The fundamental problem of shells is they are required to be two things.

1. A high-frequency REPL, which requires terseness, short command names, little to no syntax, implicit rather than explicit, so as to minimize the duration of REPL cycles.
2. A programming language, which requires readable and maintainable syntax, static types, modules, visibility, declarations, explicit configuration rather than implicit conventions.

And you can’t do both. You can’t be explicit and implicit, you can’t be terse and readable, you can’t be flexible and robust.

Shells optimize the former case, so that you can write cat beef.txt | grep "lasagna" | sort -n | uniq instead of:

with open(Path("beef.txt")) as stream:
    lines = filter(
        stream.readlines(),
        lambda line: re.match(line, "lasagna") is not None
    )
    print(set(reverse(sorted(lines))))

Which does not spark joy.

So the programming language aspect suffers: shell scripts are an unreadable nightmare of stringly-typed code resembling cyphertext.

Of course No True Scotsman would write a large and complex program as a shell script, but according to this lovely seven-line one-liner:

find / -type f -exec awk '/^#!.*sh/{print FILENAME}' {} + \
  | xargs file \
  | awk '!/ASCII text/{next} {print}' \
  | cut -d: -f1 \
  | xargs -I {} wc -l {} \
  | sort -n \
  | uniq

There are 5,635 shell scripts on my humble Ubuntu box. Of these, 79 are over one thousand lines of text, the largest being /usr/share/libtool/configure¹, a 16,394-line shell script the devil wrote². In total, there are 726,938 lines of stringly-typed shell script on my machine. This is more than I am comfortable with.

And the solution is obvious, but hard to implement because preserving backwards compatibility would require a great deal of elbow grease.

The solution is that we have one tool, but there are two things, and so there should be two tools. Shells should be terse, fast, interactive, and not too scriptable. Programs should export the terse command-line interface for use in the shell:

pandoc -t latex \
       -f markdown \
       --pdf-engine=xelatex \
       --table-of-contents \
       --toc-depth=2 \
       --resource-path=. \
       --standalone \
       input.md \
       -o output.pdf

And the typed interface, for use in a more scalable programming language:

Pandoc(
    input = (Path("input.md"), MD),
    output = (Path("output.pdf"), PDF),
    pdfEngine = XELATEX,
    resourcePath=Path("."),
    completenessMode=STANDALONE,
)

And the former can be derived from the latter, because it is a strict weakening of the typed interface.

The challenge is how to fit this into the POSIX universe where the sole entrypoint to a program is an array of strings. Historically, operating systems that have fanciful structured interfaces between programs have been left in the dust by Unix, because Unix maximizes flexibility by favoring the lowest-common-denominator interface, which is the string.

Footnotes