<hr /><i>anonymous claiming to be Joerg Sonnenberger added on 2010-12-24 23:55:38:</i><br />
I think I found the root of this problem. fossil is using ssh -T to connect to the server. This inhibits the creation of a pty and therefore sh is started with a pipe for stdin/stdout/stderr. This in turn makes libc use buffering. The commands are sent to this shell and the following POST request is obtained by sh too (due to the buffering). As a result, the remote fossil is waiting in read for data that will never arrive.

I think the best approach for this is to not do fancy things involving the shell, but switch to something like:

<verbatim>
ssh -e none -T <host> <fossil> ssh-sync <repo-path>
</verbatim>

ssh-sync differs from test-http by looping on stdin until it is closed. E.g. use ssh like the equivalent of keep-alive with HTTP.

Advantage: much less chance for a shell to mock things up. It can be easily changed to make fossil an ssh subsystem, which might be a nice option for deployment on central servers, where normal users shouldn't have shell access.