Working on building [[design/metadata]] filtered branches. Spent most of the day on types and pure code. Finally at the end I wrote down two actions that I still need to implement to make it all work: [[!format haskell """ applyView' :: MkFileView -> View -> Annex Git.Branch updateView :: View -> Git.Ref -> Git.Ref -> Annex Git.Branch """]] I know how to implement these, more or less. And in most cases they will be pretty fast. The more interesting part is already done. That was the issue of how to generate filenames in the filter branches. That depends on the `View` being used to filter and organize the branch, but also on the original filename used in the reference branch. Each filter branch has a reference branch (such as "master"), and displays a filtered and metadata-driven reorganized tree of files from its reference branch. [[!format haskell """ fileViews :: View -> (FilePath -> FileView) -> FilePath -> MetaData -> Maybe [FileView] """]] So, a view that matches files tagged "haskell" or "git-annex" and with an author of "J\*" will generate filenames like "haskell/Joachim/interesting_theoretical_talk.ogg" and "git-annex/Joey/mytalk.ogg". It can also work backwards from these filenames to derive the MetaData that is encoded in them. [[!format haskell """ fromView :: View -> FileView -> MetaData """]] So, copying a file to "haskell/Joey/mytalk.ogg" lets it know that it's gained a "haskell" tag. I knew I was on the right track when `fromView` turned out to be only 6 lines of code! The trickiest part of all this, which I spent most of yesterday thinking about, is what to do if the master branch has files in subdirectories. It probably does not makes sense to retain that hierarchical directory structure in the filtered branch, because we instead have a non-hierarchical metadata structure to express. (And there would probably be a lot of deep directory structures containing only one file.) But throwing away the subdirectory information entirely means that two files with the same basename and same metadata would have colliding names. I eventually decided to embed the subdirectory information into the filenames used on the filter branch. Currently that is done by converting `dir/subdir/file.foo` to `file(dir)(subdir).foo`. We'll see how this works out in practice..