The engine lookup and the GPU failure marker are read once per process
instead of on every dictation start and mobile catalog request, the marker
is written asynchronously, and the child reads the hotwords file with
fs/promises. The baseline moves the existing engine presence checks from
stt-worker-paths to the shared package table and adds the once-per-process
marker reads.
The speech host child can now load a second engine. A model whose files
are GGUF runs on transcribe-cpp (Metal on Apple silicon, Vulkan on
Windows and Linux, CPU otherwise); everything else stays on sherpa-onnx,
which remains the only engine on Intel Macs.
A GPU that fails to load, errors, stalls for 25 s or crashes the process
moves the dictation to a CPU-only process (TRANSCRIBE_BACKENDS=cpu) that
replays the audio the GPU had not decoded yet, so no words are lost or
repeated. A GPU that crashed or would not load is remembered per
transcribe-cpp version and OS release, and voice.speechGpu = 'off' keeps
it on the CPU.
Packaging copies transcribe-cpp, koffi and the target's native package
into resources/node_modules outside asar, fails when a target's package
is missing, and ships the MSVC runtime beside transcribe.dll on Windows.
No catalog model is GGUF yet, so nothing runs on transcribe-cpp until
the catalog change that follows.
* fix(runtime): a phone viewing a terminal on a headless runtime no longer resizes it for everyone
On orca serve / orcad, a phone that only opens a terminal now views it at the
current owner's size. It takes the size when it types or when the user taps
the phone-mode key; the desktop's take-back reclaims it. Desktop hosts keep
fitting phones on open.
* fix(runtime): keep phone claims across leave grace and held phone fit; pin the headless default for the process
A Codex chat launched codex app-server on the account home saved when the
chat was created, and nothing compared it with the selection, so after a
switch every message kept going to the old account. A chat that had hit
its usage limit kept failing with "Codex hit its usage limit" even though
the new account had quota.
Codex chats now follow the selection the way Claude chats do: the launch
prepares the selected account's home, the started process reports that
home, and the host replaces an idle chat whose account no longer matches
before its next message. Codex resumes a thread only from a rollout in its
own CODEX_HOME, so the launch links the thread's rollout into the new home
first. The chat's usage card reads the account its next turn runs on.
* fix(native-chat): clear /compact from the composer once it is sent
Conversation commands kept their draft until the host replied, while
messages release it as soon as they are on their way. Thread the
composer's onPending through the command dispatch so a sent /compact
or /clear leaves the input immediately; a refusal still hands it back.
Fixes#39122
* fix(native-chat): keep a command typed while a refused one was answering
A refused conversation command is handed back through the independent
draft merge, which replaces a lone command in the composer. With the
draft now released on send, a /clear typed while /compact was answering
was overwritten by the refused /compact. Return a refused command only
to an empty composer; anything typed since stays as typed, and the
refusal notice still says the command was not done.
* fix(native-chat): count images when a refused command checks for an empty composer
---------
Co-authored-by: mahedi-emon <113725370+mahedi-emon@users.noreply.github.com>
* feat(dictation): remove filler words from dictated text, per language
Spoken fillers (um, uh, äh, えーと, 음, 嗯 and their equivalents) are
dropped from dictated text before it is inserted, on desktop, the web
client and phones that dictate through the computer. The pass runs once
in the speech host's event sink, so every route shares it, and lives in
a new shared text pipeline that later passes join.
Only standalone sounds go, never words with meaning, and the comma after
a filler goes with it. The language comes from a single-language model
or the dictation language setting; an unknown language uses only
fillers that are no word in another language. Voice settings gain a
Remove Filler Words switch, on by default.
* fix(dictation): take the filler language from what the engine is told, and drop the comma before a final filler
* fix(dictation): keep 嗯 as a reply, clean all-caps transcripts, and read the sentence start in constant time
* refactor(dictation): one list of sentence-end marks in filler removal, and drop a duplicate off-switch test
* fix(dictation): write non-ASCII filler punctuation as escapes so the orcad bundle stays ASCII
ORCAD_CHILD_ENTRY_POINTS builds session-search-worker-entry.js, but
ORCAD_ARTIFACTS never listed it, so templates, the materializer, and the
remote installer dropped it and SSH-hosted session search failed with
'Session search worker entry not found'. Add the artifact and a test that
every built child entry is listed.
The legacy tool-row purge rescanned the whole messages table on every
writer open and foreign data_version change. Persist the scanned rowid in
meta, and install a trigger that records a floor whenever any writer
(including a downgraded build) inserts a tool row, so reused or later
rowids are still purged without full rescans.
* fix(sidebar): let the divider resize instead of scrolling the workspace list
The 4px scrollbar strip past the seam covered the visible divider pixel, so
grabbing the divider over a scrolling list dragged the scrollbar. Drop the
strip and start the resize handle on the divider; the native scrollbar inside
the sidebar is unchanged.
* fix(sidebar): trim the resize handle's outer reach by 1px
A suspended recipe VM kept getting reconnect attempts: shared-control
socket retries, health probes and SSH logins. Providers that resume a VM
on a new SSH login or connection woke it again, so suspend saved nothing
(#36363).
While a VM's record is suspended (or its resume failed) and no wake is
under way, Orca now makes no background dial to it:
- SSH: connect and the reconnect ladder refuse that VM's own SSH host,
which covers tunnels, managed-server checks, census and browser bridge.
- Shared control: no retry after close, and no socket open, including a
retry timer armed before the sleep.
- Status probe: answers asleep without dialing.
A wake (opening the workspace, Resume) is pending before it dials, so the
wake's own connect passes. Afterwards the server's control socket and
status are reopened.
A VM recorded as up whose SSH host Orca gave up reaching (the reconnect
ladder used every attempt) may have been stopped or reclaimed by its
provider. Before, opening the workspace did nothing for it (#34089). Now
it runs the recipe's resume and reconnects. If that fails, the record is
left as it was, so normal reconnects go on.
Fixes#36363
Refs #34089
The renderer's crew follow reconnect scheduler imported the shared-control
socket state module, which reaches the Node-only sealed-frame reader and
its node:zlib codec. Vite dev throws on that import in the web client,
the desktop renderer's crew panel chunk and the component fixtures.
Move the reconnect timer function next to the scheduler, its only user.
* Remove low-value ACP, hook and renderer test suites
* Remove low-value shared automation and agent test cases
* Remove low-value packaging and mobile probe test cases
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
With a per-workspace VM workspace active, New workspace showed "Run on: Local
Mac" but created the workspace inside that VM. Seed the composer from the same
project on a visible host, hold Create when the resolved target is a
per-workspace VM (which Run on never lists), and show "Choose target" instead
of another host's name when the selection is not listed.
Fixes#38669
Co-authored-by: Neil <neil@stably.ai>
Drops the logo and duplicated shortcut list for a single list with key caps, a default-agent row with an agent picker, Go to file, Mobile emulator, Reopen closed tab and a quiet Close pane.
* feat(mobile): show folder workspace Source Control on the phone
A folder workspace now has Source Control on the phone: one row per attached
worktree (repo, name, branch, ahead/behind, change count), read from the host
list's own lineage projection and each worktree's git.status. Tapping a row
opens that worktree's existing Source Control in place; Back returns to the list.
* fix(mobile): folder Source Control routing, back, hidden reads and failure states
- A desktop worktree under a server's folder is listed and opens with desktop routing and gates.
- A child opened in the dock gets Back to the folder list beside Close.
- Rows and the catalog stop reading while the screen is covered, as the single-worktree screen does.
- Offline, failed and missing-folder states; rows whose status can't be read say Unavailable.
- Server rows poll like the host list; branch shown without refs/heads/.
- Perf contract: git.status reads per signal at N and 10N rows.
* perf(mobile): folder Source Control reads summaries and rebuilds only on lineage changes
- Rows read a status summary without per-file line stats; relayed servers' rows open no desktop event stream.
- Membership resolves over a projection of lineage and label fields, so agent and terminal churn rebuilds nothing.
- A row whose re-read failed or whose host reported readError shows Unavailable.
- A connection still coming up shows loading, not offline.
- Perf contracts at N/10N rows and H/10H catalog rows.
* perf(mobile): folder membership keys on lineage only; child status reads update the folder list
Membership resolves again only when a row's lineage fields change, keying
only rows the catalog replaced; labels come from the current rows. A child
Source Control's newer status read is handed to the folder list's summary,
so Back shows what an action did without another read.
* fix(mobile): folder summary takes child reads as they land and ignores server health
Membership depends only on each relayed server's installation, so health
and relay flips no longer resolve the folder again. A child Source
Control hands each status read to the folder summary as it is accepted,
so a summary read answering just after can't hide it, and an older read
never replaces a newer summary failure.
* fix(mobile): folder summary skips restored and cached reads that could undo newer ones
Persisted status is marked restored before anyone hears of it, so a
child screen never hands an earlier launch's status to the folder list.
A cached read caught up on opening writes only an idle summary and never
cancels one in flight. Perf contracts count real lineage serializations,
committed row renders and summary publications.
* fix(mobile): clear the restored mark on any answer but the restore's own
A live answer timed before the stored record, as after a clock
correction, is still this launch's, so its reads and the folder summary
treat it as confirmed.
* fix(mobile): the restore's own write skips clearing its mark only once
A listener replacing restored data as it hears of it writes this
launch's answer, so the mark clears for it too.
* test(mobile): type folder Source Control test fixtures without assertions
* test(mobile): name mocked modules with type-only namespace imports
* feat(mobile): show folder worktrees as cards on the sheet surface
Each attached worktree is its own grouped card on the grey sheet
surface, the look the phone's option sheets use, so rows read apart
from the page.
* fix(mobile): keep folder Source Control's states on the card surface
Loading, empty, error and the list now share the sheet surface, so the
screen no longer changes background between states.
Codex's Speed row used a segmented Standard | Fast switcher while Claude and
Cursor used a single Fast chip. Speed is now always one button: one faster
tier toggles it, several (Fast and Ultrafast, or a configured Flex) open a
list from the button.
A process in Claude's group that ignored SIGTERM held /clear for the full
3 s grace; on /clear the supervisor now reaps the group after 1 s.
Linear: STA-14361
A reset that reclaims a crashed run's lock now puts that record back on release,
without overwriting a lock another start took, so the next start still reports the
crash. Starts without a rotating marker delete leftover .next staged files (one holds
a secret key); an unreadable marker leaves them for recovery by hand.
Co-authored-by: Claude <noreply@anthropic.com>
Cloning always went into <parent>/<repo name>, so a same-named repository
from another owner (or any non-empty folder of that name) could not be
cloned into the same parent. Before cloning, Orca now checks the target
on the clone host. If it is taken, the clone forms show a Folder name
field with a free suggestion (<repo>-<owner>, then <repo>-2...) and clone
into that instead. RPC repo.clone and `orca project setup-clone` accept
the folder name too.
Review rebuilt only when a git.status read changed, so a base move that left this worktree's status as it was kept the old compare (also cached 30 s). The compare entries now take git.status's gitStatusChanged signal. A status read that lands after the signal marked the shown compare rebuilds Review without re-keying worktree diffs, and a compare read that lands marked is read again.
* fix(mobile): paint host labels and relay server rows from the phone's store on relaunch
* refactor(mobile): reuse the SSH target summary schema for stored host labels
* fix(browser): import the session registry from its new profiles folder
#40002 moved browser-session-registry into sessions/profiles, and #39613 landed after it with the old path, so main no longer typechecks or builds.
* fix(test): keep the popup Electron fixture in its own worker
#40067 lets unit files share a worker's module cache. The popup compatibility fixture keeps its Electron app, mock server and temp root at module level, so after the first file the others reuse a closed browser.
* fix(native-chat): show newly installed skills in the Codex $ menu before the first message
Resting Codex chats served the background probe's skills/list from a
5-minute cache, so a skill installed in that window showed "No matching
skills", even in a new chat. Each menu open now compares a cheap stamp
of Codex's skill folders and skill/plugin config with the one taken
before the last scan, and rescans only when it changed.
Probe requests always force a reload, and probe pools are keyed by the
catalog's launch generation instead of the whole env, whose 10 s refresh
started a second Codex on every rescan.
Fixes STA-14359
* fix(native-chat): keep the structured session hook within its line limit
* perf(mobile): answer a page's ready that arrived before its bridge host
A page's ready can reach the shell before the screen has built its bridge
host (the host waits on the client and the host profile read). It used to
be dropped, so init waited for the page's next retry, up to 2 s later.
Keep the latest ready per session and answer it once the host exists.
* perf(mobile): keep one prewarmed page document while Home is shown
While native Home is on screen, boot one page document for the most
recently connected host with a cached build, parked invisibly in the
window. Opening a page from that build adopts it instead of building a
WebView, which removes the WebView start and script parse from the tap.
The spare is dropped when Home is left without a claim (2 s grace), on
background, memory pressure, its own failure and before the cache deletes
its files. It is never kept beside a shown page, and only phones with
more than 3 GiB of RAM (4 GB phones) get one. A parked page is told to
stop asking for init; the adopting view answers the ready it sent.
Builds on Jinwoo's spare pool and replay work in #32203 and #32204.
Co-authored-by: Jinwoo-H <73622457+Jinwoo-H@users.noreply.github.com>
* fix(mobile): unregister the shell's memory callbacks from the context that registered them
---------
Co-authored-by: Jinwoo-H <73622457+Jinwoo-H@users.noreply.github.com>
* feat(viewer): fit PDF previews to pane width
- Apply the saved scale on pdf.js pagesinit; a 'page-width' preset set
before page views exist was dropped, so PDFs opened at 100%.
- Refit 'page-width' when the viewer container width changes.
- Ctrl+wheel / pinch zoom stops on the fit-width scale when crossing it.
* fix(viewer): refit PDF page width once a resize settles
- Refit after the container stops resizing (150 ms) instead of every
frame of a split drag; each refit re-renders the visible pages.
- Drop the fitted-width cache: Fit to width, the wheel stop and
pagesinit fit without updating it, so returning to an old width
skipped the refit. pdf.js already no-ops a same-scale refit.
- Observe the border box so a scrollbar the refit toggles can't loop.
- Apply the latest scale preference on pagesinit, so a zoom made while
the first page loads isn't undone.
- Record the wheel stop's 'page-width' before update() can move the
current page, and read the fit scale only for zoom wheels.
* test(viewer): cover PDF wheel zoom stopping at fit width
* fix(viewer): apply the PDF wheel stop as the page-width preset
---------
Co-authored-by: stmoon <munhoney@gmail.com>
* feat(cli): list and reload installed browser extensions
* docs(browser): document extension command contracts
* docs(browser): describe extension test fixture factory
* refactor(cli): validate extension reload once and trim list output
The RPC params schema owns the ID-or-path and ID format checks. Reload
drops --profile and the constant scope field, since it always reloads
every profile where the extension is enabled. List no longer returns
icon data URLs or permission lists.
* test(browser): keep five extension reload tests at the command owner
* fix(cli): route orca extension through the packaged Linux launcher
* fix(browser): refresh Settings after an extension reload from the CLI
* test(browser): type the reload fixture list without a cast
---------
Co-authored-by: Kelvin Amoaba <97001695+AmoabaKelvin@users.noreply.github.com>
The native shell view now hosts a MobileWebShellDocument that owns the
WebView, its scheme handler or request intercept, the network block, the
bridge and the navigation policy. The view keeps layout, the back-edge
swipe and its React events. Nothing changes for a page; this is the seam
a prewarmed spare page needs.
Based on Jinwoo's page split in #32202.
Co-authored-by: Jinwoo-H <73622457+Jinwoo-H@users.noreply.github.com>
* fix(native-chat): an unanswered model or effort change fails only its message, and each agent process end is logged with its cause
A model or effort write the agent didn't answer within 30 s, or a send withdrawn while its
write was pending, closed the agent's stdin. OpenCode exits 0 on that, so the chat showed
"OpenCode stopped while this response was in progress (code 0)" and its background work
died, with nothing logged. Session setters now carry a call deadline that sends
$/cancel_request and never touches the stream; later controls and prompts wait for a
withdrawn write to settle. Every agent process end logs who ended it and why.
* fix(native-chat): name Orca as the one ending an agent it stopped after its own failure
The exit record now reads endedBy orca for a host force-close and for an ACP agent Orca
stopped after losing its connection or refusing a journal write, with that reason leading.
The withdrawn-write test no longer races its deadline.
* fix(native-chat): keep who ended an agent when its last journal write is refused
Ruled by the independent decider: the failure-close cause is recorded before any publish that
can re-enter cleanup, an ended or requested ending never takes a new one, and the ended event
reads it from before publishing. A test drives a real refused journal write on an agent's own
exit and on a lost connection.
* fix(native-chat): redact the agent process end record's reason before it is logged
The reason can carry the agent's stderr tail, and the console line is not redacted downstream.
* fix(native-chat): let Continue in Chat UI switch when Orca can't prove the terminal stopped
Continue in Chat UI closed the agent's terminal and then refused the chat
unless the host proved a clean exit, so an idle Claude on a slow or remote
host failed with "couldn't confirm Claude stopped". Any settled exit wait now
lets the chat start; the host's create still refuses a conversation another
terminal holds, using the hook rows and foreground-process checks it already
has. The wait outlasts the host's explicit-close stop.
A conversation the agent never saved (nothing sent yet) has no history to
open, so every Retry failed. After the terminal closes, that chat now restarts
in the same tab as a new conversation instead.
Tracked in #39085.
* fix(native-chat): read a Claude transcript nothing was sent in as empty, and let an unknown writer check switch
A Claude session that never got a message can leave a metadata-only
transcript. Its import found no turns and refused as unreadable, so Continue
in Chat UI still failed. A Claude file with no user or assistant record now
reads as empty, the same refusal a missing transcript gets, which the
handoff answers by starting the chat as a new conversation. Records Orca
can't decode still count as a conversation, so unreadable history keeps
failing.
A host that couldn't say whether it checked other terminals no longer stops
the switch before the close; the create checks them again.
* fix(native-chat): read an empty Claude transcript as empty through adoption, and a broken record as a possible turn
Adoption checks a transcript's identity before using it, and a metadata-only
file carries no working directory, so the empty file still refused as
unreadable. A file with no conversation adopts nothing, so it skips that
check. A record too broken to parse may be a turn, so it keeps the history's
own failure instead of reading as empty.
* fix(native-chat): only a typed record with no message reads as Claude metadata
A record shaped like a turn but missing a known type could read as metadata,
so a transcript holding a real prompt could be taken for one nothing was sent
in. Now only a JSON object with a non-empty type other than user or assistant
and no message counts as metadata; anything else may be a turn and keeps the
history's own failure.
* refactor(native-chat): narrow the Claude record type without a cast
* fix(native-chat): read a Claude meta user record as metadata when it shows nothing
A Claude session where only a local command ran, such as /rename, saves a
user record Claude flags as meta and no turn, so Continue in Chat UI still
failed for it. A meta user record the decoder can't show now counts as
metadata; one it shows, such as a tool result, still counts as a turn.
* fix(codex): recognize the 16 app-server notifications added through Codex 0.167
Users run their own Codex, so 0.167 frames already reach Orca. Unlisted notifications fell back to visible raw rows in the chat timeline; classify each one from the 0.167.0 schema instead.
* test(codex): journal no row for app-server notifications that carry provider text
* chore(claude): update the Claude Agent SDK to 0.3.318
Includes the upstream fix for omitted permissionMode (claude-agent-sdk-typescript#519), so the 'use my Claude settings' permission choice honors settings defaultMode.
* test(orcad): expect the SDK's joined --permission-prompt-tool=stdio argument
Claude Agent SDK 0.3.295+ passes named option values in the same argument as their flag.
* refactor(claude): drop the permission flag SDK 0.3.318 no longer reads
The SDK leaves an omitted permissionMode to the CLI, so the 'own' choice needs no opt-in option.
* perf(terminal): group park-exempt sleeping records once per map; index captured panes by leaf
Each mounted worktree's exemption selector rescanned every sleeping record
after a record-map change (worktrees x records), and pane reconciliation
compared every captured pane against every fallback pane (panes^2 per tab).
* perf(terminal): skip the pane index on empty fallbacks; reuse the accumulating leaf traversal
goneSessionsKeepTabMounted scanned every sleeping record per eligible tab,
so a workspace switch with thousands of tabs did tabs x records work on the
renderer thread. Build both owner sets once per source identity instead.
The host stops the old agent (about 0.6 s for Claude, up to 3 s with slow
MCP servers) before it writes the divider, so the old transcript lingered.
Linear: STA-14361