git-filter-repo/README.md

git filter-repo is a versatile tool for rewriting history, which includes capabilities I have not found anywhere else. It roughly falls into the same space of tool as git filter-branch but without the capitulation-inducing poor performance, and with a design that scales usability-wise beyond trivial rewriting cases.

While most users will probably just use filter-repo as a simple command line tool (and likely only use a few of its flags), at its core filter-repo contains a library for creating history rewriting tools. As such, users with specialized needs can leverage it to quickly create entirely new history rewriting tools.

filter-repo is a single-file python script, depending only on the python standard library (and execution of git commands), all of which is designed to make build/installation trivial: just copy it into your $PATH.

Table of Contents

• Background
  • Why filter-repo instead of filter-branch?
  • Example usage, comparing to filter-branch
  • Design rationale behind filter-repo
• Usage
  • The bigger picture
  • Examples
    • Path based filtering
    • Content based filtering
    • Refname based filtering
    • User and email based filtering
    • Callbacks
    • Using filter-repo as a library
• Internals
  • How filter-repo works
  • Limitations
    • Inherited limitations
    • Intrinsic limitations
    • Issues specific to filter-repo
    • Comments on reversibility

Background

Why filter-repo instead of filter-branch?

filter-branch has a number of problems:

• filter-branch is extremely to unusably slow (multiple orders of magnitude slower than it should be) for non-trivial repositories.

• filter-branch made a number of usability choices that are okay for small repos, but these choices sometimes conflict as more options are combined, and the overall usability often causes difficulties for users trying to work with intermediate or larger repos.

• filter-branch is missing some basic features.

The first two are intrinsic to filter-branch's design at this point and cannot be backward-compatibly fixed.

Example usage, comparing to filter-branch

Let's say that we want to extract a piece of a repository, with the intent on merging just that piece into some other bigger repo. For extraction, we want to:

• extract the history of a single directory, src/. This means that only paths under src/ remain in the repo, and any commits that only touched paths outside this directory will be removed.
• rename all files to have a new leading directory, my-module/ (e.g. so that src/foo.c becomes my-module/src/foo.c)
• rename any tags in the extracted repository to have a 'my-module-' prefix (to avoid any conflicts when we later merge this repo into something else)

Doing this with filter-repo is as simple as the following command:

  git filter-repo --path src/ --to-subdirectory-filter my-module --tag-rename '':'my-module-'

(the single quotes are unnecessary, but make it clearer to a human that we are replacing the empty string as a prefix with my-module-)

By contrast, filter-branch comes with a pile of caveats (more on that below) even once you figure out the necessary invocation(s):

  git filter-branch --tree-filter 'mkdir -p my-module && git ls-files | grep -v ^src/ | xargs git rm -f -q && ls -d * | grep -v my-module | xargs -I files mv files my-module/' --tag-name-filter 'echo "my-module-$(cat)"' --prune-empty -- --all
  git clone file://$(pwd) newcopy
  cd newcopy
  git for-each-ref --format="delete %(refname)" refs/tags/ | grep -v refs/tags/my-module- | git update-ref --stdin
  git gc --prune=now

Some might notice that the above filter-branch invocation will be really slow due to using --tree-filter; you could alternatively use the --index-filter option of filter-branch, changing the above commands to:

  git filter-branch --index-filter 'git ls-files | grep -v ^src/ | xargs git rm -q --cached; git ls-files -s | sed "s-$(printf \\t)-&my-module/-" | git update-index --index-info; git ls-files | grep -v ^my-module/ | xargs git rm -q --cached' --tag-name-filter 'echo "my-module-$(cat)"' --prune-empty -- --all
  git clone file://$(pwd) newcopy
  cd newcopy
  git for-each-ref --format="delete %(refname)" refs/tags/ | grep -v refs/tags/my-module- | git update-ref --stdin
  git gc --prune=now

However, for either filter-branch command there are a pile of caveats. First, some may be wondering why I list five commands here for filter-branch. Despite the use of --all and --tag-name-filter, and filter-branch's manpage claiming that a clone is enough to get rid of old objects, the extra steps to delete the other tags and do another gc are still required to clean out the old objects and avoid mixing new and old history before pushing somewhere. Other caveats:

• Commit messages are not rewritten; so if some of your commit messages refer to prior commits by (abbreviated) sha1, after the rewrite those messages will now refer to commits that are no longer part of the history. It would be better to rewrite those (abbreviated) sha1 references to refer to the new commit ids.
• The --prune-empty flag sometimes misses commits that should be pruned, and it will also prune commits that started empty rather than just ended empty due to filtering. For repositories that intentionally use empty commits for versioning and publishing related purposes, this can be detrimental.
• The commands above are OS-specific. GNU vs. BSD issues for sed, xargs, and other commands often trip up users; I think I failed to get most folks to use --index-filter since the only example in the filter-branch manpage that both uses it and shows how to move everything into a subdirectory is linux-specific, and it is not obvious to the reader that it has a portability issue since it silently misbehaves rather than failing loudly.
• The --index-filter version of the filter-branch command may be two to three times faster than the --tree-filter version, but both filter-branch commands are going to be multiple orders of magnitude slower than filter-repo.

Design rationale behind filter-repo (why create a new tool?)

None of the existing repository filtering tools do what I want. They're all good in their own way, but come up short for my needs. No tool provided any of the first eight traits below I wanted, and all failed to provide at least one of the last four traits as well:

1. [Starting report] Provide user an analysis of their repo to help them get started on what to prune or rename, instead of expecting them to guess or find other tools to figure it out. (Triggered, e.g. by running the first time with a special flag, such as --analyze.)

2. [Keep vs. remove] Instead of just providing a way for users to easily remove selected paths, also provide flags for users to only keep certain paths. Sure, users could workaround this by specifying to remove all paths other than the ones they want to keep, but the need to specify all paths that ever existed in any version of the repository could sometimes be quite painful. For filter-branch, using pipelines like git ls-files | grep -v ... | xargs -r git rm might be a reasonable workaround but can get unwieldy and isn't as straightforward for users; plus those commands are often operating-system specific (can you spot the GNUism in the snippet I provided?).

3. [Renaming] It should be easy to rename paths. For example, in addition to allowing one to treat some subdirectory as the root of the repository, also provide options for users to make the root of the repository just become a subdirectory. And more generally allow files and directories to be easily renamed. Provide sanity checks if renaming causes multiple files to exist at the same path. (And add special handling so that if a commit merely copied oldname->newname without modification, then filtering oldname->newname doesn't trigger the sanity check and die on that commit.)

4. [More intelligent safety] Writing copies of the original refs to a special namespace within the repo does not provide a user-friendly recovery mechanism. Many would struggle to recover using that. Almost everyone I've ever seen do a repository filtering operation has done so with a fresh clone, because wiping out the clone in case of error is a vastly easier recovery mechanism. Strongly encourage that workflow by detecting and bailing if we're not in a fresh clone, unless the user overrides with --force.

5. [Auto shrink] Automatically remove old cruft and repack the repository for the user after filtering (unless overridden); this simplifies things for the user, helps avoid mixing old and new history together, and avoids problems where the multi-step process for shrinking the repo documented in the manpage doesn't actually work in some cases. (I'm looking at you, filter-branch.)

6. [Clean separation] Avoid confusing users (and prevent accidental re-pushing of old stuff) due to mixing old repo and rewritten repo together. (This is particularly a problem with filter-branch when using the --tag-name-filter option, and sometimes also an issue when only filtering a subset of branches.)

7. [Versatility] Provide the user the ability to extend the tool or even write new tools that leverage existing capabilities, and provide this extensibility in a way that (a) avoids the need to fork separate processes (which would destroy performance), (b) avoids making the user specify OS-dependent shell commands (which would prevent users from sharing commands with each other), (c) takes advantage of rich data structures (because hashes, dicts, lists, and arrays are prohibitively difficult in shell) and (d) provides reasonable string manipulation capabilities (which are sorely lacking in shell).

8. [Old commit references] Provide a way for users to use old commit IDs with the new repository (in particular via mapping from old to new hashes with refs/replace/ references).

9. [Commit message consistency] If commit messages refer to other commits by ID (e.g. "this reverts commit 01234567890abcdef", "In commit 0013deadbeef9a..."), those commit messages should be rewritten to refer to the new commit IDs.

10. [Become-empty pruning] Commits which become empty due to filtering should be pruned. If the parent of a commit is pruned, the first non-pruned ancestor needs to become the new parent. If no non-pruned ancestor exists and the commit was not a merge, then it becomes a new root commit. If no non-pruned ancestor exists and the commit was a merge, then the merge will have one less parent (and thus make it likely to become a non-merge commit which would itself be pruned if it had no file changes of its own). One special thing to note here is that we prune commits which become empty, NOT commits which start empty. Some projects intentionally create empty commits for versioning or publishing reasons, and these should not be removed. (As a special case, commits which started empty but whose parent was pruned away will also be considered to have "become empty".)

11. [Become-degenerate pruning] Pruning of commits which become empty can potentially cause topology changes, and there are lots of special cases. Normally, merge commits are not removed since they are needed to preserve the graph topology, but the pruning of parents and other ancestors can ultimately result in the loss of one or more parents. A simple case was already noted above: if a merge commit loses enough parents to become a non-merge commit and it has no file changes, then it too can be pruned. Merge commits can also have a topology that becomes degenerate: it could end up with the merge_base serving as both parents (if all intervening commits from the original repo were pruned), or it could end up with one parent which is an ancestor of its other parent. In such cases, if the merge has no file changes of its own, then the merge commit can also be pruned. However, much as we do with empty pruning we do not prune merge commits that started degenerate (which indicates it may have been intentional, such as with --no-ff merges) but only merge commits that become degenerate and have no file changes of their own.

12. [Speed] Filtering should be reasonably fast

Usage

The bigger picture

Using filter-repo is relatively simple, but rewriting history is part of a larger discussion in terms of collaboration. When you rewrite history, the old and new histories are no longer compatible; if you push this history somewhere for others to view, it will look as though you've done a rebase of all branches and tags. Make sure you are familiar with the "Recovering from upstream rebase" section of git-rebase(1) (and in particular, "The hard case") before proceeding, in addition to this section.

Steps to use filter-repo as part of the bigger picture of doing a history rewrite are roughly as follows:

1. Create a clone of your repository (if you created special refs outside of refs/heads/ or refs/tags/, make sure to fetch those too). Note that --bare and --mirror clones are supported too, if you prefer.

2. (Optional) Run git filter-repo --analyze. This will create a directory of reports mentioning renames that have occurred in your repo and also listing sizes of objects aggregated by path/directory/extension/blob-id; this information may be useful in choosing how to filter your repo. It can also be useful to re-run --analyze after filtering to verify the changes look correct.

3. Run filter-repo with your desired filtering options. Many examples are given below. For more complex cases, note that doing the filtering in multiple steps (by running multiple filter-repo invocations in a sequence) is supported. If anything goes wrong here, simply delete your clone and restart.

4. Push your new repository to its new home (note that refs/remotes/origin/* will have been moved to refs/heads/* as the first part of filter-repo, so you can just deal with normal branches instead of remote tracking branches). While you can force push this to the same URL you cloned from, there are good reasons to consider pushing to a different location instead:

   1. People who cloned from the original repo will have old history. When they fetch the new history you force pushed up, unless they do a git reset --hard @{u} on their branches or rebase their local work, git will think they have hundreds or thousands of commits with very similar commit messages as what exist upstream (but which include files you wanted excised from history), and allow the user to merge the two histories, resulting in what looks like two copies of each commit. If they then push this history back up, then everyone now has history with two copies of each commit and the bad files have returned. You're more likely to succeed in forcing people to get rid of the old history if they have to clone a new URL.

   2. Rewriting history will rewrite tags; those who have already downloaded tags will not get the updated tags by default (see the "On Re-tagging" section of the git-tag(1) manpage). Every user trying to use an existing clone will have to forcibly delete all tags and re-fetch them; it may be easier for them to just re-clone, which they are more likely to do with a new clone URL.

   3. Rewriting history may delete some refs (e.g. branches that only had files that you wanted excised from history); unless you run git push with the --mirror or --prune options, those refs will continue to exist on the server. If folks then merge these branches into others, then people have started mixing old and new history. If users had already cloned these branches, removing them from the server isn't enough; you need all users to delete any local branches based on these refs and run fetch with the --prune option as well. Simply re-cloning from a new URL is easier.

   4. The server may not allow you to force push over some refs. For example, code review systems may have special ref namespaces (e.g. refs/changes/, refs/pull/, refs/merge-requests/) that they have locked down.

5. (Optional) Some additional considerations

   1. filter-repo by default creates replace refs (see git-replace(1)) for each rewritten commit ID, allowing you to use old (unabbreviated) commit hashes to refer to the newly rewritten commits. If you want to use these replace refs, push them to the relevant clone URL and tell users to adjust their fetch refspec (e.g. git config --add remote.origin.fetch +refs/replace/*:refs/replace/*) Sadly, some existing git servers (e.g. Gerrit, GitHub) do not yet understand replace refs, and thus one can't use old commit hashes within their UI; this may change in the future. But replace refs at least help users locally within the git CLI.

   2. If you have a central repo, you may want to prevent people from pushing old commit IDs, in order to avoid mixing old and new history. Every repository manager does this differently, some provide specialized commands, others require you to write hooks.

Examples

Path based filtering

To only keep the 'README.md' file plus the directories 'guides' and 'tools/releases/':

  git filter-repo --path README.md --path guides/ --path tools/releases

Directory names can be given with or without a trailing slash, and all filenames are relative to the toplevel of the repo. To keep all files except these paths, just add --invert-paths:

  git filter-repo --path README.md --path guides/ --path tools/releases --invert-paths

If you want to have both an inclusion filter and an exclusion filter, just run filter-repo multiple times. For example, to keep the src/main subdirectory but exclude files under src/main named 'data', run:

  git filter-repo --path src/main/
  git filter-repo --path-glob 'src/*/data' --invert-paths

Note that the asterisk ('*') will match across multiple directories, so the second command would remove e.g. src/main/org/whatever/data. Also, the second command by itself would also remove e.g. src/not-main/foo/data, but since src/not-main/ was removed by the first command, that's not an issue. Also, the use of quotes around the asterisk is sometimes important to avoid glob expansion by the shell.

You can also select paths by regular expression. For example, to only include files from the repo whose name is in the format YYYY-MM-DD.txt and is found at least two subdirectories deep:

  git filter-repo --path-regex '^.*/.*/[0-9]{4}-[0-9]{2}-[0-9]{2}.txt$'

If you want two directories to be renamed (and maybe merged if both are renamed to the same location), use --path-rename; for example, to rename both 'cmds/' and 'src/scripts/' to 'tools/':

  git filter-repo --path-rename cmds:tools --path-rename src/scripts/:tools/

As with --path, directories can be specified with or without a trailing slash for --path-rename.

If you do a --path-rename to something that was already in use, it will be silently overwritten. However, if you try to rename multiple files to the same location (e.g. src/scripts/run_release.sh and cmds/run_release.sh both existed and had different content with the renames above), then you will be given an error. If you have such a case, you may want to add another rename command to move one of the paths somewhere else where it won't collide:

  git filter-repo --path-rename cmds/run_release.sh:tools/do_release.sh \
                  --path-rename cmds/:tools/ \
                  --path-rename src/scripts/:tools/

Also, --path-rename brings up ordering issues; all path arguments are applied in order. Thus, a command like

  git filter-repo --path-rename sources/:src/main/ --path src/main/

would make sense but reversing the two arguments would not (src/main/ is created by the rename so reversing the two would give you an empty repo). Also, note that the rename of cmds/run_release.sh a couple examples ago was done before the other renames.

If you prefer to filter based solely on basename, use the --use-base-name flag (though this is incompatible with --path-rename). For example, to only include README.md and Makefile files from any directory:

  git filter-repo --use-base-name --path README.md --path Makefile

If you wanted to delete all .DS_Store files in any directory, you could either use:

  git filter-repo --invert-paths --path '.DS_Store' --use-base-name

or

  git filter-repo --invert-paths --path-glob '*/.DS_Store' --path '.DS_Store'

(the --path-glob isn't sufficient by itself as it might miss a toplevel .DS_Store file; further while something like --path-glob '*.DS_Store' would workaround that problem it would also grab files named 'foo.DS_Store' or 'bar/baz.DS_Store')

If you have a long list of files, directories, globs, or regular expressions to filter on, you can stick them in a file and use --paths-from-file; for example, with a file named stuff-i-want.txt with contents of

README.md
guides/
tools/releases
glob:*.py
regex:^.*/.*/[0-9]{4}-[0-9]{2}-[0-9]{2}.txt$
tools/==>scripts/
regex:(.*)/([^/]*)/([^/]*)\.text$==>\2/\1/\3.txt

then you could run

  git filter-repo --paths-from-file stuff-i-want.txt

to get a repo containing only the toplevel README.md file, the guides/ and tools/releases/ directories, all python files, files whose name was of the form YYYY.MM-DD.txt at least two subdirectories deep, and would rename tools/ to scripts/ and rename files like foo/bar/baz/bleh.text to baz/foo/bar/bleh.txt. Note the special line prefixes of glob: and regex: and the special string ==> denoting renames.

Finally, see also the --filename-callback from the callbacks section.

Content based filtering

If you want to filter out all files bigger than a certain size, you can use --strip-blobs-bigger-than with some size (K, M, and G suffixes are recognized), e.g.:

  git filter-repo --strip-blobs-bigger-than 10M

If you want to strip out all files with specified git object ids (hashes), list the hashes in a file and run

  git filter-repo --strip-blobs-with-ids FILE_WITH_GIT_BLOB_IDS

If you want to modify file contents, you can do so based on a list of expressions in a file, one per line. For example, with a file named expressions.txt containing

p455w0rd
foo==>bar
glob:*666*==>
regex:\bdriver\b==>pilot
literal:MM/DD/YYYY=>YYYY-MM-DD
regex:([0-9]{2})/([0-9]{2})/([0-9]{4})==>\3-\1-\2

then running

  git filter-repo --replace-text expressions.txt

will go through and replace p455w0rd with ***REMOVED***, foo with bar, any line containing 666 with a blank line, the word driver with pilot (but not if it has letters before or after; e.g. drivers will be unmodified), replace the exact text MM/DD/YYYY with YYYY-MM-DD and replace date strings of the form MM/DD/YYYY with ones of the form YYYY-MM-DD. In the expressions file, there are a few things to note:

• Every line has a replacement, given by whatever is on the right of ==>. If ==> does not appear on the line, the default replacement is ***REMOVED***.
• Lines can start with literal:, glob:, or regex: to specify whether to do literal string matches, globs, or regular expressions. If none of these are specified, literal: is assumed.
• globs and regexes are applied to each line of the file; it is not possible with --replace-text to match a multi-line string.
• If multiple matches are found on a line, all are replaced.

See also the --blob-callback from the callbacks section.

Refname based filtering

To rename tags, use --tag-rename, e.g.:

  git filter-repo --tag-rename foo:bar

This will rename any tags starting with foo to now start with bar. Either side of the colon could be blank, e.g.

  git filter-repo --tag-rename '':'my-module-'

For more general refname modification, see --refname-callback from the callbacks section.

User and email based filtering

To modify username and emails of commits, you can create a mailmap file in the format accepted by git-shortlog(1). For example, if you have a file named my-mailmap you can run

  git filter-repo --mailmap my-mailmap

and if the current contents of that file are as follows (if the specified mailmap file is version controlled, historical versions of the file are ignored):

Name For User <email@addre.ss>
<new@ema.il> <old1@ema.il>
New Name And <new@ema.il> <old2@ema.il>
New Name And <new@ema.il> Old Name And <old3@ema.il>

then we can update username and/or emails based on the specified mapping.

See also the --name-callback and --email-callback from the callbacks section.

Callbacks

For flexibility, filter-repo allows you to specify functions on the command line to further filter all changes. Please note that there are some API compatibility caveats associated with these callbacks that you should be aware of before using them.

All callback functions are of the same general format. For a command line argument like

  --foo-callback 'BODY'

the following code will be compiled and called:

  def foo_callback(foo):
    BODY

Thus, you just need to make sure your BODY modifies and returns foo appropriately. One important thing to note for all callbacks is that filter-repo uses bytestrings everywhere instead of strings.

There are three callbacks that allow you to operate directly on raw objects that contain data that's easy to write in fast-import(1) format:

  --blob-callback
  --commit-callback
  --tag-callback
  --reset-callback

We'll come back to these later because it is often the case that the other callbacks are more convenient. The other callbacks operate on a small piece of the raw objects or operate on pieces across multiple types of raw object (e.g. author names and committer names and tagger names across commits and tags, or refnames across commits, tags, and resets, or messages across commits and tags). The convenience callbacks are:

  --filename-callback
  --message-callback
  --name-callback
  --email-callback
  --refname-callback

in each you are expected to simply return a new value based on the one passed in. For example,

  git-filter-repo --name-callback 'return name.replace(b"Wiliam", b"William")'

would result in the following function being called:

  def name_callback(name):
    return name.replace(b"Wiliam", b"William")

The email callback is quite similar:

  git-filter-repo --email-callback 'return email.replace(b".cm", b".com")'

The refname callback is also similar, but note that the refname passed in and returned are expected to be fully qualified (e.g. b"refs/heads/master" instead of just b"master" and b"refs/tags/v1.0.7" instead of b"1.0.7"):

  git-filter-repo --refname-callback '
    # Change e.g. refs/heads/master to refs/heads/prefix-master
    rdir,rpath = os.path.split(refname)
    return rdir + b"/prefix-" + rpath'

The message callback is quite similar to the previous three callbacks, though it operates on a bytestring that is likely more than one line:

  git-filter-repo --message-callback '
    if b"Signed-off-by:" not in message:
      message += b"\nSigned-off-by: Me My <self@and.eye>"
    return re.sub(b"[Ee]-?[Mm][Aa][Ii][Ll]", b"email", message)'

The filename callback is slightly more interesting. Returning None means the file should be removed from all commits, returning the filename unmodified marks the file to be kept, and returning a different name means the file should be renamed. An example:

  git-filter-repo --filename-callback '
    if b"/src/" in filename:
      # Remove all files with a directory named "src" in their path
      # (except when "src" appears at the toplevel).
      return None
    elif filename.startswith(b"tools/"):
      # Rename tools/ -> scripts/misc/
      return b"scripts/misc/" + filename[6:]
    else:
      # Keep the filename and do not rename it
      return filename
    '

In contrast, the blob, reset, tag, and commit callbacks are not expected to return a value, but are instead expected to modify the object passed in. Major fields for these objects are (subject to API backward compatibility caveats mentioned previously):

• Blob: original_id (original hash) and data
• Reset: ref (name of reference) and from_ref (hash or integer mark)
• Tag: ref, from_ref, original_id, tagger_name, tagger_email, tagger_date, message
• Commit: branch, original_id, author_name, author_email, author_date, committer_name, committer_email, committer_date , message, file_changes (list of FileChange objects, each containing a type, filename, mode, and blob_id), parents (list of hashes or integer marks)

An example of each:

  git filter-repo --blob-callback '
    if len(blob.data) > 25:
      # Mark this blob for removal from all commits
      blob.skip()
    else:
      blob.data = blob.data.sub(b"Hello", b"Goodbye")
    '

  git filter-repo --reset-callback 'reset.ref = reset.ref.replace(b"master", b"dev")'

  git filter-repo --tag-callback '
    if tag.tagger_name == "Jim Williams":
      # Omit this tag
      tag.skip()
    else:
      tag.message = tag.message + b"\n\nTag of %s by %s on %s" % (tag.ref, tag.tagger_email, tag.tagger_date)'

  git filter-repo --commit-callback '
    # Remove executable files with three 6s in their name (including
    # from leading directories).
    # Also, undo deletion of sources/foo/bar.txt (change types are either
    # b"D" (deletion) or b"M" (add or modify); renames are handled by deleting
    # the old file and adding a new one)
    commit.file_changes = [change for change in commit.file_changes
                           if not (change.mode == b"100755" and
			           change.filename.count(b"6") == 3) and
			      not (change.type == b"D" and
			           change.filename == b"sources/foo/bar.txt")]
    # Mark all .sh files as executable; modes in git are always one of
    # 100644 (normal file), 100755 (executable), 120000 (symlink), or
    # 160000 (submodule)
    for change in commit.file_changes:
      if change.filename.endswith(b".sh"):
        change.mode = b"100755"
    '

Using filter-repo as a library

git-filter-repo can also be imported as a library in Python, allowing for further flexibility. Some simple examples exist in the testsuite. For this to work, the symlink to git-filter-repo named git_filter_repo.py either needs to have been installed in your $PYTHONPATH, or you need to create a symlink to (or a copy of) git-filter-repo named git_filter_repo.py and stick it in your $PYTHONPATH.

Internals

You probably don't need to read this section unless you are just very curious or you are trying to do a very complex history rewrite.

How filter-repo works

Roughly, filter-repo works by running

   git fast-export <options> | filter | git fast-import <options>

where filter-repo not only launches the whole pipeline but also serves as the filter in the middle. However, filter-repo does a few additional things on top in order to make it into a well-rounded filtering tool. A sequence that more accurately reflects what filter-repo runs is:

1. Verify we're in a fresh clone
2. git fetch -u . refs/remotes/origin/*:refs/heads/*
3. git remote rm origin
4. git fast-export --show-original-ids --fake-missing-tagger --signed-tags=strip --tag-of-filtered-object=rewrite --use-done-feature --no-data --reencode=yes --all | filter | git fast-import --force --quiet
5. git update-ref --no-deref --stdin, fed with a list of refs to nuke, and a list of replace refs to delete, create, or update.
6. git reset --hard
7. git reflog expire --expire=now --all
8. git gc --prune=now

Some notes or exceptions on each of the above:

1. If we're not in a fresh clone, users will not be able to recover if they used the wrong command or ran in the wrong repo. (Though --force overrides this check, and it's also off if you've already ran filter-repo once in this repo.)
2. Technically, we actually use a git update-ref command fed with a lot of input due to the fact that users can use --force when local branches might not match remote branches. But this fetch command catches the intent rather succinctly.
3. We don't want users accidentally pushing back to the original repo, as discussed in the section on the bigger picture. It also reminds users that since history has been rewritten, this repo is no longer compatible with the original. Finally, another minor benefit is this allows users to push with the --mirror option to their new home without accidentally sending remote tracking branches.
4. Some of these flags are always used but others are actually conditional. For example, filter-repo's --replace-text and --blob-callback options need to work on blobs so --no-data cannot be passed to fast-export. But when we don't need to work on blobs, passing --no-data speeds things up. Also, other flags may change the structure of the pipeline as well (e.g. --dry-run and --debug)
5. Selection of files based on paths could cause every commit in the history of a branch or tag to be pruned, resulting in the branch or tag needing to be pruned. However, filter-repo just works by stripping out the 'commit' and 'tag' directives for each one that's not needed, meaning fast-import won't do the branch or tag deletion for us. So we do it in a post-processing step to ensure we avoid mixing old and new history. Also, we use this step to write replace refs for accessing the newly written commit hashes using their previous names.
6. Users also have old versions of files in their working tree and index; we want those cleaned up to match the rewritten history as well. Note that this step is skipped in bare repos.
7. Reflogs will hold on to old history, so we need to expire them.
8. We need to gc to avoid mixing new and old history. Also, it shrinks the repository for users, so they don't have to do extra work. (Odds are that they've only rewritten trees and commits and maybe a few blobs, so --aggressive isn't needed and would be too slow.)

Information about these steps is printed out when --debug is passed to filter-repo.

Limitations

Inherited limitations

Since git filter-repo calls fast-export and fast-import to do a lot of the heavy lifting, it inherits limitations from those systems:

• extended commit headers, if any, are stripped
• commits get rewritten meaning they will have new hashes; therefore, signatures on commits and tags cannot continue to work and instead are just removed (thus signed tags become annotated tags)
• tags of commits are supported; tags of anything else (blobs, trees, or tags) are not. (fast-export aborts on tags of blobs and tags of tags, and simply ignores tags of trees with a warning.)
• annotated and signed tags outside of the refs/tags/ namespace are not supported (their location will be mangled in weird ways)
• fast-import will die on various forms of invalid input, such as a timezone with more than four digits
• fast-export cannot reencode commit messages into UTF-8 if the commit message is not valid in its specified encoding (in such cases, it'll leave the commit message and the encoding header alone).
• commits without an author will be given one matching the committer
• tags without a tagger will be given a fake tagger

There are also some limitations due to the design of these systems:

• Trying to insert additional files into the stream can be tricky; since fast-export only lists file changes in a merge relative to its first parent, if you insert additional files into a commit that is in the second (or third or fourth) parent history of a merge, then you also need to add it to the merge manually.

• fast-export and fast-import work with exact file contents, not patches. (e.g. "Whatever the current contents of this file, update them to now have these contents") Because of this, removing the changes made in a single commit or inserting additional changes to a file in some commit and expecting them to propagate forward is not something that can be done with these tools. Use git-rebase(1) for that.

Intrinsic limitations

Some types of filtering have limitations that would affect any tool attempting to perform them; the most any tool can do is attempt to notify the user when it detects an issue:

• When rewriting commit hashes in commit messages, there are a variety of cases when the hash will not be updated (whenever this happens, a note is written to .git/filter-repo/suboptimal-issues):

  • if a commit hash does not correspond to a commit in the old repo
  • if a commit hash corresponds to a commit that gets pruned
  • if an abbreviated hash is not unique

• Pruning of empty commits can cause a merge commit to lose an entire ancestry line and become a non-merge. If the merge commit had no changes then it can be pruned too, but if it still has changes it needs to be kept. This might cause minor confusion since the commit will likely have a commit message that makes it sound like a merge commit even though it's not. (Whenever a merge commit becomes a non-merge commit, a note is written to .git/filter-repo/suboptimal-issues)

Issues specific to filter-repo

• Multiple repositories in the wild have been observed which use a bogus timezone (+051800); google will find you some reports. The intended timezone wasn't clear or wasn't always the same. Replace with a different bogus timezone that fast-import will accept (+0261).

• --path-rename can result in pathname collisions; to avoid excessive memory requirements of tracking which files are in all commits or looking up what files exist with either every commit or every usage of --path-rename, we just tell the user that they might clobber other changes if they aren't careful. We can check if the clobbering comes from another --path-rename without much overhead. (Perhaps in the future it's worth adding a slow mode to --path-rename that will do the more exhaustive checks?)

• There is no mechanism for directly controlling which flags are passed to fast-export (or fast-import); only pre-defined flags can be turned on or off as a side-effect of other options. Direct control would make little sense because some options like --full-tree would require additional code in filter-repo (to parse new directives), and others such as -M or -C would break assumptions used in other places of filter-repo.

Comments on reversibility

Some people are interested in reversibility of of a rewrite; e.g. rewrite history, possibly add some commits, then unrewrite and get the original history back plus a few new "unrewritten" commits. Obviously this is impossible if your rewrite involves throwing away information (e.g. filtering out files or replacing several different strings with ***REMOVED***), but may be possible with some rewrites. filter-repo is likely to be a poor fit for this type of workflow for a few reasons:

• most of the limitations inherited from fast-export and fast-import are of a type that cause reversibility issues
• grafts and replace refs, if present, are used in the rewrite and made permanent
• rewriting of commit hashes will probably be reversible, but it is possible for rewritten abbreviated hashes to not be unique even if the original abbreviated hashes were.
• filter-repo defaults to several forms of unreversible rewriting that you may need to turn off (e.g. the last two bullet points above or reencoding commit messages into UTF-8); it's possible that additional forms of unreversible rewrites will be added in the future.
• I assume that people use filter-repo for one-shot conversions, not ongoing data transfers. I explicitly reserve the right to change any API in filter-repo based on this presumption. You have been warned.