README.md

sorcery

This package lets you use and write callables called 'spells' that know where they're being called from and can use that information to do otherwise impossible things.

• Installation
• Quick examples
  • assigned_names
  • unpack_keys and unpack_attrs
  • dict_of
  • print_args
  • call_with_name and delegate_to_attr
  • maybe
  • switch
  • select_from
• Rules for casting spells
  • The easy version
  • The advanced version
• How to write your own spells
  • Using other spells within spells
  • Other helpers
• Should I actually use this library?

Installation

pip install sorcery

Quick examples

See the docstrings for more detail.

from sorcery import (assigned_names, unpack_keys, unpack_attrs,
                     dict_of, print_args, call_with_name,
                     delegate_to_attr, maybe, select_from)

assigned_names

Instead of:

foo = func('foo')
bar = func('bar')

write:

foo, bar = [func(name) for name in assigned_names()]

Instead of:

class Thing(Enum):
    foo = 'foo'
    bar = 'bar'

write:

class Thing(Enum):
    foo, bar = assigned_names()

unpack_keys and unpack_attrs

Instead of:

foo = d['foo']
bar = d['bar']

write:

foo, bar = unpack_keys(d)

Similarly, instead of:

foo = x.foo
bar = x.bar

write:

foo, bar = unpack_attrs(x)

dict_of

Instead of:

dict(foo=foo, bar=bar, spam=thing())

write:

dict_of(foo, bar, spam=thing())

(see also: magic_kwargs)

print_args

For easy debugging, instead of:

print("foo =", foo)
print("bar() =", bar())

write:

print_args(foo, bar())

To write your own version of this (e.g. if you want to add colour), use args_with_source.

The packages q and icecream have similar functionality.

call_with_name and delegate_to_attr

Sometimes you want to create many similar methods which differ only in a string argument which is equal to the name of the method. Given this class:

class C:
    def generic(self, method_name, *args, **kwargs):
        ...

Inside the class definition, instead of:

    def foo(self, x, y):
        return self.generic('foo', x, y)

    def bar(self, z):
        return self.generic('bar', z)

write:

    foo, bar = call_with_name(generic)

For a specific common use case:

class Wrapper:
    def __init__(self, thing):
        self.thing = thing

    def foo(self, x, y):
        return self.thing.foo(x, y)

    def bar(self, z):
        return self.thing.bar(z)

you can instead write:

    foo, bar = delegate_to_attr('thing')

For a more concrete example, here is a class that wraps a list and has all the usual list methods while ensuring that any methods which usually create a new list actually create a new wrapper:

class MyListWrapper(object):
    def __init__(self, lst):
        self.list = lst

    def _make_new_wrapper(self, method_name, *args, **kwargs):
        method = getattr(self.list, method_name)
        new_list = method(*args, **kwargs)
        return type(self)(new_list)

    append, extend, clear, __repr__, __str__, __eq__, __hash__, \
        __contains__, __len__, remove, insert, pop, index, count, \
        sort, __iter__, reverse, __iadd__ = spells.delegate_to_attr('list')

    copy, __add__, __radd__, __mul__, __rmul__ = spells.call_with_name(_make_new_wrapper)

Of course, there are less magical DRY ways to accomplish this (e.g. looping over some strings and using setattr), but they will not tell your IDE/linter what methods MyListWrapper has or doesn't have.

maybe

While we wait for the ?. operator from PEP 505, here's an alternative. Instead of:

None if foo is None else foo.bar()

write:

maybe(foo).bar()

If you want a slightly less magical version, consider pymaybe.

switch

Instead of:

if val == 1:
    x = 1
elif val == 2 or val == bar():
    x = spam()
elif val == dangerous_function():
    x = spam() * 2
else:
    x = -1

write:

x = switch(val, lambda: {
    1: 1,
    {{ 2, bar() }}: spam(),
    dangerous_function(): spam() * 2
}, default=-1)

This really will behave like the if/elif chain above. The dictionary is just some nice syntax, but no dictionary is ever actually created. The keys are evaluated only as needed, in order, and only the matching value is evaluated.

select_from

Instead of:

cursor.execute('''
    SELECT foo, bar
    FROM my_table
    WHERE spam = ?
      AND thing = ?
    ''', [spam, thing])

for foo, bar in cursor:
    ...

write:

for foo, bar in select_from('my_table', where=[spam, thing]):
    ...

Rules for casting spells

To allow spells to find where they're being called from, you have to abide by certain laws. Otherwise they may backfire.

The easy version

Here are rules you can follow that are as easy to remember and apply as possible:

1. Use spells as simply and directly and possible. No indirection, renaming, or aliases.
2. Don't use the name of a spell for anything else.
3. Don't use the same spell more than once in a statement.
4. Never put multiple statements on the same line separated by semicolons.

The advanced version

If the above rules seem like a pain and you want to delve deeper into the magic, here is what you can actually do. You still have to avoid semicolons, but the other rules are a bit more complicated.

If you want to define a function which calls a spell for you while offering the same interface as a spell (typically __getattr__, __getattribute__, or some other boilerplate code for indirection), see the decorator no_spells.

If the spell is in its own variable, e.g. as a result of from sorcery import dict_of or d = spells.dict_of, then:

• You can use any name, as the actual value of the variable will be checked to find the call.
• You cannot use the same spell twice in the same statement, even if the statement spans several lines.

If the spell is used as an attribute of an object, e.g. sorcery.dict_of or as a method of a class, then:

• You must use the name of the original function.
• You cannot use the same attribute name (even for the same spell) twice in the same line, although you can use them twice in the same statement if they're on different physical lines.
• You cannot use the spell as an attribute of an object where it wasn't originally defined. For example, you cannot set foo.dict_of = dict_of and then call foo.dict_of(...). This is because the spell has to be in the class of the object to make the descriptor protocol work.
• For spells that are methods, you cannot use the unbound method of a class, e.g. Foo.bar(Foo(), ...).

How to write your own spells

Decorate a function with @spell. An instance of the class FrameInfo will be passed to the first argument of the function, while the other arguments will come from the call. For example:

from sorcery import spell, wrap_module

@spell
def my_spell(frame_info, foo):
    ...

will be called as just my_spell(foo).

The most important piece of information you are likely to use is frame_info.call. This is the ast.Call node where the spell is being called. Here is some helpful documentation for navigating the AST. Every node also has a parent attribute added to it.

frame_info.frame is the execution frame in which the spell is being called - see the inspect docs for what you can do with this.

At the bottom of the module where the spell is defined, add this line exactly:

wrap_module(__name__, globals())

This should also be added to any other module where you want to be able to import the spell from and use it.

Those are the essentials. See the source of various spells for some examples, it's not that complicated.

Using other spells within spells

Sometimes you want to reuse the magic of one spell in another spell. Simply calling the other spell won't do what you want - you want to tell the other spell to act as if it's being called from the place your own spell is called. For this, add insert .at(frame_info) between the spell you're using and its arguments.

Let's look at a concrete example. Here's the definition of the spell args_with_source:

@spell
def args_with_source(frame_info, *args):
    """
    Returns a list of pairs of:
        - the source code of the argument
        - the value of the argument
    for each argument.

    For example:

        args_with_source(foo(), 1+2)

    is the same as:

        [
            ("foo()", foo()),
            ("1+2", 3)
        ]
    """
    ...

The magic of args_with_source is that it looks at its arguments wherever it's called and extracts their source code. Here is a simplified implementation of the print_args spell which uses that magic:

@spell
def simple_print_args(frame_info, *args):
    for source, arg in args_with_source.at(frame_info)(*args):
        print(source, '=', arg)

Then when you call simple_print_args(foo(), 1+2), the Call node of that expression will be passed down to args_with_source.at(frame_info) so that the source is extracted from the correct arguments. Simply writing args_with_source(*args) would be wrong, as that would give the source "*args".

Other helpers

That's all you really need to get started writing a spell, but here are pointers to some other stuff that might help. See the docstrings for details.

The module sorcery.core has these helper functions:

• node_names(node: ast.AST) -> Tuple[str]
• node_name(node: ast.AST) -> str
• statement_containing_node(node: ast.AST) -> ast.stmt:
• resolve_var(frame, name: str)

FrameInfo has these methods:

• assigned_names(...)
• get_source(self, node: ast.AST) -> str

and the property file_info which returns an instance of sorcery.core.FileInfo (see the docstring of the class for what this has).

Should I actually use this library?

If you're still getting the hang of Python, no. This will lead to confusion about what is normal and expected in Python and will hamper your learning.

In a serious business or production context, I wouldn't recommend most of the spells unless you're quite careful. Their unusual nature may confuse other readers of the code, and tying the behaviour of your code to things like the names of variables may not be good for readability and refactoring. There are some exceptions though:

• call_with_name and delegate_to_attr
• assigned_names for making Enums.
• print_args when debugging

If you're writing code where performance and stability aren't critical, e.g. if it's for fun or you just want to get some code down as fast as possible and you can polish it later, then go for it.

The point of this library is not just to be used in actual code. It's a way to explore and think about API and language design, readability, and the limits of Python itself. It was fun to create and I hope others can have fun playing around with it. Come have a chat about what spells you think would be cool, what features you wish Python had, or what crazy projects you want to create.

If you're interested in this stuff, particularly creative uses of the Python AST, you may also be interested in:

• birdseye (another project of mine): a debugger which records the value of every expression
• MacroPy: syntactic macros in Python by transforming the AST at import time