As per request on c.l.p:
http://groups.google.de/group/comp.lang.python/browse_thread/thread/26c93fefefd3a100/bf1924ce28fac1ac?hl=de#bf1924ce28fac1ac
I've implemented a small patch to change the output of
repr(x) for complex variables, so that complex(repr(x))
works for any complex x. This changes the output of
repr(x) to
'<r>+<i>j'
without brackets, but leaves the string output
untouched. This change of behaviour would be in line
with int(repr(x)) and float(repr(x)) being defined for
any int or float x, repectively.
I don't know whether this patch is sensible, and
whether it breaks any current code, because (for example)
eval("5*%r" % (1+2j,))
won't work properly anymore, or whether it'd be more
sensible to change the complex constructor to also
accept a bracketed expression. I'll attach a patch to
do the latter later.

Logged In: YES
user_id=791932
The second patch (python-complex-constructor.diff) changes
the constructor to accept bracketed complex numbers which
are enclosed in a single bracket. I'd rather say this is the
better appropach to have complex(repr(x)) work, but I leave
both patches attached to this bug.
The latter patch also creates test cases testing for
formatting errors with bracketed expressions.

Logged In: YES
user_id=593130
(The current behavior is not a bug, nor is this patch
submission a bug report, so let us omit that word.)
I think your example suggests why complexes are printed in
parens, so I think enhancing complex() to accept such is
the better approach if any change is to be made.

Logged In: YES
user_id=791932
The attached patch is a revised version of the patch to the
complex constructor to accept bracketed string expressions,
which also adds documentation changes (whatsnew25).
Anyway, I personally also find this to be the "better" way,
so I've removed the repr-changing patch. And, calling it a
bug was by accident: I should've rather called it tracker
item, which is pretty synonymous for me.

I have applied this patch in 2.6 and it seem to work as intended. The test_complex.py unit tests all pass. I guess the change could be useful if you are are serializing and desearializing lists of complex numbers. I think it should be applied.
    

Works for me. I'll commit this for 2.6 (no backport) if there are no objections.

Applied as r54247. Thanks for your patch!