In version 3.2, bpo-8814 introduced copying the __annotations__ property from the wrapped function to the wrapper by default.

That would be the desired behavior when your wrapper function has the same signature than the function it wraps, but in some cases (for example, with the contextlib.asynccontextmanager function) the return value is different, and then the __annotations__ property will have invalid information:

In [2]: from contextlib import asynccontextmanager                                                                                                            

In [3]: @asynccontextmanager 
   ...: async def mytest() -> int: 
   ...:     return 1 
   ...:                                                                                                                                                       

In [4]: mytest.__annotations__                                                                                                                                
Out[4]: {'return': int}


I propose changing the behavior of wraps, to only assign the __annotations__ by default if there's no __annotations__ already in the wrapper function, that would fit most default cases, but would allow to preserve the __annotations__ of the wrapper function when the types are explicitly specified, allowing now to change the contextlib.asynccontextmanager function with the proper types (returning now an AsyncContextManager) and keep the __annotation__ valid.

I'll try to get a POC and attach to the issue, but please comment with your ideas too.

I don't understand this use-case, but would it make sense to `ChainMap` the wrapper's __annotations__ on top of the wrapped __annotations__?

Hi Terry,

That would not work in this case, as I'd want to override all annotations with the wrapper function ones if there's any, instead of merging them.

The specific use case, is a type checker (part of TestSlide testing framework), to verify that if there's any type annotations, the parameters mocked and passed to it are the expected types.

For example, the contextmanager decorator returns an actual ContextManager, wrapping whatever the wrapped function returned, so if the wrapped function annotations prevail, then there's no way if verifying that the returned type is correct.

Thanks for the ChainMap pointer though, I'll use it for sure somewhere else.

As a note, mypy does not tpyecheck the wrapper functions, probably because it would not be possible with the current code (as the typing hints get lost):

https://mypy.readthedocs.io/en/latest/generics.html?highlight=wrapper#declaring-decorators

Elaborating on the last message, given the following code:
```
  1 #!/usr/bin/env python3
  2 
  3 from functools import wraps
  4 
  5 
  6 def return_string(wrapped):
  7     @wraps(wrapped)
  8     def wrapper(an_int: int) -> str:
  9         return str(wrapped(an_int))
 10 
 11     return wrapper
 12 
 13 
 14 @return_string
 15 def identity(an_int: int) -> int:
 16     return an_int
 17 
 18 def print_bool(a_bool: bool) -> None:
 19     print(a_bool)
 20 
 21 def identity_nonwrapped(an_int: int) -> int:
 22     return an_int
 23 
 24 
 25 print_bool(a_bool=identity(7))
 26 print_bool(a_bool=identity_nonwrapped(7))
```

mypy will complain only on the last line, being unable to check properly the line 25.

I'll investigate a bit more on why mypy skips that.

Only 3.8+ for bug fixes.