Wednesday, October 28, 2026 · macOS 15.7 (24G214)

Dogfood on scott-mbp: Git power user switching over

5 repos · 29 commands · 4 broken · 5 slow · 5 filed

Git power user on macOS 15 / M3: alternated git and grit on git-built repos with MIDX+bitmaps, split index, commit-graph (changed paths), sparse-checkout cone + sparse index, linked worktrees, submodules, reftable, hooks, config includes, CRLF/filters/ident/UTF-16, and ran git gc --auto / git commit concurrently with grit commit/gc. On-disk interop is very solid (git fsck --strict clean everywhere, byte-identical trees for attribute/filter/symlink/odd-name cases, all 6 fixed issues verified). The problems are in behavior a git user relies on: a hasconfig includeIf identity bug, stale gitlinks in the index after switch/merge, missing merge/commit hooks, no ORIG_HEAD, and log -S/-G rename false positives. 5 issues filed.

os: macOS 15.7 (24G214) · arm64
machine: scott-mbp (agador.sparticus): Apple M3, 8 cores, 24 GB RAM
versions: grit 0.5.0 (installed on PATH; not used for measurements) · main eb2680ab · 2.52.0 (Apple/Homebrew)

repositories5

repo	files	commits	.git	notes
rust-lang/cargo	3,078	23,588	87 MB	also used for local clones, partial clone, gc/reflog tests
git/git	4,857	82,605	340 MB	commit-graph with changed paths written by git
scratch/adv (git-built: midx+bitmap, split index, commit-graph, submodule, worktree, hooks)	435	1,509	3 MB
octocat/Hello-World	1	3	1 MB	cloned, not otherwise used
scratch/* (attrs, fs, case, sparse, reftable, hooks, pickaxe, in-progress-ops)	—	—	—	small synthetic repos created with git and grit

commandsgrit vs. system git, wall time

command	repo	grit	git	grit/git	result	notes
grit status	rust-lang/cargo	20 ms	20 ms	1.0×	ok	with Git-written midx; 14.8 MB without
grit status	git/git	10 ms	130 ms	0.1×	ok	git first run was cold
grit log -n 20000	rust-lang/cargo	150 ms	100 ms	1.5×	memory	2.9x git's peak RSS
grit log -n 10000	git/git	70 ms	60 ms	1.2×	ok
grit log -- src/cargo/core/resolver	rust-lang/cargo	110 ms	90 ms	1.2×	ok
grit log -n 200 -- cache.h	git/git	20 ms	19 ms	1.1×	ok
grit revs --count HEAD	git/git	20 ms	10 ms	2.0×	ok
grit blame builtin/commit.c	git/git	230 ms	140 ms	1.6×	ok
grit blame Cargo.toml	rust-lang/cargo	60 ms	—	—	ok	65 MB peak
grit log -S the_repository -- builtin/	git/git	3.98 s	948 ms	4.2×	slow	4.2x; also returns 2 extra commits (renames), filed #2203
grit log -S sha1_file -n 20 -- cache.h	git/git	490 ms	133 ms	3.7×	slow	3.7x, #2203
grit log -S sha1_file -n 50	git/git	8.03 s	4.97 s	1.6×	ok	1.6x
grit log -G '^static int cmd_' -n 30 -- builtin/commit.c	git/git	280 ms	100 ms	2.8×	ok	2.8x, just under threshold
grit fsck	rust-lang/cargo	910 ms	4.19 s	0.2×	ok	detects flipped pack byte (badPackChecksum, corruptObject) like git
grit fsck --connectivity-only	rust-lang/cargo	7.50 s	540 ms	14×	slow	14x slower than git and 8x slower than grit's own full fsck; 'Checked 0 objects'. Not filed (rarely used flag)
grit fsck	git/git	9.49 s	26.72 s	0.4×	ok	only the legacy missingTaggerEntry issue, same as git
grit fsck --connectivity-only	git/git	133.90 s	2.58 s	52×	slow	52x slower than git; unfiled (see above)
grit gc (4-month-old reflog entries)	rust-lang/cargo	2.27 s	1.07 s	2.1×	ok	verified #2192
grit reflog expire --all	rust-lang/cargo	10 ms	—	—	ok
grit add (all files touched)	rust-lang/cargo	80 ms	90 ms	0.9×	ok	verified #2190, 0 loose objects
grit gc / grit repack (offline partial clone)	rust-lang/cargo	890 ms	530 ms	1.7×	ok	verified #2065
grit commit / switch / merge / stash / gc interleaved with git	scratch/adv	—	—	—	ok	fsck --strict clean; hooks ran; split index kept
grit merge (ff or real) with checked-out submodule ahead	scratch/adv	—	—	—	wrong	#2200 stale gitlink in index
grit merge with pre-merge-commit/commit-msg hooks	scratch/adv	—	—	—	wrong	#2201
grit config user.email (hasconfig includeIf)	scratch/adv	—	—	—	wrong	#2199
grit reset --hard / merge / rebase then ORIG_HEAD	scratch/adv	—	—	—	wrong	#2202
grit worktree add/switch/reset (reftable)	scratch/adv	—	—	—	ok	verified #2062
grit commit/switch/merge/reset in cone sparse index repo	scratch/adv	—	—	—	ok
grit clone (case-colliding tree)	scratch/adv	32 ms	—	—	ok	collision warning, same state as git

worked7

›git fsck --strict clean after every grit write in every scenario (loose, packed, midx, split index, sparse index, reftable, partial clone)
›grit commit/add produce byte-identical trees to git add -A + commit for CRLF/eol/filter/ident/UTF-16/symlink/exec-bit/odd-name/NFD cases
›grit gc removes stale midx/bitmaps, rewrites commit-graph with changed-path bloom data, git verify passes
›Hooks pre-commit/commit-msg/post-commit/post-checkout/post-merge/post-rewrite run; config includes (path, gitdir, gitdir/i, onbranch) and global excludesFile resolved correctly
›Cross-tool merge and conflict resolution state (MERGE_HEAD etc.) works both directions; reflog/rev syntax matches git; corruption is detected by grit fsck
›Concurrent git gc --auto / grit gc / git commit / grit commit on one repo: no corruption, clean lock failures
›Cone sparse-checkout and sparse index preserved by grit commit/switch/merge/reset

problems11

›FILED #2199: includeIf hasconfig:remote.*.url: ignored for repo-local remotes, grit commits with the wrong identity
›FILED #2200: switch / merge --ff leave stale gitlink in index when the submodule checkout is ahead (staged revert of bump)
›FILED #2201: merge skips pre-merge-commit/prepare-commit-msg/commit-msg; commit skips prepare-commit-msg; rebase skips pre-rebase
›FILED #2202: reset/merge/rebase never write ORIG_HEAD
›FILED #2203: log -S/-G list pure renames; log -S/-G with a pathspec 3-4x slower than git on git/git
›Unfiled: grit fsck --connectivity-only is 14x (cargo) to 52x (git/git, 134 s) slower than git and slower than grit's own full fsck, prints 'Checked 0 objects' (rarely used flag; should be a quick fix to share the full-fsck walk)
›Unfiled: grit status ignores diff.ignoreSubmodules / status.ignoreSubmodules (shows 'modified sub' where git is clean); grit status says 'Changed: modified sub' while grit commit says 'nothing to commit' for a dirty submodule
›Unfiled (minor UX): grit pick --abort on a git-started cherry-pick says 'branch tip moved since the pick started'; reflog selectors @{yesterday}/@{one.minute.ago} unsupported and @{<date>} uses midnight
›Unfiled (minor): grit log -n 20000 peaks 75 MB vs git 26 MB on cargo (2.9x); status peak 19.7 MB vs 7 MB with midx (absolute numbers tiny)
›Unfiled: grit config --list prints every secret value from the user's global config (same as git config --list; not a bug, but worth remembering for the --markdown/--json agent outputs)
›Unfiled: grit commit while a plain `mv README.md readme.md` case-rename is pending reports 'clean' in status but then records a rename in the commit (inconsistent, arguably better than git, too minor)

bugs filed5

#2199 #2200 #2201 #2202 #2203already reported: #1814 #1884 #1859

next time

›Real git-lfs flows over HTTP (git-lfs is not installed on this machine, and the global config has filter.lfs.required=true): push with pre-push hook, locks, post-checkout, git lfs fsck after grit commit
›Re-verify #1814 (sparse cone): switch/merge/reset/stash all kept out-of-cone files unmaterialised on this build, so it may be closeable; also test `git sparse-checkout reapply/disable` after grit operations on a sparse-index repo and grit status performance with sparse index on a large monorepo
›hasconfig/includeIf variants after #2199 is fixed (remote added after commit, worktree config, relative include paths, ~ expansion)
›Submodule workflows against a real upstream: grit switch/reset across bumps with dirty/untracked submodule content, nested submodules, submodule.recurse, status for dirty submodules vs git
›Hook parity matrix after #2201: prepare-commit-msg sources (message/merge/squash/commit), pre-push, pre-auto-gc, reference-transaction, post-checkout from worktree add; core.hooksPath from subdirectories and linked worktrees
›Large-scale concurrency: git gc --auto racing grit commit on cargo-size repos with index.lock contention and kill -9 recovery; SHA-256 repos made by git; alternates and shallow clones made by git
›fsck --connectivity-only and log -S/-G performance after fixes; blame -L ranges vs git on git/git; MIDX memory with millions of objects