A Mermaid diagram inside <div align="left"> (or right / center, or <center>) stayed centered in the Markdown preview and agent chat: DOMPurify drops inline style, and .mermaid-output's text-align: center plus the SVG's margin: 0 auto overrode the kept align attribute. Aligned containers now set inherited --mermaid-align / margin properties that the diagram output uses; unaligned diagrams stay centered.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
The Close Tab button now stacks idle and hovered icons, so the single-svg locator hit Playwright's strict mode. The spec checks every svg each toolbar button renders.
Closes#3828
Co-authored-by: Inf1mum <8135225+Inf1mum@users.noreply.github.com>
ACP discovery learned effort settings only for the session's default model, so picking another model before its first prompt lost the effort control. The shared catalog now discovers a selected model's missing settings in a short-lived ACP session (session/set_config_option, no prompt), tells unknown settings apart from verified no-effort models, shows loading, discards stale discoveries during rapid model changes, and keeps per-model failures and cooldowns separate. Idle-task model switches use the same discovery and keep the previous controls when it fails. Discovery errors keep their cause.
Closes#3789
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
Adds Debug > Trace Stream Chain: paused on a line with a java.util.stream chain, the program runs that chain once, for real, while every element is recorded at every step, and a Stream Trace tab shows it Flat (all steps side by side, an element followed through the chain) or Split. Nothing is evaluated twice; the editor stops right after the statement.
- java/jvm-inspect 1.1.0 gets openidea.jvm.stream.*: wraps each step with peek(recorder) (a class defined in the debuggee), runs the terminal operation by invokeMethod, keeps breakpoints off while the chain runs and restores them, then parks the thread. The jar is reproducible from build.py (JDT LS 1.63.0, JDK 25).
- New ide-streams crate finds the chains on the paused line (tree-sitter) and builds the plan; ide-dap drives continue, wait, step out, collect. IPC: debug_stream_chains, debug_stream_trace, debug_stream_trace_abort.
- No automatic timeout: a long chain shows progress with Stop Recording and Abort Operation. At most 50,000 records are kept; banners say when values were dropped, a parallel chain ran sequentially, or breakpoints were off.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
Adds Generate Commit Message with Agent above the commit message (Commit tool window and Commit dialog). It sends the included changes (trimmed to a budget) and the last 5 commit subjects to the agent of the project's latest task, on that task's model and in a mode that asks before editing or running anything, as a short quiet task deleted afterwards. The reply replaces the message only if the draft and the included files are unchanged; otherwise a notification says so. Without an installed agent, a notification points to Settings > Agents.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
File History offered the repository-wide Git Log actions only. Its commit menu now has Show File at Revision (the file as it was in that commit, following renames through historyPath), and Compare with Local / Compare Versions diff just that file instead of the whole tree. ⌘/Ctrl+D in the File History list diffs the selected revision of the file against its parent, or the two selected revisions, without changing Show Diff elsewhere.
Co-authored-by: <{
"message": "Not Found",
"documentation_url": "https://docs.github.com/rest",
"status": "404"
}+@users.noreply.github.com>
xterm.js 6.0 wraps pasted text in bracketed-paste markers without escaping an ESC inside it, so a pasted ESC[201~ ended the paste early and the rest was typed, Enter included. While the shell has bracketed paste on, each ESC in pasted text is shown as U+241B (as xterm.js 6.1 does); a Ctrl+V whose text holds an ESC is claimed before xterm's own listener. Covers the terminal tool window, embedded terminals and drops.
Co-authored-by: rohitdanda <10144914+rohitdanda@users.noreply.github.com>
- New Task page: a narrower box with room for a few lines, a lighter heading, and the
project, checkout and branch as one quiet line under the box. The agent shows as its
icon (name and version in its tooltip); the probe line shows only when the agent fails.
- Composer: one + menu for referencing a file, attaching an image and MCP servers; the
MCP popover opens from it or /mcp. The current-file chip turns IDE context on and off.
- Model and effort: the model button opens one searchable list of model names (pin on
hover, context size above it); effort is its own picker beside the model.
- Session: the status line moves under the box as one line — branch and worktree on the
left, changes, context use, time and terminal on the right; "Ready" is no longer shown.
With Highlight hyperlinks on, terminal URLs only got their underline from xterm's link provider while hovered, so they looked like plain text. Visible URLs now get xterm decorations in the theme's link colour with an underline (viewport-only scan coalesced per frame, rebuilt on theme or setting changes, off in the alternate buffer), in the main and embedded terminals. Link hit ranges were computed from string offsets, so wide CJK characters before a link shifted its hover and click area; both now use cell columns.
Co-authored-by: wildcatDownstairs <26499420+wildcatDownstairs@users.noreply.github.com>
Codex keeps its own project trust decision, so a project trusted in the IDE could still have its project-local .codex configuration (MCP servers, hooks, exec policies) disabled, and the Codex adapter dropped Codex's configWarning and warning notifications. Those warnings now reach the task, including ones sent before a session exists. When Codex reports the current project's .codex layer disabled, starting, resuming or forking a session first asks Trust Project in Codex / Keep Untrusted; only an explicit confirmation writes that project's trust entry through Codex's config API (with the user layer's version), Full Access doesn't bypass the question and side chats deny it.
Co-authored-by: YannABC <9603157+YannABC@users.noreply.github.com>
MCP servers can be configured at user and project level, imported from pasted JSON, tested with their tool list shown, and discovered from other agent CLIs' config files. Discovery reports only env and header key names, never values, and importing is explicit; external config files are never modified. Testing a server requires a trusted project. Each agent session can turn individual servers on or off from a picker; the overrides are stored with the task and applied when the session opens.
Closes#3608
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
A commit row closed its card on any capture-phase scroll so that scrolling the log, which moves the anchor, hides it; scrolling a long message inside the card fired the same listener. Scrolls inside the card no longer close it; the log and elsewhere still do.
Co-authored-by: wildcatDownstairs <26499420+wildcatDownstairs@users.noreply.github.com>
Expanded process runs and turn folds read nearly flat: the guide line sat left of the summary chevron, nested chevrons indented by only a few pixels, and tool rows without output yet had no chevron slot, so their icon jumped right once output arrived. The guide now drops from the parent chevron's centre, tool calls after a thought indent under it, expanded output indents with its row, and output-less tool rows reserve the chevron's width.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
Inlay hint requests only checked whether the document had changed, so an older response (e.g. for the previous viewport) could replace newer hints, or restore hints after they were disabled. A refresh or destroy now bumps a generation; work that is no longer wanted is skipped after document sync, before applying results and before marking the server unavailable.
Co-authored-by: amphetamlne <55946742+amphetamlne@users.noreply.github.com>
The list gutter in agent replies was a fixed 20px and markers sat 6px left of it without nowrap, so zoomed-in or two-digit ordered markers (10.) were cut at the row's left edge once content-visibility painted the row clipped, or wrapped under overflow-wrap: anywhere. The indent, marker gap and task-list pull-back are now in em (wider for ordered lists and for lists of 100+ items) and markers don't wrap. Adds an e2e case with zoomed single-, two-digit, bullet and task-list markers in measured rows.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
Rerun Failed Tests for several Go packages and compound configurations start runs concurrently. Each run picked a finished tab before awaiting the editor save and took it only afterwards, so two runs could pick the same tab: one replaced the other and its process never started. A tab a starting run is about to take, or was asked to reuse, is now held until it is taken.
Also fixes the agent model e2e helper: a model row's name now includes its Configure button, so the option is matched by its leading name.
The composer model picker becomes a two-page popover for every agent: page 1 lists models with search, a Pinned section, usage multiplier and description, and a summary bar for the active model; page 2 configures a model with Back, Reset to defaults, Pin, a segmented effort control and the cost category / multiplier from ACP _meta (copilotPriceCategory, copilotUsage) where present. Option defaults are kept across same-model updates for Reset. A contextTier / context_tier option is handled as per-model like effort. Agents with no effort, model options or cost metadata keep the plain select.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
The New Task page reused a probe cached for the whole session, so an updated CLI kept showing its old version. The selected agent is probed again (initialize only, no session) when the page is shown or the window regains focus, without overlapping probes.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
The completion documentation was drawn inside the lookup popup, which clipped it or let it cover the options. positionInfo now places it beside the list on the side with more room, top edges aligned, with capped width and height.
Co-authored-by: zhzhsx <1379025+zhzhsx@users.noreply.github.com>
The row under a working conversation showed a constant Working…. It now follows the turn's latest entry: Thinking…, Writing…, Reading…, Editing…, Running a command…, Running a subagent… and so on, with Working… for the rest.
Settings › Agents › Working Status adds playful words per activity, off by default. When on, they rotate from the start or after a configurable quiet time (0–600 s). Each word stays 8 s, or at least 3 s once the activity changes. Screen readers keep hearing the plain status. Users can replace or clear each activity's words.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
Toolbar and tool window icon buttons (IconButton) paint --bg-toolbar-hover across their hit target again, so hovering shows the clickable area. Disabled buttons stay transparent; pressed, selected, focus ring and icon colours are unchanged. The hover colours match the toolbar tokens already used for the pressed state in Islands Light and Dark.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
Editor font zoom scaled the Markdown preview's text but left Mermaid diagrams, embedded SVGs and raster images at 1x: Mermaid writes a pixel max-width on its svg and images were only capped at 100%. The preview now publishes --markdown-zoom-scale and records each diagram's and image's intrinsic width (SVG viewBox/width/height, data-URL SVG headers, naturalWidth), sizing them as min(base, 100%) * scale so they follow zoom in/out/reset while staying clamped to the pane at 1x. Agent chat diagrams keep scale 1; Mermaid full-screen and PNG export share the same size helper.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
Find in Conversation matched text in user-select:none chrome (summaries, roles, times), which WebKit paints over the next selectable row, and painted matches inside closed <details> that still report client rects. Those subtrees are left out of the matches, and matches in closed <details> count but are only highlighted once stepped into (which opens them).
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
Off-screen conversation rows kept their old intrinsic heights when the column switched between 900px and full width (the stream's width doesn't change), so they are re-measured when the setting changes; Agents › Chat display preferences are client settings, so remote windows no longer overwrite them with the server's.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
The hover tooltip stayed open and moved with the text while typing, covering the completion lookup. It now closes when the document or the selection changes, as Quick Documentation on hover does in IDEA.
Co-authored-by: zhzhsx <1379025+zhzhsx@users.noreply.github.com>
Following #3452, the remaining line-marker kinds kept a gutter column each (a Spring controller showed its mapping icons in a second column beside the run icon). Scripts and Docker/Compose targets join the run icons in the shared icon area (hidden with them), Spring mappings and endpoints follow, then plugin line markers, gutter actions and indexed mappings (now 14px like the others). Presses go to the pressed icon's source.
Settings › Agents › Chat › Conversation gets an opt-in toggle (agents.fitWindowWidth, off by default). When on, the conversation stream, composer, question card, user message bubbles and the New Task page drop their max-width caps and follow the pane width; when off, the readable 900 px column stays.
Co-authored-by: dengliu <5108773+dengliu@users.noreply.github.com>
Cycle Agent Effort (Option+T / Alt+T in every built-in keymap) cycles the current model's advertised effort levels while focus is in New Task or a writable task conversation, through the existing config setter; the message, caret and focus are kept and Send stays available.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
The Changes pane beside a task's conversation was fixed at 280 px. Its left edge is now a separator that drags (or moves with ←/→ when focused); the width applies to every task, is remembered across restarts, and always leaves the conversation 240 px (half of a narrower session). In the tool window the Changes view fills the window as before.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
The composer's 44 px minimum height sat on CodeMirror's scroller, while the text inside stopped at 32 px (its min-height: 100% doesn't resolve against a min-height parent), so a click in the strip below focused the bare scroller and showed no caret. The minimum height now sits on .cm-content.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
Focus Agent Input (⌘L on macOS, Ctrl+L on Windows / Linux, in every built-in keymap) returns focus to the current writable task's message input from its conversation, in an editor tab or the tool window, keeping the draft, caret and selection. It applies only while focus is inside an agent task, so the keys keep their editor and terminal meaning elsewhere; the binding is under Settings › Keymap › Agents.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
Opening a conflicted package-lock.json in the merge editor froze the app. The Your/Their panes diffed each side against the base with @codemirror/merge's default config, which gives up past ~16 KB of changed span and returns one chunk from the first change to the last, so each pane rendered the whole base as one deletion widget; syncMergeScroll ran an unbounded character diff on first scroll. Both now use the IDEA line comparison the diff viewer uses (ideaDiffConfig).
Co-authored-by: charlie-morrison <271589886+charlie-morrison@users.noreply.github.com>
With tabs placed left or right, the column could not be resized and was never narrower than 140 px. It now fits its longest name up to a limit (260 px by default); dragging its inner edge, or ←/→ on the focused edge, changes the limit for every editor group and it is remembered. The editor keeps at least 60% of each group.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
The New menu offered every language template in every project. It now uses the detected type of the project owning the selected folder for language-specific templates (Rust, Python, Go, PHP, Java/Kotlin); generic and web templates (JavaScript, TypeScript) stay in every project, as in IntelliJ IDEA Ultimate, and From Template still lists the whole catalog. Tomcat Context Descriptor is offered only when the project has a Tomcat run configuration.
Co-authored-by: shohruh-genius <7658451+shohruh-genius@users.noreply.github.com>
Right-clicking a link in Agent Task chat (and other readable panes using the shared Copy menu) only offered Copy, which copied the selected label. Links now also get Copy Link Address, as in browsers; Copy keeps copying the selection, and in-page anchors get no link item.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
Starting or forking a task in project A and switching to B before the reply arrived let A's late reply insert its task into B and replace B's New Agent Task page. startTask and forkTask now check the connection generation they started under before handling the reply or error, and again after Home closes, covering reconnects and A -> B -> A.
Co-authored-by: LunarFeller <269972968+LunarFeller@users.noreply.github.com>
The Language Servers page now groups tools by language (#3272). The search, settings and settingsPaneLayout specs still looked for the old single tool list, and the details pane directly under the body, so they failed on main.
Co-authored-by: shohruh-genius <7658451+shohruh-genius@users.noreply.github.com>
A cold start can take longer than 30 seconds before the workbench is up
(the first load of a development build, a slow machine), and the failure
page then stays over a window that did start.
Send and Start cleared the composer when the agent acknowledged the prompt, wiping text or attachments added while waiting. The acknowledgement now clears the composer only if it is still the submitted revision in the same project connection.
Co-authored-by: LunarFeller <269972968+LunarFeller@users.noreply.github.com>
Selection and chart means stopped after 12 fractional digits even when the division terminates, which could also change an integer mean (1220703130000 instead of 1220703125000). The shared decimal division now continues until the remainder is zero when the reduced denominator has only factors 2 and 5; repeating quotients keep their rounding.
Co-authored-by: LunarFeller <269972968+LunarFeller@users.noreply.github.com>
CodeMirror anchors the completion popup at the start of the word being completed, but each row starts with a kind icon, so labels sat one icon column to the right of the text. The popup is offset left by the row's leading inset (border, row margin, padding, icon and gap), kept in custom properties that the row rules use too, so the labels line up with the word.
Co-authored-by: zhzhsx <1379025+zhzhsx@users.noreply.github.com>
CodeMirror rebuilds the completion list by removing the old <ul> and appending a new one, so the advertiser bar (insert/replace hint and Sort by name), which was only added when missing, ended up above the list after the first refresh. An existing bar is now moved back behind the list.
Co-authored-by: zhzhsx <1379025+zhzhsx@users.noreply.github.com>
A metadata reply that arrived after its data source was deactivated, deleted, edited or after a project switch could bring back the Connected indicator and the cleared schemas, tables, columns or completion, and an obsolete chain could issue further requests. Requests now carry project / source / target lifetimes (requests.ts); forget() starts a new lifetime, and late replies, introspection steps and follow-up requests of an ended lifetime are dropped. Completed writes keep their result but no longer revive an old connection.
Co-authored-by: LunarFeller <269972968+LunarFeller@users.noreply.github.com>
Git Branches kept the groups and folders the user expanded or collapsed only for the session. The toggled keys, per popup kind, are now stored per project root in localStorage and loaded when the popup opens; untouched keys keep their defaults.
Co-authored-by: formiat <29867335+formiat@users.noreply.github.com>
A second click on a Git Log toolbar button reopened its menu (or left the Branch popup open): the menus opened as the global context menu at a point, which took the press on the button for an outside click. They now open as a Menu anchored to the button that toggles on the same button, covering the filter bar, Graph/View Options and the commit details menus; the Branch popup toggles too.
Co-authored-by: formiat <29867335+formiat@users.noreply.github.com>
The agent composer is now a CodeMirror editor: the text stays plain and is what's sent, while @ file references are drawn as chips (icon, name, lines) that move and delete as one piece, open on click and preview on hover; Backspace after a chip opens it as text. Quotes show as blocks, and sent messages get the same chips and quote blocks. The @ completion lists open editors first, highlights matches and shows the folder; the focused editor's file is offered above the text, a pasted project path becomes a reference, and dropping files shows a hint. Lines copied from an editor paste as a reference, terminal output as a terminal block, and code blocks show as cards. The reference trimming regex avoids lookbehind for WebKit before 16.4.
WebKit doesn't focus a clicked button, so focus moved to the editor area outside the merge editor and cleared the focused merge editor; Apply, Cancel, Accept Yours/Theirs and the toolbar actions then did nothing on a mouse click. The merge editor root is now focusable (tabindex -1) so a click inside it keeps focus there.
Tab on a database result tab stepped through every toolbar control (7 to 11 stops) before reaching the grid, and the grid then had no active cell. Now:
- a new WAI-ARIA Toolbar (src/ui/Toolbar.tsx) makes the result, plan, Output and pager toolbars one Tab stop with roving focus on arrows/Home/End;
- Tabs gets an optional focusContent prop, called on plain Tab, so Tab from a result tab focuses the grid (or plan tree / Output log);
- the grid selects its first cell when focused from the keyboard, Tab / Shift+Tab move between cells row by row and leave the grid past either end, and Tab in the cell editor commits and moves on.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
The editor, tool windows and database results painted their selected tab with the same accent, so it wasn't clear which part had focus. A focus-scope tracker now marks the editor pane, each tool window and the results view that contain the focused element; outside the focused scope the selected tab uses the existing --bg-tab-selected-inactive / --border-tab-selected-inactive tokens. Scopes are opt-in, so tab strips in dialogs and popups stay accented, and focus moving into a popup, menu or dialog keeps the current marks. IntelliJ theme import maps the inactive underlined tab colours to the same tokens.
Co-authored-by: Mr0Bread <55754540+Mr0Bread@users.noreply.github.com>
Pressing Cmd+Down or F4 in a focused revision diff did nothing when the working file was found after focus entered the pane: the keyboard dispatcher kept the navigator object registered before that lookup, while the diff replaced it. The diff now keeps the registered navigator and refreshes its handlers in place, so Jump to Source opens the working file at the mapped caret.
Co-authored-by: shohruh-genius <7658451+shohruh-genius@users.noreply.github.com>