Thursday, October 29, 2026 · Ubuntu 24.04 (Linux 6.12.94+)

Dogfood on linux-cloud: Automation and CI scripter

4 repos · 34 commands · 3 broken · 1 slow · 5 filed

As the automation/CI scripter I re-verified all six 'fixed' issues on current main (all genuinely fixed), then went after patch-id parity on real history, machine output (--json/--markdown/--filter/error kinds), env/config overrides, hooks, signals, concurrency, shallow/CI flows and grit serve token auth. Core flows are solid and fast; I filed 5 issues: two residual patch-id mismatches vs git, clone --json/--markdown leaking URL tokens, markdown output not escaping commit messages, and inconsistent JSON error kinds.

os: Ubuntu 24.04 (Linux 6.12.94+) · x86_64
machine: Cursor cloud VM, 4 vCPU Intel Xeon, 15 GB RAM
versions: grit 0.5.24 · main e84d7c9d · 2.43.0

repositories4

repo	files	commits	.git	notes
rust-lang/cargo	3,078	23,596	82 MB
git/git	4,857	82,616	324 MB
octocat/Hello-World	1	3	1 MB
scratch/patch-id, odd-names, env, hooks, race, gcrace, serve repos	—	—	—	synthetic repos under /tmp/dogfood

commandsgrit vs. system git, wall time

command	repo	grit	git	grit/git	result	notes
grit patch-id <300 cargo commits> vs git show\|git patch-id --stable	rust-lang/cargo	207 ms	917 ms	0.2×	wrong	1/400 differs (#2291)
grit patch-id <1500 commits> vs git show\|patch-id	git/git	1.25 s	5.77 s	0.2×	wrong	3/1500 differ (#2292)
grit clone file:// (local)	git/git	24.91 s	17.04 s	1.5×	ok	1.5x time, 1.8x mem; result passes git fsck
grit fetch --tags (depth-1 clone)	rust-lang/cargo	6.51 s	5.29 s	1.2×	memory	2.3x git peak RSS; not filed (similar to known fetch memory issues)
grit clone --depth 1 -b 0.80.0 file://	rust-lang/cargo	390 ms	—	—	ok
grit status --json	git/git	10 ms	10 ms	1.0×	ok
grit diff --json v2.40.0 v2.45.0	git/git	250 ms	350 ms	0.7×	ok
grit log --json -n 5000	git/git	60 ms	40 ms	1.5×	ok
grit revs --count HEAD	git/git	710 ms	560 ms	1.3×	ok
grit blame --json revision.c	git/git	820 ms	920 ms	0.9×	ok
grit merge-base v2.40.0 v2.45.0	git/git	10 ms	170 ms	0.1×	ok
grit log --json -n 20 -- builtin/log.c	git/git	10 ms	10 ms	1.0×	ok
git fsck --connectivity-only on grit-made clone	git/git	—	3.30 s	—	ok	clean
grit resolve HEAD (hyperfine -N)	rust-lang/cargo	2.1 ms	0.8 ms	2.6×	ok	2.6x; startup floor
grit log --json -n1	rust-lang/cargo	2.4 ms	1.2 ms	2.1×	ok
grit cat HEAD:Cargo.toml	rust-lang/cargo	2.3 ms	1.2 ms	1.9×	ok
grit status --json	rust-lang/cargo	13 ms	10 ms	1.2×	ok
grit diff --quiet	rust-lang/cargo	5.0 ms	2.8 ms	1.8×	ok
grit branch --json	rust-lang/cargo	2.3 ms	0.8 ms	2.9×	ok	2.9x, tiny absolute
grit tag --json (124 tags)	rust-lang/cargo	3.7 ms	0.9 ms	4.0×	ok	git number is a single-tag lookup, not equal work
grit --version	octocat/Hello-World	1.3 ms	0.3 ms	3.7×	ok	0.9 ms user time; 43 MB binary, 39,699 relative relocs
grit status --json	octocat/Hello-World	2.4 ms	1.2 ms	2.0×	ok
grit clone --json http://x:TOKEN@host/repo	scratch/serve	—	—	—	wrong	prints token (#2302)
grit push 60MB random over HTTP (grit serve, token)	scratch/serve	2.74 s	2.54 s	1.1×	ok
grit clone 120MB over HTTP (grit serve, token)	scratch/serve	6.34 s	6.92 s	0.9×	ok
grit gc x60 while 3x80 pushes	scratch/gcrace	4.10 s	—	—	ok	0 failures, fsck clean (#2210 verified)
grit reflog expire x40 while 100 commits	scratch/rl	—	—	—	ok	no entries lost (#2286 verified)
grit push --force-with-lease=<rev> x8 parallel x3	scratch/race	—	—	—	ok	1 winner, 7 'stale' each round
grit push (same new ref) x8 parallel, plain and --atomic	scratch/race	—	—	—	ok	1 winner, 7 already_exists
grit switch --detach <old tag> + SIGTERM	rust-lang/cargo	—	—	—	ok	clean cancel, nothing half-written
grit clone + SIGTERM/HUP/INT	rust-lang/cargo	—	—	—	ok	dir removed; exit 130 for all
grit commit with commit.gpgsign=true, gpg.format=ssh	scratch/sg	—	—	—	ok	signed like git
grit log/diff/show/status --json with non-UTF-8 and special file names	scratch/odd	—	—	—	ok	valid JSON; non-UTF-8 as grit+pathhex: (documented)
grit <45 commands> --json / --markdown sweep	scratch/sweep	—	—	—	ok	all JSON valid

worked8

›All six previously fixed issues are really fixed on main (pull --ff-only exit 128, patch-id empty files/slide, clone signal cleanup, reflog expire locking, gc vs concurrent pushes).
›Same-ref creation and force-with-lease races over local remotes: exactly one winner, typed 'already_exists'/'stale' reasons, repo fsck-clean.
›Cancellation on SIGTERM during switch is clean ('operation cancelled', no half-written worktree, no lock left).
›Hooks (pre-commit, commit-msg, post-commit, pre-push, pre-receive, post-receive) and GIT_PUSH_OPTION_* behave like git; nothing prompts or opens an editor or pager without a TTY, including under `script`.
›Env/config overrides (author/committer env and dates, -c, -C, GIT_DIR/WORK_TREE, GIT_INDEX_FILE, alternates, ceiling dirs, core.worktree, env -i identity errors) match git; commit.gpgsign with ssh signing works.
›Auth failures are non-interactive with documented kinds (auth_required, auth_rejected); GIT_ASKPASS, URL tokens and http.extraHeader all work against grit serve.
›Every --json output I parsed was valid JSON, and status/log/diff fields agree with git, including special file names.
›Performance on realistic repos is at or near git (status, diff, log, blame, merge-base, push/clone over HTTP); patch-id is 4-5x faster than git's pipeline.

problems11

›Filed #2291: rename similarity differs from git (94 vs 95) so diff --json similarity and patch-id differ (cargo 9d6a94a55).
›Filed #2292: patch-id differs for a plain insertion into a 437-line file (git/git t4201, commit 253118d1d99b), size dependent.
›Filed #2302: clone --json / --markdown print the token from a user:token@host URL (release and main).
›Filed #2303: show --markdown prints commit message bodies raw: unclosed ``` fence swallows the real sections, forged '## Changes' headings.
›Filed #2304: --json unknown-revision errors have no kind on 12 commands (documented revision_not_found only from merge); bundle/init I/O errors do not name the path.
›Not filed (minor): all signals during clone exit 130 (SIGTERM would be 143); other commands cancelled by SIGTERM exit 1, fetch has no handler (143, leaves tmp_pack_* like git).
›Not filed (minor): `grit log >&-` (closed stdout) exits 0 silently; git exits 128 with a write error.
›Not filed (documented): grit+pathhex: paths from status --json are not accepted by restore/cat/log/blame operands; --force-with-lease accepts only =<rev>, not git's =<ref>:<rev> (error is 'could not resolve revision main:<oid>').
›Not filed: fetch --tags in a depth-1 cargo clone uses 419 MB vs git 180 MB (2.3x); `grit gc` prints 'Running garbage collection…' on stderr even when not a TTY; no -q on switch.
›Not filed: startup floor of 1.3 ms for --version (git 0.35-0.7 ms) and 2-3x per tiny command; 58 failing readlink calls per log -n1. Under the 3x bar for equal work.
›Not filed: config get returns only the last value of a multi-valued key (no --get-all); only --list shows all.

bugs filed5

#2291 #2292 #2302 #2303 #2304already reported: #2151 #2183 #2205 #1884

next time

›Fix #2291/#2292, then rerun patch-id parity over all of cargo and git/git (--no-merges, ~80k commits) including renames with edits, copies and type changes.
›Profile the 0.9 ms user-time startup floor (clap command tree, relocations, config discovery, readlink/statx storm) and test 5000-invocation loops with perf.
›Retry kill -9 / OOM-kill during gc, repack and fetch on packed repos and check leftover tmp_pack/locks cleanup via gc; add SIGTERM during commit with hooks.
›Over real smart-HTTP/ssh servers (not only grit serve): shallow push, receive.shallowUpdate, partial clone (--blobs none) CI flows, credential helpers, 413 limits.
›Library angle: embed grit-lib in a CI tool with cancellation (ProgressSink::should_cancel on commit/add) and typed lock errors; check that the JSON error kinds are available as typed variants.
›Re-test markdown renderers (log, reflog, stash, tag -a, notes, blame) with hostile messages/paths once #2303 is fixed.