Models pass query: "" on range reads. Accept that as no search so
vault_notes_get does not fail and retry.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
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>
* feat: open saved session logs from terminal toolbar
Add one-click access to the current automatic or manual session log, flush before opening, and reuse application selection with a remembered log preference. Preserve recording controls and current-session ownership across worker routing and reconnects.
Refs #3807
* fix(logging): preserve saved logs on failed reconnect and localize failures
Address both Codex review findings with generation- and identity-checked replacement cleanup and localized user-facing error handling.
* fix(logging): guard successful main-process session replacement
Add shared inactive-log cleanup to fallback start handlers with regression coverage. Registration activation follows separately.
* fix(logging): enable log lifecycle cleanup for main-process fallback
* fix(terminal): keep log cleanup and tuck open-log into More actions
Always end a raw session-log write stream on finalize, including when a
flush already disabled it. Keep immediate snapshot retries, and pause
them only while prepare is waiting so a disk error can be reported.
Show open, reveal, and open-with in the toolbar overflow menu when
logging is active, instead of a default toolbar button.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(logging): back off failed snapshot writes and drop retired log actions
Persistent snapshot failures retry on a growing delay instead of a tight
loop, and still flush immediately when new output arrives. Hide open-log
actions after a successful reconnect retires a stopped manual log.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(terminal): keep open-log actions after automatic logging is turned off
The current automatic log keeps recording until reconnect, so the menu
stays available for that connection and clears on the next successful
connect.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(terminal): derive log menu availability from saved session logs
* fix(terminal): nest session log actions in one More actions submenu
Collapse Open, Show in folder, and the two open-with actions into a Session log submenu that follows the Terminal Encoding menu, and drop the remembered-application hint.
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>
* fix(cli): show the final server on PuTTY bastion tab titles
PuTTY jump launches kept the bastion as the TCP host and labeled the tab
with the opaque login string. Reuse the SecureCRT tabName path so an
embedded or leftover target address is the session title, without changing
where the session connects.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(cli): validate PuTTY jump tab addresses with net.isIP
The display-host check accepted malformed colon strings such as a:b:c and
rejected IPv4-mapped and zone-scoped IPv6 targets. Use Node's IP parser so
only real addresses become the tab title.
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>
* fix(terminal): close sessions when their owning renderer is destroyed
* test(terminal): keep a shared SSH transport when a copied window closes
Closing a child window must drop only the shells that window owns.
Cover several tabs on one connection split across the main window and
the copied window, for both the in-process lifetime tracker and the
worker close 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>
* fix(browser): keep the page live beside overlay side panels
The embedded host browser is a native WebContentsView that draws above
the renderer, and its occlusion check only yielded to dialogs, menus and
poppers. Opening New host / Edit host from the terminal host tree left
the page drawn over the overlay host editor panel.
Overlay-layout aside panels now carry data-native-view-occluder. Native
views cannot be clipped by a parent View on macOS, so the pane shrinks
the live page to the rectangle left beside a panel that covers a
full-height or full-width edge strip; the page stays interactive while
the editor is open. Poppers are checked against that live rectangle, so
selects inside the panel no longer freeze the page, and pictures for
poppers over the page are offset to where it sits. A panel over the
middle of the pane, or one leaving less than 160px, falls back to a
picture.
* test(browser): provide the pane geometry helper to the first-link harness
The first-link acceptance loads useHostBrowserPanel through a VM that
strips imports and injects dependencies by name, so it threw
"resolveHostBrowserPaneBesidePanels is not defined" once the hook
started using the domain helper. Load domain/hostBrowser.ts and pass
the helper in.
* fix(vault): keep the newest connection-log replays when the side store hits quota
Once the replay side store neared the localStorage quota (~48 x 1M-char
entries) every write failed with QuotaExceededError, so the stored map
froze with the oldest sessions and newer ones never survived a restart.
On a failed write, retry without the oldest unsaved replay buffers
(dropping 1, 2, 4, ... per attempt, so retries stay O(log n)) until it
fits or none are left. Storage format, per-entry cap, 50-entry cap and
log metadata are unchanged; saved logs and other windows' ids are never
dropped.
* perf(vault): start quota retries from the newest set that last fit
* fix(vault): keep the connection-log quota hint on the vault hook
The domain helper held a process-global hint, so the same logs and map
could keep a different replay set after an earlier caller. The hint now
lives in a useVaultState ref and is passed in and returned. Drop order,
saved and unknown ids, and the size estimate are unchanged.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Grok Bot <grokbot@local>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Settings, tray, and the terminal popup mount useVaultState but never
show connection-log replay. Skip the bulky side-store load and storage
handlers there so a main-window persist does not JSON.parse it in those
renderers. Main-window behaviour and the write path stay unchanged.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Unsaved logs kept up to 500×1M chars of terminalData in React state while
only 50 survive a restart. After add/update/toggle, drop terminalData from
the oldest unsaved entries beyond MAX_PERSISTED_UNSAVED_TERMINAL_DATA_ENTRIES.
List metadata stays; replay beyond 50 matches restart behaviour.
When the side-store write fails (QuotaExceeded once it nears ~50 MB), keep
the same capped merged map in memory instead of accumulating every log's
replay buffer in connectionLogTerminalDataRef.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Each connection-log write parsed the closed-session terminal-data JSON
twice, merged it twice, and JSON.stringify-compared it twice. Pass the
already-merged map through and compare keys without serializing the
multi-MB replay buffers. Storage format and in-memory state stay the same.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
The alt-screen preview tracker segmented every print span and recomputed
the last grapheme on every non-ASCII append. Skip the boundary walk when
the remainder already fits the piece cap, and segment the open grapheme
only when a later span actually rejoins it.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
ssh2 re-runs hostVerifier on every key re-exchange. When the knownHosts
snapshot for the connection doesn't contain the key (first connect with
"Add and continue", "Continue" without saving), the verifier showed the
host-key dialog again and answered asynchronously. While that answer is
pending ssh2 keeps sending channel data ahead of its NEWKEYS, and OpenSSH
drops the connection ("Bad packet length"), killing the shell, rz/sz and
SFTP on that transport. Servers or gateways with a RekeyLimit hit this
after a few MB of transfer.
Like OpenSSH, accept a rekey only if it presents the key already accepted
for this connection, and answer synchronously without prompting.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Copy style state on SGR changes rather than for each printable character. Preserve historical styles, cleared screens and cursor overwrites without dropping log content.
* docs: losslessly shrink README images and mirror Trendshift badge
Recompress local images linked from the READMEs only when the file gets
smaller and decoded pixels stay identical. Copy the Trendshift badge from
README.md into the zh-CN, zh-TW, and ja-JP READMEs.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* docs: replace README screenshots with lossless WebP
Store the ten README screenshots as lossless WebP under docs/assets/readme
and point every language README at those files. Keep them out of app.asar.
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>
* fix(sftp): Retry after a dropped connection resumes from the staged checkpoint
Retry on a failed live task always cloned it under a fresh transfer id with
checkpoint 0. After a connection drop that throws away the uploaded bytes, and
because the remote stage path is derived from the transfer id, the new attempt
stages a new .part while the old one stays orphaned on the server (cleanup runs
against the dead session).
When a top-level file failed with a session-loss error and has a checkpoint,
Retry now demotes it to interrupted/reconnectRequired and goes through the same
dedicated reconnect-resume path a restored task uses: same id, same stage,
truncated to the checkpoint. Content errors and orphaned rows keep the existing
restart-from-0 contract.
* fix(sftp): resume a dropped upload from the verified stage checkpoint
A failed upload never pauses, so Retry has a checkpoint and no SHA-256.
The bridge treated that as an unproven resume and restarted at byte 0,
truncating the same .part. Stat and trim the stage on the new connection
instead, and still restart when the .part does not cover the checkpoint.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(sftp): trim a resumed upload stage with SFTP setstat
ssh2 has no ftruncate and ssh2-sftp-client has no truncate, so a longer
.part after a reset always failed the checkpoint trim. SETSTAT {size},
or FSETSTAT on an r+ handle, shortens the stage on the live channel.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(sftp): require a stable local source before an unfingerprinted resume
A dropped upload has a checkpoint and no SHA-256. Retry kept that
checkpoint even after the local file was rewritten at the same size, and
the same predicate also resumed downloads and remote copies.
Remember size and mtimeMs when the local upload starts. On retry, keep
the checkpoint only when both still match; otherwise restart from 0.
Limit same-id resume to local-to-remote uploads.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(sftp): reject an unfingerprinted resume when prefix bytes changed
Size and mtime can both be restored after a same-size rewrite, so Retry
was appending the new suffix onto the old staged prefix. Keep the head
content sample, and a 64 KiB window ending at the published checkpoint
when that sample does not cover it. Resume only when those still match.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(sftp): resume an unfingerprinted upload only when the prefix digest matches
Hash the bytes actually acknowledged on the remote .part, in order, and
store that SHA-256 with the checkpoint offset. Retry re-reads the local
prefix and restarts when the digest or size differs, including a same-size
rewrite the head sample and checkpoint window both missed.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(sftp): cancel prefix verification and drop unused upload digests
Pass the transfer abort signal into the retry prefix re-hash so Cancel
stops the read and the task ends cancelled. Drop the in-memory digest
when an upload is cancelled, fails without a resumable checkpoint, or
the task is cleaned up.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: campaign <campaign@local>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(quit): don't freeze the main process when the user keeps unfinished transfers
Quitting with an unfinished transfer pauses it and asks for confirmation in the
renderer. Declining was reported to main as "dirty editors", so
handleBeforeQuit raised a synchronous "Unsaved changes" message box. That box
blocks the whole main process (IPC, every SFTP transfer) until it is dismissed,
names editors that do not exist, and the paused transfer was never resumed.
The renderer now reports the refusal with reason "transfers" and resumes the
tasks it paused for the prompt; main keeps the quit cancelled but skips the
native editor dialog for that case. Real unsaved editors behave as before.
* fix(quit): report a transfer refusal before resuming
A slow resume could miss the 5s dirty-check timeout and let the app
quit anyway. Reply with reason "transfers" first, then resume only the
rows that were transferring or queued before the prompt paused them.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: campaign <campaign@local>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): round-trip jump host chains in CSV import/export
CSV export dropped hostChain, so a vault imported on another machine
lost every jump host. Persist the ordered chain as an OpenSSH
ProxyJump column and bind each hop on import.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): drop inline hops from a rejected ProxyJump chain
A self-referential ProxyJump list created unresolved hops before the
chain was skipped, so the import kept those hosts. Discard inline hosts
created for that rejected chain.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): prefer a same-group hop when binding CSV ProxyJump
When an imported host and an existing vault host both match a hop
endpoint, pick the candidate in the target host's group before falling
back to the imported match.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): prefer a same-group label when binding CSV ProxyJump
Combine imported and existing label matches before pickCsvJumpCandidate so a cross-group imported label cannot beat an existing hop in the target group.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(vault): reject CSV ProxyJump chains that contain a malformed hop
A multi-hop ProxyJump cell with any invalid hop now warns and leaves the
host chain unset, instead of saving the valid hops as a shortened route.
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(sftp): remove the per-directory barrier in folder transfers
Each directory used to run its files in a private worker pool and the walk
awaited that pool before moving on, so trees of many small directories ran
at 1-2 files in flight and paid a full drain at every directory. Files now go
through one tree-wide gate sized by the transfer concurrency setting; the
walk continues while earlier directories drain and the root awaits every
batch (also when the walk throws). Entry indexes are still assigned in walk
order, so resume checkpoints are unchanged; parent directories are created
by the walk before their files are queued.
* fix(sftp): handle folder file-batch rejection when it is queued
A batch can reject while the walk is still listing later directories.
Settle that promise immediately and surface the same error from the root drain.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: campaign <campaign@local>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Files that fit one READ fanout window (64 x 32KB = 2MB) read on the shared
browse channel without taking FAST_DOWNLOAD_CHANNELS_PER_SESSION, so folder
downloads of small files run with the configured transfer concurrency
instead of strictly one file at a time. Large files still serialize (#1507).
Co-authored-by: campaign <campaign@local>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* perf(sftp): coalesce transfer lifecycle repaints to one per frame
* test(sftp): settle coalesced lifecycle paints before asserting
Wait one animation frame before reading transfer rows, and cover the final
completion, pause, cancel, and error paint plus unmount cancellation.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(sftp): do not let a stale coalesced snapshot reopen a latched pause
Pause all could leave the folder row transferring while its children
were paused. A child lifecycle publish still carried the pre-pause
parent at the same lifecycle epoch, and the store treated that equal
epoch as Resume.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: campaign <campaign@local>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
A single -o spec with col=,col= is folded into the first column's header
on procps-ng 3.3, so cwd probes never see the su'd shell.
Fixes#3820
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
#3816 sends uploads <= SMALL_UPLOAD_SHARED_MAX_BYTES (64 x 32 KiB = 2 MiB)
straight to the shared session. #3814's tests were written on a base without
#3816 and use 128 KiB / 14 B payloads, so on main:
- 'connection loss mid-upload ...' never opens the isolated channel whose
write simulates the reset; the shared write mock never answers -> hang.
- The leaked hung transfer then makes the next three upload tests time out.
- 'isolated channel ending before OPEN ...' passes alone but vacuously.
Raise those payloads above the cutoff and add a small-upload variant that
drops the connection on the shared channel.
Co-authored-by: Qi Chen <qi@local>
* fix(sftp): fail uploads promptly when the SFTP connection is lost instead of hanging
ssh2 silently drops requests issued on an SFTP channel that has already
ended, and ssh2-sftp-client keeps its stale sftp after a TCP reset. After a
reset / sshd kill / VPN drop the concurrent-isolated upload failed, then the
fallback checkpoint stat on the dead browse channel never settled, freezing
the task at 'Transferring' forever with a leftover .part.
* fix(sftp): release the path gate when a write channel ends before OPEN
A resumable staged upload marked the path gate issued, then found the
isolated channel already closed. The write-error path left a generated
stage gate held, and the resumable flow skipped the bounded fastPut wait,
so later uploads waited forever. No OPEN was sent, so release the gate
and reject with sftpChannelClosedError.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: campaign <campaign@local>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* perf(sftp): upload files that fit one fanout window on the shared session
* test(sftp): keep small shared uploads from ending the browse session
Cancel and pause of a file that fits one fanout window must stay on the
shared channel: no isolated sftp-server, no sftp.end(), and a later upload
on that same session still completes.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: campaign <campaign@local>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Fixes#3310
Borrow the live SFTP channel on single-channel OTP banners so transfers do not hang waiting for a second keyboard-interactive prompt. Channel death fails with SFTP channel closed without OTP redial; disposable upload channels may fall back once onto a healthy shared browse channel.
Carry the existing shared session-validity guard to OpenCode v2 instruction and prompt sends after startup waits. Consume rejected SSE reader cancellation during fail-closed cleanup.
Add deterministic regressions for the provider handoffs and cancellation, with real loopback HTTP/SSE smoke coverage.
Refs #3620, #3805, #3802
Share opt-in, reversible privacy masking through main-owned policy and conversation maps across Catty, integrated agents, and External MCP. Restore managed arguments before validation, approval, and execution; fail closed on retired policy state. Add shared settings, documentation, boundary and lifecycle regressions.
Validated with deterministic regressions, production desktop smoke, clean Codex review and passing cross-platform CI.
Closes#3620
* feat(ai): support OpenCode 2.x while keeping the 1.x driver
Detect the installed opencode major version. 1.x stays on the bundled SDK and the old startup banner. 2.x uses the new serve banner, HTTP Basic auth, and the session API so model lists and chat turns work.
Fixes#3799
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): stop the OpenCode 2 event stream when a prompt is rejected
A rejected prompt does not emit a terminal SSE event. Abort that stream and close the server before waiting, and keep HTTP "not found" responses from being reported as a missing CLI.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): select OpenCode 2 catalog ids and keep context optional
Aliases expose provider/catalog id rather than the provider API model id. A rejected instruction entry warns and still sends the user prompt. The entry value stays the system-prompt string, which is valid JSON for that endpoint.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): post OpenCode 2 prompts after the event stream is open
The live /api/event stream does not replay. Wait until the server accepts the subscriber and start reading it before sending the user prompt.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): refresh OpenCode version and catalog after startup races
Key the version cache by the resolved PATH binary and its mtime, so an in-place upgrade is not stuck on the previous major. While listing 2.x models, refetch after model.updated or provider.updated until the snapshot stays quiet.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): refetch the OpenCode catalog when it changes during the first read
An update that arrives while the initial model list is in flight belongs to a newer snapshot. Record the event count before that request and fetch again when it moves.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): refetch the OpenCode catalog when the settlement budget elapses during a read
A catalog read that outlasts the quiet-window budget still has to fetch the snapshot announced by an in-flight model or provider update.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): localize OpenCode 2 failures when the server dies mid-turn
A killed event stream was surfacing undici's raw "terminated" text. Map that disconnect, a closed socket, and an SSE stream that ends before the turn finishes to the existing session-failed message.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): ignore Bun version banners and hide undici transport errors
A crashed Bun binary was cached as OpenCode 1.x, and fetch or JSON-body failures still showed raw undici text. Skip Bun and Node.js runtime versions, and map those transport errors to the existing unavailable message.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): keep the OpenCode 2 config model as a qualified string
The selected provider/model stays the original string in OPENCODE_CONFIG_CONTENT. Session create still sends Model.Ref with id and providerID.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): keep caller OpenCode config and hide transport session errors
A profile's OPENCODE_CONFIG_CONTENT is merged under Netcatty's generated settings, and a server-reported fetch failure no longer reaches the chat as raw transport text.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): drop undici transport text from OpenCode 2 details
HTTP error bodies and tool failures share the same sanitizer as session errors, so fetch failed, terminated, and other side closed never reach the chat.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): keep OpenCode transport mapping to mid-stream failures
Server-authored tool and session messages are no longer rewritten just because they match a transport string. A dead reply stream still reports the localized session-failed key.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ai): restore the OpenCode 2 server username
The transport-sanitizer edit replaced the opencode username constant, so a real 2.x spawn threw ReferenceError before Basic auth. Put the constant back and lock the default client username.
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(system-manager): add container runtime switcher and kubectl pods
Keep the existing Docker tab and select docker, podman, nerdctl, or
Kubernetes from runtimes detected on the host. Kubernetes replaces the
switcher with context and namespace selectors in the same panel.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): address container runtime review findings
Move the runtime adapter into the domain layer, keep a context namespace
when namespace listing is forbidden, and emit PowerShell or cmd startup
commands for Kubernetes shell, logs, and describe.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): cover Windows container popups and stale namespaces
Run container shell and logs through the host shell's inspect wrapper on
PowerShell and cmd, and reset the namespace when a remembered Kubernetes
context is no longer present.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): parse kubectl JSON after banners and scope inspect cache
Skip bracketed banners before kubectl JSON, and include the container
runtime in image and container inspect cache keys.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): parse Podman image lists emitted as JSON arrays
Podman applies {{json .}} to the whole image slice. Accept that array,
including Names and byte sizes, and keep Docker's JSON lines.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): parse Podman stats arrays
Accept Podman's stats JSON array and its id/cpu_percent fields, and keep
Docker's one-stat-per-line output.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): list every Podman image tag
Expand each name in a Podman image's Names array into its own row so
extra tags stay visible and searchable.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): quote kubectl list and delete for the session shell
Namespace, pod, and delete commands follow PowerShell or cmd quoting when
the session shell is not POSIX, so apostrophes in a context name survive
local PowerShell.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): drop renamed runtime launch commands from shell history
Container popups now probe with _nc_rt_err and also cover podman, nerdctl,
and kubectl. History sanitizers still required _nc_docker_err, so those
launches stayed in persisted history. Confirm-dialog resets also follow the
selected runtime.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): sum Podman interface stats and quote kubectl for the executor
Podman {{json .}} stats keep per-interface counters on Network, so the
details panel now sums RxBytes and TxBytes. Background kubectl runs in
powershell.exe on local Windows even when the terminal is cmd, so those
commands are quoted for that executor instead of the interactive shell.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): remove a selected image tag instead of the shared id
Podman rows that share one image id were all deleted with rmi <id>, which
either fails while other tags remain or drops every name. Tagged rows now
run rmi repository:tag; dangling images still use the id.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): reject cmd kubectl arguments that contain percent signs
Interactive cmd.exe expands %VAR% inside double quotes, so a context such
as prod-%USERNAME% would be rewritten before kubectl sees it. Cmd popups
now refuse that argument. PowerShell quoting still preserves the percent.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* refactor(system-manager): move runtime selection into an application hook
Persisted source, context, and namespace state plus the kubectl discovery
effects now live in useContainerRuntimeSelection. DockerManagerTab only
renders the switcher and panels.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): offer kubectl shell only for running containers
Succeeded and failed pods have no running container, and a completed init
container cannot accept exec. Shell targets now come from container
statuses that are running; logs still include init containers.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): keep Windows popup commands out of shell history
PowerShell and cmd container popups inspect then run, and kubectl popups
are typed directly. History filtering only matched the POSIX sh -c wrapper,
so those launches were stored as user commands.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): mark Windows popups instead of matching kubectl syntax
Filtering kubectl --context logs and describe also dropped commands the
user typed. Windows popups now carry __NC_POPUP__, and ordinary kubectl
lines stay in history.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): reject cmd arguments that contain delayed-expansion marks
cmd.exe with /V:ON expands !VAR! inside double quotes. Contexts such as
prod!USER! are refused for cmd popups, same as percent signs.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): hide raw nerdctl and unreachable cluster errors
Only offer nerdctl after a short info probe succeeds, or when the failure is a socket permission error that sudo can still recover. Map dead containerd and unreachable kube API errors to short localized messages and keep the CLI text in the log.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): show terminating pods and ephemeral containers
A deletion timestamp is Terminating even while the phase is still Running. Ephemeral containers are included in the logs list, and a running one is a shell target.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): keep the runtime select on Kubernetes and avoid covering it
Kubernetes uses the same compact source menu as Docker, with context and namespace beside it. Shell popups open in the free part of the window so they no longer sit on top of the pod badges.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): restore the saved engine when Kubernetes disappears
A host that was on Podman before Kubernetes keeps Podman if kubectl is gone, instead of falling back to Docker and overwriting that choice.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): avoid the visible system panel when others are parked
A hidden parking-host panel can be the first system-manager match and
has a zero-size rect. Scan every match and pass the first panel that
actually has size so the shell popup still sits beside the open panel.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): keep a runtime choice across probe gaps and show podman ports
A failed nerdctl or kubectl probe was saved as the fallback engine, so the
previous choice never returned. Reconciliation now restores the stored source
when it is detected again, and automatic fallbacks are not written back.
Podman ps port mappings and exposed ports are formatted into the same display
string the container detail strip already shows.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): show a pod's Evicted reason
Pod status.reason is the eviction signal, while the phase stays Failed and
the container often reports Error. Use that reason after a container waiting
state and before the generic phase so the badge stays Evicted.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): keep an avoided shell popup inside the work area
When the source window sits partly off-screen, clamping a slot back onto
the display could cover the system panel, and the fallback then used the
raw slot origin. Place the popup in the on-screen part of the slot instead.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(system-manager): left-align the runtime selector row
The Kubernetes context, namespace, and runtime selects sat on the right of
the header, so they did not line up with the search field. Keep the same
padding, let the context select fill the remaining width, and leave the
runtime and namespace selects at their natural width.
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>
* fix(ui): use a lightning icon for quick connect
The host switcher quick-connect row now uses lucide Zap at the same size and stroke as the previous terminal icon.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ui): badge the quick-connect icon like host avatars
Reuse the DistroAvatar sm fallback classes so the Zap sits in the same 20px primary rounded box as a generic saved-host row.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ui): fill the quick-connect bolt
Zap uses fill="currentColor" so it renders solid in the primary badge, colored by the existing text-primary-foreground token.
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>