Git Magic

Safety Precautions

Always work on a disposable copy of the repo with all the remotes removed, so if you screw up, it doesn't really matter. Also, work on a separate branch. That way, you can use git reset --hard <original-branch> to get back to the starting state.

Also, make sure there are no grafts around (eg: linking to the old kdelibs history in the case of frameworks). The safest way to do this is to use fresh checkouts.

Merging in complete history

Used this for moving files from kapidox to kdoctools. Good when the history of the source repo doesn't contain too much you won't care about in the target repo. Preserves commit identities.

Create a commit in the source repo that removes any files you don't want to copy, and rearranges the remaining files to be as you want them to appear in the target repo. Then go to the root of the target repository and use git-merge-repo (from the kde-dev-scripts repository) to merge in the source repo:

/path/to/kde-dev-scripts/git-merge-repo <path to source repo>

Filtering

See mastering git filter-branch: points to extract a subproject for helpful hints.

If you want to trim the history first (which will change the commit identities), you can do something like

git filter-branch --prune-empty --tree-filter "find -type f -\! -path './.git/*' -\! -name foo.\* -delete" HEAD

This example will remove everything that does not match foo.*. Note the -path argument to find that makes sure you don't delete any of git's own files. --prune-empty will remove non-merge commits that no longer have any effect on the tree.

More complex filters are possible; doing as much as possible in the same filter-branch command is probably a good thing to aim for, especially if you want to record the original commit SHA (using --msg-filter). When I wanted to just grab the commits relevant to the two FindDocBook*.cmake files in extra-cmake-modules (which had faked history from kdelibs), I did

git filter-branch \
  --tree-filter "find -type f -\! -path './.git/*' -\! -name FindDocBook\*.cmake -delete" \
  --msg-filter 'sed -e "/^CCMAIL/d" -e "/^REVIEW/d"; echo; echo "Commit $GIT_COMMIT in extra-cmake-modules"' \
  --prune-empty \
  HEAD

This left some useless merge commits, which I eliminated with

git filter-branch --parent-filter "sed -e 's/-p 755debcea57d6ddab0cf851117fb00f3ab39af6f//' -e 's/-p 91d8128e5fc3f20a4621e2f2516ddd34b979bf01//'" --prune-empty -f HEAD

where the two commit SHAs were cut-off points: I didn't want those commits or anything preceding them (because they didn't contain the file I was interested in). These were SHAs after the first filter-branch; the reason I couldn't put them in the first filter-branch appears to be because by the time a commit was reached, its parents had already been rewritten, and so their SHAs had changed.