The Breakpoints dialog gets a JavaScript Events and Requests section: stop when a debugged page handles a mouse, keyboard, form, page or timer event, or sends a request whose URL contains a given text. They are kept per project and sent through js-debug's setCustomBreakpoints to every page session, live ones at once and later ones when they start. A page's requests wait until js-debug has answered its launch, since requests sent while it sets the session up are acknowledged and dropped. A stop on one of these says what it was for above the frames, and js-debug's Caught Exceptions filter explains that it is how errors a framework catches are stopped at. A new page session whose id was used before (after a backend restart) also gets the breakpoints.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
Debugging a Node.js program that prints a local address serving an HTML page can open that page in a browser debug session (Settings › Build, Execution, Deployment › Debugger › JavaScript, off by default). Ctrl/Cmd+Shift-clicking a URL in a console debugs it in the browser. The new web_serves_page command only asks loopback addresses, without a proxy, follows redirects only within this machine and times out after 10 s.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
The SSH/SSL tab gets Test SSH Tunnel: it logs in to the SSH host (shared SSH configuration or the settings typed in the dialog) and checks that the SSH host reaches the database host and port, without the database password or login. Success shows user@host:port and the database host:port reached; failures name the step (connecting, authenticating, reaching the database). An unknown host key offers Trust Host and Retry. The in-process SSH test server now forwards direct-tcpip channels.
A watcher batch used to re-read the whole working tree status, so a save in a large repository walked every file. The batch's paths now go to a scoped read (vcs_status_paths, gitoxide's status with a pathspec; git status with a pathspec as the fallback) and are merged into the loaded status: a path that is no longer changed drops out, a removed directory drops what is below it, new entries land in path order.
The full status still answers when the scoped one could be wrong: batches over 200 paths, a status that is not loaded yet, ignore-rule changes, a failed read, and batches that can split a rename. Git pairs a deletion with an addition across the whole index, so a pathspec read of a staged rename's target reports it as added and drops the source; batches that touch a rename, read one, or read an addition or deletion while the other side exists elsewhere take the full status. Restricted Mode keeps starting no program: the scoped read is in-process.
Co-authored-by: muutot <92625353+muutot@users.noreply.github.com>
Agents › IDE MCP Server gets a Listen address setting (mcpServer.address): 127.0.0.1 by default, or another address of this computer (LAN / VPN) or 0.0.0.0 / :: so agents on other computers can connect. Requests must still carry the bearer token, and the Host check still refuses host names other than localhost (DNS rebinding). An address that isn't an IP address, or isn't on this computer, stops the server with an error naming it. Clients on this computer keep connecting through loopback; the discovery file records the address. The settings hint and docs warn that anyone who can reach the address and has the token can use the tools.
Co-authored-by: guoyucode <29471644+guoyucode@users.noreply.github.com>
On macOS 11 and 12 other apps keep using the system's built-in WebKit (Safari 14.1 / 15.x) even after Safari is updated, so the Safari 16 interface could not start there. The bundle now declares macOS 13 as its minimum, the install docs say macOS 13, and the startup failure page no longer suggests updating Safari.
Fixes#2008
Settings › Languages & Frameworks › Language Servers looks for newer versions of the tools in the IDE's tools folder once per session and from its refresh button: release downloads resolve their latest release as an install does, npm packages ask npm view through the user's registry. A tool with a newer version gets an Update tag, the available version in its details and an Update button that installs it.
A server binary at <data dir>/server/<os>-<arch>/openidea-server was preferred over this version's download and never checked, so after an update remote windows kept installing the old server and new commands failed with -32601. The server now comes from OPENIDEA_SERVER_BINARY, the one next to the app, or this version's download (server/<version>/<os>-<arch>/); the unversioned location is an offline fallback only. A command missing on a server of another version names both versions.
On Windows a WebView2 IPC callback reaches settings_set with about 750 KiB of the main thread's 1 MiB stack already used, while validating the merged Settings needs about 250 KiB in a debug build (measured by bisecting a worker's stack size; independent of the settings file), so applying a setting could overflow the stack. The write now runs on a worker thread that the command joins, keeping writes serialized as before.
Co-authored-by: muutot <92625353+muutot@users.noreply.github.com>
FTP and FTPS deployment servers had no password field: the password was only asked for after a failed Test Connection. The Connection tab now has Password and Save password below User name, staged like a setting and written on Apply (saved, or for this session); clearing Save password forgets a saved password, and a saved one shows as "Password already stored". Test Connection logs in with a typed, not yet applied password.
Fixes#3151
Environment variable and header values of the user's MCP servers move from the user settings.toml into the credential store (AgentSecrets, as for agent environment values): on start-up, on settings import/export, and when the MCP page saves. They are filled back in when the page lists servers and when a session starts. A project's own servers keep their values in the project's .openidea/settings.toml, where the project shares them.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
Escape the page leaves unhandled reaches the window as cancelOperation:, and NSWindow leaves full screen on it, so pressing Escape once more after closing a popup dropped the IDE out of full screen. The IDE's windows now answer cancelOperation: with nothing; View › Appearance › Toggle Full Screen (⌃⌘F) leaves full screen. A page-level preventDefault was not used because several windows close on an otherwise unhandled Escape.
Fixes#3104
The IDE MCP server took a new port and token in every run, so configured HTTP clients and port forwards broke at each restart; every MCP request brought the window to the front; and an agent's edits were announced as refactorings.
- Settings › Agents › IDE MCP Server: Use a fixed port (on by default, 26371) with a fallback to a free port when it is in use or reserved, reported in Status; Keep the token (off by default) stores the bearer token in the app data, readable by this user only, with Regenerate Token. Changing either restarts the server. Access control is otherwise unchanged (127.0.0.1, bearer token, origin and host checks).
- Only confirmations, which wait for the user, raise the window.
- Agent edits stay undoable (Local History point for multi-file edits) but are not announced as refactorings.
Co-authored-by: guoyucode <29471644+guoyucode@users.noreply.github.com>
A view's html body was loaded as a srcdoc frame, which inherits the workbench CSP; its script-src 'self' blocks inline scripts, so in every build that serves the CSP html views rendered but ran no script. In the app the body is now written into a page of the openidea-plugin scheme (.frame) whose response carries only the frame's own policy, sandboxed as before. Web views already load from the scheme; notebook HTML output never runs scripts.
The app's CSP now includes `frame-ancestors 'none'`, so no page (for example a project's HTML in the built-in preview) can embed the IDE's own pages in a frame. Notebook HTML output, plugin views (srcdoc frames) and the built-in HTML preview still render in a debug bundle that serves the CSP.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
A portable Windows build (the raw exe from the portable ZIP, which has no bundle type) updated through the installer, which installed the app elsewhere instead of updating the folder it was unpacked in.
- Updater: an unbundled Windows build asks for the `windows-<arch>-portable` target. Its update is the signed portable ZIP (verified by the updater's download), unpacked into a staging folder next to the exe; only `openidea.exe` and `WebView2Loader.dll` are accepted. Each file is replaced with a backup, rolled back if any replacement fails, and the backups are removed on the next start.
- Release: the portable ZIP is signed with the updater key and `latest.json` lists `windows-x86_64-portable` and `windows-aarch64-portable`.
Co-authored-by: caojcm <43193197+caojcm@users.noreply.github.com>
The MCP server let agents read and edit code but not run a command and follow it, so they ran commands in their own shell, out of the user's sight. New tools run_command, wait_for_run, read_run_output, get_run_status, send_run_input, stop_run and list_runs run each command in a Run tool window tab _agent_<name>: the user sees it and can type into it, rerun, stop or close it, and results then carry interventions and advice so the agent stops and asks. The same name reuses its tab; a running one is refused. Every start asks the user unless tools run without confirmation. Commands run on pipes, with maxLinesCount / truncateMode, offsets to read on, and pattern waits. Finished agent tabs beyond a limit (default 1) move to an Archived (n) menu.
A call still running after a second is answered on an event stream with a progress notification every 15 s, so clients that drop a response not started within a minute keep waiting; the run tools (build_project and execute_run_configuration included) no longer cap timeout. A call is cancelled when its client goes away; the stdio bridge relays streamed answers.
Co-authored-by: guoyucode <29471644+guoyucode@users.noreply.github.com>
A Paste chosen from a menu (the terminal's context menu, Edit › Paste, Paste as
Plain Text and other paste buttons) has no paste event to read, so it used
navigator.clipboard.readText(), and the terminal tried execCommand("paste") first.
On Linux, WebKitGTK gives neither the text, so the terminal's context-menu Paste
did nothing while middle-click paste worked.
A new clipboard_text command reads the clipboard's text natively, as
clipboard_write_text already writes it: GTK on Linux, NSPasteboard on macOS, and
none on Windows, where the WebView's clipboard is still used.
readClipboardPlainText() wraps it with that fallback, and every menu paste reads
through it. On Linux the terminal pastes the native text before trying
execCommand; an image-only clipboard keeps the existing paths.
Adds an ide-vcs Backend trait for what the Commit / Local Changes view, diff viewer, annotate and log need from a repository (status, file_at / file_mode_at, diffs, line hunks, blame, block history, log, commit files, revision diffs, stage / unstage / discard, commit, pending changes), with Repo implementing it by forwarding. Git-only concepts (branches, remotes, stash, rebase, worktrees) stay on Repo, reached through as_git.
The trait has default no-op hooks for IDE file operations (files_added, files_deleted, file_moved), which the host now calls after create, copy, delete and rename so a checkout-based backend can open files for add / delete / move. The Changes-surface call sites in ide-host go through &dyn Backend. No behaviour change for Git.
Co-authored-by: arookieofc <186315048+arookieofc@users.noreply.github.com>
In a remote window the desktop took several decisions about itself from the server: it routed commands by the server's `commands` reply (a command the server left out ran locally), authorized its own commands from the server's `project_current().trusted`, and showed or ran whatever events the server forwarded, including any workbench action named by a plugin `executeCommand` event.
- Routing: commands go by `rpc::client_commands()`, the list a server of this build serves, compiled into the desktop; a test builds the server's dispatcher and compares the two.
- Trust: commands that run on the desktop follow the desktop's own decision, stored in `remote-project-trust.json` by target and root (a dev container also follows its local folder). The trust prompt appears unless the desktop's and the server's decisions agree; the answer and Revoke Trust update both. New commands `remote_trust_status` / `remote_trust_set`.
- Events: server `event`s pass an exhaustive allow-list (MCP approvals, update notices and window-group events are dropped); plugin `executeCommand` may only name `plugin.<id>.<command>`; `openProject` is dropped.
- Paths: a project root that the desktop would read as a share or device path is refused at connect.
- A remote plugin's `openUrl` and `copyToClipboard` ask first, showing the full address or the text.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
Debug builds set no default marketplace index, so the Plugins page showed only the repository's local plugins and browsing, installs, update and revocation checks couldn't be exercised without entering the index URL. The official index is now the default in every build; local plugins keep their own group.
OPENIDEA_CONFIG_DIR, OPENIDEA_DATA_DIR and OPENIDEA_LOG_DIR take an absolute path, like OPENIDEA_LOCAL_DATA_DIR, and the IDE uses <path>/<identifier> under it. Every place that resolved these directories now goes through ide_host::data_dirs (app state, invalidate caches, window zoom and window state, the MCP stdio bridge, logging, the openidea-plugin CLI), and the frontend takes the user data directory from app_info instead of Tauri's appDataDir. A relative path for the configuration or user data shows an error and exits.
Closes#2814
* feat(python): manage every interpreter from a Python settings folder
Settings › Languages & Frameworks now groups Python settings in one folder: All Interpreters (new), Dependencies (the project interpreter and its packages), Docstrings (split out of the interpreter page) and Ruff. All Interpreters lists interpreters detected on the machine, the ones other recently opened projects use and the ones added there, with a search field, PyCharm's "hide interpreters of other projects" and "show interpreter paths" toggles, and + / − to add an interpreter or remove it from the list (hidden, never deleted; a removed project interpreter falls back to the project virtualenv).
Each interpreter's source has its own icon (virtualenv, conda, uv, Poetry, pyenv/mise/asdf, system) in the list, the Add Interpreter dialog, the interpreter combo and the status bar menu. The IDE-wide list and its view options are global settings that a project cannot override; removed interpreters also disappear from the interpreter combo and the status bar popup.
* fix(python): compare interpreter paths by their own case rules and cap probes
A POSIX interpreter path kept only its own case (Linux allows `/opt/Python` and
`/opt/python` side by side); only Windows drive/UNC paths are folded now, and the
status bar uses the same key as the list. A configured `~/…` interpreter is matched
against the absolute path the backend resolves it to, so the "Project interpreter"
badge and removing it in use keep working without the frontend knowing the home
directory.
`sdk_python_project_interpreters` probes at most four recent projects at a time
instead of one thread per project, so opening All Interpreters does not start ~30
pythons at once. The hidden-but-in-use interpreter stays selectable in the combo and
status popup (custom entry / current fallback); that is now documented on `exclude`.
---------
Co-authored-by: arookieofc <arookieofc@gmail.com>
Probing a broken Python install (for example one whose python3xx.dll is gone) blocked on a modal "System Error" dialog, because Windows shows critical-error boxes for any process that fails to load a DLL. The app now sets SEM_FAILCRITICALERRORS and SEM_NOOPENFILEERRORBOX at startup, keeping inherited bits, and children inherit the mode, so such tools exit with an error code instead. Covers interpreter detection, Run, Terminal and every other tool the desktop app starts.
On Wayland with the NVIDIA driver WebKitGTK's DMA-BUF renderer fails with Error 71 unless something works around it. Turning the renderer off made WebKit paint on the CPU, so dragging a tool window stalled the UI for seconds. Setting __NV_DISABLE_EXPLICIT_SYNC=1 for the process avoids the protocol error and keeps GPU rendering. A user who already set either variable keeps their own setting; WEBKIT_DISABLE_DMABUF_RENDERER=1 still restores the old behaviour.
Co-authored-by: brunobelloni <14021404+brunobelloni@users.noreply.github.com>
Inside a <dependency>'s <groupId>, <artifactId> or <version> in pom.xml the editor suggests the coordinates installed in the local repository (Settings › Maven's, else ~/.m2/repository) and, behind them, Maven Central's: the search endpoint's groups and artifacts and the artifact's published versions, cached for the session.
The local repository answers at once; Central gets one second and otherwise finishes in the background for the next request, so an unreachable Central never holds completion. Suggestions keep their order (installed first, newest versions first), and a version counts as installed only with its .pom or a .jar, not the .lastUpdated files a failed download leaves.
Fixes#2968
Co-authored-by: Inf1mum <8135225+Inf1mum@users.noreply.github.com>
Snowflake is a built-in data source kind: connection dialog, key-pair, browser single sign-on, password or token sign-in, the database tree, SQL consoles, and Snowflake-aware SQL assistance and syntax colours. The driver speaks Snowflake's session protocol over the existing reqwest + rustls stack, with no JDBC/ODBC or native library and no new package. Metadata is read with SHOW / DESCRIBE, which need no running warehouse; introspected names are kept on disk for completion.
Co-authored-by: vvnadolskyy <12991410+vvnadolskyy@users.noreply.github.com>
Tools › Start SSH Session ran the system `ssh -t` in a terminal, which never
used the password saved in the SSH configuration and, on Windows, could not
take a typed password either.
The session now opens a login shell over the IDE's SSH connection: the
terminal manager can attach a channel (here an SSH shell) next to local PTYs,
with the same output batching, UTF-8 decoding and exit events. A password or
passphrase that isn't saved is asked for in a dialog, and an unknown host key
asks to be trusted, before reconnecting. Closing the tab hangs the shell up.
Also: a remote process closed by this side now ends without waiting for a
close that russh never forwards, and each connection attempt gets its own IPC
channel, since a failed attempt invalidates the one it was given.
Fixes#2874
Switching a theme with several windows open flickered for seconds: the plugin theme is shared by every window, and each window put back the theme its own settings named whenever it changed, before those settings caught up. Windows now act on the plugin theme only when their own selection changes, and applying the theme already shown keeps its stylesheet instead of replacing it.
Theme previews in the gear menu, Quick Switch and Settings waited 500 ms, and on themes written with hsl() or color-mix() the terminal took up to a second to read its colours: each token was resolved through a new element inserted under the root, restyling the whole page every time. Previews now show at once, and token colours resolve through one element kept for it.
The built-in One Light and One Dark themes are removed in favour of the One Dark Pro plugin; settings naming them read as Islands Light and Islands Dark.
A worktree task's Changes pane only listed the cumulative diff since the task started, so work the agent committed was mixed into the current changes. It now groups the task's work: uncommitted files compared with HEAD, and each commit made since the task started with the files it changed against its first parent. Double-clicking a commit or one of its files opens the parent-to-commit diff; commit files are history and are not rolled back. Rollback in this view restores files to HEAD. Accept still brings over the task's whole work, committed or not. View Options > Group By > Commit switches back to the flat since-start list; tasks in the project's own checkout keep that list.
Co-authored-by: panda-student <71381346+panda-student@users.noreply.github.com>
A chmod-only change is a real modification: git status lists it and a commit records the new mode, so status keeps reporting it (core.fileMode=false still hides it). The diff viewer used to show "Contents are identical" with no explanation; when both texts are equal but the git modes differ it now says "File mode changed: 100644 → 100755". The modes come from a new vcs_file_mode command (index/tree entry mode for revisions, worktree permissions for the working copy).
Closes#2853
Co-authored-by: a-rookie-of-C-language <186315048+a-rookie-of-C-language@users.noreply.github.com>
In a remote window the built-in web server ran on the client and read the local file system, so URLs for remote files were 404s. The server now lives in the host: openidea-server answers web_url with its own server on the remote loopback, and the client forwards that port over the SSH or dev container connection and returns the forwarded URL. The Host check accepts any port with 127.0.0.1 / localhost, since the forwarded port can differ.
Fixes#2915
Javadoc, KDoc, JSDoc, rustdoc and other /** and /// comments can be drawn in place as formatted documentation and switched back to source: globally (Settings › Editor › General › Appearance › Render documentation comments, off by default), per file (View › Active Editor › Render Documentation Comments, gutter menu, editor.toggleRenderedDocs) and per comment (gutter icon on hover, editor.toggleRenderedView at the caret). Comments are rescanned only once typing pauses and the Markdown renderer loads on first use.
Co-authored-by: Clarence404 <36807519+Clarence404@users.noreply.github.com>
Settings said the language server JDK needs "jdtls 21+", but recent jdtls bundles an Eclipse platform that needs JDK 25, so a JDK 21 chosen there was rejected at launch. The requirement is now read from the installed jdtls (new jdtls_required_java command, the same bundle scan as the launch check): the hint states it, flags a too-old selection, the JDK picker marks too-old JDKs, and Download JDK preselects that version.
gitoxide checks ownership of a submodule's unnormalized `<worktree>/../../.git/modules/…` path with GetNamedSecurityInfoW, which rejects paths past 260 characters, so deep submodules were skipped and the parent's status fell back to git. The app manifest now declares longPathAware, so Win32 path APIs accept long paths where the system enables them.
Co-authored-by: HL-slc <73373127+HL-slc@users.noreply.github.com>
The welcome screen showed the branch stored when a project was last opened, so a checkout made later (in or outside the IDE) stayed stale. recent_projects now reads each project's HEAD when the list is built (linked worktrees via the .git file's gitdir; a detached HEAD shows the short commit), off the async runtime, and keeps the stored value when HEAD can't be read.
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>
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>
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>
* 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>
Remote Host › Download to… asks for a local folder (the Project view's selected
folder preselected) and copies only the selected files and folders, each under
its own name, instead of recreating the server path below the project root.
Dragging items from Remote Host onto a folder in the Project view does the same.
Download to Project keeps using the mappings.
Navigate › Declaration (⌘-click, ⌘B) on {{name}} in a .http file now lists
its definitions in lookup order: the file's @name = … line, then the key in
each environment (and $shared) of the nearest http-client.private.env.json
and http-client.env.json. Dynamic variables ($uuid…) have none.
Settings › Agents › Coding Agents can add a custom ACP agent (name, id, command, arguments, environment) without writing its agent file by hand, remove one, and import the agent_servers entries of a JetBrains acp.json. Built-in and ACP Registry ids are refused, ids are validated as file stems on save and on remove, and colliding or incomplete acp.json entries are skipped with a reason.
Closes#228
Co-authored-by: a-rookie-of-C-language <186315048+a-rookie-of-C-language@users.noreply.github.com>
The Changes pane of a worktree task gets an Accept action next to Rollback: after confirmation it copies the task's changed files (all, or the selected ones) from the task worktree into the local checkout, uncommitted. A file the checkout has changed itself is left alone and listed in the notification instead of being overwritten; running and local tasks are refused.
Closes#1777
Co-authored-by: a-rookie-of-C-language <186315048+a-rookie-of-C-language@users.noreply.github.com>
Wayland has no primary monitor, so the first-map size hint fell back to scale 1 with no work-area limit and requested the saved physical size as logical. On a HiDPI output the first mapped area exceeded Mutter's 80% share and the window opened maximized before the saved geometry was restored. The hint now converts with the largest output scale and shrinks proportionally to stay within the tightest output's share; X11 keeps its primary-monitor path.
Fixes#2400
Co-authored-by: Inf1mum <8135225+Inf1mum@users.noreply.github.com>
The Profiler gets two more modes for JVM runs through async-profiler: allocations (where the program allocates, in bytes) and live objects (what is still alive at the end). A live-object profile collects garbage first: run executors get a beforeStop hook, bounded to 12 s, that runs a GC before the stop signal; a program that ends by itself says that unreachable objects may show as live.
Part 3 of #2217.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
Changes trees get Show Diff in New Window: the diff opens in its own native window (labelled diff-*, with the same capability as comparison windows), with the full diff toolbar, editable working-copy side and a save prompt for unsaved edits on close. Showing the same diff again reuses its window.
Co-authored-by: wangjunboit <33538450+wangjunboit@users.noreply.github.com>
ntbang0901
l0ng-ai
Muuandmuutot
GuoYuandguoyucode
rohitkumardandaandrohitdanda
caojcm
arookieofcandarookieofc
Bruno Belloni
Vadim LazukandInf1mum
pHandvvnadolskyy
SunZhiQiang
Clarence
HL-slcandHL-slc
Gabriel Gondim
wangjunboit