stdout and stderr usage consistency for typespec

[timotheeguerin]

From issue #6574 there seems to be some inconsistencies on what we emit to stderr.

First thing to state is that only the process status code should be used to decide if the typespec compile(or other command) fails. Assuming stderr means an error occur is an invalid behavior.

Option 1: stdout for output and stderr for system diagnostics

The general idea is that stdout should be for the output of the program and stderr for system diagnostics(not errors).
The idea with using this separation that is widely used is that you can have command like --json while still emitting logs to stderror.

Types of output we currently have

Type	Currently using	Using system above
TypeSpec diagnostics	stdout	stdout
Compiler or library crash	stderr	stdout
Tracing	stdout	stderr
Progress spinner	stderr	stderr

Diagnostics are part of the typespec output (they are basically the only output that is produced on a regular basis so it does make sense they belong to stdout.

Option 2:

Don't use stderr at all, because we can't trust Microsoft tools built on the asusmption that stderr always means an error is built

[adrian05-ms]

Hello, by any chance do you guys have an ETA for this?

I found weird behavior between typespec and rush libraries.

In our project we use rush to handle multiple packages in one repo. And with the newer versions of typespec we noticed that rush is now returning build succeed with warnings even though no warning were printed. We noticed it was because the compiler prints the "Compiling" messages using stderr as mentioned above and rush treats anything printed there as warning/ error. It's even being logged into the build error log.

What's the issue?
Rush treats the builds with warnings as failure. We can bypass that by modifying some settings on rush, but we want at some point to fail builds when warnings happen for our ci pipeline. Unfortunately, we can't do that due to the noise the compiler is generating as warning.

is it possible to move the console messages from the compiler that are not warning nor error out of stderror?

[timotheeguerin]

no we decided option 2 was not the way to go, what rush is doing is very frustrating and has been complained many times, It is again a misuse of stderr by microsoft product. We explicitly stopped using rush and moved to pnpm because of this as too many tools cause this. You can work around it by redirecting the stderr to stdout 2>&1

The issue left here is just to make sure we are consitent in our output.