git status shows all files as modified on fresh blobless-clone mount

[ThoolooExpress]

After following the Quick Start instructions (add-repo + daemon), git status inside the mount shows a lot of files as having been modified. modified. git diff shows no actual content differences for most of these files, and git update-index --refresh fixes everything.

Reproduced against cloudflare/workers-sdk:

$ git -C /tmp/workers-sdk status | grep -c "modified:"
3954
$ git -C /tmp/workers-sdk diff | wc -l
0

Root cause

Seems to be a combination of 3 separate issues:

1. batchResolveSizes silently drops all missing OIDs, leaving most files with size=0

gitstore/gitstore.go:177–219 — batchResolveSizes resolves blob sizes by running git cat-file --batch-check with GIT_NO_LAZY_FETCH=1. On a --filter=blob:none clone the server sends no blob content, so virtually every OID is missing from the local pack. The "missing" lines in the output (<oid> missing) have only two fields; the parser skips them (len(fields) < 3 → continue) and the corresponding BaseNode.SizeBytes stays 0, SizeState stays "unknown".

// gitstore/gitstore.go:201–207
scan := bufio.NewScanner(&outBuf)
for scan.Scan() {
    fields := strings.Fields(scan.Text())
    if len(fields) < 3 {
        continue          // ← "missing" lines dropped silently
    }
    ...
}

2. FUSE Getattr returns SizeBytes directly without guarding on SizeState

fusefs/merged.go:94 — Resolver.Getattr returns n.Base.SizeBytes regardless of n.Base.SizeState:

return mode, n.Base.SizeBytes, n.Base.Type, mt, nil

So ~2900 files appear as 0-byte files via every stat(2) call made to the kernel.

3. Atime and Ctime are never set in inodeAttrs

fusefs/fuse_unix.go:570–596 — inodeAttrs sets Mtime but leaves Atime and Ctime as zero time.Time{}. The Go zero time serializes to a deeply negative Unix timestamp (year ~1754 as seen by stat), so the value written into the git index when the mount is first stat-ed is guaranteed to mismatch the kernel-visible ctime on every subsequent call. With core.trustctime=true (default), git's stat-cache treats every file as "potentially modified" even when size and mtime agree.

return fuseops.InodeAttributes{
    Size:  size,
    Nlink: 1,
    Mode:  m,
    Uid:   ...,
    Gid:   ...,
    Mtime: mtime,
    // Atime and Ctime intentionally omitted → zero → garbage on stat
}

Put it all together

• File reports size=0 → git stat-cache sees mismatch vs index → git reads 0 bytes → hashes empty blob → hash mismatches blob in HEAD → "modified".
• Even for files that do have known size, bogus ctime forces a re-hash on every git status invocation (slow, and the result is only cached until the next stat-cache miss).
• Hydrator.EnsureHydrated (hydrator.go:106) can hydrate a blob on demand when a read(2) arrives, but because the reported size is 0 the kernel never asks for any bytes. The read path is never taken, and the snapshot row stays SizeState="unknown" forever.

[eersnington]

@elithrar I ran into this same issue while testing artifacts locally in docker.

I was running the example container directly with FUSE enabled:

docker run --rm -it --cap-add SYS_ADMIN --device /dev/fuse artifact-fs-example-repro sh

but after the "batch size resolution failed, files will show size 0 until hydrated" warning, some tracked blobless files initially appear as 0 bytes files through FUSE. if you run "cat", it prints nothing, and "git status" marks them modified. hydration eventually catches up, so prefetch can mask it.

environment

macOS 26.5.1 arm64
Docker Desktop 4.42.0
Docker engine 28.2.2, linux/arm64, kernel 6.10.14-linuxkit
ArtifactFS revision 083e55e

also, is a branch in my fork with an issue.md file for your agent to repro and test the issue: https://github.com/eersnington/artifact-fs/blob/repro/artifact-fs-empty-files/ISSUE.md