Keep the previous accepted wait as the in-session draft, give the help text more space, and explain invalid input without the numeric safety limit.
Co-authored-by: Taras Kornichuk <dread9ko@gmail.com>
Retain mounted controls while reconciling accepted choices, let users edit the wait before enabling notices, and clarify execution-host scope.
Co-authored-by: Taras Kornichuk <dread9ko@gmail.com>
useParkedTerminalWatcherSynchronization routed its sync through
useEffectEvent. React 19.2.x only refreshes effect-event closures for
plain function components, and the hook's caller
(TerminalPaneOverlayLayer) is memoized, so every sync used the first
render's parkedTabIds/terminalTabs. Tabs deferred at startup kept their
parked watcher (and hidden-delivery claim) after being revealed, and
tabs parked later never got a watcher.
Use the existing useLatestHandler instead, which refreshes its ref in a
layout effect before the passive sync runs.
Fixes#42020Fixes#42023
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
Add the missing zh.json entries for the Add project menu
(auto.components.sidebar.AddProjectMenu.*), the contextual-tour "New"
badge, and the collapse-all-projects tour step. Keys follow en.json
ordering and preserve the {{host}}, {{modifier}} and
{sidebar.projects.toggleCollapse} placeholders.
Deleting a browser profile with extensions forgot the extension API host
twice, and Electron's protocol.unhandle throws once the scheme is gone.
That throw skipped the partition storage wipe and the persisted extension
store cleanup.
- Only unhandle orca-extension-api while it is still registered.
- Forget the extension state, sync storage, and DNR stores in finally.
- Log a detach failure on profile delete and still wipe the partition.
Fixes#41907
Co-authored-by: Neil <neil@stably.ai>
* fix(browser): keep the user's agent-browser env and config away from desktop helpers
* fix(browser): drop an inherited agent-browser namespace from desktop helpers
* fix(browser): write the desktop helper's empty config off the main thread
* fix(browser): wait for the Orca agent-browser config before the startup sweep
* feat(daemon): capture macOS's verdict when a terminal host is refused a folder
On a confirmed app/host divergence in Documents, Desktop or Downloads, read
macOS's own record once per host incarnation and folder class: tccd decisions
and kernel sandbox denies for the host pid, host signature checks, host age
and origin, earlier good reads, time since wake signals, Full Disk Access,
fork-fallback reason, and the class root's dataless, iCloud and volume state
(read before the app-side opendir). Telemetry is enum-only in a new
daemon_pty_cwd_denial_capture event; raw filtered lines go to a capped local
file under userData logs. Every subprocess is bounded and failures are
swallowed off the spawn path.
* fix(daemon): keep the wake clock Electron-free and avoid casts in the log parser
* fix(daemon): tighten denial capture attribution and bounds
Snapshot good-read and wake context at the denial, bound the class-root
realpath, keep earlier captures when one entry is over the cap, require the
app or host in requirement failures, let an explicit pid outrank a shared copy
path, and report a capped log read as truncated.
* fix(daemon): collect Info requirement checks and claim captures before probing
Ask log show for Info records in a 30 s window ending at the denial, claim a
helper/class capture before the folder probes and release it when the app is
refused too, and report a requirement failure tied only to the shared app
identifier as uncorrelated instead of blaming this helper.
* fix(daemon): keep folder-denial capture in telemetry labels only
Drop the local diagnostics file and stop retaining raw log lines: the parser
reads them in memory to derive enums and keeps nothing.
* fix(daemon): skip the folder-denial capture when telemetry is off
Gate the folder probes and the log/codesign capture on the telemetry client's
consent check so opted-out users never have their log read or checks run.
* fix(daemon): recheck telemetry consent at every capture stage
Run every capture subprocess through a consent-checked runner (now a required
probe argument), ask again before and after the identity probe and after the
remaining probes, and free the claim when an opt-out stops a capture so a
later opt-in can still capture.
* fix(daemon): recheck consent after the access reader's lazy import
The Full Disk Access reader started after its import resolved without asking
again, and its catch swallowed the opt-out; ask after the import and let the
opt-out end the capture so its claim is freed.
* fix(daemon): report a helper's denial over a later allowed result
Overlapping tccd requests can resolve out of order; when any result for the
denied folder class is a denial, report it.
* fix(tasks): show cached GitHub rows instead of the skeleton on filter change
Filter handlers set tasksFiltering and only the forced fetch settlement
cleared it, so rows the landing run had already painted from the
per-query cache stayed hidden behind the skeleton. The landing run now
clears the flag when every selected repo (or the restored page) is
cached for the applied query; the forced refetch still runs under the
toolbar spinner and replaces the rows in place.
Rebuilds #41998 by @piotrpawlik (fixes#41997).
Co-authored-by: Piotr Pawlik <piotr.pawlik@apollo.io>
* fix(tasks): keep GitHub Tasks pages in issue-number order
materializeTaskPageItemList re-sorted every page by updatedAt while the
host search, the cross-repo aggregate, the landing pre-paint and mobile
all order by issue number. A triage edit moved old issues to the top,
and the landing probe saw a different key order and reset pagination.
Use the shared sortWorkItemsByNumber so retained pending rows keep
their number slot too.
Rebuilds #41999 by @bbingz (fixes#41987).
Co-authored-by: Bing.Z <zzb@gxsmjx.com>
* test(tasks): make landing tests revert-sensitive; scope refetch comment
Address C1 review: every landing case now fails without the cached-rows
fix, names say what they verify, and the comment no longer claims typed
search forces the refetch.
* test(tasks): replace landing test casts with annotations
* fix(tasks): hold GitHub pagination while a forced refresh is in flight
Cached rows now show during a filter switch's forced refetch, which made the
pager live; opening page 2 then got undone when the refetch replaced every
page. Disable the pager until it lands, as the phone does while revalidating.
Also drops two landing tests that covered no new behavior.
* fix(tasks): hold GitHub pagination for any in-flight landing fetch
A typed search over stale cached rows lands without a forced fetch, so the
pager was still live and its page could be undone when the landing fetch
replaced the pages. Track the landing fetch itself, as the phone holds paging
while any list load revalidates. Shares the retry-key release between the
fetch's settle paths to keep the run under the file-length limit.
---------
Co-authored-by: Piotr Pawlik <piotr.pawlik@apollo.io>
Co-authored-by: Bing.Z <zzb@gxsmjx.com>
Co-authored-by: Brennan Benson <79079362+brennanb2025@users.noreply.github.com>
* fix(browser): keep extension tabs and windows an agent sets off from taking the user's view
* fix(browser): follow the agent-lead rename for extension callers
* fix(browser): let the newest input or extension gesture decide who set an extension call off
Preserve explicit choices independently of the product default and let users
choose the quiet interval in Orchestration settings. Pause live watches
without resetting episode budgets, ordinals or admission deadlines.
Co-authored-by: Taras Kornichuk <dread9ko@gmail.com>
* fix(computer): type into Windows text fields without raising the window
Adds type-text --element-index on Windows through EM_REPLACESEL for user32,
RichEdit and WinForms edits (owner-checked, read back with WM_GETTEXT), or a
ValuePattern write merged with the TextPattern selection. Keyboard verbs on a
background window now refuse with background_unavailable before sending
anything and point at type-text --element-index. Foreground typing uses one
SendInput KEYEVENTF_UNICODE batch instead of WM_CHAR posted to the top-level
window, and get-app-state shows multiline edit text again.
* fix(computer): refuse typing into Windows list items, keep the caret on a confirmed prefix
* Give OpenCode the component fixtures MCP server
The repository registered orca-component-fixtures only in .mcp.json, so
Claude Code got the fixture tools and OpenCode did not. opencode.json
already mirrors the repo's TypeScript language server for OpenCode; it
now registers the same fixtures server, in the per-server form the rest
of the file uses, which OpenCode 1.x reads and OpenCode 2 migrates.
Tracked in #41836 (STA-15139).
Co-authored-by: Anubhav Banerjee <76914348+Anubhav-2003@users.noreply.github.com>
* Find the fixtures server from any OpenCode start folder
OpenCode reads opencode.json from the nearest ancestor but starts MCP
servers in the folder it was launched from, so a relative script path
failed outside the repository root. Walk up to the script and import it.
---------
Co-authored-by: Anubhav Banerjee <76914348+Anubhav-2003@users.noreply.github.com>
window.print() opens the system print dialog and, on Windows, a passkey request
opens Windows Hello. Both are modal windows that take the foreground, and an
agent click could set either off over the user's work. Main now tells each
page's frames whether an agent drove it since the desktop user last gave it
input, and the session frame preload refuses print() on such a page (reported
to the agent as effects.blockedPrint, pointing at pdf) and, on Windows, holds
its passkey requests unsent, still reported as passkeyRequest. The user's own
input releases the page and cancels held requests so the page can retry.
* feat(automations): record subpages in app navigation history
* feat(automations): unify page layout and redesign list, detail, and runs
* chore(i18n): remove obsolete automation page strings
* fix(automations): keep focus effects compatible with React compiler
* fix(automations): resolve run workspaces from historical ownership
* fix(automations): preserve replay and guard in-flight dispatches
* fix(automations): give the All runs link a singular default
* fix(automations): let Escape step back to the parent history entry
Escape returns to the parent page. When that page is the previous history
entry, move the index back instead of appending a duplicate, so Forward
still reaches the page that was left.
* chore(i18n): sync the All runs runtime catalog entry
* fix(automations): keep Back/Forward replays from re-recording the page left
- Record a sub-page only when the page itself changes. Clearing a replay
request could render once with the old page state and record the page
Back just left, cutting Forward history. Adds a hook-level test.
- Escape on the list closes the page without arming the step-back mode,
and closing resets it.
- Run now's busy state only clears from the open automation's runs, and a
dispatch that never answers releases the button after 60s.
* fix(automations): truncate a long automation name on the run-not-found button
* test(automations): type the history hook test's props
* fix(automations): address C1 review and UX-I1 findings
C1 (code review):
- C1-01: a Back/Forward replay that settles on another page (its run
history could not be read) replaces its history entry instead of
appending, so Forward history survives.
- C1-02: Run now's busy state is request-specific; a late failure or
acknowledgement for an earlier click no longer clears a later one.
- C1-03: the page title takes focus when the user opens Automations
(sidebar, deep link, Back/Forward); a page restored at launch or on
reload leaves focus alone.
- C1-04: returning to an off-screen run scrolls the virtualized table to
that row before focusing it.
- C1-05: a deleted-run deep link records the run it asked for, so
Back/Forward return to the same not-found page.
UX-I1:
- One "Open workspace" label on the run page.
- Run status is a dot and label everywhere; the run header uses a soft
tinted pill.
- Every failed run gets one plain-language banner. Blocked-agent reasons
are translated (trust prompt: "<Agent> is waiting for you to trust this
folder", with Open workspace), never shown as internal codes.
- A run that started off its slot shows "Scheduled … · Started …"; the
detail's Last run says "Started …" in that case.
- The three stats stay in one row down to a 480px container.
- "—" instead of "n/a" for spend and tokens.
- An automation's Runs page drops the "Run history" header and shows the
run count under the title.
- The not-found button reads "Open automation".
* fix(automations): address CodeRabbit review
- Run now tracks one request per automation, so starting another
automation's run never frees this one's busy state.
- Back/Forward to the list or Runs keeps the list search.
- While Back/Forward resolves an automation, show the loading skeleton
instead of leaving the page being left interactive.
- A new run request clears an earlier run-not-found result.
- A stale external job key no longer opens a different job.
- Spend reads "Unavailable" when run history could not be read.
- The external row's accessible name includes its host when the Host
column is shown.
* fix(automations): address internal review round 2
- A Back/Forward replay that falls back to the page just before its entry
steps back to it instead of leaving a duplicate Back step.
- The detail's Last run says "Started …" only for a run that started.
- A failed run with no recorded reason still gets a banner.
- Back/Forward to the bare Automations entry focuses the page title.
- Remove the unused status-badge variant helper.
* fix(automations): address UX-I2 findings
- UXI2-1: the detail's Last run stat shows the run's scheduled time, the
same time its rows show; only the run page labels Scheduled/Started.
- UXI2-2: a row you return to (list, rail, Runs tables) is centered when it
was off screen and gets a brief tint that fades over 1.5s, since focus
set from code after a click draws no ring. Focus survives a table that
is still re-rendering, and waits up to 3s for a reloading list.
- Trust-blocked runs use the primary Open workspace button; empty output
reads "No output."
* fix(automations): address C2 review and UX-I3 findings
C2:
- C2-01: a Back/Forward replay that cannot show its run (history read
failed) records nothing, so the entry keeps the requested run and
Forward history is untouched.
- C2-02: the list and global Runs pages focus their search field
("Search automations" / "Search runs") on user navigation; the list
search no longer autofocuses, so a restored page leaves focus alone.
- C2-03: returning to a run waits up to 5s for the table's data to list
it (cancelled when the page changes or the user focuses elsewhere),
instead of a fixed frame budget.
- C2-04: "Started" appears only with launch evidence (dispatch time, or
the start time of a completed/dispatched older record); skipped and
missed runs show their slot only.
UX-I3:
- UXI3-1: the list, detail stat, rail and run tables all show a run's
listed time: its scheduled slot, or a folded run's latest occurrence.
The usage summary carries it as lastRunListedAt; older hosts fall back.
- UXI3-2: a returned row that is already visible stays put; only an
off-screen row is centered.
* fix(automations): address CodeRabbit round 2 and UX-I4 note
- Spend reads "Unavailable" whenever the latest history read failed,
instead of totals from the retained earlier read.
- The run a table reveals on return becomes its current row, so arrow
keys continue from it.
- Back/Forward to Runs spans every host, as the Runs tab does.
- Revealing a returned run scrolls only as far as needed (align auto).
* fix(automations): address C3 review
- C3-01: "Started" needs real launch evidence. A failed run counts as
launched only if it reached a chat session or terminal; the chat
launcher's dispatchedAt on a refusal or early throw is not a launch.
- C3-02: each Run now request has its own deadline. An answer that
arrives after the request's 60s dispatch deadline ends it instead of
starting a fresh 20s wait, whichever automation is selected.
- C3-03: the returned-to row stays focused through the whole readiness
budget, including a refetch that clears the cached row and lists it
again later; the reveal is re-requested while it is missing.
- C3-04: focus and reveal use the host-qualified run key in the global
Runs table, so the same run id under two hosts returns to the right
one.
- C3-05: once the user focuses anything else, the repair stops for good.
* fix(automations): address C4 review and the rail listed time
- C4-01: a failed terminal run keeps its Started time after terminal
retention clears the pane key; terminal runs get dispatchedAt only at
dispatch. A failed chat run still needs a chat session.
- C4-02: when a refetch drops the returned row for good, focus lands on
the page search once the readiness budget ends instead of the body.
- C4-03: the run page's focus key is host-qualified, so moving between
same-id runs on different hosts focuses the new title.
- The Recent runs rail lists a folded run at its latest occurrence, the
time the list and detail show.
* fix(files): phone file opens no longer switch the desktop's view
files.open and files.openDiff passed an absent navigation target straight
to the window, which treats absence as a host switch. Phones send none,
so every phone file or diff open moved the desktop to that workspace and
the editor. Resolve an absent target to caller at the RPC; host/all from
the CLI's --focus still move the window.
* fix(mobile): Open in session selects the diff on the phone itself
The review screen relied on the desktop switching to the opened diff,
which only reached phones still following the host. Select the opened
diff for this phone before going to the session.
* fix(mobile): select only the working-tree diff and stop when the review screen closes
* test: drop type casts the changed-code gate flags
Add aria-labels to the UI zoom and terminal font size steppers and the
word separators input, name the scroll speed slider thumbs, and link the
visible Branch prefix, Arguments and Command template labels to their
fields via htmlFor/id.
Fixes#41604
* perf(pickers): window workspace comboboxes and the Cmd+J result list
The Automations and Skill install workspace comboboxes mounted one cmdk row
per workspace while open, and the Cmd+J result list mounted every capped
section row plus each See more page. cmdk moves the highlight, runs Enter and
publishes aria-activedescendant by querying mounted rows, so these lists could
not be windowed while cmdk owned selection.
Each surface now owns its highlight over the full ordered result list:
movement keys follow cmdk's bindings over every result, Enter and click share
one activation path, and the input's aria-activedescendant names the
highlighted option, which stays mounted. Rows render as plain listbox options
outside cmdk's registry and only the visible window mounts (TanStack
virtualizer; Cmd+J keeps natural layout below 50 rows). Filtering, ranking,
section caps and See more are unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* refactor(pickers): share the windowed command virtualizer and keep the compiler gate
Move the active-option pin and command-list reveal padding into one hook used
by all three surfaces, keep virtualizer consumers out of React Compiler like
the other TanStack lists, and express the dialog row styles and the flush skill
rows as command variants.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore(pickers): record the windowed command virtualizer hook's compiler bailout
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(pickers): keep virtualizer geometry across scrolls and match prior highlight/IME behavior
- Stable per-list key/size getters so scrolling and hovering no longer rebuild
every row's virtual geometry.
- Moving the pointer over the already-highlighted option does no work.
- Jump to… numbers every rendered option, disabled ones included.
- The task Cmd/Ctrl+Enter shortcut ignores IME composition keys.
- The skill install workspace picker opened with nothing chosen highlights the
first workspace again.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(pickers): let the shared virtualizer own row key identity
Callers now pass their ordered row keys; the shared hook keeps one key array
per ordered sequence and derives the key getter from it, so a list rebuilt
with the same rows (for example after an unrelated workspace metadata update)
no longer forces a full geometry rebuild, while real reorder, insert, remove
and filter changes still do.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(pickers): keep the row See more selects in view when Jump to… starts windowing
Crossing 50 rows through See more switches the list to windowed rendering;
reveal the newly highlighted row after that commit so it stays on screen as
it did before.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* perf(pickers): keep Jump to… on its original list; window only the two workspace pickers
The real-app benchmark showed no meaningful gain for the Jump to… results
(they rarely reach 50 rows), and windowing them changed the See more
behaviour at the threshold. Restore Jump to… exactly as on main and keep the
windowed list, model-owned selection and option primitive for the Automations
and Skill install workspace pickers, which measured clear wins.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ubuntu-latest now resolves to the Ubuntu 26.04 image, which installs the
amd64v3 package variant. Its GCC defaults to -march=x86-64-v3, so the
Electron-ABI node-pty in 1.4.257 shipped AVX instructions and the terminal
daemon died with SIGILL on CPUs without AVX. Pin the Linux x64 release
build to ubuntu-24.04, the image that produced the working 1.4.256.
Fixes#41637
Outline and ghost buttons in toasts and corner cards painted dark:bg-input/30
and hover:bg-accent from the app palette under popover text, and panels and
buttons drew edges from the app's --border and --input. For app themes the
notice scope now remaps --accent, --accent-foreground and --input onto the
readable secondary fill, and --border onto a popover-text edge. Dark outline
buttons get an inset edge, and --notice-border now follows the popover text.
Built-in themes paint as before.
* fix(computer-use): observe minimized, hidden and other-Space windows on macOS without a restore
macOS only enumerated on-screen windows and dropped minimized AX windows,
so an app whose window was minimized, hidden or on another Space (any
full-screen app) could not be observed, and agents fell back to
--restore-window, which activated every window of the app and switched
the user's Space.
Enumerate every window, keep the off-screen ones accessibility can still
read (other-Space windows by remote token, closed windows dropped by
their Space membership), and report their real minimized and off-screen
state. Background pointer and key input on such a window refuses before
sending anything. --restore-window now raises only the target window,
once per request, and delivery is measured from whether the target
became frontmost.
* fix(computer-use): pin the restored window and prefer visible windows by default
* fix(computer-use): keep window-index cache keys and stale-element checks across a restore
* fix(computer-use): restore only this app's resolved window and refuse a closed target before raising
* docs(computer-use): say macOS reads minimized, hidden and other-Space windows
* fix(browser): keep agent-browser exec on Orca's page and orcad Chromium headless
exec passed raw agent-browser argv through after stripping only --cdp and
--session, so connect, tab/window, close, set viewport, batch items and
global launch options could move the helper to another browser or create
targets behind Orca's tab model. Refuse them in one shared gate used by the
desktop bridge and orcad's external-Chromium provider.
orcad's external-Chromium helper inherited AGENT_BROWSER_* and read the
user's ~/.agent-browser/config.json and ./agent-browser.json, so headed or
extensions settings opened a visible window. Drop inherited settings, point
AGENT_BROWSER_CONFIG at an empty Orca-owned file and pin --headed false.
* fix(browser): refuse inspect, stream and auth in exec and match agent-browser's batch escapes
* fix(browser): retry orcad's agent-browser config write after a failure
* Preserve remote browser certificate warnings through navigation
* Apply current remote observations at the session boundary
* Document array proxy forwarding in the performance contract
* Keep remote browser failure fallbacks within current observations
* Defer remote observation reads to lifecycle callbacks
* Apply remote page observations without preempting pending outcomes
* Preserve command loading ownership across partial page updates
* Reuse browser host identity within each mirror record
* Explain transparent record access in the loading contract fixture
* Keep adoption fixture aligned with remote loading session
* Name the remote navigation fixture return contract
* fix(cli): keep Orca in the background when orca open launches it
* docs(cli): say orca open leaves the foreground alone
* fix(cli): ask a running Orca to open its window in the background
* fix(main): keep background reveals inactive on Linux maximize and drop the flag on admin restart
* fix(cli): narrow the open status before returning it
* fix(main): keep focus-existing-window free of main-window contracts for the web typecheck
* fix(cli): cap the open window request at five seconds
lint-staged 16.4.0 walks the whole worktree several times before any
linter runs. On a cold stat cache that made small commits take over a
minute. Extend the existing lint-staged patch:
- getDeletedFiles: replace the serial `ls-files --deleted` with
`diff-files -D` (parallel preload, fsmonitor bypassed) plus
assume-unchanged entries, filtered by the same lstat rule, so the
deleted-file contract is unchanged.
- getUnstagedFiles: `status --untracked-files=no`; the `??` lines were
already discarded.
- config discovery: list tracked config files only. A nested config
must be tracked or staged to apply; untracked configs at the repo
root are still found.
Run lint-staged directly from the hook; husky already puts
node_modules/.bin on PATH, so `pnpm exec` only added startup time.
* perf(github): share and pace timed pull request reads on the host
Reuse exact completed results for automatic refreshes, back off unchanged
settled pull requests, and route server revalidation through the existing
host budget while preserving foreground reads and older clients.
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* fix(github): preserve shared producer settlement and selected demand
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* test(github): cover paired receipt retries and watch fairness
* test(github): explain shared fake-clock fixture
* test(github): typecheck host and renderer integration together
* test(github): name exported fixture mock types
* Revert "test(github): typecheck host and renderer integration together"
This reverts commit 756a38787c.
* test(github): place host and renderer loops in cross-layer suite
* fix(github): preserve refresh policy and shared callers
Use the same validated freshness for cached queue reads and their follow-up. Keep forced mergeability rereads and detach cancelled consumers independently from the shared lookup.
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* fix(github): refresh settled rows when exposed again
Use the same finite foreground age to admit and consume the queued read. A recent exact result can still satisfy it, and settled rows do not acquire a recurring timer.
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* fix(github): accept only current PR read completions
Keep shared publication and retry planning on the accepted actual read. Preserve legacy own replies, terminal loading and ordinary work after a superseded lookup, and honor the accepted error gate before automatic recovery.
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* test(github): follow moved request contracts
* test(github): explain the partial event fixture type assertion
* fix(github): preserve queued deadlines and detachable PR waits
* fix(github): compare refresh wakes on the elapsed clock
* test(github): format queued refresh deadline cases
* fix(github): charge actual producers and fence retry gates
* fix(github): preserve fresh intent and producer error policy
* fix(github): retain owned cold-read retry gates
---------
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
A tab a page opened with no gesture counted as the agent's only within the
agent's 20 s idle window, so a late open (an OAuth tab finishing after the
agent went quiet) switched the user's visible tab. Track, per guest, whether an
agent drove the page since the desktop user last gave it input, carry it into
popups the page opens, and keep such opens in the background.
A popup whose owner tab was gone handed any http(s) open to the system browser,
including agent, phone and scripted ones, which took the foreground. Only the
desktop user's own click opens it there now; the rest are reported blocked.
* feat(computer-use): scroll and select rows in background Linux windows, and aim GTK 4 pointer input
Scroll moves the nearest scroll bar's value by whole viewports, handing
the scroll outward at a view's end, or pages a web document with
Component.ScrollTo, before falling back to the pointer. A click on a
list row with no action selects it through the list's Selection when
the index is proven to name that row. GTK 4 rectangles come from
WINDOW extents plus the X11 client origin, since its SCREEN extents
read [0, 0, w, h].
* fix(computer-use): keep Linux background scroll and select honest under partial and slow results
Wait for ScrollTo to move before planning the next page, report pages
already scrolled instead of falling back, let refused scroll bars fall
through, clear a multi-select list before a plain-click select, accept
only an InvalidArgs refusal as index proof, bound child walks, and
read GTK 4 origins lazily.
* fix(computer-use): scroll the innermost Linux view that overflows, and bound the scroll search
ScrollTo now runs on each ancestor that overflows its own box, innermost
first, so an inner scroller moves before the page; coordinate scrolls
start from the deepest node under the point. Cap the scroll bar search
to small containers, report a document at its end as a no-op, keep
looking past an unusable spanning child, and look up GTK 4 window names
once per operation.
* feat(settings): add per-project workspace cleanup rules to project settings
A project's settings get a Workspace cleanup section: Use global, Off, or
Custom rules seeded from the host's global rules, using the same rule rows
as the Workspace Cleanup page.
* fix(settings): compare a cleanup change against a fresh preview, not the cached one
A cached baseline can predate files that now protect a workspace, which
would hide it from the confirm. The host shares concurrent previews, so a
fresh read stays one evaluation.
* test(browser): measure agent input focus on the Windows desktop
* test(browser): show Orca inactive before the Windows holder takes the foreground
* test(browser): name the foreground window in Windows holder replies
* test(browser): lift the foreground lock before the Windows holder takes the foreground
* test(browser): run the Windows agent focus check from a dispatch-only workflow
* test(browser): skip the Windows focus checks before launch off hosted runners
* test(browser): harden the Windows focus holder and its dispatch workflow
* fix(computer): ask the user before retrying with --restore-window
Error hints, recovery steps, the skill guide and the CLI follow-up command
told agents to retry with --restore-window on their own, which raises the
target app and takes the user's screen. They now say to ask the user first,
stop recommending a restore just to observe, and fix stale per-platform
background claims.
* docs(computer): name the Linux keyboard gate and the Wayland limit
Pages outside an Orca pane (popups, serve pages) got native webContents.focus() on every agent press, which focuses and can raise their window. Emulate page focus instead, and focus natively only inside a window that already has focus.
* Add setup scripts for project group folder workspaces
* Align setup regression fixtures and catalog read validation
* Target the Advanced disclosure in group setup regression
* test: initialize project groups in worktree slice fixture
* Preserve group setup startup input across paired hosts
* Check paired group startup regression in cross-layer project
* Restore group setup settings when saving fails
* Wait for group agents to accept startup input
* Type the group startup readiness fixture foreground state
* Give group startup fixture mocks portable public types
* Reuse worktree startup for group folder setup
* Keep example shell regression while removing obsolete observer tests
* Complete launch runtime fixture PTY options
* Exercise terminal startup failure and worktree draft paste in parity tests
* Preserve repository type defaults for setup commands
Brennan BensonandTaras Kornichuk
registered2ndandNeil
OrcaWinandOrcaWin
gaokui2010
BingZandNeil
Kelvin Amoaba
Jinwoo Hong
Yasyf Mohamedali