The new system is backward compatible and will load older style plugins.
The new plugin framework allows plugins to track their dependencies, and
will auto-enable plugins as needed.
Dependencies are tracked via a class-level set named `dependencies` in
each plugin.
Plugin names are no longer tightly coupled with the plugin class name,
Pso EP8 style class names may be used.
Disabling plugins is now allowed, but ensuring proper cleanup is left to
the plugin implementation.
The use of a `post_init()` method is no longer needed for new style
plugins, but plugins following the old style will still require a
`post_init()` method.
Setting self.reconnect_max_attempts to a non-None value will limit
the number of times a connection attempt will be made before quiting
and raising a 'connection_failed' event.
The post_init() system can only reliably handle a single layer
of dependencies between plugins, but PEP plugins with XEP-0115
exceed that limit and plugins can be post_init'ed out of order. To
resolve this, we will special case XEP-0115 to be post_init'ed
first until the new plugin system with dependency tracking is
stable.
Publishes, retractions, purges, and deletions now raise the events:
- pubsub_publish
- pubsub_retract
- pubsub_purge
- pubsub_delete
In addition, custom events may be raised based on the node that
generated the notification. For example:
xmpp['xep_0060'].map_node_event('http://jabber.org/protocol/tune',
'user_tune')
will allow for using the events:
- user_tune_publish
- user_tune_retract
- user_tune_purge
- user_tune_delete
This was XEP-0237, but is now part of RFC 6121.
Roster backends should now expose two additional methods:
version(jid):
Return the version of the given JID's roster.
set_version(jid, version):
Update the version of the given JID's roster.
A new state field will be passed to the backend if an item
has been marked for removal. This is 'removed' which will
be set to True.
Each state element must have its own stanza class now. A stanza class
with an empty name field causes errors in ElementTree, even though
it works fine with cElementTree.
The payload is a dictionary of parsed cert data, as provided by
Python's getpeercert() socket method. It unfortunately does not
provide much detail beyond basic info.
This is inspired by the version from macdiesel and tomstrummer, but
their version was heavily linked with XEP-0096 and focused solely
on file transfer. This version is a more generic implementation.
Added new example for how to retrieve a Google token, following
the best case, non-browser, workflow. Other thirdparty auth
mechs (Facebook, MSN) follow a similar pattern of using an
access token.
This is mainly just useful for authenticating without using TLS.
If an access token is not provided, an attempt will be made to
retrieve one from Google.
Silently substituting the password field was nice, but for mechs
that can use either the password or an access token, it makes
things very difficult. This really only affects MSN clients since
Facebook clients should already be setting the api key.
This requires an extra credential for SASL authentication:
xmpp = ClientXMPP('user@chat.facebook.com', '...access_token...')
xmpp.credentials['api_key'] = '...api_key...'
The parsing and namespace cleaning isn't terribly expensive, but it does
add up. It was adding an extra 5sec when processing 100,000 basic
message stanzas.
Based on profiling, using around 35 stream handlers quarters the number
of basic message stanzas that can be processed in a second, in
comparison to only using the bare minimum of four handlers.
To help, we can drop handlers for stream features once the session
has started. So that we can re-enable these handlers when a stream
must restart, the 'stream_start' event has been added which fires
whenever a stream header is received.
The 'stream_start' event is a more generic replacement for the
existing start_stream_handler() method.
As part of adding this feature:
- fixed bug in update_caps() not assigning verstrings
- fixed xep_0004 typo
- can now use None as a roster key which will map to boundjid.bare
- fixed using JID objects in disco node handlers
- fixed failing test related to get_roster
Several of these bugs I've fixed before, so I either didn't push them
earlier, or I clobbered something when merging. *shrug*
