@hsbt I would say we are just doing the wrong thing here. We will fix it.

I'm not sure this is correct behavior in MRI, and I'll explain why.

In the differing case, a is not caused by b. They appear to be setting cause every time the exception is raised, which I'd consider a bug...since it would lose the original cause. An example of where this is broken:

begin
  begin
    raise 'a'
  rescue => a
    begin
      raise 'b'
    rescue => b
      p [b, b.cause]
      begin
        raise 'c'
      rescue
        raise b
      end
    end
  end
rescue
  p [$!, $!.cause]
end

And the output in JRuby 9.0.5.0 and MRI 2.3.0:

[] ~/projects/jruby $ rvm jruby-9.0.5.0 do ruby cause.rb
[#<RuntimeError: b>, #<RuntimeError: a>]
[#<RuntimeError: b>, #<RuntimeError: a>]

[] ~/projects/jruby $ ruby23 cause.rb
[#<RuntimeError: b>, #<RuntimeError: a>]
[#<RuntimeError: b>, #<RuntimeError: c>]

MRI is losing the original cause of the exception, without it being physically replaced.

I'm going to report this as a bug.

Interestingly, JRuby 9.1 (ruby-2.3 branch) appears to work like Ruby 2.3, so somewhere along the way we "fixed" this:

[] ~/projects/jruby $ jruby -v cause.rb
jruby 9.1.0.0-SNAPSHOT (2.3.0) 2016-02-13 8136dd9 Java HotSpot(TM) 64-Bit Server VM 25.60-b23 on 1.8.0_60-b27 +jit [darwin-x86_64]
[#<RuntimeError: b>, #<RuntimeError: a>]
[#<RuntimeError: b>, #<RuntimeError: c>]

Filed https://bugs.ruby-lang.org/issues/12068

If they disagree (and I will argue the point) then I guess we've already fixed this. If they agree, we'll need to reinstate 9.0.5.0's behavior.

They agreed, so we need to fix 9.1 to work like 9.0 did.