Getting notification when a long-running command finishes.

My job often involves running a command to do something, realising that it's going to take longer than I expected and going doing something else in the meantime. Conversely I'll run something that should take a long time, go do something else and find that it stopped with an error after a very short time and I didn't notice.

For a long time I've wanted to set things up so that any command that takes more than 20s to run would send me an IM when it was finished. What I didn't want was to have to prefix all my commands with some notification because that's annoying and I would probably forget in exactly the cases I need it.

It sounds pretty simple but getting it work turned out to be quite fiddly.

Delivering an instant message.

There's a program call sendxmpp that will deliver an instant message over XMPP. It's written in Perl and definitely works on Linux but probably also Mac and Windows if you really want. Every few months I would try to make it work and when I finally got it working, it would stop working again the next day.. Apparently Google's XMPP service is particularly finicky. I appear to have made it work for real now. I honestly don't know what I did differently, it may even be something that changed when Google deprecated XMPP in favour of Hangouts. As I write this, I see that it no longer works for one of my accounts but it's working fine with another! If there's a more reliable option, I'd love to know.

To make this work you'll need a second Google account because sending a message from yourself to yourself doesn't seem to work. So I created an account that I'm going to use only for sending IMs. After creating that account, I invited it to chat and made sure that the accounts can chat with each other through the normal Google chat interface.

Next I created a config file for it with the following commands (you'll need to replace some parts of the last one)

touch ~/.sendxmpp chmod 600 ~/.sendxmpp echo my.new.im.username@gmail.com;talk.google.com MyNewPassword > ~/.sendxmpp

Now to send a notification to myself I just do

echo some message | sendxmpp -t -u my.new.im.username -o gmail.com other.username@gmail.com

and the message shows up. I also get a bunch of Perl warnings but such is life.

Finally I wrapped that up as a command called notify and I can just pipe a message into that.

Spotting that a command is taking a long time to finish

I hate unix shells, they're a horrible mishmash of special cases, obscure features and crap that's only that way for backwards compatibility. What I've done is a hack in bash and it's not perfect (it seems like it would be slightly less hacky in zsh but that wouldn't fix the oddities described below). The core is this

trap bash_pre_cmd DEBUG PROMPT_COMMAND=bash_post_command

What that does is run bash_pre_cmd every time you hit enter to kick off a command and bash_post_command every time it prints the command prompt (which indicates that the command has finished, although there are various way for this not to be true, hence the oddities but in normal usage it's true). So all that remains is sensible definitions for those two commands

# Prefix things with "fergal_" to avoid name clashes. # Sets $fergal_timer to the current value of $SECONDS if it doesn't already have # a value. For whatever reason, this get called multiple times per command typed # in the shell, so we only want to start once. function fergal_start_timer () { if [[ -z $fergal_timer ]]; then fergal_timer=$SECONDS fi return 0 } # Calls fergal_timer_callback with two arguments: # 1 the number of seconds elapsed since fergal_start_timer was called # 2 the last command. # Safely handles the case where $fergal_timer was not yet set (e.g. on # startup). function fergal_end_timer () { if [[ -n $fergal_timer ]]; then fergal_timer_callback $(($SECONDS - $fergal_timer)) \ "$(fc -ln -1 | sed -e 's/^[ \t]*//')" fergal_timer= fi } function bash_pre_cmd () { fergal_start_timer } function bash_post_command () { record_eternal_history fergal_end_timer } # Notify if it takes more than 20s but always notify for "make" unless it takes 0 seconds. function fergal_timer_callback () { local time=$1 shift local command="$1" shift local notify if echo $command | grep -q "^make\b" || [[ $time -gt 20 ]]; then notify=1 fi # Never notify if it took 0 seconds. Mostly important to avoid # notifying when the user hits enter on a blank line. if [[ $time == 0 ]]; then notify= fi if [[ $notify ]]; then # sendxmpp blurts out crap on stderr even when everything works # fine, so we suppress that. (echo Completed after ${time}s; echo $command) | \ tee >(notify > ~/.notify_ouput 2>&1 || echo >&2 notify failed) fi }

You'll probably want to customize the callback for your own preferences.

Put all that into your .bashrc and cross your fingers

Oddities

If you suspend a process, that will cause the prompt to be printed and may cause a notification. If you unsuspend that process with fg then when it finishes, this code will see fg as the command that just completed. Also, we cannot notify for 0 seconds otherwise just hitting enter would cause notification and the IM would refer to the last command in the history. Interactive commands obviously take a long time and so cause useless notifications. I'll probably whitelist a few common ones like man and less. There are probably some problems too but for normal type-run-complete shell use it works just fine.

Basically shells are shitty and not very flexible. Doing this correctly would involve being able to stash away some state along with each command that's started then getting access to that state when the command finishes, it would require changes to shell.