grit://cooking/changelog← what’s cooking
Sunday, November 1, 2026 · release v0.5.28

95 changes landed on main

68 issue fixes · 27 other changes

v0.5.28 is mostly merge, push and ref-correctness fixes, custom object-store durability, and large speedups for log, diff, reset, fsck and rm on big repos.

  • ›Merge/pick: Git-compatible diff alignment and histogram default fix false conflicts and silent auto-merges (#2391, #2397, #2431)
  • ›New `grit describe` and `log --follow` for file history across renames, with Bloom-filter speedups (#2316, #2455)
  • ›Push fixes: reject moving existing remote tags, relative-path remotes, correct output labels (#2425, #2428, #2398)
  • ›Custom ObjectStore primaries: flush before ref/index publish, fixes for clean merge/pick/revert, batched writes (#2443, #2441)
  • ›Big-repo speedups: rm, clean, diff pathspec/rename, fsck, reset --hard, sparse switch, log streaming, shortlog-style grouping
  • ›Data-safety fixes: gc keeps linked-worktree staged blobs, reftable tombstones, add/mv/rm release index.lock on Ctrl-C (#2416, #2448, #2433)
$ git log 217a3104..v0.5.28242 commits on GitHub ↗300+ files +15,915 −4,972
Highlightsmeasured by the reviewers, before → after
Follow-up: Fix #2410: grit clone over https peaks at ~pack size + 85 MB RSS: git/git 402 MB vs git 96 MB (4.2x), cargo 127 vs 46 MB
Pack bytes read to obtain SHA-1 trailer on the 43,443 KiB RSS fixture2224282× better████████████████████████ 44,485,632 bytes █ 20 bytes
grit push with 1000 refspecs costs 16 ms per refspec before connecting (3.3x git); cost scales with local ref count
6,001 local refs, 1,000 shorthand specs: origin/main to branch median475× better████████████████████████ 15.3 seconds █ 0.032 seconds
switch/reset --hard in a cone sparse-index repo cost O(whole tree): 150k out-of-cone changes 3.4 s (git 0.04 s)
250k fixture reset kd wall (system Git 0.007816 s)428× better████████████████████████ 3.5 s █ 0.008 s
diff.algorithm=histogram is 10-50x slower than git and uses ~930 MB RSS on files with many repeated lines (po files)
24k-line histogram peak RSS (median of 5)61× better████████████████████████ 931,820 KiB █ 15,304 KiB
diff/show/merge/pick quadratic on many deleted+added paths with repeated basenames: 20k+20k files 11.6 s (git 0.11 s)
20k deletes + 20k adds: grit diff --stat median46× better████████████████████████ 5.0 seconds █ 0.109 seconds
issues fixed68
  • ›pick/merge silently auto-merge where git reports a content conflict (7 of 300 cargo cherry-picks onto an older branch)#2397af5742af

    This aligns Grit’s three-way merge hunk placement with Git when one side deletes a repeated tail boundary and the other inserts beside it. The review follow-up preserves exact byte identity for non-UTF-8 text while retaining indentation scoring, preventing the false conflict introduced in round one.

  • ›grit-protocol / smart_http::rpc hard-code PushGuard, no hook/cancel/outcome; grit serve hooks never see who pushed#24005c987063

    Authenticated push identity now reaches serve-side hooks, and embedders can carry in-process hooks and cancellation through protocol/HTTP sessions while receiving typed upload/receive outcomes. The round-2 commit makes advisory post-receive and post-update callback failures observable without undoing already-applied refs or rewriting the successful client response.

  • ›grit revs --all / log -n N buffer the whole walk before printing: first byte 0.8-1.3 s on git/git (git 3-17 ms)#2388a7156c54

    Plain `revs` and `log` walks now emit commit batches as the walker produces them instead of retaining the entire result before rendering. The library exposes a stream for eligible multi-token walks, while the CLI incrementally renders human, Markdown, and JSON output and preserves collected fallbacks for operations that require whole-walk state.

    revs --all first 100 bytes, main to review (14,803 commits): 226.3 ms median of 7 → 35.7 ms median of 7 (6.3× better)log -n 14000 --json first 100 bytes, main to review: 115.7 ms median of 7 → 4.9 ms median of 7 (24× better)revs --all peak RSS, main to review: 53,868 KB median of 3 → 50,632 KB median of 3 (6% better)log -n 14000 --json peak RSS, main to review: 62,832 KB median of 3 → 44,616 KB median of 3 (41% better)
  • ›grit fsck panics (slice out of range, fsck/mod.rs:692) on a reachable commit whose parent line is truncated#2386c0d44ea4

    Malformed reachable commit parent headers no longer panic fsck. Connectivity fallback parsing now uses the existing bounds-checked commit tree/parent parser, and regressions cover truncated and invalid parent lines across reachable, unreachable, loose, and packed objects.

  • ›Odb::write into a custom store probes contains() on primary and every read source before each write(); no batch hook#23943ef8f804

    Custom-primary ODB writes now go directly to the idempotent backend instead of probing the primary and every read source first. The library adds explicit flush/rollback durability barriers around pack ingest and commit publication, while the SQLite example batches writes transactionally and correctly persists or discards them.

  • ›push human output labels multi-ref pushes 'pushed refs →' and --follow-tags tag pushes 'pushed main → refs/tags/…'#2398674dd91e

    Push output now derives every successful line from that result's source ref, so multi-ref, followed-tag, `--tags`, and renamed-destination pushes identify the ref actually sent. Followed tags also make the JSON summary multi-ref, while rejected rewrites receive an actionable force-with-lease hint and strictly remote-ahead branches retain the pull-first hint.

  • ›merge: false content conflict where git merges cleanly (real git/git merge 1fbfabfa71, config.c); diff alignment differs#23917d41ce6d

    This replaces `similar`-based merge alignment with Git-compatible xdiff compaction so repeated-line edits no longer produce a false content conflict. Round 3 removes byte-key copies and shares interning across the two default-Myers side diffs, recovering a material portion of the performance cost while preserving byte-accurate histogram and patience behavior.

    Default-Myers repeated-anchor merge median (2,000 warmup; 7×20,000 release merges): 29.1 µs/merge → 39.3 µs/merge (26% worse)
  • ›grit fetch <bundle> with no refspec says "Already up to date" after ingesting objects, updates no refs#2401a4540501

    Literal bundle fetches without a refspec now resolve the implicit HEAD request before object ingestion. Headless incremental bundles fail cleanly instead of writing dangling objects and falsely reporting that nothing changed, while explicit refspec fetches continue to update their requested destination.

    grit-lib full-feature integration compile errors: 110 compiler errors → 106 compiler errors (4% better)
  • ›Windows: clone/switch/restore write symlinks as regular files even with core.symlinks=true (git creates real symlinks)#240253ed5573

    This change makes checkout-family operations honor core.symlinks on Windows: enabled repositories create actual links, while disabled repositories retain Git-for-Windows-style placeholder files. It centralizes symlink creation and preserves the target byte length in refreshed index entries so a clean checkout does not immediately report a type or content change.

  • ›grit rebase and conflicted merge leave dangling empty blob e69de29 (and empty tree) in the ODB#2403ed0d0a0a

    This change prevents rebase and conflicted merge/replay operations from writing transient empty blobs, empty trees, conflict-marker blobs, and materialized checkout trees into the real object database. It keeps genuinely referenced stage-0 auto-merge blobs durable while confining checkout-only objects to an in-memory ODB overlay.

  • ›grit log has no way to follow a file across renames (no --follow); history of a renamed file stops at the rename#231682a02794

    Adds library and CLI support for following one file's history across renames, with typed validation and Git-compatible rename detection. Round five removes heuristic basename/hyphen path guessing and revises merge lineage selection, but the merge revision currently drops valid history in the real git/git acceptance case.

    git/git builtin/commit.c followed commits (system Git reference vs branch): 607 commits → 604 commits (0% worse)
  • ›v0.5.25 linux: grit switch to far-away commit/tag is ~4x slower than git on 24k-commit repo (41ms vs 10ms)#2337ac8d1db2

    Adds a heap-free first-parent scan before the general ancestry paint walk and generation-based pruning when both commits are covered by a commit-graph. It also fixes the round-1 equality-boundary regression and adds focused regression/performance-shape tests.

    24k linear detached switch dry-run median: 33.1 ms → 32.1 ms (3% better)24k linear detached switch median: 40.8 ms → 38.3 ms (6% better)
  • ›grit mv <file> dir/ fails 'invalid path: path component is empty' (trailing slash on destination directory)#2407097d7ceb

    This change makes `grit mv file dir/` behave like `grit mv file dir` when the destination is an existing directory. It also preserves trailing-separator intent for missing destinations, introduces a typed refusal for file-to-missing-directory moves, and improves move errors and CLI hints.

  • ›grit switch: false 'untracked file would be overwritten' when a tracked symlink becomes a directory#2404325e0c50

    The wholesale switch guard now checks every added-path prefix with non-following metadata and stops when it reaches a symlink. A tracked symlink is accepted and replaced by the target directory tree; an untracked symlink is reported at the link path rather than exposing files reached through its target.

  • ›grit pull with an overlapping uncommitted change: error names no file and gives no hint (merge/switch do)#2405b8806cb1

    Pull overwrite refusals now retain their typed library error through CLI presentation. Users see the blocking paths and commit/stash guidance, while JSON and Markdown receive the same structured data as merge and switch.

  • ›grit diff and stash abort with pathless 'I/O error: Permission denied' on an unreadable directory (#2361 follow-up)#2406a71d3065

    Diff and stash worktree scans can now skip permission-denied tracked paths with a path-bearing warning while continuing over readable changes. Path-limited scans are pre-expanded correctly, including tree-only staged deletions selected by attribute, directory, mode-aware, and exclude pathspecs.

  • ›--json --filter on a 41 MB diff outcome: 1.6-2.8 s and 1.25-1.78 GB RSS (unfiltered: 0.3 s, 226 MB)#24094c23db8a

    This change removes multiple whole-document conversions from filtered JSON output. Command outcomes are serialized once, released before jaq parsing, validated for strict JSON representability, and written directly to buffered stdout while preserving the one-JSON-value contract.

    Cargo diff `.files|length` wall time: 1.7 seconds (3-run median) → 1.2 seconds (3-run median) (35% better)Cargo diff `.files|length` peak RSS: 1,244,748 KB (3-run median) → 683,088 KB (3-run median) (82% better)Cargo diff identity filter wall time: 2.8 seconds (3-run median) → 1.6 seconds (3-run median) (79% better)Cargo diff identity filter peak RSS: 1,776,000 KB (3-run median) → 691,108 KB (3-run median) (2.6× better)
  • ›grit gc / gc prune deletes blobs reachable only from a linked worktree's index (data loss, git keeps them)#2416964d4c35

    Garbage collection now treats every non-bare worktree index as an object-reachability root, so staged content in linked worktrees cannot be pruned from the shared object database. Resolve-undo entries root only regular-file blobs, avoiding attempts to resolve submodule commit IDs in the superproject ODB.

    One-worktree `gc prune --older-than now` smoke wall time: 12.1 ms → 11.6 ms (4% better)
  • ›restore, stash push, rm --cached reject ':(attr:...)' pathspecs ('did not match any file'); add/diff/log accept them#2417bd1961a0

    This makes repository-aware attribute pathspecs work consistently in restore, stash push, and rm --cached, including attributes that depend on the real Git object mode. It also recognizes long magic pathspecs in log/revs without requiring `--`, with compatibility tests and command documentation.

    Attribute restore median process time, system git reference vs reviewed grit (30 runs): 1.3 ms → 6.5 ms (81% worse)
  • ›--json errors for common CI failures have no `kind`: not a repo, unknown identity, tag exists, no merge in progress#241395319e4b

    Adds stable JSON `kind` values and structured hints for common CI failures across repository discovery, commit identity/message validation, tags, paused operations, pathspecs, stash lookup, and missing blame paths. The round-2 fix restores all session-cancellation tests to the default grit-lib suite and hardens the test classifier against the `start` helper-name false positive.

    grit-lib files classified fast by classify_tests.py: 109 files → 110 files (1% better)grit-lib files classified Git-only by classify_tests.py: 632 files → 631 files (0% better)
  • ›switch/merge ff drop index entries (and later fail) when >=16 directories are replaced by files (#2336 variant)#2415d2084c76

    This change prevents bulk and chunked checkouts from losing index entries or attempting descendant deletions after a directory has already become a file. It also corrects large-transition untracked-collision handling so switch, fast-forward merge, and hard reset can safely process directory-to-file and symlink-to-file replacements.

  • ›grit log on a terminal loses all output and exits 0 when the default pager `less` is not installed#2408cbeb673d

    This change prevents terminal `grit log` output from disappearing when the default or configured pager executable is unavailable. It preflights plain pager commands and falls back to direct terminal output, while preserving shell execution for pipelines, redirections, environment assignments, expansions, and Windows command paths.

  • ›Concurrent clones of the same bundle-serving repo race on the shared bundle cache; most fall back to full fetch#24203edaa6f4

    Adds process-level ownership for shared resumable bundle cache entries. A clone or checkout that cannot claim the shared cache lock downloads and consumes a uniquely named private bundle instead, preventing peers from corrupting or deleting the shared `.bundle`, `.part`, and `.json` files.

  • ›grit-lib: collect_referenced_object_roots + prune_loose_unreachable deletes staged blobs; no safe-roots API#24214b7b951c

    Safe maintenance root collection now protects staged objects from the main index, every linked-worktree index, and an explicitly selected GIT_INDEX_FILE. The refs-only root helper becomes internal, while docs and callers direct embedders to the repository-aware maintenance API.

  • ›rebase and history reword/squash never copy notes to rewritten commits (notes.rewriteRef ignored)#24187d1089bc

    History rewrites and amend operations now preserve configured Git notes on replacement commits. Configuration is preflighted before history movement, and notes refs use compare-and-swap against the snapshot tip so concurrent writers cannot be silently overwritten.

  • ›grit fsck in a --shared clone (alternates) takes 29 s vs 1.5 s unshared (git 6.3 s): pack dir rescanned per object#2419ba8a3b54

    This change removes repeated pack-directory and loose-object probes during alternate ODB misses while preserving external object visibility. Repository fsck now validates objects in alternates, counts them, and reports only dangling roots instead of the full unreachable closure.

    Shared-clone fsck wall time: 62.6 seconds → 5.0 seconds (13× better)Shared-clone fsck time relative to system Git: 2.9 x Git time → 0.231 x Git time (13× better)
  • ›grit rm <dir> costs ~1.1 ms CPU per removed file (even --dry-run): 2500 files 3.1 s, 50k files 59 s (git: 0.37 s / 5 s)#24242e58110c

    This change replaces repeated full-index scans during `rm` selection with sorted range queries for literal paths and one prepared matcher pass for non-literals. It also batches index-extension invalidation and empty-parent cleanup, while the round-2 commit restores canonical Unicode matching for precomposed pathspecs.

    5k-path dry-run: origin/main vs branch (50k-entry index, median of 7): 640.3 ms → 15.4 ms (42× better)5k-path dry-run: system Git vs branch (50k-entry index, median of 7): 404.2 ms → 15.4 ms (26× better)5k-path actual removal: origin/main vs branch (median of 5): 723.3 ms → 83.8 ms (8.6× better)5k-path actual removal: system Git vs branch (median of 5): 473.2 ms → 83.8 ms (5.6× better)
  • ›grit diff A B -- <path> runs rename detection on the whole diff first: 3.3 s vs git 27 ms (250k files)#2422a37035ba

    This change prunes stored-tree diffs by pathspec before rename detection and indexes exact rename candidates by object ID instead of comparing every delete with every add. It preserves attribute-aware pathspec behavior and Git’s one-sided rename semantics while substantially reducing work for narrow diffs.

    One-path diff median on 10k-file fixture (2.5k deletes + 2.5k adds): 12.7 ms → 2.7 ms (4.6× better)One-path diff slowdown versus system Git (system Git median 1.590 ms): 8.0 x → 1.7 x (4.6× better)
  • ›grit clone over https peaks at ~pack size + 85 MB RSS: git/git 402 MB vs git 96 MB (4.2x), cargo 127 vs 46 MB#24101058fe52

    Reverse-index publication now opens the pack, seeks to its trailing object hash, and reads only 20 or 32 bytes instead of loading the whole pack into a Vec. The round-5 cleanup also leaves no production RSS instrumentation or direct instrumentation environment/stderr hooks.

    grit ingest peak RSS on 43,443 KiB pack: 56,356 KiB → 12,856 KiB (4.4× better)grit/system-git peak RSS ratio: 5.2 x → 1.2 x (4.4× better)system git index-pack comparison peak: 10,932 KiB → 10,860 KiB (1% better)
  • ›diff between two commits skips inexact renames git reports: cargo home-0.5.9..HEAD 148 renames vs git's 471#24111058fe52

    Adds a grit-lib basename-driven rename phase before renameLimit, typed skip notices propagated through public diff APIs, and a thin CLI stderr notice. Round 3 tries to recover performance by bypassing the basename phase for globally unique basenames and expands duplicate-basename benchmarks.

    diff_renames threads=1 midpoint: 75.3 ms → 74.3 ms (1% better)diff_renames threads=4 midpoint: 83.6 ms → 77.0 ms (9% better)common_64 midpoint: 5.2 ms → 6.0 ms (13% worse)
  • ›Follow-up: Fix #2410: grit clone over https peaks at ~pack size + 85 MB RSS: git/git 402 MB vs git 96 MB (4.2x), cargo 127 vs 46 MB#241015b5e252

    This changes reverse-index sidecar generation to seek directly to the pack checksum trailer instead of reading the whole pack into memory. It also provides a clean replacement commit whose metadata and message satisfy the integration contract.

    Pack bytes read to obtain SHA-1 trailer on the 43,443 KiB RSS fixture: 44,485,632 bytes → 20 bytes (2224282× better)Focused pack-ingest RSS ratio, immediate repeatability runs: 1.5 grit/git ratio → 1.5 grit/git ratio (3% better)
  • ›Follow-up: Fix #2411: diff between two commits skips inexact renames git reports: cargo home-0.5.9..HEAD 148 renames vs git's 471#2411dc410cd1

    This change adds a basename similarity pre-pass before the inexact rename-limit check, allowing large commit-to-commit diffs to retain inexact renames that Git recognizes. It propagates a typed skipped-detection notice through the library and thin CLI outputs, shares rename blob data, and covers the behavior with Git-oracle and CLI regressions.

    Criterion diff_renames_adoc_similar/common_64 median: 6.0 ms → 4.4 ms (37% better)
  • ›blame -M / -C returns duplicate rows for repeated identical lines: BlameResult.lines longer than the file#24268ed4f99e

    Move/copy blame now processes each current file range once even when several tracked final lines map to the same current index. This restores the one-result-per-file-line invariant without discarding legitimate shared parent attribution.

    Issue fixture -M output rows: 4 rows → 2 rows (2.0× better)Issue fixture -M median debug CLI runtime (30 runs): 7.6 ms → 6.6 ms (15% better)
  • ›grit push ../repo says "unknown remote" for a relative local path unless it ends in .git (fetch/pull/clone accept it)#2428da9d6d56

    Push now recognizes `.`/`..`, `./…`, `../…`, and existing local Git directories as literal destinations instead of configured-remote names. This makes relative local pushes consistent with fetch, pull, clone, and system Git while leaving ordinary lone branch/refspec tokens unchanged.

  • ›grit clean (dir mode) is quadratic in tracked files: 0.7 s on a clean 100k-file tree vs git clean -n 0.04 s#2429672cd5c6

    Clean directory mode now asks the sorted index whether a directory has tracked descendants and uses an exact binary lookup for stage-0 gitlinks. This removes the index-wide scan repeated for every worktree directory without changing CLI behavior or formats.

    100k clean tree grit clean --dry-run median: 548.8 ms (8 runs after 2 warmups) → 27.8 ms (8 runs after 2 warmups) (20× better)grit/system-git median time ratio: 15.3 ratio → 0.784 ratio (20× better)system git clean -n control median: 35.8 ms (8 runs after 2 warmups) → 35.5 ms (8 runs after 2 warmups) (1% better)
  • ›grit push silently moves an existing remote tag (fast-forward or annotated->lightweight); git rejects 'already exists'#24259e26e2a8

    Plain pushes can no longer move a remote tag merely because its peeled commit is a descendant or unchanged. Existing changed tags are rejected with a stable `tag_exists` result unless `--force` is explicit, while new tags, identical tags, deletes, branches, and forced tag updates retain their intended behavior.

  • ›FilesRefStore: CAS update losing a race fails in commit() but still appends its reflog entry (phantom reflog)#2427a4d15e47

    FilesRefStore now refreshes compare-and-swap state after acquiring prepare locks and appends each update’s reflog only after its commit-time recheck succeeds. The round-3 correction ensures symbolic log-only updates lock every reflog destination, including the peeled branch whose reflog receives a mirrored entry.

  • ›merge/pick: false conflict on cargo workspaces.md where git (ort) merges cleanly; git merge-file also conflicts#243197da4583

    Tree-level three-way merges now default content alignment to histogram, matching Git's ort behavior for merge, pick, restore, stash, and rerere paths while preserving Myers semantics for the standalone merge-file API. The round-2 follow-up makes the system-Git compatibility regression use the visible test helper exports and the correct repository path type.

  • ›pick/merge silently auto-merge where git reports a content conflict (7 of 300 cargo cherry-picks onto an older branch)#23976335fb2a

    This change makes grit’s three-way file merge recognize merge-ort’s repeated-anchor EOF overlap cases, preventing picks from silently appending changes after a tail deleted on the target branch. The round-3 adjustment narrows that heuristic to repeats inside the actual one-sided tail cluster, preserving clean merges when the only earlier matching anchor lies outside the changed region.

  • ›grit log --follow / RevWalkListOptions::follow is 4.4-4.9x slower than git log --follow on git/git (2.3 s vs 0.53 s)#243572b9e6f7

    This change makes rename-following revwalks avoid repeated tree scans and duplicate rename work. It uses changed-path Bloom filters for unchanged commits, cached literal-path tree lookups for existence checks, and carries each selected parent's next follow path out of the decision step so callers do not recompute it.

    builtin/commit.c follow full median: 3,434 ms → 709.1 ms (4.8× better)builtin/commit.c follow -n20 median: 204.9 ms → 41.8 ms (4.9× better)Makefile follow full median: 2,752 ms → 577.0 ms (4.8× better)Makefile follow -n20 median: 75.1 ms → 21.0 ms (3.6× better)
  • ›log --follow / RevWalkListOptions::follow returns nothing for a path that no longer exists at the start commit#2434ff16d54e

    When `--follow` starts at a commit where the requested old path is already gone, the revwalk now continues through parents with that same path instead of stopping immediately. It therefore reaches and emits the rename/delete commit and earlier history while still returning an empty list for a path that never existed.

    Issue fixture debug CLI median (40 runs): 6.6 ms → 7.5 ms (13% worse)Fixed branch vs system Git median on issue fixture: 1.1 ms (system Git before, branch debug grit after) → 7.5 ms (system Git before, branch debug grit after) (85% worse)
  • ›rebase --exec aborts atomically on a replay conflict ('can't resolve conflicts yet') instead of stopping for --continue#24325c207079

    This change makes `rebase --exec` stop resumably on an initial replay conflict and persists exec recovery state through continue, skip, and abort. Round three narrows untracked cleanup so normal abort and successful continue preserve user files, but failed-command recovery still uses a stale baseline and can delete post-stop user files.

  • ›Objects::read of a corrupt loose/packed object returns Zlib(String)/Io(Eof) with no object id or path (CLI too)#243046788082

    Object reads now preserve actionable corruption context instead of exposing bare zlib or EOF failures. Loose failures identify the object and file; packed failures identify the object, pack and offset; repository resolution and the CLI retain that typed detail, while composite stores can still recover from another healthy copy.

  • ›diff/show/merge/pick quadratic on many deleted+added paths with repeated basenames: 20k+20k files 11.6 s (git 0.11 s)#24360ed25210

    This replaces repeated delete×add scans in basename rename preflight with hash-set membership checks and updates only the directory guess whose count changed. It adds regression coverage for repeated, disjoint basenames and a Criterion workload matching that shape.

    20k deletes + 20k adds: grit diff --stat median: 5.0 seconds → 0.109 seconds (46× better)20k deletes + 20k adds: runtime relative to system Git 2.43: 38.9 x Git → 0.845 x Git (46× better)
  • ›log -n N A..B -- path on linux with changed-path Bloom commit-graph: 1.1-3.4 s, cost tracks range size (git 0.08-0.3 s)#2438bcc923a3

    This change replaces eager range-hide painting for dense path-limited revision walks with an incremental generation-ordered hide cursor. The round-3 correction keeps deferred lower-generation frontier nodes available, preventing excluded shared roots from reappearing in later merge-parent queries.

    20,001-commit synthetic `-n 1 root..main -- watched.txt` median (debug build): 265.4 ms → 156.4 ms (70% better)Same synthetic query slowdown versus system Git (9.470 ms median): 28.0 x → 16.5 x (70% better)
  • ›Ctrl-C/SIGTERM/SIGHUP during grit add leaves a stale index.lock that blocks every later write command#24330fef76ae

    Adds cooperative interruption to add, mv, and rm so handled SIGINT/SIGTERM/SIGHUP exits through cancellation and releases index locks. The round-3 repair makes rm transactional across partial worktree deletion failures by restoring all already-removed paths and reporting incomplete rollback precisely.

  • ›switch/reset --hard in a cone sparse-index repo cost O(whole tree): 150k out-of-cone changes 3.4 s (git 0.04 s)#2437dfff8a19

    Cone sparse switch and hard reset now preserve opaque excluded subtrees instead of recursively flattening them. Added excluded trees are represented as canonical `040000` skip-worktree rows ending in `/`, so both Grit and system Git can read and round-trip the sparse index.

    250k fixture switch kd wall (system Git 0.014519 s): 0.935 s → 0.008 s (123× better)250k fixture switch main wall (system Git 0.014844 s): 0.913 s → 0.007 s (131× better)250k fixture reset kd wall (system Git 0.007816 s): 3.5 s → 0.008 s (428× better)250k fixture reset main wall (system Git 0.007832 s): 3.7 s → 0.009 s (415× better)
  • ›Follow-up: Fix #2432: rebase --exec aborts atomically on a replay conflict ('can't resolve conflicts yet') instead of stopping for --continue#2432a60cc44d

    Exec-enabled rebases now stop resumably on replay conflicts and preserve their operation state through continue/skip/abort flows. Failed exec rollback is scoped to files created by the current command, so untracked files created by the user while a rebase is stopped are no longer silently deleted.

  • ›reset --hard over 50k-150k changed files peaks at 2.4-3.4x git's RSS (297-318 MB vs 93-108 MB at 150k)#2439db306236
  • ›Follow-up: Fix #2439: reset --hard over 50k-150k changed files peaks at 2.4-3.4x git's RSS (297-318 MB vs 93-108 MB at 150k)#2439f6a04d9e

    Hard-reset checkout coalesces wholly added or removed subtrees so large tree transitions avoid materializing a leaf-level diff for every changed file. The follow-up test now makes a real system-Git compatibility comparison for nested untracked files instead of comparing grit status to a second copy of itself.

    30k/18k reset to kd elapsed: 0.361 seconds → 0.346 seconds (4% better)30k/18k reset to main elapsed: 0.323 seconds → 0.318 seconds (2% better)30k/18k reset to kd peak RSS: 35,220 KiB → 35,456 KiB (1% worse)30k/18k reset to main peak RSS: 35,684 KiB → 36,120 KiB (1% worse)
  • ›integrate_branch on a custom ObjectStore primary fails with UnsupportedObjectStore{write_local} on true merges#2442db8fb10b

    Overlay-generated merge objects now persist through the repository's configured primary object store. Custom writable primaries receive direct writes and a batch flush, while files-backed repositories continue using loose-object writes.

  • ›log --follow / RevWalkListOptions::follow returns nothing for a path that no longer exists at the start commit#2434d27b5498

    When a followed path is absent from both the starting commit and its parent, the revwalk now keeps that path active and continues into older history. Hidden absent-path commits remain hidden, while the later rename/deletion boundary and earlier path history are returned.

  • ›Clean non-ff merge, cherry-pick and revert fail on any custom ObjectStore primary (write_local UnsupportedObjectStore)#24410a22d95c

    Custom object-store repositories can now complete clean non-fast-forward merges, cherry-picks, and reverts instead of failing when their in-memory merge overlay is flushed. Files-backed repositories retain the existing loose-object persistence path, and no public API or object format changes.

  • ›ODB flush() only called by commit, stash and ingest: tag, notes, history apply update refs onto unflushed objects#24431d592948

    Adds a single durability barrier for custom buffered object stores and invokes it before refs or the index can expose newly written objects. Flush failures now roll back staged objects, including commit and pack-ingest paths, and rebase --exec no longer moves detached HEAD before its rewritten objects are durable.

  • ›Reftable auto-compaction of non-bottom tables drops a delete tombstone: deleted ref comes back, commit() returned Ok#2448191504ab

    Non-bottom reftable compaction now carries forward deletion tombstones so older tables cannot make deleted branches or tags reappear. Bottom-inclusive compaction still drops tombstones because no older table remains to shadow, preserving compact output.

  • ›No WritableObjectStore::flush before ref/index publish in annotated tag, notes, history rewrite, rebase and stage#24453dfb05b9

    Object writes to a batched custom primary are now made durable before Grit writes an index or moves refs that name those objects. A private ODB helper centralizes flush-and-rollback behavior, and all affected staging, tag, notes, history, rebase, exec, commit, and ingest paths use it.

    2,000-file add median, origin/main vs review: 38.8 ms → 37.2 ms (4% better)2,000-file add median, system Git vs reviewed Grit: 139.7 ms → 37.2 ms (3.8× better)
  • ›upload-pack sends bogus `unshallow` when re-shallowing with several wants: fetch --depth 1 pulls full history#24472d0a4a55

    This corrects upload-pack’s shallow-boundary classification for multi-tip re-shallowing and for repositories that are themselves shallow. Clients now receive only valid `shallow`/`unshallow` transitions, and traversal stops at server-side graft boundaries even though the underlying commit objects still contain parent lines.

  • ›Local-path/file:// clone and fetch into a custom store issue 2 unbatched contains() per object (390k for cargo)#2446274c154e

    Local-path and file:// transfers into custom object-store primaries avoid per-object existence round trips. Fresh stores write the reachable set directly; incremental transfers batch candidate checks, including tag-object and peeled-target validation, while retaining the existing optimized files-primary path.

  • ›diff.algorithm=histogram is 10-50x slower than git and uses ~930 MB RSS on files with many repeated lines (po files)#245132d68be8

    Reworks histogram diff indexing and traversal so repeated-line files no longer retain full-file occurrence maps through recursive regions or repeatedly hash and scan the same lines. It adds deterministic work bounds plus system-Git numstat oracles for repeated `.po`-style content.

    24k-line histogram wall time (median of 5): 1.3 seconds → 0.299 seconds (4.3× better)24k-line histogram peak RSS (median of 5): 931,820 KiB → 15,304 KiB (61× better)Wall-time multiple versus system Git median: 25.4 x → 5.9 x (4.3× better)
  • ›grit fsck --connectivity-only on cargo takes 6.6 s / 544 MB (git: 0.32 s / 141 MB) and is 8x slower than full grit fsck#2449a258013f

    Connectivity-only fsck now builds a header-only object kind map and walks only commit/tree/tag payloads, marking valid blobs reachable without inflating them. The round-2 correction preserves reachable blobs in dangling analysis, including blobs reached through trees and annotated tags.

    connectivity-only mean wall time (139,987 objects, one pack, 5 warm runs): 27.1 seconds → 1.6 seconds (17× better)connectivity-only peak RSS: 447,964 KiB → 263,248 KiB (70% better)runtime ratio versus measured system Git (0.538 s): 50.4 x → 3.0 x (17× better)
  • ›Repository::is_bare() is true for a linked worktree of a bare repository (has a work tree; git says not bare)#24548b52c19e

    Repository handles with a real work tree now report non-bare even when their common repository has core.bare=true. Worktree enumeration separately preserves Git's convention of marking the main row bare when core.bare is enabled.

  • ›Nightly test-full/coverage red: grit-lib `integration` test binary fails to compile with full-tests (115+ errors, E0639)#2444

    Updates external grit-lib integration tests to construct non-exhaustive request/options types through their supported constructors and then assign public fields. The round-2 repair restores option semantics and all compatibility tests that were accidentally dropped during the mechanical migration.

  • ›RefStore tx delete keeps the reflog: FilesRefStore then fails a nested ref create with Corrupt, Memory/Reftable accept#245017f9e61c

    Ref transactions now remove the reflog belonging to a deleted ref on memory, files, and reftable stores. This also removes stale files-backend log paths that prevented a nested ref from being created later, while preserving a newly recreated reflog in same-batch dereference/delete flows.

  • ›reset --hard over 50k-150k changed files peaks at 2.4-3.4x git's RSS (297-318 MB vs 93-108 MB at 150k)#2439325bb3ce

    This change lowers hard-reset memory by avoiding broad index/path materialization and by coalescing safe whole-subtree deletion work. Its round-two follow-up also makes the new public skip predicate compose correctly with restricted worktree diffs and adds focused coverage for that contract.

    50k-files-per-tree reset peak RSS (origin/main → reviewed branch): 71,044 KB → 75,720 KB (6% worse)50k-files-per-tree reset wall time (origin/main → reviewed branch): 1.4 seconds → 4.4 seconds (69% worse)
  • ›grit log --group-by author is ~4x slower than git shortlog -sne (cargo 0.72 s vs 0.19 s, git/git 3.0 s vs 0.77 s)#24524d3675b1

    This change makes default human and Markdown author grouping use a single count-only library walk instead of repeatedly walking and loading full commits. It accelerates packed and deltified histories through bounded prefix reconstruction and pack-order parallel counting, while the round-2 fix preserves exact author identity semantics.

    cargo 23,607 commits, grit wall time: 0.710 seconds median → 0.132 seconds median (5.4× better)cargo relative to system Git 0.1760 s: 4.0 × system Git → 0.751 × system Git (5.4× better)git.git 82,634 commits without graph, grit wall time: 2.9 seconds median → 0.413 seconds median (7.1× better)git.git without graph relative to system Git 0.7250 s: 4.0 × system Git → 0.569 × system Git (7.1× better)
  • ›RevWalk::commits/walk without a limit order commits differently from limit: Some(n), commits_iter and git rev-list#2453fac874b3

    Makes unlimited default revwalks use the same ungated committer-date ordering as bounded walks and system Git. The round-2 correction keeps object-filtered walks on the full engine, preventing the new fast path from bypassing sparse-filter parsing and validation.

    Equal-date fixture APIs exactly matching system git rev-list: 2 of 4 APIs → 4 of 4 APIs (2.0× better)
  • ›grit push with 1000 refspecs costs 16 ms per refspec before connecting (3.3x git); cost scales with local ref count#2458dc801553

    Push planning now snapshots local heads and tags once, indexes full and short names, and reuses known names/OIDs while resolving every requested refspec. This changes only grit-lib planning internals and adds focused regression and Criterion coverage; transport, CLI, public APIs, and repository formats are unchanged.

    1,001 local refs, 1,000 shorthand specs: origin/main to branch median: 2.4 seconds → 0.009 seconds (270× better)6,001 local refs, 1,000 shorthand specs: origin/main to branch median: 15.3 seconds → 0.032 seconds (475× better)6,001 local refs, 1,000 colon specs: origin/main to branch: 16.0 seconds → 0.035 seconds (462× better)1,001 local refs: system Git median to review branch median: 0.423 seconds → 0.009 seconds (48× better)
  • ›grit-lib has no describe / nearest-tag API: changelog tools must hand-roll tag peeling + walk (no CLI equivalent either)#24558c02c172

    Adds a typed nearest-tag description API to `grit-lib` and exposes it through a thin `grit describe` command with human, JSON, and Markdown output. It peels nested tags, skips non-commit tags, filters reachability in one walk, and reports the minimum exact `tag..commit` count with stable tie-breaking.

other changes27
  • ›Main keeper: make CI on main green (public-api, rustdoc)fa7a7bb7

    Refreshes the checked-in grit-lib public API surface after already-landed API additions and makes one CLI rustdoc link fully qualified. This is a mechanical CI repair with no executable behavior or Git format/protocol changes.

  • ›Main keeper: make CI on main green (rustdoc, public-api)53ed5573

    This branch contains no changes relative to origin/main. It confirms that the previously failing rustdoc link and stale public API snapshot were already corrected on main at the exact SHA under review.

  • ›Quality #73: Move logic into grit-libd2551aef

    Diff target selection now lives behind grit-lib APIs that return pathspec- and status-filtered DiffEntry rows for every DiffTarget and load opposite-side peer changes for stable hunk IDs. The CLI calls those APIs and drops its duplicate staged, unstaged, worktree, commit/range, path-filter and peer-selection helpers; --diff-filter now consistently applies to all diff modes.

    diff-10000 git-normalized grit/git ratio: 1.1 ratio → 1.1 ratio (0% better)diff-deep-leaf git-normalized grit/git ratio: 2.7 ratio → 2.7 ratio (1% worse)commit-range --stat mean: 59.0 ms → 56.5 ms (4% better)commit-range filtered --name-only mean: 9.0 ms → 9.2 ms (1% worse)
  • ›Quality #74: Deduplicatec9349e37

    Adds one fixed-width, seven-hex-digit display form to `grit-lib` and routes object-id labels throughout the library, CLI, and HTTP server through it. It removes ten named helper copies and numerous inline truncations without changing repository data, protocols, or observed output.

    log-10000 median (system Git 2.936 ms): 2.9 ms → 3.0 ms (4% worse)blame-deep-2000 median (system Git 68.376 ms): 24.3 ms → 22.4 ms (9% better)blame-large-tail-full median (system Git 71.233 ms): 30.6 ms → 30.1 ms (2% better)status-dirty-10000 median (system Git 9.921 ms): 8.1 ms → 7.9 ms (2% better)
  • ›Speed: fetch/deepen@git-git73e8d02b

    This rewrites v0/v1 smart-HTTP fetch negotiation so each stateless POST is a complete request, while the negotiator owns replay state and stops on server readiness. Shallow/deepen fetches now offer local haves, request `deepen-relative` only as a capability, and use `no-done` to avoid a final HTTP round trip.

    fetch/deepen wall: system Git → branch grit: 684.2 ms median → 180.2 ms median (3.8× better)fetch/deepen RSS: system Git → branch grit: 258.9 MiB peak → 56.6 MiB peak (4.6× better)fetch/dry-run wall: origin/main → branch: 746.9 ms median → 56.1 ms median (13× better)
  • ›Main keeper: make CI on main green (public-api)42e4affb

    Refreshes the committed grit-lib API snapshot so it records the already-landed revwalk follow-history surface. This is generated metadata only and does not alter runtime behavior.

  • ›Quality #75: Simplifyfce803fe

    Removes repository-discovery writers for GIT_TRACE_SETUP and GIT_TRACE2_PERF that existed only for the deleted upstream shell harness. It also removes a dead GRIT_DEBUG_SAFE_DIR field and replaces trace-file line-coverage assertions with direct assertions about discovered git directories, worktrees, and bare state.

    clean status, exact merge base to branch (500 order-balanced runs): 7.9 median ms → 6.0 median ms (31% better)dirty 1% status, exact merge base to branch (500 order-balanced runs): 8.4 median ms → 6.6 median ms (27% better)clean status, system Git to branch: 8.5 median ms → 6.0 median ms (41% better)dirty 1% status, system Git to branch: 9.2 median ms → 6.6 median ms (39% better)
  • ›Main keeper: make CI on main green (public-api)4c23db8a

    This review branch contains no changes: its tip is identical to origin/main. The public-api failure it was assigned to fix had already been resolved on main by refreshing the generated snapshot, so this review validates the existing fix rather than proposing another patch.

  • ›Speed: log/json-100@git-git3b3ab65b

    This removes repeated work from streamed log listings and filtered packed-ref decoration scans, batches each JSON row into one write, and regenerates hot-text layouts from both release and dist profiles. The result is byte-identical log JSON with substantially lower target latency and resident memory.

    log/json-100 release wall time: 4.6 ms → 4.1 ms (13% better)log/json-100 release peak RSS: 24.3 MiB → 18.1 MiB (35% better)log/json-100 time ratio vs system Git: 1.6 ratio → 1.4 ratio (13% better)log/oneline-100@git-git wall time: 4.0 ms → 3.6 ms (11% better)
  • ›Speed: show/head@git-git2e5c937e

    Keeps small tree/commit diffs on the caller thread and starts DiffContent workers at 16 changed entries by default. The same helper governs structured show/diff output, streaming structured diffs, and stat computation, while `ThreadPolicy::min_items_override` remains available to move the threshold.

    origin/main→branch show/head@git-git wall: 2.2 ms → 2.2 ms (0% worse)origin/main→branch show/head@git-git RSS: 9.2 MiB → 9.3 MiB (0% worse)origin/main→branch show/head@synth-10k wall: 2.7 ms → 2.3 ms (17% better)origin/main→branch show/head@synth-10k RSS: 11.0 MiB → 10.1 MiB (10% better)
  • ›Main keeper: make CI on main green (public-api)0ccc2e35

    Refreshes the committed grit-lib public API snapshot to match the API generated by the pinned cargo-public-api toolchain. There is no Rust behavior, API implementation, test, CLI, on-disk format, or wire-protocol change.

  • ›Quality #76: Library API quality21a765a3

    Config boolean, integer, and color parsing now returns structured `ConfigValueError` variants, while keyed accessors attach the offending key through `ConfigError::InvalidValue`. Callers propagate those typed failures, and 23 redundant case-variant fallback lookups are removed because config keys are already canonicalized case-insensitively.

    config/get-origin@small median wall: 1.6 ms → 1.5 ms (2% better)config/get@small median wall: 1.5 ms → 1.5 ms (3% better)config/set@small median wall: 1.7 ms → 1.7 ms (0% worse)
  • ›Main keeper: make CI on main green (fmt)21a765a3

    Reformats one function call in the worktree collision scan to rustfmt's canonical single-line layout. There is no semantic, API, on-disk format, wire protocol, test, or documentation change.

  • ›Perf regression 2026-10-31: add/all-ignorecase@synth-100k, add/all@synth-100k, add/modified@big-blobs +2315b5e252

    This change removes several accidental quadratic scans and redundant reads from status/add, worktree diff, hard reset, rebase planning, and reverse-index creation. It also batches attribute-blob fetching during large non-sparse checkouts, restoring blobless-reset correctness and substantially reducing clone/repack memory.

    status/dirty@synth-100k median (git controls 107.9/97.4 ms): 303.9 ms → 54.1 ms (5.6× better)add/all@synth-100k median (git controls 156.1/144.3 ms): 660.2 ms → 118.0 ms (5.6× better)add/all-ignorecase@synth-100k median (git controls 158.3/146.1 ms): 675.2 ms → 172.5 ms (3.9× better)add/modified@big-blobs median (git controls 225.4/216.5 ms): 169.0 ms → 147.8 ms (14% better)
  • ›Main keeper: make CI on main green (fmt)49c4e517

    Reformats one private grit-lib diff import list into rustfmt's required single-line layout. There is no behavioral, API, format, protocol, test, or documentation change.

  • ›Main keeper: make CI on main green (public-api, fmt)6f80f84f

    Refreshes grit-lib/public-api.txt to match the API currently exposed by origin/main. This is a generated snapshot-only correction; it does not alter library or CLI behavior.

  • ›Quality #77: Move logic into grit-libff57f1b0

    Fetch ref updates now expose a coarse typed kind in grit-lib, and push outcomes answer whether they contain updates or rejections. The CLI consumes those APIs for human, JSON, and Markdown rendering, eliminating duplicate counts and a stringly typed push-rejection branch while preserving output contracts.

    fetch/incremental@git-git median: 50.3 ms → 49.1 ms (2% better)push/force-with-lease@git-git median: 1,108 ms → 1,109 ms (0% worse)push/incremental@git-git median: 541.2 ms → 543.5 ms (0% worse)
  • ›Main keeper: make CI on main green (public-api)da9d6d56

    The landed commit synchronizes the checked-in grit-lib public API snapshot with the API already implemented on main, including the DiffEntriesResult return type and other accumulated surface changes. There is no remaining review-range delta because HEAD and origin/main are both 6f80f84f3.

  • ›Speed: log/oneline-100@git-git864c8d1e

    Adds a bounded pack inflate path that reads whole packed commits only through the subject needed by one-line listings, while retaining full reads for deltas, loose/custom stores, JSON, decorations, and diff output. Human/Markdown `log` and `revs` streams opt into that extent; public API, tests, docs, benchmark baselines, and hot-text layout are updated together.

    oneline-100 alternating median (system Git 3.029 ms): 3.5 ms → 3.3 ms (6% better)oneline-100 daily median (Git ratio 1.184→1.079): 3.6 ms → 3.1 ms (18% better)log/all daily median: 46.1 ms → 41.3 ms (12% better)json-100 alternating median: 3.8 ms → 3.9 ms (3% worse)
  • ›Speed: restore/hunks@large-file63ede83e

    This removes redundant work and transient allocations from partial worktree restore and the shared line-diff/apply paths. Hunk restore now avoids unnecessary hashes/stat refreshes, unified diff lines use one compact record table, and blob patching borrows line slices instead of cloning them.

    restore/hunks wall, current origin/main → branch: 15.2 ms → 12.0 ms (26% better)restore/hunks RSS, current origin/main → branch: 26.2 MB → 25.7 MB (2% better)restore/hunks wall, exact merge-base → branch: 17.1 ms → 12.0 ms (42% better)restore/hunks RSS, exact merge-base → branch: 32.1 MB → 25.7 MB (25% better)
  • ›Quality #78: Dead code and dependencies3977b882

    Removes 14 warning variants and 6 top-level error variants that no production path constructs, together with their CLI formatting and kind mappings. It also deletes several Git test-harness/environment branches that always evaluated to constants because they read `Environment::empty()`, consolidating index version selection while retaining the effective index, diff, protocol, cache-tree, and revwalk behavior.

    add/modified@synth-10k median (system git 16.707 ms): 10.3 ms → 10.4 ms (1% worse)diff/worktree@synth-10k median (system git 7.058 ms): 6.5 ms → 6.6 ms (1% worse)status/clean@synth-10k median (system git 7.196 ms): 7.1 ms → 6.6 ms (8% better)
  • ›Main keeper: make CI on main green (public-api)db306236

    Refreshes grit-lib's checked-in public API snapshot to match the library surface already present on main. It records the Cancelled error variant in generated sort order and the revwalk iterator types as non-Send/non-Sync, restoring the dedicated CI check without changing executable code.

  • ›Main keeper: make CI on main green (public-api)3344230e

    This review branch contains no changes: it points at the same commit as current origin/main. The public API snapshot repair it was intended to carry had already landed on main through d3aab95fd, so integration would be a no-op.

  • ›Speed: status/clean@synth-10k70947982

    This change reduces status startup residency by grouping hot functions by workload set, includes every alias of linker-merged functions, and deduplicates per-file attributes in the racy index scan. It refreshes the generated linker maps and records reproducible status benchmark reports and methodology.

    status clean synth-10k paired median: 7.1 ms → 6.9 ms (2% better)status clean synth-10k daily peak RSS: 11,260 KiB → 11,548 KiB (2% worse)status clean small paired median: 2.8 ms → 2.8 ms (1% better)
  • ›Main keeper: make CI on main green (test)

    Restores four local helpers used by the custom object database porcelain integration tests. This repairs the integration-test compile failure caused by an earlier conflict resolution without changing library or CLI behavior.

  • ›Speed: gc/pack-refs@synth-10kc6f34d88

    This change replaces the files-backend pack-refs HashMap rebuild with a single validated, borrowed merge of existing packed refs and sorted loose refs. It also avoids full object inflation for tag detection, reuses loaded config for lock waits, and takes the packed-refs lock before reading to prevent lost concurrent rewrites.

    gc/pack-refs time: 3.3 ms → 2.8 ms (18% better)gc/pack-refs peak RSS: 10.2 MiB → 9.7 MiB (5% better)system git gc/pack-refs control: 1.7 ms → 1.7 ms (2% better)branch/delete time: 3.9 ms → 3.5 ms (11% better)
  • ›Main keeper: make CI on main green (public-api)fc70734f

    Refreshes the committed grit-lib API inventory so the pinned public-api generator recognizes IndexPathSkipPredicate in its current sorted location and trait-object spelling. This is generated metadata only; runtime code and behavior are unchanged.

$ git log --oneline 217a3104..v0.5.28compare on GitHub ↗
fc70734fMerge branch 'factory/public-api-snapshot-a3fd'
15e6c155Refresh grit-lib public-api snapshot for IndexPathSkipPredicate
c6f34d88docs: record the pack-refs merge and its gc, clone and ref rows
cde0d658refs: pack loose refs with one borrowed merge instead of a HashMap rebuild
8c02c172Refresh grit-lib public-api.txt for describe module
99721b76Fmt describe first-parent test assertion
a30991ebDocs: note first-parent tag reachability in describe
962d4cfaFix describe tag selection and formatting
0522243eAdd RevWalk::describe and grit describe for nearest tags
c3d399adRestore custom_odb_porcelain test helpers dropped in merge
dc801553Merge branch 'factory/push-refspec-snapshot-f7f9'
1c0ce37aFix push refspec planning scanning all local refs per refspec
325bb3ceMerge branch 'factory/reset-hard-rss-6421'
70947982Merge branch 'factory/status-clean-10k'
17f9e61cMerge branch 'factory/ref-tx-delete-reflog'
3f2ba491Apply skip_index_path when restrict_to_paths is set on blob scan
acecf8e8Fixes #2439: cut hard-reset peak RSS on wide tree transitions
b3771bdbMerge branch 'factory/custom-odb-hotfix-2444'
d15113e5Fix custom_odb_porcelain helpers dropped in rebase conflict
8b52c19eMerge branch 'factory/fix-is-bare-linked-worktree'
8cce2eccFix reftable reflog visibility after ref delete
52e1e56bDocument the workload-set hot text layout and record status baselines
4380e148Regenerate the hot-text layout for the current tree
ea8ef3cfStore each racy entry's attributes once per distinct set
1cdb89daLay out hot code by workload set and list every alias of a merged function
30287d4bMerge branch 'factory/fix-full-tests-compile'
16167e43Fix full-tests migration semantics dropped in review
08ed96f9Fixes #2444: update full-tests integration for non_exhaustive API
a258013fMerge branch 'factory/fsck-connectivity-only-perf'
32d68be8Merge branch 'factory/histogram-perf-6456'
71eeaafdFix Repository::is_bare for linked bare worktrees
274c154eMerge branch 'factory/custom-primary-contains-2446'
894d09c6Fix histogram diff memory blow-up and anchor scan cost on repeated lines
d639ff3aFixes #2449: fast fsck --connectivity-only via header-only kind map
2d0a4a55Merge branch 'factory/upload-pack-unshallow-2447'
3dfb05b9Fixes #2445: flush ODB before rebase stop and exec conflict HEAD updates
51f9c780Fix reftable delete+ReflogUpdate disposition
4dfaec2cFix reftable reflog clear order in same-batch transactions
2e5aead9Fixes #2450: drop reflog when RefTransaction deletes a ref
52584e62Batch custom-primary tag validation through contains_many
939cc8cdBatch tag target checks for custom-primary local fetch
5e696ac5Fix local fetch into custom primary calling contains per object
191504abMerge branch 'factory/reftable-compaction-tombstone-e1d3'
859fa9c5Stop depth walk at server shallow boundaries
40fbeb5fFix upload-pack bogus unshallow on multi-want re-shallow
1d592948Merge branch 'factory/odb-flush-publish-6431'
0a22d95cAdd custom-primary pick, revert, and conflict overlay tests
074bfb55Fix reftable non-bottom compaction dropping delete tombstones
d27b5498Merge branch 'factory/log-follow-absent-path-6420'
bc42ec00Flush ODB before rebase --exec HEAD updates; refresh public-api
6690130eFix ODB flush before ref and index publish (Fixes #2443)
9a9f7ed3Fix log --follow when path absent at start commit
db8fb10bFix merge overlay persist on custom object store primary
f6a04d9eCompare post-grit status to baseline from system git reset
3344230eMerge branch 'factory/public-api-6409'
db306236Rebase: pass sparse hint into coalesced tree diff
d3aab95fRefresh grit-lib public-api snapshot after ApplyError and revwalk changes
c69c4b0cDerive coalesced delete paths from old tree OID after index align
7797e500Gate coalesced hard-reset deletes on wholesale safety; keep untracked
fd7d69edFixes #2439: coalesce hard-reset tree diff to cut peak RSS
a60cc44dFix exec failure rollback to keep untracked files created while stopped
a0f31aa4Fix exec restore: do not delete user untracked files on abort
d719bcddFix exec continue: preserve grit-plan and rollback failed commands
c7ff5808Fix #2432: stop resumably on rebase --exec replay conflict
dfff8a19Merge branch 'factory/sparse-switch-reset-perf-6390'
a951b4b1Fix cone sparse switch/reset staying proportional to the checkout cone
0fef76aeReview: restore worktree backups on remove_paths partial failure
93782d3fReview: fix cancel rollback, ApplyError::Cancelled, index_apply tests
995710b9Fix #2433: remove index locks when add/mv/rm is interrupted
bcc923a3Fix lazy hide frontier loss on generation cutoff
a9e35ecdFix lazy hide misclassifying include-side tips on divergent ranges
fe038937Fixes #2438: lazy hide-side paint for path-limited range revwalks
0ed25210Merge branch 'factory/diff-basename-o-n-6387'
46788082Merge branch 'factory/corrupt-object-errors-6353'
5c207079Fix exec restore: do not delete user untracked files on abort
2da61a22Fix exec continue: preserve grit-plan and rollback failed commands
15314875Fix #2432: stop resumably on rebase --exec replay conflict
8e3b3038Fix quadratic basename rename preflight on disjoint repeated basenames
ff16d54eFix log --follow when path is absent at start commit
72b9e6f7Merge branch 'factory/log-follow-perf-6382'
6335fb2aMerge branch 'factory/issue-2397-pick-conflict'
3977b882Merge branch 'factory/dead-env-knobs'
fe4c4e10Run rustfmt on warm-hint pack regression test
5af1444eFix warm-hint regression test to exercise hinted-pack shortcuts
b72a4181Address review: hinted-pack wrap, SHA-256 loose paths, public-api
9be56ceeFix #2430: name corrupt loose and packed object read failures
fce443b7Fixes #2435: speed up dense revwalk --follow with bloom and cached path lookups
97da4583Merge branch 'factory/merge-histogram-default-6357'
063eb68fDrop environment knobs read from an empty Environment
566458cbRemove Error and Warning variants nothing constructs
63ede83eMerge branch 'factory/restore-hunks-speed'
68121f65Apply rustfmt to merge_file after rebase
a4d15e47Merge branch 'factory/files-refstore-phantom-reflog'
d446aa93Fixes #2397: scope repeated-anchor guard to EOF tail base range
a6e67193Fixes #2397: align merge-ort tail conflicts without broad EOF heuristic
2cd8ac6eFixes #2397: conflict when pick appends after deleted tail
20057119test: fix merge_ort #2431 harness imports and open_grit_repo
03215c23Fixes #2431: default ort tree merges to histogram diff
864c8d1eDocument commit-prefix reads and record the one-line log rows
ce787efbRegenerate the hot-text layout for the commit-prefix read path
bef01f58Stream one-line log and revs listings with commit prefixes
915ef1d1Read commits only through their subject for one-line walks
9e26e2a8Merge branch 'factory/push-tag-reject-6335'
38b43301Lock every reflog append destination for log-only symrefs
59a2d990Address review: symref peel refresh, reflog locks, lock cleanup
3c563c63Fix FilesRefStore phantom reflog on failed CAS
312256dcReview: wire tag rejection before fetch-first; keep message crate-private
9ef18c9bFixes #2425: reject push that moves existing remote tag without force
aaded59fDocument the hunk restore speedup and record its daily rows
51321704Apply hunks to a blob without copying line bytes
372d4fd0Shrink the line diff's transient memory
210d384bSkip redundant hashing when restoring hunks on the worktree
672cd5c6Fixes #2429: use sorted index lookups in clean dir mode
da9d6d56Merge branch 'factory/push-relative-path-6345'
8ed4f99eMerge branch 'factory/blame-move-dedupe-6337'
7720efa5Fixes #2428: accept relative paths as grit push remotes
6e9b3ed5Fix blame -M/-C duplicate rows for repeated identical lines
ff57f1b0Render fetch and push outcomes from library classification
74b58a35Classify fetch and push ref updates in grit-lib
6f80f84fSync grit-lib public-api snapshot after DiffEntriesResult return type
dc410cd1Fix #2411: basename rename detection follow-up
49c4e517Run rustfmt on grit-lib diff module imports
15b5e252Record before/after daily baselines for the 2026-10-31 regressions
0721567dFetch attribute blobs with the checkout batch in chunked checkouts
8a10b5b9Read only the pack trailer when writing a .rev sidecar
037f5a84Read the index once per hard reset and look up restricted paths in order
a674218aDecide an up-to-date rebase before walking the replay range
2d0590eaHash racy hard-reset paths up to 1,024 files instead of spooling them
7bc7b774Skip hashing size-changed files and old binary blobs in worktree diffs
d60a2d85Look up gitlinks by binary search when probing for embedded repos
101be3c7Merge branch 'factory/diff-rename-basename-e669'
1058fe52Merge branch 'cursor/fix-pack-ingest-rev-rss-02ed'
ce7a31deMerge branch 'factory/ci-fmt-worktree-guard'
58425b16fix: pack ingest peak RSS by reading pack trailer for rev sidecar
10a1f60cfix: merge pathspec-limited tree diffs with DiffEntriesResult
f7fe91d4Fix rustfmt in worktree_guard untracked scan
20527877Merge latest main into rename basename branch
21a765a3Drop redundant case-variant config lookups
237a575aReturn typed errors from config value parsers
a37035baMerge branch 'factory/diff-pathspec-rename-2422'
2e58110cMerge branch 'factory/rm-dir-perf-2424'
8851056bfix: restore json_error_kinds integration module after rebase
b4e1bb02fix: restore pathspec prefilter in diff file targets
0ffd05e2fix: worktree_entries_in arity and rustfmt after rebase
29c19d2eSkip basename similarity when basenames are unambiguous
4be9dc28Address review: propagate rename-limit stderr and reuse rename blob cache
a3537f71Fix rename detection basename pass before renameLimit
c0b65931Fixes #2422: apply pathspecs before rename detection in tree diffs
ba8a3b54Merge branch 'factory/fsck-alternates-perf-02a9'
178dc366Fixes #2424: speed up grit rm directory pathspec collection
0ccc2e35Merge branch 'factory/fix-public-api-6297'
7d1089bcMerge branch 'factory/notes-rewrite-2418'
4b7b951cMerge branch 'factory/gc-index-roots-6289'
542aedb6Fix pack idx verify and Odb-owned loose negative cache
afaf27e1Fix alternates ODB pack rescan and fsck over shared object stores
832c3d6eRefresh grit-lib public-api snapshot for pathspec exports
b80d942dAddress review: preflight notes rewrite and snapshot CAS
36404830Fixes #2418: copy notes when history rewrites commits
3edaa6f4Fix concurrent bundle cache races with flock claim
b93344d7Fixes #2421: include all worktree indexes in GC reachability roots
2e5c937eMerge branch 'factory/show-diff-pool-floor'
cbeb673dMerge branch 'factory/pager-missing-less'
d2084c76Merge branch 'factory/issue-2415-bulk-dir-file'
95319e4bKeep session_cancel in default integration suite
5eb61419Fixes #2413: add JSON error kinds for common CI failures
268c350fFix large mixed checkout regression tests
f0a803caFix bulk checkout index loss on dir-to-file transitions
bd1961a0Honor index modes for attribute pathspec matching
03b6b66fFix attr pathspec matching for restore, stash, and rm
a6a06cffDocument the diff pool floor and record its daily rows
a3efc8cbDiff small commits on the calling thread
4c69c844Route env-assignment and expansion pagers through the shell
abde188cFixes #2408: fall back to terminal when pager cannot start
964d4c35Merge branch 'factory/gc-worktree-index-roots-e26c'
3b3ab65bbench: record log rows before and after the streamed JSON listing fixes
06e83553Lay out the hot code path for release builds too
de1ce859Write each streamed JSON row in one call
dc715b90log: set up a streamed listing once
433df608Scan filtered packed-refs listings in one pass
94f750f8Skip gitlink OIDs in resolve-undo reachability roots
a49b6a86Fix gc prune dropping blobs reachable only from linked worktree indexes
4c23db8aFixes #2409: stream --json --filter without extra JSON copies
a71d3065Allow clippy too_many_arguments on tree diff extras entry point
c39c3dabFix attr pathspecs in tree pathspec restrict union
3701d965Fix pathspec prefilter for tree-only staged deletions
c357c513Fixes #2406: warn and continue diff/stash on unreadable worktree dirs
fce803feRemove the shell-harness setup traces from repository discovery
42e4affbMerge branch 'factory/public-api-check'
b8806cb1Merge branch 'factory/pull-overwrite-hints-6222'
98f4f355Refresh grit-lib public API snapshot after log --follow
325e0c50Merge branch 'factory/switch-symlink-untracked-d695'
f2ce93ecFix pull overwrite errors: list paths, hint, and JSON kind
0787d1eaFix wholesale untrack scan following symlinks on added paths
097d7cebFix mv destination trailing slash and move error messages
c9349e37Merge branch 'factory/short-hex'
73e8d02bMerge branch 'factory/fetch-deepen-speed'
b9f15106Use the library short-hex helpers in grit and grit serve
361b62a6Add ObjectId::to_short_hex and objects::short_hex
d22ff8f8Document v0 smart-HTTP fetch negotiation and its benchmarks
fe220860Use no-done on stateless v0 fetches
f7613ec2Fix v0 smart-HTTP negotiation and offer haves when deepening
6d2a10a4Serve the system git http-backend in tests
ac8d1db2First-parent forward scan for is_ancestor
82a02794Fixes #2316: log --follow for file history across renames
d2551aefMerge branch 'factory/diff-target-entries'
ed0d0a0aMake issue-2403 test compile and run standalone
d0caafb3Fix mixed-merge regression test for new auto-merge blob
3a96771ePersist stage-0 merge objects from overlay on conflict stop
686496cbFix dangling empty blob and merge conflict objects
a178d909Thin grit diff onto the library's diff target entries
0a7c9253Add Diff::entries and UncommittedDiff::{entries, peer_changes}
53ed5573Merge branch 'factory/windows-symlinks-2402'
a4540501Merge branch 'factory/fix-bundle-fetch-no-refspec'
039c8f62Fix Windows build and address symlink review feedback
c9cde2caFix Windows checkout symlinks when core.symlinks is true
44cb7df2Fix bundle fetch with no refspec when HEAD is missing
fa7a7bb7Fix CI: refresh public-api snapshot and rustdoc link
7d41ce6dDrop unused similar compaction helpers after xdiff merge diffs
c9196426Borrow byte interner records; share merge Myers interning
99ee42f2Address review: byte-accurate merge diffs and Myers perf
a28f5b60Fix merge false conflicts from diff hunk alignment (Fixes #2391)
674dd91eMerge branch 'factory/push-labels-2398'
3ef8f804Merge branch 'factory/odb-write-no-probe-1c25'
c0d44ea4Merge branch 'factory/fsck-truncated-parent'
5c987063Merge branch 'factory/fix-2400-serve-hooks'
a7156c54Merge branch 'factory/revs-log-stream-ddd2'
05033dc8Record post-receive callback errors on ReceiveOutcome
506477fbFix serve push hooks and protocol serve session API
5d171c88Fix streamed revs truncated flag and JSON stream test
cbc1216eFixes #2388: stream plain revs and log walks instead of buffering
af5742afFix merge compaction keys for non-UTF-8 text blobs
7f67af69Fixes #2397: compact merge diffs like git merge-file
dbb47279Fix SQLite example flush before reopen and mutex batch state
5a50acd6Fixes #2394: skip Odb contains probes on custom-primary writes
8261e7b6Fix push human labels and diverged non-ff hints
b61f36e4Fixes #2386: avoid fsck panic on truncated commit parent lines
fac874b3Merge branch 'factory/revwalk-default-order-6470'
4d3675b1Merge branch 'factory/log-group-by-author-perf'
4e8f7757Fix count-only author grouping to use exact name/email keys
5c4c7237Fix #2452: speed up log --group-by author vs git shortlog
d9ce5495Keep object-filtered rev-list walks on the full engine
ecaf0ab4Fix unlimited default RevWalk order to match git rev-list