[GSoC Proposal/RFC] Rolling commit message writing

[GSoC Proposal/RFC] Rolling commit message writing

[Daniel Thomas]

Hello,

I looked at the ideas on the wiki ( 
https://git.wiki.kernel.org/index.php/SoC2010Ideas ) and then I remembered 
that I had an itch to scratch when it came to git.

So I have written a proposal for something which is not on that list:
https://git.wiki.kernel.org/index.php/User:Drt24/GSoC2010/

This is included below to allow for easier commenting on it.

Since this proposal is not on the ideas list I am also looking for someone 
to mentor this project.

However if it is considered that this is a silly idea (I hope it isn't) 
then I will write a proposal for the "Merge helper for LaTeX files" as 
this scratches another of my itches.

Begin proposal:

== Project Goals ==

=== What is the goal of your project? ===
To make it easier for developers to write good commit messages.

=== What you are going to try to accomplish with your project? ===
Currently it is fairly easy to write good commit messages but the effort 
that this requires of developers can be reduced still further.

This would be achieved by adding a built in system for working on a commit 
message during the stage of adding changes.

Currently this can be done manually using "-F <file>, --file=<file>" in 
git commit but we can do better than that. Specifically changes to git add 
[--interactive | -i] to add an 'am' or similar command to allow adding a 
file and then prompting for a message about the changes to that file. Also 
changes to git add [--patch | -p] to add 'm' and 'c' options to git add 
--patch to allow the addition of a message for the current hunk ('m') and 
to do a commit (prompting for review) before returning to the hunk 
currently in focus ('c').

=== How would you measure its success or failure? ===
* The creation of a good API for editing commit messages for work that is 
in the process of being added for commit.
* The integration of this API with currently existing CLI tools such as 
git add so that it fits seamlessly with developer's workflow.
* The modification of existing documentation to properly reflect this 
change.
* The addition of any additional documentation on how to use this new 
feature.
* The addition of thorough testing of this feature to 'git.git/t'.
* Compliance with SubmittingPatches and CodingGuidelines.
* Work should be undertaken with careful communication with the community.

=== The project in more detail. ===
A backend for storing the commit message that is being worked on, probably 
using a file called COMMIT_NEXTMSG in .git/ (following the standard set by 
COMMIT_EDITMSG). With a file format for describing the commit message, 
this would be a normal text file but with the use of 
"/path/to/file.extension:[line number | line-range:] <detail about this 
part of the commit>" to store the various bits of commit message as they 
are added. (alternatively it might be better to use the git notes system 
but I haven't looked into this yet)

Teaching 'git commit' about this so that it appends the contents of this 
file to the commit message supplied at the next commit if '-m' is used or 
if [-n | --nextmsg] is used then it just uses the contents of the file (of 
course [-e| --edit] would allow this to be subsequently edited.

Teaching 'git commit' --rolling or something similar which would open 
COMMIT_NEXTMSG in $GIT_EDITOR.

Teaching 'git add [-i | --interactive]' an 'am' or similar command for 
'add and prompt for message' (with care to not spoil the current 
interface.

Teaching 'git add [-p | --patch]' the 'm' and 'c' commands for adding a 
message about the changes in the current hunk and for committing changes 
added so far after being shown 'git status' being prompted as to whether 
the user wishes to continue (y/n) where 'y' shows the commit message in 
GIT_EDITOR such that writing from that commits and returns the user to the 
hunk they were looking at and 'n' takes them straight back to the hunk. 
This makes it easier to separate out multiple atomic changes from a file.

=== Schedule ===
* 3rd June begin work in earnest: (my last exam is on this day).
* 9th June finish first version of backend and API.
* 15th June finish writing tests on backend and API and any changes 
suggested by the testing or after reveiw.
* 22nd June finish teaching git commit --rolling
* 28th June finish teaching git add [-i | --interactive] the 'am' (add 
with message) command.
* 9th July finish teaching git add [-p | --patch] the 'm' (add with 
message) command.
* 12th July mid term evaluation
* 16th July finish teaching git add [-p | --patch] the 'c' (commit with 
edit returning here) command.
* 26th July finish teaching git commit [-n | --nextmsg] and to use the 
COMMIT_NEXTMSG if available.
* 2nd August write lots more tests on everything implemented so far. 
Ensure that all known bugs are properly fixed. Make any changes suggested.
* 9th August check all the documentation and make all the necessary 
changes. Write any more additional pieces of documentation necessary e.g. 
tutorials.
* 9th August suggested pencils down deadline.
* 16th August firm pencils down deadline, finish rechecking all code, 
documentation, tests and submitting patches to the mailing lists. Write 
final evaluation.
* 20th August final evaluation deadline.
* Continuing: provide any necessary support for these changes and fix any 
bugs that are found subsequently. Continue lurking on the mailing list and 
#git. Help with the continued development of git.

== Interfaces ==

===What parts of Git will you need to call?===
* editor.c
* strbuf.c
* string-list.c
* git show
* Those files which I am also modifying, see below.

=== What parts of Git might you need to change? ===
* git-add--interactive.perl
* commit.c
* builtin/{add.c,commit.c}
* Documentation/{git-commit.txt,git-add.txt}
* t/{t750?-commit*,t370?-add*} (probably just adding new files)

== About You ==

=== Projects that I have worked on ===

Last term I worked on the quebec project for my group project which is 
part of my University course. This can be seen here: 
https://www.srcf.ucam.org/quebec/tomcat/quebec/ and the code is here: 
https://www.srcf.ucam.org/quebec/quebec-0.1.tar.gz though it is not really 
in a fit state for use by anyone else and we never got around to talking 
about licenses. I wrote the User Management system, and ran the server 
(tomcat), version control (SVN after a vote by the group which had in 
general not used VCSs before) and unit testing (junit4). I also rewrote 
part of the comms system and wrote a test client for it. This project was 
also in Java. We did a lot of communication in person but we also used 
Google wave and Google docs. There were five developers in my group and we 
had a client from IBM.

Last summer (2009) I worked on the second iteration of the 
ReadYourMeter.org project (which is still in development) I wrote 8.7ksloc 
of Java and GWT. This code is available under the GPL v2 but since we 
haven't done a release yet it isn't linked to from anywhere. I was the 
only developer on this project, I communicated with my supervisor both in 
person and via email.

In Christmas 2008 I wrote a simple command line 
[https://www.srcf.ucam.org/~drt24/blog/2008/12/10-release-of-fractals/ 
fractal image generator] in SML for my own amusement.

In February 2007 I contributed a couple of minor patches to the 
Worldforge.org project (in python). Most communication was done on #lounge 
on irc.worldforge.org, but I also used the mailing list (particularly for 
the submission of patches). I had CVS commit rights but they have 
subsequently moved to Git and I have only been lurking on their mailing 
list since the summer of 2007.

[https://www.srcf.ucam.org/~drt24/onlineCVdrt24.pdf My CV.]<!-- This link 
will break once the application period is over. -->

===Do you have any prior Git experience?  Have you started to get 
involved?===

I have been using Git for over a year for various personal projects 
ranging from academic work and projects to websites I have worked on for 
various people. I haven't used it in a multi-user manner as in general it 
has been just me working on these projects in git.
I have pulled git.git and successfully compiled and installed it (which 
was pleasantly easy), I also ran the tests in git.git/t which seemed to go 
as well as it was supposed to.

I have been lurking on #git for a few days now and I have tried to provide 
support where I could for those questions which were asked when I was 
watching. I have joined the git mailing list and while I haven't been 
reading all the emails I have been reading a selection of them.

If accepted I would send a weekly progress report to the mailing list 
(unless people would object to that). I would ensure that I regularly 
pushed my changes to somewhere public so that anyone interested could look 
at them. I would ensure that I listened to advice given to me and acted 
upon it. I would start submitting patches to the mailing list as 
[PATCH/RFC] once I have finished a section of work to allow anyone to make 
suggestions on how to do it better early on in the project. I would also 
remain on #git and increase the proportion of mailing list emails which I 
read.

=== Prior contact with the Git community ===

I talked to 
[http://colabti.org/irclogger/irclogger_log/git?date=2010-03-23#l2369 
'Paradox' and 'charon' about this on #git] after 'Paradox' had asked if 
something similar was available. 'charon' mentioned that git commit 
--append --only or 'git rebase --interactiv

I posted a link to this proposal on #git a few hours before I posted it 
here but there were no responses to it.

--
Thank you,

Daniel

Re: [GSoC Proposal/RFC] Rolling commit message writing

[Avery Pennarun]

On Sun, Mar 28, 2010 at 3:32 PM, Daniel Thomas <drt24@srcf.ucam.org> wrote:
> Currently it is fairly easy to write good commit messages but the effort
> that this requires of developers can be reduced still further.
>
> This would be achieved by adding a built in system for working on a commit
> message during the stage of adding changes.
>
> Currently this can be done manually using "-F <file>, --file=<file>" in git
> commit but we can do better than that. Specifically changes to git add
> [--interactive | -i] to add an 'am' or similar command to allow adding a
> file and then prompting for a message about the changes to that file. Also
> changes to git add [--patch | -p] to add 'm' and 'c' options to git add
> --patch to allow the addition of a message for the current hunk ('m') and to
> do a commit (prompting for review) before returning to the hunk currently in
> focus ('c').

The extra prompting seems to me like it would slow down the process of
'git add -p' by making it take too many steps.  Normally when I create
a commit, I like to think of "which lines will I commit?" and "what is
the description of the whole commit?" as separate questions, whereas
this workflow would interweave the two and confuse me.

Do other people prefer it that way?

Secondly, I'm concerned that if you can't remember the description of
your entire commit by the time you commit it, then your commit is too
big.  The usual solution is to create multiple, smaller commits that
are easy to describe.  If at the end you find yourself with too many
commits, you can always join them together with git rebase.

You might also like 'git commit -v', which shows the actual diff to be
committed as part of your default commit message.  Then you can write
your message while looking at it, making it easy to describe exactly
what you changed.

Given these existing capabilities, is it still worth adding the
feature you propose?

Have fun,

Avery

Re: [GSoC Proposal/RFC] Rolling commit message writing

[Jonathan Nieder]

Hi,

Avery Pennarun wrote:

> Given these existing capabilities, is it still worth adding the
> feature you propose?

I think it might be, in some modified form.

Suppose I am working on a medium-sized change that for reasons of
bisectability or ontology has to be one commit.  In the middle of
working, I notice I needed to do something nonobvious.  Currently when
this happens, I get out a pad of paper and write it down, so I can be
sure to mention it in the commit message.

Wouldn’t it be nicer in such situations to be able to say git add
--commit-message or similar, open an editor, and write the note right
away?

In other words, it would be nice to have a file in .git that is

 - cleared on successful commit
 - automatically used to fill COMMIT_EDITMSG, like MERGE_MSG is

I suspect something like this can be achieved by the combination of
the commit.template configuration variable and a post-commit hook to
clear the file, but I haven’t tried it.

Regards,
Jonathan

Re: [GSoC Proposal/RFC] Rolling commit message writing

[Avery Pennarun]

On Mon, Mar 29, 2010 at 11:05 PM, Jonathan Nieder <jrnieder@gmail.com> wrote:
> Avery Pennarun wrote:
>> Given these existing capabilities, is it still worth adding the
>> feature you propose?
>
> I think it might be, in some modified form.
>
> Suppose I am working on a medium-sized change that for reasons of
> bisectability or ontology has to be one commit.  In the middle of
> working, I notice I needed to do something nonobvious.  Currently when
> this happens, I get out a pad of paper and write it down, so I can be
> sure to mention it in the commit message.

When I have this situation, I generally just make a temporary commit
with "git commit", then revise it using "git commit --amend" over
time.

Or else I make a series of commits, then *later* squash them all
together using 'git merge --squash' or 'git rebase -i'.

It seems like the suggested feature would encourage people to do it
the "wrong" way (not creating temporary commits, thus making it easy
to make a mistake and blow things away) just because they aren't aware
of the above options.

Is there a reason that these methods don't work for you?

Thanks,

Avery

the careless committer and fear of commitment (rebase -i vs add -p)

[Jonathan Nieder]

Avery Pennarun wrote:

> I generally just make a temporary commit
> with "git commit", then revise it using "git commit --amend" over
> time.
> 
> Or else I make a series of commits, then *later* squash them all
> together using 'git merge --squash' or 'git rebase -i'.
> 
> It seems like the suggested feature would encourage people to do it
> the "wrong" way (not creating temporary commits, thus making it easy
> to make a mistake and blow things away) just because they aren't aware
> of the above options.
> 
> Is there a reason that these methods don't work for you?

I like ‘git diff --cached’ and ‘git diff’ to show the entire list
of staged and unstaged changes.  I don’t consider this “wrong” at all.
Generally, once something becomes precious enough to be worth keeping,
it is already in presentable form and I can come up with an appropriate
commit message.  I understand that other people work differently; this
is just my personal preference.

I also use the stash heavily, which maybe mitigates the “content could
be blown away with a simple rm -f” problem.

You might also be interested in Pasky’s analysis at
http://thread.gmane.org/gmane.comp.version-control.git/138641/focus=138672

Cheers,
Jonathan

Re: the careless committer and fear of commitment (rebase -i vs add -p)

[Ramkumar Ramachandra]

Hi,

On Tue, Mar 30, 2010 at 10:32 AM, Jonathan Nieder <jrnieder@gmail.com> wrote:
> I like ‘git diff --cached’ and ‘git diff’ to show the entire list
> of staged and unstaged changes.  I don’t consider this “wrong” at all.

On a related note, my workflow is somewhat similar: I just use Magit
within Emacs, which essentially does the same thing. The staging area
in Git really helps me organize my work, and I wouldn't mind attaching
a small note to it that I can refer to while writing my commit message
(much like Jonathan's idea). However, I don't think this task is large
enough for an entire GSoC.

Also, I'm positively opposed to the idea of a complex interactive
command-line interface. If someone *did* want that kind of complexity/
interactivity, they ought to use something like Magit: a cli can't
even come close.

-- Ram

Re: [GSoC Proposal/RFC] Rolling commit message writing

[Jeff King]

On Mon, Mar 29, 2010 at 04:31:00PM -0400, Avery Pennarun wrote:

> Secondly, I'm concerned that if you can't remember the description of
> your entire commit by the time you commit it, then your commit is too
> big.  The usual solution is to create multiple, smaller commits that
> are easy to describe.  If at the end you find yourself with too many
> commits, you can always join them together with git rebase.

That was my first thought, too. I like your suggestion elsewhere in the
thread of making many small commits, and then squashing them later if
you like.

Still, I can see how people might prefer to use this as part of their
workflow, so I am not against the feature itself.

However, I am really concerned with it as a GSoC project. It just
doesn't seem like a big enough feature to take the whole summer. In the
schedule, there is a week of implementing "git add -p
--add-with-message". It seems like that should be hours, not days.

I wonder if it could be rolled into a more comprehensive proposal
concerning commit workflows (though I admit, I don't personally have any
other ideas in the area).

-Peff

PS Daniel (and other potential GSoC'ers who are reading), please don't
take this as discouragement of presenting proposals that aren't on the
ideas wiki. I really really like the concept of somebody working on
their own itch, as I think they tend to have a clearer idea of what the
solution should look like. I'm just concerned this particular proposal
isn't big enough.