Re: [PATCH 1/3] connect: let write_command_and_capabilities() send a promisor-remote reply
- From
Christian Couder <christian.couder@gmail.com>
- Date
- Nov 13, 2026, 13:10 UTC
- Message-ID
- <CAP8UFD18rTvM0mu9oSPMd3V+J9GBgc07RM91Pfh4RJqZg0s9aw@mail.gmail.com>
- In-Reply-To
- <xmqq4idlhe69.fsf@gitster.g>
On Thu, Nov 12, 2026 at 11:34 PM Junio C Hamano <gitster@pobox.com> wrote:
Show 32 quoted lines
>
> Junio C Hamano <gitster@pobox.com> writes:
>
> > Christian Couder <christian.couder@gmail.com> writes:
> >
> >> void write_command_and_capabilities(struct strbuf *req_buf, const char *command,
> >> - const struct string_list *server_options)
> >> + const struct string_list *server_options,
> >> + const char *promisor_remote_reply)
> >> {
> >> const char *hash_name;
> >> int advertise_sid = 0;
> >> @@ -754,6 +755,9 @@ void write_command_and_capabilities(struct strbuf *req_buf, const char *command,
> >> die(_("the server does not support algorithm '%s'"),
> >> the_hash_algo->name);
> >> }
> >> + if (promisor_remote_reply)
> >> + packet_buf_write(req_buf, "promisor-remote=%s",
> >> + promisor_remote_reply);
> >> packet_buf_delim(req_buf);
> >> }
> >
> > This function was designed to have a fairly generic "here we send a
> > command and then capability exchange based on the settings stored in
> > the string list" interface for extensibility. It is highly annoying
> > that we are now adding a specific option here to be sent outside the
> > framework.
> >
> > Isn't the misdesign being corrected that the "promisor-remote=..."
> > string was not sent as part of these server options list of strings?
> > Can we fix it without contaminating the API like this patch does, or
> > is it too late?I think that if we wanted to clean this up, it would need a significant refactor as there are three separate places in "connect.c" that write the client's protocol v2 capabilities:
- send_capabilities() used for `bundle-uri`, - get_remote_refs() (using send_capabilities() and its own loop) used for `ls-refs`, - write_command_and_capabilities() used for `fetch` and `object-info`.
They duplicate code for sending "server-option", "agent" and "object-format", while each sends a different set of capabilities.
The server side already has a table of capabilities (`capabilities[]` in "serve.c") with generic advertise/receive handling. The client side could have something similar for building requests.
Show 11 quoted lines
> By this I do not mean that promisor-remote=... has to be sent as one > of the server options (i.e., making the string sent to something > silly like "server-option=promisor-remote=..."). Rather, I am > wondering if the existing server_options string list should be made > into a more generalized form so that the string list can contain > elements whose keys other than "server-option" can also be > contained (like "promisor-remote"). IOW, the caller can decide to > put "promisor-remote=...promisor remote reply..." in the list, > instead of sending it as a totally separate parameter, whose > semantics is hardcoded (i.e. the fourth parameter to this function > will be sent as a value to "promistor-remote=...").
I am willing to work on a significant refactoring along what you suggest, but I think it should be a separate series from this bug fix series, especially as we might want this fix to also go to maint.
In the meantime, I can send a v2 of this series with something like a NEEDSWORK comment, for example one in front of write_command_and_capabilities() saying that it should be refactored with send_capabilities() and get_remote_refs(), so that callers build the full list of capabilities, which would remove the `promisor_remote_reply` parameter