Back to writing

#!shebangs (¶)

    Table of contents
  1. Introduction
    1. Windows file extensions
    2. Unix shebangs
  2. I don't like shebangs
    1. Developer over user (or else bye-bye checksums)
    2. Anti-Windows dogmatism
    3. PATHEXT
  3. Hedges
    1. The registry
    2. .mp3.exe
  4. Conclusion

Introduction (¶)

Windows file extensions (¶)

On Windows, the file extension of a filename is used to automatically determine how the file should be opened when you double-click on it or run it on the command line. .exe are the kings, because they are Windows's native executables. For other extensions, the shell checks the registry to find the exe associated with the extension, and runs that program with your file as the first argument, plus the rest of your arguments. For example:

If you remove the .exe extension from an executable file's name, the shell won't know what to do with it (but deeper APIs can still invoke it (one, two)). That is to say, Windows does not inspect the content of files to determine their type for the purpose of double-click launching. Every file is either an exe or has an extension which maps to an exe, or has no extension and can't be associated to a program. [1]

At no time does this prevent you from using any other executable to open a file. You can always pythonw.exe or winword.exe clock.pyw. Extensions are not restrictions, they are just shortcuts to launching your executable of choice.

[1] There is also .com, a legacy executable format which is natively executable, yet has a passthrough HKCR entry anyway, presumably because the shell is programmed to consult HKCR for anything that's not .exe. HKCR\comfile\shell\open\command\(Default)="%1" %*

Unix shebangs (¶)

On Unix operating systems, the extension is not used to help execute the file. Instead,

  1. The file must have the x permission, aka "executable bit", enabled by chmod +x filename.
  2. If the file is a native executable, it will run directly. Otherwise, the system will look for the shebang. The system reads the first few bytes of the file to figure this out.

The shebang is a line which must be at the very top of the file and must start with #!. Then, a path to an executable is given, which will actually interpret and run the remainder of the file. A Python file with a shebang might look like this:

#!/usr/bin/env python
import sys
print(sys.executable, sys.argv)

You can save this file using any extension you want, or no extension at all, and run it by calling chmod +x ./filename and then ./filename. As always, you're welcome to pass the file as an argument to any other executable if you need to use a different interpreter.

One of the principles here is that executables become abstractions, black boxes. The idea is that as a user, when you run some program on the command line, you shouldn't need to know or care whether program is a shell builtin, or an alias, or an executable on the PATH; and you shouldn't need to care what language it's written in. You could swap out a dns program written in bash with one written in Python, and your callers won't know the difference.


I don't like shebangs (¶)

And you won't find a shebang on any of the code I write.

For these arguments, suppose that the file in question is checked into version control (e.g. git), or is seeding on IPFS, or is otherwise being hash-validated, or is restored to factory condition on every software/OS update, such that "just change the shebang" is not an appropriate solution. The file is handcuffed to the interpreter specified by the shebang.

My dislike for shebangs is part principle and part practical. Principles like:

And practicals like:

I will continue to use Python as the example language for this argument because that's what I'm most familiar with, but you may extend these issues to any case where there are alternative interpreters — even if the alternatives haven't been invented yet!

Developer over user (or else bye-bye checksums) (¶)

Back in the day, Python code was written in Python 2, and it was fair to assume that the executable named python on the PATH was the Python 2 interpreter, so people wrote shebangs pointing to python expecting P2. Now, code is written in Python 3 and systems are shipping with python referring to P3, but some P2 tools are still around and in use. So now their shebang is wrong, and will be wrong for the rest of time, and the user will always need to explicitly call python2 for as long as they hang on to that file. Great, thanks.

On an extension-based system, the solution is easy. Although the file might be checked into version control, there's nothing stopping me from creating a symlink or hardlink on my PATH called program.py2 that points to the file, and an extension mapping from .py2 to python2.exe. This requires no changes to the file content itself, so it won't interfere with the file's hash.

To be clear, the file layout I would use for maintaining a Python2 and Python3 program simultaneously is:

├───cmd\ on the PATH
│   ├───a_program.py2 => D:\software\a_program\
│   └─── => D:\software\b_program\
    │   └───

So even if was written when .py implied Python2 instead of Python3, it doesn't matter because the names I actually execute are those in \cmd on my PATH. I can name my links any way I want to take advantage of the file extension mapping, while the real software is not affected whatsoever. See Master of my domain for more about my software management practices.

Sure, you can say that nobody should be using P2 programs any more, but that's moving the goalposts. Let's say someone wants to run some particular set of Python scripts through PyPy or any other implementation. What then? To give one non-Python example, here is an SO thread about the right shebang for bash. To my eyes it's a disaster:

On a shebang-based system, the choices are:

Anti-Windows dogmatism (¶)

I don't like shebangs, but it's not such a big deal — all I have to do is not use them and that's that. They pose no restrictions upon me as long as I call the interpreter myself.

What actually bothers me are the Unix elitists who in a single breath will mock Windows users as being stupid, helpless lusers, pushed around by the Big Bad File Extension Who Dictates How Files Are Executed... then write a shebang that handcuffs their file to the name of the anticipated interpreter executable. A .rose by any other extension would smell as sweet.

To demonstrate this isn't a strawman, here are some actual quotes gathered from around the 'net:

I know that files are identified by an identifier on their first line, which is why Linux files don't have file endings. But I liked file endings a lot. Isn't it nice to be able to tell what a file is/does without having to open it? In Windows I could open up a folder and immediately have a lot of information about the folder and the files within it, just from the file endings.

If you're asking if there's some kind of setting that will "show" file extensions on system files, no, that is not a thing. You seem really hung up on the idea that filename extensions are a superior method of associating type information with files. Most Linux applications just don't use this method, but have other ways of determining that information. Your question is based on the assumption that this is some kind of feature which has been disabled and can be re-enabled, which is fundamentally untrue.

I will not: I will not allow the ignorance that spews from Redmond to infect my software.

I will not: I will not cater to users with a broken OS.

I will not: I will not bother because it represents the minimal technical hurdle for Windows users.

[A developer renames his readme to readme.txt]

While it may make logical sense to some people, the ultimate implication when you see a .txt file is that the developer uses Windows. I'm not sure a lot of developers are comfortable with that.

No. File extensions are stupid and broken.

Why? What's wrong with encoding metadata about the content in the filename?

Because it then prevents me from changing the filename the way I want to. Because legacy garbage.

Interesting... most commenters here (so far) seem to be against the extension. Can someone explain why? "README" doesn't tell me if it's text, or markdown, or HTML, or anything. Having ".txt" lets me know the format without opening it first to see.

I hate them because they're an abstraction leak. The type of a file's content is different from the application that I want to use to manipulate it, which both are in turn different from the name I chose to give the file. That current file systems and the GUI representations on top of them are so badly designed as to confuse three unrelated concepts is a failing of software and as clear an indication as you could ask for of the triumph of Worse Is Better.

These two in particular point out a problem that's already solved on Windows:

You should not use an extension for executables, as they they are not interchangeable. Imagine that you have a shell script, then re-write in python, you now have to change every program that calls your script, you have leaked an implementation detail.

If you give a shell script a .sh extension, you'll have to type that .sh as part of the command name when starting it. That's the main reason why I don't like putting that extension in.

We need to talk about PATHEXT.


On Windows, there is an environment variable called PATHEXT. It contains a semicolon-separated list of file extensions that you don't have to type when running programs on the command line. For example, if .py is on PATHEXT, then you can just ask for dns on the commandline and will be a valid candidate. On the other hand, if PATHEXT is blank, then you must type the file extension for every executable, even .exe. At that point, the shell builtins will be the only extensionless commands you can run.

The order of extensions on PATHEXT indicates the order of preference for candidate files.

C:\anywhere>set PATH=C:\cmd;C:\Windows\System32
C:\anywhere>where dns
You ran!
C:\anywhere>set PATHEXT=.bat;.py;.exe
You ran dns.bat!
C:\anywhere>set PATHEXT=.exe;.bat;.py
You ran dns.exe!
C:\anywhere>where.exe dns
C:\anywhere>where.exe dns.bat
You ran!

Notice I am now calling where.exe, since .exe is no longer on PATHEXT, thus where no longer matches C:\Windows\System32\where.exe.

C:\anywhere>set PATHEXT=
C:\anywhere>where.exe dns
INFO: Could not find files for the given pattern(s).
You ran!

Regardless of the value of PATHEXT, you are always allowed to specify the full name of the file, like Even if .py is not on PATHEXT, itself is still on the PATH. PATHEXT only affects the evaluation of the bare name, dns.

And just like that, we've achieved the black box principle of executables. You can rewrite a .bat program in .py, or you can supercede a System32 exe with a script of your own. The only thing you can't do is supercede the shell builtins — dir always does dir even if you have a candidate in the cwd or on the PATH — you'll have to call those files by their extensioned names.

Between my public cmd repository and my non-public cmd folder, I've got more than 180 .py and .bat files making up my command line toolkit. Of course I'm not specifying the extensions when I run them, guys.

Hedges (¶)

It's time for me to bat for the other team and explain why Windows's solution is also bad.

The registry (¶)

The Windows registry is not a beautiful place, and the mapping of extensions to executables is not as easy as one would hope. With Windows 10, Microsoft has introduced some new APIs for dealing with extension associations, mostly because with the old APIs, inconsiderate software developers would non-optionally dive into the registry and set their program as the default for a bunch of filetypes during install, leaving the average user without a clear understanding of how to revert. However, I haven't learned these new systems because Windows refers to them as "default apps" and I have a vendetta against the word "app" so I avoid that whole interface at all costs. When I need to edit extensions, I always just open the registry and do it myself.

The most straightforward technique that I'm familiar with is to create a key HKEY_CLASSES_ROOT\.ext. You can give it a (Default) value containing the name of another HKCR key, and this will act as a redirect for the remainder of the process. Then, beneath either the .ext or the redirect destination, you create shell\open\command\(Default)="C:\myprogram.exe" "%L" %* for the most simple invocation.

Beyond that, it's all a little confusing with keys like OpenWithList, OpenWithProgids, and values like Content Type and PerceivedType (do they do anything for files without an explicit shell\open\command?), and if you want to make two extensions that run the same program but have different icons (.txt, .srt, .css → text editor) you wind up duplicating the whole shell command for each of them because you can't do multiple-inheritance style mappings. Unless I'm missing something, which is also likely.

Also, there are even more keys in HKCU\Software\Microsoft\Windows\CurrentVersion\Explorer\FileExts\.ext\UserChoice. When dealing with extensions that I didn't make myself, NirSoft's filetypesman helps me locate their hodgepodge of keys.

The point is that while I prefer extension-executable systems over shebang systems, that does not mean I'm praising the registry as it exists today.

.mp3.exe (¶)

A common criticism of Windows is that non-savvy users can be tricked into downloading some file called "Hotel California.mp3.exe" and running it because they think it's an mp3 file. This is made 1,000 times worse by the unfathomable, unforgiveable, bone-headed decision by Microsoft to hide file extensions of "known file types" by default, hiding the .exe but leaving the .mp3. It boggles the mind and stupefies the senses that this default is still in place after all these years, but it is.

On Unix, the downloaded file would not be executable until you chmod +x, which is something you can't trick most people into doing. On Windows, the .exe requires no activation and is ready to detonate as soon as you double-click on it.

That's why UAC was added to Windows, but users will Accept any dialog box that stands between them and their mp3 without reading it, so that hasn't changed much.

Conclusion (¶)

Let me re-quote two of the comments I shared above:

Because it then prevents me from changing the filename the way I want to.

The type of a file's content is different from the application that I want to use to manipulate it, which both are in turn different from the name I chose to give the file

These comments make a good point, that it would be nice to have the ability to associate files to programs without having to name them a certain way. Computing is still a relatively young field, and I wouldn't mind seeing new ways of solving this problem. I mentioned that I'd like to see some multiple-inheritance, which would simplify icon management, default programs, and human-friendly type names. But, by my insistence on maintaining file hashes, you can infer that I'd like a solution out-of-band from the file content itself. Current implementations of alternate data streams are not ready to steal this stage because they aren't transmitted when sharing the file over the internet, for better and worse.

For now, I am satisfied with file extensions (as long as I wrangle the registry myself), and I want the Unix elitists to see how I use them in my favor; that they impose no friction on my workflow at all. inb4 Stockholm syndrome.

View this document's history

Contact me: