Merge git repository in subdirectory

After getting the fuller explanation of what is going on, I think I understand it and in any case at the bottom I have a workaround. Specifically, I believe what is happening is rename detection is being fooled by the subtree merge with --prefix. Here is my test case:

mkdir -p z/a z/b
cd z/a
git init
echo A>A
git add A
git commit -m A
echo AA>>A
git commit -a -m AA
cd ../b
git init
echo B>B
git add B
git commit -m B
echo BB>>B
git commit -a -m BB
cd ../a
git remote add -f B ../b
git merge -s ours --no-commit B/master
git read-tree --prefix=bdir -u B/master
git commit -m "subtree merge B into bdir"
cd bdir
echo BBB>>B
git commit -a -m BBB

We make git directories a and b with several commits each. We do a subtree merge, and then we do a final commit in the new subtree.

Running gitk (in z/a) shows that the history does appear, we can see it. Running git log shows that the history does appear. However, looking at a specific file has a problem: git log bdir/B

Well, there is a trick we can play. We can look at the pre-rename history of a specific file using --follow. git log --follow -- B. This is good but isn't great since it fails to link the history of the pre-merge with the post-merge.

I tried playing with -M and -C, but I wasn't able to get it to follow one specific file.

So, the solution, I feel, is to tell git about the rename that will be taking place as part of the subtree merge. Unfortunately git-read-tree is pretty fussy about subtree merges so we have to work through a temporary directory, but that can go away before we commit. Afterwards, we can see the full history.

First, create an "A" repository and make some commits:

mkdir -p z/a z/b
cd z/a
git init
echo A>A
git add A
git commit -m A
echo AA>>A
git commit -a -m AA

Second, create a "B" repository and make some commits:

cd ../b
git init
echo B>B
git add B
git commit -m B
echo BB>>B
git commit -a -m BB

And the trick to making this work: force Git to recognize the rename by creating a subdirectory and moving the contents into it.

mkdir bdir
git mv B bdir
git commit -a -m bdir-rename

Return to repository "A" and fetch and merge the contents of "B":

cd ../a
git remote add -f B ../b
git merge -s ours --no-commit B/master
# According to Alex Brown and pjvandehaar, newer versions of git need --allow-unrelated-histories
# git merge -s ours --allow-unrelated-histories --no-commit B/master
git read-tree --prefix= -u B/master
git commit -m "subtree merge B into bdir"

To show that they're now merged:

cd bdir
echo BBB>>B
git commit -a -m BBB

To prove the full history is preserved in a connected chain:

git log --follow B

We get the history after doing this, but the problem is that if you are actually keeping the old "b" repo around and occasionally merging from it (say it is actually a third party separately maintained repo) you are in trouble since that third party will not have done the rename. You must try to merge new changes into your version of b with the rename and I fear that will not go smoothly. But if b is going away, you win.

git-subtree is a script designed for exactly this use case of merging multiple repositories into one while preserving history (and/or splitting history of subtrees, though that is seems to be irrelevant to this question). It is distributed as part of the git tree since release 1.7.11.

To merge a repository <repo> at revision <rev> as subdirectory <prefix>, use git subtree add as follows:

git subtree add -P <prefix> <repo> <rev>

git-subtree implements the subtree merge strategy in a more user friendly manner.

The downside is that in the merged history the files are unprefixed (not in a subdirectory). Say you merge repository a into b. As a result git log a/f1 will show you all the changes (if any) except those in the merged history. You can do:

git log --follow -- f1

but that won't show the changes other then in the merged history.

In other words, if you don't change a's files in repository b, then you need to specify --follow and an unprefixed path. If you change them in both repositories, then you have 2 commands, none of which shows all the changes.

More on it here.