Instead of raising a whole variety of exceptions, they are now wrapped
into ExtensionError derived classes.
* ExtensionAlreadyLoaded
* Raised when an extension is already loaded in Bot.load_extension
* ExtensionNotLoaded
* Raised when an extension is not loaded, e.g. Bot.unload_extension
* NoEntryPointError
* Raised when an extension does not have a `setup` function.
* ExtensionFailed
* Raised when an extension's `setup` function fails.
* ExtensionNotFound
* Raised when an extension's module import fails.
It turns out that events created in an eval command also cause
the issue described in #1506.
Ensure that events we remove are part of a module as well.
Also performs minor comment maintenance
("x", "first y", "then z") -> ("x", "y", "z")
Change the behaviour of handling iterable command_prefix types to not
silently ignore falsy prefixes and unify behaviour for all iterable
types. Add special handling of a possible TypeError in both get_prefix
and get_context for when the prefix is a different type from what is
expected.
This is powered by a dict-like class for the people who want to opt-in
to the performance downgrade for usability for majority English
speaking users.
Since it is slower than the regular dict due to the excessive calls to
str.lower along with the possibilities of gotchas for other languages
outside of English, this is kept as False for the default case.
When unloading cogs, currently we do not remove submodules from
sys.modules, meaning they will not be reloaded. Removing here
makes new imports reload from file. Of course, any already imported
modules will still hold a reference to the old module, since they
will not re-import it, and will not be forcably unloaded.
There is a counterpart for this in cogs, called __global_check_once.
This allows for predicates that would filter a command globally that
do not necessarily require rechecking in the case of e.g. the help
command such as blocking users or blocking channels.
Internally, instead of using module objects just use the `__module__`
attribute which is the same thing. From preliminary testing this seems
to work fine with both regular one-file-per-cog approaches and the
folder cog approach.
Fixes#126.
