The setImage-swapped progress frames lost the template treatment on macOS and
rendered black-on-dark once transfers started. Dial back per user decision:
the tray is now just the glyph (template image set once at construction, never
swapped) and clicking it activates the window — no menu, no progress effects.
Removes everything that only served the fancy version: progress frames +
generation, engine.aggregateProgress, the workspacesSync webview->bun channel
(schema, send proxy, panels effect, handlers) and the activate-tab action.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- AppTray: template menu-bar icon (22pt, adapts to light/dark). While
transfers run it swaps to pre-rendered frames — faded arrows + a fill bar
showing aggregate doneBytes/totalBytes of the queue (nearest 5%, throttled
500ms); back to the glyph when idle. No runtime image encoding.
- Menu: one entry per tab (checkmark on the active one) with live per-tab
transfer count and percentage (jobs mapped via session ids), an overall
progress row while transferring, Show <app>, Quit. Selecting a tab
activates the window and switches to it.
- First use of webview->bun messages: PanelsService pushes workspacesSync
(tab names, sessions, active id) via a new ready-gated send proxy; verified
end-to-end in the running app.
- engine.aggregateProgress(): totals + per-session breakdown over non-terminal
jobs (unit-verified idle/mid-transfer/idle).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two panels on the same server hold two SSH sessions by design, and every
self-copy guard compared sessionId — so same-server same-path copies slipped
through all of them, and createWriteStream truncated the destination before
the read began: the file (or every file in a folder) became 0 bytes, with the
overwrite dialog making it look like a legitimate replace.
Fix, defense in depth, keyed on FILESYSTEM identity instead of session:
- fsKey on every provider: 'local', or resolvedIP:port for SFTP (host is
DNS-resolved at connect so hostname-vs-IP aliases compare equal; username
excluded — same host:port is the same filesystem regardless of login)
- engine guard blocks equal/nested paths when fsKeys match; per-file
belt-and-suspenders check right before streams open
- probe layer for hidden aliases (different addresses, same machine): when
fsKeys differ but paths collide, a token file written via dst and read via
src decides; visible => same machine => blocked. SFTP exposes no dev/inode
(what cp/rsync use), so probing is the only deterministic test; fails open
and cleans up after itself
- UI: same-folder toast now fires across two sessions to one server, and the
overwrite dialog auto-skips (never offers Replace) when the colliding item
IS the source file
Known residual, documented in README 'Same-file protection': one filesystem
mounted at different paths on two connections is undetectable client-side.
Regression suite against the real box: file-onto-itself and folder-onto-itself
blocked with contents byte-intact, folder-into-descendant blocked, legit
same-server different-dir copies still work, local same-path still blocked,
simulated alias caught by the probe with cleanup verified.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The concurrency setting now caps total concurrent file STREAMS, not just jobs:
a resizable counting semaphore is shared engine-wide, and each job copies its
files through a small worker pool drawing from it. One folder job can use the
whole budget; several jobs share it fairly; Single stays strictly serial.
Cancel/session-death tear down every in-flight stream of the job.
Verified headless: 6-file folder at budget 3 -> exactly 3 parallel streams;
budget 1 -> serial; two folder jobs never exceed the combined budget; cancel
mid-parallel leaves no straggler streams. Also demoed a real remote error
surfacing in a failed job (upload into /proc -> server message on the row).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Transfers: Single | Multi in the options popover, with an 'At once' picker
(2/3/4/6/8) when Multi. Engine concurrency becomes an instance field applied
live: at boot from settings.json and on every saveSettings (raising the limit
starts queued jobs immediately). Verified headless: 3 slow jobs overlap at
concurrency 3, stay strictly serial at 1, clamp works.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
