* Fix CJK IME input not working (#118)
CJK (Korean, Japanese, Chinese) IME input was completely broken because
cmux never forwarded preedit/composition state to Ghostty's libghostty.
Root causes and fixes:
1. Missing preedit sync: Added syncPreedit() that calls
ghostty_surface_preedit() to notify Ghostty about IME composition
text. Called from setMarkedText, unmarkText, and after
interpretKeyEvents in keyDown.
2. Wrong composing flag: The composing flag on key events now correctly
accounts for when composition just ended (markedTextBefore was true
but markedText is now empty), preventing spurious deletions when
canceling composition.
3. Event interception during IME: Added early exits in
performKeyEquivalent, the NSWindow swizzle, and the local event
monitor (handleCustomShortcut) to avoid stealing key events while
IME has active marked text.
4. IME popup positioning: firstRect(forCharacterRange:) now uses
ghostty_surface_ime_point() for accurate cursor-relative positioning
of the IME candidate window.
* Add regression tests for CJK IME composition (#118)
31 tests covering Korean, Japanese, and Chinese IME input scenarios:
- Korean jamo combining: ㅎ -> 하 -> 한 composition lifecycle
- Chinese pinyin: multi-letter marked text and candidate selection
- Japanese hiragana-to-kanji: romaji -> hiragana -> kanji conversion
- insertText correctly commits composed text and clears marked state
- unmarkText properly clears composition state (idempotent)
- performKeyEquivalent returns false during active composition for
all key types (plain, shift, space, return, escape)
- Shortcut bypass: hasMarkedText gates the handleCustomShortcut bypass
- Multi-syllable sequences, backspace correction, and rapid transitions
- keyTextAccumulator lifecycle tests
Also adds #if DEBUG test accessors for keyTextAccumulator on
GhosttyNSView to enable unit testing the accumulator path.
Previously, --help/-h was only checked at the top level before command
dispatch. Running e.g. `cmux new-workspace --help` would execute the
command instead of printing help.
Add per-subcommand help text for all commands that take arguments/flags.
The help check runs before socket connect so it works even when cmux is
not running. Commands without dedicated help (ping, help, list-windows,
etc.) fall through to normal behavior.
* CLI: add --command flag to new-workspace
Allows running an initial command in the new workspace's terminal:
cmux new-workspace --command "cd /path && claude"
After creating the workspace, waits 500ms for the shell to initialize,
then sends the command text via surface.send_text.
Closes#120
* CLI: error on unknown flags for new-workspace
Typos like `--comand` were silently ignored. Now reports the
unknown flag and lists known flags.
* CLI: skip --command send when workspace creation fails
If new_workspace returns an error instead of "OK <uuid>",
don't attempt to send the command (which would hit the wrong workspace).
Root cause: update-homebrew.yml triggered on release:published, which fires
before softprops/action-gh-release finishes uploading assets. The workflow
downloaded a 404 page instead of the DMG and committed its SHA.
Fix:
- Change trigger from release:published to workflow_run (fires after the
release workflow completes, guaranteeing assets are uploaded)
- Add download validation with retries and file size checks
- Add SHA verification step before committing to the cask
- Add homebrew cask update to build-sign-upload.sh for local releases
- Add regression test (tests/test_homebrew_sha.sh)
- Update /release and /release-local skills with homebrew verification steps
Fixes#110
- New /pull command: pulls main and updates submodules locally
- sync-branch: never pushes automatically, asks user first
- No pushing to submodules; all changes land via PRs
Use GhosttyFlashOverlayView (hitTest returns nil) instead of plain
NSView so the overlay doesn't steal drag/mouse routing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The drop placeholder rendered in SwiftUI was hidden behind the
portal-hosted terminal surface. Add an AppKit overlay directly on
GhosttySurfaceScrollView driven by a new paneDropZone environment key
from Bonsplit so it renders above the Metal layer.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Move port scanning from shell to app-side with batching
Replace per-shell `ps -axo + lsof` scanning with a centralized
PortScanner singleton in the app. Each shell now sends lightweight
`report_tty` (once per session) and `ports_kick` (on preexec/precmd)
socket messages. The app coalesces kicks across all panels and runs a
single `ps -t <ttys> + lsof -p <pids>` covering every active panel.
Also fixes a macOS 26 Tahoe regression where `getsockopt(LOCAL_PEERPID)`
returns ENOTCONN on accepted sockets when the peer disconnects before
the handler thread starts. This was silently breaking ALL socket
commands sent via ncat --send-only. The fix captures the peer PID in
the accept loop immediately after accept(), and falls back to
LOCAL_PEERCRED (uid check) when the PID lookup fails.
* Fix PR review feedback: burst timing and auth comment clarity
- P2: burstDelays were accumulating (0.5+1.5+3+... = ~22.5s) instead of
firing at absolute offsets from burst start. Now uses burstStart anchor
so scans fire at 0.5s, 1.5s, 3s, 5s, 7.5s, 10s as intended.
- P1: Clarify LOCAL_PEERCRED fallback rationale — same security boundary
as socket file permissions (0600), does not widen attack surface.
Long-lived connections still get full descendant check via LOCAL_PEERPID.
The browser omnibar's updateNSView and controlTextDidEndEditing
were both dispatching makeFirstResponder calls without any guard
against re-dispatch. Each makeFirstResponder triggers SwiftUI's
FirstResponderObserver, which re-evaluates the view graph, which
calls updateNSView again, creating an infinite loop via the main
dispatch queue.
Fix: Add a pendingFocusRequest flag on the coordinator to prevent
re-dispatching while a focus/blur request is already in flight.
Also add nsView.currentEditor() != nil to the isFirstResponder
check so the field is recognized as focused during the transition
when the field editor (not the field itself) is first responder.
* Remove index-based CLI APIs and make commands workspace-relative
Migrate 14 CLI commands from v1 text protocol to v2 JSON-RPC, making them
workspace-relative via CMUX_WORKSPACE_ID env var fallback. Update 7 more
commands to use normalize functions instead of legacy resolvers. Remove dead
code (5 structs, 5 parsers, 3 functions). Add regression tests.
Commands migrated v1→v2: send, send-key, send-panel, send-key-panel,
new-split, new-pane, new-surface, close-surface, list-panes,
list-pane-surfaces, list-panels, surface-health, focus-pane, focus-panel.
Commands updated: move-workspace-to-window, list-workspaces,
close-workspace, select-workspace, trigger-flash, resolveWorkspaceId,
resolveSurfaceId.
* Fix CMUX_SURFACE_ID env fallback when --workspace is overridden
When --workspace is explicitly passed, don't fall back to CMUX_SURFACE_ID
from the caller's environment. The caller's surface belongs to a different
workspace, causing "surface not found" errors. Only use the env var fallback
when the workspace is implicit (from CMUX_WORKSPACE_ID or server default).
Affects: send, send-key, new-split, close-surface, trigger-flash, identify,
notify, claude-hook.
* Validate surface before close and respect -- option terminator in send
P1: close-surface now validates the surface handle exists before sending
surface.close, preventing silent fallback to focused surface when a stale
or mistyped ref is provided.
P2: parseOption now respects -- as an option terminator. send, send-key,
send-panel, send-key-panel strip the -- marker from payload args. This
prevents payload tokens like --workspace from being consumed as routing
flags (e.g. `cmux send -- echo --workspace foo` sends all text correctly).
* Fix test_send_workspace_relative: add to main() and send valid text
The test was never called from main() and sent empty string which would
raise immediately via _run_cli. Send a space character instead and add
the test to the main() call list.
* Respect --id-format in text output and resolve workspace refs across windows
Text-mode list commands (list-workspaces, list-panes, list-panels,
list-pane-surfaces, surface-health) now honor --id-format for plain output,
not just --json. Added textHandle() helper that picks ref/id/both based on
the selected format.
resolveWorkspaceId now enumerates all windows when resolving a workspace
ref, so notify/claude-hook work correctly in multi-window sessions where
the target workspace may be in a non-active window.
* Preserve escape-sequence semantics for send text
The v1 server unescaped \n, \r, \t in send payloads before injecting into
the terminal. The v2 surface.send_text handler sends text verbatim. Add
CLI-side unescapeSendText() to restore the same behavior: \n and \r map
to carriage return (Enter key), \t maps to tab. Applied to send and
send-panel commands.
* Reject malformed handles instead of passing through to server
All four normalize functions (window, workspace, pane, surface) now throw
a clear error for unrecognized handle formats instead of passing them
through. Previously, a typo like `--panel foo` would forward `foo` to the
server which would silently fall back to the focused surface.
* Allow cross-workspace surface refs in close-surface and strip plural ID arrays
Remove workspace-scoped pre-validation from close-surface so explicit
surface refs/UUIDs from other workspaces work without requiring
--workspace. The server resolves workspace from surface_id directly.
Malformed handles are already caught by normalizeSurfaceHandle.
Extend formatIDs to strip plural _ids/_refs array pairs (e.g.
surface_ids/surface_refs in pane.list output) based on --id-format,
matching the existing singular _id/_ref stripping behavior.
* Honor explicit --window over CMUX_WORKSPACE_ID env fallback
When --window is passed globally, skip the CMUX_WORKSPACE_ID env var
fallback so commands operate on the targeted window's selected workspace
instead of the caller's workspace from a different window. Affects all
migrated commands that use workspaceFromArgsOrEnv or inline env fallback.
Remove -Dxcframework-target=native from CI and release workflows,
defaulting to universal (matching upstream ghostty). The native target
produces a macos-arm64 xcframework slice that causes Xcode to link the
final binary differently (~70KB more __text), resulting in menubar and
right-click lag on M1 Max. The arm64 static libraries are byte-for-byte
identical between native and universal builds - the difference is purely
in how Xcode resolves the xcframework slice.
* Fix browser panel opening new tabs on every link click
Navigate target=_blank and window.open() links in the current webview
instead of spawning new browser tabs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Preserve cmd+click new tab behavior in createWebViewWith
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
