When an existing `headRepo` couldn't be detected, it's time to auto-fork
one. Unfortunately, an obscure Go behavior made it seem like `headRepo`
was a non-nil value, where in fact it did contain a nil pointer which
would crash the process.
This avoids ever assigning nil pointers to `var headRepo ghrepo.Interface`.
This code was put in place in preparation for a feature that never
shipped. Namely, we wanted to use the commit hash for the base branch so
we can get an accurate `git log` involving the changes in a pull
request. However, getting the commit hash from API is not the way to go
because the latest commit might not be available in the person's local
repository, and using a local tracking branch for base such as
`origin/master` works quite well in most cases without dereferencing it.
A branch is considered already pushed if its HEAD commit hash matches
the commit hash of one of the remote tracking branches:
`refs/remotes/<remote>/<branch>`
This mechanism allows the user to:
- Choose a remote as push target for their branch different than the one
that would automatically get chosen for them by `pr create`;
- Avoid a `git push` operation (and its associated git hooks) if a push
was not necessary in the first place.
This API was just fixed in github.com, but it will take a while for the
change to propagate to GitHub Enterprise installs, so guard ourselves
from false positives when querying forks.
Before: the default push target for the current branch in `pr create`
was the first repository found among git remotes that has write access.
Now: the default push target is the fork the base repo, if said fork
exists and has write access, falling back to old behavior otherwise.
This change in the default is to facilitate contributions to projects
that have a hard requirement that all pull requests (even those opened
by people with write access to that project) come from forks.
- Only check for existing PRs if `--web` wasn't given
- Fix detecting PRs from forks
- Improve order of error handling: local validation errors are handled
earlier than the ones that have to consult the API
This stubs stderr separately from stdout in command tests (before those
streams were combined) and improves test assertions around output.
Additionally, no longer use the `cmd.Print*()` family of Cobra functions
because their name sounds like the text will go to stdout, but they
write to stderr instead. Use the more explicit `cmd.ErrOrStderr()` as
output destination instead.
