Changelog · Docs

8 November 2026

Change one value in a secret and keep the rest. Open a secret from Settings → Secrets and choose Change values, which was Replace values. Each saved variable is listed by name: give one a new value, delete one or add one, and every variable you leave alone keeps its stored value. Before, saving replaced the whole set, so any variable you didn't type again stopped being supplied. See Secrets.

7 November 2026

The operations are Cabane's API. A personal access token calls the operations, query, mutate and describe, over HTTP or through the MCP server, and reads your account, your workspaces and their agents. Every other route answers it 404, as the app's own routes always did. The OpenAPI document at /api/openapi.yaml is gone, and Operation reference, in the docs' Developers tab, lists every operation in its place. A self-hosted server makes the same change when it updates to this version. First API request and Requests and responses show the calls.

Woody sets your workspace up with you. Tell Woody, in a conversation they belong to, the work you want to do in Cabane, or ask where to start. They draw on built-in recipes to suggest a setup and help you build it, explaining the pieces as you go. Tell them if you'd rather stop. See Ask your agent.

Every folder in the workspace, from a conversation's files. In Files in this conversation, the first row, All files, opens the workspace's folders beside the conversation, where you can open a file and put it in your message with Mention. While you search the list, its last row, Find “…” in all files, searches the whole workspace for your words. See Messages and mentions.

6 November 2026

Type [[ to put a file in a message. Type [[ and the file's name at the start of a word, and a list of files and conversations opens over the message box; Enter puts the one you choose in the message. The message box's search opens the same list, and a file open beside the conversation has Mention, an @, at the end of its bar. The files a conversation's messages attached or mentioned are listed beside it from Files in this conversation, in the conversation's bar or title menu. See Messages and mentions.

Every device in the workspace is visible to everyone in it. Settings → Devices now lists every device members have added, whatever its owner's sharing setting, and any member can open a device's page to read its state, harnesses and agents. When you choose Where they run for an agent, the list holds every device too: the ones you can't use are grayed out and say why, such as "only the owner can assign agents here". The setting on a device's page that decides this is now called Who can assign agents. Agents can read the workspace's devices as well, so you can ask one which agents run where and why a device isn't working. See Devices and Agent execution settings.

A folder's page shows what changed, then what it holds. Open a folder from Files and its page starts with the folder's charter when it has one, then Recently updated, then a strip of what it holds: folders first, then files, newest first. A folder inside it opens in the strip, behind a Back tile, rather than on a page of its own, and the rooms box opens with a row of tabs along the top, one per shared room. New on the page makes a folder or a file where you are: type its name in the strip and press Enter. Search files in the bar narrows the tree to the names that match as you type, and its last row searches inside files instead. See Files and folders.

A notice about a turn that didn't run says what happened, then what to do. When an agent's turn fails or can't start, the notice under your message now says what happened in one sentence, then how to fix it, and when a reply is the next step it ends with that on its own line, such as "Reply to try again." Where it names a device, it is the agent's device, with the device's name after it as a link: choose the name to open the device's page and see who owns it. See FAQ and Help.
The message box says why an agent can't run. An agent who can't run right now stays in the list of who you can write to, with a warning mark at the corner of their face, and the line beside their name says why and ends on one move: a link to the fix when you can make it, otherwise Choose another agent. Send is off while the agent you chose can't run. See FAQ and Help.

5 November 2026

Files opens on your workspace's charter and its folders. Open Files and the page starts with a card for the workspace charter: its opening paragraph, and who last changed it and when. Choose the card to read the whole charter. Below it, every top-level folder has a box listing its four most recently updated files under Recently updated, each with who changed it and when. Choose a file to open it, or the box's name to open the folder. Files kept at the top level, in no folder, have a box of their own, marked Not in a folder. The folder tree starts closed; the tree icon, Open file tree, at the top of the page, opens it.

4 November 2026

Upload has its own button in Files, and takes a whole folder. Open a folder and choose Upload, at the top of the page. Files… uploads files, and Folder… uploads a folder from your computer with the files in it; before, a folder could only be dragged in. See Files and folders.

3 November 2026

What an agent asks you now takes the place of your message box. In that conversation, your message box holds what they asked, with the answers at the right of its one field; what you type there is your own answer or a note on the one you choose. A go-ahead can name the action in the agent's words, such as Send it and Not yet, and a request to do something offers I've done it and It didn't work. With several waiting, a bar above the message box names who needs you. See Questions from agents.

2 November 2026

A request either needs to hear it's done or needs something back. When an agent asks you to do something, they say which. Tell them it's done, or type what they asked for, such as a command's output, as your answer. If you couldn't do it, send a note on its own, and the request goes back to the agent as not done, with your note. See Questions from agents.
GPT-6.1 Terra for Codex. An agent on Codex can now run on GPT-6.1 Terra, which takes GPT-5.6 Terra's place in the Model menu. An agent already set to GPT-5.6 Terra keeps it. See Harnesses and models.

31 October 2026

Agents can ask you for a go-ahead, or for something only you can do. Besides questions, an agent can now ask you for a go-ahead on one action, which you answer Yes or No, or send you a request for something only you can do, such as running a command on your computer. A note or files can go with either answer, and a note or a file alone doesn't answer a go-ahead. An agent can also attach up to five workspace files to any of the three, shown on its card. See Questions from agents.

30 October 2026

Pair a device with the code its terminal shows. The install line no longer carries a code. Run it on the device, and the terminal prints a code and a link, then waits. Type the code into Cabane where you chose Set up a new device, check which device asked, and choose Confirm: the device is paired to your account and added to that workspace. To pair a device without adding it to a workspace, use Your settings → Devices → Add a device. See Devices.

A task can let the tasks waiting on it go on before it closes. In a template, add releases as a third part of a step's line to make it the release step. Once that step is done or skipped, the tasks waiting on this one go on while the rest of its plan runs, and their waiting step names it, as in Waiting on DEAL12 · Signed. See Plans and templates.

29 October 2026

A task waits until someone starts it. Creating a task, yourself or through an agent, no longer tells anyone: it reads Not started, and Tasks lists it first, under Not started. Choose Start… on the task page, add a Note for the first step's owner if you like, and choose Start: that owner is told in the task's conversation. A task whose first step waits on another task can be set to start by itself when that task is done. See Boards and tasks and Plans and templates.

Boards belong to workrooms, and deleting replaces archiving. Each workroom can have one board, created from the room and kept there. Tasks and boards are no longer archived or put in the trash: deleting a task removes it, and its description goes to the trash as an ordinary file for 30 days. A room's archive still makes its board and tasks read-only. See Boards and tasks.

Incoming hooks. A monitor, a CI job or an alerting script can now post into one room. Choose Incoming hooks from the room's name menu to create a hook, and give its URL and secret to the outside system. Its messages read System · incoming hook. See Incoming webhooks.
Change your sign-in email. In Your settings → Account, Change email asks for the new address and your current password, then sends a confirmation link to the new address. Your email changes only when that link is opened, within 24 hours, and your old address is told once it has. See Settings and usage.

28 October 2026

Tasks are plans. A task now has an intent, saying what it's for, and a plan of steps, each owned by an agent; when one step is done, the next step's owner is told in the task's conversation. Its supervisor, an agent, owns any step nobody else does and is told when the task closes. Stages and stage owners are gone: the list groups tasks by whose move it is, and a board can start every new task from a template file. See Boards and tasks and Plans and templates.

Mark a conversation unread. On a message someone else wrote, choose More → Mark unread: the conversation is unread again from that message, with the New line above it. A mention you've already read in For you has the same choice in its ⋯ menu. See Work activity.
Agents can answer questions about Cabane. Ask an agent in your workspace how to do something in Cabane, such as connecting a device, and they can look it up in Cabane's guides. See Ask your agent.

27 October 2026

For you shows what you've reacted to. For you has a fourth column, Reactions, listing the messages you've reacted to, newest reaction first, each with your emoji. Choosing one opens its conversation at that message, and removing your last reaction to it takes it off the column. See Work activity.

26 October 2026

Search finds tasks. Search, or ⌘K, now lists tasks among its results, found by title, by reference or by what their description says. Each shows its reference, and choosing one opens the task. See Workspace overview.
Mentioning someone keeps the selected agent. When an agent is selected above the message box, mentioning another agent or a person no longer replaces it. The selected agent still receives the message, along with any agent you mention. To write only to people, remove the selected agent and mention no agent. See Messages and mentions.
The shared browser follows its tabs. A window a website opens, such as a sign-in popup, now appears in the browser panel, which switches to it. Once there is more than one tab, a row of tabs lets you choose which one to watch or control. Before, only the first tab showed. See The shared browser.

25 October 2026

An unread conversation opens at the first new message. Opening a conversation with unread messages lands on the first of them, under a New line, with the end of the message before it still in view. A link to a particular message still opens on that message, and a conversation with nothing unread opens as before. See Work activity.

24 October 2026

See which files an agent changed. At the foot of the block under an agent's reply, Changed lists the workspace files the agent wrote, as it writes them, and choosing one opens it. The block's header also shows how long the agent has been working and how many steps it has taken. See Work activity.

23 October 2026

Every person has a Loft. Each person in a workspace now has a Loft: a private workroom of their own, first under Rooms, that nobody else can join. The workspace opens on it, and setup ends with a first message to Woody ready to send there. Its person chooses its agents; nobody can rename, archive or hand it on. See Chatrooms and workrooms.

An agent's reply carries its work. What an agent did to reach its reply now sits with the reply, in one block under its message. While the agent works, the block reads In progress, with what it's doing and Stop; once it's done, All done, with the notes it wrote on the way. Open the full turn lists every step it took. See Work activity.

20 October 2026

Secrets for the agents in a room. The workspace owner can add a secret, a named set of environment variables such as a deploy key, under Settings → Secrets, and put it in a room from the room's menu. Agents working in a room that holds the secret have those variables, and the same agents working in a room without it don't. Agents on OpenCode get none. A value that shouldn't be stored in Cabane can stay on the agent's own device, behind a marker variable. See Secrets.

The workspace charter can be the owner's alone. In the workspace's Settings → General, Who edits the workspace charter can be set to Only the workspace owner; then nobody else, agents included, can change it, and it says who can. See Charters.
Password fields can be shown. On sign-in and password reset, Show password reveals what was typed; Hide password hides it again.

19 October 2026

Private rooms. Only a private room's members can see it. Its owner controls membership and settings and can transfer ownership to another person in the room. Choose Private under Visibility when creating a room; visibility cannot be changed later. See Chatrooms and workrooms.

The workspace owner can list and archive rooms they haven't joined. For the workspace owner only, Settings → Rooms lists every room except direct messages, with each private room's owner and how many people and agents are in it, and archives any of them except General and Lofts. It shows nothing inside a private room the owner isn't in.
See and end an agent's rooms from its page. An agent's page now lists the rooms it's in that the viewer can see, with Remove where they can take it out. Its owner and the workspace owner also see how many private rooms or direct messages they aren't in hold it, and Remove from all takes it out of every one without naming them. See Room membership.
A locked agent can't rewrite its own charter. When only its owner may edit an agent, the agent itself is refused too, on its charter and everything in its folder: what it's told is its owner's decision. See Charters.
Room names are unique. Creating or renaming a room to a name another room in the workspace already has, archived rooms included, is refused with "A room with that name already exists." instead of being given a numbered suffix.

18 October 2026

A room can keep its own skills. A room's folder can hold a skills/ folder, and the Skills panel on a room's charter creates skills there. In that room, a room skill is read before a workspace skill of the same name; everywhere else the workspace skill applies. See Skills.

16 October 2026

An agent works with everything on its device. An agent on a device now runs that device's harness, with its own tools and files there, always. The "Host filesystem and shell access" checkbox and the Locked/Open device policy are gone; agents that were locked gain the device the next time they run. An agent's starting folder and MCP servers stay, always on, and can be edited by whoever can edit the agent. To confine an agent, set it on the device itself: Claude Code's sandbox in the device's ~/.claude/settings.json, Codex's sandbox_mode in its ~/.codex/config.toml. See Harness sandboxes.

Long pastes become a chip. A long paste in the message box collapses into one Pasted text chip, so the box stays usable around it; choose the chip to read, edit or remove it. The message is sent with the full text, and a handle inside a paste calls nobody. See Messages and mentions.