You can learn the 2>&1 thing, but it would be nice if there was a feature of one of the greps to "slurp up" the error output.

Of course, the real problem is there's no standard, the standards that do exist are ignored, and each new command-line tool generates a new standard.

It generally isn't possible for grep to "slurp up" the error output because stderr does not get passed through pipes (by default). This is shell behavior and grep cannot do anything about it.

As a side note, some shells have started implementing a shorthand for `foo 2>&1 | grep` which is `foo |& grep`.

Probably you could do a hack like finding the program at the other end of the pipe and looking at its open file descriptors, but lots of corner cases to consider.

Right, the problem is that apparently shells haven't seen user need for selectively redirecting one or the other. Like I might want to see foo's stdout on the terminal, but pipe stderr (and only stderr) to grep, perhaps so I only see some specific error messages that I care about. `2>&1` will send stderr to stdout, and then the pipe will send everything to grep (and I believe `|&` does the same).

(To be fair, that does seem like a pretty rare use case.)

Actually, you can. If you run:

    foo 2>&1 >&- | grep ...

then you will grep only the stderr. The stdout will be discarded.

Alternatively, if you only want to grep to see the stderr, but you still want to see stdout, you can swap stderr and stdout like this:

    foo 3>&2 2>&1 1>&3 | grep ...

And in many shells, you can easily split the stdout and stderr into two separate pipelines like this:

    foo > >(grep stdout) 2> >(grep stderr)

> Like I might want to see foo's stdout on the terminal, but pipe stderr (and only stderr) to grep, perhaps so I only see some specific error messages that I care about

You mean this:

`foo 3>&1 1>&2 2>&3`

The above swaps stdout and stderr, saving + restoring the original stdout through descriptor 3. Shell redirection expressions map directly to dup2 syscalls, except the operands are reversed: 3>&1 is dup2(1, 3). Pipes map fairly simply to fork + exec. At its core the shell is a rather thin wrapper around the core Unix syscalls fork, exec, dup2, open, and close. If you look at the original shell implementation, the command parser more-or-less executes these syscalls as it goes along, left to right.

How can you grep output if there's no output?

Can you provide a concrete example of such an application?

Some applications print usage information (e.g. with --help flag) to stderr. That is a bit annoying when there is a lot of output and I just want to grep for some flag, since the first attempt will fail and then I have to try again after adding 2>&1.

Then there's `ssh-keyscan`, printing the "real" output to stdout, but also printing "comments" to stderr.

Nothing wrong with that, but it was a bit confusing the first time I saw that behavior.

If I pass the --help flag, I'm explicitly asking the program to print the help text. In this case it's appropriate for the help text to be written to standard output since the help text is the output.

If the user passes invalid options and inputs, then the program should produce no output and any error or help messages should be written to the error stream. It's important that these messages end up on the terminal when one gets the command invocation wrong while trying to pipe data into another program. If they're written to standard output, the program on the other end of the pipe will slurp all the error messages up and try to parse them as though they were valid inputs.

I think what they meant was something like:

> Unless your app doesn't generate [stdout] output,then grepping the [visible] output will not work if it's stderr and it will be confusing

With the first "output" meaning "specifically stdout", and the second meaning "what you see on the terminal".

As in, you run a command, see it printed a lot of stuff, so then you do it again but piping it to `grep` or `less`, and realize none of that output was actually going to stdout.

I see what you mean. My logic only really applies to the simpler programs which exclusively use the standard file descriptors. It gets fuzzier when the output of a program is an actual file somewhere.

The output of a package manager is the packages. The text it writes to the terminal is not part of the output, it's just the program's log. I haven't thought about what's appropriate in this situation. By unix standards package managers shouldn't even be printing anything to begin with.