You cannot make a self-referential annotation like

class Graph:
    def reverse(self) -> Graph:
       # ...

This corresponds to a common coding situation.

This has been true forever. I suggest you try python-ideas.

This would be an important fix-up if we could find some way to implement it.

The basic problem is that the class object is defined after the class definitions have been made, so the target of the Graph reference isn't known when the method definitions are being compiled.

One approach to solving this problem is to use a deferred/resolved pattern (creating a valid reference to Graph that is visible during method compilation, but gets filled-in with the other methods afterwards).

I haven't thought this through completely but think it could be done if we let the compiler write to the class dictionary (this is normally off-limits and protected by a dict_proxy).  If we let the system punch a hole in the proxy, it is possible to resolve deferred class definitions.


class Prepared(type):
    'Preload the class with a reference to itself'

    @classmethod
    def __prepare__(mcl, name, bases):
        return {name: type(name, bases, {})}

    def __new__(mcl, name, bases, mapping):
        tmpcls = super().__new__(mcl, name, bases, mapping)
        deferred_class = mapping[name]
        deferred_class.__dict__.update(tmpcls.__dict__)  # XXX need writeable dict_proxy
        return deferred_class

class Graph(metaclass=Prepared):
    def reverse(self) -> Graph:
        ...

Until and unless this gets fixed, perhaps we should document some sort of workaround.  One possibility is attached.

>The basic problem is that the class object is defined after the class definitions have been made,

I have sometimes thought the the class statement could start with binding the name to a blank object or new, blank class object. But then people might try to grab attributes (and fail!) outside of method defs.