gugu for macOS release notes

Managed setup, Apps, projects, and files

Background agent setup, Luna without membership, an improved App market, cross-device projects, spreadsheets, and conversation fixes.

Start with a managed agent

Choose Codex, Claude, or both during setup. Gugu installs and manages them in the background while you continue; notifications keep the installation progress available. New sessions start in General mode with Codex recommended. GPT-6 Luna is available without a membership and uses credits. Existing explicit preferences are preserved.

A reorganized App market

Featured Apps, categories, search, and installed Apps are together. Find an App for a file, manage its version, and enable or disable it from its details. Gugu Office and Search are delivered through the market and can provide tools to your agents.

Projects across folders and devices

Attach several related folders to one project and switch their working context together. Connect an existing local directory and browse projects on other devices. Removing a directory preserves the existing conversation history.

Work on files and spreadsheets with an agent

Chat folders now support subfolders, moving, and drag and drop. Recent files, search across conversations, and references to specific lines help you find and describe the change. Open an .xlsx in Gugu Office, select cells, and ask an agent to edit them; dependent formulas are recalculated when saved. Attachment uploads, forwards, and downloads to the agent's machine have also been updated.

Connected conversations, browsers, and computer control

Edit an earlier prompt and resume from it while Codex retains its thread. Plan approvals use a single card; consecutive tools and child agents are easier to follow. The built-in browser previews office files and imports login state from your default browser. With the Computer Use App, task previews show computer interactions.

Other improvements

• The local Gugu assistant has improved setup, dedicated settings, streaming replies, and task handling.
• Added ZCode and CodeBuddy; more agents' saved histories appear in the task timeline.
• Save and switch subscription accounts with separate login data.
• App programs and data follow their host machine; supported Apps also open on phones.
• Reorganized settings and moved the update action to the bottom rail. Windows uses a single frame and lighter scrollbars.
• Reduced repeated avatar, icon, and contact work and removed unnecessary package resources.
• Fixed duplicate Claude replies, Codex child-thread output appearing in the parent conversation, repeated draft submission, and several Windows startup and terminal connection issues.

Platform models now have a backup route to improve availability when an upstream provider runs out of capacity.

Updating

The way group files are written has changed. Update the devices you use for file collaboration to 0.2.57.On older Mac versions, the updater opens the installer. Drag the new Gugu app into Applications to complete this update; later versions continue to update in the app.

Animated bird avatars, and decks you can edit in the group folder

Agents now have an animated Gugu bird as their avatar, each with its own color and face. Open a .pptx in the group folder, pick what should change and hand it to an agent; it comes back as the same file. Two tabs can sit side by side in a task, one mode button switches between Chat, Code and Work, and Settings has a new Subscriptions page.

Agents get an animated bird

Agents without a photo now have an animated Gugu bird as their avatar. Each one has its own color and face, your agents avoid sharing a color where possible, and the bird is the same on your computer and your phone. Its body shows what it is doing: three dots hop while it works, it stretches into a "!" when it needs your approval or an answer, it slumps into a puddle on an error and bounces when it's done. On the Look page in the agent's settings you can change its color and face, or shuffle one, and nothing is saved until you click That's the one. In a direct chat with an agent, the bird in the middle of the header is bigger and hops when you hover it. Agents with a photo keep their photo.

Edit decks in the group folder

The group Workspace is now called Group folder, and Folder in direct chats. Open a .pptx there and the new Gugu Office draws every slide: click to pick a spot, drag across the slide to box an area, Shift-click to add more, then click Put in composer and @ an agent. What the agent saves is the next version of the same file, with everything you didn't ask it to touch left as it was; if you aren't picking anything at that moment, the open deck switches to the new version on its own. The tools for editing decks come with Gugu, so there is nothing else to install.

Two tabs side by side in a task

When an agent opens a file or a web page for you, it opens next to the conversation instead of covering it. You can split it yourself too: right-click a tab and choose Split right or Split down, drag a tab to an edge of the content area, or press ⌘\. Any two can sit side by side, including two files, two web pages or two terminals, and you type into whichever side you click. To go back, hover the line between the two paired tabs and click Separate views; after that, whatever an agent opens just adds a tab and doesn't take you away from what you are looking at.

One mode button: Chat, Code and Work

There is a mode button at the right end of a task's tab bar and at the right of the header in a direct chat with an agent, replacing Open in Tasks. In a task, Chat takes you to the direct chat with its agent; in the chat, picking the mode the task is in takes you back to the task. Switch to Code mode and Switch to Work mode ask first, because the turn that is running gets interrupted before the task carries on in the other mode. Switching no longer breaks up a split screen, and the task's tab stays where it was.

A Subscriptions page in Settings

Settings has a new Subscriptions page under Usage. For Claude Code, Codex and the others installed on this computer, it shows whether each runs on your subscription, your own API key or another service, the plan name, and how much of each quota is used and when it resets. It is read on this computer and nothing is uploaded. The vendor limits how often Claude's usage can be checked, so it is only checked when you click Check now.

Also improved

• A failure is reported once, with the reason: a model that needs a membership, a model that isn't available right now, credits that ran out or an expired login each get one sentence on what to do next, with a button where there is one to press (such as start a membership or switch models), instead of a raw English error. When a model isn't available, switch models in the same task and send again; no need to start a new task.
• When Claude Code in Work mode with Full Access goes into plan mode on its own, the plan now waits for your review instead of being approved automatically and carried out.
• When an agent changes a file in the group folder, it saves a new version of the same file instead of adding another file with the same name. If two agents change the same file one after the other, the second one has to fetch the latest version first, so it can't wipe out the first one's changes.
• Open tabs follow changes on disk made by an agent or another program: HTML and Markdown previews, images, PDFs and local pages in a browser tab all update without closing and reopening them.
• The Running indicator on a tab no longer overlaps the close button.
• When Gugu and your system use different light and dark themes, opening an App or a file from the group folder no longer flashes black or white first.
• The first time you install an App from a card in a group, all agents on this computer can use it by default. After you give an App to a group, tasks already running there can use it right away, without restarting. Agents can also install an App that was just added to the market, without you opening the market first.
• If an agent's connection drops, calling it in a group again reconnects it and it answers, with no need to reconnect it from the task page. A failure is said once in the group, and no Retry bar is left behind once the agent is back.
• When Codex runs on a Gugu-provided model and a reply is cut off midway, it retries within a second instead of waiting 5 minutes.
• If you signed out elsewhere or changed your password, opening Gugu takes you straight to the sign-in page instead of showing "Can't reach the server right now" indefinitely. Starting up before sign-in has finished no longer shows "Could not restore the opening task", and reloading in the first few seconds after launch no longer brings up a "Gugu can't start" dialog and quits.
• A terminal task on another machine that you are watching from this computer no longer refreshes the whole screen after sitting idle for about ten minutes.
• When an agent builds an App, it can try the page out the way a person would, clicking, typing and pressing keys, continue on the same page across checks, and take screenshots at the width the App really opens at. What it does while trying the App is real and done in its own name.

Split View, separate windows and an App market

Put up to 4 tasks side by side, and drag a task or a tab out into its own window. Apps now have a market where one click installs them. Models follow your membership tier, and machines added through Remote SSH work on your phone and other computers too.

Split View: up to 4 tasks side by side

⌘-click a task in the sidebar to open it on the right, or drag it onto the edge of the task you are looking at. Tasks you split stay together as a group in the sidebar: click a task outside the group to view it on its own, and click any task in the group to bring the whole group back.

Drag out into a separate window

A task, an App, or a single tab (terminal, browser, files) can be dragged out of the window into a separate window, or opened with right-click → Open in new window. Click Back to Gugu in its top-right corner to put it back where it was; whatever is running in its terminal keeps running.

App market

Click Apps in the far-left rail to open the market. It has featured picks, and every App has its own page showing what it does, which permissions it needs and where its code lives. One button takes you from Install to Open, and Apps with a page of their own are pinned to the rail once installed. Gugu also comes with 全能搜索 (Omnisearch): add a key for a search service, kept only on this computer, and your agents can search and read the web.

Models follow your membership tier

The model list only shows the models your tier can use: the free tier includes a few low-cost models, and a membership unlocks all of them. Hover a model to see its rate and context length, or hover Upgrade to unlock more models at the bottom to see what a membership adds. Credits decide how much you can use, not which models; if your membership ends, your tasks stay put, so switch models and keep going.

Remote SSH machines, everywhere you are

Machines added through Remote SSH now show up on your phone and your other computers, and you can start tasks on them there.

Also improved

• Settings, Usage and Send feedback moved into the avatar menu at the top left, and ⌘, opens Settings.
• Switching between your 6 most recently viewed tasks no longer reloads them.
• Delete task is now Archive task; archived tasks can be found in Task history.
• Tasks started by an agent or an App open normally after sitting idle, with their replies in place.
• A task that has gone idle no longer keeps the yellow "waiting for you" dot in the sidebar.
• Diagrams in conversations render properly.
• After a brief network drop, chats catch up on their own.
• After you change your avatar, you see the new one too.
• About shows download progress for a new version, and from the next update on only the changed parts are downloaded, so updates are much smaller.
• Claude Code can read web pages when it runs on a Gugu-provided model.

Before you update: the first launch after updating backs up this computer's data and then migrates it once; everything works as before. This version and 0.2.52 cannot connect to each other, so update all your computers and phones together. If an App an agent built for you earlier now says it can't be read, ask the agent to rewrite it for the new version.

Private chats get a workspace, and Apps open in the chat's right panel

The first start after the upgrade migrates the data on this computer once, making a full backup of the local data first; everything works as before. Private chats now have a workspace and Apps of their own, and an App opens in the chat's right panel whether it is a group, a topic or a private chat; select text in the workspace and you can cite it into the composer, to send yourself. Settings are now one place, opened from the far-left rail. When two machines run versions of Gugu that do not match, Gugu now says plainly that one needs an update. Several messages sent to Codex in a row arrive in one go, and each Hermes session can set its own thinking level, model and approval mode. A full-screen terminal opened late no longer shows a stale screen and a full-screen right panel no longer garbles the terminal; memory use is down again, and with a session page open, the memory used for drawing the screen drops from about 0.9 GB to a few tens of MB.

Sessions

• Note: on the first start after the upgrade, Gugu migrates the data on this computer once; before it does, it makes a full backup of the local data next to the original, taking about the same space. The migration finishes on its own during startup without showing anything, and the sidebar, sessions, terminals, tabs and attachments are all as they were. In the command palette, the "Recent conversations" group that repeated "Recent tasks" is gone.
• Fixed: a session that an agent started in a named project, or one created from the phone in a project, did not appear under that project but in the sidebar's Tasks section. Both now sit under their own project, and sessions misplaced this way before the upgrade are moved back during the migration.
• Improved: a child session that an agent starts in another project can now sit under the session that started it, with its project marked at the start of the row (a folder icon; hover for the project's name), the same way machines are marked. The agent decides from what you asked: sessions sent off to help it sit underneath, and sessions it opens for you to use yourself are listed on their own.
• Improved: when two machines run versions of Gugu that do not match (for example, this one is new and the other is still on 0.2.51), things used to fail only at the first action: creating a session showed an internal error and lost the draft, or the whole page crashed. Now the machine picker marks that machine "Too old", and choosing it shows "Gugu on … is too old. Update Gugu there (a Gugu server updates itself when idle)" under the composer, keeps your draft and sends nothing; the machine that has not been updated only shows that it cannot connect, and never gets halfway through an action. A Gugu server switches to the matching version right away, and the connection comes back on its own once it has.
• Fixed: a full-screen terminal program such as Claude Code could show a stale screen when its terminal view was opened late (the session ran in the background and was opened later, after a reload or reopening, or when attaching from another computer): a strip at the top froze, old frames piled up as you showed or hid the sidebar, and a multi-line paste was entered one line at a time. Attaching now gets the whole screen as it is at that moment.
• Fixed: going full screen in the right panel used to squeeze the left side narrower and narrower until it vanished: the terminal was pushed down to 2 columns and the agent re-laid out its whole screen at that width, the terminal's content was often garbled after restoring, and the whole thing was slow. Full screen now covers the left side without touching it, so restoring brings it back exactly as it was. The group's right panel works the same way.
• Fixed: a file previewed in the right panel, such as an HTML page, reloaded and jumped back to the top after switching to another tab and back. Recently viewed previews now stay loaded (up to 3 besides the one on screen) and come back where you left them.
• Improved: sessions you switch away from no longer hold on to memory: beyond the 3 most recently viewed, a session you leave lets go of its conversation and reads it again when you come back (about 0.3 seconds for the heaviest sessions), landing on the latest page: on the line where you stopped if that line is on this page, otherwise at the bottom.
• Improved: a terminal now keeps up to 10,000 lines of scrollback (100,000 before), so each full terminal uses 85–90 MB less memory and resizing the panel no longer stalls. Earlier output is in the agent's own session record.
• Improved: the session page no longer refreshes in full every second (only while the Outputs panel is open does it tick by the second), and the frosted glass under the composer is now a solid colour. Together, with a session page open, the memory used for drawing the screen no longer sits at about 0.9 GB but drops to a few tens of MB.
• Fixed: once the model a session remembered was taken off the list, the session could not start, and switching to another model did not work either; it only showed a short error in English. You can switch now: after you confirm, the session restarts on the new model and the conversation continues. The line saying it could not start now gives the reason, and the model button shows the remembered model, greyed out in the menu with a note that it is no longer among the available models.
• Improved: when you expand a turn, its tool calls are listed one by one right away, instead of being folded into "ran N shell commands" that needed another click.
• Fixed: the label of a thinking row (such as "Thought for 3s") could be cut short by the summary after it. The summary now gives way first when space runs out.
• Improved: tool names in the conversation: an App's tool shows the tool it actually called instead of the same name for every App tool, and an MCP call reads "server name · tool name", in Claude Code and in Codex alike. The approval bar shows the same name while it waits for you; in Codex it used to be a long English line.

Apps

• Added: private chats have a workspace of their own: in a private chat with a person or with an agent, the header has a Workspace button with that chat's own cloud files, the same copy on the computer and on the phone. Apps work in private chats too: installed from a card, they open in the private chat's right panel, and an agent can build an App in a private chat, keeping its data file in that chat's workspace and posting the card right there.
• Fixed: Open on an App card in a group used to jump, for any App that does not open group files, to whatever page the other side was last on (often App management), without opening the App at all. The App now opens in this chat's right panel: an App with a page of its own opens that page, and one that only opens files goes to its files; in a topic it opens in the topic pane. An App that only opens files in sessions says "Installed — it opens its files in a session." on its card.
• Added: select some text in the workspace's source or preview and "Copy · Cite" appears beside it. Cite puts the text, with a link to the file, into the current chat's composer (with line numbers when selected in the source), leaving full screen for the side panel first; you @ whoever you like and press Send yourself, so what goes out is your own ordinary message.
• Improved: the workspace file list is now a single list: Recent activity below it is gone, so the panel has one scroll bar; files are ordered by last change, newest first; and each row's second line says who created it and how long ago it changed. "Open with another App", "Rename" and "Delete" sit in a ⋯ at the end of the row (shown when the pointer is on that row), and Delete asks for "Confirm delete" in the menu. When the right panel is wide or full screen, the same list becomes a table with Name, Created by, Changed and Type, and the header stays put as you scroll.
• Note: if you picked "Always use this App" for group files, pick it once more in this version: if a file does not open with the App you chose last time, choose it again under "Open with another App" and tick "Always use this App".
• Note: the App format has been upgraded, and Apps written in the old format cannot be read in this version: App management shows them under Needs attention as Unreadable, and their cards in chat say "This App package cannot be read.". To keep using an old App, ask an agent to rewrite it in the new format.

Chat

• Improved: when picking members for a new group chat, People now come first and Agent after, so you no longer scroll past a long list of agents to find people.
• Improved: when someone asks to join a group that needs approval (for example, on behalf of an agent), the group owner and admins used to see it only after reopening New requests. The red dot on Contacts and the join request under New requests now appear straight away.
• Fixed: when searching messages in a chat, the matching word sat on a line of its own, splitting one sentence into three lines. It now stays on the same line as the words around it.
• Fixed: a message starting with "> " (such as a passage cited into the chat, or an agent's reply that opens with a quote) was taken apart into a small quote strip plus the text, and file links in it showed their raw text. It is now shown as Markdown; the composer's own reply quote works as before.
• Improved: Group announcement and Group members in group settings used to close group settings and open a separate panel with only a close button. They are now pages inside group settings, with Back at the top left to return to group settings.
• Improved: a wide table in chat or in a session now fades at its right edge while there is more to the right, and the fade disappears once you scroll to the end.

Settings

• Improved: settings are now one place. The far-left rail has two new entries at the bottom, Library and Settings: Settings opens the same settings dialog wherever you click it, grouped as Account · Usage · Activity / General · Interface · Permissions · Models / Sessions · Repository · Browser · Devices / About. The settings that used to open from the bottom of the sidebar (automatic naming, auto-approval, notifications and so on) are now under Sessions, and the theme is only under General. Library is a full page, without the session list beside it. The two entries at the bottom of the session sidebar are gone, and opening settings from a shortcut or the command palette also opens this dialog.
• Note: the Storage page is gone from settings: it measured only the worktrees sessions carry, so any session running in the project's own folder showed 0 B. To delete a session together with its worktree, delete it from the session list, which does the same thing.

Agents

• Fixed: several messages sent to Codex in quick succession while it was working went in one by one. They now arrive together at the next pause between tool calls, as with Claude Code.
• Fixed: in Codex Work-mode sessions, /compact was sent to the model as a sentence, and the model answered that it had compacted when nothing had been compacted. It now really compacts and draws the "Context compacted" divider afterwards; automatic compaction in Work mode draws the divider too.
• Fixed: Claude Code with one of Gugu's 1M-window Claude models showed the context window as 200K and compacted automatically at about 167K. It now shows and compacts by 1M.
• Improved: Hermes now connects the same way Hermes' own desktop app does: each session can set its own Thinking level, model and approval mode, both when you create it and partway through; "Send now" on a queued message goes into the current turn; questions Hermes asks you come up as options to choose from; and approvals offer Hermes' own choices, so allowing for this session and allowing always are two separate options instead of both showing as "Always allow". This needs a recent Hermes (v2026.9.24 works); with a version that does not match, you get an error saying so.
• Fixed: in a multiple-choice question waiting for your answer (such as options from Pi), every option showed as "Allow once". Each now shows its own name.
• Improved: Antigravity no longer carries a Beta badge. When the keychain is locked (for example, when signed in over SSH), Antigravity keeps its sign-in in a file, and Gugu used to report "One more step" for Code mode because of it; Gugu now asks Antigravity itself whether it is signed in and shows "Signed in".
• Improved: with only the system proxy on and TUN mode off, agents did not use the system proxy and connected to the internet directly (Antigravity, for example, went straight to Google), so they often could not connect. Gugu now adds the system proxy to the environment of agents and their installers; proxy environment variables you set yourself are left alone, and terminals you open yourself are not affected.

Other

• Improved: the App page and the Library pages no longer have a divider under their header, and the description line at the top of the App page is gone.

Extensions are now Apps, and the group workspace is one too

Extensions are renamed Apps and get their own place on the far-left rail, where you can pin the ones you use most. The group workspace is now the built-in Workspace App: people and agents in a group edit the same files together, and Apps can be shared as cards in chat and installed with one click. Hermes is added. Images open in a single viewer, and ⌘F finds text in a session.

Chat

• Improved: switching back to a chat lands on where you left off as soon as it opens, without a flash, and scrolling up through history no longer jumps now and then.
• Improved: images open in a single viewer, both in chat (bubbles, topics, chat history, image grids) and on the task page. It zooms out of the thumbnail and back into it when closed; double-click zooms in at the pointer; the arrow keys move to the other images in the chat; and long images fill the width and start at the top.
• Added: six reactions for work status: Got it 👌, On it 👀, One moment ⏳, Done ✅, Question ❓ and No ❌. The picker has two rows and shows each reaction's name on hover. When someone reacts to an agent's message, the agent is notified, as information rather than a new request.
• Added: right-click the avatar next to a message in a group chat to @ that person in the composer.
• Fixed: an @ right after a letter (for example "detail@Claude 84") did not count as a mention, so the message reached no one and the suggestion list did not open. It now counts and the list opens; an email address such as alice@gmail.com still does not. With both Claude and Claude 2 in a group, "@Claude 2" used to reach both; it now reaches only Claude 2.
• Fixed: the @ suggestion list did not scroll along as you moved through it with the arrow keys, and in group chats Esc could not close it, so the only way out was deleting the @. Both work now.
• Improved: group settings, the group announcement and direct-message chat settings now open in the right panel, side by side with the messages like the member list, instead of floating over them.
• Fixed: "⋯" → Settings in a direct-message header now opens that chat's own settings (mute, pin, search, clear). In a direct message with an agent it used to open the management page for the agent's whole account; agent settings now open from the name in the header.
• Improved: the chat list on the chat page can be collapsed, just like the task page sidebar: once collapsed, move the pointer to the left edge to peek at it and click to pin it open.
• Improved: Contacts search was rebuilt. You can search by pinyin, original name, or a full Gugu number or group number; results come in three groups (people, agents, groups) with the matching characters highlighted, and a second line says why a result matched when it was not by name. ↑↓ and Enter pick a result. When a string of digits finds nothing, you can look it up as a Gugu number directly.
• Added: Contacts → New requests now lists your request history, sent and received together, with each row's status; click a row for details.
• Fixed: the profile card of someone else's agent showed a Message button that did nothing. It is no longer shown, and if a chat cannot be opened you are told so.
• Fixed: when the last message in a chat was an image or a file, the chat list preview showed the attachment's raw text and download link. It now shows the file name.
• Improved: the hover card in the running-status bar now shows what the agent has been doing recently, in time order, instead of only what it did for this chat; a message from somewhere you cannot see shows as "A message from somewhere else". Its thinking and tool arguments and output still never appear on the card.
• Improved: group and topic names have a visible edit button next to them.
• Note: the "Replies without being named: …" line at the top of the member list is gone; each agent's row shows Active / Passive instead.

Topics

• Added: topics have a workspace of their own: the topic pane's header has a Workspace entry with the topic's own files, kept apart from the group's.
• Fixed: after picking someone from the @ list with Enter, or sending with Enter, in the topic pane, the cursor jumped back to the main composer. It now stays in the topic's composer.

Apps

• Improved: extensions are now called Apps. The far-left rail has an App entry next to Messages and Contacts, which opens My Apps and the market; the Extensions tab in the Library is gone. An App with a page of its own can be pinned to the rail from its row ("Pin to the left rail"), pinned Apps can be dragged into order, and an App with a badge shows a red dot.
• Added: the group workspace is now the built-in Workspace App, which opens in the group's right panel as a file manager. Clicking a file opens it with the App that handles that kind of file; every file can be opened with another App; when several Apps can open it you are asked which one, with an "Always use this App" option. Markdown and JSON files open in the Workspace App itself, as a preview by default when the panel is narrow. The right panel remembers, per group, which App and file were open, so they are still there when you come back.
• Added: Apps in a group open the group's single copy of a file, so members and agents can edit it at the same time and see each other's changes live; changes made offline are merged in when you reconnect.
• Added: the group's right panel can go full screen, and its divider can be dragged; the width is remembered, and dragging it very narrow closes the panel.
• Added: an App can be sent into a chat as an attachment and shows up as a card. The card can be forwarded to other groups or people or exported as a file, and Install on the card installs it right there, without switching views. The card shows Install, Update to x or Open depending on the version you have, and offers only Open when the card is older than yours, so nothing is downgraded. An App that asks for no permissions installs straight away; one that does shows a consent sheet saying who it is from, which permissions it wants, and that it has not been reviewed by Gugu.
• Improved: when the same publisher ships a new version of an App without asking for more permissions, it updates on its own, open pages reload in place, and a short notice tells you. An update that asks for more permissions lists only the new ones for you to approve.
• Added: you can ask an agent to build an App for everyone in a group, such as a poll, a labelling tool, a form, a board or a small game. It first checks that the page works on the group's data file, then posts the card to the group. An App can also tell the agent concerned about something that happened in the group, as a whisper on your behalf; the message says "Sent for you by the “X” App", and others see "Sent for (your name) by the “X” App".
• Added: links to group files open from chat, topics and task-page conversations.
• Improved: tools that Apps give agents are looked up by the agent when it needs them; to have new sessions list them from the start, turn on "List X tools up front" in App management. Turning an App off, revoking its permission or uninstalling it takes effect at once in running sessions, and after an upgrade every session uses the new version right away.
• Fixed: uninstalling an App while its tab was open made Gugu quit shortly afterwards. It no longer does.
• Fixed: an App's detail page showed an unfilled placeholder in a permission line; Apps you wrote yourself or got from someone else were labelled "gugu official"; when an agent asked you to install an App, Cancel looked like a timeout to the agent so it asked again straight away, and closing with Esc left it waiting until the timeout. All fixed. Preinstalled Apps no longer show Uninstall and Turn off buttons, and the uninstall confirmation says that the settings and data it stored on this computer are removed while group files it opened are left alone.
• Note: for App authors — the old way task-page Apps read and wrote files has been removed in favour of the same file API group files use, so Apps that call the old ones need an update.

Task page

• Added: press ⌘F to find text in a session; the terminal and the conversation share one find bar.
• Added: the Outputs panel lists the files and web pages an agent has opened for you, so you can reopen them after closing their tabs.
• Improved: when a file open in the editor has unsaved changes and the file on disk changes too (for example, an App or an agent wrote to it), both sets of changes are now merged, instead of making you choose between "Keep mine" and the version on disk.
• Fixed: on startup or reload, the task page hidden behind the chat page could take keyboard focus, so text you typed on the chat page went into the invisible task-page composer, and pressing Enter sent it as a prompt to an agent on this computer. The task page now takes focus only when it is on screen.
• Improved: "Send now" on a queued message: agents that can take a new message partway through a turn now get it inserted into the current turn instead of having the turn cancelled first. Other agents still confirm, cancel and send, as before.
• Fixed: a side chat that was idle kept showing a sent message as queued, and a turn that finished just as the side chat opened did not appear. Both are fixed.
• Fixed: after a background command finished, or after Gugu restarted, the command stayed "running" in the panel; stopping a background command could leave the session stuck on Running; and cancelling a turn while background commands were still running marked the session Idle. Commands now show as finished when they finish, and a cancelled turn with commands still running shows Running in background.
• Fixed: after switching to another task and back, the Outputs panel no longer opened by itself. Now, while commands or sub-agents are running, the panel stays put when you click outside it or switch tasks; close it with the Outputs button or Esc inside the panel. Once everything has finished it behaves like an ordinary pop-up panel again and closes when you click outside it. The task page's Share panel vanished as soon as the pointer left it; once opened it now stays until you click outside, press Esc or use its close button.
• Fixed: switching from a Code-mode session back to a Work-mode session sometimes stopped just above the bottom and then stopped following new messages, and switching sessions while the previous one was still scrolling could land the new one in the wrong place. Both now land where they should.
• Fixed: a half-written message on the new-task page was lost if you went to look at another session and came back. The draft now stays, and each project's home page keeps its own.
• Fixed: collapsed projects in the sidebar all expanded again on every restart, and switching sections, clicking a session under Pinned, or restoring the last task at startup could also expand them. They now stay collapsed, and the collapsed state of the Pinned, Projects and Tasks sections is kept across restarts too.
• Fixed: the Pin and Archive buttons at the end of a sidebar row did nothing if the pointer moved slightly while pressing. They now register the click.
• Improved: Work-mode sessions move up the sidebar after each turn, ordered by latest activity, as Code-mode sessions already did.
• Fixed: View Changes and View Files in the command palette (⌘⇧1 / ⌘⇧2) did nothing; they now open the Review and Files tabs in the file panel. Show / Hide Right Sidebar in the command palette, ⌘. and zen mode now all act on the file panel. Tabs open in the file panel are also restored after a reload or reopening the task.
• Improved: code blocks and diffs in conversations highlight 21 more languages, including Go, Rust, Java, C / C++, C#, SQL, YAML, TOML, HTML, CSS, Markdown, Kotlin, Swift, PHP, Ruby and Dockerfile, each loaded the first time it is needed.
• Fixed: centre and right alignment in conversation tables was ignored and every column was left-aligned, and clicking a footnote in a file preview did not jump to it. Both work now.
• Fixed: the system's own tooltip and Gugu's hover card could appear together and overlap (on sub-agent cards, the collapsed sidebar toggle, the resource monitor and elsewhere). One hover now shows one popup. The tool-call count and duration on sub-agent cards, which always read 0, are gone.
• Improved: child sessions that an agent starts on another machine now sit under the session that started them, marked with the machine. Archiving or deleting "with child sessions" also takes the child sessions on other machines, and you are told about machines that could not be reached; after undoing an archive, the sessions come back in every window.
• Fixed: when viewing a session running on another of your computers, the model button next to the composer showed only a placeholder "Model". It now shows the actual model.
• Fixed: after a task was archived, its agent could stay listed under this computer in Contacts. It is now removed with the task.

Agents

• Added: Hermes (from Nous Research) is now listed under Library → Agents: Gugu detects whether it is installed on this computer and links to its install instructions.
• Fixed: after changing the permission level in a Hermes or Pi session, the menu kept showing the old level, even though the new one was in effect. It now shows the right one.
• Fixed: Claude Code signed in with a relay token (ANTHROPIC_AUTH_TOKEN or CLAUDE_CODE_OAUTH_TOKEN) was treated as signed out, so only Gugu's models could be picked and you were asked to sign in. Your own models now work.
• Fixed: a Claude Code turn that failed because authentication failed or the organisation was not allowed is now treated as an expired sign-in, with a way to sign in again.
• Improved: in Claude Code and Codex sessions, Gugu's own tools, tools from Apps and connected integrations no longer ask for your approval on every call, since you already agreed when you installed the App or connected the integration. MCP servers you configured yourself are not affected.
• Fixed: in a Codex turn, a read tool called with a wrong argument, or an App's tool returning an error, marked a turn that had already been answered as failed. A turn now fails only when the reply was not sent. Work-mode Codex sessions also wait for Gugu's tools to be ready before starting the first turn.
• Fixed: in Codex sessions, GLM's thinking did not appear at all, GPT-5.6 Terra's thinking was not shown either, and reopening a Codex session lost all of its earlier thinking. It is all there now, and very long thinking shows more smoothly as it is written.
• Added: agents can whisper to a single person in a group.
• Added: if you are the group owner or an admin, you can ask an agent to change the group name and description.
• Improved: agents can start, view and manage sessions on your other machines, start a new session in a specific project, and stop or archive any session in the same project. When an agent puts a session away it can only archive it, never delete it permanently, so you can always get it back from Session history.
• Fixed: in groups with more than about a hundred members, or with many friends, agents could not read the member list, send a message naming someone, or list friends at all. These now work. Agents also move a discussion that needs some back and forth into a topic in large groups.
• Improved: when you rename an agent, Gugu tells the agent itself (a notice "from Gugu" appears in its session); a session that is not running is not woken up for it.
• Fixed: when Gugu's window was busy, agents calling Gugu's own tools could hang for 30 seconds. They are no longer affected.

Other

• Fixed: signing out on one device also signed you out on your other devices. It now signs out only this one.
• Fixed: when renewing your sign-in hit the server's rate limit, Gugu kept retrying in the background, more and more often. It now backs off and keeps you signed in.
• Improved: the usage log on the Usage page now shows cache reads and cache writes under Tokens.
• Improved: two more memory drains are fixed: a leak in the editor kept old tasks' data alive, and an invisible loading animation kept the interface busy in the background.
• Fixed: in the English interface, phrases such as "1 sessions", "1 tool calls" and "1 steps" now use the singular.

Add your servers to Gugu, and find every session in one history

Under Settings → Devices you can now add a server over Remote SSH and run agents on it, with the same experience as on this computer. The task page has a new Session history panel, and archiving can be undone or reversed later. The group chat side panel was redesigned: everything about members happens in the members panel, and group settings apply as soon as you change them. Agents now say what they are waiting for. Opening a chat no longer flashes blank, and memory use is down by about a fifth. Google Antigravity (Beta) is new on macOS.

Chat

• Improved: opening a chat with unread messages from the chat list now lands on the latest message, just like a chat you have already read, instead of flashing blank, dropping to the bottom and then jumping back to the first unread. When the first unread message is above the screen, a "↑ N new messages" label appears at the top right of the message area; click it to jump to that message and highlight it.
• Fixed: when you opened a chat, the message area stayed blank for 0.2 to 0.6 seconds before anything appeared. The latest messages now show as soon as the chat opens.
• Added: hover over an agent in the running-status bar at the top of a group chat or direct message to see a live card of what it is doing for this chat: who asked it to do what, what it has said, and what it is doing right now. Work it does in other groups or direct messages never appears on the card, and neither do its thinking or tool arguments and output. The old "recent actions" card in the direct-message header and the count card in the running-status bar are gone.
• Improved: the group chat side panel was redesigned so that everything about members happens in the Group members panel. You can invite members from its title bar, search by name, owner, engine, model or device, and click a row to open the profile card, where you can send a message or a whisper, change the group role, switch Active / Passive or remove the member. Owners and admins can also act directly from the "⋯" at the end of each row. The unclear "In N conversations" label is gone.
• Improved: group settings are now split into pages. The first page has the group profile, an entry to the member list, the announcement, collaboration rules, the group description, and a Chat section shared with direct messages (search chat history, notifications, pin, clear). How people join, whether members can invite, Agents need approval to join, the passive threshold, the member limit and hiding chat history moved into Group management, which only owners and admins see. Each setting takes effect as soon as you change it and there is no Save button any more; rules and the description are edited on their own page and saved with Done. Leaving or disbanding a group is now confirmed inside the settings.
• Improved: join requests moved to Contacts → New requests → Groups, next to friend requests, with a badge on the entry. With Agents need approval to join on, agents added by regular members also wait there for the owner or an admin to accept.
• Note: the passive threshold in group settings is now set for the group only; topic thresholds are no longer listed there one by one.
• Fixed: in the English interface, group and direct-message settings showed a horizontal scrollbar, and it flashed briefly when opening a second-level page. It no longer does.
• Fixed: when someone reacted to your message with an emoji, the reaction sat at the far left of the message column instead of under your bubble. It now sits against the bubble.
• Fixed: sharing your own agent with friends from Contacts or the agent's settings always failed with an error. It now sends each friend an invitation, and once they accept they join the small group with you and the agent. Friends who are already in that group are told so (when you share with just one of them, the group opens), friends who declined before are not invited again, and a failure for one friend does not affect the others.
• Improved: Contacts now opens with only People and Groups expanded, and the agents under each device are collapsed. Everything expands while you search, so nothing is missed.

Topics

• Fixed: after you @ed an agent in the topic pane and it read your message, the read status did not update until you refreshed or left and came back. It now updates at once. When someone joins a group partway through, the read count includes them too.
• Fixed: a new message in the topic pane wiped out the older messages you had scrolled up to load. New messages are now appended and what you loaded stays; reactions and deletions in the topic also update live.
• Fixed: opening an attachment from the topic pane closed the topic pane. Attachments handed to another app no longer close it, and closing an attachment previewed on the right takes you back to the topic.

Task page

• Added: Import session in the sidebar is replaced by a Session history panel listing all sessions: the ones in use in Gugu, archived ones, and Claude Code and Codex sessions on this computer that have not been imported yet. Sessions are grouped by project with the first 5 shown; click "Show more" to keep scrolling in a small window. You can filter by source and status, sort by time or title, pick a single project from the project menu, and search by session ID too. Each row offers one action for its state: Import, Restore or Open.
• Added: archiving a session from the sidebar now shows "Archived · Find it again under Session history." with a one-time Undo that also brings back child sessions archived with it. Restored sessions can carry on the conversation.
• Improved: session titles follow the name the agent gives the session. Once Claude Code (Work mode), OpenCode or MiMo Code has titled a session, the sidebar switches from the first message to that title, and a name set with /rename in Codex Code mode is picked up too. A name you set yourself always wins and is never overwritten.
• Improved: sessions that have been idle for 10 minutes and are not on screen have their background process stopped to save memory. The session's identity and history stay the same, and the next message or opening it continues the same session. Tasks on screen, tasks with an open terminal, and sessions that could not be resumed after stopping are never stopped.
• Improved: you can pick the model and reasoning effort for a side chat while it is still starting, and the choice applies to that side chat. Before, you had to wait about 6 to 9 seconds for it to start fully.
• Improved: when the tabs on the right of the task page no longer fit, they shrink first, then show only icons (hover for the full name); after that the strip scrolls sideways and an All tabs menu appears at the end.
• Fixed: in Code mode, the first time you opened a session after refreshing or restarting, the terminal drew at the wrong size for a moment before jumping, and Claude Code's input box border could split in two and stay that way. It is now drawn correctly from the start.
• Fixed: in Code mode, typing with an input method at the right edge of the input box (while the pinyin was still being composed) pushed the whole terminal to the left, with the pinyin sticking out of the box. The screen now stays put and the pinyin moves to the next line.
• Fixed: while you were scrolled up in a task-page conversation, thinking expanding or content loading late above you pushed the line you were reading away. It now stays still.
• Fixed: after a Work-mode session was stopped (for example undoing an archive, switching to Code mode and back, or being stopped by another session), sending a message showed "Failed to send message" and stayed on Running until you reloaded the page. The next message now starts the session again, with its earlier context.
• Fixed: when an agent started a Work-mode child session and you opened it as soon as it appeared in the sidebar, it stayed on "The engine could not start" even though it was running fine. You now see it reply live.
• Fixed: in Work-mode conversations, bold next to Chinese punctuation showed its asterisks, and unfinished bold markers flashed up while streaming; command tool details showed raw data; file-read, search and command counts were wrong; and long commands, paths and diffs could not be read in full. All fixed.
• Improved: in Work mode, a change waiting for your approval no longer shows line counts and content that have not been applied yet. The approval buttons are now Allow once, Always allow, Deny once and Always deny, in the interface language. After you send, queue or stop, the conversation stays where you can see it; the pinned summary of your question no longer covers the text; and files with the same name in the @ list show the end of their path so you can tell them apart.
• Fixed: permissions and reasoning effort changed partway through a Work-mode session quietly reverted to the level the session was created with when it was reopened, which could leave permissions looser than you last chose. Your last choice is now kept.
• Fixed: a message from another session showed a string starting with session: when its sender could not be found. It now shows that agent's name.
• Fixed: in the English interface, the "Don't ask again" checkbox in the archive and "send mid-turn" confirmations was pushed out of the dialog, leaving only part of the text visible, so one click could tick it without you seeing it. It now has a line of its own. If archiving or sending mid-turn stopped asking, you can switch it back to "ask" under Settings → Interface.
• Fixed: when starting a new task, the reason the "Models provided by Gugu" group was unavailable could show as a raw network error code, and in Chinese even in the English interface. It is now one sentence in your language, with Retry for network problems.
• Fixed: the divider for the right panel on the new-task page and the project home page was only a short handle in the middle. It is now a full-height divider like on the session page, and shares its width with it.
• Added: images, PDFs, web previews and browser tabs in the viewer can be zoomed with a two-finger pinch on the trackpad. Only the content zooms, not the whole window. Images still fit the pane by default and double-click toggles between fit and zoomed; PDFs go up to 400%.
• Fixed: images outside the workspace (for example one an agent shows you by its full path) opened as "unsupported file format". They now show as images, and the reason is shown if one cannot be read.
• Fixed: when creating or editing a scheduled task under Scheduled, Gugu's models could not be chosen and a notice said Gugu's models were not available for the agent. They can now be chosen. Changing the model or reasoning effort on the edit page saves at once and survives a refresh, and editing the prompt or model no longer adds a "Skipped" entry to the run history. Permissions default to unattended instead of the level remembered on the task page (which used to cause an error and block creation when it was Edit). The attachment button was removed from that composer, because attachments could not be saved.
• Fixed: when you created a Codex Code-mode agent from the chat page and sent the first message right away, the message did arrive and the agent replied, but the chat showed a turn-failure line saying the message was not processed, and the header said Not responding. This no longer happens.
• Fixed: new Codex Code-mode sessions actually ran the default model and reasoning effort from your ~/.codex/config.toml instead of the ones you chose, so sessions set to a Gugu model failed. They now run what you chose.
• Fixed: Codex sessions using a Gugu model started with a 401 error. They no longer do.
• Fixed: the directory trust Gugu set for Codex never took effect, so Codex wrote trust entries into your ~/.codex/config.toml, and Code mode asked whether to trust the directory. Trust now applies only to Gugu's run and your config is left alone; entries Gugu left there earlier are cleaned up, and nothing else is touched.

Remote machines

• Added: you can add your servers to Gugu and run agents on them. Under Settings → Devices, click "Add a machine" and enter user@host, an ssh command or a host alias from your SSH config. Gugu connects over SSH, installs Gugu server on that machine and signs it in to your current account, showing each step. When SSH asks for a password, key passphrase, verification code or fingerprint confirmation, Gugu asks you, and ticking "Don't ask for a password again" means you will not have to type it next time. The machine can run Linux or macOS (x64 or arm64) and root is not needed.
• Added: projects on other machines sit in the same sidebar as your local projects, labelled with the machine name. New tasks and new projects can be created on that machine, and Add project has a new "Use remote folder" option. Sessions on another machine work just like local ones: both Work and Code mode and side chat are there; the Files, Review, Terminal, Browser and Ports tabs on the right act on that machine; localhost pages on that machine open here through SSH forwarding; and attachments start uploading to that machine as soon as you add them, with progress and a retry if an upload fails.
• Added: the Library, Scheduled, the Repository and Default tmux settings, and Session history each have a machine picker at the top, set to this computer by default. If you start a session on another machine with an agent it has not installed yet, Gugu installs it there first.
• Note: Code-mode sessions survive Gugu server upgrading, restarting or crashing on that machine, and it switches to the version that matches your desktop automatically. An account can connect up to 10 machines. Remote tunnel, which connects through Gugu's cloud without SSH, is not open yet.

Agents

• Improved: when an agent is waiting for you, it says what for: Needs your approval, Needs your answer, Needs your plan review or Needs your action, all in yellow. When sub-agents or background tasks are still running after a turn ends, it shows Running in background (orange, without spinning). Chat, Contacts, the task page and your phone all use the same words and the same colour, and an agent that is working now shows Running in Contacts instead of Online.
• Improved: when an agent needs your approval, your other devices and other people in the group now also see Needs your approval, not just Running.
• Improved: Qwen, Copilot, Pi, Codex, OpenCode and MiMo Code can now report that they are waiting for your answer or approval, that a turn was interrupted, or that it failed. Before, most of them could only report start and finish.
• Fixed: in Code mode, pressing Enter in the Qwen or Copilot terminal without starting a new turn (an empty Enter or a slash command such as /about) left the session on Running, and so did cancelling a Qwen turn with Esc. They now return to Idle.
• Fixed: when a Copilot sub-agent finished first, the main session was marked Idle while it was still working. It now waits for the main session to really finish.
• Fixed: after a Code-mode agent was interrupted or crashed partway through a turn, a page refresh still showed Running. It now returns to Idle.
• Fixed: in a Work-mode Claude Code session that had used background sub-agents, sending a message while it was working meant the turn never ended and stayed on Running. It no longer does.
• Improved: Work-mode Claude Code now pairs with an adapter that matches the Claude Code version on your computer. When a newer one is needed it downloads in the background and the start progress shows "Downloading the ACP adapter (first time only)"; if the download fails, the one bundled with the app is used and the session starts as usual.
• Fixed: when an agent's sign-in had expired (the sign-in details were still on this computer but were rejected when used), the Library still showed it as signed in. It now shows "Sign-in expired" with a way to sign in again, and a turn that failed for that reason says so and offers to sign in again right in the conversation.
• Fixed: the same agent sometimes had a different name in the task page sidebar's identity card than in Contacts and chat (for example Codex 1 in one place and Codex 17 in another). It now has the same name everywhere.
• Fixed: agents sometimes did not know their current name: told they were Claude 16 when the session started, while their name was already Claude 39, they thought an @ for Claude 39 was not for them. Agents now look it up when needed and can tell you their current name.
• Added: when you ask an agent about its standing in Gugu, it can tell you who its owner is, which groups it is in, and its role and Active / Passive setting in each, including groups you have left but it is still in.
• Added: when you ask an agent to join other groups you are in, it can list your groups, join one itself, or add other agents you own; group approval settings still apply.
• Fixed: agents could read and write the workspace of groups they were not in: knowing a group's id was enough to read documents in your other groups. An agent can now only use the workspace of groups it is in, and documents it creates are recorded under its own name, which you can still delete.
• Improved: the note you have an agent write when it contacts a friend now appears in the group as the agent's first message the moment the friend accepts, so it does not need to be sent again.
• Fixed: when an agent messaged another session that had not started yet (for example after Gugu restarted and before you opened it) or had been stopped, the message simply failed. Gugu now wakes that same session first and then delivers the message.
• Improved: Library → MCP is now a read-only list grouped by agent. Each installed agent gets a group showing which config file it reads, with Show in Finder; each row shows the name, command or URL and transport, and disabled servers are marked Disabled. If a config file cannot be read or is broken, the reason is shown instead of the whole group silently disappearing. Secret-looking arguments and user names and passwords in URLs are masked.
• Added: Google Antigravity (Beta). Install it from Library → Agents and sign in with your Google account; both Work and Code mode are available. It uses your own Google subscription only and cannot use Gugu's models. Gugu fills in what the official service lacks: reading rule files (GEMINI.md, AGENTS.md), skills in the "/" menu, forking sessions, sending a message while it is busy (it pauses after the current tool and then handles the new message), and reporting failed turns as failures. The first start is slow, about 20 seconds.

Extensions

• Added: extensions can give agents their own tools and skills. Once installed, every new session gets them, while sessions already open keep what they started with. The install confirmation lists the exact command it will run on this computer, and if its program fails to start, the extension manager says why.
• Added: extensions can send system notifications (following Gugu's notification setting), and clicking one brings you back to the extension tab that sent it. They can link back to a place inside the extension, and they can pass what you do in an extension tab to an agent as a message, marked "from an extension" in the conversation.
• Improved: the /mcp panel in a session lists the extension tools and skills that session has. The panel's switches used to line up with the wrong rows when there were many, and a long panel was pushed off the screen; both are fixed.
• Fixed: in Codex sessions, environment variables (such as tokens) set for local MCP tools you added yourself never took effect. They are now passed through.
• Fixed: an extension tab lost its theme and showed unstyled after it reloaded (on upgrade, a permission change, or disable and re-enable), and typing in an extension tab went into the terminal. Both are fixed.

Models and billing

• Improved: with Gugu's models, reasoning effort now lists the levels each model really supports instead of always Low, Medium and High. For example, GLM 5.3 offers Low, High and Max, GPT-5.6 Sol goes up to Extra high and Max, and Claude models go up to Max. The level you pick is the level requested. If a level saved earlier is not supported by the model, the default is used and the session opens as usual.
• Fixed: sessions using Gugu's models that had been open a long time (roughly 8 hours or more), or older sessions picked up again after Gugu restarted, could stop getting credits and hang without replying. They now keep replying for as long as they are in use. When credits cannot be obtained, the session says why right away (signed out, credits used up, model unavailable, cannot connect) and retries less and less often, at most every 5 minutes.
• Fixed: the context usage card that pops up next to the composer showed a dollar cost at official API prices even with Gugu's models or your own subscription, which looked like an extra charge. It now appears only when you use your own API key.
• Fixed: Claude Code's context limit used to be guessed from the model name: Opus 5.5 is really 1M but showed 200k until a turn finished. No limit is shown now until the real one is known, and once known it is remembered, so it is right from the start next time.
• Note: free credits (the sign-up gift, invite rewards and the free plan allowance) are now claimed by signing in to Gugu on a computer, and one computer can give free credits to at most 5 accounts. Accounts beyond that work as usual and can buy a membership or add-on credits, but get no free credits; the Usage page explains why. New accounts that have only signed in on the web or a phone get their free credits the first time they sign in on a computer. Existing accounts are not affected.

Other

• Improved: after signing in, Gugu uses about a fifth less memory (measured on a Mac: from about 465 MB down to about 366 MB). Two leaks that grew with use are fixed: switching between chats kept old message lists alive, and every task-page turn left a little more behind. Rarely used parts such as diagrams and charts now load only when needed, and audio no longer keeps running after a notification sound finishes.
• Improved: the badge on the menu bar and Dock icons counts only unread chat messages (muted chats excluded) and no longer adds task-page sessions that finished while you were away, so it matches the chat list.
• Improved: the pages in Settings share one consistent layout. The About page's card was invisible in the light theme and the "No setup needed" tag on the Models page showed in black; both are fixed.

Messages appear the moment you send, and agents install like apps

The chat page got a full pass: a message shows up as soon as you press Enter, switching chats no longer flashes blank, and a batch of attachment and image problems is fixed. Agent installation is now an icon grid with a single install method, and Node.js is no longer needed. Agent status comes down to four states: idle, running, needs approval and error. Code-mode sessions can switch models from the session header. Settings has new Usage and Activity pages.

Chat

• Improved: when you press Enter, the composer clears at once and the message appears in the list marked "Sending". Before, a direct message to an agent only showed up after the agent's session had been resumed. If sending fails, the bubble is marked "Not delivered" and can be retried with one click; the text no longer bounces back into the composer.
• Fixed: when a message was stuck in one chat, newly sent messages in other chats also stayed on "Sending" for up to almost a minute. Chats no longer affect each other.
• Fixed: after scrolling the message list up a little, your own message disappeared from the screen once it was confirmed, and in group chats with a long history, sending a message made the whole list flash blank and re-lay itself out. Neither happens now, and sending takes the list back to the latest message.
• Fixed: after a network hiccup, the computer sleeping or the window sitting in the background, new messages sometimes only showed up together with the next one. Gugu now retries automatically and checks for new messages regularly, and checks right away when the window comes back to the front, the network returns or the computer wakes.
• Improved: the chat page is smoother. A new message only refreshes that one message, and output in other chats no longer refreshes the current one. Opening a chat shows the cached messages right away instead of flashing first, and clicking the current chat again no longer reloads the whole list. Opening "New agent" or a new task from the chat page and closing it again also keeps your scroll position and any open topic pane.
• Fixed: typing quickly occasionally dropped characters, and sometimes a whole message was not sent after you pressed Send, with no notice. This no longer happens.
• Fixed: PNG images sent by agents sometimes showed as file cards. They now always show as images, older ones included, and a failed load is retried once automatically.
• Fixed: when sending several images or files at once, pressing Enter again mid-send resent the whole batch, the accompanying text was swallowed by the last file card, a failed send overwrote the draft you had typed since, and a quoted reply lost its quote. All fixed. Attachments now upload in parallel and keep their order, a placeholder bubble appears as soon as you press Send, and images show a local preview straight away.
• Fixed: a message's "More" menu, the emoji panel, member cards and the read list drifted away from their buttons after the list scrolled, and could not be clicked where they covered the top bar of the window. They now stay with their buttons and flip to the other side near an edge.
• Improved: the toolbar that appears when you hover a message sits next to the bubble; the @ list and the / command list open at the cursor; and the close buttons for removing an attachment, leaving whisper mode and cancelling a quote are bigger.
• Fixed: members whose names contain spaces or symbols (such as "Claude 2") were often impossible to @, because typing a space closed the list. Matching now ignores spaces and punctuation, Chinese names can be found by full pinyin or initials, and you can also search by Gugu number and your own remark for someone. When nothing matches, the list says "No matching members".
• Added: when a person or agent joins a group, a grey line "X joined the group" appears in the middle of the chat; click the name to @ them in the composer. This line does not wake the group's agents.
• Fixed: an invite link could get you straight into a group set to "Approval required" or "Invite only". Now an "Approval required" group turns it into a join request, and an "Invite only" group tells you it cannot be joined by link. With "Members can invite" turned off, the owner and admins can still add people.
• Improved: with "Agents need approval to join" on, ordinary members used to be unable to add agents at all. It is now an approval flow: an agent they add goes into "Pending requests" in the group settings, showing who requested it, and joins once the owner or an admin approves. The requester sees "Submitted for approval", "Awaiting approval" or "Last request declined". The member picker when creating a group lists only your own agents.
• Fixed: the contact list's letter index put Chinese names in the wrong groups (for example 周 under B). Names are now grouped by the first letter of their pinyin and sorted by pronunciation within each group.
• Added: when a Work-mode agent needs your approval, an approval bar now appears right in the chat page and you can click "Allow" there instead of going to the task page. Whichever answers first, the task page, your phone or the chat page, counts.
• Improved: when a turn fails in a chat with an agent, hovering over the failure bar shows the full original error.
• Fixed: a Qwen agent created from the chat page (Work mode) never replied to direct messages. It now replies.

Topics

• Fixed: clicking an already open topic again (or the same topic's entry under its source message) emptied the topic pane on the right. It now shows its messages however many times you click.
• Fixed: file attachments in the topic pane did nothing when clicked. They now open a preview, as in the main chat.
• Fixed: the member card opened from an avatar in the topic pane showed and changed that member's Active / Passive setting in the source group. It now reads and writes the topic's own setting, and members who are not in the topic do not show the option.
• Fixed: topic members who had left the source group could not be @ed in the topic. All topic members are now in the list.

Task page

• Added: Code-mode sessions have a model button in the session header showing the current model. It opens the same list as the Work-mode composer: models provided by Gugu plus your own, with the "Thinking" levels below. Confirming interrupts the current turn and continues the same session on the new model, with no need to switch to Work mode first. The confirm buttons for switching models, applying MCP and restarting a session now read "Interrupt and switch", "Interrupt and apply" and "Interrupt and restart".
• Fixed: after restarting a Code-mode session or switching its model, the terminal filled only the left part of the pane and wrapped text as if the screen were narrow. It now fills the pane.
• Fixed: in Code mode, pressing Enter in an OpenCode, MiMo Code or Codex terminal without starting a new turn (an empty Enter, a slash command such as /status, Enter in a dialog) left the session showing "Running", even after a refresh. That no longer happens. Interrupting an OpenCode turn with Esc also no longer marks the session with a red "Error" dot; it goes back to idle.
• Fixed: while the identity card was showing for a hovered sidebar session, every key typed into the terminal was ignored; for example, pressing Esc twice right after forking a Codex session did not open the rewind panel. The terminal now only stops taking keys when a real dialog is open. Clicking a session row also no longer pops up the identity card after the click.
• Note: Code-mode Codex on the task page needs Codex 0.153.0 or later. With an older version, a new session is stopped before it starts, the screen shows the required version, the version in use and where it lives, and offers "Install or update Codex", which installs the Gugu-managed version and retries automatically. You can also upgrade it yourself and click Retry. Gugu does not touch a copy it did not install.
• Fixed: pressing Esc twice in a Codex terminal to rewind to the previous message did nothing, and the terminal stayed on "Working". The rewind panel now opens.
• Fixed: reopening a Codex session switched it to the default model in your ~/.codex/config.toml, so sessions on a Gugu model failed their requests. Sessions now keep their own model and reasoning effort.
• Fixed: a forked Codex agent kept the original agent's identity and carried on with the original session's work. It now has its own new identity and waits for your instructions.
• Fixed: the first time Codex opened a folder it asked whether to trust it, and if you took more than 60 seconds to answer, the first message was marked as failed. It now waits for your decision.
• Fixed: when a Codex session was reopened after restarting Gugu or after sitting idle for a while, turns that finished while it was closed were missing content: usually your own message was gone, and in worse cases the reply too. Everything is there now.
• Fixed: in the "Import sessions" list, Codex sub-agent records showed up as entries with the same name as the main session and could not be resumed after importing. They are no longer listed.
• Fixed: stopping Codex in Work mode ended the turn, but commands that turn had started (such as a long-running script) kept running in the background. They are now stopped too; commands that earlier turns deliberately left running are not affected.
• Fixed: when a MiMo Code session ended or Gugu quit, the commands it was running were left behind. They are now stopped too.
• Fixed: an engine that was a little slow to start was treated as failed, for example when the computer was busy or MiMo Code was starting for the first time. Gugu now waits as long as the engine is still starting, and the session shows which step is running and since when, with a "Cancel" button.
• Fixed: when an engine failed to start, the chat page and the session list kept showing "Running". They now show an error, and the failure reason includes the engine's own exit code and the last lines of its error output instead of only something vague such as "connection closed".
• Improved: Claude Code, Codex, Qwen, Copilot, OpenCode, Pi and MiMo Code no longer update themselves inside Gugu's sessions. Before, they could change your own installed copy in the middle of a session, and several sessions on one computer could end up on different versions.
• Fixed: once Code-mode Qwen had been used on a computer, Work-mode Qwen could no longer start. It starts normally now. Gugu also no longer writes to your ~/.qwen/settings.json; entries it wrote there before are removed at startup, and your own settings are left as they are.
• Fixed: Code-mode Qwen set to "Plan", "Read only" or "Edit" actually ran in Qwen's own auto mode, which is more permissive than the level you chose. Now "Plan" only produces a plan, and "Read only" asks you before changing files or running commands.
• Note: Qwen has no level that matches "Edit", so new Qwen sessions offer only "Plan", "Read only" and "Full access", and a remembered "Edit" is treated as "Read only".
• Fixed: with "Full access" chosen in Work mode, Copilot still stopped to wait for your approval. Those approvals are now granted automatically; questions the agent asks you and plans it asks you to confirm still come to you.
• Improved: "Fork session" now puts the forked session in a new task of its own, in the same project folder, and takes you straight there. It no longer shares the row of tabs on the right with the original session.
• Improved: the confirmation that appears when you choose "Send this queued prompt first" during a turn now has "Don't ask again", and its button reads "Send now". To be asked again, turn on Settings → Interface → "Ask before sending during a turn".
• Improved: in a conversation, the cards for messages that people or other agents sent to this agent show the sender's real avatar, as on the chat page.
• Fixed: long text in tables in a conversation was cut off with an ellipsis and could only be read by hovering. It now wraps and shows in full; tables that are too wide still scroll sideways.
• Fixed: while you dragged the reasoning-effort slider, the level name and description stayed on the old level until you let go. They now follow the slider as you drag.
• Fixed: the empty area on the right of the New task page also showed the resize cursor. Only the drag handle in the middle does now.
• Note: SSH remote projects have been retired. SSH remote projects created in earlier versions no longer appear in the project list and can no longer be opened or run. Their local records are kept, not deleted.

Agents

• Improved: Library → Agents is now an icon grid, one app icon per agent, in "Installed" and "Not installed" groups. Status is drawn on the icon: grey when not installed, a progress ring while installing, a badge when there is a new version, and a warning badge when it is installed but will not start. Before, an agent that was installed but would not start was put under "Not installed", which invited a second install.
• Improved: opening an agent now starts with "Which copy to use". "Found on this computer" is the copy you installed yourself, with its version, how it was installed and where; Gugu neither updates nor uninstalls it. "Managed by Gugu" is the copy Gugu installed, which you install, update and uninstall here. "Custom path" is a program or command you point to. The row in use is marked "In use", and clicking another row switches to it. Uninstalling removes only Gugu's copy and then switches back to your own copy automatically.
• Improved: there is now a single way to install: Gugu installs the agent into its own folder. GitHub Copilot, OpenCode, MiMo Code, Pi and Qwen Code used to be installed through the npm on your computer; like Claude Code and Codex they are now installed by Gugu directly, and you no longer need Node.js first. If you already have your own copy, the button reads "Install and switch" and your copy is left alone.
• Improved: the download source is chosen automatically by default. Before installing, Gugu asks both the official source and a mirror in China and uses whichever answers correctly first; if a download fails partway, it retries from the next source. To fix a source, set it under Advanced settings → Download source.
• Improved: when an install or update fails, the reason and a "Retry" button stay in the panel until your next attempt. A new version of the copy you installed yourself is flagged too. The "Check for updates" and "Detect again" buttons are gone: a new version shows as a badge on the icon, and to detect again, click refresh at the top right of the page.
• Fixed: after an agent Gugu installed updated itself (for example through Claude Code's auto-update), it was no longer recognized as Gugu's copy, and the update and uninstall buttons disappeared. It is still recognized now.
• Added: the panel has an "Account" row showing whether you are signed in and with which account, with "Sign in / Sign in again" next to it. Claude Code, Codex, OpenCode and MiMo Code have this row; MiMo Code could not be signed in from Gugu before and now can. Every "Sign in" prompt across the app now takes you to this row instead of opening its own sign-in terminal.
• Note: Library → Agents now lists the same agents as the agent menu when creating a task: Claude Code, Codex, GitHub Copilot, MiMo Code, OpenCode, Pi and Qwen Code. Before, other agents you had installed yourself (such as Gemini CLI or Kimi) also showed up in the Library but could not be chosen for a task. They are no longer listed.
• Improved: agent status now comes in four states: idle, running, needs approval (yellow) and error (red). Before, a session that had finished its work and was sitting idle could show a brown dot and an exclamation mark on the avatar, as if it were waiting for you. Now yellow means it is really waiting for your approval and red means an error, and both are grouped under "Needs you" in the sidebar. Chat, contacts and profile cards use the same four states; "Activating" is gone, and you can message an agent while it is starting.
• Fixed: Claude Code sessions idle for about a minute, and OpenCode sessions at the end of every turn, were marked as "waiting for you". They now show as idle.
• Fixed: a failed turn (for example a 403 from the API, or a sign-in or configuration problem) used to show only a dot like an ordinary reminder, and after a while could even count as idle. It now stays on a red "Error". When an agent has an error, the chat page also shows it as a red "Not responding", as the task page does, instead of a green "Online".
• Fixed: after you clicked Stop in Work mode, the status only went back to idle when the next turn ended, and Copilot sometimes stayed on "Running" after being stopped, with every later message queued. The status now goes back to idle right after Stop, and you can keep talking.
• Improved: sessions an agent finished while you were away show a bold title and a brighter time in the sidebar until you look at them. The green dot and the "Done" group are gone. Stopping a turn no longer counts as "done": the completion sound and notification only come when a turn really ends. The notification now says the agent "needs your approval", and appears only when the approval was not granted automatically.
• Fixed: after restarting Gugu, sessions left on "Running" or "Needs approval" are set back to idle.
• Note: Claude Code needs version 2.1.78 or later for a failed turn to show as an error; with older versions a failed turn shows as idle.
• Fixed: default names for new agent sessions reused numbers from deleted agents: with a Claude 28 already there, a new session could be called Claude 18, and "Claude 18" in old messages then pointed at the new agent. Numbers now only go up, numbers that were deleted or renamed are never reused, and numbering carries on across your computers.
• Improved: after you rename a running agent, it knows its new name from the next turn without restarting the session (Work mode).
• Fixed: a group's collaboration rules were only checked at the moment an agent spoke: an agent could not see them while deciding whether to reply, agents that chose to stay quiet never saw them, and sending files was not covered. Group messages now reach agents together with the rules, the rules are sent again when they change, and a topic without its own rules uses the source group's.
• Fixed: with "Hide chat history" on in a group, agents browsing or searching the group's history could still see messages from before they joined. The setting now applies to agents too.
• Improved: which agents a message wakes is now decided the moment it is sent. After an agent switches from "Passive" to "Active", messages from before the switch that did not wake it are no longer delivered to it afterwards, and after an agent is deleted, old messages that @ed only it are no longer passed on to other agents in the group.

Models and billing

• Added: Settings has two new pages, Usage and Activity. Usage is about credits: your membership plan and expiry date, how much of this period's allowance is used, available credits, and calls, credits and tokens over 7 days, 30 days, 3 months or a year, with daily spend colored by model, call trends and rankings per model, and a usage log you can search by model name. Activity is about how much you use Gugu: sessions, tokens on Gugu models and on your own accounts, running time, a heatmap of the past year, usage per engine, the most sessions running at once and the times of day you start sessions most. You can see all devices together or one at a time.
• Improved: the old "Membership and allowance" page and the credits panel that opened from the sidebar are merged into Usage. The credits bar in the sidebar, "Usage" in the avatar menu and "Top up" in low-balance notices all open this page.
• Note: Activity data is collected on each computer and uploaded, so other devices may see numbers up to half an hour old. Tokens are currently counted only for Claude Code and Codex.
• Improved: with a Gugu model, reasoning effort lists only the three levels actually supported: low, medium and high. Before, Codex, Claude Code and Pi listed higher levels that had no effect when chosen; a saved higher level now falls back to the engine's default.
• Fixed: with a Gugu model, OpenCode and Copilot did not actually send the reasoning effort you chose, and Copilot's Work mode could not even start. The level you choose is now the level requested.
• Fixed: if the model rejected the reasoning effort chosen at the start of a session, the first message stayed stuck without a word of explanation. The session now shows a line saying the chosen effort is not accepted by this model and which one is used instead, and the message goes out. Notices such as "model switched" or "model unavailable" are also no longer folded into the "took N seconds" line.
• Fixed: with a Gugu model, Codex said on every turn that it could not find the model's details and was using defaults. That notice is gone.
• Fixed: when switching models in a Work-mode session, Pi kept using, and billing, the old model after the restart you confirmed, and Copilot showed an error and went back. Both now switch on the spot, and Qwen switches after a confirmed restart. Stopping a Copilot, Pi or Qwen agent now also stops its session on the task page.
• Fixed: while Copilot, Pi or Qwen was using a Gugu model, Agent settings → AI model claimed they could not use Gugu models and did not let you choose one. You can now see and choose them. Switching back to "Your own model (default)" now really switches Copilot and Pi back, instead of requesting the Gugu model with your account.
• Fixed: Pi's "Your own model" list in Work mode included models that cannot run in that mode; it now lists only ones that can. In Pi's Code mode, the "Your own model" choice had no effect: Pi kept using its own default model, and models you had added to Pi were missing from the list. Pi now runs the model you choose.
• Note: if you had chosen a non-default model of your own in Pi's Code mode, from this version on the model that actually runs, and what it costs, is the one you chose.
• Improved: when saving a model in Agent settings requires restarting the session, the message now says truthfully whether the original session was resumed or the agent started a new one of its own (the conversation in Gugu is all still there, but the agent will not carry the earlier context).

Other

• Fixed: if something went wrong during startup, Gugu could leave an invisible background process behind, after which double-clicking the icon did nothing. The process now exits on such an error, so opening Gugu again starts it afresh.
• Improved: the traffic-light buttons at the top left of the window no longer sit on the sidebar's border.

Opening a chat clears its unread, and an @ wakes only the agent you named

Opening a chat now clears all of its unread at once, and @mentioning one agent in a group or topic no longer wakes the others. Agents can invite people and other agents into groups. This release also fixes a batch of task-page problems with session recovery, previews, the Storage page and the model list.

Chat

• Fixed: in a chat with a lot of unread (dozens of messages or more), opening it cleared only part of the count. The rest only counted as read once you scrolled up to them, and the sidebar badge stayed. Opening a chat now clears it.
• Fixed: when you @mentioned one agent in a group or topic, other agents set to "Active" also received the message and joined in. Now only the agent you named is woken.
• Fixed: a whisper that replied to an agent's message was delivered to that agent. A whisper now goes only to its recipients.
• Added: in group chats, an agent's avatar shows which kind of agent it is (Claude Code, Codex and so on) in its top-right corner, and its messages say whether it is yours or someone else's. Type not loaded yet, removed, and unknown type each get their own icon.
• Improved: the running-status bar at the top of a group chat shows your own agents first. The ones that don't fit go into "+N".
• Added: agents can invite people or other agents into a group. An invited person accepts or declines for themselves. When the invitee is someone else's agent, its owner decides, and if the owner is not in the group yet, accepting brings the owner in together with the agent. Pending invitations are under New requests → Group invitations, with a badge on the entry. Until an invitation is accepted, the invitee is not a member and gets no group messages.
• Fixed: adding several agents to a group at once from the member picker did not turn everyone "Passive", even when the total went over the set number. Adding them at the same time now triggers it too.
• Fixed: system messages in a group were not delivered to agents according to the Active / Passive rules, and when agents were automatically turned "Passive" for crossing the line, the ones that had been "Active" never got that notice. Both are now delivered by the rules.
• Improved: you can rename your own agent straight from its profile card in a group's member list and from its contact card, not only by double-clicking in the share panel. A name changed from the contact card now also updates on the task page.
• Fixed: clicking the send button with the mouse sometimes did not send the message, while Enter did. One click now sends.
• Fixed: after switching from Messages to Tasks and back, an open topic pane was closed, the message list had scrolled elsewhere, and an unsent draft in the topic was gone. Now everything is where you left it. Messages that arrive while you are away still count as unread until you actually see them.

Topics

• Added: the topic member panel has "Add topic members". Pick people from the source conversation, confirm, and an invitation message in the topic pulls them in.
• Improved: when you type @ in a topic, the topic's participants are listed first and the rest of the source conversation after them. "@all" only notifies the topic's participants and does not pull in people from the source conversation who have not joined.
• Improved: a topic's agent limit (past it, every agent turns "Passive") is now set in the source group's settings, in a new "Topic thresholds" section. A topic can still have its own value. Before, a background refresh could wipe out a number you were halfway through typing. It no longer does.
• Improved: the topic link under the source message shows how many people actually take part, and updates when someone joins or leaves.

Task page

• Fixed: after an agent's Code-mode session exited, or after restarting Gugu without opening the task page yet, direct messages and group @mentions to that agent failed or got no reply. Now the original session is resumed first and then answers. This covers Claude Code, OpenCode, Qwen and Copilot.
• Fixed: quitting Gugu treated the task page's agents as deleted: their direct messages were archived and they came back with a new identity on the next launch. Agents and their direct messages now stay as they were.
• Fixed: Qwen and Copilot Code-mode sessions on a Gugu model answered the first message, but every direct message after that failed. After the session restarted, Copilot stopped at a session picker and Qwen resumed a different session. Both now resume their own session.
• Fixed: in Claude Code's side chat, the first message sometimes went out on your local account, was refused, and the chat showed as running without ever replying. The first message now uses the Gugu model.
• Improved: with a Gugu model, Claude Code sometimes took an extra 20–30 seconds before its first answer. The first turn in Work mode now answers in about 8 seconds, and the first side-chat message in about 5.
• Fixed: if your own ~/.curlrc had options such as -i, Claude Code on a Gugu model never replied and showed no error. Your curl settings no longer affect it.
• Improved: the access key for Gugu models no longer shows up in the system's process list and is no longer written to a file on disk.
• Fixed: with a Gugu model, Codex's /status showed a smaller context window than the model really has (for example 258K for a 1M model). It now shows the real window.
• Fixed: after a failed turn, clicking "Reconnect" turned the failure reason back into a generic message. The real reason and the model service's original text now stay after reconnecting.
• Fixed: when creating a project or task, the agent in the composer was empty even though several agents were installed, and you had to pick one again. It now uses, in order: what you chose this time, the one you last started, the default in Settings, Codex, then any available agent. It is left empty only when none is available.
• Fixed: clicking an in-page link such as a table-of-contents entry in a web preview turned the whole preview white. It now jumps to that spot. Audio, video, embedded pages, fonts and styles pulled in by CSS now show in previews too. The preview for HTML attachments in chat is fixed the same way: in-page links no longer blank it, and external links no longer load the outside page inside the preview.
• Fixed: the tab strip on the right of the New task page and the project home page still had the old style. It now matches the tabs everywhere else.
• Improved: hovering over a session in the sidebar shows an identity card with the session name, session ID, agent name and agent ID; click a row to copy it. "Show the identity card on hover" in Settings controls it. The right-click menu adds "Rename agent" next to "Rename session", which changes the name others use to @ it. The share panel now keeps only the Gugu number, the invite link and the poster / public options.
• Fixed: the conversation minimap left out the turn in progress and inputs added later in the same turn. Both now get marks.
• Fixed: Settings → Storage scanned the same project folder over and over, spun for minutes and pushed Gugu's memory use up to several GB. The reclaimable total was also multiplied by the repeats (4 GB shown as 246 GB). It now finishes in seconds, the total is real, and scanning stops when you leave the page.
• Fixed: with a very large working folder (tens of thousands of files), Gugu stopped responding for more than two minutes after launch. Indexing now takes a few seconds and the app stays usable meanwhile.
• Fixed: with "Enable tmux" on, quitting Gugu also ended the tmux sessions of your tasks. They now stay and are picked up again on the next launch.

Models and billing

• Improved: the model list follows the models Gugu actually serves right now. New models appear on their own, and models that are temporarily not served are no longer listed. Before, they stayed in the list and failed when chosen.
• Fixed: a running session could lose models it was able to switch to partway through, with a "no access to this model" error. Models you could switch to when the session started now stay available for the whole session.
• Fixed: with a Gugu model, you occasionally got an "account not usable" error. At busy times requests from many users were rate-limited together; each client is now counted on its own.

Agents

• Improved: agent settings open in a dialog in the middle of the screen instead of a panel on the right. Long content scrolls inside the dialog, so the buttons at the bottom stay reachable.
• Fixed: in agent settings → Install → source, once you had pointed Gugu at an agent yourself, clicking "Point at it myself…" again did nothing, and the path you had entered could not be seen or changed. It now opens the input with your last value, and that row in the menu is marked "Using".
• Fixed: after an agent stopped a session with its tools, the session still read as "working" or "idle", so you could not tell it had stopped. It now shows as stopped.
• Note: Code-mode Claude Code sessions need Claude Code 2.1.69 or later. With an older version Gugu tells you before starting the session.

Other

• Improved: the whole app was brought in line with one design standard: corner radii, colors, shadows and divider lines now follow one set, and blue is used only for running, connected and selected.
• Fixed: after opening the task page once, a batch of elements across the app that should be round or oval turned into pill shapes.

Topics, rebuilt as separate discussions

A topic is no longer a reply panel inside a group. It is a separate discussion with its own members that only notifies the people in it. This release also fixes long-history paging, session recovery, update installs and a batch of agent problems.

Topics

• Added: topics are rebuilt as separate discussions. Choose "Open topic" on a message and it is created in one step, named after the start of that message (you can rename it). This works in direct messages too. Whispers and retracted messages cannot start a topic.
• Added: a topic opens in a split pane on the right while the source chat stays on the left, still readable and writable. The source message shows a topic link with the topic name, message count, who opened it and when it was last active.
• Added: a topic has its own members. Anyone who can see the source chat can see the topic, but its messages only notify participants. You become one by posting in it or being @mentioned into it.
• Added: topics are nested under their source chat in the chat list. Each chat shows its 3 most recent topics and folds the rest into "More · N", which opens History → Topics with search and a "Mine" filter.
• Added: inside a topic, the read ring next to a message lists who has and has not read it, same as in the main chat.
• Added: when a group or topic has more agents than a set number (5 by default, adjustable in group settings and in the topic's member panel), all of them switch to passive. A system message explains what happened and offers a one-click undo.
• Improved: agent response modes go from three to two, passive and active. The old "adaptive" mode behaved the same as active.
• Improved: the members button at the top of a group chat now opens the member list, with people and agents in separate sections. It used to open group settings.
• Removed: the old in-group reply panel. Replies posted in it were moved back into the main chat as replies to the original message, so nothing is lost.
• Removed: the summary shown at the top of group chats, together with its red "summary generation failed" bar.

Chat

• Fixed: scrolling up in a long group chat stopped at the earliest loaded message. You can now page in both directions, and the message you are reading stays in place.
• Fixed: "Jump to latest" now fetches the real latest messages. Jumping to unread and the new-message count after reopening a chat are more accurate, and a new message arriving while you read history no longer moves your position.
• Fixed: agents with the same name were indistinguishable in the @ menu. Each agent now shows its owner, and picking one sends only to that one. Previously the other same-named agents were added as recipients too.
• Fixed: when an agent was only listening in a group and the group piled up 80 or more unread messages that did not mention it, later direct messages and @mentions never reached it. They now do.
• Fixed: when asked in a group to pass a message to another agent, an agent sometimes sent it to itself and reported it as delivered. Self-sends are now refused.

Task page

• Fixed: switching a Claude Code session from Code mode to Work mode for the first time wiped all prior context, leaving a blank panel and no error. History now replays in full, and switching back and forth keeps it.
• Fixed: sessions saved with a model alias (such as opus) failed to restore, and a full-screen error hid the history. They now restore. Saved history shows first, a failure is one row in the message list, the composer stays, and you can retry.
• Fixed: a failed turn always showed the same "the prompt could not be delivered" line whatever the cause. It now names the real cause, and you can expand it to read the original error text from the model service.
• Fixed: after history replay, tool calls were reduced to a name and a status with no details. Arguments, output, exit codes, failure reasons and attachments are now kept across app restarts, and calls made by subtasks open too.
• Fixed: a Codex session in Work mode could crash the task page mid-turn.
• Fixed: "Rename" in the sidebar context menu flashed an input and lost it. Renaming now uses a dialog and saves only on confirm.
• Fixed: API keys and proxies exported in ~/.zshrc worked in your terminal but were missing inside task sessions, so Codex reported a missing environment variable. Terminal, Code mode and Work mode sessions now all see them.
• Improved: a plain folder can now be a project and run tasks. A new blank project only creates the directory and no longer runs git init, and an existing project keeps working if its .git is gone.

Models and billing

• Fixed: in sessions started on a Gugu-provided model, the in-session model menu showed only one entry under "Your Codex models" and only Default for Claude Code. It now matches the new-task page.
• Fixed: switching a Codex session from a Gugu-provided model to your own subscription failed to reopen the session and reverted the model button. It now switches and keeps the conversation. The same switch in OpenCode is fixed as well.
• Fixed: a Claude Code Work session started on Gugu Claude Sonnet 5 flipped its model panel to Opus 5 and sent one extra, billed Opus request every turn. The panel now stays on the model you chose and the extra request is gone.
• Fixed: during a long Codex turn on a Gugu model, putting the computer to sleep for a while ended the turn with a timeout or "401 Invalid token". The turn now finishes after wake.
• Fixed: if your own ~/.claude/settings.json contained a static API key, Gugu-model sessions used that key and got a 401. Gugu-model sessions are no longer affected by that key.
• Fixed: with OpenCode on a Gugu model, the new-task page offered Low, Medium and High thinking levels but the session showed a fourth slot, and reset threw an error. Both places now match and reset returns to Default.
• Improved: after a membership or credit change lands, an open session can continue without being reopened. It could previously be refused against the old balance until the next renewal.

Agents

• Added: Pi is now in the engine list when creating an agent, and the task page supports it in Work mode (chat UI), not only the terminal UI. Its running status is real, a message sent while it is busy is picked up at the next tool call, and installing Pi while the app is open needs no restart. One limitation: to use Pi's own models in Work mode, the key has to come from an environment variable (one set in your shell config reaches the session). Work mode cannot see a login made with pi login in your terminal. Gugu-provided models are not affected.
• Fixed: Pi asked for approval on every tool call and ignored the "auto-approve by default" setting. It now follows the setting.
• Removed: the "Sign in to use Pi models" entry in Pi's model list. It only opened an empty terminal, so sign in to Pi from your own terminal instead. Sign-in for Claude Code, Codex, Gemini, Kimi, MiMo Code, OpenCode, Qoder and Qwen is unchanged.
• Fixed: MiMo Code could create a session but rejected the first message. It now replies normally on Gugu models.
• Fixed: in Work mode, OpenCode and MiMo Code sessions never received the permission level you chose. They now run at the level you pick, and OpenCode's Plan, Edit and Full Access levels actually take effect. Codex sessions can change permissions mid-session; the change applies on the next turn, and the UI shows the new value at once with a queued notice.
• Fixed: after a Codex session was rebuilt in the background or forked, its permissions fell back from Full Access to restricted with networking off. Resume and fork now keep the level you chose.
• Fixed: opening a second Gugu on the same computer could kill the Codex engine the first one was using. The first restart then waited 30 seconds and failed, and only the second worked. Engines still in use are left alone, and one restart returns you to the same conversation.
• Fixed: starting Codex inside a task session brought up its own update menu, which swallowed your first input. It no longer appears. Update Codex from Gugu's Agents page.
• Fixed: two cases where Codex sessions misjudged a turn. A turn could hang forever after you added input mid-turn, and one failed tool call marked the whole turn failed even when a later call succeeded.
• Fixed: the agent details panel in the Library reopened by itself on your next visit or after a restart.

Other

• Fixed: an update had finished downloading but nothing in the UI let you install it, and "Check for updates" only answered "new version found". Reopening the app also deleted the downloaded package. "Restart to update" in the banner and "Restart to install" on the About page now stay available, and the package is kept.
• Improved: the network requests Gugu makes in the background while idle dropped from 266 to 18 per two minutes. The trade-off is that a "Retry" pressed on another device can take up to 30 seconds to run on this one, where it used to take under 2.
• Fixed: the login page, crash page and menu bar icon still used the old icon. They now use the new one.
• Note: older Gugu builds once wrote a notify entry into your ~/.codex/config.toml. This release removes that one entry at startup and leaves every other byte of the file, including a notify line of your own, untouched.

Six more agents can use Gugu-provided models

OpenCode, Qwen, Kimi and three more agents can use Gugu-provided models, and release notes move to a searchable, shareable web page.

• Fixed: OpenCode, Qwen, Kimi, Copilot, MiMo Code and Qoder can now use Gugu-provided models.
• Added: release notes live at gugu.nextmind.space/help/release-notes/, with contents, search and a link per version; after an update Gugu opens that version's entry.
• Added: each platform has its own release notes, so changes don't mix across platforms.
• Fixed: release notes failing to load on phone and web; failures now give the real reason, not a generic network error.
• Fixed: beta builds that never offered an update.
• Added: a help center covering setup, quickstart, troubleshooting, billing, FAQs and release notes.
• Improved: sidebar row actions appear only on hover or keyboard focus and show which agent they belong to.
• Improved: agents send files by location instead of having the model rewrite the whole file.
• Removed: three sidebar switches in Settings that did nothing.

A slider for thinking levels, and release notes that show up

Thinking levels move to a slider, every Gugu-provided model has been checked, and release notes are there after you update.

• Added: a marked slider for thinking levels, so the current level and the range are visible at a glance.
• Improved: when an agent is waiting for your reply it shows "Waiting", no longer confused with thinking or idle.
• Improved: Gugu-provided models now connect to each provider's official API directly; eight models that did not work were removed, and the remaining fifteen were each verified.
• Fixed: an agent's first greeting after joining a group chat did not show.
• Fixed: release notes were missing after installing a new version, and a version without notes could be marked as read.
• Fixed: the minimap on the right of a conversation did not line up with the messages.

Installing an agent finally shows its progress

Installation tells you what it is doing, can be stopped, and clearly shows what is already installed.

On Windows, agents installed through npm used to fail every single time. On other systems they installed, but you were left staring at a frozen screen — no idea how far along it was, no way to stop it, and no way to tell afterwards which ones were already installed.

That is all fixed: which agent installs, how far it has got, and whether to stop — all of it is yours to control, and what you already have installed is visible at a glance.

• Fixed: update buttons that did nothing, installations that never reopened the app, duplicate prompts and missing release notes; beta and stable now update separately.
• Added: browser-authorized services such as Linear and Notion are available directly inside sessions.
• Improved: the task page turns a dragged session into an @ mention, turns a dropped folder into a path, and can restart Code-mode sessions.
• Improved: an empty installation now says "nothing selected yet" instead of showing an opaque error.
• On mobile: system permissions are grouped on one Permissions page, and background reminders are delivered.

Updating now takes one restart

Gugu prepares the update in the background; confirm once and it replaces the app and reopens itself after a single restart.

Gugu now downloads new versions in the background without interrupting you. Only once the download is finished does the app show "A new version is ready" and a single "Restart and update" button.

That one press is all it takes. No quitting Gugu first, no dragging the installer into Applications, no reopening it yourself — Gugu swaps in the new version after it exits and relaunches on its own.

• Check manually: Settings → About still has “Check for updates”.
• Release notes: after an update, this page explains what changed; the table of contents jumps to earlier releases.

Release process maintenance

This release only fixes how Gugu is published; nothing changes in use.

• Note: this release only fixes how Gugu is published. Nothing changes in use.

Authorize a service, then use it in your session

Services such as Linear and Notion work in a session right after browser authorization, and mobile gets a permissions page and splash screen.

• Added: services such as Linear and Notion can be used in a session right after you authorize them in the browser.
• On mobile: a permissions page and splash screen put first-run permissions in one place.
• Improved: Gugu no longer mixes up which device an agent runs on and who owns it.

The extension marketplace moves to the cloud

The extension marketplace lists extensions from the cloud, and sessions start with the CLI you chose.

• Added: the extension marketplace gets its list from the cloud, so every extension installs from one place.
• Improved: sessions start with the CLI you chose instead of quietly switching to another copy.

Sessions remember, and open faster

Session history survives restarts, startup and the chat list are faster, and status dots stay accurate.

• Added: session history is saved on your computer and survives a restart.
• Improved: faster startup and chat list; workspaces open sooner.
• Fixed: status dots wrong after error recovery, and unneeded dashes in interface text.

Sidebar, attachments and themes come together

A Codex-style sidebar, local copies of attachments, an unread badge on the menu bar icon, and light, dark or system themes.

• Improved: sidebar expand and collapse rebuilt after Codex, and switching workspaces is steadier.
• Added: attachments keep a reusable local copy, and the menu bar icon shows an unread badge.
• Added: light, dark and follow-system themes.

Per-session MCP, and switching models mid-session

MCP can be turned on or off per session, the model can change mid-session, and subagents no longer start in the wrong folder.

• Added: turn MCP on or off for one session without affecting others.
• Added: switch Gugu models mid-session; the next turn uses the new model.
• Fixed: subagents starting in the wrong folder.

A redesigned task page, and agent progress on your phone

A redesigned task page, running agents on your phone, and gpt-6-astra in the model list.

• Improved: the task page is redesigned, with input, messages and status shown consistently.
• Added: view and look after running agents from your phone.
• Added: gpt-6-astra in the model list.

Pinned task messages stay put

Switching sessions no longer duplicates, mixes up or loses the task page's pinned message.

• Fixed: the task page's pinned message repeating or showing in the wrong session after switching.
• Improved: your messages keep their order at the top, and looking back no longer jumps.

One place for purchases, refunds and membership credits

You can buy, upgrade and request refunds on the web, and membership allowance converts straight into credits.

• Added: buy, upgrade membership and request refunds on the web, and view past refunds.
• Improved: each period's membership allowance converts directly into credits, so the balance shown is what you can actually use.

Prices and membership tiers match everywhere

Prices and membership tiers now read the same on desktop and on the web.

• Improved: prices and membership tiers now read the same on desktop and on the web.

Extensions arrive

Extensions are added, and tasks return to their original session more reliably.

• Added: extensions. A task is checked before it returns to its original session, so it never lands in the wrong one.

An accurate model list, and sign-ins that no longer overwrite each other

Model names and availability match what is actually offered, and signing in on several devices no longer overwrites one device's sign-in with another's.

• Improved: model names and availability in the list match the models actually offered.
• Fixed: when signed in on several devices, one device's sign-in no longer overwrites another's.

Release process maintenance

This release only changes how Gugu is published; nothing changes in use.

• Note: this release only changes how Gugu is published. Nothing changes in use.

Sessions and tasks finally connect

Jump between the session page and the task page, and a redesigned composer.

• Added: jump between the session page and the task page without losing the task's context.
• Improved: a composer redesigned after Codex's, with clearer attachment and send states.

The task page starts more reliably

Fixes the terminal and other components failing to load, which kept the task page from starting.

• Fixed: the terminal and other components sometimes failed to load, so the task page could not start.

The task page no longer crashes on open

Fixes the task page crashing at launch and sitting on "startup took too long"; remote projects can be connected, and problem reports include diagnostic details.

• Fixed: the task page crashing at launch and staying on "startup took too long".
• Added: remote projects; problem reports include diagnostic details.
• Improved: a failed task explains why instead of leaving only a red dot.

Fix for the task page failing at launch

Fixes the task page failing as soon as it started after installing 0.2.25.

• Fixed: after installing 0.2.25, the task page failed as soon as it started.

The task page becomes the default home

Open Gugu and hand a task to a coding agent on your computer; messages render LaTeX too.

• Added: the task page is the default home; give Claude Code or Codex a task, no project needed.
• Added: project and workspace tabs, task archiving, and inline and display LaTeX.
• Fixed: the startup session list, empty-session order, split-message resends, and deleted agents' state across devices.
• Improved: sign-ins are kept per device; failed commands hide sensitive details and keep terminal output out of chat.

Local agents use the tools already on your Mac

Claude Code uses your installed copy, Codex keeps your settings, downloads and failures are clear, and the installer is smaller.

• Added: Claude Code uses your installed copy and Codex keeps your ~/.codex settings, so nothing downloads twice.
• Improved: agent downloads show "X MB received"; failures explain why and can be retried.
• Fixed: Linux updates, long member lists, leftover processes, duplicate delivery and permission scope.
• Improved: the installer shrank from about 135 MB to 32 MB; setup failures and UI errors use plain language.

Cloud agents show what they are doing

Cloud agents show what they are doing, feedback takes images, and group invites can add several people.

• Added: cloud agents show activity like reading, editing and searching, not just online/offline.
• Added: paste images into feedback, drafts are kept; multi-select group members.
• Fixed: quoted replies, missed cloud-agent replies, the payment window, group-name editing and multi-agent status.
• Improved: no one can probe emails one by one to see who uses Gugu.

Feedback and session import arrive

Import a local Claude Code or Codex session as a cloud agent, and report problems from the sidebar.

• Added: a sidebar feedback entry that attaches your version and platform.
• Added: import a local Claude Code or Codex session as a cloud agent, keeping its context.
• Improved: exact search by full email, adding people by QR code, and one entry point for adding contacts.
• Fixed: adding agents, the "running" status, group chat summaries and import errors; a download with no progress for 60 seconds is reported as stuck.

Clearer rules for who answers in group chats

An @mentioned agent always gets the message, agents decide for themselves on messages that name no one, and Linux auto-update works again.

• Added: an agent you @mention always gets the message and answers; @all doesn't oblige every agent; on unaddressed messages, each agent decides whether to reply.
• Fixed: quoted-message previews, and agents replying to each other until the group flooded.
• Fixed: Linux auto-update not finding new versions.

Update progress, and accurate credits

Update downloads show a percentage, device choices stay on their own device, and the sidebar shows your real credit balance.

• Added: percentage progress for update downloads instead of a bare spinner.
• Fixed: a choice made on another device leaking into the device picker.
• Improved: the sidebar credit balance follows your real balance instead of a stale number.

Messages no longer disappear silently

Unsent messages go out once a brief network or login drop ends, agents no longer stick on "running", and a daily What's up briefing begins.

• Added: messages that could not be sent during a brief network or login drop go out once it ends.
• Fixed: agents stuck on "running".
• Added: a daily What's up briefing that brings changes worth your attention to you.

Workspace renames and clearer cloud-agent permissions

Rename workspace files and resolve edit conflicts; cloud agents check permissions and privacy before joining groups.

• Added: rename workspace files and choose how to resolve local/cloud edit conflicts.
• Improved: cloud agents check permissions and privacy before joining a group; the creation panel fits phones.
• Fixed: wrong beta UI states, and unexplained cloud-agent creation failures.

One creation flow, and a Discover tab on mobile

Agents are created the same way everywhere, mobile gets bottom navigation and Discover, and failed macOS updates can be retried.

• Improved: creating an agent works the same from every entry point.
• Added: bottom navigation on mobile, with Discover and Me tabs.
• Fixed: macOS updates that could hang; a failed update can be retried.

A quieter message list and fuller Markdown

Fewer repeat notices, agents show their real device, and Markdown tables and strikethrough return.

• Improved: fewer repeated notices, and the empty and sending states no longer cover each other.
• Fixed: agents show their real device, so cross-device actions reach the right one.
• Added: Markdown tables and strikethrough display again.

Smoother device management and agent creation

Device management, a device choice for new agents, and image-upload failures that say why.

• Added: device management and a settings panel; pick the device a new agent runs on.
• Improved: a single creation flow with clearer device status.
• Fixed: image uploads that failed silently.
• Released: an AppImage installer for Linux.

Agent status matches in direct and group chats

Direct and group chats now show the same agent status.

• Improved: an agent's start, running and finished states match in direct and group chats.
• Fixed: wrong status dots and misattributed replies when several agents ran at once.

Call an agent by name, and use your own Claude CLI

Say an agent's name in chat to send it the message, use your own Claude CLI, and get Intel or Apple-silicon Mac installers.

The version-by-version history starts here; earlier development builds had no formal release and are not listed.

• Added: say an agent's name in chat (no @ needed) to send it the message; use your own Claude CLI; push notifications and a group-chat entry point.
• Released: macOS installers for Intel and Apple silicon.

An installer for Apple silicon

Apple-silicon Macs can download and install Gugu directly, and the Android test build gets a fixed download address.

• Released: a macOS installer for Apple silicon.
• Released: the Android test build now has a fixed download address.

Mac and Android ship together for the first time

macOS and Android installers are both available, from one place.

• Released: macOS and Android installers.
• Improved: both installers download from the same place.

The first Gugu desktop release

The first Gugu desktop version you can download and install.

• Released: the Gugu desktop app is available to download.