Restored SSH tabs reconnect after app restart but skipped the host startup command because the boot effect closed over a pre-vault host snapshot. Prefer hostRef for SSH when the dial target matches.
Fixes#3905
A pending tool switch stored an intermediate focused-tool map in the ref.
An urgent open on another tab then updated only state, and the identity
guard never copied that tab back, so a later click on the old tool was
ignored. Committed entries now catch up without dropping a write that is
still ahead of the last render.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
A keep-count change during victim drain was still applied from the count
captured when the leave started, so a zero setting could soft-hide the
leaving tab with no later timer to correct it. Retention now uses the live
count, and an already soft-hidden tab re-enters hibernate when the limit hits
zero.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(terminal): soft-hide the tab that just left when hibernate keep is full
Evict the oldest soft-hidden session to a full snapshot, and keep the
session the user just left on hideRuntimeOnly. Disconnected sessions,
alternate screen, and inline images keep their existing branches. Wake
replay order is unchanged.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(terminal): full-hibernate when keep-count eviction is refused
A soft-hidden session with inline images consumes the eviction request
and stays mounted. Soft-hiding the tab that just left anyway let that
protected session pin the oldest slot, so every later tab was retained
too. Walk oldest-first, keep the leaving tab soft-hidden only after one
victim actually leaves the set, and full-hibernate it when none can.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(terminal): retain a leaving tab only after eviction hibernates
Waking a soft-hidden victim clears its marker before fullHibernateRuntime
finishes. Treating that as success soft-hid the tab that just left even
when drain, snapshot, or alternate-screen failure restored the victim.
Await hibernate completion outside the pure keep-count helper, and put a
failed victim back in the soft-hidden set so it stays a candidate.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(terminal): reserve hibernate slots outside domain policy
Keep the keep-count decision pure. The application walk reserves the
leaving tab before awaiting a victim, rechecks that slot before
soft-hide, and does not restore a victim that became visible or had
hibernate turned off while the upgrade was in flight.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* docs(terminal): keep activeTabId in workspace ctx equality
Solo tab switches already leave TerminalPanesHost equal when the sessions
ref is unchanged. Reveal flush and repaint still run on the visible pane,
so removing activeTabId would not move the measured settle time.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* docs(terminal): use an ASCII arrow in the workspace ctx comment
AGENTS.md keeps comments ASCII-only. The keep-activeTabId note used a
Unicode arrow.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* perf(terminal): commit side-panel tool switches in one render
Icon clicks write the focused tool and the layout together through
selectSidePanelTool. External open-tab writes still reconcile in an effect,
and that effect skips when the layout is already the selection.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(terminal): merge pending side-panel writes across urgent renders
Icon switches inside startTransition used to publish absolute maps from
refs, and the next render copied the last committed state back over
those refs. A second tab's switch could then replace the queued map and
drop the first tab's selection and split. Patch each tab onto the queued
map, and leave a pending ref in place until that render commits.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Split the notes tree into a memo pane whose equality tracks notes, groups,
selection, editing, and expansion. Hide still flushes the draft through the
existing ref and leaves the editor mounted.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Notes and SFTP already park behind a hidden class. History returned null,
so each Notes to History switch remounted the virtual list. Keep that tree
mounted, subscribe to shell history only while it is visible, and remeasure
the list when it is shown again.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(sftp): preserve successful uploads with backup cleanup warnings
Warn only on confirmed backup cleanup permission refusal after final publication. Preserve actionable warnings across transfer events, history restoration and completion ordering without masking deletion or commit failures.
Narrow follow-up to #3830; related #3879.
* fix(sftp): retain folder cleanup warnings before child completion
* feat(serial): support local Windows named pipe consoles
* test(serial): synchronize pipe acceptance and cover saved-host form
* fix(serial): preserve named-pipe paths with spaces when editing hosts
* test(serial): verify Windows pipes through native Electron IPC
* test(serial): keep Electron listener callbacks in renderer context
* test(serial): cover busy pipe cancellation and bounded slow-reader writes
* test(serial): construct single-instance fixture pipe path correctly
Preserve the original PR #3885 commits and integrate latest main.
Resolve local shell defaults, acknowledge committed workspace insertion, and preserve background readiness codes.
Fixes#3886
Sub-threshold tab pointer moves no longer lose activate: dragstart is not cancelled; reorder still arms at ≥4px; mouseup/dragend activate when click was cancelled and no reorder started.
Fixes#3886
Primary-button mousedown on plugin/editor/session/log-view top-tab close buttons now preventDefault+stopPropagation so HTML5 drag on the parent tab cannot swallow the click.
The list-models pipeline already served LKG on failure; this wires the
rest of the LKG-first, manual-refresh policy end to end.
Main: the list-models IPC accepts a force flag — manual refresh always
skips the in-memory cache and the persisted entry and goes live, and a
forced success overwrites the persisted catalog.
Renderer: the panel passes force only from the retry banner and the
picker's refresh button; panel-visibility and send-preflight loads stay
non-force and replay the renderer cache, so with a persisted catalog
main answers them without spending the live-fetch budget.
Picker: a footer refresh icon-button next to the status chip, visible
whenever any fetched catalog (live or last-known-good) or the degraded
warning is present; the spinner row remains the in-flight feedback.
Tooltip localized as ai.chat.refreshModelList in en/es/ru/zh-CN/zh-TW.
Fresh installs are unchanged (no persisted entry → live fetch as
before); existing users stop seeing jitter warnings on panel open and
pick up new CLI models via the refresh button.
A failed live catalog fetch serves the persisted last-known-good catalog
with a warning, and every TTL-expired automatic load re-attempted live —
so a backend whose discovery keeps failing re-armed the renderer's
「暂时无法更新实时模型列表」 banner forever, even though the persisted
models worked and selecting one never touched the banner state.
Serve the persisted last-known-good catalog without a live attempt on
every non-force load (the LKG-first policy, decided at the single
fetchSdkModelCatalog choke point): automatic loads now return a clean,
warning-less serve, so the next automatic load after the renderer's TTL
clears a stale banner. Warnings are reserved for an explicit force
refresh that fails, and for a first fetch with neither cache nor
persisted entry. The persisted current-id remap moves into a helper
shared by both serve paths.
Do not fall back to a fresh GUI login when a dedicated connection rejects.
Capture capability selection before awaiting so source cleanup cannot change
the failure policy. Preserve ordinary reusable-channel fallback and cover
both paths with real bridge regression tests.
Preserve PR #3875's original author commits and scoped TERM-SSHD transfer policy.
Copy persisted filename encoding into dedicated file connections. Capture only
host identities actually accepted by the terminal route, including jump hops,
and fail closed on later unverified changes without orphan host-key prompts.
Add first-trust, changed-key, route-scope and encoding regressions.
Deduplicate terminal reveal and focus recovery while preserving changed-layout and explicit resume/fullscreen recovery. Add behavioral regressions.
Related to #3825; this rendering optimization does not establish that the original Windows/macOS issue is resolved.
A managed tool call that races bridge shutdown was still in flight, so
the socket error rejected it with "TCP connection lost". Managed
children do not reread discovery; report that race as not connected
and leave external in-flight wording unchanged.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
When NETCATTY_EXTERNAL_MCP_DISCOVERY_FILE is set, rpcCall recovers on demand via ensureConnected(): re-read discovery, reset TCP buffer, connect to current port, auth/verify with refreshed token. Concurrent callers coalesce. Close/error reject in-flight RPCs without replaying. Managed SDK children keep fail-fast.
Fixes#3871
* feat(shortcuts): toggle the host list with Ctrl/Cmd+Shift+B
Ctrl/Cmd+B already switches broadcast mode and stays available to a focused terminal. The new shortcut reuses the existing host-tree toggle.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(shortcuts): let the host list chord through Monaco and idle terminals
Editor focus inside Monaco can toggle the host list. The chord is consumed only when that toggle runs, matching broadcast.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* feat(terminal): split compose-bar send target from toolbar broadcast
Toolbar broadcast still fans out keystrokes, paste, and terminal-path
snippets. The compose-bar indicator is now a toggle for whether Enter
sends to broadcast peers or only the active session. Turning the toolbar
switch on or off resets that target to follow it; later flips stay
independent.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(terminal): keep active-only compose sends off the toolbar fan-out
The focused workspace executor omitted broadcast, so executeSnippetCommand
fell back to the toolbar switch and still wrote every peer. Pass
broadcast: false on that path.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* perf(vault): store connection-log replay blobs in IndexedDB
Move unsaved connection-log terminal replay out of the multi-megabyte
localStorage blob into one IndexedDB row per log id. Persist only the
changed ids, keep the newest 50, and signal other windows with a tiny
revision key.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): keep connection-log replay across boot races and IndexedDB failure
Re-read connection-log metadata after the IndexedDB load before pruning,
return the capture promise through TerminalLayer, and write the unpruned
log blob when the side store rejects the commit.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): keep unmigrated connection-log replays until IndexedDB commits
Copy legacy replay ids that are missing from IndexedDB before deleting
the old blob, and do not treat an uncommitted legacy map as the diff
baseline.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): migrate legacy replays without clearing newer IndexedDB rows
Insert missing legacy ids in one transaction instead of replaceAll, so a
replay committed after the empty read is not wiped.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): rebase the IndexedDB fallback onto the latest connection logs
When the side store rejects a close, attach this window's replay to the
connection-log blob currently on disk instead of writing the stale snapshot.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): rebase queued connection-log writes onto the latest logs
A persist task can run after another window changes the log list. Write
the rebased list instead of the snapshot captured when the task was queued.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): keep queued connection logs after a remote snapshot
A storage event can replace memory with another window's logs before the
queued write runs. Keep ids from the queued snapshot unless this snapshot
deleted them.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): restore only locally changed queued logs
A queued persist was putting back every id from the pre-queue snapshot,
which undid another window's delete of a log this recipe never touched.
Track the ids the recipe added or replaced, and reload IndexedDB during
boot when the revision key moves across the migrate await so that peer
replay is hydrated before this window marks the revision seen.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): rebase connection logs after the side-store await
A non-pruned write stored the pre-await snapshot and could replace logs
another window saved during the IndexedDB round trip. Rebase immediately
before that write, shrink the quota fallback until the newest replay fits,
and leave the revision unconsumed when IndexedDB cannot be read.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): ignore same-window connection-log echoes
The deferred localStorage notification replays this window's own metadata
write. Applying it replaced connectionLogsRef with an older blob and dropped
logs that were added or finalized while that write was still queued.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): do not restore a queued log deleted from memory
A queued add or update was written back when the id was already gone from
memory, which let another window keep that row after a later local delete.
Rebase now applies a changed id only while memory still has it.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Attach close preparation clears the popup session ref, then unmount
cleanup called closeSession. That emits netcatty:exit to the home
renderer and can kill the shared PTY. Skip that second close.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
A timed-out SFTP OPEN used to pin the connection forever when its callback
never arrived. Wait up to 2s, then force-clear the gate and admit a new open.
Late channels are still closed. After TIMEOUT or PENDING, skip the fastPut
reopen and upload on the shared channel.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
canRetryFromCheckpoint and the upload prefix digest missed ECONNABORTED,
socket hang up, ERR_SFTP_CHANNEL_CLOSED, and client-socket. Those Retries
restarted at byte 0 and stranded the staged .part.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
New-file uploads have a null existingMode, so restoreRemoteMode is a no-op.
Recheck the destination only after a real mode restore and keep one
pre-rename safety lstat otherwise.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>