Following up on jglick's comment in JENKINS-64831:

I suppose this could be dealt with by having the stdio mode use a slightly more complicated protocol (new Channel.Mode?) whereby each packet is printed as text-ish with a distinctive or even random prefix and a trailing newline. Thus any unrelated messages printed to stdout would just be ignored.

I suppose this is possible and could even be a variant of the existing Channel.Mode.TEXT mode that is split up into line-based frames with a prefix for each line. One downside to that scheme is that Base64-encoding the data would decrease efficiency.

One question I have is whether anything else besides SSHLauncher (from SSH Build Agents) is using standard input and standard output as the streams behind a ClassicCommandTransport. That is the only use case I am aware of and it seems like the primary one. If so, perhaps an alternative would be an SSH-specific solution that creates another SSH channel for Remoting I/O, leaving the primary channel as-is for standard input and standard output. The dedicated Remoting SSH channel would effectively function as its own socket whose input and output streams would back the ClassicCommandTransport. Since Windows now supports AF_UNIX domain sockets, OpenSSH supports Unix domain socket forwarding as of 6.7, and Java 17 has native support for UnixDomainSocketAddress, perhaps we could create a Unix domain socket on the controller, forward it to the agent over SSH (not sure if Trilead supports this yet), and pass the socket path to Remoting, which would create its ClassicCommandTransport using the input and output of that socket instead of standard input and standard output (which would remain untouched). Thus we would not need to write our own framing protocol but could rely on the built-in multiplexing offered by the SSH protocol itself.

whether anything else besides SSHLauncher (from SSH Build Agents) is using standard input and standard output

https://github.com/jenkinsci/command-launcher-plugin/blob/669d7ccb7c317ad2ba51a77dd5b2b30c88891d85/src/main/java/hudson/slaves/CommandLauncher.java#L170 I think; presumably less common than SSHLauncher.

perhaps we could create a Unix domain socket on the controller […]

I would not advise spending so much time on this, especially not if it winds up making agent launchers harder to understand. I think we could simply delete StandardOutputSwapper and deprecate or delete StandardOutputStream with few real consequences. Launcher already calls System.setOut which ought to redirect any output from Java code, so the only problem would be stray text sent to stdout by native code—which we would really like to get rid of from the agent JVM anyway.

Base64-encoding the data would decrease efficiency

FWIW I did not suggest that. Would suffice for each packet-ish to emit some magic byte sequence, then packet length, then raw (binary) packet, then newline. The reader can then consume data starting with the magic sequence up the length (including any interior 0x0A) and discard other lines that look like text. Whether that is practical as a Mode depends on whether the mode translation stream (analogous to BinarySafeStream) is being wrapped around a reasonably buffered output, or whether Command.writeTo is going to be writing one byte at a time. ClassicCommandTransport is actually not ideal for this purpose, but in fact that transport is not used any more since the two sides would negotiate up to ChunkedCommandTransport. The related classes (esp. ChunkHeader actually already implement most of what we need and it might just be possible to make ChunkedInputStream.readHeader tolerate textual junk. But again, probably not worth the effort.

I would not advise spending so much time on this, especially not if it winds up making agent launchers harder to understand.

I do not see why it would make agent launchers harder to understand, but I do not plan on spending any time on it at all.

Am I the only one bothered by the legacy "slave" references all over the place?

dborin - I think so. Not an issue for me at least.