Mapping one chat API onto another

Every chat API expresses the same handful of ideas: a transcript, an instruction that sits above the transcript, a ceiling on how much comes back, a description of the tools the model may call, and a stream of fragments on the way out. Because the ideas are the same, moving between two of them looks like a renaming exercise, and for about eighty per cent of the surface it is. The interesting part is the other twenty.

These pages walk one concept at a time across the shapes each provider uses, and they spend most of their length on the part that does not survive the translation — the field with no counterpart, the value that means something subtly different on the other side, the default that differs, the signal one API gives you and the other does not. If you are writing the adapter by hand, that residue is the entire job; the renames write themselves.