* feat!: redesign bcrypt creation API and narrow Argon2 verification options
* chore: keep major API changes scoped to bcrypt
* fix(bcrypt): preserve Node 10 and 12 runtime compatibility
* fix(bcrypt): support cancellation without global AbortController
* fix(bcrypt): polish the 2.0 API and packaging before release
- Settle the public Promise before removing the abort listener, and remove
it with the same options, so throwing or capture-flag EventTargets cannot
hang a call or leak the listener.
- Keep signal.reason as a non-enumerable AbortError cause, so callers can
tell a TimeoutError or custom reason apart.
- Throw TypeError for option values of the wrong type and keep RangeError
for out-of-range or malformed values.
- Copy only the 72 password bytes bcrypt reads for async work, and wipe
those copies on drop; replace the unreachable!() salt branch.
- Declare package exports so internal files cannot be deep-imported.
- Stop declaring the WASI package as an optional dependency, so installs do
not pull it and a missing native binary fails loudly.
- Leave the version bump to the release step, add the 2.0.0 changelog, and
preapprove this repo's own @node-rs/* packages.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(bcrypt): keep the event loop alive while waiting for AbortSignal.timeout
AbortSignal.timeout() uses a timer that does not keep the event loop
alive, and the controlled backend has no native work that would. The
test passed only when other tests kept the worker busy, and failed with
"Promise returned by test never resolved" on the Linux and WASI jobs.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore: resolve bcrypt-previous from the registry
With the bcrypt workspace back at 1.10.9, yarn's transparent workspaces
link bcrypt-previous (npm:@node-rs/bcrypt@1.10.9) to the workspace
itself whenever the lockfile is regenerated, so the previous-release
tests would compare the new code with itself. No workspace depends on
another, so only the workspace: protocol needs to link workspaces.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(bcrypt): regenerate bindings with @napi-rs/cli 3.10.5
main moved @napi-rs/cli from 3.9.0 to 3.10.5. Regenerate the bcrypt
loaders and type declarations from native and wasm32-wasip1-threads
builds so the committed files match what the release build produces.
binding.js still parses as ES2019 for Node 10 and 12.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bcrypt): keep custom signal errors from breaking settled or aborted calls
A signal whose removeEventListener throws made a successful call raise an
unhandled rejection, which exits modern Node, and on abort it skipped the
native cancellation. A signal whose reason getter throws stopped the abort
listener before it settled, so the call stayed pending or resolved after
the abort. Ignore cleanup failures once the call has settled, and read the
optional reason defensively.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(bcrypt): add parseOptions and Node-style error codes for 2.0
Bundle two more contract changes into the 2.0 break so the error and
rehash-on-login story is complete at release:
- Thrown errors carry Node-style codes: TypeError ERR_INVALID_ARG_TYPE,
RangeError ERR_OUT_OF_RANGE (including native InvalidArg translations),
and AbortError ABORT_ERR, matching Node's own AbortError.
- parseOptions(hash) returns { version, cost } through the verifier's
parser (bcrypt::HashParts), so every hash verify accepts is parseable,
including +4 costs and imported 2x labels; hashes verify always rejects
throw RangeError. Mirrors argon2's parseOptions in this repo.
- AbortSignalLike.removeEventListener declares the options argument the
adapter actually passes.
- Drop unused blowfish and quickcheck dependencies from the bcrypt crate
and the workspace.
Native and wasm32-wasip1-threads bindings are regenerated. api.cjs still
parses as ES2019 for Node 10 and 12.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bcrypt): tidy parseOptions review nits
- Use one options type for add/removeEventListener on AbortSignalLike,
matching the comment that the same object is passed to both.
- Give parseOptions its own arity message; it never had positional
options to drop.
- Cover the 60-byte non-ASCII parser branch in the test, which the
previous 61-byte input missed.
- Mention in the README that parseOptions throws for hashes verify
always rejects.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(bcrypt): give the backend contract error a code
A stale platform package now fails at import time with
code ERR_BCRYPT_INCOMPATIBLE_BINARY, so every error the package throws
sits inside the documented ErrorCode contract.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(bcrypt): switch hashing backend to bcrypt-rust
Replace the bcrypt 0.19 dependency with bcrypt-rust 1.0, keeping the crate
name so imports stay valid. The new crate is a drop-in for every API used
here except one deliberate difference: its HashParts parser is strict and
rejects the noncanonical '+4' cost spelling the old backend accepted.
Verify and parseOptions now use a lenient parser that mirrors the old
backend byte-for-byte (60 ASCII bytes, 2a/2b/2x/2y markers, u32::parse for
cost), then hand the canonicalized hash to bcrypt::verify so the cost gate
and constant-time compare stay in the library.
~20% faster than the C++ backend on hash and verify.
* bench: make benchmark comparisons 1:1 and migrate argon2 to tinybench
bcrypt: uniform call signatures via per-impl adapters, a fixed salt so the
hash suite measures hashing only, a real sync compare for verify (it
previously timed an unawaited async call with a hash inside the loop), one
cost across suites, all five implementations in every suite, and a compact
sorted table with latency, rme, ops/s, and a ratio to the baseline.
argon2: replace the hand-rolled performance.now() harness with tinybench
fixed-iteration runs, keep the per-iteration tag assertion and the
native-sync / native-async / js-wasm grouping, and adopt the same report
format.
* bench: narrow tinybench TaskResult before reading stats
TaskResult is a discriminated union; latency/throughput exist only on
completed and aborted-with-statistics. CI type-checks benchmark files
through packages/argon2/tsconfig.json, unlike oxnode which strips types.
* perf(bcrypt): fast-path genSaltSync options; note wasm salt overhead in bench comment
genSaltSync previously ran every call through options() + creation(),
allocating a fresh Object.create(null) copy even for {} and {cost}. Now
undefined uses a frozen empty options object and reusable options shapes
validate in place — identical error paths preserved for non-plain or
extra-key objects. Saves ~200ns of JS wrapper overhead per call.
* deps: bump bcrypt-rust to 1.0.1
Picks up the crypt_blowfish fused-round kernel shape
(Brooooooklyn/bcrypt-rust@f65cd90): bcrypt() at cost 10 on x86_64
drops ~19% (138.2ms -> 112.6ms vs vendored C at 108.7ms); arm64
unchanged.
* perf(bcrypt): cut genSaltSync overhead; pool OS entropy per thread
genSaltSync now serves 16-byte salts from a 256-byte thread-local pool
refilled by SysRng (direct getrandom), amortizing one syscall over 16
salts without any userspace PRNG in between. The salt string is built
without format! machinery and the bcrypt base64 engine is constructed
once in a LazyLock instead of per call.
api.cjs gains a specialized fast path for the {cost,version} shapes:
getOwnPropertyNames/getOwnPropertySymbols cover the same key space as
Reflect.ownKeys at half the cost, the sync() closure is inlined, and
anything the fast path cannot prove valid funnels into a shared full-
validation path so error types, codes, and messages stay identical.
x86_64 sandbox (30k-iteration amortized loop, best of reps):
@node-rs/bcrypt ~380-560ns vs wasm OpenBSD ~550-730ns
per-call min: 210-250ns vs 320-370ns; bench.js min 290ns vs 860ns.
* deps: bump bcrypt-rust to 1.0.2
Picks up the xor_late asm boundary (Brooooooklyn/bcrypt-rust#8): LLVM no
longer reassociates the fused round into two serial xors after the S-box
loads. On x86_64 EPYC the raw kernel now beats crypt_blowfish C
(~54.6ms vs ~54.9ms at cost 10) and JS hashSync/verifySync beat node
bcrypt (C++); arm64/wasm32 untouched via cfg fallback. CodSpeed: -4% on
all single-hash benches.
---------
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Bump the workspace jieba-rs dependency from 0.10 to 0.11.0. No source
changes are required; the addon builds and all tests pass.
Refresh the README benchmark tables with a fresh run of @node-rs/jieba
against nodejieba@3.5.8, and print a label before each tinybench table
so the suites can be told apart in the output.
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* perf(argon2): validate UTF-8 with simdutf8 before owning the string
Check the JS byte slice with simdutf8, then build String via
from_utf8_unchecked so from_utf8 does not walk the bytes again.
* perf(argon2): decode PHC Base64 with a LUT instead of the engine
Salt and tag fields are tens of bytes. A 256-entry table plus leftover-bit
checks beats the generic base64 crate, and on aarch64 argon2-rust's SIMD
would not even run for a 16-byte salt.
* perf(argon2): decode PHC fields with argon2-rust SIMD Base64
Use the new encode_base64/decode_base64 entry points. decode_phc still
owns param order and data=. Temporarily depends on the local
argon2-rust tree until 1.1.0 is on crates.io.
* feat(argon2): consume argon2-rust 1.1 PHC APIs
Drop the local PHC parser, encoder, and compare. Hash and verify go
through hash_encoded_with_ad, decode_phc, and verify_with_ad.
Require argon2-rust 1.1 from crates.io; the sibling path dep is gone.
* feat(argon2): switch backend to argon2-rust
Replace RustCrypto argon2 with argon2-rust 1.0.0 from crates.io.
Keep the public JS API. Verify accepts node-argon2 PHC param order
and associatedData. The bench pins the same raw params, salt, and
tag across implementations.
* fix(argon2): keep Node 22 tests and bound verify threads
Skip the node:crypto.argon2Sync comparison on runtimes that lack it.
Cap worker threads at available parallelism so a PHC p= value cannot
spawn one OS thread per lane. Allocate tag buffers with try_reserve
instead of vec! so a huge outputLen is a JS error, not a process abort.
* fix(argon2): skip node-argon2 interop when its native addon is missing
node-argon2 0.45 has no darwin-x64 prebuild. Intel macOS CI runs x64
Node on an ARM host, so a static import crashed the whole spec file.
* test(argon2): compare against node:crypto argon2
Drop the node-argon2 native addon from tests. Interop uses the
builtin KDF, which is available on every CI host that has it.
* chore: upgrade all npm deps
* fix: update tsconfig for TypeScript 7
Remove deprecated downlevelIteration and importsNotUsedAsValues options,
and explicitly include node types required by TypeScript 7.
* fix: download test artifacts into package dirs
@napi-rs/cli 3.8 requires all configured targets when running
napi artifacts. Test jobs only have one platform, so download
bindings directly into the workspace like the package template.
* fix: download bindings into packages/ for tests
upload-artifact uses packages/ as the common root, so the artifact
contains argon2/*.node not packages/argon2/*.node. Download into
packages/ so binaries land where the loaders expect them.
* fix: add WASI .d.cts files required by @napi-rs/cli 3.8
napi artifacts now requires a complete WASI loader set including
*.wasi.d.cts. Also upload the full WASI support files from the wasm
build job for the publish path.
* fix: upload only wasm artifacts from WASI builds
WASI loaders and types stay committed in the repo and updated with
the codegen; CI only needs the built .wasm binaries.
renovate[bot]