* fix(mobile): keep Android split screen from restarting or squeezing the app
One UI gives a split-screen pane its own density. Android recreated
MainActivity on that change, and React Native seeds its display metrics
from the application context, which keeps the full-screen density, so
the pane was laid out at the wrong scale. A config plugin declares
density in configChanges and re-seeds the metrics from the activity.
* fix(mobile): catch React Native's rotation re-seed before the frame draws
A one-shot listener added on configuration change missed a cold start or
a 180-degree flip at a non-zero rotation, where React Native's root view
re-seeds from the application context without a configuration change.
* fix(mobile): publish re-seeded display metrics in one step
Text measurement reads the holder off the UI thread, so build the
pane-density screen metrics before publishing them.
* fix(mobile): show workspace images in session Markdown tabs
* fix(mobile): show workspace images restored by undo or paste
* fix(mobile): keep session image loader current and inside the workspace
* fix(mobile): classify entity-encoded image URLs by their decoded scheme
* refactor(mobile): classify editor images with the shared Markdown href router
* test(mobile): expect workspace image markup in the code-block selection check
* fix(mobile): save iPhone library photos as JPEG named by their bytes
The iOS photo picker handed back the asset's HEIC bytes as-is, and the host
named every pasted image .png, so agents read a HEIC file as a PNG and failed.
Ask the picker for the compatible representation (JPEG), and name pasted
images by their signature so the extension always matches the bytes.
* fix(mobile): keep unknown pasted bytes on the .png name
A .heic name falls outside the shared image MIME table, so chat previews
and agent prompts would drop the image. The picker now sends JPEG, so
leave HEIC detection out.
* refactor(mobile): keep the pasted image extension list module-private
The iOS web shell cancelled every subframe navigation, including the sealed about:srcdoc frames that HTML previews and chat visual cards render in, so they stayed blank. Allow exactly about:srcdoc subframes, let only a main-frame load spend the shell-load allowance, and never fail the shell document over a subframe response. Chromium never asks the Android shell about about: subframes, so Android already renders them.
* fix(native-chat): stop async question Submit from queuing duplicate answers
When the agent is mid-turn, answering a Codex async question queues the
answer as an ordinary message. The card stayed up until a transcript row
retired it, which a queued message never posts until the turn ends, so
every Submit queued another copy. Hide the occurrence once its answer is
queued, on desktop and phone.
* fix(native-chat): steer async question answers into the running turn
An answer to the agent's own mid-turn question belongs to that turn.
Queuing it behind the follow-up setting held it until the turn ended,
after the agent had already proceeded without it. Send it with the
steer intent whenever the follow-up resolver would queue a plain send,
on desktop and phone. Stop and a waiting command still queue, and the
queued fallback still hides the card.
* fix(native-chat): hold async questions while a queued answer waits
Replace the mount-scoped hide on a queued answer with one derived from
the queue: a question stays hidden while a queued card quotes it, across
remounts, and returns if that card is deleted or edited away. Lock
Submit until the send settles, and keep a returned answer out of the
composer draft.
* fix(native-chat): keep structured session under the line cap
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
UIKit's split-keyboard update crashes breaking our constraints on
keyboardLayoutGuide, so the probe reads keyboardWillChangeFrame instead.
The hinge interaction goes too: UIKit already relays out a view that
read reserved regions in layoutSubviews when they change.
Fixes#36742
* feat(mobile): collapse sub-agent lists by default and group idle sub-agents
Mobile follow-up to #36936: per-parent child disclosure with +N count and a
live-state dot when collapsed, and a quiet 'N idle' group for two or more
idle children once expanded.
* fix(mobile): keep sub-agent disclosures reachable and remembered
- Hit slop extends only downward so toggles no longer steal taps from the row above
- Disclosure state survives workspace list remounts via a bounded module cache
- Collapsed signal considers all descendants, not just direct children
The free-slot probe fills every desktop slot, so end()'s final chunk forward
is refused by construction. Since #36922 reserved native slots, that final
read now succeeds and its forward is refused, failing Mobile verify under Bun.
Desktop RPCs and on-device native verbs shared the bridge's 64 in-flight request slots on both page and shell. Unanswered desktop calls could hold all of them, refusing native.media.pick and native.media.release; refused releases stranded staged handles until expiry, so later picks hit native_media_handle_cap. Desktop calls now stop at 60 so native verbs always have the last 4.
* fix(mobile): run the repo's archive hook when deleting a workspace from the phone
Mobile's delete sent worktree.rm without runHooks, so the host skipped the
orca.yaml archive hook that desktop runs. The phone now does what desktop's
ensureHooksConfirmed('archive') does: it asks to trust an unapproved shared
archive script, sends runHooks with desktop's reply timeout, and when the hook
fails it shows the host's error with Delete Anyway, which resends with
allowFailedArchiveHook.
repo.hooks now reports archiveTrust (hash and content of the shared orca.yaml
archive script, or null when none needs trust). Older hosts omit it, and the
phone then keeps skipping the hook rather than run a script it cannot check.
* fix(mobile): share one Modal for delete drawers and pin each prompt to its host
Delete's confirm, archive trust prompt and hook failure now share one
BottomDrawerModalHost so iOS never presents over a closing drawer. A pending
prompt or failure keeps the clients it started on, the row hides as soon as
Delete is tapped, repo-wide trust counts server-owned repos, and folder repos
report no archive trust since their removal runs no hook.
* fix(mobile): recheck archive trust before Delete Anyway and count server repos once
Delete Anyway goes back through the trust decision, so a script changed since
approval is asked about again. Repo-wide trust counts each server repo once
rather than once per workspace, and restoring a row never duplicates it.
* fix(mobile): honour repo-wide archive trust only with the desktop catalog loaded
Until the catalog loads, a same-id repo on another host cannot be ruled out,
so only an approval of that exact script counts. The trust drawer's actions
take the pending request synchronously so a double tap cannot both run and
skip the hook.
* fix(mobile): drop archive prompts and failures from another desktop
A trust prompt or hook failure belongs to the desktop it started on. After a
host switch it is no longer shown or acted on, so a retry can never pair one
desktop's trust with another desktop's repo catalog.
* refactor(mobile): one archive trust check and an optional drawer busy flag
* fix(mobile): file delete-failure strings under the host screen overlays
Also pins the host's archive hash as a literal digest so the node test no
longer imports renderer code.
* fix(mobile): carry delete haptics and automations onto the renamed hook trust drawer
The phone's liveness probe read a full status.get (~12 KB, ~2.3 KB gzipped)
and threw it away on every idle and foreground probe. It now sends
`status.get {probe: true}` and the desktop answers `{runtimeId}` alone,
skipping the machine-name wait. status.get keeps accepting any params, as
when it declared none; older desktops ignore the flag and answer the full
status, which still proves liveness.
terminal.list takes `includePreview: false`. The phone's session inventory
and automation pane lookup never read the 300-char output tail, so they ask
the desktop to drop it; older desktops strip the unknown key and keep
sending it.
Compressed terminal binary frames now use gzip level 1. On recorded-style
terminal output it keeps 93-99% of the level-6 saving (57-95% of bytes
saved) at 1.2-4x less main-thread CPU; text replies keep the default level,
where level 1 costs ~45% more bytes on the wire.
* fix(mobile): guard catalog snapshots against late replies and trim the snapshot paths
A slow unchanged or delta reply that predates a client switch (A→B→A)
could re-point the paired-server snapshot token at rows it no longer
matched; the next row delta then rebuilt onto the wrong base. Pending
reads now carry the token they asked with, and admit drops any non-full
reply whose token is no longer held.
Single-flight reads with one trailing join move into
WorktreeCatalogSnapshotClient.read(), so paired-server rows get the same
protection home had; home's own copy goes away. Server rows use one read
path for old and new desktops and drop snapshots for servers that leave.
The host screen asks for a first page only while nothing is shown and
the full read is not failing, and drops it on an invalid reply. The
shell no longer seeds the page with a status proven on a connection
that has since dropped.
Cleanup: drop exactActivity (unreleased; home only ranks by coarsened
activity and its exact rows defeated delta reuse), fold the legacy opt-in
check into the snapshot resolvers, remove unused worktree.ps result
types and a duplicate page subscription, and re-record the affected
RPC goldens.
* fix(mobile): keep server-row snapshots per client and skip first pages for loaded empty hosts
A shared per-host snapshot let a slow read on a replaced client reclaim
ownership from the new one, discarding the new client's reply. Snapshots
are now kept per client, and a delta rebuild retry runs only for the
current owner. An empty host that already loaded no longer asks for a
first page on every poll.
* fix(mobile): stop page bounce and input zoom; URL keyboard for new browser tabs
* feat(mobile): smoother motion, press feedback everywhere, reduced motion for sheets
Working rings and the tool-run pulse run as compositor animations on the
page. Sheets, the right panel, the page stack and the pager cross-fade under
Reduce Motion. The right panel shares the sheet's drag physics. The composer
cross-fades Send/Stop/Save and the mic, jump-to-latest and prompt cards ease
in, banners ease their height, and every Pressable shows a pressed state.
Pairing gets a success beat, and the clean state fades in after a commit.
The phone gave a direct socket 5 s from open to authenticated, while
the desktop keeps an unauthenticated socket for 10 s. On a slow link
(a handshake is two round trips, so ~2.5 s RTT is enough) the phone
dropped the socket the desktop would have accepted, then redialed into
the same deadline forever: the 'Can't connect' loop a test documented.
Both ends now share PAIRED_E2EE_HANDSHAKE_DEADLINE_MS (10 s). A link
that needs 6 s to authenticate now connects on the first dial; one that
never authenticates still escalates the same way, 5 s later. Relay
dials are unaffected: they keep their own stage budgets and race the
direct dial after 2.5 s.
The very first time a host is opened there is no cached page, so a
single failed dial (a 12 s connect deadline or a 5 s handshake on a
slow link) put up the offline screen, 'Connect to this host to download
the latest app', while the phone was still redialing and often seconds
from connecting.
The shell's gates now carry whether the dial is still within the
attempts the connection label calls connecting (the same three it
waits before warning). A first open keeps checking through them; when
they run out the gate verdict changes, the flow restarts and the
offline screen shows as before. A cached page still opens at once, as
today.
The download-failure decision moves to mobile-web-shell-download-failure.ts
unchanged, to keep the session reducer under the size cap.
Opening a host with nothing cached waited for the whole worktree
catalog before drawing a row. On a real desktop with 1,084 workspaces
that is ~153 KB compressed, about 1.2 s at 1 Mbps.
The first read now also asks for the host's first 60 rows by its own
ranking (pinned, unread, active, most recent output), ~17 KB on the same
desktop, and draws them while the list is still empty. The full catalog
replaces them as soon as it lands, as it replaces cached rows today. The
first page is never marked loaded or cached, and a failed full read
clears it so the screen still reports the failure instead of a partial
host. Hosts of every version accept the limit parameter.
Opening a host ran status.get twice in a row: once for the shell's
compatibility gate, then again from the page's own gate, which held the
host screen (and with it the first worktree.ps) for one more round trip.
Over relay that round trip is the slowest part of opening a host.
The shell now hands the page the status its gate proved, in init. The
page's gate starts from it, computing its own compatibility verdict from
those fields, and still reads the host itself; its own answer replaces
the shell's when it lands. Reconnects keep today's behavior: the page
re-reads before trusting the host again.
The shell's gate handling moves to use-mobile-web-shell-gates.ts so the
session hook stays under the size cap.
Mixed versions: init's hostStatus is optional and salvaged, so older
shells omit it and the page asks as before; older pages ignore it.
The host screen re-read repo.list every minute and on every repo or SSH
event. Each reply carried every repo's uploaded icon inline as a data URL
(up to 400 KB each, barely compressible), although repos rarely change.
repo.list now takes an optional afterSnapshotId and answers unchanged
when the caller already holds the exact reply. The id is content-derived
(the hashing moves to content-snapshot-id.ts, shared with worktree.ps).
The phone keeps the last catalog per client, reuses it on unchanged
replies and still refreshes SSH host labels and health.
repo.list declared no params, so older hosts ignore the field and answer
the full catalog as before; callers that send nothing get the byte-exact
legacy reply. Icons still render from the same data, at the same size.
The eight native Settings screens use the shared NavBar and grouped rows,
Source Control's tabs move into the bar's segmented control, and Accounts,
Troubleshoot, the custom key editor, the Tasks pickers and drawers and the
workspace details sheet drop their hand-rolled rows for ListRow.
The host screen re-read every paired server's full worktree catalog on
each 3 s poll, through the desktop's mobileRelay.hosts.worktrees, even
when nothing changed: the same cost worktree.ps had before row deltas.
mobileRelay.hosts.worktrees now answers snapshot and row-delta reads with
the same resolver as worktree.ps (made generic over row type). The phone
keeps one WorktreeCatalogSnapshotClient per server, which now takes the
catalog request it snapshots, so server rows reuse the same token,
rebuild and copy-on-admit logic as desktop rows.
The params schema is strict, so the phone sends the snapshot fields only
to a desktop advertising mobile.relay-host-worktree-deltas.v1; older
desktops keep getting plain reads. Callers that send neither field, and
servers whose rows are not fetched yet, get the plain reply.
* feat(mobile): polish press feedback, sheet motion, states, spacing and layout jumps
Every control presses and disables the same way, with one haptic rule.
Sheets follow the finger and settle without bounce, pickers stack instead of
swapping, and screens slide over a dimmed, shifted parent. Loading, empty,
error and offline states share one view with plain-language errors, rows and
labels line up on one grid, and screens stop jumping as data arrives.
* fix(mobile): settle voice, tasks and accounts states; keep host snapshot current
Voice settings with no paired desktop shows the connect state instead of
loading forever. A failed task settings read after a successful probe retries
hydration. The shared host-list snapshot drops on invalidation and ignores
stale passes. Accounts shows a re-pair state on auth failure, and busy
action-sheet rows keep their disabled state for screen readers.
Two paths showed Disconnected while the phone was still connecting:
- A relay dial started from idle (after a background suspend, or when no
session was live) left the client disconnected until the relay hello,
because the replacement session starts in connecting without announcing
it. The migration now forwards the replacement's initial dial state.
- The host screen page never learned which path the shell was dialing, so
a direct dial that failed over to relay read as Reconnecting, then Can't
connect and the unreachable empty list, while relay was still working.
Bridge connection snapshots now carry the active and pending paths and
the relay verdict, and the shell sends a connectionPath frame when only
those change, so the page shows Connecting via Relay like the native
screens. Path frames fire path listeners, never state listeners, so
nothing on the page mistakes them for a reconnect.
Mixed versions: the snapshot fields are optional, so older shells omit
them and the page keeps its previous defaults. Pages announce
connectionPathFrames in ready; older pages never receive the new frame.
* Show agent message cards in phone chats
* test: type the phone markdown mock with its component props
* Fix phone agent message cards and sender navigation
* Open native chats on the phone in place
* Preserve desktop host and agent when opening chats
* Separate phone cards from shared chat opening foundation
* Show phone agent cards through shared sender opening
* Preserve chat opening and captured lifecycle contracts
* test: cover sender focus and unconfirmed chat close outcomes
* test: exercise structured worker CLI lifecycle through runtime socket
* test: use checked parked mail observation in CLI integration
* test: validate provider notification correlation identities
* test: construct checked lifecycle and navigation fixtures
* Keep cross-program session proofs in their test program
* Localize mobile chat opening feedback and complete runtime contracts
* Use the mocked text component identity in History feedback proof
* Temporarily expose grown E2E type diagnostics
* Keep E2E timer and fetch fixtures in their runtime domains
* Reconcile decreased E2E wait pattern counts
* Preserve live test clocks and expose remaining type diagnostics
* Keep mobile session data types independent of native views
* Complete structured session wire surface fixtures
* Expose existing grown E2E type diagnostics temporarily
* Preserve checked wire fixture tab mock signatures
* Preserve worker redrive when automatic close is refused
* Honor automatic close admission before settlement and retirement
* test: retain History reveal and close publication race
* Match automation reveal callback to activation result
* Align CLI integration fixture with current settings and output
* Prevent retained chats from reopening removed workspaces
* Import notification types directly from their owner
* Unify local removal fences for accepted worktree aliases
* Keep cross-surface authority tests in the combined test project
* Initialize coordinator mail test workspace ownership
* Register the retained automation chat test workspace
* Use checked inputs in worker start failure test
* Preserve the automation test repository host literal
* Open phone card senders through shared caller navigation
* Carry the phone sender deadline through shared admission
* Finish empty phone card measurement and consumer coverage
* test(mobile): target native hosts in image-only card regression
* fix(mobile): fade folded agent message edges
The connection snapshot schema, the shell's snapshot builder and the
page's not-ready guard move to bridge-connection-snapshot.ts and
BridgeConnectionCache.require(). No behavior change: the envelope
re-exports the schema, so every frame parses and serializes as before.
Prepares the next change, which adds connection path fields to the
snapshot, without pushing the bridge modules over the size cap.
Every JSON reply and stream event reached the phone as base64 of sealed
UTF-8, a third larger than the JSON itself, and terminal snapshots (up to
512 KB per attach) went out uncompressed. Opening a host on a real desktop
moved ~2 MB, about 17 s on a 1 Mbps link.
Phones now advertise transport.compressed-frames.v1. For those sockets the
desktop gzips text frames of 1 KiB or more, and terminal frames of that
size, and seals them as binary frames tagged with a new kind byte, the
payload kind and the inflated length. The phone inflates them (native
gunzip when available) back into the same text or binary frame before
routing. Screencast frames are never compressed, and the relaying desktop
never offers the capability to paired servers, whose frames it rewrites.
On the same desktop, opening a host now moves ~180 KB.
Desktop CPU is unchanged: gzip then seal of the 1.5 MB catalog takes the
same time as seal then base64 of the plain one. Setting
ORCA_DISABLE_MOBILE_COMPRESSED_FRAMES=1 on a desktop turns compression off
without a release. Older phones never advertise the capability and older
desktops ignore it, so both keep plain frames.
* Open native chats on the phone in place
* Preserve desktop host and agent when opening chats
* Preserve chat opening and captured lifecycle contracts
* test: cover sender focus and unconfirmed chat close outcomes
* test: exercise structured worker CLI lifecycle through runtime socket
* test: use checked parked mail observation in CLI integration
* test: validate provider notification correlation identities
* test: construct checked lifecycle and navigation fixtures
* Keep cross-program session proofs in their test program
* Localize mobile chat opening feedback and complete runtime contracts
* Use the mocked text component identity in History feedback proof
* Temporarily expose grown E2E type diagnostics
* Keep E2E timer and fetch fixtures in their runtime domains
* Reconcile decreased E2E wait pattern counts
* Preserve live test clocks and expose remaining type diagnostics
* Keep mobile session data types independent of native views
* Complete structured session wire surface fixtures
* Expose existing grown E2E type diagnostics temporarily
* Preserve checked wire fixture tab mock signatures
* Preserve worker redrive when automatic close is refused
* Honor automatic close admission before settlement and retirement
* test: retain History reveal and close publication race
* Match automation reveal callback to activation result
* Align CLI integration fixture with current settings and output
* Prevent retained chats from reopening removed workspaces
* Import notification types directly from their owner
* Unify local removal fences for accepted worktree aliases
* Keep cross-surface authority tests in the combined test project
* Initialize coordinator mail test workspace ownership
* Register the retained automation chat test workspace
* Use checked inputs in worker start failure test
* Preserve the automation test repository host literal
Home re-read the full worktree catalog for every connected host on each
focus, reconnect and change event just to show counts and a resume row.
It now keeps a snapshot per host and client, so repeat reads are
unchanged replies or row deltas. One read runs per host at a time, and a
refresh asked mid-read joins a single trailing read, so concurrent reads
cannot discard each other's delta and no refresh settles with older rows.
Home ranks and shows exact activity, so it asks for it: worktree.ps takes
exactActivity to skip poll coarsening, and acceptRowDelta alone now opts a
first read into snapshots. Home's first read sends no snapshot id, so a
host without row deltas answers its exact legacy catalog and every later
read stays legacy; home's data is unchanged on every host version.
A working agent's heartbeat changes the catalog snapshot id on nearly every
3 s poll, so the host screen re-downloaded the whole catalog each time
(1.5 MB for ~1000 workspaces, ~26 MB/min on a real desktop). Callers that
hold the base snapshot now opt in with acceptRowDelta and receive only the
changed rows; unchanged rows arrive as indexes into the base catalog.
The host keeps only row hashes for the 16 most recent snapshots, so an
evicted base falls back to the full catalog. Older hosts ignore the
parameter and older phones never send it.
* feat(mobile): expand Mermaid diagrams into a pinch-zoom viewer; image lightbox zooms the image, not the page
* fix(mobile): expand Mermaid labels with nonbreaking spaces instead of a blank image
* feat(mobile): draw SVG files with a Preview/Source toggle; pinch, drag and double-tap to zoom images
* test(mobile): render an SVG preview under the shipped shell policy and pinch it
* fix(mobile): double-tap after a drag at fit zooms in instead of re-fitting
* fix(mobile): pinch follows the fingers, Back resets zoom, and writable SVG artifacts get a preview
* fix(mobile): keep a writable SVG's save failure visible in Preview