Use turtle if you want to write light-weight and maintainable shell scripts.

turtle embeds shell scripting directly within Haskell for three main reasons:

• Haskell code is easy to refactor and maintain because the language is statically typed
• Haskell is syntactically lightweight, thanks to global type inference
• Haskell programs can be type-checked and interpreted very rapidly (< 1 second)

These features make Haskell ideal for scripting, particularly for replacing large and unwieldy Bash scripts.

This tutorial introduces how to use the turtle library to write Haskell scripts. This assumes no prior knowledge of Haskell, but does assume prior knowledge of Bash or a similar shell scripting language.

If you are already proficient with Haskell, then you can get quickly up to speed by reading the Quick Start guide at the top of Turtle.Prelude.

If you are on Windows, the easiest way to follow along is to install Git for Windows and use the Git Bash program that it installs to get a fully featured Unix-like environment.

For all operating systems, the recommended way to compile and run the following examples is to download the stack package management tool by following the instructions here:

https://github.com/commercialhaskell/stack

... and then run the following instruction anywhere outside of a Haskell project:

$ stack install turtle

This tutorial will mostly focus on using Haskell as a scripting language. The first two lines of each script below contain boilerplate instructions so that stack will load and run the script. This helps ensure that a script will run on any computer that has a stack executable, as stack can install a Haskell compiler if one is not already present. If you are curious about how these two lines work, they are described here:

https://github.com/commercialhaskell/stack/blob/master/doc/GUIDE.md#script-interpreter

If you want to make a Windows script independently executable outside of a Git Bash environment, you can either (A) compile the script into an executable or (B) run these two commands from a cmd shell with administrator privileges to make all *.hs scripts executable:

assoc .hs=Haskell
ftype Haskell="C:\path\to\stack.exe" "%1" %*

Let's translate some simple Bash scripts to Haskell and work our way up to more complex scripts. Here is an example "Hello, world!" script written in both languages:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script
 
                                    -- #!/bin/bash
{-# LANGUAGE OverloadedStrings #-}  --
                                    --
import Turtle                       --
                                    --
main = echo "Hello, world!"         -- echo Hello, world!

In Haskell you can use -- to comment out the rest of a line. The above example uses comments to show the equivalent Bash script side-by-side with the Haskell script.

You can execute the above code by saving it to the file example.hs. If you are copying and pasting the code, then remove the leading 1-space indent. After you save the file, make the script executable and run the script:

$ chmod u+x example.hs 
$ ./example.hs
Hello, world!

If you delete the first two lines of the program, you can also compile the above code to generate a native executable which will have a much faster startup time and improved performance:

$ # `-O2` turns on all optimizations
$ # `-threaded` helps with piping shell output in and out of Haskell
$ stack ghc -- -O2 -threaded example.hs
$ ./example
Hello, world!

You can even run Haskell code interactively using ghci, which is an interactive REPL for Haskell. You can either use ghci by itself:

$ stack ghci
<ghci links in some libraries>
Prelude> :set -XOverloadedStrings
Prelude> import Turtle
Prelude Turtle> echo "Hello, world!"
<ghci links in some libraries>
Hello, world!
Prelude Turtle> :quit
$

From now on I'll omit ghci's linker output in tutorial examples. You can also silence this linker output by passing --ghc-options -v0 to stack ghci.

You'll already notice a few differences between the Haskell code and Bash code.

First, the Haskell code requires two additional lines of overhead to import the turtle library and enable overloading of string literals. This overhead is mostly unavoidable.

Second, the Haskell echo explicitly quotes its string argument whereas the Bash echo does not. In Bash every token is a string by default and you distinguish variables by prepending a dollar sign to them. In Haskell the situation is reversed: every token is a variable by default and you distinguish strings by quoting them. The following example highlights the difference:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

                                    -- #!/bin/bash
{-# LANGUAGE OverloadedStrings #-}  --
                                    --
import Turtle                       --
                                    --
str = "Hello!"                      --STR=Hello!
                                    --
main = echo str                     --echo $STR

Third, you have to explicitly assign a subroutine to main to specify which subroutine to run when your program begins. This is because Haskell lets you define things out of order. For example, we could have written our original program this way instead:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

main = echo str

str = "Hello, world!"

Notice how the above program defines str after main, which is valid. Haskell does not care in what order you define top-level values or functions (using the = sign). However, the top level of a Haskell program only permits definitions. If you were to insert a statement at the top-level:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

echo "Hello, world!"

... then you would get this error when you tried to run your program:

example.hs:7:1: Parse error: naked expression at top level

You can use do notation to create a subroutine that runs more than one command:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

                                    -- #!/bin/bash
{-# LANGUAGE OverloadedStrings #-}  --
                                    --
import Turtle                       --
                                    --
main = do                           --
    echo "Line 1"                   -- echo Line 1
    echo "Line 2"                   -- echo Line 2

$ ./example.hs
Line 1
Line 2

do blocks can use either use the indentation level to control their duration or they can use curly braces and semicolons. To see the full rules for do syntax, read: http://en.wikibooks.org/wiki/Haskell/Indentation.

Some commands can return a value, and you can store the result of a command using the <- symbol. For example, the following program prints the creation time of the current working directory by storing two intermediate results:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script
 
                           -- #!/bin/bash
import Turtle              --
                           --
main = do                  --
    dir  <- pwd            -- DIR=$(pwd)
    time <- datefile dir   -- TIME=$(date -r $DIR)
    print time             -- echo $TIME

$ ./example.hs
2015-01-24 03:40:31 UTC

The main difference between = and <- is that:

• The <- symbol is overloaded and its meaning is context-dependent; in this context it just means "store the current result"
• The = symbol is not overloaded and always means that the two sides of the equality are interchangeable

do notation lets you combine smaller subroutines into larger subroutines. For example, we could refactor the above code to split the first two commands into their own smaller subroutine and then invoke that smaller subroutine within a larger subroutine:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

                            -- #!/bin/bash
import Turtle               --
                            --
datePwd = do                -- datePwd() {
    dir    <- pwd           --     DIR=$(pwd)
    result <- datefile dir  --     RESULT=$(date -r $DIR)
    return result           --     echo $RESULT
                            -- }
main = do                   --
    time <- datePwd         -- TIME=$(datePwd)
    print time              -- echo $TIME

The refactored program still returns the exact same result:

$ ./example.hs
2015-01-24 03:40:31 UTC

We can also simplify the code a little bit because do notation implicitly returns the value of the last command within a subroutine. We can use this trick to simplify both the Haskell and Bash code:

datePwd = do      -- datePwd() {
    dir <- pwd    --     DIR=$(pwd)
    datefile dir  --     date -r $DIR
                  -- }

However, keep in mind that the return statement is something of a misnomer since it does not break or exit from the surrounding subroutine. All it does is create a trivial subroutine that has no side effects and returns its argument as its result. If you return an expression, you're just giving it a new name:

do x <- return expr  -- X=EXPR
   command x         -- command $X

-- Same as:
command expr         -- command EXPR

In fact, the first line is equivalent to let x = expr, which more closely mirrors the equivalent Bash syntax:

do let x = expr      -- X=EXPR
   command x         -- command $X

-- Same as:
command expr         -- command EXPR

Also, for a subroutine with a single command, you can omit the do:

main = do echo "Hello, world!"

-- Same as:
main =    echo "Hello, world!"

Notice how the above Haskell example used print instead of echo. Run the following script to find out what happens if we choose echo instead:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

import Turtle

main = do
    dir  <- pwd
    time <- datefile dir
    echo time

If we run that we get a type error:

$ ./example.hs

example.hs:8:10:
    Couldn't match expected type `Line' with actual type `UTCTime'
    In the first argument of `echo', namely `time'
    In a stmt of a 'do' block: echo time
    In the expression:
      do { dir <- pwd;
           time <- datefile dir;
           echo time }

The error points to the last line of our program: (example.hs:8:10) means line 8, column 10 of our program. If you study the error message closely you'll see that the echo function expects a Line value (a piece of text without newlines), but we passed it 'time', which was a UTCTime value. Although the error is at the end of our script, Haskell catches this error before even running the script. When we "interpret" a Haskell script the Haskell compiler actually compiles the script without any optimizations to generate a temporary executable and then runs the executable, much like Perl does for Perl scripts.

You might wonder: "where are the types?" None of the above programs had any type signatures or type annotations, yet the compiler still detected type errors correctly. This is because Haskell uses "global type inference" to detect errors, meaning that the compiler can infer the types of expressions within the program without any assistance from the programmer.

You can even ask the compiler what the type of an expression is using ghci. Let's open up the REPL and import this library so that we can study the types and deduce why our program failed:

$ stack ghci
Prelude> import Turtle
Prelude Turtle>

You can interrogate the REPL for an expression's type using the :type command:

Prelude Turtle> :type pwd
pwd :: MonadIO io => io Turtle.FilePath

For right now, ignore all occurrences of MonadIO and just read the type as:

Prelude Turtle> :type pwd
pwd :: IO Turtle.FilePath

We will cover MonadIO later on.

Whenever you see something of the form (x :: t), that means that 'x' is a value of type 't'. The REPL says that pwd is a subroutine (IO) that returns a FilePath. The Turtle prefix before FilePath is just the module name since the FilePath exported by the turtle library conflicts with the default FilePath exported by Haskell's Prelude. The compiler uses the fully qualified name, Turtle.FilePath, to avoid ambiguity.

We can similarly ask for the type of datefile:

Prelude Turtle> :type datefile
datefile :: Turtle.FilePath -> IO UTCTime

datefile is a function whose argument must be a FilePath and whose result is a subroutine (IO) that returns a UTCTime. Notice how the input argument of datefile (which is a FilePath) is the same type as the return value of pwd (also a FilePath).

Now let's study type of echo to see why we get the type error:

Prelude Turtle> :type echo
echo :: Line -> IO ()

The above type says that echo is a function whose argument is a value of type Line and whose result is a subroutine (IO) with an empty return value (denoted '()').

Line is a wrapper around Text and represents a Text value with no internal newlines:

newtype Line = Line Text

Now we can understand the type error: echo expects a Line argument but datefile returns a UTCTime, which is not the same thing. Unlike Bash, not everything is Text in Haskell and the compiler will not cast or coerce types for you.

The reason print worked is because print has a more general type than echo:

Prelude Turtle> :type print
print :: Show a => a -> IO ()

This type signature says that print can display any value of type 'a' so long as 'a' implements the Show interface. In this case UTCTime does implement the Show interface, so everything works out when we use print.

This library provides a helper function that lets you convert any type that implements Show into any other type that implements IsString:

-- This behaves like Python's `repr` function
repr :: (Show a, IsString b) => a -> b

You could therefore implement print in terms of echo and repr:

 print x = echo (repr x)

... which works because Line implements IsString

You can use ghci for more than just inferring types. ghci is a general-purpose Haskell shell for your system when you extend it with turtle:

$ stack ghci
Prelude> :set -XOverloadedStrings
Prelude> import Turtle
Prelude Turtle> cd "/tmp"
Prelude Turtle> pwd
FilePath "/tmp"
Prelude Turtle> mkdir "test"
Prelude Turtle> cd "test"
Prelude Turtle> touch "file"
Prelude Turtle> testfile "file"
True
Prelude Turtle> rm "file"
Prelude Turtle> testfile "file"
False

You can also optionally configure ghci to run the first two commands every time you launch ghci. Just create a .ghci within your current directory with these two lines:

:set -XOverloadedStrings
import Turtle

The following ghci examples will all assume that you run these two commands at the beginning of every session, either manually or automatically. You can even enable those two commands permanently by adding the above .ghci file to your home directory.

Within ghci you can run a subroutine and ghci will print the subroutine's value if it is not empty:

Prelude Turtle> shell "true" empty
ExitSuccess
Prelude Turtle> shell "false" empty
ExitFailure 1

You can also type in a pure expression and ghci will evaluate that expression:

Prelude Turtle> 2 + 2
4
Prelude Turtle> "123" <> "456"  -- (<>) concatenates strings
"123456"

This works because ghci automatically wraps anything that's not a subroutine with print. It's as if we had written:

Prelude Turtle> print (2 + 2)
4
Prelude Turtle> print ("123" <> "456")
"123456"

Haskell performs global type inference, meaning that the compiler never requires any type signatures. When you add type signatures, they are purely for the benefit of the programmer and behave like machine-checked documentation.

Let's illustrate this by adding types to our original script:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

import Turtle

datePwd :: IO UTCTime  -- Type signature
datePwd = do
    dir <- pwd
    datefile dir

main :: IO ()          -- Type signature
main = do
    time <- datePwd
    print time

The first type signature says that datePwd is a subroutine that returns a UTCTime:

--         +----- A subroutine ...
--         |
--         |  +-- ... that returns `UTCTime`
--         |  |
--         v  v
datePwd :: IO UTCTime

The second type signature says that main is a subroutine that returns an empty value:

--      +----- A subroutine ...
--      |
--      |  +-- ... that returns an empty value (i.e. `()`)
--      |  |
--      v  v
main :: IO ()

Not every top-level value has to be a subroutine, though. For example, you can define unadorned Line values at the top-level, as we saw previously:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

str :: Line
str = "Hello!"

main :: IO ()
main = echo str

These type annotations do not assist the compiler. Instead, the compiler independently infers the type and then checks whether it matches the documented type. If there is a mismatch the compiler will raise a type error.

Let's test this out by providing an incorrect type for 'str':

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

str :: Int
str = "Hello!"

main :: IO ()
main = echo str

If you run that script, you will get two error messages:

$ ./example.hs

example.hs:8:7:
    No instance for (IsString Int)
      arising from the literal `"Hello, world!"'
    Possible fix: add an instance declaration for (IsString Int)
    In the expression: "Hello, world!"
    In an equation for `str': str = "Hello, world!"

example.hs:11:13:
    Couldn't match expected type `Line' with actual type `Int'
    In the first argument of `echo', namely `str'
    In the expression: echo str
    In an equation for `main': main = echo str

The first error message relates to the OverloadedStrings extensions. When we enable OverloadedStrings the compiler overloads string literals, interpreting them as any type that implements the IsString interface. The error message says that Int does not implement the IsString interface so the compiler cannot interpret a string literal as an Int. On the other hand the Text, Line and FilePath types do implement IsString, which is why we can interpret string literals as Text, Line or FilePath values.

The second error message says that echo expects a Line value, but we declared str to be an Int, so the compiler aborts compilation, requiring us to either fix or delete our type signature.

Notice that there is nothing wrong with the program other than the type signature we added. If we were to delete the type signature the program would compile and run correctly. The sole purpose of this type signature is for us to communicate our expectations to the compiler so that the compiler can alert us if the code does not match our expectations.

Let's also try reversing the type error, providing a number where we expect a string:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

str :: Line
str = 4

main :: IO ()
main = echo str

This gives a different error:

$ ./example.hs

example.hs:8:7:
    No instance for (Num Line)
      arising from the literal `4'
    Possible fix: add an instance declaration for (Num Line)
    In the expression: 4
    In an equation for `str': str = 4

Haskell also automatically overloads numeric literals, too. The compiler interprets integer literals as any type that implements the Num interface. The Line type does not implement the Num interface, so we cannot interpret integer literals as Line strings.

You can invoke arbitrary shell commands using the shell command. For example, we can write a program that creates an empty directory and then uses a shell command to archive the directory:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script
 
                                             -- #!/bin/bash
{-# LANGUAGE OverloadedStrings #-}           --
                                             --
import Turtle                                --
                                             --
main = do                                    --
    mkdir "test"                             -- mkdir test
    shell "tar czf test.tar.gz test" empty   -- tar czf test.tar.gz test

If you run this program, it will generate the test.tar.gz archive:

$ ./example.hs
ExitSuccess
$ echo $?
0
$ ls test.tar.gz
test.tar.gz

Like ghci, the runhaskell command running our script prints any non-empty result of the main subroutine (ExitSuccess in this case).

The easiest way to learn a new command like shell is to view its documentation. Click on the word shell, which will take you to documentation that looks like this:

shell
    :: Text         -- Command line
    -> Shell Line   -- Standard input (as lines of `Text`)
    -> IO ExitCode  -- Exit code of the shell command

The first argument is a Text representation of the command to run. The second argument lets you feed input to the command, and you can provide empty for now to feed no input.

The final result is an ExitCode, which you can use to detect whether the command completed successfully. For example, we could print a more descriptive error message if an external command fails:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

main = do
    let cmd = "false"
    x <- shell cmd empty
    case x of
        ExitSuccess   -> return ()
        ExitFailure n -> die (cmd <> " failed with exit code: " <> repr n)

This prints an error message since the false command always fails:

$ ./example.hs
example.hs: user error (false failed with exit code: 1)

You should also check out the proc command, which is less powerful but safer since it decreases the likelihood of code injection or malformed commands:

proc
    :: Text         -- Program
    -> [Text]       -- Arguments
    -> Shell Line   -- Standard input (as lines of `Text`)
    -> IO ExitCode  -- Exit code of the shell command

Most of the commands in this library do not actually invoke an external shell or program. Instead, they indirectly wrap other Haskell libraries that bind to C code.

Also, some people prefer that subprocess runners throw exceptions instead of returning an ExitCode. procs and shells are the exception-throwing variations on proc and shell.

This library provides type-safe string formatting utilities, too. For example, instead of writing this:

 cmd <> " failed with exit code: " <> repr n

... you could format the string using printf style instead:

 format (s%" failed with exit code: "%d) cmd n

What's neat is that the compiler will automatically infer the number of arguments and their types from the Format string:

$ stack ghci
Prelude Turtle> :type format (s%" failed with exit code: "%d)
format (s%" failed with exit code: "%d) :: Text -> Int -> Text

The compiler deduces that the above Format string requires one argument of type Text to satisfy the s at the beginning of the format string and another argument of type Int to satisfy the d at the end of the format string.

Note that this is also the idiomatic way to convert a FilePath to Text:

format fp :: FilePath -> Text

For the very common case where you echo a formatted value, you can use printf:

>>> printf ("Hello, "%s%"!\n") "world"
Hello, world!

If you are interested in this feature, check out the Turtle.Format module for more details. For more complex string formatting needs, check out the text-format library.

The turtle library provides support for streaming computations, just like Bash. The primitive turtle streams are little more verbose than their Bash counterparts, but turtle streams can be built and combined in more ways.

The key type for streams is the Shell type, which represents a stream of values. For example, the ls function has a streaming result:

Prelude Turtle> :type ls
ls :: Turtle.FilePath -> Shell Turtle.FilePath

That type says that ls takes a single FilePath as its argument (the directory to list) and the result is a Shell stream of FilePaths (the immediate children of that directory).

You can't run a Shell stream directly within ghci. You will get a type error like this if you try:

Prelude Turtle> ls "/tmp"

<interactive>:2:1:
    No instance for (Show (Shell Turtle.FilePath))
      arising from a use of `print'
    Possible fix:
      add an instance declaration for (Show (Shell Turtle.FilePath))
    In a stmt of an interactive GHCi command: print it

Instead, you must consume the stream as it is generated and the simplest way to consume a Shell stream is view:

view :: Show a => Shell a -> IO ()

view takes any Shell stream of values and prints them to standard output:

Prelude Turtle> view (ls "/tmp")
FilePath "/tmp/.X11-unix"
FilePath "/tmp/.X0-lock"
FilePath "/tmp/pulse-PKdhtXMmr18n"
FilePath "/tmp/pulse-xHYcZ3zmN3Fv"
FilePath "/tmp/tracker-gabriella"
FilePath "/tmp/pulse-PYi1hSlWgNj2"
FilePath "/tmp/orbit-gabriella"
FilePath "/tmp/ssh-vREYGbWGpiCa"
FilePath "/tmp/.ICE-unix

You can build your own Shell streams using a few primitive operations,

The first primitive is empty, which represents an empty stream of values:

Prelude Turtle> view empty  -- Outputs nothing
Prelude Turtle>

Another way to say that is:

view empty = return ()

The type of empty is:

empty :: Shell a

The lower-case 'a' is "polymorphic", meaning that it will type check as any type. That means that you can produce an empty stream of any type of value.

The next simplest function is return, which lets you take any value and transform it into a singleton Shell that emits just that one value:

Prelude Turtle> view (return 1)
1

Another way to say that is:

view (return x) = print x

The type of return is:

return :: a -> Shell a

Notice that this is the same return function we saw before. This is because return is overloaded and works with both IO and Shell.

You can also take any subroutine (IO) and transform it into a singleton Shell:

Prelude Turtle> view (liftIO readline)
ABC<Enter>
Just (Line ABC)

Another way to say that is:

view (liftIO io) = do x <- io
                      print x

The type of liftIO is:

liftIO :: IO a -> Shell a

Once you have those primitive Shell streams you can begin to combine them into larger Shell streams. For example, you can concatenate two Shell streams using (<|>):

view (return 1 <|> return 2)
1
2

Another way to say that is:

view (xs <|> ys) = do view xs
                      view ys

The type of (<|>) is:

(<|>) :: Shell a -> Shell a -> Shell a

In other words, you can concatenate two Shell streams of the same element type to get a new Shell stream, also of the same element type.

Let's try using (<|>) on two real streams:

Prelude Turtle> view (ls "/tmp" <|> ls "/usr")
FilePath "/tmp/.X11-unix"
FilePath "/tmp/.X0-lock"
FilePath "/tmp/pulse-PKdhtXMmr18n"
FilePath "/tmp/pulse-xHYcZ3zmN3Fv"
FilePath "/tmp/tracker-gabriella"
FilePath "/tmp/pulse-PYi1hSlWgNj2"
FilePath "/tmp/orbit-gabriella"
FilePath "/tmp/ssh-vREYGbWGpiCa"
FilePath "/tmp/.ICE-unix"
FilePath "/usr/lib"
FilePath "/usr/src"
FilePath "/usr/sbin"
FilePath "/usr/include"
FilePath "/usr/share"
FilePath "/usr/games"
FilePath "/usr/local"
FilePath "/usr/bin"

Finally, note that Shell implements the IsString interface, so a string literal will type-check as a Shell that emits a single Text value:

Prelude Turtle> view "123"
"123"
Prelude Turtle> view (return "123")  -- Same thing
"123"
Prelude Turtle> view ("123" <|> "456")
"123"
"456"
Prelude Turtle> view (return "123" <|> return "456")  -- Same thing
"123"
"456"

This library also provides the select function for conveniently emitting a list of values:

Prelude Turtle> view (select [1, 2, 3])
1
2
3

We can use select to implement loops within a Shell:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

                                    -- #!/bin/bash
{-# LANGUAGE OverloadedStrings #-}  --
                                    --
import Turtle                       --
                                    --
example = do                        --
    x <- select [1, 2]              -- for x in 1 2; do
    y <- select [3, 4]              --     for y in 3 4; do
    liftIO (print (x, y))           --         echo \(${x},${y}\);
                                    --     done;
main = sh example                   -- done

That will print every permutation of 'x' and 'y':

$ ./example
(1,3)
(1,4)
(2,3)
(2,4)

This is because Shell behaves like a list comprehension, running each following command once for each element in the stream. This implies that an Shell stream that produces 0 elements will short-circuit and prevent subsequent commands from being run.

-- This stream emits 0 elements but still has side effects
inner :: Shell a
inner = do
    x <- select [1, 2]
    y <- select [3, 4]
    liftIO (print (x, y))
    empty

outer :: Shell ()
outer = do
    inner
    liftIO (echo "This step will never run")

If you want to run a Shell stream just for its side effects, wrap the Shell with sh. This ensures that you don't alter the surrounding Shell's control flow by unintentionally running subsequent commands zero times or multiple times:

outer :: Shell ()
outer = do
    sh inner
    liftIO (echo "Now this step will exactly once")

This uses the sh utility instead of view. The only difference is that sh doesn't print any values (since print is doing that already):

sh :: Shell a -> IO ()

This trick isn't limited to select. You can loop over the output of any Shell by just binding its result. For example, this is how view loops over its argument:

view :: Show a => Shell a -> IO ()
view s = sh (do
    x <- s -- `x` ranges over every output of `s`
    liftIO (print x) )

You can also loop over a stream in a one-liner, still using do notation. Just insert semi-colons between statements:

Prelude Turtle> -- for file in /tmp/*; do echo $file; done
Prelude Turtle> sh (do file <- ls "/tmp"; liftIO (print file))
FilePath "/tmp/.X11-unix"
FilePath "/tmp/.X0-lock"
FilePath "/tmp/pulse-PKdhtXMmr18n"
FilePath "/tmp/pulse-xHYcZ3zmN3Fv"
FilePath "/tmp/tracker-gabriella"
FilePath "/tmp/pulse-PYi1hSlWgNj2"
FilePath "/tmp/orbit-gabriella"
FilePath "/tmp/ssh-vREYGbWGpiCa"
FilePath "/tmp/.ICE-unix"

You can filter streams using Control.Monad.mfilter, like this:

>>> view (select [1..10])
1
2
3
4
5
6
7
8
9
10
>>> view (mfilter even (select [1..10]))
2
4
6
8
10

This works because mfilter's implementation is equivalent to:

mfilter predicate stream = do
    element <- stream
    if predicate element then return element else empty

In other words, mfilter loops over each element of the stream and only returns the element if the predicate is True

There are other ways you can consume a Shell stream. For example, you can fold the stream using predefined Folds from Control.Foldl:

Prelude Turtle> import qualified Control.Foldl as Fold
Prelude Turtle Fold> fold (ls "/tmp") Fold.length
9

Prelude Turtle Fold> fold (ls "/tmp") Fold.head
Just (FilePath "/tmp/.X11-unix")

Prelude Turtle Fold> fold (ls "/tmp") Fold.list
[FilePath "/tmp/.X11-unix",FilePath "/tmp/.X0-lock",FilePath "/tmp/pulse-PKd
htXMmr18n",FilePath "/tmp/pulse-xHYcZ3zmN3Fv",FilePath "/tmp/tracker-gabriella
",FilePath "/tmp/pulse-PYi1hSlWgNj2",FilePath "/tmp/orbit-gabriella",FilePath 
"/tmp/ssh-vREYGbWGpiCa",FilePath "/tmp/.ICE-unix"]

You can also compute multiple things in a single pass over the stream:

Prelude Turtle> fold (select [1..10]) ((,) <$> Fold.minimum <*> Fold.maximum)
(Just 1,Just 10)

If you are interested in this feature, check out the documentation in Control.Foldl.

turtle comes with built-in support for the standard text streams.

For example, you can write to standard output using the stdout utility:

stdout :: Shell Line -> IO ()
stdout s = sh (do
    txt <- s
    liftIO (echo txt) )

stdout outputs each Line value on its own line:

Prelude Turtle> stdout "Line 1"
Line 1
Prelude Turtle> stdout ("Line 1" <|> "Line 2")
Line 1
Line 2

Another useful stream is stdin, which emits one Line value per line of standard input:

stdin :: Shell Line

Let's combine stdin and stdout to forward all input from standard input to standard output:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

                                    -- #!/bin/bash
{-# LANGUAGE OverloadedStrings #-}  --
                                    --
import Turtle                       --
                                    --
main = stdout stdin                 -- cat

If you run that it will continue to echo lines until you signal end of input using Ctrl-D:

$ ./example.hs
ABC<Enter>
ABC
Test<Enter>
Test
42<Enter>
42
<Ctrl-D>
$

You can also read and write to files using the input and output utilities:

Prelude Turtle> output "file.txt" ("Test" <|> "ABC" <|> "42")
Prelude Turtle> stdout (input "file.txt")
Test
ABC
42

You can embed external shell commands as streams within your Haskell program.

For example, suppose that we want to use the system's built in ls command. We can just run:

Prelude Turtle> stdout (inshell "ls" empty)
.X11-unix
.X0-lock
pulse-PKdhtXMmr18n
pulse-xHYcZ3zmN3Fv
tracker-gabriella
pulse-PYi1hSlWgNj2
orbit-gabriella
ssh-vREYGbWGpiCa
.ICE-unix

This works because type of inshell is:

inshell
    :: Text    -- Command line
    -> Shell Line  -- Standard input to feed to program
    -> Shell Line  -- Standard output produced by program

This means you can use inshell to embed arbitrary external utilities as first class streams within your Haskell program:

Turtle Prelude> stdout (inshell "awk '{ print $1 }'" "123 456")
123

You should also check out the inproc command, which is less powerful but safer since it decreases the likelihood of code injection or malformed commands:

inproc
    :: Text        -- Program
    -> [Text]      -- Arguments
    -> Shell Line  -- Standard input to feed to program
    -> Shell Line  -- Standard output produced by program

Using inproc, you would write:

Turtle Prelude> stdout (inproc "awk" ["{ print $1 }"] "123 456")
123

You can transform streams using Unix-like utilities. For example, you can filter a stream using grep.

Prelude Turtle> stdout (input "file.txt")
Test
ABC
42
Prelude Turtle> stdout (grep "ABC" (input "file.txt"))
ABC

Let's look at the type of grep:

grep :: Pattern a -> Shell Line -> Shell Line

The first argument of grep is actually a Pattern, which implements IsString. When we pass a string literal we just create a Pattern that matches the given literal.

Patterns generalize regular expressions and you can use this table to roughly translate several regular expression idioms to Patterns:

Regex      Pattern
=========  =========
"string"   "string"
.          dot
e1 e2      e1 <> e2
e1 | e2    e1 <|> e2
e*         star e
e+         plus e
e*?        selfless (star e)
e+?        selfless (plus e)
e{n}       count n e
e{m,n}     bounded m n e
e{0,n}     upperBounded n e
e?         optional e
[xyz]      oneOf "xyz"
[^xyz]     noneOf "xyz"

Here are some examples:

Prelude Turtle> -- grep '^[[:digit:]]\+$' file.txt
Prelude Turtle> stdout (grep (plus digit) (input "file.txt"))
42
Prelude Turtle> -- grep '^[[:digit:]]\+\|Test$' file.txt
Prelude Turtle> stdout (grep (plus digit <|> "Test") (input "file.txt"))
Test
42

Note that turtle's grep subtly differs from the traditional grep command. The Pattern you provide must match the entire line. If you want to match the interior of a line, you can use the has utility:

Prelude Turtle> -- grep B file.txt
Prelude Turtle> stdout (grep (has "B") (input "file.txt"))
ABC

You can also use prefix or suffix to match the beginning or end of a string, respectively:

Prelude Turtle> -- grep '^A' file.txt
Prelude Turtle> stdout (grep (prefix "A") (input "file.txt"))
ABC
Prelude Turtle> -- grep 'C$' file.txt
Prelude Turtle> stdout (grep (suffix "C") (input "file.txt"))
ABC

sed also uses Patterns, too, and is more flexible than Unix sed:

Prelude Turtle> -- sed 'sCD/g' file.txt
Prelude Turtle> stdout (sed ("C" *> return "D") (input "file.txt"))
Test
ABD
42
Prelude Turtle> -- sed 's/[[:digit:]]/!/g' file.txt
Prelude Turtle> stdout (sed (digit *> return "!") (input "file.txt"))
Test
ABC
!!
Prelude Turtle> import qualified Data.Text as Text
Prelude Turtle> -- rev file.txt
Prelude Turtle> stdout (sed (fmap Text.reverse (plus dot)) (input "file.txt"))
tseT
CBA
24
Prelude Turtle>

You can also use Patterns by themselves to parse arbitrary text into more structured values:

Prelude Turtle> let pair = do x <- decimal; " "; y <- decimal; return (x, y)
Prelude Turtle> :type pair
pair :: Pattern (Integer, Integer)
Prelude Turtle> match pair "123 456"
[(123,456)]
Prelude Turtle> data Pet = Cat | Dog deriving (Show)
Prelude Turtle> let pet = ("cat" *> return Cat) <|> ("dog" *> return Dog) :: Pattern Pet
Prelude Turtle> match pet "dog"
[Dog]
Prelude Turtle> match (pet `sepBy` ",") "cat,dog,cat"
[[Cat,Dog,Cat]]

See the Turtle.Pattern module for more details if you are interested in writing more complex Patterns.

Sometimes you may want to acquire resources and ensure they get released correctly if there are any exceptions. You can use Managed resources to acquire things safely within a Shell.

You can think of a Managed resource as some resource that needs to be acquired and then released afterwards. Example: you want to create a temporary file and then guarantee it's deleted afterwards, even if the program fails with an exception.

Turtle.Prelude provides two Managed utilities for creating temporary directories or files:

mktempdir
    :: FilePath          -- Parent directory
    -> Text              -- Directory name template
    -> Managed FilePath  -- Temporary directory

mktemp
    :: FilePath                    -- Parent directory
    -> Text                        -- File name template
    -> Managed (FilePath, Handle)  -- Temporary file

You can acquire a Managed resource within a Shell with using:

using :: Managed a -> Shell a

... and here is an example of creating a temporary directory and file within a Shell:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

main = sh (do
    dir       <- using (mktempdir "/tmp" "turtle")
    (file, _) <- using (mktemp dir "turtle")
    liftIO (print file) )

When you run the above script it will print out the name of the temporary directory and file:

$ ./example.hs
FilePath "/tmp/turtle15976/turtle15976"

... and you can verify that they were deleted afterwards:

Turtle Prelude> view (find (has "turtle") "/tmp")
Turtle Prelude> -- No results

As an exercise, try inserting an exception and verifying that the temporary: file and directory are still cleaned up correctly:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

main = sh (do
    dir       <- using (mktempdir "/tmp" "turtle")
    (file, _) <- using (mktemp dir "turtle")
    liftIO (print file)
    liftIO (die "Urk!") )

To learn more about Managed resources, read the documentation in Control.Monad.Managed.

If you are sick of having to type liftIO everywhere, you can omit it. This is because all subroutines in turtle are overloaded using the MonadIO type class, like our original pwd command where we first encountered the the MonadIO type:

Prelude Turtle> :type pwd
pwd :: MonadIO io => io Turtle.FilePath

This means this command is overloaded to run in any context that implements the MonadIO interface, including:

• IO (obviously)
• Shell
• Managed

You can tell if a type constructor like Shell implements MonadIO by clicking the link to the type constructor and looking for the instance list. There you will see a list of instances like:

Monad Shell
Functor Shell
MonadPlus Shell
Applicative Shell
Alternative Shell
MonadIO Shell
MonadManaged Shell
...

These instances represent the overloaded functions associated with Shell and we can see from the list that Shell implements MonadIO so we can use pwd (or any other subroutine in this library) within a Shell.

However, not all subroutines in the Haskell ecosystem are overloaded in this way (such as print), so you will still occasionally need to wrap subroutines in liftIO.

All Managed operations are also overloaded in turtle, meaning that you can omit the using command. For example, we could change our last example to:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle

main = sh (do
    dir       <- mktempdir "/tmp" "turtle"
    (file, _) <- mktemp dir "turtle"
    liftIO (print file)
    die "Urk!" )

Any command that is generalized over the MonadManaged interface can run in the following contexts:

• Managed (obviously)
• Shell

The Turtle.Options module lets you easily parse command line arguments, using either flags or positional arguments.

For example, if you want to write a cp-like script that takes two positional arguments for the source and destination file, you can write:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

-- cp.hs

{-# LANGUAGE OverloadedStrings #-}

import Turtle
import Prelude hiding (FilePath)

parser :: Parser (FilePath, FilePath)
parser = (,) <$> argPath "src"  "The source file"
             <*> argPath "dest" "The destination file"

main = do
    (src, dest) <- options "A simple `cp` utility" parser
    cp src dest

If you run the script without any arguments, you will get an auto-generated usage output:

$ ./cp.hs
Usage: cp.hs SRC DEST

... and you can get a more descriptive output if you supply the --help flag:

$ ./cp.hs --help
A simple `cp` utility

Usage: cp.hs SRC DEST

Available options:
  -h,--help                Show this help text
  SRC                      The source file
  DEST                     The destination file

... and the script works as expected if you provide both arguments:

echo "Test" > file1.txt
$ ./cp.hs file1.txt file2.txt
cat file2.txt
Test

This works because argPath produces a Parser:

argPath :: ArgName -> Optional HelpMessage -> Parser FilePath

... and multiple Parsers can be combined into a single Parser using operations from the Applicative type class since the Parser type implements the Applicative interface:

instance Applicative Parser

You can also make any argument optional using the optional utility provided by Applicative:

optional :: Alternative f => f a -> f (Maybe a)

For example, we can change our program to make the destination argument optional, defaulting to stdout if the user does not provide a destination:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle
import Prelude hiding (FilePath)

parser :: Parser (FilePath, Maybe FilePath)
parser = (,) <$>           argPath "src"  "The source file"
             <*> optional (argPath "dest" "The destination file")

main = do
    (src, mDest) <- options "A simple `cp` utility" parser
    case mDest of
        Nothing   -> input src & stdout
        Just dest -> cp src dest

Now the auto-generated usage information correctly indicates that the second argument is optional:

$ ./cp.hs
Usage: cp.hs SRC [DEST]

$ ./cp.hs --help
A simple `cp` utility

Usage: cp.hs SRC [DEST]

Available options:
  -h,--help                Show this help text
  SRC                      The source file
  DEST                     The destination file

... and if we omit the argument the result goes to standard output:

$ ./cp.hs file1.txt
Test

We can use the optional utility because the Parser type also implements the Alternative interface:

instance Alternative Parser

We can also specify arguments on the command lines using flags instead of specifying them positionally. Let's change our example to specify the input and output using the --src and --dest flags, using -s and -d as short-hands for the flags:

#!/usr/bin/env stack
-- stack --resolver lts-10.2 script

{-# LANGUAGE OverloadedStrings #-}

import Turtle
import Prelude hiding (FilePath)

parser :: Parser (FilePath, FilePath)
parser = (,) <$> optPath "src"  's' "The source file"
             <*> optPath "dest" 'd' "The destination file"

main = do
    (src, dest) <- options "A simple `cp` utility" parser
    cp src dest

This now lets us specify the arguments in terms of flags:

$ ./cp
Usage: cp.hs (-s|--src SRC) (-d|--dest DEST)

$ ./cp --help
A simple `cp` utility

Usage: cp.hs (-s|--src SRC) (-d|--dest DEST)

Available options:
  -h,--help                Show this help text
  -s,--src SRC             The source file
  -d,--dest DEST           The destination file

$ ./cp --src file1.txt --dest file3.txt
$ cat file3.txt
Test

You can also provide subcommand functionality such as the following example which pretends to increase or decrease the system volume:

{-# LANGUAGE OverloadedStrings #-}

import Turtle

data Command = IncreaseVolume Int | DecreaseVolume Int deriving (Show)

parser :: Parser Command
parser
    =   fmap IncreaseVolume
            (subcommand "up" "Turn the volume up"
                (argInt "amount" "How much to increase the volume") )
    <|> fmap DecreaseVolume
            (subcommand "down" "Turn the volume down"
                (argInt "amount" "How much to decrease the volume") )

main = do
    x <- options "Volume adjuster" parser
    case x of
        IncreaseVolume n -> printf ("Increasing the volume by "%d%"\n") n
        DecreaseVolume n -> printf ("Decreasing the volume by "%d%"\n") n

This will provide `--help` output at both the top level and for each subcommand:

$ ./options --help
Volume adjuster

Usage: options (up | down)

Available options:
  -h,--help                Show this help text

  Available commands:
    up
    down

$ ./options up --help
Turn the volume up

Usage: options up AMOUNT

Available options:
  -h,--help                Show this help text
  AMOUNT                   How much to increase the volume

$ ./options up 10
Increasing the volume by 10

See the Turtle.Options module for more details and utilities related to parsing command line options. This module is built on top of the optparse-applicative library, which provides even more extensive functionality.

By this point you should be able to write basic shell scripts in Haskell. If you would like to learn more advanced tricks, take the time to read the documentation in these modules:

• Turtle.Prelude
• Turtle.Format
• Turtle.Pattern
• Turtle.Shell
• Control.Foldl
• Control.Monad.Managed

If you have more questions or need help learning the library, ask a question on Stack Overflow under the haskell-turtle tag. For bugs or feature requests, create an issue on Github at https://github.com/Gabriella439/turtle/issues

This library provides an extended suite of Unix-like utilities, but would still benefit from adding more utilities for better parity with the Unix ecosystem. Pull requests to add new utilities are highly welcome!

You can also turn turtle scripts into runnable scripts using Nix, like this:

#! /usr/bin/env nix-shell
#! nix-shell -i runghc --packages "ghc.withPackages (x: [ x.turtle ])"

{-# LANGUAGE OverloadedStrings #-}

import Turtle

main = echo "Hello, world!"

... or create an auto-reloading script like this:

#! /usr/bin/env nix-shell
#! nix-shell -i "ghcid -T main" --packages ghcid "ghc.withPackages (x: [ x.turtle ])"

{-# LANGUAGE OverloadedStrings #-}

import Turtle

main = echo "Hello, world!"

These are the most frequently asked questions from new users:

Question: My program hangs when I run a subprocess that reads from standard input. What do I do?

Answer: Make sure you compile your program with the -threaded flag

Question: How do I convert FilePath to Text?

Answer: Use (format fp)

Question: My program prints some extra output every time it starts. How do I remove it?

Answer: Compile your program and run the executable instead of interpreting the program.

Question: How do I transform a (Pattern a) into a (Pattern [a])?

Answer: Use many or some (both are from Control.Applicative and re-exported by Turtle)

Question: Why are star / plus not the same as many / some?

Answer: Because [Char] is a String, which is not the same thing as Text. String is deprecated in favor of Text in modern Haskell code, primarily for performance reasons and also because Text provides better support for Unicode.

Question: Some Haskell libraries still use String. How do I convert back and forth between String and Text?

Answer: Use Data.Text.pack and Data.Text.unpack

Question: What's the easiest way to fail with a descriptive error message if a subprocess command like proc/shell returns a non-zero exit code? code?

Answer: Use (procs cmd args input) or (proc cmd args input .||. die "Descriptive error message") (or shell / shells, respectively)

Question: How do I close a resource that I acquired?

Answer: Use runManaged, sh, or (<|>) (all resources acquired in the left stream will close before beginning the right stream). Alternatively, use with to acquire a resource for a limited scope.

Question: How do I use turtle to run another shell as a subprocess?

Answer: Use system in conjunction with the process library, like this:

Turtle.system (System.Process.proc "/bin/sh" []) empty