* fix(sidebar): keep the host pill legible on the selected workspace card
* fix(sidebar): keep nested lineage cards out of the selected host pill rule
* fix(sidebar): leave the host pill label on its theme-validated color
* fix(sidebar): derive the selected host pill border from the validated sidebar foreground
* fix(ghostty): import the primary font-family from a repeated fallback list
* fix(ghostty): report dropped font-family fallbacks on import
* fix(ghostty): append font-family across config files like Ghostty
* fix(mobile): keep readable automations when a newer host sends unknown values
The phone read the host's automation list, run history and editor inventories all-or-nothing, so one row with a value this build doesn't know (a newer agent, surface or host kind) blanked the whole Automations screen. Read each list row by row with the shared salvage reader and the desktop reader's ownership rules, keep automations whose agent this build doesn't know (labelled by id, validated by the host), and read an unknown default agent as unset.
Tracked in STA-13196.
* fix(mobile): keep automations with a newer surface, summary or run host
Plan check follow-up: the phone never writes the surface, so an unknown one keeps its row and skips the phone-side argument pre-check like an unknown agent; an unreadable last-run summary reads as none, as on desktop; a run's host id is display-only, so a newer host kind keeps its run.
Tracked in STA-13196.
* fix(mobile): keep an unknown saved agent out of writes and never read an unreadable run or pane as gone
* fix(mobile): keep the keyboard open when a tab tap opens its details card
Tracked in STA-13199.
* fix(mobile): keep the keyboard open when the session title opens its details card
The session title opens the same details card as a tab, but sat outside every
keep-focus container, so its tap moved focus off a focused composer or terminal
input and closed the soft keyboard. Attach the existing keep-focus helper to the
title button so the tap leaves the field focused, as tab taps already do.
* fix(mobile): keep the keep-focus helper safe where there is no DOM
The session title now attaches the keep-focus helper, and the title renders in
renderers without a DOM, where checking against HTMLElement throws. Check that
HTMLElement exists first, as other mobile code already does.
* test(mobile): use a type-only namespace import for the react-native mock type
* Keep chat pickers usable while options recover
* Show only the current process's reported permission and effort after a restart
When a native chat's agent restarted, the first frame at the new process could still show the old process's reported permission (for example "Full access" on an Own chat) as pickable, then flip to unknown a moment later. A reported value now counts only while the options answer that carried it belongs to the current process, on desktop and phone.
Also updates the restart tests: a pick made while the new process's options read is pending or failing goes to the new process on both desktop and phone.
* fix(cli-runtime): preserve advisory inspection results on close failure
* fix(appimage): keep close errors from overriding verified inspection results
The AppImage identity and stable launcher readers closed their read-only
handle in an unguarded finally, so a close error replaced the verified
result and broke CLI status, startup crash capture and launcher upgrades.
Attempt the close once and discard its error, as the CLI runtime check does.
Also load the real closeSync in the node:fs mock factory instead of inside
the test body.
* test(cli-runtime): drop type assertions from close-failure fixtures
* Clarify Hand Off when detected agents share the source usage limit
* Base the Hand Off family explanation on offered agents
Ignore agent ids this build does not know, count chats offered without
their CLI, and say the source-family agents are available rather than
detected. Keep Agent terminology consistent with each catalog.
* test: feed the unknown detected agent id as parsed wire data instead of a type assertion
* fix: explain the empty Hand Off list from the same host rows the picker shows
The usage-limit explanation read desktop preferences and the offered list, while the picker rows use the host's preferences. A host that disables every agent while the desktop enables the source agent then claimed the source family was available. Derive the explanation from the settled picker rows before the family filter, counting only rows that can open.
* fix(terminal): recognize labelled Claude composer frames
* fix(terminal): recognize Claude's default simulated cursor in the composer
Claude Code's default caret hides the native cursor and parks it at a
painted caret cell, so its composer (labelled or bare) was never detected:
the draft leaked into screen reads and automatic mail was not held.
Accept a hidden cursor on the Claude prompt only inside a complete composer
frame with the cursor right of the glyph. On the normal buffer the parked
cursor suffices (Claude drops the inverse caret paint on focus-out); on an
alternate or unknown buffer the cursor cell must also be inverse. Both facts
are read per call from xterm, only while the cursor is hidden. The closing
rule now comes from the row where the continuation walk stopped, removing
the extra rows-below scan. Codex and boxed composers keep rejecting hidden
cursors.
* test(terminal): scrub worktree name from blurred Claude fixture
* Make goal command entry consistent across supported chat agents
* Keep the composer goal mode binding local
* Open the goal chip only for the provider's built-in /goal
Carry the provider's built-in marker on slash-command rows (Claude's
initialize descriptors report it) through to the picker, and enter goal
mode from a picked /goal only when the session has host goals or the row
is the provider's own built-in. A user command or skill named goal keeps
its normal invocation, and bare /goal without host goals is sent as
typed again. Host goal objectives resolve pasted text before the /goal
prefix is stripped, as before.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Mark Grok's built-in slash commands from its initialize response
Grok lists its own built-ins in the initialize response of each process.
Carry that onto the session command rows as the existing built-in marker,
so picking Grok's /goal opens the goal chip like Claude's. A row that
Grok tags as a skill or workflow never takes the marker, since a session
can close the built-in and let a same-named skill take its name.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Feed the Grok built-in command test a recorded initialize response
Replace the hand-trimmed initialize reply with Grok 1.0.70's recorded
initialize result, keeping its seven built-in command rows and removing
host-identifying metadata.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Type the composer send result as boolean in the IME composition test
The structured composer send now resolves to whether it was admitted.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Continue in Chat UI opens the chat before closing the terminal, and stops a CLI still starting
The chat tab now opens right after the terminal and is selected before the terminal closes; its create waits for proof the CLI exited, so the workspace never sits empty while a large conversation loads. The terminal is kept only when the host proves its shell runs nothing, so a CLI clicked before its first status update is stopped instead of left writing beside a chat that can't send.
* Continue in Chat UI asks before stopping another job, shows a chat already opening, and keeps a failed chat in place
A job started in the shell after the agent quit now goes through the terminal's running-process prompt instead of being stopped silently. A pane not bound to its terminal on this client is treated as unverifiable and closed, like any unverifiable read. When a chat for the same conversation is already opening, Continue shows it and keeps the terminal rather than handing the terminal into a launch it doesn't own. After the terminal closed, a failed start stays in the chat that took its place and says why once, including when the CLI's exit couldn't be confirmed, so the workspace is never left blank. A launch that ended before its tab opened reports its failure again, and a tab picked during the close stays selected.
* Continue in Chat UI asks before stopping a different agent, and answers a chat still opening without joining it
Only the conversation's own agent is stopped without the running-process prompt; another agent started in that shell is asked about like any other job. A chat already opening the same conversation is found before the new chat starts and shown (or named, from the tab menu) with the terminal kept, so an unconfirmed chat's create is never re-sent beside the running CLI. The in-chat message for an unconfirmed stop no longer points at a host connection. Test timestamps are fixed values.
Port the xterm, webgl, search, serialize and image source patches to upstream
08829c7. Where upstream fixed the same bug (#6204 mouse NaN guard, #6226 warm-up
cancel on dispose, #6235/#6237 serialize width clamp) take upstream and keep only
Orca's remaining delta (domContainer warm-up gate, clipped wide-cell blanking).
Bump mobile's webgl subset patch alongside. Drop serialize fuzz seed normal:5,
which now round-trips.
* fix(agent-status): match OSC 7501 spec length limits and progress parsing
* fix(agent-status): keep a capped OSC 7501 body whose ST splits across chunks
* fix(plugins): give no catalog release entry for a commit install would reject
The release entry printed by validate:plugin and orca plugin release-entry hashed the
committed HEAD tree without running install's file checks, so a commit with only
orca-plugin.json (declared theme files left uncommitted) got an entry whose install fails.
Run install's validatePluginInstallTree on the materialized tree, shared with install staging.
* refactor(plugins): return early per release-entry rejection; drop the temp tree path from the note
* fix(plugins): check the release entry the way install reads it
Read the committed manifest with install's 1 MiB bound, and round-trip the entry through
the catalog release schema and pluginCatalogReleaseError, so a field a catalog drops (e.g.
a contribution title over 160 chars) gives a note instead of an entry install refuses.
* fix(plugins): name the field and drop temp paths in release-entry notes
* Preserve typing during worktree chat handoff
* Keep chat drafts editable during worktree creation handoff
Transfer the live draft immediately and defer deleting its saved source
until the destination write succeeds, including after retries.
* Fix worktree creation handoff when an agent prompt is ready
Allow approval and question prompts to receive the handoff while
preserving unsent drafts and waiting for text composition to finish.
* Fix worktree creation handoff when chat cannot show an editor
* Prevent stalled worktree chat handoffs while preserving IME input
* feat(sidebar): render linked issues and tasks as compact reference rows
Issues, Linear and Jira tasks in the workspace hovercard now use the same
compact two-line row as PRs: state glyph, single-line title, provider mark,
identifier, source, labels summary, and a row actions menu.
* refactor(sidebar): share row menu shell and task actions across hovercards
- WorktreeCardRowActionsMenu backs both PR and task row menus
- Card tasks are expressed as attachments so single and linked rows share
labels, sources and the task menu
- getWorktreeCardTaskActions replaces duplicated copy-link and dismiss
wiring in WorktreeCardMeta and the activity hovercard
* fix(editor): release closed diff tabs' Monaco models under tab groups
Tab groups key each diff view by `${file.id}::${tab.id}`, and DiffViewer
derived its model-path owner prefix from that scoped key. Closed-tab cleanup
sweeps `diff:*:<file.id>` prefixes, so it never matched and every closed diff
tab leaked its original and modified TextModels. Each live model holds a
listener on Monaco's language-service change event, which surfaced as the
"potential listener LEAK" renderer error and an eventual renderer OOM.
DiffViewer now takes the open file ID as the model owner for its path prefix,
keeping the tab-scoped key for the per-view path segment and view-state cache.
* fix(editor): release Changes-mode diff models and pairs superseded across remounts
Changes mode gives edit tabs a DiffViewer whose models the edit-close path
never swept, so each closed Changes-mode tab leaked an original/modified
TextModel pair. ChangesModeView now passes the file ID as model owner, and
closing an edit tab that mounted a Changes view sweeps that owner's diff
prefixes (ordinary edit closes still skip registry enumeration).
DiffViewer remembers each view's last model paths per owner, so a remount
after content changed while unmounted disposes the superseded pair instead
of holding it until the file closes.
---------
Co-authored-by: m4air <m4air@m4airs-MacBook-Air.local>
* docs(agents): drop casting section; casting lint now explains the policy
The changed-code casting scan already enforces no type assertions (as const allowed)
and SAFETY:-annotated suppressions. Its diagnostics now print the allowed
alternatives and the exact suppression form, so AGENTS.md no longer needs to.
* docs(agents): oxfmt AGENTS.md so CI formatting check passes
* Bound agent journal queue before reading control binds
Limit startup queue growth by bytes and operation count, while preserving
backpressure and allowing final Pi events to drain after exit.
* Test read pause and resume for large Claude messages
* Adjust Chat UI announcement rollout and show a nonblocking card
- Assign new profiles to an 80/20 control and treatment experiment,
with treatment split between all users and eligible harnesses.
- Show terminal-default users a corner card and record targeting
assignments to measure outcomes.
* Refine the Chat UI announcement card layout and copy
* Add live switching demo to Chat UI launch tip experiment
* fix test
* Shorten Chat UI announcement telemetry documentation
* Add a swapped-button cohort to the Chat UI launch tip
- Assign 10% of new enrollments to the new variant while preserving existing assignments.
- Track live-switching link exposure and simplify the demo copy.
* Adjust Chat UI launch tip rollout and switching options
- Add a centered dialog cohort and independently vary button emphasis.
- Target users who have not switched and allow switching back to terminal.
- Record live-switching exposure only when the link is visible.
* Validate persisted chat tip emphasis before initializing telemetry
* Target chat UI launch tips using the current default surface
Freeze eligibility at enrollment while preserving the historical audience.
* perf(browser-pane): stop the client-hosted page position loop when nothing changes
The shared position driver requested an animation frame every frame for as
long as any client-hosted page or docked DevTools was shown, keeping the
renderer and compositor at display refresh rate while the app sat idle.
Frames now run only in a short settle window after something that can move a
pane or open an overlay: DOM mutations, viewport resize/scroll/DPR changes,
transitions and animations, input/focus/drag events, font and resource loads.
Finite running animations extend the window; endless spinners do not. The
DevTools dock also resyncs directly on drag passthrough changes.
* perf(browser-pane): settle position syncs in two frames and ignore terminal/editor DOM churn
Busy-case measurement showed the 250 ms settle window kept syncs at 60/s
under any steady DOM traffic (sidebar rows re-rendering every 200 ms,
xterm scrollbar style writes while output streams). Every DOM change
already wakes the driver, so two frames per wake are enough; mutations
confined to .xterm/.monaco-editor can't move panes and no longer wake it.
* perf(browser-pane): recheck client-hosted page position once a second without frames
A pane move with no DOM mutation, event or CSS animation (a style sheet
edit, a script-driven animation) left the overlay stale until the next
wake. A frameless 1 s recheck bounds that; unchanged rects write nothing,
so an idle app still requests no frames.
The stale-review shimmer animates background-position on the main thread for
as long as review data stays stale, once per sidebar card. It now steps 30
times per sweep and pauses when its card is out of view. Pending-check
spinners also pause offscreen, and pending checks, the phone-driver dot and
the notebook busy pulse stop under reduced motion.
* Read the Claude account route off the main thread for chat reads and sends
Chat option reads, model-list reads and every send resolved the selected
Claude account through the router's synchronous state-file read. Route them
through an async twin that shares the same login cache. A relaunch for a
removed account no longer prepares (and recreates) its folder.
* Chats follow an account switch for worker results too
A coordinator chat that had started any worker kept answering every
message on the account it used before a switch, until every worker it
started finished and every result was read. Every new turn, including a
worker's result, now launches on the selected account. A reply already
running finishes where it started, and a chat whose own background work
is still running keeps its process until that work ends.
This removes the per-account result tracking that held the old account
(the parent-work query, the response index and its queued and
usage-limit carve-outs, the input account recipe on the startup attempt,
and the read-only journal restore that only fed it), plus the journal
index and change-feed fields only it read. A relaunch can no longer
prepare, and so recreate, a removed account's folder. The record still
saves the account each launch actually ran on, for Session History.
* Chats reading the Claude account at the same time share one read of its state file
* Test worker results after an account switch through the real account choice, next to a user's message
* Pass the mailbox pointer's sender snapshot in the account-switch worker test
* Import worker custody helpers once in the process listing, and build the expected homes without a shared object
* fix(antigravity): avoid built-in skill files in repositories
* fix(antigravity): say which runtime releases write the skill file
* test(antigravity): read the recorded 1.4.1 model and Effort options
* test(antigravity): type the recorded option key
* feat(mobile): give Tasks and Automations their own sections on Home
Home's bottom bar squeezed Tasks, Automations and New workspace into
three equal buttons whose labels cut off on phones. Tasks and
Automations now each get a section under Desktops, matching the old
Tasks card, and the bar keeps New workspace at full width. The
desktop's phone preview is redrawn to match.
* fix(mobile): neutral Tasks subtitle and a preview that shows both sections
The Tasks row said "Issues and pull requests", GitHub's wording, to GitLab,
Linear and Jira users too; it now reads "Browse your issues and reviews".
Section subtitles wrap to two lines so translations and large text aren't
cut. Each section now owns its desktop picker.
The desktop phone preview cut off the new sections in its narrow frame and
in shorter windows. It now shows two Jump back in rows, and drops the extra
rows when the frame is narrow so Tasks and Automations stay in view.
* chore: drop unrelated pnpm-lock.yaml churn from the main merge
* fix(terminal): send exactly one mode-2031 report per flip, including after reconnect
xterm 6.1's built-in DECSET 2031 handling sent a second ?997 report on every
appearance flip alongside Orca's own push, and a flip that happened while a
pane's stream was down was never delivered. Orca panes now own mode 2031 and
?996n, the push no longer waits on the phone-fit override, and onConnect
catches the pane up through the existing dedupe.
Closes#10913
* fix(terminal): catch up legacy relay reattach and forget 2031 on process replacement
* fix(terminal): keep 2031 with the viewer holding the grid and catch up on release
* fix(terminal): re-send 2031 after another viewer's hold and keep owed reports across repeated DECSET
* fix(terminal): tint the scrollbar thumb from the terminal foreground so it shows on light themes
* fix(terminal): dim the foreground-tinted scrollbar thumb on dark themes
* fix(terminal): keep the scrollbar thumb at least as visible as the legacy grey on muted dark themes
* fix(terminal): document the opaque cap on the scrollbar contrast floor
* docs(settings): point the Chat or CLI choice at Settings → Agents
The Settings → Chat → Providers page was removed; the default and the per-agent choice now live in Settings → Agents.
Tracked in STA-13195.
* docs(settings): say the Chat | CLI choice is for supported agents on this computer
Offscreen or clipped working rings dropped their CSS animation through one
shared IntersectionObserver; revealing a ring restarts it and the existing
animationstart anchor keeps every ring on the shared phase.
The spinner bench gains an always-on comparison variant and reports
compositor frame counts.
* test(e2e): expand child agent lists in spinner benchmark fixture
Sub-agent lists start collapsed since #36936, so the spinner benchmark's
working subagent rows never mounted and the census assertion failed.
* test(e2e): wait on child-agent disclosure expansion instead of sleeping
The diff-review dock has no next-step bar, so its bottom action now offers
Resolve with <agent> for a conflicting pull request, with the launch notice.
In the hub the bar shows its agent launch notice on the Pull request tab
too, since the removed box was the one that showed it there, and an Unlink
failure reaches the bar's notice now that the PR tab has no footer.