Skip to content

Adopt 17 open upstream eclipse.platform fixes and enhancements - #40

Merged
hsoliwal merged 36 commits into
masterfrom
m3/upstream-prs-20261116
Nov 16, 2026
Merged

hsoliwal merged 36 commits into
masterfrom
m3/upstream-prs-20261116

Conversation

@hsoliwal

Copy link
Copy Markdown
Owner

Brings 17 open upstream eclipse-platform/eclipse.platform PRs into the fork. Each one is ready for review upstream and green on upstream CI.

Conflicts with fork work were resolved by reading both sides. Java files were merged member by member, keyed by signature. Where the fork had only re-formatted code, upstream's change wins; where the fork had added behaviour, both are kept. Bundle-Version bumps take the higher version. Upstream's Java 22 unnamed variables were renamed to the fork's Java 21 form (unused).

Verification on 45b74ad:

  • Tree loss guard: no fork path dropped, and every upstream PR path is present. Any differences from upstream PR content are the Java 21 rename only.
  • All 77 changed Java files parse at Java 21.

Not in this PR (each has a conflict that needs its own reading): 2203 2934 3059

iloveeclipse and others added 30 commits September 2, 2026 13:59
Broken file links were not present in resource model before, now they
are (at least with Linux x86_64 LinuxFileHandler or Java NIO
PosixHandler, in the future hopefully on all Unix
platforms).

For "broken" link overlays we need an additional patch in Platform UI.

Fixes eclipse-platform#2797
Preserve breakpoint check state while doing filtering in Export/Import
Breakpoints
Introduce IWatchExpressionCellEditorFactory, a new adaptable API that
allows debug targets to plug in language-aware editing for watch
expressions creations.
Ctrl+click (Cmd+click on macOS) on any Run As / Debug As shortcut entry
now opens the Run/Debug Configurations dialog.
On macOS the tree selects on mouse down, and opening the modal "unsaved
changes" prompt at that point corrupts the tree. Skip selection changes
while the mouse is pressed and handle them on release.

Fixes eclipse-jdt/eclipse.jdt.debug#909
The "still initializing" timer and the initialization job could both
create the contents of the same compare editor input. One of the two
content trees stayed alive when the editor switched to another input,
and its viewer later diffed against the disposed input.

Add a regression test and a DisplayUtils helper for driving the event
loop in tests.

Fixes eclipse-platform#123
…aracter

getContents() decoded the whole accumulated byte buffer with the stateful
StreamDecoder every time the read thread had appended data, so bytes the
decoder had already consumed were fed to it again. When a multibyte
character was split across two reads and getContents() was called after
each read without flushContents() in between, the retained lead byte was
followed by the first byte of the buffer again and the decoder emitted
U+FFFD.

Decode only the bytes not yet fed to the decoder and accumulate the
result; flushContents() and setBuffered() reset the accumulated text and
the fed-byte count as before. Adds a regression test to
OutputStreamMonitorTests and bumps org.eclipse.debug.tests.

The defect was found by an automated LLM-based code reviewer in an
academic study at Jeonbuk National University; the fix and the test were
written and verified by the authors with the assistance of Claude.

Fixes eclipse-platform#3058

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A combining mark has no cell of its own. It was dropped, which lost
data: an accent typed as a separate mark, a Hangul syllable sent as
conjoining jamo, a Japanese voicing mark. More than marks go with the
character before them: a zero width joiner and what it joins, a skin
tone, a presentation selector, the second regional indicator of a flag.
What a reader takes for one character is a grapheme cluster, and where
one ends is now decided by the Unicode rules (UAX #29) through
java.text.BreakIterator, over the text as it is written, the cluster
written last going first, so a cluster goes on across writes but not
across anything else done in between: a cursor move, an erase, an edit
or a scroll. A cluster takes the cells its code points would,
at least one and at most two, summed as Windows Terminal sums them: a
mark takes none, VS16 makes it wide. Grown wide on the last cell it
moves whole to the next line, as a wide character does; in insert mode
its second cell is made room for. A cluster that starts with a code
point taking no cell goes onto the one before it, or takes a cell of
its own when there is none. Left out, before the clusters are found,
are control characters, the format characters no cluster needs (a zero
width space, bidirectional marks, overrides and isolates), and what
only hides text: tag characters outside the tag sequence of a flag
(UTS eclipse-platform#51), a variation selector that does not follow a character, a
second prefix sign. None is drawn, and in copied text they could make
a command read differently from how it runs or carry text no one saw. A cluster keeps at most 32 chars (the longest emoji sequence
is 15), so a flood of marks cannot grow one cell without bound.

How many cells each character takes was worked out again, from the
chars around it, wherever it was needed: in the emulator, the
selection, copying, the word under the pointer, the canvas and the
renderer, each with its own conditions. The text data now keeps it.
TerminalTextDataStore records, per line, the cells that are the second
of a character (written with it in one call: a wide character and its
filler, a surrogate pair, or a cluster made wide) and the clusters
beside them; lines of one-cell characters keep nothing extra. Reading
it is a lookup (getCellWidth, getCluster, isSimpleLine, new default
methods of ITerminalTextDataReadOnly, hence the version bump), and
everything that reads cells goes through it and through Cells, which
puts characters back together; the old inference in CharWidth is gone.
The emulator writes every character through one method, moves
characters whole when cells are inserted or deleted, and keeps one
record of the cluster written last.

The cells hold what is drawn, the character composed with its marks
(NFC) where it has such a form; the cluster is kept as the program
wrote it, and it is what is drawn over its cells (a flag, a family, a
keycap as one) and what selecting and copying give.

A narrow character beyond the BMP (a mathematical letter, U+1F321, a
regional indicator alone) takes one cell, as East Asian Width and
Windows Terminal have it; it took two because its two chars could not
share a cell, and is now kept beside the line as a cluster is.

Text of one-cell characters is drawn as before, from the line's chars.
Runs of wide characters are written in one go, the boundary rules are
consulted only for characters they could join (found once per
character from BreakIterator itself), and character widths of the BMP
are looked up. Writing CJK text is about 1.8 times as fast as before,
and drawing it faster too; other text is drawn at least as fast and
written up to about 12 percent slower, the cost of finding clusters.

Tests: CellsTest, and a randomized test that mixes writes (split at
any point), cursor moves, erases, insertions, deletions, scrolling,
resizing and the alternate screen, checking after every step that the
cells, the selection, the copy and the drawn text agree, and that
written text is copied back unchanged.
ProcessConsole took the file supplied to a process's standard input only
from the launch configuration (Common tab > Input File). Nothing outside
the configuration could change it, so a plugin that wants every launch to
read the same input had to edit configurations or write to the process
itself, and then mixed its input with a configuration's own file.

ProcessConsole now first looks for ATTR_CAPTURE_STDIN_FILE among the
attributes of the process's launch and uses the configuration's only if
the launch has none. A listener can set it when the launch is added,
before any process exists. A launch without a configuration can supply
input this way too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The compiler reported a potential null access of the configuration. Read
ATTR_APPEND_TO_FILE where the configuration is not null.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rm does

Keys that Claude Code, Codex and other terminal programs use, checked
key by key against xterm:

- Shift+Tab sends back tab (CSI Z), as xterm and Windows Terminal do,
  and no longer moves the focus out of the terminal; Claude Code
  switches modes with it.
- Ctrl, Shift and Alt with Delete carry the modifier (CSI 3 ; 5 ~ and
  so on); every Delete used to arrive as a plain Delete.
- A real Ctrl+Alt keeps its Alt: ESC before the control character, as
  in xterm (Ctrl+Alt+V pastes an image in Codex on WSL). AltGr, which
  SWT reports as Ctrl+Alt with the character it types, types that
  character as before.
- Ctrl+2 to Ctrl+8 send NUL, ESC, FS, GS, RS, US and DEL; on Windows
  they sent the digits.
- On the alternate screen Alt and Ctrl with the arrows go to the full
  screen program even where Eclipse has a binding for them (Alt+Up
  edits a queued message in Codex); at the shell prompt Alt+Up still
  maximizes the view.
A program that redraws its screen in many writes can bracket each
screen with CSI ? 2026 h and CSI ? 2026 l. The view polls the text
every 50 ms and used to paint whatever it found, often a screen half
drawn, which shows as flicker and as a cursor jumping over the screen.

While a program is between the two, the canvas model does not look at
the text, and the next poll after the end paints the finished screen.
The check and the copy of the text are one locked step, so a screen
cannot begin between them, and a scroll or a resize waits like a poll:
the view moves with the next copy, not over the lines held back.
A program that never ends a screen, or begins the next one right after,
does not freeze the view: it is held back at most 200 ms
since it last looked, on a clock that does not move with the time of
day. A full reset or a disconnect ends the mode, and
the last poll then shows what the program had drawn. As with DECTCEM,
the flag is kept on the control and copied to the new canvas model
under the same lock when the tab is moved to another view.

Split out of eclipse-platform#2902 as asked in its review.
The UI callback of the job acted on whatever input the editor had by then, so the canceled job of a replaced input closed a reused compare editor while the new input was still initializing.

Assisted-by: multiple AI agents and layers of automated tooling 🤖
Allow callers to clear unified diff annotations, code minings, toolbars and listeners without changing the editor document or supplying a baseline. Restore navigation actions and diff folding, and prevent pending computations from reinstalling the display.
Reserve and paint glyph-aware code-mining heights so emoji fallback fonts do not overlap source lines. Keep click-through text aligned and fill source-row backgrounds across increased line spacing.
This commit adds a "Preview" button on the String Substitution pref
page. It is enabled for a single selection and opens a small
dialog showing the variable's name, value, description and contributing
plug-in, each with a bold title. This will be useful for readability
when there is long values
Programs that draw their own screen - htop, ncurses menus, and lately
CLIs with clickable options and wheel scrolling - ask the terminal for
mouse events with DEC modes 1000 (presses and releases), 1002 (also
drags) and 1003 (every move), in xterm's legacy encoding or the SGR one
(mode 1006). The terminal ignored the request and kept every click for
its own selection, so such programs could not be clicked on or scrolled.

The canvas now hands presses, releases, moves and the wheel to the
control while a program has asked for them, with the alt and control
keys folded into the button as xterm does. A click the program takes is
the program's alone, as in xterm and GNOME Terminal: the middle-click
does not also paste, the right-click opens no context menu (the menu
key still does), and neither a double-click nor the copy on release
acts on it. The canvas reports it in a listener that runs before every
other and stops the event there. Shift keeps the mouse, the wheel
included, for the terminal, which is the convention every terminal uses
to still allow a selection while a program is listening; a button held
down drags for the program only if its press went there. A wheel notch
is three of the lines SWT counts, however its events split them, the
rest counting toward the next notch in the same direction. Positions are
the program's screen, counted from its first line, not from the top of
the view: with history above the screen the view rarely starts on a
whole line, and a click over the history the user scrolled back to is
the terminal's, not the program's. The legacy encoding reports nothing
past column or row 223, as xterm, rather than bytes it has no room for,
and a release in it keeps the alt and control bits of its press. A
report is sent as bytes, not through the terminal's text encoding, so a
column or row past 95 stays the one byte the legacy encoding means. A
press that went to the program takes the terminal's selection away and
becomes the selection anchor, so a later shift-drag starts from where
the user last clicked rather than from a selection made long before. A release
goes wherever its press went, whatever the keyboard state by then, so
press and release always reach the same side.

This builds on the DEC private mode parsing change (every parameter in
a set/reset), since ncurses turns the mouse on with CSI ? 1006 ; 1000 h.

A reset, which runs before every connection, turns mouse reporting and
SGR mouse encoding off, and nothing is reported while the terminal is
not connected: a program cut off while it listened left the reports on
for the next connection, and a click in between threw from writing to
an output stream that no longer existed. The modes are written by the
reader job and read on the display thread, so they are volatile, as is
application cursor keys, which had the same gap.

MouseReport.mouseReport is public, as the test bundle calls it across
the bundle boundary.
Resolved by recipe: 4 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 1 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 0 file(s); provenance index and policy recomputed.
Resolved by recipe: 1 file(s); provenance index and policy recomputed.
Resolved by recipe: 1 file(s); provenance index and policy recomputed.
@hsoliwal
hsoliwal merged commit 8beaba7 into master Nov 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Stack trace detection do not work in colored console