Handling typing.Annotated

[SimonHeybrock]

Currently using typing.Annotated as parameters or for type hints in providers leads to mostly cryptic exceptions. One could consider supporting it, but it would require using __annotations__ (or inspect.get_annotations in newer Python) instead of typing.get_type_hints.

One could probably get this too work, but I fear it may be more brittle. Therefore, we should carefully consider whether support for Annotated is desirable. If not, we should attempt to raise clearer error messages.

[jl-wynen]

More details from separate issue:

from typing import NewType, Annotated
import sciline as sl

Int = NewType('Int', int)
Str = NewType('Str', str)
AStr = Annotated[Str, 'metadata']


def foo(x: Int) -> AStr:
    return Str(str(x))


def bar(s: Str, a: AStr) -> list[str]:
    return [s, a]


pl = sl.Pipeline([foo, bar], params={Int: Int(3)})
# Can compute results:
print(pl.compute(Str))  # -> 3
print(pl.compute(list[str]))  # -> ['3', '3']

# But the annotation is lost in the graph:
g = pl.get(Str)
print(list(g._graph))  # -> [__main__.Str, __main__.Int]
g = pl.get(list[str])
print(list(g._graph))  # -> [list[str], __main__.Str, __main__.Int]
print(list(g._graph[list[str]].arg_spec.keys()))  # -> [__main__.Str, __main__.Str]

# Raises
# sciline.handler.UnsatisfiedRequirement: ('No provider found for type', typing.Annotated[__main__.Str, 'metadata'])
pl.compute(AStr)

[jl-wynen]

One could probably get this too work, but I fear it may be more brittle. Therefore, we should carefully consider whether support for Annotated is desirable. If not, we should attempt to raise clearer error messages.

Disallowing this would mean that annotated types can never be used with Sciline. This may be too limiting.