Thanks to the guys who pointed me towards the XML Pull Parser. It's not quite what I'm after, but it'll provide a good source of material to steal from while writing my own.

Just to state the problem better, imagine a protocol based loosely on Jabber, where each side is sending the other a stream of XML fragments. The parser would be receiving chunks like “<foo:bar baz="quux">Fish</f”, and it'd need to parse as much as it could, trigger any events based on that parsing (in this case an opening tag and a text section), and then save the remainder until another chunk is delivered. The XML Pull Parser wouldn't quite work because you'd call nextTag, and it'd choke on the incomplete end-tag.

Luckily, I can cheat, write a parser that does just enough to divide the incoming stream into separate (but well-formed) XML messages, and farm the rest off to a full-featured parser.

State machines. Joy.