clean-push

Automated git flow to produce safe, neat (rebased + squashed) pull-requests.

Have you ever been frustrated with git because:

• Your long-lived work branch commit-history has become too messy when the time has come to publish your changes?
• You had duplicate or partly overlapping commits because of history-changing squashes & rebases?
• A commit has been reverted, but instead of having an empty delta you had both commit + reverted changes?
• When you tried to rebase, you got into "rebase hell" (continue, skip, abort) loop, which seemed to never end?
• Syncing your repo with origin/master created more conflicts than expected?
• A push after a merge unexpectedly included someone-else's changes due to a messy merge with master + a 3-way diff against an already existing PR?
• You wanted to reuse an existing branch for additional work, but there was too much legacy in its commits so you were forced to start a new one and replay all your changes again?
• You resorted to (git diff + git apply) or (stash + pop) or (a loop of git cherry-picks from prior work) but found the multi-step process too complex/involved and error-prone?
• You created a pull-request, and noticed a small error, now you want to redo the PR with some small fix, but hesitate because it would mean too much work all over again? (think git commit --amend but for pushes)

If you answered "yes" to any of the above?

clean-push may be just the script you need.

clean-push implements the following:

• A full sync with the main branch before starting a push leading to a pull request
• Simple conflict detection, like merge does
• Consolidated/simple conflict resolution (unlike with git rebase)
• The pull-request is clean:
  • It has only one delta (diff) vs main
  • It is rebased on top of the latest HEAD of main
  • It doesn't contain unwanted merge commits
• For safety: most risky git operations which might mess-up your work, are done on a temporary/throwaway branch.
• Your feature branch is atomically modified once (with one git reset) after all issues have been resolved and the single delta vs main looks good.
• You get a chance to edit your pull-request message in your favorite editor and make it look even better.
• You can re-edit all the messages in a long list of commits without the complexities introduced by the fix-up syntax of git rebase -i
• The pull-request on github looks exactly like you wrote it in your editor: the first line becomes the title of the pull-request
• If you want to fix-up anything, you can just repeat the call to clean-push and the new push will override the previous instead of appending to it
• You never get into "rebase-loop hell", because git rebase isn't used anywhere
• Works both on Linux and Mac OS-X
• Protected from being called from a git hook (a nested call which may cause damage)
• Ability to pause & allow you to edit intermediate git steps before executing them. A common use-case for me is to be able to add --no-verify to the end of a git commit or git push sub-command in order to skip some long duration hooks.

Crucial terminology

In this text, there are repeated references to two distinct branches.

The 1st is the main branch you have branched from to do your development. This branch can be referred to by several names, among them:

• master
• main
• parent branch (although git doesn't really support parent/child relationships between branches, as branch names are just refs, i.e aliases for commit-ids)
• CICD branch (if we auto-deploy from it)
• dev or the 'development branch'
• The source of truth
• When this branch is used to release from, it may be called the release branch
• On github.com in settings/branches they call this the "default branch"

Similarly, different people may refer to the ephemeral branch they're developing on, as any of:

• work branch
• topic branch
• bug-fix branch
• feature branch

As far as clean-push is concerned, there are only two branches. In this text, they are referred to as:

• main or master
• work branch

Regardless of what they are actually called.

clean-push queries the current branch (from which it was called) in runtime for the work branch actual name.

Figuring-out the main branch actual name is harder. clean-push tries to check the following for existence, in order:

• "$1" (first, arg, if passed on the command line)
• dev
• staging
• main
• master

Better heuristics for figuring out the later are welcome (please open a github issue if you know a more reliable and elegant solution).

Usage:

While working on your work/topic branch, you can run any of:

    clean-push [<main-branch-name>]

    4-way-diff [<main-branch-name>]

The argument is optional. If you don't provide it the script(s) will try to guess it. See also the Customizing behavior section below.

4-way-diff

4-way-diff is a handy script that provides a quick, non-destructive, (read-only) view of the 4-way state for full situational awareness.

It tells you which of the 4 copies is not in-sync by performing the full circle of comparisons:

• local work branch vs its 'git index' (is current branch fully committed a.k.a: clean?)
• local work branch vs remote/pushed work branch
• local work branch vs remote master
• remote work branch vs local copy of master (the 'tracking' branch for master)

Try it and it can help you to quickly diagnose what may still not be in-sync.

Here's for example the message you get when something isn't committed yet:

[full diff comes here]

4-way-diff: (1) work branch not clean. Need to commit (or stash) locally

Here's the message you get when your branch is fully committed but has not been pushed yet:

[full (reverse) diff comes here]

4-way-diff: (2) work branch: local != remote. Need to push

Here's the message you get when your branch changes are committed, pushed, but not yet merged to the main branch on the remote server:

[full (reverse) diff comes here]

4-way-diff: (3) local work != remote main. Need to remote-merge

Here's the message you get when your branch changes are committed, pushed, and merged remotely, but your local main branch is now a step behind (because the merge was only done on the remote):

[full diff comes here]

4-way-diff: (4) local work != local main. Need to pull in local main
To fix:
        git checkout main && git pull && git checkout work

And finally, the message you get when everything is in-sync:

(1) work branch is clean, cool
(2) work branch: local == remote, no need to push, cool
(3) remote work doesn't exist (already merged?)
(4) local work == local main, no need to pull in main, cool

Customizing behavior

The following default behaviors can be changed in both clean-push and 4-way-diff:

1. Name of "upstream" (git remote repository location)

2. Names of "main" branch-names to try

3. If your remote is not called origin, you may change the variable you should change the default value of the UPSTREAM variable in the code, from:

UPSTREAM=${UPSTREAM:-origin}

to:

UPSTREAM=${UPSTREAM:-<your_upstream_name>}

2. In order to figure out the "main" branch from which you started and want to merge into, the code makes a few possible guesses.

If these values don't work for you, you may change the value of MAIN_BRANCH_NAMES=( ... ) in the code.

You may also pass an explicit "main" branch name argument to any of the scripts which will disable the guessing altogether, e.g.:

    # Run a clean-push vs 'cicd-branch'
    clean-push cicd-branch

HOWTO

First, make sure the two utilities:

clean-push
4-way-diff

are always in your $PATH. I simply have ~/bin in my $PATH and have them both copied there, like this:

mkdir -p ~/bin

# Make sure the two scripts are executable:
chmod a+rx clean-push 4-way-diff

cp clean-push 4-way-diff ~/bin

# Make sure you have ~/bin in your PATH.
# This has to be in your ~/.bashrc or ~/.profile
export PATH=~/bin:$PATH

Now go to your repo (cd .../your/repository) Assuming you start in branch 'main':

# Create the 'work' branch.
git checkout -b mydev

# Add/remove files and make changes as you would normally do in git...

# Once: ready to push, make sure everything is committed:
git commit -a

# Now that your repo is 'clean', run the clean push
clean-push

Similarly, to compare to remote, just run:

4-way-diff

That's all there is to it!

Caveats

clean-push is intended for flows:

• With a single common main branch (main, master, or similar).
• Where developers want to be in sync with each other as much as possible
• Where the main branch is the source of truth (e.g. used for continuous/automatic CICD and releases)
• Encourages fast-development by many developers on the main branch
• Detects conflicts as early as possible by frequent merging for other developer branches

If your flow:

• Encourages multiple long-lived diverging separate development tracks
• Rarely merges
• Has multiple forks like production vs "next gen". Extreme case: what you develop today will be seen by customers only in 3 years, if at all.
• Doesn't use continuous integration and deployment (CICD)

Then clean-push may be useful for work on one fork, but may not be useful beyond it.

Credits:

clean-push is based on method (1) of this excellent page by Lars Kellogg-Stedman

Before discovering that page, I tried several unsatisfactory solutions to the problem, all of which fell short on some aspect of the problems described above.

Thanks to Steve Malmskog & Jordan Bucholtz for their early testing & contributions.

Portability notes

clean-push is a simple bash script.

It was developed and tested using:

• bash: 4.4.20, 5.0.17
• git: 2.17, 2.25.1

Verified on Linux (Ubuntu 18.04, and 20.04) & Mac OS-X (10.15, Catalina).

License

Written-by: Ariel Faigon, 2021

Released under the MIT licence.