* refactor(ipc): group handlers and PTY support by owner
* chore(test): prune retired IPC timeout baseline
* fix(types): include worktree path settings in web project
Review-note state was a dependency of the react-markdown components map, so
every note change produced new component types and React remounted the whole
preview, resetting scrollTop. Annotated block components (p, pre, blockquote,
table, li, headings) are now module-level and read annotation renderers from
a context, and links read worktree routing at click time, so the map stays
stable across note changes. Find highlights are re-applied when annotations
change because the block wrapper can detach the highlighted ranges.
Fixes#43317
* fix(terminal): show the real Windows error when node-pty cannot load conpty.node
* fix(terminal): keep the first load error across the retry and cap the retry wait per module
When a branch's PR is merged and new commits land on the same branch,
the Checks tab kept showing "Pull request already exists" with Open
Review pointing at the merged PR. The hosted-review cache read and the
create-eligibility lookup both ignored the worktree HEAD, so a merged
review stayed current until its cache entry expired.
Filter stale merged reviews out of the Checks tab's hosted-review cache
read, and pass the worktree HEAD into create eligibility (Checks,
Source Control, Create PR intent, and main's create preflight) so the
branch lookup re-reads merged reviews once HEAD moves.
* fix(native-chat): hide overlays of hidden chats and show the agy model
Open popovers, menus and dialogs portaled from a chat tab stayed on screen,
pinned to the top-left, after switching tabs: a hidden Activity does not hide
portals beneath an already-hidden host. Each chat pane now portals overlays into
a body-level container that hides with the pane.
Antigravity reads its model from user-level config only, so the configured
default it reports is now recorded and new chats show the model instead of
"Model".
* test: stub the act environment without a cast
* feat(workspace-cleanup): delete Orca's own chats of deleted workspaces with their history
* fix(workspace-cleanup): close a deleted chat's shown tab the usual way and leave its rows to the next tick
* fix(workspace-cleanup): keep a chat worker that still owes a dispatch
* chore(workspace-cleanup): baseline the forgotten-chat journal SQL, which runs in the journal workers
* fix(workspace-cleanup): follow the tab table's move after rebasing onto main
* fix(workspace-cleanup): check the gone folder after the catalog re-read, nearest the delete
The host-side readiness wait slept 250 ms between checks, so a launch
found orcad ready on average ~125 ms after it published readiness, and
up to 250 ms later; its fractional-sleep probe also slept 250 ms before
the first check. Every activation, recovery and rollback launch waits
in this loop.
Hosts with fractional sleep now test for the readiness file every
50 ms with the shell's builtin `[ -s ]`, which starts no process. The
checks that do (the deadline, a complete readiness line, a launch that
already exited) still run every 250 ms, and on every iteration once the
file has bytes. Hosts with integer-only sleep keep the 1 s cadence, and
Windows keeps its own wait.
Claude-Session: https://claude.ai/code/session_01GR9M8uNzGJ1Gaam7ML1Sk3
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Pi native chat launched Pi with PI_CODING_AGENT_DIR pointing at the chat
account dir but never wrote Orca's managed extensions there. Only the PTY
env path called installManagedExtensions, so orca-agent-status.ts was
missing (or stale after an upgrade) for chat-launched Pi and subagent rows
never appeared.
piRpcLaunchPart now refreshes the managed extensions in the chat agent dir
when agent status hooks are enabled for Pi. The gate is read from the
host's current settings via a shared structuredAgentSettingGates helper.
Fixes#43291
Workspace sleep counted a released chat as at rest only with exit evidence
or a never-started record. An adopted conversation carries a provider
handle, so it never looked never-started, and with no agent ever run it
had no exit evidence either; sleep reported it as still running. Gate on
the lease not possibly writing instead (owner exited or lease never
acquired), which does not depend on the provider handle.
Fixes#42703
* fix(wsl): wait through a stuck orcad install lock and say what Add waits for
On a local WSL distro the orcad install waiter no longer gives up at the
15-minute step timeout just before stale takeover becomes possible: lock
age sets its deadline (capped at 22 minutes), so a relaunch waits only
the rest of the window. The Add WSL host dialog and host row say what
the connect waits for, clear once the lock is taken, and say when Orca
stops waiting if takeover would come too late. SSH keeps its timing.
A failed acquire keeps its fence token (a lost reply may follow a
committed create or steal), and install cleanup removes the lock only
while it still names this install's token, failing (and keeping the
token) when the lock cannot be inspected.
Fixes#43019
* fix(wsl): show only the added distro's progress in the Add dialog
Every window hears every Add's progress, so the dialog now applies only
events for the distro it is adding (case-insensitive); another window's
Add can no longer clear or overwrite its lock-wait text.
Continue in Chat UI refused Claude conversations whose saved lines name the
directory the chat started in, e.g. after moving a worktree and copying its
chats into the new workspace's Claude project folder. Claude resumes by the
cwd's project bucket, so a transcript in the workspace's own bucket (including
Claude Code's hashed name for paths over 200 characters, compared without
case on Windows) now belongs to that workspace.
Fixes#43242
Co-authored-by: Claude <noreply@anthropic.com>
The hourly database maintenance pass now deletes each Run whose work all
settled and that nothing touched for 24 hours, one Run per short write
transaction at a tenth of the thread, unless a coordinator it bound or a
reader of its unread mail is still live. It also deletes acknowledged relay
items (keeping each stream's newest) and worker rows whose Dispatch is gone.
Co-authored-by: Yasyf Mohamedali <yasyf@poetic.com>
Hidden worktree surfaces (prewarm, background mount measurement, browser
paint retention) and the hidden workbench stay laid out at opacity 0 and
inert. Chromium collects -webkit-app-region from every box except
visibility:hidden ones, and the later region wins, so a hidden strip's
window-drag area stacked after the visible surface swallowed clicks on the
visible pane's tabs, close buttons and "+" whenever no terminal had focus.
Hide drag strips inside inert subtrees so they drop out of the region list.
Fixes#43150
* fix(wsl-hosts): offer the WSL host for folders on a drive mapped to a distro share
A drive letter mapped to \\wsl.localhost\<distro> (net use W: ...) named a
distro folder Orca only recognised in its UNC spelling, so adding W:\home\...
never offered the WSL host. Main now resolves a drive's share (bounded, cached
per drive) and the add-project offer and sidebar drop check that spelling.
Refs #40648
* fix(wsl-hosts): share one drive lookup, keep late answers, skip duplicate offers
- A slow or hung drive keeps its one in-flight lookup cached: later adds wait
on it (bounded) instead of starting another, and a cold distro's late answer
serves the next add. Only a failed lookup is retried.
- Dot segments are normalized; ones climbing above the drive root resolve to nothing.
- A drive path already saved as a project skips the offer, so accepting can't
add a second copy at the UNC spelling.
- The offer takes the caller's generation, so a closed or reopened dialog never
shows an older add's offer.
- Sidebar drops take the busy guard before the lookup.
Refs #40648
* fix(wsl-hosts): only the newest offer shows; a deduped add explains instead of moving nothing
- Each offer, close, reopen and host change starts a new request; a slower
older lookup never replaces a newer add's offer.
- Registration matches by real path, so accepting can land on a project already
saved as W:\. The move now uses the repo the add answered with, and when that
is a drive-spelled row it says so instead of opening a move that can't take it.
Refs #40648
* chore(wsl-hosts): follow main's moved shared constants and execution-host modules
Refs #40648
Plugin workers now get GIT_TERMINAL_PROMPT=0 so git fails cleanly instead of
prompting. The full-access warning and the authoring docs now say a worker can
use saved logins and SSH keys and that the environment filter is not a sandbox.
Adds tests for dropped code-loading variables and Windows case folding.
Undo and redo restore known text and move the cursor without an explicit
reason, so blame stayed hidden or showed the previous line's value until the
cursor moved again.
Remote (SSH or paired-host) file links only offered "Download & open with
default app". Add a "Download…" row that saves a copy to a user-picked path
without opening it. Folder links are saved over SSH when the host supports
recursive transfer and reported otherwise; paths that resolve to this
computer are reported as already local. Hidden in paired web clients, which
have no local disk.
Shares the downloaded-file toast and SSH folder download with the file
explorer's Download action.
Fixes#43038
Co-authored-by: Claude <noreply@anthropic.com>
The sweep now also deletes Codex rollouts whose workspace folder is gone,
under the same folder rule. A rollout is deleted under every name it has
across the host's Codex homes (sessions and archived_sessions, .jsonl and
.jsonl.zst), only when every copy is idle, its last turn ran in the gone
folder and Codex holds no writer lock for the thread. History link passes,
backfill and resume skip a rollout while it is being deleted, and no longer
leave a dangling symlink when the source is gone.
Fixes#16776
Co-authored-by: Max Hsu <maxmilian@gmail.com>
Once dragged or resized, the floating workspace keeps its saved rect, so
after a monitor or display-scale change it can sit mostly off-screen with
no way back short of clearing app storage. The trigger's right-click menu
(status bar icon or floating button, which stays reachable when the panel
does not) now has "Reset Floating Workspace Position and Size": it clears
the saved bounds and maximize flag, and a mounted panel returns to the
default anchored layout immediately. An unmounted panel picks up the
default layout on its next mount.
Fixes#42721
Co-authored-by: Shaukat Ibrahim <shaukati@amazon.com>
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
The rollout walker warned on every opendir error after its caller's handler
had already classified it, so each pending backfill date without a day folder
logged ENOENT at startup. The index heal logged a stack trace for every pass
on a machine without Codex. Absence is now an empty listing, and a missing
CLI aborts the heal quietly for a later retry, like the Codex model catalog
probe already does.
Co-authored-by: Bensonluo <18586615+Bensonluo@users.noreply.github.com>
Co-authored-by: Pr1p <48266347+Pr1p@users.noreply.github.com>
* feat(editor): git blame for the cursor line
* fix(editor): blame review fixes
Co-authored-by: AnddyAgudelo <andiagudelo@gmail.com>
Co-authored-by: Matt Tuusberg <tuusberg@gmail.com>
Co-authored-by: eric <1034850788@qq.com>
Co-authored-by: 5hseok <110245258+5hseok@users.noreply.github.com>
* fix(editor): blame follows the owning repo, focused pane and visible editor
* fix(editor): blame review fixes for compiler and lint
* fix(editor): blame keeps the status bar steady and resends nothing when HEAD is unchanged
* fix(editor): inactive blame panes do no work; read worktrees through the record accessor
* fix(editor): released blame hides; split panes never take over; local IPC passes knownHeadOid
* fix(editor): blame baseline tracks shallow history, uses source lines, and drops stale results after HEAD moves
* fix(editor): blame hides the old baseline while a HEAD-change reload is pending
* fix(editor): blame baseline is the raw blob; only real HEAD moves drop it
* fix(editor): only a failed focus reload keeps blame
---------
Co-authored-by: AnddyAgudelo <andiagudelo@gmail.com>
Co-authored-by: Matt Tuusberg <tuusberg@gmail.com>
Co-authored-by: eric <1034850788@qq.com>
Co-authored-by: 5hseok <110245258+5hseok@users.noreply.github.com>
Plugin workers run with an allowlisted environment that dropped
XDG_RUNTIME_DIR, DBUS_SESSION_BUS_ADDRESS and SSH_AUTH_SOCK, so git
credential helpers backed by libsecret and ssh-agent could not reach the
user's session. Allowlist these socket pointers and strip Chromium's
`disabled:` bus marker so GLib falls back to $XDG_RUNTIME_DIR/bus.
Fixes#43189
Frames only got the colour tokens and --orca-radius, so a plugin panel could
not match a user's chosen app font. Add --font-sans and --font-mono to the
panel design token allowlist; the existing theme sync already refreshes frames
in place when the root style changes, so a font change now restyles them too.
Native chat visuals already listed both tokens and now get them from the
shared allowlist.
Fixes#42575
Co-authored-by: Claude <noreply@anthropic.com>
* fix(editor): give new files the owning host's line ending, add End of Line setting
A file with no line breaks yet took the ending of the OS the window runs on,
so a Windows client wrote CRLF into files on WSL hosts and Linux SSH hosts.
Such models now adopt the ending of the host that owns the file (LF on Linux
and macOS, CRLF on native Windows), and Settings > General > Editor gains an
End of Line choice (Auto, LF, CRLF). Files with line breaks keep theirs.
Fixes#43216
* fix(editor): resolve paired-host files by endpoint and keep new-file EOL through undo
- Resolve a file's host from its endpoint-relative authority, so a file on a
WSL host or paired server reads that host's OS instead of this client's.
- Only a WSL host's own files fall back to Linux; SSH targets behind it do not.
- Track whether a model was loaded with line breaks, and keep a new model on
its chosen ending through undo, late host reports, and setting changes.
- A model without line breaks adopts the ending of external text written in.
- Read remote OS facts directly so the selector builds no translated strings.
- Docs describe Auto by where the workspace runs (Windows projects in WSL get LF).
* fix(editor): keep undo valid across late EOL conversion; respect text written while unmounted
- Convert typed lines with an undoable pushEOL step instead of a bare setEOL,
whose stale offsets corrupted Undo. Only a new choice (late host report or
setting change) converts multiline text, so Undo of the conversion sticks.
- Record outside text with line breaks on the model from the content-sync
paths, so a remounted editor never converts text the user did not type.
* fix(editor): never convert a new file's line endings once its lines are saved
A clean buffer with line breaks matches the file on disk, so it is treated as
an existing file: changing End of Line later no longer converts it, marks it
dirty, and lets autosave rewrite it. Only typed, unsaved lines are converted.
* fix(editor): treat a new file as saved only when its bytes match disk
The dirty flag ignores Markdown trailing whitespace, so a clean flag did not
prove typed line breaks were saved. Compare the model text with the buffer's
last known disk signature instead. Also type the line-ending state by its
model shape for the anti-slop lint, localize the LF/CRLF labels, and note in
the docs that autosaved lines keep their ending if the host reports later.
* fix(editor): only set a new file's line ending while it has at most one line
Converting text that already has line breaks kept running into cases where
saved or in-flight lines were rewritten. The editor now sets the ending only
while the model has at most one line, at mount and whenever the choice
changes, and re-applies it when Undo returns to one line. Text with line
breaks is never converted. This removes the multi-line conversion, its undo
step, the per-model state and the disk-signature check.
* fix(editor): keep an existing file's line ending after it is emptied
A model that already had line breaks when the editor mounted, or got them
from a programmatic or outside write, is flagged as an existing file and
never has its ending set again, so emptying and retyping it keeps the file's
ending. Line breaks the user typed do not set the flag, so Undo back to one
line still restores the chosen ending. Scope the docs sentence on outside text.
* fix(editor): carry a buffer's line ending through model rekeys and renames
A replacement model was built from text and language alone, so an emptied
or one-line existing file lost its ending and existing-file status. The rekey
carry now copies both to the model it creates. A rename remounts the editor
under a new view id, so the rename records the old model and the remount
carries an unsaved draft the same way. Existing models at the destination are
never touched.
* fix(editor): hand a renamed buffer's line ending over as a snapshot
The rename handoff stored the old model's address, which model cleanup could
dispose before the renamed tab mounted, and records were never cleared. The
rename now snapshots { ending, existing-file } from the old model, only for a
tab with an unsaved draft, keyed by the new tab id. The tab-close cleanup
drops it, and the remount applies it only to the model it builds from the
draft. Same-view re-keys still carry from the live model.
* feat(wsl-move): let users collapse and hide moved-project entries
The sidebar's moved-to-WSL list now has a collapsible header (persisted),
and an entry without local chats can be hidden (persisted). Undo stays
reachable from the project's menu on the WSL host; an entry whose host no
longer lists the project always shows, since it is the only way back.
Fixes#42628
* fix(wsl-move): match Undo to the clicked host row, keep list fields desktop-only, unhide on Undo
- The workspace menu's Undo follows the clicked row's own host, not the first
registration of a repo id listed on several hosts.
- Both list fields are pairing-local, so paired reads and writes never carry
them and a host's strict ui.set never sees them.
- Undo forgets its projects' hidden entries in main, so moving one again to
the same distro lists it even if the next startup read fails.
- Header uses text-2xs and the chevron honors reduced motion.
Refs #42628
* fix(wsl-move): resolve folder hosts alike for hide and Undo, prune only on the desktop
- A folder's host comes from its record, falling back to its project group's
host, in both the hide rule and the right-click Undo.
- The startup prune is skipped on a web client, whose moved list is always
empty.
Refs #42628
* fix(wsl-move): never let unhiding moved entries block the undo reload
Refs #42628
* fix(wsl-move): read folder records through the catalog selector
Refs #42628
* fix(native-chat): point a refused WSL-folder chat at WSL hosts with a one-click path
When a chat is refused because a project opened from Windows runs under WSL,
the toast and chat notice now offer "Run <distro> as a host" (turns on the
setting, adds the distro, then opens the move for that project) or "Move to
<distro> host…" when the host already exists. The CLI refusal keeps its code
and message and adds data.nextSteps (and wslDistro) pointing at the same path.
All of it shows only while this desktop can add WSL hosts.
Refs #42126
* fix(native-chat): guard the WSL host setup and match the CLI hint to the app
- One setup per distro (case-insensitive) shared by the toast and the chat
notice; setups run one at a time so a failed add always restores the setting
it read. The notice button shows Setting up… and is disabled meanwhile.
- The CLI hint now uses the same gate as the app: WSL host service present,
new work accepted, server files shipped, and no Windows-drive folder (the
drive check moved to shared code). Its steps skip what is already done:
with the host added and the setting on it says only Move to <distro> host….
- Tests for the toast wording and wiring, the resume-in-chat toast, the
notice's WSL gate, and repeated or concurrent clicks.
* fix(wsl-hosts): share one per-distro run and queue between the add-project offer and chat setups
runWslDistroAsHost now joins a run already in flight for the same distro and
queues runs for other distros, so the add-project offer and a chat refusal's
setup can never interleave their setting restores.
* fix(wsl-hosts): run host setups for different distros side by side
A hung add or deploy for one distro no longer holds setups for every other
distro. Runs in flight share the setting instead: the first records it, and
the last to finish puts it back only if none of them added its host. A run
that starts while that restore is saving waits for it before reading the
setting. Same-distro calls still join the run in flight.
* fix(wsl-hosts): import GlobalSettings from its new settings path
* fix(wsl-hosts): start a new WSL host on this desktop's agent permission
A newly added WSL host kept the product default, Full access, even when
this desktop chose otherwise, so Claude workers on it launched with
--dangerously-skip-permissions and stopped on Claude's one-time Bypass
Permissions warning. Add now copies the local choice onto a host still at
the default, and the readiness failure names that warning and the setting.
Refs #42126
* fix(wsl-hosts): seed a new host's permission in one conditional host write
The read-then-write seed could overwrite a choice another client saved on
the host in between. The host now applies a capability-gated
seed-permission mutation inside its own settings write, only while it is
still at the product default. The blocked-message test replays the
captured Claude transcript.
Refs #42126
* fix(wsl-hosts): keep a synchronously throwing seed from failing Add
Refs #42126
* refactor(main): organize agent session and tab lifecycle owners
* fix(ci): carry moved backend test and mobile routing metadata
* test: retire copy assertions in headless notification test
* feat(agent): one agent picker card with agent tabs on top
Dropdown pickers show agent icon tabs, then search, the agent's models with More models, one effort and Fast row, and an Open in Chat | CLI switch. The + menu opens the same card without tabs, followed by the prompt and launch buttons. Settings and automation forms keep Permissions and Arguments as fields below the picker.
* fix(agent): address picker review findings
- Release the dropdown pane's picked setup when it closes, so a reopened
picker starts from what the Permissions/Arguments fields saved.
- Skip the trigger summary with a trigger label and memoize it on the draft.
- Remove Arguments chips from the text being typed.
- List Blank Terminal and other non-agent rows in search and reach them
with ArrowDown.
- A query naming an agent lists the agent first and picks it without a model.
- Agent tabs: roving tabindex, arrow keys, Enter/Space, Shift+Tab from search.
- Modern automations show one Arguments field.
* fix(agent): keep automation extra arguments supplemental and retranslate the trigger summary
- Restore the automation editor's supplemental Extra agent arguments field as main had it,
so saved extras stay on top of inherited host arguments.
- Recompute the memoized trigger summary when the UI language changes.
* fix(agent): keep the picker's effort and Fast row on one line
Raise the agent dropdown's minimum width to 26rem, which fits Codex's widest
measured effort row plus Fast (404px), and cap it at the available width
with 8px collision padding so it still fits narrow windows.
* fix(agent): key the picker summary on UI language without a lint suppression
* fix(agent): name the model in the closed picker trigger
Before a host's listing loads, resolve the trigger's model label from the cached host listing,
then the release's catalog, then Claude's id-based listing name, and the id only last.
* fix(agent): name live-listed models in the closed trigger and fit the text-recipe effort row
- The closed trigger reads the agent's last listing an open picker showed, so models only the
live listing names (Codex gpt-6-astra) keep their label and effort.
- Text actions label their inherit-effort choice Default, with the inherited level as its tooltip,
so the row fits one line.
* fix(agent): name a closed trigger's model from the user's pick, and keep the inherit-effort name
- Replace the per-host listing cache with a bounded record of picked models, written only at
pick time and keyed by agent, host and model; mounted closed triggers re-render when it changes.
- The text action's Default effort keeps "Same as default · <level>" as its accessible name.
* perf(agent): re-render only the pickers showing a newly picked model
Each picker subscribes to its own agent, host and model's picked row; an unchanged write is
skipped, and the subscription is hoisted to module scope.
* fix(agent-picker): address CodeRabbit review threads
- Dropdown pane derives from the controlled value when there is no form source, so a kept setup no longer shows a stale choice
- Model search drops setup rows that repeat a listed model
- Agent-only picker search matches extra choices (e.g. Custom command)
* fix(agent): label the chat/CLI switch "Open as"
Without --worktree, a reuse start resolved the caller's workspace and was refused with terminal_worktree_mismatch whenever the terminal lived elsewhere. An explicit --worktree keeps today's behaviour, including the refusal.
Clicking a legend entry in the Claude, Codex, and Muse daily usage charts
now hides or shows that token type, and the bars rescale to the visible
types across the ten days shown. At least one type always stays visible.
The Claude and Codex charts share a new UsageDailyTokenChart component.
Codex and Muse reasoning tokens are no longer double counted in the daily
bars, and the Output tooltip keeps the full output count.
Fixes#43222
Co-authored-by: Neil <neil@stably.ai>
* feat(cli): list and read agents by their Orca session ID with orca agent list/show/read
Live terminal and chat agents are listed, shown and read by Orca session ID on the selected host; sleeping native agents can be shown and read by ID. Chat reads walk the chat's own bounded history pages back to the latest /clear.
* fix(cli): agent show names a live agent and its last reply as agent list does; help says sleeping terminal agents aren't listed yet
* Bound agent reads and scoped workspace inventories
* fix(cli): agent read carries the conversation's state, so a running chat turn reads as working
agent read now reports the same state agent list/show report, from the same status
rows, in text (a leading state line) and --json (state). A turn still running no
longer reads like a finished chat with no reply.
* Fix agent conversation read consistency and limits
* fix(cli): agent read derives its state from the history it returns
agent read no longer samples live status. A chat's state comes from the
journal snapshot that produced its entries: an open turn or an accepted
send with no turn yet reads working; done needs an explicitly successful
latest turn whose reply is the newest entry shown. A terminal agent's
state comes from the provider turn markers in the same transcript bytes:
a prompt with no end after it reads working, done needs an explicit end
(Claude stop_reason, Codex task_complete) with the reply as the newest
entry. Anything else omits state. The journal lane wait added for the
live state is removed.
* Tighten agent read snapshot evidence
* Keep agent read done for finished turns
A dead owner now only voids working. Codex completion is proved by the turn's own
decoded reply rather than row adjacency, and Claude's by its end row's own reply,
without needing the prompt inside the read window.
* Adapt agent commands to reorganized shared contracts
* Warn on incomplete agent lists, read sleeping transcript paths, and retry unanswered provider id lookups
* Follow main's moved shared modules
* Bound live transcript recovery to conversation panes
* Show sleeping terminal agents where agent list lists them
agent list lists a finished terminal agent while its pane sleeps, since main keeps
that status row. agent show now counts the same sleeping place, and the help says
which sleeping agents are listed. Messaging still ignores sleeping panes.
* Clarify sleeping agent discovery and messaging help
* Say sleeping terminal agents can't be messaged and note the WSL read exception
* Follow the moved CLI argument boundary after rebasing
* Show plain Orca session IDs in agent command help
* Expect plain UUID Orca session IDs in the transcript source test
* Follow main's moved shared modules after rebasing
* Name Orca session IDs <id> in agent help, since older IDs keep their agent prefix
* Follow main's moved agent-session modules after rebasing
Drop the address field's 3px focus ring for a hairline border, size the
toolbar buttons like other pane headers (24px, 14px muted icons), and
replace the Annotate split button with one icon. Grab and Draw move to
the icon's right-click menu and the top of the ⋯ menu.