I am curious, how do you handle latency issues for on demand access? I saw you use FUSE (and FSKit), and from my experience it is pain to make filesystems in userspace work on-demand over WAN because a) latency, and variable RTT; and b) you can't saturate the wire and aggressively read ahead things, otherwise native apps will freeze, lag, or just make the UX unpleasant, especially if there are too many placeholder files, or large files with random jumps in them.
I think what makes FUSE/FSKit great here is that agents usually only need to see the file metadata + read a handful of actual files, rather than some applications needing read many things. If you're doing huge rewrites, this is a problem, but most tasks are usually somewhat small. Definitely is a problem that I've ran into though, we do cache aggressively to try to solve some of this, but it'll never be as fast as reading/writing directly to disk. We have benchmarks [1] if you want to take a look at how we're testing some of the performance there.
[1] https://oak.space/oak/benchmarks
Couldn't run your benchmarks, as I did not create an account, and a bit of a different beast that I am comparing against (P2P distributed filesystem), but these are my numbers and setup, and the on-demand part lines up with what I have observed:
Setup: a Linux box on the other side of Romania (compared to where I am living) reading from a Windows box in Singapore (~200 ms RTT)
- reading 1 MiB of a 1 GB remote file pulls only 16 MiB (~98% avoided) - this is because of my fine tuning optimization choice - first byte approx: 2.3s - git-LFS repos also clone cold over the mount byte-perfect (separate Mac - Linux run on a ~20 ms RTT)
The thing that I do differently is that my metadata is eagerly pushed, as I optimized for content streaming.
And 100k-file tree mounts I did not test yet.
But my goal was to have instant file access for generic files between apps, and peer to peer, supports also Windows :D
here is the tool: https://github.com/KeibiSoft/KeibiDrop
That unfortunately doesn’t match my experience at all. My Claude often runs rg in the repo attempting to find things that need to be changed. And of course Claude still needs to invoke the build tool to ensure the change can be compiled, which necessarily involves reading almost every single file at least for a fresh checkout? Or did you envision the build tool being completely remote?