Christian Couder's series fixes what he calls a brown paper bag bug in the protocol v2 "promisor-remote" capability, found by Matt Cote during the NVIDIA hackathon after the Gerrit User Summit 2026. The client sends a "promisor-remote=..." reply with the ls-refs and bundle-uri commands, but not with fetch.
Over file:// or ssh:// this works by luck. The same upload-pack process handles every request and remembers the reply from ls-refs. Servers that handle each request on its own, such as Git over HTTP or JGit (and so Gerrit), never see the reply. It also fails when only object IDs are fetched, because then no ls-refs request is made.
The patches say that send_capabilities() is the only place that writes the reply, while fetch requests are built by write_command_and_capabilities() via send_fetch_request() in fetch-pack.c. One patch adds a promisor_remote_reply argument to write_command_and_capabilities(); all callers pass NULL at first, so behaviour does not change. Another makes fetch use it, and a third documents in gitprotocol-v2 that the client should send the reply with fetch too.
Junio Hamano objected to the API change. He said the function was designed with a generic interface, in which a command and its capabilities are sent from settings held in a string list, and that adding a specific option outside that framework is "highly annoying". He asked whether the real misdesign is that the promisor-remote string was never sent as part of the server-options list, and whether it can be fixed without contaminating the API or whether it is too late.
In a follow-up he clarified that he does not mean the string should be sent as a "server-option=promisor-remote=..." value. He suggested generalizing the existing server_options string list so that it can hold elements with keys other than "server-option", and letting the caller put the promisor-remote reply there. The excerpts show no reply from Couder yet, so the design remains open.