* Fix macOS Chinese IME full-width punctuation and window-switch input
Two macOS IME regressions in the terminal:
- Full-width punctuation (，。？！) was sent as half-width ASCII. With
kitty keyboard on, xterm encodes+sends the ASCII symbol on keydown and
preventDefaults it, dropping the IME's committed glyph from the input
event. Bypass those punctuation keydowns/keypress/keyup so the native
input pipeline runs, then forward the committed glyph to the PTY.
- After switching windows, the IME stayed stuck in ASCII with no way back
to CJK. macOS+Chromium leaves a stale NSTextInputContext on the still
focused helper textarea after reactivation (electron#32307/#34952).
Force a blur -> next-frame refocus on window focus to rebuild it.
Both paths are macOS-scoped where relevant and covered by unit tests.
* Tighten macOS IME punctuation handling
---------
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* fix(browser): keep webview guests alive across worktree switches
Electron <webview> guests are destroyed when their DOM parent is removed.
When switching worktrees, BrowserPane chrome unmounts and previously tore
down the webview, forcing a reload on return.
Introduce persistent page viewports inside overlay slots so webviews stay
in a stable parent without keeping the full React chrome mounted. On
worktree switch, park hidden viewports instead of destroying guests;
explicit close paths still call destroyPersistentWebview.
Fixesstablyai/orca#6385
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(browser): harden persistent webview viewports
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
- Centralize and align auto-merge eligibility logic across web and
mobile clients.
- Use the `enablePullRequestAutoMerge` GraphQL mutation instead of
`gh pr merge --auto` to prevent immediate merges on clean branches.
- Fall back to `gh pr merge --auto` when a merge queue is required on
the base branch.
- Hide the auto-merge control when only optional checks are pending.
Previously, the AI Vault only blocked resume actions for workspaces with
an active SSH connection (checking `connectionId`). This allowed resume
actions to run on runtime-owned workspaces, which are non-local but do
not use SSH.
To resolve this:
- Introduce `getAiVaultResumeWorkspaceTargetStatus` to classify targets
based on both `connectionId` and `executionHostId`.
- Restrict AI Vault panel actions, session resume checks, and drop layers
to local workspaces.
- Preserve the `executionHostId` on project groups during normalization to
ensure runtime-owned groups retain their execution host status across
persistence reloads.
Avoid forcing the active view back to the creation panel when background
worktree preflight completes after the user has already navigated away.
- Add a `revealCreationSurface` option to `continueBackgroundWorktreeCreation` to skip switching views.
- Check both `activeView` and `activePendingCreationId` to determine if the creation UI is actually visible.
- Use this option when kicking off background creation for GitHub work items.
Windows runs managed hooks through a shell, and #6078 wrapped the
launcher in `powershell -EncodedCommand` to survive spaces in profile
paths. Combined with the inner PowerShell Invoke-WebRequest post, every
hook spawned two PowerShell processes (~300ms startup each), so a hook
took ~650ms+ and fired up to 6x per turn. Codex 0.140 renders that as
lingering "Running <event> hook" rows. The earlier RC worked around it
by deleting SessionStart/UserPromptSubmit/Stop, which loses lifecycle
status fidelity.
Keep all six Codex events and make them fast instead:
- Codex runs hooks as `cmd.exe /C <command>` and forwards our string
verbatim when it has no spaces/quotes, so emit the bare .cmd path for
cmd-safe profiles (zero shell startup) and fall back to the encoded
PowerShell launcher only for spaced/metachar paths (#6078 robustness).
- Replace the inner PowerShell post with curl.exe (Windows 10 1803+),
posting the same form fields as the POSIX hook and reading the raw
payload from stdin via `--data-urlencode payload@-` so UTF-8 (e.g. CJK)
survives without code-page translation.
Result: the common Codex case is 0 PowerShell (~70-150ms vs ~650ms);
spaced-path profiles drop to 1.
Claude runs hooks through Git Bash, so its launcher must stay
PowerShell-encoded (a bare path is what breaks it), but its inner post
moves to curl.exe too, cutting Claude from two PowerShell startups to
one.
Validated against the real Codex 0.140 binary (interactive TUI + exec):
all managed hooks fire, render briefly, and clear with no lingering
rows; the listener receives every post in 1-6ms.
Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(automations): track exact run terminal targets
Record terminal pane keys and PTY ids on automation dispatches, persist them across run updates, and propagate them through runtime launch results.
Use the recorded pane and PTY for completion detection, session reuse, prompt submission, and View run navigation so split-pane tabs reopen the actual run terminal.
Keep legacy runs compatible by falling back to workspace resume when exact terminal identity is absent.
* chore: address automation targeting review cleanup
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Jinwoo-H <jinwoo0825@gmail.com>
Co-authored-by: Orca <help@stably.ai>
Update the pill and badge styles for GitHub PR statuses and check runs
to use muted background washes with theme-appropriate text colors. This
replaces solid high-contrast backgrounds with soft backgrounds, and
fixes low-contrast text in light mode by using darker foreground colors.
* fix(ports): stop cursor-move redraws corrupting advertised dev-server hosts
The Live Ports panel showed `localhst:5199` (and opened a broken URL with the
missing `o`) for a dev server that printed `http://localhost:5199`.
AdvertisedUrlWatcher cleans raw PTY bytes with stripTerminalControls, which
*deletes* ANSI/cursor sequences instead of emulating them. CLIs that redraw a
line differentially step the cursor forward over characters already on screen
(`ESC[1C`) rather than reprinting them; a real terminal renders the underlying
glyph, but deleting the move splices the two text runs together and drops the
skipped cell — `http://localh` + `ESC[1C` + `st` collapses to `localhst`.
`new URL()` then accepts the corrupted, plausible-looking host verbatim, so it
gets cached, displayed, copied, and opened.
Neutralize horizontal cursor moves (forward/back `C`/`D`, absolute column `G`,
position `H`/`f`) to a space before the CSI delete pass, so the URL candidate
matcher breaks at the seam and can't fuse a corrupted hostname. Worst case we
skip that frame and fall back to the kernel bind, which is the safe default.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs(ports): document stripTerminalControls
Add a docstring to the exported helper to clear CodeRabbit's docstring
coverage warning and explain why horizontal cursor moves are neutralized.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(ports): invalidate cursor-redrawn URL candidates
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Jinwoo-H <jinwoo0825@gmail.com>
Co-authored-by: Orca <help@stably.ai>
Jinwoo HongandOrca
Brennan Benson
Neil
github-actions[bot]
Jinjing
lvfenandJinjing
