TextureAtlas.clearTexture() wipes every page and both cache maps but never
bumped _pageLayoutVersion, so it silently broke the invalidation contract that
#6042 introduced for shared atlases.
The atlas is shared by every terminal with an equal config (CharAtlasCache),
but WebglRenderer.clearTextureAtlas() only clears the *calling* renderer's
model. Every other owner keeps vertex data pointing into the rows that were
just wiped, and GlyphRenderer.beginFrame() has no way to notice, so those
terminals keep drawing from stale texture coords until something else happens
to invalidate them. Once the clearing terminal rasterizes new glyphs into the
freed rows, the siblings render whatever landed there instead of their own
text.
Embedders hit this whenever they force a redraw on one terminal while others
are alive; VS Code calls Terminal.clearTextureAtlas() on OS resume, which is
why a stale terminal there stays garbled even after opening a new one (the new
terminal re-acquires the same cached atlas) and only recovers once the whole
panel is closed and the last owner drops the atlas.
Bump _pageLayoutVersion on clear so sibling renderers rebuild their models on
their next frame, matching what page merges and evictions already do.
The added test fails on master (terminal B's row renders terminal A's flood
glyphs) and passes with the fix.
* fix(addon-search): advance highlight scan by buffer match size
_highlightAllMatches() stepped to the next find() using term.length
and col + 1 when staying on the same row. Search results expose size
in buffer columns (wide glyphs, wrapped matches), which select() and
decorations already use.
Advance with col + size and wrap row/col when crossing the terminal
width so highlight-all does not skip, duplicate, or stop early on
non-ASCII matches.
Co-authored-by: Cursor <cursoragent@cursor.com>
* test(addon-search): regression for highlight-all wide char advance by size
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
* test(addon-search): move highlight-all regression to integration tests
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>
Upgrade actions/checkout, setup-node, upload-artifact, download-artifact,
and codeql-action to their latest major versions that run on Node 24.
Replace NathanFirmo/wait-for-other-job with native needs: build for
test-unit and test-integration jobs.
Fixes#6006
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
Remove early return when ybase===0 && y===0 so clear() always
clears markers, resets buffer lines, and refills the viewport.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
* fix(selection): drag-scroll end row matches last visible line
Use ydisp + rows - 1 for selectionEnd when scrolling down during drag,
consistent with finalSelectionEnd for select-all.
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
* fix(selection): reset start to origin when start row is trimmed
When handleTrim removes the selection start row, clamp to [0,0] and
refresh; do not keep a stale column on row 0.
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
* 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>
* fix(DomRenderer): skip missing buffer lines during row render
renderRows and link underline refresh stopped updating when any
viewport row lacked buffer data, leaving stale DOM on rows below.
Match ImageStorage.render behavior: clear the row element when
present and continue with remaining rows.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(DomRenderer): clear blink state when skipping missing lines
When a viewport row has no buffer line, renderRows clears the row DOM
and continues but did not reset _rowHasBlinkingCells for that row.
That could leave the text blink manager thinking blinking cells remain
after the row was emptied.
Clear blink state for the row when skipping a missing buffer line,
including during link underline refresh.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Character joiner scanning already iterated cells with
getTrimmedLength(), but trailing attribute runs and
_stringRangesToCellRanges used buffer cols. On short lines that
could run joiners on a single trailing cell and extend open ranges
through empty tail columns, affecting ligature/join rendering.
Use trimmed length consistently for trailing-run detection, the
cell-range walk, and the final range end column.
Co-authored-by: Cursor <cursoragent@cursor.com>
Mocha 11.7.x pulls in serialize-javascript 6.0.2, which is affected by
code injection via RegExp.flags and Date.toISOString (CVE incomplete fix).
Force the patched 7.x release via npm overrides.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
* Add InputHandler unit tests for ISO-2022 charset mapping
InputHandler unit tests always used MockCharsetService, so regressions in
print()'s ASCII remapping, G0/G1 designation (ESC ( ) * +), and SO/SI
level switching were never caught at unit level.
Add an ISO-2022 character sets suite wired to real CharsetService and
BufferService that verifies DEC line drawing on G0 (ESC ( 0), G1 after
ESC ) 0 + SO, and charset/glevel restoration via ESC 7 / ESC 8 before
printing.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add BufferService.scroll unit tests for isUserScrolling
BufferService.scroll() adjusts ybase and ydisp differently depending on
whether the circular buffer is full and whether isUserScrolling is set,
but no unit test asserted those branches. InputHandler and other callers
use MockBufferService.scroll, which throws, so real scroll viewport logic
was only exercised indirectly.
Add BufferService.test.ts covering full-buffer scroll with the user
scrolled up (ydisp decrements), partial-buffer scroll while scrolled up
(ydisp unchanged), and full-buffer scroll at the viewport bottom
(ydisp tracks ybase).
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add InputHandler unit tests for DECOM cursor positioning
CUP with DEC origin mode (DECOM) offsets rows by the scroll region top
and clamps within DECSTBM margins, but unit tests only toggled the mode
for save/restore. Playwright covered the combination; headless unit tests
did not.
Add a cursor positioning test matching the Playwright DECOM case: with
margins and origin on, CUP places the cursor relative to scrollTop, clamps
to the margin, and uses absolute rows when origin is off.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add WindowsMode unit tests for wrap heuristic edges
ConPTY does not set isWrapped on soft-wrapped lines; updateWindowsModeWrappedState
infers wrapping from the last cell of the previous row. Browser Terminal tests
exercise this only through full writes, with no direct coverage of the helper
or negative cases (trailing space or null).
Add WindowsMode.test.ts asserting isWrapped is set for a full non-space line,
cleared for trailing whitespace, and cleared when the previous line ends empty.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add OscParser.reset unit test for mid-payload abort
OscParser.reset() must call end(false) on active handlers when aborted
during PAYLOAD, but no test invoked reset() directly. EscapeSequenceParser
tests mock OSC, so production reset/cleanup was unverified at unit level.
Add a reset describe block asserting handlers receive END false after a
partial payload, then a subsequent OSC sequence completes normally.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add DcsParser.reset unit test for mid-payload abort
DcsParser.reset() must call unhook(false) on hooked handlers when aborted
mid-DCS, matching OscParser's end(false) cleanup. No DcsParser test called
reset() directly, leaving the abort path unverified at unit level.
Add a reset describe block asserting UNHOOK false after partial put, then a
subsequent DCS sequence completes with UNHOOK true.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add InputHandler print tests for grapheme joining
print() uses UnicodeService.shouldJoin to merge combining marks, but
InputHandler tests never exercised that path. The U+200B infinite-loop
regression test used mocks and had no assertions, so it would pass even
if print() misbehaved.
Add tests for e + combining acute in one parse and across two parses.
Strengthen the U+200B regression test to use BufferService and assert the
buffer does not grow spurious rows.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add CharsetService unit tests for glevel and designation
CharsetService only updates the active charset when setgCharset targets the
current glevel; most InputHandler tests use MockCharsetService instead. A
regression in setgLevel or inactive-slot designation would not be caught
without direct service tests.
Add CharsetService.test.ts covering inactive designation, setgLevel
activation, active-slot updates, and reset clearing state.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add reflowCursorLine unit tests for ConPTY resize behavior
reflowCursorLine lets resize reflow wrapped lines that contain the cursor
(needed when ConPTY does not). No test referenced the option or
reflowLargerGetLinesToRemove's cursor skip branch.
Add BufferReflow tests for skip vs reflow when the cursor is inside a
wrapped block, and Buffer.resize tests showing merge on widen only when
reflowCursorLine is enabled.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add EscapeSequenceParser reset test for async pause recovery
reset() sets ParserStackType.RESET so the next parse continues in GROUND
at chunkPos + 1 without resuming a paused async handler. Existing async
tests only verified reset lifts FAIL after improper continuation, not
mid-chunk recovery with trailing printable data.
Add a test that pauses on async CSI, calls reset(), then continues parsing
and asserts XY is printed once and the stack returns to NONE.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add Buffer.resize unit test for ConPTY row growth
When rows increase on Windows ConPTY, Buffer.resize appends lines without
decrementing ybase or ydisp so the viewport stays stable while ConPTY
repaints. Buffer tests covered the non-Windows scroll-up path but never
passed windowsPty into options.
Add a ConPTY row-growth test asserting ybase, ydisp, and line count when
windowsPty is configured.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add Buffer reflow test for ConPTY buildNumber threshold
Buffer disables reflow on ConPTY when buildNumber is below 21376, but no
unit test asserted that gate. Column shrink without reflow leaves wrapped
content on a single line instead of splitting across rows.
Add a test comparing shrink behavior at build 21375 versus 21376, plus
retain unwrap-on-widen coverage for supported ConPTY builds.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add ApcParser.reset unit test for mid-payload abort
ApcParser.reset() calls end(false) on active handlers when aborted mid-APC,
matching OscParser and DcsParser, but no ApcParser test invoked reset()
directly.
Add a reset describe block asserting END false after partial put and a
clean subsequent APC sequence.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add InputHandler unit test for REP with grapheme clusters
repeatPrecedingCharacter uses parser.precedingJoinState and reprints the
full cluster from the buffer, but only Playwright covered e + combining
acute with CSI b. A regression could break REP for joined sequences without
failing unit tests.
Add a print-suite test asserting CSI 2 b triples e + acute in one parse.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add InputHandler unit tests for REP no-op cases
repeatPrecedingCharacter returns early when parser.precedingJoinState is
zero, including after another escape sequence clears join state. Only the
positive grapheme REP case was unit tested.
Add tests for bare CSI b and for REP following SGR so regressions cannot
spuriously duplicate buffer content.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add DecorationManager unit test for wrapped search matches
DecorationManager splits matches across wrapped rows with per-segment x
and width, but only SearchEngine had unit coverage; decoration geometry
was left to Playwright.
Add DecorationManager.test.ts using a real Terminal, a wrapped-buffer
match from SearchEngine, and assertions on registerDecoration options
for both row segments.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add BufferService.scroll unit test for DECSTBM margins
When scrollTop is non-zero, scroll() shifts lines inside the margin with
shiftElements instead of pushing scrollback. Existing BufferService tests
only covered scrollTop === 0.
Add a test with margins on rows 1–3 that verifies rows outside the region
stay fixed and lines inside the region scroll correctly.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add BufferService.scrollLines unit tests
scrollLines adjusts ydisp and toggles isUserScrolling for viewport
scrolling, but only scroll() had unit coverage. Regressions in the
user-scroll flag or ydisp clamping would affect the whole viewport.
Add tests for scrolling up, clamping at ydisp 0, and clearing
isUserScrolling when reaching the buffer bottom.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add DecorationManager overview ruler dedup unit test
Multiple highlight matches on the same buffer line should only register
one overview ruler marker; later matches on that line omit ruler options.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Assert overview ruler on each line of wrapped search match
Wrapped highlight decorations span two buffer lines; each segment
should register overview ruler options unlike same-line dedup.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(buffer): bound combined-cell copy in copyCellsFrom
When copying a column range, combined-string metadata was copied for
every source column >= srcCol with no upper limit. Partial copies could
pull _combined entries from outside [srcCol, srcCol + length) into the
destination and corrupt cells beyond the copied span.
Only copy combined keys within the same range as the cell data loop.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(buffer): clear stale _combined entry in BufferLine.setCell
setCell updated combined data when IS_COMBINED_MASK was set but left
orphaned _combined[index] entries when writing a non-combined cell.
Delete the map entry in that case so later combined copies and lookups
stay aligned with cell content flags.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(buffer): clear stale extended attrs in BufferLine.setCell
setCell stored extended attributes when HAS_EXTENDED was set but did
not remove orphaned _extendedAttrs[index] when overwriting with a
cell without extended background data. Delete the entry in that case,
matching the combined-cell cleanup added for setCell earlier.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(buffer): clear dest extended attrs in copyCellsFrom
When copying a column range, extended attributes were copied from
source cells with HAS_EXTENDED but stale _extendedAttrs on the
destination were left for cells without extended data. Delete dest
entries in that case, matching setCell behavior and the bounded
_combined copy loop.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(buffer): clear extended attrs in setCellFromCodepoint
The input-handler fast path set extended data when HAS_EXTENDED was
present but left orphaned _extendedAttrs entries when overwriting
with non-extended cells. Delete the map entry in that case, matching
setCell and copyCellsFrom.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(buffer): clear stale combined data in setCellFromCodepoint
When the input handler overwrites a cell that previously held a
grapheme cluster (IS_COMBINED_MASK), setCellFromCodepoint updated
_data and _extendedAttrs but left this._combined[index] intact.
setCell already deletes _combined for non-combined cells; this path
did not, so getString and renderers could show the old cluster string
after a normal codepoint write.
Add delete this._combined[index] at the start of setCellFromCodepoint,
matching setCell's non-combined branch and the extended-attrs cleanup
already performed in the same method.
Co-authored-by: Cursor <cursoragent@cursor.com>
* perf(buffer): only delete stale combined/extended map entries when present
Co-authored-by: Daniel Imms <Tyriar@users.noreply.github.com>
* fix(buffer): treat combined/extended maps as flag-gated caches
_flags in _data are authoritative; sparse maps are only read when
IS_COMBINED_MASK or HAS_EXTENDED is set. setCellFromCodepoint no longer
touches _combined. loadCell clears CellData side fields when flags are off.
copyCellsFrom copies map entries per-cell from src flags only.
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>
Graciliano Monteiro PassosandDaniel Imms
kkd927andDaniel Imms
dependabot[bot]andPretaSOFT
Bob LeeandDaniel Imms
Gevik BabakhaniandDaniel Imms
Ark0NandDaniel Imms
KindaBrazy
Anthony Kim
Anthony Kim
Enes Uysal
Paul Horn
jerch
