Changelog

What's new

Every release that changed something you'd notice, newest first. Releases that only moved things around under the hood aren't listed: there are a lot of those, and none of them are news.

Every release, including the ones that aren't here, is on GitHub with its full commit list and downloads. Read what we promise never breaks.

October 13, 20261.334.0

Standing grant section headings no longer show a item count or a description line under them.

October 12, 20261.333.1Breaking

The Agents page is removed; use This device to set up this PC and write your first task.

Set up this PC and your first task from This device instead of the Agents page.

October 12, 20261.333.0Breaking

Anything parsing GrantRevokeSchema loses additionalProperties, properties, required and type; use the current sandbox-contract grant schemas instead.

You can sign in with a one-time code sent to your email, not only with Google.
The editor shows the mist wallpaper by default until you choose another option or none.
In the desktop app, sign-in and setup now let you return to This computer when there is no workspace to go back to.
You can hear replies read aloud in chat, use hands-free voice with spoken answers, and get spoken replies in Discord voice channels.
Bring back appears when the sandbox has changes to copy home; opening a folder shows progress while its file list is read, including on large or WSL folders.
A sandbox setup started while this PC still needs Docker now waits for this PC's setup and carries on by itself once it is ready.
The context chip at the top of a chat lists the system prompt and what went with your first message; that message no longer shows a separate context paperclip.
Open needs refresh right after you grant access or save a change, even if a check was already running.
The free trial now shows how many messages you have left today, a progress bar, when it resets, and clearer notes on the model button.
The sandbox switcher shows connection status on the bottom-left and the restart spinner on the bottom-right, like other rail icons.
Resuming after a stale background notice no longer makes the agent reply twice to the same message.
The selection highlight on a session card starts moving as soon as you select it instead of sitting still for a moment first.
Selecting a session card replays the selection shimmer from the start of its sweep.
The open session card gets a soft highlight and a gentle moving shimmer so it’s easier to see which chat is active.
Rename an open persona from its heading, the same way you rename the sandbox and model setups.

October 12, 20261.332.0Breaking

Settings and API consumers must read modelSetups and roleSetups; modelDefault and modelRoles are gone and migrate when settings load.
Clients that read or write sandbox settings must set limitMoveCarryUnder explicitly; the wire contract no longer defines a default on read or write schemas.

You choose which repository to focus on from a menu in the page header instead of a list in the side column.
Agent turns finish sooner after background servers hit their limit, those servers actually stop when reported, and work can land if you cancel the job or…
Map holdout in agent behaviour settings again compares conversations that open with the project map to those that open without it.
When this machine is busy, new chats can suggest starting on another of your machines that has room, or move there automatically if you choose that in settings.
Sandbox Access lists conversation permissions with live or archived status, icons, and buttons to revoke one allowance or all for a chat.
New tabs open right away so you can type an address immediately, window runs on the tab strip show as an icon or initial, and searches from the address bar…
Pinned models without a chosen effort level now show that the model's default effort is used.
You assign each background job to a named model setup and drag jobs between setups instead of using a default list plus per-job lists.
Choosing always on a subagent spawn prompt clears other pending prompts on that agent for this turn, and messages within your agent tree no longer trigger…
If your repository checks file changes after you turned checks on, checks you already accepted keep running; only new or changed ones wait until you accept…
Plan limits groups blocked accounts under Needs action, with a Reconnect link to the Models tab for signed-out accounts.
Agent Jobs now starts from default models you can share across jobs, with separate lists for jobs you move off the default.
Choose when a message sends from the menu on the Send button; the button shows the time or agent you picked.
On Windows, open a new terminal and run intentic with a folder path (such as intentic .) to open that folder in Intentic.
Agent instructions now say which providers do not use your system prompt and keep their own.
You can show only one repository from the file tree or Project commands in the palette; the separate Projects page on the rail is gone.
When your allowance is spent, an empty composer or the continue strip can schedule the turn for when the limit resets instead of offering Continue that fails…
Agent cards show when scheduled messages will send; archiving an agent cancels those bookings; Edit puts waiting messages back in the composer.
Devin account sign-in can finish on its own when Intentic runs on your computer or in a browser that catches the redirect; otherwise paste the URL or code…
Connect a Devin account to run Devin's agent and pick models from what your Devin plan offers.
The Devices board shows each computer as a tile with its operating system icon and lets you filter by live, needs attention, or offline.
Expanded rows use a lighter highlight so the panel background and texture stay visible instead of a flat overlay.
The editor reloads once when the dev server re-bundles dependencies under an open tab, instead of breaking side panels or views.
Buttons and forms keep showing progress and errors consistently while work is running, including after you leave the screen and come back.
Runners built from the same sandbox overlay as yours are no longer marked outdated just because their image name differs.
Host connect setup keeps a stable SSH tunnel across single-use codes, and ic progress lines align with emoji and CJK in the terminal.
Chats on a remote runner now warn you before you send why the shield cannot run there, and docs-style placeholder phone numbers no longer trigger false…
Taking back your own waiting message puts its words and attachments back in the composer instead of removing them, and drops the chat from the list when that…
Notices, tool calls, and checklists in chat line up in one column, and settled prompts no longer stay highlighted like the one still waiting for your answer.
In the run dialog, pick which computer to use once at the top, then start each listed program with its own Run button.
Copy-paste install and device commands are now single-line URLs; terminal tools show clearer step-by-step progress.
An account that hit a usage limit stays unavailable until it serves again, even when another account on the same provider was refused later.
Draft cards on the agent board stay visible when you open them while chat lives in another window, instead of disappearing briefly.
On your phone, swipe between an agent's chat and their pending changes; the header shows how many changes are still to land.
In Privacy Shield activity, peek at each masked image before masking and as the provider saw it.
Unused per-model capacity rows no longer duplicate your main weekly limit when that limit already covers every model.
Chat buttons and links are easier to tap on a phone, and the composer bar lays out better on narrow screens.
Sandbox and agent settings panels reflow instead of clipping when the window or side panel is narrow.
When you schedule a message for later, you can pick a day on a calendar and enter a time instead of the browser’s date-and-time control.
On phone you can swipe rows aside to archive or take back a message, flick bottom sheets closed, swipe from the left edge to go back, and tap the active tab…
Office files you can change open ready to edit instead of opening read-only until you tap Edit.
You can now stage and review changes by scoped areas instead of just whole repositories.
New Windows setups now run sandboxes on Intentic's engine by default, even if Docker Desktop is installed.
Chat no longer shows an account as ready at the limit while the service still treats it as spent until capacity returns or the cooldown ends.
Copying sandbox volumes with only a container name no longer requires you to also pass a helper or image when that container provides one.
You can use the Intentic engine as Docker in a WSL distro and copy images, volumes, and stopped containers between Intentic and Docker Desktop.
When you edit a chat message, the notice bar now uses a card-style look with a subtle shadow.
Your first visit includes a getting-started checklist and the agents board in three lanes from the start, with hints for what to do next.
On Windows, copying files from sandboxes and other Docker work against the Intentic engine is much faster while the engine is running.
The device page names Intentic's engine when your sandboxes use it, and containers left behind after moving engines no longer appear in the sandbox list.
Your computer can put sandboxes you are not using to sleep to save memory; opening one wakes it again.
Desktop updates can install while first-run engine and image downloads pause and resume on the next launch.
Help for when to suggest a session summary is clearer, and German, Spanish, and French now show those settings.
Help text for automation clocks, idle chat warming, low-memory holds, and limit handoffs is shorter and easier to read.
Memory and process breakdown help text appears when you hover section headings instead of always below them.
Sandbox resource metrics use MiB and GiB and can show a separate memory breakdown from process totals by role or session.
When a chat stops because the model allowance ran out, you can choose to keep the full session, trim older tool output, or start fresh from a summary before…
On Windows you can move your sandboxes to Intentic's own container engine from This device so they run without Docker Desktop.

October 9, 20261.331.0

On codebase health, each hotspot row names the suggested refactor and the tab uses a clearer stats strip and table layout.
Usage limits shown at 100% no longer block your account when the provider only reports whole percents until it actually refuses a turn.
Dropping a large folder onto your workspace into a local sandbox finishes much faster when the folder has lots of small files.
Refactor actions on the Health tab now have a model picker, and refactors use their own Refactors model list in settings.
Commit-message drafts now announce when they start, are ready, or fail, and a reload no longer reports a draft that was never seen to fail.
Chat install requests update when you change tools on the Environment card, and drafts you discard in Changes are treated as declined.

October 9, 20261.330.0Breaking

The Settings toggle and the terminal.toggleWorkTerminals command for showing work terminals are gone; reveal a work terminal from its own surface instead.
Integrations that relied on whisperReady in listener status must use GET /speech/status to know when dictation is ready.

Large folder drops on the Windows desktop app copy reliably again, and interrupted copies no longer leave corrupted partial files in the workspace.
Very long chats open on the latest messages and stay smoother while you scroll or while the assistant is still replying.
After setup you land on the file tree with chat ready; Projects stays in More until you pin it, visit it, or open a project.
Push progress in the changes review area uses softer text and stacks the extra detail line only on small screens.
On Windows, very large folders dropped onto the workspace are copied into the sandbox by the app instead of blocking the page while every file is read for…
Hover highlights on rows, chips, tabs, and buttons now fade in at the same pace instead of some changing instantly.
Review panel agent filters show the active choice with a colored chip only, without a checkmark.
Selected origin filters in the changes review bar look subtler and match other on-state chips instead of appearing bolder.
Dropping very large folders into the workspace no longer freezes the browser, and import progress stays visible while files are scanned and prepared.
In the changes review panel, the origin filter you have selected is easier to spot thanks to a check mark and stronger chip styling.
On first launch you can set up this PC from Agents, use the built-in Windows container engine, queue a task while setup runs, and use Repair for local…
Help for running tests on your machine is translated in German, Spanish, and French, with shorter wording in Polish.
The help text for turning pages in chat on or off is shorter and clearer.
Docker Desktop starts in the system tray when you prepare your environment, without opening its welcome or sign-in screens first.
Dependencies opens on findings and area comparisons, a reduced package graph, and an Inside tab that stacks one package by its declared layers.
Icons in extension, environment, and account lists now show as simple gray glyphs instead of boxed badges.
Messages held because the sandbox is low on memory can send automatically once memory stays free, or you can leave them waiting until you choose Send.
Grok and Gemini coding agent turns now run on OpenCode 2.
Voice input can show download progress and keep listening in your browser until the sandbox speech model is ready.
You can optionally name a version before saving, undo all changes from one contributor at once, and save or back up from the changes panel instead of the…
The Unsent badge no longer shows draft age or preview text on hover; only the Unsent label is visible on the chip.
The privacy shield trial note now explains more briefly that requests reach Intentic and the model vendor and personal data leaves your device.
Fixed-size provider marks in agent privacy settings no longer appear inside a rounded bordered box.
While a commit message is being drafted, you see which model is writing and live progress inside the message field, with commit and sync on one bar below.
Preview and browser status chips no longer show hover cards that repeat what the chip label already says.
Finished agent cards show a closed ring on the session icon; context use appears when you hover the icon instead of on the ring.
A folder opened from the folder menu now replaces the current one without the window going blank.
Idle local models now unload as intended, freeing memory, and the app uses less CPU while it runs under load.
Uncommitted changes in the sidebar line up chevrons, file status marks, and row actions in consistent columns.
Recent projects on an empty folder are grouped by Windows or WSL, with a short path and last-used time on each row.
Switch which conversation's files you see using search and lane-grouped lists instead of a flat menu.
Pipeline job cards show the workflow that called each job above its name, with status and duration on a line beneath.
Long job names in the pipeline graph wrap onto a second line instead of being cut off with ellipsis.
You can run repository-declared programs on connected computers from the workspace or a terminal, with an optional isolated Windows Sandbox run.
Documentation now explains how to list your extension in the gallery using a repository topic, with no account or upload step.
Browser tabs from every open window appear on the tab strip, grouped by window, and you can close a tab without bringing that window to the front.
Pipeline run diagrams size each column to its job names instead of using one fixed width for every card.

October 8, 20261.329.0

Device setup can request a new code when yours expired, start pulling the sandbox image while Docker installs, and use the CLI shipped with the desktop app…
The chat panel footer no longer shows whether the sandbox is online or a Manage link.
Push and pull commit counts now appear on the Sync button instead of in separate pills beside it.
When landing is blocked, use Fix on the review header to have the agent resolve conflicts; if the agent is mid-turn, the fix waits until that turn finishes.
You can let a connected computer run programs your sandbox builds and sends, when you turn that permission on for the device.
In sandbox configuration, Models and Agent now appear above Environment and Secrets so you can set up a model first.
The pipelines screen now uses the full screen width so horizontal run stages have more room to display.

October 8, 20261.328.0

Setup restarts behave like a normal Windows restart so other apps can save work, and when the workspace is unreachable the app shows a reconnecting screen and…

October 8, 20261.327.0Breaking

GPU acceleration for local models is no longer configurable; models run on CPU only. Existing configurations with the gpu field will be automatically migrated.
Clients requiring automation models or access:GET /automations.agent must switch to the updated automation contract and routes.
leanGuidance settings are no longer accepted; remove them from configuration and use the default agent guidance instead.
Clients using front-wire must migrate to the netd wire contract to keep communicating with the sandbox.

Pipeline run graphs now match GitHub’s job columns and stacking order, including jobs the run has not reported yet.
The number of staged files now appears on the Commit button instead of in a separate line beside it.
In review changes, the commit message area is now at the top, above agent and source filter chips.
The menu for opening extra browser windows now says “More windows” instead of “Apps and desktop.”
In review, choosing a session or your changes now stages those files; Commit only records what is staged, not hidden rows still shown as unstaged.
Stage, commit, and review outgoing work from the Changes sidebar without opening a separate commit page.
Setup shows how far the Docker Desktop download has come and when it is waiting for you to answer Windows.
Setup checklist labels appear in your language when you resume installation after restarting the app.
On the commit page, tap chips to stage or unstage groups of changes; the commit includes whatever is staged in the changes list, with each repo’s history…
In Sandbox Models, each provider tile shows how many accounts or models you have in the top corner; the text under the name only appears while something is…
With Changes open, the main pane is a commit page where you pick which files to include, amend unpushed commits, and request a fresh commit message from an…
The browser’s + button now opens a web tab directly, and switching tabs keeps the selected page and keyboard focus in sync.
Browse live apps, web pages, and desktop windows together in one tabbed view.
Office documents that use common stand-in fonts now show bold and italic correctly instead of distorted or mis-wrapped text.
Open tabs in your own browser window and use it freely alongside agent browsers.
Browse the web in the agent’s browser, search from the address bar, and switch between open windows.
Cursor chats can now stay protected by masking personal data in prompts, file reads, tool output, and MCP responses.
See your browser connection, site permissions, and activity in a refreshed popup that matches the app and works offline.
See live apps, browsers, desktop windows, public ports and VPN status in the desktop status bar.
Open a thought and its run's band stays visible, giving you a direct way to reveal the calls.
See the command a permission request is waiting on before you allow it.
Run agents in isolated execution domains to better protect the host system.
See account names for paired sandboxes and identify pairings that no longer belong to your account.
Find Pipelines beside Workspace in the work band.
Find the live app alongside Browsers and the terminal, including when it’s already open.
See each sandbox update or rollback’s progress and output while it runs.
Failed model tile address updates are now reported in the console.
Existing files are kept safe during sync setup, and setup progress stays visible while the computer reconnects.
Dependency and security checks now include packages the workspace installs separately.
Keep documentation search visible while scrolling the sidebar.
Choose the exact kept build to roll back to when multiple builds share a version.
Reload or share a link to return to the same open model or persona.
Fully reverted changes no longer produce a misleading “Changes landed” announcement.
Booked retries and account moves now survive a daemon restart.
CI can reuse typechecks already completed in a sandbox, so pushes spend less time waiting for those checks.
Git pushes no longer fail with "Permission denied (publickey)" after a restart when the saved SSH public key file was stale.
When a page the agent showed throws an error in your chat, the agent is told and you can ask it to fix the page.
SSH key setup no longer fails with "Permission denied" when an older public key file exists; the stored key is now always used.
CI pipelines now cancel superseded measurement jobs, freeing runner capacity for newer pushes; test suites on the fleet draw workers from a shared memory pool…
See the held-message notice only when its transcript line is out of view, and send only the messages held by that notice.
Search results and file tiles now keep consistent spacing regardless of name length or display mode.
Context menus with submenus no longer clip the submenu when opened.
Scripted page previews now render correctly when their content measures its size as it loads.
The More menu keeps its rows in the same order when you visit a section and marks your current section.
Pages with charts or other content that size themselves on load now render at the correct width.
Set up automations from a need and watch events to trigger agent actions.
Published page links are now copied to your clipboard more reliably.
Interactive pages in chat are now included in sandbox configuration tracking and extension capabilities.
Extensions can now access their own settings and event streams through dedicated API routes without declaring permissions.
Long-running operations now properly cancel when requested, eliminating resource leaks and allowing responsive UI updates during stops and aborts.
Restore deleted sandboxes from Devices ▸ Recently deleted instead of Sandbox settings.
The quick bar now opens a scratch pad for writing messages:read the full conversation by opening the chat instead of expanding it inline.
Local model cards now show download size, memory use, and context window as short facts instead of a sentence, making it easier to compare options at a glance.
Manage connected accounts and sign-ins from each provider’s panel.
Agents can now show interactive pages in chat and collect answers directly from them.
Address bars and URL displays now use a unified AddressField component that's editable in browsers and read-only with copy/open buttons elsewhere.
AI account setup is now in Sandbox ▸ Models instead of a separate connect view, unifying model and account management.
Sign-in attempts now appear in a dedicated card at the top of the connect view with one attempt running at a time, and failed or expired sign-ins now report…
Switch between List and Calendar views to plan automations by the week, see what runs when, and click any free time to schedule a one-time wake.
Model pin list now shows the rank order explanation as a tooltip on hover instead of as a repeated caption, reducing visual clutter.
"Open with Intentic" now appears in Windows 11's native context menu for folders and documents, in addition to the classic "Show more options" menu.
The add secret button is now in the page's action bar instead of inside the secrets list, and the list only shows when it contains secrets or while adding a…
Text that appears upside down in photos is now read correctly more often, by reading it both ways and keeping the reading that is more confident overall.
Pictures with nothing on them or that repeat earlier ones now hide behind a "Show" button in the turn, keeping the strip focused on novel content.
Files in a folder can now be grouped by when they last changed instead of by kind, making it easy to find recently saved or downloaded files.
When a spent message limit holds a turn, queued messages now show when they'll be sent with the turn instead of appearing ready to send immediately.
Local windows can now switch to another folder or file from the folder's name at the top of the file tree, with quick access to recently opened places.
Names are now masked more precisely:single words are no longer masked even if they're in the name lists, only full names (first name with non-ambiguous…
Find and manage values the privacy shield should never mask, including adding several at once by pasting a list.
Clarify what personal information is protected from untrusted providers.
View which commands the safety judge evaluates without leaving the settings page.
Shield descriptions no longer say unsupported runtimes are refused.
Extensions can now declare broker rules to keep credentials out of agents’ hands.
Fenced conversations now stay within their allowed workspace areas, including during verification and job handling.
Pressing a chat control now highlights the card or panel it points to, even when it is already visible.
Explain when a sandbox is unreachable from other devices and let you expand details about missing container settings.
Sandbox updates, recovery, and workspace handling are more reliable.
Avoid false redactions of ordinary words, network addresses, and role-based contact details.
Agents now get richer workspace context and more precise code search results.
Extension rows now show clearer details, route usage, and separate update offers from update policies.
See your last disk scan after restarting, scan again for current sizes, or stop a scan in progress.
You can end a turn while a device approval is pending and resume automatically when the owner answers.
Fix responses from proxied backends hanging when compression is applied.
Learn how areas limit what people and their agents can access in a workspace.
Connect SSH keys without exposing private keys to the agent; signing is handled by the sandbox and protected by credential rules.
Conversations limited to selected workspace areas now run in isolated sandboxes, or are refused if the sandbox cannot be built.
Questions asked before a restart now return in place with the work already completed, instead of appearing twice.
Connectors can now securely broker credentials to approved hosts and endpoints.
Expand a provider failure to view its full message, links, and copyable support details.
Edits to installed extensions now point you to the source checkout so your changes can be reviewed and kept.
Schedule messages independently so sending one early no longer releases the others.
More of the interface is translated, with visible text consistently drawn from the locale files.
Typing a space at the end of a Markdown line no longer makes its block split or re-render under the caret.
Press Enter in code and HTML blocks to start a new line without leaving the block.
Personal information is better protected when agents process messages, images, and shared content.
Show loading placeholders for images and diagrams until they’re ready.
Edit Markdown with rendered pictures and HTML, and follow links with Ctrl/Cmd-click.
See each chat’s last-turn proof beside its status in the chat rail.
Find more detailed guides and comparisons for coding agents, sandboxes, and parallel-agent tools.
Privacy notices now explain before sending when a provider can’t be covered by the privacy shield.
Try extension changes in a sandbox before pushing, and see where an agent’s work lives.
Your last selected repository or category is restored when you return to a rail.
A single attached note now appears as a bare paperclip without a redundant “1” count.
Chats for agents whose work you land now close automatically in other windows.
Loading reviews and agent conversations now shows a simple status line instead of guessed content.
Notification links now open conversations directly in the chat surface where you use chat.
Explore a calmer demo workspace and see updated screenshots across the site.
Choose workspace areas with clearer role-specific access guidance when inviting or updating members.
Memory warnings now account for full swap and sustained memory pressure.
Sandboxes recover more reliably after interruptions and keep their host and runtime state in sync.
See when the sandbox desktop has open windows, and get guidance for opening a program when it’s empty.
You can attach local folders to machine sandboxes and receive landed work from agents.
Browser profiles can now sign in using a browser launched in owner mode.
Drag the chat tile off the rail to open the chat in its own window.
The chat history action is no longer shown in the chat rail.
Clearing chats now briefly shows each chat closing before the lane closes up.
Windows and WSL now avoid duplicating sandbox updates, backups, and cleanup.
Connect an Android phone to let agents interact with its screen, apps, files, and notifications.
Sandboxes with agents mid-turn are no longer restarted automatically; you can choose to restart them now or wait.
Sandbox restarts and stalled operations are handled more reliably, helping conversations and services recover without hanging.
See when the agent’s desktop tools are paused while you take control of the desktop.
Review what the privacy shield masked in each request, and see the original values behind tokens.
Privacy shield settings no longer show empty explanatory text.
Loading placeholders now gently shimmer to make in-progress requests easier to recognize.
Stopping a message before it starts no longer leaves you stuck if the server still accepted that turn:you can continue from where you left off.
In a sandbox you can open Desktop to watch the agent’s screen, use the keyboard and mouse yourself, then hand control back.
Find projects and signed-in AI accounts on this computer, then connect them to your sandbox with fewer steps.

October 5, 20261.326.0

In a folder window you can choose to work with an agent; the app sets up a sandbox on your computer while you keep working in that folder.
While a page loads, its placeholder matches how that page looked the last time you opened it in this sandbox.
Setting up on your PC now says when a restart or sign-out is required and continues from your last progress when you try again.

October 4, 20261.325.0

Right-clicking in-app navigation links no longer shows a system link menu that can open pages outside the desktop app.
The stop-command control on an agent card now explains that stopping helps the work finish and land.

October 4, 20261.324.0Breaking

Built-in connectors no longer include Komodo or its deployment alert; add the deployments extension to connect Komodo again.

On narrow agent cards, status text stays on one line and shows in full when you hover it.
If the OpenCode runtime does not exit when you stop it, it is force-stopped so a new session can start.
Hitting the sandbox control socket with curl or another HTTP client no longer shuts down the sandbox; you get a short message that the socket is not for HTTP.
Commands waiting for terminal input show on the agent card and chat panel, and you can stop them so landing is not blocked behind a prompt.
Agent card and subagent row times keep updating after a run finishes, without the whole agents board redrawing every second.
Archiving an agent closes its chat tab in every open window, including separate chat windows, instead of leaving it stuck open.
Usage meters on sandbox storage and billing animate their fill growing from empty when you open those views.
Under Settings → Appearance you can set Animations to System, On, or Off so menus and panels open instantly when you prefer less motion.
Comparison pages now include frequently asked questions about each alternative, with updated summaries checked against official sources.
The agents board and chat list feel less cramped, with clearer separation between lanes and cards.
Chat and fleet session rows now share the same rounded corners and shading that fits the panel or board they sit on.
Agent board cards use softer corners and a subtle diagonal background instead of a flat fill.
Setup resume text reflects a run that stopped early, and a registry "denied" pull is handled as login credentials rather than a network fault.
After an update downloads, your sandbox can restart on its own when nothing is running and nobody is using the editor; change that on the update card.
WhatsApp can message people by name, read older chat history when the phone is online, quote-reply and react, and remember chats after a restart.
In the sanctum theme, secondary, warning, and danger buttons now look like dark stone plaques instead of light-outlined translucent panels.
Claude plugin hook access can prompt for your approval before a run, and the chat may show live status from enabled plugins.
The composer and pinned prompt stay at the bottom when you expand or scroll long thoughts, errands, or notes in the transcript.
WhatsApp linking shows a pairing code only after the server accepts the request, and shows a clear message when linking is refused.
Long chats can drop older tool output before each model turn, and agent behaviour settings may show how much context that freed.
Project sync skips each machine’s local cache under .intentic so those folders no longer block two-way sync.
Google models in the picker follow the translator catalog; ones your account cannot call are dropped after Google refuses the request.
Connected device rows now show updated unreachable link counts after you forget unreachable links, without waiting for the agent to reconnect.
The working indicator stays at the bottom of the chat when notices appear under the assistant reply while it is still streaming.
Google chats and helpers only start when your chosen model is available on your enabled Google accounts, not just listed globally.
The popped-out chat window no longer shows your background wallpaper or frosted styling on the pinned message area.
Floating chat with a wallpaper shows a frosted band behind the pinned prompt that fades at the bottom, and scrolling text is no longer covered by an opaque…
New Google channel models show up for agent turns after the OpenCode runtime is idle, without stopping turns already running.
Popped-out chat windows now show your chosen wallpaper, with a frosted pinned prompt instead of a solid card background.
Chat turns stay in place when an agent resumes after sign-in renewal or your workspace syncs, instead of jumping or redrawing.
Agent changes are now automatically verified at the end of each turn, ensuring errors are caught before you move on.
The application now automatically attempts to reconnect to your platform account when returning to the workspace window.
In the desktop app, you can open account settings, update your profile, and sign out from a window that is showing a folder on your computer.
Letting a provider read one conversation shows shorter tooltip text that no longer explains names and emails reaching the provider.
Selected sidebar rows and session cards show a cleaner, theme-matched highlight line.
Selected sessions get a soft primary-colored edge and tint; cards no longer show a separate left bar for needing your attention:lane and status still indicate…
Subscription capacity appears in the side rail when chat is docked wide in the main window, not only in a floating panel.
Wide diagrams can be dragged horizontally to pan; diagram rendering no longer makes the page scrollbar jump.

October 3, 20261.323.0Breaking

Images are now masked locally; the "withhold" and "read" options in the privacy shield policy have been removed and replaced with automatic masking.

Install or update extensions from Discover without reloading the editor; the detail panel shows the outcome and can open Installed.
Failed agent setting saves show why; the push toggle resets when enable is refused; the desktop app no longer offers web push.
You can schedule a message for a specific time or send it after another agent's work has landed.
The desktop app shows agent status on its tray and taskbar icon and can alert you when something needs you or a turn finishes while you use another app.
Setup sandbox attach errors now follow your selected language instead of always showing English.
Updating an extension from Discover updates your existing install instead of adding a duplicate, and the update badge no longer appears when you are already…
File sync with a sandbox should use much less CPU on your device; sandbox state copies update about once a minute instead of every few seconds.
When an agent stops waiting for your answer, the alert that replaces the waiting one no longer plays a sound or vibration.
Agent cards no longer keep showing a scheduled retry after you cancel a paused agent or the retry is dropped.
You are not asked to land again when main has already changed files after your last successful land.
Orchestrator cards move to Finished after you've moved past subagent stops; Clear and retention archive those subagents together with the parent.
You can allow a provider to read the current chat unmasked from the strip above the composer without trusting it in every conversation.
On the setup page you can see which account you're signed in as and sign out; an unfinished draft sandbox from that visit is removed first.
Project synchronization now automatically handles main-line branches, ensuring your environment stays consistent with repository updates.
Custom sandbox definitions apply on first boot, and long workflow gate checks no longer time out at the edge.
Permission prompts can allow everything in this conversation (hard rules still ask each time); secret host-guard prompts can allow a secret for the whole…
Commands using shell variables to set destinations now avoid approval cards if the variable is set to a fixed address earlier in the same line.
Shadows on interface cards now correctly appear even when the card is set to a flat style.
The devices list now shows a card for each sandbox with its file and port status clearly visible at a glance.
A title now appears above the activity timeline.
The privacy shield settings are now visible in the sandbox overview.
The privacy shield now detects and masks personal data in images directly on your machine before they are sent to any provider.
You can now collapse multi-step tool calls in the transcript by clicking the indicator in the gutter.
Assistant tool runs now show as a visual band in the chat spine that grows based on activity, replacing the previous numerical count.
Failing main-line branches are now grouped into a single block with a lane showing their related pipeline runs.
The remote browser view now renders sharper text and no longer flashes when the page settles or updates.
You can now access your machines and device sync settings directly from the main navigation rail.
The fix control for failing main branches is now centralized in the repository banner, simplifying the interface for branch-wide breakages.
You can now see the status of agent updates on the Devices board and rail even after leaving the device page.
Migration of hosted machines no longer fails when a background disk snapshot is already in progress.
You can now attach files, such as screenshots, to your answers when typing in your own words.
The privacy image pack is now installed automatically when you enable image reading or name modeling in your privacy settings.
The chat panel no longer overlaps the window close buttons when positioned in the corner of the editor.
Syntax highlighting is now available for TOON files.
HTML preview warnings appear next to the Preview control with details on hover, and the copy action works when the pointer is over the rendered page.
Continue after a stopped turn no longer adds a message to the chat; the agent picks up with a short notice instead.
Applying a capability change closes the need card right away instead of leaving it open until the service catches up.
The option to show hidden technical files in the workspace tree stays at the bottom of the panel so you are less likely to turn it on by mistake while opening…
Popover menus and context menus show lighter border and divider lines.
Panels and cards are outlined with a light shadow instead of a border.
Paired devices now appear in a single-column list that matches other sandbox screens instead of a two-column card grid.
You can turn on a privacy shield so sensitive details are masked before prompts go to model providers you do not trust.
The Fix button shows that logs are being read while a fix starts, and leaving the pipelines view no longer opens the agent list when that finishes.
Each paired computer on the Devices board appears as its own card with clearer sections for environments and sandboxes.
In Agent settings, subscription accounts show the email you signed in with as the row title, and Grok via Claude Code is labeled when it appears next to a…

October 1, 20261.322.0

The model name in the chat composer stays visible on narrower windows when you pick a schedule time or persona.
Side panel actions use shorter labels like “Show beside” and “Fill middle.”
Enter with an empty composer continues a waiting turn; the scheduled-time send control no longer activates without typed text.
Agent settings for repo checks after edits now live under Tools with the heading "Checks after edits" instead of under finishing work.
Chat capacity now shows "<1% left" when a small amount of allowance remains instead of looking fully used or unavailable.
In a popped-out chat window, Past chats no longer appears on the tab rail; open history from the main chat view or the docked header instead.
On this computer, the sandbox menu no longer lists separate ways to add a project or another sandbox:use Add sandbox to set up more.
When you sign in from the desktop app, you get a clearer “continue in your browser” screen and can reopen the browser or tap Not now to cancel.
When chat is on the side, what you open beside it can use the full workspace until you toggle to show the main section again.
The side panel places opened files and previews beside the chat at full height, with a draggable divider between the two columns.
Project folders are now kept in live sync with your computer automatically, reflecting all edits and deletions in real time.
Projects whose sandbox runs in Docker on your computer can sync without an SSH tunnel, and the desktop app can ask whether to create the sandbox here or on a…
If saving spending caps on a capability tile fails, the retry message appears in your chosen language.
Maintainers can now manage wallet spending limits, device pairings, and sandbox connections for sandboxes they are invited to.
The deleted sandboxes page no longer shows how long items are kept for recovery or a count on the recently deleted list.
Connected account rows line up more evenly, and sign-in details no longer repeat your email’s default organization name.
Choosing Move after hitting a usage limit now switches you to your fallback account immediately instead of leaving the chat waiting on a Move label.
The Secrets group for your own sandbox secrets no longer shows a count next to its label.
You can now see your account and switch between sandboxes in any local window, with Alt+1–9 shortcuts to quickly pick the one you need.
In the desktop app, the place switcher on local folders lists your sandboxes and opens the workspace on the one you pick.

October 1, 20261.321.0

Drop files anywhere on a folder window to copy them into that folder.

October 1, 20261.320.0

While a sandbox setup is running, the app window title in the taskbar and app switcher shows that a setup is in progress.
You can now use `iq verify <file>` to check that the file paths, code references, and line numbers in an answer actually exist in your repository.
The computer you installed on no longer appears twice in the device list when its host connection never reported which machine it is.
Docker and local model servers start again after a sandbox restart instead of staying unavailable.
When a model's safeguards stop a turn, you can retry on the same model without the blocked reply or continue on a different model.
With a secret’s host limit on, curl to an allowed site piped into jq or chained with && can run without an approval card.
While a checkout rebuild is in progress on the sandbox update card, the progress bar uses the full width and “Having trouble?” moves to the bottom of that…
When an agent works on its own branch, commits there are copied onto the conversation branch at the end of the turn so Land can take them; review shows when…
Opening the sandbox update card starts the download on your machine and shows progress while the sandbox keeps running.
Chat spine row markers stay faint in the margin until you hover, open a run, or a step fails.
On wide chat panels, turn icons and spine markers sit more clearly in the side gutter instead of crowding the message rows.
Attached context and agent activity use the same left-margin markers, and the vertical line down each turn is gone.
Expanded disclosure rows no longer close when you click the empty column beside the content; use the chevron or row header instead.
The app opens straight into your folder in the editor; sandboxes and setup live under This device in the sidebar instead of a separate home screen.
Hidden tool steps now appear as a count on the left of each reply; hover for a breakdown by type or click to expand thinking and steps.
The sandbox update card leads with what's new and one update action, with optional steps and developer notes available when you open them.
Agent cards for rate-limited turns can show Send again when you can retry, clearer “back in” and resend timing, and where a booked move is headed.
UTF-16 files such as desktop.ini open as text instead of binary and stay read-only so saving does not overwrite them in the wrong encoding.
Thinking and collapsed-run controls in the chat sidebar look like simple icons instead of bordered chips.
Plan limits show as separate provider cards, with overall capacity and alerts kept visible beside them on wider panels.
On desktop, rows in many settings groups now share the same height, and section spacing is consistent across sandbox and settings pages.
Subagent trays under agent cards and in chat open and close more smoothly, and rows below slide to make room instead of jumping.
Needs you brings together agent requests, paused chats, and extension approvals in one inbox, with quick actions for many items without opening each section.

September 30, 20261.319.0Breaking

Sidecars are no longer supported; any configuration or tooling relying on them for workspace file status or caching must be removed or updated.

When your sandbox will not connect, recovery shows what failed and lets you run a fix from the desktop app.
Approve or reject a waiting plan from the bar above the composer; those buttons no longer appear on the plan card in the transcript.
Lists and detail panels no longer use a vertical line on the left; hierarchy and status show through indentation, tint, and icons instead.
Desktop sign-in completes successfully again instead of showing an incomplete sign-in link.
Selected session cards no longer show a colored outline; selection is shown by the card background only.
Selected session cards no longer show an extra inset ring or colored shadow; selection is shown by the border and background only.
Collapsing an empty child tray on an agent card no longer leaves a thin line visible under the card.
Trays under agent cards open and close smoothly when you move between cards instead of animating partway or stuttering.
When an account has no quota left, your message stays queued until quota resets instead of sending immediately and showing a usage limit error.
Subagent rows under an agent card stay folded until you focus that card; other cards show how many subagents they started instead of listing them all.
Savings and holdout comparisons in agent settings now show as a table with matching outcome columns and translated group labels.
The pipeline run side panel shows run details on one compact row so the workflow graph has more room.

September 30, 20261.318.0Breaking

GET /system/subagents/{id}/transcript and SubagentIdParamSchema are gone; use fleet roster and agent children instead of fetching subagent transcripts.
Stop parsing FindingSchema, push-checks APIs, pre-push-fix runs, and CI red findings or resolved/dismissed decisions; align to the updated contract lock.

You can rename or remove a persona from its tile on the personas page without opening its settings first.
Persona settings use square tiles; changes to the open persona save when you select a different one.
The API tokens section in settings no longer shows a count next to the heading.
Attention, Active, and Finished lane headers no longer show the small coloured status dots beside their titles.
Agent runs stop staying blocked for days after your API quota has already reset on the provider side.
The approvals page can use the same appearance wallpaper as workflows, automations, projects, and pipelines.
The wallpaper you pick in Appearance now shows on Pipelines pages, same as Workflows, Automations, and Projects.
The sandbox switcher no longer shows a subtitle under Add another sandbox about separate files and accounts.
More audio and video formats open in the media player, and files from a floating chat window preview with the built-in viewers instead of showing preview…
The sandbox can periodically bring in what others pushed to your tracked branches without you pulling, and you can turn that off under finished work settings.
The safety judge settings screen no longer shows the extra paragraphs that explained off, watch, and on behavior.
Choosing Automatic for project installs no longer shows an extra explanation under the option.
Persona avatars share one clay companion; body color stays with the persona and the item it holds reflects its name or role.
Hover View prompt to see which prompts are replaced and which get the instructions added.
Hover Restart sandbox on the version card to see what a quick restart does; the extra row saying the rest of the page still matches your sandbox is no longer…
When your sandbox is behind on some features, what still works is explained more clearly and shown like the other comparison rows.
The commit for the file you're reviewing shows on the diff toolbar; tap or hover it to see all related commits and open the graph.
In agent review, committed work is summarized in a pill in the list header; open it to browse commits.
Agent review commit history opens in a floating list that is not clipped by the panel and can sit beside or below the line on small screens.
Child rows in a tray line up model, effort, and time in columns so you can scan them more easily.
Subagent type and provider appear when you hover the child title on the board row and chat bar, not beside the title.
On the agent board, wider child rows show model and tier on the same line as the title instead of underneath it.
Switching accounts while a turn is waiting on a question now sticks, so your next message and Continue run on the account you picked.
On desktop, double-clicking an agent card opens that agent's changes view.
Agents running while you are away can leave permission requests open for you to answer instead of failing automatically.
Booked usage-limit retries now leave Attention while still showing when they’ll resume.
Chat stays responsive while syntax highlighting and long transcripts load on mobile.
Stopped turns no longer stay on Stopping forever, and messages work again after the local agent server has crashed.
The Credits page and its link in the site footer have been removed.
Each persona keeps the same illustrated face when you rename it, with new companion artwork in chat and lists.
Agents can install project dependencies during a turn when you allow it, and review lists new packages before you land their changes.
Wallpaper backgrounds keep their natural color instead of looking grey and washed out.
A background you pick in Appearance now also shows behind Workflows, Automations, and Projects while you scroll those pages.
On a phone, Settings and Sandbox open as section lists you drill into, row actions live behind ⋯, and primary actions like new agent or upload use a floating…
In Appearance you can pick a background for the agents board, including Mist that follows your light or dark theme.
On the agents board, each child row shows its kind next to the title instead of on a separate line under it.
When a recipe is waiting, rebuild from checkout stays on the Sandbox tab only; the environment card now directs you there instead of showing the same action…
Anyone's tile can show a small cluster of agent faces when agents are listed, and the new chat button no longer changes width when you switch agents.
On the agent board, child task rows only show a type label when you picked something specific, not Claude’s generic default.
Popped-out chat windows can show linked files and app previews in a panel beside the conversation.
Files and app previews can open in a side panel beside the main view, and extensions can add their own side tabs.
The bottom of the chat sidebar now uses one matching pill to start a new agent or chat, including the agent’s picture when you’re viewing one agent’s chats.
Personas sit in a grid above your chats; pick one to narrow the list instead of switching to a separate Personas view.
When a specific persona is selected, starting a new chat now automatically creates it as that persona.
In chat, personas appear as a grid of face tiles; use Clear above a persona’s chats to close their finished chats at once.
You can now restrict which hosts an agent is permitted to send secrets to.
Browser tab notification sounds are softer wooden chimes instead of simple electronic tones.
Subagent tray rows and the subagent bar show the model and thinking tier at a glance, including what was actually served for in-process runs.
Child rows and subagent chat bars show which model and effort tier each child uses; hover the label for full details.
After you release a held message, it no longer keeps showing in the waiting queue under the turn that is running it.
The browser tab can show when agents need you, when the sandbox is offline, and when a turn finished while you were away, with optional sounds under Settings…
The sandbox update card no longer leaves extra empty space at the bottom when “Having trouble?” is already shown next to Rebuild.
Scrolling up in a chat no longer snaps the transcript back to the bottom when new messages or layout updates arrive.
Capacity usage bars now ease to new levels after a refresh, briefly showing what was just used instead of jumping in one step.
Manage local projects, browse and edit files, and bring agent work back into the desktop app.
Deleting your account and signing in again with the same Google account no longer resets your hosted standing, abuse strikes, or free hour usage.
Improved contextual help tooltips across the interface for clearer guidance.
Collapsing expanded finished work on the agents board keeps the finished section in view instead of leaving you scrolled far down the page.
Agents waiting on a usage limit stay in Attention with a clear resend time instead of looking active, unless the turn is already moving to another account.
Switching accounts while waiting on a usage limit updates the countdown to the account you picked, and the retry runs there.
In the chat sidebar’s Personas view, personas stay in a fixed list; choose one to see only that persona’s chats instead of expanding and collapsing groups.
When you open a folder on your computer, sandbox-only actions are hidden or explained instead of failing without context.
Account sign-in that ends on a page on your computer can finish by itself when the desktop app or browser extension is connected, without pasting a link.
On checkout-built sandboxes, “Having trouble?” now sits next to rebuild instead of alone at the bottom of the card.
Rebuild from checkout on the Environment tab only shows when a recipe is waiting; use Update on the Sandbox tab to pick up new code otherwise.
Held status messages no longer show a dotted underline when more detail is available on hover.
You can open a folder on your computer in the desktop app to edit it there, or keep a local folder in sync with your workspace during setup.
You can open an in-process subagent's transcript inside its parent's chat from the tray or the tool card.
Paid hosted sandboxes get their own monthly awake hours; free try hours are shared across your other hosted machines.
The agent seal hover card no longer shows the footnote about where run data comes from and CI after push.
Hovering an agent card’s proof seal opens a detail panel with the full check command and clearer status text instead of a clipped tooltip.
The Claude model picker refreshes with your account’s full catalog after you sign in, even if you opened it before connecting.
After you connect any model, the free Google option no longer leads the list or opens automatically on the connect screen.
Dragging to select text in a rename field no longer changes file selection when you release over another row or empty space.
Pasting a copied folder from your file manager now attaches or uploads its files instead of an empty file named after the folder.
PDF-derived text drops repeating headers, footers and page numbers; highlighted and editor-style code blocks keep every line when converted to markdown.
Right-click a file in Changes to open its diff or file, reveal it in the explorer, stage or discard, or copy paths.
Spawned subagents apply their finished edits to your checkout; use merge when their work was held back because it clashed with yours.
When your allowance is used up, you can write now and the message sends when quota reopens, or send sooner via another account or a weekly reset.
Sandbox logos in the switcher and menus show as a plain picture or initial without a bordered tile.
Subagents no longer have their own chat list or page; see started, spawned, and running ones on each agent’s card on the fleet board.
The account checkmark in chat now matches which account sends your next message, and reconnect prompts only appear for that account.
Pushing no longer shows leftover check debt or runs a pre-push fix agent; use local verify and CI to catch failures.

September 28, 20261.317.0Breaking

The capability offer API has been removed; agents must now use the new needs-based interaction flow to request and manage permissions or credentials.

You can now roll back to previous sandbox versions or recover healthy state from the version overview card if an update fails.
Sandbox storage shows how much space you can free and lists the largest categories first, with smaller ones behind show more.
You can edit office files in ONLYOFFICE in your browser after a one-time download, or choose the Docker document server under Editor engine in extension…
Version drift can flag when agent connection requests differ, and API docs list those routes under Needs.
File syncing now works reliably on Windows PCs even without Developer Mode enabled.
Agents can now ask for the specific environment or credentials they need to complete tasks directly in the chat interface.

September 28, 20261.316.0Breaking

Stop parsing Mainline schemas and offload landCheck; read push-checks state and the updated workspace contract instead.

The mainline board and land-check agent setting are removed; use Pipelines and push checks to see what is left before you push.
You can download files, folders, or a multi-selection from the workspace; folders and several items at once save as one ZIP.
Nested usage limits now show as on hold until the parent allowance resets instead of looking ready to use.
The device agent waits up to five minutes before rebooting a distro that WSL shut down, and the CI fleet skips passes while WSL disk maintenance is running.
On a device page, file sync and port mirroring each show one clear status for all sandboxes, and bulk port controls are labeled turn all on or off.
Isolated Cursor conversations now run the agent in your worktree so edits and shell commands target the agent branch.
Stop a subagent or a scheduled retry with agents cancel; wait returns the child's full report and treats paused retries differently from finished work.
Chat warns when your sandbox must be updated before you can switch accounts on a turn held for quota.
When the sandbox holds your message, it appears at the bottom of the chat as not sent, with send anyway or raise memory options.

September 27, 20261.315.0

Hosted sandboxes are reached the same way as sandboxes on your own machine, through a connection to the edge, not through Fly’s public routing to the machine.
Folder cover on the home screen no longer shows a file name, size, and Open button above the document.
Sandbox environment actions appear at the top of the card, and dev checkouts show one rebuild option instead of two similar ones.
Delegated work from Codex, Cursor, and OpenCode runs now appears on the subagent roster, nested under the tool call that started it.
Main line explains each column under its title, and pre-push findings appear in their own column instead of mixed with land check results.
Chat capacity and plan usage now show shorter limits nested under the allowances they draw from, and dim nested room when the outer limit is used up.
Open Main line from the rail for a full check board; Google accounts that need verification show how to fix them, and chat no longer repeats the same failure…
When you show the same file in every folder, use the file explorer to move between folders; the home area shows that folder’s copy of the file.
Picking a different account in the model list only switches billing; a stopped or limit-held turn stays waiting until you choose to continue.
Subagents started in a chat show as subagent cards that stay up to date after the turn ends, and spawned ones can open their own conversation.
Open the full main-line land check from the rail tile or the board status bar instead of a panel under the board.
From workspace home you can pick a file name and read that file in each folder instead of the tile view.
Document outline headings show readable titles without markdown source markers.
The document outline no longer shows how many headings it contains.
Closing a popped-out panel window no longer leaves the main window waiting several seconds for the panel to reappear.
Warnings and errors now use the same circular exclamation icon in badges, notifications, the file editor, and subagent status instead of mixed triangle or X…
Warning badges on view chips show a clearer exclamation mark without a duplicate shape around it.
Warning and status icons in the editor stay clearer at the smallest sizes.
When a helper agent needs you, its parent's card moves to Attention; stopped or failed helpers fold into summary rows under that parent.
Pipeline run titles stay on one line instead of wrapping.
Stopping a turn now updates the chat panel's status immediately, matching the board's card.
Floating panels no longer appear in two windows at once when you open or restore a panel window.
The agents badge counts only the Attention lane, not unread updates on finished work.
The chat account picker no longer shows a link to view usage and headroom.
Environment contents refresh right away after you approve or edit a proposal, without leaving and reopening the tab.
The highlight outline on agent cards is no longer cut off in the expanded agents list.
The repair-after-landing setting now explains that failed checks after landing send a fix agent, including when main CI keeps failing.
The help text for repairing failures after work lands is now a brief summary instead of a long description.
Chat usage bars now show longer limits like weekly above shorter session limits so remaining capacity is easier to read.
Paste an agent id in the board or chat filter to find that agent, scroll to it, and press Enter to open it.
Conversations with a message still in the composer are no longer removed from the board by the idle archive sweep.

September 26, 20261.314.0

While supervising a spawned agent, wait returns early when something is said into your turn, and you hear when its work lands or when land or check failures…
Push rows hide the branch when it matches your repo’s usual one, and an outdated sandbox is told to update to see what a push left.
Sandboxes now detect when they need an update and offer guidance on how to update them.
When an agent is blocked by your uncommitted changes, you can now open the Changes panel directly from the agent card to resolve the conflict.
Device error reporting now identifies if your local agent installation is outdated, and sandbox peer identity migrations are more robust.
Sandbox live metrics show machine load as busy cores with a trend, and group smaller memory kinds behind an expandable line.
The chat capacity strip uses a compact warning when readings are stale, and shows remaining plan capacity as usage left.
Allowance meters now shift from green toward amber and red as you use more of your limit, not only when you hit the warning level.

September 26, 20261.313.0Breaking

Clients must use MainlineStatus reds and pushed instead of pushes, and Finding without kind, state, check, settledAt, or measuredAt on pushes.
Agent and archive payloads no longer include keepWarm.auto, promptCache.keepable, or ended reasons limited, rewrote, midnight, or moved; read keepableUntil…
Tools that still parse the removed front-wire or tunnel wire schemas must upgrade to the current sandbox-contract manifests; the old tunnel envelope framing…
The fix-dependency-breakage automation is removed; land-check routing handles red installs instead of waking that chore.
Parsers that read per-capability mcp from extension wire payloads must use manifest contributes.tools for agent tool servers instead.
`pnpm verify:turn` is removed; run scoped tests and typecheck on what you changed until post-land verify runs.
Extension and connector processes must use INTENTIC_EXTENSION_TOKEN instead of INTENTIC_PANEL_TOKEN to reach the daemon.

When you approve starting a child agent, you can review its task and change the model, account, or effort before it starts.
The chat side panel names remaining allowance once at the top and shows each limit as a bar with a percentage beside it.
Agents started by another now appear under that agent’s card on the board, and you can archive or restore them together with the parent.
Choosing an account by hand sends your next message on that account even when the chat already showed it, and seats your org turned off are rechecked…
Plan and chat usage now shows how much allowance is left and meters drain as you use it, instead of showing percent used.
Waking a hosted sandbox while a restart or rebuild is still running may ask you to try again in a moment.
Loopback and edge TLS certificates can complete even when your server cannot reach the zone's nameservers over the network.
You can set aside push check findings you do not plan to fix, and leftover push issues show with other reds on the main line.
On mobile, renaming a file in the workspace tree confirms the new name once the rename finishes.
Complexity warnings now distinguish score increases from harmless function edits.
Logs for sandbox containers are now accessible through the CLI.
Hosted sandboxes now automatically verify that they can upgrade their stored data before they start on a new version.
Chats on an account that lost Claude Code access or can no longer serve are moved to an available account instead of failing on every message.
You can now restrict which extensions a persona has access to in its persona powers settings.
You can now use the `keep` tool to leave background commands running after a turn ends, and land commits now automatically use structured subjects derived…
Keep-warm controls sit in one settings group, and the app shows simpler messages when a held cache stops warming.
Hosted sandboxes that lack a tunnel connection get the correct edge settings the next time they wake, without using Restart.
An agent that will wake on its own stays shown as still active even when its card lists no watch.
Automated messages in chat now show as from the sandbox with a short reason, instead of looking like you sent them.
Batch actions on several sandboxes now show live progress (for example, "Stopping 1 of 2") until they finish.
If your machine id file exists but cannot be read, the app now reports that error instead of silently creating a new machine identity.
File and home previews use Quick Look wording instead of Peek.
Main-line CI repair remembers its red streak across sandbox restarts.
The capability editor previews extension-provided agent tools for a card, and extension backends restart only when server code or enabled backends change.
Heavy builds and tests share the queue based on the program that actually runs, not only the exact bash line the agent typed.
On a device page you can tick several sandboxes and start, stop, restart, update, or remove them together from the bar above the list.
Your PC and its WSL distros are treated as one machine in the device list, and renaming a device moves every OS install on that PC together.
Signing in again as the same person keeps your existing account and chats instead of creating a new one.
Agent cards no longer show a separate red failure line; details are in the corner seal tooltip, and on the board you can click the seal to open the fix…
Agent cards and chat tabs show check status as one corner seal; hover for details, and only red failures stay spelled out on the card.
Continuing after a usage limit on another connected account now moves the chat and resumes in one step.
The agent panel memory gauge now warns at the same memory level the sandbox uses when it holds your turn.
The mainline panel shows which lands a red check blames and lists each failure with its name and file path.
The editor can open a terminal over WebTransport when the sandbox edge supports it, instead of always using a separate WebSocket connection.
Persona cards from older releases keep which devices they may use after you update.
Sandbox start, stop, restart, and resource limits on your device now run through ic, and you can save a new resource share for the next restart without…
Stored files this version cannot read are reported and preserved instead of being overwritten or replaced with new defaults.
Chat and connection screens now say whether to sign in again, ask an admin, or wait when an account cannot run a turn.
A daemon restart no longer drops lands waiting for the main-line check or loses hold state for a failing check.
Sandbox updates now show every file conversion the new version will run on the update card, not only ones the running daemon had already opened.

September 25, 20261.312.0Breaking

Integrations that relied on the assign `to` field matching the removed wire pattern must validate assignee targets themselves; the API no longer applies that…

Sandbox API documentation now includes an Offloaded work section for running heavy commands on remote runners.
When your agent needs a site, the extension can open its panel with the reason; you can decline for 15 minutes, and recent actions show as plain sentences.
In plan mode, the assistant can save its plan to the plan file instead of having those writes blocked.
You can send heavy agent commands and post-land checks to a runner on one of your machines instead of running them in the sandbox.
Browser sign-in profiles open again after rebuilding the sandbox instead of staying blocked as already in use on another machine.
After you push, any issues the push check still reports stay on the Main line until a later check clears them, you dismiss them, or you assign an agent to fix…
The Agents board status bar shows whether main passes and what is checking; opening Main line shows queue, check, result, and history in columns.
Extension capabilities can declare an MCP backend path so each granted card gets its own MCP server endpoint.
Hover over screenshot tiles at the end of a turn to see a larger preview without opening the viewer.
Removing or renaming a device now also revokes or renames its paired WSL and other OS installs, so old installs cannot reconnect on their own.
The agents status bar now shows one expanded panel at a time; tapping another section switches to it instead of showing two panels side by side.
Geek metrics can only be turned off under Settings ▸ Appearance, not from the agents board.
Main line status shows only on the board, not at the bottom of the chat sidebar.
Main-line status and sandbox metrics now sit in a bottom status dock you can expand into resizable panels on the board and in chat.
Chat lane headers no longer show how many conversations are in each section.
Reconnecting Claude no longer leaves a duplicate revoked entry in your account list.
You can bring back deleted files and folders from the notification or with Undo for up to a day.
Sandbox CPU, memory, and disk now appear in a slim bar at the bottom of the Agents board; open it for full metrics or hide them from there.
Accepting a sandbox update migrates your saved settings and workspace data to the new version instead of leaving them unreadable.
After a daemon restart, continuing a Cursor chat clears a stuck run instead of dropping the session.
Reconnecting Claude with the same email and organization keeps your existing account and chats instead of adding another row.
The agents board and chat tabs show mainline integration and land-check status for your agents.
Continuing a chat no longer switches it to another account unless you picked that account in the composer.
Stuck usage meters no longer drag down the age shown for other accounts, and accounts that can't be re-read are called out with the provider's message.
Leaving an Office file and opening it again brings the same editor back right away; after the first setup, the document server starts in seconds instead of…
After a reload, the terminal connects right away instead of waiting on a QUIC handshake that may not complete.
Sandbox terminals should stay connected and feel more responsive when your network drops or slows down briefly.
Save for next restart only appears on devices whose agent can defer the change; older agents no longer restart the sandbox when you meant to save.
When a turn ends, servers the agent left running for you stay up until you stop them from chat or preview.
Switching accounts or resuming in a fresh session no longer duplicates recent chat at the end of the transcript or misplaces restart notices in your message.

September 24, 20261.311.0Breaking

Stop expecting AgentArchiveSchema on sandbox agent archive responses; update clients to the current archive API without that schema.
intentic.dev/maker no longer works; use intentic.dev/desk for the same light product page and installer handoff.

You can save sandbox memory and CPU changes for the next restart without stopping the sandbox now.
When you add a file to a folder that groups siblings under package.json, the explorer expands that group so the new file stays visible.
While editing markdown, raw HTML blocks appear as syntax-highlighted code on every line instead of plain unformatted text.
Repository check settings now show running, off, or paused at a glance, with each command grouped by when it runs.
Closed Mermaid code blocks in the markdown editor show the diagram until you click to edit the source.
Chats can stay pinned under a persona on the rail, and a closed archived chat no longer reopens when a new message arrives on its thread.
While Docker starts, the device window shows how long you've waited and when it will give up, and it no longer hangs when Docker isn't responding.
After a deploy or while the dev server refreshes dependencies, the editor reloads once instead of leaving diagrams or syntax highlighting broken.
File diffs from a repository’s history or file list open in the side pane so the list stays on screen.
The file picker and chat composer respond faster as you type, and the agent board flickers less when status timers update.
If your organization disabled subscription access on an account, you'll see ask an org admin for a seat instead of being told to sign in again.
Environment blocks that install apt packages whose name is not a command now show as active when that package is installed.
You can keep an idle agent's prompt cache warm from the session menu or chat until you return, so the next reply stays cheaper.
Browser find-in-page now opens collapsed chat notes and prompts when the match is inside them.
If the sandbox restarts mid-turn, sending another message continues from your original request and progress, not only the latest background update.
The edition for people who do not write code is now at intentic.dev/desk instead of /maker.
Small spreadsheet and table file comparisons load immediately instead of showing a comparing state first.
Large spreadsheet and CSV diffs no longer freeze the editor while the two versions are compared.
Spawned agents wait when memory is low; a child killed for memory can be continued with send instead of failing outright.
Pushes open their own terminal instead of waiting behind every conversation's checks, and end-of-turn checks no longer time out while queued.
You can open a terminal in a folder from the workspace file tree context menu.
Reloading a popped-out panel window no longer briefly shows that panel back in the main window.
The floating chat bar stays open when you switch sidebar views and only closes when you click the main workspace area.
Agents waiting on a watch or unfinished background job no longer show as ready to land until that wake completes.
Empty trial chats now ask whether you already pay for AI or run a model locally, with a link to connect and lift the daily cap.
The floating chat bar stays open while you select text in the preview, and you can minimize it or open the full chat from the top buttons.

September 24, 20261.310.0

Draft conversations the daemon never filed no longer show a loading state while hydrate runs.
Background commands that never actually start are closed when the turn ends or you restart, instead of staying stuck as running.
The workspace home folder grid stays smoother while you drag panes and no longer flashes a horizontal scrollbar when you switch folders.

September 23, 20261.309.0Breaking

The wire contract no longer offers what it did under StopTurnSchema. Anything reading those surfaces must stop relying on them
Parsers of agent events, attach frames, or turn facts must stop using removed wire variants and read retries on breaks and errors instead of nested outage…

In agent Instructions you can turn on short guidance and keep part of conversations on the long form to compare how they perform.
Chat search, code-change stats, and large lists should load faster and the sandbox should stay responsive while those run.
Chat, settings, and usage screens show readable model titles instead of raw model id strings.
The workspace can show when installed dependency versions are behind the lockfile, not only when packages are missing.
Sync can replay unpushed commits onto upstream when histories diverge; conflicts leave your branch unchanged.
Creating or reconnecting a sandbox redeems setup codes on the platform that issued them, so codes from dev or staging no longer fail as invalid.
If your browser extension isn't connected, the assistant now tells you to pair it again instead of acting as if it can use that browser.
Restore recently deleted sandboxes from Sandbox settings under Recently deleted instead of the sandbox menu.

September 23, 20261.308.1

Docker Desktop’s WSL distros no longer show up as environments you can connect to on Windows.

September 23, 20261.308.0

Hosted sandbox awake hours are counted correctly when you reconnect, delete a workspace, or several actions happen at the same time.
The Projects view refreshes when you add or remove a repository, and finished background commands show on the job card right away instead of after a delay.
If landing your agent’s work breaks checks on the main branch, the app can open a follow-up turn to help fix it.
We have reorganized the project structure to improve maintainability and performance, resulting in a cleaner and faster developer experience.

September 23, 20261.307.0

Activity, engine installs, and enrolled hosts in Connect update as soon as the sandbox changes instead of on a fixed refresh timer.

September 23, 20261.306.0

Extension listing security reviews can run for a long time without the connection dying from idle timeouts.

September 23, 20261.305.0Breaking

Device list and report JSON no longer use gap no-agent; treat unreadable agent state with unreported or the other gap values instead.
Bundles and APIs that still send dependencyFreshness or savings.dependencies will fail validation; delete those fields from settings and savings payloads.
Panel list JSON no longer has a vitest field on each panel; use tests instead when deciding which repos expose runnable suites.
Use agent.routeChat and ChatRouteSchema instead of routeModel, POST /personas/route, ModelRouteSchema, and PersonaRouteAskSchema; noticeWait is chatRoute only.
Sending the dev-reload device command no longer works; use dev-restart instead to rebuild and restart the sandbox from your checkout.
Manifests and APIs no longer accept the host capability, HostConfigSchema, or the removed credential gate schemas; use device capabilities instead.
Stop parsing `card` on capability offers, recommendations, attach frames, and agent or transcript events; use the request-based offer and attach payloads…

Searches that include an underscore or percent sign now match those characters exactly instead of acting like wildcards.
When a message retry is far in the future, the chat shows the scheduled date and time instead of a long hour countdown.
Commands the agent runs in the background now appear in the chat with live status, and you can see how many are still running and open the terminal to watch…
The marketing site now includes comparison pages for Devin, GitHub Copilot, and OpenAI Codex alongside the existing comparisons.
Upgrading from a device in the sandbox updates every agent on that computer, including WSL distros registered for your work.
When a fix or other message the sandbox started is blocked before it runs, Send again reruns that task with its original prompt.
After an agent lands work, you can review leftover scratch files in agent review and choose to add them to your tree or remove them.
Turn on geek metrics under Settings ▸ Appearance to see live sandbox and per-agent CPU and memory on the Agents board.
Runner setup in the hub can keep streaming on slow installs, and failed or interrupted runs show clearer messages instead of a generic error.
You can install or update extensions that haven't passed the registry security audit after you confirm you understand the code wasn't checked.
Agent cards and reactions update right away when you act on the board instead of waiting for the server to catch up.
You can set sandbox memory above the previous automatic limit; only the computer’s actual RAM caps what the sandbox can use.
The device detail panel no longer shows an empty block when there are no sandbox groups listed.
Intentic’s default instructions now follow your installed Claude Code preset for the model you use, with harness-only sections removed.
Chat screenshots load faster; you can view them at actual size and download the original file instead of the preview.
Screenshots and images from agent tools appear at the end of each turn, and you can open them larger and step through every picture in the chat.
You can hide system prompts in chat, and the agent can wait on background jobs and child agents instead of polling on a timer.
The dev sandbox rebuild panel no longer shows extra lines about which device should run the rebuild; the copyable command is still there.
Show clearer translated status messages and recover more reliably when sandbox or model services are unavailable.
Opening the folder picker no longer re-fetches the workspace tree if it's already loaded in the explorer.
When the sandbox is low on memory, your message is held with a warning; you can send again to run it anyway or raise the memory limit.
Sandbox no longer checks package versions against registries or shows dependency savings in agent settings.
The Apps panel can discover and run test suites in repos that use Bun as well as Vitest.
Drop sandbox links that stopped answering from Devices; brief sandbox outages no longer force pairing your device again.
The assistant prefers your connected browser for signed-in pages and still lists it when your computer or the app was offline; on Windows it opens an…
When your sandbox is behind the app, the overview now says Restart sandbox instead of Reload sandbox so it is not confused with refreshing the page.
Inviting someone no longer shows an empty areas chooser when the sandbox has no named areas.
Sandbox browser sessions are less likely to fail permanently when a virtual display is still running but its socket file was removed.
Agent review warns when a checkout was left on another branch, and isolated agents can switch branches in their own copy without an extra approval prompt.
Full-repo verification skips instead of starting when it cannot get a queue slot before the wait deadline.
Sandbox sync warnings now name the app and daemon, show affected product areas, and explain that reload is faster than rebuilding the whole image.
Rebuilding from checkout shows named steps and layer progress; the full build log stays hidden until you choose to show it.
Chat titles in the session list now show the conversation name from the first message instead of the full raw prompt text.
Land tells you when there was nothing new to apply, and still applies the conversation’s work when the checkout isn’t on the agent branch.
Ctrl- and Shift-click open links in the desktop app, stale updates recover when opening terminals or models, and composer effort bars are easier to click.
Watch and dev-server commands no longer wait in the heavy queue, and jobs that hang stop blocking other work after a time limit.
Pasted text with curl-style @ paths or other non-workspace @ tokens no longer blocks sending your message.
When a turn is refused for low sandbox memory, you can raise the memory cap from the chat notice and change resources from the sandbox overview.
Tools with fixed-length array arguments work on local models instead of breaking every tool on that turn.
On a device page you can select several sandboxes and remove them from that machine at once; git clone folders are no longer treated as build output you can…
The sandboxes heading on a device page no longer shows a number beside the label.
Credential setup hints and secret storage lines keep correct spacing around highlighted terms and generated-secret wording.
Commit-message helpers on small models get a trimmed diff instead of failing, and automated runs that hit an error no longer say your message is waiting to…
When Intentic isn't reachable, the screen keeps checking and takes you back automatically when it's up, including after you reload the page.
Dismissing a push or pre-push result keeps the pull and push commit counts visible next to that summary instead of hiding them.
Connected machines and capability setup in the app are labeled as devices instead of hosts.
Agent turns wait when memory is tight including swapped pages, and long-unused Office and browser sessions are stopped to free resources.
Models with a small context window can still run a turn; the chat explains when instructions were shortened or swapped for a shorter prompt.
While a reply scrolls beneath it, a pinned prompt stays as a single fading line instead of a full-width bar.
You can mint a provisioning token in Settings and let an agent create additional sandboxes on your account, with chat approval for each one.
Starting or recreating a sandbox no longer fails when its Docker network was removed while the sandbox was stopped.
Agent board column headers no longer show how many items are in each lane.
The active sandbox shows whether it is online through the corner placement icon color instead of a separate status dot.
The desk workspace is now the maker home view, and guest-facing chat is visitor chat instead of front desk.
Tier moves and provisioning use the correct machine size, and you can start hosted plan checkout without sending a request body.
The connect screen shows Starting and Ready for local models and opens chat only after the model is actually serving, not while it is still loading.
Resolving agent conflicts no longer switches to a different account and resets your chat session.
Sandbox areas open in expandable rows, save as you edit, and you can add a short description before granting access.
Mobile sheets can be dragged or dismissed with Back, and workflow details remain usable on narrow screens.
See exactly what each chat was told before your messages, including its base prompt and added guidance.

September 21, 20261.304.0

Model lists no longer shrink when one account cannot access the REST catalog.
Access grants now hand members the assistants working in their assigned areas.

September 21, 20261.303.0Breaking

SliceFolderSchema, SliceIdParamSchema, SliceSchema, and SlicesListSchema are removed; use the area schemas instead.
Users can no longer add service or integration capabilities; use extension-provided CLI connectors such as Stripe instead.
Agent recovery policies consolidate into a single per-ending setting; resumeAfterLimit, moveAfterLimit, and resumeAfterOutage endpoints are removed in favor…

Members can now change files only within granted folders without gaining publishing or secret access.
Fenced conversations now keep their transcripts and checklists separate from other conversations.
Choose hosted machine sizes with different prices and monthly hour allowances.
Find agents, files, terminals, and commands from one searchable palette.
Workspace access is now organized and granted through named areas instead of slices.
Connect Stripe with a key-only CLI connector and use it directly from agent conversations.
Desk users now see only pages they can open and can chat with providers already connected to the sandbox.
Item count badges are no longer shown in capability, VPN, or netdisk lists.
Connect accounts or download a model locally from one guided screen.
Persona avatars now use the Clay illustration style.
Image downloads now retry transient failures and explain whether the problem is the network, credentials, or package.
Automation schedules now run on the workspace owner’s chosen time zone and display that clock clearly.
Shorter session-limit recovery guidance is easier to scan.
Session cards now show owner names without colored dots.
Agents now carry a monthly-updated brief about this sandbox into each turn.
Desk access is now granted only after at least one assistant is selected.
Read-only menus no longer show an unnecessary separator when no actions are available.
Codex models now show descriptions and their available reasoning levels.

September 20, 20261.302.0

Sandbox setup no longer fails when a generated code looks like a command-line flag.
Edit Auto's guidance directly from its model row.
Temporarily unavailable models now show when they will be available again.
Browser sessions now start only when you use a browser tool, reducing turn startup time and resource use.
Desktop setup no longer loses the selected sandbox, rejects certain codes, or shows false scrollbars.
Show one clear status for each computer using its worst environment status.
New chats no longer suggest a vendor sign-in before you choose a model.
Markdown files with YAML frontmatter now display their metadata cleanly instead of as a heading.
Workspace data now refreshes after reconnects, and clicks recover correctly after interrupted drags.
You can tell Auto which work to prioritize, accounts to conserve, and when to use the strongest model.
Workspace owners can invite members to chat with selected assistants without exposing the rest of the workspace.
The site now highlights doing more work with less AI waste on the subscriptions you already have.

September 20, 20261.301.0

Mobile artwork now keeps both figures visible instead of cropping one out.
Make the add sandbox and sandbox settings icons easier to see.
See a clearer, better-fitting homepage layout on phones.
The effort control no longer appears while Auto selects the effort level.
Persona faces are now larger and consistently sized in rows, pickers, and chat controls.
The effort control now shows Auto and cannot be changed while automatic mode is active.
The Auto model picker now uses a shorter explanation.
See running CI pipelines and their jobs update sooner.
The README now explains the workspace, setup, review flow and runtime boundaries.
Reopening long conversations now loads faster while keeping delegated tool calls available when opened.
Reloading a conversation no longer duplicates messages from an active or recently completed turn.
Diff reviews now keep file actions together and place reading options in a compact menu.

September 20, 20261.300.1

Cursor turns now try another connected account when one account has exhausted the requested model.
Show sleeping Windows and WSL environments together as one computer.
Switching Google accounts now opens the account chooser instead of retrying the same account.
Remove methodology tooltips from the agent settings panels.
The source filter now fits cleanly inside the activity filter bar without a doubled border.
Keep the navigation action visible without JavaScript and avoid a brief flash on landing pages.
See clear guidance when the app and sandbox use mismatched contracts.
The header action now appears and disappears correctly as the navigation bar changes size.
Setup screens now show less redundant explanatory text.

September 19, 20261.300.0Breaking

Agent tier selection (fast model routing and prompt complexity judgment) has been removed; turns now use the pinned model without automatic downgrade or hold…

When you open a project from the file list, the workspace browser and search now stay within that project instead of showing files from other projects or the…
When connecting a device, if it hasn't been renamed yet, you're now offered to name it after its hostname for easier identification in tools.
Device pages now show information once instead of repeating it across multiple sections, making the interface cleaner and easier to scan.
Middle-click now closes tabs, pills, and cards across the editor, matching the gesture browsers and most editors use.
intentic now emphasizes token savings, safety, and agent alignment as core site guarantees rather than user actions.
File sync now detects sandbox-side changes twice as fast, reducing latency from 10 seconds to 2 seconds.
Experiment measurements now show side-by-side bar charts comparing treated and control arms, making it easier to see the measured effect at a glance.
Markdown documents in settings now scroll independently, keeping save buttons visible while reading long text.

September 19, 20261.299.0

The landing pages now feature the browser access option as the primary call-to-action, with app download as secondary.
The home page hero plate now has lighter, more refined visual depth with better control over the veil effect on inner pages.
You can now access the browser workspace directly from the main buttons on the desk page.
The welcome header has been removed from the landing pages for a cleaner look.
The site images have been optimized to load more quickly.
Visual typography now includes more font weights for better text contrast and styling.

September 19, 20261.298.0

Model pins now correctly omit settings that are not supported by the selected model or provider.
The board's Everyone/Mine filter now hides when you're the only person with access, since both options say the same thing.

September 19, 20261.297.0

Deleted sandboxes now stay recoverable for 7 days: use `ic sandbox restore` to bring one back with its files and history intact, or `ic sandbox trash delete`…
You can finish checkout in your browser and return to the app without losing your billing page.
Primary actions now use engraved gold buttons with hover, focus, and pressed states across the site and app.
Filter the board by everyone, your sessions, or another member’s sessions with matching identity markers.
Show more context when a review is held for changes.
Stop repeating old setup, reachability, and placement alerts in daily admin notifications.
Users can see who owns a session, filter sessions by owner, and hand sessions over to someone else.
Readers can now explore a document-focused edition of intentic built around their files and everyday work.

September 19, 20261.296.0

Make effort meter rungs easier to click without changing their compact appearance.
Connection screens now identify unresponsive extension routes, and moving a hosted sandbox asks before deleting its machine and data.
See whether each sandbox runs in Intentic's cloud, on your computer, or on another machine.
Readers coming from the desk edition now see document-focused messaging and a workspace instead of a code-focused one.
Choose Auto to let the app select a model for each new chat based on your message and available allowance.
Browser and chat views no longer crash when panels or canvases disappear, and effort controls are easier to press.
The sync watcher now backs off when a sandbox is unreachable, reducing unnecessary DNS lookups and TLS handshakes, and periodically cleans up stray port…

September 19, 20261.295.0

Folders with thousands of files now open and scroll smoothly; the desk and tree no longer build components for files you can't see.
Open marks now stay visible while scrolling through long tool calls, positioned below the prompt that opened them.
Job names in pipeline graphs now link directly to the job's page on the forge.
Reaction buttons now appear in the card header with other actions instead of below the marks, and marks are displayed more compactly in the stats row.
Mark conversations with emoji reactions that everyone on the sandbox sees, with names of who left each mark.

September 18, 20261.294.0

The platform now checks that the configured Stripe price is valid and can be sold when the server starts, reporting any issues at boot instead of waiting for…
The interactive demo's default mode is now labeled "Curated" to better reflect its curated selection of agents and extensions.

September 18, 20261.293.0

User interface strings are now available in German, Spanish, French, and Polish translations.
Automations can now trigger once at a specific date and time, then automatically disable themselves.
You can now mount network file shares into your sandbox to let the agent work with files on your local or remote servers.
Workspace state directory now consistently manages OnlyOffice JWT secrets and related files across sandbox rebuilds.

September 18, 20261.292.0

Row titles and associated controls are now vertically aligned for a more consistent appearance.
Update buttons are now simplified to fit together, and release notes are displayed more cleanly with a more concise summary.
Finished chats now sort by time rather than by what they owe, so recently completed sessions appear at the top of the finished fold regardless of pending work.
The setup page now automatically attempts to restart stopped hosted machines instead of showing a button, with automatic recovery for common issues like idle…
Window drag operations no longer leave text selections behind on the desktop app.
The extension gallery now features the logs viewer, scrub metadata tool, and SaldeoSMART financial management extension.
The setup screen now displays a cleaner, more compact header when preparing your sandbox.
The setup page now survives being reloaded mid-installation, letting you refresh without losing your sandbox or needing to re-run the install command.
Display headlines now render as flat color text instead of textured stone, improving readability and reducing visual complexity.
The finished lane now strictly respects its display cap, ensuring consistent scrolling behavior for finished agents.

September 18, 20261.291.0

Sync status now displays in the user's language when no upstream branch exists.
Plan limits on the chat rail now account for rate-limited accounts, showing held accounts separately so the age reflects only accounts that can actually take…
Persona briefs now display below names in the sandbox personas list instead of in a separate column.
Finished chats you've already read now close automatically when you move focus, unless you pin them.
Sync status now displays ahead and behind commit counts as visual pills, making it easier to see upstream synchronization status at a glance.
Agent cards now move to Active immediately when a turn is sent, instead of waiting for the server to confirm.
Questions raised inside custom tools now appear immediately instead of waiting for the agent's next message.
Terminal now handles Ctrl+Backspace, Cmd+arrows, and Home/End keys correctly for line editing in shell sessions.
Environment contents now correctly detect prefix-installed npm modules like @cursor/sdk by reading their package.json manifest instead of looking for a…
Selected cards that need action are now shown at the top of the finished lane instead of being moved to the bottom.
The unstage button now displays a minus icon instead of an undo icon for better visual consistency.
The sandbox overview no longer shows an online status badge when the environment is already running.
Your subscription plan badge now appears directly next to your profile information in the account menu.
Agent memory import controls are now located in their own dedicated settings section.
File icons and workspace grouping now support a wider variety of document, spreadsheet, and archive formats.
You can now run JavaScript code in Cursor agent sessions, with your established security rules automatically protecting those scripts.
The settings interface for agent instructions and memory has been streamlined and simplified.
Agent instructions are now correctly applied to Cursor sessions, and steering messages are delivered without restarting the active turn.
You can now open and view agent conversations even while the agent is busy performing an action like landing.
Skills and engine versions now display clearer status labels to indicate their installation source and recommendations.
Rebuilding your environment from a local checkout now clearly states that it also applies any pending recipe changes, helping you distinguish it from a simple…
Environment install counts are no longer displayed in section headers.
The app now starts Docker Desktop automatically when you open it after a restart, if your sandbox is hosted on this machine.
Word documents now show diffs as marked redlines instead of two side-by-side versions.
Browser sessions now handle page dialogs like alerts, confirms, and prompts so agents can interact with them.
Documents in Word, OpenDocument, and RTF formats now preview their first page in the file peek card.
Build output directories no longer disappear from the Changes list while the daemon is running checks.
A new guide explains why a full page reload can show older code than HMR displayed, with steps to verify and fix the issue.
Browsers now have a client-side address bar for navigation, back button, and can handle JavaScript dialogs that pages open.
Date and number formatting now updates immediately when you change the language, matching translated text updates.
You can now tell a device to stop mirroring a single port number that something else on that machine already uses, while keeping all other ports mirrored.
The phone preview now displays the device name, viewport dimensions, and scale percentage below the frame, with refined hardware styling and a material-based…
Dialogs and media player controls now render more consistently and respond faster to state changes.
The chat quick bar now opens smoothly with an improved animation that doesn't blur text or interfere with autocomplete lists.
Save and discard buttons now appear in the changes panel header for faster access instead of at the bottom.
Changed documents (Word, PDF, spreadsheets) now show tracked changes over their text; images can be compared with swipe or onion-skin overlays.
The chat quick bar now displays the agent's identity with a visual context window indicator ring instead of a status icon.
The composer no longer shifts position when you hover over the transcript peek handle.
Social share cards now display the app in light mode with a live screenshot, matching the product's visual identity.
The rail no longer empties itself when tile groups are rebuilt with identical but non-identical objects.
UI text is now consistently translated across desktop, web, and extension interfaces in multiple languages.

September 17, 20261.290.0

You can now hover over a handle at the top of the chat quick bar to read recent messages without opening the full panel.
The floating chat composer now displays cleanly without extra borders or status information, making it easier to focus on writing your message.
The chat composer now floats as a pill over the page instead of taking up space at the bottom, so it no longer covers content when you open it.
You can now compose messages from any area of the app using a quick bar at the bottom, without navigating to the chat panel.
Condition watches can now be disarmed individually from their transcript row, and watch wake messages are now displayed as notices rather than user messages.
Asset uploads to GitHub that fail transiently are now retried automatically instead of failing the entire release.

September 17, 20261.289.0

When you re-measure plan limits and some accounts are rate-limited, the chat panel now shows which ones can't be read yet and when they'll be available again.
The desk view is now always shown when you have files in your workspace, and the appearance setting to toggle it has been removed.
Agents can now query PostHog session replays, events, errors, and feature flags with a personal API key.
Mobile surfaces now properly restore scroll position after testing, preventing layout shifts from affecting subsequent measurements.
Preview performance on the sandbox's own machine is now hundreds of milliseconds faster, and ONLYOFFICE can auto-start the document server with the sandbox to…
Bottom sheets, graphs, and code viewers now scroll smoothly under your finger on phones, and unnecessary controls are hidden to make room.
Document server initialization now waits for the OnlyOffice entrypoint to finish all setup steps before allowing documents to open, preventing failures when…

September 17, 20261.288.0

OnlyOffice and other embedded editors can now frame their own pages within the preview proxy.
Office documents (docx, xlsx, pptx and variants) now open in ONLYOFFICE Docs for full editing and saving when the extension is enabled.
Device agent updates now work for distros on Windows PCs, which previously had no way to be updated from this page.
The interface now speaks English, German, Spanish, French, and Polish, picking your browser's language automatically.

September 17, 20261.287.0

When a device operation fails, you can now see the command to run directly on the machine to accomplish the same thing.
Agent and chat cards now consistently show the same status word and color in all views:why something needs attention, that it has unread updates, or nothing…
You can now open zip and tar archives in your workspace and browse their contents without extracting them first.
Fix CI turns now load cards, plan mode, terminal hand-off, and your signed-in accounts, giving you the full environment to diagnose and fix the pipeline.
Archive extraction now works reliably in CI environments by conditionally testing zip support based on tool availability.
Builds no longer re-fetch the knowledge engine from GitHub when its commit hasn't moved, reducing rate-limit pressure on shared CI runners.
You can now right-click an archive file (.zip, .tar, .tar.gz, .tar.bz2, .tar.xz, .tar.zst, .gz, .bz2, .xz, .zst) to extract it into a new folder beside it.
Onboarding no longer fails when postgres is still initializing; it now waits for the database to be ready before attempting migrations.
Dragging files and folders in the workspace is now smoother and no longer causes tab freezes in Brave.
The desk now supports renaming, creating files and folders, deleting entries, and drag-drop:the same file management as the tree.
Picture and video files now display thumbnails in the desk tiles and quick-look card, with videos playing silently on hover.
Browse folders as large tiles in the main area when nothing is open; toggle in Appearance settings under Desk.
Task checklists now display at both the beginning and end of each turn, showing which tasks moved on without finishing so you can see the complete picture of…
Renaming sandboxes, personas, and accounts now happens in a unified field that stays compact and never moves the page layout.

September 17, 20261.286.0

The editor no longer flashes white when loading in dark mode, showing the correct background color from the first paint.
Session titles now display cleanly without action words, while work categories (fix, audit, redesign) are read from stored metadata for board tinting.
The device status summary display has been removed from the Devices tab.
Mouse input events in xdotool commands no longer wait for pointer movement, improving responsiveness of mouse actions.

September 17, 20261.285.0

Chat marks now stay visible when the lane grows wider, instead of being cut off by the edge.
Image pull failures are now properly distinguished between registry issues and local image store problems, improving release robustness.
Build output changes are now detected even when writes happen faster than file watching can track them.
Command output now uses consistent phrasing for counts (e.g., "1 resource" vs "3 resources" instead of "resource(s)") and aligns data tables to readable…
Fenced code blocks in the markdown editor now display with syntax highlighting and language labels, matching the rendered view, while keeping the source…
Tap a push notification to land straight in the conversation; attach photos and files from the composer with a paperclip; Chat is now a tab on your phone.
Checks marked as scoped now run instantly when you edit a file, giving you feedback before you finish writing the line.
Run `pnpm try:onboarding` to build and test the installer on your PC before publishing, with options to check readiness, skip rebuilding, or teardown afterward.

September 16, 20261.284.0

Sign-in panels now show "waiting for you" before sending to the provider and "signing in…" after, making the flow clearer.
Chat and subagent lists now display as clean card columns without recessed backgrounds, reading as unified lists rather than separate sections.
Tabs in the workspace and chat now appear as underlines instead of bordered pills, reducing visual clutter where they function as headers.
Audio files can now be attached to messages and play inline with a waveform visualization, just like images show thumbnails.
Chat rail lanes now display as a flat list with headers labeling cards instead of recessed sections, creating a cleaner visual hierarchy.
The Devices tab now offers a "Clear build output" button for directories where ignored files are preventing the sandbox's deletions from syncing, and the…

September 16, 20261.283.0

The app now uses a consistent typeface across all interface themes instead of switching fonts between light and dark modes.
Marketing screenshots can now pin which extensions appear in the sidebar independently of the demo fixture density, allowing cleaner product showcase images.
The application now uses your system's default browser instead of defaulting to Google Chrome.
The AI account connection indicator now correctly identifies connected providers, ensuring the agent configuration screen loads as expected.

September 16, 20261.282.0Breaking

Extension install fields are no longer accepted in manifest schemas; define extension sources directly within your local configuration instead.

You can now use "iq read <symbol>" to view code definitions directly without needing to know line numbers or perform multiple search steps.
Makers can now see, review, and save their changes in a dedicated panel.
Visual alignment of icons in the interface has been corrected.
Device cards no longer display empty gaps when they have no environments, warnings, or logs to show.
The interface now displays a clear status and progress notification whenever you perform an action that restarts your sandbox, such as a rebuild or update.
Clicks and mouse movements no longer accidentally trigger in the top-left corner of the browser window while the page is still loading.

September 16, 20261.281.0

The workspace now correctly shows which changes are being applied during a land, instead of incorrectly claiming the workspace is clean.
You can now see progress indicators on hub tabs while long-running tasks, such as container builds or sandbox restarts, occur in the background.

September 16, 20261.280.0

The AI now automatically resumes requests that stop due to errors or timeouts instead of forcing you to press Continue manually.
You can now set the appearance to System to have the app automatically match your operating system's light or dark mode.
Forms now properly ignore repeated button clicks while processing, and secret updates are protected against accidental overwriting.
A warning is now shown if you attempt to require passkeys before adding your own.
Agent cards now glide smoothly into new positions when they move between lanes.
File editing is now always enabled, removing the separate edit mode and providing persistent save controls.
Desktop app links that open in a new window now work correctly.
The add model button in automations now correctly anchors to the list for improved screen reader navigation.
Disabled toggle switches now correctly show their active status instead of appearing inactive.
Screenshots from agent tool runs are now correctly handled and no longer cause chat errors when using free trial models.

September 16, 20261.279.1

The rebuild log will no longer falsely report a failed build if the read command times out due to high machine load.
The app now stores browser profiles and trash files directly under the .intentic/ directory instead of a host/ subdirectory.

September 16, 20261.279.0

Your current task checklist is now visible in the side rail, staying on screen while you chat.
Assistant thinking, tool history, and meta-notes are now displayed in the margins to keep conversation threads clean and readable.
You can now define custom responses or approval requirements per sender or role for your message-listening automations.

September 16, 20261.278.0Breaking

Configurations and manifests containing sandboxRemove no longer parse; use the sandboxes permission to manage and remove sandboxes instead.

The temple scene now shows clearer carvings and more readable figures at a glance.
The desktop setup window now matches the visual style of the web setup experience.
Thinking, errands, and context added to messages now appear as compact, expandable edge pills.
The build illustration now distinguishes your material from the AI agents doing the work.
Removing sandboxes now uses the existing manage-sandboxes permission instead of a separate removal permission.
Repository rows now open one management panel containing Git, Docs, and Health tabs.
Desktop sign-in now guides you through browser authentication and handoff with clearer progress and recovery actions.
See the same agent status, actions, and sandbox details in the desktop app and web workspace
Repository files inside `.intentic` no longer appear as deleted after syncing.
File sync now correctly handles multiple operating systems on one computer, letting you sync folders directly to Linux WSL distros or Windows based on where…
The editor now loads before your dev servers start, making the interface responsive sooner when opening a workspace.

September 15, 20261.277.0

The app now automatically releases unused remote environments when you arrive, allowing you to seamlessly set up your local computer instead.

September 15, 20261.276.0

Agent sessions now reliably continue even if you close the original window while the first message is sending.
File sync now correctly handles multiple operating systems on one computer, letting you sync folders directly to Linux WSL distros or Windows based on where…
You can now remove installed extensions, which also deletes any connections configured from their cards.
The "Copy session name" action now copies the clean session identifier without the agent prefix, ready to be pasted into the CLI or Quick Open.

September 15, 20261.275.0

Popped-out chat, terminal, and preview panels now open in their own separate windows instead of browser tabs.
The chat composer now correctly loads available AI providers for members who lack full access to the sandbox's underlying connection settings.
The dashboard now ignores brief connection drops when restarting a device's agent, preventing temporary warnings while the machine reconnects.
Machines with WSL distributions now appear as a single device card with separate connection states for Windows and each Linux distribution.
Markdown tables now automatically wrap long paths and align numeric columns for better readability.
You can now rebuild or reload your sandbox from the Windows side of a PC without needing to connect its Linux side directly.

September 15, 20261.274.0

Commands like dev rebuild and reload now correctly target the Linux environment on a PC instead of failing when the Windows side is selected.
The "Pause syncing" button now correctly identifies when a sandbox is unreachable and avoids triggering errors for inactive connections.

September 15, 20261.273.0

If a part of the editor fails to load, a notification now appears offering to reload the page and fix it.
Row information and controls now intelligently wrap to fit available space instead of getting squashed in narrow views.
Machines running multiple environments now list each one separately in the device board, making it clearer which environment you are operating.
You can now install extensions using a branch or release name instead of a full commit hash.
Window dragging and maximization are now more reliable, and the interface no longer flickers on startup.
Agent device status and task descriptions are now shown as icons with hover details for better clarity.

September 15, 20261.272.0

You can now clear local conflicts directly from an agent's card when your own edits prevent it from finishing its work.
Scrollbars now use a more consistent default appearance throughout the interface.
You can now toggle between different phone and tablet frames when previewing your app.

September 15, 20261.271.0Breaking

The acceptEdits mode has been removed; replace it with the manual mode to request confirmation per tool or bypassPermissions for automatic execution.

Setting up your sandbox now happens within the existing window instead of opening a second one.
Window title bars no longer briefly flash grey when switching between views.
The desktop app is now available for download and automatic updates through the Microsoft Store.
You can now identify turns where the assistant spoke but performed no actions by filtering for "silent" turns in the diagnostics report.
The desktop app's setup and requirements screens now correctly guide you through restarts and account changes, and match your workspace's light or dark theme.
The file upload and document viewer systems have been reorganized to improve maintainability.

September 15, 20261.270.0

Workspace trees and browser displays now load faster by sharing resources across concurrent requests.
Agent cards now show your progress through checklist tasks as a ring of ticks around their status icon.
You can now open and preview OpenDocument, Rich Text, and EPUB files directly in the workspace.
File renames, deletions, and uploads are now processed concurrently, and tabs automatically follow files that are renamed.
Presentations now display in the file viewer, and opening previously viewed documents is near-instant.
The command palette now groups related actions together and shows recently used commands first.
You can now disconnect devices directly from their rows in the Capabilities list.
The indicator for when a model is thinking now uses a calm breathing animation instead of a spinning wheel.
The setup window now drags by any part of its header, not just small gaps between buttons.
The app now shows a splash screen during startup, preventing a blank screen when loading over a slow connection.
Files you upload now appear instantly in the explorer as you drop them, before they finish sending to the server.
Tooltips now have a clearer border and improved elevation to stand out against the interface.
Developers can now switch to a production deployment within their local admin panel to view live data.

September 14, 20261.269.0

Project cards now show readable summaries, stable colours, and direct links to running previews.
See live progress and waiting status while files are rendered to text in the background.

September 14, 20261.268.0

Navigation now uses clearer names for product, documentation, and comparison sections.
Readers arriving from the light desk now keep its light look and maker-friendly sandbox defaults in the app.
Register passkeys, use them to sign in, and require passkey verification for sandbox access.
Text weights now appear visually consistent across light and dark skins.
The app and site now keep their intended typography even without network access.
Reload a sandbox when its host is unknown if exactly one device is connected.

September 14, 20261.267.0

Site text now preserves the intended weight range and headline color across skins.
Explain machine, sandbox, persona and project in one visual guide.
Headlines now use a quieter, screen-focused serif across the site.
Choose separate reading and headline fonts from the Type panel, with Public Sans now used by default.
Use @ in the composer to search files or change persona, placement, model, and effort settings
Choose a more readable typeface and see clearer desk pages.
The site now shows tailored product imagery in both dark and light skins.
Opening a project now filters repository views and gives new agents a project-scoped identity.
Empty markdown files now show where to start typing and remain clickable across the editor pane.

September 14, 20261.266.0Breaking

Device sandbox inventories now arrive on Device.sandboxes; clients must read that optional field instead of Device.report.sandboxes.

Opening a project now narrows the workspace, agents, checks, and other project views to that project.
Offer a light reading view for readers who prefer a paper-like design.
Visitors can now receive approval and human replies after the live chat connection closes.
Browse workspace repositories as projects, create new ones, and open each in its own workspace tree.
See when a conversation's cached context is about to expire.
Choose a maker-friendly workspace that highlights projects, drafts, accepted changes, and easy restore points.

September 14, 20261.264.0

See a new visual explanation of how plans, agents, and approvals become your software.
Keep review results visible while changes refresh in the background.

September 13, 20261.263.0

Workspace can now show readable text and archive member listings for supported binary files.
Choose which setup notes each persona receives before your messages are sent to the model

September 13, 20261.262.0

Hosted setup no longer reports startup transitions as failures and now detects stale edge deployments.

September 13, 20261.261.0

Chat file previews use a consistent compact button, and capacity reset details take up less space.
Badges now animate in and out so new counts are easier to notice.
Folded message details now use tighter spacing and align more cleanly with their headers.
Browser MCP tools that require approval now run successfully.
Show when a sandbox was removed instead of waiting indefinitely for it to reconnect.

September 13, 20261.260.0

Attachments now appear as distinct file objects instead of empty-looking message bubbles
The terminal now automatically matches your current color theme and skin.
The chat capacity side panel now shows how long remains until your usage limits reset.

September 13, 20261.259.0

Preview attached files with names and sample lines in the composer and chat.
Accounts that cannot run turns are listed with why, and no longer count toward the total in the capacity ratio.
Resources for a conversation stay protected while an agent turn is still registered, so they are not cleaned up as idle too early.
Installer setup now stops instead of hanging after a terminal command failure.
App previews open inside the editor on first run instead of a blank frame when the dev server blocks embedding.
Signing in again to a subscription account you already connected no longer leaves the sign-in spinner stuck.
The last stop before a turn ends re-runs verification so a check that now passes is accepted without another follow-up.
The agents board and Synthesize use the same side-by-side conversations as your popped-out chat, and your workflow choice is restored with your tabs.

September 12, 20261.258.0

You can switch to your own computer while hosted setup is still starting; setup won't reopen the hosted workspace if provisioning finishes afterward.
When you pop the chat into its own window, the whole transcript scales up evenly so prompts, answers, and buttons stay visually balanced.

September 12, 20261.257.0Breaking

Sandboxes no longer take dev-rebuild as a sandbox op; rebuilding a dev sandbox from source now goes through the dev-rebuild device command instead.

Resolving agent conflicts uses that agent's model, and landing agent changes no longer fails when dependency folders were mirrored into the worktree.
When you pop chat into its own window, transcript messages use larger text that is easier to read.
On the sandbox personas page, matching new chats to a persona is controlled with a switch instead of Off and On buttons.
The short explanation at the top of the sandbox personas page is removed.

September 12, 20261.256.0

Sandboxes you built from a local checkout can be rebuilt from that code in the editor, and updating the release channel no longer replaces that image unless…

September 12, 20261.255.0

Safety settings no longer repeat the same notice about actions that always require your confirmation.
The agent safety rules panel no longer shows a count next to the heading, and command patterns stay aligned beside the machine verdict columns.

September 12, 20261.254.2

Keyboard navigation is now easier to follow with clearer focus indicators on buttons, links, and sidebar items.

September 12, 20261.254.1

Hosted sandboxes should start from the warm pool in seconds again after a sandbox image update, instead of timing out and rebuilding cold.

September 11, 20261.254.0

Repositories can declare commands to run before a push or when an agent turn touches them; you turn each declaration on in agent settings before it runs.

September 11, 20261.253.0

Light mode looks warmer and clearer, with visible row highlights and easier-to-read text.

September 9, 20261.252.0

Run sandbox rebuilds, sync setup, and agent restarts directly from the app instead of copying commands to a terminal.

September 9, 20261.251.0

Icons in the ports list and account menu look updated, and port origins use the same symbols as their sections.
Dismissing a failed pre-push check keeps the result on Review; you can reopen it or push again without rerunning checks if nothing changed since.
Chat rail session rows show how full the context is around the avatar and highlight unfinished work when a turn stopped mid-task.
After a fix agent stops, you can continue where it left off or start a fresh attempt from the run button.
When a fix for a failed check or pipeline already has an attempt, you can continue that attempt or start over with a new one.
Active agent cards stand out more on the board, finished ones look quieter, and context use shows as a ring around each agent’s icon.
Pipelines, workflows, and deployment tiles can show a running spinner on the sidebar while work is in progress, including at the same time as failure counts.

September 8, 20261.250.0

Steering an automated turn lets it ask for approval again; unattended runs explain missing accounts are intentional, not failed logins.
Model pickers use the same thinking and speed toggles and no longer offer a Default choice that left those settings unset.
Unfinished work shows as a dot on the status icon with details in the tooltip instead of a separate label, and outdated step counts clear after a session…
Icons in the editor, shared links, code views, and diagrams now load offline with the same look, without an external icon service.
You can close every chat in Attention or Active from that lane’s Clear button, not only finished chats.
Renaming a chat title no longer saves to the wrong chat when you switch tabs or open a chat from another window, and the rename field closes if that chat…
When a model becomes unavailable, your chat is moved to another model temporarily. If the model becomes available again, your chat is automatically restored…
The Devices tab no longer marks machines as gone quiet while a cached list is on screen or a fresh read is still loading.
Standing workspace rules live in AGENTS.md and are applied on every turn for every model, not only when a runtime discovers CLAUDE.md on its own.
You can turn your own skills off without losing the saved text; only built-in tools use the skills list in settings.

September 8, 20261.249.0Breaking

personaRouting is only true or false on the wire; off, suggest, and auto strings in settings or manifest JSON must be replaced with false or true.
Stage, commit and discard now take a described scope instead of a path list; reload any tab left open on an older version before using them.
Automations now require a models ladder listing which providers and models may run them; agent, harness, and model fields are removed from the schema and API.
automation triggers via ingestKey and polling are removed; use control tokens to authorize callers instead.

The desktop app now prevents external websites from triggering sandbox operations like recreate, and setup links opened from outside the app only prompt once…
Devices now displays as a board of cards for each machine, with deep-linking by machine and the pairing form in a modal, instead of folding machines into…
Rules can now sample a fraction of matching occasions with the sample condition, and the turbo build cache is now shared across CI runners and sandboxes for…
New chats are matched to a persona when you send the first message:the chat shows the reading in the transcript instead of a chip on the composer, and the…
Agent cards show Landing while work merges into the workspace, and starting another land waits instead of conflicting or timing out early.
Subagent cards show each child's chosen model, including tier names like Sonnet, when a spawn named one instead of inheriting the parent's.
Code comments now follow a consistent single-line format stating what code cannot do, improving readability and maintainability across the codebase.
Icons in the rail, menus, and lists now use Intentic’s custom line drawings instead of the previous Remix-style set.
Desktop rail icons for workspace, subagents, terminal, and capabilities look updated.
Reattaching or hydrating a chat no longer duplicates messages that were already on screen.
Desktop sidebar rail icons are slightly larger for easier recognition at a glance.
The desktop rail, More menu, and mobile tabs now use custom navigation icons instead of the previous generic icons.
Starting an agent from the caret menu now uses a confirm button and can set thinking and speed before the run starts.
Task lists no longer repeat when usage limits block a retry, and the weekly session reset option only appears when your account can actually claim it.
In the desktop app, move and window buttons live in the app’s top bar instead of a separate system title bar above the workspace.
Clicking an agent link in pipelines, docs, or acceptance reports opens that conversation in the docked chat panel instead of leaving your current view.
Devices and extensions reconnect when the sandbox goes quiet on the link, and setup immediately says when a public address cannot be reached because no edge was
Stop showing Connect for providers that manage their own credentials and avoid failed connections for provider IDs containing slashes.
Close the installer and continue watching its progress in your workspace
Wallet spending limits update when you save the capability while signed in, and invited viewers no longer receive your sandbox connect token.
Home page product screenshots now extend slightly wider than the text column above them.
The pricing page now makes clear that you use and pay for your own AI accounts on both self-hosted and hosted options.
Icons in the pricing page feature lists line up more cleanly with the text next to them.
When a usage limit stops a turn, you can move it to another account, see what each continue option costs in tokens, and optionally keep the same provider sessio
Commit messages with misplaced markers or spacing now reach Git hooks correctly, and long notes no longer end mid-word.
Breaking commit messages with a scope are now accepted instead of being rejected for marker placement.
Re-running device or desktop sync setup from your sandbox card replaces a local agent that no longer understands the command.
Hosted sandboxes no longer shut down for idle if you connect or start activity while the idle check is running.
After resuming a conversation, the task list and step completion count stay in sync with work from earlier turns.
Dragging pane edges no longer selects text on the page, and resizable columns stay the correct width beside the icon rail.
Auto-drafted commit subjects are no longer cut at 80 characters mid-word, and common commit-msg hook rejections are repaired before you commit.
Bare iq files lists the workspace; search answers name symbol definitions and show where the match occurred; confident vs ambiguous compares top files, not two
The hosted plan can now be tested end-to-end without a Stripe account, and the platform validates the webhook secret at boot to catch misconfigurations before t
Scheduled and chore automations now run in isolated worktrees so changes stay on a branch until you land them.
Access and API token screens use the same status chips as elsewhere, with invite and mint controls sized to match the member list.
Maintenance shows when a chore was already answered by an agent; those rows lose the warning style, sort below untouched ones, and no longer count as due until
The app no longer shows a permission prompt for local sandbox access on desktop if your browser has already granted or denied the permission.
Python type checking no longer reports undefined-name errors twice when both ruff and pyright run.
When the hosted fleet is full, setup says no machines are free and points you to run on your own computer instead of showing a provider error.
Your plan now appears as a chip next to your name on Settings ▸ Profile, matching the account menu.
Install and connect script links on intentic.dev deploy with the site instead of staying stale or 404.
The account menu shows your plan as a badge beside your name; use Settings for billing instead of the old hours or renewal link.
With the safety judge off, a shared model list for a job group no longer overwrites the judge or shows “jobs differ” because of it.
Fixing the same push check failure again continues the existing fix conversation instead of starting a second agent.
Sync conflicts on a paired device now list the affected files on the device card, and Fix with agent starts a turn to help reconcile the two copies.
Shell and panel sessions show the code-style terminal icon in pills and recent work, matching the sidebar terminal tile instead of the monitor icon.
Settings Billing shows your plan, slots, and monthly awake hours; waking a free hosted sandbox is blocked once the month's hours are used until you upgrade or
The What gets stopped list is a table comparing this sandbox and your devices side by side, with command patterns shown as highlighted snippets instead of grey
Scheduled automations can wait until a chosen number of new sessions have run since the last wake; until then, due runs skip and the row shows how far along you
Terminal tabs no longer open to a blank pane or drop keystrokes typed before the session pane is ready.
On Agent Models, each group can show one shared model list for all its jobs or separate rows per job, with a switch in the group heading.
A terminal tab stays on its own session when other jobs or agents open windows, so it no longer goes blank mid-run.
Checked checkboxes and radio buttons show a visible tick again instead of a mark that blended into the background.
You can remove or rename unrecognized settings keys from sandbox manifest warnings with one click instead of editing the file by hand.
The in-app services catalog and spend-approval waiting state on agents are gone; paying external APIs still works through the agent wallet.
After Copy or Paste from the terminal menu, the keyboard stays on the terminal so your next keystrokes go to the shell.
If you're on the hosted plan, setup shows your hosted option as always on instead of free.
Standard sandboxes include Python tooling; .py edits get ruff and pyright feedback, and pytest tests count toward turn-end test checks like TypeScript tests.
Invites now say whether you join as viewer, collaborator, or maintainer; without maintainer access, the workspace and changes panel stay read-only with a clear
Terminal search (Ctrl+F) now works over the full scrollback, and the scrollbar shows where matches are in history.
The onboarding setup now works correctly in CI environments where the job container cannot access the host's Docker socket.
Dismissed runtime installs no longer clutter the Environment card; they're now hidden by default with one click to review them, and the maintenance panel reorga
The workspace is reorganized for improved organization and clarity: shared contracts, utilities and extension APIs moved into a dedicated `_shared/` directory,
Markdown documents can now be edited with a unified reading and writing surface across the app, with live formatting markup hidden until you type into a block.
Agent roster now automatically reloads if it detects it missed an update from the daemon, instead of remaining frozen until you reload the page.
Control token scope picker now displays concise action descriptions instead of full technical documentation, making it clearer what permissions each scope grant
Automations now use control tokens for programmatic access instead of ingestKey credentials, with improved security and audit trails.
Areas opened from the More menu now show they'll disappear when you leave, with a right-click option to keep them pinned.
The verbs section on the landing page now displays only an eyebrow label above the verb list instead of a large heading, simplifying the visual hierarchy.
Safety policy now edits in a markdown-aware surface with syntax highlighting, growing with your text instead of fixed rows.
Blog page headings now scale responsively across device sizes for better readability.

September 6, 20261.248.0Breaking

Context shelves are removed; sessions now carry a static context defined on the persona card. Existing shelves must be migrated to the card's `context` field an

The website now has a pricing page explaining that intentic is free, with a navigation bar organized by buyer journey instead of builder features.
Chore scope information (what doesn't apply, what hasn't been measured) now appears as a single expandable line under the list instead of a strip above it, and
End-to-end tests now pull the latest sandbox image on each run instead of reusing cached versions, fixing stale image issues in long-lived test environments.
Deployments status tally now appears on the title row alongside Automations and Pipelines, improving the reading order for operations monitoring.
Sandbox and development environment now use Node 24.20.0, pnpm 12.3.4, and updated tools including cloudflared 2026.8.3 for improved stability and security.
Automations can now include a dreaming session that periodically reviews past sandbox runs and suggests improvements to personas, capabilities, environment, or
The model selection interface now uses a sticky header with bulk actions that stay visible while scrolling through the job list.
Model roles that were previously configured automatically now require explicit selection: unset roles no longer fall back to a default model.
Closing your last chat no longer shows a selected agent on the board while your chat window sits empty.
The message composer now grows to fit your message without losing sight of the conversation, automatically adjusting to your window size.
Skill descriptions are now validated to fit within the token budget, ensuring they don't add unnecessary overhead to every agent turn.
Error messages about failed folder opens now clear when you switch workspaces or close the folder, preventing stale errors from appearing on unrelated operation
The sort-by-added toggle has moved to Settings › Appearance; it's no longer on the Changes panel and agent review bars, reducing clutter while keeping the prefe
Sandboxes provisioned via `ic sandbox connect` are now reachable on their public addresses; previously they would start successfully but answer 502 on connectio

September 6, 20261.247.0Breaking

Extension capability fields that gated on a when object of key and value must now write the condition as text, such as auth == 'key'.

Search results now show each file once with later hits indented underneath, reducing output by ~15% while preserving all line numbers and matched content.
Sandbox usage can compare conversations with and without the project map using how many directory listings the first turn needed.
When a chat is blocked by a full session window, you can spend your once-a-week reset to reopen that account immediately and continue, without waiting for the t

September 4, 20261.245.0

Effort meter and reset button no longer shift position when selecting a reasoning effort level.
Measure the token cost of finding and reading symbols across the codebase to optimize for agent efficiency.
Edit-rule scripts are now tracked in version control and survive sandbox rebuilds, fixing a day-long outage where linting and byte validation were silently off.
Turn verification now reports all problems found instead of stopping after the first one, so you get a complete list of issues to fix at once.
Machines running agent builds too old to report their version now show why their sandboxes can't be managed, instead of silently offering no buttons.

September 4, 20261.244.0

The safety judge can now be turned off, or set to watch mode to learn what it decides without interrupting commands; you can also choose which AI model judges c
Environment rebuilds on hosted sandboxes now build overlays on Fly; press Rebuild now on the Environment card and the tools are baked in within minutes.
Extensions and registry browse are now one section with pills to switch between what you have installed and what's published, keeping your search across both.
Engine update, revert, and channel changes now persist across tab switches, and updates show live progress with automatic polling.
npm package publishing now automatically retries on transparency log conflicts, preventing incomplete publishes when the same version is published multiple time
The Max reasoning effort tier is now available when starting agent runs from buttons that don't have their own thinking control, fixing cases where it was incor
Agents can now be required to get approval from a named person before using specific credentials, letting you put high-risk secrets behind human gates rather th
The subagent pane now shows controls in a consistent footer location matching the chat interface, removing redundant header information.
Translator engine downloads now use the correct architecture names (amd64 for x64, aarch64 for arm64) that upstream publishes.