The usual workflow is: 1. clone or fetch; 2. make and commit changes in git as desired; 3. run dgit build or dgit sbuild, or generate the source and binary packages for upload some other way; 4. do pre-upload tests of the proposed upload; 5. run dgit push.
The suite's git tip is left on the local branch dgit/suite ready for work, and on the corresponding dgit remote tracking branch. Also, the origin remote will be set up to point to the package's dgit-repos tree for the distro to which suite belongs.
In more detail: dgit push checks that the current HEAD corresponds to the .dsc. It then pushes the HEAD to the suite's dgit-repos branch, makes a signed git tag, edits the .dsc to contain the dgit metadata field, runs debsign to sign the upload (.dsc and .changes), pushes the signed tag, and finally uses dput to upload the .changes to the archive.
For a format `3.0 (quilt)' source package, dgit push may also have to make a commit on your current branch to contain quilt metadata. It will do this automatically if necessary. You can explicitly request that dgit do just this dgit quilt-fixup.
dgit push always uses the package, suite and version specified in the debian/changelog and the .dsc, which must agree.
If dgit push fails while uploading, it is fine to simply retry the dput on the .changes file at your leisure.
It is not normally necessary to run dgit quilt-fixup explicitly; where necessary it is done as part of dgit push.
If you are lucky the other uploaders have also used dgit and integrated the other relevant git history; if not you can fetch it into your tree and cherry-pick etc. as you wish.
However the first time dgit is used it will generate a new origin commit from the archive which won't be linked into the rest of your git history. You will need to merge this.
If last upload was in fact made with git, you should usually proceed as follows: identify the commit which was actually used to build the package. (Hopefully you have a tag for this.) Check out the dgit branch (git checkout dgit/suite) and merge that other commit (git merge debian/version). Hopefully this merge will be trivial because the two trees should be the same. The resulting branch head can be merged into your working branches (git checkout master && git merge dgit/suite).
If last upload was not made with git, a different approach is required to start using dgit. First, do dgit fetch (or clone) to obtain a git history representation of what's in the archive and record it in the remotes/dgit/dgit/suite tracking branch. Then somehow, using your other git history plus appropriate diffs and cherry picks from the dgit remote tracking branch, construct a git commit whose tree corresponds to the tree to use for the next upload. If that commit-to-be-uploaded is not a descendant of the dig remote tracking branch, check it out and say git merge -s ours remotes/dgit/dgit/suite; that tells git that we are deliberately throwing away any differences between what's in the archive and what you intend to upload. Then run dgit push to actually upload the result.
dgit maintains a pseudo-remote called dgit, with one branch per suite. This remote cannot be used with plain git.
The dgit-repos repository for each package contains one ref per suite named refs/dgit/suite. These should be pushed to only by dgit. They are fast forwarding. Each push on this branch corresponds to an upload (or attempted upload).
However, it is perfectly fine to have other branches in dgit-repos; normally the dgit-repos repo for the package will be accessible via the remote name `origin'.
dgit push will also (by default) make signed tags called debian/version and push them to dgit-repos, but nothing depends on these tags existing.
dgit push can operate on any commit which is a descendant of the current dgit/suite tip in dgit-repos.
Uploads made by dgit contain an additional field Vcs-Dgit-Master in the source package .dsc. (This is added by dgit push.) This specifies a commit (an ancestor of the dgit/suite branch) whose tree is identical to the unpacked source upload.
Uploads not made by dgit are represented in git by commits which are synthesised by dgit. The tree of each such commit corresponds to the unpacked source; there is an origin commit with the contents, and a psuedo-merge from last known upload - that is, from the contents of the dgit/suite branch.
dgit expects repos that it works with to have a dgit remote. This refers to the well-known dgit-repos location (currently, the dgit-repos project on Alioth). dgit fetch updates the remote tracking branch for dgit/suite.
dgit does not (currently) represent the orig tarball(s) in git; nor does it represent the patch statck of a `3.0 (quilt)' package. The orig tarballs are downloaded and kept in the parent directory, as with a traditional (non-gitish) dpkg-source workflow.
To a user looking at the archive, changes pushed using dgit look like changes made in an NMU: in a `3.0 (quilt)' package the delta from the previous upload is recorded in a new patch constructed by dpkg-source.
If you are the maintainer of a non-native package, you currently have two sensible options:
Firstly, you can regard your git history as primary, and the archive as an export format. For example, you could maintain topic branches in git and a fast-forwarding release branch; or you could do your work directly in a merging way on the dgit/suite branches. If you do this you should probably use a `1.0' format source package. In the archive, the delta between upstream will be represented in the single Debian patch.
Secondly, you can regard your quiltish patch stack in the archive as primary. You will have to use other tools besides dgit to import and export this patch stack. For `3.0 (quilt)' packages, dgit has to do more work to work around some braindamage in way dpkg-source handles changes made to this format. See also the BUGS section. We recommend against the use of `3.0 (quilt)'.
If your suite is part of a distro that dgit already knows about, you can use this option to make dgit work even if your dgit doesn't know about the suite. For example, specifying -ddebian will work when the suite is an unknown suite in the Debian archive.
To define a new distro it is necessary to define methods and URLs for fetching (and, for dgit push, altering) a variety of information both in the archive and in dgit-repos. How to do this is not yet documented, and currently the arrangements are unpleasant. See BUGS.
Debian Policy needs to be updated to describe the new Vcs-Dgit-Master field (and to specify that it is an RC bug for that field to refer to an unavailable commit).
The method of canonicalising suite names is bizarre. See the --existing-package option for one of the implications.
dgit push should perhaps do `git push origin', or something similar, by default.
Debian does not have a working rmadison server, so to find out what version of a package is in the archive, or to canonicalise suite names, we ssh directly into the ftpmaster server.
The mechanism for checking for and creating per-package repos on alioth is a hideous bodge. One consequence is that dgit currently only works for people with push access.
Debian Maintainers are currently not able to push, as there is not currently any mechanism for determining and honouring the archive's ideas about access control. Currently only DDs can push.
dgit's representation of format `3.0 (quilt)' source packages does not represent the patch stack. Currently the patch series representation cannot round trip through the archive. Ideally dgit would represent a quilty package with an origin commit of some kind followed by the patch stack as a series of commits followed by a pseudo-merge (to make the branch fast-forwarding). This would also mean a new `dgit rebase-prep' command or some such to turn such a fast-forwarding branch back into a rebasing patch stack, and a `force' option to dgit push (perhaps enabled automatically by a note left by rebase-prep) which will make the required pseudo-merge.
If the dgit push fails halfway through, it should be restartable and idempotent. However this is not true for the git tag operation. Also, it would be good to check that the proposed signing key is available before starting work.
dgit's handling of .orig.tar.gz is not very sophisticated. Ideally the .orig.tar.gz could be transported via the git repo as git tags. Doing this is made more complicated by the possibility of a `3.0 (quilt)' package with multiple .orig tarballs.
`3.0 (quilt)' packages have an additional difficulty: if these are edited in the most normal way, and then fed to dpkg-buildpackage, dpkg-source will add extra quilt patch metadata to the source tree during the source package build. This extra metadata is then of course not included in the git history. So dgit push needs to commit it for you, to make sure that the git history and archive contents are identical. That this is necessary is a bug in the `3.0 (quilt)' format.
There should be an option which arranges for the `3.0 (quilt)' autocommit to not appear on your HEAD, but instead only in the remote tracking suite branch.
There should at the very least be some advice in the manpage about how to use dgit when the signing key is not available on the same machine as the build host.
The option parser requires values to be cuddled to the option name.
dgit assumes knowledge of the archive layout. There appears to be no sane way to find the path in the archive pool of the .dsc for a particular suite. I'm assured that the archive layout is a `well known algorithm' by now.
--dry-run often does not work with fetch, even though this is a logically plausible request. (It fails, instead.)