Levi Neuwirth
e050b6dddd
B4's harness measured intermittent AlreadyLocked on reopening a store it had just dropped -- up to 29 retries over 147ms, 7 failures in 40 runs, reproducing on a single-shard engine. StoreEngine::drop was not the cause: it closes every channel and joins every writer, and looked correct in every single-process test. An flock is held by the open file description, not by a descriptor. A concurrently forked child transiently inherits that description, and FD_CLOEXEC closes it at exec, not at fork -- so during that window the parent closing its own descriptor releases nothing. Scope 3.1 makes AlreadyLocked a refusal and never a wait, so every consumer that closes and reopens a root -- a recovery drill, an in-place restart, the Phase 2 migrator -- could be refused with no defined retry. Releasing must therefore be an explicit act. lock_root returns an RAII RootLock that issues LOCK_UN in Drop before the file closes, and both owners -- RecoverySession and ShardDrive -- hold that guard. The unlock lives in sys.rs beside try_lock_exclusive so the locking syscalls stay in the one funnel. StoreEngine needs no change: it holds a RecoverySession, never a File, so it inherits the fix. Drop is the only release path. A release() and a file() accessor were written and deleted before landing: neither had a caller, and an uncalled second way to release a lock is exactly the decoy the charter names. A failed LOCK_UN in Drop cannot be returned and must not be swallowed, so it increments a counter the regression asserts unchanged. The rejected alternative was fixing this in StoreEngine::drop alone. That leaves ShardDrive and the drive's one-shot session exposed and makes correctness depend on a descriptor lifetime that fork can extend. The regression is synchronized rather than timed: the child forks while the lock is held, signals ready on one pipe, and blocks on a second until after the parent has released and attempted its reopen, so the inherited descriptor is provably open across the whole window and the reopen is asserted on its first attempt. With the explicit unlock removed it fails 10/10; as landed it passes 40/40. Measured under load -- 200 close-reopen cycles against 72,255 concurrent forks -- 0 refusals, worst case 1 attempt and 10.8ms; the same load kills the pre-fix behaviour within 0.05s, so the load reproduces the defect rather than merely being weak. Contract review 2026-07-28-B records the amendment. Two carry-forwards are recorded in scope 6.6: the verify-store-recovery SIGKILL cycles still drive the journal seam rather than submit, so kill -9 never lands inside a real publication and the acknowledged-crash-recovery criterion is only partly earned; and B4's bounded reopen retry must become a one-attempt assertion now that the defect it compensates for is gone. scripts/check-phase1.sh GATE_EXIT=0. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .github/workflows | ||
| bench | ||
| crates | ||
| deploy | ||
| doc | ||
| scripts | ||
| .gitignore | ||
| Cargo.lock | ||
| Cargo.toml | ||
| LICENSE | ||
| README.md | ||
README.md
LeVCS
A distributed version control system with first-class federation, signed authority chains, and a cascading merge engine. Content-addressed by BLAKE3, signed with Ed25519, and built to fix what git can't.
Status: v0.1.0 — protocol substrate complete, workflow surface deferred. The object model, federation API, merge cascade, and instance server all work end-to-end. There is no PR review surface, issue tracker, or web UI yet — those are the next layer up. See
doc/technical-report.mdfor a full framing of where this project is and why.
What's different from git
- Identity is in the protocol. A repo's membership is a versioned, signed authority object with explicit roles (Reader/Contributor/Maintainer/Owner). Force-push enforcement and push authorization are protocol-level, not server policy.
- Federation is first-class. Every repo has a global
repo_id(BLAKE3 of its genesis authority); instances mirror each other in three storage modes (full / release-only / metadata-only). - Merge is a cascade, not a line-level diff. Per-file dispatch to a handler ranked by aggressiveness: textual fallback, format-aware (JSON / YAML / TOML / XML / Markdown / prose), tree-sitter for source code (Rust, Python, JS/TS, Go, C/C++, Java, Ruby, Bash), and wasm-sandboxed plugins for the long tail.
- BLAKE3, not SHA-1. Tree-hashed, ~5 GiB/s on a laptop, 32-byte IDs everywhere.
- Releases are signed objects, not mutable name pointers.
For a deeper comparison and context, see the technical report.
Building
LeVCS is a Rust workspace. You'll need a recent stable toolchain (workspace MSRV is 1.75) and a C compiler for the tree-sitter grammars.
cargo build --release
Two binaries land in target/release/:
levcs— the user-facing CLI.levcs-instance— the federation HTTP server.
Install them somewhere on PATH:
sudo install -m 0755 \
target/release/levcs target/release/levcs-instance \
/usr/local/bin/
Quick start (single user, local only)
# Generate an identity key (stored in $XDG_CONFIG_HOME/levcs/keys.toml).
levcs key generate --label me
# Create a repository wherever you have files to track.
mkdir /tmp/demo && cd /tmp/demo
echo "hello" > a.txt
levcs init --key me
levcs track --all
levcs commit -m "first commit"
levcs log
That's a fully working LeVCS repo. Branch and merge:
levcs branch feature/x
echo "more" >> a.txt
levcs commit -m "wip"
levcs branch main
levcs merge feature/x
If a merge produces conflicts, drop into the resolution TUI:
levcs merge --resolve
Cut a release:
levcs release v0.1.0 --notes "first release"
Hosting an instance
To dogfood the federation surface, run levcs-instance on a VPS behind
nginx or Caddy. The full walkthrough is in
deploy/README.md: build, systemd unit, reverse
proxy templates, firewall, and the laptop-side bootstrap.
The compressed version:
sudo cp deploy/levcs-instance.service /etc/systemd/system/
sudo cp deploy/instance.toml.example /etc/levcs/instance.toml
sudo $EDITOR /etc/levcs/instance.toml
sudo systemctl enable --now levcs-instance
# ... then drop deploy/Caddyfile.example into /etc/caddy/Caddyfile
From your laptop, point the local repo at the instance and push:
levcs instance --set https://levcs.example.com/levcs/v1
levcs push refs/branches/main
The first push to a fresh instance auto-inits the repo with your genesis authority. Subsequent pushes are role-checked against the authority chain.
Repository layout
crates/
levcs-core/ Object model (Blob/Tree/Commit/Release/Authority),
hash, store, refs, repository abstractions.
levcs-identity/ Authority objects, Ed25519 keys, signing/verify.
levcs-merge/ Cascade engine, format and tree-sitter handlers,
plugin runtime, merge records.
levcs-protocol/ Pack codec, wire types, request signing, P2P.
levcs-client/ Thin HTTP client over the federation API.
levcs-instance/ Axum HTTP server (the federation peer).
levcs-cli/ The `levcs` user-facing CLI.
levcs-tui/ Conflict-resolution terminal UI.
deploy/ Production deployment artifacts (systemd, Caddy, nginx).
scripts/ Reproducible benchmark and ops scripts.
doc/ Technical report and architecture docs.
.github/workflows/ CI configuration.
Testing
cargo test --workspace
Runs the full suite — unit tests, integration tests, federation end-to-end tests including the three-instance "dogfood" scenario, the merge conformance corpus, and property-based fuzz tests. ~194 tests at v0.1.0; full run is well under a minute on a modern laptop.
Useful subsets:
# A single crate's tests
cargo test -p levcs-merge
# A specific integration test
cargo test -p levcs-instance --test dogfood
# Property tests only
cargo test -p levcs-merge --test proptest_textual
Benchmarks
Microbenchmarks live in each crate's benches/ directory and use
criterion. A reproducible runner with metadata capture and optional
flamegraph generation is at scripts/bench.sh:
scripts/bench.sh --quick # smoke test (~ a minute total)
scripts/bench.sh # full criterion run (~ a few minutes)
scripts/bench.sh --flamegraph # generate per-bench SVG flamegraphs
scripts/bench.sh --bench pack_codec # one bench only
Output goes to bench-results/<host>-<UTC-timestamp>/ with a parsed
summary.txt, criterion's HTML reports, and a metadata.txt capturing
rustc version, kernel, CPU, and git rev for run-to-run comparison.
Headline numbers on a Ryzen 7 laptop:
- Pack decode at 10 × 1 MiB entries: ~2.3 ms (4.3 GiB/s).
- Blob serialize + BLAKE3 at 1 MiB: ~190 µs (5.1 GiB/s).
- Textual three-way merge of a 100 KiB document: ~4.6 ms.
Pack encoding is the throughput floor at ~380 MiB/s — bottlenecked by zstd level 3 on incompressible data.
Documentation
doc/technical-report.md— Distribution document. What LeVCS is, how to use it, and how it differs from / improves upon git. Targets technical evaluators and the workflow-spec reader.deploy/README.md— Comprehensive VPS deployment walkthrough.spec/— The protocol specification and trust-root revision. Currently kept private; ask the maintainer for a copy.
Contributing
This is a young project. The most useful contributions right now are:
- Trying it. Run
levcs initon a real project, push to a local instance, and report friction. - Workflow design. The next major piece of work is the workflow spec — PR/review surface, issues, CI conventions. Discussion welcome.
- Plugin handlers. The wasm plugin protocol exists; concrete handlers (e.g. protobuf, SQL migrations) are needed to validate it.
- Tightening CI. The
fmtandclippyGitHub jobs are informational; flipping them to gating would close a small but real quality gap.
Please open an issue or reach out before starting non-trivial work so we can coordinate.
License
Released under the Apache License 2.0 — see LICENSE for the
full text.
Citation
If LeVCS supports academic work, please cite the v0.1.0 release. A formal citation entry will land with the workflow spec; in the meantime a repository-URL reference is fine.