sley

An embeddable Rust Git-equivalent library with a thin, byte-compatible CLI wrapper.

The pinned compatibility oracle is upstream Git 2.55.0. Repository semantics live in typed engine crates and the sley::Repository facade; the CLI owns argv/environment setup, terminal integration, dispatch, and Git-identical byte rendering. The curated conformance harness runs the oracle and Sley separately and compares exact TAP cells.

This repository is not a complete Git replacement yet. Tracked gaps are explicit in PARITY.md.

cargo build -p sley-cli --bin sley --release
cargo test --workspace
./target/release/sley version

Run a command through the CLI:

cargo run -p sley-cli --release -- status --short
cargo run -p sley-cli --release -- log --oneline -20

The original 2026-07-10 Git v2.55.0 baseline ran 891 enrolled scripts on an M1 Pro, but its Sley side used a generated /bin/sh shim for every command while Git ran directly. That candidate-only process launch makes the historical aggregate and tail timings unsuitable for release performance claims. Its equal-work classification remains useful: only 425 scripts did demonstrably equal work, and failing/short-circuiting rows remain excluded.

The harness now launches Sley directly under the installed git name. In the first corrected three-pair shard, exact t7004-tag.sh (231/231 cells) measured 5.744 s for Git and 5.297 s for Sley: a 1.084× paired-median speedup, with the selected-run aggregate, median, p95/p99, and wall-time comparisons all passing. After reference-fsync policy and backend-selection caching, a five-script refs shard retained exact work in three alternating trials and measured a 0.796 Sley/Git aggregate ratio, 1.265× median speedup, and 0.79 p95/p99 ratios; every script beat Git, including t1460-refs-migrate.sh at 1.144×. After repack/bitmap optimization, exact t5333-pseudo-merge-bitmaps.sh improved from 10.783 s to 8.081 s on Sley, but Git's 6.633 s median still leaves Sley at 0.821× and that shard's performance gates remain red. Exact t7063-status-untracked-cache.sh is now 1.987× faster on Sley; exact t3311-notes-merge-fanout.sh is 1.045× faster in the latest sample but remains close enough to the 1.05× threshold to require the dedicated Linux certification run.

The first clean-oracle, direct-launch full SHA-1 matrix completed all 891 scripts without an abort or timeout. Git v2.55.0 established 881 passing scripts plus 10 legitimate skip-all scripts; Sley reported 496 passing, 385 failing, and the same 10 skip-all scripts. Exact per-cell comparison is the release measure: 513/891 scripts match every oracle-applicable cell, while 378 remain incomparable. A one-shot equal-work diagnostic over the 496 scripts that also passed end-to-end has Sley ahead in aggregate, median, p95, and p99, but it is not an alternating paired run and is not a certification result.

The latest integrated eight-wave rerun raised the exact result to 556/891 and the raw result to 538 passing, 343 failing, and 10 skip-all scripts (30,103/32,357 assertions, 93%). A subsequently verified MIDX regression repair (t5319, 98/98) and credential-engine closure (t0300, 56/56) establish a current verified exact floor of 557/891. t0301-credential-cache.sh was exact before its daemon-latency change, but the managed sandbox now rejects its Unix socket bind, so it is not carried into that floor until an unsandboxed rerun. On the 538 exact end-to-end rows in the integrated run, the non-alternating equal-work diagnostic measured a 0.749 Sley/Git aggregate ratio, 1.227× median paired speedup, and 0.723/0.696 p95/p99 ratios. These are development diagnostics only: at least 334 correctness gaps remain and the five-pair Linux certification has not run.

After caching repeated parent-directory metadata probes in status, a fresh 21-case, byte-identical common-command sample passes both gates: 0.811 Sley/Git geometric mean and no row over 1.05. The former blocker, large-repository status --short, now measures 22.87 ms for Sley versus 23.17 ms for Git. Eight-wave wall time is still a separate, not-yet-certified measure.

This is targeted development evidence, not the required five-pair dedicated Linux certification. crates/sley-testkit/UPSTREAM_TIMINGS.md records both the superseded baseline and the corrected measurement protocol.

Only byte-identical, equal-work rows are eligible for Git/Sley performance claims. Re-run the local benchmark surfaces with:

python3 scripts/bench-human-common.py
./scripts/bench-vs-git.sh
cargo bench -p sley-bench

See the testkit timing document for the three-trial nightly and five-trial Linux certification commands.

Behavior lives in typed engine crates; the sley::Repository facade is the embedder entry point; sley-cli is a thin argv/setup/dispatch shell. Prefer table-driven CLI options (sley-options) and shared engines over per-command handlers (ADR 0001). Large payloads stream as pull Read/Write with cooperative cancel and sideband demux as a Read adapter (ADR 0002). Clustered upstream/parity work by engine is tracked in docs/ROADMAP_ENGINES.md.

Apache-2.0