+I think this branch is now enough to be useful. It adds the following:
+
+If the `transient` plugin is loaded, `$srcdir/.ikiwiki/transient` is added
+as an underlay. I'm not sure whether this should be a plugin or core, so
+I erred on the side of more plugins; I think it's "on the edge of the core",
+like goto.
+
+Pages in the transient underlay are automatically
+deleted if a page of the same name is created in the srcdir (or an underlay
+closer to the srcdir in stacking order).
+
+With the additional `ready/transient-tag` branch,
+`tag` enables `transient`, and if `tag_autocreate_commit` is set to 0
+(default 1), autocreated tags are written to the transient underlay.
+There is a regression test.
+
+With the additional `transient-autoindex` branch,
+`autoindex` uses autofiles. It also enables `transient`, and if
+`autoindex_commit` is set to 0 (default 1), autoindexes are written to
+the transient underlay. There is a regression test. However, this branch
+is blocked by working out what the desired behaviour is, on
+[[todo/autoindex_should_use_add__95__autofile]].
+
+> I wonder why this needs to be configurable? I suppose that gets back to
+> whether it makes sense to check these files in or not. The benefits of
+> checking them in:
+>
+> * You can edit them from the VCS, don't have to go into the web
+> interface. Of course, files from the underlays have a similar issue,
+> but does it make sense to make that wart larger?
+> * You can know you can build the same site with nothing missing
+> even if you don't there enable autoindex or whatever. (Edge case.)
+
+>> I'm not sure that that's a huge wart; you can always "edit by
+>> overwriting". If you're running a local clone of the wiki on your laptop
+>> or whatever, you have the underlays already, and can copy from there.
+>> Tag and autoindex pages have rather simple source code anyway. --s
+
+> The benefit of using transient pages seems to just be avoiding commit
+> clutter? For files that are never committed, transient pages are a clear
+> win, but I wonder if adding configuration clutter just to avoid some
+> commit clutter is really worth it.
+
+>> According to the last section of
+>> [[todo/auto-create_tag_pages_according_to_a_template]], [[chrysn]] and
+>> Eric both feel rather strongly that it should be possible to
+>> not commit any tags; in [[plugins/autoindex/discussion]],
+>> lollipopman and [[JoeRayhawk]] both requested the same for autoindex.
+>> I made it configurable because, as you point out,
+>> there are also reasons why it makes sense to check these
+>> automatically-created files in. I'm neutral on this, personally.
+>>
+>> If this is a point of contention, would you accept a branch that
+>> just adds `transient` and uses it for [[plugins/recentchanges]],
+>> which aren't checked in and never have been? I've split the
+>> branch up in the hope that *some* of it can get merged.
+>>
+>>> I will be happy to merge transient-recentchanges when it's ready.
+>>> I see no obstacle to merging transient-tag either, and am not
+>>> really against using it for autoindex or aggregate either
+>>> once they get completed.
+>>> I just wanted to think through why configurability is needed.
+>>> --[[Joey]]
+>>
+>> One potentially relevant point is that configuration clutter only
+>> affects the site admin whereas commit clutter is part of the whole
+>> wiki's history. --[[smcv]]
+
+> Anyway, the configurability
+> appears subtly broken; the default is only 1 if a new setup file is
+> generated. (Correction: It was not even the default then --[[Joey]])
+> With an existing setup file, the 'default' values in
+> `getsetup` don't take effect, so it will default to undef, which
+> is treated the same as 0. --[[Joey]]
+
+>> Fixed in the branches, hopefully. (How disruptive would it be to have
+>> defaults take effect whenever the setup file doesn't set a value, btw?
+>> It seems pretty astonishing to have them work as they do at the moment.) --s
+
+>>> Well, note that default is not actually a documented field in
+>>> getsetup hooks at all! (It is used in IkiWiki.pm's own `getsetup()`, and
+>>> the concept may have leaked out into one or two plugins (comments,
+>>> transient)).
+>>>
+>>> Running getsetup at plugin load time is something I have considered
+>>> doing. It would simplify some checkconfig hooks that just set hardcoded
+>>> defaults. Although since dying is part of the getsetup hook's API, it
+>>> could be problimaric.
+>>> --[[Joey]]
+
+autoindex ignores pages in the transient underlay when deciding whether
+to generate an index.
+
+With the additional `ready/transient-recentchanges` branch, new recent
+changes go in the transient underlay; I tested this manually.
+
+Not done yet (in that branch, at least):
+
+* `remove` can't remove transient pages: this turns out to be harder than
+ I'd hoped, because I don't want to introduce a vulnerability in the
+ non-regular-file detection, so I'd rather defer that.
+
+ > Hmm, I'd at least want that to be dealt with before this was used
+ > by default for autoindex or tag. --[[Joey]]
+
+ >> I'll try to work out which of the checks are required for security
+ >> and which are just nice-to-have, but I'd appreciate any pointers
+ >> you could give. Note that my branch wasn't meant to enable either
+ >> by default, and now hopefully doesn't. --[[smcv]]
+
+* Transient tags that don't match any pages aren't deleted: I'm not sure
+ that that's a good idea anyway, though. Similarly, transient autoindexes
+ of directories that become empty aren't deleted.
+
+ > Doesn't seem necessary, or really desirable to do that. --[[Joey]]
+
+ >> Good, that was my inclination too. --s
+
+* In my `untested/transient` branch, new aggregated files go in the
+ transient underlay too (they'll naturally migrate over time). I haven't
+ tested this yet, it's just a proof-of-concept.
+
+> I can confirm that the behavior of autoindex, at least, is excellent.
+> Haven't tried tag. Joey, can you merge transient and autoindex? --JoeRayhawk
+
+>> Here are some other things I'd like to think about first: --[[Joey]]
+>>
+>> * There's a FIXME in autoindex.
+>>
+>> > Right, the extra logic for preventing autoindex pages from being
+>> > re-created. This is taking a while, so I'm going to leave out the
+>> > autoindex part for the moment. The FIXME is only relevant
+>> > because I tried to solve
+>> > [[todo/autoindex should use add__95__autofile]] first, but
+>> > strictly speaking, that's an orthogonal change. --s
+
+>> * Suggest making recentchanges unlink the transient page
+>> first, and only unlink from the old location if it wasn't
+>> in the transient location. Ok, it only saves 1 syscall :)
+>>
+>> > Is an unlink() really that expensive? But, OK, fixed in the
+>> > `ready/transient-recentchanges` branch. --s
+
+>> >> It's not, but it's easy. :) --[[Joey]]
+
+>> * Similarly it's a bit worrying for performance that it
+>> needs to pull in and use `Cwd` on every ikiwiki startup now.
+>> I really don't see the need; `wikistatedir` should
+>> mostly be absolute, and ikiwiki should not chdir in ways
+>> that break it anyway.
+>>
+>> > The reason to make it absolute is that relative underlays
+>> > are interpreted as relative to the base underlay directory,
+>> > not the cwd, by `add_underlay`.
+>> >
+>> > The updated `ready/transient-only` branch only loads `Cwd` if
+>> > the path is relative; an extra commit on branch
+>> > `smcv/transient-relative` goes behind `add_underlay`'s
+>> > back to allow use of a cwd-relative underlay. Which direction
+>> > would you prefer?
+>> >
+>> > I note in passing that [[plugins/autoindex]] and `IkiWiki::Render`
+>> > both need to use `Cwd` and `File::Find` on every refresh, so
+>> > there's only any point in avoiding `Cwd` for runs that don't
+>> > actually refresh, like simple uses of the CGI. --s
+
+>> >> Oh, right, I'd forgotten about the horrificness of File::Find
+>> >> that required a chdir for security. Ugh. Can we just avoid
+>> >> it for those simple cases then? (demand-calculate wikistatedir)
+>> >> --[[Joey]]
+
+>> * Unsure about the use of `default_pageext` in the `change`
+>> hook. Is everything in the transientdir really going
+>> to use that pageext? Would it be better to look up the
+>> complete source filename?
+>>
+>> > I've updated `ready/transient` to do a more thorough GC by
+>> > using File::Find on the transient directory. This does
+>> > require `File::Find` and `Cwd`, but only when pages change,
+>> > and `refresh` loads both of those in that situation anyway.
+>> >
+>> > At the moment everything in the transientdir will either
+>> > have the `default_pageext` or be internal, although I
+>> > did wonder whether to make [[plugins/contrib/album]]
+>> > viewer pages optionally be `html`, for better performance
+>> > when there's a very large number of photos. --s
+
+>> >> Oh, ugh, more File::Find... Couldn't it just assume that the
+>> >> transient page has the same extension as its replacement?
+>> >> --[[Joey]]
+
+--------------------------
+
+## An earlier version
+