* test(lsp): smoke-test everyday code intelligence per language with real servers
A new ide-host integration test drives completion, go to definition, Find
Usages, hover, diagnostics with a quick fix, an intention, formatting and
rename on each language's fixtures/ project, through the ops the app's
commands call and the language server the app starts by default:
TypeScript (vtsls), Python (basedpyright + Ruff), Rust (rust-analyzer),
Go (gopls), Java (jdtls) and Kotlin (kotlin-lsp). A failure names the
language and the operation.
The tests are ignored by default; a Language smoke workflow installs pinned
server versions and runs one language per job on changes to crates/,
languages/ and fixtures/. Adds fixtures/java-app (Maven, JDK 21) for Java.
* test(lsp): rename TypeScript from a usage and check Rust errors while typing
Covers the fixes the smoke test turned up: a TypeScript rename started at a
call site in another file renames the declaration (#1301), and rust-analyzer's
own diagnostics and fixes arrive without saving (#1306).
Current Kotlin LSP releases launch with bin/intellij-server; the
kotlin-lsp.sh / kotlin-lsp.cmd wrappers the installer picked now print a
deprecation warning and are announced to go away, which would leave a
fresh install without a server. The installer prefers the launcher and
still finds the wrappers in older releases.
rust-analyzer reports its own analysis (a missing field, a wrong type) only
to clients that pull diagnostics; it pushes just `cargo check`'s, so in the
editor a Rust type error showed up only after saving. The client now
advertises `textDocument/diagnostic` and pulls an open file's diagnostics
from servers with a `diagnosticProvider`: when it opens or its server
starts, after edits (once typing pauses), on save (and other open files
for servers with inter-file dependencies) and on
`workspace/diagnostic/refresh`. Pulled and pushed diagnostics are kept
apart and shown, and given to code actions, together; a newer pull
supersedes an older reply, a server-cancelled pull is retried, and closing
a file drops what was pulled.
Renaming an imported function at a call site renamed only the local binding,
leaving `import { formatMoney as formatPrice }` and the declaration as it
was: both TypeScript servers default to aliases on rename. vtsls gets
`typescript.preferences.useAliasesForRenames = false` (and the JavaScript
one), typescript-language-server `providePrefixAndSuffixTextForRename = false`,
in every language that starts them.
Unversioned files in the commit view could be added or ignored but not deleted. Their context menu now offers Delete…, as IDEA's Changes view does: it asks first and moves the files to the Trash, only when every selected file is unversioned.
gopls sends a document version for every file of a rename, including files
the editor never opened. Applying a workspace edit rejected any versioned
file without an open document as changed, so a Go rename touching a closed
file always failed with "changed; request the refactoring again". For a file
that isn't open, the preconditions captured with the edit now decide; an
edit without them is still rejected.
Show History picked its repository with repoOf, which routes a nested repository's own root to the enclosing repository (right for the commit view's submodule pointer), so a submodule folder listed its gitlink history in the parent repository. A repository's own root now opens that repository's history, as in IDEA; other paths still use the innermost repository.
The inline cell editor is a single-line input that drops a value's line breaks on the first change, and it committed on every Enter; the Record view's fields had the same problem. Multi-line values now open in the multi-line editor, Shift+Enter opens it for any cell (from the grid or the inline editor) and Cmd/Ctrl+Enter applies it; CRLF values keep CRLF; the grid marks line breaks with a muted ↵; the Record view uses a text area for multi-line values.
Merged project windows are separate native windows that take turns being visible. Hidden members kept their own earlier frame and monitor, so the first switch after merging, and any switch after moving or resizing the window, moved or resized the window on screen; a maximized member could appear on another display. The group now copies the selected window's frame and display to the hidden members after a selection and after the visible window is moved or resized (debounced), skips copies when the frame already matches, and leaves macOS simple fullscreen to its own path, so switching projects only changes which window is visible.
Co-authored-by: xtr-hub <242808782+xtr-hub@users.noreply.github.com>
#1309 made the Find in Files, file scope, index and walk_files walkers require git inside a repository, so an outer repository's .gitignore stops at a nested repository. External TypeScript project discovery (ide-lsp) and Lombok detection (ide-sdk) still walked with require_git(false) and could miss a nested repository's sources the same way; they now use the same check. Adds tests for walk_files (nested repository with its own .gitignore and info/exclude, a plain folder inside the outer repository, the outer repository, a folder outside any repository) and for scope_files.
Co-authored-by: jaw52 <53564781+jaw52@users.noreply.github.com>
The walkers used require_git(false), and with it the ignore crate never treats a .git directory as a boundary, so an outer repository's .gitignore hid a project that is its own repository inside an ignored folder (Find in Files found nothing). Walkers now require git when the walked folder is inside a repository, so rules stop at a nested repository's root as in git; folders outside any repository still honour their .gitignore. Covers the Find in Files engine, the file scope, the index walk and ide-vfs walk_files.
New Branch dropped its suggestion when validateBranchName rejected it, and the current branch always exists, so a plain New Branch from HEAD always opened empty; the test never seeded the local branches. The suggestion is kept, the dialog shows that the branch exists from the start and Create stays disabled until the name changes; other prefilled prompts (Rename Branch) keep showing errors only after editing.
The find bar belonged to the pane and every bar shared one query, and its effect re-ran the search and moved the caret whenever the active tab changed, so switching files dragged the previous file's search into the next one and lost its caret. Each tab now has its own find bar with its own query and options, as with IDEA's per-editor search session. The last search only pre-fills a newly opened bar (or the selection does) and drives Find Next without a bar; showing a bar again only highlights, and the caret moves only when the query or an option changes in that bar. Bars of closed tabs are dropped.
Co-authored-by: jaw52 <53564781+jaw52@users.noreply.github.com>
The JVM run-target scanner only reads the file's own text, so a class like `class App extends AtomStarter {}`, whose superclass (in the project or a library JAR) declares `static main`, got no gutter Run, no Ctrl+Shift+R and no file-level Run, although the JVM launches it. Classes with an `extends` clause and no main of their own are now candidates: the superclass chain is walked through the language server's type hierarchy, each class supertype's source (file, decompiled or source-jar document) is checked for a static main, and the answer is cached per class and `extends` clause, so the gutter, Run from context, launch and debug agree. Interfaces are skipped (a static interface main is not inherited); negative answers are asked again when the server settles. `main(java.lang.String[] a)` is accepted as well.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
On WebKit, CodeMirror positions tooltips absolutely inside the editor, but the tooltip space still reported the whole window as the right bound. CodeMirror never shifted a tooltip near the editor's right edge to the left, the box was squeezed into the space left of that edge, and the next measurement read the squeezed width, so it could end as a one-character column. The tooltip space is now the editor's own horizontal extent (clamped to the window) and tooltips size to their content, which fixes hover, inline blame, debugger values, signature help and hints together.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
"Show pinned tabs in a separate row" was ignored while "Multiple rows" was on, because the pinned row explicitly excluded horizontal layouts with multiple rows. The two options are independent, so the pinned row now shows whenever some but not all tabs are pinned, and each row wraps on its own. With the rows split, dropping a tab on the pinned row pins it and keeps it in the pinned group, and dropping it on the other row unpins it.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
Scroll containers set cursor: text, and the webview applies an element's cursor to its scrollbar, so scrollbars showed the I-beam. The containers now keep the arrow and the text inside them gets the I-beam: the editor (.cm-content instead of .cm-scroller, which also covers the diff and merge editors), the source preview, the Run, Build and Debugger consoles, the Git console, the Database output and the Services log.
Co-authored-by: ppxb <17081240+ppxb@users.noreply.github.com>
The workbench sets user-select: none on body and nothing restored it for editable elements, so a contenteditable such as CodeMirror's .cm-content inherited it. In WebKit a contenteditable with user-select: none cannot hold a selection, so typing, paste and drag-and-drop into the editor could be dropped. input, textarea and [contenteditable="true"] now get user-select: text next to the cursor: text rule they share; the selection lock still overrides it.
Co-authored-by: nguyenmuoi157 <33494100+nguyenmuoi157@users.noreply.github.com>
Activating a tool window focuses its `[data-autofocus]` element or first
content field one frame later, else its body. A lazily loaded view such as
the Terminal renders its field after that, so the first ⌥F12 left focus on
the body. When no target exists yet, the tool window now watches its body
and focuses the target when it appears, unless focus has moved on.
Adds a `stubDesktop` fixture that switches the browser-mode workbench to
desktop mode over a stubbed IPC layer: tests answer the commands they need,
push Channel messages and IdeEvents, and assert on the commands invoked.
Unstubbed commands fail as unavailable; documents keep the in-memory backend.
New flows covered:
- editing: modified marker and ⌘S save, ⌘R Replace All with one undo step
- code intelligence: completion, ⌘B to another file, Find Usages, in-place
rename, diagnostics with an ⌥⏎ quick fix
- VCS: ⌘K commit with message validation and default inclusion
- run: ⌃R output streaming and exit, ⌘F2 stop
- debug: breakpoint, stop with frames and variables, step over, resume
- search: Find in Files results opening at the match line; the Search
Everywhere file case no longer needs fixme
- terminal: typed command reaches the shell and its output renders
- project: Close Project to Welcome, trust prompt, open a recent project
java-usages, kotlin-bridge and the patched Lombok agent are JDT LS bundles and a Java agent embedded into the app, not OpenIDEA plugins. Move them to java/ so plugins/ holds only plugins, and update the paths that embed or mention them.
Also fix the stale build path in the hello fixture and the contributor guide's layout (there is no templates/ directory; starter templates live in the openidea-plugin CLI).
Opening a popup (intentions, completion) moved focus out of the editor and
autosaved it. The save cleanup could change the text, such as trailing
spaces on another line, so the quick fix the popup listed no longer matched
and the first attempt to apply it failed.
While focus is in a popup the editor is not saved; it is saved when focus
moves on from the popup to anything but the editor.
An edit's preconditions hashed disk contents plus unsaved buffers. Saving a
buffer moved its text from one to the other and changed the hash, so a
quick fix, rename or refactoring listed just before a save (autosave of a
freshly typed scratch, for one) failed with "changed; request the
refactoring again", and only the next attempt worked.
Unsaved buffers now stand in for their files' bytes in the hash, so saving
doesn't change it; edits, external writes and file operations still do.
The Available tab always showed a "There is no online plugin repository"
notice under the search field when no index was set. Release builds use
the official index by default, so the notice only appeared in development
builds, where the empty URL field above it already said the same thing.
The URL field itself sat at the top of the list although most people never
change it.
- The notice is gone, and the search placeholder reads "Search plugins"
because the list holds marketplace plugins as well as local ones.
- With a repository configured, the empty details panel says "No plugin
selected" instead of claiming there is no online repository.
- The marketplace index URL is edited from the gear menu's "Plugin
Repository…" dialog; leaving it empty restores the default.
A new Java scratch was an empty file that couldn't run until a class and
main method were typed in. Languages can now define `scratch-template` in
`languages/<id>.toml`; a new scratch gets that content with the caret at
`$END$`.
- Java, Kotlin, Go, Rust, C and C++ get a runnable `main`.
- Other languages still start empty.
jdtls still completes and navigates JDK classes in a non-project file such
as a scratch; it only skips semantic checks and project classes. The banner
said navigation and completion don't work, and in a Maven or Gradle project
it offered a build reload that can't bring a file outside the project in.
- The banner now says only syntax errors are reported and only JDK classes
resolve.
- The Maven / Gradle reload is offered only for files inside the open
project (or one of its workspace projects).
`taskkill /T /F` only starts terminating the tree, and the caller then
waits for its direct child. A grandchild could still be exiting, with its
files open, when `kill_tree` returned or `Peer::exited` resolved. The
descendants are now opened before the kill and waited for (up to 5 s)
after it. A process counts as a descendant only if it started after its
parent, so a reused parent id doesn't pull in unrelated processes.
Closes#1270
A provider call that starts a plugin instance (or reads files) can run past
the edit-path soft deadline on a slow runner and returns nothing; the tests
now ask again, as the editor does. The AI completion test cancels the slow
request only once the server has it, and measures from the cancel. The
SourcePreview test waits for the effect that marks the selection scope.
Closes#1269
- `openidea-plugin new` makes a Rust plugin depend on the SDK at the tool's
release tag (git) unless the plugin is inside the source tree the tool was
built from or `--sdk` is given; the path written before only resolved on
the machine that built the tool.
- The command template compiles again: `run_command` returns
`openidea_plugin::Error`. A test now runs `cargo test` on the generated
plugin against the SDK.
- plugins/README.md: installing the tool with `cargo install --git`, the SDK
dependency, where plugins can be installed (welcome screen and trusted
projects), and where to look when a plugin misbehaves.
* fix(lsp): resolve newly installed Python packages without a restart
Packages installed into an excluded .venv (uv add, uv pip install) never
reached the Python language servers: the watcher reports changes in excluded
folders only as excluded_dirs, and didChangeWatchedFiles forwarded nothing but
the server's own source files. basedpyright kept its failed import resolution
until restarted.
Language servers get a watched-files list of globs. Files matching it, and
excluded folders matching it, are forwarded too; the Python servers watch
their configuration and site-packages.
* fix(lsp): let servers register the files they watch
Clients now support dynamic registration of workspace/didChangeWatchedFiles,
with relative patterns. Changes on disk go to each server as its registered
watchers match them, including folders in excluded trees (a package installed
into .venv, go.mod rewritten by go get). Folders outside the project that
relative patterns name (a virtualenv elsewhere, a path dependency) get their
own watchers. Servers that register nothing keep hearing of their languages'
files.
Registrations sent right after initialized are no longer dropped while the
server is still being marked ready.
Replaces the per-server watched-files list.
The webview focuses a clicked button and the focus stays there: across
project window-tab switches, and popups hand it back to their trigger. The
next keystroke switches the engine to keyboard modality, so that button got a
focus ring (window tabs, the title-bar project switcher). A pressed control is
now marked and keeps no ring until Tab or an arrow key moves focus onto it.
Focus itself is unchanged, since actions resolve their context from it.
Closes#1211
Folders now appear in the composer's `@` completion next to files, from a new
`find_dirs` index query over the folders that hold indexed files (kept
incrementally in the index, ~0.5 ms on 100k files). Enter or a click
references the folder as `@dir/` and sends it as a resource link; Tab keeps
completing inside it.
Closes#1202
Closes#969
## What and why
On Windows, deleting an agent task and its worktree could fail with `os error 32` while the agent's process tree still held the directory or its files open. Deletion now waits up to 5 seconds for the agent process to be killed and reaped, then retries folder removal for transient locks. This also covers persistent ACP agents such as CodeBuddy that cannot reload their session and therefore stay alive while idle.
## How it was verified
- Windows desktop acceptance through the Tauri MCP bridge, using a deterministic ACP agent with a real PowerShell descendant holding a file open. Closed the conversation tab, then deleted the task with worktree removal enabled. A separate helper retained another file lock for 1.2 seconds: the pre-fix implementation showed `os error 32`; the fixed implementation deleted the task and directory, with both agent processes gone.
- Also verified direct deletion, archive then delete, and deletion with worktree removal unchecked. Processes exited in all cases; branches were preserved, and unchecked removal preserved the folder.
- `cargo test -p ide-jsonrpc -p ide-agent -- --test-threads=2`: 155 passed, 5 real-agent tests ignored. The test runner used Git Bash on PATH and a process-local `core.autocrlf=false`, matching CI's LF fixture setup; no global Git settings were changed.
- `cargo clippy -p ide-jsonrpc -p ide-agent --all-targets -- -D warnings`, `cargo fmt --all --check`, and `git diff --check`: passed.
- Full workspace (Windows/macOS/Linux), frontend and E2E CI: [final PR run](https://github.com/l0ng-ai/openidea/actions/runs/41819945536).
- Windows regression coverage now checks deletion while waiting for permission and while idle without `session/load`, actual sharing violations, delayed file-lock release, and `Peer::exited()` waiting for a descendant's file handle to close. Existing Unix process tests remain enabled.
- Evidence (local, not committed):
- Before: `C:/Users/thoma/repo/pr1240-evidence-20261104/before.png`
- After: `C:/Users/thoma/repo/pr1240-evidence-20261104/after.png`
- Recording: `C:/Users/thoma/repo/pr1240-evidence-20261104/after.gif`
- Runtime results: `C:/Users/thoma/repo/pr1240-evidence-20261104/runtime-cases.json`
- Real signed-in CodeBuddy and native Claude/Codex were not exercised. Native Claude/Codex use an in-process connection, so their `exited()` resolves immediately and removal relies on the bounded retry. Shared-worktree fork protection remains unchanged.
Enumerate every locally installed Windows font face, including per-user font files. Correct glyph-width probing and collection-file enumeration, and exclude cloud fonts that are not installed.
Closes#1241
* feat(plugins): list community plugins from a registry in the marketplace
The marketplace index can now carry plugins of other publishers, published
from their authors' repositories through the l0ng-ai/openidea-plugins
registry.
- `openidea-plugin index` with a key replaces only that publisher's plugins
(it used to drop every other entry); without a key it indexes each package
against the publishers the index already lists, checking the key and the
id prefix instead of only the signature itself.
- `openidea-plugin merge` takes another index's publishers, plugins and
revocations in place of every publisher but `--keep`; the other index can't
list or revoke the kept publishers' plugins.
- `openidea-plugin check` accepts a `.oiplugin` file.
- Releases merge the registry's index into plugins-index.json and attach the
Linux x64 tool; plugins-index.yml refreshes the community plugins in the
latest release hourly or when the registry signals a change.
- plugins/README.md and CONTRIBUTING.md describe official and community
plugins.
* feat(plugins): offer the newest plugin version this IDE can load
Index entries record the plugin API each version needs (`engine`: the
manifest's `engines.openidea` narrowed to the API its component was built
for), and an index may list several versions of a plugin. Settings › Plugins
offers the newest version this IDE can load; with none, the plugin is shown
as incompatible with the API it needs, and downloading it is refused.
`openidea-plugin index` keeps, of each plugin, the newest version per API
requirement, so users of an older IDE keep the last version built for them;
`--download-base` makes downloads absolute, and the release workflow points
them at the release that published them. Scratch directories are now unique
per call.
* feat(plugins): file marketplace plugins under categories
`plugin.toml` takes `categories` (up to 3 of languages, themes, icon-themes,
linters, vcs, ai, frameworks, tools, other). A plugin that declares none is
filed under what its contributions imply (themes, icon themes, languages),
else under other; categories an IDE doesn't know are ignored, and `check`
warns about them. Index entries carry the categories, and Settings › Plugins
› Available filters by them and shows them in a plugin's details. The
official plugins declare theirs.
* feat(plugins): let a marketplace index point to where it moved
An index with `moved` sends the IDE to the index at that URL (relative or
absolute, up to three hops; an https index can only move to https). Downloads
resolve against the index actually read, so a marketplace can change hosts
while released IDEs keep the address they were built with.
* fix(plugins): full-width category filter and aligned incompatibility note
* fix(plugins): translate the AI category in Chinese
* feat(plugins): keep a plugin's signer across updates and announce updates
- Index format: fields are only added; an IDE reading a newer `format`
still applies its revocations and `moved`, lists no plugins and asks to
be updated, instead of failing to read the index.
- Key continuity: an installed plugin updates silently only to a package
signed with the same key, or a key it endorsed (`KeyEntry.endorsement`,
`index --previous-key`, `openidea-plugin endorse`). Another signer, or an
unsigned package replacing a signed one, is shown on the install page and
needs the user's confirmation (`Consent.signerChange`).
- Official key: release builds carry the official publisher's public key
(`OPENIDEA_OFFICIAL_PLUGIN_KEYS`, derived from PLUGIN_SIGNING_KEY);
`openidea.*` plugins must be signed with it or a key it endorsed,
whatever the index lists.
- Updates: the daily marketplace check announces new versions once, with
Update All (also in the gear menu and as an action). An update that asks
for nothing new installs with the grants the installed version has.
- Index entries need `engine`; `categories` beyond the first three are
ignored instead of rejecting the manifest.
l0ng-ai
xtr-hubandxtr-hub
Jaw
ntbang0901
ppxbandppxb
nguyenmuoi157
Marius Ciepluch