BUGS

* The ControlMaster protocol code doesn't deal with STDIN very well.

If you use the ControlMaster socket to request a shell, you end up with your
terminal being in a non-sane state. You can run 'stty sane' or 'reset' to
bring you to sanity.

Using 'scp' with the ControlMaster will actually transfer the file, but when
you're done you have to hit ^C in order to exit.

Requesting a TTY works as you'd expect, you can do: 'ssh -t somehost top'
and it will work.

In most of these cases though, once you drop back to you shell, the first
character you type at a prompt will get eaten by the io.Copy() code in
controlmux.go. This is annoying. I try to make this less annoying by only
dealing with STDIN if the user requests a shell, tries to scp, or requests
a TTY.

* GNU Readline is not thread-safe

Only the first session to SSH into MetaSSH gets to experience the glory of GNU
Readline.  Everybody else gets some off-brand readline, because all hell breaks
loose if I try to access the C readline library from multiple goroutines.

* SSH Server feature doesn't require authentication

Considering this thing lets you run commands on a bunch of hosts, maybe it should
require some authentication. Or if we want to keep it auth-less it should at the
very least not listen on 2222 but have its own ControlMaster socket in the private
~/.ssh area that allows you to talk to it.

* Assumes single user/auth method works for all hosts

MetaSSH makes the bogus assumption that every host it needs to SSH into uses
the same auth credentials. It should read usernames and keys from the JSON
blob, and prompt the user for each private key it may need to unlock before
starting the server.

* Connect command wrong when there are zeo targets

If you say 'connect' without first having picked any targets, the command
happily tells you that it connected to one host, when in fact it hasn't.

* Quantiles can vary wildly from run to run

The docs for the quantile library state:
"Package quantile computes approximate quantiles over an unbounded data stream
within low memory and CPU bounds. A small amount of accuracy is traded to
achieve the above properties."

I'm not sure if we're seeing that "small amount of accuracy" or if I'm just
using the library wrong. I'm betting there is a bug here somewhere.

* ControlMaster protocol implementation is half-assed

1) It has a hardcoded naming scheme for the sockets.
2) Protocol version numbers are not checked.
3) MuxCAliveCheck should actually verify the backend is alive.

* Quantiles seem very broken when concurrency is limited to 128 (agent)

* The run command seemingly runs without any run arguments.

* When a timeout for a run command is exceeded, it should kill the process.