Two opens of one window could each find no current project and build a full project (index, language servers, watchers); the second replaced the first without closing it, leaving two jdtls on one workspace. open_project and open_light now hold a per-window lock from the check to the insert.
* perf(vfs): list directories off the main thread and reuse the repository lookup
Opening a folder on a network share (\\wsl.localhost\<distro>\...) is slow
because every listing ran on the main thread and walked up to the repository
root and ran a repository discovery again.
- fs_read_dir runs on the blocking pool.
- GitignoreStack::cached keeps the repository root and the repo-level rules
for a few seconds. An edit of .git/info/exclude is still seen at once (its
mtime and size are checked) and every .gitignore / .ignore is read fresh.
Refs #916
* feat(remote): open WSL projects through a WSL transport
A folder of a WSL distro (\\wsl.localhost\<distro>\... or \\wsl$\...) now opens
as a remote project whose openidea-server runs inside the distro, so the
terminal, SDK discovery, run, language servers and git are native Linux.
- RemoteTarget (ssh | wsl) replaces the SSH configuration id in the remote
launch, connection, session and recents; old recents are read as SSH and a
WSL target never inherits the legacy trust migration.
- remote_wsl: probe the distro, stream the server into ~/.openidea/server and
start it with `wsl.exe -d <distro> --exec <bin> --stdio`. Distro names are
validated and every command goes through --exec with positional arguments.
- wsl_distros / wsl_resolve commands; File > Open, the CLI and recents route
WSL folders to the distro and focus the window that already has them.
- Remote Development dialog: SSH/WSL choice with the detected distros; the
status bar shows `WSL: <distro>` / `SSH: <host>`.
- ide-jsonrpc: skip the NUL left at the start of a line by UTF-16 output of
wsl.exe, which made the first message of the server disappear.
Not run on Windows yet: only a stand-in for wsl.exe is tested.
Refs #925, #327, #721
* fix(remote): satisfy clippy in WSL output decoding
* style(run): format run types
* fix: regenerate bindings and drop untranslated WSL catalog entry
* test(i18n): list WSL as an untranslated product name
* fix(run): pick build wrappers and shell quoting by the execution host
A remote window runs commands on its Linux or macOS server, so a Windows
desktop no longer selects gradlew.bat, mvnw.cmd or Windows quoting for it.
* fix(welcome): open legacy WSL UNC recents and sessions through the distro
Recents and the restored session called the local openProject for a
\\wsl.localhost folder saved before WSL projects, which fails with OS
error 123. They now resolve it like the other entry points.
* fix(windows): keep WSL files out of remote project launches
path_launch tested the WSL share before is_dir, so a first-instance CLI
launch with a WSL file opened a remote project instead of LightEdit.
* fix(remote): run-command OS comes from the remote host, not an assumption
The server's OS was dropped after connect, so a restored session had none and
isWindows() assumed every remote host is Unix. Keep it on the connection and
the session, and resolve it through hostOs().
---------
Co-authored-by: NguyenBang <nguyenbang@NTBANG.local>
Co-authored-by: NguyenBang <nguyenbang@Mac.lan>
Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>
A terminal that starts but never shows a prompt left nothing in the log: the
spawn was logged at debug level only. The spawn (shell, pid, size), the first
output (after how long, and whether it is ConPTY's cursor position query) and
the first input (and whether it is the cursor position report) are now logged
at info level, once per session, so a user's log tells whether the shell never
wrote, or its query was never answered.
Refs #1224
The action was only enabled while the Pull Requests tool window was connected to
a repository, so from Find Action or a shortcut it did nothing in a project that
wasn't connected yet (gh not installed or not logged in, or several GitHub
remotes to choose from). It is now available in any open project: connected, it
forgets the choice and shows the repository selector as before; not connected,
it opens the Pull Requests tool window, which shows the selector or how to log in.
Fixes#1464
With many outgoing commits the Push dialog grew to the full screen height and its
whole body scrolled besides the changed-files tree, because the body was not a flex
container and the commit list's flex had no effect. The single- and multi-repository
Push dialogs now open at most 720px tall with the commit list filling the rest and
scrolling on its own; Dialog gets a maxHeight that applies until the user resizes.
Closes#1485
Co-authored-by: jaw52 <53564781+jaw52@users.noreply.github.com>
With API 1.1 the host provides ui.copy-to-clipboard, so that case only failed
on its signature. Use a name that never exists so it keeps covering a missing
function.
v0.1.36 shipped API 1.0.0, so its WIT is now the snapshot api_frozen checks
every later version against. World items are compared without their @version,
since a later minor renames them and wasmtime resolves the release's names to it.
Linked builds are saved normalised (canonical, forward slashes), but unlinking
compared the paths it was given verbatim, so a build outside the project could
not be unlinked on Windows (C:\… vs C:/…), nor through a symlinked path. Unlink
now normalises each path the way linking saves it. The Gradle linkage tests also
create the platform's wrapper (gradlew.bat on Windows) and compare canonical
paths without the \\?\ prefix.
In some YAML files over about a kilobyte, a mapping that follows comment lines
lost its highlighting from the second comment to the end of the mapping. The
YAML parser (@lezer/yaml with @lezer/lr) can give that mapping node an inverted
range when it falls on a tree buffer boundary (1024 entries by default), and
highlighting skips the whole node. The YAML mode now parses into buffers of
65536 entries, so ordinary YAML files stay in one buffer and never hit the
boundary.
Refs #809
Each selected row kept its own rounded corners, so a contiguous
multi-selection showed notches between rows. Rows whose neighbour is
also selected now drop the radius on that side.
The Changes view's Delete… for unversioned files had no key. When every selected
file is unversioned, Delete/Backspace now opens the same confirmation (files go to
the Trash) and the menu item shows the platform's delete key; Delete on a changelist
group still deletes the changelist.
Closes#1477
Co-authored-by: jaw52 <53564781+jaw52@users.noreply.github.com>
Show Diff (Cmd+D) in the Commit window or the Git window's Local Changes opened the
file still selected in the Project view, which keeps its selection while another view
has focus, instead of the change selected in the tree. Show Diff and Compare with
Branch or Revision now take the action's paths first, then the changes tree's
selection while it has focus, then the Project view's selection, then the active file.
Co-authored-by: DomKing <7285294+DomKing@users.noreply.github.com>
Sync with Deployed to… opened the comparison in its own native window on desktop,
but the default-server item, which the context menu shows by the server's name
(and as the only entry when one server is configured), still called the in-page
dialog, so the comparison was confined to the project window. Both actions now
share the same opener: a native window on desktop with a notification when it
can't open, the in-page dialog in the browser build.
Fixes#1459
A bottom tool window lists the connected Android devices (adb devices -l) and
streams the selected one's log (adb logcat -v threadtime, the last 200 lines first),
parsed into time, pid, level, tag and message in a virtualised list that follows the
end until scrolled up. The filter box takes Android Studio's logcat expressions
(tag:, message:, pid:, level:, -term, |, parentheses, regex) and remembers recent
ones. The new ide-android crate runs adb, which it finds through ANDROID_HOME /
ANDROID_SDK_ROOT, the usual SDK locations or PATH; the stream goes over a Channel
in batches (300 lines, or once the oldest waiting line is 60 ms old) and stops with
its request. The window is listed under the stripe's overflow until first opened.
Co-authored-by: payailk <80739895+payailk@users.noreply.github.com>
- One composer button shows the model and its effort; its popup has the searchable model list and a stepped effort slider over the levels the agent advertises.
- With an on/off permissions option (Allow All), the mode menu carries it as a switch, and new tasks start with it on like the other bypass settings.
- Recognise ACP session-mode URIs when picking mode icons.
- The Stop action in a running session is an icon button.
Dart projects that depend on the Flutter SDK were debugged through
`dart debug_adapter`, which cannot start a program importing dart:ui and has no
hot reload. A session now uses `flutter debug_adapter` when the program's or the
working directory's pubspec.yaml depends on the Flutter SDK, or when an attach
configuration asks for it (add-to-app: vm-service-uri and device map to the
adapter's attach arguments). Hot Reload and Hot Restart are sent over DAP and
appear in the toolbar of Dart sessions. Dart files also get a debug configuration.
Co-authored-by: payailk <80739895+payailk@users.noreply.github.com>
Debug session tabs had no close button, and ending a session only marked it
terminated, so every session stayed in the strip and Cmd+W fell through to hiding
the window. Each tab can now be closed: a running session asks before it is
terminated (the Run tab's question, shared through the workbench), then the
session, its children and its console are dropped and the tab on the left is
selected.
Co-authored-by: payailk <80739895+payailk@users.noreply.github.com>
check_component only compared import names and the API version, so a component
built from unreleased WIT (a function added under @1.0.0) or missing the plugin
world's exports passed `openidea-plugin check` and installing, then failed when
the IDE started it. It now links the component against the host's linker and
reports the missing or mismatched function with what to do.
Closes#1468
Right-clicking a line or the line numbers of a diff pane now offers Annotate with
Git Blame / Close Annotations. Each side is blamed on its own history: a revision on
that revision (at the path the file had there), the index on HEAD with its own text,
and the working file like the editor's annotations, refreshed after edits and when
HEAD moves. Compare Files, Local History and plain-text sides don't offer it. The
divider between the panes, which had no menu, acts on the pane focused last.
Fixes#1443
Configurations created for a main class, and the Spring Boot applications
detected in a module, set their working directory to the module directory.
A relative path in the VM options (such as --module-path=app/server/target/dependency)
then resolved against the module and the JVM failed to start. They now leave the
working directory empty, so the process starts in the project root, which is what
an empty working directory means elsewhere. The module is still stored separately
and keeps supplying the classpath and the debugger's sources.
Fixes#1197
Plugins get three host calls in API 1.1: vcs.remotes (under vcs.read) lists the
repository's remotes, ui.copy-to-clipboard and ui.open-url act on the window that
invoked the plugin's command.
Remote URLs are returned without the user:password part they may carry, and the
two UI calls fail when there is no invoking window instead of reaching every window.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
The Profiles node listed only the POMs' profiles; it now also lists those of the user settings.xml (the configured file, else ~/.m2/settings.xml), checked by default when <activeProfiles> names them or they are activeByDefault, and deactivated with -P!id when unchecked.
Fixes#1442
The shim for WebKit's reordered taps (#1072) took a press stamped at the same time as the
release before it for a stale one, and finished it with a synthetic pointerup, mouseup, click and
dblclick a tick later, even after the engine had delivered that press's own release. A fast
double-click whose second press shared the first release's time step was activated twice: on
macOS, double-clicking a Gradle task started it twice and the second start asked to stop the
first ("is not allowed to run in parallel"), which is why the linked Gradle E2E failed on main.
Only a press stamped before the release is stale now, and a press the engine releases before the
tick is left alone.
A project-tab group goes fullscreen through simple fullscreen, which is_fullscreen() doesn't report, so with the default "prefer tabs in full screen only" a project opened over a fullscreen group came up as a floating window. The new-project tab target now also counts simple fullscreen.
Fixes#1438
Welcome › Projects can put recent projects into named groups: Move to Group (other groups, New Group…, Remove from group) on a project, Rename / New / Remove Group on a group (removing keeps its projects). Groups are listed first, fold with click, Enter or ←/→, and searching lists matches without groups. Groups live in recent-groups.json beside recent.json; grouped projects are exempt from the 30-project cap and leave their group when removed from recent projects.
Fixes#1420
The shim for WebKit bug 219670 also ran in Chromium on macOS (the e2e browser), where a scripted double-click stamps its second press like the release before it; the shim took that for a reordered tap and dispatched an extra click and dblclick, activating a tree row twice and timing out e2e/linkedGradle.spec.ts on a parallel-run prompt. It is now installed only in WebKit.
Importing a PhpStorm .icls applied Java/Kotlin keys and PHP-only colours to every language, so PHP variables, strings and punctuation came out wrong. Java, Kotlin and PHP attributes are now kept per language (Color Scheme lists them under their own nodes, and editors scope them by data-language), explicit attribute inheritance is resolved, an explicit TextAttributes value resets omitted bold/italic, the light/dark parent follows the scheme's background, and PHP classes, method declarations, fields, brackets and separators get their own tags. Selecting a language attribute in Settings previews that language's colours.
Co-authored-by: marianomd <274989+marianomd@users.noreply.github.com>
Adds a search bar to query result grids (match case / words / regex, previous/next occurrence, Filter rows) and a pager for console results (first/previous/next/last, page size). Later pages re-run the read-only query with a skip, which every driver reads past without decoding, so only one page crosses IPC; Last counts rows with the statement wrapped so a trailing line comment can't swallow the closing parenthesis.
Closes#1431
Co-authored-by: kidp2h <40386174+kidp2h@users.noreply.github.com>
The selection is drawn in a layer behind the text, so any range with its own background (find
matches, warnings in light schemes, usages, semantic backgrounds) hid it: selected text there
kept its yellow or other highlight. Selected text is now also marked in the innermost span with
the selection colour, in text editors, notebook cells and the VCS viewers.
Agent conversations: an editor tab mounts only while it is active, so switching back to a
task rebuilt the whole conversation, parsed and sanitized every message again, and fetched
all entries again, which replaced them and rendered everything a second time. Loaded entries
are already kept current by the task's events, so they are no longer fetched again; rendered
Markdown is cached by text; and a conversation opened at its end renders its last 40 entries
first and the earlier ones in batches behind it (all at once when a scroll position or Find
needs them).
Git log: with branch-aware layout (#1076) a commit's lane was its branch's global layout
index, so a repository with many branches drew lanes dozens of columns apart, joined by
diagonals across the row. Each row boundary and node level is now packed side by side in
layout order; adjacent rows share their boundary, so lines stay joined. The column header row
is hidden by default and can be shown from View Options › Columns › Column Headers, where
columns are resized and reordered.
An interpreter that detection doesn't look for (a virtualenv in a subfolder such as backend/.venv, or one set by path) was missing from sdk_detect, so the status bar and interpreter menu showed its raw path. sdk_detect for Python now adds the project's interpreter when detection didn't find it, named by the same rules, and a project-local .venv in a subfolder is named after that folder.
Fixes#1428
Closing Settings while an ACP Registry agent was installing cancelled the install, and Install was offered again while a previous attempt could still hold the agent's folder. Installs now outlive the page as a background task of the project windows (stoppable from the status bar); the backend refuses a second install of the same agent and marks it installing in the list, which a reopened page follows until it ends.
Fixes#1430
Closing a window, closing or switching a project, or quitting while Run/Debug sessions are still running now asks before terminating them (Terminate / Cancel / Don't ask again, sharing run.confirmTerminate with the Run tab prompt). Run publishes its live sessions through a RunningProcessesProvider contribution, so the close guard collects them without importing the run feature; the prompt comes before the unsaved-changes step so cancelling never discards edits.
Closes#1433
Co-authored-by: ppxb <17081240+ppxb@users.noreply.github.com>
Spring route sources came back from the host with native separators, so on Windows every source failed the prefix check against the forward-slash project root and Search Everywhere showed no endpoints. RouteSource.path is now an IpcPath and the index compares by path segment; the add-to-ignore result also returns IPC paths.
Fixes#1422
IMEs that commit without composition events (Doubao on macOS) inject their text asynchronously while the keyCode-229 keydown that started the commit is still pending. xterm ignores an insertText event while a keydown is pending, and its deferred textarea diff misses text that lands after its snapshot, so a fast run of digits lost characters (1234567890 arrived as 138). On macOS the terminal now delivers those commits itself (only commits during a pending IME keydown, and composition input without compositionstart); real compositions, keys xterm already sent and input without an IME keydown stay xterm's. Linux and Windows keep the fcitx5 ImeCommitGate. Keystrokes are written to the PTY as they come instead of per animation frame, and those typed while the shell is still starting are sent once it exists.
Co-authored-by: DomKing <7285294+DomKing@users.noreply.github.com>
The overview lane sat over the scrollbar with pointer-events: none and is 12px wide, the native scrollbar 10px: a press on the lane's edge, or where marks covered the thumb, reached the text layer and dragged a text selection instead of scrolling. The lane now owns the vertical scrollbar, as IntelliJ's error stripe does: the native one is hidden, the lane's thumb drags (with the same geometry it is drawn with), and a press above or below the thumb pages the editor. Marks still jump to their line.
Co-authored-by: IsGuibo <40512175+IsGuibo@users.noreply.github.com>
feat(maven): add POMs outside the root reactor as Maven projects
Add Maven Projects only re-linked POMs Maven had already discovered: picking a pom.xml inside the
project that the root POM doesn't list as a module (an independent build kept out of the reactor)
closed the dialog and changed nothing. Such POMs are now kept in the project's Maven settings
(linkedProjects) and loaded as independent top-level projects, after the root reactor, by the
static model, the sync, the Java language server import and the content roots; each runs its goals
in its own directory. A picked POM that the root reactor lists stays its module. The dialog takes
several POMs at once, and unlinking a linked project drops the link instead of hiding it.
Opening a folder of independent Gradle builds selected only the first build. Link Gradle Projects… (pick a build's settings or build file) and Unlink on a build's root let an ordinary project window manage several independent builds without changing their scripts. The build model holds a list of Gradle builds; each keeps its own wrapper, working directory, environment, task routing and sync cache, keyed by the owning build directory. The links persist in the project's Gradle settings (linkedProjects); unlinking leaves files and saved run configurations alone. Existing includeBuild support is unchanged, and an explicitly linked build takes precedence over a composite that also includes it. Workspace windows keep Link/Unlink disabled.
Co-authored-by: gudcks0305 <84506708+gudcks0305@users.noreply.github.com>
fix(vfs): report files deleted on macOS again
Since the watcher queues raw notify events (#1412), deleting a file on macOS was never reported, so
its symbols stayed in the index and the tree kept it (the project smoke tests failed for every
language). FSEvents delivers a deleted file's coalesced history flags as create, remove, then
modify events. Create + remove merged into a transient path, the following modify turned it back
into a creation, and a creation of a path that no longer exists is dropped. A modify no longer
revives a transient path; only a create does, and the disk decides as before.
* fix(tools): find Python on Windows without python3.exe
Installing debugpy on Windows failed with "cannot run python3.exe … (os error 3)": the tool
manager looked for `python3` first, and on Windows that only finds the App Execution Alias in
WindowsApps, which can't start without the Store's Python. Windows installs `python.exe` and the
`py` launcher instead. On Windows the interpreter is now looked up as `python`, `py`, then
`python3`, and a WindowsApps alias only counts when it runs. The debugpy adapter's default
interpreter is `python` on Windows too. macOS and Linux keep `python3` first.
* fix(tools): keep searching past broken Python aliases
Spanish (es, with es-MX / es_MX system locales) joins the UI languages under Settings › Appearance & Behavior › System Settings › Language and Region, with the key catalog (locales/es.json) and the source-text catalog (locales/text/es.json, 6180 entries).
Co-authored-by: FizzyKey8 <81670057+FizzyKey8@users.noreply.github.com>
Themes imported from IntelliJ drew a lighter gutter strip and kept the built-in blue for the selected tab's border and completion matches. The converter read the classic UI's GUTTER_BACKGROUND, where the new UI paints EDITOR_GUTTER_BACKGROUND or else the editor's background; it read EditorTabs.underlineColor, where Islands borders the selected tab with underlinedBorderColor; and it didn't map CompletionPopup.matchForeground. Scheme attributes written as `<value />` (no colours, not inherited) now show as plain text instead of keeping the built-in colours.
Co-authored-by: ldg030201 <100822592+ldg030201@users.noreply.github.com>
On macOS, while an editable element has focus and a third-party input
method (reproduced with WeChat's WeType) is active, WebKit can deliver a
tap as pointerup, mouseup, pointerdown, mousedown (WebKit bug 219670).
The tap's click is lost, and because the engine still considers the
button down, the next tap starts with a bare mousedown and no
pointerdown. Taps on tree rows, tabs and tool window buttons needed a
second tap.
A press stamped no later than the pointerup just before it is a
reordered tap: the shim finishes it with a synthetic pointerup, mouseup
and click (and dblclick for a double) after the page's handlers ran, and
dispatches the skipped pointerdown when a primary mouse press arrives
with no pointer down. The next tap's own click is kept. Installed on
macOS only; untrusted events and other pointer types are ignored.
Verified in a WKWebView with WeType, alternating taps on two targets:
main 19/20 taps clicked their target, with this change 20/20, every press
ran the pointerdown handlers and setPointerCapture worked; with the ABC
layout the shim dispatches nothing.
Refs #760, #724
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
On macOS, with tap to click and a third-party input method active, WebKit
can deliver a tap's mousedown after the button is already up (WebKit bug
219670). Tauri's drag-region script starts the native drag on any
mousedown, so AppKit ran its drag loop with no release left to end it and
the window followed the pointer until the next click.
On macOS the title bar now uses its own drag region: a drag starts only
when the mousedown has the primary button down, and the new
window_start_drag command checks NSEvent.pressedMouseButtons before
start_dragging. Double-click zoom stays on mouseup and is cancelled by
movement. The handlers are installed once per page. Windows and Linux
keep data-tauri-drag-region.
Refs #760
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
A Java file outside a Maven / Gradle build ran as `java File.java` (JEP 330), which launches the file's first top-level type on JDK 21 and earlier. A file whose main is in a later class failed with "can't find main(String[]) method in class: Parent", and the gutter Run of a second or nested class's main launched the first class. Such files now compile with the JDK's own `javac -g -d <class dir>` and run `java -cp <class dir> <main class>`; a main in the first class, and compact source files, still use the source launcher. Debug passes the JDWP agent through JDK_JAVA_OPTIONS, which only the java launcher reads.
Co-authored-by: ldg030201 <100822592+ldg030201@users.noreply.github.com>
#1176 brought in quick-xml 0.41 next to ide-host's 0.42, so ide-host's dependency in Cargo.lock must name its version; without it cargo rewrites the lock on every build and --locked builds fail.
Window destruction and the close_project command detach the project at once and let its language servers, kernels and debug sessions shut down in the background; opening a project waits for pending shutdowns and exit waits for them with a bound.
Closes#1414
On a fresh install, Settings › Plugins › Available showed no marketplace plugins until the user refreshed by hand: opening the page asked for the listing without a refresh, which only returned the cached index, and a first run has none. The listing now reads the index when none has been read yet; later opens keep using the index already read, and an explicit refresh reads it again.
l0ng-ai
Jaw
DomKingandDomKing
payailk
Deng Liu
Mariano Dupontandmarianomd
Nguyen Thinh
ppxbandppxb
黄贵波
yoo hyeong chan
FizzyKey8
Lee Dong Geon