My Editing Workflow

I recently considered that my day-to-day editing cycle might be of some interest, so here it is.

πŸ”— Vim’s quickfix

I have been using quickfix in anger for many years. The gist of it is that given a simple (but configurable if you absolutely must) format of filenames, cursor positions, and extra text, you can create a list that vim can help you easily iterate over. I often use it with fugitive’s :Ggrep functionality, to iterate over lines matching a pattern, but it’s really easy to wire into anything else.

I have a little command to run all of our Go tests and add any error positions (usually compilation failures but they could be actual failing tests) to the quickfix. It’s this easy:

command! -nargs=* GoTest execute 'cexpr system("go test ' <args> '")'

I can run :GoTest (or often :GoTest ./... to test all packages).

Any lines that look like this will be added to the list:

some/file:123:23: busted code here
other/file:312:2: more code here

Then you can use :cnext, :cprev to go forward and back, or :copen to show the window… Or best, use unimpaired which maps [q and ]q to the next or previous item in the quickfix list.

Awesome! With this I can now bounce back and forth between all these errors.

πŸ”— minotaur

It’s great to be able to iterate over all the changes you need to make, but I want a faster feedback cycle. With my tool minotaur I can easily test my code on each save to disk. The most direct usage is:

$ minotaur . -- sh -c 'go test ./...'

(go test is wrapped with shell because minotaur passes the changed files to the script, which go test doesn’t understand, so we wrap to ignore.)

That works great, and I use that pattern all the time to automate little scripts based on file events. It works on all three major operating systems and the interface is about as simple as it could get.

πŸ”— gotest

I have a thin wrapper atop minotaur though to get slightly nicer output. Here’s all the code, since it has some neat tricks:

#!/bin/bash

test() {
   echo ========== $(date) ==========

   # unpack GOTESTARGS back into [email protected]
   eval "$GOTESTARGS"
   set -- "${s[@]}"

   go test "[email protected]"
   echo ==================================================
}

if [ "$1" = "test" ]; then
   shift
   test
else
   # Serialize into string
   GOTESTARGS=$(
      s=("[email protected]")        # temporary array
      set | grep ^s=  # `set` serializes a named array
   )
   export GOTESTARGS

   # reexec the current script; this works from minotaur and
   # ensures the reexec code is working on the first run
   "$0" test
   leatherman minotaur . -- "$0" test
fi

This code does a few interesting things:

  • It serializes all arguments into a single string, so they can be passed via env var
  • It deserializes those args back into [email protected]
  • It reΓ«xec’s itself immediately so the user running it sees output at startup

(I have to mention here that I wouldn’t have been able to figure this out without help from my friend Ingy dΓΆt Net.)

I think being able to do this serialize/deserialize trick is really powerful, since keeping args as a list of tokens is really useful in Unix systems.

πŸ”— …but then I got nervous

While writing the above I considered: am I proud of the above because I solved real problems, or because I solved problems of my own doing? With a few added, trivial features in minotaur the above code becomes:

#!/bin/sh

exec minotaur -report -run-at-start -suppress-args \
        . -- go test "[email protected]"

And if I get real and just change minotaur such that the common case is default, it’s even simpler:

#!/bin/sh

exec minotaur -report . -- go test "[email protected]"

In theory the other interface is as simple as it gets, which means it is maximally composeable. That’s great, but these extra little features (about 30 more lines of code in total) make working with minotaur easier to use…


When I rewrote minotaur last I was striving for something that would be able to limit script runs based on which files changed. I still think that’s worth doing, but in general it’s vastly more work than it is to just eat the time.

For many years I’ve tried to have interfaces that do exactly one thing. A part of me still thinks that’s the best abstraction, but a growing part of me thinks that if an abstraction does exactly one thing it’s not even really an abstraction, it’s a wrapper.

One of my favorite abstractions of all time is SQL. If SQL were written in these terms it would be an inscrutable, inefficient pipeline. I love being able to throw together little ad-hoc systems, but if these things are so off-the-cuff, shouldn’t they be easy and fun? Should they really require knowledge of how to serialize lists of tokens in bash?

And for things that are not ad-hoc, for things that are built to last, surely they should be written in a style that is not weird or surprising. When are thin wrappers actually appropriate? I don’t know. I am coming around to the idea that interfaces should generally have lots of functionality.

Thanks to Matthew Horsfall for reviewing this post.


(The following includes affiliate links.)

I recently read A Philosophy of Software Design. I really enjoyed it and will likely have a whole blog post about it, but in short it has informed some of the opinions above. I suggest reading it.

Totally unrelated but another book I recently read was BPF Performance Tools. I intensely enjoyed this book and will definitely publish an article about it. It may not be for everyone, but if you are interested in deep visibility of software on Linux, this is a great set of tooling to invest in.

Posted Mon, Feb 3, 2020

If you're interested in being notified when new posts are published, you can subscribe here; you'll get an email once a week at the most.