This is the place to start your practical exploration of sarge.

Installation and testing

sarge is a pure-Python library. You should be able to install it using:

pip install sarge

for installing sarge into a virtualenv or other directory where you have write permissions. On Posix platforms, you may need to invoke using sudo if you need to install sarge in a protected location such as your system Python’s site-packages directory.

A full test suite is included with sarge. To run it, you’ll need to unpack a source tarball and run python setup.py test in the top-level directory of the unpack location. You can of course also run python setup.py install to install from the source tarball (perhaps invoking with sudo if you need to install to a protected location).

Common usage patterns

In the simplest cases, sarge doesn’t provide any major advantage over subprocess:

>>> from sarge import run
>>> run('echo "Hello, world!"')
Hello, world!
<sarge.Pipeline object at 0x1057110>

The echo command got run, as expected, and printed its output on the console. In addition, a Pipeline object got returned. Don’t worry too much about what this is for now – it’s more useful when more complex combinations of commands are run.

By comparison, the analogous case with subprocess would be:

>>> from subprocess import call
>>> call('echo "Hello, world!"'.split())
"Hello, world!"

We had to call split() on the command (or we could have passed shell=True), and as well as running the command, the call() method returned the exit code of the subprocess. To get the same effect with sarge you have to do:

>>> from sarge import run
>>> run('echo "Hello, world!"').returncode
Hello, world!

If that’s as simple as you want to get, then of course you don’t need sarge. Let’s look at more demanding uses next.

Finding commands under Windows

In versions 0.1.1 and earlier, sarge, like subprocess, did not do anything special to find the actual executable to run – it was expected to be found in the current directory or the path. Specifically, PATHEXT was not supported: where you might type yada in a command shell and have it run python yada.py because .py is in the PATHEXT environment variable and Python is registered to handle files with that extension, neither subprocess (with shell=False) nor sarge did this. You needed to specify the executable name explicitly in the command passed to sarge.

In 0.1.2 and later versions, sarge has improved command-line handling. The “which” functionality has been backported from Python 3.3, which takes care of using PATHEXT to resolve a command yada as c:\Tools\yada.py where c:\Tools is on the PATH and yada.py is in there. In addition, sarge queries the registry to see which programs are associated with the extension, and updates the command line accordingly. Thus, a command line foo bar passed to sarge may actually result in c:\Windows\py.exe c:\Tools\foo.py bar being passed to subprocess (assuming the Python Launcher for Windows, py.exe, is associated with .py files).

This new functionality is not limited to Python scripts - it should work for any extensions which are in PATHEXT and have an ftype/assoc binding them to an executable through shell, open and command subkeys in the registry, and where the command line is of the form "<path_to_executable>" "%1" %* (this is the standard form used by several languages).

Chaining commands

It’s easy to chain commands together with sarge. For example:

>>> run('echo "Hello,"; echo "world!"')
<sarge.Pipeline object at 0x247ed50>

whereas this would have been more involved if you were just using subprocess:

>>> call('echo "Hello,"'.split()); call('echo "world!"'.split())

You get two return codes, one for each command. The same information is available from sarge, in one place – the Pipeline instance that’s returned from a run() call:

>>> run('echo "Hello,"; echo "world!"').returncodes
[0, 0]

The returncodes property of a Pipeline instance returns a list of the return codes of all the commands that were run, whereas the returncode property just returns the last element of this list. The Pipeline class defines a number of useful properties - see the reference for full details.

Handling user input safely

By default, sarge does not run commands via the shell. This means that wildcard characters in user input do not have potentially dangerous consequences:

>>> run('ls *.py')
ls: cannot access *.py: No such file or directory
<sarge.Pipeline object at 0x20f3dd0>

This behaviour helps to avoid shell injection attacks.

There might be circumstances where you need to use shell=True, in which case you should consider formatting your commands with placeholders and quoting any variable parts that you get from external sources (such as user input). Which brings us on to ...

Formatting commands with placeholders for safe usage

If you need to merge commands with external inputs (e.g. user inputs) and you want to prevent shell injection attacks, you can use the shell_format() function. This takes a format string, positional and keyword arguments and uses the new formatting (str.format()) to produce the result:

>>> from sarge import shell_format
>>> shell_format('ls {0}', '*.py')
"ls '*.py'"

Note how the potentially unsafe input has been quoted. With a safe input, no quoting is done:

>>> shell_format('ls {0}', 'test.py')
'ls test.py'

If you really want to prevent quoting, even for potentially unsafe inputs, just use the s conversion:

>>> shell_format('ls {0!s}', '*.py')
'ls *.py'

There is also a shell_quote() function which quotes potentially unsafe input:

>>> from sarge import shell_quote
>>> shell_quote('abc')
>>> shell_quote('ab?')
>>> shell_quote('"ab?"')
>>> shell_quote("'ab?'")

This function is used internally by shell_format(), so you shouldn’t need to call it directly except in unusual cases.

Passing input data to commands

You can pass input to a command pipeline using the input keyword parameter to run():

>>> from sarge import run
>>> p = run('cat|cat', input='foo')

Here’s how the value passed as input is processed:

  • Text is encoded to bytes using UTF-8, which is then wrapped in a BytesIO object.
  • Bytes are wrapped in a BytesIO object.
  • Starting with 0.1.2, if you pass an object with a fileno attribute, that will be called as a method and the resulting value will be passed to the subprocess layer. This would normally be a readable file descriptor.
  • Other values (such as integers representing OS-level file descriptors, or special values like subprocess.PIPE) are passed to the subprocess layer as-is.

If the result of the above process is a BytesIO instance (or if you passed in a BytesIO instance), then sarge will spin up an internal thread to write the data to the child process when it is spawned. The reason for a separate thread is that if the child process consumes data slowly, or the size of data is large, then the calling thread would block for potentially long periods of time.

Passing input data to commands dynamically

Sometimes, you may want to pass quite a lot of data to a child process which is not conveniently available as a string, byte-string or a file, but which is generated in the parent process (the one using sarge) by some other means. Starting with 0.1.2, sarge facilitates this by supporting objects with fileno() attributes as described above, and includes a Feeder class which has a suitable fileno() implementation.

Creating and using a feeder is simple:

import sys
from sarge import Feeder, run

feeder = Feeder()
run([sys.executable, 'echoer.py'], input=feeder, async=True)

After this, you can feed data to the child process’ stdin by calling the feed() method of the Feeder instance:


If you pass in text, it will be encoded to bytes using UTF-8.

Once you’ve finished with the feeder, you can close it:


Depending on how quickly the child process consumes data, the thread calling feed() might block on I/O. If this is a problem, you can spawn a separate thread which does the feeding.

Here’s a complete working example:

import os
import subprocess
import sys
import time

import sarge

    text_type = unicode
except NameError:
    text_type = str

def main(args=None):
    feeder = sarge.Feeder()
    p = sarge.run([sys.executable, 'echoer.py'], input=feeder, async=True)
        lines = ('hello', 'goodbye')
        gen = iter(lines)
        while p.commands[0].returncode is None:
                data = next(gen)
            except StopIteration:
            feeder.feed(data + '\n')
            time.sleep(0.05)    # wait for child to return echo

if __name__ == '__main__':
        rc = main()
    except Exception as e:
        rc = 9

In the above example, the echoer.py script (included in the sarge source distribution, as it’s part of the test suite) just reads lines from its stdin, duplicates and prints to its stdout. Since we passed in the strings hello and goodbye, the output from the script should be:

hello hello
goodbye goodbye

Chaining commands conditionally

You can use && and || to chain commands conditionally using short-circuit Boolean semantics. For example:

>>> from sarge import run
>>> run('false && echo foo')
<sarge.Pipeline object at 0xb8dd50>

Here, echo foo wasn’t called, because the false command evaluates to False in the shell sense (by returning an exit code other than zero). Conversely:

>>> run('false || echo foo')
<sarge.Pipeline object at 0xa11d50>

Here, foo is output because we used the || condition; because the left- hand operand evaluates to False, the right-hand operand is evaluated (i.e. run, in this context). Similarly, using the true command:

>>> run('true && echo foo')
<sarge.Pipeline object at 0xb8dd50>
>>> run('true || echo foo')
<sarge.Pipeline object at 0xa11d50>

Creating command pipelines

It’s just as easy to construct command pipelines:

>>> run('echo foo | cat')
<sarge.Pipeline object at 0xb8dd50>
>>> run('echo foo; echo bar | cat')
<sarge.Pipeline object at 0xa96c50>

Using redirection

You can also use redirection to files as you might expect. For example:

>>> run('echo foo | cat > /tmp/junk')
<sarge.Pipeline object at 0x24b3190>
^D (to exit Python)
$ cat /tmp/junk

You can use >, >>, 2>, 2>> which all work as on Posix systems. However, you can’t use < or <<.

To send things to the bit-bucket in a cross-platform way, you can do something like:

>>> run('echo foo | cat > %s' % os.devnull)
<sarge.Pipeline object at 0x2765b10>

Capturing stdout and stderr from commands

To capture output for commands, just pass a Capture instance for the relevant stream:

>>> from sarge import run, Capture
>>> p = run('echo foo; echo bar | cat', stdout=Capture())
>>> p.stdout.text

The Capture instance acts like a stream you can read from: it has read(), readline() and readlines() methods which you can call just like on any file-like object, except that they offer additional options through block and timeout keyword parameters.

As in the above example, you can use the bytes or text property of a Capture instance to read all the bytes or text captured. The latter just decodes the former using UTF-8 (the default encoding isn’t used, because on Python 2.x, the default encoding isn’t UTF-8 – it’s ASCII).

There are some convenience functions – capture_stdout(), capture_stderr() and capture_both() – which work just like run() but capture the relevant streams to Capture instances, which can be accessed using the appropriate attribute on the Pipeline instance returned from the functions.

There are more convenience functions, get_stdout(), get_stderr() and get_both(), which work just like capture_stdout(), capture_stderr() and capture_both() respectively, but return the captured text. For example:

>>> from sarge import get_stdout
>>> get_stdout('echo foo; echo bar')

New in version 0.1.1: The get_stdout(), get_stderr() and get_both() functions were added.

A Capture instance can capture output from one or more sub-process streams, and will create a thread for each such stream so that it can read all sub-process output without causing the sub-processes to block on their output I/O. However, if you use a Capture, you should be prepared either to consume what it’s read from the sub-processes, or else be prepared for it all to be buffered in memory (which may be problematic if the sub-processes generate a lot of output).

Iterating over captures

You can iterate over Capture instances. By default you will get successive lines from the captured data, as bytes; if you want text, you can wrap with io.TextIOWrapper. Here’s an example using Python 3.2:

>>> from sarge import capture_stdout
>>> p = capture_stdout('echo foo; echo bar')
>>> for line in p.stdout: print(repr(line))
>>> p = capture_stdout('echo bar; echo baz')
>>> from io import TextIOWrapper
>>> for line in TextIOWrapper(p.stdout): print(repr(line))

This works the same way in Python 2.x. Using Python 2.7:

>>> from sarge import capture_stdout
>>> p = capture_stdout('echo foo; echo bar')
>>> for line in p.stdout: print(repr(line))
>>> p = capture_stdout('echo bar; echo baz')
>>> from io import TextIOWrapper
>>> for line in TextIOWrapper(p.stdout): print(repr(line))

Interacting with child processes

Sometimes you need to interact with a child process in an interactive manner. To illustrate how to do this, consider the following simple program, named receiver, which will be used as the child process:

#!/usr/bin/env python
import sys

def main(args=None):
    while True:
        user_input = sys.stdin.readline().strip()
        if not user_input:
        s = 'Hi, %s!\n' % user_input
        sys.stdout.flush() # need this when run as a subprocess

if __name__ == '__main__':

This just reads lines from the input and echoes them back as a greeting. If we run it interactively:

$ ./receiver
Hi, Fred!
Hi, Jim!
Hi, Sheila!

The program exits on seeing an empty line.

We can now show how to interact with this program from a parent process:

>>> from sarge import Command, Capture
>>> from subprocess import PIPE
>>> p = Command('./receiver', stdout=Capture(buffer_size=1))
>>> p.run(input=PIPE, async=True)
>>> p.stdin.write('Fred\n')
>>> p.stdout.readline()
'Hi, Fred!\n'
>>> p.stdin.write('Jim\n')
>>> p.stdout.readline()
'Hi, Jim!\n'
>>> p.stdin.write('Sheila\n')
>>> p.stdout.readline()
'Hi, Sheila!\n'
>>> p.stdin.write('\n')
>>> p.stdout.readline()
>>> p.returncode
>>> p.wait()

The p.returncode didn’t print anything, indicating that the return code was None. This means that although the child process has exited, it’s still a zombie because we haven’t “reaped” it by making a call to wait(). Once that’s done, the zombie disappears and we get the return code.

Buffering issues

From the point of view of buffering, note that two elements are needed for the above example to work:

  • We specify buffer_size=1 in the Capture constructor. Without this, data would only be read into the Capture’s queue after an I/O completes – which would depend on how many bytes the Capture reads at a time. You can also pass a buffer_size=-1 to indicate that you want to use line- buffering, i.e. read a line at a time from the child process. (This may only work as expected if the child process flushes its outbut buffers after every line.)
  • We make a flush call in the receiver script, to ensure that the pipe is flushed to the capture queue. You could avoid the flush call in the above example if you used python -u receiver as the command (which runs the script unbuffered).

This example illustrates that in order for this sort of interaction to work, you need cooperation from the child process. If the child process has large output buffers and doesn’t flush them, you could be kept waiting for input until the buffers fill up or a flush occurs.

If a third party package you’re trying to interact with gives you buffering problems, you may or may not have luck (on Posix, at least) using the unbuffer utility from the expect-dev package (do a Web search to find it). This invokes a program directing its output to a pseudo-tty device which gives line buffering behaviour. This doesn’t always work, though :-(

Looking for specific patterns in child process output

You can look for specific patterns in the output of a child process, by using the expect() method of the Capture class. This takes a string, bytestring or regular expression pattern object and a timeout, and either returns a regular expression match object (if a match was found in the specified timeout) or None (if no match was found in the specified timeout). If you pass in a bytestring, it will be converted to a regular expression pattern. If you pass in text, it will be encoded to bytes using the utf-8 codec and then to a regular expression pattern. This pattern will be used to look for a match (using search). If you pass in a regular expression pattern, make sure it is meant for bytes rather than text (to avoid TypeError on Python 3.x). You may also find it useful to specify re.MULTILINE in the pattern flags, so that you can match using ^ and $ at line boundaries. Note that on Windows, you may need to use \r?$ to match ends of lines, as $ matches Unix newlines (LF) and not Windows newlines (CRLF).

New in version 0.1.1: The expect method was added.

To illustrate usage of Capture.expect(), consider the program lister.py (which is provided as part of the source distribution, as it’s used in the tests). This prints line 1, line 2 etc. indefinitely with a configurable delay, flushing its output stream after each line. We can capture the output from a run of lister.py, ensuring that we use line-buffering in the parent process:

>>> from sarge import Capture, run
>>> c = Capture(buffer_size=-1)     # line-buffering
>>> p = run('python lister.py -d 0.01', async=True, stdout=c)
>>> m = c.expect('^line 1$')
>>> m.span()
(0, 6)
>>> m = c.expect('^line 5$')
>>> m.span()
(28, 34)
>>> m = c.expect('^line 1.*$')
>>> m.span()
(63, 70)
>>> c.close(True)           # close immediately, discard any unread input
>>> p.commands[0].kill()    # kill the subprocess
>>> c.bytes[63:70]
'line 10'
>>> m = c.expect(r'^line 1\d\d$')
>>> m.span()
(783, 791)
>>> c.bytes[783:791]
'line 100'

Displaying progress as a child process runs

You can display progress as a child process runs, assuming that its output allows you to track that progress. Consider the following script, progress.py:

import optparse # because of 2.6 support
import sys
import threading
import time

from sarge import capture_stdout

def progress(capture, options):
    lines_seen = 0
    messages = {
        'line 25\n': 'Getting going ...\n',
        'line 50\n': 'Well on the way ...\n',
        'line 75\n': 'Almost there ...\n',
    while True:
        s = capture.readline()
        if not s and lines_seen:
        if options.dots:
            msg = messages.get(s)
            if msg:
        lines_seen += 1
    if options.dots:
    sys.stderr.write('Done - %d lines seen.\n' % lines_seen)

def main():
    parser = optparse.OptionParser()
    parser.add_option('-n', '--no-dots', dest='dots', default=True,
                      action='store_false', help='Show dots for progress')
    options, args = parser.parse_args()
    p = capture_stdout('python lister.py -d 0.1 -c 100', async=True)
    t = threading.Thread(target=progress, args=(p.stdout, options))
    while(p.returncodes[0] is None):
        # We could do other useful work here. If we have no useful
        # work to do here, we can call readline() and process it
        # directly in this loop, instead of creating a thread to do it in.

if __name__ == '__main__':

When this is run without the --no-dots argument, you should see the following:

$ python progress.py
....................................................... (100 dots printed)
Done - 100 lines seen.

If run with the --no-dots argument, you should see:

$ python progress.py --no-dots
Getting going ...
Well on the way ...
Almost there ...
Done - 100 lines seen.

with short pauses between the output lines.

Direct terminal usage

Some programs don’t work through their stdin/stdout/stderr streams, instead opting to work directly with their controlling terminal. In such cases, you can’t work with these programs using sarge; you need to use a pseudo-terminal approach, such as is provided by (for example) pexpect. Sarge works within the limits of the subprocess module, which means sticking to stdin, stdout and stderr as ordinary streams or pipes (but not pseudo-terminals).

Examples of programs which work directly through their controlling terminal are ftp and ssh - the password prompts for these programs are generally always printed to the controlling terminal rather than stdout or stderr.


In the subprocess.Popen constructor, the env keyword argument, if supplied, is expected to be the complete environment passed to the child process. This can lead to problems on Windows, where if you don’t pass the SYSTEMROOT environment variable, things can break. With sarge, it’s assumed that anything you pass in env is added to the contents of os.environ. This is almost always what you want – after all, in a Posix shell, the environment is generally inherited with certain additions for a specific command invocation.


On Python 2.x on Windows, environment keys and values must be of type str - Unicode values will cause a TypeError. Be careful of this if you use from __future__ import unicode_literals. For example, the test harness for sarge uses Unicode literals on 2.x, necessitating the use of different logic for 2.x and 3.x:

if PY3:
    env = {'FOO': 'BAR'}
    # Python 2.x wants native strings, at least on Windows
    env = { b'FOO': b'BAR' }

Working directory and other options

You can set the working directory for a Command or Pipeline using the cwd keyword argument to the constructor, which is passed through to the subprocess when it’s created. Likewise, you can use the other keyword arguments which are accepted by the subprocess.Popen constructor.

Avoid using the stdin keyword argument – instead, use the input keyword argument to the Command.run() and Pipeline.run() methods, or the run(), capture_stdout(), capture_stderr(), and capture_both() functions. The input keyword makes it easier for you to pass literal text or byte data.

Unicode and bytes

All data between your process and sub-processes is communicated as bytes. Any text passed as input to run() or a run() method will be converted to bytes using UTF-8 (the default encoding isn’t used, because on Python 2.x, the default encoding isn’t UTF-8 – it’s ASCII).

As sarge requires Python 2.6 or later, you can use from __future__ import unicode_literals and byte literals like b'foo' so that your code looks and behaves the same under Python 2.x and Python 3.x. (See the note on using native string keys and values in Environments.)

As mentioned above, Capture instances return bytes, but you can wrap with io.TextIOWrapper if you want text.

Use as context managers

The Capture and Pipeline classes can be used as context managers:

>>> with Capture() as out:
...     with Pipeline('cat; echo bar | cat', stdout=out) as p:
...         p.run(input='foo\n')
<sarge.Pipeline object at 0x7f3320e94310>
>>> out.read().split()
['foo', 'bar']

Synchronous and asynchronous execution of commands

By default. commands passed to run() run synchronously, i.e. all commands run to completion before the call returns. However, you can pass async=True to run, in which case the call returns a Pipeline instance before all the commands in it have run. You will need to call wait() or close() on this instance when you are ready to synchronise with it; this is needed so that the sub processes can be properly disposed of (otherwise, you will leave zombie processes hanging around, which show up, for example, as <defunct> on Linux systems when you run ps -ef). Here’s an example:

>>> p = run('echo foo|cat|cat|cat|cat', async=True)
>>> foo

Here, foo is printed to the terminal by the last cat command, but all the sub-processes are zombies. (The run function returned immediately, so the interpreter got to issue the >>>` prompt *before* the ``foo output was printed.)

In another terminal, you can see the zombies:

$ ps -ef | grep defunct | grep -v grep
vinay     4219  4217  0 19:27 pts/0    00:00:00 [echo] <defunct>
vinay     4220  4217  0 19:27 pts/0    00:00:00 [cat] <defunct>
vinay     4221  4217  0 19:27 pts/0    00:00:00 [cat] <defunct>
vinay     4222  4217  0 19:27 pts/0    00:00:00 [cat] <defunct>
vinay     4223  4217  0 19:27 pts/0    00:00:00 [cat] <defunct>

Now back in the interactive Python session, we call close() on the pipeline:

>>> p.close()

and now, in the other terminal, look for defunct processes again:

$ ps -ef | grep defunct | grep -v grep

No zombies found :-)

About threading and forking on Posix

If you run commands asynchronously by using & in a command pipeline, then a thread is spawned to run each such command asynchronously. Remember that thread scheduling behaviour can be unexpected – things may not always run in the order you expect. For example, the command line:

echo foo & echo bar & echo baz

should run all of the echo commands concurrently as far as possible, but you can’t be sure of the exact sequence in which these commands complete – it may vary from machine to machine and even from one run to the next. This has nothing to do with sarge – there are no guarantees with just plain Bash, either.

On Posix, subprocess uses os.fork() to create the child process, and you may see dire warnings on the Internet about mixing threads, processes and fork(). It is a heady mix, to be sure: you need to understand what’s going on in order to avoid nasty surprises. If you run into any such, it may be hard to get help because others can’t reproduce the problems. However, that’s no reason to shy away from providing the functionality altogether. Such issues do not occur on Windows, for example: because Windows doesn’t have a fork() system call, child processes are created in a different way which doesn’t give rise to the issues which sometimes crop up in a Posix environment.

For an exposition of the sort of things which might bite you if you are using locks, threading and fork() on Posix, see this post.

Other resources on this topic:

Please report any problems you find in this area (or any other) either via the mailing list or the issue tracker.

Next steps

You might find it helpful to look at information about how sarge works internally – Under the hood – or peruse the API Reference.

Comments powered by Disqus