ISTM the simplest approach would be to just set self->args in BaseException.__new__() (like in Georg's patch) but to ignore the possibility that the user might later set self.args to something stupid "wrong":

diff -r 51ac5f06dd04 Objects/exceptions.c
--- a/Objects/exceptions.c      Tue Jul 24 03:45:39 2012 -0700
+++ b/Objects/exceptions.c      Tue Jul 24 22:12:49 2012 +0100
@@ -44,12 +44,17 @@
     self->traceback = self->cause = self->context = NULL;
     self->suppress_context = 0;

-    self->args = PyTuple_New(0);
-    if (!self->args) {
-        Py_DECREF(self);
-        return NULL;
+    if (!args) {
+        args = PyTuple_New(0);
+        if (!args) {
+            Py_DECREF(self);
+            return NULL;
+        }
+    } else {
+        Py_INCREF(args);
     }

+    self->args = args;
     return (PyObject *)self;
 }

Certainly it will not work for all cases (like calling a base classes' __init__ with different arguments), but it does cover the *very* common case where __init__() is defined but does not call the base classes' __init__().

Such a patch is minimally invasive and, as far as I can see, would not break currently working code.

Would this be acceptable for a bugfix release?