GitHub Actions, nix-shell and coloured output

jacg · February 10, 2021, 6:04pm

GitHub Action logs show pretty coloured output.

When executing a project’s tests in a GHA with nix-shell --run <our test suite>, the output is monochromatic. Doing exactly the same locally, I get pretty colours in the test output.

The monochromatic test output on the GHA makes me really appreciate the colours: I want them back!

How can I persuade nix-shell to pick up whatever it is in the GHA environment that makes the colours appear?

abathur · February 12, 2021, 4:08pm

Curious about this as well. I have some CI jobs that run color-using nix-shell demos, and one of the downsides I’ve found about moving to GA from travis-ci is losing that colored output.

ryantm · February 12, 2021, 6:49pm

Typically I see test suites look at TERM to decide if the can output color. I’d print that out, outside and inside your shell to see what’s going on.

abathur · February 13, 2021, 7:21am

This may not be a perfect proxy for the OP’s case since I just stuck it in a nix-build.

Outside of the execution (i.e., as an entry in the .yml), it is dumb, but inside my test.sh (running inside the nix-build) it is xterm-256color as I’d expect.

jacg · February 13, 2021, 2:07pm

Hmm, I seem to get dumb both inside and outside:

- name: Check $TERM outside nix-shell
  run: echo $TERM
- name: Check $TERM inside nix-shell
  run: nix-shell --run "echo $TERM"

gives the outputs

Run echo $TERM
  echo $TERM
  shell: /bin/bash -e {0}
  env:
    NIX_PATH: nixpkgs=channel:nixos-unstable
dumb

and

Run nix-shell --run "echo $TERM"
  nix-shell --run "echo $TERM"
  shell: /bin/bash -e {0}
  env:
    NIX_PATH: nixpkgs=channel:nixos-unstable
[... irrelevant noise ...]
dumb

FWIW, my test runner is pytest. I guess I should try running it outside of the nix-shell and see if it produces any colour.

jacg · February 13, 2021, 3:03pm

pytest showed no colours in the GitHub action, both outside and inside the nix-shell.

In the case of pytest I have managed to get around the problem with

nix-shell --run "PYTEST_ADDOPTS=--color=yes <run my tests>"

domenkozar · February 14, 2021, 12:18pm

As hinted, I’d try:

...
runs-on: ubuntu-latest
env:
  TERM: xterm-256color

jacg · February 17, 2021, 5:16pm

That seems to result in monochrome pytest output.

cdepillabout · February 18, 2021, 4:46am

I don’t know if this would be helpful, but I recently reviewed a change to a library that tries to smartly determine whether to print in color or not.

Basically, the fix was to look at the output file descriptor for whether it is a terminal device. If it is, then print in color. If it is not, then don’t print in color.

Here’s the line in question:

github.com

cdepillabout/pretty-simple/blob/d1025af063ee0fcd70a1126a510faabbf2fc370c/src/Text/Pretty/Simple/Internal/Printer.hs#L192


      
          -- | Given 'OutputOptions', disable colorful output if the given handle
          -- is not connected to a TTY.
          hCheckTTY :: MonadIO m => Handle -> OutputOptions -> m OutputOptions
          hCheckTTY h options = liftIO $ conv <$> tty
            where
              conv :: Bool -> OutputOptions
              conv True = options
              conv False = options { outputOptionsColorOptions = Nothing }
          
          
    tty :: IO Bool
              tty = hIsTerminalDevice h
          
          
-- | Parse a string, and generate an intermediate representation,
          -- suitable for passing to any /prettyprinter/ backend.
          -- Used by 'Simple.pString' etc.
          layoutString :: OutputOptions -> String -> SimpleDocStream Style
          layoutString opts =
            annotateStyle opts
              . layoutSmart defaultLayoutOptions
                {layoutPageWidth = AvailablePerLine (outputOptionsPageWidth opts) 1}
              . indent (outputOptionsInitialIndent opts)

Here’s the library function:

https://hackage.haskell.org/package/base-4.14.1.0/docs/System-IO.html#v:hIsTerminalDevice

I’m not actually sure what function is being called under the hood, but my guess is something like isatty():

It could be that GitHub Actions is redirecting output in some way that isatty() doesn’t return the same thing that you were seeing on TravisCI. Although it seems like you’ve figured out that forcing color output works.

abathur · February 20, 2021, 6:02am

I poked at this a little bit, and it seems like:

GA isn’t running tests with a tty, while travis-ci is
Nix commands seem to filter ansi escapes if !tty (not certain I’m interpreting this correctly; I’m not really familiar with the Nix source)
I found a suggestion for hacking this with the script command, but it only works for the linux builds.

It doesn’t look like GA has any short run plan to change this, so it seems like the “solution” is for Nix commands to have a flag or env that explicitly enables passing through escapes. (see also: https://no-color.org)

With the caveat that I don’t know exactly why travis-ci and GA disagree here and realize they may have some compelling reasons, it strikes me as incredibly sad/stupid that every person/team/project who encounters this has to waste time understanding the issue, hunting down which tool in their toolchain is eating the sequences, and fiddling flags/envs (or filing issues…)

brendanzab · December 30, 2021, 11:05am

Heh, thanks to the above, I managed to get the colours on my run commands back with:

defaults:
  run:
    shell: 'script -q -e -c "export TERM=xterm-256color; bash {0}"'

Only problem was that I then ended up getting a bunch of spammy output regarding as the packages downloaded:

I’m guessing github doesn’t allow for the rewriting of previous lines?

abathur · December 30, 2021, 6:21pm

For others reading later: the above will only work on linux (bsd script doesn’t have the -e or -c args), so it won’t be a suitable default if you have macos in your build matrix.

Also, I noticed a new comment since I last read that thread suggesting the above may mark some failing builds as successful?

https://github.com/actions/runner/issues/241#issuecomment-842566950

blaggacao · July 28, 2023, 1:36am

This approach claims to be cross platform: https://github.com/Yuri6037/Action-FakeTTY/tree/master/script