Jarvis, More Over-Engineered Than Ever Before

I first introduced Jarvis in a blog post almost three months ago. Since then, he's undergone a few upgrades.

He was originally hacked together as a series of modules that "magically" worked together. Now, don't get me wrong, he did exactly what I wanted him to do. Building new plugins was super simple: drop a python file in his plugins directory and he'd scan through and install them at startup. So long as they had a few specific functions, everything would "just work".

Like I say, he did exactly what I wanted him to... originally. But I have big plans for him and it was starting to get unwieldy to make him do what I wanted him to do the way I wanted him to do it.

Enter the magic of Python and metaprogramming.

Now, many of my friends already know how much I love Python. They also may have heard me cackling wildly as I implemented the new Jarvis re-design. Here's why:

In Python, all functions are first-class objects.

Not impressed yet? Fair enough, there are plenty of languages which can make the same claim. That's OK, we'll get there.

One of the cool things about treating functions as objects is you can do object-ey things with them, for example:

1
2
3
4
def i_am_a_function(arg0, arg1):
    do_something()

i_am_a_function.is_an_object = 'Absolutely true'

And given that everything is duck-typed in Python, things like attributes on functions can be whatever we want them to be.

Another cool thing we can do with functions, since they're first class, is to pass them to each other. In addition to passing them around as normal parameters, eg.

1
2
3
4
5
6
7
8
def bar(fn):
    fn()
    print 'bar'

def foo():
    print 'foo',

bar(foo)  # prints 'foo bar'

Python also allows us to use special decorator syntax. If we were to re-write the previous example with decorators, we'd have:

1
2
3
4
5
6
7
8
9
def bar(fn):
    fn()
    print 'bar'

@bar
def foo():
    print 'foo',

foo()  # also prints 'foo bar'

There are a few big advantages to this, as I see it. The fact that the caller doesn't need to know which functions to chain and in what order is a big one, but even more important is that we are beginning to develop a framework for doing some very fun things.

Now, let's put those two tidbits together to come up with something cool. Here's a snippet from Jarvis, of the class from which all Plugins inherit.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
class Plugin(object):
    # snip

    @staticmethod
    def on_message(msg):
        def on_message_decorator(func):
            func.regex = re.compile(msg)
            return func
        return on_message_decorator

    # snip

See what's going on here? We've defined a function, which takes some parameter and returns a different function with an attribute set to the regex compiled form of that parameter. In other words, on_message is a decorator which can be applied to a function which accepts a single parameter (the decorator, not the function); the decorator will then store some data on a new attribute in the wrapped function.

That's all well and good, but how do we use that?

First I need to explain a bit about metaclasses: now, in addition to functions being objects, classes are also objects. Just as other objects are instances of a particular class, classes themselves are instances of a metaclass.

What does this mean? Well, it means that by subclassing from type (the base / default metaclass), we can change how classes are instantiated.

In Jarvis, we have

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
class PluginMetaclass(type):
    def __new__(mcs, name, bases, namespace, **_kwargs):
        result = type.__new__(mcs, name, bases, dict(namespace))
        result.response_fns = [fn for fn in sorted(namespace.values())
                               if hasattr(fn, 'regex')]
        return result

class Plugin(object):
    __metaclass__ = PluginMetaclass

    # snip

This metaclass really doesn't do much, only one line besides calling its parent. But that one line is the basis of all the magic.

In our metaclass, we define the response_fns member of the resulting class as a list of all functions of that class with the 'regex' attribute set. Recall this is the attribute we set in our decorator.

Since the __new__ function is called as the class is instantiated, the resulting class object will always have this attribute set.

Finally, we define a method in which our base class can make use of this new attribute:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
class Plugin(object):
    # snip

    def respond(self, ch, user, msg):
        for fn in self.response_fns:
            regex_match = fn.regex.match(msg)
            if not regex_match:
                continue

            fn(self, ch, user, regex_match.groups())

    # snip

What have we done here? Basically, we've defined a function that, when called, will search through all methods of that class which have the 'regex' attribute set, attempt to apply the regex, and call the method only if a match is found.

This allows us to -- very easily -- define plugins which listen and respond to specific phrases. How easily? Well...

1
2
3
4
class CashPool(Plugin):
    @Plugin.on_message(r'(.*) (\w+) sent \$([\d\.]+) to ([, \w]+)\.?')
    def send_money(self, ch, user, groups):
        # snip

The above function listens for any instances of users sending money to each other, recognizes the event, and splits the event into groups (eg. from_user, to_user, value, etc)... all before even calling the function. Writing new plugins is now incredibly simple, all of the template-ey message-parsing work can be done in a single line.

And if that's not cool, I don't know what is.

Jarvis lives on my Github and is very proud to be celebrating his first birthday. As always, I'm happy to see contributions in any form!