* fix(sidebar): unstick interrupted panel drags so the bottom panel keeps tracking layout
A width/height/corner drag whose pointer stream ends without pointerup
(pointercancel, or pointer capture lost prematurely — e.g. a trackpad
gesture interrupting the drag) left draggingWidth/draggingBottom/
draggingCorner stuck true: onPointerUp guards on hasPointerCapture and
never ran, and no onPointerCancel/onLostPointerCapture handler existed.
The stuck drag state pauses center-column measurement (draggingRef)
permanently, so centerRect goes stale and the bottom panel keeps frozen
inline left/right edges — it no longer expands when the right sidebar
collapses or the app rail toggles until a full reload.
Fix: mark each drag uncommitted on pointerdown and committed in
pointerup before releasing capture; the new pointercancel /
lostpointercapture handlers abort an uncommitted drag by reverting the
DOM and layout variables to the store's committed sizes and clearing
the dragging state, which resumes measurement.
* fix(sidebar): clamp abort-drag rollback to collapsed/narrow geometry (P2 review)
abortDrag wrote the committed width even when the right panel is
collapsed or the viewport is narrow, where the layout-push effect had
already written 0 to --dsh-sidebar-width — an interrupted drag then left
the app shell squeezed by an invisible sidebar. Mirror the layout
effect's clamping: write 0 when narrow or the panel is closed.
---------
Co-authored-by: Menghuan1918 <menghuan2003@outlook.com>
The close settled on auxclick delivery, which the browser redirects to the
nearest common ancestor of the press/release targets when they differ — any
drift, middle-button autoscroll scroll, or tab-list reflow between press and
release swallowed the close (the strip hides its scrollbar, so Chrome's
autoscroll indicator is inert here and only added confusion).
Now the press target is recorded on middle mousedown (preventDefaulted to
disarm autoscroll at the source) and the close settles on the first middle
mouseup OVER that same tab — release-position semantics matching VS Code
(microsoft/vscode#101028) and user expectations for Chrome tabs
(crbug/40679924): pressing on a tab and releasing elsewhere cancels.
Left-button activate/drag and wheel scrolling are untouched
(tests/tab-bar-middle-click.spec.tsx).
Reads the aionui-panel settings namespace's rightPanel provider choice: while
it resolves to 'aionui-panel', the whole sidebar does not mount (host
settings.get reports externalDisable, the client mount gate skips mounting
and the link/open-path/turn-tail takeovers step aside). Live via the
settings/document-updated push (best effort; deployments without the
'remote' service fall back to boot-time evaluation). No new user-facing
setting — the aionui panel card remains the single source of truth.
* feat(editor): merged explorer mode — path input, docked file tree, global name search
New editorExplorer pref (default on, editor card gear): the editor tab gains
a path-input header and a toggleable right-docked tree panel (per-tab state
in tab.meta.treeOpen; file tabs default closed, the seeded path-less home
tab defaults open). Fresh sessions seed that empty editor tab instead of
the explorer tab when merged mode is on; the standalone explorer tab stays
in the + menu unchanged.
- Extract FileTree.tsx from ExplorerView (shared by both hosts)
- Host fs.search route: recursive case-insensitive name search rooted at
the session cwd, budget-capped (200 matches / 100k visited), skips .git,
never descends symlink dirs
- EditorHost reads the pref reactively (boolean snapshot) and keeps the
pre-merge layout verbatim when off
* feat(settings): select row type for declarative settings; editorExplorer as iconed select
- SidebarSettingToggle gains type 'select' with options {value,title,desc,icon}
and multi (default false; stores the picked values as an array in options
order). Any option carrying an icon renders big-icon option cards in the
dropdown and the icon in the closed anchor; icon-less rows are one-line text.
- SelectRow in SideCardSection rides the primitives Menu (portal); single-pick
commits and closes, multi toggles and stays open. Wired for both host-prefs
toggles and plugin-owned pluginToggles via the new onSelectValue seam.
- The editorExplorer merged-mode pref is now an iconed select (merged vs
separate) instead of a switch; the stored value stays boolean.
- Advertise the capability as 'settingSelect' in SIDEBAR_FEATURES.
* feat(editor): unify explorer into the editor surface, resizable tree dock, hoisted toolbar
- One tree implementation, two presentations: the shared TreePanel (global
name search + refresh + FileTree) docks as the editor tab's right panel
and — in merged mode — fills the explorer tab's window; the pref-off
explorer keeps its legacy header shell.
- The docked panel drag-resizes from its left edge (160-480px clamp),
persisted as tab.meta.treeWidth alongside treeOpen.
- The text editor's preview/edit/save toolbar moves into the path-input
header row: FileViewerProps gains internal toolbar fields ('host' mode
skips the viewer's own row, reports EditorToolbarState and registers
EditorToolbarControls); non-merged mode is byte-identical to before.
* feat(editor): single files window with in-place switching; explorer descriptor removed
- editor tab replaces explorer on the + menu (order 10, Files); persisted
explorer tabs migrate to the editor home tab via sanitizeState
- editorExplorer now only controls open behavior: merged = in-place
updateTab, independent = openSidebarFile; both modes share the merged
chrome (path input + hoisted preview/edit/save toolbar + tree toggle)
- file tree context menu gains 'open in new tab' / 'open to the side'
(mintTabId avoids the id safety net; insertLeafAt splits right)
- tree dock resize uses pointer capture (touch-action none, 6px handle)
- defaults: openByDefault off, sidebar width 35%
- file tree clips horizontal overflow (no sideways scroll)
- e2e: mount lane arms the editor-chunk wait before goto and clicks the
tree row's left edge (hover @-reference button sits mid-row); drag lane
expands the panel first (collapsed push writes 0px), waits for the
slide-in to settle, and picks the leftmost col-resize strip; 'Files'
title assertions require draggable to skip the bottom panel's welcome
cards
* feat(editor): split mode turns the path-less window into the standalone explorer
In independent (split) mode a path-less files window renders ONLY the
tree panel (search + FileTree, full-window form) — no path input, no
hoisted toolbar, no tree toggle. File clicks open per-path editor tabs
(full chrome, tree docked closed by default) through openSidebarFile,
unchanged. Merged mode is untouched (chrome + docked tree, in-place
switching). TreePanel's previously reserved 'full' prop now has its
consumer.
* catalog: add dsh-video-preview to the file-previewer recommended plugins
Adds the video preview plugin (dsh-video-preview) to the "add preview plugin" modal catalog, with zh/en descriptions. Data integrity is guarded by tests/plugin-list.spec.ts.
* test: make add-plugin-modal window.open count catalog-driven
The test pinned the viewer catalog to a single entry (2 "Open:" buttons).
Adding a second recommended plugin (dsh-video-preview) renders 4 buttons
and broke the hardcoded assertion. Derive the expected count from
builtinViewerPlugins so adding/removing plugins keeps the test honest.
- Sidebar.tsx: useLayoutEffect for the layout-push variables so cleanup+setup
land in the same pre-paint frame (no full-width flash on drag release);
applyDrag writes the width only while the panel is open (a closed right
panel no longer squeezes the conversation when dragging the bottom height);
flush the final drag frame before stopDragScheduling on all three handles;
sync centerRect.right on width/corner release so the bottom panel does not
jump.
- index.ts/pty-manager.ts: close a session's UI terminals immediately on
session/disposed (they used to linger until reconnectGraceMs expired).
- sidebar.module.css: no highlight on the resize handles while dragging
(local preference; the corner hover hint stays).
* feat: configurable terminal shell with Windows pwsh auto-probe
Windows previously hardcoded powershell.exe (5.1) in defaultShell(), ignoring PowerShell 7 installs. Resolve the shell in priority order: explicit shell config -> DSH_SIDEBAR_SHELL -> PATH/known-dir probe for pwsh.exe -> inbox 5.1 fallback. The same resolved shell feeds both the UI terminal tabs (PtyManager) and the model-facing terminal_* tools (AgentPtyRegistry).
defaultShell() now accepts injectable platform/env/exists options so the Windows chain is unit-testable on the ubuntu CI, which never executes win32 branches. The probe also checks ProgramW6432 for 32-bit Node, and machines without PowerShell 7 keep the 5.1 fallback.
Refs: #46
* fix: trim resolved SHELL and document ProgramW6432 in resolver contract
- POSIX SHELL is validated with trim() but was returned untrimmed; trailing whitespace could leak into the spawned executable path. Return the trimmed value like the explicit/config and Windows/env branches.
- ShellResolutionOptions.env comment now lists ProgramW6432, which the Windows probe prefers over ProgramFiles on 32-bit Node.
- Test pins the SHELL trimming behavior.
Map the .cs extension to the built-in csharp StreamLanguage from
@codemirror/legacy-modes/mode/clike (already a dependency, no new packages),
so .cs files open with syntax highlighting in the sidebar editor.
The open-path interception reroutes EVERY chat-side file open into the
sidebar editor, including the stock "Show in folder" gesture: the
ui-deliverables row passes "." (resolved to "<cwd>/." by the chat view),
so clicking it opened a bogus editor tab for the workspace directory
(host error: '"..." is a directory').
- openpath-intercept: detect the folder-reveal gesture ("." / "./" /
"<cwd>/.") and route it to a new revealInExplorer dep instead of the editor.
- intercept: revealInExplorer switches to the explorer tab (expanding the
panel), expands the produced files' parent directories, and highlights /
scrolls to them; the intercepted produced-files row now also renders its
own "Show in folder" button when files are hidden.
- produced-files: selectProducedFiles reads the engine Turn data
(owner.turn.data.get("deliverables")) — the same source ui-deliverables
uses — with the node replica as fallback. The old code read owner.nodes,
which the real TurnTailOwnerProps ({turn, seq, openFile}) does not carry,
so the takeover never claimed the chain.
- state: transient `revealed` set (never persisted) plus a revealPaths
reducer that expands ancestor directories.
- ExplorerView: highlight + scroll revealed rows; Sidebar/tabs/service
thread `revealed` through TabComponentProps.
- tests: folder-gesture routing, engine-data selector, wiring seat, zh/en copy.
Add a side-card preference 'browserAllowedLoopback' (comma-separated
host or host:port entries) that lets the sidebar browser navigate to
explicitly trusted loopback addresses. Bare hosts (localhost,
127.0.0.1) match every port; host:port entries match exactly.
Allowlisted local addresses additionally get the 'allow-same-origin'
sandbox token in the iframe: local dev servers (Vite etc.) need a real
origin for their module/HMR/fetch pipeline and would otherwise render
blank in the opaque-origin sandbox. The page stays cross-origin to the
GUI and to every other site.
The server-side browser.probe route mirrors the same allowlist so
probed local servers are not rejected as 'local addresses are not
probed'.
Terminal tabs resolved the shell from the dsh process environment only,
falling back to a hardcoded /bin/bash — service-managed launches without
SHELL therefore ignored the account's login shell (e.g. zsh). They also
spawned with an empty argv, so even a resolved zsh ran as a non-login
shell and skipped ~/.zprofile.
defaultShell() now falls back to the account's login shell (passwd) after
$SHELL, and both registries spawn POSIX shells with -l so the tab behaves
like a terminal emulator (loads ~/.profile / ~/.zprofile). Windows keeps
powershell.exe with no login flag.
Fixesomdsh-dev/DSH-better-sidebar#68
Some servers (e.g. aliyun consoles) answer HEAD without the
X-Frame-Options / Content-Security-Policy headers that only their
GET response carries. The embeddability probe then wrongly reported
the site as embeddable, leaving the plain iframe to surface the
browser's misleading "refused to connect" error (the site actually
forbids framing via frame-ancestors).
Retry once as GET when both embed signals are absent (the body is
discarded — only headers matter), mirroring the existing 405/501
HEAD fallback. The client then shows the friendly embed-refusal
panel with an "open in browser" action instead of the misleading
connection error.
Atlas Zeng