expect: Scripts for Controlling Interactive Processes

Don Libes
National Institute of Standards and Technology [email protected]

ABSTRACT: Contemporary shells provide minimal control (starting, stopping, etc) over programs, leaving interaction up to users. This means that you cannot run some programs non-interactively, such as passwd. Some programs can be run noninteractively but only with a loss of flexibility, such as fsck. This is where the toolbuilding philosophy of UNIX begins to break down. expect crosses this line, solving a number of long-standing problems in the UNIX environment. expect uses Tcl as a language core. In addition, expect can use any UNIX program whether or not it is interactive. The result is a classic example of a little language synergistically generating large power when combined with the rest of the UNIX workbench. Previous papers have described the implementation of expect and compared it to other tools. This paper concentrates on the language, primarily by presenting a variety of scripts. Several scripts demonstrate brand-new features of expect. Keywords: expect; interaction; POSIX; programmed dialogue; shell; Tcl; UNIX

Reprint of Computing Systems , Vol. 4, No. 2, University of California Press, Berkeley, CA, November 1991.

1. Introduction
fsck, the UNIX le system check program, can be run from a shell script only with the y or n options. The manual [1] denes the y option as follows: Assume a yes response to all questions asked by fsck; this should be used with extreme caution, as it is a free license to continue, even after severe problems are encountered. The n option is safer, but almost uselessly so. This kind of interface is inexcusably bad, yet many programs have the same style. ftp, a le transfer program, has an option that disables interactive prompting so that it can be run from a script. But it provides no way to take alternative action should an error occur. expect is a tool for controlling interactive programs. It solves the fsck problem, providing all the interactive functionality non-interactively. expect is not specically designed for fsck, and can handle ftps errors as well. The problems with fsck and ftp illustrate a major limitation in the user interface offered by shells such as sh, csh, and others (which will generically be referred to as the shell in the rest of the paper). The shell does not provide a way of reading output and writing input from a program. This means the shell can run fsck but only by missing out on some of its useful features. Some programs cannot be run at all. For example, passwd cannot be run without a user interactively supplying the input. Similar programs that cannot be automated in a shell script are telnet, crypt, su, rlogin, etc. A large number of application programs are written with the same fault of demanding user input. expect was designed specically to interact with interactive programs. An expect programmer can write a script describing the dialogue. Then the expect program can run the interactive program non-interactively. Writing scripts for interactive programs is as simple as writing scripts for non-interactive programs. expect can also be used to automate parts of a dialogue, since control can be passed from the script to the keyboard and vice versa.

2. A brief overview of expect

The implementation and philosophy of expect is described at length by Libes [2]. Briey, scripts are written in an interpreted language. (A library is available for C and C++ programmers but it will not be further discussed in this paper.) Commands are provided to create interactive processes and to read and write their output and input. expect1 is named after the specic command which waits for output from a program. The language of expect is based on Tcl. Tcl is actually a subroutine library, which becomes embedded into an application and provides language services. The resulting language looks very much like a typical shell language. There are commands to set variables (set), control ow (if, for,

1. For readability, times roman bold is used for display of le or program names, helvetica for keyword or other language elements, and courier for literal strings or code fragments.

Don Libes

continue, etc), and perform the usual math and string operations. Of course, UNIX programs can be called (exec). All of these facilities are available to any Tcl application. Tcl is completely de-

scribed by Ousterhout [3][4]. expect is built on top of Tcl and provides additional commands. The spawn command invokes a UNIX program for interactive use. send sends strings to a process. expect waits for strings from a process. expect supports regular expressions and can wait for multiple strings at the same time, executing a different action for each string. expect also understands exceptional conditions such as timeout and end-of-le. The expect command is styled after Tcls case command which matches a string against a number of other strings. (Whenever possible, new commands were modeled after existing Tcl commands so that the language remained a coherent set of tools.) The following denition of expect is paraphrased from the manual page [5]:
expect patlist1 action1 patlist2 action2 . . .

waits until one of the patterns matches the output of the current process, a specied time period has passed, or an end-of-le is found. If the nal action is null, it may be omitted. Each patlist consists of a single pattern or list of patterns. If a pattern is matched, the corresponding action is executed. The result of the action is returned from expect. The exact string matched (or read but unmatched, if a timeout occurred) is stored in the variable expect_match. If patlist is eof or timeout, the corresponding action is executed upon end-of-le or timeout, respectively. The default timeout period is 10 seconds but may, for example, be set to 30 by the command set timeout 30. The following fragment is from a script that involves a login. abort is a procedure dened elsewhere in the script, while the other actions use Tcl primitives similar to their C namesakes. expect "*welcome*" break \ "*busy*" {print busy; continue} \ "*failed*" abort \ timeout abort Patterns are the usual C-shell-style regular expressions. Patterns must match the entire output of the current process since the previous expect or interact (hence the reason most are surrounded by the * wildcard). However, more than 2000 bytes of output can force earlier bytes to be forgotten. This may be changed by setting the variable match_max.
expect actually demonstrates the best and worst of expect. In particular, its exibility comes at

the price of an occasionally confusing syntax. The pattern-lists can contain multiple patterns except for keyword patterns (e.g., eof, timeout) which must appear by themselves. This provides a guaranteed way of distinguishing them. However, breaking up the lists requires a second scan, which can interpret \r and \n as whitespace if not correctly quoted. This is exacerbated by Tcl providing two forms of string quoting: braces and double quotes. (If unambiguous, Tcl does not require strings to be quoted at all.) There is a separate section in the expect manual page to explain this complexity. Fortunately, a healthy set of examples seems to have held back complaints. Nonetheless, this aspect will be probably be revisited in a future release. For readability in this paper, scripts are presented as if double quotes sufced.

expect: Scripts for Controlling Interactive Processes

Characters can be individually quoted with a backslash. Backslashes are also used to continue statements, which otherwise are terminated at the end of a line. This is inherent to Tcl. Tcl also continues scanning when there is an open brace or double-quote. In addition, semicolons can be used to separate multiple statements on a single line. This sounds confusing, but is typical of interpreters (e.g., /bin/sh). Nonetheless, it is one of the less elegant aspects of Tcl.

3. callback
It is surprising how little scripting is necessary to produce something useful. Below is a script that dials a phone. It is used to reverse the charges so that long-distance phone calls are charged to the computer. It is invoked as expect callback.exp 12016442332 where the script is named callback.exp and +1 (201) 644-2332 is the phone number to be dialed. # first give the user some time to logout exec sleep 4 spawn tip modem expect "*connected*" send "ATD[index $argv 1]\r" # modem takes a while to connect set timeout 60 expect "*CONNECT*" The rst line is a comment. The second illustrates how a UNIX command with no interaction can be called. sleep 4 will cause the program to block for four seconds, giving the user a chance to logout since the modem will presumably call back to the same phone number that the user is already using. The next line starts tip using spawn so that tips output can be read by expect and its input written by send. Once tip says it is connected, the modem is told to dial the number. (The modem is assumed to be Hayes compatible, but it is easy to expand the script to handle others.) No matter what happens, expect terminates. If the call fails, it is possible for expect to retry, but that is not the point here. If the call succeeds, getty will detect DTR on the line after expect exits, and prompt the user with login:. (Actual scripts usually do more error checking.) This script illustrates the use of command-line parameters, made available to the script as a list named argv (in the same style as the C language). In this case, element 1 is the phone number. The brackets cause the enclosed text to be evaluated as a command, and the result is substituted for the original text. This is similar to the way backquotes work in csh. This script replaced a 60K program (written in C) that did the same thing.

4. passwd & conformance testing

Earlier, passwd was mentioned as a program that cannot be run without user interaction. passwd ignores I/O redirection and cannot be embedded in a pipeline so that input comes from another program or le. It insists on performing all I/O directly with a real user. passwd was designed

Don Libes

this way for security reasons, but the result is that there is no way to test passwd non-interactively. It is ironic that a program so critical to system security has no way of being reliably tested. passwd takes a username as an argument, and interactively prompts for a password. The following expect script takes a username and password as arguments, and can be run non-interactively: spawn passwd [index $argv 1] set password [index $argv 2] expect "*password:" send "$password\r" expect "*password:" send "$password\r" expect eof The rst line starts the passwd program, with the username passed as an argument. The next line saves the password in a variable for convenience. Like the shell, variables do not have to be declared in advance. In the third line, expect looks for the pattern password:. The asterisk allows it to match other data in the input, and is a useful shortcut to avoid specifying everything in detail. There is no action specied, so expect just waits until the pattern is found before continuing. After receiving the prompt, the next line sends a password to the current process. The \r indicates a carriage-return. (All the usual C conventions are supported.) There are two expectsend sequences because passwd asks the password to be typed twice as a spelling verication. There is no point to this in a non-interactive passwd, but the script has to do this because passwd assumes it is interacting with a human that does not type consistently. Lastly, the line expect eof searches for the end-of-le in the output of passwd and demonstrates the use of keyword patterns. Another such pattern is timeout, used to denote the failure of any pattern to match in a given amount of time. Here, eof is necessary only because passwd is carefully written to check that all of its I/O succeeds, including the nal newline produced after the password has been entered a second time. This script is sufcient to show the basic interaction of the passwd command. A more complete script would verify other behaviors. For example, the following script checks several other aspects of the passwd program. Complete prompts are checked. Correct handling of garbage input is checked. Process death, unusually slow response, or any other unexpected behavior is also trapped. spawn passwd [index $argv 1] expect eof timeout "*No such user.*" "*New password:" send "[index $argv 2]\r" expect eof timeout "*Password too long*" "*Password too short*" {exit 1} {exit 2} {exit 3} \ \ \

{exit {exit {exit {exit

4} 2} 5} 5}

\ \ \ \ 5

expect: Scripts for Controlling Interactive Processes

"*Retype new password:" send "[index $argv 3]\r" expect timeout "*Mismatch*" "*Password unchanged*}" "\r\n" expect timeout "*" eof

{exit 2} {exit 6} {exit 7} {exit 2} {exit 6}

\ \ \ \ \

This script exits with a numeric indication of what happened. 0 indicates passwd ran normally, 1 that it died unexpectedly, 2 that it locked up, and so on. Numbers are used for simplicity expect could just as easily pass back strings, including any messages from the spawned program itself. Indeed, it is typical to save the entire interaction to a le, deleting it only if the command under test behaves as expected. Otherwise the log is available for further examination. This passwd testing script is designed to be driven by another script. This second script reads a le of arguments and expected results. For each set, it calls the rst script and then compares the results to the expected results. (Since this task is non-interactive, a plain old shell can be used to interpret this second script.) For example, a data le for passwd could look like this: passwd.exp passwd.exp passwd.exp passwd.exp passwd.exp passwd.exp 3 0 5 5 6 4 bogus fred fred fred fred fred abledabl abcdefghijklm abc foobar ^C abledabl bar -

The rst eld names the regression script to be run. The second eld is the exit value that should match the result of the script. The third eld is the username. The fourth and fth elds are the passwords to be entered when prompted. The hyphen is just a placeholder for values that will never be read. In the rst test, bogus is a username that is invalid, to which passwd will respond No such user. expect will exit the script with a value of 3, which also appears as the second element in the rst line of the regression suite data le. In the last test, a control-C is actually sent to the program to verify that it aborts gracefully. In this way, expect can be used for testing and debugging interactive software, such as required by IEEE POSIX 1003.2 (Shells and Tools) conformance testing. This is described in more detail by Libes [6].

5. rogue & pseudo-terminals

UNIX users are familiar with processes connected to other processes by pipes (e.g. a shell pipeline). expect uses ptys (pseudo-terminals) to connect spawned processes. Ptys provide terminal semantics so that programs think they are performing I/O with a real terminal.

Don Libes

As an example, the BSD adventure game rogue runs in raw mode, and assumes a characteraddressable terminal exists at the other end of the connection. expect can actually be programmed to play rogue using the human interface that comes with it. rogue is an adventure game which presents you with a player that has various physical attributes such as a strength rating. Most of the time, the strength is 16, but every so often maybe one out of 20 games you get an unusually good strength of 18. A lot of rogue players know this, but no one in their right mind restarts the game 20 times to nd those really good congurations. The following script does it for you. for {} {1} {} { spawn rogue expect "*Str: 18*" break \ "*Str: 16*" close wait } interact The rst line is a for loop, with the same control arguments as in C. rogue is started, and then the strength checked to see if it is 18 or 16. If it is 16, the dialogue is terminated via close and wait (which respectively closes the connection to the pty and waits for the process to exit). rogue reads an end-of-le and goes away, after which the loop is restarted, creating a new game of rogue to test. When a strength of 18 is found, control breaks out of the loop and drops down to the last line of the script. interact passes control to the user so that they can play this particular game. Imagine running this script. What you will actually see is 20 or 30 initial congurations y across your screen in less than a second, nally stopping with a great game for you to play. The only way to play rogue better is under the debugger! It is important to realize that rogue is a graphics program which uses Curses. expect programmers must understand that Curses does not necessarily create screens in an intuitive manner. Fortunately, it is not a problem in this example. A future enhancement to expect may include a builtin terminal emulator in order to support the understanding of character graphics regions.

6. ftp
The rst script actually written with expect did not print out hello world. Instead, it did something much more useful. It ran ftp without user interaction. ftp is a program which performs le transfer over TCP/IP networks such as the Internet. The ubiquitous implementation requires the user to provide input for all but the most simple uses. The script below retrieves a le from a host using anonymous ftp. The hostname is the rst argument to the script. The lename is the second argument. spawn ftp [index $argv 1] expect "*Name*"
expect: Scripts for Controlling Interactive Processes

send "anonymous\r" expect "*Password:*" send [exec whoami] expect "*ok*ftp>*" send "get [index $argv 2]\r" expect "*ftp>*" Dedicated programs have been written to perform background ftp. While they use the same underlying mechanism as expect, their programmability leaves much to be desired. Since expect provides a high-level language, you can customize it to your needs. For example, you can add: persistence if the connection or transfer fails, you can retry every minute, hour, or even aperiodic intervals that depend on other factors such as user load. notication you can be notied upon transmission via mail, write or any other mechanism of your choice. You can even be notied of failure. initialization each user can have their own initialization le (e.g., .ftprc) in a high-level language for further customization, much like csh uses .cshrc.

expect could do many more sophisticated things. For example, it could use McGill Universitys Archie system. Archie is an anonymous telnet service that provides access to a database describing the contents of the entire Internets anonymous ftp repositories. Using this, a script could ask Archie where a le is, and then download it to your system. This requires only a few more lines at the beginning of the ftp script above. No known background-ftp programs provide even one of the features mentioned above, no less all of them. In expect, the implementation is trivial. Persistence requires a loop in the expect script. Notication is an exec of mail or write. An initialization le can be read with one command (source .ftprc does just the right thing) and can use any expect command. Although these features can be added by hooks into existing programs, there is still no guarantee that everyones needs will have been met. The only way to have such condence is to provide a general-purpose language. A good solution would be to integrate Tcl, itself, directly into ftp and other applications. Indeed, that was the original intent of Tcls design. Until this is done, expect provides much of the benet of Tcl to many applications without any rewriting at all.

7. fsck
fsck is yet another example of a program with an inadequate user interface. fsck provides almost no way of answering questions in advance. About all you can say is answer everything yes or answer everything no. The following fragment shows how a script can automatically answer some questions yes, and the rest no. The script begins by spawning fsck, and then answering yes to two types of questions, and no to everything else. for {} {1} {} { expect \ eof 8
Don Libes

break \

"*UNREF FILE*CLEAR?" "*BAD INODE*FIX?" "*? " }

{send "y\r"} \ {send "y\r"} \ {send "n\r"}

In the next version, the two questions are answered differently. Also, if the script sees something it doesnt understand, it executes the interact command which passes control back to the user. The user keystrokes go directly to fsck. When done, the user can exit or return control to the script, here triggered by pressing the plus key. If control is returned to the script, it continues automated processing where it left off. for {} {1} {} { expect \ eof "*UNREF FILE*CLEAR?" "*BAD INODE*FIX?" "*? " }

break \ {send "y\r"} \ {send "n\r"} \ {interact +}

Without expect, fsck can be run non-interactively only with very reduced functionality. It is barely programmable and yet it is the most critical of system administration tools. Many other tools have similarly decient user interfaces. In fact, the large number of these is precisely what inspired the original development of expect.

8. Controlling multiple processes: job control

expects concept of job control nesses some of the usual implementation difculties. Two issues are involved: The rst is how expect handles classic job control, such as occurs when you press ^Z at the terminal. The second is how expect handles multiple processes. The answer to the rst issue is: Ignore it. expect doesnt understand anything about classic job control. For example, if you spawn a program and then send it a ^Z, it will stop (courtesy of the pty driver) and expect will wait forever. In practice, however, this is not a problem. There is no reason for an expect script to ever send a ^Z to a process. It doesnt have to stop a process, per se. expect simply ignores a process, and turns its attention elsewhere. This is expects idea of job control and it works quite well. The user view of this is as follows: When a process is started by spawn, the variable spawn_id is set to a descriptor referring to that process. The process described by spawn_id is considered the current process. (This descriptor is exactly the pty le descriptor, although the user treats it as an opaque object.) expect and send interact only with the current process. Thus, to switch jobs all that is necessary is to assign the descriptor of another process to the variable spawn_id. Here is an example showing how job control can be used to have two chess processes interact. After spawning them, one process is told to move rst. In a loop, a move is sent from one process to the other, and vice versa. The read_move and send_move procedures are left as an exercise for the reader. (They are actually very easy to write, but too long to include here.)

expect: Scripts for Controlling Interactive Processes

spawn chess set id1 $spawn_id expect "Chess\r\n" send "first\r" read_move spawn chess set id2 $spawn_id expect "Chess\r\n" for {} {1} {} { send_move read_move set spawn_id $id1 send_move read_move set spawn_id $id2 }

;# start player one

;# force it to go first

;# start player two

Some applications are not like a chess game where players alternate moves in lock step. The following script implements a spoofer. It will control a terminal so that a user will be able to log in and work normally. However, whenever the system prompts for either password or login, expect begins recording keystrokes until the user presses return. This effectively collects just the logins and passwords of a user without the usual spoofer problem of seeing Incorrect password try again. Plus, if the user connects to another host, those additional logins will be recorded also!2 spawn tip /dev/tty17 set tty $spawn_id spawn login set login $spawn_id log_user 0 for {} {1} {} { set ready [select $tty $login] case $login in $ready { set spawn_id $login expect {"*password*" "*login*"} { send_user $expect_match set log 1 } "*" ;# ignore everything else set spawn_id $tty; send $expect_match } ;# open connection to ;# tty to be spoofed ;# open connection to ;# login process

2. The usual defense against a spoofer is to disallow write access so that the spoofer cannot open public terminals to begin with.

10

Don Libes

case $tty in $ready { set spawn_id $tty expect "*\r*" { if $log { send_user $expect_match set log 0 } } "*" if $log { send_user $expect_match } set spawn_id $login; send $expect_match } } The script works as follows. First connections are made to a login process and terminal. By default, an entire session is logged to the standard output (via send_user). Since this is not of interest, it is disabled by the command log_user 0. (A variety of commands are available to control exactly what is seen or logged.) In a loop, select3 waits for activity from either the terminal or the process and returns a list of spawn_ids with pending input. case executes an action if a value is found in a list. For example, if the string login appears in the output of the login process, the prompt is logged to the standard output and a ag is set so that the script will begin recording the users keystrokes until a return is pressed. Whatever was received is echoed to the terminal. A corresponding action occurs in the terminal half of the script. These examples have demonstrated expects form of job control. By interposing itself in a dialogue, expect can build arbitrarily complex I/O ow between processes. Multiple fan-out, multiplexed fan-in, and dynamically data-dependent process graphs are all possible. In contrast, the shell makes it extraordinarily difcult just to read through a le one line at a time. The shell forces the user to press control characters (^Z, ^C) and keywords (fg, bg) to switch jobs. These cannot be used from shell scripts. Similarly, the shell running non-interactively does not deal with history and other features designed solely for interactive use. This presents a similar problem as with passwd earlier. Namely, that it is impossible to build shell scripts which regressively test certain shell behavior. The result is that these aspects of the shell will inevitably not be rigorously tested. Using expect, it is possible to drive the shell using its interactive job control features. A spawned shell thinks it is running interactively, and will handle job control as usual. Not only does it solve the problem of testing shells and other programs that handle job control, but it also enables the shell to handle the job for expect when necessary. Processes to be manipulated with shell-style job control can be backed with a shell. This means that rst a shell is spawned, and then a command is sent to the shell to start the process. If the process is suspended by, for example, sending

3. select calls poll() on USG systems and, in retrospect, should have been called something less biased and more meaningful.

expect: Scripts for Controlling Interactive Processes

11

a ^Z, the process stops and control returns to the shell. As far as expect is concerned, it is still dealing with the same process (the original shell). Not only is expects approach exible, it also avoids duplicating the job control software that is already in the shell. By using the shell, you get the job control of your choice since you can pick the shell to spawn. And should you need to (such as when testing), you really can drive a shell so that it thinks it is running interactively. This is also useful for programs that change the way they buffer output after detecting whether they are running interactively or not. To further pin things down, during interact, expect puts the controlling terminal (the one expect was invoked from, not the pty) into raw mode so that all characters pass to the spawned process verbatim. When expect is not executing interact, the terminal is in cooked mode, at which time shell job control can be used on expect itself.

9. Using expect interactively

Earlier were shown scripts that are used interactively with interact. interact essentially gives a user free access to the dialogue, but sometimes ner control is desired. This can be achieved using expect which can read from the standard input just as easily as it reads from a process. A predened spawn_id maps to the standard input and the standard output. Alternatively, the commands expect_user and send_user perform I/O with the standard input and the standard output without changing spawn_id. The following script reads a line from the standard input for a given amount of time. This script (named timed_read) can be called from, for example, a csh script as set answer = timed_read 30. #!/usr/local/bin/expect -f set timeout [index $argv 1] expect_user "*\n" send_user $expect_match The third line accepts any newline-terminated line from the user. The last line returns it to the standard output. If nothing is typed before the timeout, nothing is returned. The rst line allows systems that support the #! magic to invoke the script directly (without saying expect before the script name) if its execute permission is set. Of course a script can always be invoked explicitly, as expect script. Options preceded by a c ag are executed as commands before any in the script. For example, an expect script can be traced without reediting by invoking it as expect -c "trace ..." script (where the ellipsis represents a tracing option). Multiple commands may be strung together on a single script line or within braces, separated by semi-colons. Naturally, this extends to the c argument. For example, the following command runs program foo for 20 seconds. expect -c "set timeout 20; spawn foo; expect"

12

Don Libes

Once the timeout is set and the program is spawned, expect waits for either an end-of-le or the 20 seconds to pass. If the end-of-le is seen, the program has (almost certainly) exited, and expect returns. If the timeout has passed, expect returns. In either case expect exits, implicitly killing the current process. It is educational to try and solve these last two examples without using expect. In both cases, the usual approach is to fork a second process that sleeps and then signals the original shell. If the process or read nishes rst, the shell kills the sleeper. Passing pids and preventing the background process start message is a stumbling block for all but the most expert shell programmers. Providing a general approach to starting multiple processes this way complicates the shell script immensely. Invariably, the programmer writes a special-purpose C program.
expect_user, send_user, and send_error (for writing to the standard error) are frequently used in longer expect scripts which translate a complex interaction from a process to a simple one for the user. In [7], Libes describes how adb could be securely wrapped with a script, preventing a system administrator from needing to master the intricacies of adb, while at the same time dramatically lessening the likelihood of a system crash due to an errant keystroke.

A simpler example is automating ftp to retrieve les from a personal account. In this case, a password must be supplied. Storing the cleartext password in a le should be avoided even if the le permissions are heavily restricted. Supplying passwords as arguments to a script is also a security risk due to the ability of ps to retrieve them. A solution is to call expect_user at the beginning of the script for each password that the script must supply later. The password will be available to the script (and only to the script), even if it has to retry ftp every hour. This technique is useful even if the information is to be entered immediately. For example, you can write a script which changes your password on every machine on which you have an account, whether or not the machines share a common password database (or even run UNIX). By hand, you might have to telnet to each machine and then enter the new password. With expect, you enter the password once and let the script do the rest of the work.
expect_user and interact can also be mixed in a single script. Imagine debugging a program that

only fails after many iterations of a loop. An expect script could drive the debugger, setting breakpoints, running the program for the appropriate number of loops, and then returning control to the keyboard. It could also alternate between looping and testing for a condition, before returning control.

10. Programming expect interactively

expect may be programmed interactively. For example, if expect is run with no arguments, it prompts for commands. This is similar to what one normally does when interactively using a shell. This interactive mode may also be entered by pressing a user-dened string during interact. Once prompted by the interpreter, you can type expect commands which are executed immediately. You can call dened procedures, perform job control, or even recursively invoke interact. For example, suppose you are running a script to automate fsck. You answer some of the questions yourself, and then decide that the rest should all be answered yes. You can escape from interact

expect: Scripts for Controlling Interactive Processes

13

to the expect interpreter and invoke a procedure to answer the remaining questions without further interaction from you. This can be made as complex as you like. The arguments to interact are actually string-action pairs. (The default action is to invoke the interpreter interactively.) This generalized mechanism can support all the usual styles of escapes. such as tips ~-prexed commands or cshs single-character job control keys. Actions may be any expect command. As an example, the following line maps the strings ~y, ~a, and the ^C and ^Z characters. interact \ ~y {yes} \ ~a {send "[exec date]"; send_user "hello world"} \ \Cc {exit} \ \Cz {exec kill -STOP 0} When ~y is typed, a procedure called yes is invoked. This could further automate the fsck interaction just described, so that the user does not have to explicitly start the interpreter and type yes. ~a invokes a more complex action. When typed, hello world is seen at the terminal and the current date is sent to the process as if the user had typed it. The other pairs exit or suspend an expect session while interacting with a spawned process. (With no map, the characters would be passed uninterpreted to the current process.) Appropriate maps can simulate csh-style job control or much fancier actions. For instance, ^Z could pass control to the interactive expect interpreter analogous to what ^Z does in the shell or it could change jobs to a spawned shell and resume the interaction. An unrealistic but amusing application of character mapping is the following script which runs a shell with a Dvorak keyboard. For brevity, only lowercase letters are mapped. proc dvorak {} { interact ~q {return continue} q {send } w {send ,} e r {send p} t {send y} y u {send g} i {send c} o p {send l} s {send o} d f {send u} g {send i} h j {send h} k {send t} l x {send q} c {send j} v b {send x} n {send b} , . {send v} / {send z} \; {send s} z {send \;} }

~d {} {send {send {send {send {send {send {send {send {send \

\ .} f} r} e} d} n} k} w} -}

\ \ \ \ \ \ \ \ \

log_user 0 scan [exec printenv SHELL] "%s" shell spawn $shell log_user 1 send_user "~d for dvorak, ~q for qwerty (default)\n" send_user "Enter ~ sequences using qwerty keys\n" interact ~d dvorak ~q {} 14
Don Libes

This script has two interacts. The user switches between them by typing ~d (for Dvorak) and ~q (for qwerty). The Dvorak translation occurs in the procedure dvorak dened with proc. Within dvorak, an interact gives each character an action that corresponds to sending its Dvorak counterpart instead. Nothing has to be sent to the user, since the character will be echoed (if necessary) by the current process. The return continue action for ~q causes the Dvorak interact to return the value continue to its caller. interacts caller happens to be an earlier interact (at the bottom of the script) which evaluates the continue and literally continues. This isnt anything magical. They are just Tcl commands that are appropriately handled. The script chooses the desired shell by examining the SHELL environment variable. Since printenv appends a newline to the end of its output, this has to be stripped off and is done here by scan an equivalent to scanf in the C programming language. This script is excessive and is not at all what this feature of interact was intended for. Nevertheless it works and demonstrates a number of interesting aspects.

11. Non-interactive programs are controlled differently

Some interactive programs have non-interactive alternatives. However, it is often the case that these alternatives are controlled in a way quite unlike the original interactive program. Thus, you need to learn two ways of doing things: interactively and non-interactively. For example, suppose you want to locate a printer server. This is described by the rm value in the printcap le. Interactively, you might use an editor, or even, more, to search the le for the correct printer and then begin scanning for the rm eld. To automate this, you must switch to a completely different program, such as awk. Alternatively, you could just translate the interaction you were doing by hand into send/expect sequences. The following fragment does exactly this. It was used by a larger script that manipulated printers by running lpc, the interactive interface to the BSD line printer system. spawn ed /etc/printcap expect {*\n} ;# discard character count send "/$printer/\r" for {} {1} {} { expect "*\r\n*:rm=*\n*" { # found rm, now get value set i [string first :rm= $expect_match] scan [range $expect_match [expr $i+4] end c] \ "%\[^:\]" server break } "*\r\n*\\\r\n" { # look at next line of entry send "\r" } "*\r\n*\n" { # no more lines in entry - give up 15

expect: Scripts for Controlling Interactive Processes

break } } This script uses ed although any editor could be used. First ed is directed to search for the printer. Once the printer is found, returns are sent to get the successive lines until the value is either located or no more lines remain. Using a specialized tool such as awk might seem like a better alternative, except if you arent familiar with awks style of processing. While the same claim could be made about expect, this script illustrates the idea that (ignoring syntax differences) you can automate a procedure you know how to do interactively by simply translating it into send/expect sequences.

12. Is expect too fast?

The previous example demonstrated how expect can use an editor to read a le. expect has simpler ways of reading les. For instance the command send [exec cat /etc/motd] writes the contents of /etc/motd to the current process. Calling a UNIX program to read a le may not seem like a fast method but it is a lot faster than having a user type it in. In a window environment, cutting and pasting is an alternative, but this takes a large amount of time also. Realistically, blazing speed is hardly needed in a program that simulates users. The speed of expect operations is described by Libes [2]. One side-effect not discussed is that expect can overrun input buffers designed for human typists. send supports a slow option (send s) specically to avoid this problem. It is controlled by parameters which describe the number of bytes to send atomically and a length of time to wait between each packet.
send also supports a simulation of actual human typing speed (send -h) according to a modi-

ed Weibull distribution [8], a common statistical tool to simulate interarrival times. The algorithm is driven by a random number generator and several user-chosen parameters. The parameters describe two average character interarrival times (default and word endings), minimum and maximum interarrival times, and a variability shape. Errors are not simulated as this can be done by the user directly. Simplistic errors may be generated by embedding typing mistakes and corrections (if desired) in a send argument. A more sophisticated approach could use an expert system as a coroutine.

13. Security
The passwd script shown earlier was designed solely to be used for conformance testing. Many system administrators want such a script to embed in a comprehensive adduser script, which would set up everything that a generic new user needs including an account and password. Unfortunately, calling the passwd script from another script reopens the very problem that the passwd program was designed to solve. Passwords should not be used as arguments to programs because they can be seen by ps and similar programs.

16

Don Libes

The solution is to have the expect script generate the passwords directly. This closes the hole, while at the same time forcing the use of computer-generated passwords which are generally more difcult to guess than human-generated passwords. This technique does not extend to programs such as telnet, ftp, su, etc., where a human really does need to provide the password. The solution is to have the expect script prompt for the password interactively via expect_user. In contrast to the program (or shell script) prompting when the password is needed, expect can prompt at the beginning of a script for all the passwords that will be needed. Even if the same password is used in several programs, the user need only enter it once since the script will remember it until it is needed. Often, it is convenient to run such scripts in the background. Starting processes asynchronously from the shell, however, prevents them from reading keyboard input. Thus expect scripts must be started synchronously. The fork and disconnect commands are used later to move expect into the background. For example, the following script reads the password, disconnects itself from the terminal, sleeps for one hour, and then goes on to execute commands that require a password. system stty -echo send_user "Password: " expect_user "*\n" send_user "\n" system stty echo scan $expect_match "%s" pass if [fork]!=0 exit disconnect exec sleep 3600 spawn su expect "*Password:*" send "$pass\r" # more commands follow ;# disable echoing

;# echo newline ;# strip off terminating \n

This script begins by disabling echo so that the password can be typed unseen. Unlike exec which manipulates its standard I/O so that it is accessible to expect, the system command does no manipulation, thereby allowing stty to affect the terminal.
fork literally causes expect to fork. Like the UNIX system call of the same name, it returns the

child process ID to the parent. Since the parent has nothing else to do, it immediately exits. The shell will detect this as normal program termination. Meanwhile, disconnect breaks the association between the child process and the terminal so that the rest of the script can continue immune to the user logging out. This paradigm provides a secure way of starting long-running background processes which require passwords. This works well with security schemes such as MITs Kerberos system. In order to run a process authenticated by Kerberos, all that is necessary is to spawn kinit to get a ticket, and similarly kdestroy when the ticket is no longer needed. Before expect, there was no way to achieve such results. The choice was either inexibility or insecurity. expect has made this choice unnecessary, and given us the best of both worlds.
expect: Scripts for Controlling Interactive Processes

17

14. Conclusions
expect provides a means of automating interactive programs. There are a great many such programs in the UNIX domain that lack non-interactive alternatives. expect leverages off of these programs with only a small amount of programming effort. expect solves a variety of problems with programs that 1) dont run non-interactively ( rlogin, telnet); 2) know theyre running interactively and change their behavior (csh, rn); 3) bypass stdio and open /dev/tty (crypt, passwd); 4) dont provide their full functionality non-interactively (fsck, ftp); or 5) dont provide the friendliest user interface (adb, rogue). All of the new noninteractive versions that result can now be usefully called from shell scripts because they can return meaningful error codes and no longer require user interaction. expect provides help even when you want to run programs interactively. If they lack a programmable interface, you can partially automate the interaction and then share control. Of course, the ideal solution is to rewrite the application with a programmable front-end. For new applications, there is no excuse not to use Tcl. It is small, efcient, easy to use, and probably sufces for 90% of all tools. Building Tcl into an application will always be better than an after-the-fact solution like expect. But for tools which dont warrant the Tcl library, or are too old to be rewritten, expect is a fast solution. expect is actually quite small. On a Sun 3, the current version is 64k. This includes the entire Tcl language. expect has few built-in functions. For example, expect doesnt have a communications protocol, nor does it know about sophisticated le access methods. It doesnt need to. It can invoke another program to do the work. At the same time, this gives you the exibility of using any software you already have. Do you need to communicate with a serial line? Use tip, cu, or kermit. With a TCP socket? Use telnet. You make the choice. This building block philosophy is very much in keeping with the UNIX tradition of hooking small programs together to build larger ones. In this respect, expect functions as a new kind of glue, much like the shell itself. Unfortunately, shell job control was designed only with interactive use in mind and cannot automatically control interactive processes. expects job control is generalized and has no such restriction. The two forms of job control do not interfere and can be used together. While expect only runs on UNIX, it can be useful in managing non-UNIX sites as long as they are networked to a UNIX host. Via telnet or tip, a script can login and play its usual interactive games. My site has scripts that do exactly this on VMS and Symbolics Lisp machines. Our VMS wizards would rather avoid UNIX entirely, but they know a timesaver when they see it.

15. Acknowledgments
This work was supported by the National Institute of Standards and Technology (NIST) Automated Manufacturing Research Facility (AMRF). The AMRF is funded by both NIST and the Navy Manufacturing Technology Program.

18

Don Libes

Thanks to Scott Paisley who wrote the callback script. John Ousterhout is responsible for Tcl, without which expect would not have been written. John also critiqued expect as well as the rst paper about it. I am indebted to him. Several people made important observations or wrote early scripts while I was still developing the command semantics. Thanks to Rob Densock, Ken Manheimer, Eric Newton, Scott Paisley, Steve Ray, Sandy Ressler, Harry Bochner, Ira Fuchs, Craig Warren, Barry Warsaw, Keith Eberhardt, Jerry Friesen, and Dan Bernstein. Thanks to Mike Gourlay, Clem Cole, Andy Holyer, and Alan Crosswell for help in porting expect to various UNIX platforms. Thanks to Steve Simmons, Joe Gorman, and Corey Satten for xing some of the bugs. Finally, thanks to K.C. Morris, Chuck Dinkel, Sue Mulroney, and the anonymous Computing Systems reviewers, who gave me extensive suggestions on improving this paper. Certain trade names and company products are mentioned in order to adequately specify procedures and equipment used. In no case does such identication imply recommendation or endorsement by the National Institute of Standards and Technology, nor does it imply that the products are necessarily the best available for the purpose.

16. Availability
Since the design and implementation was paid for by the U.S. government, expect is in the public domain. However, the author and NIST would like credit if this program, documentation or portions of them are used. expect may be ftped anonymously as pub/expect/expect.shar.Z from ftp.cme.nist.gov. Request email delivery by mailing to [email protected]. The contents of the message should be (no subject line) send pub/expect/expect.shar.Z. As of August, 1991, over 2500 sites had retrieved expect.

17. References
[1] [2] [3] [4] [5] [6] fsck, UNIX Programmers Manual, Section 8, Sun Microsystems, Inc., Mountain View, CA, September, 1989. Don Libes, expect: Curing Those Uncontrollable Fits of Interaction, Proceedings of the Summer 1990 USENIX Conference, Anaheim, California, June 11-15, 1990. John Ousterhout, Tcl: An Embeddable Command Language, Proceedings of the Winter 1990 USENIX Conference, Washington, D.C., January 22-26, 1990. John Ousterhout, tcl(3) overview of tool command language facilities, unpublished manual page, University of California at Berkeley, January 1990. Don Libes, expect User Manual, to be published as NIST IR 744-91, National Institute of Standards and Technology, Gaithersburg, MD. Don Libes, Regression Testing and Conformance Testing Interactive Programs, Proceedings of the Summer 1992 USENIX Conference, San Antonio, Texas, June 8-12, 1992.
expect: Scripts for Controlling Interactive Processes

19

[7]

Don Libes, Using expect to Automate Systems Administration Tasks, Proceedings of the Fourth USENIX Large Installation Systems Administration (LISA) Conference, Colorado Springs, Colorado, October 17-19, 1990. Norman Johnson, Samuel Kotz, Continuous Univariate Distributions, Vol. 1, Houghton Mifin Co, New York, NY, 1970.

[8]

20

Don Libes