The docs layout mounts two SearchTriggers (sidebar and mobile nav), each
with its own window keydown listener, so Ctrl/Cmd+K opened two
SearchDialogs. Their body scroll locks restored out of order, leaving
body overflow stuck at 'hidden' after closing. Route the hotkey to the
first visible trigger so only one dialog mounts.
Fixes#42903
Co-authored-by: Claude <noreply@anthropic.com>
* fix(quick-open): match file names across Unicode normalization forms
Quick Open compared the typed query and the listed paths code unit by
code unit. macOS often stores Hangul and accented names decomposed (NFD)
while keyboards type them composed (NFC), so those files never matched.
Compose both sides before matching; the returned path stays as listed.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
* fix(tab-bar): compose new-tab file queries like the prepared file index
* fix(quick-open): recompose comparison keys after lowercasing
---------
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
Co-authored-by: Kelvin Amoaba <97001695+AmoabaKelvin@users.noreply.github.com>
* fix(chat): show deferred Claude permission picks
* fix(chat): clarify deferred permission picks
* fix(mobile): keep deferred permission picks readable on phones
Narrow mobile values show a compact 'saved · next message' label, accessible
names keep the full saved/running sentence, and the permission drill-in
explains the running mode in a wrapping caption.
* fix(mobile): narrow deferred permission summary and drop contradictory full-access name
Narrow to select options inside the deferred summary for the mobile typecheck,
use the plain model label while a pick is deferred, and assert labels instead
of English copy in the picker tests.
* fix(mobile): explain a deferred permission pick in the home row subtitle
QA on a 402 pt iPhone still clipped the one-line 180 pt row value. Show the
saved mode as the value and wrap the running mode in the row's subtitle.
* test(mobile): drop a type assertion in the picker test
* fix(worktree): refresh the base without merging into a checked-out branch
Creation already starts from the fetched remote-tracking commit. Keep Local
Main Up to Date now moves the local branch only when no worktree has it
checked out, with a compare-and-swap update-ref, so creation never takes
another checkout's index.lock or a checkout admission slot.
* fix(worktree): recheck checkout ownership right before moving the base ref
* fix(worktree): let Git refuse a checked-out base branch when fast-forwarding it
* fix(worktree): keep the base ref update a compare-and-swap, checked under one admission grant
Revert the local fetch: it overwrote FETCH_HEAD, failed when the file
transport is disallowed and protected nothing extra on Git before 2.35.
The final ownership check now runs with update-ref under one admission
grant, and a branch a rebase or bisect has reserved counts as checked out.
* fix(worktree): settle routing before the base-ref grant and count reserved branches as in use
* fix(worktree): check base-branch ownership from Git's files right before update-ref starts
Holding one admission grant across commands could deadlock on WSL routing
probes. The write now takes its own admission and, once admitted, rereads
worktree HEAD and rebase/bisect files before Git starts.
* fix(worktree): treat any worktree HEAD naming the base branch as in use, and recheck before the WSL fallback
* fix(worktree): fast-forward a checked-out base in its own clean checkout after the create returns
Creates stay on the fetched remote ref and never wait on local main. Once a
create succeeds, a detached, admission-exempt job (one per repo, coalesced)
fast-forwards a checked-out base in its owner checkout with the ff-only safety
flags, only when that checkout has no tracked changes; Git refuses a held
index.lock and overwritten untracked files. It has no deadline, so it is never
killed mid-checkout. Not-checked-out bases keep the compare-and-swap update-ref.
Restores the setting copy, and drops the unused relay outcome parsers.
* fix(worktree): skip checked-out history walks during create, defer sparse creates' base update, add Windows long paths
The create-time refresh now stops at a checked-out base before rev-list, so
only the admission-exempt background job walks its history. Sparse creates
start the update after their own checkout. The owner merge gets Windows long
paths like the create's checkouts, and no longer holds the app-wide ref
maintenance pause while a filter stalls.
* fix(sidebar): stop experimental worktree cards growing an icon-only details row
On the experimental cards with Details row on, link/issue/port icons moved
into the details row and opened it by themselves, so cards with nothing
else for that row got a second line that was blank except for icons on the
right. Icons now join the details row only beside another detail, and
otherwise stay on the title line.
* test(sidebar): type the details-row fixture properties and drop an unused import
MobileNativeChatMessage keyed each prose block by a running offset that its proseRow closure advanced. Paced prose calls that closure from its own renders (reveal steps, the scroll growth hold), so every such render continued from the last total and handed the code block a new row key, which dropped its wrap choice. Key the block by its index instead.
Advertise ACP form elicitation for Antigravity so its ask_question tool
sends forms with a per-question q<N>_other write-in field instead of an
interaction_ permission, which can only select an offered option. Pair
<key>_other companions with their choice field, and drop the server's
count-only form message so the card falls back to the first question as
its heading. Servers without form support keep the interaction_ path.
Fixes#42879
Let users name, rename and reset labels for saved Claude and Codex accounts without changing credentials or restarting agents. Keep the sign-in email visible across Settings, account switchers, chat usage, CLI and phone surfaces, and identify the account in removal confirmations.
Tracked in #21182.
Co-authored-by: Wooseong Kim <innocarpe@gmail.com>
* Remove finished-agent retention and keep sleeping status with panes
* Fix legacy pane projection and sleeping-agent regression fixtures
* Cover sleeping display-only status in parked terminal coverage
* Keep pane conversation identity only across continuous agent observations
* Remove duplicate import and keep sleep fixtures type checked
* Preserve conversation boundaries across replays and inherited clocks
* Cover conversation map stability across a large status batch
* Bound conversation capacity work to changed pane owners
* Pass capacity classifiers their layout input directly
* Finalize conversation provenance once per status batch
* Keep paired snapshot tests focused on observable outcomes
* Type the conversation provenance test fixtures explicitly
* Consolidate agent session imports after main merge
* Preserve provider replacement evidence when resetting conversation continuity
* Separate retention deletion from sleeping agent display
* Restore migration row bucket comparison
* Skip unusable pane identities before title ownership lookups
* Refresh conversation titles when their owning pane changes
* Update completion test sound mock after module move
* Keep finished agent status with sleeping panes and update agents by key
* Forward sleeping pane status and scope paired tab updates
* Preserve keyed selectors and compiler invalidation through final checks
* Keep bound store hooks visible to the compiler and measure published links
* Install provenance measurement before factory capture
* Use a static store hook and name presentation metadata once
* Preserve keyed freshness and atomic agent status publication
* Preserve remote folder dashboard regression after suite retirement
* Model pane wake before live identity reobservation in regression
* Share keyed row transactions and finish main integration
* Align runtime fixtures with keyed agent status reads
* Keep status projections addressed and activity summaries current
* Retire empty workspace status summaries after their source leaves
* Use the standard store fixture for parked terminal graph tests
* Align status fixtures with app types and a fixed test clock
* test: capture stored terminal render failure details
* Read terminal titles and completion prompts from keyed agent status
* Compose foreground agent state in worktree test stores
* test: complete lineage list agent state fixture
* Keep sleeping completion when retiring process authority
* Run durable agent-hook status tests with Node SQLite
* Preserve sleeping Done when renderer retires its process
* Update terminal fixtures for process retirement
* Spike addressed renderer agent status publication
* Join status only against admitted mobile membership
* Scope status membership reconciliation and observe bindings by key
* Cover status admission and scoped replica recovery
* Name the status replica runtime owner type
* Read sleeping pane resume sources by key
* Bind sleep capture fixture to its live panes
* Bound status replica retirement and recover interrupted publication
* Pass the pane tree to the sleep capture fixture graph
* fix(status): skip non-UUID leaves, release phone frames on headless restore, avoid unbound tree scan
- Graph status bindings and replica membership skip leaves that can't carry a status row instead of throwing.
- The pre-seed phone gate only holds frames for the renderer generation that owes the seed; restoring headless authority redelivers the held worktrees.
- Status capture skips the leaf count when the tab has no terminal, and counts saved leaves only up to two.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(status): release held phone frames whenever the unseeded renderer loses the graph
Desktop builds without a headless fallback mark the graph unavailable but keep the renderer generation, so the pre-seed gate stayed shut until a window reopened.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* perf(status): reuse the capped saved-leaf count per immutable layout
Status churn on an unmounted tab with only a tab-level terminal no longer descends a deep saved split tree on every change.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Preserve accepted completion identity through sleep
* Pass a nullable title to the status classifier
* Resolve accepted host status once and preserve delivery order
* Drop published sleep status when its host owner retires
* Use the supported remote completion test envelope
* Preserve published activity when filling completed status
* Update imports for main module moves
* Remove an unreachable native chat owner comparison
* Read host and unread card inputs by address instead of per root
A host-scoped sidebar card re-read every root's row and acknowledgement on
any store write, so one card with many sleeping roots paid for other
workspaces' updates. Card host filtering, the summary pill's unread state
and the workspace unread gate now first compare the workspace's projected
row revision and the changed keys of the tab-keyed layout/PTY maps and
the acknowledgement record. Only an addressed change re-reads the roots.
Acknowledge and unacknowledge annotate their writes so readers can skip
keys they don't show.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Correct sleeping pane test imports after module moves
* Combine duplicate subscription test imports
* Record addressed acknowledgement deletions on row drops
* Type subscription fixture key set for string readers
* Keep accepted updates complete after observer failures
* Seed the send picker with its SSH terminal identity
* Correct status publication imports after preload module move
* 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>
* Keep a host card's unchanged agent rows after filtering out another host's pane
Re-lineage clones every row once a host-scoped card filters a row out, so an unrelated pane's status write rebuilt the card's row array and re-ran its signal and rollup memos. Reuse unchanged rows as main's sidebar row selector does.
* Document the runtime stub cast in the status admission test
* Await graph publication in the phone projection status tests
* Share the renderer status binding check and move graph tab-id validation out of the window graph
Main's async graph publication grew session-window-graph past the 300-line limit; the binding predicate now has one owner beside its capture path.
* Seed the Copy Session ID shortcut test through keyed agent status
* Restore the reviewed status binding checks and move the remount check instead
Keeps the capture path's lookup behind its UUID and changed-worktree guards; the window graph stays under the line limit by moving the PTY remount check to the graph promotion base.
* Seed the graph payload partition test through keyed agent status
* fix(agent-status): retry status publication after a failed or superseded graph sync
An inactive sender now asks for a graph sync on each status change (base got this
retry from agentStatusEpoch in the sync key), and a superseded ack parks the sender
until the next accepted graph instead of stopping it.
* fix(agent-status): don't stage status changes while a snapshot is pending
A parked sender kept tombstones for every removed occupant; the next snapshot
recaptures every row and main clears its replica on snapshot.
* fix(agent-status): don't start renderer status publication in paired web clients
The web stub answers every batch with superseded; with superseded now re-arming
on each accepted graph, a web sender would resend snapshots and request graph
syncs for nothing.
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
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 <alpha-eng@stably.ai>
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>
Folder workspaces were only rendered (and listed by `orca worktree ps`) when
their owning project group had a parentPath. Manual groups created without a
base folder have parentPath null, so their folder workspaces disappeared. Gate
on the group being visible instead of on it having a base folder.
Fixes#42865
Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(terminal): drop saved scrollback of deleted worktrees in history GC
* fix(terminal): keep session folders behind dangling links, WSL paths, empty mount points and saved tabs
* fix(terminal): keep session folders of renamed worktrees and unproven mounts
* refactor(terminal): read the ENOENT code through the shared errno helper
* fix(terminal): match sibling worktree parents without case folding
* fix(terminal): count only real sibling directories and recheck before pruning
* fix(terminal): recheck each gone worktree once per parent probe and honor abort while pruning
* fix(terminal): re-probe the proving sibling per worktree and stop pruning once cancelled
The turn bar appeared and ticked while the user's message still said
"Sending…", because the host reports working as soon as it owes a send.
Once the turn opened, its clock also counted from the send press
(requestedAt), so the bar settled already showing the send time.
The live clock and "Worked for" now count from the provider turn start,
and neither the bar nor the activity line shows while the live turn's
opening message is still sending. Mobile has no "Sending…" label, so it
waits on the same pending submission to keep live and settled times equal.
* fix(native-chat): import the moved Codex async question contract
#42459 imported codex-async-question-input from its pre-#42620 path, which
breaks the orcad template build on main.
* fix(native-chat): keep the status projection within its line limit
#42459's awaitsAnswer field took structured-agent-session-projection.ts to
301 lines; fold the attention early return into the status ternary.
* feat(browser): remember "open another app" choices per site
The prompt offers an unchecked "Always allow <site> to open <scheme> links".
A saved choice opens the app without asking only on the desktop user's own
fresh click, one app per click. Pages that can't be named (sandboxed or srcdoc
frames, no referrer) are remembered for the tab's site; loads the browser
started itself are never remembered. Only https, loopback and local-file sites
qualify. Choices live in the existing site permissions, so they are per
browser profile, listed and revoked in Settings > Browser > Site permissions.
* test(browser): don't pin the platform-dependent app name
* fix(browser): spend the click a prompt answered, and let agent clicks ask
* fix(status-bar): usage follow-ups for label order, Fable-only rows and spend tone
- Compact status bar puts named pool labels first, like Detailed (Cursor Models 64%).
- A roster row with only a Fable/MCP reading shows that value in place of Not checked.
- Fable figures in the roster carry a title naming the weekly limit, following Used/Left.
- The extra-usage spend bar exposes its tone as data-usage-tone.
- Tests pin a Fable-only reading through failed and in-flight refreshes.
* fix(status-bar): title the Fable figure on the roster's Detailed line
Adds managed Windows x64/arm64 Antigravity Chat runtimes, keeps compatible host agy selection, and makes sign-in prove the saved token works on a fresh server (Google can grant fewer scopes than requested, leaving a token that can never refresh).
Builds on Lami's Windows host-CLI admission work in #40344.
Co-authored-by: Lami <154405627+Lqm1@users.noreply.github.com>
* fix(session): drop a removed workspace's saved session state everywhere
Workspace removal is now driven by one per-field table, main's terminal-membership
rebase no longer re-writes empty rows for worktrees the window stopped listing, and
the post-scan local prune also drops saved session state of worktrees whose folders
are gone (keeping unsaved drafts, archived, active and otherwise-owned workspaces).
* fix(session): address review pass 1 on deleted-workspace session cleanup
Match Windows path casing for protected metadata, filter cleanup candidates before
probing paths, batch recency deletion, count only references removal deletes, and
keep paused WSL moves out of repo removal.
* fix(session): keep repo removal's existing WSL freeze guard placement
* fix(session): count only row owners removal can delete
* refactor(session): let repo removal rely on shared owner removal
Drop repo removal's per-field loops that the shared owner removal already covers,
remove type assertions from the new tests, and lower the layout-writer baseline
(net four fewer main-process writers; the new cleanup module adds two).
* refactor(session): read session record rows through a type guard
* fix(session): drop default-tab flags of worktrees a loaded repo no longer lists
QA found the window kept these flags for deleted workspaces and saved them back
after main removed them, so cleanup kept re-selecting those workspaces.
* fix(session): address CodeRabbit on deleted-workspace cleanup
Match unsaved-draft owners path-equivalently, drop rejected unified tab and group
records when the rebase skips a dropped worktree, and pin the POSIX test fixtures'
platform.
* fix(session): address council review on deleted-workspace cleanup
Keep default-tab flags any worktree list still shows, read only eligible
candidates' files for the draft guard, drop a dropped worktree's active-tab entry
in the rebase, and save only the local session after cleanup.
* fix(session): keep a workspace marked archived without an archived checkout
* fix(session): follow the moved workspace-scope module after merging main
SwiftUI controls advertise only AXAttributedDescription but answer
AXDescription, so treat AXDescription as readable in that case and give
Calculator keys and System Settings sidebar buttons their labels.
Render custom actions ("Name:X\nTarget:..\nSelector:..") as just their
name, and match perform-secondary-action on the exact pretty name first so
a custom "Delete" is not shadowed by AXDelete.
Refs #42828
Co-authored-by: Neil <neil@stably.ai>
* fix(notifications): hold Codex completion while its goal keeps running
Codex starts the next goal turn by itself, so a successful turn that
settles while the thread goal is active is not the end of the work.
The SQL observer now reports when the newest goal row is an active
Codex goal on a live generation, and the completion feed holds success
until that stops. Claude goals finish within one turn and never hold.
Co-authored-by: BingX <22970633+bing-xbx@users.noreply.github.com>
* fix(notifications): announce a goal turn that ends on an open async question
---------
Co-authored-by: BingX <22970633+bing-xbx@users.noreply.github.com>
* fix(git): retire native process groups after abnormal stops
* fix(git): drain worker output after group retirement
* test: retain descendant cleanup until liveness assertion passes
* feat(mobile): manage MCP servers from /mcp in chat
/mcp in a Claude or Codex chat on the phone opens an MCP servers sheet with
the same status, sign-in, reconnect and on/off controls as desktop. The
phone never runs the agent, so sign-in finishes by pasting the address the
browser lands on. The sheet is offered where the host reports
agent-session.mcp.v1 and the chat's agent manages MCP servers, and it stays
open through a reconnect so a sign-in in progress isn't lost.
* feat(mobile): show /mcp as a tray on the composer instead of a sheet
/mcp now opens the MCP servers list as a tray docked on the composer's top
edge, the same box the goal strip uses. The header shows the counts (or the
sign-in in progress) and a close button; rows are one line with the status
dot, a truncating status, Sign in / Reconnect and the switch. The list is
capped at 40% of the chat's visible height, which shrinks with the
keyboard, so the paste field, tray and composer stay on screen.
* feat(browser): free idle serve browser tabs under Memory Saver and restore them when used again
On orca serve, agent-opened browser tabs kept a renderer process each for as
long as the host ran. Serve tabs now follow Settings > Browser > Memory Saver
(on by default, 4 hours): a saved tab with no command, capture or viewer for
that long gives back its renderer. It stays listed and its next use starts it
again under the same page id with its back/forward history. Pages that
page-less commands route to, or that hold request routes, a HAR capture, a
dialog, popups or what the desktop Memory Saver keeps, are not freed.
* fix(browser): apply Memory Saver changes to idle serve tabs and close freed tabs fully
* test(browser): spend the first offscreen window before counting frames under Xvfb
* fix(browser): remember an all-requests route so idle freeing keeps its page
* fix(native-chat): keep chats an older Orca saved after a downgrade
After upgrading to the SQLite chat store, going back to an older Orca and
upgrading again hid everything the older build wrote: chats it started never
appeared, and turns it added to chats already converted never showed.
Each chat now records the original journal revision it was converted from.
On every launch, before the store opens, chats whose original moved on are
kept: a chat this file hasn't converted joins as itself and converts from the
current original; a converted chat the older build continued is kept as a
second chat beside it holding the older build's history. Nothing is deleted
or merged, and the original journal is never written. Attachments the older
build claimed stay claimed so the sweeper keeps their files.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(native-chat): retry chats an older Orca saved until they are kept
- A failed chat leaves the original unabsorbed, so the next launch retries it
instead of taking its revision as shown.
- Every chat gets a baseline revision at the first copy (record-only chats
too), so a chat deleted here stays deleted unless the older build writes it.
- Chats converted before revisions were kept are filled in once, and only
while the original is unchanged.
- The original's tabs and claims are read once per pass, and the pass is
skipped while the original's file identity, now including its WAL's, is
unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(native-chat): fill in revisions only for an untouched original, and show kept chats
- Chats converted before revisions were kept are filled in only when the
original's identity matches exactly (a clean WAL); any other case runs the
full pass, which can only add chats.
- A chat new to this file gets a tab; one it already holds keeps its
visibility here. A file that never recorded its tab index records it with
every chat shown, so a kept chat isn't hidden behind client-derived tabs.
- Claims are grouped without copying arrays.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(native-chat): show every chat an older Orca wrote, even one closed here
A chat closed in the new build and then continued in the older build stayed
hidden; it now gets its tab back. The test's row helper uses a type guard
instead of a cast.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore: keep the audit-tools lockfile as on main
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(native-chat): reread kept-chat tab state after a failed chat; no extra stat
A chat whose transaction rolled back left the pass's in-memory tab state
ahead of the file, so a later chat could skip recording the tab index. The
pass now rereads it after a failure. The original's absence comes from the
receipt's own stat instead of a separate sync existence check.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* test: keep the chat store's original-journal provenance when seeding record state
The seed helper cleared all store metadata, including the receipt and per-chat
revisions only the chat store writes, so a later launch looked like a file that
had never recorded them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(native-chat): recheck the original before taking its revisions as current
Revisions filled in for files converted before revisions were kept now commit
only if the original's identity is unchanged after reading them, so a write
by another process mid-pass is absorbed next launch instead.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(native-chat): manage MCP servers from /mcp in Claude and Codex chats
/mcp in a Claude or Codex chat now opens an MCP servers panel instead of
sending the text to the agent, where SDK mode only printed a one-line
summary. The panel lists each server's status and lets you sign in,
reconnect, and turn servers on or off.
Claude goes through the Agent SDK's control requests on the running chat,
including the runtime-only mcp_authenticate and mcp_oauth_callback_url.
Codex goes through app-server: mcpServerStatus/list, mcpServer/oauth/login
and config/batchWrite, with short-lived app-servers for chats at rest.
Sign-in opens the browser; when the agent runs on another host, the user
pastes the address the browser lands on.
* feat(native-chat): show /mcp as a tray on the composer instead of a dialog
/mcp now opens a tray attached to the composer's top edge, built on the
same tab strip as the goal and worker strips, so it stacks under them.
The header sums up connected, sign-in and failed servers; rows are one
line each with the action and switch on the right, and sign-in details
sit inline under their row. Esc or the close button returns focus to the
draft, and switching away from the chat closes it.
The modal wiring is gone, and the usage modal is back to its own request
type now that nothing shares it.
ACP agents can ask the user to open a sign-in page with URL-mode
elicitation (elicitation/create mode "url", then elicitation/complete).
Orca now advertises URL elicitation to ACP chats and shows the request as a
"Sign-in requested" card with the link's host, Open sign-in page and
Decline. Open replies accept once the system browser opens the page; the
card collapses to "Sign-in finished" when the agent sends completion.
Grok forwards its MCP servers' elicitations through its own
x.ai/mcp/elicit extension, which Orca refused before. URL requests now use
the same card and form requests use the existing question card.
* fix(browser): keep session cookies across restarts
Electron never writes cookies without an expiry to a partition's cookie database, so sites that sign in with one sign out every time Orca restarts. Orca now saves each browser profile's session cookies (sealed with the secret store) and puts them back at launch, before restored tabs load.
* perf(browser): restore and save session cookies without a renderer per profile
Launch restore opened a hidden CDP renderer for every browser profile with
saved cookies, two at a time, before the first window. With 51 profiles that
added 1.7 s to launch. Unpartitioned cookies now go back through Electron's
cookies API; only partitioned ones, or ones whose scope the jar already holds,
still use CDP.
Saves opened a hidden renderer per profile every 30 s while a site rotated a
session cookie, with no limit across profiles: 51 profiles meant 51 renderers
at once. A save where each changed scope's one cookie was only set again now
takes the new value from Electron's cookie list. Other saves read through CDP,
at most two at a time. A save also no longer reschedules itself after its
profile is deleted.
* fix(browser): keep the session cookie file in step with failed writes and reads
A save recorded its cookies as saved before writing them, so a failed write
made the retry skip the write. An unreadable file read as empty, so a later
save with no session cookies skipped the write and left the old file to be
restored on a later launch. Tracking now starts before the file is read, so
a cookie change during a slow read is not missed.
* fix(browser): hold session cookie saves until the launch read is in
Tracking starts before the saved file is read, so a save could run against an empty baseline and the late read could then replace a newer one. Saves now wait for the read, and quit leaves a file that is still being read untouched.
* fix(i18n): translate 1,302 missing Japanese UI strings
Fixes#42027
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(i18n): correct Japanese usage error grammar
* fix(native-chat): resolve moved async question module in journal status
* style(i18n): format merged Japanese settings groups
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Neil <neil@stably.ai>
* fix(computer): click macOS windows that refuse focus with --restore-window
A raised window that never becomes the focused window (Archicad's CEF start window) failed every --restore-window click with window_not_focused, because the click fence required the target to be focused as well as topmost at the point. A real click reaches the topmost window whether it is focused or not, so when the target stays unfocused the click now goes ahead while the target is topmost at the point, and is refused before anything is sent when another window covers it. Afterwards the app the user had in front comes back if the raise or the click left the target app in front without the target window.
* fix(computer): restore the front app only after a click on a window that refused focus
* fix(computer): keep the target app in front when an unfocused-window click opened something
* fix(computer): count small popups and re-read the front app last before restoring it
Electron's input-event only observes the main frame's widget, so a click in an
out-of-process iframe never released a quieted prompt, funded a popup or ended
an agent's lead. Mouse activation now comes from before-mouse-event, which
fires for every frame.
* feat(computer): add opt-in flags for smaller computer-use screenshots
Every observe and action screenshot came back as a full-resolution PNG, so a Retina window cost about four times the pixels it needed. --screenshot-scale caps pixels per window point, --screenshot-max-edge caps the longer side, and --screenshot-format jpeg re-encodes it, on macOS, Windows and Linux. Defaults are unchanged.
* fix(computer): treat requested screenshot sizes as ceilings and skip repeat encodes