GlyphRenderer sizes its GL texture array and fragment-shader sampler array to
exactly TextureAtlas.maxAtlasPages at construction, but _createNewPage pushes
past that cap whenever a 4-page same-size merge is impossible (heterogeneous
sizes, e.g. 1024+512+512+512). The next render throws TypeError: Cannot read
properties of undefined (reading 'version') and rendering freezes — the
single-terminal breakage vector for long heavy-CLI sessions (vscode#322756).
The test builds the renderer with maxAtlasPages=4 (same code path as the
production 8/16/32 values), churns distinct glyphs, and asserts no pageerror,
pages.length <= texture capacity, and an intact pinned header. RED today:
4x TypeError from GlyphRenderer.render.
TextureAtlas.beginFrame() returned a consume-once boolean, so after a page
merge only the FIRST renderer to draw rebuilt its vertex model. Terminals
sharing the atlas via CharAtlasCache kept stale texture-page indexes/coords
and drew garbled glyphs on their next partial render (vscode#322756).
Replace the flag with a monotonic TextureAtlas.pagesVersion, bumped on page
merges and overflow-page creation. Each GlyphRenderer tracks its own
last-seen value in beginFrame(), so every terminal sharing the atlas rebuilds
its model independently. setAtlas() resets the tracker to force a rebuild
when the atlas instance changes.
WebglSharedAtlasGarble now passes (was diff ~130/255 vs tolerance 14); full
webgl addon suite 74/74 green.
Two terminals with equal configs share one TextureAtlas (CharAtlasCache).
When the atlas merges pages it raises a consume-once clear-model flag
(TextureAtlas.beginFrame); only the first terminal to render consumes it and
rebuilds its vertex model. Every other terminal keeps stale texture-page
indexes/coords and draws wrong glyphs on its next partial render even though
its buffer is intact.
The test churns terminal A to force merges, then gives terminal B a tiny
partial update: B's untouched header row garbles (signature diff ~130/255 vs
tolerance 14) while B's buffer still reads HEADER_REF. A DOM-renderer control
runs the same workload without GPU acceleration and passes: the corruption is
exclusive to the WebGL path.
RED on current code by design; the fix lands separately.
generateColoredAsciiFlood emitted a trailing \r\n after its final cell, so a
flood whose last line lands on the bottom row scrolled the screen. In the
alt-buffer repro this destroyed the pinned header row in the BUFFER, making
the 255/255 signature diff a content change, not renderer corruption. The
single-terminal merge path is actually healthy: with the scroll removed the
same 30-chunk merge-heavy workload renders the header with zero drift.
- generateColoredAsciiFlood: no linefeed after the final cell
- WebglReproExperiment: now a green guard; asserts row-1 buffer text first so
a content bug can never masquerade as corruption again
- WebglVisual: accurate comments; erase rewritten lines so leftover prefill
text doesn't look like corruption
- WebglSingleTermStress: replace stale EXPECTED-RED comment (test passes)
* fix(addon-webgl): skip missing buffer lines in _updateModel
_updateModel force-unwrapped buffer.lines.get(row), which can be
undefined during trim or resize. That threw and left stale glyphs on
rows below the gap.
Clear the WebGL model row and glyph cells, reset blink state, and
continue with remaining viewport rows, matching DomRenderer and
ImageStorage handling.
Co-authored-by: Cursor <cursoragent@cursor.com>
* refactor(addon-webgl): extract _nullModelCell helper
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
The webglcontextlost section repeated "fired" ("fires the ... event
fired on the canvas"). Drop the extra word.
Co-authored-by: Cursor <cursoragent@cursor.com>
Correct "transferred", remove duplicate "is", and use "one-off"
instead of "once of" in developer-facing comments.
Co-authored-by: Cursor <cursoragent@cursor.com>
The merging-pages index sort used (a, b) => a > b ? 1 : -1, which
returns -1 when a === b and violates the comparator contract. Use
(a, b) => a - b for ascending numeric order.
Co-authored-by: Cursor <cursoragent@cursor.com>
CustomGlyphRasterizer parsed large-arc and sweep flags with parseInt
without a radix when drawing SVG arc paths. Use base 10 for 0/1 flags.
Co-authored-by: Cursor <cursoragent@cursor.com>
The import/no-cycle refactor had moved IBuffer, IBufferSet, and related
cell/line types into src/common/Types.ts, with buffer/Types.ts only
re-exporting them. That inverted the natural module boundary and made
buffer code depend on the top-level types hub.
Define buffer-owned types in buffer/Types.ts instead:
- CharData, IExtendedAttrs, IAttributeData, ICellData, IBufferLine
- IMarker, IBuffer, IBufferSet, BufferIndex
Restore UnderlineStyle to buffer/Constants.ts (used by IExtendedAttrs)
and remove those definitions from Types.ts.
buffer/Types.ts imports only what it needs from outside the buffer layer:
ICircularList and ICharset from Types, IDisposable from Lifecycle,
IEvent from Event, and UnderlineStyle from Constants. Types.ts does not
import buffer/Types.ts, so the previous Types <-> buffer cycle stays broken.
Update imports across common, browser, headless, addons, and tests to use
common/buffer/Types for buffer types rather than common/Types.
Co-authored-by: Cursor <cursoragent@cursor.com>
Replace the primary lint pipeline with oxlint and oxlint-tsgolint for type-aware TypeScript rules, preserving existing rule behavior from eslint.config.mjs via @oxlint/migrate and targeted jsPlugins (@stylistic/eslint-plugin, eslint-plugin-jsdoc as jsdoc-js, oxlint-plugin-eslint for no-restricted-syntax and object-curly-spacing).
Keep a hybrid ESLint pass (eslint.config.naming.mjs) only for @typescript-eslint/naming-convention until tsgolint implements it (oxc-project/tsgolint#186). Legacy eslint.config.mjs and eslint.config.typings.mjs remain for reference but are no longer used by npm scripts.
Add .oxlintrc.json for src/addons/demo/test and .oxlintrc.typings.json for typings/ (syntax-only, no --type-aware). Update lint, lint-fix, lint-api, and lint-changes to run oxlint first, then naming ESLint. Platform binaries install via oxlint-tsgolint optionalDependencies on Linux, macOS, and Windows.
Fixes https://github.com/xtermjs/xterm.js/issues/5690
Co-authored-by: Cursor <cursoragent@cursor.com>
Replace direct console.log/warn/error calls in the WebGL addon with
ILogService so messages respect the terminal logLevel and custom logger.
Fixesxtermjs/xterm.js#5921
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
configEquals omitted deviceMaxTextureSize, deviceCellWidth, and
deviceCellHeight, which could cause the char atlas cache to reuse an
atlas built for a different GPU texture limit or cell size.
Fixes#5915
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
Register vertex array objects for disposal in GlyphRenderer and
RectangleRenderer, matching existing buffer and program cleanup.
Fixes#5918
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
Dispose _onRemoveTextureAtlasCanvas in TextureAtlas.dispose() alongside
_onAddTextureAtlasCanvas to avoid leaking listeners when the atlas is torn down.
Fixes#5917
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
Three related disposable-registration gaps that leak xterm Terminal
instances when the host application unmounts a Terminal. Evidence +
full writeup at juspay/kolu#606.
1. addon-webgl/WebglRenderer._cursorBlinkStateManager was declared
`= new MutableDisposable()` without the `this._register(...)`
wrapper that every sibling MutableDisposable in the class uses.
CursorBlinkStateManager runs a setInterval for the cursor blink
and its `dispose()` correctly clears it — but dispose() was never
called because the disposable wasn't registered. The live timer
kept the renderer (and thus the Terminal) alive past host
unmount.
2. common/TaskQueue.DebouncedIdleTask had no `dispose()` method.
Added one that clears the internal IdleTaskQueue (cancelling any
pending idle callback).
3. browser/services/RenderService._pausedResizeTask (a
DebouncedIdleTask) was not registered. Wrapped with
`this._register(...)` so RenderService.dispose propagates.
Verified in a downstream app (kolu) with a repro that toggles Focus
vs Canvas layout modes:
- Pre-fix: 24/28 disposed xterm Terminals retained past host unmount
+ forced GC.
- Post-fix: 0/28 retained.
Retention chain (before fix):
DOMTimer → ScheduledAction → V8Function
→ CursorBlinkStateManager._renderCallback → WebglRenderer
→ Terminal
Anthony Kim
jerch
Victor Sych
Sridhar Ratnakumar