Vim is a shell command, and its fast startup supports that use-case: shell tasks, whether ad-hoc (interactive) or orchestrated (pipeline, script), are cheap and thus frequent.

Yet Vim’s startup story is relatively unpolished. We expect shell tools to consume standard input (“stdin”) and emit to standard output (“stdout”)—but Vim supports this awkwardly, at best. The endeavor is never mentioned in Vim tutorials, including the “Unix as IDE” hymnals. And it is puzzled out of Vim’s documentation only by careful inspection.

Vim is positioned as a script host (Vimscript, if_python, …) , but not as a participant. Shell users expect their tools to compose, like this:

# Does not work!
$ printf 'a\nb\nc\nb\n' | vim +'g/b/norm gUUixx' +2 +'norm yy2p' | tr x z

Why doesn’t that work? What can we do instead?

Let’s talk about -s-ex

The goal is to penetrate Vim with input, manipulate it non-interactively, and produce output consumable by other shell tools.

Sending text input to Vim requires the explicit - file.

$ echo foo | vim -

Working non-interactively is less obvious. Vim’s testsuite does something like this:

$ vim -es -u NONE -U NONE -i NONE --noplugin -c ... -c "qall!"

But what is -es? Not merely the combination of -e and -s, it is a special “silent mode” described at :help -s-ex:

Switches off most prompts.
...
The output of these commands is displayed (to stdout):
    :print
...
Initializations are skipped.

So -es does not draw the UI, and we can emit text to stdout using :print.

$ echo foo | vim - -es +'%p' +'qa!'
Vim: Reading from stdin...
foo

:%p prints the entire buffer and :qa! ensures that Vim quits. In Vim version 8, that “Vim: Reading from stdin…” message can be avoided with --not-a-term.

Note that -es and -se are not equivalent, the Vim parser quite literally expects -e to precede -s:

$ echo foo | vim - -se +'%p' +'qa!'
Garbage after option argument: "-se"

A similar order-sensitivity befalls the - file argument: vim - -es behaves differently than vim -es -! The former consumes stdin as text, while the latter activates stdin as Ex commands.

If you run into trouble, use -V1 to reveal why -e isn’t working:

$ printf 'foo\n' | vim -es
# No output. Non-zero error code.
$ echo $?
1

$ printf 'foo\n' | vim -es -V1
Entering Ex mode.  Type "visual" to go to Normal mode.
:foo
E492: Not an editor command: foo

So now we can light up our tinsel:

$ printf 'a\nb\nc\nb\n' | vim - -es --not-a-term +'g/b/norm gUUixx' +2 +'norm yy2p' '+%p' '+qa!' | tr x z
a
zzB
zzB
zzB
c
zzB

Yay! We did it. Wait, you’re going home already…?

Ugly sweater party

Apparently Vim thought this was an ugly sweater party. Vim’s sweater has - and Reading from stdin... and forty-four --help options. Let’s learn more about Vim before seating it next to Grandpa vi.

Input at startup can take these forms:

  • user input (“keyboard”)
  • Ex commands
  • text

By default even in non-interactive mode (-es) Vim treats input as “commands”, a tradition from Grandpa vi. Note that “commands” in vi parlance means general user input (starting from Normal-mode).

With -e (and -es) Vim treats input as Ex commands: those entered at the : prompt, or “statements” in Vimscript.

$ printf "put ='foo'\n%%s/o/X\n%%print\n" | vim -es

fXo

Finally the - file tells Vim to slurp input as plain text into a buffer. Then commands can be given with -c or --cmd.

There’s something I want to announce at the dinner table: if you specify the - file, then other file arguments are not allowed. :help vim-arguments characterizes these independent “editing ways”:

Exactly one out of the following five items may be used to choose how to start editing:

Should you try to invoke multiple “editing ways”, Vim will leave the table. For example, asking Vim to read both - and a.txt is asking too much:

$ echo foo | vim - a.txt
Too many edit arguments: "a.txt"

Santa POS is coming to town

Gathering from the POSIX vi specification we find these directives regarding non-interactive (“not a terminal”) cases:

Historically, vi exited immediately if the standard input was not a terminal. …

If standard input is not a terminal device, the results are undefined. The standard input consists of a series of commands and input text, …

If a read from the standard input returns an error, or if the editor detects an end-of-file condition from the standard input, it shall be equivalent to a SIGHUP

  • Input is always interpreted as user input, even if non-interactive.
  • When stdout is not a terminal, Vim may behave as it likes (“undefined”).
  • EOF (that is, closed input stream) means quit.

POSIX conspicuously omits the -e and -s startup options. Traditional vi lacks -e, but nvi implements both and calls out POSIX’s incongruence with “historical ex/vi practice”.

We’ve uncovered the origin of Vim’s eagerness to consume stdin as something alive instead of something inert: ’twas always done that way.

$ echo foo | vim
Vim: Warning: Input is not from a terminal
Vim: Error reading input, exiting...
Vim: Finished.

Vim must warn about stdin-as-commands because (1) it’s potentially destructive and (2) it’s almost always accidental (does anyone actually use this feature?).

Vim exits after stdin closes (EOF), as prescribed by POSIX. (Party trick: use vim -s <(echo ifoo) to convince it to keep running.)

POSIX does not mention -E, a variant of -e ignored by Vim’s own testsuite. -e invokes getexmodeline whereas -E invokes getexline. The distinction is not useful, and in Nvim they both invoke getexline.

Under the tree: Neovim

Nvim version 0.3.1 features some improvements to the workflow described above. The documentation and manpage were rewritten.

Nvim now treats non-terminal stdin as plain text by default (the explicit - file is not needed):

$ echo foo | nvim

That means Nvim never pauses to display a warning. Nvim also allows multiple “editing ways”:

$ echo foo | nvim file1.txt file2.txt

It also works with -Es (but not -es), and it exits automatically (no +'qa!' needed):

$ echo foo | nvim -Es +"%p"

If you ever want to execute stdin-as-commands, use -s -:

$ echo ifoo | nvim -s -

With these improvements one can use nvim -es as one might use python -c or perl -e. For example, I use it in my .bashrc to configure the shell depending on the Nvim version:

if nvim -es +'if has("nvim-0.3.2")|0cq|else|1cq|endif' ; then
  ...
fi

Eggnog

The ergonomics for delivering data to Vim at invocation-time are essentially unchanged from vi—owing, yes, to POSIX deference and our old friend backwards compatibility—but perhaps primarily to inertia.

It seems that few people actually care about the traditional behavior: the precise behavior of -es for example was broken in Nvim for years but no one complained. And Vim’s own codebase (including testsuite) does not use -E. If no one uses a feature, it might be ok to change it.

Merry Textmas! This holiday, when you’re with your loved ones, think of Vim.


This article was featured in the Vim advent calendar.