v10.0.0-alpha.5
Alpha 5 is a types release. The WebGPU resource hooks now carry three's exact node generics, the stale TSL augmentations are gone, and the built declarations are compiled by a strict consumer fixture on every build. Alongside it: the reconciler oddities that strict-TypeScript consumers surfaced while working around the old types, and the single-canvas depth-attachment fix.
useUniformanduseUniformskeep three's exactUniformNode<TNodeType, TValue>generics and the input's exact keys. Anumberuniform is aNode<'float'>, so TSL math,mix()andbloom()accept it without a cast;.valueis typed; an unknown key is an error (#3769, #3886, #3887).ScopedStoredistinguishes nested scopes from object-valued and callable (Fn()) leaves at runtime, and.scope(key)accepts an explicit schema generic for cross-call access.- The stale
ShaderNodeObjectandthree/tslFnaugmentations are gone. three's ownFnoverloads, including theFn(fn, 'void')statement-call form compute kernels use, are reachable again with R3F installed. Matrix2is an accepted uniform value (mat2), matching three'suniform()overloads.StorageLikeincludesStorage3DTextureandStorageArrayTexture, souseGPUStorageaccepts whatcompute.mdxdocuments.once()is typed as the value the prop declares:translate={once(0, 0, 5)}andcenter={once()}typecheck without a cast.- On the
/webgpuentry,Canvas'sonCreatedreceivesWebGPURootState, so WebGPU-only renderer members no longer need aninstanceofnarrow. pnpm verify-typescompiles a strict consumer fixture against the built declarations withskipLibCheck: false, for every entry.- Reader-mode
useUniforms,useNodes,useBuffersanduseGPUStorageaccept an explicit schema:useUniforms<{ blurAmount: number }>()returns typed nodes for values registered elsewhere, instead ofUniformNode<unknown>. extend(THREE)accepts a whole module namespace without a cast; only its constructors are registered.raycaster={{ params: { Points: { threshold: 0.2 } } }}typechecks:paramsis a partial, matching the runtime merge into the raycaster's defaults.eventSourceaccepts any DOMElement, so anSVGRenderer's<svg>no longer needs a cast.<primitive>keeps typed pointer events (onClick={(e) => ...}inferse) and accepts arbitrary props for the wrapped object.- The examples compile under strict TypeScript without a single
as unknown ascast; the remainingas anysites are third-party gaps (three'sBackend,@use-gesture) and are commented as such.
useTexture'sonLoadreceives the keyed record for record inputs, matching its declared type. It was handeduseLoader's positional array, sotextures.map.wrapTthrew at runtime.- Prefixed elements (
<threeLine>) no longer throw "ThreeLine is not part of the THREE namespace" on their first prop update. - A removed prop resets to the class default for every class. Any class whose constructor takes arguments (every material, geometry and camera) previously had the prop set to
0, turning a removedcolorblack. fromRef(ref, transform)maps the resolved sibling before assignment, so props that need the sibling wrapped are declarative:lightsNode={fromRef(lightRef, (l) => lights([l]))}.- Six-file
.hdrcube sets load throughHDRCubeTextureLoaderinuseEnvironment,<Environment>and the Canvasbackgroundprop, with linear color space. - A lone primary
<Canvas id>owns the renderer's default canvas target, so its depth attachment follows the layout instead of the element's construction size (#3905). - Canvas honors the
resize.debounceprop after the initial measurement (#3881). useRenderPipelinecallbacks receive aWebGPURenderer-typed state, may not return the hook-ownedscenePass, and the hook's return narrowsrenderPipelineonisReady(#3901).- The Pointcloud example's fragment shader handles point colors correctly.
- The demo registry is the one table of contents; the app derives its groups from it. Adds
WebGPUFogandWebGPUPrimaryOnly, drops stale entries, and the examples cleanup from #3883.
no-clone-in-looponly flags.clone()calls. A loop variable namedcloneused asclone.positionno longer trips it.
v2.12.0
- The language encoding flag (general purpose bit 11) is now set only when the encoded filename or the encoded comment of an entry holds a byte outside printable ASCII, where it used to be set on every entry. Nearly every archive zip.js writes therefore differs from v2.11.4 byte for byte. Nothing is decoded differently, printable ASCII being spelled identically in UTF-8 and in Code Page 437, and it is how every other writer decides the flag. Control characters are excluded deliberately: the Code Page 437 table maps 0x01 to 0x1f and 0x7f to the IBM graphic characters rather than to the control characters themselves, so a name or a comment holding one of them keeps the flag and keeps round-tripping
EntryMetaData#filenameUTF8andEntryMetaData#commentUTF8returned byZipWriter#add()now report the flag that was actually written, where they always reportedtrue. They agree with the values read back from the archive, which the previous behaviour contradicted for every entry whose name is printable ASCIIHttpReadernow throws anERR_HTTP_STATUSerror for any status outside 2xx, where it accepted everything below 400. A 304 answered by a caching proxy to a conditional request was read as a successful empty response, and the archive failed later with an unrelated end of central directory error- The
checkPasswordOnlyoption combined withpassThroughset totruenow verifies the password. It reported every password as valid, wrong ones included, and streamed the whole entry instead of stopping after the first block: passing the encryption stage through left nothing able to raise the sentinel that means "the password is correct" ZipDirectoryEntryImportOptions#passThroughand thepassThroughoption ofZipDirectoryEntryExportOptions#readerOptionsare typedbooleaninstead ofboolean | "compressed". The filesystem API has always refused"compressed"at runtime, the types now say so and code passing it stops compiling
- New
passThroughvalue,"compressed", which passes the compression stage through and still runs the encryption stage. It is what re-encrypting an entry without recompressing it needs: reading with it decrypts and does not inflate, writing with it encrypts and does not deflate.passThroughis a depth rather than a pair of switches, and encryption is the outer stage - New
entryoption onZipWriter#add(), which takes the entry read from another archive and derives the ten or so values a copy needs from it. Copying an entry by forwarding the obvious subset used to produce silent corruption, an encrypted entry landing markedencrypted=falseover untouched ciphertext. The oracle the option is built on is byte identity: an archive copied entry by entry with{ passThrough: true, entry }is the archive it was read from - New
checkLocalFilenameoption, which selects whether the filename of the local header is compared, independently of whether a difference throws. The two axes were conflated:strictness: "strict"compared the filename and rejected, everything else did neither, so a mismatched filename could never be reported as a warning.{ checkLocalFilename: true, checkLocalDirectory: false }now reports it onEntryMetaData#warnings ZipDirectoryEntry#exportZip()accepts aZipWriterinstance, the symmetric counterpart ofZipDirectoryEntry#importZip()accepting aZipReader. The entries are added to that writer and the archive is left open, so the caller closes it, can add entries of its own before or after, can export several trees into one archive, and can readZipWriter#warnings- New
ZipWriter#warningschannel, reporting the adjustments the writer makes silently: an entry stored because no deflate implementation is reachable, a date clamped to the MS-DOS range with no extra field left to carry the original value. Each reason is deposited once, with the filename of the first entry it applied to ZipReaderandZipWriterimplementSymbol.asyncDispose, soawait usingfinalizes the archive on scope exit- New
outputSizeproperty on the errors thrown by the codecs, holding the number of bytes that reached the writer before the failure. It is what a caller salvaging an archive after a mid-entry failure needs to know ZipWriter#add()acceptsnullas the reader of an entry with no content- Every user-visible error message is now exported as a constant,
ERR_ABORTEDincluded, and an audit in the test suite keeps it that way across the fourteen builds configure()checks the type ofbaseURI,workerURIandwasmURI, which used to fail much later and quietly
- The codecs of a reader stream are now released on demand and on cancel, instead of being held until the whole archive was read
- A failure raised on the main thread while a codec task runs over the worker message protocol now carries
outputSize, so the writer advances by the bytes that really reached it. It did not, and the central directory of an archive salvaged after such a failure was written short of the real positions, by 655360 bytes in the measured case - Terminating the workers no longer leaves the WebAssembly module torn down while an operation that reloaded it believes it is loaded, which silently substituted the native gzip path for the rest of the process
preventCloseno longer makesFileEntry#arrayBuffer()andZipReaderStreamhang forever. Both build their writable internally, exactly like theWriterinstances the option is documented to ignore- A copy made with the
entryoption now derives the deflate level bits, the language encoding flag and the unix extra field type of the source entry. It dropped the level bits, cleared bit 11 on an ASCII-named source that carried it, and rewrote a 0x7855 record as a 0x7875 one - The refusal to change the last modification date of a ZipCrypto entry no longer fires under
passThrough: "compressed", where the entry is encrypted again or not encrypted at all and the date is free to change ZipDirectoryEntry#exportFileSystemHandle()now refuses the"compressed"value ofpassThroughspelled at the top level, as it already did throughreaderOptions. It wrote the raw deflated bytes into the extracted files insteadZipDirectoryEntry#getExportedSize()isasync, so its option validation errors reject instead of throwing synchronously out of a method declared to return a promise- The
ERR_INVALID_UNCOMPRESSED_SIZEerror now carriesoutputSize, which the most common corruption shape did not set - An entry written with
passThroughand no crc32 to declare is refused instead of being written with a wrong one, and AE-2 is written when encrypting passthrough content - The local header and the central directory now agree on the extended timestamp record
- The
extendedTimestampremark no longer claims that EPUB allows an extra field on itsmimetypeentry. OCF states that there must not be one, which is what pins the byte offset ofapplication/epub+zipso a reader can sniff the format without parsing the archive, and epubcheck reports it as an error. ODF and EPUB need the same pair of options - Twelve other claims corrected against what the code does, among them: the precedence of the values the
entryoption reads, which beat the options of theZipWriterrather than being defaults against them; the condition under whichWARNING_CLAMPED_LAST_MODIFICATION_DATEappears; the combinations under whichLocalDirectory#rawFilenameis defined; the status codesERR_HTTP_STATUScovers; whatERR_INVALID_URIvalidates; and the determinability ofgetExportedSize()for names holding a slash, obsolete since implicit directories were exempted - The options of
ZipWriter#close()are documented as belonging to the caller closing the archive when aZipWriteris passed toexportZip(),globalCommentandsignCentralDirectoryincluded - The
entryoption points at the branch it selects for an encrypted source: it suppliesencrypted, so copying such an entry withpassThroughset totruewrites the ciphertext as-is and keeps the password it was encrypted with, while a password in scope encrypts the other entries only. Re-keying is what"compressed"is for
- The whole lint script runs in CI,
deno lintincluded, and the fidelity harness runs intest-ci - New fixture and test covering a Windows Explorer zip64 archive: a Deflate64 entry too large for 32 bits, in an archive carrying no zip64 end of central directory, with the two headers sentinelling different size fields
- New tests for the passthrough stages, the copy entry option, the async dispose path, the reader stream release, the writer warnings and the filesystem export into a
ZipWriter - The type tests are checked to be registered in a tsconfig, and the API export surface is audited across every build
Full Changelog: https://github.com/gildas-lormeau/zip.js/compare/v2.11.4...v2.12.0
@tanstack/solid-start-client@2.0.0-rc.7
- Updated dependencies [
a09b50c]:- @tanstack/solid-router@2.0.0-rc.7
@tanstack/solid-start@2.0.0-rc.7
- Updated dependencies [
a09b50c]:- @tanstack/solid-router@2.0.0-rc.7
- @tanstack/solid-start-client@2.0.0-rc.7
- @tanstack/solid-start-server@2.0.0-rc.7
@tanstack/solid-start-server@2.0.0-rc.7
- Updated dependencies [
a09b50c]:- @tanstack/solid-router@2.0.0-rc.7
v16.4.0-canary.20
- test(turbopack): remove dead top-level task unmarks: #98246
- Fix flaky successive HMR changes test: #97610
- docs: Remove version label from Turbopack in Pages Router: #98316
- turbo-tasks-backend: fix strongly consistent read hanging on a canceled task: #98312
Huge thanks to @lukesandberg, @sokra, and @icyJoseph for helping!
10.4.0
- destroy event; Transformer, Tween, pointer captures and stage timers release destroyed nodes (Anton Lavrevov)
- size the buffer canvas of cache() and toCanvas() to the target, allocate it lazily and release it (Anton Lavrevov)
- memoize isCSSFiltersSupported() and release its probe canvas (Anton Lavrevov)
- stop the frame interval of a Sprite on destroy() (Anton Lavrevov)
- keep descendant caches on clearCache() and release the previous canvases on re-cache (Anton Lavrevov)
- setAbsolutePosition() under a zero-scale ancestor no longer writes NaN (Anton Lavrevov)
- negative radius and cornerRadius no longer throw and blank the layer (Anton Lavrevov)
- batchDraw() no longer freezes the layer after a throwing draw (Anton Lavrevov)
- Blur keeps the colour of semi-transparent pixels and survives large radii (Anton Lavrevov)
- Tween ownership bookkeeping (Anton Lavrevov)
- clamp a negative cornerRadius in drawRoundedRectPath (Anton Lavrevov)
- hasShadow() follows shadowOffsetX and shadowOffsetY changes (Anton Lavrevov)
- cache() and transform batching clean up after a throw (Anton Lavrevov)
- split text into graphemes with Intl.Segmenter (Anton Lavrevov)
- toObject() copies attributes instead of mutating them, accepts frozen values and skips methods named like custom attributes (Anton Lavrevov)
- drawHitFromCache() honours hitCanvasPixelRatio (Anton Lavrevov)
- Path arc bounds, SVGO arc flags, empty data, and segment-relative getPointAtLength (Anton Lavrevov)
- Transformer follows the grabbing pointer, guards zero-size boxes, and resolves its anchors directly (Anton Lavrevov)
- Tween of component attributes, of array attributes the node lacks, and to() no longer mutates its params (Anton Lavrevov)
- absolute getters and derived caches of nodes under a cached container, fire(type, null), non-finite child rects (Anton Lavrevov)
- Canvas size equals the truncated bitmap at a fractional pixel ratio and keeps its logical size; setSizeIfChanged() with NaN (Anton Lavrevov)
- drag with several pointers releases only the lifted pointer; startDrag() without an event ends on release (Anton Lavrevov)
- Transformer suppresses the hit graph of its own layers only (gate moved to Layer.shouldDrawHit), forwards getClientRect(config), honours top in getAbsoluteTransform() (Anton Lavrevov)
- RegularPolygon without sides, Sprite getSelfRect() and guards for a missing animation or frame (Anton Lavrevov)
- report a non-finite number at the core aggregation points instead of silently drawing nothing (Anton Lavrevov)
- Tag.getSelfRect() grows by exactly the pointer for a left or right pointer (Anton Lavrevov)
- eleven one-line bugs (Anton Lavrevov)
- layer.size() no longer resizes the layer canvas; getChildren() returns a copy (Anton Lavrevov)
- review follow-ups for the polishing batch (Anton Lavrevov)
- Transformer follows node changes made between pointer moves of a resize (Anton Lavrevov)
- JSDoc that contradicted the code, wrong names and types, examples that did not run (Anton Lavrevov)
- rewrite the unreleased changelog in the usual short style (Anton Lavrevov)
- keep a running grapheme index for charRenderFunc instead of recounting previous lines per character (Anton Lavrevov)
- upload filter chain result once and stop reallocating the filter canvas (Anton Lavrevov)
- release the per-layer canvases of a stage export and skip the client rect when the box is given (Anton Lavrevov)
- resize a canvas in one pass (Anton Lavrevov)
- drop the two Maps every node allocated (Anton Lavrevov)
- wrap fixed-width text in linear time (Anton Lavrevov)
- resolve edge pixels of the hit graph directly and bound the search (Anton Lavrevov)
- relayout listeners of Text and TextPath on the prototype, listener cache keyed by class (Anton Lavrevov)
- getClientRect() no longer scans the subtree at every nesting level (Anton Lavrevov)
- drag reads pointer positions once per stage, skips the redraw on a plain click, and resets the drag position per drag (Anton Lavrevov)
- hoisted component getter names, single-array Transform, direct drawImage, hoisted TypedArray (Anton Lavrevov)
- Animation keeps its running animations in a Set; an animation stopping itself no longer skips the next one; setLayers() copies its argument (Anton Lavrevov)
- hygiene (Anton Lavrevov)
- keep JSDoc in the published .d.ts files (Anton Lavrevov)
- size budget of 48 KB, CI runs size, import and type checks on Node 22 (Anton Lavrevov)
- release script runs the tests, drops the empty cdn-link commit and the hard-coded directory (Anton Lavrevov)
- named exports and prepublishOnly (Anton Lavrevov)
- drop the named exports and the separate UMD entry (Anton Lavrevov)
- remove dead Node internals and hoist the non-bubbling event list (Anton Lavrevov)
- build the browser tests from a clean Parcel cache (Anton Lavrevov)
- drop the unregistered fillLinearRadial* declarations, build ellipses with context.ellipse() (Anton Lavrevov)
- remove dead code (hit fill/stroke duplicates, lineDash fallbacks, trace branch, unused globals, addDeprecatedGetterSetter) (Anton Lavrevov)
- red/green/blue registered once with the shared component validator, alpha uses afterSetFilter (Anton Lavrevov)
- remove dead code; cloneObject() copies nested arrays and typed arrays; shadow probe releases its canvas (Anton Lavrevov)
- resolve every TODO comment in src and tests (Anton Lavrevov)
- remove dead code, dead files and unused devDependencies (Anton Lavrevov)
- ad94d9d: update CHANGELOG (Anton Lavrevov)
- f142b6e: update CHANGELOG (Anton Lavrevov)
- 7ee3232: update CHANGELOG (Anton Lavrevov)
- c9f3830: update CHANGELOG (Anton Lavrevov)
- 0f88c28: update CHANGELOG (Anton Lavrevov)
- declarations that contradicted the runtime or the docs (Anton Lavrevov)
- ac581a4: update CHANGELOG with new version (Anton Lavrevov)
- e51c16d: build for 10.4.0 (Anton Lavrevov)
electron v45.0.0-alpha.5
Note: This is an alpha release. Please file new issues for any bugs you find in it.
This release is published to npm under the alpha tag and can be installed via npm install electron@alpha, or npm install electron@45.0.0-alpha.5.
- File System Access permission requests and the
file-system-access-restrictedevent are scoped to the requesting document, grants are reset when the origin's last page closes, and write access works in in-memory sessions. #53690 (Also in 43) - Fixed
Trayicons not appearing when running inside Flatpak or Snap sandboxes on Linux. #53654 (Also in 44) - Fixed
app.getGPUInfo('complete')stalling the GPU process for several seconds on some Windows systems. #53645 (Also in 44) - Fixed
session.setPermissionCheckHandlerreceiving the top-level origin and a nullwebContentsforhidandusbchecks made from a subframe. #53679 (Also in 43, 44) - Fixed
webContents.on(),removeListener()andremoveAllListeners()throwing "Object has been destroyed" forconsole-messagelisteners after the WebContents was destroyed. #53495 (Also in 42, 43, 44) - Fixed a crash on Linux when
process.envwas written while another thread was reading the environment, and a memory leak when a worker thread exits. #53508 (Also in 42, 43, 44) - Fixed a possible crash on Windows when a file dialog was shown for a window that was being closed at the same time. #53586 (Also in 42, 43, 44)
- Fixed a renderer crash when the main process sent IPC to, or a page navigated, a same-process
window.open()child whosecontextIsolationdiffered from its opener's. #53538 (Also in 42, 43, 44) - Fixed a spurious
node_initerror being logged when creating aShadowRealmwith Node.js integration enabled. #53578 (Also in 44) - Fixed crashes in
setDisplayMediaRequestHandlerwhen the granted frame had been destroyed or another tab was granted by id. #53673 (Also in 42, 43, 44) - Fixed several non-functional DevTools integrations: the Security panel's "View certificate" button, Ctrl+wheel zooming inside DevTools, and DevTools keyboard shortcuts (e.g. F8) while the inspected page has focus. #53471 (Also in 43, 44)
- Fixed window state not being persisted when a window entered or left fullscreen, or was maximized or unmaximized, without also being moved or resized. #53577 (Also in 44)
openExternalpermission requests started by a frame that has since gone away are attributed to that frame's origin rather than to the navigating page. #53696 (Also in 42, 43)pointerLockandkeyboardLockpermission requests now report the requesting frame, andexecCommand('paste')requires user activation in the frame that calls it. #53692 (Also in 44)
- Electron's internal JavaScript bundles are no longer listed in
require('module').builtinModulesor loadable throughprocess.getBuiltinModule(). #53624 - Fixed the app becoming unresponsive when a page has a very large number of
app-regiondraggable elements. #53596 (Also in 43, 44) - Improved throughput of responses that
protocol.handlehandlers return straight fromnet.fetch. #53379 (Also in 44)