process.run kept at most 4 MiB of stdout and silently dropped the rest while reporting success, so large structured output (kubectl -o json, cargo metadata) reached plugins cut off mid-document. The limit is now 16 MiB, as for file reads, and stdout beyond it fails the call with an explicit error instead of returning truncated output. stderr is still truncated without failing.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
Root markers skipped the path itself, so a module folder whose own folder holds the build file (what Run/Debug's classpath and launch queries pass) resolved to the enclosing root, and with per-module jdtls a query started a second instance; a directory now counts as its own ancestor. A reconfigure that switches the jdtls workspace layout restarted the instance it had just created, because its settings were compared before the start recorded them; servers created in the same pass are skipped.
knip now ignores the generated Django parser terms like the other template parsers'. The progress-replacement test started its barrier task while the released search could still be measuring indexing, so it sometimes saw Indexing; the barrier now starts after the search. The local download test passes '/'-separated remote paths, as the local file system lists them on Windows.
git has no line counts for binary files, so a task's changes list showed +0 −0 for each binary the task produced. The list now shows only non-zero counts, as the run summaries in the same view do.
When a maximized tool window unregistered (its plugin was disabled, uninstalled or crashed), the layout kept its side maximized with nothing left to show there, which hid the editor and every other tool window and left the workbench blank; Hide Active Tool Window and Maximize did nothing. Unregistration now clears a maximized side that has no registered visible window, along with a stale focused or last-active window (focus returns to the editor), and visibility, context keys and Hide Active Tool Window ignore such a stale maximized side.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
process.run, process.run-command and net.fetch were given the call's whole remaining wall-clock budget, so a subprocess or request that timed out returned to Wasm already past the deadline and the next epoch check trapped the instance. Three slow external calls then counted as three crashes and disabled the plugin. Blocking host I/O now gets the budget minus a 250 ms unwind window, and a failed I/O call near the deadline gets one 250 ms extension per call so the guest can handle the error. Pure Wasm loops still trap at the deadline.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
The File tab followed only the active text editor, so with a notebook active it said no file was open; notebook diagnostics are also stored per cell document. It now follows the active tab's path for other editor kinds and collects the notebook's cell documents (percent-encoded and Windows drive paths included).
jdtls builds launches with ECJ, which rejects some code javac accepts (for example a generic method call used as an if condition), so such a Maven project could not be run from the IDE. Settings › Build, Execution, Deployment › Maven › Delegate IDE build/run actions to Maven makes Automatic and Build (Make) launches of Maven modules compile with Maven, as Rebuild already did. Classes JDT wrote for sources with errors look up to date to Maven's incremental compiler, so a delegated launch JDT reports errors for recompiles everything and javac's result decides whether it starts. Editor diagnostics still come from jdtls. Off by default.
Co-authored-by: GGVidal <52431033+GGVidal@users.noreply.github.com>
Notebook cells had no language server support. Servers that advertise notebookDocumentSync (basedpyright, Ruff) now receive the notebook through LSP 3.17 notebookDocument/didOpen, didChange and didClose, with one vscode-notebook-cell document per cell, so names defined in one cell resolve in the others. Cell editors get completion, hover, diagnostics and go to declaration; a location in another cell scrolls to and selects it in that cell, and cell diagnostics show under the notebook in Problems.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
Edits to .openidea/data-sources.toml and the global data-sources.toml made outside the IDE (git pull, branch switch, a text editor) were only read when the project opened. A notify watcher on the project root, .openidea and the global data directory now reloads both files after a short settle delay, closes connections whose source changed or was removed, and emits DbSourcesChanged so the Database tool window refreshes. Saves from the IDE update the on-disk snapshot so they don't echo back as external changes.
Co-authored-by: brunobelloni <14021404+brunobelloni@users.noreply.github.com>
jdtls' bundled ECJ rejects a generic method call used directly as an if/while/for/do or ternary condition (Type mismatch: cannot convert from Object to boolean) where javac infers boolean; no ECJ release accepts it. Document the case and the workaround (assign to a boolean first) in the Java limitations, en and zh. Refs #2624.
Co-authored-by: Inf1mum <8135225+Inf1mum@users.noreply.github.com>
The test waited 2 s for its Python stub agent to start and to exit. Under a
parallel test run the interpreter's start-up alone can take longer, so the test
failed now and then. Both are polls for a condition, so a longer bound costs
nothing when it passes.
The splash covers the page and swallows raw mouse presses until it has faded
out. Only one spec waited for it, so the tool window resize test sometimes
pressed on the splash instead of the gap (3 of 15 runs locally). The workbench
fixture now waits for it to be removed.
The full-screen view kept a copy of the diagram as it was when opened, so a theme
change (or an edit) redrew the inline diagram but left the open view in the old
colours. The view now replaces its copy in place whenever its block is redrawn,
keeping the current pan and zoom; Download saves the newest drawing.
The generated compiler-defaults.prefs was passed to jdtls as a bare path in
java.settings.url. jdtls parses that setting as a URI whenever a watched file
changes, so a data directory with a space (macOS 'Application Support') logged a
URISyntaxException on every file event. Pass a file: URI instead; reading the
prefs back accepts both forms.
When HEAD moved after a local task started, a file was treated as the user's doing only if its bytes on disk equalled the start commit's. Under core.autocrlf the working-tree bytes (CRLF) never match the committed blob (LF), so every such file was kept in the task's changes and rollback could revert the user's own reset or checkout.
Tracked files now count as changed only when git's own worktree diff against the start commit lists them (path or rename source), so git's line-ending and other filters apply. Files left untracked by a mixed reset are byte-compared with CRLF on disk matching LF committed. Tests run with core.autocrlf off and on, plus a mixed-reset case.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
When several actions share a shortcut, dispatch took the first enabled one in preset order, so an action with no keyWhen (Edit Commit Message, enabled wherever VCS is) could take Eclipse's Alt+Shift+R from Rename File in the project view. At the same sequence, a binding whose keyWhen matches the current context now runs before one without a matching keyWhen; otherwise the preset/override order decides as before.
Fixes#2815
Co-authored-by: Inf1mum <8135225+Inf1mum@users.noreply.github.com>
Selecting a test in the results tree showed only its failure details, and a
suite or file showed nothing. Vitest's `stdout | file > suite > test` and
`stderr | …` blocks now become the output of that test (file-level logs go to
the file), and selecting a suite or file shows its own output followed by its
tests' output, as in IntelliJ IDEA.
Also in the Vitest parser: the default reporter's `❯ Suite (n)` lines now open
that suite, so a failed test is no longer listed twice (once under its suite
from the FAIL header and once directly under the file), and a failure's details
end at Vitest's `⎯⎯[1/1]⎯` separator instead of running into the summary.
Fixes#2803
The Run and Debug consoles stayed empty: their scroll container also carried
the ANSI palette class, which became `display: contents` when the palette moved
to the color scheme, so the container lost its box and its virtualized lines
never rendered. The marker class served no purpose any more and is removed.
A long commit subject pushed the author and time out of the Update Info and
Compare with Branch commit lists, since the row could not shrink below its
content. The subject now ends in an ellipsis with the full message in a tooltip,
hash, author and time keep their width, and both dialogs can be resized with the
list filling the window.
Fixes#2642
The project view handled Shift+F6, F6, F4, Delete, copy/cut/paste, Copy Path and New in its own onKeyDown, so unbinding or rebinding those actions in the keymap had no effect there (a cleared Rename File still renamed on Shift+F6, a rebound one was ignored, and Eclipse's Shift+Alt+R opened the file rename). Those keys now go only through the global dispatcher and the preset bindings; the view keeps just the strokes no preset binds (Cmd+Z undoing the last file refactoring, Ctrl+Enter's New popup on macOS). The Eclipse presets bind Delete Project Item explicitly, since the hardcoded Delete used to cover for them.
Fixes#2801
Co-authored-by: Inf1mum <8135225+Inf1mum@users.noreply.github.com>
Python discovery looked for Poetry virtualenvs only in macOS's cache folder and for uv Pythons only in ~/.local/share/uv/python/*/bin/python3. It now follows each tool's own locations and overrides: Poetry's POETRY_VIRTUALENVS_PATH / POETRY_CACHE_DIR or the platform cache dir, and uv's UV_PYTHON_INSTALL_DIR or its data dir (%APPDATA%\uv on Windows, XDG data dir elsewhere, the legacy macOS folder), probing both the Windows (python.exe at the install root) and Unix (bin/python3) layouts.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
mermaid 12.1.0 still asks for katex ^0.16, which GHSA-238p-pmpm-9mq7 affects (fixed in 0.18.2). A pnpm override resolves mermaid's katex to 0.18.10; mermaid only uses renderToString, which is unchanged. An e2e test checks that KaTeX math in mermaid labels still renders. The override can go once mermaid is 12.3+.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
Deployment servers get two more types next to SFTP and local folders: FTP and
FTPS (explicit AUTH TLS, certificates checked against the system trust store).
The Connection tab takes host, port, user name (empty logs in anonymously) and
passive mode; the password is asked for on connect like an SSH password and is
kept for the session or saved in the OS credential store.
The FTP filesystem backs upload, download, sync, automatic upload, the remote
host browser and remote editing. Listings use MLSD when the server has it, else
LIST with MDTM for exact times, so sync compares timestamps reliably; uploads
keep timestamps with MFMT when "Preserve files timestamps" is on.
Closes#1200
Pin / Unpin Tab only acted on editor tabs and was disabled while a tool window
had focus, so a shortcut couldn't pin a Find or Usages tab. With a tool window
focused it now toggles the pin of that window's selected content tab, through
the same content tab registry Close Tool Window Tab uses; Find and Usages tabs
support it.
Closes#2786
A tool window's Options menu gets a View Mode submenu with Dock Pinned (the default), Dock Unpinned and Undock, as actions `toolwindow.viewMode.dockPinned|dockUnpinned|undock` (category Window). Unpinned and undocked windows hide when focus moves to the editor or another tool window, not to a popup, menu or dialog; an undocked window lies over the editor at its side without resizing it, and its inner edge resizes it. The mode is kept per window in the project layout, saved with Tool Window Layouts and reset by Restore Default Layout.
A menu button in a tool window now takes focus when clicked: WebKit doesn't focus clicked buttons, so the menu returned focus to the editor on close and hid a window just switched to Dock Unpinned or Undock.
Co-authored-by: formiat <29867335+formiat@users.noreply.github.com>
kmp-lsp (tree-sitter based, no JVM or Gradle import, low memory) was only an
alternative for Java. It is now also the last fallback for Kotlin, after
JetBrains' kotlin-lsp and fwcd's kotlin-language-server, so Settings >
Language Servers > Language server for Kotlin can pick it; the managed
install and its jar indexer sidecar are unchanged.
A server that is an alternative for several languages showed only the first
language's picker on its own page; it now shows one picker per language.
Closes#1017
When tsserver hits its heap limit, Node aborts it (exit code 134 on Windows,
SIGABRT elsewhere) and vtsls restarts it, so code intelligence silently pauses
or misses results (references, go to definition) and the cause only shows in
the log. vtsls's "TSServer exited" log line for such an exit now raises an
error notification, once per session, naming typescript.tsserver.maxTsServerMemory
and with a button that opens Settings > Language Servers.
Closes#1685
JSON.parse keeps only the last of two equal keys, so a catalog could carry
duplicate entries (one of them silently unused) without any test noticing.
The text catalog test now parses every message and source-text catalog with
the TypeScript JSON parser and lists keys defined more than once.
Double-clicking a file no editor understands (`.xlsx`, `.pdf`, ...) opened the read-only hex view, because the hex provider claims every binary file. Settings > Editor > File Types gains "Files opened in associated applications" (Office documents and PDFs by default); matching files open with their system application through a new `fs_open_with_system` command, which requires a trusted local project window and a file inside the project or one of its content roots. Remote windows and failed launches fall back to the existing viewers. External open handlers are a new workbench extension point asked before the file editor providers.
Co-authored-by: panda-student <71381346+panda-student@users.noreply.github.com>
Covers Download to… in the Remote Host view: the Project view selection preselects the folder picker (its parent folder for a selected file), cancelling the picker downloads nothing, and the chosen folder receives the remote paths.
Co-authored-by: GGVidal <52431033+GGVidal@users.noreply.github.com>
On Windows `uri_to_path` rejected any `file:` URI host other than `localhost`, so network-share results (`file://server/share/a.ts`) from language servers were dropped from diagnostics and renames and opened as read-only virtual documents; such hosts now map to `\\server\share\...`, and `path_to_uri` writes UNC paths in the same RFC 8089 form. Lower-case or escaped drive letters (`file:///c%3A/x`) are upper-cased in `uri_to_path` and `to_ipc`, so the UI no longer treats `c:/x` and `C:/x` as different files. Non-Windows behaviour is unchanged.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
The atomic save renames a temp file over the target, and Windows refuses that while another process (antivirus or indexer scan, sync client, a program reading its config) has the file open without delete sharing, so the save failed with "Access is denied" or a sharing violation. The rename is now retried for up to about a second on Windows errors 5, 32 and 33; other errors and other platforms fail at once.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
`WorkspaceEdit::deduplicate` dropped any edit already present for the file, including repeats inside one `TextDocumentEdit`, so a server edit inserting the same text twice at one position (two closing brackets, two blank lines) lost one insert. It now only drops edits an earlier alias entry for the same file already contributed and keeps every edit of one entry; the result stays idempotent.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
Reloads after an external change (`reload_from_disk`) and after undoing a workspace edit (`reload_resource_buffers`) always re-ran encoding detection, so a file the user had reloaded or converted to a legacy encoding came back as mojibake and later saves used the guessed encoding. Documents now record that the encoding was chosen and both reload paths decode with it; files whose encoding was never chosen are still detected again.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
Broadcasts called their listeners in a bare loop, so one throwing handler kept the rest (including the catch-all "*" IDE event handlers) from seeing the event. The IDE event bus, local diagnostics, endpoint-change and navigation-modifier broadcasts now go through `notifyEach`, which logs a throwing listener and moves on. `onIdeEvent` wraps each subscription so unsubscribing one of two identical handlers keeps the other.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
Tauri runs commands declared without `async` on the main thread, so Local History, shelf, change-list and run-configuration reads and writes, `.env` loading and the line diff of two arbitrary texts blocked input, painting and every other IPC call while they ran. They are now async and run on the blocking pool; IPC signatures are unchanged. `vcs_changelists_save` stays synchronous so its fire-and-forget saves land in order. A guard test keeps these commands async.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
gitoxide's status stops at the first IO error, so one unreadable directory failed the whole status and the Commit window showed no changes, where git warns and lists the rest. Status now asks git (porcelain v2) when gitoxide fails; a test checks the parser against gitoxide on a full status matrix.
Fixes#2769
Double-click, Enter or Show Diff on a file in the Git log's changes opened a new editor tab each time. It now opens the editor's preview tab, as the Commit window does, and while that tab shows one of the pane's diffs, selecting another file shows it there. Pinned tabs are left alone.
Fixes#2768
Delve in DAP mode builds the debuggee as __debug_bin<digits> on every platform; the post-session sweep only accepted that form on Windows, so the binary stayed in the project on macOS and Linux. Closing the last session tab also left an empty Debug (or Run) tool window open; both now hide with their last tab.
Fixes#2766
Next/Previous Difference also dispatched a scrollIntoView effect, which set scrollTop in the same measure as the smooth scroll and cut the animation off. The smooth scroll already centres the caret, so the effect is removed.
Fixes#2764
Right after typing in a Java file, Show Usages answered for the wrong symbol and Go to Declaration could do nothing. Usages, Find Usages and hierarchy now wait for the editor's edits to sync before sending the caret position, and the java-usages commands (usages, implementations, hierarchy, configuration) wait for jdtls to reconcile queued document changes before resolving a position (bundle 1.13.0, rebuilt reproducibly from source). File-system events (e.g. the build writing target/classes) now only drop cached navigation answers instead of discarding an in-flight Go to Declaration.
* feat(hotkeys): warn when the OS grabbed an IDE shortcut (#1876)
Probe single-stroke Alt bindings at startup with RegisterHotKey and
notify once per session when another app owns one, with an action that
opens Keymap for reassignment.
Co-Authored-By: Claude Code <noreply@anthropic.com>
* fix(hotkeys): satisfy clippy::collapsible_if on the F-key parse
Co-Authored-By: Claude Code <noreply@anthropic.com>
---------
Co-authored-by: Claude Code <noreply@anthropic.com>
Show Usages showed nothing until its popup opened (300 ms after the first usages, or when the whole answer came), so a slower search looked like the shortcut had done nothing. It now shows "Searching for usages…" at the caret until the popup or the answer replaces it, and pending hints (such as Go to Implementation's) appear after 50 ms instead of 250 ms.
A filtered log connects each visible commit to the nearest matching ancestor along every merge path, and in merge-heavy histories one row could collect dozens of dashed connections, each getting its own lane, so the graph column stretched the log horizontally. pruneFilteredEdges now drops a dashed connection whose target the commit's other connections already reach (shallowest first), so reachability is unchanged with one lane instead of one per hidden merge path. Real parents, collapsed fragments and unfiltered histories are untouched; parent/child navigation follows the same pruned graph.
Fixes#2670
Co-authored-by: Inf1mum <8135225+Inf1mum@users.noreply.github.com>
Clicking a completion appended it to the command line and re-filtered the list, so repeated clicks kept accumulating goals. A click now selects the row and shows its command line in the input; another click switches to that row; the list filters only on typing; Enter or a double-click runs it. Goals and tasks already on the command line are no longer offered again.
Co-authored-by: PLZY <+PLZY@users.noreply.github.com>
The selection-occurrence cache keyed matches by absolute buffer row and relied on onRender to invalidate edited lines, so a full scrollback trimming from row 0, CSI 3 J or clear() shifted cached margin and scrollback rows without invalidation, and a selection trimmed off row 0 left its boxes behind. The cache now anchors a marker at row 0 while populated (any top trim disposes it and clears the cache), invalidates active-screen rows across baseY advances on write, and clears the overlay when a trim removes the selection.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
Adds a Django Template language: .html/.txt files in a templates folder of a project with manage.py (new [[framework-files]] detection), a Lezer grammar for tags, variables, filters and comments with HTML mounted between them, block folding/indent/matching, {% %} / {{ }} / {# #} pairing, end-tag insertion and completion, per-line {# #} comments, HTML language server and Emmet in the HTML parts, a banner offering the language for HTML files with Django tags elsewhere, a fixture and docs. The built-in HTML syntax check skips Django templates, whose tags the HTML grammar reports as errors.
Refs #2581
Co-authored-by: andreabasile <7715018+andreabasile@users.noreply.github.com>
gopls reports available upgrades only after its "Check for upgrades" code lens runs, never reports deprecated modules, and checks vulnerabilities only on demand. Opening a go.mod now runs the upgrade check once per file and session, gopls' vulncheck setting is on for imports (Go, go.mod, go.work), and `go list -m -u` marks deprecated requirements on open and save. go.mod highlighting also covers godebug, tool and ignore.
Deng Liuanddengliu
l0ng-ai
Gabriel Gondim
rohitkumardandaandrohitdanda
Bruno Belloni
Vadim LazukandInf1mum
Alexanderandformiat
SunZhiQiang
arookieofcandClaude Code
zhangdagouandPLZY
Andrea Basileandandreabasile