Even if this library can't be used for everything, it will allow a whole lot more use cases that are currently possible.
The main issue is that Github do not have CORS support for http access to repos (though they do support CORS for access to repos via their proprietary API) and I don't think the other big public hosters like Bitbucket or Google Code support CORS either.
On the subject of big pack files, sure you'd want to stream them to storage instead of trying to store in memory (for instance JGit has this as an option) but I agree Tim that to start with, even supporting small-medium size repos would be a huge win.
(Essentially it's a protocol based on fast-import to allow access a remote repository without cloning or receiving a pack.)
Also, there are proxies for converting TCP into websocket that should work with the Git protocol, and at least Chrome has some kind of native TCP socket from extensions.
[1]: https://github.com/e-butik/Gittern/
However, there are technical considerations. For example, pack files can get pretty large (indeed version 2 packfiles supports files > 4 GB), and I doubt browsers are optimized for random access to large files (so you wouldn't have to load the entire file in memory). Another potential problem would be pushing and pulling. The dumb HTTP transport is obviously possible to use (but rarely supported anymore), and the smart HTTP transport should be able to get working, but for implementing the git protocol, or git+ssh you'd need regular sockets, which AFAIK browsers don't provide.
I do however wish the author the best of luck with this project, and since I see that he's here, I will leave with links to a few resources I found useful:
* http://git-scm.com/book/en/Git-Internals
* http://opensource.apple.com/source/Git/Git-26/src/git-htmldo...
* http://www.opensource.apple.com/source/Git/Git-17/src/git-ht...
* http://git.rsbx.net/Documents/Git_Data_Formats.txt