Minimal example:

>>> import os
>>> somekey = 'random'
>>> try:
...   os.environ[somekey]
... except KeyError as e:
...   print(repr(e))
...   somekey == e.args[0]
... 
KeyError(b'random',)
False

Tested in Python 3.3.1 on Debian

Well, it is not that it is converting it in the exception, it is that it is converting it in order to do the lookup.  It is an interesting question whether or not its value in the exception should be considered a bug.

"It is an interesting question whether or not its value in the exception should be considered a bug."

I consider it as a bug because it is very surprising to have a different key on the error. Attached patch fixes the issue.

Sounds fine to me.

Is there a reason you are not using 'with assertRaises' in the test?

> Is there a reason you are not using 'with assertRaises' in the test?

The test is not testing the exception class, but the first argument of
the exception. How do you test that using assertRaises()? Especially
to check "err.args[0] is missing".

with self.assertRaises(KeyError) as cm:
    os.environ[missing]
self.assertEqual(cm.excecption.args[0], missing)

(I don't know why assertRaises returns itself rather than just returning the exception in the with, but that's the API).

New changeset 72df981e83d3 by Victor Stinner in branch '3.3':
Close #17702: os.environ now raises KeyError with the original environment
http://hg.python.org/cpython/rev/72df981e83d3

New changeset ea54559a4442 by Victor Stinner in branch 'default':
(Merge 3.3) Close #17702: os.environ now raises KeyError with the original
http://hg.python.org/cpython/rev/ea54559a4442

New changeset b0c002fa4335 by Victor Stinner in branch '3.3':
Issue #17702: use assertRaises() for the unit test
http://hg.python.org/cpython/rev/b0c002fa4335

New changeset cc6c5b5ec4f2 by Victor Stinner in branch 'default':
(Merge 3.3) Issue #17702: use assertRaises() for the unit test
http://hg.python.org/cpython/rev/cc6c5b5ec4f2

This patch introduces a bit ugly traceback. Shouldn't there be “raise from None” or “raise from previous_exc” instead of simple raise?

"This patch introduces a bit ugly traceback. Shouldn't there be “raise from None” or “raise from previous_exc” instead of simple raise?"

Oh, I see.

>>> os.environb[b'10']
Traceback (most recent call last):
  File "/home/vstinner/prog/python/default/Lib/os.py", line 648, in __getitem__
    value = self._data[self.encodekey(key)]
KeyError: b'10'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/vstinner/prog/python/default/Lib/os.py", line 651, in __getitem__
    raise KeyError(key)
KeyError: b'10'

Attached patch adds "from None".

It's also in __delitem__.

New changeset 26c049dc1a4a by Victor Stinner in branch '3.3':
Close #17702: On error, os.environb now removes suppress the except context
http://hg.python.org/cpython/rev/26c049dc1a4a

New changeset 01f33959ddf6 by Victor Stinner in branch 'default':
(Merge 3.3) Close #17702: On error, os.environb now removes suppress the except
http://hg.python.org/cpython/rev/01f33959ddf6

> This patch introduces a bit ugly traceback.

It is now fixed, thanks for the report!