Changelog

Claude Fable 5.2 and chat references

Reference other chats by name, let agents read the history of your other chats, and choose Claude Fable 5.2, with more dependable file links and a leaner Python setup.

New

• Claude Fable 5.2 — pick it from the model picker. Its effort levels are Low, Medium, High, Extra high and Max. Claude Fable 5.1 is no longer offered; chats that used it need another model selected before they can continue.
• Reference chats by name — type @ in the composer and choose a chat by its title. The reference follows the chat if you rename it.
• Agents can read your other chats — an agent can page through the messages of any of your chats in the same organization, including its sub-agents, to pick up where another conversation left off.

Improved

• Default model — new chats start on Claude Sonnet 5.5.
• File links in chat — links to files now open files inside or outside your project, with optional line and column targets such as #L12, #L12C5 or :12:5. A missing image now shows what went wrong instead of an unexplained box.
• Python extras install faster and smaller — packages a script asks for are installed on their own instead of copying the whole managed Python environment, which could take several gigabytes of disk for the first script that needed one.

Branch conversations and work across devices

Branch a conversation to try another approach, let agents coordinate across chats, and approve desktop work from your phone, with clearer usage reports and more reliable simulation viewing.

New

• Fork a conversation — choose Fork in a chat's sidebar menu to continue from its completed history in a new chat. The original stays intact; worktree chats get a separate worktree with your staged, unstaged and untracked changes.
• Agents across chats — ask an agent to find and message another of your chats in the same organization, create a new chat, or fork its conversation to pursue another direction.
• Approve desktop work from your phone — Approve and Decline buttons now work for Qubis's own tool requests on iOS and Android. Update desktop Qubis too so it can carry out the approved action.
• Start the container engine from chat — when a simulation needs an installed engine that isn't running, the agent can offer a card to start it and continue.

Improved

• Usage — Settings → Usage adds a 24-hour view, custom date ranges, and Tokens, Cost and Agent calls selectors for both the chart and model breakdown. Compare personal and organization usage, with clearer totals and large token counts.
• Files in chat — files sent by the agent now have an Open button in desktop Qubis, including files from sub-agents. The card shows download progress, failures and expired links.
• Fewer unexpected tabs — the agent is guided to link generated files in chat so you can choose when to open them.
• Questions and progress — new question cards offer explicit choices while still accepting a typed answer, and the agent is guided to explain what it is doing as it works.
• Engineering reports — the agent has better guidance and tools for including mesh, convergence and field views in CFD and FEA reports, with results from unconverged runs labelled provisional.
• Larger Ferro models — GPU solves can use the graphics card's supported buffer sizes instead of a smaller fixed limit, allowing larger meshes on capable hardware.
• Return after sign-in — the browser's sign-in completion page now offers Return to Qubis on systems with Qubis's app links registered.

Fixed

• CAD models could drift off screen while orbiting or disappear when zooming close to a surface.
• A failed Ferro solve could take down Qubis. Solves now run separately, and failures appear in the viewer.
• Markdown images failed to render when their file paths contained spaces.
• Plugins such as monday.com repeatedly asked you to sign in after their access token expired. Sign in once after updating to save the credentials needed for future refreshes.
• Chats hosted on another computer could appear local when both computers used the same folder path. Host identification now distinguishes them once the host reconnects its chat tools.
• Tools that ran automatically could leave unactionable approval cards on other devices.
• Python and CAD scripts could fail during environment setup even when they did not need extra packages.
• A simulation whose case folder wasn't mounted could send the agent looking for nonexistent logs; the failure now explains the missing mount.
• Closing Qubis on Linux Wayland could crash during shutdown.

Setup on demand and answers from anywhere

Sign in and start working straight away: Qubis installs local tools when a task needs them, questions and approvals can be answered from any of your devices, and getting from signup to a subscription takes one click.

New

• Local tools when you need them — the setup screen after sign-in is gone. When a task needs a program you don't have, such as a container engine or gmsh, the agent asks in the chat with a card to install it or skip it for now, then carries on where it stopped.
• Answer from any device — the agent's questions and approvals for Qubis's own actions can now be answered from another computer or your phone, not only on the computer that asked. The first answer wins.
• Open Qubis from signup — the page at the end of signup has an Open Qubis link that brings Qubis to the front and moves it past setup, without pressing Check again.

Improved

• Subscribing — when an account runs out of credits, the chat stays on screen and a bar above the composer offers Subscribe or Buy credits, whichever your plan allows. Other access notices name the step that unblocks you, Free plans show the credits left beside the context meter, and the plan dialog says what a seat includes.
• Sign-in and plugin connection pages — the browser pages at the end of a sign-in or plugin connection share one Qubis design, and the plugin page lists each step of the connection and what happened.
• GovCloud billing — GovCloud builds no longer offer plans or credit purchases, which failed with a server error there.

Fixed

• The Python and CAD tools failed with a uv error on computers with an older uv installed, because Qubis used that copy instead of the one it ships with.
• On macOS, a standard (non-administrator) user couldn't install an update; Qubis now asks for an administrator's approval and installs it.
• A stopped Colima was reported as missing, and Local tools offered to install a container engine instead of starting Colima.
• Windows: opening Settings → Local tools or enabling cloud runs flashed a burst of console windows.

Files from the agent and Python environments

The agent can now hand you a file, run Python with the packages a script needs, and show its edits as the git changes they are; Settings has a new layout and GovCloud builds say so.

New

• Files from the agent — the agent can send you a file from the computer hosting the chat. It appears in the chat as a card you can download or open, and stays available for seven days.
• Python dependencies and project environments — the agent's Python runner takes extra packages or a requirements file, and pandas, matplotlib, CoolProp and Cantera are in the base environment. Declared dependencies are saved beside the script in .prism-python.json and reused on later runs. For an existing Python project, a different Python version, or an editable install, the agent can create or reuse an isolated project environment instead.
• Several data sources in one NVPOST scene — a post-processing configuration takes any number of sources, and each source's settings live in the Scene configuration instead of the menu bar. Mirror is removed.
• GovCloud builds are labelled — the GovCloud channels show an amber GOVCLOUD tag beside the wordmark and on the sign-in screen, the window title reads Qubis (Gov), and About names the channel.
• MCP server instructions — a plugin's own usage guidance, which some servers send when they connect, now reaches the agent instead of being dropped, so plugins that rely on it behave as their authors intended.

Improved

• Diff pane — the Diff pane shows the project's git changes, driven by git rather than by watching files. Edit approvals moved into chat cards, so they still appear in a folder with no repository, and the pane hides when there is nothing to diff.
• Settings — one layout for every page. General gathers Appearance, Layout, 3D viewport and Storage; Agent is split into Models, Permissions, Chat and Custom instructions; the sidebar is grouped; bundled tools moved to Local tools; and search or the command palette opens a page and scrolls to the section.
• Tool output — tool results show a bounded preview instead of a raw dump, with input collapsed and a raw view and exact copy a click away. Connector cards name the action and the arguments that matter.
• Background chats — a follow-up queued on a chat sends when that chat's turn ends, even while you are reading another chat. Before, it waited until you came back.
• Lighter on the harness — Qubis lets go of the feeds and MCP hosts of chats it is not showing and that are not running, instead of keeping every chat a session touched attached until quit.
• Windows setup — WSL installs without the shim that broke it, Podman without a machine is recognised by the solver daemon, and the Start Podman button reports progress.

Fixed

• Cmd + / Cmd − / Cmd 0 zoomed whichever viewer tab the pane tree considered focused, even with the keyboard in the chat composer or terminal. Zoom now follows keyboard focus.

Intel cloud machines, Podman and a lighter terminal

Pick Intel cloud machines where your environment offers them, set up local tools with Colima or Podman, and run terminals that no longer weigh the app down.

New

• Intel cloud machines — the cloud machine picker offers Intel machines alongside AMD and ARM wherever your environment has them, and the new-run form starts on the architecture your environment uses by default.
• Colima and Podman — local tools setup offers Colima on macOS and Podman on Linux and Windows next to Docker Desktop, so a Docker Desktop licence is no longer the only route.
• Mark as Unread — right-click a chat to mark it unread; it goes back to bold until you open it again, and the mark survives a restart.

Improved

• Terminal — terminal panes copy only the visible rows instead of the whole scrollback on every event, bursts of output redraw once, and link hover no longer scans the screen on every mouse move, so busy terminals stay light.
• Responsiveness — moving the mouse no longer rebuilds the whole window, and chat list updates at launch arrive in batches, so Qubis stays smooth with many chats open.
• Run log — a cloud run's log line names the machine's architecture, such as running on a C-48 machine (intel).
• Agent instructions — the meshing, code-CAD and Python guidance the agent reads no longer contradict each other about how to run Python and when to open a model in the viewer.

Fixed

• Windows: Qubis stayed closed after an in-app update; it now reopens once the installer finishes.
• Slack and other broker-managed MCP connections kept sending an expired token after it was rotated, so calls failed with "Auth required" while Plugins still showed the server as ready. The connection now refreshes its token on every request, and one that needs you to sign in again says so.
• Cmd-click (Ctrl-click) in the terminal could open a file:// or other non-web link printed in the output; only http, https and mailto links open now, and the link is re-checked at click time.
• Run File handed a path with ; or $(…) in its name to the shell unquoted.
• The sub-agent page showed the live reasoning text twice, under the header and in the transcript.
• A cloud environment whose default architecture had no machine your plan could use left the new-run form on an architecture nobody could pick.

Agents pane, remote chats and tab search

Follow and talk to sub-agents from their own pane, pick up a chat running on another computer, and find any open tab with Cmd+P.

New

• Agents pane — sub-agents show up as one-line rows in the chat and get a page of their own in the Agents pane, with a codename, the task they were given, the tools they ran and their output.
• Message a sub-agent — write to a sub-agent from its page, and ask the agent about one by its codename.
• Remote chats — a chat running on another of your computers is marked remote; you can read it, send to it, take it over here, and see when its computer is offline.
• Approvals on any computer — cloud job approvals and most tool approvals can be answered from any computer you're signed in on, and the first answer wins. Approvals for Qubis's own actions are still answered on the computer that asked.
• Tab search — Cmd+P (Ctrl+P on Windows and Linux) searches every open tab, listing the ones that don't fit in the tab bar first.

Improved

• Editing the queue — with the composer empty, ↑ pulls the queued prompts back into the composer as one block to rewrite and re-queue, and ↓ selects a single queued message to edit or steer on its own.
• Crowded tab bars — tabs shrink to fit before any are hidden, the rest sit behind a +N button, and hovering a tab shows its full path.
• Windows tab bar — each pane's tabs get their own row under the title bar, so more of them fit.
• STL and OBJ files — surface files show exactly what's in them, and open in post-processing alongside your results.
• Cloud job approval card — shows the machine, solver image, hourly rate and command, with Approve and Reject always in reach.
• Office Suite setup card — says whether LibreOffice or Microsoft Office is handling documents.
• Theme picker — Settings lists themes by name, the same as the command palette.
• Quitting — Qubis closes straight away instead of waiting for a running solve to finish.
• Tooltips — every button's tooltip has the same look, and Search in files shows its shortcut.

Fixed

• Long code identifiers in a chat table ran into the next column.
• Folders full of untracked files slowed down the file explorer's git status.
• A brief network blip showed a connection error; it now appears only if reconnecting keeps failing.
• The agent described itself as a coding agent.
• The pane toggle and maximize button showed on a chat with no folder, where they did nothing.
• Ferro's Add Bolt and Bolt Margins tools are gone: bolts were never actually modelled, so their results couldn't be trusted.

Steering, approval cards and editor fixes

Steer a running turn from the composer, approve tools from a cleaner card, and get a long list of code editor fixes.

New

• Steer from the composer — while the agent is working, Cmd+Enter (Ctrl+Enter on Windows and Linux) sends your message into the running turn instead of queueing it behind.
• Markdown lists in the composer — Enter continues a list and Tab indents it.
• Engineering review decks — the presentations skill builds design-review style decks with agenda and bottom-line-up-front slides, status-coloured tables and editable charts.

Improved

• Approval cards — the title and Always allow / Approve / Reject sit on one bar above the call's details, and you can fold a card away.
• HTML files — .html reports open in Qubis's browser instead of showing their markup.
• Question cards — the title and context are selectable, so you can copy them.
• Copying chat text — copying from inside a list item or quote leaves out its bullet or marker, and the selection highlight lines up with what gets copied.
• macOS installer — the DMG opens to a laid-out drag-to-install window, and Qubis has a new macOS icon.
• Windows — launching Qubis a second time focuses the window that's already open.

Fixed

• Quitting Qubis while the agent waited on a question or approval ended the turn; the card is still there when you reopen.
• Sub-agents could lose their progress cards after a reconnect.
• The code editor's language server did nothing for files with a space in their path, and hover, clicks and Cmd+hover landed on the wrong column in long files.
• Undo in the code editor went one keystroke at a time; it now undoes a word at a time, and the completion popup filters as you type.
• The Browse button for a custom runtime catalog in Settings → Solvers stopped responding once a path was set.

Schematic viewer and cloud jobs

Qubis opens KiCad and Altium schematics, and the agent can run jobs on cloud machines after you approve them.

New

• Schematic viewer — open KiCad (.kicad_sch) and Altium (.SchDoc) sheets in Qubis without either tool installed, and select parts and nets by clicking or by asking the agent.
• Cloud jobs — with Remote Compute enabled for your organization, the agent can run a command on a cloud solver image. You approve each job's command, machine and hourly rate, and each file upload and download, even in Bypass mode.

Improved

• Altium boards — the agent can read a board's stackup, tracks, vias, pours, net classes and design rules for impedance, return-path, power and thermal analysis.
• Organization picker — choosing an organization at sign-in uses a dropdown, sorted alphabetically.

Centered chat and loading

The New Chat view stays centered while Qubis refreshes conversations.

Fixed

• The New Chat composer and Qubis mark could sit far to the left on wide windows; they now stay centered while conversations load.

Plugin updates and engineering reports

Installed plugins now update from the Qubis catalog, and the agent can write engineering reports.

New

• Plugin updates — when the catalog has a newer version of a plugin you've installed, Qubis shows "Update available" and updates it in place. Whether you had the plugin turned on or off is kept.
• Engineering reports — the agent can write structural, CFD, trade study, algorithm description and technical memo reports from Qubis's report templates, and preview them in Qubis.

Improved

• Altium net labels — the agent can find net labels across a whole Altium project that miss their wire or sit badly on it, fix the ones that can't change connectivity, and check the netlist afterwards.

Fixed

• Connecting to Altium could fail with "Already connected" after a failed handshake.
• Altium tools run side by side could fail with "Adapter job … is unresolved".

Annual plan and a clearer billing page

Pay for Qubis yearly with credits that refill every month, and see your team's seats and usage at a glance.

New

• Annual plan — pay for Qubis once a year. Your credits still refill every month, and Qubis shows exactly when the next refill lands.
• Updated Deepseek Flash 4.1 — Added support for the latest Deepseek Flash checkpoint

Improved

• Billing page — a more compact seat card, and usage for every member of your organization. New seats have no monthly limit unless an admin sets one.
• Paused accounts — if your account is paused, Qubis says so as soon as you open it, in a notice that replaces the composer.
• macOS dialogs — file pickers and alerts follow your macOS light or dark setting.

Fixed

• Docs links from the agent's skills weren't clickable.
• Zoomed images could spill outside their pane on machines without GPU rendering.
• Long names in the NVPOST scene card overlapped the visibility switch.

Agent questions and background tasks

The agent can ask you questions mid-task, and everything running in the background now shows up in one place.

New

• Agent questions — the agent can stop and ask you up to four questions at once: multiple choice, yes/no or free text. The card sits above the composer, so you can pick an option, type your own answer, or skip.

Improved

• Background tasks — background commands and solver runs now show in the same card as subagents, and you can cancel each one on its own.
• Activity icons — agents and tasks that are running in the background get a slower, quieter animation, so you can tell them apart from the agent that's working right now.

Fixed

• Hiding the right pane (Cmd+L) with the explorer or search docked no longer pops them open as a floating window.

Improved Plugin UI & Reliability Fixes

Improved plugin catalog UI, introduced MATLAB and KiCad plugins, and refreshed mesh viewers.

New

• Plugin search — the Plugins page has one search box, a Plugins / Skills switch, and filters for category and status.
• MATLAB plugin — works with MATLAB and Simulink, with nothing to install inside MATLAB.
• KiCad plugin (beta) — review, check and edit KiCad boards from the agent.
• Drop preview — dragging in the file explorer shows exactly where the file will land.

Improved

• Altium plugin — the agent can run any Altium menu command (DRC, Update PCB, autorouting, outputs) and set board outlines.
• Catalog plugins — GitHub, Linear, Notion, Atlassian and Sentry now install from the catalog. If you already had them, they carry over and stay on or off as you left them.
• STL and MSH viewers — redesigned to match the STEP and NVPOST viewers.

Fixed

• New installs start in Accept Edits mode, not Bypass, so shell commands ask first.
• Onshape exports from the agent work again.
• Improved agent responses to failed tool calls.
• File search no longer shows duplicates from nested git checkouts.
• The bug report form warns you when the description is too long, instead of failing on send.
• Error on corrupted OpenFOAM case instead of crashing.

Better CAD imports and smoother conversations

Faster STEP imports, easier-to-read long prompts, and more reliable chat interactions — alongside the improvements from 0.5.1 and 0.5.2.

Improved

• STEP imports — models open faster, and files containing a mix of solids, standalone sheets and free faces retain more of their geometry. Incomplete previews show missing-face warnings instead of silently losing geometry.
• Long prompts — large messages stay in expandable input cards with Show more / Show less controls, so you can preview, expand and edit the full prompt.
• Image attachments — large images are resized before sending and rotated to the correct orientation. Oversized image decodes are rejected to limit memory use, and images that exceed the message budget are skipped with an explanation.
• Tooltips — hints scale with text zoom, and shortcut labels match the action and your operating system.

Fixed

• Editing and resending an earlier message works while a response is streaming, including just after pressing Stop.
• Returning to the chat from Settings, Plugins or Solvers puts keyboard focus back in the composer.
• A failed optional plugin connection no longer blocks sending a message while it reconnects in the background.
• Word-delete shortcuts work in multi-line text fields, including bug reports, custom instructions and the plugin skill editor.
• Very small CAD models, including sub-millimetre geometry, can be selected in the viewer.
• The Usage chart no longer leaves a model-breakdown overlay covering the bars; daily breakdowns remain available on hover, with a colour key below the chart.

Bug fixes and new plugins

New

• Microsoft 365 connectors — connect your work account to use Outlook (search and read mail, save drafts and replies, list and create calendar events), Microsoft Teams (read and post in chats and channels) and OneDrive & SharePoint (search and read files).

• Slack — connect a Slack workspace and let the agent search, read and post in your channels, and read your direct messages. Search covers the public channels you are in.

• Altium Designer (Beta) — Windows only. With Altium open, the agent can read your schematics and boards, cross-probe a part or net, pull the design out as data, and review it against datasheets before fabrication.

• Google Calendar and Gmail — list your calendars and events, find a free time across attendees and change events; and search, read and draft Gmail. Both run on Google's own servers, which are still in preview, so they may not connect for every account yet.

• Report a bug — a bug button in the threads sidebar footer.

• Models — Claude Sonnet 5.5 joins the model picker, with a 1M-token context and adaptive thinking. DeepSeek V4 Pro has been retired; DeepSeek V4.1 Flash is unchanged.

Improved

• Thread age — each sidebar row now shows how old the thread is.
• Recent projects — the workspace chip above the composer has a dropdown of recent projects for one-click switching.
• CAD Viewer — .step and .stp files open in an improved CAD Viewer tab

Fixed

• Improved stability and reliability of the new agent harness.
• Stop now takes effect on screen immediately, and a model change the backend rejected is retried instead of being dropped.
• If Qubis was force-quit, the next launch could stall for up to two minutes before your plugins were ready; they now reconnect immediately.
• If Qubis quit unexpectedly, the next launch now says so and links to the log.
• On Windows, running the prism CLI with no Qubis open spawned hundreds of copies of itself.
• Numpad keys did not work in the agent chat on Windows.
• A plugin waiting on a sign-in read as "Added"; it now shows "Not signed in", and six plugins show their proper names again.
• On a plugin's sign-in row, two Cancel buttons could appear at once.
• The agent could not copy skill templates into the workspace.
• Collapsing a conversation group did not reset its "Show more", so reopening it stayed fully expanded.
• Light themes: the composer border was too faint, and the tab drop-zone preview was invisible.

Reworked command palette and Smartsheet

A rebuilt command palette, a Smartsheet connector, and solver runs in the background-tasks list — plus fixes for the @-mention popup, the diff view, and long chats with many images.

New

• Smartsheet — read and write sheets and workspaces by connecting a Smartsheet API token. Requires a Smartsheet Business, Enterprise, or Advanced Work Management plan.
• Command palette — Cmd+K is rebuilt: filter pills for agents, commands, files, plugins and skills, and every shortcut a row shows now runs that row.
• Background solver tasks — solver runs, structural (Ferro) solves and runtime installs appear in the background-tasks pill at the bottom of the chat, with progress.

Fixed

• The @-mention popup flickered and collapsed while typing.
• Chats with many attached images stopped sending once they hit the model provider's size limit; the oldest images are now left out of the replayed request instead.
• The thread diff view saturated a CPU core on long threads.
• The built-in browser on macOS reported an outdated Safari version, which broke sign-in on some sites.

Plugin catalog and split terminals

Browse and install plugins (including Onshape) from inside Qubis, tile terminals and browsers next to your files, and pick from a wider set of models in a simpler picker.

Highlights

Plugin catalog

Browse, install and remove plugins from the redesigned Plugins page. Each plugin has its own page with a preview of its skills, and new plugins appear without a restart. New in the catalog: Onshape, GitLab, LangSmith, monday.com, Context7 and DeepWiki. You can also ask the agent to install or manage plugins for you — it asks for your approval first.

Onshape integration

Connect your Onshape account through Qubis, then let the agent browse your documents and export STEP files straight into your project. Your Onshape credentials are never stored on your machine.

Split panes for terminals and browsers

Terminals and browser tabs now split, move and resize just like file tabs, so you can keep a shell, a docs page and a model open side by side.

New

• More models — Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5, GPT-6, DeepSeek V4.1 Flash (the new default), DeepSeek V4 Pro, Kimi K3, GLM 5.3 FlashX and Grok 4.7, with provider logos in the picker.
• Model and effort picker — model and reasoning effort are now one dropdown that always shows the effort in use.
• Usage by model — Settings → Usage breaks each day's tokens down by model, with tokens, calls and cost.
• Plugin mentions — @-mentioned plugins show as inline chips with their icons, in the composer and in sent messages.
• Background commands — watch live output from running background commands in the chat panel, and cancel them from there.
• Profile menu — your avatar and name in the threads sidebar, with Account, Sign out, Docs and Settings.
• Switching computers — if your account is already connected on another machine, Qubis asks before moving the connection to this one.
• Rounded corners on Windows 11 — the Qubis window now matches native Windows 11 windows.

Improved

• Large STEP assemblies — much faster to import, and reopening them is near-instant.
• Long chats — long conversations scroll smoothly.
• Background threads — the agent keeps running solvers, browser actions and other app tools while you work in a different chat.
• Many connected plugins — the agent looks up plugin tools as it needs them, so hundreds of tools no longer crowd the conversation.
• Plugin activity — repeated calls to the same plugin collapse into one expandable row.
• @-mention popup — spans the full composer, lists enabled plugins and recent files first, and uses the same fuzzy matching as quick open.
• Composer layout — attach, permission mode and background commands on the left; model and context usage on the right.
• Attached images — show as previews in the chat instead of a list of filenames.
• Agent behavior — checks its saved or exported files before saying it's done, and gives Windows and Linux shortcuts on those platforms.
• Plugin updates — a newer bundled plugin replaces an older installed copy, and missing plugins reinstall automatically.
• Updates on macOS — explains how to fix an update that fails because Qubis is running from a read-only location.
• Windows title bar — dragging works reliably and the window buttons are the standard Windows ones.
• macOS app icon — now the same size as other apps in the Dock.
• Model lineup — older models (Claude Opus 4.8, Sonnet 4.6, Haiku 4.5, GPT-5.5, Grok 4.3) have been retired in favor of the newer versions above.

Fixed

• The built-in browser failed to start on Windows when Qubis was installed in Program Files.
• Japanese and other non-Latin text showed as boxes in the Windows editor.
• Bold links showed their asterisks as part of the link.
• Long file paths overlapped the controls in the thread diff header.
• The PDF page-number box lost focus as soon as you clicked it.
• STL/OBJ meshes were clipped in tall, narrow panes, and Reset View didn't fix it.
• Windows showed macOS (⌘) shortcut hints in Settings and the file explorer.
• The error banner's Dismiss button was hard to click and could be pushed off-screen.
• Session search showed an internal name instead of "sessions".
• The custom instructions box in Settings → Agent showed as a white box in dark themes.
• Ctrl+C didn't dismiss the agent's question card on Windows and Linux.