DOKK / manpages / debian 10 / elvish / elvish-language.7.en
elvish-language(7) Miscellaneous Information Manual elvish-language(7)

This document describes the Elvish programming language. It tries to be both a specification and an advanced tutorial; if it turns out to be impossible to do these two things at the same time, this document will evolve to a formal specification, and more readable tutorials will be created.

Examples for one construct might use constructs that have not yet been introduced, so some familiarity with the language is assumed. If you are new to Elvish, start with the learning materials (/learn).

Note to the reader. Like Elvish itself, this document is a work in progress. Some materials are missing, and some are documented sparingly. If you have found something that should be improved -- even if there is already a "TODO" for it -- please feel free to ask on any of the chat channels advertised on the homepage (/). Some developer will explain to you, and then update the document. Question-driven documentation :)

Elvish source code must be UTF-8-encoded. In this document, character is a synonym of Unicode codepoint (https://en.wikipedia.org/wiki/Code_point) or its UTF-8 encoding.

Also like most shells, Elvish uses whitespaces -- instead of commas, periods or semicolons -- to separate constructs. In this document, an inline whitespace is any of:

A space (ASCII 0x20) or tab (ASCII 0x9, "\t");
A comment: starting with # and ending before the next newline or end of file;
Line continuation: a backslash followed by a newline.

A whitespace is either an inline whitespace or a newline ("\n").

Like most shells, Elvish has a syntax structure that can be divided into two levels: a statement level and an expression level. For instance, on the expression level, "echo" is a quoted string that evaluates to echo; but on the statement level, it is a command that outputs an empty line. This distinction is often clear from the context and sometime the verb used. A statements executes to produce side effects; an expression evaluates to some values. (The traditional terms for the two levels are "commands" and "words", but those terms are quite ambiguous.)

The most common data structure in shells is the string. String literals can be quoted or unquoted (barewords).

There are two types of quoted strings in Elvish, single-quoted strings and double-quoted strings.

In single-quoted strings, all characters represent themselves, except single quotes, which need to be doubled. For instance, '*\' evaluates to *\, and 'it''s' evaluates to it's.

In double-quoted strings, the backslash \ introduces a escape sequence. For instance, "\n" evaluates to a newline; "\\" evaluates to a backslash; invalid escape sequences like "\*" result in a syntax error.

TODO: Document the full list of supported escape sequences.

Unlike most other shells, double-quoted strings do not support interpolation. For instance, "$USER" simply evaluates to the string $USER. To get a similar effect, simply concatenate strings: instead of "my name is $name", write "my name is "$name. Under the hood this is a compound expression.

If a string only consists of bareword characters, it can be written without any quote; this is called a bareword. Examples are a.txt, long-bareword, and /usr/local/bin. The set of bareword characters include:

ASCII letters (a-z and A-Z) and numbers (0-9);
The symbols -_:%+,./@!;
Non-ASCII codepoints that are printable, as defined by unicode.IsPrint (https://godoc.org/unicode#IsPrint) in Go's standard library.

The following are bareword characters depending on their position:

The tilde ~, unless it appears at the beginning of a compound expression, in which case it is subject to tilde expansion;
The equal sign =, unless it is used for terminating map keys or option keys, or denoting assignments or temporary assignments.

Unlike traditional shells, an unquoted backslash \ does not escape metacharacters; use quoted strings instead. For instance, to echo a star, write echo "*" or echo '*', not echo \*. Unquote backslashes are now only used in line continuations; their use elsewhere is reserved will cause a syntax error.

The three syntaxes above all evaluate to strings, and they are interchangeable. For instance, xyz, 'xyz' and "xyz" are different syntaxes for the same string, and they are always equivalent.

Elvish does not have a separate number type. For instance, in the command + 1 2, both 1 and 2 are strings, and it is the command + that knows to treat its arguments as numbers. This design is driven by syntax -- because barewords are always treated as strings, and digits are barewords, we cannot treat words like 1 as number literals. At some point the language may get a dedicated number type (or several number types), but they will likely need to be constructed explicitly, e.g. (number 1).

Lists and maps are the basic container types in Elvish.

Lists are surround by square brackets [ ], with elements separated by whitespaces. Examples:

~> put [lorem ipsum]
▶ [lorem ipsum]
~> put [lorem
        ipsum
        foo
        bar]
▶ [lorem ipsum foo bar]

Note that commas have no special meanings and are valid bareword characters, so don't use them to separate elements:

~> li = [a, b]
~> put $li
▶ [a, b]
~> put $li[0]
▶ a,

Maps are also surrounded by square brackets; a key/value pair is written &key=value (reminiscent to HTTP query parameters), and pairs are separated by whitespaces. Whitespaces are allowed after =, but not before =. Examples:

~> put [&foo=bar &lorem=ipsum]
▶ [&foo=bar &lorem=ipsum]
~> put [&a=   10
        &b=   23
        &sum= (+ 10 23)]
▶ [&a=10 &b=23 &sum=33]

An empty map is written as [&].

Variables are named holders of values. The following characters can be used in variable names (a subset of bareword characters):

ASCII letters (a-z and A-Z) and numbers (0-9);
The symbols -_:~. The colon : is special; it is normally used for separating namespaces or denoting namespace variables;
Non-ASCII codepoints that are printable, as defined by unicode.IsPrint (https://godoc.org/unicode#IsPrint) in Go's standard library.

In most other shells, variables can map directly to environmental variables: $PATH is the same as the PATH environment variable. This is not the case in Elvish. Instead, environment variables are put in a dedicated E: namespace; the environment variable PATH is known as $E:PATH. The $PATH variable, on the other hand, does not exist initially, and if you have defined it, only lives in a certain lexical scope within the Elvish interpreter.

You will notice that variables sometimes have a leading dollar $, and sometimes not. The tradition is that they do when they are used for their values, and do not otherwise (e.g. in assignment). This is consistent with most other shells.

A variable can be assigned by writing its name, =, and the value to assign. There must be inline whitespaces both before and after =. Example:

~> foo = bar

You can assign multiple values to multiple variables simultaneously, simply by writing several variable names (separated by inline whitespaces) on the left-hand side, and several values on the right-hand side:

~> x y = 3 4

Use a variable by adding $ before the name:

~> foo = bar
~> x y = 3 4
~> put $foo
▶ bar
~> put $x
▶ 3

Variables must be assigned before use. Attempting to use an unassigned variable causes a compilation error:

~> echo $x
Compilation error: variable $x not found
[tty], line 1: echo $x
~> { echo $x }
Compilation error: variable $x not found
[tty], line 1: { echo $x }

If a variable contains a list value, you can add @ before the variable name to get all its element values. This is called exploding the variable:

~> li = [lorem ipsum foo bar]
~> put $li
▶ [lorem ipsum foo bar]
~> put $@li
▶ lorem
▶ ipsum
▶ foo
▶ bar

(This notation is restricted to exploding variables. To explode arbitrary values, use the builtin explode (/ref/builtin.html#explode) command.)

When assigning variables, if you prefix the name of the last variable with @, it gets assigned a list containing all remaining values. That variable is called a rest variable. Example:

~> a b @rest = 1 2 3 4 5 6 7
~> put $a $b $rest
▶ 1
▶ 2
▶ [3 4 5 6 7]

Schematically this is a reverse operation to variable explosion, which is why they share the @ sign.

You can prepend a command with temporary assignments, which gives variables temporarily values during the execution of that command.

In the following example, $x and $y are temporarily assigned 100 and 200:

~> x y = 1 2
~> x=100 y=200 + $x $y
▶ 300
~> echo $x $y
1 2

In contrary to normal assignments, there should be no whitespaces around the equal sign =. To have multiple variables in the left-hand side, use braces:

~> x y = 1 2
~> fn f { put 100 200 }
~> {x,y}=(f) + $x $y
▶ 300

If you use a previously undefined variable in a temporary assignment, its value will become the empty string after the command finishes. This behavior will likely change; don't rely on it.

Since ordinary assignments are also a kind of command, they can also be prepended with temporary assignments:

~> x=1
~> x=100 y = (+ 133 $x)
~> put $x $y
▶ 1
▶ 233

Temporary assignments must all appear before the command. As soon as something that is not a temporary assignments is parsed, Elvish no longer parses temporary assignments. For instance, in x=1 echo x=1, the second x=1 is not a temporary assignment, but a bareword.

Note: Elvish's behavior differs from bash (or zsh) in one important place. In bash, temporary assignments to variables do not affect their direct appearance in the command:

bash-4.4$ x=1
bash-4.4$ x=100 echo $x
1

Elvish has lexical scoping. Scopes are introduced by lambdas or user-defined modules.

When you use a variable, Elvish looks for it in the current lexical scope, then its parent lexical scope and so forth, until the outermost scope:

~> x = 12
~> { echo $x } # $x is in the global scope
12
~> { y = bar; { echo $y } } # $y is in the outer scope
bar

If a variable is not in any of the lexical scopes, Elvish tries to resolve it in the builtin: namespace, and if that also fails, cause an error:

~> echo $pid # builtin
36613
~> echo $nonexistent
Compilation error: variable $nonexistent not found
  [interactive], line 1:
    echo $nonexistent

Note that Elvish resolves all variables in a code chunk before starting to execute any of it; that is why the error message above says compilation error. This can be more clearly observed in the following example:

~> echo pre-error; echo $nonexistent
Compilation error: variable $nonexistent not found
[tty], line 1: echo pre-error; echo $nonexistent

When you assign a variable, Elvish does a similar searching. If the variable cannot be found, it will be created in the current scope:

~> x = 12
~> { x = 13 } # assigns to x in the global scope
~> echo $x
13
~> { z = foo } # creates z in the inner scope
~> echo $z
Compilation error: variable $z not found
[tty], line 1: echo $z

One implication of this behavior is that Elvish will not shadow your variable in outer scopes.

There is a local: namespace that always refers to the current scope, and by using it it is possible to force Elvish to shadow variables:

~> x = 12
~> { local:x = 13; echo $x } # force shadowing
13
~> echo $x
12

After force shadowing, you can still access the variable in the outer scope using the up: namespace, which always skips the innermost scope:

~> x = 12
~> { local:x = 14; echo $x $up:x }
14 12

The local: and up: namespaces can also be used on unshadowed variables, although they are not useful in those cases:

~> foo = a
~> { echo $up:foo } # $up:foo is the same as $foo
a
~> { bar = b; echo $local:bar } # $local:bar is the same as $bar
b

It is not possible to refer to a specific outer scope.

You cannot create new variables in the builtin: namespace, although existing variables in it can be assigned new values.

A function literal, or lambda, is a code chunk surrounded by curly braces:

~> f = { echo "Inside a lambda" }
~> put $f
▶ <closure 0x18a1a340>

One or more whitespace characters after { is required: Elvish relies on the presence of whitespace to disambiguate lambda literals and braced lists. It is good style to put some whitespace before the closing } as well, but this is not required by the syntax.

Functions are first-class values in Elvish. They can be kept in variables, used as arguments, output on the value channel, and embedded in other data structures. They can also be used as commands:

~> $f
Inside a lambda
~> { echo "Inside a literal lambda" }
Inside a literal lambda

The last command resembles a code block in C-like languages in syntax. But under the hood, it defines a function on the fly and calls it immediately.

Functions defined using the basic syntax above do not accept any arguments or options. To do so, you need to write a signature.

A signature specifies the arguments a function can accept:

~> f = [a b]{ put $b $a }
~> $f lorem ipsum
▶ ipsum
▶ lorem

There should be no space between ] and {; otherwise Elvish will parse the signature as a list, followed by a lambda without signature:

~> put [a]{ nop }
▶ <closure 0xc420153d80>
~> put [a] { nop }
▶ [a]
▶ <closure 0xc42004a480>

Like in the left hand of assignments, if you prefix the last argument with @, it becomes a rest argument, and its value is a list containing all the remaining arguments:

~> f = [a @rest]{ put $a $rest }
~> $f lorem
▶ lorem
▶ []
~> $f lorem ipsum dolar sit
▶ lorem
▶ [ipsum dolar sit]

You can also declare options in the signature. The syntax is &name=default (like a map pair), where default is the default value for the option:

~> f = [&opt=default]{ echo "Value of $opt is "$opt }
~> $f
Value of $opt is default
~> $f &opt=foobar
Value of $opt is foobar

Options must have default values: Options should be optional.

If you call a function with too few arguments, too many arguments or unknown options, an exception is thrown:

~> [a]{ echo $a } foo bar
Exception: need 1 arguments, got 2
[tty], line 1: [a]{ echo $a } foo bar
~> [a b]{ echo $a $b } foo
Exception: need 2 arguments, got 1
[tty], line 1: [a b]{ echo $a $b } foo
~> [a b @rest]{ echo $a $b $rest } foo
Exception: need 2 or more arguments, got 1
[tty], line 1: [a b @rest]{ echo $a $b $rest } foo
~> [&k=v]{ echo $k } &k2=v2
Exception: unknown option k2
[tty], line 1: [&k=v]{ echo $k } &k2=v2

User-defined functions are also known as "closures", because they have closure semantics (https://en.wikipedia.org/wiki/Closure_(computer_programming)).

In the following example, the make-adder function outputs two functions, both referring to a local variable $n. Closure semantics means that:

1.
Both functions can continue to refer to the $n variable after make-adder has returned.
2.
Multiple calls to the make-adder function generates distinct instances of the $n variables.
~> fn make-adder {
     n = 0
     put { put $n } { n = (+ $n 1) }
   }
~> getter adder = (make-adder)
~> $getter # $getter outputs $n
▶ 0
~> $adder # $adder increments $n
~> $getter # $getter and $setter refer to the same $n
▶ 1
~> getter2 adder2 = (make-adder)
~> $getter2 # $getter2 and $getter refer to different $n
▶ 0
~> $getter
▶ 1

Variables that get "captured" in closures are called upvalues; this is why the pseudo-namespace for variables in outer scopes is called up:. When capturing upvalues, Elvish only captures the variables that are used. In the following example, $m is not an upvalue of $g because it is not used:

~> fn f { m = 2; n = 3; put { put $n } }
~> g = (f)

This effect is not currently observable, but will become so when namespaces become introspectable (https://github.com/elves/elvish/issues/492).

Indexing is done by putting one or more index expressions in brackets [] after a value.

Lists can be indexed with any of the following:

A non-negative integer, an offset counting from the beginning of the list. For example, $li[0] is the first element of $li.
A negative integer, an offset counting from the back of the list. For instance, $li[-1] is the last element $li.
A slice $a:$b, where both $a and $b are integers. The result is sublist of $li[$a] up to, but not including, $li[$b]. For instance, $li[4:7] equals [$li[4] $li[5] $li[6]], while $li[1:-1] contains all elements from $li except the first and last one.

Both integers may be omitted; $a defaults to 0 while $b defaults to the length of the list. For instance, $li[:2] is equivalent to $li[0:2], $li[2:] is equivalent to $li[2:(count $li)], and $li[:] makes a copy of $li. The last form is rarely useful, as lists are immutable.

Note that the slice needs to be a single string, so there cannot be any spaces within the slice. For instance, $li[2:10] cannot be written as $li[2: 10]; the latter contains two indicies and is equivalent to $li[2:] $li[10] (see Multiple Indicies).

Not yet implemented: The string @. The result is all the values in the list. Note that this is not the same as :: if $li has 10 elements, $li[@] evaluates to 10 values (all the elements in the list), while $li[:] evaluates to just one value (a copy of the list).

When used on a variable like $li, it is equivalent to the explosion construct $li[@]. It is useful, however, when used on other constructs, like output capture or other

Examples:

~> li = [lorem ipsum foo bar]
~> put $li[0]
▶ lorem
~> put $li[-1]
▶ bar
~> put $li[0:2]
▶ [lorem ipsum]

(Negative indicies and slicing are borrowed from Python.)

NOTE: String indexing will likely change.

Strings should always be UTF-8, and they can indexed by byte indicies at which codepoints start, and indexing results in the codepoint that starts there. This is best explained with examples:

In the string elv, every codepoint is encoded with only one byte, so 0, 1, 2 are all valid indices:
~> put elv[0]
▶ e
~> put elv[1]
▶ l
~> put elv[2]
▶ v
In the string 世界, each codepoint is encoded with three bytes. The first codepoint occupies byte 0 through 2, and the second occupies byte 3 through 5. Hence valid indicies are 0 and 3:
~> put 世界[0]
▶ 世
~> put 世界[3]
▶ 界

Strings can also be indexed by slices.

(This idea of indexing codepoints by their byte positions is borrowed from Julia.)

Maps are simply indexed by their keys. There is no slice indexing, and : does not have a special meaning. Examples:

~> map = [&a=lorem &b=ipsum &a:b=haha]
~> echo $map[a]
lorem
~> echo $map[a:b]
haha

If you put multiple values in the index, you get multiple values: $li[x y z] is equivalent to $li[x] $li[y] $li[z]. This applies to all indexable values. Examples:

~> put elv[0 2 0:2]
▶ e
▶ v
▶ el
~> put [lorem ipsum foo bar][0 2 0:2]
▶ lorem
▶ foo
▶ [lorem ipsum]
~> put [&a=lorem &b=ipsum &a:b=haha][a a:b]
▶ lorem
▶ haha

Output capture is formed by putting parentheses () around a code chunk. It redirects the output of the chunk into an internal pipe, and evaluates to all the values that have been output.

~> + 1 10 100
▶ 111
~> x = (+ 1 10 100)
~> put $x
▶ 111
~> put lorem ipsum
▶ lorem
▶ ipsum
~> x y = (put lorem ipsum)
~> put $x
▶ lorem
~> put $y
▶ ipsum

If the chunk outputs bytes, Elvish strips the last newline (if any), and split them by newlines, and consider each line to be one string value:

~> put (echo "a\nb")
▶ a
▶ b

Note 1. Only the last newline is ever removed, so empty lines are preserved; (echo "a\n") evaluates to two values, "a" and "".

Note 2. One consequence of this mechanism is that you can not distinguish outputs that lack a trailing newline from outputs that have one; (echo what) evaluates to the same value as (print what). If such a distinction is needed, use slurp (/ref/builtin.html#slurp) to preserve the original bytes output.

If the chunk outputs both values and bytes, the values of output capture will contain both value outputs and lines. However, the ordering between value output and byte output might not agree with the order in which they happened:

~> put (put a; echo b) # value order need not be the same as output order
▶ b
▶ a

Exception capture is formed by putting ?() around a code chunk. It runs the chunk and evaluates to the exception it throws.

~> fail bad
Exception: bad
Traceback:
  [interactive], line 1:
    fail bad
~> put ?(fail bad)
▶ ?(fail bad)

If there was no error, it evaluates to the special value $ok:

~> nop
~> put ?(nop)
▶ $ok

Exceptions are booleanly false and $ok is booleanly true. This is useful in if (introduced later):

if ?(test -d ./a) {
  # ./a is a directory
}

Exception captures do not affect the output of the code chunk. You can combine output capture and exception capture:

output = (error = ?(commands-that-may-fail))

Tildes are special when they appear at the beginning of an expression (the exact meaning of "expression" will be explained later). The string after it, up to the first / or the end of the word, is taken as a user name; and they together evaluate to the home directory of that user. If the user name is empty, the current user is assumed.

In the following example, the home directory of the current user is /home/xiaq, while that of the root user is /root:

~> put ~
▶ /home/xiaq
~> put ~root
▶ /root
~> put ~/xxx
▶ /home/xiaq/xxx
~> put ~root/xxx
▶ /root/xxx

Note that tildes are not special when they appear elsewhere in a word:

~> put a~root
▶ a~root

If you need them to be, surround them with braces (the reason this works will be explained later):

~> put a{~root}
▶ a/root

Wildcard patterns are patterns containing wildcards, and they evaluate to all filenames they match.

We will use this directory tree in examples:

.x.conf
a.cc
ax.conf
foo.cc
d/
|__ .x.conf
|__ ax.conf
|__ y.cc
.d2/
|__ .x.conf
|__ ax.conf

Elvish supports the following wildcards:

? matches one arbitrary character except /. For example, ?.cc matches a.cc;
* matches any number of arbitrary characters except /. For example, *.cc matches a.cc and foo.cc;
** matches any number of arbitrary characters including /. For example, **.cc matches a.cc, foo.cc and b/y.cc.

The following behaviors are default, although they can be altered by modifiers:

When the entire wildcard pattern has no match, an error is thrown.
None of the wildcards matches . at the beginning of filenames. For example:
?x.conf does not match .x.conf;
d/*.conf does not match d/.x.conf;
**.conf does not match d/.x.conf.

Wildcards can be modified using the same syntax as indexing. For instance, in *[match-hidden] the * wildcard is modified with the match-hidden modifier. Multiple matchers can be chained like *[set:abc][range:0-9]. There are two kinds of modifiers:

Global modifiers apply to the whole pattern and can be placed after any wildcard:

nomatch-ok tells Elvish not to throw an error when there is no match for the pattern. For instance, in the example directory put bad* will be an error, but put bad*[nomatch-ok] does exactly nothing.
but:xxx (where xxx is any filename) excludes the filename from the final result.

Although global modifiers affect the entire wildcard pattern, you can add it after any wildcard, and the effect is the same. For example, put */*[nomatch-ok].cpp and put *[nomatch-ok]/*.cpp do the same thing.

On the other hand, you must add it after a wildcard, instead of after the entire pattern: put */*.cpp[nomatch-ok] unfortunately does not do the correct thing. (This will probably be fixed.)

Local modifiers only apply to the wildcard it immediately follows:

match-hidden tells the wildcard to match . at the beginning of filenames, e.g. *[match-hidden].conf matches .x.conf and ax.conf.

Being a local modifier, it only applies to the wildcard it immediately follows. For instance, *[match-hidden]/*.conf matches d/ax.conf and .d2/ax.conf, but not d/.x.conf or .d2/.x.conf.

Character matchers restrict the characters to match:
Character sets, like set:aeoiu;
Character ranges like range:a-z (including z) or range:a~z (excluding z);
Character classes: control, digit, graphic, letter, lower, mark, number, print, punct, space, symbol, title, and upper. See the Is* functions here (https://godoc.org/unicode) for their definitions.

Note the following caveats:

Multiple matchers, they are OR'ed. For instance, ?[set:aeoiu][digit] matches aeoiu and digits.
Dots at the beginning of filenames always require an explicit match-hidden, even if the matcher includes .. For example, ?[set:.a]x.conf does not match .x.conf; you have to ?[set:.a match-hidden]x.conf.
Likewise, you always need to use ** to match slashes, even if the matcher includes /. For example *[set:abc/] is the same as *[set:abc].

Writing several expressions together with no space in between will concatenate them. This creates a compound expression, because it mimics the formation of compound words in natural languages. Examples:

~> put 'a'b"c" # compounding three string literals
▶ abc
~> v = value
~> put '$v is '$v # compounding one string literal with one string variable
▶ '$v is value'

Many constructs in Elvish can generate multiple values, like indexing with multiple indices and output captures. Compounding multiple values with other values generates all possible combinations:

~> put (put a b)-(put 1 2)
▶ a-1
▶ a-2
▶ b-1
▶ b-2

Note the order of the generated values. The value that comes later changes faster.

NOTE: There is a perhaps a better way to explain the ordering, but you can think of the previous code as equivalent to this:

for x [a b] {
  for y [1 2] {
    put $x-$y
  }
}

In practice, you never have to write (put a b): you can use a braced list {a,b}:

~> put {a,b}-{1,2}
▶ a-1
▶ a-2
▶ b-1
▶ b-2

Elements in braced lists can also be separated with whitespaces, or a combination of comma and whitespaces (the latter not recommended):

~> put {a b , c,d}
▶ a
▶ b
▶ c
▶ d

(In future, the syntax might be made more strict.)

Braced list is merely a syntax for grouping multiple values. It is not a data structure.

Braced lists are evaluated before being compounded with other values. You can use this to affect the order of evaluation. For instance, put *.txt gives you all filenames that end with .txt in the current directory; while put {*}.txt gives you all filenames in the current directory, appended with .txt.

TODO: document evaluation order regarding tilde and wildcards.

The command is probably the most important syntax construct in shell languages, and Elvish is no exception. The word command itself, is overloaded with meanings. In the terminology of this document, the term command include the following:

An ordinary assignment, introduced above;
An ordinary command, introduced in this section;
A special command, introduced in the next section.

An ordinary command consists of a compulsory head, and any number of arguments, options and redirections.

The head must appear first. It is an arbitrary word that determines what will be run. Examples:

~> ls -l # the string ls is the head
(output omitted)
~> (put [@a]{ ls $@a }) -l
(same output)

The head must evaluate to one value. For instance, the following does not work:

~> (put [@a]{ ls $@a } -l)
Exception: head of command must be a single value; got 2 values
[tty], line 1: (put [@a]{ ls $@a } -l)

The definition of barewords is relaxed for the head to include <, >, * and ^. These are all names of numeric builtins:

~> < 3 5 # less-than
▶ $true
~> > 3 5 # greater-than
▶ $false
~> * 3 5 # multiplication
▶ 15
~> ^ 3 5 # power
▶ 243

Arguments (args for short) and options (opts for short) can be supplied to commands. Arguments are arbitrary words, while options have the same syntax as map pairs. They are separated by inline whitespaces:

~> echo &sep=, a b c # seq=, is an option; a b c are arguments
a,b,c

Redirections are used for modifying file descriptors (FD).

The most common form of redirections opens a file and associates it with an FD. The form consists of an optional destination FD (like 2), a redirection operator (like >) and a filename (like error.log):

The destination fd determines which FD to modify. If absent, it is determined from the redirection operator. If present, there must be no space between the fd and the redirection operator. (Otherwise Elvish parses it as an argument.)
The redirection operator determines the mode to open the file, and the destination fd if it is not explicitly specified.
The filename names the file to open.

Possible redirection operators and their default FDs are:

< for reading. The default FD is 0 (stdin).
> for writing. The default FD is 1 (stdout).
>> for appending. The default FD is 1 (stdout).
<> for reading and writing. The default FD is 1 (stdout).

Examples:

~> echo haha > log
~> cat log
haha
~> cat < log
haha
~> ls --bad-arg 2> error
Exception: ls exited with 2
Traceback:
  [interactive], line 1:
    ls --bad-arg 2> error
~> cat error
/bin/ls: unrecognized option '--bad-arg'
Try '/bin/ls --help' for more information.

Redirections can also be used for closing or duplicating FDs. Instead of writing a filename, use &n (where n is a number) for duplicating, or &- for closing. In this case, the redirection operator only determines the default destination FD (and is totally irrevelant if a destination FD is specified). Examples:

~> ls >&- # close stdout
/bin/ls: write error: Bad file descriptor
Exception: ls exited with 2
Traceback:
  [interactive], line 1:
    ls >&-

If you have multiple related redirections, they are applied in the order they appear. For instance:

~> fn f { echo out; echo err >&2 } # echoes "out" on stdout, "err" on stderr
~> f >log 2>&1 # use file "log" for stdout, then use (changed) stdout for stderr
~> cat log
out
err

Elvish does not impose any ordering of arguments, options and redirections: they can intermix each other. The only requirement is that the head must come first. This is different from POSIX shells, where redirections may appear before the head. For instance, the following two both work in POSIX shell, but only the former works in Elvish:

echo something > file
> file echo something # mistaken for the comparison builtin ">" in Elvish

Special commands obey the same syntax rules as normal commands (i.e. syntactically special commands can be treated the same as ordinary commands), but have evaluation rules that are custom to each command. To explain this, we use the following example:

~> or ?(echo x) ?(echo y) ?(echo z)
x
▶ $ok

In the example, the or command first evaluates its first argument, which has the value $ok (a truish value) and the side effect of outputting x. Due to the custom evaluation rule of or, the rest of the arguments are not evaluated.

If or were a normal command, the code above is still syntactically correct. However, Elvish would then evaluate all its arguments, with the side effect of outputting x, y and z, before calling or.

The del special command can be used to either delete variables or elements of lists and maps. Operands should be specified without a leading dollar sign, like the left-hand side of assignments.

Example of deleting variable:

~> x = 2
~> echo $x
2
~> del x
~> echo $x
Compilation error: variable $x not found
[tty], line 1: echo $x

Example of deleting element:

~> m = [&k=v &k2=v2]
~> del m[k2]
~> put $m
▶ [&k=v]
~> l = [[&k=v &k2=v2]]
~> del l[0][k2]
~> put $l
▶ [[&k=v]]

The and special command evaluates its arguments from left to right; as soon as a booleanly false value is obtained, it outputs the value and stops. When given no arguments, it outputs $true.

The or special command is the same except that it stops when a booleanly true value is obtained. When given no arguments, it outpus $false.

TODO: Document the syntax notation, and add more examples.

Syntax:

if <condition> {
    <body>
} elif <condition> {
    <body>
} else {
    <else-body>
}

The if special command goes through the conditions one by one: as soon as one evaluates to a booleanly true value, its corresponding body is executed. If none of conditions are booleanly true and an else body is supplied, it is executed.

The condition part is an expression, not a command like in other shells.

Tip: a combination of if and ?() gives you a semantics close to other shells:

if ?(test -d .git) {
    # do something
}

However, for Elvish's builtin predicates that output values instead of throw exceptions, the output capture construct () should be used.

TODO: add more examples.

Syntax:

while <condition> {
    <body>
} else {
    <else-body>
}

Execute the body as long as the condition evaluates to a booleanly true value.

The else body, if present, is executed if the body has never been executed (i.e. the condition evaluates to a booleanly false value in the very beginning).

Syntax:

for <var> <container> {
    <body>
} else {
    <body>
}

Iterate the container (e.g. a list). In each iteration, assign the variable to an element of the container and execute the body.

The else body, if present, is executed if the body has never been executed (i.e. the iteration value has no elements).

(If you just want to capture the exception, you can use the more concise exception capture construct ?() instead.)

Syntax:

try {
    <try-block>
} except except-varname {
    <except-block>
} else {
    <else-block>
} finally {
    <finally-block>
}

Only try and try-block are required. This control structure behaves as follows:

1.
The try-block is always executed first.
2.
If except is present and an exception occurs in try-block, it is caught and stored in except-varname, and except-block is then executed. Example:
~> try { fail bad } except e { put $e }
▶ ?(fail bad)

Note that if except is not present, exceptions thrown from try are not caught: for instance, try { fail bad } throws bad; it is equivalent to a plain fail bad.

Note that the word after except names a variable, not a matching condition. Exception matching is not supported yet. For instance, you may want to only match exceptions that were created with fail bad with except bad, but in fact this creates a variable $bad that contains whatever exception was thrown.

3.
If no exception occurs and else is present, else-block is executed. Example:
~> try { nop } else { echo well }
well
4.
If finally-block is present, it is executed. Examples:
~> try { fail bad } finally { echo final }
final
Exception: bad
Traceback:
  [tty], line 1:
    try { fail bad } finally { echo final }
~> try { echo good } finally { echo final }
good
final
5.
If the exception was not caught (i.e. except is not present), it is rethrown.

Exceptions thrown in blocks other than try-block are not caught. If an exception was thrown and either except-block or finally-block throws another exception, the original exception is lost. Examples:

~> try { fail bad } except e { fail worse }
Exception: worse
Traceback:
  [tty], line 1:
    try { fail bad } except e { fail worse }
~> try { fail bad } except e { fail worse } finally { fail worst }
Exception: worst
Traceback:
  [tty], line 1:
    try { fail bad } except e { fail worse } finally { fail worst }

Syntax:

fn <name> <lambda>

Define a function with a given name. The function behaves in the same way to the lambda used to define it, except that it "captures" return. In other words, return will fall through lambdas not defined with fn, and continues until it exits a function defined with fn:

~> fn f {
     { echo a; return }
     echo b # will not execute
   }
~> f
a
~> {
     f
     echo c # executed, because f "captures" the return
   }
a
c

TODO: Find a better way to describe this. Hopefully the example is illustrative enough, though.

When using a literal string as the head of a command, it is first resolved during the compilation phase, using the following order:

1.
Special commands.
2.
Functions defined on any of the containing lexical scopes, with inner scopes looked up first. (This is exactly the same process as variable resolution).
3.
The builtin: namespace.
4.
Should all these steps fail, it is resolved to be an external command. Determination of the path of the external command does not happen in the resolution process; that happens during evaluation.

You can use the resolve (/ref/builtin.html#resolve) command to see which command Elvish resolves a string to.

During the evaluation phase, external commands are then subject to searching. This can be observed with the search-external (/ref/builtin.html#search-builtin) command.

A pipeline is formed by joining one or more commands together with the pipe sign (|).

For each pair of adjacent commands a | b, the output of a is connected to the input of b. Both the byte pipe and the value channel are connected, even if one of them is not used.

Command redirections are applied before the connection happens. For instance, the following writes foo to a.txt instead of the output:

~> echo foo > a.txt | cat
~> cat a.txt
foo

All of the commands in a pipeline are executed in parallel, and the execution of the pipeline finishes when all of its commands finish execution.

If one or more command in a pipeline throws an exception, the other commands will continue to execute as normal. After all commands finish execution, an exception is thrown, the value of which depends on the number of commands that have thrown an exception:

If only one command has thrown an exception, that exception is rethrown.
If more than one commands have thrown exceptions, a "composite exception", containing information all exceptions involved, is thrown.

Adding an ampersand & to the end of a pipeline will cause it to be executed in the background. In this case, the rest of the code chunk will continue to execute without waiting for the pipeline to finish. Exceptions thrown from the background pipeline do not affect the code chunk that contains it.

When a background pipeline finishes, a message is printed to the terminal if the shell is interactive.

A code chunk is formed by joining zero or more pipelines together, separating them with either newlines or semicolons.

Pipelines in a code chunk are executed in sequence. If any pipeline throws an exception, the execution of the whole code chunk stops, propagating that exception.

Exceptions have similar semantics to those in Python or Java. They can be thrown with the fail (/ref/builtin.html#fail) command and caught with either exception capture ?() or the try special command.

If an external command exits with a non-zero status, Elvish treats that as an exception.

Flow commands -- break, continue and return -- are ordinary builtin commands that raise special "flow control" exceptions. The for and while commands capture break and continue, while fn modifies its closure to capture return.

One interesting implication is that since flow commands are just ordinary commands you can build functions on top of them. For instance, this function breaks randomly:

fn random-break {
  if eq (randint 2) 0 {
    break
  }
}

The function random-break can then be used in for-loops and while-loops.

Note that the return flow control exception is only captured by functions defined with fn. It falls through ordinary lambdas:

fn f {
  {
    # returns f, falling through the innermost lambda
    return
  }
}

Namespace in Elvish helps prevent name collisions and is important for building modules.

Prepend namespace: to command names and variable names to specify the namespace. The following code

e:echo $E:PATH

uses the echo command from the e namespace and the PATH variable from the E namespace.

Names of namespaces can contain colons themselves. For instance, in $x:y:z the namespace is x:y. More precisely, when parsing command names and variable names, everything up to the last colon is considered to be the namespace.

A convention used in articles is when referring to a namespace is to add a trailing colon: for instance, edit: means "the namespace edit". This is remniscent to the syntax but is not syntactically valid.

The following namespaces have special meanings to the language:

local: and up: refer to lexical scopes, and have been documented above.
e: refers to externals. For instance, e:ls refers to the external command ls.

Most of the time you can rely on the rules of command resolution and do not need to use this explicitly, unless a function defined by you (or an Elvish builtin) shadows an external command.

E: refers to environment variables. For instance, $E:USER is the environment variable USER.

This is always needed, because unlike command resolution, variable resolution does not fall back onto environment variables.

builtin: refers to builtin functions and variables.

You don't need to use this explicitly unless you have defined names that shadows builtin counterparts.

Namespaces that are not special (i,e. one of the above) are also called modules. Aside from these special namespaces, Elvish also comes with the following modules:

edit: for accessing the Elvish editor. This module is available in interactive mode.

See reference (/ref/edit.html).

re: for regular expression facilities. This module is always available. See reference (/ref/re.html).
daemon: for manipulating the daemon. This module is always available.

This is not yet documented.

You can define your own modules with Elvishscript but putting them under ~/.elvish/lib and giving them a .elv extension. For instance, to define a module named a, store it in ~/.elvish/lib/a.elv:

~> cat ~/.elvish/lib/a.elv
echo "mod a loading"
fn f {
  echo "f from mod a"
}

To import the module, use use:

~> use a
mod a loading
~> a:f
f from mod a

The argument to use is called the usespec and will be explained in more details below. In the simplest case, it is simply the module name.

Modules are evaluated in a seprate scope. That means that functions and variables defined in the module does not pollute the default namespace, and vice versa. For instance, if you define ls as a wrapper function in rc.elv:

fn ls [@a]{
    e:ls --color=auto $@a
}

That definition is not visible in module files: ls will still refer to the external command ls, unless you shadow it in the very same module.

It is often useful to put modules under directories. When importing such modules, you can control which (trailing) parts becomes the module name.

For instance, if you have the following content in x/y/z.elv (relative to ~/.elvish/lib, of course):

fn f {
    echo 'In a deeply nested module'
}

It is possible to import the module as z, y:z, or x:y:z:

~> use x/y/z # imports as "z"
~> z:f
In a deeply nested module
~> use x/y:z # imports as "y:z"
~> y:z:f
In a deeply nested module
~> use x:y:z # imports as "x:y:z"
~> x:y:z:f
In a deeply nested module

To be precise:

The path of the module to import is derived by replacing all : with / and adding .elv. In the above example, all of x/y/z, x/y:z and x:y:z have the same path x/y/z.elv and refer to the same module
The part after the last / becomes the module name.

Namespace imports are also lexically scoped. For instance, if you use a module within an inner scope, it is not available outside that scope:

{
    use some-mod
    some-mod:some-func
}
some-mod:some-func # not valid

Modules are cached after one import. Subsequent imports do not re-execute the module; they only serve the bring it into the current scope. Moreover, the cache is keyed by the path of the module, not the name under which it is imported. For instance, if you have the following in ~/.elvish/lib/a/b.elv:

echo importing

The following code only prints one importing:

{ use a:b }
use a:b # only brings mod into the lexical scope

As does the following:

use a/b
use a:b
February 8, 2019 Elvish 0.12