(Please, help me adjust title and tags.)
When I run connmanctl I get a different prompt,
enrico:~$ connmanctl
connmanctl>
and different commands are available, like services, technologies, connect, ...
I'd like to know how this thing works.
I know that, in general, changing the prompt can be just a matter of changing the variable PS1. However this thing alone (read "the command connmanctl changes PS1 and returns) wouldn't have any effect at all on the functionalities of the commands line (I would still be in the same bash process).
Indeed, the fact that the available commands are changed, looks to me like the proof that connmanctl is running all the time the prompt is connmanctl>, and that, upon running connmanctl, a while loop is entered with a read statement in it, followed by a bunch of commands which process the the input.
In this latter scenario that I imagine, there's not even need to change PS1, as the connmanctl> line could simply be obtained by echo -n "connmanctl> ".
The reason behind this curiosity is that I'm trying to write a wrapper to connmanctl. I've already written it, and it works as intended, except that I don't know how to properly setup the autocompletion feature, and I think that in order to do so I first need to understand what is the right way to write an interactive shell script.
How can I capture the input/ output from a script in realtime (such as with tee), but line-by-line instead of character-by-character? My goal is to capture the input typed into the interactive prompts of a script only after backspaces and auto-completion have finished processing (after the RETURN key is hit).
Specifically, I am trying to create a wrapper script for ssh that creates a timestamped log of commands used on remote servers. The script, which uses tee to redirect the output for filtering, works well, but the redirected output gets jumbled with unsubmitted characters whenever I use the backspace key or the up/down keys to scroll through my remote history. For example: service test stopexitservice test stopart or cd ..logs[1Pls -al.
Perhaps there is a way to capture the terminal's scrollback and redirect that like with tee?
Update: I have found a character-based cleanup solution that does what I want most of the time. However, I am still hoping for an answer to this question (which may well be msw's answer that it is very difficult to do).
In the Unix world there are two primary modes of handling keyboard input. These are known as 'raw' in which characters are passed from the terminal to the reading program one at a time. This is the mode that editors (and such) will use because the editor needs to respond immediately when you press a key.
The other terminal discipline is called 'cooked' which is the line by line behavior that you think of as the bash line by line input where you get to backspace and the command is not executed until you press return. Ssh has to take your input in raw, character-by-character mode because it has no idea what is running on the other side. For example, if you are running an editor on the far side, it can't wait for a return before sending the key-press. So, as some have suggested, grabbing shell history on the far side is the only reasonable way to get a command-by-command record of the bash commands you typed.
I oversimplified for clarity; actually most installations of bash take input in raw mode because they allow editor like command modification. For example, Ctrl-P scrolls up the command history or Ctrl-A goes to the beginning of the line. And bash needs to be able to get those keys the moment they are typed not waiting for a return.
This is another reason that capturing on the local side is obnoxiously difficult: if you capture on the local side, the stream will be filled with Backspaces and all of bash's editing commands. To get a true transcript of what the remote shell actually executed you have to parse the character stream as if you were the remote shell. There also a problem if you run something like
vi /some_file/which_is_on_the_remote/machine
the input stream to the local ssh will be filled with movement commands snippets of text including backspaces and so on and it would be bloody difficult to figure out what is part of a bash command and what is you talking to the editor.
Few things involving computers are impossible; getting clean input from the local side of an ssh invocation is really, really hard.
I question the actual utility of recording the commands that you execute on a local or remote machine. The reason is that there is so much state which is not visible from a command log. As a simple example here's a log of two commands:
17:00$ cp important_file important_file.bak
17:15$ rm important_file
and two days later you are trying to figure out whether important_file.bak should have the contents you intended or not. Given that log you can't answer that simple question. Even if you had the sequence
16:58$ cat important_file
17:00$ cp important_file important_file.bak
17:15$ rm important_file
If you aren't capturing the output, the cat in the log will not tell you anything. Give me almost any command sequence and I can envision a scenario in which it will not give you the information you need to make sense of what was done.
For a very similar purpose I use GNU screen which offer the option to record everything you do in a shell session (INPUT/OUTPUT). The log it creates also comes with undesirable characters but I clean them with perl:
perl -ne 's/\x1b[[()=][;?0-9]*[0-9A-Za-z]?//g;s/\r//g;s/\007//g;print' < screenlog.0
I hope this helps.
Some features of screen:
http://speaking-my-language.blogspot.com/2010/09/top-5-underused-gnu-screen-features.html
Site I found the perl-oneliner:
https://superuser.com/questions/99128/removing-the-escape-characters-from-gnu-screens-screenlog-n
Every once in a while I have to fire up a GUI program from my terminal session to do something. It usually is Chrome to display some HTML file are some task alike.
These programs however throw warnings all over the place and it can actually become ridiculous to write anything so I always wanted to redirect stderr/stdout to /dev/null.
While ($PROGRAM &) &>/dev/null seems okay I decided to create a simple Bash function for it so I don't have to repeat myself everytime.
So for now my solution is something like this:
#
# silly little function
#
gui ()
{
if [ $# -gt 0 ] ; then
($# &) &>/dev/null
else
echo "missing argument"
fi
}
#
# silly little example
#
alias google-chrome='gui google-chrome'
So what I'm wondering about is:
Is there a way without an endless list of aliases that's still snappy?
Are there different strategies to accomplish this?
Do other shells offer different solutions?
In asking these questions I want to point out that your strategies and solutions might deviate substantially form mine. Redirecting output to /dev/null and aliasing it was the only way I know but there might be entirely different ways that are more efficient.
Hence this question :-)
As others have pointed in the comments, I think the real problem is a way to distinguish between gui vs. non-gui apps on the commandline. As such, the cleanest way I could think of is to put this part of your script:
#!/bin/bash
if [ $# -gt 0 ] ; then
($# &) &>/dev/null
else
echo "missing argument"
fi
into a file called gui, then chmod +x it and put it in your ~/bin/ (make sure ~/bin is in your $PATH). Now you can launch gui apps with:
`gui google-chrome`
on the prompt.
Alternatively, you can do the above, then make use of bind:
bind 'RETURN: "\e[1~gui \e[4~\n"'
This will allow you to just do:
google-chrome
on the prompt and it would automatically append gui before google-chrome
Or, you can bind the above action to F12 instead of RETURN with
bind '"\e[24~": "\e[1~gui \e[4~\n"'
To separate what you want launched with gui vs. non-gui.
More discussion on binding here and here.
These alternatives offer you a way out of endless aliases; a mix of:
Putting gui in your ~/bin/, and
Binding use of gui to F12 as shown above
seems the most ideal (albeit hacky) solution.
Update - #Enno Weichert's Resultant Solution:
Rounding this solution out ...
This would take care of aliases (in a somewhat whacky way though) and different escape encodings (in a more pragmatic rather than exhaustive way).
Put this in $(HOME)/bin/quiet
#!/bin/bash -i
if [ $# -gt 0 ] ; then
# Expand if $1 is an alias
if [ $(alias -p | awk -F "[ =]" '{print $2}' | grep -x $1) > 0 ] ; then
set -- $(alias $1 | awk -F "['']" '{print $2}') "${#:2}"
fi
($# &) &>/dev/null
else
echo "missing argument"
fi
And this in $(HOME)/.inputrc
#
# Bind prepend `quiet ` to [ALT][RETURN]
#
# The condition is of limited use actually but serves to seperate
# TTY instances from Gnome Terminal instances for me.
# There might very well be other VT emulators that ID as `xterm`
# but use totally different escape codes!
#
$if $term=xterm
"\e\C-j": "\eOHquiet \eOF\n"
$else
"\e\C-m": "\e[1~quiet \e[4~\n"
$endif
Although it is an ugly hack, I sometimes use nohup for a similar needs. It has the side effect of redirecting the command's output, and it makes the program independent of the terminal session.
For the case of running GUI programs in a desktop envinroment it has only little risk for resource leaks as the program will anyway end with window manager session. Nevertheless it should be taken into account.
There is also the option of opening another terminal session. In Gnome for instance, you can use gnome-terminal --comand 'yourapp'.
But this will result in opening many useless terminal windows.
First, it's a great idea to launch GUI applications from the terminal, as this reduces mouse usage. It's faster, and more convenient in terms of options and arguments. For example, take the browser. Say you have the URL in the clipboard, ready to paste. Just type, say, ice (for Iceweasel) and hit Shift-Insert (to paste) and Enter. Compare this to clicking an icon (possibly in a menu), wait for the window to load (even worse if there is a start up page), then click the URL bar (or hit Ctrl-L), then Ctrl-V... So I understand you desire for this to work.
But, I don't see how this would require an "infinite" list of aliases and functions. Are you really using that many GUI applications? And even so, aliases are one line - functions, which may be more practical for handling arguments, are perhaps 1-5 lines of (sparse) code. And, don't feel you need to set them up once and for all - set them up one by one as you go along, when the need arises. Before long, you'll have them all.
Also, if you have a tabbed terminal, like urxvt (there is a Perl extension), you'd benefit from moving from "GUI:s" to "CLI:s": for downloading, there's rtorrent; for IRC, irssi; instead of XEmacs (or emacs), emacs -nw; there is a CLI interface to vlc for streaming music; for mail, Emacs' rmail; etc. etc.! Go hunt :)
(The only sad exception I've run across that I think is a lost cause is the browser. Lynx, W3M, etc., may be great from a technical perspective, but that won't always even matter, as modern web pages are simply not designed with those, text-only browsers in mind. In all honesty, a lot of those pages look a lot less clear in those browsers.)
Hint: To get the most out of a tabbed terminal, you'd like the "change tab" shortcuts "close" (e.g., Alt-J for previous tab and Alt-K for next tab, not the arrow keys that'll make you reach).
Last, one solution that'll circumvent this problem, and that is to launch the "GUI:s" as background processes (with &) in your ~/.xinitrc (including your terminal emulator). Not very flexible, but great for the stuff you always use, every time you use your computer.
Ok, so I kept thinking that the one function should be enough. No aliases, no bashisms, no nonsense. But it seemed to me that the only the way to do that without possibly affecting regular use, such as expansions and completions, was to put the function at the end of the command. This is not as easy as one might first assume.
First I considered a function tacked onto the end in which I would call, say, printf "%b" "\u" in order to cut the current line, plug in another printf, paste it back in, and a quote or two at the beginning, then do what little is needed at the end. I don't know how to make this work though, I'm sorry to say. And even if I did, I couldn't hope for any real reliability/portability with this method due to the varying ways shells interpret escape sequences, not to mention the terminal emulators they run in. Perhaps stty could offer a way forward along these lines, but if so, you won't find it here... now, anyway.
I eventually instead resorted to actually copying the current command's /proc/{PID}/cmdline to a variable, then (shamefully) killing it entirely, and finally wrapping it as I pleased. On the plus side, this is very easily done, very quickly done (though I can imagine arguing its 'efficiency' either way), and seems mostly to work, regardless of the original input, whether that be an alias, a variable, a function, etc. I believe it is also POSIX portable (though I can't remember if I need to specify the kill SIGNALS by name for POSIX or not), and is definitely no nonsense.
On the other hand, its elegance certainly leaves much to be desired, and, though it's probably not worth worrying about, it does waste entirely a single PID. And it doesn't stop completely the shell spam; that is, in my shell I have enabled background jobs reporting with set and so, when first run, the shell kindly informs me that I've just opened and wasted a PID in two lines. Also, because I copy the ../cmdline instead of interfacing directly with the 0 file descriptor, I expect pipes and ; and etc to be problematic. This I can, and likely will, fix myself very soon.
I will fix that, that is, if I cannot find a way to instead make use of, as I suspect can be done,SIGTSTP + SIGCONT by first suspending the process outside a subshell then within one continuing it after redirecting the subshell's output. It seems that this works unreliably for reasons I haven't yet discovered, but I think it's promising. Perhaps nohup and a trap (to effectively rehup it, as it were) is what is needed, but I'm not really sure how to put those together either...
Amyway, without further ado, my semi-aborted, backwards gui launcher:
% _G() { (
_gui_cmd="$(tr '\0' ' ' </proc/"$\!"/cmdline )" ;
kill -9 "$\!" ;
exec eval "${_gui_cmd} &" )
&>/dev/null 2&>1
}
% google-chrome-beta --disk-cache-dir="/tmp/cache" --disk-cache-size=100000000 &_G
[1] 2674
[1] + 2674 killed google-chrome-beta --disk-cache-dir="/tmp/cache" --disk-cache-
%
So another problem one might encounter if attempting to do similar is the order of expansion the shell assumes. Without some trick such as mine you're sure to have some serious difficulty expanding your function before the prior command snags it as an argument. I believe I have guarded against this without unnecessary redundancy by simply tacking my _G function call onto the original command's & background intstruction. Simply add &_Gto the tail-end of the command you wish to run and, well, good luck.
-Mike
P.S. Ok, so writing that last sentence makes me think of what might be done with tee.
I am trying to understand the code in this page: https://github.com/corroded/git-achievements/blob/gh-pages/git-achievements
and I'm kinda at a loss on how it actually works. I do know some bash and shell scripting, but how does this script actually "store" how many times you've used a command(im guessing saving into a text file?) and how does it "sense" that you actually typed in a git command? I have a feeling it's line 464 onwards that does it but I don't seem to quite follow the logic.
Can anyone explain this in a bit more understandable context?
I plan to do some achievements for other commands and I hope to have an idea on HOW to go about it without randomly copying and pasting stuff and voodoo.
Yes on 464 start the script, everything before are helping functions. I dont know how it gets installed, but I would assume you have to call this script instead of the normal git-command. It just checks if the first parameter is achievement, and if not then just (regular) git with the rest parameters is executed. Afterwards he checks if an error happend (if he exits). And then he just makes log_action and check_for_achievments. log_action just writes the issued command with a date into a text file, while achievments scans for that log file for certains events. If you want to add another achievment you have to do it in this check_for_achievments.
Just look how the big case handles it (most of the achievments call the count_function which counts the # usages of the function and matches when a power of 2 is reached).
I sweated over the question above. The answer I'm going to supply took me a while to piece together, but it still seems hopelessly primitive and hacky compared to what one could do were completion to be redesigned to be less staticky. I'm almost afraid to ask if there's some good reason that completion logic seems to be completely divorced from the program it's completing for.
I wrote a command line library (can be seen in scala trunk) which lets you flip a switch to have a "--bash" option. If you run
./program --bash
It calculates the completion file, writes it out to a tempfile, and echoes
. /path/to/temp/file
to the console. The result is that you can use backticks like so:
`./program --bash`
and you will have completion for "program" in the current shell since it will source the tempfile.
For a concrete example: check out scala trunk and run test/partest.