grit://cooking/changelog← what’s cooking
Monday, November 2, 2026 · release v0.5.29

101 changes landed on main

72 issue fixes · 29 other changes

grit v0.5.29 fixes many correctness bugs (non-UTF-8 paths, shallow clones, case-insensitive filesystems, rerere), adds cancellation and HTTP timeouts, and speeds up describe, notes, merge and diff.

  • ›describe rewritten as one walk, replacing per-tag history walks that took 52 s on git/git (#2461)
  • ›fetch/clone now honor SIGINT/SIGTERM and http.lowSpeedLimit on stalled servers; push, rebase, worktree add and hooks cancel cleanly
  • ›Non-UTF-8 paths fixed in add, commit, clean, stash, switch and merge; non-UTF-8 tag messages accepted
  • ›Fixed macOS case-insensitive ref and directory handling, shallow-clone push/show/log/blame, and receive-pack updateInstead
  • ›New: rerere in merge/pick/rebase, pull --rebase fork-point, refs/replace in log/diff, switch -c upstream, stash drop prints commit id
  • ›Faster: merge renames, notes add, restore --staged, diff output, local push memory, first status after commit; json --filter memory cut
$ git log bd1c65f8..v0.5.29219 commits on GitHub ↗300+ files +15,131 −5,382
Highlightsmeasured by the reviewers, before → after
grit-lib: dropped Repository handles keep pack/idx mmaps alive (thread-local cache): 200 opens = 16 GB VmSize
VmSize growth after 200 fresh open/read/drop cycles on a 178 MB pack356903× better████████████████████████ 35,690,260 KiB █ 100 KiB
http.lowSpeedLimit/Time never fires on a silent or stalled-body HTTP server: fetch/clone hang forever (git: 5 s)
Absolute timeout gap versus system Git (5.016s)351× better████████████████████████ 5.3 seconds █ 0.015 seconds
switch_repository ignores Cancel until the work is done: pre-set cancel on a 150k-rename switch takes 6 s
Pre-set cancellation latency on 15k-file, 30-directory rename switch (base → branch median, 5 runs)248× better████████████████████████ 499.1 ms █ 2.0 ms
grit restore --staged is O(paths x index size): re-staging/unstaging 64k paths takes 8 s (git 0.06 s)
64k staged removals median (system Git control: 0.032379s)106× better████████████████████████ 6.9 seconds █ 0.065 seconds
macOS (core.ignorecase=true): reset --hard/switch removing a directory is quadratic: 64k files 14 s (git 2.3 s)
32k hard-reset median user CPU (system Git median: 0.012s)100× better████████████████████████ 7.0 seconds █ 0.070 seconds
issues fixed72
  • ›grit switch -c <name> <remote-branch> sets no upstream (git does by default; grit branch does)#2460e4a44da4

    Creating a branch with `grit switch -c work origin/main` now records `origin/main` as its upstream and reports that relationship in human, JSON, and Markdown output. Users can force or suppress the behavior with `--track` and `--no-track`, while automatic mode follows the existing branch tracking configuration.

  • ›--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)#24094b09e375

    Filtered JSON command outcomes are now converted directly from serde data into jaq values, with repeated object keys sharing one allocation. Filter results are still written directly as compact JSON, and unfiltered JSON continues through serde_json unchanged.

    Cargo .files|length wall time: 1.0 seconds median of 4 alternating runs → 0.763 seconds median of 4 alternating runs (34% better)Cargo .files|length peak RSS: 503,514 KiB median → 449,606 KiB median (12% better)Cargo identity filter wall time: 1.4 seconds median of 4 alternating runs → 1.1 seconds median of 4 alternating runs (21% better)Cargo identity filter peak RSS: 499,250 KiB median → 450,054 KiB median (11% better)
  • ›grit push from a shallow clone to a remote that lacks the history fails 'object not found: <boundary parent>'#2463f97e90d1

    Pushes from shallow clones now terminate object enumeration at local shallow boundaries instead of reading absent parents. Wire pushes announce those boundaries to receive-pack, while local-path pushes classify required shallow roots, honor `receive.shallowUpdate`, expose the temporary shallow file to hooks, and persist accepted boundaries.

  • ›receive-pack on case-insensitive FS (macOS) accepts refs Foo and foo; later reads of Foo return foo's value#24719565a23c

    On files-backed repositories that compare paths case-insensitively, ref creation now checks existing loose and packed names plus sibling creates under ASCII case folding. A colliding receive-pack create is refused before it can write a shadowing loose ref, and the wire status reports `reference already exists`.

    Non-colliding receive-pack create, 10k packed refs, debug grit median: 158.4 ms → 157.0 ms (1% better)Same 10k-ref scenario: system git vs review grit debug: 30.9 ms → 157.0 ms (80% worse)
  • ›skip-worktree: grit restore . destroys the local edit; stash/reset clear the flag so grit commit commits it#2466d9cca2c3

    Preserves user-managed skip-worktree entries through restore, stash, and reset so local-only files are not exposed, staged, overwritten, or deleted unexpectedly. The round-two follow-up closes the remaining cross-tree reset cases, including atomic refusal when deleting a dirty skipped file.

  • ›Cancelled/SIGINT'd reset --hard that deletes a directory leaves the files deleted (no rollback), HEAD/index unchanged#246582664480

    Makes hard-reset cancellation transactional when a whole tracked directory is removed as one coalesced checkout operation. It restores every tracked file and any dirty worktree bytes after cancellation without treating genuinely untracked nested files as rollback data.

  • ›rerere.enabled is ignored: grit merge/pick/rebase/commit never replay or record resolutions in rr-cache#2468e645ac63

    Grit now participates in Git's shared rerere cache throughout merge, pick, revert, and rebase conflict workflows. It replays prior resolutions at conflict stops, records newly resolved postimages on commit/continue, clears abandoned state on abort, and reports rerere actions in structured and rendered CLI output.

  • ›serde-models: no Serialize for blame/describe/tag/worktree results; CommitOutcome date is a raw 'ts tz' string#247421b00c58

    Exposes blame, describe, tag-list, and worktree result models directly through the optional `serde-models` feature. It also makes commit outcome identity timestamps consistently RFC 3339, adds committer and full message data to `CommitOutcome`, and simplifies CLI commit rendering to consume those library fields.

  • ›receive-pack: receive.denyCurrentBranch=updateInstead moves the ref but not the worktree/index; dirty trees not refused#2470a6e5a8de

    This change gives receive.denyCurrentBranch=updateInstead its own typed policy and deploys accepted pushes into the checked-out worktree and index. It refuses dirty, unmerged, staged, hook-declined, and untracked-overwrite cases before moving the ref, while preserving warn/ignore behavior and handling linked worktrees.

  • ›refs/replace honored by grit cat but ignored by log, show, revs and diff: inconsistent history view vs git#247313c36255

    History operations now consistently apply `refs/replace` while retaining the original displayed object IDs. The round-2 follow-up makes replacement discovery work through loose refs, packed-refs, and the repository's reftable backend, and fixes the remaining commit-diff path.

    Loose replace-ref fixture rev count (system Git expects 2): 4 commits → 2 commits (2.0× better)Packed replace-ref fixture rev count (system Git expects 2): 4 commits → 2 commits (2.0× better)Packed fixture `grit revs --count HEAD` median over 20 warm runs: 6.6 ms → 6.6 ms (0% better)
  • ›grit add / commit fail 'I/O error: No such file or directory' while any untracked file has a non-UTF-8 name#2464d4395a83

    Untracked-cache scanning and staging now preserve non-UTF-8 path bytes end to end instead of opening a lossy replacement path. The round-3 correction recognizes reversible pathhex directory markers in raw-byte space, expands their contents, and prevents malformed empty-name trees.

  • ›reftable: load_cached_stack_tables self-deadlocks on a missing table (hangs forever; whole grit serve process wedged)#24771f83c3ba

    Reftable readers no longer hold the process-global stack-table cache lock while reading `tables.list`, opening or memory-mapping table files, or retrying after compaction. A missing table now returns an error instead of recursively locking the same mutex and wedging every later reftable request in the process.

  • ›Commit walks in a --shared clone (alternates) are ~3x slower than the same repo unshared and than git (#2419 residual)#2472b4630230

    This change gives each pack store reusable thread-local read state so primary-to-alternate object lookups stop discarding pack hints on every object. It also avoids repeated loose probes for a known-empty primary, loads commit graphs from alternate object directories, and preserves external pack discovery after an empty-primary miss.

    Cargo shared log --oneline -n 5000 mean: 129.8 ms → 113.4 ms (14% better)Cargo shared revs --all mean: 388.6 ms → 352.3 ms (10% better)Cargo unshared log --oneline -n 5000 mean: 54.5 ms → 60.4 ms (10% worse)Cargo unshared revs --all mean: 194.8 ms → 247.3 ms (21% worse)
  • ›grit serve has no graceful shutdown: SIGTERM kills in-flight pushes and clones immediately#247871fddd33

    Adds bounded graceful shutdown to the HTTP service used by `grit serve`. The first shutdown signal closes the listener and drains active requests, a configurable grace period caps that drain, and a second signal terminates immediately.

  • ›grit log/status print 7-char commit ids that grit itself rejects as ambiguous (git/git: 121 such commits)#24763ed1a2b7

    Human and Markdown commit identifiers now honor core.abbrev and extend through collisions, so copied identifiers resolve to the displayed object instead of becoming ambiguous in large repositories. The round-two follow-up exposes the helper at the canonical commit API path and removes the Markdown page's per-row abbreviation setup.

    grit log -n 1000 --markdown median on packed git.git: 9.3 ms (origin/main vs branch; 2 warmups, 9 runs) → 13.1 ms (origin/main vs branch; 2 warmups, 9 runs) (29% worse)branch Markdown log vs equivalent system Git median: 12.6 ms (system Git vs branch; 2 warmups, 9 runs) → 13.1 ms (system Git vs branch; 2 warmups, 9 runs) (4% worse)ambiguous reproduction id length: 7 hex characters (origin/main vs branch) → 10 hex characters (origin/main vs branch) (43% better)
  • ›Cancelling start_rebase during apply takes ~1.1 s to return (finishing takes 0.4 s; switch rollback 30 ms)#2475765230dc

    Rebase apply rollback now records each old tree entry’s object ID and mode in its journal, then restores clean tracked paths through the existing batched parallel checkout writer. Cancellation is checked before ref publication, and all rebase-session apply paths expose cancellation through the documented Library(Cancelled) shape.

    same-fixture flag-to-return cancellation latency (single release run): 1.1 ms → 0.947 ms (17% better)release perf guard cancelled rebase vs completion (branch fixture): 10.1 ms → 5.1 ms (98% better)
  • ›SIGTERM during a commit hook is ignored until the hook ends; a 2nd SIGTERM leaves a stale index.lock#2412afef1051

    Commit operations now observe cancellation while pre-commit and commit-message hooks are running. Cancellation force-stops the hook child, becomes the typed top-level cancellation error, and lets index-lock guards clean up before the CLI exits; merge commit hooks use the same path.

  • ›grit describe / RevWalk::describe on git/git takes 52 s (git: 46 ms): one full history walk per candidate tag#2461856d98f0

    This change replaces per-tag reachability counts with a single commit-graph-backed describe walk and reuses a stable within-bit when tagged commits are re-popped. The round-3 panic fix works, but the overall new walk becomes much slower than current main on the real repository this issue targets.

    git/git describe HEAD median wall time (review branch is a lower bound because run had not completed): 65.6 seconds → 300 seconds (78% worse)
  • ›grit diff --stat prints 'Bin' for binary files without 'old -> new bytes' sizes#2339e1e84dfc

    Binary diff-stat rows now preserve both side lengths and show them in human, Markdown, and JSON output. The round-2 follow-up completes the no-index path and regenerates every shared schema affected by the new fields.

  • ›Follow-up: Fix #2461: grit describe / RevWalk::describe on git/git takes 52 s (git: 46 ms): one full history walk per candidate tag#2461e64aa63c

    Replaces candidate-by-candidate history work in describe with one collected target-history walk and batched tag-cone propagation, while retaining exact Git-compatible commit distances. The final follow-up derives candidate eligibility from that same reachability set, including first-parent semantics, instead of performing separate ancestry walks per tag.

    grit describe --tags HEAD median on 82,668-commit/1,012-tag git.git clone without commit graph: 53.8 seconds → 2.4 seconds (22× better)
  • ›fetch/clone against a stalled HTTP server ignore SIGINT/SIGTERM/SIGHUP and library Cancel forever (no read timeout)#2469

    This change makes HTTP discovery and stateless upload-pack waits cancellation-aware by running blocking client calls on workers and polling cancellation on the owner thread. It also adds receive timeouts/low-speed handling, CLI signal coverage, and stalled-server library regressions.

  • ›status/add warn 'nested repository ... recorded as a gitlink' for a symlink to a repo dir (it is stored as 120000)#2482007f4674

    Symlink worktree entries are now rejected before embedded-repository resolution, so a link to `.` or another repository is treated as an ordinary symlink. Add outcomes also report an embedded gitlink only when classification or the resulting index entry actually says mode 160000.

  • ›merge/pick/rebase: rename on one side + delete on the other merges cleanly and resurrects the file (git: conflict)#248346b3f3f9

    This fixes a silent data-resurrection bug in three-way merges. When the incoming side renames a file that the current side deleted, grit now retains Git-compatible stage 1 and 3 entries at the new path and stops merge/replay operations as conflicted instead of committing the renamed file as clean.

  • ›describe (main, one-walk rewrite) reports 'no annotated tag reachable' when a tag is reached at two depths via a merge#2479198b5794

    Replaces the stale-key/re-pop describe traversal with a reachable-DAG settlement that finalizes each commit once. Tags reached through multiple merge depths are retained, exact range-count distances are computed, and exponential merge-ladder reprocessing is eliminated.

    61-commit ladder, 1 annotated tag, in-process debug mean: 0.592 ms/call → 0.720 ms/call (18% worse)61-commit ladder, 61 annotated tags, in-process debug mean: 6.3 ms/call → 6.0 ms/call (5% better)
  • ›grit clone <typo'd local path> reports success: creates an empty repo that clones itself#2480e21d02e4

    Local clone sources are now resolved and validated against the caller’s working directory before grit claims or initializes the destination. A typo therefore returns the existing typed remote-unreadable error instead of manufacturing an empty repository that points to itself; bundles and genuine empty repositories continue through normal clone handling.

  • ›merge/pick: false content conflict at end of file where git (ort) and git merge-file merge cleanly (cargo tests/*.rs)#248439377cce

    This change stops clean EOF appends from being turned into false conflicts when the other side has a large earlier deletion. It replaces a broad repeated-line suffix heuristic with a narrow, hunk-local anchor-slide check and updates regression coverage to distinguish Myers (`merge-file`) behavior from histogram/tree-merge behavior.

  • ›Follow-up: Fix #2469: fetch/clone against a stalled HTTP server ignore SIGINT/SIGTERM/SIGHUP and library Cancel forever (no read timeout)#24691b465a82

    Makes smart-HTTP fetch and clone waits interruptible without retaining borrowed cancellation pointers. Discovery and POST work run on HTTP workers while the owner thread polls a scoped Cancel, and normal configured HTTP behavior remains available for slow servers.

    Two-second delayed advertisement: origin/main vs review branch: 2.0 seconds → 2.0 seconds (0% better)
  • ›macOS (core.ignorecase=true): reset --hard/switch removing a directory is quadratic: 64k files 14 s (git 2.3 s)#2485060ed70e

    Scheduled-delete membership during case-insensitive wholesale directory checks is now backed by an ASCII-folded hash set instead of scanning every deletion for every disk entry. Both reset’s checkout safety walk and switch’s worktree-guard walk use the same internal lookup, reducing the reported quadratic CPU cost to near system-Git wall time.

    32k hard-reset median wall time (system Git median: 0.421s): 7.5 seconds → 0.544 seconds (14× better)32k hard-reset median user CPU (system Git median: 0.012s): 7.0 seconds → 0.070 seconds (100× better)32k detached-switch wall time (system Git: 0.433s): 7.6 seconds → 0.736 seconds (10× better)32k detached-switch user CPU (system Git: 0.012s): 7.1 seconds → 0.156 seconds (45× better)
  • ›grit describe / RevWalk::describe on git/git takes 52 s (git: 46 ms): one full history walk per candidate tag#246128d2f777

    Describe now has two bounded strategies: a generation-ranked backward walk when a usable commit graph exists, and an exact fixpoint computation over the reachable subgraph when it does not. The fallback removes the per-candidate full-history distance loop that caused the round-2 regression while retaining exact winner and tie-break behavior.

    git/git no-graph describe HEAD median latency: 1,787 ms (origin/main → review branch, 9 interleaved warm runs) → 1,810 ms (origin/main → review branch, 9 interleaved warm runs) (1% worse)git/git no-graph describe HEAD median peak RSS: 114,804 KiB (origin/main → review branch, 5 warm runs) → 114,760 KiB (origin/main → review branch, 5 warm runs) (0% better)
  • ›grit stash drop discards a stash without printing its commit id; stash apply/pop cannot take a commit id to recover it#24862ad9f917

    Stash drop now returns and displays the removed stash commit id, making accidental drops recoverable. Stash apply, pop, branch, and show accept typed stash references so commit ids can be used where appropriate, with versioned JSON schemas and updated human/Markdown output.

  • ›merge --check ignores -X ours|theirs: predicts a conflict that the real merge -X resolves (34 of 500 git/git merges)#2491ca3b1681

    `merge --check` now predicts the command actually requested when `-X ours` or `-X theirs` is present. Both attached and detached HEAD paths forward the already-built merge options, and a system-Git comparison locks in the resulting tree IDs.

    Warm merge-check runtime: 7.7 ms median, debug build, n=30 → 7.8 ms median, debug build, n=30 (2% worse)
  • ›grit gc in a repo with no commits writes a commit-graph that grit fsck then reports as corrupt (git accepts it)#2487c9d7c585

    This change makes empty-history commit-graphs internally consistent. The reader accepts a valid zero-entry graph already accepted by Git, while both filesystem writer entry points leave the graph untouched when there are no reachable commits; the CLI and docs describe the no-op clearly.

  • ›macOS/Windows: grit clone aborts and deletes the clone when remote has branches/tags differing only in case#24909a42a354

    This change lets a clone preserve branch and tag names that differ only by ASCII case by writing the initial ref batch to packed-refs. Later fetches continue past colliding loose refs with a typed rejection and reftable guidance, while receive-pack remains strict and rejects a same-batch Foo/foo push without storing either ref.

  • ›push killed by SIGTERM/SIGHUP leaves a partial tmp_pack in the remote's tmp_objdir-incoming dir; gc never reclaims it#24923a92e708

    Push and receive-pack now cooperate with SIGINT/SIGTERM/SIGHUP cancellation, including while deflating a large streamed blob, so interrupted local/SSH pushes drop their receive quarantine before refs can change. Garbage collection also recognizes aged abandoned receive quarantines, while the round-2 adjustment preserves upload-pack’s existing immediate first-signal termination behavior.

  • ›grit push of a whole repo to a local bare repo peaks at 2.1-2.3x git's RSS (git/git 1.46 GB vs 713 MB)#24930956b6eb

    Streams pack index rows alongside locally built pack bytes, releases sender-side reachability/cache state before ingest, and installs non-thin packs without rescanning their bodies. Local copy and push delta-search windows now default to 32 MiB, with correctness and Linux RSS regression coverage added.

    Local push peak process-tree RSS on 120.65 MiB packed fixture: 371,920 KiB → 310,928 KiB (20% better)Local push elapsed time on 120.65 MiB packed fixture: 6.8 seconds → 3.5 seconds (98% better)Peak RSS ratio versus system git (384,960 KiB): 0.966 ratio → 0.808 ratio (20% better)
  • ›pull --rebase / rebase ignore fork-point: commits a teammate dropped by force-push are re-applied (git drops them)#2488f8019286

    Grit now uses reflog-aware fork-point boundaries for implicit rebase and pull --rebase, so commits removed by a teammate’s force-push are not resurrected. It adds explicit fork-point controls, library APIs, Git-compatibility coverage, and matching documentation.

  • ›Ctrl-C during worktree add / rebase / history kills grit mid-write: half-built worktree, stale index.lock#2489393571fb

    Interrupt-aware cancellation now covers worktree creation, history rewrites/rebase, cleanup, maintenance, bundles, checkout, and checkin while preserving rollback and lock cleanup guarantees. The round-3 correction makes bundle output transactional with respect to an existing destination: only the temporary file is removed on failure, and replacement occurs only after a successful write.

  • ›grit-lib API clarity: 71 placeholder rustdoc lines on public items, two list_branches and two FileDiff types#24946057055c

    Replaces generated or copy-pasted rustdoc on public diff, command-runner, diagnostics, rerere, gitmodules, and commit-graph items with descriptions of actual semantics. It also explains which branch-list and file-diff APIs embedders should choose, documents rev resolution error behavior, updates the library guides, and adds a source-level regression guard.

  • ›Nightly test-full/coverage red: grit-lib `integration` test binary fails to compile with full-tests (115+ errors, E0639)#2444285dbe51

    Adds a small ConnectOptions constructor that selects Git protocol version while defaulting private cancellation state, then migrates all affected external integration tests to it. This restores nightly full-tests compilation without changing transport behavior or wire bytes.

  • ›merge moves the branch and writes its reflog in separate RefStore transactions; reflog failure is swallowed#2503c9019d6c

    Branch movement and its branch/HEAD reflog records now succeed or fail together through the established ref-store commit path. A reflog backend error is returned to the caller and leaves the branch tip unchanged instead of silently reporting a successful merge with missing history.

  • ›grit-lib: dropped Repository handles keep pack/idx mmaps alive (thread-local cache): 200 opens = 16 GB VmSize#24959fd677c5

    Dropping the final repository object store now removes that store's pack/index mappings from the dropping thread's live cache and saved snapshots. Weak lifetime tokens also prevent a dead store from being saved again when a thread switches cache contexts, while preserving the existing hot read path.

    VmSize growth after 200 fresh open/read/drop cycles on a 178 MB pack: 35,690,260 KiB → 100 KiB (356903× better)200-cycle elapsed time: 54 ms → 30 ms (80% better)
  • ›grit-lib: References::update("HEAD") silently detaches HEAD and no ref update writes a reflog (core.logAllRefUpdates)#249739278ce3

    References::update now follows symbolic HEAD to the checked-out branch instead of detaching it, and repository-scoped ref updates create reflogs according to core.logAllRefUpdates. Callers also gain explicit direct-update and compare-and-swap APIs, while round two makes HEAD and branch reflog decisions independent when only one existing log should be preserved.

  • ›grit diff (human output) issues one tiny write per line: 49 MB diff to a file 2.5 s (1.75 s sys) vs git 0.9 s#2501a3f0fa38

    This change routes streamed diff bodies through a locked stdout writer that uses a 64 KiB buffer for pipes/files while leaving terminals immediate. It also writes the plain gutter directly and applies the same buffering policy to patch-text and no-index patch output, with regression coverage for batching and byte parity.

    15.30 MB human diff median wall time (5 runs): 0.246 seconds → 0.181 seconds (36% better)Pipe reader read calls: 13,408 reads → 284 reads (47× better)Pipe reader average chunk size: 1,141 bytes/read → 53,868 bytes/read (47× better)
  • ›macOS: switch/merge/pick refuse 'untracked file would be overwritten' on a case-only rename when 32+ paths change#2499f9f397f0

    Fixes false untracked-overwrite refusals for case-only renames when a large delete set selects the wholesale scan. The guard resolves the filesystem’s actual spelling once through a per-scan cache, then checks that spelling against case-aware tracked and scheduled-delete lookups while retaining real collision detection.

  • ›--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)#24095dd76f7e

    This change preserves the direct typed serde-to-jaq JSON filter pipeline after the rejected bytes/reparse experiment and documents why that invariant matters. It adds a large diff-shaped unit test showing the direct conversion remains faster than serializing JSON bytes and parsing them back.

    Unfiltered JSON median time: 0.412 seconds → 0.360 seconds (14% better)Unfiltered JSON median peak RSS: 186,388 KB → 186,232 KB (0% better).files|length median time: 0.775 seconds → 0.750 seconds (3% better).files|length median peak RSS: 475,324 KB → 478,264 KB (1% worse)
  • ›Single-writer-slot RefStore: updates to disjoint refs fail with LockHeld after 100 ms (backoff starves)#2500f2075894

    Adds a public FIFO gate for ref backends with a single in-process writer slot and threads lock-wait budgets through `RefStore::prepare_for`. Memory-backed and wrapped stores now serialize contending writers fairly, release admission on commit/abort/drop, and skip timed-out tickets so disjoint updates no longer starve.

    2,000 uncontended in-process repository ref updates: 40.4 ms median over 10 release-mode samples → 40.6 ms median over 10 release-mode samples (1% worse)2,000 sequential ref updates: system Git processes vs branch grit-lib in-process: 1,556 ms → 40.6 ms (38× better)
  • ›bundle unbundle of an incremental bundle without its prerequisites says 'corrupt object: N deltas unresolved'#2498436ac746

    Incremental bundle unbundle now validates all prerequisite commits before touching the pack. Missing prerequisites produce a typed library error carrying every `(oid, subject)` pair and a CLI error listing the object ids; the destination object database remains untouched.

    Missing-prerequisite unbundle median latency (10 fresh repositories, debug CLI): 13.1 ms → 4.9 ms (2.6× better)Latency multiple versus system Git (system Git median 1.081 ms): 12.1 x → 4.6 x (2.6× better)
  • ›upload_pack on a repo with an injected RefStore never advertises HEAD (git clone over v2 checks out nothing)#2505b34c5e65

    Serving now reads HEAD and repository tips from the repository's injected RefStore when refs are not disk-backed. That restores v0 HEAD/symref and v2 ls-refs HEAD/symref-target advertisements and makes related upload/receive policy checks see memory-only refs.

  • ›reference-transaction hook gets all-zero old oid on the HEAD line when the repo uses an injected RefStore#2504e9d2da16

    Reference-transaction hook lines for log-only symbolic HEAD updates now resolve HEAD through its target ref, so subsequent commits report the real previous tip instead of an all-zero oid. The round-2 test parser correctly ignores the hook script's trailing delimiter, allowing the end-to-end regression to validate prepared and committed phases.

  • ›fetch/clone against a stalled HTTP server ignore SIGINT/SIGTERM/SIGHUP and library Cancel forever (no read timeout)#24698f5e243d

    Makes clone’s bundle-uri discovery and smart-HTTP waits interruptible by moving blocking HTTP operations to workers polled by the caller. It also wires configured low-speed windows into those waits and adds library/CLI regressions for stalled clone cancellation, signals, delayed responses, and process-environment hygiene.

    Progressing response duration observed before outcome (system git vs branch grit): 1.9 seconds → 1.0 seconds (47% worse)HTTP connections for progressing response (system git vs branch grit): 1 connections → 2 connections (50% worse)
  • ›serve::receive_pack cannot accept a push into a custom ObjectStore primary (quarantine is files-only), undocumented#250672741fd1

    Receive-pack can now accept pushes into repositories backed by a custom writable object store and custom ref store. Incoming objects remain isolated in a memory overlay through validation and pre-receive, then are copied to the configured primary at the same publication point used by files quarantine migration.

  • ›branch/tag create, switch -c and log --json cost ~25 us per loose ref: 20k loose tags: branch zz 570 ms vs git 4 ms#25021933df53

    Ref creation under core.ignorecase no longer loads every loose ref body. The replacement scans packed ref names and walks only relevant loose-ref directory components, including nested case-folded path collisions such as Foo/bar versus foo/baz; deletes avoid the create check entirely.

    20k loose tags: grit branch create median (system git 1.244 ms): 168.9 ms → 2.2 ms (78× better)20k loose tags: grit tag create median (system git 1.336 ms): 220.4 ms → 17.2 ms (13× better)20k loose tags: grit tag delete median (system git 1.409 ms): 69.3 ms → 2.0 ms (34× better)
  • ›switch_repository ignores Cancel until the work is done: pre-set cancel on a 150k-rename switch takes 6 s#2509a3a8a7af

    Switch now checks cancellation before taking the destructive path, during large tree-transition planning, and before each progressively sized wholesale-directory removal batch. A pre-set cancellation therefore exits before touching the worktree, while cancellation that arrives later still uses the existing journal rollback guarantees.

    Pre-set cancellation latency on 15k-file, 30-directory rename switch (base → branch median, 5 runs): 499.1 ms → 2.0 ms (248× better)Uncancelled 15k-file, 30-directory rename switch (base → branch median, 5 runs): 526.0 ms → 525.5 ms (0% better)Uncancelled rename switch comparison (system Git → branch Grit median, 5 runs): 419.0 ms → 525.5 ms (20% worse)
  • ›main regression: first status after grit commit/stash is 6 s on 250k files, 24 s on 500k (release: 0.35 s)#2507469bdcec

    This change replaces repeated full-index scans during untracked-cache refresh with raw-byte indexed lookups. It handles sorted indexes through the existing index search and unsorted indexes through one precomputed sorted snapshot, preserving non-UTF-8 path correctness while restoring first-status performance.

    250k-file first `grit status --short` wall time: 7.2 seconds → 0.370 seconds (20× better)250k-file clean status control: system Git vs reviewed branch: 0.279 seconds → 0.370 seconds (25% worse)
  • ›grit notes add / RepositoryNotes::add is O(total notes): 35 ms vs git 1.8 ms with 20k notes (19x), 25 MB vs 6 MB#24961db3a0a3

    This change replaces whole-notes-tree scans and rewrites for single-note reads, adds, previews, and removals with a path-local fanout walker and bottom-up tree splice. It keeps full scans for listing and falls back to the existing tree materializer for irregular layouts.

    500-note in-process RepositoryNotes::add: 1.9 ms/op → 0.696 ms/op (2.8× better)Branch embedded benchmark: grit vs system Git: 1.8 ms/op (system Git before, grit after) → 0.934 ms/op (system Git before, grit after) (95% better)
  • ›merge/pick/revert of a commit that renames N dirs of files is quadratic: 150k renames 57 s (git 9 s)#2508d1cb18b5

    This change makes merge, pick, and revert exact-rename augmentation deduplicate known rename pairs with one exact-value HashSet lookup per candidate instead of scanning the complete pair vector repeatedly. It adds a repeated-basename directory-move compatibility test against system Git and a Criterion benchmark for the hot path.

    Criterion exact rename merge (1,000 renames, point estimate): 11.9 ms → 8.5 ms (40% better)Criterion exact rename merge (2,000 renames, point estimate): 33.8 ms → 21.8 ms (55% better)
  • ›Follow-up: Fix #2469: fetch/clone against a stalled HTTP server ignore SIGINT/SIGTERM/SIGHUP and library Cancel forever (no read timeout)#24698fb317bd

    Removes whole-request low-speed deadline behavior from PendingHttp waits and locks that behavior in with unit and real SmartHttpTransport discovery regressions. Healthy responses can now run longer than http.lowSpeedTime as long as the body reader observes sufficient throughput, while cancellation remains polled.

    Healthy slow-discovery elapsed time reached before result: 1.0 seconds → 7.2 seconds (7.2× better)
  • ›Follow-up: Fix #2496: grit notes add / RepositoryNotes::add is O(total notes): 35 ms vs git 1.8 ms with 20k notes (19x), 25 MB vs 6 MB#249639bc6a47

    Notes add/remove use path-local tree splices so existing fanout trees no longer require materializing every note. This follow-up fixes the remaining boundary case by routing the 256th insertion from a 255-note flat root through the established fanout writer and locks it down with a Git-compatible regression.

    system git notes add control (~500-note fixture): 1.8 ms/op → 1.9 ms/op (8% worse)
  • ›Truncated or corrupt .git/index panics (exit 101) or aborts on huge entry count instead of returning an error#25189d6e97a1

    This hardens index parsing so truncated entries and attacker-controlled entry counts cannot slice beyond input or request enormous Vec allocations. Corrupt files now return typed IndexError results, while valid Git-written EOIE/IEOT indexes continue to load.

  • ›Annotated tag with non-UTF-8 message breaks clone, gc, show, tag --json and fsck (git fsck is clean)#2510e46cd794

    Annotated tag message bodies no longer make a repository corrupt merely because they contain invalid UTF-8. Callers receive a lossy Unicode view, while serialization uses preserved raw bytes so tag object IDs and signatures remain unchanged.

  • ›grit clean --force silently keeps untracked non-UTF-8-named files ('Removed nothing', exit 0); --dry-run lists them#2514f599c413

    Clean now resolves repository-relative storage strings back to raw work-tree paths before deletion, so non-UTF-8 files and directories shown by dry-run are actually removed by force mode. The round-2 correction preserves the safety boundary around nested repositories by testing cumulative path prefixes rather than isolated components.

    Median clean latency, 10 debug runs: 6.3 ms → 7.2 ms (13% worse)
  • ›grit restore --staged is O(paths x index size): re-staging/unstaging 64k paths takes 8 s (git 0.06 s)#25167567cc72

    This change makes staged restore classify all selected paths first and update the index in one batch. It retains unchanged entries and their stat data, preserves split-index positions on replacements, and records resolve-undo while dropping conflict stages exactly as the former sequential path did.

    64k staged removals median (system Git control: 0.032379s): 6.9 seconds → 0.065 seconds (106× better)8k staged re-add median (system Git control: 0.101465s): 0.986 seconds → 0.146 seconds (6.8× better)
  • ›--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)#24093c94c1c5

    Adds a fast CLI integration test that creates and modifies 400 files, runs a two-commit JSON diff through `.files|length`, and verifies the filtered result is 400. It places coverage for the already-landed memory fix in the normal test suite instead of behind `full-tests`.

    cargo diff `.files|length` median wall time: 0.937 seconds (origin/main vs branch; 3 warm runs) → 0.990 seconds (origin/main vs branch; 3 warm runs) (5% worse)cargo diff `.files|length` median peak RSS: 452,868 KiB (origin/main vs branch; 3 warm runs) → 454,896 KiB (origin/main vs branch; 3 warm runs) (0% worse)cargo diff `.` median wall time: 1.4 seconds (origin/main vs branch; 3 warm runs) → 1.2 seconds (origin/main vs branch; 3 warm runs) (9% better)cargo diff `.` median peak RSS: 456,664 KiB (origin/main vs branch; 3 warm runs) → 453,096 KiB (origin/main vs branch; 3 warm runs) (1% better)
  • ›RepositoryBundles::create truncates the existing bundle before it can fail: a failed backup destroys the previous one#2515cc0fc888

    Bundle creation now stages output in a sibling `.bundle.part`, flushes it, and atomically replaces the destination only after successful pack generation. The handle gains cancellable/progress-aware and options-based creation APIs, missing objects become a typed bundle error, and the CLI calls the repository handle instead of lower-level bundle code.

  • ›stash with a non-UTF-8 file name saves a bogus 'grit+pathhex:' path, leaves the file modified; pop adds a stray file#25120db665d4

    Stash creation and restoration now retain raw non-UTF-8 path bytes instead of accidentally treating reversible storage strings as literal filenames. The round-2 addition makes untracked restore preflight atomic for non-UTF-8 file/directory collisions before any tracked worktree or index mutation.

  • ›A syntax error in .git/config or ~/.gitconfig makes every command and Repository::discover say 'not a git repository'#25196f7f32c1

    Repository discovery and open now return typed config errors instead of collapsing malformed local, global, or included configuration into “not a repository.” The CLI unwraps those config errors through its context chain and renders the existing fatal config diagnostic; invalid core.bare values are now validated when repository setup reads that key.

  • ›http.lowSpeedLimit/Time never fires on a silent or stalled-body HTTP server: fetch/clone hang forever (git: 5 s)#2511c1a133d5

    HTTP low-speed limits now interrupt socket reads that receive no bytes, covering silent servers and servers that stall after response headers. Clone also stops immediately when its optional bundle-URI probe fails for low speed, avoiding a second full timeout window before normal discovery.

    Silent clone timeout, origin/main to reviewed branch: 10.3 seconds → 5.0 seconds (2.0× better)Absolute timeout gap versus system Git (5.016s): 5.3 seconds → 0.015 seconds (351× better)
  • ›switch and merge refuse on a clean tree when a changed tracked file has a non-UTF-8 name (false 'uncommitted changes')#251339dbb1ca

    Repository paths carried as reversible storage strings are now decoded before filesystem and index access across switch, merge guards, restore, stash, checkout preview, and snapshot restoration. Human restore output is quoted appropriately, while machine-facing paths retain reversible encoding.

  • ›show/log -p/--stat/diff/blame fail 'revision <oid> not found' on the boundary commit of any shallow clone#25209b788474

    Shallow commits now behave as effective roots across revision navigation, commit diffs, patch/stat logs, pickaxe, and blame. Blame exposes root/boundary attribution consistently in human and stable JSON output, while grafts continue to override shallow metadata.

  • ›stash, reset and merge clear the assume-unchanged flag (git ls-files -v 'h' -> 'H'); git keeps it#252132e26b9e

    Preserves the index assume-unchanged bit through reset, merge, and stash when the resulting stage-0 entry still identifies the same path, mode, and object. If the blob or mode changes, the bit is dropped, matching system Git while leaving skip-worktree behavior unchanged.

  • ›macOS (core.ignorecase): switch/merge --ff/pick keep the old-case directory name after a case-only directory rename#2522110efa5e

    This change makes case-only directory renames materialize the target spelling on case-insensitive filesystems during switch, fast-forward merge, pick, and restore. The round-three fix removes the restore planner's remaining Cartesian destination scan by building and reusing a folded hash lookup.

  • ›grit serve closes with a large request body unread after an early 401/404: nginx answers 502 instead#252732fdd8c1

    Early smart-HTTP and upload-routing errors now consume unread request bytes before completing the response, preventing reverse proxies from replacing useful 401/404/413 statuses with 502. Decode errors use a bridge between Hyper's async body and the blocking protocol reader so the remainder can be drained with a 15-second idle timeout and 1 GiB ceiling.

other changes29
  • ›Main keeper: make CI on main green (public-api)d81dac56

    Refreshes the committed grit-lib public API snapshot to match three already-landed ref-store and receive-pack surface changes. It records one changed function signature, one new validation function, and one new refusal enum variant; runtime code is untouched.

  • ›Main keeper: make CI on main green (public-api)fff21b20

    There is no branch change to integrate. The author correctly detected that current main already contains the refreshed grit-lib public API snapshot and avoided creating a redundant commit.

  • ›Quality #79: Move logic into grit-libe192feb9

    Adds a reusable grit-lib view of how a GitButler workspace stands against its target, excluding the bookkeeping workspace commit and reporting each applied head's ahead count. Status now calls that API and only maps the typed result into output, with unchanged human/JSON/Markdown behavior.

    status-dirty-1000 Grit median: 2.9 ms → 2.8 ms (2% better)status-clean-1000 Grit median: 2.7 ms → 2.7 ms (0% worse)status-after-checkout-1000 Grit median: 5.7 ms → 5.6 ms (1% better)status-dirty-10000 Grit median: 6.5 ms → 6.4 ms (0% better)
  • ›Quality #80: Deduplicate3d65fce7

    Adds `ConfigSet::bool_or` for boolean settings that intentionally fall back on unset or malformed values, plus accessors for four heavily reused core booleans. About 70 readers across grit-lib and one CLI presentation helper now share that implementation, deleting three private helpers and duplicate case-variant lookups without changing behavior.

    add/modified median (system git 21.538 ms): 14.6 ms → 12.8 ms (13% better)commit/small median (system git 36.740 ms): 22.9 ms → 22.7 ms (1% better)diff/worktree median (system git 8.290 ms): 7.8 ms → 8.1 ms (3% worse)
  • ›Speed: commit/dry-run@synth-10k9f52af64

    Commit preview now derives its file list directly from the index-to-parent-tree cache-tree walk instead of probing the ODB for every index row, rehashing staged worktree files, and flattening the full parent tree. It also fixes tracked non-UTF-8 staging paths and uses a byte-exact fallback so previews and real commits report the same stored path names.

    commit/dry-run@synth-10k wall (system git after: 11.826 ms): 33.2 ms → 8.0 ms (4.1× better)commit/dry-run@synth-10k peak RSS: 15.8 MiB → 12.8 MiB (24% better)system git target control: 11.7 ms → 11.8 ms (1% worse)add/dry-run@synth-10k wall: 10.4 ms → 10.6 ms (2% worse)
  • ›Main keeper: make CI on main green (rustdoc)c4dba287

    This patch replaces one invalid public-to-private rustdoc link with prose and corrects two stale intra-doc references in the CLI and HTTP server. It restores the strict workspace rustdoc gate without changing executable code or public behavior.

  • ›Main keeper: make CI on main green (rustdoc)e1e84dfc

    There is no integration payload: the submitted branch is byte-for-byte identical to current origin/main at c4dba287f. The rustdoc repair had already landed through commit 524bddb27 and is green on main.

  • ›Quality #81: Simplify4470e0f6

    Shrinks grit-lib by removing APIs with no callers and confining test-only collection helpers to tests. It also removes an unread xdiff field/lifetime and an unused reftable wrapper, leaving production behavior unchanged.

    diff-10000 median (exact base 71fddd336 → branch): 7.7 ms → 6.6 ms (18% better)blame-deep-2000 median: 41.7 ms → 40.9 ms (2% better)blame-large-tail-full median: 34.3 ms → 33.8 ms (1% better)blame-large-mid-churn-full median: 11.8 ms → 11.6 ms (1% better)
  • ›Main keeper: make CI on main green (public-api)47c3e687

    Refreshes the committed grit-lib API inventory so it describes the API already shipped on main. The generated file now includes tree diff byte sizes and the expanded parallel diff-stat row, while normalizing field order.

  • ›Main keeper: make CI on main green (public-api)060ed70e

    Refreshes the committed grit-lib API snapshot so it records the two blob-size fields on structured tree diffs and the two added u64 values in parallel diff-stat rows. An equivalent commit has already landed on main, so the reviewed Mesh branch now has no unique diff.

  • ›Speed: switch/branch@synth-10k2d6fca44

    This change teaches the hot-text profiler about ordinary small branch switches that use serial checkout writes, then deterministically searches for a denser function-group order. It regenerates the library and CLI linker fragments and records the measured dist baselines; runtime semantics are unchanged.

    switch/branch@synth-10k wall ms (system Git 22.542 ms): 17.5 ms → 16.4 ms (7% better)switch/branch@synth-10k peak RSS (system Git 10.19 MiB): 13.1 MiB → 12.4 MiB (6% better)switch/branch@synth-100k wall: 85.8 ms → 86.2 ms (0% worse)switch/delete-heavy@wide-48k wall: 120.8 ms → 117.3 ms (3% better)
  • ›Main keeper: make CI on main green (public-api)46dadafb

    There is no review diff to integrate. The reviewed branch points at da7c6d56f, where the public-api snapshot repair had already been merged, and current origin/main is ahead of that commit.

  • ›Quality #82: Library API quality2e117b0b

    Replaces the unrestricted string-backed config error escape hatch with specific, documented error variants carrying keys, values, reasons, paths, and environment indices. It also separates submodule layout/state failures into `SubmoduleError`, preserves the exact malformed section input during renames, and keeps CLI fatal-prefix behavior stable.

    status dirty 10k median: 6.5 ms → 6.6 ms (1% worse)status clean 10k median: 6.2 ms → 5.3 ms (17% better)status after checkout 10k median: 38.7 ms → 36.9 ms (5% better)commit 10k median: 110.0 ms → 100.6 ms (9% better)
  • ›Main keeper: make CI on main green (public-api)6057055c

    Refreshes the committed grit-lib API snapshot so its ordering exactly matches the pinned cargo-public-api generator. It only repositions six existing function signatures; no exported item, implementation, test, or runtime behavior changes.

  • ›Speed: branch/merged@many-refs85f9bca2

    Regenerates the two linker layout fragments so frequently executed library and CLI startup code is packed into fewer resident code windows. It also records the profiling diagnosis and commits before/after daily benchmark reports; there is no Rust source, public API, on-disk format, wire protocol, or command behavior change.

    branch/merged release time: origin/main to reviewed layout: 2.3 ms median (20 runs each) → 2.0 ms median (20 runs each) (13% better)branch/merged release peak RSS: origin/main to reviewed layout: 11.6 MiB → 8.4 MiB (38% better)branch/merged time: system git to reviewed grit: 1.3 ms median (20 runs each) → 2.0 ms median (20 runs each) (33% worse)branch/list release time: 2.0 ms median (20 runs each) → 2.0 ms median (20 runs each) (2% worse)
  • ›Quality #83: Move logic into grit-lib22553eae

    Adds a grit-lib batch API for per-commit insertion/deletion totals and makes `grit log` consume the library's existing batch diff API directly. The CLI no longer reads commits, tracks merge slots, or folds totals itself; output behavior remains under the same presentation mapping.

    grit-cli/src/commands/log.rs lines: 1,703 lines → 1,646 lines (3% better)
  • ›Main keeper: make CI on main green (test, public-api, clippy)a46e45ab

    Adds the missing `rebase_run_request` import used by the history rebase CLI. This restores compilation and therefore unblocks the test and clippy CI jobs without changing runtime behavior.

  • ›Quality #84: Dead code and dependencies6083ca2b

    This removes nine backend-specific reftable functions that only reopened ReftableRefStore and forwarded to its existing API. It also deletes an unused RpcChannel trait, its sole implementation, and the private post_round helper reachable only through that implementation; tests now exercise the canonical RefStore and generic reflog/ref APIs directly.

    public reftable free functions: 17 functions → 8 functions (2.1× better)
  • ›Main keeper: make CI on main green (clippy, test)6083ca2b

    This review branch contains no changes relative to current origin/main. The missing `rebase_run_request` import was already delivered by another factory branch, so this branch now only confirms the repaired main commit.

  • ›Main keeper: make CI on main green (public-api)a32db43d

    Updates the committed grit-lib API snapshot so ConnectOptions::with_protocol_version appears under its explicit inherent impl block, exactly as the pinned generator emits it. No Rust source, behavior, formats, protocols, tests, or CLI surfaces are changed.

  • ›Main keeper: make CI on main green (public-api)f2075894

    This verifies that the already-landed public API snapshot refresh repaired the red main CI job. The review branch itself has no commits or diff against origin/main, so there is no additional patch to integrate.

  • ›Main keeper: make CI on main green (rustdoc)4b2cfb44

    Updates one `Bundle::unbundle_with_progress` rustdoc sentence so its prerequisite-check reference resolves to the public bundle-fetch entry point. This removes the strict rustdoc failure without changing runtime behavior or public API shape.

  • ›Speed: pick/one@synth-10k8cfec59e

    This change teaches the generated hot-text linker layout about pick, revert, and merge, grouping their commonly executed code into resident-code-friendly sections. It updates both generated linker fragments and documents the measured Linux x86_64 dist results without changing Git behavior, public APIs, or repository formats.

    pick/one@synth-10k wall time: 54.9 ms → 55.0 ms (0% worse)pick/one@synth-10k peak RSS: 19.0 MB → 16.3 MB (17% better)pick/one@synth-100k wall time: 136.9 ms → 134.2 ms (2% better)pick/one@synth-100k peak RSS: 34.7 MB → 31.8 MB (9% better)
  • ›Main keeper: make CI on main green (public-api, rustdoc)b0818e71

    Refreshes grit-lib’s committed public API snapshot to reflect the ref-store backend signatures and helpers that already exist in the implementation. This is a generated-snapshot correction only; it changes no Rust behavior or repository format.

  • ›Speed: reset/undo-commit@git-git9f409237

    This change batches the packed commit-header reads used to seed orphan/reachability paints, reuses pack and inflate state, and avoids faulting most of large pack-index OID tables through positioned row-window reads. It also moves the first parallel chunk onto the caller thread and records primed linkage in one lock acquisition; no public API or on-disk format changes.

    reset/undo-commit grit median: 7.0 ms → 6.5 ms (7% better)reset/undo-commit grit peak RSS: 24.6 MiB → 19.1 MiB (29% better)reset/undo-commit system-git control median: 2.2 ms → 2.1 ms (3% better)reset/hard-delete-heavy grit median: 219.6 ms → 220.9 ms (1% worse)
  • ›Speed: branch/list@many-refs9f409237

    This branch intentionally makes no source changes. It records that branch listing itself is already cheap and that the remaining tiny-command gap is dominated by process startup; independent review found the reviewed head and current-main dist executables byte-for-byte identical.

    branch/list@many-refs time ratio vs git: 1.5 x git (20 runs/tool) → 1.5 x git (20 runs/tool) (4% better)branch/list-json@many-refs time ratio vs git: 1.1 x git (20 runs/tool) → 1.4 x git (20 runs/tool) (18% worse)branch/list@many-refs-loose time ratio vs git: 1.4 x git (20 runs/tool) → 1.4 x git (20 runs/tool) (7% better)branch/merged@many-refs time ratio vs git: 1.4 x git (20 runs/tool) → 1.4 x git (20 runs/tool) (3% worse)
  • ›Main keeper: make CI on main green (public-api, test)49df9944

    This change makes HTTP clone/fetch stop when a server accepts a connection but never sends response headers and low-speed settings are configured. It connects the existing low-speed window to ureq's response-header receive timeout while retaining the existing body-throughput reader after headers arrive.

    HTTP stall regression duration: 5 seconds → 2.0 seconds (2.5× better)
  • ›Main keeper: make CI on main green (test)cc0fc888

    When HTTP low-speed limits are configured, the ureq client now applies the low-speed window to receiving response headers. This closes the silent-server gap before the existing body-rate reader begins; the reviewed commit is already present in origin/main, leaving no mesh diff to integrate.

    Silent-stall regression duration: 5 seconds → 2.0 seconds (2.5× better)
  • ›Main keeper: make CI on main green (public-api)a400174f

    Refreshes the committed grit-lib API inventory to exactly match the API generated by the repository's pinned cargo-public-api 0.52.0 toolchain. It records one blame field, the EffectiveCommitParents surface, and the generated field ordering for TagData; runtime code is untouched.

$ git log --oneline bd1c65f8..v0.5.29compare on GitHub ↗
a400174fMerge branch 'factory/fix-public-api-6877'
32fdd8c1Merge branch 'factory/serve-discard-body-79db'
877721e3Refresh grit-lib public-api snapshot
e8f02fa1Bound raw body drain after decode errors via body bridge
525f5961Fixes #2527: drain large POST bodies before early HTTP responses
110efa5eRun cargo fmt on ureq_client timeout match
d81935b8Use memihash lookup for restore icase collision planning
67a51ab4Address review: batch restore icase collisions and fix test build
e345a66dFix case-only directory rename leaving wrong on-disk spelling
32e26b9eFixes #2521: preserve assume-unchanged across reset, merge, and stash
7c938bcbAdd PERF-SETTLED.md, the record of settled performance questions
9b788474Review: restore references_update_head tests and update blame JSON schema
dcf31e37Fix shallow boundary handling for show, log, diff, and blame
39dbb1caMerge branch 'factory/fix-non-utf8-switch-6820'
43f18d53Fixes #2513: decode stored paths for switch, merge, and restore guards
c1a133d5Merge branch 'factory/http-low-speed-per-read-14ce'
6f7f32c1Merge branch 'factory/fix-config-parse-errors-2519'
0db665d4Merge branch 'factory/stash-non-utf8-2512'
1ff0e88aAbort clone on low-speed bundle probe instead of re-fetching
5ab600fcFix HTTP low-speed timeout on silent and stalled-body servers
cc0fc888Merge branch 'factory/bundle-atomic-create-6828'
06c21579Stash untracked preflight: compare path collisions on raw bytes
96112727Fixes #2512: stash non-UTF-8 paths use raw index bytes
3c94c1c5Merge branch 'factory/fix-json-filter-2409'
e0253ae4Fix config parse errors during repository discovery
ca2670ebUse atomic_replace for bundle install on Windows
9280b9a2Fix atomic bundle create and typed ObjectNotFound
e46cd794Merge branch 'factory/fix-tag-nonutf8-2510'
f599c413Merge branch 'factory/clean-non-utf8-2514'
9d6e97a1Merge branch 'factory/fix-index-corrupt-panic'
f45f307aFix grit clean --force skipping non-UTF-8 untracked paths
7567cc72Fix restore --staged index updates scaling with batch removal
008ea5ecFixes #2409: integration test for large diff --json --filter
270c151dFixes #2510: accept non-UTF-8 annotated tag messages
67720673Fixes #2518: return IndexError for truncated or corrupt index files
49df9944http: abort silent stalls at lowSpeedTime before response headers
9f409237Merge branch 'factory/reset-orphan-tip-reads'
b0818e71Merge branch 'factory/fix-rustdoc-public-api'
39bc6a47notes: splice on add; fan out when flat root reaches 255 entries
e5d28885Refresh public-api.txt for ref store backend exports
d1cb18b5Merge branch 'factory/fix-2508-merge-rename-quadratic'
8fb317bdMerge branch 'factory/http-low-speed-progress-6780'
1db3a0a3Fixes #2496: splice notes tree on add instead of full rewrite
58bafd43Add HTTP low-speed regressions for PendingHttp discovery path
8cfec59eDocument the pick, revert and merge hot-text layout with dist benchmarks
02de9d94Lay out the pick, revert and merge code path in the hot text sections
4b2cfb44Merge branch 'factory/fix-rustdoc-bundle-link-6785'
9f949c3fFixes #2508: linearize same-blob rename augment in merge_trees
54925ac9Document the batched tip header reads for undo-commit
81f310c6Fix broken rustdoc link in Bundle::unbundle_with_progress
469bdcecFix unsorted UNTR tracked-under probe to avoid per-dir index scan
6aab1243Fixes #2507: use index lookup for UNTR tracked-under probe
a3a8a7afFixes #2509: honour cancel during switch planning and wholesale checkout
1933df53Merge branch 'factory/loose-ref-casefold-2502'
72741fd1Merge branch 'factory/receive-pack-custom-primary-2506'
8f5e243dMerge branch 'factory/fix-clone-http-stall-2469'
e9d2da16Merge branch 'factory/ref-tx-head-old-oid-012d'
b34c5e65Merge branch 'factory/serve-injected-head-898f'
13ab3191Read many hidden tip commits' headers in one batch
436ac746Merge branch 'factory/bundle-unbundle-prereqs'
04a702a4Fixes #2506: accept receive-pack pushes into a custom object primary
30bb4d5dFix #2469: interruptible clone HTTP and low-speed stalls
f2075894Merge branch 'factory/single-writer-gate-2500'
e9a72406Fixes #2505: read HEAD and ref tips from injected RefStore in serve
fc4e1d65Reject ignorecase D/F collisions on nested ref paths
560cea2cFix loose-ref casefold check scanning every ref on create
5dd76f7eMerge branch 'factory/fix-json-filter-memory-2409'
2feaa09dTest: connect prerequisite ref via grit ref store
18adbc30Test: connect prerequisite ref before incremental unbundle
204970a3Fix bundle unbundle missing-prerequisite diagnostics
10c36b98Fix #2504: resolve symbolic HEAD for reference-transaction log-only lines
a32db43dMerge branch 'factory/ci-green-public-api-6753'
2acd602cReview: lazy ref_lock_wait, public-api snapshot, docs and tests
0094f441Fixes #2500: FIFO SingleWriterGate and RefStore::prepare_for
17b21e05Fixes #2409: keep direct serde→jaq --filter path
f9f397f0Merge branch 'factory/fix-2499-case-rename-untracked-8f83'
bde254d4Refresh grit-lib public-api snapshot for ConnectOptions::with_protocol_version
39278ce3Merge branch 'factory/fix-2497-head-update-reflog'
9fd677c5Merge branch 'factory/pack-tls-drop-e9e3'
c9019d6cMerge branch 'factory/merge-reflog-atomic-6738'
42784539Update grit-lib public-api snapshot after rebase onto main
781679acFix worktree guard false untracked hits on case-only renames
a3f0fa38Fixes #2501: buffer grit diff human output on pipes
eccbb952Fixes #2495: release pack TLS mappings when PackStore drops
c21daa56Fixes #2503: atomically update branch and reflog on merge
59a0191dFixes #2497: independent HEAD vs branch reflog when peeling HEAD
243b63a3Fixes #2497: update public-api snapshot for new ref handle methods
eaeeb192Fixes #2497: follow HEAD symrefs and write reflogs from References::update
285dbe51Fix #2444: add ConnectOptions::with_protocol_version for full-tests
6083ca2bRemove the unused RpcChannel transport trait
1c2d730eRemove reftable free functions that only wrap RefStore
a46e45abMerge branch 'factory/ci-green-6700'
22553eaeMerge branch 'factory/log-line-totals'
67b7ad8dFix missing rebase_run_request import in history rebase CLI
85f9bca2docs: record where branch --merged spends its time against git
cfbbc87bRegenerate the hot-text layout for the current tree
c83675c0Merge branch 'factory/public-api-snapshot-6698'
6057055cMerge branch 'factory/api-rustdoc-clarity-2494'
06f695e6Refresh grit-lib public-api snapshot for fork-point API ordering
1f7ff236log: take per-commit diffs and line counts from grit-lib
80398d77diff: add Diff::commit_line_totals for per-commit first-parent line counts
393571fbPreserve pre-existing bundle file on create cancel
9e7e8389Address review: rebase exec signals, clean exit 130, bundle create cancel
cbcd69b6Fix interrupt handling for worktree add, rebase, and history
2e117b0bMerge branch 'factory/typed-config-errors'
f8019286Merge branch 'factory/fork-point-pull-6662'
0956b6ebMerge branch 'factory/push-memory-2493'
46dadafbRefresh grit-lib public-api snapshot after stash and transport API changes
645d382cFixes #2494: clarify grit-lib public rustdoc and API cross-links
3a92e708Merge branch 'factory/push-interrupt-cleanup-2492'
e276c905Document typed config and submodule errors in the config guide
2917c5b1Replace stringly ConfigError::Other with typed config and submodule errors
2d6fca44Merge branch 'factory/switch-branch-lean'
edc10cb4Apply rustfmt to merge_base_git_compat test
918919e5Test: two merge bases oracle for try_reflog_fork_point
79d13ea7Address review: null reflog OIDs, test attrs, full merge bases
86812391Fix pull/rebase fork-point after force-push
c96b66c9Fixes #2493: bound local push pack memory
9a42a354Merge branch 'factory/clone-ignorecase-refs-2490'
aa6b49e6Reorder integration test modules for rustfmt
2928641aInstall InterruptCancel only for receive-pack
e8ded92dFixes #2492: cancel push on signals and prune stale quarantine
c9d7c585Merge branch 'factory/fix-empty-commit-graph-2487'
ca3b1681Merge branch 'factory/merge-check-favor-2491'
b16ea7ccScope ignorecase case-pair packed refs to clone only
f02d6b49Fix clone/fetch on ignorecase when remote refs differ only by case
2ad9f917Merge branch 'factory/stash-drop-oid-6660'
da7c6d56Merge branch 'factory/public-api-6669'
4e8e0d63Document the small-switch hot-text layout with dist benchmarks
c2873cfbRegenerate the hot-text layout with the small-switch profile
5d7f29b3Profile small switches and search the hot-text order
dcc7d244Fix empty-repo commit-graph gc and fsck mismatch
060ed70eMerge branch 'factory/ignorecase-wholesale-delete-6655'
0a3db526Fix stash drop oid output and apply by commit id
b236f499Fix merge --check to honor -X ours|theirs
086a5035Refresh grit-lib public-api snapshot for diff stat and transport API
fe0522a5Fixes #2485: O(1) ignorecase lookup for wholesale directory deletes
1b465a82Merge branch 'factory/http-fetch-cancel-lifetime-6617'
39377cceMerge branch 'factory/fix-2484-eof-merge'
e21d02e4Merge branch 'factory/clone-missing-local-d665'
47c3e687Merge branch 'factory/public-api-6646'
ad9fbc85HTTP cancel: scoped owner borrow and drop 1s header cap
c903f33aHTTP cancel: replace stack fat pointer with scoped owner guard
a2416566Integrate HTTP fetch cancel and stall interrupt fix (6536)
198b5794Merge branch 'factory/describe-kahn-6625'
46b3f3f9Merge branch 'factory/rename-delete-merge-2483'
949338caFix clone when local source path is missing (Fixes #2480)
007f4674Merge branch 'factory/fix-symlink-gitlink-warn-2482'
3a760f7dFix HTTP fetch/clone hang on stalled servers (Fixes #2469)
b176f90fCI: refresh grit-lib public-api snapshot for diff stat sizes
32a52c2dAdd describe diamond regressions for merge tag peels
66a954e0Fixes #2484: stop false EOF merge conflicts when anchor is a real boundary
4fa1fd6dFixes #2483: conflict on theirs rename vs ours delete
e248d6f5Fixes #2482: do not treat symlink paths as nested repositories
4470e0f6Move test-only helpers into their test modules
39c89f2cRemove grit-lib public functions that nothing calls
e64aa63cMerge branch 'factory/describe-perf-6590'
24ec802frevwalk: fast exact RevWalk::describe without commit-graph
e1e84dfcFix no-index binary stat sizes and regenerate JSON schemas
8d20de07Update grit-lib public API snapshot for SideBytes and diff stat sizes
9a700374Fix diff --stat binary byte sizes (Fixes #2339)
c4dba287Merge branch 'factory/rustdoc-abbrev-link'
9f52af64Merge branch 'factory/commit-dry-run-speed'
bb9edfcbdocs: record the commit preview speedup and its daily rows
710f8dafcommit: list non-UTF-8 paths in a preview in stored form
aba4feecadd: decode stored paths when staging tracked worktree changes
e0aa6b05commit: preview without probing the object store per index entry
524bddb2Fix rustdoc intra-doc links for workspace -D warnings
3d65fce7Read boolean config through ConfigSet::bool_or and the core accessors
1608f57cconfig: add ConfigSet::bool_or and core boolean accessors
e192feb9status: use GitButlerWorkspace::standing for the workspace view
fca703bbgitbutler: add GitButlerWorkspace::standing
856d98f0revwalk: one backward walk for describe instead of per-tag counts
fff21b20Merge branch 'factory/public-api-6587'
33ed55e7Refresh grit-lib public API snapshot for core.abbrev helper
afef1051Fixes #2412: honour cancel while commit hooks run
765230dcMerge branch 'factory/rebase-cancel-rollback-6565'
a963e6e7Fix slow rebase apply cancel rollback (Fixes #2475)
3ed1a2b7Merge branch 'factory/fix-display-oid-abbrev'
71fddd33Merge branch 'factory/graceful-serve-shutdown'
b4630230Merge branch 'factory/alternate-read-perf-6541'
d81dac56Merge branch 'factory/public-api-6573'
eaabe728Address review: canonical commit API and batched Markdown abbrevs
bdce026dFix display oid abbrev to follow core.abbrev and extend on collision
3bed7d10Fixes #2478: graceful shutdown for grit serve
96b39af2Keep reprepare on empty-primary miss after pack warmup shortcut
e0439fabFix alternate ODB reads thrashing pack TLS and empty-primary probes
f2c39244Refresh grit-lib public API snapshot for ref-store changes
1f83c3baFix reftable stack cache deadlock on missing table
d4395a83Merge branch 'factory/non-utf8-untracked-6537'
13c36255Merge branch 'factory/replace-refs-revwalk-2473'
a6e5a8deMerge branch 'factory/update-instead-receive-pack'
21b00c58Merge branch 'factory/serde-models-2474'
0cc71675Expand pathhex directory markers when staging collapsed untracked dirs
24d94438Fix collapsed untracked directory markers for non-UTF-8 dir names
c9913feaFixes #2464: preserve non-UTF-8 bytes in untracked scan and commit dry-run paths
5a282c99serde-models: Serialize blame, describe, tag, and worktree results
e645ac63Merge branch 'factory/rerere-porcelain-6533'
409744fbApply rustfmt after rebase onto main
cbb975aeLoad replace refs from ref store including packed-refs and reftable
c240cf91Honor refs/replace in revwalk, navigation, log, diff, and blame
bfdcca25cargo fmt: receive_pack_update_instead_untracked test
382cea30Refuse updateInstead deploy when untracked paths would be overwritten
8eba488aFix receive.denyCurrentBranch=updateInstead push-to-checkout deploy
82664480Merge branch 'factory/fix-reset-hard-cancel-rollback-6523'
d9cca2c3Merge branch 'factory/skip-worktree-2466'
86ba38f4Fixes #2468: squash rerere postimage and completion rerere CLI output
a505049fFixes #2468: wire rerere into merge, pick, rebase, and commit
9565a23cMerge branch 'factory/case-insensitive-ref-create'
f97e90d1Merge branch 'factory/shallow-push-2463'
4b09e375Merge branch 'factory/json-filter-memory'
322e5b40Fix hard reset cancel rollback for coalesced directory deletes
02b58e8fFix cross-tree skip-worktree reset edge cases
f296a888Fix skip-worktree handling in restore, stash, and reset
5b36a8faFix receive-pack ignorecase ref collisions (Fixes #2471)
c974aaedFix shallow clone push when remote lacks ancestry
e4a44da4fix: set upstream when grit switch -c starts from a remote branch
26b35f13Fixes #2409: cut --json --filter memory via direct jaq Val build
28d2f777Fixes #2461: describe uses ranked one-pass walk with commit-graph