by specifying undef as manager param.
This bit me earlier this week when I wasn't paying enough attention and included some logic in my call() method *after* calling response_cb(). There are other explicit returns in the code samples here. I think it wouldn't hurt to make it obvious that response_cb()'s return value needs to be returned by the Middleware.
If this option isn't set, references to function arguments in the stack trace are retained. This can lead to destructors being called twice if an exception is thrown from a destructor, like in DBIx::Class::Storage::TxnScopeGuard, for example.
When multiple exceptions are thrown, store all stack traces either by refaddr or string value. Then try to find the trace matching the caught exception. Make sure to only store the first stack trace for each exception so that rethrown exceptions are ignored.
The content filter in ErrorDocument never set the $done variable to a true value in the non-subrequest case. As a result, it kept returning empty strings at the end of input, not undef like it should. This is actually harmless because response_cb copes with this behavior.
The Catalyst test body_fh.t was exploding because it gets this 'can' method on it's object. The latest Test::Builder alpha also calls ->can(...) on test packages, in this case causing an error since the package is not a hashref, but instead is just a string. This will do the expected behavior for plack, but not explode if someone calls 'can' on the package.