Automatically rewrite git history for open source release

Question

I'm currently releasing several projects as open source. Typically the complete source is provided as ZIP archive or checked in at an open source repo. This makes analysis by ohloh difficult.

In case the software has been developed in a non-public repository, the complete history is available. However, I do not want to have the full history released.

I want to use git for reaching one of the two possibilites:

(i) One commit per author: There should be one commit per author (with the commit date the final release date). Each commit contains the lines of code, which finally made it into the final version.

(ii) Original commits with only the final code lines: In this variant, the number of commits itself are preserved. Each commit is modified in a way that only the lines, which finally made it into the final version, are preserved and all other ones are deleted.

Has anyone implemented one of the variants yet? Variant (i) seems to be doable using git-blame and some scripting.

Answer 1

git-oss-releaser is a solution for option (i).

git-oss-releaser converts a given git repository to a git repository only containing the files of the last commit and commits resembling git blame output for each file. The original history is lost.

usage: git-oss-releaser.py [-h] repoDir outDir

Positional arguments:

• repoDir: The repository to transform. May also be a subdirectory of a repo.
• outDir: The directory where the new repo should be created. Has to be empty.

Optional arguments:

• --name NAME The user.name to use for committing the files. Defaults to git's global user.name.
• --email EMAIL The user.email to use for committing the files. Defaults to git's global user.email.
• --date DATE The date to use for commits. Defaults to the date the last commit.

Note that git distinguishes author and committer at a commit. The author is taken using git blame, the committer data is taken from the global user.name and user.emailor the given configured --name and --email.

DEBUG mode can currently only be enabled in the code.

Limitations

• Works on git repositories without any untracked files only
• Empty lines are assigned to "git-oss-releaser" and not the first or last author adding these empty lines
• Repository has to contain at least one non-binary file
• Commit date is derived from non-binary files only
• Tested under git for windows only

Answer 2

(i) One commit per author

I guess that's logically not possible: Suppose you've got a sequence of commits like this:

• Commit: A, Author: Alpha
• Commit: B, Author: Beta
• Commit: C, Author: Alpha

If commit C depends on anything done in B, you cannot reorder and squash A and C anymore.

(ii) Original commits with only the final code lines

For that you can use 'git filter-branch --tree-filter'. Beware that the following script might eat kittens, because I've only tested it on a simple test repository. You've been warned:

git filter-branch --prune-empty --tree-filter '
    # directory which contains the final version of the project
    FINAL_VERSION="$DIRECTORY_WITH_REPOSITORY_OF_FINAL_VERSION"

    # directory which contains the filtered version of the repository
    FILTER_DIR="$(pwd)"

    # apply the current commit in to the final version in reverse mode,
    # ignore the rejects
    cd "$FINAL_VERSION"
    git show "$GIT_COMMIT" > /tmp/original.patch
    cat /tmp/original.patch | patch -p1 -t
    git diff > /tmp/filtered.patch

    # reset the FINAL_VERSION to the original state.
    git reset --hard
    git clean -f -d -x

    # apply the patch which contains the lines which could be reversed on
    # the filtered version
    cd "$FILTER_DIR"
    # revert the last commit
    patch -p1 -t < /tmp/original.patch

    # apply the filtered patch
    patch -p1 -t < /tmp/filtered.patch
    # remove the rejects by the modified patch 
    find -name "*.orig" -o -name "*.rej" | xargs rm -rf
' previousRelease..HEAD    

(this assumes that you've tagged the branching point with "previousRelease". you also have to adapt the FINAL_VERSION variable.)