How Korandje made "with" agree it-with its subject

Korandje, the language of Tabelbala in southwestern Algeria, requires the comitative preposition "with" to agree in person and number, not with its object, but with its subject (strictly speaking, with its external argument):

ʕa-ddər ʕ-indza xaləd, I-went I-with Khaled.
nə-ddər n-indza xaləd, you-went you-with Khaled.

This seems to be vanishingly rare worldwide. The nearest parallels I have encountered are ones in which the comitative is expressed using a serial verb, but a closer look at the syntax and morphology of Korandje shows that indza is indeed a preposition, not a verb or a noun. Perhaps most strikingly, when you relativise on its object, you pied-pipe not only the preposition but the agreement marker on it too:

ʕan bạ-yu ʕ-indz uɣudz əgga ʕa-b-yəxdəm
my friend-s I-with whom PAST I-IMPF-work
"my friends with whom I was working"

Its historical source, proto-Songhay *ndá "with, and, if", was also a preposition, and did not display agreement. Comparative data makes it possible to reconstruct how this change took place: it developed out of a strategy, common in Berber and found in some Songhay languages, of expressing "I went with Khaled" as "I went, I and Khaled", which seems to be the result of reinterpretation of a postverbal subject as part of the adjacent comitative phrase. This development in turn provides the first attested way to reverse the well-known grammaticalisation chain "with" > "and". If you want to know more, read my article, which has just been published:

"How to make a comitative preposition agree it-with its external argument: Songhay and the typology of conjunction and agreement". In Paul Widmer, Jürg Fleischer, and Elisabeth Rieken (eds.), Agreement from a diachronic perspective, Berlin: De Gruyter, pp. 75-100, 2015. (offprints available on request - just email me.)

Here's the abstract:

This article describes two hitherto unreported comitative strategies exemplified in Songhay languages of West Africa – external agreement, and bipartite – and demonstrates their wider applicability. The former strategy provides the first clear-cut example of a previously unattested agreement target-controller pair. Based on comparative evidence, this article proposes a scenario for how these could have developed from the typologically unremarkable comitative and coordinative strategies reconstructible for proto-Songhay, in a process facilitated by contact with Berber. The grammaticalisation chain required to explain this has the unexpected effect of reversing a much better-known one previously claimed to be unidirectional, the development COMITATIVE > NP-AND.