* perf(terminal): submit at once to a shell idle at its prompt
* test: name the captured app transcript by role
* fix(terminal): require host shell-foreground proof and a clean prompt cycle before skipping the submit settle
* fix(terminal): keep the settle for a pane without an identity
* fix(terminal): count a failed idle proof toward the submit settle
* fix(terminal): refuse a fast submit cancelled during the idle proof
* fix(terminal): time the idle proof with a monotonic clock
* test: settle writer idle proofs by hand instead of timed promises
* test: remove low-value suites and cases
* test: shrink flaky-pattern counts after case removals
* test: remove empty suite and stale release workflow fixture
Concise task-list rows carry id, status, title and ids only; specs and results stay in --full (407 KB to 11.5 KB text on a 138-task Run). Concise gate-list rows drop resolutions and timestamps.
* Share chat options read ownership across desktop and phone
* Start a new chat context's options read at once
Reads were coalesced on one queue for the whole view, so after a reconnect, provider restart, phone path change or chat switch the current read waited behind the previous context's request, up to the transport timeout. Each attachment now has its own queue: reads still coalesce to one active and one trailing request within a context, while a new context reads immediately and the old context's reply stays inert. Restores the shared option refresh contract and the reopen test to their original expectations.
* Let desktop options tests wait for the shared read queue
The shared read queue starts a read one microtask after it is scheduled, where the old desktop runner started it synchronously. The selection lifetime and context usage tests now wait for each read to start before asserting on it, keeping every original assertion, including one active and one trailing refresh. The paired server test compares transport targets by value, since the options view memoizes an equal target per environment.
* Keep chat support projections stable between options answers
Desktop and phone rebuilt the conversation, image and command support objects every render, and a host that omits commands produced a fresh empty list each time. Project them once per accepted answer.
* Import options wire types once in the selection reads performance test
* Keep chat support visible across a controller change
Replacing the selection controller, as a desktop re-pair does, cleared the last accepted support, so slash commands, the goal row and the context meter disappeared until the next options answer. On the phone, returning to a chat before the other chat answered showed fallbacks. The support stays tagged with its session and every consumer already filters by session, so it is now kept, as before the shared owner.
* Coalesce phone options rereads across turns and status changes
Each new turn or host status change re-attached the phone's options reader, so a quick run of turns sent one parallel options request per turn. These changes now reread within the current context, as on desktop: one request runs and at most one waits behind it, and only the latest answer publishes. A status change that arrives while a read is in flight now waits for that read, which the transport bounds.
* Restore the phone status wake assertion and type the burst test commands
A refused pick is dropped rather than kept as a finished operation, and the path change replaced the options record, so no revalidation read is in flight when the status wake arrives. The wake therefore reads at once, as the original test asserted. The burst test's superseded answer now carries a real conversation command.
orca orchestration send --to dispatch:<id> now works whatever the worker's state: a working worker reads it now; a finished worker still open gets it as a follow-up Task in place; a closed worker is reopened in its saved conversation, or started fresh with the earlier task and report when no conversation is saved. Unread mail at worker_done becomes a follow-up automatically. An older paired server refuses send to a finished worker with an update message.
* test(wsl): time the lane's first output without the agent-paste Enter settle
terminal.send with enter:true waits 500 ms before Enter so agent TUIs do
not take it as part of a paste; no keystroke pays that, but every lane
first-output number did. Send the probe with its Enter inside the text,
record time to the shell prompt as its own field, and poll at 25 ms.
Stopping orcad now checks the pid's command line names the active slot
and tolerates an orcad that already exited.
* test(wsl): count only orcad-cold runs that stopped a live orcad
A run whose stop found no live orcad for the active slot measured a warm
reconnect; keep it out of orcadCold, report how many were attempted, and
fail if none stopped one. Rename time-to-prompt to first screen output,
since a login banner may come first.
* test(wsl): count an already-exited orcad as stopped; name first output plainly
stopOrcadKeepingDistro now tells an exited orcad (no pid, or no such
process) apart from a live pid whose command line does not name the
slot: the first counts as stopped, only the second is excluded from
orcadCold and reported as excludedNotStopped. Rename the first-screen
field to timeToFirstOutputMs.
* test(wsl): count orcad as down only on evidence, never on a missing record
GONE now needs a known active slot and a numeric pid file whose process no
longer exists; a missing or malformed record answers UNKNOWN, which is
not counted as stopped.
worker-start and task-create take --spec-file <path> (or - for stdin). worker-start --keep-open keeps a worker open after it finishes in one step, and worker-keep-open replaces worker-retain (kept as a hidden alias). Concise text output escapes terminal control characters in worker-written values.
Once Orca has typed a wake line into an agent terminal, it now always sends the Enter, even if the mail was read during the short delay before it. Previously the Enter was skipped and the typed line stayed in the composer, where it counted as a user draft and held later wakes.
#31617 made terminal pane owner resolution read state.unifiedTabsByWorktree, but this test's mocked store never had the field, so shell contracts fail on main with "Cannot read properties of undefined (reading 'wt-1')".
* feat(worktree): open the native chat while its worktree is still being created
When native chat is the default, creating a worktree now shows the chat
pane straight away instead of a spinner panel. Its composer works while
git runs. A message sent before the worktree exists shows as the first
user message with a step card under it (prepare, fetch base branch,
create worktree with checkout %, start the agent), and the chat launch
sends it as its first message once the worktree is ready. Text left
unsent moves into the real chat's composer. Failures mark the step that
failed and keep Retry and Dismiss.
The renderer holds the message and hands it to the existing launch
prompt path, so hosts need no protocol change. The terminal route,
folder workspaces, SSH and VM recipes keep the old panel.
* fix(worktree): keep a held chat message when the host declines the chat
A background create whose host declines native chat opens the agent
terminal from the submit-time plan, which never saw a message sent
while the worktree was being created. That message now lands unsent in
the agent's input. The pending composer draft is also only deleted
once it has been copied into the new chat.
Sheets slide on iOS's curve over their own height, ease when their content
changes, and paint a lighter surface with grouped rows and a drawn switch.
Create worktree, the model picker, the GitHub issue sheet, file preview,
Agent Session History and the task list take the redesign draft's layouts
at the app's smaller type sizes.
Git checks out with one sequential worker by default, so creating a worktree in a large repo spends most of its time writing files on one core. Every create-path command that writes the working tree (worktree add, the prepared-checkout and tip-refresh resets, and the sparse checkout) now passes -c checkout.workers=0, which gives one worker per core. On a 41,924-file clone on a 14-core Mac, git worktree add dropped from 3.8-8.3s to 2.4-3.0s. If checkout.workers is set at any scope, read with git config --get on the same host and repo as the checkout, Orca leaves it alone. Git before 2.32 does not know the key and ignores it.
* fix(chat): collapse settled permission notices
* Keep notice scroll addresses unique per rendered row
* test(native-chat): cover notice scroll bounds and shared classification
* fix(mobile): preserve notice summary long-press actions
* Preserve mobile notice selection and caret rendering
* Use type imports in mobile notice web tests
* Resolve real SVG test host from mobile dependencies
* fix(relay): cap HTTP request bodies before they are buffered, including chunked ones
Every relay POST route checked only the Content-Length header before calling req.json(), so a chunked request with no Content-Length was buffered and parsed whole. /v1/resolve does that before any authentication. Enforce Hono's bodyLimit app-wide at RELAY_PROTOCOL_LIMITS.maxHttpBodyBytes (4 KiB, the limit every route already declares), with the cell reserve and demote routes keeping their own 16 KiB limit; push already does the same.
* test(relay): prove a 6 KiB reserve body reaches its route
* fix(acp): separate recovered history from outside activity
* Fix repeated recovered history divider corrections
* Compose ACP replay targets before checked settlement
* Explain ACP scaling fixture timeout
* Complete ACP fixture types and runtime English catalog
* Type exported ACP fixture warning mock
* Restore ACP history without an extra notice
The runtime stages shipped node:24.22.1-alpine (Alpine 3.24.2) as built, so CVE-2026-85091 (zlib) and CVE-2026-107891 (musl) were present in all three images. Move to node:24.22.2-alpine (Alpine 3.24.3) and apk upgrade the runtime stage, as the api and auth images already do, so later Alpine fixes land at build time.
Agent Dashboard terminal previews had no xterm linkHandler, so OSC 8
hyperlinks fell back to xterm's confirm() + window.open() default, which
Electron blocks. Route OSC 8 links through the same Mod+click activation
as plain-text URLs (opening via shell.openUrl) and show the destination
as a native title on hover, since previews have no pane tooltip.
Co-authored-by: alpenshtokk <wawesandways@gmail.com>
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
The copied .deb install command now passes
-o Binary::file::APT::Sandbox::User=root to apt. The _apt sandbox user
can't read the package in the user's private ~/.cache, so apt already
fell back to root and printed an "unsandboxed as root" notice; presetting
that fallback for the file: method removes the notice without changing
effective privilege, and network methods stay sandboxed.
Fixes#35641
Co-authored-by: Neil <neil@stably.ai>
Five more runs stop only orcad in the guest, as its idle exit does, then
reconnect with the distro still up. startupGate.orcadCold reports them
against the warm bar, marked not gated, with orcad's startup medians.
Orchestration commands print small flat fields in text and --json by default; --full prints the previous output. Failures exit non-zero with error: <code>: <message> on stderr.
Adds Slatewave to the popular dark extended catalog. Slatewave is a calm slate-and-teal palette with an existing theme family across Alacritty, Kitty, Ghostty, WezTerm, iTerm2, VS Code, Neovim, and Obsidian (https://getslatewave.com).
The ANSI mappings mirror the upstream Slatewave palette exactly, so terminal output stays visually aligned with the editor themes in the same family.
* feat(plugins): show public install counts in the plugin catalog
* fix(plugins): tolerate the empty stats placeholder; count bundled official plugins
* test(plugins): pin install-stats test clock
* feat(mobile): compact top bar, bottom search and filter chips
Every list screen gets one compact bar (back, centred title, at most two
buttons, the rest in a ⋯ pull-down), with search moved to a wide bar at
the bottom and filters as chips. The workspaces page shows the host name
at the top of the list and folds it into the bar as you scroll.
* test(mobile): pass the Tasks bar's accessory in the host subtitle test
* fix(mobile): stretch filter chips to fill the row; follow moved header controls in render checks
Writes the stack height and composer clearance vars only on the card stack and
toaster wrapper, skipping unchanged values, so a composer resize no longer
restyles the whole document.
Fixes STA-13053
* fix(plugins): protect every security and destructive string from language packs
Language packs could rewrite trust copy outside a hand-kept list of app
areas (catalog-link dialog, browser-extension Allow/Deny, Node mirror).
Protected keys are now generated from structure: every key in a file that
renders a security surface, a destructive confirmation, or a native message
box, or that carries a // @protected-copy pragma, plus same-directory copy
modules. verify:localization-catalogs fails when the list is stale.
Pack strings are also capped at max(64, 4x English) characters and may not
contain control or bidi characters. Validation stays strict; loading drops
only the offending entries.
* chore(plugins): resync protected copy keys with main
* fix(plugins): cap pack strings in shared strict checks and pin new trust copy
* fix(plugins): keep the packaged CLI off the English catalog
* chore(plugins): resync protected copy keys with main
* chore(plugins): resync protected copy keys with main
When the native exit watch fails to register, the child never reports an
exit, so a later kill() could signal a pid Windows has since reused. Mark
such a child untracked: kill() refuses it and the warning says so. No exit
is faked; the WSL holder already releases through its own pid and
creation-time check.
* fix(plugins): accept plain repo, release and tag page URLs in Git install
The Git URL install dialog refused anything without a #ref suffix, so pasting a
repository or GitHub release link failed with "Add a #tag or #commit". It now
reads owner/repo, repository URLs with or without .git, SSH URLs and GitHub /
GitLab release, tag, tree and commit pages, pins the newest version tag (or the
default branch's commit when there is none) with git ls-remote, shows that
version and commit, and installs only that exact commit.
* fix(plugins): let the Git install dialog close while a version check runs
Top-layer content (showModal dialogs, popovers, fullscreen elements)
paints above the overlay's max z-index and a modal dialog makes it inert,
so hover and clicks reached the page instead of the picker. The outline
now lives in the top layer and is raised again when new top-layer content
opens; the overlay reads hover and blocks trusted presses at window
capture and hands real pointer picks to awaitClick.
Fixes#28657