Re: git-am applies commit message diffs

From: Patrick Steinhardt <ps@pks.im>
Date: Feb 10, 2026, 14:22 UTC
Message-ID: <aYs_P8QujA6mL81-@pks.im>
In-Reply-To: <CA+P7+xrNycJHTyJwn9AQcJLG0dDAE7KrTvWTHBi+CiQUqK8p5A@mail.gmail.com>

On Mon, Feb 09, 2026 at 06:16:35PM -0800, Jacob Keller wrote:

Show 50 quoted lines

> On Mon, Feb 9, 2026 at 7:59 AM Patrick Steinhardt <ps@pks.im> wrote:
> >
> > On Fri, Feb 06, 2026 at 04:03:58AM -0500, Jeff King wrote:
> > > On Fri, Feb 06, 2026 at 09:18:50AM +0100, Matthias Beyer wrote:
> > >
> > > > That said, I am no expert in either C or the git codebase at all, but
> > > > from what I saw from reading the git-am codebase, it looks like it tries
> > > > to find the patch by looking for three dashes on a line with a linebreak
> > > > behind ("---\n").
> > >
> > > Yes, that is how the split is made.
> > >
> > > > From what I read, it looks for that from the first line.
> > > > What I would think of here is looking for that "patchbreak" from the
> > > > _end_ of the email rather than from the top, that would have prevented
> > > > this issue, right?
> > >
> > > The patch itself may legitimately contain "---" on a line by itself (it
> > > would indicate that the line "--" was removed from a file). That would
> > > confuse your parser, including in a way that we end up only applying
> > > part of the diff (everything before that fake "---" becomes commit
> > > message, and everything after becomes cover-letter material up to the
> > > next "diff" line).
> > >
> > > I suspect it also creates corner cases with cover-letter material
> > > (between the "---" and the diff itself) that itself contains any "---"
> > > marker.
> > >
> > > I don't think there is a way to unambiguously parse the single-stream
> > > output that format-patch produces. This is a reasonably well-known
> > > gotcha (at least around here). E.g., some earlier discussions:
> > >
> > >   2024: https://lore.kernel.org/git/ca13705ae4817ffba16f97530637411b59c9eb19.camel@scientia.org/
> > >   2022: https://lore.kernel.org/git/d0b577825124ac684ab304d3a1395f3d2d0708e8.1662333027.git.matheus.bernardino@usp.br/
> > >   2015: https://lore.kernel.org/git/CAFOYHZC6Qd9wkoWPcTJDxAs9u=FGpHQTkjE-guhwkya0DRVA6g@mail.gmail.com/
> > >
> > > There are probably more, but it's actually a tricky thing to search for
> > > in the archive, so I stopped digging. ;)
> >
> > Maybe we can't parse it unambiguously. But what we _can_ detect is that
> > a patch is ambiguous in the first place, right? So maybe we could extend
> > git-am(1) to bail by default with a hint that tells the user that:
> >
> 
> I think it might make sense in a breaking change to update format
> patch and git am to have an "unambiguous" mode which would allow
> somehow to unambiguously distinguish between commit message contents
> and patch data. I'm not 100% sure how to do this, and it likely
> requires some sort of breaking changes to both tools to allow
> distinguishing properly between the two points.

That is worth a thought indeed. I guess one of the biggest questions here is whether we can introduce such an unambiguous mode in such a way that old Git clients/patch(1) would continue to understand them. I wouldn't mind much if they would still misinterpret the ambiguous parts. But if so, we could make this unambiguous mode the default without a breaking change.

This is all pure speculation though, I have no idea whether such a backwards-compatible and forwards-safe mode exists.

Show 5 quoted lines

> Obviously if you're sending the contents together, a malicious user
> could edit the formatted patch to move or copy whatever the
> "signifier" for patch vs commit separator is... but at least we'd
> prevent the cases where someone accidentally includes diffs without
> intending to.

Well, if we had such an unambiguous mode I would say that eventually, Git should start to refuse patches that have been generated without this mode by default.

Patrick