What's new

• Branch a chat, or lay it out as a canvas:

  • Answers have a new Continue from here action that starts a branch from that message; everything before the fork stays the same as in the original chat. Branches sit under the main thread in the left panel, and Tangu knows how many branches the chat has and where each one got to. From the last answer in a branch, Bring back returns you to the original thread with that answer attached to the input as a quote.

  • By the way opens on the right of the chat and keeps what you asked there; Ask in sidebar on selected text is now part of By the way. Under each answer you see the branches, side questions and child sessions that started from it.

  • Switch to Canvas at the top of the chat column. Each turn becomes a question card and an answer card, stacked vertically by default or laid out side by side. Long answers show a preview on the card: double-click a card to open a reading pane on the right, and use ↑ ↓ to move to the previous or next turn of that branch. Hover the ⊕ under an answer to branch off from there and keep asking, or the ⊕ beside a question to ask it another way. Cards and empty space have right-click menus: branch, copy, By the way, show only this branch, collapse the cards below (they fold into "+N"), and rename or archive a branch. The canvas follows a new answer while it's written, and if you scroll away, an arrow takes you back.

• Child sessions and task cards: Tangu can open several visible child sessions to split up the work (up to 4 at a time). Each one hands its result back to the main chat when it's done, and the main chat carries on. In the left panel, child sessions sit under their parent with a status at the end of the row: running, waiting for you, or done but not yet viewed. The answer that started them lists them underneath, and a row that's waiting for you offers Reply or Let Tangu answer. When an answer ends with a question, the chat is marked as waiting for you. Task cards have a new Start child session action.

• Muse's chat and the inbox each have a job: anything Muse needs you to answer or decide, the to-dos it notes down and background actions that need your approval all come to your chat with Muse; finished work still goes to the inbox. The badge on the Tangu icon counts chats with new messages or waiting for you, and turns a stronger color when one is waiting. A chat that starts waiting for you also sends a system notification. In Settings → Notifications, letters now come as a Daily digest: one a day, or none.

• Cards in chat do their own math: for answers such as a mortgage calculator, splitting a bill, unit conversion, vocabulary flashcards or step-by-step guides, change a number in the card and the result updates on the spot, without asking the model again. Cards can show large figures and tables, and a follow-up question carries the values you changed in the card. Web cards in chat are now called Intelligent UI everywhere. Phones still show the previous cards for now.

• Chat:

  • Thinking and tool calls are combined into one Process line above the answer. Collapsed, it's a one-line summary (such as "Thought for 8s · Ran 2 commands · Read 1 item"); expanded, it lists every step.

  • Attachments show as tiles: images as square thumbnails, other files as a strip with their type and size. Sent attachments sit above the message, and a message with only attachments no longer draws an empty bubble. Very long prompts fold automatically; click Show all to read the rest (copying, editing and what the model receives still use the full text).

  • Click an image in a chat to view it large right there: switch left and right, zoom with the scroll wheel or a pinch, double-click for actual size, copy or download it, and press Esc to close.

  • Sent messages rise out of the input box; with Reduce motion turned on in your system, they just fade in.

  • No need to drag an image in again to edit it: when this turn has no image, image editing uses the one from the previous message, and it can keep working on the image it just generated. Images dragged into a chat become image attachments (local chats still get the file path too). Editing and generating images in chats that already contain images uses less quota.

  • Thinking levels now work with DeepSeek V4.1 Flash, Gemini 4 and StepFun. When a model doesn't support the level you picked, it uses the nearest one (lower first, then higher), and the level actually used is shown next to Thinking effort.

• Penzor Cloud:

  • A new Penzor Cloud Space on the Ribbon opens on a usage overview. The left panel lists what you keep in the cloud under Vault sync, Amadeus Cloud, App files, Exports and Trash. You can download, rename, move, move to trash, browse earlier versions and restore (synced vaults are view and download only), and synced files show which device wrote them.

  • My drive: upload files and whole folders, create folders, or drag files straight in from Finder. Files can be up to 50 MB each and count toward the 10 GB for app files. Uploads and downloads queue in the background and keep going when you switch Spaces; a floating card at the bottom right shows progress and lets you pause, cancel and retry. The Ribbon icon shows a progress ring, and quitting while something is still transferring asks first.

  • The Sync page gathers the switches, status and syncing devices for your notes vault, Agents & memory and plugin data. Once you sign in and finish onboarding, a "Sync this computer to Penzor Cloud" card appears with all three checked; click Start syncing to turn them on. If you're already signed in, you'll see the card once after updating.

  • When you sign in on another computer and the cloud already has a vault with files that differ from this one, vault sync waits for you on the Sync page: Merge both, Use the cloud copy (this computer's differing files move to a backup folder next to the vault first, nothing is deleted), or Don't sync notes on this computer yet. If you run out of space, smaller files go first, files that don't fit stay on this computer, and you're told once.

• Guide: a new Guide Space is Forsion's manual. Chapters unlock level by level, the tasks are things you do anyway, and whatever you've done already counts. Each chapter has a diagram under its intro that shows where to look. Take me there switches to the right place and rings the spot to click on the actual screen; you click it yourself, and the ring moves on to the next step. Ask Tangu in the right panel is Tangu's own chat, with the current chapter attached above the input as a quote. Each new level brings an inbox letter with things to try next.

• Me: the Agents list has a new Me row at the top. It opens the profile all your agents share about you, in three sections: Collaboration, Portrait and Memory. Collaboration shows what you ask of every agent on top and what you and your agents have agreed on below. You can edit each one directly, with a change history and undo. Collaboration update cards in a chat take you there with View handbook or Revise, and an agent's Collaboration tab links to it as well. With Agents & memory sync on, this profile syncs across your computers, and if two computers change it, the edits are merged line by line first. Schedules your agents set up sync too, and each automatic task still runs only on the computer that scheduled it.

• Home:

  • New theme backgrounds: Orbit, Sun path, Paper and Glyphs, and Forsion rings is back. Orbit and Sun path drift slowly; turn the motion off in the wallpaper panel or in Settings → Appearance → Home wallpaper motion. Glyphs fills the screen with faint characters that ripple out from the cursor as you type, and also when you click an empty spot or send a message.

  • Recent-activity cards now sit below the input box (desktop): 4 by default, 1 to 6 from the right-click menu. The calendar card shows your next event and the inbox card your unread count, and the greeting and date share one line.

  • The achievements panel is split into Cloud achievements and On this device; cloud achievements add up across your account and come with rewards to claim. When an achievement is ready to claim, the trophy on the Ribbon shows a dot (on "⋯" when the trophy is tucked in there).

  • The Hack2Gate poster is gone, and Images only uses the picture you upload.

  • After you open another view from Home, switching away and back, or clicking the Home icon again, returns to Home itself; other tabs stay open in the background.

• Colors: theme colors are now layered by role. Buttons and actions use the theme color, cues such as selection, focus and progress use the highlight color, and hover is always a neutral gray. Settings → Appearance → Colors has a new Highlight row: follow the theme color, pick one of six built-in colors, or choose your own. The Coral, Lavender and Clear blue backgrounds were redone as Clay, Dusk and Mist (if you had one selected, you get its new version), and Sage is new. In light mode, primary buttons in Classic and Ink are one step lighter.

• Interface:

  • Switching Spaces has a transition: click the Ribbon and the workspace slides up or down in the Ribbon's direction while the Ribbon itself stays put. With Reduce motion on, it just fades. Not on phones yet.

  • One switcher for view modes: Chat / Canvas, Note / Canvas, the calendar's Day / Week / 3 days / Month and similar switches all sit at the far left of the view's top row. ⌥⌘V (Ctrl+Alt+V on Windows) moves to the next mode, and "View mode" is searchable in the command palette. In a narrow view only the icons remain, and narrower still it turns into a dropdown. Code Studio, workspace files and the file preview window use it too.

  • The Agent Desk and Mini cards have floating capsules instead of top bars. With the Agent Desk on the right expanded, two capsules sit on top: the item name on the left, which also switches items, and the view buttons, "⋯" and Clear / Collapse on the right; content scrolls underneath. A Mini card's top bar now matches the phone: the Space and a title dropdown (to switch chats or notes) on the left, New, Show in main window and Close on the right.

• The right panel follows the main view: whichever kind of view is in front in the main area, the right panel switches to the set that goes with it: details, files and outline for a chat; chat, outline, backlinks and graph for a note; chat for PDFs, images, whiteboards and the like; to-dos for the calendar. Opening another file of the same kind leaves the right panel alone. Views you put in the right panel yourself and ones the Space pins are never closed, and ones you close don't come back after a restart.

• Calendar and to-dos:

  • The calendar has a new layout: a mini month and your list of calendars on the left (click a color square to show or hide that calendar), to-dos on the right. Events are light blocks with a colored line on the left, and past ones fade. Calendar Spaces whose layout you never changed switch to the new one; customized layouts stay as they are.

  • Click an empty slot and type a title right there to create a one-hour event; press and drag to set its length, or drag across the all-day row for an event spanning several days. Dragged events snap to 15 minutes, ⌥-drag makes a copy, and moving or deleting an event can be undone with ⌘Z. Shortcuts: C for new, T for today, D / W / 3 / M to switch views.

  • Clicking an event opens an edit card beside it. Quick add for to-dos moved to the top and understands phrases like "tomorrow 3pm" or "next Monday", and you can drag a to-do straight into the time grid.

  • Events have a right-click menu (or press Shift+F10, or use Event actions on the edit card): copy, cut, clone (⌘D), paste, delete and copy summary, with undo and redo for the last 20 steps. Cloning an event that is a note copies the note's text too. Drag the top or bottom edge of a selected event to change its start or end time.

  • On phones, tapping an event opens a drawer at the bottom with its details, precise time editing and delete. Tapping near the edge of a short event used to count as a drag and open nothing; it opens the drawer now. Drag handles appear only once an event is selected, and the start and end times update live while you drag.

  • Opened as its own tab, the to-do view is the full version: filter by Today / Scheduled / All and by list at the top, choose which list a new to-do goes into and press N to add one, drag a to-do onto a group such as Tomorrow to change its date (undoable), move with ↑ / ↓ and check it off with Space. In the calendar's right panel and in Mini cards it's the compact version.

• Whiteboards:

  • A row of view buttons at the bottom: Grid (⌘'), Snap to objects (⌥S), View mode (⌥R), Zen mode (⌥Z) and Full screen, plus a pen mode button in front when you use a stylus. Turning on the grid also opens its settings: spacing, major line interval, direction (squares, horizontal lines only or vertical lines only) and strength. Whiteboards zoom to fit everything when they open. View mode's shortcut and right-click menu used to do nothing; they work now. On phones these are in the menu.

  • Drawing with a stylus keeps up better: fast strokes no longer drop points, and pressure lines up point by point. If your palm touches the screen mid-stroke, the stroke is no longer cut short or deleted, and you can start drawing with your palm already resting on the screen.

• Notes:

  • The @ mention panel has categories: All, Agents, Dates and Notes; switch with ← / →.

  • forsion:// links in a page open inside the app (in the editor, hold ⌘ / Ctrl while you click).

  • Typing or deleting no longer makes fold arrows twitch or the content below jump.

  • If you keep typing while a note or folder is being moved, your text no longer lands in a conflict copy in the old folder, and an old copy of the note no longer reappears where it used to be.

  • On the Cloud side, synced vault sections show which devices sync them ("This device" for this one) and come before the Cloud vault section. Right-click a top-level folder to show it as a synced vault, or remove a section left behind by older versions.

• Settings:

  • Settings → Models is reorganized into three parts. Default models are grouped into chat, image, voice and translation, each showing where its value comes from, with one click to go back to following it; read-aloud, voice call and voice input models moved here too. Model library lists every model in six types, and cloud models can be hidden on this computer. Under Providers, you can set each model's type and Text only flag, or type in a model name.

  • Web search is now an ordered list: connected search sources are tried in the order you set. Drag to reorder and test each one; a key is saved only once its test passes. DeepSeek is a new search source (bring your own key).

  • New Settings → Devices, which opens from the device pill on the Ribbon: switch between this computer, the cloud and other devices, and turn on Allow other devices to connect, all in one place. Paired devices are now called Visiting devices and are folded away by default. Computer Use moved into the AI group next to Browser control, and searching "strict background" finds it.

  • In Settings → Shortcuts, recording a key that another command already uses shows a note under the row, and nothing changes until you click Reassign here; the other command no longer loses its key silently. System-reserved keys such as ⌘Q or copy and paste are blocked, and keys the note editor would take first (⌘B for bold, for example) can be saved with a warning. The Keys toggle in the search box lists the commands that use whatever key you press. Clashes that already exist show as "N conflicts" in the toolbar and can be filtered. Global shortcuts, such as the one for the Mini window, pause while you record so they can't grab the keys.

  • Settings → Forsion Cloud is regrouped: the sidebar now has Account, Membership & usage, Backpack, Submissions, and Help & diagnostics, with tabs under Account, Membership & usage, and Help & diagnostics. Every page can be found in Settings search.

• Git: the project's Git tab has a new Remotes card that lists every remote, which account it uses and where pushes go now. Each row offers Push here from now on and Push here once, and the Push button names where it pushes. The three publish buttons are now one Add remote… dialog, which also works when the project already has a remote: new remotes are named after their site, and the existing origin is left alone. Remotes that aren't called origin are recognized too. When the account Forsion brings is turned away (no permission, or repository not found), the push or pull is tried once more with this computer's own Git credentials. In Settings, Accounts shows which account each site uses, Commit identity has its own panel that marks which level is in use, and the old Git panel is now called Commits and pushes.

• Store and plugins:

  • A plugin's page in the Store has Overview, Comments and Versions. You can comment and reply, and authors get comments in their inbox. Forsion version and platform are checked before installing. In a chat, you can ask Tangu to search the Store; results come as plugin cards that install with one click.

  • The Submissions page adds a detailed description, tags, Forsion requirements and supported platforms, and plugin names may repeat.

  • A plugin's details page lists the capabilities it asks for (such as "read and write your databases, read and write @ tasks in notes, read the foreground window"). In Settings → Hooks, hooks that come with a plugin are marked "From plugin · observe only".

• Open files from your system: PDF and Markdown files can be opened with Forsion (Open With in Finder, dropping onto the Dock icon, or double-clicking on Windows). Files outside your notes vault open in a temporary tab for preview; Markdown can be edited there and is saved back to the original file without being imported into the vault. Files inside the vault open as notes as usual.

• The Chrome extension is now Forsion Clip:

  • Tangu for Chrome has been merged into the Forsion Clip web clipper, so one extension does both: clip pages into your vault, and let Tangu keep working in the Chrome you use, in its own "Tangu" tab group in the background. The extension ID and folder are unchanged: after updating, click Reload on Chrome's extensions page (or restart Chrome) and your existing pairing is kept.

  • Installing: in Settings → Browser control, click Open extensions in Chrome and turn on Developer mode, then drag the Forsion Clip card from Settings onto the extensions page. You can also copy the path or open the extension folder and load it by hand.

  • Pairing: once installed, Clip asks to connect and Forsion shows a prompt. Click Allow and turn on to set up clipping and browser control in one go (this also turns on Allow other devices to connect); if you'd rather not, you can still paste the connection code by hand. The Settings card shows whether Clip is installed and connected, and which version. When a Forsion update brings a newer Clip, a connected Clip switches to it on its own (keep Developer mode on in Chrome).

• Faster: long chats open much faster and no longer stutter while an answer streams in. Opening or closing the sidebar on a long note takes a fraction of a second instead of over a second, and typing is smoother. Code blocks are highlighted only when they scroll into view.

• Terms updates: when the Terms of Service or Privacy Policy change, a card after you sign in explains what changed. Usually you just click Got it; when you need to agree again, you choose Accept or Decline and sign out. Once you've confirmed on one device, the others don't ask again. Windows that pop up on their own (terms, cloud sync, What's new, plugin announcements) now appear one at a time, and wait while onboarding is open.

• Updates:

  • When a new version is available, a two-column What's new window opens: versions on the left (earlier versions included) and that version's post from the website's release notes on the right. If the website can't be reached, the changelog is shown instead. Open it anytime from Settings → About. While the app is open, Forsion also checks for updates every 6 hours and after the computer wakes from sleep.

  • A new version is offered only once it is on the website's download page. In mainland China the installer downloads from the website by default, elsewhere from GitHub, and it switches to the other one if a download stalls; you can also switch while downloading, or pin one in Settings → About → Download from. On macOS, Download installer starts this version's .dmg download in your browser instead of opening the Releases page.

• Android:

  • Quitting while you're in Tangu no longer loses your open tabs on the next launch, and when Start in is set to another Space, that Space no longer opens with the previous Space's tabs and panels.

  • On tablets and on phones in landscape, Spaces move to a full-height bar on the left: Home comes first, pinned Spaces take priority when they don't all fit and the last slot is All, and your avatar sits at the bottom; in portrait the bottom Dock stays. The title capsule moves into the main area, and Home is centered on tablets. Whiteboards fill the whole screen height with their menu and toolbar in the capsule row, and with a finger the pan hint now says to drag or pinch with two fingers.

  • In the image viewer, pinch to zoom, swipe left or right to switch images and drag down to close; system back closes it too. Downloaded images are saved to Downloads.

  • The "⋯" in the top bar shows a dot when there's an achievement to claim, and the dot on Tangu also counts new messages and waiting chats with Muse.

  • The view mode switcher sits between the two top-bar capsules (with the native top bar, it's under View mode in "⋯"), and the notes bottom bar no longer has a canvas button.

  • Standalone whiteboards and databases in subfolders no longer show "File missing" when opened.

  • Installing a plugin from the Store that doesn't support phones, or needs a newer app version, now tells you why instead of showing an English error code.

• Windows:

  • Quickly dragging a Mini card's edge or corner to resize it no longer sometimes does nothing.

  • Mini cards now animate when they fold against the screen edge and when they expand, no longer show square corners after a theme change, and stay on their own monitor when folded against an edge shared with another monitor.

  • Note paths are now written one way, so the same note is no longer treated as two: tab following, unsaved drafts, pins and calendar settings all line up, and records from older versions convert automatically.

  • Approvals: in Ask mode, only read-only cmd forms such as dir and type skip approval, and every other command needs approval (git status too). Muse now reads its own logs and files with a file-reading tool that needs no approval, instead of queuing a string of approval cards.

• Web (takes effect once the web version is updated):

  • Opened in a phone or tablet browser, the web version now looks and works like the Android app: a bottom Dock in portrait and a bar on the left on tablets in landscape. An iPad opened or refreshed in landscape gets this layout too, instead of the desktop one.

  • Your Forsion account page, the Store (membership, points, items and plugin submissions), announcements and Penzor Cloud are now on the web, and Penzor Cloud can browse, upload and download files.

• Plugins (for authors): the forsion-plugin skill is now 1.59.0.

  • Views and UI: a plugin view can close itself from its own content with view.close() (the same API for desktop floating panels and Web overlays); ctx.ui.mountSettingsPage / mountSettingsItem draw settings pages like the host's from data, laid out by control on phones; ctx.ui.setViewModes puts a view mode switcher in a plugin view's top row; ctx.registerToolResultRenderer gives your engine half's tools native result cards that can move between the chat and the Agent Desk; ctx.ui.contextMenu adds opts.keepOpen and a handle with update(items), so checklists can be ticked several times in a row, and long menus scroll (up to 200 items); ctx.ui.mountModelPicker and per-call model choice are new too; a setting with secret: true shows as a password field (masked, not encrypted).

  • Layout and Ribbon: rightCompanions on registerView declares the right panel set a view wants when it's in front; ctx.app.setSpaceDot / setSpaceProgress put a dot or a progress ring on your Space in the Ribbon; ctx.spaces.mountIcon follows the Space, so re-registering the same id with a new icon redraws it in place; ctx.closeView(id, { collapse: true }) collapses a whole side; ctx.app.queuePopup puts your own pop-ups in the host's queue; an item in a Space recipe's layout.main can set "mainPage": true to be the page you land on in that Space.

  • Data and events: ctx.databases / ctx.noteTasks read and write databases and @ tasks in notes; the calendar tier of ctx.calendar (overlay calendar sources, calendar settings, .ics management), ctx.ui.mountRecord, openNote jumping to a heading, and notifications with action buttons; ctx.app.onVaultEvent reports creates, deletes, renames and outside changes in the vault; a command receipt can carry a result for the model; "syncData": true in the manifest syncs plugin data with the account across desktop and phone; plus ctx.spaces, ctx.tangu.mountComposer, ctx.app.openSpace / openLink / watchAccountName, ctx.wallpaper.bing and ctx.achievements.

  • Engine plugins: the manifest can declare a bundled stdio MCP server (mcpServers) and observe-only hooks that never block, started and stopped with the plugin; the tool context exposes presentDesk / runId; a plugin's details page lists the capabilities its manifest declares.

• Fixes:

  • DeepSeek V4.1 Flash couldn't see images you sent (they were turned into a text description first).

  • In long chats and long notes, jumping from the outline or a search result got pulled back to the bottom or landed in the wrong place.

  • Editing an image without an attachment made the agent retry the same error over and over; it now says that a path isn't an attachment and asks you to attach the image again.

  • The pointer over an inbox letter's body was always a grab hand, and inbox list subtitles showed backslashes, as in run\_bash.

  • On macOS, shortcuts with ⌥ failed because ⌥ changed the character typed; they now go by key position, and modifier keys in shortcut hints follow the system menu order (⌥⇧⌘).

  • View on trace notes, evolution update cards and notifications only opened the top of Tangu's details; it now scrolls to the matching entry and highlights it briefly.

  • With the calendar open in a tab, switching to another tab and back made the week view jump to August and the month view to a different month.

  • In the English interface, calendar event cards and the new-event card showed dates in Chinese.

  • English chats were sometimes given Chinese titles and summaries.

  • In notes, pressing ← or Backspace at the start of a line inside a fold block split the text out of the block, so collapsing it only hid the paragraphs before.

• Reference apps and windows: press ⌘ twice in any app to send the current window into a chat or note as a reference, with a screenshot of that window attached (this needs Screen Recording permission; without it only the reference is sent). It works while another app is in front, and the reference goes to the cursor, then a visible input card, then a new chat. Change the shortcut to double ⇧, double ⌥ or a regular key combination in Settings → Shortcuts.

• One selection menu: selecting text in notes, PDFs and chats brings up the same menu, with the current view's own tools first; you can change the order in Settings. Translations show as a card and use machine translation by default.

• Translation model: Settings → Models has a new Translation row under the default models for a dedicated translation model (Forsion Cloud provides machine translation models). Leave it empty to use the current chat model.

• Peek at a Space: rest the pointer on a Space icon in the left bar to bring up that Space's left panel and use it directly. Items you open there open in a new tab next to the current one, without switching Space. Items that open a view can be dragged to the tab bar or the main area.

• New chat:

  • The bar above the input card: the project pill is aligned, the opening note joins this bar, and you can add extra folders to the chat.

  • The run location pill: choose whether the chat runs on this computer or on another computer in Devices; offline devices are folded away by default.

  • A new chat opened from Notes lands in the notes vault when no project is chosen. A new chat's model no longer gets stuck on the default helper model.

• Talk to Muse directly: you can chat with Muse in its own chat, and it notes down what you ask of it first. When the Agent Desk on the right is idle, it shows a small Muse Space card.

• Cards in chat: whether a card appears depends on what the answer contains, and the card is the answer. The setting for how many cards appear moved to Settings → Models.

• Prompt suggestions are on by default and stay until the next turn starts. The switch is in Settings → Advanced → UI & sessions.

• An agent's Growth page is split into Memory, Log and Evolution, with the log folded by day.

• Git: push and pull with your linked GitHub account; Settings has a new Accounts page and a commit identity; connect an existing remote, publish a project to GitHub, or clone a project from a Git URL.

• Notes:

  • File wikilinks open in a split beside the current note; right-click a file link for the file menu; embedded files show their source when the cursor enters them, as in Obsidian; wikilinks have a new look; images and videos open inside the app.

  • Selecting across lines by dragging is more accurate; switching tabs keeps the scroll position; ⌘-click opens a new tab; clicking in the gap between blocks puts the cursor in the nearest block.

• Projects: both AGENTS.md and CLAUDE.md in the same folder are read; chats in the notes vault show project details too.

• Inbox: message rows have two lines and the reading pane shows the sender on top; messages can be grouped and folded by sender; action buttons appear at the end of a row on hover.

• Plugin files are written to Plugins/<plugin name> in the notes vault by default.

• Cloud read-aloud: when none of your own providers has a read-aloud model, read-aloud goes through Forsion Cloud.

• First-run guide: Computer History is now part of it (off by default; you read and confirm what it records before turning it on). The flow, layout and motion were reworked, and you can click back to any step you have visited. After an update you only see the steps that are new in this version.

• Android:

  • Home no longer shows the previous Space's list. System back: the first level goes to Home, and back on Home once more leaves the app; pages opened from a list go back to the list; an edge back swipe moves the page with your finger (Android 14 and later).

  • Input bars and toolbars at the bottom rise and fall with the keyboard; Home cards can be edited with a long press; multi-select sheets tick in place; swiping a row and letting go spring back naturally; presses show a ripple.

  • The inbox uses the phone layout; notification cards sit below the top capsules; the item in "⋯" is called Command palette again.

• Windows: attachment links follow a note when you move it, so images no longer break; Mini cards fold against the screen edge when dragged there; project details show instruction files as paths inside the project; fixed escaping in a few Git settings.

• Computer Use updated to 0.7.1: double ⌘, double ⇧ and double ⌥ gestures (used by app references) and capturing a single window.

• Plugins (for authors): the text selection menu has a plugin interface; ctx.tangu.translate is new; a bundle with an engine plugin reaches its engine plugin even when its install folder name differs from the manifest id (2.13.3 and earlier required them to match); plugin files go to Plugins/<plugin name> by default.

• Fix: in a product made of plugins only, the Space you were last in turned blank after a reload.

• Profile: Settings → Forsion Cloud has a new Profile page. It is the same profile as in the web account center. Change your photo, nickname, bio and links in place, feature your works, change the cover and avatar frame, and make the profile public at a web address if you like. Your email, phone number, membership expiry, points and quota never appear on it.

• Cards in chat show up on their own: answers such as an itinerary, a checklist, a pick-one choice or a list of steps now come as cards even when you did not ask for one, and actions inside a card have transitions. When a card's content occasionally lands in the reply as raw text, it is corrected automatically.

• Quotes and Ask in sidebar: a message can carry several quotes, and they stay as cards after you send it. Ask in sidebar on selected text now opens a side chat in the right sidebar instead of interrupting the current chat.

• Mini cards can be resized: drag an edge or a corner, from 260×160 up to 640×900. A card reopens at the size you last used.

• PDF: with a plugin that takes over PDFs installed (Better PDF, for example), PDF links in notes, links with a page number and quotes in chat all open in the plugin, following Default open with.

• Notes: the hover preview card closes after you follow a wikilink to another page. [[ references in the chat input fall back to fuzzy matching when no name matches exactly.

• Interface: all dashed borders are gone. Dragging a file into the chat area tints the whole area and lights up the input box; the Files panel, Image Studio and placeholder buttons follow the same treatment.

• Android:

  • Opening a chat, going back, swiping back and switching Spaces are smoother; long chats render in batches.

  • Home is a single page: it has no tabs and always opens on Home itself; chats and notes you open from Home go to their own Space.

• Computer Use updated to 0.6.3:

  • Apps that can only be driven from the screenshot, such as WeChat: results now state that coordinates are pixels in the screenshot; scrolling and dragging by coordinates use real input, and scroll distance uses the same unit as coordinates; with Mos installed, scroll direction and distance are no longer altered.

  • When a window is on another desktop, the error explains why and how to bring it back.

  • Strict background, Show what the agent is doing, Managed browser and Browser tools in the plugin settings used to have no effect when changed; they now apply.

• Plugins (for authors): ctx.ui.mountMarkdownEditor has a new modes: false option that hides the Edit / Source / Preview row; a file type that takes over .pdf receives opens with a page number or quote through file.onGoto; Mini panel layouts should follow their container instead of assuming 320×420.

• Fix: after a truncated emoji appeared in a chat, every later message in that chat failed to send.

• The new tab page is organized by what you want to do: search, recently used, new, open (one tile per Space) and the sidebar panels of the current Space. Clicking another Space's tile opens its view in the current tab; the browser and the terminal open in place too.

• Visuals in chat: charts and small tool cards in a reply render while they are being generated, and buttons inside a card can go back and change the answer. The agent can also return native document cards, calendar and to-do cards, and app cards. Choose how many cards you get in Settings.

• Voice calls can go through Forsion cloud: once you sign in to your Forsion account, you can make voice calls without configuring a call model yourself; calls are charged by duration against your membership quota. The mobile app can make calls too, with the call shown as a bar at the top of the page while chat keeps working. Requires a call model enabled on the server.

• Prompt suggestions (off by default): after a turn ends, the empty input box shows a likely next message in gray; press Tab to accept it. Turn it on in Settings → Appearance. Desktop only.

• The agent can tell how much is left: ask it how much room this conversation has left, or about your Forsion account's quota, membership and points. Anything that acts on your account asks for your confirmation every time.

• The agent tells you what it noted, right away: when the agent writes a memory or an evolution record, a receipt line appears under the reply; changes to the collaboration notes come as a separate request card. What the background review writes is attached after the reply it belongs to, and items left for you to confirm can be accepted or discarded in place. A new chat mentions above the input box what was noted in the background since last time.

• Collaboration notes are now written for you to read: each item is something you do, in words you have used. Notes written by an older version show “Have the agent rewrite it” next to them; your corrections go to memory and no longer change the notes.

• Project details have a new Pull action. Once Forsion Git is enabled on the server, you can also publish a project to it in one step; the app fetches the push and pull credentials for you.

• Support for the Claude 5.5 generation: Haiku 5.5 is recognized; Sonnet 5.5 cannot turn thinking off, so choosing Off falls back to the lowest level.

• Screenshots and images:

  • Computer Use no longer resends every old screenshot each turn and keeps only the latest 3; large images are downscaled before they enter the conversation. Long tasks run faster and are interrupted less often.

  • When the main model has no image input, the interface says so while a screenshot is being transcribed; Computer Use is unavailable in that case and explains why.

• macOS: detached windows and Space windows no longer have the empty title strip above the tab bar. The installer window's background shows what to do if the app will not open. An app icon you chose stays after you quit.

• Windows: the default splash no longer flashes by. An app icon you chose stays on the shortcut after you quit.

• Android:

  • The top bar and the Dock are now floating glass capsules. The Dock has five slots with the icon above the name; full-screen lists and the home page use a one-handed layout: one main button plus gestures, and swipe actions show icons only.

  • Chat page: the page no longer jumps when the keyboard opens; the send button moved to the right of the input box and modes moved into the model menu; wide tables scroll sideways inside their own frame.

  • Spaces bundled with a plugin now appear; a plugin's own files read and write correctly in a cloud vault; plugins that only support desktop are marked Desktop only in the Store.

  • The home screen icon follows the icon chosen in the app.

• Plugins:

  • When an installed plugin's Space does not appear, the plugin card and the Sandbox panel say why.

  • Open settings on an installed plugin in the Store goes straight to that plugin's own settings page.

  • For authors: ctx.ui.mountBoard / boardToSvg add a whiteboard handwriting layer over a plugin's own content; list sources support hierarchy, dragging and renaming.

• Notes:

  • After you move a note or a folder elsewhere, or move blocks across folders, the image and attachment references inside are recalculated and no longer break.

  • The trash of a cloud vault now lists deleted files and can restore them.

  • Saving in a cloud vault no longer corrupts image addresses; notes already saved that way display correctly when opened.

  • Database column widths fit their content more accurately; the first click into a long note no longer stutters.

• Fixes:

  • An unsent draft in the input box is kept per conversation and no longer carries over when you switch conversations or start a new chat.

  • In the Inbox, a reply that arrived after an approval only lit the red dot and did not appear in the list.

  • With Settings open on another page, jumping to the same location again from elsewhere did nothing.

  • In dark mode, the input card of the sidebar chat did not read as a card.

  • With interface zoom other than 100%, the whiteboard layer did not line up with the content under it.

  • Icon buttons on the Inbox reading page and the Button block in notes never had their styles applied.

  • On the web and mobile, sending a message more than 30 minutes after a conversation started failed. Takes effect after the server is updated.

  • When a web fetch only gets a page title, the agent says so and switches to the browser.

  • Feedback attachments include the whole conversation instead of only the latest 500 messages.

• The default Ribbon order is now Tangu → Notes → Calendar → Inbox → Agents → Coding → Images.

• Smaller local conversation database: intermediate streaming data of finished conversations is kept for 7 days only.

• The Market is now the Store. Agents and Spaces are listed under Plugins. A new Membership & points section in the sidebar works once you sign in to your Forsion account: it shows your current plan and points balance, and lets you pick a plan and pay, buy point packs and redeem items with points.

• Dock a chat beside an app (macOS). Run “Dock a chat beside an app” from the command palette and pick a window on screen. The chat panel docks beside it and follows when you drag, move or resize that window. Messages in the panel reference the app by default, so the agent can look at it and control it directly.

• Computer History now works on Windows. Once you turn it on in Settings, it records app switches, window titles, browser URLs and typed text. Password fields, password managers and private windows are never recorded.

• Windows: dragging a Space icon off the Ribbon no longer opens a window when you press Esc or release it elsewhere. On machines without hardware acceleration or with slow rendering, the “Tree shadow” splash shows a static scene instead of stuttering.

• The current Space is no longer highlighted in the Ribbon's recently used group. Only its icon on the bar marks where you are.

• Fixed: the web version always failed its connection check, which left the input box disabled.

• The built-in Computer Use is updated to 0.6.2.

• A native shell on Android. The top bar, bottom navigation bar and half-screen sheets now use native system controls: opening a Space shows its list first, tapping an item opens the main area, and Back returns to the list. The avatar menu switches between local and cloud, switches Units and opens Settings. You can also install and run external plugins on your phone. Other phone improvements: badges and haptic feedback on the bottom bar, a long-press menu on messages, voice input, Share to Forsion from other apps, the + button split into Camera, Photos and Files, dropdowns shown as half-screen sheets, times on session rows and a clearer offline message. System notifications tell you when a background task finishes or needs your approval, and command confirmations can be denied or allowed right on the notification.

• Open a whole Space in its own window. Right-click it on the Ribbon, ⌘/Ctrl-click it, or drag its icon off the Ribbon. Each Space gets one window, and its position and size are kept across restarts. The middle of the Ribbon now lists recently used Spaces; set how many in Settings → Appearance, or turn the list off.

• A new default splash screen, Tree shadow. Window light and branch shadows on a wall with a line of Fusang poetry, in daylight for light mode and moonlight for dark mode, with the version number in the lower left. The earlier tree-mark animation is still there as Splash media → Classic icon. You can also use your own splash media and app icon in Settings → Appearance → Splash & icon. On Windows, starting in dark mode no longer flashes a light background first.

• Each Space can have its own appearance. In Settings → Space → Space appearance, set the design language, accent color, background color and light or dark mode for each Space. Tinting the shell from content is now off by default and can be turned on in Settings.

• Interface: the layers of the dark theme have been retuned, and there is a new Ink color scheme. Primary buttons in dark mode are no longer a block of white. Two new appearance switches, Relaxed text and Gentle transitions, are on by default, and Dim surroundings is off by default. The Ribbon returns to its place after you scroll it, which you can turn off; its command area now keeps five items in view instead of four, and expanding or collapsing it is animated. The slash menu no longer shifts when you switch categories, and the model menu's submenu is no longer taken over by rows you pass on the way.

• Hand work to agents from a note: three new blocks — Page instructions, Agent task and Reusable prompt. Type @Agent to insert a task to hand off; it runs in the background while the note stays open. AI commands in the slash menu are grouped into Writing, Instructions, Tasks and Prompts.

• PDF: the built-in PDF reader is now read-only and reloads in place when the file changes. Plugins can take over .pdf files, and Settings lets you choose the default opener. Built-in PDF annotation is removed in this version and will come back as a plugin.

• Image Studio now uses a project-and-detail layout: project navigation on the left, a launchpad or the canvas in the main area, and the conversation on the right beside the properties.

• Links: web links in the app open in your system browser by default, and web citations in chat follow the same setting. You can switch back to the built-in browser in Settings.

• Voice: Qwen-Audio-Realtime is available for calls. You can choose the cloning model for read-aloud voices, and the app guides you through recording a sample and checks it locally. Fixed voice sample uploads timing out from outside mainland China.

• Memory and self-improvement:

  • Memory is split into project and global scopes. When a project's memory is full, the model tidies it before saving, and sentences that were merged or dropped can be restored one by one in the project details.

  • Working notes are now called the Evolution Record. Agents can write to it themselves, and you accept or discard risky entries one at a time. Agents put away tools and skills they rarely use, and Muse reviews usage weekly.

  • Background reviews start as soon as something goes wrong: you correct the agent, you stop it, tools fail repeatedly, or it says it finished when it did not.

  • Automatic recall in a project session no longer brings in what was said in other projects. Sessions for team members, discussions and sub-agents follow the project they belong to.

• Agents can read and change some settings on this device. Every change takes effect only after you confirm it. Credentials, approval levels and permission settings are not available to agents.

• Plan mode is no longer available in team mode. Its menu item is dimmed and says why.

• More reliable subscription sign-in: tokens renew automatically when they are used, including for image generation, image editing and read aloud.

• Cloud sync: agent-level collaboration notes (HUMAN.md) are now synced, and log sync no longer loses the cloud header. Syncing across devices takes effect once the server is updated.

• Windows: running commands no longer opens a console window, and no child processes are left behind after a timeout, a cancel or a stop.

• Plugins (for authors):

  • ctx.tangu.mountChat mounts the native conversation in a plugin's own view, and ctx.app.trash moves files in the vault to the trash.

  • A Space can declare pinned Views (pinned) and use its own icon file (iconFile).

  • Plugins can take over a file type with override: true. Only .pdf is open for now.

  • Mount APIs return a { render, dispose } handle. Extend View has a new close reason, layout, so layout changes are no longer reported as the user closing the view.

  • During voice calls the host provides the audio level through agentStatus.speechLevel, which plugins can use for lip sync.

  • Local plugin task APIs: ctx.tangu.request calls routes of an engine plugin in the same bundle, ctx.tangu.openSession opens a task session, ctx.registerScopedRoutes registers ownership-checked routes, ctx.paths.dataDir gives a stable data directory, and the runs APIs create and track background tasks. Tools provided by plugins can be called through the desktop app's outward MCP, on this device only.

• Unit: the command line can update a plugin while Unit is running. A new optional management interface shows status, registers release packages and runs updates; it is off by default and is turned on in unit.json.

• Fixed the layout being saved to the wrong place when startup temporarily fell back to another Space. A plugin Space set as the startup target is now opened when it finishes loading late.

• Fixed the height of plugin views opened directly in the bottom panel, and stopped a chosen skill's full text from being pasted into the conversation.

• Real-time voice calls, like GPT Live. When the input box is empty, the send button starts a call. The call runs in the Mini card with the agent's avatar, a hang-up button and a microphone picker. Hand work off to the agent during the call and it reads the result back when it's done. You can also type during a call and the text goes straight into it, and voice transcripts are corrected with what the real-time model actually heard. If the service drops the connection with an error, the call reconnects instead of ending, and results from work handed off during the reconnect aren't lost. Calls replace hands-free auto-send.

• Voice settings are split into Voice calls, Voice input and Read aloud, with separate voices for calls and for reading aloud.

• Notes: browse and restore earlier versions from ⋯ → Version history. Type footnotes directly, click a superscript to jump to its definition and hover to preview it. :emoji: shortcodes autocomplete as you type, and ⋯ → Copy as Markdown is new. Collapsible blocks can hold code, tables and other special content, and code blocks detect their language automatically.

• Canvas: objects snap and show guides when you drag or resize them near other objects. Shapes, frames and connections can be copied and pasted, and Mod+D duplicates them. Change the arrowheads on either end of a connection, drag an end to reconnect it, or drag a new connection out of a card's edge handle. Dragging a note from the sidebar onto the canvas adds an embedded card (hold Alt for a link card), and the context menu has Add note…. Canvases can be exported one way to JSON Canvas.

• Turn a conversation into a note: save a whole conversation as a note in one step, and add interactive blocks (```forsion-sketch) to notes. Charts and flowcharts in chat have a clearer look.

• A calmer interface: a lighter Ribbon, the status bar off by default, frame icons that light up only on hover, and an option to hide chat avatars. Scroll the Ribbon's Space and command areas to see more icons; scrolling follows your input and snaps when you stop, and … now opens on click. Spacing and type rhythm are tuned throughout, and message action rows show the time.

• Sessions and onboarding: Historian picks an emoji icon for each session, new sessions that haven't started hide the Agent Desk, and onboarding lets you choose the default background agent.

• Backpack and submissions: the avatar menu in the lower left has a new Backpack, where you can view and use items and vouchers from Forsion Cloud. You can submit plugins and manage your submissions right in the store, or from Settings → Forsion Cloud → Plugin submissions.

• Easier updates: core plugins check for updates together, the plugin market adds an npm channel, and each installed market plugin can update automatically. When an update is ready, a Restart to update entry stays on the Ribbon. If tasks are still running, Forsion asks whether to restart later or quit and update now.

• Plugins:

  • Plugins can declare required plugins. A plugin can't be turned on until the plugins it needs are installed and running. When a required plugin stops, the plugins that depend on it pause, and they resume when it comes back. Settings lists what's missing and links to it in the market.

  • Turning engine plugins on or off, updating them and uninstalling them no longer restarts the backend. A few packaging styles still need a restart, and Forsion says so when they do.

  • Plugin views can dock in the native bottom panel and switch between a launch layout and a project layout, like Coding Studio.

• More reliable agents:

  • Unattended runs such as Muse and automations no longer wait for answers or approvals that nobody is there to give.

  • After the engine restarts, interrupted tasks are no longer rerun from the start, and replies written before the interruption are kept. A normal exit waits for running tasks to wrap up.

  • Runs no longer retry once your subscription quota is used up, and Codex tells you when its quota resets.

  • Tool arguments with broken JSON are no longer run as empty arguments, and calls to tools that don't exist suggest similar tool names.

• Models: Codex direct connections support GPT-6.1 Sol.

• Computer Use 0.6.1: turning the plugin off and on repeatedly without restarting the engine no longer piles up process exit listeners.

• Fixed example messages in theme previews showing a date in 1970.

• New: HUMAN.md, a collaboration handbook between you and your agents. Write down how you like to work together in a readable, editable handbook with two levels: the agent level holds working preferences that apply across projects (for example, "When there is a design choice, give me two comparable options first"), and the project level holds only the goals, division of work and acceptance criteria of the current project (for example, "I do the visual review before this release"), stored in the project's .tangu/HUMAN.md (a HUMAN.md already in the project root or in .forsion/ is still read). When you give clear feedback, the agent updates the handbook itself, sparingly and with immediate effect, and leaves an update card in the chat that you can view, edit, undo or expand to see the reasoning; the card is still there after you reopen the session. The "Collaboration" tab on agent details and project details lets you edit the handbook directly, read the raw text, see the file location and browse the change history (the last 40 versions, kept on this device and never written into the project repository).

• Handbook boundaries: what you say right now always takes priority; the handbook only describes how to work together and grants agents no tool permissions; saving is version-checked, so it never silently overwrites your manual edits (on a conflict your draft is kept and you review the latest version before merging); undo only applies to an update that is still the latest version; and agents that share memory do not share each other's handbooks.

• New thinking level: Ultra. The thinking-level slider gets an eighth notch after Max (you can also type /think ultra): thinking is always at its highest, context goes up to the model's limit, and the agent proactively hands work that can run in parallel to several sub-agents at once. It is available only in local work sessions; web, Chat mode, external engines and team sessions don't have it. Each time you turn it on, a dialog explains that token use can be several times the usual and offers "Also switch this session to Full access" (unchecked by default, since parallel sub-agents each wait for your approval) and "Don't show this again". Sub-agent usage also counts toward the per-run cost cap.

• Approvals now wait in a tray above the input box, and the agent keeps working while you decide. Pending confirmations stack above the input box: the oldest is expanded and ready to approve, the rest are listed as rows (click a row to expand it), and the tray can be collapsed; the chat itself keeps a single line such as "N actions are waiting for your approval". When an action needs your approval, the agent parks it, carries on with the steps that don't depend on it, and runs it at the next step once you approve; it only stops and waits when it has nothing else to do. The agent's questions and plan approvals use the tray too (the plan text stays in the chat, the decision buttons are in the tray). After you move to the next card, its buttons are briefly on cooldown so a double click can't approve the next item by mistake. Sub-agents, external engines, the terminal TUI, message channels and Muse still wait for each approval in turn.

• You can keep sending commands and messages while a task is running, and they queue. /compact, /refine and the messages sent after them wait in the area above the input box and run one after another when the current run ends (click × to remove one, ↑ to take it back into the input box); before, /compact during a run only showed "Compression failed". The terminal TUI and /retry in the desktop app queue the same way.

• New: Computer history (macOS only, off by default). Turn it on in Settings → Computer history to let Forsion record what you do on this Mac: app switches, window titles, URLs, changes to the text in the focused input field, and the names of controls you click and keyboard shortcuts you press. It never records raw keystrokes or screenshots, and password managers, Keychain, secure input fields and private windows are excluded automatically. Records stay on this Mac for 7 days; you can pause, exclude apps or sites, or clear them. Today's records show as a 20-minute timeline with real app icons, and you can browse the last 7 days. Agents read them through the read_computer_history tool (visible only when the feature is on, and unavailable in remote sessions, message channels and teams), and Muse also looks at a summary of the last few hours. Clearing the records does not delete conversations or Muse journal entries that already quoted them. Requires Computer Use 0.6 and the macOS Accessibility permission.

• Tangu can work in the Chrome you are already using, and there is a new Tangu for Chrome extension. Agents can now list, read and drive the tabs you already have open in Chrome, Edge or Brave (new browser_tabs tool; it needs remote debugging turned on in the browser, which asks "Allow remote debugging?" once per new connection). With the extension installed (Settings → Browser has three installation steps), reading, clicking and typing happen in background tabs without taking focus or showing the debugging banner, pages Tangu opens itself go into an orange "Tangu" tab group and need no approval, and you pair only once. Clicks and typing in your own tabs still need approval, and unattended runs such as Muse and automations never touch your browser.

• Documents work out of the box. The desktop app now ships Python with 15 pinned libraries for office and data work (python-docx, python-pptx, openpyxl, pandas, matplotlib, reportlab and more), so agents no longer run pip install on the spot when handling docx, pptx, xlsx and PDF files. It also bundles a LibreOffice conversion engine, so read_document reads doc, docx, xls, xlsx, ppt, pptx, odt, ods and odp by converting to PDF first and parsing real pages, with no separate LibreOffice install. Git is bundled on Windows and macOS too (Linux still uses the system git), and Windows includes the VC++ runtime the engine needs.

• Project details can now run Git actions: create a repository, commit (it lists every change first, then writes the message with the session's model and lets you edit it), create a branch (with a prefilled prefix) and push. Settings → General → Git sets the branch prefix (default tangu/), commit message guidelines and whether rewriting remote history is allowed, and agents follow them when you ask them to create a branch or write a commit; agents never create a repository or commit on their own. If a repository's own configuration can run programs (filters, hooks and the like), Forsion doesn't read its Git status until you choose "Trust and continue".

• Agents can offer to turn what they build into a creation. Web pages and tools made in a chat used to have no way into Creations. Now an agent shows a card when it fits: "Make it a creation" creates a folder in your Creations location and moves the session there to carry on, and "Add to Creations" takes in an existing project in place without copying it (project details has "Add to Creations" too). In any Space, a creation now saves a version after each run (desktop only).

• Generative AI right in your notes. Select text and use "AI ▾" in the toolbar, type /ai, or press Space on an empty line (turn it on in Settings → Notes → Behavior) to improve writing, fix spelling and grammar, shorten, expand, summarize, translate, continue, or give a custom instruction. The result streams into a preview first and nothing is written until you confirm (and the write can be undone). The selection toolbar and the ⠿ block menu also have "Ask Tangu", which hands the selection to the side chat, and every assistant answer in a chat has "Insert into note" to put it back into a note in one step. When Tangu edits a note you have open, the changes are marked in the text and a bar at the bottom shows "Tangu changes: N" so you can review each one, revert all or keep them.

• Links and formatting in notes, a big batch of improvements (benchmarked against Notion and Obsidian): the backlinks panel lists each mention with its context and turns unlinked mentions into links in one click; [[ completion puts recently used notes first and supports #heading, #^block and aliases; #tags render as clickable pills with completion; renaming a note on web and mobile now rewrites [[links]] across the whole vault; ==highlight== and %%comment%% preview live, and <kbd>, <sub> and <sup> are editable in place; the note ⋯ menu adds full width, small text and a page font (default, serif or monospace), and you can lock a page; cursor and scroll position and fold state are remembered per note on this device.

• Editing and canvas improvements: in-page find supports replace, replace all, match case and regular expressions (source mode too); tables support continuous editing with Enter, Tab and Shift+Enter, and the block menu has a "Table" section; Cmd/Ctrl+D duplicates a block, Shift+↑↓ extends the selection by block, and Alt+Enter follows the link at the cursor; the status bar now counts visible text only and adds word and selection counts. On the canvas, cards, shapes, frames and connectors can have colors, dragging to the edge of the view pans automatically, and holding Space pans while Cmd/Ctrl + =/-/0 zooms. A batch of save and sync fixes for problems that could lose or overwrite content also landed.

• Muse and background agents are easier on you: Muse decides on its own how long to sleep based on your routine (any activity from you wakes it, and rule-triggered wake-ups and due schedule items still wake it); Muse and automations that use the cloud default background model have their own background quota, so they no longer eat into your chat quota, and when it runs out they pause by default unless you choose to continue on the main quota (requires server 2.3.21); each Muse cycle now starts a new session, so the opening context no longer snowballs from cycle to cycle and each cycle costs fewer tokens; the Muse Space previously failed to render and now does; and Spaces that agents build for themselves can read and update their own data (to-dos, schedule, Library) and render from it.

• The Automations Space is now a node canvas. First-run templates (daily reminder, agent briefing, notify on a new table row, one-click agent) get you started; the canvas can be dragged, panned, zoomed and auto-arranged, and each node edits a trigger or a notification / agent / database / tool step; new rules are saved paused; notification titles and bodies can insert date, time and trigger-row variables; each run has an expandable ledger with step summaries and full errors (a trial run really executes); and deleting an automation can be undone for 5 seconds instead of asking for confirmation.

• Context ring details: click the context ring in the input box to see a color-coded segmented bar (the striped end is the space reserved for auto-compaction), how much room is left before auto-compaction and a "Compact now" button; click the header to expand a breakdown (history, tools and this turn, system prompt, memory, project and files, skills and more). When signed in to Forsion, a "Forsion quota" section at the bottom shows today's and this week's usage as percentages, a countdown to the daily reset at 00:00 Beijing time and the weekly reset time, and turns to a warning color when 15% or less is left.

• Images, audio and video show up directly in chat: a paragraph in an assistant message that contains only ![[…]] is rendered in place as an image or a media player, for both vault-relative paths and absolute paths outside the vault (local desktop sessions only; files over 50 MB fall back to a reference bar).

• Tangu can operate your Android phone (experimental). After you turn on "Allow Tangu to operate this phone" in Settings → Advanced → Experimental features in the Android app (and confirm in the system dialog), Tangu can use system capabilities in conversations started on the phone, such as opening apps and maps; it asks once more before opening apps other than the browser and maps. Only tasks started on the phone itself can drive the phone, accessibility-based screen reading and tapping is not included yet, and remote drivers such as the computer, WeChat or Muse are not supported.

• Settings has a new top-level page, "Forsion Cloud", so you can take care of your account without leaving the desktop app: Account (photo, nickname, linked phone and email, membership tier), Quota & points (today's and this week's quota, quota reset cards, points and top-ups, background quota, invite and redemption codes), Security (re-link, change password, sign out everywhere, delete account), Usage and Feedback. AI quota is shown only as the percentage left. These pages update with Forsion Extend.

• Announcements are new, and automatic points deduction is retired. When Forsion posts an important notice, a card pops up when you open the desktop app and doesn't come back once you choose "Got it" or close it; links in announcements open in your browser. Using a quota reset card now plays a quota-restored animation. "Pay with points automatically" is gone: when your AI quota runs out, points are no longer spent to keep going, and points can still be exchanged for reset cards (updates with Forsion Extend 0.7; requires server 2.3.26).

• Computer Use: fixed agents having no computer-use tools at all on a fresh Windows install (the bundled plugin was off by default and the helper lacked the VC++ runtime); the macOS helper is now signed with a fixed certificate so updates no longer drop its Accessibility permission, and after upgrading from an older version you need to grant Accessibility to tangu-computer-use once more in System Settings; the automatic Mini window no longer follows the cursor, now sits in the top-right corner where you can drag it, and returns to the Forsion window you started from when the run ends.

• The first-run guide has a "Local environment" step again, and Settings has a "Local environment" page. The guide lets you choose the official source or a mainland China mirror (recommended by time zone, system language or IP region) and checks local tools such as npm, git and docker; the page under Settings → General puts the developer tools list first and lets Tangu install what's missing. One-click install on Windows no longer hangs (install commands run non-interactively, and a download page is offered when there is no package manager).

• A large batch of interface polish from two rounds of UI/UX review: in Settings, the backend mode card only edits a draft and switching takes a button that states the consequences, instant toggles no longer overwrite settings you are editing, items that need a restart appear in a sticky save bar, and settings can be searched; "Restore default layout" can be undone; dates and times are formatted consistently and labels follow one wording and punctuation style; the Calendar and Creations no longer flash a skeleton on first open; in-app notifications can have a custom duration; the Inbox left column can be resized; the first day of the week in Calendar follows the language (Monday in Chinese, Sunday in English, adjustable in Calendar settings). After installing or uninstalling a plugin in the Settings window, the main window follows right away without a refresh, and a "Restart backend" button appears when one is needed.

• The changelog now has an English edition that follows the interface language (a version missing from the English file shows its Chinese text).

• A NoExtend desktop installer is now available on macOS, Windows and Linux alongside the standard installer with Forsion Extend. Both variants keep the same app identity and local data. NoExtend has its own update feed, so later updates stay on that variant.

• Permissions for Load in Forsion and desktop shortcuts now also check when the folder was created, so a folder deleted and recreated at the same path no longer inherits the earlier permission. Permissions saved by older versions lack this record and need to be confirmed once: opening such a shortcut asks you to choose Add to desktop again in Creations, and the Sandbox panel in the Coding Studio asks you to choose Load again.

• The device page (this computer's Forsion opened from another device) is safer. The following now only work on this computer itself: installing plugins; changing external engines, messaging channels, model providers, web search, hooks and approval rules; permanently deleting sessions or messages in a session; listing read-aloud voices; and emptying the notes trash or permanently deleting items from it. Trying them on the device page shows "This action is only available on that device itself, not over a remote connection". Archiving sessions still works.

• Tasks started from the device page run at most in the "Auto edit" approval mode by default, even when "Full auto" is selected. On the device page, approval cards can only be approved or rejected as is: the command can't be edited and "Always allow this session" isn't offered. Sessions from the device page can't use the filesystem root, the home folder or a folder holding protected configuration as their working folder.

• The host files the device page can browse are limited to the working folder, the notes vault and the project folders this computer already has; after the update, sessions whose folder was changed from the device page, and project sessions started by device-page tasks, no longer widen what can be read. A project you later open elsewhere on this computer by other means (for example from the command line) becomes browsable on the device page once you pick its folder with "Add local project" on this computer. Right after updating, the device page may briefly report that the engine isn't ready.

• Tasks started from the device page also can't create or change this computer's agents, skills, working notes, long-term memory, schedules or automations. Those changes affect every later conversation on this computer, so make them on the computer itself. Conversations driven from the device page aren't written into long-term memory by Historian either.

• On the device page, a session's approval mode can only be made stricter, not looser; a session that never had one can only be set to "Read only".

• Device-page sessions also can't use an app's own data folder as their working folder (~/Library on macOS, AppData on Windows, ~/.config on Linux and similar). Projects in cloud drives (Dropbox, OneDrive, Google Drive and others) and in iCloud Drive are unaffected.

• Agents can no longer read Forsion and Tangu configuration and secret files, including config.json, .env, MCP and model provider settings, the WeChat channel token, plugin settings, the agent browser's sign-in data and Forsion's app data folders. This applies to conversations on this computer too; reading them from the command line needs your approval.

• Shortcuts (symlinks) in an agent's private Library folder that point outside it can no longer be read.

• The local engine now uses its own local token instead of your Forsion account token.

• MCP: the server names dev and anything starting with dev_ are reserved for built-in developer tools. Existing servers with those names stop loading and show an error in their status; renaming them brings them back. Error and text results from MCP tools are marked as external data before they reach the model, and images they return are treated as untrusted content.

• This computer's device credentials are now encrypted. The pairing credentials behind "Allow other devices to connect" and the token for an external engine are now kept in the system keychain (macOS Keychain, Windows DPAPI, or GNOME Keyring or KWallet on Linux) instead of in plain text in a configuration file, and they no longer appear in the configuration the app's interface can read. They move over automatically the first time you start the new version.

• On macOS, after an update the system may ask for keychain access the first time Forsion reads these credentials. Choose "Always Allow". If you deny it, this computer won't switch to a new device identity and won't fall back to storing them in plain text; it only stops connecting to your other devices for now (direct LAN connections keep working). The system won't ask again while Forsion is running, so the notice offers "Restart Forsion": restart and choose "Always Allow" to recover. If the saved credentials can no longer be decrypted (for example, the keychain was reset), click "Retry", or confirm "Re-register this device" in the system dialog (your other devices will need to trust it again).

• If the token for an external engine can't be read, the external connection panel in Settings › Connection shows a notice. You can retry, restart, or enter the token again.

• On Linux without a usable system keyring (for example, no GNOME Keyring or KWallet installed), these credentials are stored as before and the device switcher shows a notice. Remote sessions can't be turned on in this state.

• What encryption covers: the Windows and Linux keychains aren't isolated per app, so other programs running as the same user may still read them. It protects against the credentials being read straight from a configuration file or carried off in a backup or sync copy.

• If you go back to an older version of Forsion, this computer registers again as a new device and your other devices need to confirm it again. You'll also need to re-enter the token for an external engine.

• The LAN pairing prompt, the repeated-crash dialog, the download-complete notification and the titles of system file pickers now follow the interface language. They used to always be in Chinese.

• Approval cards now say which device a request came from. When a task started from another device (a phone or the device page) asks for approval, the card shows its source under the title: a device registered to your account and recognized by the server appears as "From a remote session · device name" (the name it was registered with; renaming it later doesn't change this. If that name has only invisible characters, it shows "a registered device"); clients that can't be tied to a specific device, such as the device page in a browser or the web app, appear as "an unidentified client on your account", and LAN-paired devices and peer-to-peer connections are labeled too. Tasks started on this computer show nothing extra. Recognizing specific devices requires server 2.3.24.

• The device switcher no longer lists phones. A phone only drives your computers and doesn't accept connections, so it used to show up as a device that was always "offline".

• The session workspace (the files panel, attachments you send and artifacts you download, including when you use this computer remotely from the device page or your phone) only reads and writes files inside the workspace: shortcuts (symlinks) that point outside it are never listed, downloaded or written through, even when the whole session folder has been replaced by one, and a named pipe in the workspace no longer makes requests hang. Workspaces synced to the cloud skip such shortcuts too when pulling attachments and sending artifacts back.

• New "Allow remote sessions" switch, in Settings › Remote sessions and under "Allow other devices to connect" in the device switcher. When it's off, other devices (a phone, another computer, the device page in a browser) can still view sessions, answer approvals and stop tasks, but can't start or continue sessions on this computer, upload attachments, or hand a Muse to-do over to Muse; turning it off doesn't stop tasks that are already running. Until now, having "Allow other devices to connect" on was enough for other devices to start tasks here; from this update they need this switch too.

• If "Allow other devices to connect" was on when you updated, "Allow remote sessions" is turned on as well and "This account's browsers and web app" is marked as allowed in advance, so the device page keeps working. To tighten this, revoke it in Settings › Remote sessions › Allowed devices: browsers, the web app and P2P connections can then only view, answer approvals and stop tasks, and this computer won't ask again until you choose Allow in the same place. (If you weren't signed in to Forsion when you updated, you'll confirm it once on this computer the first time the device page in a browser starts a task.) On new installs, and if you never turned on connecting, it's off by default.

• The first time another device asks to run sessions on this computer, this computer asks you to confirm, showing the name the device was registered with, its type and its registration date. Devices that aren't in your account's device list are refused without a prompt. Device pages opened in a browser, the Forsion web app and older app versions can't be told apart, so they share one entry, "This account's browsers and web app", which is confirmed once the first time it asks. P2P connections from the device switcher use this entry too, but they never ask: to let a P2P connection run sessions, allow this entry in Settings › Remote sessions on this computer (or open the computer once through Relay). After you decline, you won't be asked again for 10 minutes; if no one answers a prompt within 2 minutes, it won't come back for a minute. Every allowed device is listed in Settings and can be revoked at any time. When a request is refused, the device page now says why (declined, revoked, that computer isn't signed in, the device isn't in your account, too many pending requests and so on) instead of always saying it's waiting for confirmation.

• New "Highest approval mode for remote sessions" (Readonly, Auto edit or Full auto; Auto edit by default). Sessions started from other devices, including the device page, never go above it, and it can only be changed on this computer. Choosing Full auto requires ticking "I understand the risk" and confirming: commands in remote sessions are then no longer checked one by one, and approvals are no longer guaranteed to come from a person.

• This switch only covers Agent sessions: while "Allow other devices to connect" is on, paired devices and devices signed in to the same account can still browse files in this computer's working folders and read and write the vault from the device page, as before.

• Remote sessions can't be turned on while the device credentials aren't stored encrypted (for example, on Linux without a usable system keyring). Clearing desktop data also resets the remote session switch and the allowed devices.

• This computer now shows a system notification when a remote session needs your approval. When a task running here from your phone or the device page needs an approval or an answer, you get a notification such as "Remote session from Pixel 9 needs your approval" even if the Forsion window is closed. It only names the session and the tool, never the command itself, and has no "Approve" button: click it to open the session and decide there. New requests in the same session within 30 seconds are grouped into one notification, and the notification is withdrawn once the request is handled elsewhere. Tasks started on this computer are still only shown in the window.

• If nobody responds within a minute, your phone's inbox gets a reminder. When a remote session's request has been waiting on this computer for a minute, your account receives an inbox message "An agent is waiting for you" (it only contains the computer's name and the number of items, no session content). Open the inbox on your phone to see which computer is waiting and how many items there are. Opening that session straight from the message comes in a later release, once your phone can open sessions that live on your computer; until then, "Open session" in the message only works for sessions your phone already lists. There is at most one reminder per session every 10 minutes, and reminders are archived after 24 hours. While the phone app is in the background or closed, the reminder doesn't pop up on its own; you'll see it the next time you open the app (real push notifications come in a later release). Requires server 2.3.24.

• The session list shows a "waiting for you" dot: sessions with an approval or a question waiting for you get a small warning-colored dot on the corner of their icon (it takes priority over "running" and "unread"), and hovering shows how many. It also lights up for sessions you haven't opened and sessions started from another device.

• Actions that touch protected configuration can only be approved on the computer running them. Actions that would change Forsion's local configuration or credentials can now only be rejected from a phone, the device page or a messaging channel, not approved (they could be before); the approval card explains why. You can still approve them on this computer. In a messaging channel (such as WeChat) you're told right away to approve it on the computer, and once you do, the task's result is still sent back to that chat.

• Messages with an expiry date in the phone inbox are now archived automatically once they expire, the same as on desktop, and the inbox is refreshed as soon as you return to the app.

• New emergency stop. Press ⌃⌥⇧. (Ctrl+Alt+Shift+. on Windows and Linux), choose "Stop all remote tasks" in the menu bar icon, or click "Emergency stop" in Settings › Remote sessions to immediately end every task on this computer that you didn't start at the keyboard: remote sessions from your phone, the device page or other computers, tasks from messaging channels such as WeChat, and Muse and automations. Background processes they started are ended too, and to-dos you handed to Muse from your phone that haven't been finished go back to pending. Conversations you started on this computer are not affected. It works offline as well, and a notification tells you what was stopped.

• After an emergency stop, remote access stays locked, even across restarts. While locked, other devices can only view sessions and stop tasks. They can't start, continue, approve or change anything, pair new devices or open direct connections, and your phone and the device page say "Remote access to that computer is locked. Unlock it on the computer itself". Muse, watch rules and Agent schedules on this computer pause too; anything that came due runs after you unlock. You can only unlock on this computer. Choose "Unlock…" in the menu bar or in Settings and verify it's you. Macs with Touch ID use Touch ID. Without Touch ID, administrator accounts enter the login password, and standard accounts, Windows and Linux confirm in a dialog. Clearing desktop data doesn't remove the lock.

• The menu bar shows remote sessions. While a remote session runs on this computer, the menu bar icon shows "Remote" and the menu says "Remote session running · device name" (and notes when an approval is waiting for you); while locked it shows "Locked". A notification appears when a new device starts a remote session (at most once every 10 minutes per device).

• You can change or turn off the emergency stop shortcut in Settings › Remote sessions (pressing the current shortcut while recording a new one doesn't trigger an emergency stop). If another app has taken the shortcut, both the menu bar and Settings say "Emergency stop shortcut unavailable", and you can still stop from the menu bar.

• While a remote task is running, this computer won't go to sleep when idle, regardless of the "Stay awake while sessions run" setting (closing the lid still puts it to sleep).

• On your phone, you can now choose to run Tangu on one of your computers (that computer needs "Allow other devices to connect" turned on). New chats then run in that computer's real folders (in its default workspace unless you pick another), and the Agents and models offered for new chats are that computer's. Approvals, progress and artifacts come back to the phone through the Forsion cloud. Settings, the inbox and automations stay on the phone. Your choice is remembered the next time you open the app.

• Each conversation stays on the computer it was started on. Switching where to run later only affects new conversations: messages, approvals, stopping and files for existing ones still go to their original computer, including after you reopen the app. Conversations on different computers can run at the same time, and one computer going offline doesn't affect the others.

• When that computer is offline, its Forsion engine isn't running, it rejects the connection, or it can't confirm this device, a notice above the input box says which one. A task that was running while the connection dropped isn't marked as failed, and it picks up where it left off once the computer is back. The phone reconnects on its own when the connection returns, with no need to retry. For problems that aren't retried automatically, such as this device not being recognized or its credentials being rejected, the notice has a Retry button. If that computer refuses only one action (for example, it's still waiting for you to allow this device there), only that action fails, and viewing sessions and receiving new messages keep working.

• While running on another computer, approvals can only be approved or rejected as is, and hard-deleting sessions, rewinding messages, restoring code checkpoints and external engines aren't available. Image thumbnails load through the relay, and attachments are limited to about 9 MB per send.

• Choose where to run in the Android app. The devices sheet at the bottom of the left drawer now has three parts: "Run on" lists Forsion cloud and every computer on your account, with whether each can be used right now (available, engine not running, offline, waiting for that computer to allow this phone, declined, remote sessions turned off, and so on); "Open device screen" moves into a second, collapsible part and works as before; "This phone" shows this phone's registration. Phones no longer appear in the list.

• The first time you pick a computer, the app registers this phone as a device on your account and asks that computer to allow it. The prompt appears on the computer, and the phone shows "Allow this phone on "computer"" for up to two minutes. Registration only happens the first time you pick a computer, so the feature adds no device if you never use it. You can remove the registration under "This phone"; computers stop trusting this phone after that. Removing it switches back to Forsion cloud first, and the phone isn't registered again in the background; that only happens the next time you pick a computer.

• The phone's device identity is kept in storage encrypted by the Android Keystore. It never passes through the web layer, is never logged, and is left out of system backups and device transfers. Each Forsion account gets its own identity. After reinstalling the app or clearing its data, the phone registers again with a new identity and has to be allowed on your computers again.

• Requests to your computer get this phone's identity attached by the app at the system level, so the computer knows which phone is acting. If the identity can't be obtained, the request is not sent. A lost connection or a temporary server error is treated as offline and picks up again once it recovers, an expired sign-in asks you to sign in again, and an outright refusal shows "This phone couldn't verify its identity". With an older server the app shows "The server is too old to run from a phone" and doesn't add this phone to your device list. Reloading the app also closes any open connections to your computers.

• When you use this computer remotely from the device page or your phone, reading the background agent settings now only returns whether Historian and Muse are on and how often they run. Historian's custom and default prompts, the folders Muse may read, the models both use, Muse's active hours and token budgets, and notification and escalation settings are no longer sent to the remote side. This does not cover Muse's permission level, heartbeat interval or persona, which the remote side can still read (the device page's Muse page shows the permission level). On the device page, Settings › Agents › Background agents is now read-only and says these can only be set up on this computer (it used to show every setting, but changes could not be saved).

• Read-only git commands such as git status now ask for approval when the workspace contains a planted, fake git directory. Before, an agent could use such a directory to run commands as you without any approval, and a remote session could use it to raise its own approval limit. Read-only git commands still run without approval in normal repositories, git worktrees, submodules and folders that are not in any repository.

• System notifications, prompts and the tray menu now use the app's language. When the app picked Chinese from your system language or region without you choosing a language in Settings, remote session approval notifications, the "Remote session request" prompt and the tray menu could still show up in English (and the other way round). Before the Forsion window opens, for example in the tray menu right after launch, the last app language is used.

• In the phone's "Run on" list, cases where that computer won't show a prompt at all (it isn't signed in to Forsion, it can't find this phone in its account, too many requests are waiting, nobody answered the last prompt, and so on) now say why right away, instead of showing "Allow this phone on "computer"" and waiting for two minutes. Only a computer that chose Don't allow shows "You can ask again in 10 minutes". If the prompt on the computer closes without an answer, the phone stops waiting and asks you to tap again. If a computer couldn't be reached earlier, tapping it again shows its current answer instead of staying on "Can't reach it right now". When the prompt on the computer times out, the phone says nobody answered instead of still asking you to allow it there.

• When that computer refuses something while you run tasks on it from your phone or use its device page, the message also gives the actual reason instead of always saying it's waiting for that computer to allow this.

• Agents on your computer can now find files you attach from your phone (or the device page). The file already reached that computer, but in sessions that run in a folder on the computer the agent didn't know where it was and had to search the whole disk. Now the file's location on that computer is handed to the agent with your message, just like a file dragged in on the computer, and the agent reads it from there. When you reopen the session, a file tag for it shows at the start of the message. Small files picked with "Add › File" in sessions that run in a folder on the computer work the same way. If a file uploads from your phone but the message itself isn't sent, the file isn't attached to the next message you send on the computer.

• Conversations run on this computer from your phone or the device page no longer flow automatically into background agents and other conversations on this computer: the recent session titles, daily log summary and activity that Muse sees leave them out, and the related past excerpts that local conversations attach automatically no longer come from them (you can still find them when you ask an agent to search your history in a local conversation). Remote conversations also no longer write to an agent's daily log (neither Historian nor the agent itself), in line with them not writing long-term memory. A conversation from this computer that you continued or renamed on your phone is treated the same way: when you carry on with it on the computer, Historian no longer writes daily log entries, long-term memory or working-note suggestions from it (titles and summaries still update). A note you add when rejecting a background agent's pending action on your phone is no longer written to that agent's daily log.

• When you use this computer remotely from the device page or your phone, the agent list you get no longer includes each agent's persona and developer instructions (both fields are empty when you open an agent's profile on the device page; agents could already only be changed on that computer), and the approval mode shown is the one that actually applies to remote tasks. The Muse status you get no longer includes local folder paths, the run budget or error details.

• See and download a remote session's files on your phone. After you switch Tangu on your phone to one of your computers, the Workspace panel on the right starts with a "Session files" group that lists the attachments and outputs this session has on that computer, including attachments you sent from the phone, each with a download button. That computer's project folders are still listed below it. Files here can't be deleted while you run on another computer.

• The download button in the Android app now actually saves the file: it goes to the phone's Downloads folder, and a message shows the name it was saved under (files with the same name are numbered automatically). Before, tapping it did nothing. The phone can save up to 50 MB at a time and tells you when a file is larger; this requires Android 10 or later.

• Your phone now shows where an approval was answered. When a task started from your phone is approved or rejected directly on the computer, the outcome line in the phone's chat reads "Approved on the host computer (computer name)" or "Rejected …"; approvals answered on another device show that device's name, and the ones you answered on this phone aren't labeled. Collapsed approval cards also show the computer name, and connection notices use the computer's name instead of "your computer".

• While your phone waits for a computer to allow it, the "This phone" row already shows its registered name instead of "Not registered yet".

• After the connection to a computer drops and reconnects, replies that were already shown are no longer occasionally shown a second time.

• See your computers' sessions in the phone sidebar. Below the session list in the left drawer, each computer on your account now has its own section. Each lists that computer's recent sessions (including ones started on the computer itself) and shows its current state: online, offline, agent engine starting, engine not running, remote sessions turned off, waiting for that computer to allow this phone, and so on, each shown differently. An offline computer shows only its name and state, and its sessions come back once it's online again. Tap a session to open it on the phone and carry on; approvals, stopping and files go to that computer, and they keep going there after you switch new sessions to the cloud or another computer. These sessions are fetched live while the app is open and aren't stored on the phone. From the phone you can rename and archive them but not delete them for good. Until the phone has picked a computer once (so it isn't registered yet), each section only shows the computer's state; sessions appear after that.

• New chats show where they will run (in Chat mode too), right above the input box: the cloud or a computer, plus that computer's current state. Tap it to open the "Run on" sheet. The place you pick is remembered: new chats go back to that computer by default only while it's online and available, and use the cloud otherwise. Sessions created on a computer no longer carry the phone's notes vault path, cloud project name, external engine, or image and vision model settings.

• "Open session" in the inbox reminder about an agent waiting for you now opens that session directly, without first switching the phone to that computer.

• When a computer refuses to create a session (for example, remote sessions are turned off there), the message now says why instead of only saying the session couldn't be created.

• With "Allow other devices to connect" on, the desktop app now reports its agent engine state (running, starting, not running) to your account, so the phone can tell "the computer is online but its engine isn't up" from "ready to use". The desktop main window, the web app and the device page don't show these sections or the "Run on" pill.

• On a Mac, commands approved in a remote session can no longer change this computer's Forsion settings. When a task started from your phone or the device page runs commands on this Mac (run_bash and background processes), they run inside a system write protection. They can't write Forsion settings and sign-in data (including the highest approval mode for remote sessions, the allowed devices list and the emergency stop state), agent configuration, skills, plugins and hooks, .tangu folders in projects, shell startup files such as ~/.zshrc, dotfiles in your home folder such as ~/.gitconfig, ~/.config or ~/Library/LaunchAgents, and neither can any program the command starts. The workspace, agents' Library folders and cache folders such as ~/.npm stay writable. Before, with the host sandbox off (the default), a command you approved on your phone could rewrite these files directly, for example to raise the remote session's own approval limit, and the command itself (python3 fix.py, or npm install with an install script) may not have made that obvious. Tasks started on this computer are not affected, and with the host sandbox on, its own sandbox applies as before.

• In remote sessions, tools that set up their own sandbox (such as swift build or swift package) can't start it inside this write protection and fail; the tool result suggests adding --disable-sandbox or running them on this computer. If the system's write protection isn't available, commands from remote sessions are not run and the reason is shown; they never run unprotected. Background processes started on the computer without this protection no longer accept input from a remote session. Each remote command takes about 9 ms longer to start.

• Forsion now reads a project's Git status (the Git status attached when a task starts, and the project details panel) inside this write protection no matter where the task was started, and so do Git commands an agent may run without approval (such as git status). A Git filter or similar program a remote command left in the repository still runs, but can't change the files above; if the protection isn't available, Forsion skips these Git reads and such commands need approval. On every platform, git status and similar commands need approval when the repository's own configuration sets up programs Git would run (filters, Git LFS extensions, fsmonitor, hooks, external diff and the like). On Linux and Windows, Forsion no longer reads the Git status of a repository that configures a filter or a Git LFS extension, and the project details panel shows no Git information for it.

• Linux and Windows don't have this protection: a command you approve on your phone can change Forsion settings on this computer, including the highest approval mode for remote sessions, and the command may not make that obvious. Settings › Remote sessions now says so. On Linux, setting "Local command and file sandbox" to "Workspace writes only" prevents it; Windows doesn't have such a sandbox yet, so only approve commands you understand.

Forsion 2.12

2.12.1 (2026-09-30)

• Fixed conversations failing when a plugin or custom tool has an invalid definition. Invalid tools are now isolated so other tools remain available, and feedback logs identify the affected tool and its source.

2.12.0 (2026-09-29)

• New: HUMAN.md, a collaboration handbook between you and your agents. Write down how you like to work together in a readable, editable handbook with two levels: the agent level holds working preferences that apply across projects (for example, "When there is a design choice, give me two comparable options first"), and the project level holds only the goals, division of work and acceptance criteria of the current project (for example, "I do the visual review before this release"), stored in the project's .tangu/HUMAN.md (a HUMAN.md already in the project root or in .forsion/ is still read). When you give clear feedback, the agent updates the handbook itself, sparingly and with immediate effect, and leaves an update card in the chat that you can view, edit, undo or expand to see the reasoning; the card is still there after you reopen the session. The "Collaboration" tab on agent details and project details lets you edit the handbook directly, read the raw text, see the file location and browse the change history (the last 40 versions, kept on this device and never written into the project repository).

• Handbook boundaries: what you say right now always takes priority; the handbook only describes how to work together and grants agents no tool permissions; saving is version-checked, so it never silently overwrites your manual edits (on a conflict your draft is kept and you review the latest version before merging); undo only applies to an update that is still the latest version; and agents that share memory do not share each other's handbooks.

• New thinking level: Ultra. The thinking-level slider gets an eighth notch after Max (you can also type /think ultra): thinking is always at its highest, context goes up to the model's limit, and the agent proactively hands work that can run in parallel to several sub-agents at once. It is available only in local work sessions; web, Chat mode, external engines and team sessions don't have it. Each time you turn it on, a dialog explains that token use can be several times the usual and offers "Also switch this session to Full access" (unchecked by default, since parallel sub-agents each wait for your approval) and "Don't show this again". Sub-agent usage also counts toward the per-run cost cap.

• Approvals now wait in a tray above the input box, and the agent keeps working while you decide. Pending confirmations stack above the input box: the oldest is expanded and ready to approve, the rest are listed as rows (click a row to expand it), and the tray can be collapsed; the chat itself keeps a single line such as "N actions are waiting for your approval". When an action needs your approval, the agent parks it, carries on with the steps that don't depend on it, and runs it at the next step once you approve; it only stops and waits when it has nothing else to do. The agent's questions and plan approvals use the tray too (the plan text stays in the chat, the decision buttons are in the tray). After you move to the next card, its buttons are briefly on cooldown so a double click can't approve the next item by mistake. Sub-agents, external engines, the terminal TUI, message channels and Muse still wait for each approval in turn.

• You can keep sending commands and messages while a task is running, and they queue. /compact, /refine and the messages sent after them wait in the area above the input box and run one after another when the current run ends (click × to remove one, ↑ to take it back into the input box); before, /compact during a run only showed "Compression failed". The terminal TUI and /retry in the desktop app queue the same way.

• New: Computer history (macOS only, off by default). Turn it on in Settings → Computer history to let Forsion record what you do on this Mac: app switches, window titles, URLs, changes to the text in the focused input field, and the names of controls you click and keyboard shortcuts you press. It never records raw keystrokes or screenshots, and password managers, Keychain, secure input fields and private windows are excluded automatically. Records stay on this Mac for 7 days; you can pause, exclude apps or sites, or clear them. Today's records show as a 20-minute timeline with real app icons, and you can browse the last 7 days. Agents read them through the read_computer_history tool (visible only when the feature is on, and unavailable in remote sessions, message channels and teams), and Muse also looks at a summary of the last few hours. Clearing the records does not delete conversations or Muse journal entries that already quoted them. Requires Computer Use 0.6 and the macOS Accessibility permission.

• Tangu can work in the Chrome you are already using, and there is a new Tangu for Chrome extension. Agents can now list, read and drive the tabs you already have open in Chrome, Edge or Brave (new browser_tabs tool; it needs remote debugging turned on in the browser, which asks "Allow remote debugging?" once per new connection). With the extension installed (Settings → Browser has three installation steps), reading, clicking and typing happen in background tabs without taking focus or showing the debugging banner, pages Tangu opens itself go into an orange "Tangu" tab group and need no approval, and you pair only once. Clicks and typing in your own tabs still need approval, and unattended runs such as Muse and automations never touch your browser.

• Documents work out of the box. The desktop app now ships Python with 15 pinned libraries for office and data work (python-docx, python-pptx, openpyxl, pandas, matplotlib, reportlab and more), so agents no longer run pip install on the spot when handling docx, pptx, xlsx and PDF files. It also bundles a LibreOffice conversion engine, so read_document reads doc, docx, xls, xlsx, ppt, pptx, odt, ods and odp by converting to PDF first and parsing real pages, with no separate LibreOffice install. Git is bundled on Windows and macOS too (Linux still uses the system git), and Windows includes the VC++ runtime the engine needs.

• Project details can now run Git actions: create a repository, commit (it lists every change first, then writes the message with the session's model and lets you edit it), create a branch (with a prefilled prefix) and push. Settings → General → Git sets the branch prefix (default tangu/), commit message guidelines and whether rewriting remote history is allowed, and agents follow them when you ask them to create a branch or write a commit; agents never create a repository or commit on their own. If a repository's own configuration can run programs (filters, hooks and the like), Forsion doesn't read its Git status until you choose "Trust and continue".

• Agents can offer to turn what they build into a creation. Web pages and tools made in a chat used to have no way into Creations. Now an agent shows a card when it fits: "Make it a creation" creates a folder in your Creations location and moves the session there to carry on, and "Add to Creations" takes in an existing project in place without copying it (project details has "Add to Creations" too). In any Space, a creation now saves a version after each run (desktop only).

• Generative AI right in your notes. Select text and use "AI ▾" in the toolbar, type /ai, or press Space on an empty line (turn it on in Settings → Notes → Behavior) to improve writing, fix spelling and grammar, shorten, expand, summarize, translate, continue, or give a custom instruction. The result streams into a preview first and nothing is written until you confirm (and the write can be undone). The selection toolbar and the ⠿ block menu also have "Ask Tangu", which hands the selection to the side chat, and every assistant answer in a chat has "Insert into note" to put it back into a note in one step. When Tangu edits a note you have open, the changes are marked in the text and a bar at the bottom shows "Tangu changes: N" so you can review each one, revert all or keep them.

• Links and formatting in notes, a big batch of improvements (benchmarked against Notion and Obsidian): the backlinks panel lists each mention with its context and turns unlinked mentions into links in one click; [[ completion puts recently used notes first and supports #heading, #^block and aliases; #tags render as clickable pills with completion; renaming a note on web and mobile now rewrites [[links]] across the whole vault; ==highlight== and %%comment%% preview live, and <kbd>, <sub> and <sup> are editable in place; the note ⋯ menu adds full width, small text and a page font (default, serif or monospace), and you can lock a page; cursor and scroll position and fold state are remembered per note on this device.

• Editing and canvas improvements: in-page find supports replace, replace all, match case and regular expressions (source mode too); tables support continuous editing with Enter, Tab and Shift+Enter, and the block menu has a "Table" section; Cmd/Ctrl+D duplicates a block, Shift+↑↓ extends the selection by block, and Alt+Enter follows the link at the cursor; the status bar now counts visible text only and adds word and selection counts. On the canvas, cards, shapes, frames and connectors can have colors, dragging to the edge of the view pans automatically, and holding Space pans while Cmd/Ctrl + =/-/0 zooms. A batch of save and sync fixes for problems that could lose or overwrite content also landed.

• Muse and background agents are easier on you: Muse decides on its own how long to sleep based on your routine (any activity from you wakes it, and rule-triggered wake-ups and due schedule items still wake it); Muse and automations that use the cloud default background model have their own background quota, so they no longer eat into your chat quota, and when it runs out they pause by default unless you choose to continue on the main quota (requires server 2.3.21); each Muse cycle now starts a new session, so the opening context no longer snowballs from cycle to cycle and each cycle costs fewer tokens; the Muse Space previously failed to render and now does; and Spaces that agents build for themselves can read and update their own data (to-dos, schedule, Library) and render from it.

• The Automations Space is now a node canvas. First-run templates (daily reminder, agent briefing, notify on a new table row, one-click agent) get you started; the canvas can be dragged, panned, zoomed and auto-arranged, and each node edits a trigger or a notification / agent / database / tool step; new rules are saved paused; notification titles and bodies can insert date, time and trigger-row variables; each run has an expandable ledger with step summaries and full errors (a trial run really executes); and deleting an automation can be undone for 5 seconds instead of asking for confirmation.

• Context ring details: click the context ring in the input box to see a color-coded segmented bar (the striped end is the space reserved for auto-compaction), how much room is left before auto-compaction and a "Compact now" button; click the header to expand a breakdown (history, tools and this turn, system prompt, memory, project and files, skills and more). When signed in to Forsion, a "Forsion quota" section at the bottom shows today's and this week's usage as percentages, a countdown to the daily reset at 00:00 Beijing time and the weekly reset time, and turns to a warning color when 15% or less is left.

• Images, audio and video show up directly in chat: a paragraph in an assistant message that contains only ![[…]] is rendered in place as an image or a media player, for both vault-relative paths and absolute paths outside the vault (local desktop sessions only; files over 50 MB fall back to a reference bar).

• Tangu can operate your Android phone (experimental). After you turn on "Allow Tangu to operate this phone" in Settings → Advanced → Experimental features in the Android app (and confirm in the system dialog), Tangu can use system capabilities in conversations started on the phone, such as opening apps and maps; it asks once more before opening apps other than the browser and maps. Only tasks started on the phone itself can drive the phone, accessibility-based screen reading and tapping is not included yet, and remote drivers such as the computer, WeChat or Muse are not supported.

• Settings has a new top-level page, "Forsion Cloud", so you can take care of your account without leaving the desktop app: Account (photo, nickname, linked phone and email, membership tier), Quota & points (today's and this week's quota, quota reset cards, points and top-ups, background quota, invite and redemption codes), Security (re-link, change password, sign out everywhere, delete account), Usage and Feedback. AI quota is shown only as the percentage left. These pages update with Forsion Extend.

• Announcements are new, and automatic points deduction is retired. When Forsion posts an important notice, a card pops up when you open the desktop app and doesn't come back once you choose "Got it" or close it; links in announcements open in your browser. Using a quota reset card now plays a quota-restored animation. "Pay with points automatically" is gone: when your AI quota runs out, points are no longer spent to keep going, and points can still be exchanged for reset cards (updates with Forsion Extend 0.7; requires server 2.3.26).

• Computer Use: fixed agents having no computer-use tools at all on a fresh Windows install (the bundled plugin was off by default and the helper lacked the VC++ runtime); the macOS helper is now signed with a fixed certificate so updates no longer drop its Accessibility permission, and after upgrading from an older version you need to grant Accessibility to tangu-computer-use once more in System Settings; the automatic Mini window no longer follows the cursor, now sits in the top-right corner where you can drag it, and returns to the Forsion window you started from when the run ends.

• The first-run guide has a "Local environment" step again, and Settings has a "Local environment" page. The guide lets you choose the official source or a mainland China mirror (recommended by time zone, system language or IP region) and checks local tools such as npm, git and docker; the page under Settings → General puts the developer tools list first and lets Tangu install what's missing. One-click install on Windows no longer hangs (install commands run non-interactively, and a download page is offered when there is no package manager).

• A large batch of interface polish from two rounds of UI/UX review: in Settings, the backend mode card only edits a draft and switching takes a button that states the consequences, instant toggles no longer overwrite settings you are editing, items that need a restart appear in a sticky save bar, and settings can be searched; "Restore default layout" can be undone; dates and times are formatted consistently and labels follow one wording and punctuation style; the Calendar and Creations no longer flash a skeleton on first open; in-app notifications can have a custom duration; the Inbox left column can be resized; the first day of the week in Calendar follows the language (Monday in Chinese, Sunday in English, adjustable in Calendar settings). After installing or uninstalling a plugin in the Settings window, the main window follows right away without a refresh, and a "Restart backend" button appears when one is needed.

• The changelog now has an English edition that follows the interface language (a version missing from the English file shows its Chinese text).

• A NoExtend desktop installer is now available on macOS, Windows and Linux alongside the standard installer with Forsion Extend. Both variants keep the same app identity and local data. NoExtend has its own update feed, so later updates stay on that variant.

• Permissions for Load in Forsion and desktop shortcuts now also check when the folder was created, so a folder deleted and recreated at the same path no longer inherits the earlier permission. Permissions saved by older versions lack this record and need to be confirmed once: opening such a shortcut asks you to choose Add to desktop again in Creations, and the Sandbox panel in the Coding Studio asks you to choose Load again.

• The device page (this computer's Forsion opened from another device) is safer. The following now only work on this computer itself: installing plugins; changing external engines, messaging channels, model providers, web search, hooks and approval rules; permanently deleting sessions or messages in a session; listing read-aloud voices; and emptying the notes trash or permanently deleting items from it. Trying them on the device page shows "This action is only available on that device itself, not over a remote connection". Archiving sessions still works.

• Tasks started from the device page run at most in the "Auto edit" approval mode by default, even when "Full auto" is selected. On the device page, approval cards can only be approved or rejected as is: the command can't be edited and "Always allow this session" isn't offered. Sessions from the device page can't use the filesystem root, the home folder or a folder holding protected configuration as their working folder.

• The host files the device page can browse are limited to the working folder, the notes vault and the project folders this computer already has; after the update, sessions whose folder was changed from the device page, and project sessions started by device-page tasks, no longer widen what can be read. A project you later open elsewhere on this computer by other means (for example from the command line) becomes browsable on the device page once you pick its folder with "Add local project" on this computer. Right after updating, the device page may briefly report that the engine isn't ready.

• Tasks started from the device page also can't create or change this computer's agents, skills, working notes, long-term memory, schedules or automations. Those changes affect every later conversation on this computer, so make them on the computer itself. Conversations driven from the device page aren't written into long-term memory by Historian either.

• On the device page, a session's approval mode can only be made stricter, not looser; a session that never had one can only be set to "Read only".

• Device-page sessions also can't use an app's own data folder as their working folder (~/Library on macOS, AppData on Windows, ~/.config on Linux and similar). Projects in cloud drives (Dropbox, OneDrive, Google Drive and others) and in iCloud Drive are unaffected.

• Agents can no longer read Forsion and Tangu configuration and secret files, including config.json, .env, MCP and model provider settings, the WeChat channel token, plugin settings, the agent browser's sign-in data and Forsion's app data folders. This applies to conversations on this computer too; reading them from the command line needs your approval.

• Shortcuts (symlinks) in an agent's private Library folder that point outside it can no longer be read.

• The local engine now uses its own local token instead of your Forsion account token.

• MCP: the server names dev and anything starting with dev_ are reserved for built-in developer tools. Existing servers with those names stop loading and show an error in their status; renaming them brings them back. Error and text results from MCP tools are marked as external data before they reach the model, and images they return are treated as untrusted content.

• This computer's device credentials are now encrypted. The pairing credentials behind "Allow other devices to connect" and the token for an external engine are now kept in the system keychain (macOS Keychain, Windows DPAPI, or GNOME Keyring or KWallet on Linux) instead of in plain text in a configuration file, and they no longer appear in the configuration the app's interface can read. They move over automatically the first time you start the new version.

• On macOS, after an update the system may ask for keychain access the first time Forsion reads these credentials. Choose "Always Allow". If you deny it, this computer won't switch to a new device identity and won't fall back to storing them in plain text; it only stops connecting to your other devices for now (direct LAN connections keep working). The system won't ask again while Forsion is running, so the notice offers "Restart Forsion": restart and choose "Always Allow" to recover. If the saved credentials can no longer be decrypted (for example, the keychain was reset), click "Retry", or confirm "Re-register this device" in the system dialog (your other devices will need to trust it again).

• If the token for an external engine can't be read, the external connection panel in Settings › Connection shows a notice. You can retry, restart, or enter the token again.

• On Linux without a usable system keyring (for example, no GNOME Keyring or KWallet installed), these credentials are stored as before and the device switcher shows a notice. Remote sessions can't be turned on in this state.

• What encryption covers: the Windows and Linux keychains aren't isolated per app, so other programs running as the same user may still read them. It protects against the credentials being read straight from a configuration file or carried off in a backup or sync copy.

• If you go back to an older version of Forsion, this computer registers again as a new device and your other devices need to confirm it again. You'll also need to re-enter the token for an external engine.

• The LAN pairing prompt, the repeated-crash dialog, the download-complete notification and the titles of system file pickers now follow the interface language. They used to always be in Chinese.

• Approval cards now say which device a request came from. When a task started from another device (a phone or the device page) asks for approval, the card shows its source under the title: a device registered to your account and recognized by the server appears as "From a remote session · device name" (the name it was registered with; renaming it later doesn't change this. If that name has only invisible characters, it shows "a registered device"); clients that can't be tied to a specific device, such as the device page in a browser or the web app, appear as "an unidentified client on your account", and LAN-paired devices and peer-to-peer connections are labeled too. Tasks started on this computer show nothing extra. Recognizing specific devices requires server 2.3.24.

• The device switcher no longer lists phones. A phone only drives your computers and doesn't accept connections, so it used to show up as a device that was always "offline".

• The session workspace (the files panel, attachments you send and artifacts you download, including when you use this computer remotely from the device page or your phone) only reads and writes files inside the workspace: shortcuts (symlinks) that point outside it are never listed, downloaded or written through, even when the whole session folder has been replaced by one, and a named pipe in the workspace no longer makes requests hang. Workspaces synced to the cloud skip such shortcuts too when pulling attachments and sending artifacts back.

• New "Allow remote sessions" switch, in Settings › Remote sessions and under "Allow other devices to connect" in the device switcher. When it's off, other devices (a phone, another computer, the device page in a browser) can still view sessions, answer approvals and stop tasks, but can't start or continue sessions on this computer, upload attachments, or hand a Muse to-do over to Muse; turning it off doesn't stop tasks that are already running. Until now, having "Allow other devices to connect" on was enough for other devices to start tasks here; from this update they need this switch too.

• If "Allow other devices to connect" was on when you updated, "Allow remote sessions" is turned on as well and "This account's browsers and web app" is marked as allowed in advance, so the device page keeps working. To tighten this, revoke it in Settings › Remote sessions › Allowed devices: browsers, the web app and P2P connections can then only view, answer approvals and stop tasks, and this computer won't ask again until you choose Allow in the same place. (If you weren't signed in to Forsion when you updated, you'll confirm it once on this computer the first time the device page in a browser starts a task.) On new installs, and if you never turned on connecting, it's off by default.

• The first time another device asks to run sessions on this computer, this computer asks you to confirm, showing the name the device was registered with, its type and its registration date. Devices that aren't in your account's device list are refused without a prompt. Device pages opened in a browser, the Forsion web app and older app versions can't be told apart, so they share one entry, "This account's browsers and web app", which is confirmed once the first time it asks. P2P connections from the device switcher use this entry too, but they never ask: to let a P2P connection run sessions, allow this entry in Settings › Remote sessions on this computer (or open the computer once through Relay). After you decline, you won't be asked again for 10 minutes; if no one answers a prompt within 2 minutes, it won't come back for a minute. Every allowed device is listed in Settings and can be revoked at any time. When a request is refused, the device page now says why (declined, revoked, that computer isn't signed in, the device isn't in your account, too many pending requests and so on) instead of always saying it's waiting for confirmation.

• New "Highest approval mode for remote sessions" (Readonly, Auto edit or Full auto; Auto edit by default). Sessions started from other devices, including the device page, never go above it, and it can only be changed on this computer. Choosing Full auto requires ticking "I understand the risk" and confirming: commands in remote sessions are then no longer checked one by one, and approvals are no longer guaranteed to come from a person.

• This switch only covers Agent sessions: while "Allow other devices to connect" is on, paired devices and devices signed in to the same account can still browse files in this computer's working folders and read and write the vault from the device page, as before.

• Remote sessions can't be turned on while the device credentials aren't stored encrypted (for example, on Linux without a usable system keyring). Clearing desktop data also resets the remote session switch and the allowed devices.

• This computer now shows a system notification when a remote session needs your approval. When a task running here from your phone or the device page needs an approval or an answer, you get a notification such as "Remote session from Pixel 9 needs your approval" even if the Forsion window is closed. It only names the session and the tool, never the command itself, and has no "Approve" button: click it to open the session and decide there. New requests in the same session within 30 seconds are grouped into one notification, and the notification is withdrawn once the request is handled elsewhere. Tasks started on this computer are still only shown in the window.

• If nobody responds within a minute, your phone's inbox gets a reminder. When a remote session's request has been waiting on this computer for a minute, your account receives an inbox message "An agent is waiting for you" (it only contains the computer's name and the number of items, no session content). Open the inbox on your phone to see which computer is waiting and how many items there are. Opening that session straight from the message comes in a later release, once your phone can open sessions that live on your computer; until then, "Open session" in the message only works for sessions your phone already lists. There is at most one reminder per session every 10 minutes, and reminders are archived after 24 hours. While the phone app is in the background or closed, the reminder doesn't pop up on its own; you'll see it the next time you open the app (real push notifications come in a later release). Requires server 2.3.24.

• The session list shows a "waiting for you" dot: sessions with an approval or a question waiting for you get a small warning-colored dot on the corner of their icon (it takes priority over "running" and "unread"), and hovering shows how many. It also lights up for sessions you haven't opened and sessions started from another device.

• Actions that touch protected configuration can only be approved on the computer running them. Actions that would change Forsion's local configuration or credentials can now only be rejected from a phone, the device page or a messaging channel, not approved (they could be before); the approval card explains why. You can still approve them on this computer. In a messaging channel (such as WeChat) you're told right away to approve it on the computer, and once you do, the task's result is still sent back to that chat.

• Messages with an expiry date in the phone inbox are now archived automatically once they expire, the same as on desktop, and the inbox is refreshed as soon as you return to the app.

• New emergency stop. Press ⌃⌥⇧. (Ctrl+Alt+Shift+. on Windows and Linux), choose "Stop all remote tasks" in the menu bar icon, or click "Emergency stop" in Settings › Remote sessions to immediately end every task on this computer that you didn't start at the keyboard: remote sessions from your phone, the device page or other computers, tasks from messaging channels such as WeChat, and Muse and automations. Background processes they started are ended too, and to-dos you handed to Muse from your phone that haven't been finished go back to pending. Conversations you started on this computer are not affected. It works offline as well, and a notification tells you what was stopped.

• After an emergency stop, remote access stays locked, even across restarts. While locked, other devices can only view sessions and stop tasks. They can't start, continue, approve or change anything, pair new devices or open direct connections, and your phone and the device page say "Remote access to that computer is locked. Unlock it on the computer itself". Muse, watch rules and Agent schedules on this computer pause too; anything that came due runs after you unlock. You can only unlock on this computer. Choose "Unlock…" in the menu bar or in Settings and verify it's you. Macs with Touch ID use Touch ID. Without Touch ID, administrator accounts enter the login password, and standard accounts, Windows and Linux confirm in a dialog. Clearing desktop data doesn't remove the lock.

• The menu bar shows remote sessions. While a remote session runs on this computer, the menu bar icon shows "Remote" and the menu says "Remote session running · device name" (and notes when an approval is waiting for you); while locked it shows "Locked". A notification appears when a new device starts a remote session (at most once every 10 minutes per device).

• You can change or turn off the emergency stop shortcut in Settings › Remote sessions (pressing the current shortcut while recording a new one doesn't trigger an emergency stop). If another app has taken the shortcut, both the menu bar and Settings say "Emergency stop shortcut unavailable", and you can still stop from the menu bar.

• While a remote task is running, this computer won't go to sleep when idle, regardless of the "Stay awake while sessions run" setting (closing the lid still puts it to sleep).

• On your phone, you can now choose to run Tangu on one of your computers (that computer needs "Allow other devices to connect" turned on). New chats then run in that computer's real folders (in its default workspace unless you pick another), and the Agents and models offered for new chats are that computer's. Approvals, progress and artifacts come back to the phone through the Forsion cloud. Settings, the inbox and automations stay on the phone. Your choice is remembered the next time you open the app.

• Each conversation stays on the computer it was started on. Switching where to run later only affects new conversations: messages, approvals, stopping and files for existing ones still go to their original computer, including after you reopen the app. Conversations on different computers can run at the same time, and one computer going offline doesn't affect the others.

• When that computer is offline, its Forsion engine isn't running, it rejects the connection, or it can't confirm this device, a notice above the input box says which one. A task that was running while the connection dropped isn't marked as failed, and it picks up where it left off once the computer is back. The phone reconnects on its own when the connection returns, with no need to retry. For problems that aren't retried automatically, such as this device not being recognized or its credentials being rejected, the notice has a Retry button. If that computer refuses only one action (for example, it's still waiting for you to allow this device there), only that action fails, and viewing sessions and receiving new messages keep working.

• While running on another computer, approvals can only be approved or rejected as is, and hard-deleting sessions, rewinding messages, restoring code checkpoints and external engines aren't available. Image thumbnails load through the relay, and attachments are limited to about 9 MB per send.

• Choose where to run in the Android app. The devices sheet at the bottom of the left drawer now has three parts: "Run on" lists Forsion cloud and every computer on your account, with whether each can be used right now (available, engine not running, offline, waiting for that computer to allow this phone, declined, remote sessions turned off, and so on); "Open device screen" moves into a second, collapsible part and works as before; "This phone" shows this phone's registration. Phones no longer appear in the list.

• The first time you pick a computer, the app registers this phone as a device on your account and asks that computer to allow it. The prompt appears on the computer, and the phone shows "Allow this phone on "computer"" for up to two minutes. Registration only happens the first time you pick a computer, so the feature adds no device if you never use it. You can remove the registration under "This phone"; computers stop trusting this phone after that. Removing it switches back to Forsion cloud first, and the phone isn't registered again in the background; that only happens the next time you pick a computer.

• The phone's device identity is kept in storage encrypted by the Android Keystore. It never passes through the web layer, is never logged, and is left out of system backups and device transfers. Each Forsion account gets its own identity. After reinstalling the app or clearing its data, the phone registers again with a new identity and has to be allowed on your computers again.

• Requests to your computer get this phone's identity attached by the app at the system level, so the computer knows which phone is acting. If the identity can't be obtained, the request is not sent. A lost connection or a temporary server error is treated as offline and picks up again once it recovers, an expired sign-in asks you to sign in again, and an outright refusal shows "This phone couldn't verify its identity". With an older server the app shows "The server is too old to run from a phone" and doesn't add this phone to your device list. Reloading the app also closes any open connections to your computers.

• When you use this computer remotely from the device page or your phone, reading the background agent settings now only returns whether Historian and Muse are on and how often they run. Historian's custom and default prompts, the folders Muse may read, the models both use, Muse's active hours and token budgets, and notification and escalation settings are no longer sent to the remote side. This does not cover Muse's permission level, heartbeat interval or persona, which the remote side can still read (the device page's Muse page shows the permission level). On the device page, Settings › Agents › Background agents is now read-only and says these can only be set up on this computer (it used to show every setting, but changes could not be saved).

• Read-only git commands such as git status now ask for approval when the workspace contains a planted, fake git directory. Before, an agent could use such a directory to run commands as you without any approval, and a remote session could use it to raise its own approval limit. Read-only git commands still run without approval in normal repositories, git worktrees, submodules and folders that are not in any repository.

• System notifications, prompts and the tray menu now use the app's language. When the app picked Chinese from your system language or region without you choosing a language in Settings, remote session approval notifications, the "Remote session request" prompt and the tray menu could still show up in English (and the other way round). Before the Forsion window opens, for example in the tray menu right after launch, the last app language is used.

• In the phone's "Run on" list, cases where that computer won't show a prompt at all (it isn't signed in to Forsion, it can't find this phone in its account, too many requests are waiting, nobody answered the last prompt, and so on) now say why right away, instead of showing "Allow this phone on "computer"" and waiting for two minutes. Only a computer that chose Don't allow shows "You can ask again in 10 minutes". If the prompt on the computer closes without an answer, the phone stops waiting and asks you to tap again. If a computer couldn't be reached earlier, tapping it again shows its current answer instead of staying on "Can't reach it right now". When the prompt on the computer times out, the phone says nobody answered instead of still asking you to allow it there.

• When that computer refuses something while you run tasks on it from your phone or use its device page, the message also gives the actual reason instead of always saying it's waiting for that computer to allow this.

• Agents on your computer can now find files you attach from your phone (or the device page). The file already reached that computer, but in sessions that run in a folder on the computer the agent didn't know where it was and had to search the whole disk. Now the file's location on that computer is handed to the agent with your message, just like a file dragged in on the computer, and the agent reads it from there. When you reopen the session, a file tag for it shows at the start of the message. Small files picked with "Add › File" in sessions that run in a folder on the computer work the same way. If a file uploads from your phone but the message itself isn't sent, the file isn't attached to the next message you send on the computer.

• Conversations run on this computer from your phone or the device page no longer flow automatically into background agents and other conversations on this computer: the recent session titles, daily log summary and activity that Muse sees leave them out, and the related past excerpts that local conversations attach automatically no longer come from them (you can still find them when you ask an agent to search your history in a local conversation). Remote conversations also no longer write to an agent's daily log (neither Historian nor the agent itself), in line with them not writing long-term memory. A conversation from this computer that you continued or renamed on your phone is treated the same way: when you carry on with it on the computer, Historian no longer writes daily log entries, long-term memory or working-note suggestions from it (titles and summaries still update). A note you add when rejecting a background agent's pending action on your phone is no longer written to that agent's daily log.

• When you use this computer remotely from the device page or your phone, the agent list you get no longer includes each agent's persona and developer instructions (both fields are empty when you open an agent's profile on the device page; agents could already only be changed on that computer), and the approval mode shown is the one that actually applies to remote tasks. The Muse status you get no longer includes local folder paths, the run budget or error details.

• See and download a remote session's files on your phone. After you switch Tangu on your phone to one of your computers, the Workspace panel on the right starts with a "Session files" group that lists the attachments and outputs this session has on that computer, including attachments you sent from the phone, each with a download button. That computer's project folders are still listed below it. Files here can't be deleted while you run on another computer.

• The download button in the Android app now actually saves the file: it goes to the phone's Downloads folder, and a message shows the name it was saved under (files with the same name are numbered automatically). Before, tapping it did nothing. The phone can save up to 50 MB at a time and tells you when a file is larger; this requires Android 10 or later.

• Your phone now shows where an approval was answered. When a task started from your phone is approved or rejected directly on the computer, the outcome line in the phone's chat reads "Approved on the host computer (computer name)" or "Rejected …"; approvals answered on another device show that device's name, and the ones you answered on this phone aren't labeled. Collapsed approval cards also show the computer name, and connection notices use the computer's name instead of "your computer".

• While your phone waits for a computer to allow it, the "This phone" row already shows its registered name instead of "Not registered yet".

• After the connection to a computer drops and reconnects, replies that were already shown are no longer occasionally shown a second time.

• See your computers' sessions in the phone sidebar. Below the session list in the left drawer, each computer on your account now has its own section. Each lists that computer's recent sessions (including ones started on the computer itself) and shows its current state: online, offline, agent engine starting, engine not running, remote sessions turned off, waiting for that computer to allow this phone, and so on, each shown differently. An offline computer shows only its name and state, and its sessions come back once it's online again. Tap a session to open it on the phone and carry on; approvals, stopping and files go to that computer, and they keep going there after you switch new sessions to the cloud or another computer. These sessions are fetched live while the app is open and aren't stored on the phone. From the phone you can rename and archive them but not delete them for good. Until the phone has picked a computer once (so it isn't registered yet), each section only shows the computer's state; sessions appear after that.

• New chats show where they will run (in Chat mode too), right above the input box: the cloud or a computer, plus that computer's current state. Tap it to open the "Run on" sheet. The place you pick is remembered: new chats go back to that computer by default only while it's online and available, and use the cloud otherwise. Sessions created on a computer no longer carry the phone's notes vault path, cloud project name, external engine, or image and vision model settings.

• "Open session" in the inbox reminder about an agent waiting for you now opens that session directly, without first switching the phone to that computer.

• When a computer refuses to create a session (for example, remote sessions are turned off there), the message now says why instead of only saying the session couldn't be created.

• With "Allow other devices to connect" on, the desktop app now reports its agent engine state (running, starting, not running) to your account, so the phone can tell "the computer is online but its engine isn't up" from "ready to use". The desktop main window, the web app and the device page don't show these sections or the "Run on" pill.

• On a Mac, commands approved in a remote session can no longer change this computer's Forsion settings. When a task started from your phone or the device page runs commands on this Mac (run_bash and background processes), they run inside a system write protection. They can't write Forsion settings and sign-in data (including the highest approval mode for remote sessions, the allowed devices list and the emergency stop state), agent configuration, skills, plugins and hooks, .tangu folders in projects, shell startup files such as ~/.zshrc, dotfiles in your home folder such as ~/.gitconfig, ~/.config or ~/Library/LaunchAgents, and neither can any program the command starts. The workspace, agents' Library folders and cache folders such as ~/.npm stay writable. Before, with the host sandbox off (the default), a command you approved on your phone could rewrite these files directly, for example to raise the remote session's own approval limit, and the command itself (python3 fix.py, or npm install with an install script) may not have made that obvious. Tasks started on this computer are not affected, and with the host sandbox on, its own sandbox applies as before.

• In remote sessions, tools that set up their own sandbox (such as swift build or swift package) can't start it inside this write protection and fail; the tool result suggests adding --disable-sandbox or running them on this computer. If the system's write protection isn't available, commands from remote sessions are not run and the reason is shown; they never run unprotected. Background processes started on the computer without this protection no longer accept input from a remote session. Each remote command takes about 9 ms longer to start.

• Forsion now reads a project's Git status (the Git status attached when a task starts, and the project details panel) inside this write protection no matter where the task was started, and so do Git commands an agent may run without approval (such as git status). A Git filter or similar program a remote command left in the repository still runs, but can't change the files above; if the protection isn't available, Forsion skips these Git reads and such commands need approval. On every platform, git status and similar commands need approval when the repository's own configuration sets up programs Git would run (filters, Git LFS extensions, fsmonitor, hooks, external diff and the like). On Linux and Windows, Forsion no longer reads the Git status of a repository that configures a filter or a Git LFS extension, and the project details panel shows no Git information for it.

• Linux and Windows don't have this protection: a command you approve on your phone can change Forsion settings on this computer, including the highest approval mode for remote sessions, and the command may not make that obvious. Settings › Remote sessions now says so. On Linux, setting "Local command and file sandbox" to "Workspace writes only" prevents it; Windows doesn't have such a sandbox yet, so only approve commands you understand.

Download Markdown

Forsion 2.11

2.11.4 (2026-09-23)

• New "PROJECT details". In the Tangu Space, when the current session belongs to a local project folder (including the Tangu default folder, but not the vault), the details panel on the right turns into a project page, just like TEAM details: "Agents" lists the agents and teams that have worked in this project, and you can start a new session with any of them or set one as the project default; "Git" shows the branch, how far it is ahead of / behind upstream, changes and recent commits, and opens the project in a terminal with one click.

• Under "Configuration" on the project page, you can edit the project instruction file AGENTS.md (create .tangu/AGENTS.md with one click if it doesn't exist, or let Tangu read the project and generate it for you), manage project skills, view approved plans, and set the project's default agent or team, model, thinking effort and approval mode. These defaults stay on this device and are never written into the project repository; they are applied automatically when you start a new session in this project.

• New side chat (/btw): ask a quick side question with the current session's context. The answer appears in a floating panel that follows the session and is never written into the main conversation, and you can ask even while the main task is running. Type /btw, click "Ask aside" in the text-selection toolbar, or press ⌘; (Ctrl+; on Windows) to open it; if an answer is useful, click "Quote in chat" to attach it to the main conversation's input box without overwriting the draft you're writing. Side chat counts toward your quota too; the web and mobile apps require a server update.

• Shell code blocks in chat can now be run directly. bash / sh / zsh code blocks get a "Run" button that opens a new terminal in the bottom panel to execute them (commands that ask for a password still take input as usual); when it finishes, the output and exit code are sent back to the session automatically and the agent picks up from there. Note that the output is sent to the model as-is. Shell code blocks in notes can be run too, but their output isn't sent back. Desktop only.

• Databases now have record bodies and page previews. Each record can hold a body of text: click the open button in the first column to view and edit it in a side, centered or full-page preview, with previous / next navigation, and the open mode is remembered per view. Also new: list view, view duplication and search across properties. Databases that contain record bodies can no longer be edited by older versions of Forsion, so update all your devices together.

• Smoother database editing and a more consistent interface: cells support typing directly, arrow-key navigation, continuous editing with Enter / Tab / Shift+Tab, and Esc to cancel, and confirming a candidate in a Chinese input method no longer submits by mistake; single-select / multi-select menus can be searched, navigated with the keyboard, and used to create new options; editing a database embedded in a note no longer exposes its Markdown source. Font sizes, selection outlines and icons across the grid, menus, toolbars and record pages now follow one system; auto-fit, grouping and row folding move into the "Layout" section of view settings, and the cover picker has been reorganized.

• Global skill library. The long skill list in Settings becomes a "Global skills" library: grouped by source and searchable, it lets you read SKILL.md and its bundled files, and supports creating, importing from a folder, duplicating, editing, disabling and deleting skills (deletions keep a backup); skills that ship with the app are read-only, but you can duplicate one and edit the copy. Skill details now render their body as Markdown, and code blocks there have only a copy button, no "Run" button.

• Agent management moves into the Agents Space. Creating agents, setting the default, reordering and deleting all happen in the Agents Space, and the roster is now a collapsible, resizable left sidebar; each agent's skill panel separates its own skills from inherited ones, with Auto, Custom or Off modes, and you can disable a specific skill for a single agent. Settings for background agents (Historian / Muse) now take effect only after you click "Save", and your input is kept if saving fails.

• Agents can borrow each other's skills. Coding's web-building skill can now be borrowed on demand by other agents (for example, to build a web page in the Tangu Space), and agents also know which teammates they can hand work to or discuss things with. Lent skills carry a "Shared" tag in the list; for your own agents, add shared: true to a skill's SKILL.md to share it.

• Redesigned onboarding: instead of carrying over settings one by one, it now comes down to a few key choices — connecting services, default model, appearance, working directory and optional device permissions — with entry points to everything else on the finish page. The default model picker is grouped and searchable just like the chat box's model menu, and you get a live preview while picking a style, color scheme and font for the appearance.

• A more compact, consistent interface throughout: font sizes, corner radii and spacing everywhere follow a single hierarchy, and menus and popovers use lighter, closer shadows and open with a subtle fade-and-expand. The raised / flat toggle in "Appearance" stays, with raised as the default.

• Support for GPT-6 Sol / Luna and Claude Opus 5.5. Signing in with a Codex subscription now lets you see and pick GPT-6 Sol / Luna, and thinking effort is sent at the levels each model actually supports (Sol / Luna can turn thinking off); when you connect Claude Opus 5.5 with your own API key, thinking is always on, choosing "Thinking off" no longer errors, and thinking summaries are visible between tool calls.

• Context window capped at 272k by default. Models with a 1M context now use only up to 272k by default; when you need a longer context, switch "Context limit" in the chat box's model menu to the maximum (applies to all sessions with that model; adjustable on desktop only). If you had set a low auto-compaction threshold, compaction will kick in earlier accordingly.

• Feedback replies now arrive in your Inbox. After you submit feedback, each reply is delivered to your Inbox as a message; open it to see the full exchange for that feedback and all its replies, and reply directly, with images, text or JSON files attached. Desktop only; requires a server update.

• Agents no longer stuff work logs into long-term memory: a single memory over 300 characters is rejected with a hint to record it in the work log instead, updating an old entry edits it in place rather than appending one "correction" after another, and Dream memory consolidation no longer keeps failing when there are many memories. Long entries already saved are not shortened automatically; you can trim them manually in the memory panel.

• Session titles arrive faster: within seconds of sending a message, the session is titled from what you said, without waiting for the reply to finish; titles you've changed manually are never overwritten.

• For your own agents, you can rename the folder directly in the header of the details page, and references in related sessions, teams and channels are updated along with it; built-in agents, agents synced to the cloud, agents provided by plugins, and agents that are currently running can't be renamed.

• The mode menu no longer lists every agent in one long row: switching agents now lives in a "Current agent ›" submenu to the right of "Standard mode". The old "Group chat mode" entry has been renamed "Team mode" everywhere, with its description updated to reflect how parallel work happens now.

• Opening skill details from a sidebar such as Tangu details now temporarily covers the current content and offers a "Back" button, and your scroll position and drafts are still there when you return; in the main area it opens in a temporary tab alongside.

• When the chat column is very narrow, the input card switches to a compact style (shorter placeholder, icon-only stop button); in short conversations the last message sits close to the input card instead of across a big empty gap.

• After a reset card is used successfully, a quota restore animation plays: the card lights up, today's and this week's quota refill in turn to their actual values, and the number of remaining cards is shown.

• Images that sit on their own line in a note can be dragged directly, with no need to find the ⠿ handle; a single click still selects the image.

• When you ask Tangu to "Merge conflicts" and a conflict copy is only a subset of the original with no unique content, it now deletes the copy directly (canvas and table notes included) instead of leaving you to delete it by hand. Sync also no longer creates a conflict copy when another device rewrites the same content verbatim.

• Themes: the built-in design language "Zhi" is retired, and anyone using it moves back to Genesis; the color scheme of the same name is now called "Clear blue". Theme cards and Glass settings now have full English text, and agent portraits with hard-to-read text under some color schemes have been fixed.

• Fixed: in direct chats with an agent, the approval mode / thinking effort shown in the input box didn't match what was actually in effect (for example, showing "Approve for me" while actually asking every time per the agent's "Read-only" setting). The input box now shows the mode actually in effect, and /status notes whether the approval mode comes from the session, the agent or the default.

• Fixed: when dragging blocks in columns, blocks couldn't be dropped into a shorter column, the indicator line lingered after release, horizontal and vertical lines appeared at the same time, ⠿ couldn't drag after an image was selected, and handles of right-column blocks were out of reach. Blocks now land exactly where the indicator line is drawn, and releasing with no indicator line doesn't move anything.

• Fixed: dragging a block or file above or below a divider did nothing or landed one position off; dragging into a collapsed section made the block "disappear" on the spot, and that section now expands automatically. After selecting a block with ⠿, pasting or uploading a file inserts it after that block instead of at the end of the document.

• Fixed: after opening a note from the file tree in the left sidebar, the tree highlight, status bar word count and outline stayed on the note shown at startup; on first entering Note Space, the tree highlighted a note that wasn't actually open; closing the current editor made the highlight jump back to the start page. The sidebar chat not automatically referencing the current note in the main area is fixed as well.

• Fixed: with smooth caret on, pressing Backspace at the start of the lower of two empty lines left the caret stuck in place.

• Fixed: when selecting text near the left edge of a narrow sidebar, the selection toolbar was partly cut off.

2.11.3 (2026-09-21)

• New "Creations" Space. What you build in the Coding Space now has a home: web apps and Forsion plugins are listed here. Click to open one, or run it in a separate window, go back to the Coding Space to keep editing, rename it or move it to the trash. Work already published online carries a "Published" badge. If you don't need it, you can turn it off in "Settings → Plugins".

• Add creations to your desktop. Choose "Add to desktop" in Creations to put an icon on your desktop; double-click it to bring up Forsion and open that creation in its own window (installed version only). The desktop icon keeps working even if you rename or move the project folder.

• Web creations' local data no longer resets on restart. Previously, the preview address changed every time Forsion started, so data a creation stored in the browser (localStorage, IndexedDB) was lost. Each creation now has a fixed address, and previewing it in the Coding Space and opening it from Creations use the same data.

• You can now build Forsion plugins in the Coding Space. New projects get a "Forsion plugin" starting point, and plugin projects get a Sandbox panel in the bottom toolbar: click "Load in Forsion" to install the plugin straight from the project folder into the running Forsion, with automatic reload on save; setup errors, view mount errors and the plugin's own console output are collected in the panel and can be sent back to the chat with one click so Coding can keep fixing. Note that this is a "development load", not an isolated environment: the plugin runs in the real app against your real vault, with the same permissions as installed plugins; plugins that declare custom file types need to be installed before testing.

• Coding Studio version history now uses git (requires git on your machine). After each conversation turn, if the project has changes, Forsion saves a version automatically, named after what you said in that turn; you can also save a named version manually at any time. Before restoring an older version, the current state is saved as a backup first, so a restore can itself be undone. Without git, the "Versions" panel prompts you to install it; previously saved snapshots can still be found and restored at the bottom of the panel. If the project is already your own git repository, Forsion only shows its history and never modifies it.

• Coding now knows how to handle plugin projects: it reads the Forsion plugin development skill first and guides you to load and debug with the Sandbox panel.

• With "Full access" selected in a team session, members no longer ask for approval every time. Previously, team members kept the approval mode from their own settings (such as "Auto edit"), or the mode in effect when the team started working, so even after switching to Full access, members running commands or writing files still asked you for approval again and again, and file writes were often flagged as "Write outside workspace". Members now always follow the approval mode currently selected for the team session.

• Switching the approval mode takes effect immediately. After you switch modes in the input area, a running task follows the new mode from its next action, without waiting for the turn to end; switching in a team session applies to all members at once.

• In a team member's sub-chat, the approval mode shown and changed is the team's, with a title stating that "changes apply to the whole team" (previously it showed the member's own setting, which had no effect).

• An approval mode that failed to save is now reported honestly: for example, if the engine is temporarily unreachable, the input area falls back to the mode actually in effect and lets you know, instead of showing an unsaved mode while tools still run without approval.

• Session settings are now saved per item: changing thinking effort, plan mode, approval mode and so on in different windows no longer writes settings just changed elsewhere back to old values.

• Installing GitHub-hosted plugins from the Market is more reliable: downloads try a direct connection and several proxy mirrors in turn (mirrors first when "Use China mainland mirrors" is on), and each address has a timeout so it no longer spins forever; the install button shows connection / download progress, and failures explain the reason right in the Market along with network suggestions.

• Action results in Settings and Feedback are now visible: the success or failure of actions such as uninstalling a plugin / Space or rescanning plugins shows at the top of the panel instead of passing silently; "Send test notification" pops up a real notification card in the main window.

• "Reset layout", "Uninstall Space" and "Diagnose in chat" run from the Settings / Feedback windows now act on the main window (previously they only changed the Settings window itself); achievement unlock pop-ups triggered in Settings also play in the main window. Command buttons in plugin details are now offered only in the main window's ⌘K.

• The backend no longer spawns a pile of engine processes when the engine exits during startup. Previously, if the engine silently exited before becoming ready, the desktop would stack two restarts at once, leaving multiple orphan processes and ending up showing only "Crashed"; it now retries along a single chain, and won't relaunch the engine while quitting, clearing data or installing an update. The engine writes to the backend log which startup phase it stopped at, visible in feedback exports.

• With Docker installed but not running, startup no longer prints two warnings with stack traces. This check is no longer printed when the sandbox isn't using Docker; when a Docker sandbox is in use, it reports a single line with the reason it can't connect.

• Agents that share the default memory are no longer asked for approval every time as "Write outside workspace" when writing material to their own Library.

• The timeline in feedback exports now includes the reason for each approval and the mode in effect at the time, making "why is it asking me again" easier to investigate.

• In the English interface, failure reasons for cloud sync and for uninstalling Spaces / plugins no longer show the original Chinese text.

• Fixed: the line-number column of code diffs in approval cards overflowed the card and covered the "Approve" button, making it unclickable.

2.11.2 (2026-09-21)

• New built-in "Image Studio": create an Image Studio from the main interface to generate, edit and organize images. It supports text-to-image and reference-image editing, cropping, rotating and flipping, color adjustment, layers and blending, transparent backgrounds, scene extension, and exporting the whole board. The canvas has a full context menu ("Download image" / "Download original" / "Export selection" / "Add to chat" / "Layer order" / "Lock" / "Delete"); downloaded images include your current edits, while "Download original" keeps the original file.

• Image generation now lets you pick a rendering quality, and sizes no longer come out wrong. The quality level (low / medium / high, with higher levels on some models) is passed all the way from the interface to the model; ratios such as 16:9 are converted using the size table each model family actually supports, so you no longer get cropped images like 1024×576. New image models from Seedream and Qwen are supported.

• Image generation is billed per image. When a single request asks for multiple images, pricing follows the number of images actually produced (previously only one request's worth was charged); different quality levels can have different prices. Requires server 2.3.15.

• Early warnings when your quota is running low. The top of the Chat View warns in three steps at 15% / 10% / 5% remaining, based on whichever of today's or this week's quota is tighter, with a separate notice when it runs out, plus "Upgrade membership" and "Use a reset card" entry points; you can dismiss the current level, and it reappears when you reach the next, tighter one. The notice disappears as soon as quota is restored. Requires server 2.3.15.

• A dot-matrix wave animation while generating images: the generate_image row in progress is no longer a plain tool audit card but a full dot-matrix wave that follows the theme colors; click it to play Snake (arrow keys / WASD / touch swipes, you can wrap through the left and right walls, click again to exit). With the system's "Reduce motion" on, it falls back to a static dot matrix. Only the running row is replaced; other tools running in parallel display as usual.

• Changing theme / font / zoom / smooth caret / interface language in Settings now updates other windows immediately. Settings has been a separate window since 2.11.1, and until now these changes only applied to the Settings window itself.

• Plugin setup guides are now checklists that can actually be verified. Previously, almost every plugin popped up a card after installation, mostly restating its instructions, and clicking "Finish setup" merely closed the card. Now only plugins that genuinely need setup before they can work show a card (for example, a countdown with no date set, calls with no server configured, or computer use without macOS permission); the card lists "what's still missing" item by item for you to fill in on the spot, and the host actually checks whether each requirement is met. Other plugins' instructions move to "How to use" on the plugin's details page, with nothing left out.

• Databases get "Row folding". The table view toolbar gains "Row folding": check the fold keys (only rows with the same values in these columns fold together), and optionally choose a time column and time window (in minutes, default 30, aligned to your local clock); rows within the same window collapse into a single summary row: numbers are summed, columns with more than one distinct value list the values they contain and their counts, and dates show "earliest – latest"; click a summary row to expand it back into individual rows. Sorting applies to the summary values; it can be combined with property grouping (folding within groups) and has no effect in hierarchy view. It only changes presentation, not data, and the configuration is saved with the view. Plugin tables can use it too via TableSpec.fold (ctx.table.caps.fold).

• The auto-compaction threshold for long conversations is now adjustable (Settings → Models). With a 1M context window, the default trigger was so high that compaction almost never happened; you can now set a percentage at which compaction starts. The context progress bar updates as soon as compaction finishes, without sending another message.

• Teams no longer say the same thing twice. Repeated speech by a member within the same activation is blocked; a Chinese sentence ending in 。DONE is correctly recognized as finished; and when there's nothing new, the next member is no longer woken up for nothing.

• Diagnostic attachments in the Feedback panel can be exported to a local file: click "Export log file" before submitting to keep a copy yourself, and attachments in the admin Feedback Center have their download button back (requires server 2.3.15).

• Settings and Feedback panels now attach to the session you're viewing: exporting session logs from "Advanced › UI & sessions" no longer exports the first session in the list.

• Agent Desk supports a 3D companion avatar (with the Live3D plugin); avatars are bound per agent with an adjustable idle pose, and the companion is captured when an agent takes a screenshot.

• Live island on Android: while a session is running, progress shows in the status bar / live island, and it tucks away automatically when done.

• The three command palette entries in Developer options (mobile UI preview, live activity log view, foreground window sampling) now appear in ⌘K as soon as you turn them on, without restarting the app.

• Fixed: the Node bundled in the Windows installer didn't include npm, making external engines (codex) sit idle for 30 seconds at startup.

• Fixed: the "Paste as" menu occasionally flew to the top-left corner of the window.

• Fixed: malformed requests are no longer reported as server errors (requires server 2.3.15).

2.11.1 (2026-09-19)

• Cloud sync per-file limit now scales with your membership tier: Free 5 MB, Plus 10 MB, Pro 500 MB (note body text stays at 5 MB). Large attachment uploads and downloads no longer get cut off by a fixed 2-minute timeout.

• New sixth panel type: Floating Panel. Settings, Market, Achievements and Feedback no longer take over the main window; on macOS / Windows they open in separate windows you can drag, resize, minimize and close, and on the web they appear as a fixed, centered floating panel. Floating and Mini Panel are both open to plugins. Auxiliary windows no longer replay the startup animation.

• Attachments can be copied straight out: select an image or attachment in a note and copy it, then paste it into external chats, documents and folders (on macOS both the native image and the file are written). Images and attachments stay rendered when the cursor passes over them, and the source only opens when you hover </>; image dragging, "Move to new column" and resize handles are fixed as well; the outline relationship graph no longer resets its layout on regular saves or cloud sync.

• Editor save fixes: strikethrough turning into literal ~~ after a while, the second marker failing when italic sits right next to bold, and [[links]] after a code block in a list being saved as broken \[\[ links are all fixed. Chinese punctuation directly next to ** / * / ~~ is now parsed with CJK-friendly rules (e.g. a bold label ending in a full-width colon, followed immediately by more text, now bolds correctly), consistently in the editor and chat bubbles.

• Search, tags and wikilinks now read the decoded body text: character references no longer make search miss words, and a multi-character Chinese tag is no longer cut down to its first character.

• Agent details reorganized: "Growth" combines memory and evolution, there is a separate "Schedule" tab, and built-in skills are collapsed by default; a new "Evolution" tab shows working notes, candidates awaiting review and nomination reminders, and skills you created yourself carry a "Self-built" badge. TEAM supports image avatars, and names and bios are edited in place in the details header.

• Agents take better care of themselves: after upgrading, self-evolution auto mode, Dream automatic memory consolidation, Muse and Historian are on by default (if you turn them off later, that choice is kept); changes to an agent's name or bio take effect on the next turn.

• Thinking effort now defaults to "Medium" across the board (Arioso / Aria presets, the chat default and Muse's background cycle); Muse's default token budget goes from 100k to 1M.

• Fixed: Muse's schedule was always empty: the original Muse instructions written at install time by older versions are upgraded once, so Muse can schedule tasks normally.

• Grok Build (xAI subscription login) compatibility: requests now go through Grok Build's CLI proxy and headers, and old credentials are migrated automatically.

2.11.0 (2026-09-17)

• The session sidebar gets a new "Orbits" view. Direct chats with an agent, external engines, teams and projects share one list sorted by latest message time, and the ones you use often can be pinned with right-click "Pin to top". The old sidebar is renamed "Sessions (legacy)", and you can switch back at any time.

• Teams can now work in parallel. Pick two or more agents above the input box to enter team mode, or set up a long-term team. Members start working at the same time, each in their own sub-session, while the team session only coordinates; Team Desk shows what each member is doing and lets you open the details. Members who say they're done stop speaking, and the session wraps up only once everyone is done; when a member needs approval, the request goes straight to the main chat, and public outputs sync back to the main session.

• In a direct chat you can @ a project to dispatch a task: the agent opens a new session in that project to do it.

• New Agents Space: view agent and team details and assign members to teams; sub-sessions can be viewed in full in the side panel.

• Built-in personas are now Arioso, Aria, Recita, Coding and Muse, and the old built-in agent roster entry is removed.

• Realtime voice supports hands-free conversation: pause after you finish speaking and it sends automatically, no button needed; the call continues when you switch to another view and hangs up when you open another existing session. If sending fails, what you said goes back into the input box.

• Context compaction for long conversations is rebuilt: the compaction point is saved, so reopening a session no longer compacts again; compaction kicks in early when nearing the limit, and when the context overflows it compacts automatically and retries. Prompt cache hit rates are higher too, so you use fewer tokens.

• This turn's stats appear while waiting for a reply: elapsed time, token count and thinking time; the blinking cursor at the end of streaming output is gone.

• Feedback panel upgraded: open it by typing /feedback or clicking the feedback icon at the bottom of the Ribbon; choose a feedback type and optionally attach diagnostics (version, UI state, errors, runtime timeline and usage) and the current session's contents.

• Prevent sleep while sessions are running (Windows / macOS, off by default): once enabled in "Settings → General", your computer won't go to sleep from inactivity while a session is running, and returns to normal once everything finishes.

• Home wallpapers add a Hack2Gate poster.

• Open file tabs follow renames and deletions: after you rename, move or delete a file in the tree, open PDF, image, whiteboard, database, plugin file and dashboard tabs follow it to the new path or close automatically, including tabs collapsed in the sidebar and in other windows; tabs for attachments deleted along with a note, or files deleted along with a reference block, close too.

• Dashboard fixes: no more endless loading skeleton after a rename or delete, icons no longer vanish en masse after the file structure changes, and multi-block column dragging is supported; dashboard tabs also support back / forward, appear in Recent and are highlighted in the tree.

• Switching to another Space and back no longer loses the file view; when focus leaves the main area, the current main-area tab is still identified correctly.

• Cloud sync guards against accidental deletion: a large burst of deletions in a short time is paused first and only synced after you confirm (requires server 2.3.12). Fixed: per-item sync silently disappearing after upgrading to 2.9.9; the same account now reconnects to its original sync automatically.

• When you have Tangu merge sync conflict copies, changes from both sides are kept; canvases and tables are not merged automatically.

• Multiple windows or processes changing settings at the same time no longer overwrite each other; the external MCP server no longer hangs on startup.

• Muse's token budget is counted by billed usage, excluding cache hits; background calls no longer refresh the current session's context progress bar.

• Mobile: creating whiteboards and dashboards in a local vault and pasting images or files in the editor work again; files and feature views opened in the single-column layout are recorded in back history and Recent.

• Line breaks in the Inbox are finally line breaks. A single line break in server broadcasts and agent messages used to be merged into the same line (standard Markdown semantics); messages now display line by line, like a letter. Blank lines still separate paragraphs.

• The Inbox supports images. Illustrations in the body of Forsion broadcasts now display directly.

• Android now gets update notifications. It checks for a new version at launch and shows an "Update" page when one is available, and "Download update" fetches the installer directly from an address reachable in mainland China; previously Android had no update prompt at all, and "Check for updates" in settings only opened the website.

Download Markdown