Previously they were printing the original command that was run, and
telling the user to rerun it. However, the command sometimes would not
match the exact command that was run (for example, when using the
`--local` flag), and in any case, it's simpler and more useful to print
the underlying error anyways.
https://github.com/rubygems/rubygems/commit/5bc0d51b58
In case we have a corrupted lockfile that claims to support a platform, but
it's missing platform specific gems for it, bundler has a check that
detects the situation and forces a re-resolve. The result of this check
is kept under the `@locked_specs_incomplete_for_platformn` instance
variable in `Definition`.
The installer, however, calls `Definition#nothing_changed?` before this
instance variable has been filled, so the result of it is actually
incorrect here since it will claim that nothing has changed, but
something has changed (locked specs are incomplete for the current
platform).
The consequence of this incorrect result is that the installer thinks it
can go on without re-resolving, resulting in the incomplete resolution
from the lockfile being used, and in a crash being triggered due to
that.
The solution is to make sure the `@locked_specs_incomplete_for_platform`
instance variable is filled before `nothing_changed?` gets called.
Moving it to `initialize` makes the most sense, not because it's the
best place for it (we can refactor this later), but because all of the
other "outdated definition" checks are already set there.
https://github.com/rubygems/rubygems/commit/708afdd789
The other way, in particular matching a substring in the gemspec
summary, is brittle and no longer used since Ruby 2.0.
This needed rewriting the specs that depended on that way.
https://github.com/rubygems/rubygems/commit/059dbfa971
The spec was just faking an installed gemspec without any installed gem
backing it up, resulting in `bundle install` claiming that the gem was
already installed when it was not.
https://github.com/rubygems/rubygems/commit/c35531d1c7
In 2021, Ruby 2.5 and older are EOL.
We can set the default required Ruby version to 2.6.0 to
encourage people to use newer Ruby.
If the command is executed with old Ruby, it falls back to 2.3.0.
It's still possible to create a gem for older Ruby just by changing
two lines of code (one in gemspec and another is in rubocop.yml).
We want to update the original repo, not a fresh one.
This went undetected because the `git commit` command inside the
`update_git` method ignores failures, and in this case it was failing
because all files are untracked in the new repo.
I will fix that later but for now fix the typo.
https://github.com/rubygems/rubygems/commit/c889f1d715
The glob information was not specified in the string representation for
a source, which led to non-deterministic behaviour when generating the
lockfile, since sources are sorted by this value.
https://github.com/rubygems/rubygems/commit/493b880abc
Previously, if an autoload failed (the file was loaded, but the
constant was not defined by the autoloaded file). Ruby will try
to autoload again if you delete the autoloaded file from
$LOADED_FEATURES. With this change, the autoload and the
constant itself are removed as soon as it fails.
To handle cases where multiple threads are autoloading, when
deleting an autoload, handle the case where another thread
already deleted it.
Fixes [Bug #15790]
The error had not be caught be specs because `bundle install` was
returning a zero exit code when plugin installation errors happened. So
I fixed that issue too.
https://github.com/rubygems/rubygems/commit/90cde87856
