grit://cooking/dogfood← what’s cooking
Sunday, November 8, 2026 · Windows 11 Pro 10.0.26200 (Build 26200)

Dogfood on win-x64: Edge cases

6 repos · 33 commands · 1 broken · 5 slow · 5 filed

Dogfooded grit on Windows 11 x64 with the edge-case scenario: unicode and space paths, CRLF/autocrlf/gitattributes, symlinks, exec bits, empty repo, detached HEAD, .gitignore patterns, Windows-hostile tree entries, long paths, locked files, large binaries, plus cargo and Hello-World for timings. Correctness and git interoperability are very good: everything grit wrote passed git fsck --strict and matched git's view. I filed 5 issues: backslash tree entry accepted on checkout, two large-blob memory issues (fsck, restore with autocrlf=true), a Windows status slowdown on cargo, and a --stat alignment bug.

os
Windows 11 Pro 10.0.26200 (Build 26200) · x86_64
machine
win-x64 self-hosted worker (virtual machine, 1 CPU package on a Debian/EDK II host, 16 GB RAM)
versions
grit none installed on PATH · main e1bbc863 · git version 2.55.0.windows.5 (system config: core.autocrlf=true, core.symlinks=false, core.fscache=true)
repositories6
repofilescommits.gitnotes
rust-lang/cargo3,08123,66480.9 MBfull clone, git-made copy used for timings; pack 74.5 MiB over the wire
octocat/Hello-World130 MB5 commits on all branches
scratch/uni repo ünï1130 MBunicode/space/emoji/NFD names; also CRLF files
scratch/big-binary (grit-big)21401.6 MB400 MB random + 1.5 GB zero file, loose blob for the random file
scratch/fs1111 MBsingle 1 GB zero file
scratch/crlf2, ign, bin1, mrg, clone1, win-bad/wb-*810 MBsmall repos for gitattributes/CRLF, gitignore, binary/BOM/UTF-16 files, merge/stash/hooks, symlink/exec bits, hand-built hostile-name trees
commandsgrit vs. system git, wall time
commandrepogritgitgrit/gitresultnotes
grit init / status / log / diff / show / stash (empty repo)scratch/empty———okexit codes and messages sensible
grit commit (9 unicode/space files)scratch/uni repo ünï67 ms——okgit fsck --strict clean
grit mv / rm / restore / blame / show --stat / diff --jsonscratch/uni repo ünï25 ms——ok--stat alignment bug #2567
grit add/commit/restore with autocrlf and gitattributesscratch/crlf235 ms——okmatches git --eol output
grit clone (symlink/exec repo)scratch/srcgit138 ms——ok
grit ignore check / add --dry-run / cleanscratch/ign35 ms——okmatches git ls-files -o --exclude-standard
grit switch --detach / commit / tag / reflogscratch/ign30 ms——ok
grit clone (hostile tree entry back\slash.txt)scratch/win-bad78 ms79 ms1.0×wronggrit accepts and writes back/slash.txt, git rejects: #2563
grit clone (case-colliding tree)scratch/src-case127 ms119 ms1.1×oksame collision warning and state as git
grit clone (304-char path)scratch/longp145 ms123 ms1.2×okgit fails with Filename too long, grit works
grit switch with locked filescratch/win140 ms——okclear error, repo state intact
grit merge --check / merge / merge --abort (CRLF+unicode)scratch/mrg45 ms——ok
grit commit with failing and CRLF pre-commit hookscratch/mrg82 ms——ok
grit stash / pop (unicode, untracked)scratch/mrg50 ms——ok
grit push / clone / fetch / pull / worktree add+remove (unicode+space paths, file:/// URL)scratch/mrg61 ms——ok
grit clone https://github.com/octocat/Hello-Worldoctocat/Hello-World737 ms702 ms1.0×ok
grit clone https://github.com/rust-lang/cargorust-lang/cargo15.03 s11.56 s1.3×oknetwork-bound; 3 runs: grit 15.0-26.5 s vs git 11.3-13.5 s; not filed (network noise, memory under bar)
grit statusrust-lang/cargo130 ms50 ms2.6×slow2.6-2.8x slower, filed #2566 (below the 3x label, but over the 100 ms/50 ms/10% bar)
grit log -n 50rust-lang/cargo15 ms27 ms0.6×ok
grit log --json -n 1000rust-lang/cargo30 ms50 ms0.6×ok
grit diffrust-lang/cargo53 ms42 ms1.3×ok
grit blame src/util/workspace.rsrust-lang/cargo89 ms130 ms0.7×ok
grit fsckrust-lang/cargo1.53 s5.85 s0.3×okgrit 3.8x faster
grit revs HEAD --countrust-lang/cargo125 ms167 ms0.8×ok
grit describerust-lang/cargo213 ms126 ms1.7×slow1.7x slower than git describe (annotated tags); open duplicates #2462/#2467
grit commit (400 MB random + 1.5 GB zero files)scratch/big-binary2.16 s22.29 s0.1×okgit add+commit; grit 10x faster, far less memory
grit fsck (1.5 GB + 400 MB blobs)scratch/big-binary3.46 s4.51 s0.8×memory#2564
grit fsck (single 1 GB blob)scratch/fs11.72 s2.39 s0.7×memory#2564
grit gc / repack (1.9 GB, loose)scratch/big-binary8.09 s14.01 s0.6×okgrit faster; git gc peak figure is of the launcher shim so not comparable
grit status after 1-byte edit of 400 MB filescratch/big-binary351 ms1.86 s0.2×ok
grit diff (400 MB binary)scratch/big-binary1.09 s3.33 s0.3×ok
grit switch --detach HEAD~1 (rewrites 400 MB file)scratch/big-binary1.50 s1.43 s1.1×ok
grit restore big_zero.bin (1.5 GB, autocrlf=true)scratch/big-binary2.17 s2.70 s0.8×memory2.0x git's peak; #2565. With autocrlf=false: 9.8 MB
worked8
  • ›Everything grit wrote (unicode/space/emoji/NFD names, CRLF, gitattributes, symlink and exec modes, large blobs, merges, stash, worktrees) passed git fsck --strict and matched git status/ls-files/ls-tree
  • ›CRLF handling matches git exactly across core.autocrlf true/false and text/eol/binary/-text attributes, including the lone-CR and mixed cases
  • ›Windows hostile names (reserved device names, trailing dot/space, NTFS streams, GIT~1, .git aliases, ..\x and .git\hooks traversal) are rejected like git; case-collision warnings match git
  • ›Long (>260 char) paths work where git fails with 'Filename too long'; a file held open by another process gives a clear error and leaves HEAD, index and work tree consistent
  • ›.gitignore evaluation matches git including negation, anchors, **, escapes, unicode and core.ignorecase
  • ›Commit, status after edit, gc and fsck on 400 MB-1.5 GB binaries are 2-10x faster than git with much less memory
  • ›--json output (status, diff incl. binary and no-newline markers, log, describe, clean, add --dry-run) was well-formed and consistent in everything I tried; unicode survives unescaped
  • ›Branch/tag name validation refuses refs that system git cannot read on Windows (aux, con, nul.x, a:b, trailing dot)
problems13
  • ›FILED #2563: tree entry with a backslash in its name is checked out as a subdirectory on Windows (git rejects it); status then shows untracked back/ and a following commit duplicates the entry
  • ›FILED #2564: grit fsck holds whole blobs in memory (1 GB file = 1.06 GB peak vs git 10 MB)
  • ›FILED #2565: restore/checkout of a 1.5 GB file with core.autocrlf=true (Windows default) peaks at 3.1 GB vs git 1.5 GB; streams with autocrlf=false
  • ›FILED #2566: grit status on cargo (3081 files, Windows) is 130 ms vs git 47 ms (2.8x)
  • ›FILED #2567: --stat misaligns the bar with CJK/emoji file names
  • ›NOT FILED: grit describe on cargo is 213 ms vs git 126 ms (existing open issues #2462/#2467)
  • ›NOT FILED: grit clone of cargo over the network is 15-26 s vs git 11-13 s (about 3.5 s of non-network time: pack prep, indexing, resolving, checkout); network noise makes the ratio unreliable and memory is under the bar
  • ›NOT FILED: `grit add top.txt` for a file named Top.txt exits 0 and stages nothing, with no message (git does exactly the same); confusing for agents but at parity
  • ›NOT FILED: failed `grit switch` over a locked file says 'cannot remove' although the file is being modified, and its hint talks about a partially restored work tree when nothing moved; minor wording
  • ›NOT FILED: `grit commit` aborts for a nested repo with no commits, 'embedded git repository with no commit checked out'; same as git, but commit auto-stages so a stray nested repo blocks every commit
  • ›NOT FILED: local-path clone copies unreachable loose objects (same as git --local); the summary counts only reachable objects
  • ›NOT FILED: grit gc on a 1.9 GB loose repo peaks at 965 MB; git's real peak could not be measured via the shim and is probably smaller, needs a re-test on Linux
  • ›Could not search issues with gh (not logged in); I used the unauthenticated GitHub search API instead, which rate-limited partway, so duplicate checks for the backslash and stat-alignment reports were by title keywords only
bugs filed5
#2563 #2564 #2565 #2566 #2567already reported: #2462 #2467
next time
  • ›Re-run the large-blob memory cases (fsck, restore with autocrlf=true, gc/repack) on Linux to see which are Windows-only and measure git's peak properly
  • ›Find what status spends its kernel time on in Windows (per-file CreateFile vs directory enumeration) and test status on git/git and a 50k-file tree
  • ›Symlinks proper: run with Developer Mode on so core.symlinks=true, test symlink to dir, dangling links, symlink-to-file type changes
  • ›NTFS 8.3 short-name aliases of .gitmodules/.gitattributes (GITMOD~1) and junctions/reparse points inside the work tree
  • ›Partial clone/--blobs, sparse checkout, submodule-looking gitlinks, and bundle round trips with unicode ref names
  • ›Concurrent operations (two grit processes, stale index.lock/HEAD.lock handling, antivirus-held files) on Windows