Thursday, October 29, 2026 · Windows 11 10.0.26200 (NTFS, Defender on)

Dogfood on win-x64: New-CLI author

7 repos · 28 commands · 2 broken · 5 slow · 5 filed

As a new-CLI author on win-x64 I built a scratch front end on grit-lib's public API (op-log/undo, paged log, fetch progress, integrate with cancel, held-handle repack) and ran the CLI on cargo, git/git, and 22.5k/200k/5k-file synthetic repos. Library and CLI are largely solid on Windows (typed lock/in-use errors, rollback on locked files, long-lived handle survives repack, status/diff on 200k files faster than git), but I found a regression of #2190 (grit add rewrites packed blobs as loose: 71x slower), a case-collision status/diff mismatch, slow cancel rollback, and gaps for building undo on the library. Of the six fixed issues, five verified; #2278 still has 60% silent time.

os: Windows 11 10.0.26200 (NTFS, Defender on) · x86_64
machine: win-x64 worker WIN-H6JSI3IMAJG: Intel i5-12400, 16 GB RAM
versions: grit grit 0.5.24 (cargo build --release -p grit-cli) · main 1d054b51 · git version 2.55.0.windows.5 (peak memory measured on mingw64\bin\git.exe, not the cmd wrapper); system gitconfig core.autocrlf=true core.symlinks=false

repositories7

repo	files	commits	.git	notes
rust-lang/cargo	3,078	23,596	81 MB
git/git (bare)	4,857	82,623	325 MB	1020 refs; file count is ls-tree of HEAD
octocat/Hello-World	1	3	0 MB	used for repack with open handle
scratch/wide-22.5k	22,500	1	13 MB	150 dirs; branch other changes every file (git fast-import)
scratch/flat-200k	200,000	1	52 MB	all files in one directory
scratch/dup1-5k	5,000	1	2 MB	branch mod packed with git gc; repro for #2328
scratch/small (stk, cf, lock1, coll3, oplog3, racy)	30	6	0 MB	stacked-branch, conflict, lock and case-collision fixtures

commandsgrit vs. system git, wall time

command	repo	grit	git	grit/git	result	notes
grit switch other (22.5k files)	scratch/wide-22.5k	2.93 s	5.46 s	0.5×	ok	grit 1.9x faster
grit reset --hard <other> (22.5k files)	scratch/wide-22.5k	11.57 s	5.56 s	2.1×	ok	2.1x slower and 3.9x slower than grit's own switch; duplicate of #2311 (open)
grit stash (22.5k modified files)	scratch/wide-22.5k	69.61 s	7.58 s	9.2×	slow	9x slower; cause largely #2328 duplicate loose writes; also starts background gc --auto
grit stash pop (22.5k files)	scratch/wide-22.5k	3.29 s	6.59 s	0.5×	ok	50 s when a background gc --auto from the preceding grit stash is still running
grit add (22.5k modified, blobs already stored)	scratch/wide-22.5k	35.05 s	2.41 s	15×	slow	#2328
grit add (5k modified, blobs only in a pack)	scratch/dup1-5k	10.92 s	153 ms	71×	slow	#2328: 5000 duplicate loose objects, 71x slower
grit status (clean)	scratch/flat-200k	498 ms	766 ms	0.7×	ok
grit diff	scratch/flat-200k	243 ms	330 ms	0.7×	ok
grit fetch origin (local path, full)	rust-lang/cargo	13.73 s	12.84 s	1.1×	ok	git peak is the parent process only (children not counted); silent progress gaps are #2278
library Remote::fetch local path with progress sink	rust-lang/cargo	14.07 s	—	—	wrong	60% of the run emits no progress events (#2278)
grit status	rust-lang/cargo	122 ms	45 ms	2.7×	ok	2.7x on a single cold run, under the 3x line
grit log -n 1000 --oneline	rust-lang/cargo	26 ms	44 ms	0.6×	ok
grit diff HEAD~50 HEAD	rust-lang/cargo	62 ms	71 ms	0.9×	ok
grit fsck	rust-lang/cargo	1.81 s	5.92 s	0.3×	ok
grit revs HEAD~5000..HEAD	rust-lang/cargo	302 ms	167 ms	1.8×	ok
grit merge-base HEAD~500 HEAD~3000	rust-lang/cargo	67 ms	144 ms	0.5×	ok
library paged log, page 5000 (LogWalkCursor + load rows)	rust-lang/cargo	229 ms	276 ms	0.8×	ok
library full walk + load_commits_with_decorations	git/git (bare)	1.46 s	997 ms	1.5×	ok	page 10000: 778 ms / 101 MB
grit log --json -n 100000	git/git (bare)	1.25 s	997 ms	1.3×	ok	84 MB of JSON incl. full messages; git ran --format=%H%x09%an%x09%B
grit fsck	git/git (bare)	16.48 s	41.05 s	0.4×	ok
grit blame Makefile	git/git (bare)	1.34 s	1.08 s	1.2×	ok
grit log -n 1000 --oneline -- builtin/log.c	git/git (bare)	2.99 s	578 ms	5.2×	slow	5.2x; -n 200 -- Makefile 413 vs 89 ms (4.6x); duplicate of #1619 (macOS) now seen on Windows; no commit-graph in the clone
grit rebase main (3-branch stack, update-refs)	scratch/small (stk, cf, lock1, coll3, oplog3, racy)	—	—	—	ok	all three branches moved (#2213)
grit rebase conflict, --continue, history undo	scratch/small (stk, cf, lock1, coll3, oplog3, racy)	—	—	—	ok	#2274 fixed; stop leaves a dangling empty blob/tree (harmless)
grit switch with exclusively locked file	scratch/small (stk, cf, lock1, coll3, oplog3, racy)	127 ms	—	—	ok	typed 'file is in use' message, nothing moved
grit switch collide; grit status; grit diff	scratch/small (stk, cf, lock1, coll3, oplog3, racy)	—	—	—	wrong	#2329: status clean, diff shows A.txt, no warning
library integrate_branch with FlagCancel at 500..2500 ms	scratch/wide-22.5k	11.91 s	2.94 s	4.1×	slow	#2330: cancel@2500 returns after 11.9 s vs 2.9 s uncancelled; all runs roll back cleanly
grit repack -f with library handle open	octocat/Hello-World	109 ms	62 ms	1.8×	ok	held Repository re-read objects after the pack was replaced

worked8

›Typed, readable errors for locked files (file in use) and held index.lock (LockHeld) with full rollback; no Access denied surprises
›A long-lived Repository handle survives another process replacing packs on Windows
›grit status/diff on a 200k-file directory is faster than git (0.5 s vs 0.77 s, half the memory)
›grit fsck is 2.4-3.3x faster than git on cargo and git/git
›rebase --update-refs on multi-commit stacks, conflict stop/continue and history undo
›Cancellation always leaves HEAD, index and worktree consistent with no stale lock
›serde-models gives a usable StatusModel and result types; ProgressSink/Cancel/Progress traits are easy to implement
›Short-name into_ref for integrate_branch and the paged LogWalkCursor are pleasant APIs; docs on switch/commit/integrate are thorough

problems14

›Filed #2328: grit add/commit/stash rewrite blobs already in a pack as loose objects (regression of closed #2190)
›Filed #2329: case-colliding paths on Windows: no switch warning, grit status clean but diff/git show modified
›Filed #2330: cancelling a 22.5k-file checkout takes up to 9.4 s to return (rollback ~3x forward)
›Filed #2331: no snapshot/restore API for undo logs; stash create fails mid-merge; apply_commit rejects identical untracked; Error serde loses fields
›Filed #2332: API hygiene: hidden pub fields, Vec-only walks, 14 Windows warnings, destructive symlink restore on non-unix, C expression in a rustdoc comment
›#2278 reopened via verify_fix: local fetch is silent for 60% of the run
›Not filed (duplicate #2311): grit reset --hard is 2.1x git and 3.9x grit switch on 22.5k files
›Not filed (duplicate #1619): path-limited log on git/git is 4.6-5.2x slower than git on Windows too
›Not filed (minor UX): grit stash errors 'Committer identity unknown' where git stash works without a configured identity
›Not filed (minor): a conflicted rebase stop leaves a dangling empty blob and empty tree in the ODB
›Not filed (minor): grit-lib example gritx-status hashes work-tree files by writing them with odb().write (pollutes the repo) although hash::hash_object exists
›Not filed (unverified): grit-lib depends on the `sha1dc` crate while AGENTS.md says not to implement sha1dc; I did not check whether it is used
›grit stash starts a background `gc --auto` that competes with the next command (stash pop 3.3 s -> 50 s); folded into #2328 (see also open #1859)
›Release build of grit-cli takes 7.5 min on this machine (i5-12400), grit-lib alone 4 min

bugs filed5

#2328 #2329 #2330 #2331 #2332already reported: #2311 #1619 #1798 #1859 #2142

next time

›Test symlink restore on Windows with core.symlinks=true (Developer Mode): history undo / rewrite rollback of a tracked symlink (restore_worktree_path non-unix branch)
›Run the throttled HTTPS fetch of git/git against grit-http-server and measure ReceivedObjects progress and server RSS (not reached this session)
›Custom ObjectStore plus injected RefStore end to end: commit, merge, fetch, export to disk, git fsck
›Re-run #2328 on Linux/macOS and with gc.auto triggered, and measure grit commit and stash with already-stored content on a 100k-file repo
›Cancel latency of commit/stash/merge sessions after #2142 and of reset --hard rollback with parallel restore
›Build a jj-style op log once a conflicted-state snapshot exists; check undo of stash pop and pick with in-progress operations
›Windows: core.longpaths and 260-char paths through the library, reserved names after #2257, concurrent create_commit from two processes