Add support for custom parameter types

[ananis25]

I want to extend the CLI app with custom data types not supported by Typer/Click. Right now, the Typer app raises an error on encountering an unsupported type.

The solution you would like

We could add a method to the Typer object to register custom data types along with a function to convert terminal input string to an object of the desired type.

A possible API could be:

app = typer.Typer()
app.register_custom_type(_type, _deserializer)

Here is the monkey-patched solution I use right now without providing a registration API. If this seems like a useful feature, I'll try refine this and submit a PR.

import datetime
from typing import Any, Callable, Optional

import typer
import click


_get_click_type = typer.main.get_click_type
REGISTERED_TYPES = {
    datetime.date: lambda string: datetime.datetime.strptime(string, '%Y-%m-%d').date()
}

class TyperCustomParam(click.ParamType):
    name = 'CustomParameter'
    def __init__(self, typ: type, deserializer: Optional[Callable] = None): 
        self.name = typ.__name__
        self._deserializer = typ if deserializer is None else deserializer

    def convert(self, value, param, ctx): 
        try:
            return self._deserializer(value)
        except Exception as E: 
            self.fail(
                f"couldn't serialize {value} to an instance of type {self.name}, error: {E}"
            )

def supersede_get_click_type(
    *, annotation: Any, parameter_info: typer.main.ParameterInfo
) -> click.ParamType:
    if annotation in REGISTERED_TYPES:
        return TyperCustomParam(annotation, REGISTERED_TYPES[annotation])
    else:
        return _get_click_type(annotation=annotation, parameter_info=parameter_info)

typer.main.get_click_type = supersede_get_click_type

Alternatives

Alternatives would be to either modify the original routine or create a wrapper function to accept a Click supported datatype and convert later. Letting users create custom types directly is preferred since that lets us create CLIs without touching existing code.

Thank you for looking into this. Similar to FastAPI, typer is a really nice library.

[jackric]

I need this feature (actually for datetime.date like your example) - so please make PR!

[ananis25]

Hey @jackric, the method used above requires a global container, which makes me uneasy. You could probably use this snippet in your code and import the typer module from it.

[jackric]

@ananis25 I think a nice approach would be to register custom types with a new function - either on typer or a typer.Typer object:

import typer
typer.register_type(datetime.date, lambda string: datetime.datetime.strptime(string, '%Y-%m-%d').date())

import typer
app = typer.Typer()
app.register_type(datetime.date, lambda string: datetime.datetime.strptime(string, '%Y-%m-%d').date())

typer.main.get_click_type would then check an internal store of registered types before its default behaviour.
For bonus points, allow multiple deserializing functions to be tried in order. They could be specified as a list to register_type, or multiple calls to register_type

[ananis25]

If I understand correctly, the above snippet does exactly this, isn't it? I'd leave it to the maintainers to see if they really want to add this.

[sylann]

I agree with the solution proposed by @jackric.
Also, to add on top of that:

There are lots of "wrapper types" that can be called on a string input and give an augmented object that is still a string. Many more are just variants of already handled types.

So maybe if Typer did just try to call the annotated type by default, it would be a quick and easy way to solve many problems with not so much work.
It could even be enabled (or disabled) with an option in the Typer instanciation. In case some users somehow expect to get an error on unhandled types.

--

I think it should at least be possible to skip an unhandled type with a mere warning. To let users properly annotate their functions without having to use some crazy black magic.

[tmr232]

One thing that is missing when registering custom types (or at least not explicitly clear) is the usage when I just want to add a converter to an existing type. For example - parsing a hexadecimal number. I want to extend the parsing, but the result type is still a number.

The options I see for this are either:

@app.command()
def command(number: int = Typer.Argument(converter=lambda x: int(x, 0)))

Or

class HexInt(int): pass

# Fill in registering custom type

@app.command()
def command(number: HexInt)

I am not sure what I think about either.

[jackric]

My 2 cents: declaring the HexInt class/type feels wrong, because later the type system says number is HexInt, when we really want to think of it as an int.

[paulo-raca]

I wrote a PR that implements this, please let me know what you think: #443

[dsully]

+1 would love to see @paulo-raca 's PR merged. Thanks!

[tiangolo]

Thanks all! Support for this was added in #583, the new docs are here: https://typer.tiangolo.com/tutorial/parameter-types/custom-types/, it's available in Typer 0.8.0, just released. 🎉