The `readline.set_completer_delims` doesn't play well with others because
it assumes that only it ever allocates or modifies the
rl_completer_word_break_characters buffer.  If other programs modify this
value, for example changing it from heap allocated space to stack
allocated space, the results can be catastrophic.

To remind you, the function essentially works as:

    set_completer_delims(PyObject *self, PyObject *args)
    {
        // ...
        free((void*) rl_completer_word_break_characters;
        rl_completer_word_break_characters = strdup(break_chars);
        // ...
    }

where `break_chars` is the user provided string.

Take, for example, R as another programs which changes the readline
completer strings.  When an embedded R instance is initialized (say, using
`r2py`) something similar to the following takes place::

    static void
    set_rl_completer_word_break_characters(const char *new)
    {
        static char[201] buffer;
        strncpy(buffer, new, 200);
        rl_completer_word_break_characters = buffer;
    }

    static void
    initialize_embedded_R(...)
    {
        // ...
        set_rl_completer_word_break_characters(...);
    }

As you might expect the next trip through `readline.set_completer_delims`
after initializing R will be catastrophic when we attempt to free a stack
allocate buffer.

I think we should consider modifying the `readline.set_completer_delims`
to store the allocated buffers in the module state::

    set_completer_delims(PyObject *self, PyObject *args)
    {
        // ...
        free(_readlinestate_global->break_chars);
        rl_completer_word_break_characters = strdup(break_chars);
        _readlinestate_global->break_chars = rl_completer_word_break_characters;
        // ...
    }

This would prevent the segfault and memory leaks, and would render weird
hacks (like https://bitbucket.org/lgautier/rpy2/commits/408bae913653 in
the r2py code) unnecessary.

Thanks for reporting this. Do you want to contribute a proper patch? You'll find instructions at http://docs.python.org/devguide/

Patch attached. I implemented this by adding a 'static char *' which
holds the memory we allocate. I did not use the PyState machinery.

New changeset 55c7295aca6c by Antoine Pitrou in branch '2.7':
Issue #17289: The readline module now plays nicer with external modules or applications changing the rl_completer_word_break_characters global variable.
http://hg.python.org/cpython/rev/55c7295aca6c

New changeset df0afd3ebb70 by Antoine Pitrou in branch '3.3':
Issue #17289: The readline module now plays nicer with external modules or applications changing the rl_completer_word_break_characters global variable.
http://hg.python.org/cpython/rev/df0afd3ebb70

New changeset 0f65426009e2 by Antoine Pitrou in branch 'default':
Issue #17289: The readline module now plays nicer with external modules or applications changing the rl_completer_word_break_characters global variable.
http://hg.python.org/cpython/rev/0f65426009e2

Thanks for the patch! I made a variable name a bit shorter and also added some error checking on the strdup() result.