disposeAll() only requests the pty kill; the bash child exits
asynchronously via onExit, so deleting the temp HOME right after
races the still-running shell (it may be writing history files) and
fails with ENOTEMPTY on CI (seen on both #114 and #120).
Wait for the spawned handle to report exited (bounded 2s) before
removing the directory, with a short retry loop as a fallback.
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.
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
Two failure modes of a sidebar render crash, fixed together:
1. Layout-push leak (root cause): the layout-push effect wrote
--dsh-sidebar-width/--dsh-sidebar-height on document.documentElement
without a cleanup, so ANY unmount of the Sidebar (error-boundary swap,
plugin disable, HMR) left the variables behind and layout.css kept
squeezing #root with a stale margin — "the sidebar cannot be hidden"
until a full reload. The effect now removes both properties on cleanup
(the CSS fallback 0px restores the layout).
2. Crash containment (UX): the root error boundary previously replaced the
WHOLE sidebar — toggle cluster included — with an error strip on any tab
crash. A new shared RenderBoundary (src/client/RenderBoundary.tsx) now
scopes containment per tab: a crashing viewer/editor shows a dismissible
strip inside its own pane while the toggle cluster, the other tabs, and
the panel stay alive. The descriptor is rendered as a real element
(createElement) so even top-level component throws land in the boundary,
not in TabContent's own render. The root boundary remains as the last
resort for Workbench/shell errors.
Tests (tests/sidebar-crash.spec.tsx, jsdom + real Sidebar/store harness):
- unmount clears both CSS variables (red before the fix: 307px leaked)
- a crashing tab is contained: strip in-pane, collapse button alive,
layout push still live
- the retry button recovers a tab whose crash has since been fixed
Full suite: 425 tests pass; tsc --noEmit clean.
node-pty@1.1.0 builds the spawn-helper executable only for OS=="mac"
(binding.gyp); linux uses forkpty directly and ships no helper at all.
The test's non-Windows premise was wrong for linux, and no candidate
exists there to assert on.
- checkout with fetch-depth: 0: the git log paging test needs the full
repository history (default shallow clone yields one commit)
- configure a global git identity on the runner: src/git.ts deliberately
never sets one and revert/cherry-pick rely on ambient config
- tests/unit.spec.ts: node-pty@1.1.0 ships a prebuilt spawn-helper only for
darwin/win32; on linux node-gyp compiles it into build/Release. Align the
assertion with ensureSpawnHelper's candidate list instead of hardcoding
the prebuilds/ path (the Linux test run was red for this reason)
The panel width lived in the per-session state (dsh-sidebar:v1:<id>), so a
drag in one conversation never carried to another: switching conversations
snapped the panel back to that session's own width. The width is a layout
preference, not session content, so it now persists under a cross-session
key (dsh-sidebar:v1:width):
- every persist writes the current width to the global key;
- loadState prefers the global width over a session's persisted one;
- a cache-hit session switch adopts the global width.
Backward compatible: with no global key present (first run / older data)
the behavior is unchanged. Adds a regression test covering drag-in-A ->
switch-to-B and drag-in-B -> switch-back-to-A.
ymac