grit://cooking/dogfood← what’s cooking
Tuesday, November 10, 2026 · Ubuntu 24.04 (Linux 6.12)

Dogfood on linux-cloud: New-CLI author

3 repos · 30 commands · 2 broken · 4 slow · 5 filed

As a new-CLI author I re-checked the six fixes, built a tiny front end (st/save/stack/fsck/blame/walk) against grit-lib's public API with the serde-models feature, audited cancel/progress and JSON coverage, and ran a structure-aware mutation fuzzer over pack .idx, commit-graph, MIDX and split/untracked-cache index files. Verdict: the common library paths work well (status/commit/branch/blame/fsck-cancel are quick and typed), but two 'fixed' issues (#2390, #2563) were silently reverted on main by commit 5e28d0386, and I filed five new findings: a 22x slower RevWalk object enumeration, two corrupt-file robustness bugs (allocator abort, quadratic fsck), and two API gaps (no cancel/progress on read-side ops, serde-models covering only 12 of 43 result types).

os
Ubuntu 24.04 (Linux 6.12) · x86_64
machine
Cursor cloud VM, 4 vCPU, 15 GB RAM
versions
grit 0.5.33 · main 41d768c8 · 2.43.0
repositories3
repofilescommits.gitnotes
octocat/Hello-World131 MBalso used as the tiny repro for MIDX/bloom aborts and upload-pack bombs
rust-lang/cargo3,08623,70382 MBfuzz target (copies with commit-graph, MIDX, split index, untracked cache), perf matrix
git/git4,86682,797326 MBblame re-check, rev-list --objects, cancel audit
commandsgrit vs. system git, wall time
commandrepogritgitgrit/gitresultnotes
grit log -n 20 --onelinerust-lang/cargo10 ms10 ms1.0×ok
grit statusrust-lang/cargo8.0 ms8.6 ms0.9×okhyperfine warm; first run after clone 130 ms vs 10 ms
grit revs --count HEADrust-lang/cargo110 ms140 ms0.8×ok
grit blame --rev HEAD Cargo.tomlrust-lang/cargo90 ms110 ms0.8×ok
grit describerust-lang/cargo180 ms100 ms1.8×okgit/git describe not re-run (known #2467/#2462)
grit diff 0.80.0..HEADrust-lang/cargo310 ms320 ms1.0×ok
grit fsckrust-lang/cargo1.42 s5.85 s0.2×ok4x faster than git; grit lists all dangling objects, git only roots
grit log -n 50 -- src/cargo/corerust-lang/cargo10 ms10 ms1.0×ok
grit merge-base HEAD~3000 HEAD~2500rust-lang/cargo60 ms60 ms1.0×ok
grit clone (local path)rust-lang/cargo190 ms160 ms1.2×ok
grit bundle create --allrust-lang/cargo870 ms1.05 s0.8×okgit bundle verify accepts it
grit fsck (corrupt idx + MIDX)rust-lang/cargo960.00 s3.85 s249×slow#2579, killed after 16 min
grit log -n 1 (MIDX num_packs=0xffffffff)octocat/Hello-World———crash#2578 SIGABRT 103 GB alloc; git fatal exit 128
grit log -- README (bloom num_hashes=0xffffffff)octocat/Hello-World———crash#2578 SIGABRT 16 GB alloc
grit upload-pack --stateless-rpc (18.8M duplicate wants)octocat/Hello-World3.05 s10.97 s0.3×memory#2390 regressed on main (fix reverted by 5e28d0386)
grit log -n 20 --onelinegit/git0.0 ms0.0 ms—ok
grit statusgit/git4.6 ms6.3 ms0.7×okhyperfine warm
grit revs --count HEADgit/git420 ms580 ms0.7×ok
grit blame --rev HEAD Makefilegit/git1.10 s1.07 s1.0×ok#2572 no longer reproduces on main
grit blame --json builtin/commit.c vs git blame --line-porcelaingit/git———ok4 of 1993 lines differ (was 41, #2574)
grit log -n 50 -- builtin/commit.cgit/git40 ms40 ms1.0×ok
grit diff v2.0.0..HEADgit/git860 ms520 ms1.7×ok
lib RevWalk::walk(objects=true)git/git81.15 s3.79 s21×slow#2577, 21x
lib RevWalk::walk(objects=true)rust-lang/cargo3.80 s440 ms8.6×slow#2577, 8.6x
lib RevWalk::walk (commits only)git/git610 ms——ok
lib Diff::blame(Makefile) cancel latencygit/git1.12 s——okreturns 917 ms after cancel request; no cancel parameter (#2582)
lib Maintenance::fsck cancel at 400 msrust-lang/cargo408 ms——okreturned 8 ms after cancel; full run 1657 ms
grit serve + curl v2 fetch / git clonescratch/5x300KB———ok#2568 verified
grit bundle create/unbundle chainscratch———ok#2569 verified; git bundle verify and fsck agree
stk st / save / stack (lib front end)scratch———okStatusModel and CommitOutcome serialize with serde-models; commit carries change-id; works with global ssh signing config
worked6
  • ›Status, commit, branch listing, ahead_behind and conflict kinds are easy to use from a 40-line program and serialize with serde-models (StatusModel, CommitOutcome have change_id, per-file status, conflict_details)
  • ›fsck is 4x faster than git on cargo (1.4 s vs 5.85 s) and cancels within 8 ms
  • ›Warm status/log/merge-base/path-limited log on cargo and git/git are at parity with git; bundle create is faster than git and verified by git
  • ›bundle chains (#2569), grit serve v2 fetches (#2568) and stash pop of deletions (#2570) work
  • ›Index, split-index, untracked-cache, FSMN and commit-graph generation/edge chunks survived 650+ index and ~100 commit-graph mutations with no panic, hang or abort
  • ›grit-lib public module paths and docs: quickstart, rustdoc and examples compile cleanly against the 0.6.0 tree
problems7
  • ›#2390 and #2563 were reverted on main by commit 5e28d0386 (fixes and regression tests removed); reopened via verify_fix
  • ›#2577 RevWalk::walk with objects=true is 8-22x slower than git rev-list --objects (81 s on git/git)
  • ›#2578 corrupt MIDX num_packs / bloom num_hashes abort the process with 100 GB / 16 GB allocations
  • ›#2579 corrupt idx + MIDX makes fsck run >16 min (per-lookup MIDX chain rebuild)
  • ›#2582 no CancelToken/ProgressSink on RevWalk walk/describe/blame; fsck progress phases are mislabeled; two progress traits
  • ›#2583 serde-models covers only 12 of 43 audited result types; every CLI command re-declares view structs
  • ›Not filed (minor): #2574 residual 0.15% blame divergence on blank/brace lines; content/docs/library-quickstart.md still says grit-lib = "0.5.0" while the crate is 0.6.0; StatusOptions/CommitRequest/DescribeOptions/RevListOptions are plain pub-field structs (RevListOptions has ~45 rev-list-flag fields) rather than non_exhaustive builder structs; grit fsck with a MIDX whose pack lacks its .idx reports 'No integrity problems' (git: failed to load pack entry) - too minor; first status after clone 130 ms vs 10 ms (cold index refresh, warm is parity); 'grit stash push' has no -q (CLI choice); mdoc: lib.rs says legacy flat paths private but revwalk::describe/rev_list modules are private while their option structs are only reachable via different public paths (grit_lib::revwalk::DescribeOptions vs revwalk::rev_list::RevListOptions), which made first-contact discovery slower
bugs filed5
#2577 #2578 #2579 #2582 #2583already reported: #2467 #2462 #2572
next time
  • ›Re-check #2390 and #2563 after the revert is repaired and audit the remaining hunks of commit 5e28d0386 (grit-lib/src diff, 160+ files) for other lost fixes
  • ›Re-run describe on git/git (still known 46 s) including --tags, --first-parent and tag tie-breaks against git describe
  • ›Extend the fuzzer to the pack body (delta chain cycles, bad ofs-delta, oversized object headers), pack .rev/.mtimes, bitmaps, and reftable tables with a git >= 2.45 binary; fuzz through the library API directly with catch_unwind so panics are counted
  • ›After #2579 is fixed, check the negative-lookup cost with a MIDX plus a missing (not corrupt) .idx and with incremental MIDX chains
  • ›Custom ObjectStore + injected RefStore end to end through merge/rebase/push and linked worktrees (#2127/#2128/#2131 still open)
  • ›Build the undo-log / stack restack front end against the History handle with conflicts, and check that every outcome type becomes Serialize once #2583 lands
  • ›Case-insensitive filesystem (ciopfs) semantics for switch/checkout/worktree add; Windows/macOS cancel latency of read-side ops once they accept a CancelToken