Shift+Enter inserted a hardBreak inside one paragraph, while paste, prompt
recall and restored drafts already used one paragraph per line. ProseMirror
handles triple-click, Mod-click and macOS Ctrl+A/Ctrl+E per paragraph, so on
a message typed with Shift+Enter they all acted on the whole message.
A small appendTransaction plugin now splits every hardBreak into a new
paragraph in the same undo step, so Shift+Enter, rich-text drops and HTML
pastes all produce paragraph lines. Saved drafts with hard breaks are split
into paragraph lines when the editor adopts them. The placeholder and @agent
spacing follow the per-paragraph shape. The NativeChatLineBoundaryKeys
override from #42409 is no longer needed and is removed.
Fixes#42606
Co-authored-by: Neil <neil@stably.ai>
* fix(window): native reloads and window close wait for the durable session checkpoint
Menu, keyboard (including Force Reload), DevTools and in-app reloads now ask
the renderer to run the same checkpoint preparation as the in-app restart,
including the main-side durable flush and its retry-then-degrade budget.
A renderer that reports a failed save keeps its document, matching the
in-app reload. A renderer that never answers is reloaded after a bounded
deadline (quit's frozen-renderer bound plus the checkpoint flush deadline),
as before. Rich and landing-page window close wait for the staged
checkpoint to reach disk and cancel when it fails.
* fix(window): name a failed close checkpoint and close on the second failure
A durable-write failure now reports its reason in the canceled-quit toast,
and a repeated close after it proceeds as before the wait existed, so a
broken disk cannot trap quit.
* fix(window): close waits for a pending reload checkpoint and a renderer exit cannot strand quit
A close requested while a reload's durable checkpoint runs now waits for that verdict and replaces the reload. A renderer that exits after acknowledging a quit close no longer leaves quit waiting for a confirmation that cannot come.
* fix(window): settle reload/close races found in review
A prepare request that fails to send no longer leaves later reloads skipping the checkpoint. Any close a dying renderer was answering now finishes, and every cancel clears it. A reload no longer discards a quit close the renderer is still answering, and a reload replaced by a close restores the page's ordinary unload guards. In-app reload failures show the same toast as menu reloads.
* fix(window): track one close per window that quit only upgrades
Stale cancels are recognised from the acknowledgement order, the custom X gets its own request id, a reload retires an ordinary close, and a close deferred behind a reload bypasses a renderer that died meanwhile.
* fix(window): an overlapping close joins the pending checkpoint instead of counting its failure twice
* fix(window): a frozen renderer cannot hold Force Reload for the full save deadline, and an unreceived close no longer destroys the window on a later crash
* fix(window): a tracked quit always finishes when its renderer dies; an ordinary close counts as received once any of its requests was acknowledged
When Claude answered a startup control request with an error, Orca dropped
its words, so the chat showed "Claude couldn't start. Send your message to
try again." with no Details. Attach Claude's answer as the diagnostic;
SDK-tagged transport failures (exits, spawn errors) keep their stderr.
* perf(layout): skip unused normalization reports in pane commands (#42615)
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* Clarify cleanup pinning and ignored-file protection (#42616)
* test: remove 55 low-value suites and assertion driver (#42605)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(skills): skip Claude sync scratch folders during skill discovery (#42610)
Claude Code parks replaced skill copies in `.trash` and in-progress syncs
in `.staging` under its skills roots. Discovery walked into both, so each
synced skill showed up as duplicates. Prune those directory names in the
native file walk and in the generated WSL `find` command.
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
* test: remove 78 low-value suites and trim redundant cases (#42621)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* test: remove 53 low-value suites and trim redundant cases
---------
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Co-authored-by: BingZ <zzb@gxsmjx.com>
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
Claude Code parks replaced skill copies in `.trash` and in-progress syncs
in `.staging` under its skills roots. Discovery walked into both, so each
synced skill showed up as duplicates. Prune those directory names in the
native file walk and in the generated WSL `find` command.
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
* fix(computer): retry Linux computer use with the system python3 when PATH's lacks PyGObject
* fix(computer): retry from venvs that share the system python binary
* refactor(computer): reuse isExecutableFile for the system python check
When a linked file's macOS default app is Cursor, VS Code or VS Code
Insiders, hand it `--goto path:line:column` with
`open -n -g -a <app> --args`, the same LaunchServices hand-off the bundled
CLI ends in, minus the CLI's Node startup and `--env` forwarding. The
editor sees a desktop (non-Dock) open, so it reuses the file's window as
a CLI open would. An empty env keeps Orca's variables out of the editor
and lets a cold start resolve the login-shell env. `open` returns once
the app has launched, like shell.openPath, so there is no timer; a
non-zero exit or spawn error falls back to shell.openPath. Other
platforms, other apps, links without a line, and files main reveals
instead of opening are unchanged.
The IPC location reuses RuntimeFileOpenPosition. Moves the positive
safe-integer guard to src/shared so this and its two existing copies
share one definition.
Fixes#29508
Co-authored-by: Takumi Hara <takumihara1226@gmail.com>
The main BrowserWindow and its webContents logged MaxListenersExceededWarning on
every boot because normal subsystem fan-out exceeds Node's default of 10. Raise
the per-event warning threshold to 30 so the warning only fires on real leaks.
Fix the two leaks that the warning was masking:
- plugin panel push: each open added a 'destroyed' listener to the sender and
close never removed it. Track the release hook per session (via
addFanoutListener) and release it on close.
- pending VM checkouts: the drop handler removed its 'render-process-gone'
listener but not its 'destroyed' listener, so each renderer crash leaked one.
Fixes#42594
Co-authored-by: Neil <neil@stably.ai>
After a rejected chord (conflict or invalid), the Shortcuts recorder keeps
listening, but keys only reach it while it has focus. Clicking elsewhere left
the row stuck on "Press keys..." and kept setShortcutRecorderFocused(true),
so main-process window shortcuts stayed suspended.
While recording, a capture-phase window click listener now ends recording
when focus has moved off the recorder. Window/app switches keep recording as
before. The row error stays mounted (and copyable) after click-away; starting
any recorder clears stale errors. While recording, a rejected chord now shows
"Press another shortcut, or Esc to cancel." next to the error.
Fixes#42398
Co-authored-by: Neil <neil@stably.ai>
When orca-computer-use was already installed and current, a leftover
folder under the former name (computer-use) that Orca does not own --
edited or unrelated user content -- still produced a `leave` op, so the
Computer Use settings panel showed "Needs attention" with a Retry that
could never clear it.
When the current-name target is `keep`, drop former-name leaves whose
reason is `edited` or `other`; those folders are the user's own skills,
not blockers. Leaves Orca still owns but cannot retire (for example an
unwritable released copy) keep surfacing as attention.
Fixes#42408
Co-authored-by: Neil <neil@stably.ai>
* fix: show chat model and effort in composer tooltip
* fix: keep disabled model picker tooltip inspectable
* fix: retain composer picker focus and tooltip description
Chromium does not wheel-scroll an overflowing <fieldset>, so on short
windows or high UI scaling the Edit Quick Command dialog body could not
be scrolled down to the Advanced section. Move the overflow onto a plain
wrapper div and keep the fieldset for the disabled state and key handling.
Fixes#42576
* feat(native-chat): preview image files a reply links to
Agents usually name the images they make as links, so the chat showed only
text. A paragraph's authored links to image files on the chat's host now get
the image under that paragraph; the link stays as written.
- Select on the destination the author wrote: the plain-text path linkifier
marks the links it adds and keeps the authored URL of links it re-encodes,
so the transcript and whole-message image discovery pick the same images.
- Web, protocol-relative and network paths never preview; previews read the
literal routed host path through the existing chat image path.
- Build each child list once, so many previews stay linear.
Co-authored-by: Vitalii Kulish <mancunianez@gmail.com>
* perf(native-chat): mount chat image previews as they near the viewport
Each Markdown image preview is a full component (lightbox, actions, read
lease), so a reply naming hundreds of images mounted them all at once and
stalled its row. Mount each preview the first time it nears the viewport
and keep it mounted, so an open preview never loses its trigger.
Real Chromium, 1000 image links in one reply: mount long task 343 -> 150
ms, heap at mount 105 -> 72 MB; reads, decoded images, concurrent reads
and blob URLs were already bounded and stay so.
* fix(native-chat): release deferred image placeholder observers
* test(native-chat): provide a complete visibility observer entry
---------
Co-authored-by: Vitalii Kulish <mancunianez@gmail.com>
Ports the edge halves of two release-branch changes onto main, which production
edge deploys come from (prod edge = ba3f991):
- Ingress mode (#41986): signed ingress header (INGRESS_SECRETS) and token mode
for a cloud load balancer (INGRESS_TOKENS secret, INGRESS_ORIGINS,
INGRESS_SOURCE_CIDRS). All or nothing: a partial or malformed setup turns the
mode off and logs relay_edge_ingress_config_invalid; it never throws. Requests
from the configured egress ranges that fail the trust rule get 403. Admin calls
are matched first. Adds cloud/apps/relay-ingress (stateless proxy) and the
relay-contract ingress-header module.
- POST /__admin/verify (#42538, edge half): the director's check for a phone
resume token or invite that exists only on the edge. Hash only, read-only: an
object that never held the host (or no longer holds it here) answers not_here
before anything is written; otherwise the token's jti is its only write.
Replay and jti_full are 409. Adds hash-input variants to RelayCredentialStore
(shared with the edge; existing methods delegate, behaviour unchanged).
Kept from main: drain, single-use admin tokens (jti), signing-key rotation,
JWKS-outage tolerance, keepalive auto-response, health probe. The place label
keeps main's '<continent>/<colo>' and, only behind a verified ingress, appends
'/<country>/<ingress name>'.
No user effect until the ingress vars/secret are set or the director calls verify.
* fix(native-chat): list installed Antigravity skills in the slash picker
Antigravity ACP reports built-in commands and no skills, and the chat had no picker profile, so typing / never scanned skill folders. Keep the scan of its own skill folder and shared agent skills.
* fix(skills): discover Antigravity CLI user skills
The Antigravity server also loads skills from ~/.gemini/antigravity-cli/skills,
which no discovery root covered, so skills placed there never reached the
Antigravity Chat slash picker.
---------
Co-authored-by: Wooseong Kim <innocarpe@gmail.com>
* fix(native-chat): keep the answer visible when a Stop hook fails
A failed Stop hook follows a completed answer, but the turn fold treated
its red row as the turn's final answer, hiding the reply behind the
Worked-for disclosure. Mark failed Stop-hook rows with a dedicated
presentation so they stay visible beside the answer instead of
displacing it; turns that genuinely fail still end on their error.
Fixes#42487
* fix(native-chat): identify Claude Stop hooks from hook_event
hook_name carries the matcher-suffixed display name (for example
Stop:guard), so comparing it to 'stop' dropped the hook-failed
presentation for Stop hooks with a matcher. Read the bare event name
from hook_event and keep hook_name for display; a matcher-suffixed
hook_name without hook_event still counts so hook_name-only frames
keep working.
* fix(native-chat): read failed hooks from their frame and fold only mid-turn ones
Every failed hook, not just Stop, completes the turn, and a failing
PostModelRequest hook also lands after the answer. The fold now spots a
failed-hook row by the hook frame every such row already carries, so it
also works for rows from older hosts and rows already stored. Failures
that end a turn stay visible below the answer; earlier ones fold as work.
---------
Co-authored-by: Kelvin Amoaba <97001695+AmoabaKelvin@users.noreply.github.com>
An ACP form elicitation's message is context for the whole form, but each question card and the answered receipt repeated it. Only the first question now carries it, plus any field with no title or description of its own.
* feat(mobile): ship Android as dev.orca.app and check updates by install source
Android becomes dev.orca.app, the package already live on Google Play; iOS keeps
com.stably.orca.mobile. The move screens and the "Orca (old)" label stay keyed to the old package,
so they never show in the new app.
The app asks Android who installed it. A Play install gets its updates from Play, so it runs no
GitHub check, hides "Check for updates" and points the update wall at Google Play. Any other
install checks GitHub for android-v releases; the old package keeps mobile-android-v.
* ci(mobile): release dev.orca.app to Play and publish Google's signed APK
An android-v tag (or a manual run) builds the AAB with the upload key and sends it to Play's
internal track. It then downloads the universal APK Google signed for that versionCode, fails unless
its V3.0 signer is Google's app signing key, and publishes it as the android-v GitHub pre-release.
Play and GitHub installs are the same signed app. A manual run with publish_github_release off only
validates the bundle with Play.
The orca.dev APK bump job is gone: the site builds its link from mobile-android-v, so it stays on
the last old-app APK, which sends people to the new app.