larryhastings/appeal

Command-line parsing library for Python 3.

Copyright 2021-2023 by Larry Hastings

Quickstart

import appeal
import sys

app = appeal.Appeal()

@app.command()
def hello(name):
    print(f"Hello, {name}!")

app.main()

Here's a simple fgrep utility:

import appeal
import sys

app = appeal.Appeal()

@app.command()
def fgrep(pattern, *files, ignore_case=False):
    if not files:
        files = ['-']
    print_file = len(files) > 1
    if ignore_case:
        pattern = pattern.lower()
    for file in files:
        if file == "-":
            f = sys.stdin
        else:
            f = open(file, "rt")
        for line in f:
            if ignore_case:
                match = pattern in line.lower()
            else:
                match = pattern in line
            if match:
                if print_file:
                    print(file + ": ", end="")
                print(line.rstrip())
        if file != "-":
            f.close()


if __name__ == "__main__":
    app.main()

Overview

Appeal is a command-line argument processing library for Python, like argparse, optparse, getopt, docopt, Typer, and click. But Appeal takes a refreshing new approach.

Other libraries have complicated, cumbersome interfaces that force you to repeat yourself over and over. Appeal leverages Python's rich function call interface, making it effortless to define your command-line interface. You write Python functions, and Appeal translates them into command-line options and arguments.

Appeal provides amazing power and flexibility--but it's also intuitive, because it mirrors Python itself. If you understand how to write Python functions, you're already halfway to understanding Appeal!

Appeal has only one dependency, my big library.

Appeal is currently only supported for POSIX platforms (UNIX, Linux, BSD, OS X, etc). It might work on Windows but this has not yet been tested.

A New And Appealing Approach

Appeal isn't like other command-line parsing libraries. In fact, you really shouldn't think of Appeal as a "command-line parsing library" per se. And, although you work with Appeal by passing in functions for Appeal to call, you shouldn't think of these functions as "callbacks".

Appeal lets you design APIs callable from the command-line. It's just like any other Python library API--except that the caller calls you from the command-line instead of from Python. Appeal is the mechanism converting between these two domains: it translates your API into command-line semantics, then translates the user's command-line back into calls to your API.

This raises another good point: the API you build using Appeal also often makes for a very nice automation API, allowing your program to also be used as a library by other programs with minimal effort.

Basics

Taxonomy

Let's start by establishing the terminology we'll use for command-lines, based on command-line idioms established by POSIX and by popular programs. Here's a sample command-line, illustrating all the various types of things you might ever see:

% ./script.py --debug add --flag ro -v -xz myfile.txt
  ^           ^       ^   ^      ^  ^  ^   ^
  |           |       |   |      |  |  |   |
  |           |       |   |      |  |  |   argument
  |           |       |   |      |  |  |
  |           |       |   |      |  |  multiple short options
  |           |       |   |      |  |
  |           |       |   |      |  short option
  |           |       |   |      |
  |           |       |   |      oparg
  |           |       |   |
  |           |       |   long option
  |           |       |
  |           |       command
  |           |
  |           global long option
  |
  program name

Command-lines are a sequence of strings separated by whitespace. The meaning of each string can depend both on the position of the string and the characters in the string itself.

An argument is any whitespace-delimited string on the command-line that doesn't start with a - (minus sign). Unless it's an oparg--which we'll talk about in a minute--the meaning of an argument is defined by its position. For example, if you ran:

fgrep WM_CREATE window.c

WM_CREATE and window.c would be arguments; the first argument, WM_CREATE, would be the string you wanted to search for, and window.c would be the name of the file you wanted to search.

A command is a special kind of argument some programs use to specify what function you want the program to perform. A good example of a program that uses commands is git; when you run git add or git commit, add and commit are both commands. The command is always the first argument to a program that uses them.

If a string on the command-line starts with a - (minus sign), that's an option. There are two styles of option: short options and long options.

Short options start with a single dash, -. This is followed by one or more individual characters, which are the short option strings. In the above example, we specify two sets of short options: the first is -v, the second is -xz. You can combine options togther, and it's the same as specifying them separately. We could have said -vxz, or -v -x -z; these both do the same thing. When we talk about short options, we say the word "dash" followed by the letter. For example, -v would be pronounced "dash v".

Long options start with two dashes, --. Everything after the two dashes is the name of the option. In the above example, we can see one long option, --flag. Again, when we talk about long options, we say the dashes out loud, followed by the words from the option. For example, --flag would be pronounced "dash dash flag".

Both types of options can optionally take one (or more) arguments of their own. An argument to an option is called an oparg. In the above example, the long option --flag takes the oparg ro.

Finally, there are global options and command options. Global options apply to the entire program, are always available, and are specified before the command. Command options are command-specific, and appear after the command. Global options can be long options or short options; command options can be long options or short options, too.

Remapping Python To The Command-Line

Now let's consider a Python function call:

#      positional parameters |  keyword-only parameters
#            |        |      |      |
#            v        v      |      v
def fgrep(pattern, filename, *, ignore_case=False):
    ...

We can draw some similarities between Python function calls and command-lines.

For example, they both support arguments where position is significant. A command-line argument is similar to a Python function positional parameter, in that they're both identified by position.

Python function calls and command-lines also both support arguments identified by name. A command-line option is similar to a Python keyword-only parameter.

Note: In Python, all parameters after *args or * are keyword-only parameters. You can only pass in arguments to those parameters by name. If you called fgrep yourself in Python, you could only pass in a value for ignore_case like this:

result = fgrep(p, f, ignore_case=True)

This leads us to the fundamental concept behind Appeal. With Appeal, you write a Python function, and tell Appeal that it represents a command. Appeal examines the function, translating its parameters into command-line features. Positional parameters become command-line arguments, and keyword-only parameters become options.

(Technically, Appeal translates both positional parameters and positional-or-keyword parameters into arguments. For the sake of clarity and consiseness, I'll always refer to these collectively as positional parameters.)

Hello, World!

Let's see Appeal in action, with our first example. In all our examples we're going to assume your program is called script.py. Let's say script.py looked like this:

import appeal
app = appeal.Appeal()

@app.command()
def hello(name):
    print(f"Hello, {name}!")

app.main()

If you now ran python3 script.py help hello, you'd see usage information for your hello command. It'd start like this:

usage: script.py hello name

Already, a lot has happened! Let's go over it piece by piece:

• We created an Appeal object called app. This object will handle processing the command-line and calling the appropriate command function.
• We decorated the function hello() with @app.command(), a method call on our Appeal object. This tells Appeal that hello() should be a command, using the name of the function as the command string, and translating the function's parameters into the command-line parameters. So our command-line command is called hello. We call a function decorated with @app.command() a command function.
• Our hello() command function takes one positional parameters, name. Therefore, our hello command on the command-line takes one positional argument, which we identify as name in the usage string.
• Appeal also automatically created simple help for our program, displaying usage information. Usage shows you what command-line options and arguments the command will accept.

So! If you ran this command at the command-line:

% python3 script.py hello world

Appeal would call your hello() function like this:

hello('world')

and you'd be rewarded with:

Hello, world!

The return value from your command function is the return code for your program. If you return None or 0, that's considered success; returning a non-zero integer indicates failure. (And if your function exits without a return statement, Python behaves as if your function ended with return None.)

Default Values And *args

Let's change up our example, and add an optional parameter:

import appeal
app = appeal.Appeal()

@app.command()
def fgrep(pattern, filename=None):
    print(f"fgrep {pattern} {filename}")

app.main()

Now our command is called fgrep, and it takes two parameters. The second one, filename, is optional, with a default value of None.

You can of course specify both parameters yourself. Running this:

% python3 script.py fgrep WM_CREATE window.c

results in Appeal calling your fgrep() function like this:

fgrep('WM_CREATE', 'window.c')

But you can also omit the filename parameter. If you run this command at the command-line:

% python3 script.py fgrep WM_CREATE

Appeal would call fgrep() like this:

fgrep('WM_CREATE', None)

Actually that's not 100% accurate. When Appeal builds the arguments to call your fgrep() function, it only passes in the arguments you passed in on the command-line. So actually Appeal calls your fgrep() function like this:

fgrep('WM_CREATE')

And it's Python that sets the filename parameter to None.

What else can Appeal command functions do? Well, they can have a *args parameter. Naturally, a command function that takes *args (internally called a var_positional parameter) can accept as many positional arguments as the user wants to supply. Here's a demonstration:

import appeal
app = appeal.Appeal()

@app.command()
def fgrep(pattern, *filenames):
    print(f"fgrep {pattern} {filenames}")

app.main()

Now the user could pass in no filenames, one filename, fifty filenames--as many as they want! They'd all be collected in a tuple and passed in to fgrep() in the filenames parameter.

Options, Opargs, And Keyword-Only Parameters

Now let's examine what Appeal does with keyword-only parameters. Let's add three keyword-only parameters to our example:

import appeal
app = appeal.Appeal()

@app.command()
def fgrep(pattern, *filenames, color="", number=0, ignore_case=False):
    print(f"fgrep {pattern} {filenames} {color!r} {number} {ignore_case}")

app.main()

Now the fgrep command-line usage looks like this:

usage: script.py fgrep [-c|--color] [-n|--number int] [-i|--ignore-case] pattern [filenames]...

Again, a lot just happened.

First, I'll remind you, keyword-only parameters are presented as options on the command-line. Appeal automatically took each keyword-only parameter, added '--' to the front of the parameter name, and turned that into an option. (Also, if the parameter name has any underscores, Appeal turns those into dashes.)

Second, Appeal also automatically uses the first letter of a keyword-only argument as a short option. So the color keyword-only parameter becomes both the --color and -c options. When running your program, the user can use -c or --color interchangably. The same goes for -i and --ignore_case, and for -n and --number.

(What if you have two keyword-only parameters that start with the same letter? The first one gets the short option. If we added a keyword-only parameter named credit to the end of fgrep()'s parameter list, Appeal would map color to --color and -c, but only map credit to --credit.)

Third, options are always optional. (As a pedantic wag might put it--"the clue's right there in the name.") Therefore, in Appeal, keyword-only parameters to command functions must always have a default value. (Python programmers usually have default values for their keyword-only parameters anyway, so this requirement isn't a big deal.)

Fourth, notice that --color takes an argument, or oparg. Appeal noticed that the color parameter had a default value of ""--its default value is a str. So Appeal infers that you want the user to supply an oparg to --color. If the user specifies --color on the command-line, it must be followed by an oparg, and Appeal will take the string off the command-line and pass it straight into the color parameter.

Fifth, --number also takes an oparg, but it has a default of 0. Appeal infers from that that --number should be an int. Appeal automatically converts the string from the command-line into a Python object for you, using the type of the default value. (Appeal did that for --color too--except --color takes a str, so no conversion is necessary.) When the user provides an oparg to --number on the command-line, it must be followed by an oparg; Appeal will take that oparg, pass it in to int, then take the return value from int and pass it in to the number parameter.

Finally, ignore_case has a default value of False. Boolean values for options are a special case: they don't take an oparg. All they do is negate the default value. So if the user specifies -i once on the command-line, Appeal would pass True in to the ignore_case parameter.

(By the way, a default value of None is a second special case. If a positional or keyword-only parameter has a default value of None, Appeal behaves as if the type of the default is str. It consumes an argument or oparg from the command-line and passes it in unchanged to that parameter.)

Let's put it all together! If you ran this command at the command-line:

% python3 script.py fgrep -i --number 3 --color blue WM_CREATE window.c

Appeal would call fgrep() like this:

fgrep('WM_CREATE', 'window.c', color='blue', number=3, ignore_case=True)

And if you ran this command at the command-line:

% python3 script.py fgrep --color green boogaloo

Appeal would call fgrep() like this:

fgrep('boogaloo', color='green')

The Global Command, Subcommands, And The Default Command

Many programs that support "commands" also have "global options". Global options are options specified on the command-line before the command. For example, in the example command-line at the top of this document, script.py takes a --debug option specified before the command--which makes it a "global option".

Appeal supports global options, too. It's simple: write your command function like normal, but instead of decorating it with Appeal.command(), decorate it with Appeal.global_command(). Appeal will process all those options before the command, and call your global command function.

Appeal.global_command() also gets used for programs that don't use "commands". Although the "command" command-line paradigm is popular these days, most programs don't bother with them. For example, ls, grep, and... hey! python itself! None of these programs support commands, but they all support command-line arguments and options.

Naturally, Appeal supports this behavior. Simply decorate one function with Appeal.global_command() and don't add any command functions.

On the flip side of this coin, Appeal also supports subcommands. This is a common feature of command-line parsing libraries, though it's rarely-used in practice. The idea is, your command can itself be followed by another command.

To add a subcommand to your Appeal instance, just decorate your command function with two chained command calls, specifying the name of the existing command in the first call, like so:

@app.command()
def db(...):
    ...

@app.command("db").command()
def deploy(...):
    ...

This adds a deploy subcommand under the db command. So now the whole command-line looks something like this:

script.py [global arguments and options] db [db arguments and options] deploy [deploy arguments and options]

Finally, what should Appeal do if your program takes commands, but the user doesn't supply one? That's what the default command is for. The default command is a command function Appeal will run for you if your Appeal instance has commands, and the user doesn't supply one. For example, if script.py has ten different commands, but the user just runs

script.py

without any arguments, Appeal would run the default command.

If you don't specify a default command, Appeal has a built-in default default command. The default default command raises a usage error, which means it prints basic help information and exits.

To specify your own default command, just decorate a command function with the Appeal.default_command() decorator. For example, if you wanted your program to run the status command when the user didn't specify a command, you could do this:

@app.default_command()
def default():
    return status()

Notice that the default command doesn't take any arguments or options. It simply can't accept any, by definition.

(If the user specified options without a command, they'd be considered "global options" and would be processed by the global command. And if the user specified an argument, that would automatically be the name of the command to run.)

And yes, subcommands can have a default command too:

@app.command('db').default_command()
def db_default():
    return db_status()

Annotations And Introspection

Python 3 supports annotations for function parameters, meant to conceptually represent types. Appeal supports annotations too; they explicitly tell Appeal what type of object a parameter requires. For example:

import appeal
app = appeal.Appeal()

@app.command()
def fgrep(pattern, *filenames, id:float=None):
    print(f"fgrep {pattern} {filenames} {id}")

app.main()

Here id has a default value of None, but it also has an explicit annotation of float. If the user uses --id on the command-line, it must be followed by an oparg, which Appeal will convert to a Python object by calling float. (And, as you can see, the annotation and the type of the default don't necessarily have to agree... although it's usually a good idea.)

Although annotations are meant to represent types, Appeal actually accepts any callable--it can be a type, or a user-defined class, or just a regular function. Appeal calls these annotations converters.

Here's how Appeal decides on the converter for a parameter, from highest-priority to lowest-priority:

• If the signature for that parameter has an annotation, Appeal uses the annotation as the converter.
• If the signature for that parameter doesn't have an annotation, but does have a default value, Appeal will use type(default) as the converter in most cases. The exceptions:
  • If type(default) is NoneType, Appeal will use str instead.
  • If type(default) is bool, and the parameter is a keyword-only parameter, Appeal will use a special internal class that implements the special-case "negate the default" behavior for options with boolean default values.
• If the signature for that parameter lacks both an annotation and a default value, Appeal uses str as the converter.

Converters are surprisingly flexible. For example, Appeal will introspect the converter for a keyword-only parameter and map all its positional arguments into opargs. That's how Appeal supports options that take multiple opargs: you simply annotate the keyword-only parameter with a converter that takes multiple arguments. Appeal will also pay attention to the annotations for the converter's own arguments, and use those to convert the strings from the command-line into Python objects.

Let's tie it all together with another example:

import appeal
app = appeal.Appeal()

def int_and_float(integer: int, real: float):
    return [integer*3, real*5]

@app.command()
def fgrep(pattern, *filenames, position:int_and_float=(0, 0.0)):
    print(f"fgrep {pattern} {filenames} {position}")

app.main()

Here, Appeal would introspect fgrep(), then also introspect int_and_float(). The resulting usage string would now look like this:

usage: script.py fgrep [-p|--position integer real] pattern [filenames]...

--position takes two opargs. Appeal would call int on the first one and float on the second one. It would then call int_and_float() with those values, and the return value of int_and_float() would be passed in to the position parameter on fgrep().

So now if you ran:

% python3 script.py fgrep -p 2 13 funkyfresh

Appeal would call:

fgrep('funkyfresh', position=[6, 65.0])

Finally, let's change the example to demonstrate something else: although converters can be any callable, user-defined classes work fine too. And Appeal can correctly infer the type based on the default value for any type. So consider this example:

import appeal
app = appeal.Appeal()

class IntAndFloat:
    def __init__(self, integer: int, real: float):
        self.integer = integer * 3
        self.real = real * 5

    def __repr__(self):
        return f"<IntAndFloat {self.integer} {self.real}>"

@app.command()
def fgrep(pattern, *filenames, position=IntAndFloat(0, 0.0)):
    print(f"fgrep {pattern} {filenames} {position}")

app.main()

This example behaves essentially the same as the previous example in this section, except the formatting of position is slightly different. But the command-line usage is exactly the same! Appeal inferred the converter for position based on the type of its default value, then introspected that type to determine how many opargs it should consume from the command-line and how to convert them.

An important note about annotations

If you use static type analysis in your project, your static type analyzer may not enjoy analyzing Python code using Appeal. Static type analyzers are designed to understand "type hints", a means of specifying static type information introduced in Python 3.5 with the typing module. But Appeal doesn't use type hints, and there are some ways Appeal uses annotations that static type analyzers may not like.

Fortunately, there are ways to get static type analyzers to work alongside Appeal.

First, you can decorate your Appeal command functions and converters with @typing.no_type_check(). This should only be necessary if you use functions as annotations; if you only ever use types and classes, this shouldn't be necessary.

Second, if you're using Python 3.9 or newer, you can use typing.Annotated with your annotations. typing.Annotated allows you to specify an ordered list of values, and static type hints only ever use the first value. Appeal also handles typing.Annotated, but Appeal only ever uses the last value. This makes it easy--you can have both types of annotations, side by side, and both static type checkers and Appeal are perfectly happy.

Converter Flexibility

You can use almost any function as an annotation... within reason. Appeal will introspect your annotation, determine its input parameters, and call it to convert the command-line argument into the argument it passes in to your command function.

For example, what if you wanted an option that accepted a string which gets broken up based on a delimiter substring? This is a common idiom for configure scripts on UNIX-like platforms; for example, Python's own configure script supports this option:

--with-dbmliborder=db1:db2:...

Happily that's easy to do in Appeal. Just write a converter function that accepts a string, breaks it into substrings however you like, and returns the list.

Although... you don't need to bother! Appeal also provides a converter that does it for you, called appeal.split(). You pass in as many delimiter strings as you want, and appeal.split() will split the command-line across all of them. (If you don't specify any delimiters, appeal.split() will split at every whitespace character.)

Specifying An Option More Than Once

One thing you might have noticed by now: the interfaces you've seen only allow Appeal to handle command-lines where an option can be specified either zero times or one time. What if you want the user to be able to specify an option three times? Or ten? That's what the MultiOption class is for. MultiOption objects are converters that allow options to be specified multiple times.

MultiOption isn't useful by itself; it's only an abstract base class. To make use of it you'll need to use a subclass--or create your own.

This time, let's start with some examples. Appeal provides three useful subclasses of MultiOption: counter, accumulator, and mapping.

First, let's look at counter. counter simply counts the number of times an option is specified on the command-line. This is a somewhat common idiom for "verbose" options; a program that supports -v to mean verbose may allow you to specify -v more than once to make it more verbose. Here's how you'd do that with Appeal:

import appeal
app = appeal.Appeal()

@app.command()
def fgrep(*, verbose:appeal.counter()=0):
    print(f"fgrep verbose={verbose!r}")

app.main()

If the user ran

% python3 script.py fgrep

Appeal would call

fgrep()

allowing Python to pass in the default value of 0 to verbose. And if the user ran

% python3 script.py fgrep -v --verbose -v

Appeal would call

fgrep(verbose=3)

accumulator handles options that take a single oparg. It remembers them all and returns them in a single array. Like so:

import appeal
app = appeal.Appeal()

@app.command()
def fgrep(*, pattern:appeal.accumulator=[]):
    print(f"fgrep pattern={pattern!r}")

app.main()

If the user ran

% python3 script.py fgrep --pattern three -p four --pattern fiv5

Appeal would call

fgrep(pattern=['three', 'four', 'fiv5'])

What if you don't want strings, but another type? Using crazy science magic from the future, accumulator is actually parameterized. You can say:

import appeal
app = appeal.Appeal()

@app.command()
def fgrep(*, pattern:appeal.accumulator[int]=[]):
    print(f"fgrep pattern={pattern!r}")

app.main()

and now the opargs to --pattern will all be converted using int.

You can even specify multiple types as arguments to the parameterized version of accumulator, separated by commas. The option will then require multiple opargs and convert them to the types specified.

mapping is like accumulator except it returns a dict instead of a list. An option annotated with mapping() consumes two positional arguments from the command-line; the first one is the key, the second one is the value. (You can also parameterize mapping the same way you parameterize accumulator, though you can only specify exactly two types.)

Of course, you can also subclass MultiOption to make your own converter classes with custom behavior. MultiOption subclasses can override these three methods:

class MultiOption:

    def init(self, default):
        ...

    def option(self, ...):
        ...

    def render(self):
        ...

Well, actually, subclasses are required to override option() and render(). But init() is optional.

If you then specify a subclass of MultiOption as an annotation on a keyword-only parameter of an Appeal command function, several things happen:

• If that option is specified one or more times on the command-line, Appeal will instantiate exactly one of these objects and call its init() method.
• Every time the user specifies that option on the command-line, Appeal will call the option() method on the object.
• After finishing processing the command-line, Appeal will call the render() method on the object, and pass the value it returns as the argument to that keyword-only parameter.

The most powerful part of this interface: you can redefine option() to suit your needs--it supports the same sort of polymorphism as annotations do. Appeal will introspect your option() method to determine how many opargs to consume from the command-line, and how to convert them.

Let's demonstrate all this with another example. If you want your option to take two opargs, with one being an int and the other being a float, you would define option() in your subclass as:

class MyMultiOption(appeal.MultiOption):

    def option(self, a:int, b:float):
        ....

Every time the user specified your option, it would take two opargs, and they would be converted into an int and a float before calling your option() method. It's up to you to decide how to store them, and how to render them into a single value returned by your render() method.

MultiOption is a subclass of a general Option class. Option behaves identically to MultiOption, except it only permits specifying the option once on the command-line, which means it will only your option() method once. You usually don't need to bother with making subclasses of Option--it's usually better to just use a class directly, like our class IntAndFloat example. The only feature you get by subclassing Option is, you get the default value for the parameter passed in to your constructor.

(The downside of subclassing Option and MultiOption is that it makes exporting your Appeal API as an automation API a little less convenient for the user, because your users will have to construct these objects and feed values into them by calling the option method.)

Data Validation

What if you want to restrict the data the user provides on the command-line? That's simple, just use a converter! Appeal provides a couple sample converters for data validation, but it's easy to write your own.

The classic example is a parameter where you can only use one of a list of values. For that, you can use Appeal's validate() converter. For example, this command restricts the direction parameter to one of six canonical directions:

import appeal
app = appeal.Appeal()

@app.command()
def go(direction:appeal.validate('up', 'down', 'left', 'right', 'forward', 'back')):
    print(f"go direction={direction!r}")

app.main()

You can pass in an explicit type using a type= named argument to validate(); if you omit it, it uses the type of the first argument.

Appeal also has a built-in range validator called validate_range(). It takes start and stop arguments the same way Python's range() function does. Note that validate_range() differs from Python's range() in one subtle way: values equal to stop are allowed.

If you prefer, you can "clamp" the value the user passed in to the range, by supplying the argument clamp=True to validate_range(). In that case, if the value the user specifies is outside the range, validate_range() will return the closest value of either start or stop.

(That's why validate_range() allows the value to be equal to stop. clamp would be annoying to use if stop itself was an illegal value--particularly if the types were floats.)

Appeal validation functions are straightforward to write. So, if these are insufficient to your needs, you can easily write your own. Take a look at the implementations of validate() and validate_range() inside Appeal to see one way to do it!

Multiple Options For The Same Parameter

Some programs have a set of options on their command-line that are mutually exclusive. Consider this simple-minded command-line:

go [--north|--south|--east|--west]

That is, you want the user to be able to "go" in one of those four directions, but only one. How would you do that in Appeal?

Easy. You simply define multiple options that write to the same parameter. All the behavior you've seen so far is using the default way of mapping keyword-only parameters to options. But actually Appeal allows you to make your own mappings. You can map a parameter as many ways as you want, even using different converters!

To manually define your own options, use the Appeal.option() method on your Appeal instance. It's a decorator you apply to your command function. The first argument is the name of the parameter you want the option to write to. After that is one or more options you want to map to this parameter. Appeal.option() also takes default and annotation keyword-only parameters, allowing you to specify respectively the default value or annotation for this option.

Here's a simple example of how to implement the above go command with Appeal:

import appeal
app = appeal.Appeal()

@app.command()
@app.option("direction", "--north", annotation=lambda: "north")
@app.option("direction", "--south", annotation=lambda: "south")
@app.option("direction", "--east",  annotation=lambda: "east")
@app.option("direction", "--west",  annotation=lambda: "west")
def go(*, direction='north'):
    print(f"go direction={direction!r}")

app.main()

All these annotations return a string. But actually you can return any type you want--and you can even map multiple annotations that return different types to the same parameter. You can even annotate with a MultiOption to allow specifying that option multiple times!

Note that, whenever you use the option() decorator to map your own options onto a parameter, Appeal won't add its default options for that parameter. It'll only have the options you explicitly set. Which means, for example, that in the sample code above, there aren't any short options for the options we created. -n won't work, only --north.

One final thing. Your command function can accept **kwargs too. The only things that will go into it are options you create with Appeal.option(), which map to parameters that don't otherwise exist.

Recursive Converters

You already know that you can pass in a converter that takes multiple arguments, and Appeal will consume multiple arguments from the command-line to fill it. And if the arguments to that converter have annotations, Appeal will call those functions to convert the command-line argument into the type your converter wants.

But what if you did... this?

import appeal
app = appeal.Appeal()

def int_float(i: int, f: float):
    return (i, f)

def my_converter(i_f: int_float, s: str):
    return [i_f, s]

@app.command()
def recurse(a:str, b:my_converter=[(0, 0), '']):
    print(f"recurse a={a!r} b={b!r}")

app.main()

The my_converter() parameter i_f is a positional parameter with an annotation that, itself, takes two positional parameters.

Would it surprise you to know--yes, it actually works!

Converters have been fully recursive this whole time. Actually this fact has been hiding in plain sight all along--all the examples using int_and_float() are recursive too, because int_and_float() has parameters annotated with int and float. Of course, those functions only take a single string argument; my_converter() takes two annotated positional parameters. But the principles remain the same.

Still, this is a more complex situation than we've seen before. recurse takes a positional parameter b that has a default value, but its converter takes multiple positional parameters, and one of those also has a converter that takes multiple positional parameters. How does Appeal map this to the command-line?

Appeal "flattens" the tree of converter functions into a linear series of arguments and options. In this case the usage string would look like this:

recurse a [i f s]

This tells you the recurse command takes either one or four command-line arguments. That optional group of three command-line arguments has a special name in Appeal: it's an "argument group". Technically, Appeal views this command-line as taking two "argument groups": the first group is required, and consumes one command-line argument; the second group is optional, and consumes three command-line arguments.

(We actually saw our first "argument group" in the second example in the Annotations And Introspection section above, but that time the argument group was an oparg.)

Now let's add an option and see what changes:

import appeal
app = appeal.Appeal()

def int_float(i: int, f: float):
    return (i, f)

def my_converter(i_f: int_float, s: str, *, verbose=False):
    return [i_f, s, verbose]

@app.command()
def recurse2(a:str, b:my_converter=[(0, 0), '', False]):
    print(f"recurse2 a={a!r} b={b!r}")

app.main()

Now the usage looks like this:

recurse2 a [i [-v|--verbose] f s]

Notice the way Appeal renders it in the usage string--the options aren't created until after the first argument in the optional argument group. This may seem strange but that's how it works. That's how it has to work.

Why? From a high conceptual level, Appeal doesn't know that you've "entered" the optional argument group until it sees the user supply the first argument for that group. So it doesn't create the options defined in that group until after the first argument.

This high conceptual level corresponds exactly to how Appeal calls your function. Consider, if the user runs this command:

recurse2 xyz

Appeal calls your function like so:

recurse2('xyz')

Since Appeal never called my_converter(), it can't map --verbose. It can only map --verbose once it knows it's going to call my_converter(), and that only becomes true the moment you supply that second command-line argument.

Once you do supply that second command-line argument, you have to supply two more, for a total of four.

recurse2 pdq 1 2 xyz

Appeal calls your function like so:

recurse2('pdq', my_converter(int('1'), float('2'), xyz))

recurse2 pdq 1 2 xyz

And in this example, you can supply the -v or --verbose anywhere after the second parameter. So if your command-line looks like this:

recurse2 pdq 1 2 xyz -v

Appeal calls recurse() like this:

recurse2('pdq', my_converter(int('1'), float('2'), xyz, verbose=True))

Take a look back at all the examples in this document, and consider that anywhere you specify a function or type, you can pass in nearly any callable you like.

For example, the parameterized version of mapping isn't limited just to simple types. If you used mapping[str, int_float] as the annotation for a keyword-only parameter, that option would consume three arguments on the command line: a str, an int, and a float, and the dictionary would map strings to 2-tuples of ints and floats.

Now you're starting to see how powerful Appeal's converters really are!

Now Witness The Power Of This Fully Armed And Operational Battle Station

But recursive converters are just the beginning!

Buckle your seatbelt, Dorothy--because Kansas is going bye-bye.

--Cypher, "The Matrix" (1999)

Options that map other options

What if you did... this?

import appeal
app = appeal.Appeal()

def my_converter(a: int, *, verbose=False):
    return [a, verbose]

@app.command()
def inception(*, option:my_converter=[0, False]):
    print(f"inception option={option!r}")

app.main()

Woah, that works too! We've created an option that itself takes an option. If you run fgrep --option, you can now also specify -v or --verbose--but only after you've specified --option.

In case you're wondering: Appeal.option() must decorate the function that takes the parameter you're mapping an option to. So if you want to define explicit options for the verbose parameter to my_converter in the above example, you'd decorate my_converter with Appeal.option() calls, not inception. (This also means, everywhere you use my_converter as a converter, it will behave the same, including taking the same options.)

Multiple options that aren't MultiOptions

We're just getting started! How about this:

import appeal
app = appeal.Appeal()

def my_converter(a: int, *, verbose=False):
    return [a, verbose]

@app.command()
def repetition(*args:my_converter):
    print(f"repetition args={args!r}")

app.main()

That works too, and I bet you're already guessing what it does. This version of weird accepts as many int arguments as the user wants to specify on the command-line, and each one can optionally take its own -v or --verbose flag.

Positional parameters that only consume options

I'll give you one more crazy example:

import appeal
app = appeal.Appeal()

class Logging:
    def __init__(self, *, verbose=False, log_level='info'):
        self.verbose = verbose
        self.log_level = log_level

    def __repr__(self):
        return f"<Logging verbose={self.verbose!r} log_level={self.log_level}>"

@app.command()
def mixin(log:Logging):
    print(f"mixin log={log!r}")

app.main()

Can you guess what usage for mixin looks like? (Probably!) It looks like this:

mixin [-v|--verbose] [-l|--log-level str]

Even though log is a positional parameter, it doesn't consume any positional arguments on the command-line. The Logging converter only adds options! This is what object-oriented programmers might call a "mix-in". With the Logging converter, you can add logging options to every one of your commands, without having to re-implement it each time. (Though in most cases it's probably better to add such options to a global command function.)

Internally this works exactly like you'd expect. Since the log parameter consumes no command-line arguments, Appeal will always call its converter. Specifying any of the options will set arguments for that call. And the resulting Logging object will be passed in as the argument to log.

What's really going on here is that, from Appeal's perspective, there's no difference between a "command function" and a "converter". A command function is just a converter that happens to be mapped to a command. So anything you can do with a command function, you can do with a converter too. A converter can define options, it can be decorated with app.option() (or app.parameter() which we haven't discussed yet), it can have accept any kind of parameter defined by Python, and any parameter can use (almost) any converter. And those converters can recursively use other converters.

Really, anything can be used with anything:

• Converters for positional parameters can take positional parameters, or keyword-only parameters, or *args, or **kwargs.
• Converters for keyword-only parameters can take positional parameters, or keyword-only parameters, or *args, or **kwargs.
• Converters for *args can take positional parameters, or keyword-only parameters, or *args, or **kwargs.
• Command functions can use any converter.
• The global command function can use any converter.

By now you can see the expressive power Appeal gives you. Of course, you'll rarely use only a fraction of that power. But it's reassuring to know that, whatever command-line API metaphor you want to express, it's not just possible in Appeal--it's easy.

Classes, Instances, And Preparers

Maybe you've noticed--all the examples so far have used standard Python functions as Appeal commands. What about method calls? Can you use those for commands? The answer is, yes of course! But it's slightly more complicated.

Appeal's whole purpose in life is to call functions by pulling data from the command-line. Whenever it sees a positional parameter on a function, it thinks "okay, I'm gonna have to supply an argument to that". So if you map an unbound method call to a command:

class MyApp:
    @app.command()
    def sum(self, *operands: int):
        return sum(*operands)

Appeal would see the self parameter and think "aha! I need to pass a string in there!" We need to prevent Appeal from seeing that parameter in the first place.

There are two major techniques to handle this. The first is straightforward, if a bit inflexible: create the instance of your class first, then call app.command()() on the bound instances. Like this:

app = appeal.Appeal()
class MyApp:
    def sum(self, *operands: int):
        return sum(*operands)

o = MyApp()
app.command()(o.sum)
app.main()

Since you pass in the already-bound method to Appeal, it doesn't even see the self parameter in the signature. (The signature of a bound method doesn't include the self parameter.)

This works fine... but maybe it looks a little weird. We're no longer decorating functions (or methods), instead we're calling the decorator function directly and passing in the bound method. It also restricts us to one instance of MyApp per Appeal instance, which might be restrictive.

The other technique uses a little magic to provide a convenient and familiar-looking interface. Appeal.app_class() gives you two decorators; you use one to decorate your class, and the other to decorate methods in the class. Appeal will instantiate your class for you, and use your __init__ method as your app's "global command" to handle global options!

import appeal

app = appeal.Appeal()
app_class, command_method = app.app_class()

@app_class()
class MyApp:
    def __init__(self, *, verbose=False):
        print(f"MyApp init verbose={verbose!r}")
        self.verbose = verbose

    def __repr__(self):
        return "<MyApp>"

    @command_method()
    def add(self, a, b, c):
        print(f"MyApp add self={self!r} a={a!r} b={b!r} c={c!r} self.verbose={self.verbose!r}")

app.main()

Behind the scenes, this uses a CommandMethodPreparer object to handle late-binding the method to the object. Since Appeal.app_class() is a little inflexible, you may want to use these objects directly. You can create one manually by calling Appeal.command_method(). Here's an example showing how to use one:

import appeal

app = appeal.Appeal()
command_method = app.command_method()

class MyApp:
    def __init__(self, id):
        self.id = id

    def __repr__(self):
        return f"<MyApp id={self.id!r}>"

    @command_method()
    def add(self, a, b, c):
        print(f"MyApp add self={self!r} a={a!r} b={b!r} c={c!r}")

my_app = MyApp("dingus")

p = app.processor()
p.preparer(command_method.bind(my_app))
p.main()

This is the first time you're seeing the Processor object. All the runtime information for processing a command-line lives in the Processor object; in fact, Appeal.main and Appeal.process are both thin wrappers over their equivalent methods on the Processor object. Moving all the runtime information into the Processor object lets you process multiple command-lines with the same Appeal object, even simultaneously!

The CommandMethodPreparer object is at the core of how Appeal handles late-binding of methods to objects. First, you decorate the method calls of your class with this object. You then call the bind method on that object to pass in the instance of that class you want to bind those methods to--though app_class() takes care of that for you. bind() returns a callable you pass in to Processor.preparer, which binds the method to that instance before Appeal calls it.

Under the covers, CommandMethodPreparer wraps the method with a functools.partial object, passing in a placeholder object for the self parameter. Then command_method.bind() replaces the placeholder for the real instance. For maximum compatibility, it actually uses getattr() to bind the instance to the method.

Writing Help

Appeal automatically generates usage for your command functions. But it's up to you to write the documentation explaining what those commands and arguments and options actually do.

There's very complete notes on how to write documentation in Appeal, see appeal/notes/writing.documentation.txt in the Appeal source distribution. In a nutshell, you write docstring in a particular way, and Appeal can mechanically parse them and combine them together. So you document each converter separately, and Appeal smooshes all these bits of documentation together to produce the help for your command function.

(One note: the main help for your program should be the docstring for your Appeal instance's global command.)

Appeal's latest superpower: reading config files

Appeal allows for friction-free command-line APIs. You write your command function, point Appeal at it, and whoosh! now you've got a command-line interface. But there are other interfaces users may want to use to configure your program. Now Appeal can work with those too.

For example, your program may read configuration from environment variables. Some programs launch an editor; for example git will open an editor when committing a revision. Traditionally on UNIX-based platforms this is configurable using two environment variables, VISUAL and EDITOR in that order of preference.

Appeal doesn't need to add explicit support for environment variables, as Python already has an easy-to-use interface. For example, here's how to support the environment variables configuring your editor:

However, many programs also support a configuration file, also called an "rc file" on UNIX. By convention settings in such a config file usually take precedence over environment variables. For example, you can configure what editor git uses for commits with a value called core.editor stored in a config file.

As of 0.6, Appeal has support for reading data from configuration files. Note that Appeal doesn't read the data files itself; you already have a library for that. Instead, Appeal has a generic mechanism for reading data from either an iterable or a mapping-- either a list or a dict.

The first step is to read in the values from the configuration file, and produce a dict or dict-like object. You can use any library you like. For example, the tomli library works well for TOML files. JSON and YAML parsers also work nicely. And this facility works especially well with my Perky file format. Though that's just a coincidence, as they were designed separately, years apart. Honest!

(You can also use configparser to read your INI config file, but this doesn't mesh well with Appeal's model. Better support for reading INI files is a possible future direction for Appeal.)

Once you've got a dictionary containing your configuration information, you can get Appeal to read from it using a single method call:

    Appeal.read_mapping(self, callable, mapping)

Simply pass in the callable you want called, and the mapping--the dict--you read from your config file. Appeal will read the names of the callable's parameters, pull values out of the mapping using those names, and pass those values in to a call to the callable.

Of course, any mapping will work. But this method works particularly well with classes decorated with dataclasses.dataclass. In just a few lines, you can define a class to contain your configuration information, read it out of a file, and populate the class with values of all the correct types!

In a lot of ways, this works very similarly to Appeal when it's processing a command-line. For example:

• Appeal will use the annotations and default values to convert the values from the dictionary into the correct types.
• Parameters with default values are optional; parameters without default values are required.

But there are differences too:

• You can use positional-only, positional-or-keyword, or keyword-only arguments. However, var-positional (*args) and var-keyword (**kwargs) are unsupported.

Let's bring all of this together with an example. Let's say we're writing a hypothetical program that may launch an editor. Our sophisticated program has five ways to decide what program to run for the editor. In decreasing order of importance:

• Command-line options '-e' and '--editor' specify the editor to use for this instnace.
• The config file ~/.myprogramrc is a Perky file, and it can contain an editor value.
• If the user has set an VISUAL environment variable, use that.
• If the user has set an EDITOR environment variable, use that.
• The default value is /usr/bin/vi.

Here's sample Python code implementing those semantics:

    default_editor = os.environ.get("VISUAL",
        os.environ.get("EDITOR", "/usr/bin/vi"))
    @dataclasses.dataclass
    class ConfigFile:
        editor:str=default_editor

    d = perky.load(os.path.expanduser("~/.myprogramrc"))
    app = appeal.Appeal()
    config_file = app.read_mapping(ConfigFile, d)

    @app.global_command()
    def global_command(*, editor=config_file.editor):
        print(f"editor = {editor}")

    app.main()

Note: using os.path.expanduser and a hard-coded filename like this is no longer considered best practice. You should use platformdirs to define the paths to your config files.

Nesting

Appeal's config file reader supports reading values from nested dicts. This maps directly onto nested function calls in annotations. If an annotation takes two or more parameters, the name of the parameter with that annotation will be used as the name of the nested dict.

Since that probably wasn't clear--sorry!--an example would probably help. Consider this example dictionary:

    d = {
        'a': 33,
        'b': {
            'verbose': True,
            'color': 'blue',
        },
    }

Here the value of 'b' is a nested dict. If we want Appeal to read a dict with this shape, it will have to descend into that nested dict. Appeal does that by default when a parameter has an annotation, and the annotation takes two or more parameters. Here's sample Python showing how to read this dict using Appeal:

    def read_b(verbose=False, color='black'):
        return (verbose, color)

    def config_file(a: int, b: read_b):
        return (a, b)

Because read_b is an annotation taking multiple parameters, Appeal will assume the value of 'b' is a nested dict, and will get the values of 'verbose' and 'color' from that dict.

If you don't want this behavior, you can disable it by decorating the annotation function with the unnested method on the Appeal object. If we change the code to the following:

    @app.unnested()
    def read_b(verbose=False, color='black'):
        return (verbose, color)

    def config_file(a: int, b: read_b):
        return (a, b)

Appeal won't descend into a nested dict named 'b'. In this case the dictionary would have to be be completely flat, like this:

    d = { 'a': 33, 'verbose': True, 'color': 'blue' }

Iterables

Appeal can also read from iterables inside the dictionary. The parameter that accepts an iterable should be annotated with a subclass of MultiOption. Appeal will instantiate the MultiOption and use the MultiOption protocol to fill the object.

For example, if your config file dict looked like this:

    d = {
        'color': 'blue',
        'lines': [
            'line 1',
            'here is line 2',
            'and finally, line 3',
            ]
    }

Appeal could map it to this callable:

    @dataclasses.dataclass
    class ConfigFile:
        lines: appeal.accumulator
        color:str = ''

If the MultiOption option takes multiple parameters, then the list must contain dictionaries. For example, this Python code:

    class Resolutions(appeal.Multioption):
        def init(self, default=None):
            self.default = default
            self.values = None

        def option(self, width:int, height:int, depth:int):
            if self.values is None:
                self.values = []
            self.values.append((width, height, depth))

        def render(self):
            if self.values is None:
                return self.default
            return self.values

    @dataclasses.dataclass
    class ConfigFile:
        resolutions: Resolutions
        color:str = ''

Would be able to read this mapping:

    d = {
        'color': 'orange',
        'resolutions': [
            {'width': 1280, 'height': 1024, 'depth': 24},
            {'width': 1600, 'height': 1200, 'depth': 16},
        ],
    }

read_iterable

In addition to Appeal's read_mapping method, Appeal also supports a read_iterable method. The API is almost identical:

    Appeal.read_iterable(callable, iterable)

However this function is much simpler. The iterable should itself be an iterable of iterables. Appeal will call the callable you specify once for every nested iterable. The result will be appended to a list, and read_iterable will return that list.

As always, Appeal will handle converting values using the annotations you specify. Unlike read_mapping, here you may pass in a function accepting *args, in which case Appeal will process an arbitrary number of trailing arguments. Also unlike read_mapping, there's no support for testing--neither nested dictionaties nor (further) nested lists. (Like when processing the command-line, when reading values from the iterable, nested annotations get flattened.)

CSV files

Finally, Appeal has special support for reading CSV files. This may seem like an odd thing to support--nobody uses CSV files as config files. But CSV files were the proof-of-concept for Appeal's config file support, and it's proved useful in another project, so for now it's staying in. There's a special method for reading CSV files:

    Appeal.read_csv(self, callable, csv_reader, *, first_row_map=None)

You pass in your callable, and a fresh csv.reader object. Appeal will read the rows out of the CSV object, passing in the strings into the callable, and append the result to a list. The return value is that list.

If first_row_map is false, read_csv will ignore the first line of the CSV file (the "column names" line) and pass in the values from the CSV file by position. If first_row_map is true, read_csv will use the rows from the first line of the CSV file as keys in a dictionary, populate the values with each subsequent row, and will pass the arguments by name.

In other words, if first_row_map is false, Appeal calls

    callable(*row)

for every line after the first line in the CSV file. And if first_row_map is true, Appeal calls

    d = {key: value for key, value in zip(column_headers, row)}
    callable(**d)

for every line after the first line in the CSV file.

API Reference

Appeal(help=True, version=None, positional_argument_usage_format="{name}", default_options=default_options)

Creates a new Appeal instance.

If help is true, Appeal automatically adds help support to your program:

• Adds hard-coded -h and --help options that print basic help.
• If your Appeal instance has any commands, and you haven't defined a help command, automatically adds a help command.

If version is true, it should be a string denoting the version of your program. Appeal will automatically add version support to your program:

• Adds hard-coded -v and --version options that print the version string.
• If your Appeal instance has any commands, and you haven't defined a version command, automatically adds a version command which prints the version string.

positional_argument_usage_format is the format string used to format positional arguments for usage. The only valid interpolations inside this string are {name}, which evaluates to the name of the parameter, and {name.upper()}, which evaluates to the upper-cased name of the parameter. So if you want your usage string to show arguments or opargs as <name> or NAME, you can achieve that by setting positional_argument_usage_format to <{name}> or {name.upper()} respectively.

default_options is a callable, called when a keyword-only parameter for a command function or a converter doesn't have any options explicitly mapped to it. The purpose of default_options is to call Appeal.option() one or more times to create some default options for that keyword-only parameter.

The API for a default_options callable should be:

default_options(appeal, callable, parameter_name, annotation, default)

• appeal is the Appeal instance.
• callable is the command function or converter the parameter is defined on.
• parameter_name is the name of the keyword-only parameter that does not have any explicitly defined options.
• annotation is the annotation for this parameter. This may be explicitly set on the function, or it may be inferred from the default parameter.
• default is the default value for this parameter. Since Appeal requires that keyword-only parameters must always have default values, this may never be inspect.Parameter.empty.

The return value of default_options is ignored.

The default value of default_options is Appeal.default_options(), documented below.

Appeal.command(name=None)

Used as a decorator. Returns a callable that accepts a single parameter callable, which must be a callable.

Adds the callable as a command for the current Appeal instance. If name is None, the name of the command will be callable.__name__.

(Doesn't modify callable in any way.)

Appeal.global_command()

Used as a decorator. Returns a callable that accepts a single parameter callable, which must be a callable.

Sets the global command for this Appeal object. This is the command that processes global options before the first command function.

Can only be set on the topmost Appeal object. (You can't call app.command('foo').global_command().)

(Doesn't modify callable in any way.)

Appeal.default_command()

Used as a decorator. Returns a callable that accepts a single parameter callable, which must be a callable.

Sets the default command for this Appeal object. The default command is run when your Appeal instance has subcommands, but the user doesn't supply the name of a command on the command-line.

Your default command function must not take any parameters.

(Doesn't modify callable in any way.)

Appeal.option(parameter_name, *options, annotation=empty, default=empty)

Used as a decorator. Returns a callable that accepts a single parameter callable, which must be a callable.

Maps an option on the command-line to the parameter parameter_name on the decorated function. All subsequent positional parameters are options, like --verbose or -v. (Thus, they must be strings, either exactly two characters long, or four or more characters long.)

annotation is the converter that will be used if this option is invoked. If no explicit annotation is supplied, Appeal.option() will default to type(default).

default is the default value for this option. Since this parameter only comes into play if the user specifies this option, a default value here is nearly useless. But it does have two uses:

• If the type of the annotation is a subclass of Option, this default value will be passed in to Option.init().
• If no annotation is specified, the annotation defaults to type(default).

It's illegal to call Appeal.option() without specifying a value for either annotation or default.

Raises AppealConfigurationError if any option has already been mapped inside this Appeal instance with a different signature.

(Doesn't modify callable in any way.)

Appeal.parameter(self, parameter_name, *, usage=None)

Used as a decorator. Returns a callable that accepts a single parameter callable, which must be a callable.

Allows for configuration of a positional (or positional-or-keyword) parameter on a command function or converter. parameter_name is the name of the parameter; it must be a parameter of the decorated callable.

Currently the only supported configuration is usage, which specifies the string that will represent this parameter in usage information.

(Doesn't modify callable in any way.)

Appeal.main(args=None)

Processes a command-line and calls your command functions. Stops at the first failure result and passes it in to sys.exit(). Catches usage errors; if it catches one, displays usage information. The implementation calls Appeal.process().

Appeal.process(args=None)

Processes a command-line and calls your command functions. Stops at the first failure result and returns that result. Doesn't catch any errors. Useful mainly for automation, particularly for testing, and as the main driver underlying Appeal.main().

Appeal.default_options()

Appeal.default_long_option()

Appeal.default_short_option()

These functions create the default options for a keyword-only parameter. They're all valid callbacks for the default_options parameter for the Appeal() constructor. Appeal.default_options() is the default value for that parameter.

Appeal.default_long_option() creates the option --{modified_parameter_name} with the default annotation and default value. modified_parameter_name is parameter_name.lower().replace('_', '-').

Appeal.default_short_option() creates the option -{parameter_name[0]} with the default annotation and default value.

Appeal.default_options() creates both.

In all three cases, if the function isn't able to map at least one option, it raises an AppealConfigurationError.

Notes on the default option semantics:

• When Appeal.default_option() converts a keyword-only parameter into a long option and a short option, Appeal copies off the first character as the short option, and then runs a conversion function on the string. The conversion function lowercases the string and converts underscores into dashes. So for the the keyword-only parameter Define, Appeal.default_option() would (attempt to) create the two options -D and --define. For the keyword-only parameter block_type, it would attempt to create -b and --block-type.

• What if you have multiple keyword-only parameters that have the same first letter? Only the first mapping succeeds. So if you use def myoptions(*, block_type=None, bad_block=None) as an Appeal command, -b will map to block_type. If you want it to map to bad_block, just swap the two keyword-only parameters so bad_block is first, or explicitly define your options by decorating your function with Appeal.option(). (As of some recent version, Python guarantees it will maintain the order of keyword-only parameters when introspecting a function--and it was accidentally true in every version of Python before that explicit guarantee anyway.)

AppealConfigurationError

An exception. Raised when the Appeal API is used improperly.

AppealUsageError

An exception. Raised when Appeal processes an invalid command-line. Caught by Appeal.main(), which uses it to print usage information and return an error.

AppealCommandError

An exception. Raised when an Appeal command function returns a result indicating an error. (Equivalent to SystemExit.) Caught by Appeal.main(), which uses it to print usage information and return an error.

Reference

The library inspects the parameters of your function and uses those for the arguments, options, and opargs of your subcommand:

• Positional-only and positional-or-keyword parameters (parameters before *, or *args,) map to positional arguments. This:

  @app.command()
  def fgrep(pattern, file, file2=None):
      ...

  would take two required command-line arguments, "pattern" and "file", and an optional third command-line argument "file2".

• Keyword-only parameters map to options. They must have a default value. The name of the parameter is the name of the option, e.g. this subcommand accepts a --verbose argument:

  @app.command()
  def foo(*, verbose=False):
      ...

• If an argument to your function has an annotation, that value is called to convert the string from the command-line before passing in to your function. e.g.

  @app.command()
  def foo(level:int):
      ...

  would call int on the string from the command-line before passing it in to level.

• If a parameter to your function doesn't have an annotation, but does have a default value, it behaves as if you added an annotation of type(default_value). e.g.

  @app.command()
  def foo(level=0):
      ...

  would also call int on the string from the command-line before passing it in to level.

  • Keyword-only parameters with a bool annotation or a boolean default value are special: they don't take an argument. Instead, they toggle the default value.

  • Parameters with a default value of None and no annotation are also slightly special, in that they take a str argument (as taking a NoneType argument doesn't make sense).

  • Appeal automatically adds single-letter options for keyword-only parameters when possible. Since keyword-only parameters maintain their order in Python*++*, Appeal gives the single-letter shortcut to the first parameter that starts with that letter. e.g.

    @app.command()
    def foo(*, verbose=False, varigated=0):
        ...

    -v would map to --verbose, not --varigated.

Putting it all together: if you wanted to write an "fgrep" subcommand with a "usage" string like this:

fgrep [-v|--verbose] [--level <int>] pattern [ file1 [ file2 ... ] ]

you'd write it as follows:

@app.command()
def fgrep(pattern, *file, verbose=False, level=0):
    ...

++ This is now guaranteed behavior in current Python, and even in the Python 3 series before that, it was always true anyway.

Appeal And POSIX Utility Semantics

The POSIX standard defines command-line behavior for all POSIX utility commands, in 1003.1, Chapter 12, currently at revision POSIX.1-2017:

https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap12.html

Appeal isn't a perfect match for POSIX semantics; it disallows some things POSIX allows, and allows some things POSIX disallows.

• As per required POSIX semantics (1003.1-2017, Chapter 12), options can never be required. It therefore follows that in Appeal, keyword arguments to command functions must always have a default.
• The POSIX standard makes no mention of "long options", so it's not clear whether or not the standard permits them. (Presumably they will be permitted in a future standard.)
• POSIX requires that options that accept/require multiple opargs should accept them as a single string with either spaces or commas separating the opargs. Appeal supports this behavior with appeal.split. But it also permits options that consume multiple separate opargs from the command-line.
• POSIX requires that all options be specified before any positional arguments. Appeal doesn't enforce this, and will happily consume options and positional arguments in any order. In fact, "subcommands" require permitting options after positional arguments for anything beyond the simplest possible subcommand support.
• POSIX requires that, if an option (short option) has a single optional argument (oparg), its argument must be concatenated to the option. For example, if -f takes an optional argument, and you want to specify the argument guava, you must write this as -fguava, no other spelling is permissible. While Appeal supports this spelling, it also supports -f=guava and -f guava. More importantly, if you specify -f on your command-line (and not -f=<something> or -f<something>), Appeal will consume the next argument on the command-line as an oparg, which is what POSIX definitely does not want. I feel Appeal's consistency is more important than supporting this syntactic hack. Note that the oparg is still optional, so if -f is the last thing on your command-line, that will achieve this "option with default value" behavior.

Additional Subtle Features And Behaviors

• You can specify options and arguments in any order on a command-line, Appeal doesn't care. If you want Appeal to stop recognizing arguments starting with dashes as options, specify -- (two dashes with nothing else). All subsequent strings on the command-line will be used as arguments, even if they start with a -.
• Many built-in types are not introspectable. If you call inspect.signature(int) it throws a ValueError. Appeal has special-cased exactly five built-in types: bool, int, str, complex, and float.
• Accumulator actually allows parameterizing multiple types, separated by commas. Accumulator[int, float] will take two opargs each time the option is specified, and the first will be an int and the second will be a float. The list returned will contain tuples of ints and floats.
• You can't call main() on an Appeal object more than once. The Appeal() instance you use has internal state that changes when you execute its main() method.
• Information about a particular converter is localized to a particular Appeal() instance. If you decorate a converter with @app.option(), every place inside that Appeal() instance that you use that converter will also pick up the changes you made with @app.option().
• You shouldn't call usage() until you've added all the commands, options, and parameters information into your Appeal object. Why? Because, for example, usage() computes the default options for keyword-only parameters that haven't gotten any explicitly defined options. But if you then define one of those options, Appeal will throw an error at you.
• Almost any callable can be a converter--but not every function. There are two limitations. First, as already mentioned, in order for a function to be a legal converter, every keyword-only parameter must have a default value. The second requirement is more specific: in order to use a function as a converter for a *args* parameter, somewhere in the annotations tree under that function, some function must take a required positional parameter.

Finally, the UNIX make command has an interesting and subtle behavior. The --jobs and -j options to make specify how many jobs to run in parallel. If you run make without any parameters, it runs one job at a time. If you run make -j 5, it runs five jobs at a time. But! If you specify make -j, where -j is the last thing on the command-line it runs as many jobs at a time as it wants. In a way, the -j option has two default values.

Can you do this with Appeal? Naturally! Simply specify your keyword-only parameter with both an annotation and a default value, then design the annotation function to take one argument that also has a default value. Like so:

def jobs(jobs:int=math.inf):
    return jobs

@app.command()
def make(*targets, jobs:jobs=1):
    ...

Restrictions on Appeal command functions:

• You may not use inspect.Parameter.empty as a default value for any keyword-only parameter to a converter or command function.
• The converter for a var_positional (*args) parameter must require at least one positional argument.

Changelog

next version under development

0.6.3 2024/09/06

• Bugfix for usage. If conversaion fails for a command-line parameter, Appeal now prints a context-specific error, followed by usage for the command that failed. Fixes #18.
• Bugfix for read_mapping. Previously you couldn't have two parameters with the same name anywhere in the annotations tree for a mapping function, and now you can.

0.6.2 2023/10/12

• Presentation change: if you run a program without arguments, runs no-argument help instead of usage. This prints out both usage information and a list of commands, which seems more useful. That's how most modern programs do it (e.g. git, hg).
• Minor API change: renamed Appeal's custom exceptions, to remove the word Appeal. So, for example, AppealUsageError is now simply UsageError. I added aliases so the old names still work; I'll remove them eventually, but I promise to keep the old names around for at least one year.
• Fixed usage generation, added tests.
• Fixed a bug where using a "simple type" (e.g. bool, float) as an annotation for an option would cause Appeal help to raise an exception. Fixes #15.

0.6.1 2023/07/22

• Fixed 3.6 and 3.7 support--some equals-sign-in-f-strings uses crept in.
• Added GitHub Actions integration. Tests and coverage are run in the cloud after every checkin. Thanks to Dan Pope for gently walking me through this!
• Fixed metadata in the pyproject.toml file.
• Added badges for testing and supported Python versions. (Didn't add the coverage badge yet... it's too embarassing!)

0.6 2023/07/20

A huge upgrade!

• A new feature: Appeal can now read configuration files! Check out the new APIs Appeal.read_mapping, Appeal.read_iterable, and even Appeal.read_csv. This was a massive undertaking and involved a big overhaul of the compiler.

• The biggest change to existing behavior: Appeal now early-maps options. (See issue #3.) In short: when options are only defined in an optional group, they get provisionally mapped (made available) before the first argument in that group. Using that option enters the group just like specifying the first argument in that group. You'll see the difference in usage; an optional group that mapped an option used to look like [a [-v|--verbose] b c] but now looks like [[-v|--verbose] a b c].

• Appeal now handles multiple short options smashed together (e.g. -ace) identically to them being specified separately (e.g. -a -c -e). This caused an observable change in behavior regarding when child options get unmapped.

  • Appeal only permits using child options in a limited context: it must be after the parent option is executed, it must be after the parent option has consumed all its required or optional opargs, and it must be before any top-level positional argument or option mapped before the parent option was executed. But Appeal was lax about enforcing these rules when using multiple short options smashed together (e.g. -ace); it would handle all the options and then unmap child options as needed. The good news: Appeal now enforces these rules here too. (The old behavior seems to have been intentional on my part--what was I thinking?!)

• The usage message raised for an unknown option is now much better. If the option is defined anywhere in the program being run, it prints a different message telling you it can't be used here, but also tells you where it can be used. For example, if you use option -x, but that's a child option mapped by --parent, the message would say -x can't be used here, it must be used immediately after --parent.

• Renamed Appeal.argument to Appeal.parameter. This was one of those "what was I thinking?" moments. The function affects the parameter, not the argument. The old name still works but will be removed before 1.0.

• short_option_concatenated_oparg is now more strictly enforced: it's only permitted for short options that have exactly one optional oparg, as specified by POSIX.

0.5.9

• Improved the error message generated when you have a required parameter after a VAR_POSITIONAL parameter. (This command-line can never succeed, because the VAR_POSITIONAL consumes all remaining arguments on the command-line, which means the subsequent required parameter can never be satisfied.) Fixes #6.
• Changed README to use absolute instead of relative links for images, which means images should now render properly on the Appeal page at PyPI. Thanks for the PR, Hugo!
• Switched a bunch of internal classes to use Python "slots". Hopefully a tiny memory and speed optimization.
• Remove unused / unneeded stuff:
  • unused partial_replace* functions
  • unused Charm bytecode instructions jump_relative and load_o_option
  • unneeded CharmInterpreter register option (only used in one error message, obviated by instead using the program name directly)
• Now that Appeal depends on big anyway, switch to big's PushbackIterator..
• Add peephole optimizer step for jump-to-jump optimization. Honestly this was probably unnecessary, as I don't think Appeal's compiler can even generate code with jumps-to-jumps (yet).

0.5.8

• Fixed up the "name" of the program for options. We used to take the name of the command, add all the option strings, and join it together with commas, as in 'command, -o, --option'. Now it looks like 'command -o | --option'.
• Fixed presentation bug: if you didn't have enough positional arguments for your command function, but you ever invoked an option on the command-line, the usage text would contain the name of the last option invoked (aka the last Charm program run). Added a regression test for this.
• Cleaned up implementation a little: instead of using mystery lists on the CharmInterpreter stack and context_stack, I now use instances of bespoke CharmStackEntry and CharmContextStackEntry classes.

0.5.7

• Rewrite the technology behind accumulator[...] and mapping[...]. It previously used exec(), which was limiting; for example, you couldn't use your own types or converters. The new implementation should be much more robust; it now manually defines an explicit signature for the option() method of the subclass it creates.
• This fixed a regression, where you couldn't use a locally-defined class (e.g. IntFloat) as one of the types in the square brackets for accumulator and mapping. Added a test for this.

0.5.6

• Fix formatting for usage when you have a global command and subcommands.

0.5.5

• Add support for typing.Annotated, new in Python 3.9.
• Add dependency to my big library. This gives Appeal a much better implementation of multisplit, and I plan to switch to the big word wrapper and columnizer functions, which are a... "big" improvement over what's in Appeal right now.
• Rename SingleOption to just Option. (The name SingleOption is now deprecated, but I'll leave it as a redundant name for Option... for now.)

0.5.3

• Fix compatibility back to Python 3.6.

0.5.2

• Fix compatibility with Python 3.11. Python's inspect.Parameter object no longer allows a name that happens to be a keyword, which was a minor inconvenience (Appeal used to use lambda here sometimes).

0.5.1

• Fixed regression, issue #5. If you didn't supply enough required parameters, you'd get a TypeError instead of a proper usage error.

0.5

• Initial release!