sized ones that communicate a story.

Keeping that metadata in tickets means that it can’t be read within ‘git blame’. It’s also all too easy for it to be lost entirely when changing ticketing systems, transferring codebases between teams or companies, etc.

> Majority of commit messages are low quality and would benefit significantly from a good summary of what was done.

Why? At that point doesn't the diff speak for itself?

Not to mention, the best commit messages are about the code that isn’t there.

As you say, we know what’s there. I need the author to tell me why it’s there, and perhaps what other alternatives were abandoned because they didn’t work.

You know, so I don’t waste a day finding out for myself why you didn’t just do ${obvious}.

Well, that is exactly what the code does right now: https://github.com/suwi-lanji/auto-commit/blob/40c1fc34adce5....

I would probably have it look at the files and try to generate a nice summary of “why”, but for those minor things where you add a parameter or fix whitespace I think this is OKish.

True, but you can infer the why from what changed for a lot of cases

e.g. - Add types for X, Y Z

if the PR goal is to make types more strict, that message is clear.

I feel like the quality will be worse than if the engineer really put some thought into it, but the problem is, commits are annoying to write.

A lot of people do “wip” or do a worse than average job.

Having a summary of what changed is still better than that.

Edit: if you feed more context about what you’re trying to develop it will probably be able to infer.

I’d rather have «wip» commit messages, and thus forced to open the diff, than worrying of the message being hallucinated.

I’m totally cool with people using this as an initial draft and then manually tweaking it though.

Commits are annoying to write is in the same space as "variable names are annoying to write".

Communicating intent should be trivial or, if it's not, put effort into it because communication with that future dev may be essential .

Now, maybe some bullet points may suffice and quickly be reshaped by AI to make it more succinct and clear but you still should make the effort. The more you do it, the more it becomes an easy part.

The problem is when you copy and paste the same info in slightly different ways in many places and I can appreciate some form of suggestion around "you missed explaining why you want to change this part".

There's a space for AIb but it isn't "so that I don't need to think".

I've never worked that way.

In my mind, commit messages should be a quick summary of what. The basic motivation should be clearly labeled as a feature, bugfix, etc., but that's all. Commit messages are for quick browsing and a summary of what changed, not extensive justification.

The why is too important to be put in a commit message. The product why belongs in a linked issue that describes the bug or use case in full detail. Meanwhile, the technical why's (why this particular solution as opposed to alternatives) belong in the code itself as comments.

The first line (or atleast, the first 80 characters) should be a quick summary - so you can quickly browse via git blame.

But the actual commit message should consist of History/Motivation/Context - so that someone who's going through the blame can understand why a certain change was made, and what the context was.

Linus had a good template for this, which makes a lot of sense: https://gist.github.com/finalfantasia/bd0070673ca27e5f7473

If I want history, motivation and context to actually be read at some point I put it in tests, code comments and README/docs.

If I want my history, motivation and context to be ephemeral I put it in a commit message.

It still perplexes me why people obsess over commit messages while the places where people are actually looking when they have these questions are neglected.

I don't particularly want a commit log at the top of every function, TBH. If I'm looking for the history and context of a change, I want to look at the commit log. If I want a description of the current version, I look in the docs (and yeah, sometime I can only piece together why something is the way it is now by looking through how it's changed. But documenting each change has a much better chance of addressing questions like "why did it work and now it doesn't" then trying to get the docs perfect will.)
>I don't particularly want a commit log at the top of every function

If thats what you got from my comment then you misinterpreted it.

Nobody wants docs to look like commit messages. They want them to be relevant to the context.

{Tests, Code Comments, Documentation} are 3 distinct places to trawl through when quickly going through git blame.

The commit message is one place - and gives the author an opportunity to speak directly with a future developer over the place-in-time-context that this change was made.

Git blame is something you use to ask "what the fuck"? Hence why it's got the tongue in cheek moniker "git blame".

If devs are constantly asking "what the fuck?" all over the code base then that's usually coz tests, code comments, docs and code quality were all badly neglected.

Better commit messages are a band aid over that gaping wound.

You're focusing a bit much on the Developers bit. It's not just Developers who are working through the code-base.

As a lame example - Incident Response/SRE will also be trying to get their heads around changes being made - especially if they're responding to an outage, and trying to figure out what change broke production - and why it was made.

Not everyone will know every bit of the project as intimately as the Dev team - and having a good commit message will help any unfamiliar response team mitigate, or escalate accordingly.

Issues are not part of the source code repository. The issue tracker may not be there anymore at some point in the future, and it also may not be accessible when you work offline.

I agree about the technical justifications belonging in the source code itself as comments.

Also when the issue is a bug the information you'll find in the tracker is usually about the symptoms and says nothing about the fix.
> The product why belongs in a linked issue that describes the bug or use case in full detail.

That's nice until you switch git hosts or project management systems and the context is lost. Commits live forever, use them.

At least a small summary of "why" would've been so helpful many times in the legacy codebase I maintain instead of "fix tests", "format", "FOO-123"