
 * More testing
 * Write test scripts

 * Complete LWP5 compatibility
 * Finish the authentication hooks (digest)
 * LWP::Conn::HTTP improvements
     - bundling of multiple request in a single syswrite() when
       pipelining is likely to improve performance.
     - non-blocking gethostbyname (Net::DNS?)
     - minimize headers sent for second and following requests.
     - know something about idempotent methods
     - always pass response data through sinks (?)
     - support HTTP/1.0 Keep-Alive (?) (make more sense for a server)
     - try to record HTTP version number supported by the server
       in the LWP::Server object (or URI::Attr).
 * Make a single request able to override persistent connection
   behaviour (make a request always processed on its own connection)
 * LWP::Conn::FILE could support PUT and Content-Ranges
 * Other LWP::Conn implementations: FTP, NNTP, Gopher, MAILTO
 * Support HTTPS, TLS.  This might need restructuring of the LWP::Conn:HTTP.
   Perhaps the socket should be a member and not the connection object
   itself.
 * gzip/compress sinks
 * deflate parameters
 * tuning - make things go faster
 * documentation - as always

 * Delayed spooling (for robots)
 * RobotUA
 * Caching support
 * Some applications

 * Move to same eventloop as Tk uses.  There appears to be a problem because
   the Tk fileevent does not allow a socket to be both readable and writable
   at the same time.  Grahams Event.pm does not work with current IO modules
   so I have not tried to use it yet.  Compatibility with Win32 is also a
   concern.
   
