* Chat UI leaves Experimental: each agent's Chat | CLI choice decides launches
Remove the experimentalNativeChat switch and every gate on it. The Chat
settings page, Default Agent Chat | CLI choice, + menu Chat/CLI rows,
source-control surface choice, Antigravity chat section and orchestration
chat-error switch are always present. What a new launch opens comes only
from the per-agent choice (no saved choice opens a chat) plus the existing
support checks.
- One lookup, nativeChatPreferenceFor, keyed by agent id so host-registered
agents read the same way; unreadable settings still open a terminal.
- A relayed desktop stamp (nativeChatDefault) maps onto the launched agent;
an explicit surface still wins. Old phones' createSupport keeps admitting
by the host's per-agent choice.
- settings.get stops publishing the retired switch keys; nothing in the
repo reads them from a host.
- Startup chat sync and hydration run when some agent can open as a chat
or the machine already holds chats.
- The retired key is stripped on load. The one-time upgrade tip from the
chat upgrade is retired with its audience record; a new record, captured
once from the raw profile, marks which existing profiles had Chat UI off.
- The window-side AI-button tab path, which only ran with the switch off,
is deleted.
* Keep profiles that had Chat UI off on the terminal
Existing users who never turned Chat UI on would otherwise see new agent
launches open as chats, since an agent without a saved choice now opens as
a chat. The one-time load step that records the new audience also sets
every known agent to CLI for those profiles, in the same save and only
when the record was absent, overwriting saved chat choices the old switch
masked. Profiles that had Chat UI on and new profiles are untouched.
Browser profiles keep their own settings, so the web client does the same
once: a stored switch that is not on seeds CLI choices, and dropping the
key marks the step done.
Agents first seen after this step open as chats.
* Agents without a saved Chat | CLI choice open as their CLI
An agent with no saved choice now opens in the terminal for everyone,
including fresh installs and agents first seen later. The one-time load
step carries what launches opened while Chat UI was experimental: with
the switch on, every known agent without a saved choice is saved as
chat; with it off, chat choices the switch masked are saved as CLI.
Existing profiles join the launch tip audience whether the switch was on
or off. Browser profiles get the same carry from their saved key.
The startup chat restore and host agent list now run only when some
agent is saved as chat or the runtime holds chats.
* Preserve owned terminal launches and local chat choices
* Clarify chat and terminal support
* Explain worker CLI integration timeout budget
* Use checked annotations for chat graduation test fixtures
* Store agent surface default separately from per-agent overrides
* Preserve explicit Chat intent across default changes
* Preserve current browser settings during host replies
* Correct browser settings factory and regression types
* Name exported launch fixture mock types and clarify agent defaults
* Preserve settings fallbacks and complete launch test lifecycles
* Check allocated tab identity before launch fixture interpolation
* Register combined integration project for compiler parity probes
* Use checked settings in the launch test fixture
* Preserve retired Chat preference data privately
* Align chat E2E fixtures and capture pane lifecycle evidence
* Use typed raw settings in retention contracts
* Consolidate transcript diagnostic imports
* test: align chat fixtures with live pane and native menu contracts
* test: capture relay tab selection loss diagnostics
* test: wait for owned shells before synthetic chat fixtures
* test: preserve all-host visibility in SSH observer
* test: bind synthetic transcript fixtures before provider seeding
* test: assert selected chat owner after narrow image preview
* ci: allow fifth compiler project to complete comparison
* Type the web-terminal fixture's chat map without an assertion
* Right-click the selection itself in the chat copy spec
* fix(agent-status): settle a Claude pane when a prompt hook blocks the prompt
Claude fires no Stop after a UserPromptSubmit hook blocks the prompt, so the
pane stayed Working: 2 minutes until the OSC 7501 fallback, or until the next
turn when Claude's idle landed before Orca's own prompt hook. A prompt now
claims the program's working run, and a claimed run that goes idle without a
result or user input settles the turn the way a plain Stop would, in either
order. Input during the run (Escape, Ctrl+C) leaves the verdict to interrupt
inference, and relayed panes keep the relay's adjudication.
* fix(agent-status): keep a typed prompt from claiming an earlier rested run
Any input since the run began now withdraws its rest from the blocked-prompt
settle, so the keystrokes of the next prompt stop a stale idle run from being
claimed if that prompt's hook lands before its own working report.
A first WSL deploy downloads the pinned Linux Node archive (about 55 MB)
while the user waits, and fails on a network that blocks it. Once the WSL
hosts flag is open and the setting is on, fetch it in the background into
the cache the deploy reads, through the same materializer (mirror, proxy,
hash check). Once per run, deduped, never boots a distro; a failure is
logged and the deploy's download path is unchanged.
* A chat rename also renames the agent's own conversation
Renaming a native chat saved the name only in Orca, so the agent's own resume list (claude --resume, Codex's resume picker) kept showing its old title for the same conversation.
After a rename is saved, the host now gives the chat's current agent conversation the same name, the way the agent's own rename does: Claude through its host rename request (or its title line when no Claude is running), Codex through thread/name/set (live, or from a short-lived app-server under the chat's account), Pi through set_session_name while it runs. A new agent conversation in the same chat (first message, /clear, rewind) takes the name when it starts; resuming one never does, so a rename typed in the agent's own CLI is kept. Names Orca chose are never pushed. Failures are logged and never affect the rename.
* Preserve provider naming across rewind, close, and configured launches
* Use a fixed timestamp in session title fixture
* Format session rename regressions and simplify setup
* Resolve idle naming without conversation launch admission
* Assert configured PATH in idle naming regression
* Fix Codex session rename lint errors
* Correct provider rename regression fixtures
* Remove redundant rename traversal test after corpus validation
* Restore filesystem rename scaling contract
* Reuse Codex close delegate for disposal
* Join accepted provider naming work before runtime shutdown
* Correct the provider naming shutdown test fixture
* Preserve explicit conversation names across Claude startup
* Settle proven retiring writers before naming follow-up
* Preserve configured Codex rename invocation and align integration fixtures
* Fix construction and fixtures exposed by CI
* fix(chat): reconcile naming owner on late Claude exit
* fix(chat): align naming fixtures with async contracts
* test: separate Claude initialization results from native fixture exits
* test: type Claude naming fixture launch options
* test: await original journal close completion
* fix: persist Claude prompt names with session titles
* fix(agents): accept launch-time --model and --effort for grok
Grok's session option catalog already maps model and reasoning effort to
-m / --reasoning-effort, but it did not opt into worker launch
preferences, so orchestration worker-start rejected --model/--effort for
grok. Opt the catalog in and list Grok in the CLI help and docs.
Fixes#36199
* test(orchestration): use an agent without launch preferences in the federation rejection test
Editor tabs called the built-in getFileTypeIcon table directly, so an
enabled plugin file-icon theme changed the explorer but not the tabs.
Plain file tabs now render the shared theme-aware FileTypeIcon, and
PluginContributionIcon accepts a ReactElement fallback so the icon
component type stays stable across renders. Simulator tabs pick the
phone icon from their language instead of their label, so they keep it
under a theme.
Fixes#36326
Co-authored-by: Neil <neil@stably.ai>
One-shot POSIX text generators (chat naming, commit/branch/PR text)
passed a bare CLI name to the supervisor, whose PATH-only lookup missed
version-manager installs (e.g. nvm) that native chats already find when
the GUI process PATH omits them. Resolve bare commands with the existing
guest-aware installed-CLI resolver, searching relative and empty PATH
entries from the agent's launch cwd so project recipe CLIs keep
precedence, and share that PATH resolution with Codex structured launch.
Fixes#35942
Co-authored-by: Neil <neil@stably.ai>
Git rejects a ref component ending in .lock, so creating a worktree from an
issue whose title ends that way (e.g. "... mise.lock") failed with
"is not a valid branch name". computeValidatedBranchName now rewrites a
trailing .lock to -lock in the branch name only, so the worktree folder
keeps the requested name.
Fixes#36404
Co-authored-by: Neil <neil@stably.ai>
* fix(terminal): only the Windows console-limit failure keeps a dead terminal open
#16045 retained every non-zero local PTY exit and only used the Git Bash
console-capacity marker to pick the overlay's label, so an ordinary shell
exiting after a failed command left a dead pane behind on macOS and Linux too.
Claim the exit for the capacity overlay only when all of it holds: a native
local Windows ConPTY pane, a non-zero code, the same PTY whose fresh spawn was
observed, no user input on that PTY (an unsettled write counts), and the exact
MSYS marker in that PTY's own output. Every other exit falls through to the
unchanged teardown path.
Per-PTY interaction rides on the existing per-PTY process-exit record, marked at
the input-forward chokepoint so a late write acknowledgement can never label a
replacement PTY.
STA-5604
* test(terminal): pin the two unpinned arms of the Git Bash capacity gate
Gate-level ablation of the six-clause capacity condition found the six
conjuncts each individually pinned, but two neighbouring arms free:
- the `session.deps.onPaneProcessDied` presence check. The callback is
optional on PaneConnectionDeps; without the guard a capacity exit calls
`undefined` and throws out of the exit handler, leaving the pane neither
retained nor torn down.
- the per-PTY keying in `noteTerminalInputForPty`. A generation installs its
exit state when its callbacks are captured, before the transport rebinds
its PTY id; a write in that window would be credited to the incoming PTY
and silently suppress its capacity overlay.
Both new tests fail with their own arm reverted and pass with it restored.
No production change.
* Preserve startup input intent across terminal connection
* Keep terminal input attempts with their launch across remounts
* Load the real capacity-test transport during suite setup
* Match the paste-input test result type to its API
node-pty falls back to WinPTY below Windows build 18309. WinPTY never sends
ConPTY's DA1 announcement, so every plain terminal was classified as having no
input reader and the healthy shell was killed and respawned. On v1.4.244 that
discard killed the WinPTY agent twice, and winpty's native kill throws
'Pty seems to have been killed already', failing every plain terminal spawn.
Fixes#36317
Match src/href/srcset attribute names case-insensitively so links written
as SRC, HREF, or SrcSet are validated like lowercase ones, while URL case
is preserved.
Fixes#36211
Co-authored-by: Neil <neil@stably.ai>
The local agent Environment editor returned null when an agent had no
default env, override, or draft, so only goose showed it while empty and
users had no UI path to add a first variable. Always render it, dropping
the showEmpty/hasDefaultEnv plumbing.
Fixes#35997
Self-hosted Azure DevOps Server instances capped below REST 7.x reject
every api-version=7.1 call with VssVersionOutOfRangeException, so PR
detection, checks, review state and PR creation silently failed.
Step down a fixed 7.1 -> 6.1 -> 5.1 ladder (plus the existing -preview
retry) on version rejections, cache the negotiated version per origin,
and bound retries. PR creation shares the same negotiation.
Fixes#36423
Co-authored-by: Neil <neil@stably.ai>
Team-limited (enterprise) Cursor accounts report the member's included
allowance as `individualUsage.overall` instead of an enabled `plan` pool,
so the monthly meter showed nothing. Fall back to `overall` when `plan` is
absent or disabled. Team-wide `teamUsage.onDemand` stays unmapped because
it is not this member's allowance.
Co-authored-by: Neil <neil@stably.ai>
CodeBuddy was the only catalog agent with neither an iconUrl nor a bundled
favicon, so AgentIcon fell through to Google's favicon service, which returns
a generic globe for codebuddy.ai. Map it to the already-shipped
src/shared/agent-icons/codebuddy.png (also used by mobile).
Fixes#36259
Co-authored-by: Neil <neil@stably.ai>
* fix(ai-vault): clear the Session History search when its workspace changes
The query was panel state with no tie to the workspace or project it
searched, so the sidebar panel kept filtering the next workspace with it
and showed no matches. Clear it when the Workspace or Project scope's
subject changes; All and user-picked scope changes keep it.
* refactor(ai-vault): pick the query subject with a scope lookup
* Show recorded session accounts in History details
* Fix session account lint failures and extract presentation
* Isolate session account cards from unrelated settings updates
* Keep session account labels tied to available current roster evidence
* Add English session account labels
* Share recorded session account observation and isolate optional stream evidence
* Verify healthy account and usage consumers during optional registry loss
* Bound session account comparisons and acquisitions
* Use nonmutating reversal in account peer tests
* Type the account subscription replay fixture without assertions
* Repair account directory types and integration fixture imports
* Give account details fixture a portable return type
* fix: release account streams on transport abort
* test: name desktop account stream IPC event type
* chore: temporarily expose e2e compiler diagnostics
* test: type account stream unsubscribe as callable
* fix(terminal): link a URL wrapped inside a table cell
A markdown table rendered by an agent TUI wraps a long URL onto extra rows
of its own cell, leaving the sibling cells blank. Two gates in the
hard-wrapped HTTP builder rejected that layout, so the link stayed cut at
the first row and opened a truncated URL:
- a changed prefix under a second left frame was always read as an
adjacent table cell, and
- only the padding up to the URL's own right frame is cell-local, so a
populated sibling cell to the right killed the row.
Treat a blank gutter between the frames as the same cell continuing, and
only require the URL's own cell padding to be blank. A blank-gutter
continuation is also enough evidence to join a two-row URL, which the
previous three-row minimum ruled out.
* fix(terminal): join a two-row URL in the first table column
A URL that starts in the left table column continues through the
unchanged-prefix branch, so no blank sibling cell marks its continuation
row and the row-count guard dropped the second row.
Count the blank cells right of the URL as continuation evidence for that
layout too, and keep single-column tables on the three-row minimum so a
complete URL that happens to end at the cell border is not glued onto the
next table row.
* fix(terminal): join a table-cell URL only when its row's other cells are blank
A wrapped table cell leaves every other cell on its continuation rows blank,
so that one rule now decides a two-row join in any column. This replaces the
adjacent-cell gutter count and the first-column 0.95 border threshold. Checking
only the adjacent cell glued a next table row like `│ 2 │ │ notes │` onto
the URL above it.
* fix(terminal): end a table-cell URL at a row whose right-hand cells are filled
Below the URL's first row, a filled cell to its right starts a new table row. Text beside an overlay with no table frame is not a sibling cell either.
* fix(terminal): end a table-cell URL where the cells left of it change
Every continuation row of a wrapped cell shows the same cells to its left, so a row that fills them again (a repeated label) starts a new table row.
* fix(terminal): apply table-cell rules only to rows that keep the table frame
An overlay's gutter can show a frame beside its first row only; those rows vary by design and keep the overlay join.
* fix(terminal): count cells right of a URL as a table only when they close with a frame
A split pane or screen text beside an overlay ends without a closing frame and is not a sibling cell.
---------
Co-authored-by: ZheNing Hu <adlternative@gmail.com>
The Plugins status bar button always navigated to Settings > Plugins, so a
second click left the page open. Route the click through a new 'toggle'
settings navigation intent: Settings, which owns the active pane, closes
itself (via the unsaved-changes guard) when that pane is already showing and
otherwise navigates as before. The Mod+Shift+X shortcut stays open-only.
Fixes#36415
Since #36218 a lone update that grows access opens straight into its own review,
and several such updates list with Review buttons instead of joining Update all.
* Add host-owned model and effort defaults to agent settings
* Unify terminal defaults and raw argument precedence
* Keep agent model defaults typed through creation
* Preserve quick-create prompt through preparation
* fix: complete execution-host preference boundary and CI contracts
* Preserve raw options until a replacement is emitted
* Complete launch choice ownership and consumer contracts
* Align launch contracts with compiler boundaries and consolidate proof
* Name exported fixture mocks and verify ordered chip removal
* Preserve typed choices across compound legacy publication edits
* Preserve non-emitted composed launch choices through legacy edits
* Preserve rollback omissions and settle agent launches before success
* Carry terminal draft ownership through launch completion
* Keep inline draft delivery independent of terminal projection
* test: consolidate launch ownership contracts
* test: check web launch fixture array types
* test: align launch fixtures and conformity records
* test: import the shell-ready marker from its declaring module
* test: give caller routing the web host tab's completion contract
* fix(plugins): stop painting a disabled plugin's terminal theme
* fix(plugins): tell the clicking window when a plugin is disabled or removed
* test: merge duplicate import
* fix(native-chat): a dropped screenshot is written once, straight into the chat's store
A screenshot dragged from its macOS thumbnail into a local chat was read into
memory, or spilled to a private copy in the OS temp folder for a batch over
50 MiB, then moved into the chat's attachment store; when the temp folder and
the store were on different disks that move was a second full copy.
The page now reads the dropped File's bytes, the way the common pattern does,
and the store writes them once into a hidden part file in its own folder and
renames it into place. Main still decides which dropped paths are such files
and leaves them unread; a file over the store's 50 MiB limit keeps its checked
private copy, and a store that is unavailable, refused or failing gets the
received bytes as a private copy instead. The in-memory tier, the spill and
the cross-disk import are gone.
Tracked in STA-11413.
* test(native-chat): type the dropped-bytes mock instead of asserting
* fix(native-chat): own dropped files through completion and drain
* fix: keep failed drop cleanup outside payload drain
* fix: include drop integration dependencies in web typecheck
* test: run main file and plugin integration contracts in Node
* fix(i18n): preserve runtime entries after removing untracked settings (#31852)
* fix(native-chat): write dropped screenshots straight into the chat store
A local chat's drag-temp files stay open after their checks and are
streamed once into a pending file in the chat's attachment folder, then
renamed into place. This removes the in-memory tier, the temp-folder
spill and the cross-disk copy that made large cross-disk drops slower.
A file over the attachment limit, or a drop with no store, keeps the
private temp copy made from the same open handle. A file rewritten after
the drop is refused as changed.
* fix(native-chat): keep checked drop source open for fallback
* test(native-chat): make dropped source rewrite deterministic
* Use bounded reads for small checked drop files
* Register approved checked drop copier line exception
* Keep checked drop source copying in a focused module
* Finish buffered drop writes after proven bytes persist
* Capture budgeted drop bytes before waiting for the chat store
* Assert the capture test store getter resolves successfully
---------
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
xterm's WebGL atlas warm-up pre-rasterized default-color ASCII without a DOM
container. When it ran before the first glyph was drawn (a blank restored
pane), the glyph canvas was still detached, so Chromium ignored the page's
-webkit-font-smoothing: antialiased and baked heavier glyphs into the cache
until something cleared the atlas (e.g. an image paste).
Warm up through the terminal element and skip it while that element is not
connected; skipped glyphs are drawn on demand.
* fix(mobile): lead workspace-list agent rows with the conversation name
* fix(mobile): name a same-tab child once its parent row is gone
* refactor(runtime): share the structured chat tab id parser
* test(runtime): build the row-name summary without a cast
* perf(runtime): read each worktree's session once per worktree.ps pass
* plugins: tell once when an automatic update waits for review
When the auto-updater holds back an update because it asks for more access
or moved repository, show one toast per plugin version (grouped when several
wait). Review opens the existing per-plugin update review or the Updates list;
it never installs or approves. Told versions persist across restarts, kill-listed
plugins are skipped, and the copy is protected from language packs.
* test(plugins): wait on the update pass instead of sleeping in the review toast e2e
A plugin syntax only applied to files Orca showed as plain text, so plugins
could not color MATLAB .m (Objective-C), OCaml .ml/.mli (F#), Julia .jl,
JSP .jsp/.jspf (HTML stopgap), or Haskell .hs and Erlang .erl/.hrl, which
Orca maps to ids it never registers a grammar for.
Now an enabled plugin syntax wins for any file name or extension it
declares, except files Orca opens as TypeScript, JavaScript, JSON, CSS,
SCSS, Less, HTML or Markdown (their language drives completions,
diagnostics or Markdown features). JSP stays overridable. Ties between
plugins go to the lower syntax id instead of load order. The editor and
both diff views share the hook, so they agree.
The profile-state backup temporary name embedded the full 64-char SHA-256
owner scope, producing 191-195 byte names that fail with ENAMETOOLONG on
eCryptfs homes (namemax=143). Use an 8-char base64url prefix (48 bits) of the
digest instead; orphan cleanup still reclaims files written with the legacy
full-hex scope.
Fixes#33936
* feat(mobile): give every corner one of three shapes
Corners had drifted to 17 different radii, so a screen mixed 6 pt
buttons, 14 pt cards, 12 pt groups and 4 pt badges. Every corner now
comes from the shape of what it is:
- control (12): anything you tap — buttons, icon buttons, fields,
pickers, segmented controls, key caps.
- surface (16): anything that holds content — cards, groups, sheets,
banners, code boxes. Dialogs and the composer box use floating (24).
- pill: badges, chips, status pills, dots, avatars and the composer's
own round buttons.
- tile (10) for icon squares and mark (4) for inline code and check
squares; nested shapes take the parent's radius minus the gap.
Every rounded style also asks for the continuous (squircle) corner
curve; iOS draws it, Android and the web ignore it.
* fix(mobile): keep circles and pills on the circular corner curve
The continuous curve squares off a corner radius that reaches half the element's size, so the FAB, dots, avatars and the composer's round buttons stopped being round on iOS.
* fix(mobile): put the anchored detail card on the surface radius
* fix(mobile): keep the circular curve on unclipped bordered styles
React Native strokes an unclipped visible border on a circular path while it fills the background with the continuous curve, so the two edges drifted apart at the corners.
Orchestration-created chat workers are now classified as chats by their Orca session ID, so worker-start --terminal reuses a finished chat worker the same way it reuses a terminal worker. Dispatch admission and the coordinator census share one new-work rule (kept open, release in progress, held by a finished Dispatch, actively assigned pane), and the structured session binding is keyed by the worker's root session.
* fix(native-chat): drop a cancelled image's space from a hidden chat's draft
* test(native-chat): name the hidden image cancel test for what it checks
* refactor(native-chat): read the previous node instead of a flag when dropping a chip's space
* test(native-chat): expect a handed-back edit to drop the kept chip's space
Close focused the editor through Tiptap's focus command, which waits a frame. The Find input unmounts first, so a fast Mod+F hit the body and the editor-scoped shortcut ignored it; hidden Linux E2E windows never ran that frame at all. Focus the ProseMirror view synchronously before Find unmounts.
When two desktops pair both ways, B mirrored A's browser tab and republished
it in its session inventory under A's workspace and page ids. A applied B's
copy onto its own workspace, so A's local page became B-hosted and showed a
dead Remote browser pane (selector_not_found from B).
Browser rows whose page runs on a paired server now carry
hostRuntimeEnvironmentId, as editor rows do since #33752, so B serves them to
its phone but not to other runtimes.
Fixes#36386