---
url: https://clusto.app/docs/changelog.md
description: >-
  Follow notable Clusto additions, behavior changes, and fixes, including
  changes available before the next release.
---

# Changelog

All notable changes to this project are documented here.

Format follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).

## \[Unreleased]

## \[1.15.43] - 2026-11-16

### Changed

* **Team shown as an icon beside the organization name in the Factory rail** - In the Factory project rail, the Team entry is now a small people icon on the same line as the organization name instead of a separate row below it. The icon keeps its waiting-on-you count badge, and clicking it opens the team as before.

### Fixed

* **Company employees read GitHub issues instead of asking you how** - With `company.agents` on, the Lead, QA and Growth employees in **Factory > Team** used to stop and ask how to read GitHub issues, because their sandbox has read-only clones and no forge token. The daemon now reads the open issues of each employee's repositories (the company repository and its product repositories) with the factory account and lists them in the employee's turn brief: on the first turn of the day, all open issues (up to 30, most recently updated first, with labels, comment count and link), and on later turns only issues updated since. A repository whose issues cannot be read is named in the brief instead of failing the turn. Employees still hold no token. The starter Lead role for new companies now points at that list; the role files of existing companies are not changed, so edit your hq repository's Lead role if you want it to mention the list.
* **Factory runs no longer sit on "CI running" after CI passed** - When GitHub was read in the few seconds between the factory opening a pull request and CI registering its checks, the empty check list was trusted for 10 minutes, so the run stayed in CI wait long after CI went green (and the GitHub view showed no checks). An empty check list on a recently updated pull request is now reread on every poll, so the run moves on within a minute of CI finishing.
* **Empty projects stay folded in the workspace tree** - A project with no sessions no longer expands when you click it, and it no longer shows a "No sessions yet" line. Its expand arrow is hidden, and the Expand and Collapse actions are left out of its menu.

## \[1.15.42] - 2026-11-14

### Changed

* **Much less tool output in agent terminals** - The agent terminal and the expanded activity in Chat view now show a successful command or read as just its header line with an "N lines" toggle; tap it to see the whole output. Failed calls still show their last lines. Two or more consecutive routine tool calls (reads, commands, searches and the reasoning between them) fold into one "+N tool calls (Bash, Read)" row, with the newest call still visible; tap to show them. Edits, todos, plans, subagents and errors stay visible. Edit diffs fold after 16 rows (was 40) and can be collapsed again. `Ctrl+O` still expands everything.

* **New session button on every project row** - Phones now show the + next to the ... menu on each project in the workspace tree, so starting a session is one tap.

* **Factory picker groups projects under their organization** - On phones the Factory scope picker no longer lists a separate "Team" option: picking the organization opens its team, and its projects sit indented one level below it.

### Fixed

* **Attach a photo or a device file in agent chats and terminals** - The paperclip in an agent's prompt, in both Chat and Terminal view, in the assistant and in headless agent sessions, opened only the node context picker, so a phone could not take a photo or pick an image from the device. It now opens the full attach menu everywhere: **Take photo**, **From this device**, and **From a node** where the chat can link context. Headless agent sessions also show their live output again instead of an empty transcript.
* **Phone bar names the agent you just started** - An agent started from the New dialog opens in the Workspace pane, and the phone's top bar showed "No project" and "No session" for it. The bar now shows the agent's project and name, and the previous and next arrows step from it like from a terminal.
* **An agent shows as Terminal or Chat the same on every device** - The Terminal or Chat view is now a setting of the agent itself and travels with the chat across the cluster, so desktop, Android, iOS and the web app show an agent the same way with the same sidebar icon. Before, the choice was saved per device, so an agent started as a Terminal on Windows could appear as a Chat on Android. Switching the view on any device changes it everywhere. On the first run after updating, each device carries its old choices onto its chats once, only for chats that have no view yet. A chat with no explicit view still follows its project's default view, then the global default in **Settings**. Free on every plan.

## \[1.15.41] - 2026-11-14

### Added

* **Agents can create Clusto projects** - Ask a chat or terminal agent to "create a project in Clusto for C:\Repos\SlotLaps" and it calls the new `clusto_project_create` MCP tool (`name`, absolute `dir`, optional `description` and `run_as_user`); `clusto project new` from inside such a session works too. With the limited session credential the project is created only on the agent's own node, for a folder its own OS account owns, and every Run as on it must be that same non-root account; otherwise the agent gets a refusal message and you create the project in the app. Editing and deleting projects stay with you, and the Clusto Assistant chat does not get the tool. On Windows, folder ownership checks now accept a folder you own outside your home directory, such as `C:\Repos\app`. Free on every plan.
* **Subagent cards in chats** - When an agent delegates to a subagent (Claude Code's Agent or Task tool, opencode's task tool, Codex spawn_agent), the Chat view's expanded activity and the Terminal view show a card with the subagent type, description, live status, tool count, time and tokens, last tool or progress summary, and its final reply as Markdown. Its steps fold under **Show N steps**, parallel subagents group under their own cards, and **Open subagent** shows the full steps. "N subagents running" appears above the usage strip (Chat view) and in the prompt status line (Terminal view); tap it to see what each one last did. Headless Claude chats now also receive subagent narration and reasoning. Codex shows the card and status but not nested steps yet.
* **Messages between chats** - With the `chat.messaging` flag on, type `/send <chat> <text>` in a chat to send a message to another of your chats by name, and agents can message your other chats too. A message from another chat is shown with a dashed outline and a **From** link that opens the sender. Untick the accept-messages checkbox on the **General** tab of **Chat settings** to stop a chat from receiving them.
* **Teammate chats** - With the `chat.teams` flag on, right-click a chat in the sidebar and pick the add teammate entry, or type `/teammate`, to start another chat in the same workspace with a name, role, model, opening prompt and an optional worktree. Teammates are indented under their lead in the sidebar, and a status-line segment such as "3 teammates, 1 waiting" lists them with their state. It also offers **Show team** to lay them out, and the sidebar menu can stop the whole team. With `chat.messaging` on as well, start a message with `@name` to send it to that teammate.

### Changed

* **Company pages moved to the organization's Team row in Factory** - The company (Team, Channel and the roster settings) is shared by every project of an organization, so it no longer sits in the tab row beside one project's board. Each organization in the Factory rail has a **Team** row (with the waiting-on-you count, shown even before it has projects); picking it shows **Team**, **Channel** and **Settings** tabs. A project's tabs are now only **Board**, **Insights** and **Settings**, and the Settings tab no longer has a Project/Company switcher. A project with several boards has a chevron in the rail that folds its boards out as rows (the fold state is remembered), and on desktop the strip above the columns just names the shown board with its menu and **New board**. On a phone the header project picker is grouped by organization and lists "RobotaWorks team" (the organization name plus "team") above its projects; the separate company picker is gone. The company pages stay behind the `company.agents` flag (Labs).
* **Simpler New session dialog** - The dialog is narrower and asks only for the essentials: target, **Project**, **Shell** or **Agent** (now with icons), and the shell or AI provider. For an agent, **Model**, **Effort** and the **Terminal**/**Chat** view moved behind **More options**, whose collapsed row names the current choices, and the project description line is gone. The dialog still opens with what you last used in the project.
* **Quieter model picker in New terminal** - The agent terminal's model picker no longer shows the model count and Refresh line, and the app asks a node for its model list at most every 30 minutes instead of every 5.
* **Chat settings is now a dialog with tabs** - The per-chat **Chat settings** panel (the composer's sliders button, or **Chat settings** in F1) opens as a centered dialog instead of a right-side sheet, with **General**, **MCP servers** and **Agent** tabs. In MCP servers every checkbox is always clickable: the chat follows the node configuration until you untick a server, then runs only the ticked ones, and **Use node configuration** switches back. The footer buttons are **Make default for new chats**, **Cancel** and **Save**, and help text is shorter.
* **One-row status line in the Terminal view** - The status line under the prompt in the Terminal view now stays on a single row and clips instead of wrapping at any width, and the key hint that shared the row is gone.

### Fixed

* **Phone switcher bar lists agents** - The bar at the top of the Workspace on a phone (previous and next arrows, project dropdown, session dropdown with its position pill) listed only terminal sessions, so agents were missing and a project holding only agents did not appear at all. It now shows the same projects and rows as the sidebar Workspace tree, agents included: picking an agent opens it, the arrows step through a project's terminals and agents together, and picking a project opens its first terminal or agent. An agent opened on a phone shows the same bar above it, and its **New** entry opens the New dialog on the agent's project. Agent rows have no rename or close buttons; those stay in the agent's own menu.
* **New agent is focused and highlighted** - an agent started from the New session dialog now takes the keyboard focus in its prompt box and shows as the selected row in the sidebar, as a new terminal does.
* **Ctrl+Shift+T starts what you last created** - In the unified workspace (`workspace.unified`), Quick new session no longer always starts a plain shell. It starts what the New dialog would start with a plain Enter: the kind of the focused pane first (a focused agent gives another agent with the same provider and view, Terminal or Chat), otherwise the project's last Create. If the last thing you created in a project was an agent in the Terminal view, Ctrl+Shift+T starts another agent there with the same provider, model, effort and view, shown in the focused pane, and a shell project still gets a shell. From the Chats view it now starts another agent in that agent's project instead of a plain terminal. Outside the unified workspace it still starts the last-used shell, and Split with new session and restarting a lost terminal still start a shell. The shortcut labels now read "Quick session (active project, last shell or agent)" and "Quick session (same project, same kind)".
* **Subagent output no longer disturbs the reply** - A background subagent's late output no longer opens an empty extra turn, and subagent text no longer replaces the chat's own reply or leaks into the live streaming reply, the sidebar's last-line preview or "copy last code block".
* **The Linux CLI installs its own daemon version** - `clusto install` with the Linux `clusto` CLI (x64 and arm64) set up the daemon of the previous release, because the CLI was built before its embedded `clustod`. The CLI is now built after the daemon, so a fresh install gets the matching version without waiting for the node's self-update.
* **Enter sends company messages** - In Factory company threads (the Messages tab of an employee and the shared channel), Enter now sends the message and Shift+Enter adds a new line, matching the chat composer. On touch keyboards Enter still adds a new line and the send button sends. Previously only Ctrl+Enter sent.
* **SSH agent connections use your keys** - SSH connections set to **Default keys** (SSH agent) now run `ssh` as the node's session user (`default_session_user`, the same account its terminals use) when the daemon runs as root or as the Windows LocalSystem service. Before, `ssh` ran as root or SYSTEM and offered that account's agent and keys, the server refused them, and the SSH terminal opened and vanished at once (the app then showed "session ... owner is not known yet"). Saved-key, stored-key and password connections are unchanged.
* **Shells running an agent show the agent icon** - A shell terminal where you start Claude Code, Codex or another agent CLI by hand now shows the agent icon in the sidebar and its pane tab while the agent runs, and goes back to the shell icon when it exits. Ctrl+T from such a pane offers a new agent instead of a shell.
* **Missing check command stops dispatching** - A factory project with no `check` command (neither `check` in `.clusto/factory/factory.toml` on the default branch nor **Check command** in the project's settings) used to start every Ready card, fail it at setup and move it to Review / "Needs you", so with Auto or Full auto card after card bounced there. Once a project's rules fail to load, the dispatcher now starts no more of its cards until the project's settings change or you retry a card (**Retry as a new run**) and the rules then load. The banner now reads "set a `check` command: Check command in the project's settings, or `check` in .clusto/factory/factory.toml on the default branch".

## \[1.15.40] - 2026-11-13

### Added

* **Numbered markers for attached files** - Every file you attach to a chat or agent terminal message (clipboard paste, **Files from this device**, **Photo library** or **Take photo**) now inserts a marker at the caret: `[Image #1]` for images, `[File #2]` for other files. The number is the file's position in the message, so you can write "look at \[Image #1] then compare it with \[Image #2]". The pending file chips and the file chips and thumbnails in the transcript show the same `#N`. Removing a pending file removes its marker and renumbers the later ones. The agent gets the files as a list numbered `#1`, `#2` and so on, and headless agent sessions get each marker replaced by the uploaded file's path in place (files without a marker are still appended after the text). The agent-side numbering needs the next release. Free on every plan.
* **Reasoning effort in the chat status line** - The status line under a chat's message box (and under the prompt in the Terminal view) now shows the chat's reasoning effort right after the model, for example "High effort" or "Max effort", or "Default effort" when the chat has no pinned level and the provider default applies. It is on by default; switch it off or move it in **Settings > Chats > Conversations > Status line** with the new **Reasoning effort (thinking)** switch. Free on every plan.
* **Mark single alerts or one group read in Attention** - Each group card (a session, chat, project or kind of alert) now has a **Mark all in ... read** check button in its header (it names the group) while it has unread alerts, and every unread row has its own check button that marks just that alert read without opening it. Both clear the alerts from the unread count and sync across your devices through the existing read marks. Free on every plan.
* **clusto doctor removes leftover cluster addresses** - A roster address that no known node owns (an address-only entry a retired machine left behind, such as an old build server's hostname) used to be dialed by every node forever, and nothing could remove it because Remove node needs a node. `clusto doctor` now reports each one as a WARN line ("*address* is in the cluster roster but no known node answers there ..."), and `clusto doctor --fix` asks "Remove *address* from every node's roster? \[y/N]" and removes it on every node. The removal is remembered and gossiped with the cluster config, so a node that was offline drops the address when it reconnects and no peer can sync it back. An address a known node holds, or one with a live link, is refused: remove the node instead. Needs the admin token, like the other doctor repairs. Free on every plan.
* **Edit the company roster in the app** - With the `company.agents` Labs flag on, **Factory > Settings > Company** is now a roster editor instead of a read-only list: **Add employee**, **Edit** and **Remove** work in the app, with no need to edit `.clusto/company/company.toml` on GitHub. The dialog has Name, Address (the employee id, filled from the name and fixed once saved), **What it does** (the role prompt), an **Observe / Propose / Act** autonomy picker, a schedule picker (weekdays or every day at a time, every few hours, only when messaged, or a custom cron or `every <n>m|h`), repository chips, a daily budget, turns per day and a model. Saving commits the roster and any changed role prompt to the company repository's default branch with the factory account's token and reloads the roster at once; a protected default branch that refuses direct commits makes the save fail with a message. A removed employee's role file and chat history stay, at least one employee must remain, and a role prompt over 20,000 bytes is edited in the repo, not in the app. The page is simpler (Add employee and Pause/Resume up front, Reload roster and the GitHub links in a menu), and the F1 palette has **Add employee** and **Edit *name***. Hand-editing `company.toml` still works, but a roster saved from the app is rewritten in a normalized form that drops comments other than the standard header. Free on every plan.

### Changed

* **Factory project settings show the plan instead of failing on save** - With the `factory.core` flag on, **Merge on their own up to** in a Full auto project no longer ends in "Project not saved: ... is a Pro feature". On Personal, T1 to T3 are shown with a **Pro** badge and disabled, with a short line and a **See plans** link (no purchase link on iOS). Paid plans see a hint to check the sampling audit before raising the ceiling. A project already set above T0 stays editable if the plan lapses, and then merges T0 changes only.
* **Agents start headless only** - With `workspace.unified` on, the New dialog's **Agent** option offers only the **Terminal** and **Chat** views, so Model and Effort always apply. The TUI choice is gone: to use an agent's own interface, start a Shell and run its CLI, or use **Open in TUI** on an agent. Ctrl+T from a shell running an agent CLI now opens a headless agent of that provider.
* **Quieter toasts on phones** - On a phone-width screen, toast messages now appear at the top instead of covering the composer and bottom bar, show one at a time instead of an expanded stack, and clear after 2.5 seconds. Swipe a toast up to dismiss it early. Desktop is unchanged.
* **Drag sidebar rows on touch screens** - With the `workspace.unified` Labs flag on, hold a sidebar row for a moment on a phone or tablet, then drag it: drop it on a folder to file it there, on **Drop here for the top level** to take it out of its folder, or between a project's sessions to reorder them. A quick swipe still scrolls, a tap still opens the row, and holding still without moving opens the row menu; the list scrolls by itself near its top and bottom edges. This replaces the phone-only **Move to folder** button on agent chat rows and the **Move to folder** item in session and agent row menus; the F1 palette command **Move *name* to folder** still opens the folder picker. The Chats view without the flag is unchanged. Free on every plan.
* **Large pastes collapse in the prompt box** - Pasting more than 800 characters or more than 10 lines into the prompt box of an agent terminal or a chat now shows a `[Pasted text #1 +42 lines]` token instead of the whole text, the way the Claude Code TUI does. The full text is sent to the agent when the message goes out, deleting the token drops the paste, and the collapsed paste survives leaving the chat and restarting the app with the draft. In the Terminal view, a long prompt also recalls as a token with Up and Down, and a long sent prompt in the transcript shows its first 3 lines plus a "\[+N lines]" toggle that `Ctrl+O` (expand all output) also unfolds. Free on every plan.
* **Denser layout on phones** - On touch screens narrower than 768px (or landscape phones) the whole interface is scaled to 0.875, covering text, spacing and buttons, so a typical phone fits about as much as a 411px-wide screen and screens feel less crowded. The terminal font size is unchanged.
* **No horizontal scrollbars on touch screens** - Horizontal scrollbars are now hidden on touch screens. Tab strips and chip rows scroll by swiping, which removes the stacked bars under the Factory tabs.
* **Compact chat composer and status line on phones** - On a phone-width screen the prompt box uses the transcript text size with a one-line placeholder, and only Stop shows while a reply runs and the box is empty (the queue and steer button returns once you type). The agent chip shows a short model name such as "Opus 5.5", and the status line stays on one line with just the bars and percents until you tap it. Tool calls, their output and reasoning also render a step smaller so long commands wrap less.
* **One menu button per sidebar row on phones** - With the `workspace.unified` Labs flag on, project, terminal and agent rows in the sidebar tree show a single "..." button on a phone instead of separate window, mute, pin and close buttons. It opens the same menu a long-press opens. Section headers likewise keep + and put manage and new folder in a menu. Desktop keeps its hover buttons.
* **Folder picker fits phones** - On a phone the folder picker is a full-height sheet with a smaller header and a single footer row of Cancel and Add. Pick the open folder with the checkbox in the path bar instead of a separate Select this folder button.
* **Tighter navigation sheet on phones** - The phone navigation sheet now shows its menu entries as compact rows that line up with the tree headers above them. The node, RTT and session counts share one line, and the footer icons, including Settings, fit on one row.
* **Compact Factory card header on phones** - On a phone, a Factory card's header uses less space: smaller action buttons, tighter spacing and a title clamped to two lines. Tap the title to show all of it. The board's column tabs are slimmer too.

### Fixed

* **Agents in a workspace pane stay live when you switch away** - A pane now keeps the last five agents it showed open in the background. They keep receiving output while another one is on screen, and switching back is instant, at the same scroll position, instead of reloading and redrawing the whole conversation from the top.
* **Pane actions button sits at the top of an agent in the Terminal view** - In the workspace and the Chats view, the pane's actions button stayed 64px down the pane, where the chat header would be, even when the agent showed the Terminal view with no header. It now takes the top corner unless the chat shows its header.
* **Pro, Team and trial plans now apply to your nodes for the Software Factory** - Every node used to run with Personal limits whatever your license, so a paying user got only 2 agent slots and could not raise a project's merge ceiling above T0. The desktop and mobile app now hands its signed license to the node it connects to (on connect, after activation and after online re-validation). The node checks the clusto.app signature, keeps the paid license that runs longest and passes it to its peers, including a peer that comes back online later. Paid plans get 8 agent slots per runner node (Personal 2) and Full auto may merge up to T3 (Personal T0). A Personal device never downgrades a cluster, and a cancelled plan stops applying at its own deadline. The web app cannot hand over a license, so the plan reaches nodes from a desktop or mobile app. The factory ships behind `factory.core`.
* **Chats with running subagents no longer show as done on other devices** - when an agent ended its turn while background subagents or shells kept working, only a device connected to the node running the agent kept the spinner; every other device showed the chat as finished. The node now shares the task count with the cluster and records when the tasks end, so every device shows the chat as working until they finish.
* **Pinned agents stay in their project** - in the unified workspace tree (`workspace.unified`), pinning a terminal or agent used to move it out of its project to the end of the Pinned section, so under a pinned, expanded project it looked like a stray last row. A pinned terminal or agent now stays in its project (or folder) and moves to the top of it. The Pinned section holds pinned projects and folders.
* **`/resume` shows the agents to pick from** - sending `/resume` on its own (the send button on a phone, or Enter after closing the picker) used to jump straight to one earlier agent; it now opens the list. The list is ordered by each agent's last answer instead of a sync timestamp that made every agent look equally recent, shows more than eight entries, and a dedicated-folder chat lists only the agents of its own folder rather than every dedicated chat. In the `/` picker, Enter or a tap now runs the command as in the CLI (`/compact` no longer needs a second Enter), and Tab completes it so you can add an argument.
* **Project search puts the best match first** - typing a project name in the project picker now lists an exact or prefix name match above projects that only match on their folder path or description, instead of ordering everything by how often you use it.
* **Terminal row menu has Rename, Open in new window and Close** - in the unified workspace tree, right-clicking (or long-pressing) a terminal session now offers the same actions as its hover buttons, so they are reachable on touch screens.
* **Esc interrupts the agent in the Terminal view again** - In a chat shown in the Terminal view, Esc (and answering "2" to a paused tool call) did nothing to the running turn, so the agent kept working. Only the Stop button reached the agent. Esc now interrupts the turn the same way Stop does and keeps the conversation.
* **Back closes an open Factory card on phones** - On Android and iOS, pressing the system back button (or swiping back) with a Factory card open returned you to the previous view instead of the board. Back now closes the card and leaves you on the board; a card opened from inside another card also goes straight back to the board.
* **Attach files and photos in headless agent sessions** - The prompt composer of a headless agent session (an `sdk` mode session opened from the session list, in the Terminal or Chat agent view) had no paperclip, so on desktop and mobile you could not attach files from the device or take a photo. It now has the same attach menu as chats: files from this device, **Photo library** (iOS) and **Take photo**. Picked, pasted or photographed files upload to the node that runs the agent (the daemon attachments folder terminal pastes use, 10 MB per file) and their paths are added to the prompt on their own line, so the agent reads them with its file tools. If an upload fails, the prompt is not sent and your text returns to the composer. **From a node...** linking is not offered in this composer.
* **Notes sync errors no longer fill the daemon log with HTML** - When the S3 server for a synced Notes vault answered with an error page (for example an nginx 502), the warning carried the whole page several times over. It is now one line: the error chain up to the first markup or line break, capped, plus the root cause (for example "Connection refused") when the cap cut it off.
* **Factory card tabs on phones** - The card detail tabs (Overview, Chat, Changes, Runs) now stay on one row. The "Runs" tab no longer wrapped onto a second line and painted over the card text.

## \[1.15.39] - 2026-11-12

### Added

* **New workspace from just a name** - With the unified workspace on (`workspace.unified` Labs flag), you can create a workspace by typing a name; there is no folder to pick. The node makes a folder in its chat workspaces root and the result is a project whose files sync across nodes (it shows the **Synced** badge in Locations). Use the **New workspace** button next to **New Project** on the Locations page (Projects tab) or the F1 command **New workspace**. The dialog asks for the Name and, in a multi-node cluster, the Node that creates the folder (new agents in the workspace start there). When it is created, the New dialog opens on **Agent** in the new workspace, so you go from a name to a running agent. It replaces the old New chat workspace flow, which had no equivalent after the migration. Free on every plan.
* **Chat status line with model, reset times and plugin segments** - The usage strip under an agent chat's message box (and the line under the prompt in the Terminal view) is now the status line. It shows the current model (what the running agent reported, else the chat's pinned model), the 5-hour and weekly limits with their reset time inline ("5h 42% resets 14:00"), context fill, cost and credits left, and token totals (off by default). The Claude and OpenCode cost is now labelled "~$0.42 API": what the turns would cost at API list prices; OpenRouter still shows the amount charged. Plugin views with `placement = "status"` add their own segments. In **Settings > Chats > Conversations > Status line** (renamed from Usage strip) each segment has a checkbox and up/down buttons to reorder it, and plugin segments are listed as "`<view title>` (plugin `<name>`)". With the limits switched off, a tap on the line still shows them. New F1 command **Customize the chat status line** opens that tab; the toggle is now **Show/Hide the status line in chats** and the eye-menu entry is **Status line**. Free on every plan.
* **Built-in slash commands in the agent prompt** - In the prompt of a headless agent, in both the Terminal view and the Chat view, Clusto now answers a few CLI commands itself instead of sending them to the agent. `/new` and `/clear` start a fresh agent in the same pane with the same provider, model, effort, permissions, MCP servers, node, run-as user, workspace and linked context (the old agent stays in the list and a reply in progress keeps running). `/resume` turns the picker into a list of the earlier agents of the same workspace, newest first, and typing after it filters by title; `/resume <text>` + Enter opens the first match. `/rename <name>` renames the agent (alone it opens the rename dialog) and `/model` opens the chat settings. Built-ins are listed first in the `/` picker and shadow an agent command of the same name; every other `/command` still goes to the agent as typed. `/new`, `/clear` and `/resume` appear only where the pane can switch chats (Workspace and Chats view panes). The F1 palette has **Start a fresh agent in this pane** and **Resume an earlier agent in this pane**. Free on every plan.
* **Preview attached files before sending** - Files attached in the chat composer can now be opened in the preview before the message is sent: tap or click the file name on its chip.

### Changed

* **Agent terminals close like terminal sessions** - An agent shown in the Terminal view now offers **Close** instead of **Delete chat**, closes at once when idle, and only asks "Close '*title*'?" with **Close anyway** while it is still working, the same as a PTY terminal. The Chat view keeps its delete confirmation.

* **Low-memory banner floats over the terminal** - The "node is low on memory" / "keeps losing its peer links" banner now floats over the top of the terminal instead of pushing it down, and closing it once hides it for good on that device. Node doctor and the Cluster node cards still show the same warnings.

* **Headless agent Terminal view drops the chat chrome** - An agent shown in the Terminal view (for example one started from the New dialog in the unified workspace) no longer shows the chat header bar, the composer's Notes and Chat settings buttons or the feature tips. Notes, open in window, the view switch, Open in TUI, the message view menu, rename and delete now sit in one **Agent actions** (three-dot) menu at the end of the prompt line, next to Send, and stay in the F1 palette. Switching back to the Chat view restores the normal header. It makes the headless agent feel like the TUI and work reliably on a phone.

* **Terminal view prompt looks like a CLI** - In an agent chat's Terminal view the prompt is now a CLI-style line between two rules at the terminal font size, instead of a rounded box at a larger browser font. The plan usage (5h, Week, context, cost) moved from a separate band above the prompt into a status line under it, next to the key hint, and a new Terminal-view chat opens with a single dim line at the top instead of a centered card. The agent's reply also stands apart from its work now: tool calls print smaller with a tinted tool name, their output is dimmed (errors stay red) and reasoning is the faintest, while the reply keeps full contrast. The prompt and usage changes apply only to the Terminal view; the tool styling also applies to the expanded work in the Chat view.

* **Terminals and chats look alike in the sidebar** - In the unified Workspace tree and the Terminals and SSH sections, every row now starts with an icon for its kind: a bot for an agent terminal (a headless agent shown in the Terminal view, or an agent CLI such as Claude, Codex or OpenCode in its own TUI), a terminal icon for a plain shell, and the speech bubble for a chat in the Chat view. Notes, Assistant, Design and Factory chats keep their own icons. Terminal rows use the regular interface font instead of the terminal font, and their waiting, working and done status now sits at the end of the row like a chat's.

* **Project chats keep pasted files out of git status** - A chat started in a project (the agent runs in the project's own folder) now stores pasted files in `.clusto/chat-files/<chat id>/` inside the project. That folder holds its own `.gitignore`, so the files no longer show up as untracked in `git status`. They used to land in an `attachments/` folder. Other chats keep using `attachments/chats/<chat id>/`.

* **Pane actions button moves with the terminal corner** - The three-dot pane actions button in a terminal pane now sits at the top-right corner when nothing else is there, and moves down below the session label, scroll position, find bar, key bar or a read-only / offline banner only while one of them shows. It used to keep a fixed gap at the top even when that corner was empty.

### Fixed

* **Agents in a synced project now sync their files** - An agent started from the New dialog in a synced project (a workspace, including former chat workspaces turned into projects by the migration) now runs as a replicated workspace chat in that folder, so its files sync across nodes like the migrated chats do. Before, it ran as a plain project chat and the folder did not replicate.
* **Pasted files preview from any node** - A pasted image or file in a chat failed to preview with "not a regular file" when the app was connected to a different node than the one running the chat. The node that stores a pasted file now records itself and the file's full path, and the preview reads the file from there.
* **One copy menu on phones** - selecting text on Android or iOS showed the system Cut/Copy/Paste bar and a second Clusto copy toolbar on top of each other. Only the system bar shows now; terminals keep their own selection controls.
* **Esc sends your queued message right away** - When you queue a message while an agent is replying and then press Esc (or Stop), the queued message now goes out as soon as the turn ends. Before, it could sit in the queue until the next chat refresh, most often for an agent whose chat was not the one in focus.
* **Pins and sidebar folders sync across devices** - Pinned terminals, chats, projects and sidebar folders (also your per-chat Terminal or Chat view and the unified workspace folders) now sync through account cloud sync. A pin change uploads within about 1.5 seconds, and another signed-in device picks it up live when the app is shown or its window is focused, without a restart. Settings sync also merges per key against the last synced state, so a change made on this device is no longer dropped when another device synced at the same time. Needs cloud sync enabled and unlocked in Settings > Account; free on every plan.
* **Agents stop jumping around in the workspace tree** - With the unified workspace on, agents under a project now stay in place, newest created first, instead of moving to the top every time one gets a reply. Dragging them into your own order still works as before.
* **Ctrl+Shift+T beside an agent opens a terminal** - In the unified workspace (`workspace.unified`), Ctrl+Shift+T on an agent starts a quick terminal (last shell) in the agent's project instead of a new chat, and a focused agent pane in the Terminals workspace starts it in the agent's project instead of the most recently used one. With the flag off, Ctrl+Shift+T in Chats still makes a quick chat in the same workspace.
* **Notes fit on a phone** - On a phone the Notes header is now one row (vault, node and a three-dot menu holding Search, New note, Chat, Pull, Push and vault settings), the tab strip shows only when two or more notes are open, the editor bar shows the file name instead of the full path with larger buttons, and the line-number gutter is gone. While the keyboard is up, the header, tabs and Files/Editor switcher hide so the note gets the screen.
* **Ctrl+A in a note's preview selects only the note** - In Notes, Ctrl+A in preview mode (or in the preview pane of split view after clicking into it) now selects the rendered note text instead of the whole app window. With the source editor focused it still selects the Markdown source.
* **Ctrl+T starts the same kind of session as the focused pane** - In the Workspace, Ctrl+T (New terminal) now preselects what the focused pane holds: from an agent running in its own TUI (Claude, Codex, OpenCode) the dialog opens on Agent with that provider in the TUI view, from a headless agent on its provider and Terminal or Chat view and project, and from a shell on Shell. Before, it opened on whatever was last created in the project, so Ctrl+T from a TUI agent could start a headless chat. An empty pane still uses the project's last setup.
* **Middle-click closes rows in the Workspace tree on Windows** - a wheel click on a session row closes it and on a chat row asks to delete it, as on Linux; on the Windows app it did nothing. Pinned rows still ignore middle-click.
* **Chats no longer show done while the agent is still working** - A Claude chat (including a headless agent in the Terminal view) could show the done check in the sidebar while subagents or background shells it started were still running, or while Claude replied on its own after a background task finished. The sidebar now shows the working spinner in both cases, and Stop also closes a turn the agent started on its own. You can still send a message as soon as the turn itself is over; only the status icon reflects the background work, and the background-task spinner shows only on the node that runs the agent.

## \[1.15.38] - 2026-11-12

### Added

* **Saved SSH connections are rows in the workspace tree** - With `workspace.unified` on, every saved SSH connection is its own row in the Workspace sidebar, next to projects, with its terminal sessions listed under it. **+** connects (opens a new SSH terminal) and the gear edits the connection. The row menu has Connect, Edit, Move to folder, Collapse/expand, Close all sessions (only that connection's) and Delete connection (removes the saved connection, not the server). Connection rows can be filed into sidebar folders like projects, and idle ones hide with the hide-empty-projects setting (Show inactive brings them back). The separate SSH sidebar section no longer appears in this mode: **New SSH connection** and **Manage SSH credentials** are in the F1 palette, and the New session dialog's SSH target still works. SSH sessions opened before this change, or whose saved connection was deleted, stay in an **SSH** group.
* **Calendar block for plugin views** - A plugin view can show a `calendar` block: one to twelve months from a `start` month, with marked `days` (each with a tone, an optional label read on hover and by screen readers, `outline` for a ring such as a public holiday and `dot` for a day that needs attention). Give it `select` buttons and the user picks a range by clicking a day, dragging across days or Shift+click to stretch it; on touch, tap two days or long-press and drag; with the keyboard, arrows move, Enter picks, Shift+arrows stretch and Escape clears. A bar under the selection shows the range and the buttons, and pressing one posts its action with its `values` plus `from` and `to` (YYYY-MM-DD, both included). Weekends are dimmed and today is underlined. See the Plugins docs page.

### Changed

* **Migrate all profiles for the unified workspace** - In **Settings > Labs > Unified workspace** (`workspace.unified`), the **Migrate** button is now **Migrate all profiles**: approve once and the connected profile migrates right away, while every other saved connection profile migrates automatically the first time any of your devices connects to it. For each profile the app turns the cluster's `workspace.unified` override on if it is off, then runs the same migration (node database backed up first as `data.sqlite3.pre-unified-<timestamp>.bak`, nothing deleted). The daemon checks still apply (every connected node on the same version, no chat answering); a refused profile tries again on its next connect. The approval is the synced setting `clusto-unified-workspace-migration`, so it follows the account through cloud sync (per device without it). The panel shows **Approved for all profiles** with a collapsible Profiles list marking each profile **Migrated** or **Waits for connect**, a toast announces each profile that moves, and **Preview migration** still works for the connected profile. The panel also shows in Labs once the account approved, even if this cluster's flag is still off.

* **Full auto factory tiers up to T3** - A Full auto project can now approve plans on their own up to any tier (T0 to T3) and merge changes on their own up to any tier, where it used to stop at T1 plans and T1 merges. Merging anything above T0 on its own stays a Pro capability, and a production deploy is still never automatic.

* **Attention is a compact card list** - Each session, chat or category is one card with its name, node and alert count in the heading. Rows are a single line: what happened ("Finished", "Session is asking"), the message and the time, without repeating the session name on every row. A card shows its three newest alerts and folds the rest behind **Show older**, so a busy session no longer fills the page.

* **Locations drops the Workspaces tab once you have migrated** - With `workspace.unified` on and the cluster migrated, **Locations** (`/projects`) no longer shows the **Workspaces** tab, because the migration turned every chat workspace into a project. Old `?tab=workspaces` links open **Projects** and the F1 "Show chat workspaces" command is hidden. Projects whose folder syncs across nodes (former chat workspaces) show a **Synced** badge next to their name. Before the migration the Workspaces tab shows a hint with an **Open migration** button that leads to **Settings > Labs > Unified workspace**. With the flag off nothing changes.

* **Inactive projects are hidden in the workspace tree** - With `workspace.unified` on, a project shows in the sidebar only while it has a terminal session or an agent that is working, waiting for input, has an unread answer, is open, or was used in the last 24 hours. Pinned projects and the built-in Notes and Scratch always show. Use the **Show inactive projects (N)** row at the bottom of the tree, **Settings > Navigation > Hide inactive projects** (formerly "Hide projects with no sessions") or the F1 palette to bring them back. Every device starts from the trimmed list once; the old setting is not carried over.

* **Settings > Chats becomes Settings > Agents after the migration** - With `workspace.unified` on and the cluster migrated, the Chats settings tab is named **Agents** and its defaults card reads **New agents**, since those model, effort and permission defaults pre-fill the New dialog for every agent session. Old `?tab=chats` links still open it. With the flag off nothing changes.

* **One New dialog starts shells and agents** - With `workspace.unified` on and the cluster migrated, the New dialog (New session: Project / SSH / This PC) starts shells and agents in one place. Picking **Agent** shows AI provider, Model and Effort in one row, then the view (Terminal / Chat / TUI); Model and Effort apply to the headless Terminal and Chat views. **Create** starts the agent directly in the focused Workspace pane instead of opening the separate New chat dialog as a second step. Every "New chat" entry point (F1 "New chat", the project page, the **+** on a project in the sidebar, an empty pane, chat workspace pages) opens this dialog on Agent in that project, and Scratch and Notes projects start agents in the node's chat folder as before (Notes links the vault as context). The dialog remembers the last setup per project on this device (Shell or Agent, provider, model, effort, view, shell); a project never used before falls back to its default agent and view, then the last setup used anywhere, then **Settings > Agents > New agents**. The old New chat dialog stays only for the pre-migration layout and design sessions.

* **Plugin Update dialog is redesigned** - **Settings > Plugins > Update** reads the plugin as soon as it opens, with no "Read plugin" click. It shows the source as a summary (for example "Folder on Contabo" and the path), the version change (0.2.0 to 0.3.0), contributions marked **New** and the **Removed** ones. When nothing changed it says "Already up to date" and the button becomes **Reinstall**. The trust warning appears only at install or when an update adds new code paths. Git plugins keep a branch, tag or commit field with **Check again**. In the Install dialog "Read plugin" is now **Preview**. The plugin list shows a count summary ("2 views, 1 time connector, 1 account") and folder plugins name their node. Plugins settings and dialogs are now translated, including Czech.

### Fixed

* **New terminal starts agents directly in the project** - With the Labs flag `workspace.unified`, New terminal > Agent with the Terminal or Chat view no longer opens a second New chat dialog or turns the project folder into a replicated chat workspace (which re-owned the repository and copied up to tens of thousands of files to every node, making Create take about 40 s). The agent now runs in the project's own folder on the project's node as the project's run-as user, so the repo's `CLAUDE.md` and `.claude` settings load, and the Working Directory field is hidden for headless agents. The daemon now refuses a non-synced project folder as a chat workspace, so update the daemon and app together.
* **Right-click menus no longer pile up** - Right-clicking a second row (sidebar rows, the Terminals and Chats trees, files, notes) while a menu was open used to leave the first menu open underneath the new one, and choosing an item closed only the top menu, so one seemed to stay open. Now each right-click (or long-press on touch) closes the open menu and opens only the menu of the item under the pointer, and choosing an item closes it.
* **Folder plugins install across operating systems** - Updating or installing a folder plugin no longer fails with "a local plugin path must be absolute" when the app is connected to a node with a different OS than the plugin's node (for example the app on a Windows node and the plugin folder on a Linux node). The check accepts Unix, drive and UNC absolute paths whatever the connected node runs.
* **Working indicator stays on while Claude's background shells run** - a terminal session no longer shows as done while a background command Claude started is still running; the node watches the session's processes and clears the working state once they exit.

## \[1.15.37] - 2026-11-11

### Added

* **Unified workspace** - Behind the Labs flag `workspace.unified` (Settings > Labs, off by default, free on every plan), Terminals and Chats become one **Workspace** sidebar tree: a Pinned section for any item, nestable folders you make yourself (layout only, deleting one moves its contents up), every project listed even without sessions (Settings > Navigation can hide empty ones), and built-in Scratch and Notes projects. A new dialog starts a Shell or an Agent in a project, and an Agent shows as Terminal, Chat or TUI. **Settings > Labs > Unified workspace > Preview migration** shows what would move and **Migrate** backs up the node database first (`data.sqlite3.pre-unified-<timestamp>.bak`), turns named chat workspaces into projects and puts every chat in one; nothing is deleted and turning the flag off brings the old views back. Drag terminals and agents to reorder them inside a project; on touch, use Move up / Move down in the row menu. See the Unified Workspace docs page.
* **Terminal view for chats** - Every chat can show as **Terminal** (an activity log like the Claude Code TUI: one line per tool call, condensed output with expand, inline diffs with Open diff, ANSI colors, pinned todo list (Codex plans show there too), numbered permission prompts answered with 1/2/3, Up/Down prompt history, Ctrl+O expand all, `[` and `]` between turns) or **Chat** (answers with the agent's work folded into one summary row per turn). Switch from the chat header or F1; the choice syncs across devices and the default is in Settings > Conversations. Team and task agent sessions (the session detail page) now use the same composer as chats, with prompt history and slash commands, and get the same Terminal/Chat view switch.
* **Open in TUI** - The chat header and F1 have **Open in TUI** (not for OpenRouter): it ends the headless agent and runs the provider's own TUI (`claude --resume`, `codex resume`, `opencode --session`) on the same conversation in a terminal. The chat takes no messages while it is open and shows a banner with Show terminal; after you close it, sending in the chat continues the same conversation. When you come back (close the terminal, then send, retry or reopen the chat), the Claude and Codex turns made in the TUI are copied into the chat transcript once, under the note "Continued in the TUI. The turns below were made there." (or "No turns were made there."). OpenCode TUI turns are not copied yet, and a `/clear` inside the TUI starts a new conversation that is not copied. Reopening a chat while its TUI terminal is open no longer starts a second agent.
* **Take a factory card** - A card can be taken by whoever works it outside the factory: a chat, a terminal session or you by hand. The board shows who holds it, the factory never runs a taken card and no other agent or session can take it. On a card use **Take**, **Release**, **Open chat** / **Open terminal** and, on folder projects, **Work in a chat**, which takes the card, starts a chat on the project's folder and sends the brief; all are in the `F1` palette. Agents use the `clusto_factory_take_card` and `clusto_factory_release_card` tools, and the CLI has `clusto factory card take` and `clusto factory card release [--column]`. A hold ends when the chat is deleted or the session ends, or when the card moves to Backlog, Ready or Done. Ships behind the `factory.core` flag. Free on every plan.
* **Lite factory projects** - A project can be Lite: a plain kanban board where the factory never runs cards and you work them in your own chats and terminals. It is always Manual and the run settings are hidden. Turn it on in the project's Settings or in the Add project wizard; a project with a running card must be stopped first. Ships behind the `factory.core` flag. Free on every plan.
* **Plugin tabs in the Time view** - a plugin view with `placement = "time"` shows as a tab of the Time view instead of a sidebar page.
* **Local mode** - The Windows and Linux desktop app now works on this computer with no node installed. Choosing **Simple** on first launch, or **Set up Clusto on this computer** on the Connect view, runs a node inside the app: nothing is installed system-wide, there is no administrator, UAC or pkexec prompt and no download, so it works offline. The app pairs with it and saves a **This computer** profile. The node lives in the app's local data folder (`local-node`), listens on loopback only (port 9178 or the first free one in 9180-9500) and runs standalone. It starts with the app and stops when you quit; terminals keep running in their holder processes and reattach on the next start, and a crashed node restarts by itself (it gives up after 5 exits in 2 minutes). It updates with the app, never on its own. The Settings node card **Clusto node on this computer** shows it as **Run by this app** with **Start**, **Stop** and **Restart**. Builds without the bundled CLI still use the elevated installer. Free on every plan.
* **Finished-turn check mark on terminal sessions** - Sessions in the sidebar Terminals tree and the mobile session switcher show a green check mark when the agent finished its turn and is neither working nor waiting for an answer. It clears when the next turn starts. Claude Code's own idle star glyph is no longer shown in the session label.
* **Choose what Attention shows and counts** - The Attention header has a new **Filter** button that opens a **Show and count** popover with one row per kind of alert (Waiting for your answer, Terminal agents, Chat replies, Tasks, Factory, Reports, GitHub, Azure DevOps, Other). **Show** lists the kind in Attention and **Count** adds it to the bell and Attention badges; a hidden kind never counts. **Reset to all** restores the default and a dot on the button marks a changed filter. The choice syncs to all your devices with account cloud sync, and the popover is also in the `F1` palette as **Attention: choose what to show and count**. Free on every plan.
* **Zoom the time calendar with Ctrl+scroll** - In the Time view's Calendar tab, hold Ctrl and scroll (or pinch on a trackpad) over the grid to step through the hour-height levels. The time under the pointer stays in place and the page itself no longer zooms. The zoom buttons show a "Ctrl+scroll" tooltip.
* **Pin terminals and chats** - Pin a terminal with the pin button on its sidebar row (hover on desktop, always visible on touch) or `F1` **Pin or unpin terminal** / **Pin terminal *name***; pin a chat with **Pin to top** in its right-click (long-press) menu or `F1` **Pin chat *title***. Pinned rows move to the top of their folder or workspace group and show a pin icon. A pinned terminal cannot be closed and **Close all** keeps it, and a pinned chat cannot be deleted (nor can a workspace holding one) until you unpin it. Pins are stored in app preferences (`clusto-pins`) and sync across your devices. Free on every plan.
* **Hide weekends in the time calendar** - A new **Weekends** button in the Calendar toolbar hides or shows Saturday and Sunday, remembered per device. With weekends hidden the wide layout shows Monday to Friday, phone paging stays within weekdays, and **Today** on a weekend jumps to Friday.
* **Restart a stalled terminal session** - When a terminal session stops responding, a banner appears on it with a **Restart process** button. It restarts the session's process in place; if the node refuses, an error toast says why.
* **Heavy-load banner on terminal sessions** - A session now shows a banner when the node it runs on is under heavy load, so a slow terminal or agent is explained instead of looking frozen.
* **Vacation view in the Task request plugin** - The Task request plugin has a new vacation view for planning and reviewing time off requests. Its browser extension was updated to match.

### Changed

* **Long chats scroll smoothly at any length** - The chat transcript is virtualized, so a chat with thousands of messages scrolls as smoothly as a short one. The **Show earlier messages** button is gone because the whole chat is always there to scroll. Browser find (Ctrl+F) only searches the part of a very long chat that is on screen.
* **Esc and Stop interrupt instead of ending the agent** - In every chat, Esc or the Stop button interrupts the running turn and keeps the agent: Claude, Codex and OpenCode keep the conversation and take your next message at once (OpenRouter still stops the turn, and so does a sandboxed run-as account). Each OpenCode chat turn now runs through a short-lived local OpenCode server, which adds a few seconds to the start of a turn. An interrupted turn shows "Interrupted" instead of an error, and sending a new direction while the agent works no longer restarts it.
* **Model and permission mode apply live** - Changing a Claude chat's model or permission mode now applies to the running agent without a restart.
* **Quick log links set the project** - `clusto://time/quick-log` and `clusto://time/start` take `&project=<name>` and book on the time project of that name (exact match first, else ignoring case; archived projects are skipped). When no project matches, the entry is booked without a project and the toast names the project that was not found.
* **Design mode moved to Settings > Labs** - Chat design mode is now behind the `chat.design` flag, off by default. With it off, **New design session**, the **Chat | Design** switch in New chat and the design preview pane are hidden. A chat already in design mode keeps working, and its Chat settings still show the Design mode switch so you can turn it off. Turn the flag on in Settings > Labs to get everything back.
* **One-command Linux server install** - The Linux install script (`curl -fsSL https://dl.clusto.app/install.sh | sh`) now starts the `clusto install` setup wizard itself on your terminal, through `sudo` when you are not root, like the Windows script. Before, it only installed the CLI and you ran `sudo clusto install` as a second step. With no terminal, no `sudo`, or `CLUSTO_SKIP_WIZARD=1`, it only installs the CLI and prints the next step; if the wizard fails it runs `clusto doctor` and exits with an error.
* **Claude Code scrolls smoothly in Clusto terminals** - Each node's AI setup now has a **Claude Code smooth scrolling** check that sets `wheelScrollAccelerationEnabled` to `false` and `CLAUDE_CODE_SCROLL_SPEED` to `1` in `~/.claude/settings.json`, so Claude Code's fullscreen view moves a steady 3 lines per mouse notch on every platform instead of jumping on fast scrolls. Each value is written only when the key is absent, so one you set yourself is kept (set `wheelScrollAccelerationEnabled` to `true` to get acceleration back), and nothing is written while the node's AI setup is off.
* **Terminals stay up to date in the background** - The desktop app now keeps up to 32 terminal sessions (newest first, on online nodes) streaming in the background and fetches their screen snapshots one at a time, so switching to any of them opens instantly from memory instead of showing "Fetching session history...". A session that loses its in-memory copy (resized from another device, or a very busy stream overflowing the buffer) is re-warmed quietly, at most once per 30 seconds. The indicator now only appears on a blank terminal or when you choose **Reload full history**; background refreshes of a terminal that already shows content (reconnect, take-over) are silent. Phones are unchanged and keep fewer streams.
* **Sixteen features left Settings > Labs and are now on for everyone** - These no longer have a switch: opt-in usage data (still asks for your consent first, with Settings > Privacy), the GitHub view, the Azure DevOps view, user plugins (Settings > Plugins and the `clusto plugin` commands), automatic Agent Docs sync (each node's mode still applies: two-way, receive-only or manual), GitLab accounts (Settings > GitLab), the Time view and timer, local terminals on a desktop PC without a node (including the Explorer **Open a local terminal** option), tearing out layouts and node terminals into their own window, Notes multi-vault (the vault switcher and creating vaults; a second vault still needs Pro, Team or a trial), 24-bit color in new terminals (`COLORTERM=truecolor`), terminal answers while a pane is unfocused, auto-continue after a dropped Claude Code response (at most 3 times in 10 minutes, cancelled when you type), limited credentials for sessions that run as another non-root account, refusing plaintext LAN credential routes, and terminals that survive a node restart. Only `factory.core`, `factory.issue_sync` and `company.agents` remain in Labs.
* **Plaintext LAN credential routes are now refused** - The client port no longer serves credential routes over unencrypted HTTP to any network other than loopback, Tailscale and the networks listed in `[transport] plaintext_networks`; those requests get `403`. There is no warn-only mode anymore. An old app that reaches a node only over a LAN address must be updated from clusto.app, or the network added to `plaintext_networks`.
* **Terminals survive restarts by default** - New terminal sessions on Linux and Windows nodes run in their own holder process, so a node update, daemon restart or crash no longer ends them. Terminals already running when the node first updates still end once.
* **Factory project Settings redesigned as cards** - The Settings tab is now General, Autonomy, Agent pipeline (Commands / Agents / Limits), Runner, Delivery, Dry run, Holdout scenarios, Advanced and Danger zone. Autonomy is a three-way choice (Manual / Auto / Full auto) with a one-line explanation each, and both tier ceilings are always visible: in Full auto you pick "Approve plans on their own up to" (T0 or T1) and "Merge on their own up to" (T0, or T1 on Pro); in Manual and Auto the plan ceiling is fixed at T1 and a "Merge T0 changes on their own" setting (`auto_merge_t0`) replaces the merge tier.
* **Every factory.toml key is a form field** - Setup, check, test (slow suite) and deploy commands, plan first, clean-context review, author and reviewer model (a picker on the runner node), extra instructions, budget per card (USD), check and CI fix attempts, largest change (lines), open pull requests at once and intake label (repository projects only), and holdout pass rate (%). An empty field uses the repository's value or the default, shown as the placeholder; on/off fields offer repository-or-default, On and Off. The fields edit the project's Clusto-side rules, which win key by key over `.clusto/factory/factory.toml`. The raw text, recipes and risk map moved under Advanced. Editing a field rewrites that text and drops its comments, and if the text does not parse the fields lock and Save is blocked until it is fixed.
* **Mark done and Delete card in the Factory card header** - A card whose run stopped waiting on you (for example the agent made no commits because the issue was already fixed, attempts or budget ran out, or a plan awaits approval) now has a **Mark done** button in the card header. It ends the run and moves the card to Done without a merge, and an open pull request stays open. Before, the Move menu only offered Merge for Review to Done, which read wrong when there was nothing to merge. A Blocked card (for example after Abandon) can be marked done too, from the Move menu, the card header or by dragging it to Done. **Delete card** now also shows in the card header for Review and Blocked cards, and deleting a card whose run waits on you abandons that run first, so it is one step. The F1 palette has **Mark done** for the selected card when it applies.

### Fixed

* **File preview stays open on the way to it** - Hovering a file path in the terminal opens a preview card, but when more paths were printed below the hovered one, moving the pointer down into the card crossed another path and the card closed the moment the pointer reached it. The card now stays open while the pointer is inside it.
* **Right-click keeps the selection in Claude Code** - In Claude Code's fullscreen mode, and in other programs that track the mouse (tmux, vim), right-clicking selected text to open the terminal menu used to also send the click to the program, which copied the text and cleared the selection (or pasted when nothing was selected). The right-click now only opens Clusto's menu.
* **Claude Code keeps its spinner while background tasks run** - A session that ended its turn while background shells or sub-agents were still running used to drop its working spinner. The daemon now reads Claude Code's per-task status records (OSC 7501 child records) and keeps the session marked working until the last background task finishes.
* **Folder projects now show their GitHub issues** - A factory folder project ("On your nodes") never linked its Main board to GitHub, so a checkout with 12 open issues could show a single hand-filed card. When you create a folder project, the daemon now reads the checkout's `origin` remote; if it is a github.com repository that one of your GitHub accounts (any purpose) lists, the Main board links to it with auto import on, so open issues land in Backlog and closed ones move cards to Done. Existing folder projects are not relinked: use the board's `...` menu > **Edit board** > **Link GitHub issues**, pick the repository and turn on **Import new issues on their own**, or run **Import issues** to pull them in at once. The link is read-only, so the factory never comments on or closes those issues.
* **Node Doctor CPU headroom works on Windows and macOS** - These hosts have no Linux pressure data, so the row used to show "Not Installed" with nothing to do. Without it, the check now averages CPU use over the last 15 minutes: 90% or more is an error ("Close or lower the priority of the heaviest processes, or move heavy jobs to another machine"), anything lower is Ready with the measured value. Linux keeps its rule (40% or more of the time waiting for a CPU). Checks that do not apply to a node, or have no CPU samples yet, now carry the badge "Not available" instead of "Not Installed".
* **Node Doctor finds a Hive-era node on this computer** - Settings > Node Doctor now shows a "This computer's node" row when the PC the app runs on still runs the old `hived` daemon instead of `clustod`, even when that daemon is too old to report its own "Daemon install" check. The row is detected from the OS service manager (Windows services or systemd). **Fix** (and **Fix all**) upgrades the node in place after one administrator prompt (UAC on Windows, pkexec on Linux): it fetches the current `clusto` CLI and runs its install repair, and the new daemon switches the node to `clustod` and keeps its data, token and sessions. On Linux without pkexec the row shows a command to run instead: `curl -fsSL https://dl.clusto.app/install.sh | bash && sudo /usr/local/bin/clusto doctor --fix`.
* **New terminals keep keyboard focus** - A terminal started from a menu (such as the sidebar's right-click menu) no longer loses focus to the menu's button a moment after it opens, and a This PC terminal from the New terminal dialog now opens focused.
* **No raw colour codes in Factory command output** - The Factory dry run result, card check output and the CI logs sent to agents no longer show terminal colour codes such as `[2m[22m[39m`. The daemon strips ANSI escape sequences from command output before storing it. An existing dry run result keeps its old text until the dry run runs again.
* **Checks with fine-grained GitHub tokens** - A fine-grained GitHub token can never read check runs (GitHub offers the Checks permission only to GitHub Apps), so the pull request detail and the Factory item's Changes tab failed with "Resource not accessible by personal access token", Re-run failed checks failed the same way, and the poller showed no checks, which let the factory treat a red CI as green after 15 minutes. When check runs answer 403, Clusto now reads the jobs of the commit's GitHub Actions runs instead (Actions read permission), so checks, job logs, Re-run failed and the factory CI gate work with these tokens. Checks from non-Actions CI apps stay invisible to a fine-grained token.
* **Large pastes no longer stall terminal typing** - Pasting a large block of text into a terminal is now sent in small ordered pieces that never split a character or an escape sequence. Keys typed right after a big paste are no longer held up behind it, and pasted brackets and mouse reports arrive intact.
* **Faster connect to a node on the same computer** - When your home node runs on the same machine as the app, its direct address is now tried at once instead of waiting behind the Clusto Tunnel. The app connects to it sooner at startup and after a reconnect.

## \[1.15.36] - 2026-11-10

### Added

* **Agents can book and edit time entries** - The `clusto` MCP server every agent gets has four new time tracking tools, so you can ask an agent to "log 2 hours on project X this morning" or "fill in yesterday": `clusto_time_entries` lists the account's entries in a date range with the total and the time projects, `clusto_time_entry_create` books a finished entry (`start` plus `end` or `duration_minutes`), `clusto_time_entry_update` changes one by `entry_id`, and `clusto_time_entry_delete` removes one. Three more tools manage time projects: `clusto_time_project_create`, `clusto_time_project_update` (rename, color, client, archive) and `clusto_time_project_delete`. Times are RFC 3339 (without an offset they are read in the node's time zone). Agents use them only when you ask and change only the entries of the account that owns the running timer; time connectors and imports stay in the app. Needs the `time.tracking` Labs flag, free on every plan.
* **Every agent can read the docs and the cluster** - `clusto mcp` now serves fifteen read-only tools to every agent that were only in Clusto Assistant chats before: `clusto_docs_search` and `clusto_docs_read` (the embedded Clusto docs), `clusto_guide`, `clusto_cli_help`, `clusto_mcp_status`, `clusto_cluster_status`, `clusto_cluster_config` (secret values are never returned), `clusto_list_sessions`, `clusto_list_projects`, `clusto_list_teams`, `clusto_list_reports`, `clusto_read_report`, `clusto_github_prs`, `clusto_azure_devops` and `clusto_daemon_log` (tokens redacted). Agent session credentials may now send the matching daemon requests (cluster configuration with the cluster token masked, GitHub and Azure DevOps accounts and snapshots, the daemon log).
* **Agents can manage tasks, notes and factory cards** - `clusto mcp` adds `clusto_task_create`, `clusto_task_update`, `clusto_task_delete`, `clusto_task_trigger`, `clusto_task_cancel`, `clusto_task_pause` and `clusto_task_resume`, with the same rule as the `clusto task` CLI: a session credential changes only tasks that run as its own OS account on its own node. `clusto_notes_move` moves or renames a note or folder (never overwrites) and `clusto_notes_delete` removes a note or an empty folder. `clusto_factory_cards` lists a project's cards (optionally one column) and `clusto_factory_move_card` moves a card to a column when you ask. The Clusto Assistant still has no task write tool and uses action cards.
* **Notes, time tracking and every agent tool from a shell** - New commands run the `clusto mcp` tools without MCP, against the local daemon like `clusto lessons`: `clusto notes vaults|ls|search|read|write|mv|rm` (`--vault`; `write` takes `--content`, `--file` or stdin, and `--base-sha256` to update), `clusto time status|start|stop|ls|add|edit|rm` and `clusto time project add|edit|rm` (for example `clusto time add "standup" --start 2026-11-09T09:00 --minutes 30 --project Clusto`), and `clusto tool ls [--json]` and `clusto tool call <name> '<json>'` to run any tool by name (for example `clusto tool call clusto_github_prs '{}'`).
* **Install the node without sudo or administrator rights** - `clusto install --user` installs the node daemon for the current account only: no root on Linux/macOS, no administrator prompt on Windows. It installs to `~/.local/bin` (Linux/macOS) or `%LOCALAPPDATA%\Programs\Clusto` (Windows) and starts at boot or sign-in through a systemd user unit with lingering (Linux), a LaunchAgent (macOS) or an `HKCU` Run entry (Windows). Sessions and chats run as that account only, a machine has either a system install or a user install, and `clusto doctor` shows the Linux unit as "clustod user service". Self-update, restart and survive-restart terminals work as before.
* **Node Doctor flags a starved host and flapping peer links** - Node Doctor (and `clusto doctor`) now has two more checks per node: "CPU headroom" (Linux: errors when processes waited for a CPU 40%+ of the time over the last 15 minutes, and suggests lowering heavy jobs' priority, e.g. a systemd drop-in with CPUWeight=20 and Nice=10 for a CI runner, or moving them) and "Peer links" (errors when this node had to reset the link to a peer 3+ times in the last hour, names the peer, and suggests checking CPU headroom on both ends and the network path with `tailscale ping <node>`).
* **Markdown preview and images in Factory card bodies** - The card body ("What should change") in the card pane and in the New card dialog now has Write and Preview modes. Preview renders the Markdown like the GitHub issue pane (headings, lists, code, tables, links and images; remote images are click-to-load). A card with a body opens on Preview, an empty one on Write, and a read-only card shows only the rendered preview. In Write mode you can paste, drop or pick images with the **Add image** button (PNG, JPEG, GIF or WebP, up to 10 MB each). Each image is stored on the project's runner node in the daemon's attachments folder, never in the repository, and stays readable by the account the project's agents run as; the body gets a `![name](/absolute/path)` line. Preview shows these images inline and tapping one opens it full size. When a run starts, the brief lists the card's attached images and tells the agent to open them. An agent running inside a factory machine sandbox (`[sandboxes]`) cannot see these files yet.
* **Initialize company repository** - Behind the `company.agents` release flag (with `factory.core`), a GitHub factory account whose company repository (default `hq`, or the account's Company repository setting) does not exist yet or has no `.clusto/company/company.toml` now shows its Team page with a "Set up the company repository" callout instead of an empty page or a "Repository not found" error. **Initialize company repository** (on the callout, on the company Settings empty state, in the Factory **...** menu and in the F1 palette) creates `<org>/<repo>` as a private repository when it is missing, commits a starter roster with one `lead` employee and its role prompt `roles/company/lead.md`, and reloads the roster. It refuses when a roster already exists and never overwrites one. Creating the repository needs the factory token's Administration (write) permission on the organization and committing needs Contents (write); without Administration, create the empty repository on GitHub and press Initialize again.
* **Terminals survive a node update or restart** - Behind the `terminal.survive_restart` release flag (node scope, off by default; turn it on in Settings > Labs), every new terminal session on that node runs in its own small holder process (`clustod pty-hold`, `clusto-pty` in `ps`). A node update (`clusto update-node`, Update All), a daemon restart or a daemon crash no longer ends it: when the daemon is back the session re-attaches with the same id and tab, its scrollback intact (including output produced while the daemon was down, up to the last 4 MiB), and full-screen programs such as Claude Code repaint. It works on Linux and Windows, not macOS. Terminals started before the flag was turned on, and every terminal after a machine reboot, still end. `clusto doctor` shows a "Terminal holders: N running" line on Linux.
* **Node Doctor repairs a node stuck on hived** - Settings > Node Doctor shows a "Daemon install" row for any node still running as `hived`: an error with the reason the switch to `clustod` is skipped, or "Fix restarts the daemon" when nothing blocks it. The node's **Fix** button (also **Fix all**) now runs the repair on that node remotely: it stages the cutover rollback backup, moves stale leftover binaries aside, and when nothing blocks the cutover any more restarts the daemon, which then switches to `clustod` and keeps its data, token and sessions. The node is back after about a minute. Nodes already on `clustod` show no row.
* **Open a chat workspace in Files** - The workspace page has an **Open in Files** button, also the F1 command **Open workspace *name* in Files**, which opens the workspace folder in the Files view on the node that serves it.
* **Edit a chat workspace** - The workspace page (Chats > a workspace folder) has a new **Edit** button, also the F1 command **Edit workspace *name***. **Node** pins every chat in the workspace to the chosen node (agents restart there with their next message, workspace files replicate there) and new chats start there; it works for any workspace folder, including ones never created through New workspace. **Folder** moves the workspace to an absolute path on that node (Browse picker): an existing folder such as a repo is used as it is, with **Copy instructions and default context** (on by default) copying the workspace's instructions and default context into it, while a new or empty folder receives the whole workspace with its files (same disk only). Every chat in the workspace follows.

### Changed

* **Tailscale gets CPU priority on Linux nodes** - clustod now raises `tailscaled.service`'s CPUWeight and IOWeight to 1000 (drop-in `/etc/systemd/system/tailscaled.service.d/zz-clusto-priority.conf`, written at daemon start, never restarts tailscaled). Peer links run over Tailscale, so on a busy host its packets used to wait behind builds long enough for peers to reset the link.

* **Feature tips: X hides for now, Don't show again retires** - The X on a feature tip (the lightbulb hints) now hides it only until the next app start. A new **Don't show again** button on every tip retires it for good. Retired tips, including ones retired by using the feature, are saved in your profile settings (a synced `clusto-` key carried by account cloud sync and cluster preferences), so they follow you to other devices instead of being device-local. Turning tips off entirely is still per device, and **Settings > Appearance > Feature tips > Show tips again** also brings back tips hidden for now.

* **Auto-continue reacts faster** - after "Connection lost mid-response", Clusto now types `continue` 2 s after the cut instead of 5 s, so it resumes before you would reach the keyboard; typing first still cancels it. Each scheduled, cancelled or failed auto-continue is now written to the node log (`terminal.auto_continue`).

* **Shorter What's new** - The What's new dialog after an update now lists only the headline of each change (up to 6 per version, then "and N more") instead of the full notes, with a **Full release notes** link to that version's section of the online changelog.

* **Attention view shows only the notification history** - The Attention view (`/attention`) no longer has tabs. The "Now" tab (the live list of agents grouped Needs you, Ready to review, Running and Idle) is removed, and the view shows the 7-day notification history with **Mark all read** in the header. The bell, `Ctrl/Cmd+Shift+A` (Open attention list) and `Ctrl/Cmd+I` (Open notification history) all open this page. The Node Doctor tip that showed when Needs you was empty is gone too.

* **Taking over a terminal is instant and reports its result** - Taking over a terminal that is open on another device no longer redraws it. The node now answers every take-over with an explicit result, so a refused or failed take-over is reported at once instead of silently reverting.

* **Typing stays responsive on a slow link** - Keystrokes travel as compact binary frames (about 25 bytes instead of 99), and the node sends a keystroke echo ahead of bulk output, so typing stays responsive while another terminal floods. `Ctrl+C` on a flooding terminal stops the stale output within a second, and after a long outage a busy terminal is rebuilt from a screen snapshot instead of replaying megabytes.

* **Idle apps stop receiving cluster status** - The full cluster status is no longer pushed to every client every 10 seconds. Live node metrics flow only while the cluster view or the Settings cluster card is open, and the sidebar shows "Nodes online n/m" instead of the age of the newest peer heartbeat.

* **Chat replies stream with far less data** - A reply streams as compact deltas (about 12 times fewer bytes per token) only to devices that have the chat open. Other devices get a status update at most once a second.

* **Opening a chat never takes over its running turn** - Take-over now happens only when you send a message or retry, after a 20 second grace for the node running the chat. A banner shows while that node is out of reach.

* **Long chats open with the most recent messages** - A long chat shows its latest messages first and a **Show earlier messages** button (or scrolling up) brings in 30 more at a time.

* **Faster app start** - The web entry script is about half its size. The terminal engine, Markdown rendering, the Czech locale and dialogs load on demand.

* **Node reconnects send only what changed** - Peer nodes reconnect by comparing digests and sending only the changed records (about 12 kB instead of about 15 MB per reconnect). Peer links no longer reset when healthy: a 5 second ping, any received traffic as a sign of life, and a threshold that scales with the measured round trip.

* **Terminal recording writes about 10 times less to disk** - Output is stored in flushes of up to 30 seconds or 64 kB. A normal daemon restart loses nothing; a hard crash can lose up to 30 seconds of recorded scrollback.

* **Lower idle CPU** - Background loops on the node run only when something changed, and Agent Docs sync reads files only after a change.

* **Smoother terminals on desktop and phones** - The app batches terminal output once per frame. Phones stop streaming terminals that are not on screen and resume them when you return.

* **Terminal diagnostics show echo time and frame gaps** - Settings > Terminal > Diagnostics has a new **Measured on this device** section with the sampled keystroke echo time and frame gaps. The numbers stay on the device.

### Fixed

* **Same project on two nodes now shows on both** - Adding a project with the same name and folder as an existing one on another node (for example a repo checked out at the same path on two PCs) silently returned the other node's project, so the new node's project never appeared. The node is now part of a project's identity, so each node gets its own project.

* **Working spinner on nodes with an old config** - A node whose `config.toml` still held the agent detection rules an earlier release wrote out (Claude matched on braille, which current Claude Code no longer shows) never marked its Claude sessions as working, so the sidebar spinner stayed off. The daemon now replaces such a frozen default list with the current rules on startup; a list you edited is kept.

* **Windows node no longer stuck on hived** - An early aborted Hive-to-Clusto cutover could leave `clustod.exe` and `clusto.exe` next to `hived.exe` in `C:\Program Files\Hive`, and every later cutover was skipped with "both ...hived.exe and ...clustod.exe exist". The daemon now renames such leftovers to `<name>.stale-<timestamp>` before the cutover check, only while it runs as `hived` with its rollback backup staged, no cutover in progress and no `clustod` service registered. The node then switches to `clustod` on its next restart or update.

* **Closing a feature tip no longer shows the next one** - After you close, turn off or use the tip on screen (for example the one at the bottom after creating a session), no other tip appears until the next app start.

* **Terminal no longer stuck read-only after a node blip** - when the node that owns a session drops off for a few seconds and comes back, the open terminal reclaims and reattaches on its own instead of staying on "Session unclaimed - read-only" until you tap Claim.

* **One notice instead of a stack of relay errors** - when a request has to go through another node and that link is down or stalled, the app shows a single "*node* is not answering right now" notice that replaces itself, instead of a raw `failed to forward to peer: ...` toast per keystroke.

* **iOS attach menu opens each picker directly** - In the iOS app, the paperclip attach menu (terminal key bar, terminal context menu and chat composer) now lists **Photo library**, **Take photo**, **Files** and **From a node...**, and each opens its picker straight away. Before, choosing **Files from this device** opened a second iOS sheet (Photo Library / Take Photo / Choose Files) on top of the app menu. iPhone HEIC photos picked from the library are converted to JPEG so agents can read them. Android and desktop are unchanged.

* **Find in notes works** - Ctrl+F in an open note did nothing. It now opens a find and replace bar in the editor (Enter / F3 for the next match, Shift+Enter for the previous one, Esc to close), and from preview-only mode it switches to split view first. Ctrl+Shift+F also opens the vault-wide search again while a note is open. **Find in note** is in the F1 palette too.

* **Note link suggestions match the app theme** - the list that opens after typing `[[` in a note now uses the app's colors, font and highlight instead of the editor's stock look, and shows more notes at once.

* **Large chat workspaces replicate** - A chat workspace over 16 MiB stopped replicating, and creating a chat in one failed with "chat workspace exceeds the 16 MiB replication limit". Workspaces now replicate as a file list with a SHA-256 hash per file, and each node copies only the files whose hash it does not hold, from a node that has them, in 1 MiB chunks written straight to disk and checked against the hash. Workspaces up to 2 GiB (512 MiB per file) now follow the chat to every node without the memory spikes the old whole-workspace copy caused, and an unchanged file is never sent again. A copy interrupted by a restart resumes on its own. A workspace over the new limit still gets its chat; only its replication is skipped.

* **Node resource charts survive restarts** - Each node now saves its last hour of CPU, pressure, memory, swap, network and disk samples to `host-metrics.json` in its data directory once a minute and reloads them at startup, so a daemon restart loses about a minute of history instead of the whole hour. The app (desktop, mobile and web) also asks each online node for its saved hour when it connects, and again when a peer comes back online, so the Cluster node card sparklines show the past hour immediately instead of starting empty.

* **Attention notifications follow what you have read** - A chat notification ("Reply is ready") reads by itself once that chat is read, on any device, and a report notification reads once the report is read or archived. Reading an Azure DevOps pull request or work item in its view also reads its notifications. What is on screen reads its notifications too, including ones that arrive while it stays open or while the window was in the background: a focused terminal reads its session notifications, and an open GitHub pull request or issue or Azure DevOps item reads its own. **Mark all read** in Attention (and the F1 command **Mark all notifications read**) now clears everything: every notification, unread chat, Azure DevOps pull request and work item across all accounts, report in the feed and terminal bell marker in the sidebar. It shows whenever any of those has something unread, and a toast says so if reports cannot be marked. The GitHub nav badge counts reviews waiting on you and is not read state, so it stays.

* **Dragging Factory cards between columns works again** - Every drop was refused, for example Blocked to Ready. Dragging now uses pointer events instead of native drag and drop: the mouse wheel scrolls lists and the board while you drag, holding a card near the board's left or right edge or a column's top or bottom edge scrolls that way, and Escape cancels the drag.

* **No project picker above the Backlog quick add** - On the All projects board the Backlog quick add no longer shows a project picker at the top of the column, where it read like a board switcher. The **Add to** project picker now appears next to the Add buttons once you type a task.

* **Deleted chat workspaces stay deleted** - Deleting a named chat workspace removed its folder only on the node that owns it. Connected to another node, the app still listed it as an empty workspace, because that node kept the copy its replicated chats ran in. A deleted workspace's leftover copies on other nodes are no longer listed.

* **Faster Factory board switching and updates** - Board pushes no longer deep-proxy every run, unchanged cards are not re-rendered, and subtask and dependency counts are computed once per update instead of once per card.

* **Clicking into a long note no longer scrolls it to the middle** - Focusing the source pane of a long note, in any editor mode, jumped the note to its middle. The helper that keeps a focused field above the on-screen keyboard ran on every focus and centred the editor's whole content, which is as tall as the note. It now reveals only the line holding the caret, and never scrolls for a field taller than the visible area.

* **No doubled spaces in the iOS terminal** - Space, punctuation or the double-space period right after a suggestion or autocorrect no longer sends an extra space to the terminal.

* **Chat workspace edits are no longer missed by replication** - A quick edit to a chat workspace file that left its size unchanged could be missed by replication. It is now detected and replicated.

## \[1.15.35] - 2026-11-09

### Changed

* **Smaller replicated chat workspaces** - A replicated chat workspace is now limited to 8 MiB per file and 16 MiB per workspace, down from 32 MiB and 128 MiB, so a full fetch stays quick on slow links. A larger workspace stops replicating with a warning, as before.

### Fixed

* **A busy chat workspace no longer floods the cluster** - A chat whose agent kept writing files into its workspace (one downloaded about 15,000 map images) made every node resend the whole workspace, up to 128 MiB, to every other node several times a second. That ran the busiest node out of memory, dropped node links every 30 seconds and drove heavy network use on client PCs. Workspace changes now replicate as only the added, changed and removed files, at most once every 5 seconds per workspace, and a node that missed a change fetches the whole workspace once.
* **Nodes stop dialing stale peer addresses** - Nodes kept dialing addresses with random ports picked up from old connections. The roster now prunes them once a node announces its real peer port, and the config file is rewritten only when a sync actually changes something.
* **Nodes no longer drop each other every 30 seconds when a reply is slow** - A forwarded request that took longer than its timeout (a GitHub or Azure DevOps call, a large reply) was treated as a dead link: the node disconnected its peer, both sides resent their full cluster state, and the retried request timed out again. Clients relayed through that node saw constant lag, and the extra state traffic pushed a node's memory until it was killed. A timeout now counts against a link only when the peer has sent nothing since the request went out.
* **The app returns to your home node** - When the node the app normally connects to was unreachable, the app connected through another node and then made that node its new home, so it stayed one extra hop away for good. It now keeps its home node, checks every 2 minutes whether it answers again, and moves back once it does. The web app is unchanged.
* **Expired Claude sign-in shows the sign-in card** - when Claude's login on a node has expired and cannot be refreshed, the chat now shows "Claude is signed out" with a Sign in button instead of a generic "Something went wrong" error.

## \[1.15.34] - 2026-11-08

### Added

* **Several boards per project in the Factory** - a factory project can now have more than one board. Every project has a **Main** board, and existing cards are on it after the upgrade. A row above the columns switches boards (tabs on desktop, a picker on phones), shows a linked repository as a badge and how many cards need you. The board menu has **Edit**, **Import issues** (linked boards) and **Delete** (not the Main board, and only once it has no cards); **New board** adds one. The project still owns the repository, rules, runner, autonomy and run limit, and agents pick Ready cards from all of its boards. Behind `factory.core`, free on every plan.

* **Link a Factory board to GitHub issues** - **Edit board > Link GitHub issues** takes any GitHub account from Settings > GitHub and a repository (`owner/name`), with an optional **Auto import new issues** and a label filter. **Import issues** lists the repository's open issues (up to 500, newest updated first) with search and checkboxes; each one becomes a Backlog card titled "#N title" with the issue body as its brief and its labels, and issues already on the board are greyed out. Auto import checks every 10 minutes on the node that hosts the GitHub account. A card's overview has **Link issue** / **Unlink** and **Move to board**, both off while a run holds the card, and a card whose issue is closed on GitHub moves to Done on its own. A link only reads: Clusto does not comment on or close issues through it, except for a forge project's Main board on its own repository, which keeps commenting on and closing the issue after the run merges. On a folder project, the run's brief carries the issue and asks for a `Fixes owner/repo#N` line in the commit message, so pushing the delivered branch closes the issue. `clusto factory card add` takes `--board`, and `clusto factory board` prints each board's columns. GitHub only for now.

* **Open a previewed file in a tab** - the file preview card has an **Open in tab** button that opens the file as an editor tab beside the focused terminal or chat. It works in the Terminals and Chats views and is not shown for PDFs, which the card renders itself.

* **First-launch choice: Simple or Advanced** - the first time the app starts it asks whether to run Clusto on this computer (Simple) or to connect to other machines (Advanced). Simple sets up a node on this computer for you.

* **Notes folder colors** - right-click a folder in the Notes tree (or use its menu, or F1 > **Folder color...**) to give it one of nine named colors. The color is stored in `.clusto-folders.json` at the vault root and syncs between nodes like a note.

* **Clone a repository as a project** - the GitHub view has a **Clone** button on repository rows, and **New project** has a **Clone from URL** option. Pick HTTPS or SSH and a node and folder; the clone uses that node's OS user and their own git credentials (`gh auth`, an SSH key or Git Credential Manager), never your Clusto GitHub token. The CLI has `clusto project clone <url> --path <dir>`.

* **Project repository links** - the project dialog has a **Repository link** field that ties a project to a GitHub repository (`github:owner/name`) explicitly. It overrides the git remote match and works offline. `clusto project update --link` sets it; an empty value clears it.

* **Session colors and title templates** - **Settings > Appearance > Session colors and titles** sets a session title template (`{conversation}`, `{project}`, `{node}`, `{script}`, `{user}`) and color rules that match on node, command, user or project (first match wins). Colors show as a stripe in the sidebar, a dot on pane tabs and a dot in the mobile switcher. Projects can also set their own color and template. Settings sync between your devices.

* **Clusto terms glossary** - a short explanation of node, cluster, session, chat, project, working directory, workspace, team and task opens once after the first connect. Reopen it from F1 > **What do Clusto terms mean?** or the Connect page button. First-run screens also have tap-to-explain term hints, and **Learn more** opens the new Concepts page in the docs.

* **Learn-by-using feature tips** - a small tip appears once at a time for features you have not tried (command palette, split panes, sending files to a terminal, and more), never before you have seen the glossary. Closing a tip or using the feature retires it on that device. Turn tips off or on in **Settings > Appearance > Feature tips**, or with F1 > **Turn off feature tips** / **Show feature tips again**.

* **More feature tips and a Tips list** - ten tips in total, including phone-only tips for the command palette button and the terminal key bar, plus tips for chat files and queue, the assistant after a failed task, Node Doctor, note wiki links and the Clusto Explorer menu. The glossary dialog has a **Tips** list.

* **Command palette button on mobile** - phones show a search button in the header that opens the command palette.

* **Attention view across nodes** - the new **Attention** view (`Mod+Shift+A`, or the bell in the sidebar header and mobile header, or the Attention menu entry on Android) groups what needs you right now: **Needs you** (with the agent's question shown inline and a Copy button), **Ready to review**, **Running** and **Idle** (agent sessions only, 10 shown plus **Show all**). The bell shows a count. `Mod+I` opens the notification center.

* **Notification history** - the notification center has a **History** tab with alerts grouped by session, chat, project or category. Read state is shared across the cluster, answered prompts are marked **Answered**, and history keeps 7 days or 500 alerts per node.

* **Team tab beside the conversation** - a chat or terminal session that started a team now has a **Team** tab next to the conversation with the roster, each member's state and a tap to open its session. It opens by itself once on desktop when a team is created.

* **Latency overlay for terminals** - **Settings > Terminal > Diagnostics** (or F1) turns on an overlay that shows where typing time goes (total, key, invoke, native, echo, parse, paint) plus events per second, KB per second, long tasks and the active renderer. A **Renderer** override picks Automatic, WebGL or DOM, and the Logs view can export the `terminal-latency` line.

* **Run as for tasks** - a task can run as a chosen OS account. The account resolves from the task, then the project for this node, then the project's all-nodes entry, then the node default session user, then the daemon account. The task list shows the **Run as** account and where it came from, and `clusto task new` / `task update` take `--user`.

* **Continue a task run's Claude session from its report** - **Discuss in chat** on a report posted by a finished Claude task run has a **Continue the run's session** choice that resumes the run's own Claude conversation in a chat, so you can reply and keep working. It needs the same OS account and never runs as root, and only your reply is sent. `clusto report post --agent-session <id>` records the conversation for other callers.

* **Lessons limited to chosen nodes or accounts** - **Choose nodes...** when approving a lesson, or **Limit** on an approved row, restricts it to some nodes and OS accounts. It still syncs everywhere but stays hidden from agents elsewhere.

* **May control this node (directional trust)** - **Configure Node > May control this node** limits which nodes of the cluster may start or drive sessions, tasks and chats on a node, change its settings, update or restart it, or send it plugins, Agent Docs, lessons and other records that act there. The node can still control the others, and the dialog warns when an allowed controller can itself be controlled by a node you left out. A limited node no longer accepts the cluster token from the app or the CLI, so pair with that node's own token. The CLI has `clusto node controllers [--node ID] [--set a,b | --any]`, and the setting is `controllers` under `[cluster]`. Off by default; metadata such as sessions and projects still replicates.

### Changed

* **Project issue mirroring is now the Main board's link** - the project setting "Mirror the repository's issues onto the board" is gone. The same behavior is now the Main board's GitHub link (**Edit board > Link GitHub issues**).
* **Shortcuts dialog lists paste and attachment shortcuts** - the Keyboard shortcuts dialog now lists the paste and attach shortcuts (text, files and screenshots, Alt+V for a clipboard image, drag and drop), is translated to Czech, shows Cmd and Option on macOS and reflects custom bindings.
* **One checklist for settings that take several nodes** - Notes **Kept on**, plugin nodes, agent doc nodes and forge standby nodes now share the same node checklist. **Kept on** is available without the `notes.multi_vault` flag; the flag now gates only the vault switcher and creating more vaults.
* **Easy local setup in the Daemon setup tab** - the Daemon setup tab has a one-click setup for this computer, and F1 has a global **Set up Clusto on this computer** command.
* **One New terminal entry point** - `Mod+T` and the sidebar **+** open one **New terminal** dialog with Project / node, SSH and This PC (desktop, `terminal.local`) options. The palette has **New terminal**, **New SSH terminal** and **New terminal on this PC**; typing `user@host` offers a prefilled new SSH connection.
* **Folder browse on every working-directory field** - Tasks, Teams, Add worker, Adopt session and Knowledge backup now have a folder button that browses the node. SSH private-key selection uses the shared file picker with a node switcher, and first-run screens say "node" instead of "daemon".
* **Browse button for program and shell command fields** - custom shell, project default shell, node default shell, shell list Program and the team agent custom command now browse for the program on their node; picking one keeps any arguments already typed.
* **Faster typing and less redraw work in terminals** - the terminal does less work per key and per redraw.

### Fixed

* **File preview for Windows paths with backslashes** - hovering or tapping a path like `C:\Users\...\guide.md` printed by an agent on a Windows node no longer sticks on "Loading preview...". The same path now also opens in an editor tab.
* **Factory dry run shows that it is running** - starting a dry run in the Add project wizard or project settings now shows a running status until the result arrives, instead of nothing at all while setup and the check run on the runner node.
* **Option+V image paste works on macOS** - pasting a clipboard image into a terminal with Option+V now works on macOS.
* **Stricter cluster trust** - a node's identity and a forge account's host list can no longer be changed by an unrelated peer; **Trust new identity** now applies on the node where you run it.
* **Task and run updates stay on allowed nodes** - live task and run updates no longer reach nodes that are left out of a limited Notes vault.
* **Escape is no longer delayed in Windows run-as sessions** - Escape used to wait about 50 ms before reaching a run-as session on Windows; it is now sent at once.

## \[1.15.33] - 2026-11-06

### Added

* **One-click node setup from the Connect view** - on Windows and Linux, when the app is disconnected and no node runs on this PC, the Connect view shows a **Set up Clusto on this computer** card. One button installs the Clusto node as a background service and connects the app to it, with no wizard and no pairing code to copy. The OS asks for administrator rights once (Windows UAC, or the Linux pkexec prompt); the app then pairs with the new node, saves a profile named after the machine and connects. On Linux without pkexec or a polkit agent the card shows a copy-paste command and connects by itself once the node is up. If a node already runs here but no saved profile reaches it, the card offers **Connect**, which opens Add Profile prefilled with the address and node identity. A first run with no saved profiles now lands on the Connect view, the Explorer menu's **Set up this PC as a node** uses the same flow, and the F1 palette has **Set up Clusto on this computer**. `clusto install` gets `--pairing-code-file <path>` and `install.ps1` honours `CLUSTO_PAIRING_CODE_FILE`. Free on every plan; macOS is not offered.
* **GitHub Projects board** - the GitHub view has a new **Projects** tab that shows a GitHub Projects (v2) project as a kanban board, with a picker for your own projects and your organizations' projects. The columns are the project's "Status" field options plus "No status". Drag a card to another column (desktop) or use the card's **Move to** menu (phone, keyboard) to set its status on GitHub; the move shows at once and rolls back with a message if GitHub refuses it. It needs the token scope `project` (read-only `read:project` is enough to view); without it the board shows which scope is missing and the fix: `gh auth refresh -h github.com -s project` for a gh CLI token, or connecting GitHub again on clusto.app, which now asks for Projects access, for **Use my Clusto GitHub sign-in**. Boards stop at 1000 items and project lists at 100 per owner, and say when they are cut. A project without a "Status" field is shown read-only, and cards cannot be reordered within a column. An issue or pull request card opens in the GitHub view, other items open on GitHub. Behind the `github.view` flag, free on every plan.
* **Drag and drop on the Azure DevOps sprint taskboard** - in the Overview tab's Sprint taskboard widget, drag a work item card to another state column in its own lane to change its state in Azure DevOps (moving a card between people's lanes is not supported). Every card also has a **Move to** menu for touch and keyboard, and a failed move snaps back with an error. The board now shows every state of the work item types on it, including empty ones, like the Azure DevOps taskboard.
* **Change history for GitHub issues and Azure DevOps work items** - the detail pane now has **Comments** and **History** tabs under the description, each with a count, and the tab you pick is remembered on the device. History is a timeline, newest first, of who changed what and when (relative time, full date on hover): value changes show old and new, labels, tags, assignees and links show as added and removed chips, and edited rich text collapses to a "Text edited" button that opens a word-level diff. Azure DevOps history comes from the work item's newest 200 revisions, GitHub history from the issue events; older entries load 30 at a time with **Show older changes**. The F1 palette has **Show history of issue #N** / **Show comments of issue #N** in the GitHub view and **Show history of work item N** / **Show comments of work item N** in the Azure DevOps view.
* **Waiting status for Gemini CLI and GitHub Copilot CLI, and a device row in Node Doctor** - every node now wires the `?` marker and notification for Gemini CLI (hook entries in `~/.gemini/settings.json`) and GitHub Copilot CLI (a Clusto-owned `~/.copilot/hooks/clusto.json`) on its own, next to Claude Code, Codex and opencode; Node Doctor gets **Gemini CLI waiting status** and **GitHub Copilot CLI waiting status** checks. Cursor has no hook that fires while it waits, so it keeps the copyable instructions, now a **Copy setup instructions** button in the Node Doctor card footer instead of the separate **Other AI tools** card. **Settings > Node Doctor** also has a **This device** row at the top: system notification permission, alert popups, alert sound, Android background checks and muted sessions, plus **Send test notification**; **Fix all** (and the F1 palette) fixes this device too.

### Changed

* **Reorder Overview widgets by dragging** - the widgets on the GitHub and Azure DevOps Overview tabs can be reordered by dragging the grip handle in each widget's header, on the overview itself or in the **Customize** dialog. The up and down buttons in **Customize** remain for touch and keyboard.
* **Factory cards open in a dialog** - on desktop, clicking a card on the Factory board now opens its details (overview, chat, changes, runs, and actions such as Move, Merge and Address review) in a centered dialog over the board instead of a resizable panel on the right. Close it with the X or Escape; Escape inside a text field does not close it, so typed edits are kept. Phones keep the full-screen card view.
* **Factory tabs** - the Factory view (`factory.core`) now has tabs under its header: Board, Team, Channel, Insights and Settings. Team and Channel show only while `company.agents` is on, and the selected tab is remembered. The left rail is the scope for every tab: All projects, then projects grouped by owner, with forge projects under their organization (and a Team link plus a waiting-on-you count when that factory account has a company) and folder projects under "On your nodes"; on a phone the scope is the project picker in the header. Insights and project Settings are full pages instead of side sheets, and with a company in scope Settings has a Project | Company switch that shows the company roster (employees, schedules, budgets, repositories, Open the roster file, Reload roster, Pause/Resume). The Team page now shows a "Needs you" strip, an employee grid with each employee's repositories, and the factory cards in progress as "Factory runs", and a thread opens as a right pane on desktop. The "..." menu keeps only Refresh, New card, Add project and company Pause/Resume/Reload, and the F1 palette gained "Show factory" tab commands and a "company roster" command per company.
* **Easier sideways scrolling on the Factory board** - the board's horizontal scrollbar is now wider with a visible track, and a plain mouse wheel over column headers, the gaps between columns or a column with nothing to scroll moves the board sideways. A column whose cards overflow still scrolls up and down under the wheel, and Shift+wheel works everywhere.

### Fixed

* **No "Started by" badge on your own sessions** - a session or chat you started no longer shows your own name as its owner when the app launched offline or before clusto.app answered. The app now remembers which account it was last signed in to until you sign out.
* **Snap apps and `systemctl --user` work in run-as terminals on Linux** - when the daemon runs as root and starts a session as another user who has no login session or lingering, it now starts that user's systemd manager first. Snap apps such as `btop` no longer fail with "is not a snap cgroup", and the session gets `XDG_RUNTIME_DIR` and `DBUS_SESSION_BUS_ADDRESS`.

## \[1.15.32] - 2026-11-06

### Added

* **Factory projects with a kanban board** - the Factory view now has a project rail and a board per project with the columns Backlog, Ready, In progress, Review, Done and Blocked. Add a project with the **Add project** wizard: a forge repository, a local folder on any node, or a new repository (clone a URL or `git init`). The wizard detects the stack (Node with pnpm, yarn, bun or npm, Rust, Go, Python, Make), suggests setup, check and test commands, and runs a dry run in a fresh clone. Repositories outside your organization work through fork pull requests (GitHub only). Per project you set the runner node, delivery (pull request, fork pull request, local branch that fast-forwards a target branch once the folder has no uncommitted changes, or push branch), autonomy (Manual, Auto, Full auto), WIP limit, weight and pause, plus a rules overlay over the repository's `factory.toml` and `risk.toml` that **Publish rules** writes back as a pull request or a branch. Folder projects run as a user you choose, never root, in their own clone; running as your own user needs an explicit acknowledgement. Full auto approves plans up to T1 and merges T0 by default; T2, T3 and production deploys never run on their own. Agent slots are shared across projects by weight (2 on Personal), with card dependencies. A plan can split a card into up to 12 subtask cards, and `factory.toml` has a new optional `test` key that runs after `check`. Issue sync imports labelled issues as cards, comments on and closes the issue on merge (Azure DevOps: comment only) and moves a card to Done when its issue is closed. From the terminal, `clusto factory projects`, `board`, `card add` and `card move`; agents get `clusto_factory_projects` and `clusto_factory_file_card`, which files a Backlog card labelled "proposed" for you to accept. Behind the `factory.core` flag, projects and the board free on every plan.
* **Node Doctor in Settings** - **Settings > Node Doctor** (System group) lists every cluster node with its AI setup status (All Good, N Issues, Off, Offline or Error), expands to the per-node check list, and has **Fix** per node plus **Fix all** and **Check again**. An unreachable peer shows its error in its row. F1 has **Open Node doctor** and **Fix agent setup on all nodes**. Session (agent) credentials can only read their own node's setup.
* **Pick several days and give the AI your own rules in Suggest** - the day picker in **Suggest** now has a toggle for each day of the shown week plus **Weekdays**, **Whole week** and **Clear** shortcuts, so one search covers the days you pick and leaves out the days in between. With **Name the entries with AI** on, a new **Instructions for the AI** box (up to 4000 characters, saved and synced across your devices) takes your own naming and filtering rules, for example "Track only work for Acme. Name entries B#(work item id) (work item title). Skip personal projects." The AI also sees your open Azure DevOps work items and running-sprint board items from every Azure DevOps account the node holds, pull request titles that match a block's git branch, and your most-used entry names from the last 28 days, so it can reuse your exact naming. Blocks your instructions say not to track appear unticked at the bottom under **Left out by your instructions**, and chips per folder with total time let you hide a folder's suggestions, remembered across devices. AI descriptions can now be up to 160 characters. Behind the `time.tracking` flag, free on every plan.
* **Timer controls for agents** - every agent Clusto sets up now has `clusto_timer_status`, `clusto_timer_start` and `clusto_timer_stop` through `clusto mcp`, so you can ask Claude Code, Codex or Gemini CLI to start a timer on a task, say what is running, or stop it. The timer shows in the Time view like one started in the app. Editing or deleting entries stays in the app.
* **Auto-continue for Claude Code in terminals** - when Claude Code in a Clusto terminal session stops with "API Error: Connection lost mid-response" (or "The response stopped arriving", "Server error mid-response", "Connection lost before a response was produced", "Your computer went to sleep mid-response") or ends on an overloaded API, Clusto types `continue` and Enter for you. It waits 5 s after a dropped connection and 30 s after a server error or overloaded API, cancels if you type in that terminal meanwhile, and stops after 3 automatic continues per session in 10 minutes. Auth, billing, usage-limit and context errors are never continued. Behind the `terminal.auto_continue` flag (cluster scope, off by default; turn it on in **Settings > Labs**). Claude Code's managed hooks gain a `StopFailure` hook for it.
* **Hide and show the keyboard on touch terminals** - the terminal key bar on a touch device has a Hide/Show keyboard button before the chevron. The terminal then uses the full height, and the keyboard stays down through paste, attach, reload and returning to the app until you tap the terminal or press Show.
* **Close menu for Notes tabs** - right-click a Notes tab (long-press on touch) for Close, Close others, Close tabs to the right, Close tabs to the left and Close all. F1 has matching **Close other note tabs**, **Close note tabs to the right** and **Close all note tabs** commands.
* **Install an exact release from the CLI** - `clusto update --version X` and `clusto update-node --version X` install that exact signed release, and `clusto update-node --cluster [--version X | --match-node NAME]` brings every node to one version or to the version of a named node. Each release publishes its own `manifest-<version>.json` and the 10 most recent releases stay installable; a node already newer is left alone, and a retired version reports that only the 10 most recent releases are installable.

### Changed

* **Factory moves cards by drag and drop** - the four stat tiles and the Board, Inbox, Team, Channel, Holdouts and Metrics tab bar are gone. Drag a card from Backlog or Ready to In progress to start it now, back to Backlog or Ready to stop the run, from Review to Done to merge (or mark a plan review or a run waiting on you done), from Review to Ready to send it back with an optional note (the agent continues on the open pull request), and to Blocked with a note; dragging inside a column reorders it except in Review and Done. On a phone a column switcher replaces the side-by-side columns. Metrics became **Insights** and Holdouts moved into the board menu and project settings, and the Team and Channel panels sit under **Company** in the project rail. Retry now starts a new run of the same card and keeps the open pull request. Existing factory repositories become projects in Auto mode, so labelled issues keep starting on their own, and existing items become cards. Behind the `factory.core` flag.
* **Settings regrouped and tidied** - **Settings** is now General (Account, Appearance, Terminal, Chats, Notes, Notifications, This device) and System (Cluster, Node doctor, Updates, Daemon setup, Privacy, Labs). Appearance keeps Theme (CSS variables fold away unless the theme is custom), Right sidebar and Date and time. The new **Terminal** section holds terminal font size, padding and session label, the terminal key bar and the on-screen keyboard; the new **Notes** section holds the notes font size and spelling check. **Client** is now **This device** (old `?tab=client` links still work), and **Copy AI setup instructions** moved to **Node doctor > Other AI tools**. Under **Chats > New chats**, model and effort are one table with a row per provider and one **Refresh models** button. Every row lines its control up in one column, and sliders show the slider, number and reset on one line.
* **Update All restarts the selected nodes together** - **Cluster > Update All** now opens a list of the outdated online nodes with a checkbox each, all ticked; untick nodes you do not want to update now. The run has two phases: every selected node first downloads and stages its daemon and CLI, and only when all of them are ready do they restart together, the connected node last. Before, each node restarted as soon as its own download finished, so the cluster went through several waves of disconnects. A node that fails to stage keeps its version and is reported; the others still restart. A staged update whose restart signal never arrives (the driving node died) is discarded after 15 minutes. Updating a single node with its own **Update** button still restarts it right away.
* **App install asks first when nodes are older** - **Settings > Updates > Install on this device** now asks "Update your nodes first?" when any cluster node still runs an older version than the app update, with an **Install app anyway** button, because a newer app sends requests older nodes do not understand. Update the nodes first and the app last.
* **What's new opens by itself after an update** - the first start after an update now opens a "Welcome to vX" dialog with the release notes straight away, instead of a short toast with a button that was easy to miss while the app connected.
* **Lighter chat resync between nodes** - When nodes reconnect, they now exchange chats without their transcripts and workspace files, then fetch only the content a node is missing. A resync that used to move about 70 MB on a large cluster now moves far less, so chats and workspaces appear sooner after a restart or a network change.

### Fixed

* **Scheduled tasks fire on time** - cron and interval tasks now fire within about a second of their slot instead of up to about 100 s late, and a timed pause ends on time. After a clock jump or a resume from sleep a task is at most 30 s late and runs once, with no backfill.
* **Closing the last Notes tab no longer drops unsaved edits** - closing the last Notes tab with unsaved changes used to discard them silently; it now asks first. Cancelling the prompt no longer leaves a note open without a tab.
* **Clear errors for stale agent session credentials** - a session credential issued by another node, or before a token or node id change, now gets an error telling you to restart the agent session instead of "invalid token", and a credential from another host has its own message. `clusto chat mcp` and `clusto team mcp` no longer fall back to a token on disk when `CLUSTO_TOKEN` is missing; they stop with an error that names it.
* **Claude mid-stream cuts are explained in chats** - a chat that ends with "Connection lost mid-response", "The response stopped arriving" or a similar Claude mid-stream cut is now classified as a network error instead of unknown, so the failure card explains it correctly.
* **Drag to split panes works on Windows** - dragging a terminal, chat or tab onto a pane showed the split zone but did nothing on release. A drag that ends over a pane now lands even when the drop event is missing. Torn-out windows also stop intercepting drags natively, the same as the main window.
* **Lower daemon memory after resyncs** - the daemon caps glibc malloc arenas (Linux) and, when a peer sends its full state on reconnect, rewrites only the chats that changed instead of every chat with its transcript. Memory on a long-running node used to climb with every resync and sit at 2-3 GB.
* **The `?` waiting marker is wired automatically for Claude Code, Codex and opencode** - every node now sets up the waiting-for-input marker as part of its AI setup (on start, after knowledge changes and every 5 minutes), so there is no manual hook setup any more. Claude Code gets Clusto's own hooks in `~/.claude/settings.json` (permission and question prompts set waiting, `Stop` and `UserPromptSubmit` clear it; your own hooks are never touched, and `idle_prompt` stays unwired). opencode gets a Clusto-owned plugin at `~/.config/opencode/plugins/clusto.js` that reports permission and question requests. Codex reports waiting itself. The copyable prompt for Claude Code's questions is now the question text. Turning Knowledge off on a node removes the hooks and the plugin. The AI setup page has new checks: Claude Code, Codex and opencode waiting status.
* **opencode gets the Clusto MCP server without editing its config** - the same plugin registers the `clusto` MCP server through opencode's `config` hook, so `opencode.jsonc` is never rewritten, and a `clusto` entry you wrote wins. The AI setup page shows it as the **opencode MCP server** check.
* **New terminals open at the right size** - a new session used to start a few rows too tall and columns too wide, then visibly snap to the pane size as soon as it appeared. The app now measures the terminal grid the same way xterm does, so the shell starts at its final size, and a brand-new terminal stays hidden until it fits its pane before fading in.
* **Chats moved into a workspace show up there on every device** - chats dragged into a workspace before 1.15.31 could stay in the loose list on other nodes, so the workspace looked missing when the app connected through them. Nodes now settle on the workspace by themselves after the update, copying the chat's files into its folder, with no need to drag the chats again.
* **The cluster settles faster after a node restarts** - after a node restarts (for example during an update), the cluster reconnects with far fewer follow-up disconnects, so typing into a terminal on another node no longer lags right after an update. Three causes are fixed. When a peer link switched from the direct connection to Clusto Tunnel, the old link stopped reading at once, so replies and state updates still in flight were lost, requests timed out after 30 s and two timeouts disconnected the peer again (another full resync); the old link now drains for 10 s before closing. A restarted node can never answer requests sent to its previous process, yet the sender waited the full timeout and counted it against the new, healthy process; nodes now announce a per-process instance id in the peer handshake and a reconnect from a new instance fails those requests immediately. A restarting node also dialed its peers before loading its own state from the database, so its first anti-entropy snapshot was empty; it now dials only after seeding. The catch-up syncs sent on join (OpenRouter, GitHub, Azure DevOps, GitLab, cluster feature settings, node commands) now go only to the peer that joined instead of every peer, so a cluster restart no longer multiplies them.

## \[1.15.31] - 2026-11-05

### Added

* **Standby nodes for GitHub, Azure DevOps and GitLab accounts** - an account can now list standby nodes in Settings. While its host node is offline (a laptop that sleeps, say), the first online standby polls instead, so the view stays live and actions keep working. A token or PAT you save once is copied to the host and its standbys, and only to them; an az CLI login stays on the node that ran it. Each account row in Settings shows which node is polling right now, warns when it failed over or when no host is online, and the view says when a standby is polling.
* **What's new after an update** - the first start after an update shows an "Updated to vX" notice with a **What's new** button that opens the release notes of every version this device has not run yet, with a link to the full changelog. A fresh install skips it. Reopen it any time from F1 (**What's new in Clusto**) or by clicking the version at the bottom of the sidebar. The changelog now has a section per release, written automatically on every version bump.
* **Pan and zoom in the Mermaid diagram dialog** - a diagram now opens at natural size fitted to the dialog instead of squeezed to its width. Drag to move it, scroll or pinch to zoom around the pointer, double-click to zoom in, or use the toolbar to zoom, show actual size and fit; with the diagram focused, `+`, `-`, `0`, `1` and the arrow keys do the same. The dialog also grows to nearly the full window.
* **Usage strip in chats** - a one-line strip above the message box of every chat and of the Assistant shows what the provider reports: Claude Code's 5-hour and weekly plan limits, context fill and estimated cost; Codex's plan windows, credit balance and context fill; the cost OpenRouter charged and the API key's credits left; OpenCode's estimated cost. A figure turns amber at 75% and red at 90%, and a new chat on the same account shows the latest limits before its first turn. Turn it or its fields on or off in **Settings > Chats > Conversations > Usage strip**, the chat's eye menu or F1.
* **Download folders and selections as one zip** - Download on a folder, or bulk Download of several entries, in the Files view and the project file explorer now produces one zip that the node holding the files builds, instead of one save dialog per file with folders skipped. A single file still downloads as itself. The node refuses a build that would leave less than 1 GiB free and deletes the archive once it is downloaded.
* **Drag and drop panes and tabs** - drag a session from the Terminals or SSH list, or a chat from the Chats list, onto a pane: the middle shows it there and an edge splits the pane. Dragging one that is already in a pane moves it, so you can rearrange a layout. File and diff tabs can be reordered, dragged onto another pane's tab strip, or moved into a pane of their own by dragging them to an edge, with the split button at the end of the tab strip or with `Ctrl+Alt+Shift+Arrow`. Dragging needs a mouse; touch screens keep the buttons.
* **Files and diffs open as tabs in Chats** - a file opened from the right sidebar in Chats opens as an editor tab on the chat's pane, where you can edit and save it, for project files and for files in any other chat folder. A changed file in the Git tab opens as a diff tab instead of a dialog, and the Git tab is now on by default for a chat in a project root. A chat without its own workspace shows the project or linked folder it was started on in the Files tab. PDFs still open in the preview card.
* **New chat in the same workspace** - in Chats, `Ctrl+T` opens New chat with the focused chat's workspace, node and project already picked, and `Ctrl+Shift+T` starts a quick chat there without the dialog. Splitting a chat pane and the empty-pane picker preselect the same workspace, and while two or more chats are in panes, each gets a blue dot in the Chats sidebar.
* **Workspace instructions** - the chat workspace page has an **Instructions** section above Default context that edits `.clusto/instructions.md` in the workspace folder. Every chat in the workspace loads it on its next turn, and an agent asked to set up context for the workspace writes to the same file. Only workspaces with their own folder have instructions, and the workspace needs a chat before you can edit them.
* **Chats waiting for your answer** - a chat whose turn is paused on a tool permission or a question shows an amber dot in the Chats sidebar instead of the working spinner, and its workspace header counts the chats waiting on you.
* **Suggest time entries from activity** - **Suggest** in the Time view header (or F1 **Suggest time entries from activity**) finds work you did not time in Claude Code transcripts, Codex sessions and your git commits on every online node, and proposes entries for a day or the whole week that never overlap entries you already booked. The connected node's Claude CLI can name them. You review, edit and pick the entries to add; nothing is saved before that. Needs the `time.tracking` Labs flag.
* **Earlier entries as you type a time description** - the timer bar and the time entry dialog suggest earlier entries from the last 90 days that match the description or project name, and picking one fills in its project, tags and links too. Needs the `time.tracking` Labs flag.
* **The factory resolves merge conflicts after you approve** - **Approve and merge** now marks the item **Approved** and the factory lands it: when the pull request conflicts with the default branch, the agent merges the latest default branch in and resolves the conflicts (up to 2 attempts), and the change merges once the check, AI review and CI are green again, without asking you again. New review comments withdraw the approval, and an `auto_merge_t0` item takes the same path. Needs the `factory.core` Labs flag.

### Changed

* **Azure DevOps read state syncs across devices** - pull requests and work items you read (or mark unread) on one device now show the same on every device, instead of each device keeping its own read marks. The marks replicate through the cluster, and what a device had read locally before is carried over the first time it connects.
* **Chats open at once** - each device keeps the chats you open in a local cache (about 48 MB, oldest dropped first), so a chat you have opened before appears immediately, even after restarting the app, and the node sends only the messages written since. The node also sends stored history before it starts the agent. Deleting a chat or a connection profile removes its cached copy.
* **Lighter chat history** - an image a tool returned (such as a screenshot the agent read) is replaced by an `[Image]` placeholder when a chat's history is replayed, instead of sending the full base64 data to the app and printing it in the tool row.
* **Progress messages between tool calls show by default** - the short status text an agent writes between tool calls is visible again on long turns; thinking and tool calls stay folded, and a display choice you already saved still applies.
* **Chat requests float over the conversation** - permission and question requests sit in an overlay just above the message box instead of a second scrolling area, so they stay in view wherever you scroll. The chevron minimizes the overlay to one line, and a new request opens it again. Your own messages now render as Markdown too, unless **Markdown source** is on.
* **New workspace picks a node and a project** - the New workspace dialog in Chats asks for the node new chats in the workspace run on, and an optional project whose folder becomes default context of every chat in it. Named workspaces replicate across the cluster, so every device lists them before they hold a chat.
* **Agent chats from GitHub and Azure DevOps can go into a workspace** - the dialog that starts an agent chat from an issue, pull request or work item has a **Chat workspace** picker. The chat is listed in that workspace while the agent still works in the project's checkout, and a workspace started from the same project is preselected.
* **`Ctrl+Shift+Arrow` opens a pane at the edge** - with no pane in that direction, the shortcut splits the focused pane toward the arrow and starts a quick session in Terminals, or a quick chat in the same workspace in Chats. In an empty chat composer it now works as a pane key.
* **Company moved into the Factory view** - the Company view is now the **Team** and **Channel** tabs of the Factory view on a GitHub factory account, and its Inbox is merged into the factory Inbox: employee questions and proposed tasks come first, then plans, merges, stopped items and audits. The old `/company` address opens the Team tab, and F1 **Go to Factory** replaces Go to Company. Needs the `company.agents` Labs flag.

### Fixed

* **Running timer pill spacing** - the sidebar timer pill now fills the row with even padding, a compact stop button and the elapsed time right-aligned, and its labels are translated. Needs the `time.tracking` Labs flag.
* **Reports no longer reopen the deep-linked report or jump on expand** - a `/reports/<id>` link now opens and marks the report once, so returning to Reports or a feed update no longer scrolls back to it. Marking a report read no longer shifts its content, collapsing a report brings its header back into view, and a report you open stays in the Unread feed until the next load.
* **Factory tier badge spacing** - the tier badge shows "T3 Designed" again instead of running the code and name together.
* **Moving a chat into or out of a workspace reaches every node** - a chat dragged from the loose chat list into a workspace folder, or back out, stayed in its old place on every node but the one that ran the move. Replicas now take the new workspace kind; a chat that already diverged is fixed by moving it again after the update.
* **Lower daemon CPU on projects with nested dependency folders** - project and chat file watchers walked and watched nested `node_modules` and other skipped folders below the top level, and the chat watcher rebuilt itself after every batch of changes, which spiked CPU on large repositories. Those folders are now skipped at any depth, and new folders gain watches as they appear.
* **Azure DevOps sprint boards are complete** - the sprint taskboard read at most 200 work items per organization. It now reads up to 1000 in batches and says when a sprint has more.
* **Azure DevOps Markdown descriptions keep their formatting** - a description or comment written with the Markdown editor lost its line breaks in the work item pane and in the Implement with agent prompt. Markdown now passes through as written, and HTML tables become Markdown tables.
* **Notes settings no longer time out when two views load them** - a second request a few seconds after the first dropped the node's answer, so both showed "Timed out loading notes settings".
* **`clusto new` and `clusto open` keep a Windows path** - run on Linux or macOS against a Windows node, the CLI joined `C:\Users\...` onto its own working directory. An absolute path of any OS is now passed as given.
* **Codex context fill counts the latest request** - the daemon measured a Codex conversation's context from the thread's running token total, which grows with every request, so it overstated how full the window was.
* **Quieter daemon log** - broadcasts to a peer link that just closed, iroh's network monitor, a memory alarm on a daemon that stays past the threshold (now every 30 minutes unless memory keeps growing), a failed move of the old notes vault and a plugin install that keeps failing the same way no longer repeat a warning every minute or retry. A failing Agent Docs check is logged again only when its detail changes or after 6 hours, and S3 notes sync errors show their full cause.
* **Reconnecting no longer types stale replies into the shell** - output replayed after a reconnect had its old terminal queries (cursor position, device attributes, colors) answered again, which could type text like `^[[12;1R` into the prompt. The node now strips those queries from the replay, and replayed output can no longer overwrite your clipboard.
* **Programs in an unfocused split pane get their terminal replies** - a program in a terminal you own but whose pane was not focused waited for cursor and device replies that never came, until it timed out. Behind the `terminal.owner_replies` flag.
* **Reopened terminals keep accents, emoji, links and underline styles** - the restored screen of a terminal dropped combining marks and joined emoji, OSC 8 hyperlinks, curly, dotted and colored underlines and the program's cursor shape.
* **Cached terminal screens expire** - screens a device keeps for instant reopen are removed when the session ends and after 24 hours, and saving one no longer reads the whole cache.
* **Steadier daemon memory** - a stalled chat capture could make a session's output buffer grow without limit, and every session kept up to about 16 MB of already saved output in memory. The buffer is now capped (the reader logs the gap and continues) and the in-memory copy stays under about 2 MB.

## \[1.15.30] - 2026-11-03

### Fixed

* **Pasted chat attachments open again** - files pasted into a chat are stored relative to the chat workspace, and the message bubble tried to load them by that relative path, so thumbnails and previews failed. The app now resolves them against the chat's workspace folder on its node.

## \[1.15.29] - 2026-11-02

### Added

* **Company: persistent AI employees** - with the `company.agents` Labs flag (needs `factory.core`), a factory GitHub account gets a team of AI employees read from a roster in the organization's company repository (`.clusto/company/company.toml`, repository `hq` unless set in **Settings > GitHub**). Employees wake on a schedule, message each other and you, ask questions and file tasks that the software factory turns into pull requests, each with its own autonomy level, daily budget and turn limit. A new **Company** view has Team, Channel and Inbox tabs, and you can message any employee or factory task worker directly.
* **Date and time format** - **Settings > Appearance > Date and time** picks the date format (System, DD.MM.YYYY, DD/MM/YYYY, MM/DD/YYYY or ISO) and 24-hour or 12-hour time for every date the app shows: time tracking, task runs and pauses, reports, file times, sync times and node build dates. Date fields in the time entry dialog and the Pause task **Resume at** field are typed in the chosen format. The setting syncs with your other settings, and F1 has **Change date and time format**.
* **Files and diffs open as tabs on the terminal's pane** - in Terminals, a file opened from the right sidebar opens as an editor tab on the focused terminal's pane, and a changed file in the Git tab opens as a diff tab there, instead of replacing the pane. The terminal keeps running under the tabs. Tabs belong to the session, follow it into other layouts, survive an app restart and close with the session; the diff header has **Open in editor**.
* **Move chats between workspaces** - drag a chat in the Chats sidebar onto a workspace folder, or use **Move to workspace** in its context menu, the move button on phones or the F1 **Move chat** command. Only the chat moves: its agent restarts in the new folder with your next message, primed with the chat history. Notes vault chats, the Assistant and factory chats cannot move.
* **New empty chat workspace** - a **New workspace** button in the Chats header and an F1 command create an empty named workspace folder on the connected node, ready for new or moved chats.
* **Chat read state syncs across devices** - reading a chat on one device marks it read on every device in the cluster, and F1 **Mark all chats read** clears every unread answer at once. Existing chats start as read after the upgrade.
* **Attachment previews in chat bubbles** - images sent with a message show as inline thumbnails in the bubble, and clicking any attachment (image, PDF, text or code) opens a preview dialog with a **Download** button on desktop and mobile. File previews now also cover PDF.
* **Two-way Toggl sync** - the new Toggl Track plugin pushes finished Clusto time entries to Toggl and pulls Toggl changes (new entries, edits and deletions) back on every interval run and **Sync now**, with the newer side winning a conflict. Time connectors can declare `pull = true` to become two-way. Needs the `plugins.user` and `time.tracking` Labs flags.
* **Delete a time entry from its dialog** - the Edit time entry dialog has a **Delete** button with a confirmation that names the entry and its duration.

### Fixed

* **Run-as commands with captured output work on Windows again** - since 1.15.18 the run-as relay treated piped handles as a console and failed, which broke the Azure DevOps az CLI probe, SDK chats and one-shot commands for run-as users. The relay now uses the pseudoconsole only when stdin and stdout are both consoles.
* **The terminal follows a restore and maximize at once** - when a window restored and re-maximized a few times in a row, the terminal stayed at the restored width for half a minute. The resize hold now ends as soon as the window settles at a larger size.
* **The iOS app icon is accepted by App Store Connect** - the iOS icons are flattened to RGB, because App Store Connect rejected the alpha channel in the Clusto icon.

## \[1.15.27] - 2026-11-01

### Added

* **Split a pane with a new session in one shortcut** - `Ctrl+Shift+D` (`Cmd+Shift+D` on macOS) splits the focused pane along its longer side and starts a quick session in the new pane on the same node and project with the last-used shell; an empty focused pane is filled instead. In the Chats view the same chord starts a quick chat. The shortcut is a rebindable palette command.
* **Tear out layouts and nodes into their own windows** - on desktop, move a Terminals pane layout (**Move layout to new window**, `Ctrl+Shift+O` or F1) or one node's Terminals tree and workspace into a separate window, and move it back later. Torn-out windows still open at exit come back with their position and size on the next start, and a session shows in one window at a time. Behind the `workspace.tear_out` release flag (off by default).
* **Waiting-for-input badges** - a session whose agent is waiting on you shows an amber question marker on its row, a count on its project folder (also while collapsed) and a count on the Terminals and SSH nav items, across every online node. F1 **Go to next waiting session** cycles through them. Codex is covered too: 0.166 and later report it natively, and the daemon detects the "Action Required" title of Codex 0.156 to 0.160.
* **Import projects from Claude Code and Codex** - each node scans the folders its session user's Claude Code and Codex already know (`~/.claude.json`, `~/.claude/projects`, Codex `config.toml` projects and recent sessions) and turns the ones you pick into projects, skipping missing folders and existing projects. Open it from the empty Projects state, the New Project dialog, Settings > Daemon Setup or F1, or run `clusto project import`. Admin only.
* **Support tickets** - bug reports, ideas and questions sent from the app (Contact support, F1, Settings > Account or the Assistant) get a ticket code (`CL-...`). Confirmation and reply emails arrive in Czech or English, and an opt-in email tells you when the release with the fix is published.
* **Czech app interface** - the app follows the system language: Czech on a Czech system, English everywhere else. Czech translations are added screen by screen, so an untranslated screen still shows English.

### Changed

* **Open session as user preselects the node's default user** - the dialog picks the node's `default_session_user` once the user list loads, unless you already chose someone.
* **Assistant action cards leave the tray once you move on** - an applied or declined action card is no longer shown in the Assistant's action tray after you send your next message.
* **No purchase links in the iOS app** - the iOS app hides the pricing link in Account, the **See plans** button in Notes and the paid-plan wording on the license gate, as App Review requires. Free account creation links stay.

### Fixed

* **The daemon starts on Windows while `clusto mcp` is running** - updating the CLI next to `clustod` renamed the new binary onto `clusto.exe`, which Windows refuses while any process runs it, so the service exited after every update. The running file is now moved aside first, and a failed CLI install at startup is only a warning.
* **The Linux AppImage no longer turns grey on Arch and similar distributions** - the AppImage now bundles its GStreamer plugins, so the first notification sound no longer crashes the web view on hosts without Debian's plugin path.
* **No gap at the bottom of the iOS app** - the web view no longer subtracts the status bar and home indicator a second time, which left the layout about 82 points short.
* **The iOS app shows the Clusto icon** - builds carried Tauri's template icon instead of Clusto's.

## \[1.15.26] - 2026-10-31

### Added

* **Live Git tab and auto-open diffs of changed files** - the right sidebar's Git tab now shares the daemon's project file watcher, so an edit, a build or an agent write refreshes the change list and any open diff on its own, for any agent CLI and for projects on other nodes. With **Open diffs of changed files** (Settings > Appearance > Right sidebar, on by default, desktop only, also in F1) the sidebar opens the Git tab on the diff of the file that just changed.

### Changed

* **Untracked files show their contents as a diff** - the Git tab shows a new file as a whole-file added diff instead of an empty "No diff".

### Fixed

* **An open diff no longer goes blank after a status refresh** - the diff you have open is fetched again, and file watches are restored after a reconnect.

## \[1.15.24] - 2026-10-29

### Added

* **Plugins can use your Azure DevOps accounts** - a plugin that declares `accounts = ["azure_devops"]` in its manifest makes its Azure DevOps REST calls through the accounts in **Settings > Azure DevOps** (PAT or az CLI login), via a token-scoped plugin API on its node, so it never sees or stores a credential. The install dialog lists the accounts a plugin declares, and calls that change something are recorded in the account's audit log with the plugin id. User plugins remain behind the `plugins.user` Labs flag.

### Fixed

* **A node already updating shows as updating on every client** - the cluster status now reports a self-update in progress, both for a node updating itself and for a peer updated through another node, so node cards on other devices or after a reload show it as installing instead of offering **Update** again. A second update request for a node that is already updating is refused instead of starting a parallel one.

## \[1.15.22] - 2026-10-26

### Fixed

* **The desktop app no longer hangs on a blank window at startup** - in 1.15.21 the Explorer menu queue watcher and its store kept triggering each other, so the window stayed blank.

## \[1.15.21] - 2026-10-25

### Added

* **Windows Explorer menu and local terminals** - right-click a folder, a drive or empty space in a folder and choose **Clusto > Open terminal here** or **Add as Clusto project**; the installer adds the menu for the current user and **Settings > Client > Explorer menu** turns it off or on. When this PC has no node in the connected cluster, a dialog offers to set it up (installer plus pairing code) or add its node to the cluster. With the new `terminal.local` release flag (app scope, off by default) it can also open a local terminal, a shell the desktop app runs itself with no node, listed in a new **Local terminals** view. See [Explorer menu and local terminals](/explorer-menu).
* **Time calendar** - the Time view has a **Calendar** tab that draws the week as a grid of entries in their project colors, with a line for the current time. Drag on empty space to book a range, click to book your quick log length, drag a block to move it, or drag its bottom edge to resize; drags snap to 15 minutes. On a phone it shows one day (three on a wider screen) and you swipe through the week. Part of the `time.tracking` Labs flag.
* **Settings view for plugin time connectors** - a time connector can name one of its plugin's views (`view = "..."`) that holds its settings, rendered under the connector in the Time view's **Sync** tab instead of in the sidebar. Plugin forms gain a masked `password` field kind for secrets, and plugins receive `CLUSTO_VERSION`.

### Changed

* **Ctrl+C asks twice when the line has unsent input** - without a selection, the first `Ctrl+C` on a terminal line holding text you typed or pasted but have not submitted shows "Press Ctrl+C again to interrupt", and a second press within 2 seconds sends it. This keeps a slip next to `Ctrl+V` from wiping a message in Claude Code. On an empty line one press interrupts at once.

### Fixed

* **The close-session warning no longer lists an agent's MCP servers on Windows** - npm shims an agent starts through `cmd.exe /d /s /c` (such as `npx @playwright/mcp`) are no longer reported as running commands.

## \[1.15.19] - 2026-10-24

### Added

* **Time tracking** - a new **Time** view with a timer bar (description, a project picker that creates projects as you type, tags, live elapsed time), entries grouped by day with Continue, Edit and Delete, and **Add entry** for past time. A **Week** dashboard has stat tiles, hours per project and a Mon-Sun timesheet, a **Projects** tab sets color, client and archive, and a running-timer pill sits in the sidebar footer or phone header. Entries replicate to every node. Free on every plan, behind the `time.tracking` release flag (Settings > Labs, off by default).
* **Time links, quick log and Toggl import** - an entry can link Azure DevOps work items and plugin connector tasks through the **Link** picker or `B#123` / `TR#456` in the description, and an open Azure DevOps work item gets **Start timer** (also in F1). **Quick log** books a short entry that ends now, also from `clusto://time/quick-log` and `clusto://time/start` links. The **Sync** tab imports from Toggl without duplicates and runs plugin time connectors (`[[time_connectors]]` in `clusto-plugin.toml`).
* **Warning before closing a busy session** - closing a session from the Terminals sidebar, the mobile switcher, `Ctrl+Shift+W`, F1 **Close session** or the project page now asks its node what still runs inside it. If the agent is working or waiting, a program reports progress, or shell jobs or agent-started commands are running, a dialog lists them with **Close anyway** and **Cancel**. An idle session closes at once.
* **Factory item tabs** - the item pane in the Factory view has **Chat** (read and reply to the item's agent in place), **Changes** (the pull request's description, checks and diff) and **Details** tabs. The tab you pick stays picked as you move between items.
* **TLS setup progress on node cards** - turning on browser TLS for a node shows provisioning and restarting progress on its card, and the card's actions are disabled while an update or TLS setup runs.

### Changed

* **Removing a node is permanent until you adopt it again** - every daemon records the removal in `[[cluster.removed_nodes]]` and passes it on in the cluster config sync. A removed node stays out across restarts and fleet updates, and a peer that was offline during the removal forgets it on its next connect. Adopting the node again lifts the removal on every node.

### Fixed

* **Scrolling Claude Code in a Windows run-as terminal no longer types escape codes** - a mouse wheel report split across two console input batches reached the inner console as plain keys (`;15M;34M...`). The relay now holds a cut-off sequence until the rest arrives, or for at most 50 ms, so a lone Escape still goes through.
* **Run-as sessions check the working folder first** - a run-as session whose default folder the target user cannot open (such as the daemon user's private home) used to die with a raw shell error. An implicit default now falls back to the user's home with a warning, and an explicit folder fails with a message naming the user, the folder and the node.
* **`systemctl --user` works in run-as sessions** - run-as shells on Linux now get `XDG_RUNTIME_DIR` and `DBUS_SESSION_BUS_ADDRESS` for the target user when that user's runtime dir and session bus exist.
* **A blank name clears a node's display name** - renaming a node to an empty name now removes the name everywhere instead of storing an empty string, and `clusto rename-node <id> ""` clears it from the CLI.
* **Chat composer focus follows its pane** - moving pane focus with `Ctrl+Shift+Arrow` no longer leaves typing and Enter going to the previously focused chat.
* **Factory chats stay out of every node's Chats list** - a chat's owner and factory link now replicate to nodes that already held the chat, so factory chats no longer show up in other nodes' Chats lists.

## \[1.15.18] - 2026-10-23

### Added

* **Software factory verification: risk tiers, reviewer and holdouts** - every factory change gets a risk tier, T0 to T3, from the repository's `.clusto/factory/risk.toml`, built-in sensitive paths and diff signals; the agent and reviewer may raise a tier but never lower it. A clean-context reviewer agent reads the task and diff before the pull request opens and writes an explain-back card into its description, owner-written holdout scenarios run after green CI in a throwaway copy of the factory machine, and an optional plan step makes T2/T3 work wait for your approval. T0 auto-merge is opt-in (`auto_merge_t0`). Behind the `factory.core` release flag.
* **Software factory on Azure DevOps and GitLab, with recipes** - the factory now works through GitHub, Azure DevOps or GitLab accounts bound to their organization or group, opening pull or merge requests, following their CI and taking review threads as feedback. `[recipe.<name>]` profiles in `factory.toml` override plan, review, models, budgets and size limits and add instructions to every brief; an item names its recipe, or intake picks one from the issue's labels.
* **Factory Inbox, Holdouts and Metrics tabs** - the Factory view gains an **Inbox** (plans to approve, green pull requests to merge from your phone with their explain-back card, stopped items to retry or abandon, pending audits), a **Holdouts** editor per repository, and **Metrics** (merged by tier, open pull requests, audit finding rate, CI fixes per merge, rework, spend per repository and week). Every action has an F1 command.
* **Factory chats belong to their item** - a chat the factory creates links to its item, stays out of the Chats list unless this device turns on **Show factory chats**, and shows a banner with **Open in Factory** and **Take over** / **Hand back**. While the factory drives the item the chat's composer is locked; Take over hands the item to you, and Retry hands it back.
* **GitLab accounts** - Settings > GitLab adds gitlab.com or self-managed accounts, whose host node polls a group's projects, open merge requests and pipelines for the factory. The token stays on the host node and is pinned to its base URL. Behind the `gitlab.accounts` release flag.
* **Prefilled token links** - **Create token** in the GitHub settings opens the fine-grained token page with the name, a one-year expiry and the permissions the GitHub view and the factory need already filled in. The Azure DevOps settings link opens the PAT page of your organization, and the Azure DevOps settings card gains a **Purpose** select so an account can be marked for the factory.
* **Remembered view state** - the sidebar, F1 and mobile nav reopen the page you last visited in each section, the Settings tab is kept in the URL, each chat keeps its unsent composer draft, and filters, search and sort persist on the device in the Sessions, Locations, Tasks, Logs, Teams, Agent Docs, Factory, GitHub and Azure DevOps views.

### Fixed

* **Run-as programs on Windows ignored every key** - a program started for another Windows user (such as Claude Code) read keys from its own console instead of the terminal, so it rendered and then ignored all input, Ctrl+C included. The run-as relay now hosts a pseudoconsole in the user's session and passes input, output and resizes through it.
* **Settings > Labs saves failed with "unknown feature flag: company.view"** - a stored override for a retired flag no longer blocks Labs. Only overrides the catalog still knows are shown, and the stale entry drops on the next save.
* **Factory items on Claude stayed in their first state** - a Claude chat keeps one process for the whole conversation, so the factory never saw a turn end. A Claude result line now ends the turn.
* **No doubled space before a pasted image chip in Claude Code** - dropping or pasting files into a terminal running Claude Code sends image paths last with nothing after them, so the `[Image #N]` chip no longer gets an extra space in front.

## \[1.15.17] - 2026-10-22

### Added

* **Overview dashboards in the GitHub and Azure DevOps views** - both views now open on an **Overview** tab of widgets (review requested or needs my vote, failing workflows or pipelines, my pull requests, my issues or work items, running builds, repositories), each showing its first 5 items with a **View all** button that opens the matching tab already filtered. **Customize** picks, reorders and sizes the widgets, and the layout syncs with your account settings. Both views stay behind their `github.view` and `azure_devops.view` release flags.
* **Azure DevOps sprint taskboard** - the Azure DevOps Overview has a taskboard for the running sprint, with one lane per assignee (yours first and open, the others collapsed to a count and remaining hours) and one column per state in workflow order. Cards show priority, parent link, remaining work, tags and the work item type color, with Sprint and Work item type filters.
* **Azure DevOps read state and grouping** - the Pull requests and Work items tabs show their share of the nav badge as an "N new" badge, an **Unread** filter shows only unread items, and a **Group** picker splits the list by state, by type, or not at all.
* **Chat pane picker and pane shortcuts** - an empty chat pane now offers a picker as in Terminals: chats already in another pane are listed with their pane number, the rest most recent first, then **New chat** and **New design session**. `Ctrl+Tab` / `Ctrl+Shift+Tab` and `Ctrl+1` to `Ctrl+9` focus panes, `Ctrl+W` closes the focused pane (the chat keeps running), and `Ctrl+T` opens a new chat in the focused pane.
* **Software factory machine and organization** - a node's `[sandboxes]` config can map an OS user to an Incus system container, so the factory's steps, chats, git and terminal sessions for that user run inside it. A factory GitHub account now requires an **Organization**, and the start dialog lists only that organization's repositories. Behind the `factory.core` release flag.

### Changed

* **Company view removed** - the GitHub-label agent Company view is gone in favor of the software factory, together with its settings, notifications, `[companies]` config section, the `company.view` flag and the Assistant's company status tool. Stored Company alerts are dropped from the notification log.
* **Notes vaults are a stricter boundary** - a node that keeps a vault now refuses everything forwarded from a node the vault leaves out, not only Notes: terminals, chat turns, tasks, the file explorer, git and plugins, while plain status reads still get through. A left-out node gives its sessions and its `clusto` MCP server a credential limited to that node, and chats of a node that keeps the vault no longer replicate to a left-out node. Behind the `notes.multi_vault` release flag.

## \[1.15.16] - 2026-10-22

### Fixed

* **A flaky connection no longer blanks the app** - after the first successful connect, an automatic reconnect keeps sessions, chats and the current view on screen instead of covering them with the connecting overlay. The open chat reloads after a drop so entries written during the gap fill in, and search text and filters survive the reconnect.
* **Slow mobile links are no longer dropped by the resume probe** - the connection probe after resume or a network change now waits four measured round trips, between 5 and 10 seconds, instead of a fixed 2 seconds, so a link with a 1-2 s round trip is no longer torn down and replayed.
* **Touch selection in terminals copies what is highlighted** - the terminal dropped its selection behind the long-press overlay when the keyboard closed, the buffer switched or input reached the PTY, so the handles stayed up but **Copy** copied nothing. The overlay now re-applies the selection whenever the terminal reports none.

## \[1.15.15] - 2026-10-21

### Added

* **Software factory** - send an open GitHub issue, a red CI run on the default branch or a typed task from the new **Factory** view, or with **Send to factory** on a GitHub issue page. An agent fixes it in its own clone as the unprivileged `clusto-factory` user, the repository's `.clusto/factory/factory.toml` check command verifies it, and the daemon pushes a branch and opens a pull request as a dedicated factory GitHub account (Settings > GitHub, **Purpose: Factory**), looping on CI failures until you merge; the factory never merges. Behind the `factory.core` release flag (Settings > Labs, off by default).
* **Split panes in Chats** - the Chats view splits into panes like Terminals so you can read several chats side by side: `Ctrl+Alt+Arrow` splits, `Ctrl+Alt+Enter` maximizes, `Ctrl+Alt+W` closes the other panes and `Ctrl+Shift+Arrow` moves focus, with the same Pane actions menu and F1 commands. Closing a pane never stops or deletes its chat, and the layout is remembered per device.
* **Read and unread in Azure DevOps** - pull requests and work items show an unread dot and bold title until you open them, and turn unread again when they change. **Mark all read**, **Mark unread** and matching F1 commands manage it, and the nav badge counts only unread items. Part of the Azure DevOps view behind the `azure_devops.view` release flag.

### Changed

* **Redesigned node cards in Cluster** - every node, the connected one included, uses one card layout: status dot, name, badges (**This node**, **Offline**, **Identity changed**, **Browser TLS**) and a **...** menu with **Node settings**, **Enable browser TLS**, **Restart node** and **Remove node** (**Leave cluster** on the connected node), each destructive action confirmed first. **Connect here** and **Session as user** sit in a bottom action row, and each metric shows its sparkline under the live reading.
* **Per-node update progress** - a node behind the latest manifest shows a **vX.Y.Z available** strip with an **Update** button. Once an update starts, the card reads **Updating to vX.Y.Z** and then **Restarting to vX.Y.Z** until the node reports the new version; a node that has not come back after 10 minutes gets a warning, and **Update All** shows how many nodes are behind.
* **Company chats with any agent role** (the Company view was removed in 1.15.17) - the Company view's **CEO** tab is now **Chats**: **New chat** picks any role from the repository's `roles/` folder and opens a `role:<role>` issue that starts that agent. Items and chats gain **Close** and **Reopen**. The Company view stays behind the `company.view` release flag.
* **Reports read like documents** - report bodies get document spacing, a readable line length and data-style tables, and a leading heading that repeats the report title is dropped because the card header already shows it.

### Fixed

* **Mermaid diagrams lifted from a terminal render with trailing text** - Claude Code does not always leave a blank row after a diagram, so the following text was read as diagram source and failed to parse. The diagram dialog now drops up to 30 trailing lines until the source parses and says how many it left out.
* **Git works on projects owned by another account under a root daemon** - a project with no run-as user now runs git as the non-root owner of its folder, so git no longer refuses the repository for "dubious ownership" or leaves root-owned files in its `.git`.
* **SSH sessions start in the node's browse home** - a saved SSH connection now opens in the node's default browse folder, so the Files sidebar lists a real folder. A Windows service has no `$HOME` and fell back to `.`.

## \[1.15.14] - 2026-10-21

### Added

* **Multiple Notes vaults** - keep several named vaults (for example Work and Personal), each with its own folder, S3 settings and open tabs, picked from a vault switcher in the Notes toolbar. A vault's **Kept on** checklist chooses which nodes hold a copy, and nodes outside that scope never receive its bucket settings, credentials or notes. A second vault needs Pro, Team or a trial. Behind the `notes.multi_vault` release flag (Settings > Labs, off by default); an existing `[notes]` section migrates into a vault named **Notes** with the same folder and bucket.
* **Agents can pick a Notes vault** - the new `clusto_notes_vaults` MCP tool lists the vaults a node keeps, and the notes tools (`clusto_notes_*` and the chat tools `clusto_list_notes`, `clusto_search_notes`, `clusto_read_note`, `clusto_write_note`) accept an optional `vault`. Without it they use the node's first vault.
* **Chats in workspace panes** - a pane in the Terminals split-pane workspace can hold an agent chat next to terminals, SSH sessions and the editor. The empty-pane picker lists recent chats after the sessions and has **New chat** beside **New terminal**, and the F1 palette adds **New chat in focused pane**. Closing the pane keeps the chat.
* **Right sidebar in Chats and SSH** - the Files/Git/Run sidebar now also appears in Chats and for SSH sessions, with its own panel set per view (Terminals: Files, Git, Run; Chats and SSH: Files). Choose panels from the sliders button in the sidebar header or **Settings > Appearance > Right sidebar**, and toggle it with **Toggle right sidebar** (Ctrl/Cmd+Shift+E).
* **Azure DevOps sprints** - the Work items tab gets a **Sprint** picker (every sprint, current sprint, or one sprint), a sprint badge per item and a summary line per running sprint with days left and "N of M done". **Whole sprint** lists every item of the running sprints grouped by state with assignees, and the F1 palette has matching commands.
* **Quick new chat** - an F1 command that starts a chat at once with the remembered provider and model, your **Settings > Chats** defaults and **No workspace**, and opens it with the composer focused.
* **Ctrl+Enter in create dialogs** - `Ctrl+Enter` (`Cmd+Enter` on macOS) creates from any field in the New chat, New session and SSH connection dialogs.
* **Paste an SSH connection string** - the SSH connection editor fills host, username and port from a pasted `user@host:port`, `ssh://user@host:port` or a whole `ssh -p 2222 -i key user@host` command. New connections get a **Save & connect** button.

### Changed

* **Slimmer New chat and New session dialogs** - New chat has a **Chat | Design** switch at the top and Provider, Model and Effort on one row, with Start from project, Workspace, Node and Run as behind a **More options** row that summarizes the current choices. New Session asks only for **Project** and **Shell** up front and moves working directory, run-as user, name and arguments behind **More options**; Name is now optional.
* **Simpler SSH authentication choice** - authentication is a three-way choice: **Default keys**, **Private key** (From vault, Paste key or File on node) or **Password** (Type it or From vault). The editor defaults to **From vault** when the vault already holds a matching credential, and Remote directory and Extra ssh arguments move under **More options**.
* **Azure DevOps hand-off uses your own checkout** - Review, Fix failing build and Implement with agent now start a chat in the project's existing checkout instead of creating a worktree, and Clusto runs no git command itself. The brief tells the agent how to fetch and diff without disturbing your branch and to leave changes uncommitted.
* **Agent dialogs remember the project** - the GitHub and Azure DevOps agent dialogs preselect the project you last started from for each repository (and each Azure DevOps project for work items), synced with your other preferences.
* **Compact chat composer toolbar** - provider, model, effort and permissions sit behind one chip that names the model, and linking and browsing notes share a single Notes menu, so the composer toolbar stays on one row with the send button at its end.
* **Smoother wheel scrolling in terminal apps** - like native terminals, Clusto now sends one wheel report per line scrolled (three per mouse-wheel notch, trackpads by travel), so Claude Code's fullscreen view follows the gesture instead of moving one line per event.
* **Terminal resize keeps text at full size** - while a terminal you own waits on its own resize (a split drag, sidebar toggle or window resize), the grid stays at full size pinned top-left and clipped by the pane instead of zooming out and snapping back. Layout animations are skipped when the OS asks for reduced motion.

### Fixed

* **Images pasted into Claude Code on Windows attach as images** - the saved file path no longer carries the `\\?\` prefix, which made Claude Code paste it as plain text instead of an `[Image #N]` attachment.
* **Mermaid detection stops at a list under the diagram** - a Markdown list drawn right below a Mermaid code block in a terminal is no longer read as part of the diagram source.
* **Worktree fetches can no longer hang on a credential prompt** - the daemon's `git fetch` for a worktree never asks for credentials (including Git Credential Manager windows) and gives up after 90 seconds, and a project folder that is not a git checkout is refused with a message naming the folder.

## \[1.15.13] - 2026-10-20

### Fixed

* **Azure DevOps az CLI sign-in finds the right tenant** - an organization belongs to one Microsoft Entra tenant, so when the login's default tenant reaches no organization, the host now tries every other tenant the login holds and keeps the first one that works. A tenant that demands multi-factor sign-in (`AADSTS50076`) is skipped until you run `az login --tenant <tenant id>` for it, and the app waits up to 2 minutes for the check.
* **New terminals on a phone fit the screen** - a terminal opened right after a dialog closed the keyboard now refits once the keyboard is down instead of keeping its first size estimate until an unrelated resize, and that estimate now subtracts the terminal padding.

## \[1.15.12] - 2026-10-18

### Changed

* **Enter adds a new line on phones and tablets** - on a touch-first device Enter in the chat composer adds a new line and the send button sends. Desktops keep Enter to send and Shift+Enter for a new line, and Ctrl+Enter (Cmd+Enter on a Mac) sends on every device.
* **Smooth mouse-wheel scrolling in terminals** - a notched mouse wheel on desktop now animates terminal scrollback instead of jumping several rows per click. Trackpads and touch scrolling stay 1:1.

### Fixed

* **The chat composer fits a phone** - the agent settings and host-specific chips stay on one row that scrolls sideways instead of wrapping, and the **Browse notes** button has an icon like its neighbors instead of reading as a stray word.

## \[1.15.11] - 2026-10-17

### Added

* **`clusto doctor` checks the cluster** - between the install and AI setup checks it now lists peers whose identity changed (every link to them is refused), stale entries an offline peer leaves when its host was reinstalled or re-adopted under a new node id, and connected peers on another Clusto version. `--fix` asks `[y/N]` at a terminal before each repair: trust the new identity, remove the stale entry, or update every outdated node.

### Changed

* **Cloud sync and team profiles: Pull or Push instead of "Sync now"** - Settings > Account no longer has a two-way **Sync now** button on the personal Cloud sync card or the Team profiles card, since it was unclear which side won. Each card now offers **Pull from cloud** (this device takes the cloud copy as is, dropping unsynced local edits and local-only profiles) and **Push to cloud** (the cloud copy becomes this device's state, and other devices or teammates follow on their next sync, including deletions), both behind a confirmation. Background sync is unchanged (newest edit wins). The F1 palette's "Sync team profiles now" is replaced by matching pull and push commands for cloud sync and each team vault.
* **Pull from cluster asks before replacing this device's settings** - Settings > Cluster > **Pull from cluster** now compares the cluster copy with this device first. If anything differs, a dialog lists how many settings would change (changed, new, removed), names them, and says when the copy was saved, with Cancel as the default; if nothing differs it says so and does nothing. Confirming keeps a backup of this device's settings, and **Undo last pull** in the same card puts them back. Device-specific settings (battery saver, on-screen keyboard mode, panel widths and split ratios, open note tabs, chat read markers) are no longer pushed or pulled by cluster preferences or account cloud sync, so a copy saved on a phone can no longer resize or throttle a desktop.
* **Terminals can advertise 24-bit color** - with the `terminal.truecolor` release flag on (node scope, off by default), terminal sessions start with `COLORTERM=truecolor` (unless the session sets its own), so tools that detect color depth (Claude Code, chalk, bat, delta) draw full color instead of quantizing to the 256-color palette. 1.15.9 and 1.15.10 set it unconditionally; it is behind the flag while a Windows report of agent TUIs freezing is checked.

### Fixed

* **A revoked GitHub token, Azure DevOps PAT or expired az CLI login no longer floods the daemon log** - when GitHub (`401`, or a `403` that is not a rate limit) rejects the token of a GitHub account or a Company, or Azure DevOps (`401` on every organization) rejects an account's credential, or the az CLI on the host has no usable login, the host stops polling it every interval, logs the rejection once instead of a warning per poll, and retries after 2 minutes with the wait doubling up to 30 minutes. The view says the token was rejected and should be updated (for az: sign in again with `az login` as the named OS user on the host node, then refresh); saving a new token or the refresh button retries at once, and the first good poll resumes normal polling.
* **Agent Docs sync no longer repeats the same warning every 5 minutes** - the daemon logs a failing AI setup check once and again only when its detail changes (repeats go to debug), while the app and `clusto doctor` show the status as before. Instructions that render to nothing on a node (for example a Linux-only body on Windows) are skipped as "Agent Docs has no instructions for (tool) on this node; nothing written" instead of reported as an error, and the existing file is left untouched. A write refused because an instruction file is a symbolic link (such as `~/.claude/CLAUDE.md` linked into a dotfiles checkout) now names the link and its target and says to replace the link with a regular file or set Agent Docs sync to Manual on that node; Clusto still never writes through it.

## \[1.15.10] - 2026-10-15

### Added

* **Terminal programs can notify, report progress and report their own status** - the daemon now reads the escape sequences other modern terminals (Ghostty, kitty, iTerm2, Windows Terminal) honor, for every terminal session and whether or not a client is attached. `OSC 9`, `OSC 777;notify` and kitty `OSC 99` raise a Clusto notification with the program's own text (at most one a second per session). `OSC 9;4` progress shows as a thin bar under the session in the sidebar and clears itself after 15 s without an update. Program Status Protocol (`OSC 7501`) reports drive the working spinner, the `?` waiting marker (permission, question or sign-in) and the finished or failed notification, and once a program sends one the title and footer heuristics stand down for that session; the daemon answers the `OSC 7501;?` feature query.
* **Terminal bell** - a bell (BEL) in a terminal you are not looking at marks the session with a bell icon in the sidebar until you focus it, and on desktop flashes the taskbar entry when Clusto is in the background (at most every 5 s, never for a muted session).

### Changed

* **The last Hive names are now Clusto** - the built-in chat MCP server is `clusto-chat-context` and its tools are `clusto_set_chat_title`, `clusto_list_context`, `clusto_read_context_file`, `clusto_list_notes`, `clusto_search_notes`, `clusto_read_note` and `clusto_write_note`. The team MCP server is `clusto-team` (tools `mcp__clusto-team__*`), the bundled team skill is `clusto-team` (the daemon removes the old `hive-team` skill folder), and team worker branches are named `clusto/team-<team>/worker-<n>`. The daemon's `/metrics` names are `clusto_*` (`clusto_daemon_uptime_seconds`, `clusto_sessions_total`, `clusto_cluster_peers_connected`). Stored chat transcripts are migrated to the new markers, and app settings saved under the old `hive` keys move to `clusto` keys automatically on first start. Prompts, permission allowlists or scrapers outside Clusto that name the old servers, tools or metrics need the same rename.

## \[1.15.9] - 2026-10-15

### Added

* **Agent Docs sync across nodes** - user-scope instructions, skills, subagents and commands now follow you to every node with no review step. Each node captures edits, new items and deletions from its tools' folders, writes what the cluster holds for its OS, account and installed tools, and removes items deleted elsewhere. When both sides changed the newer edit wins and the losing file is copied to `~/.clusto/agent-sync/replaced/` first; files whose template uses `clusto:if` blocks or non-path variables are left for review instead of captured. Status shows as **Agent Docs sync** on the Knowledge AI setup card and in `clusto doctor`; Each node has a sync mode, `[knowledge] agent_sync`: two-way (default), receive only, or manual, so a machine with a separate work login can stay apart and still take single items on demand. Automatic sync ships off behind the `agent_docs.sync` release flag (Settings > Labs); the mode can be set before the flag turns on.
* **Binary files in skills** - images, fonts and audio inside a skill sync and back up byte for byte (up to 1 MiB per file, 8 MiB per skill).

### Fixed

* **A restarted node no longer loses Agent Docs edits to a stale peer** - a node now seeds its Agent Docs records and template variables into cluster state at startup, so a peer's older snapshot can no longer overwrite them, and a deleted skill, subagent or command is a replicated tombstone that an offline node cannot bring back.

## \[1.14.50] - 2026-10-07

### Added

* **Chat workspaces are optional** - a Chat now starts with **No workspace** and runs in a shared `clusto-chats/clusto-common` folder on its node, whose files stay on that node while the transcript still replicates. The New chat dialog marks the Workspace picker optional and offers No workspace, Create a new workspace, Pick a folder myself, and Existing workspaces. Workspace, Node, and Run as move behind a collapsed **More options** line with a summary. Design sessions still always get their own folder. **Start from project** now only links the project as context. In the Chats sidebar, workspace groups come first and chats without a workspace are listed flat under the Chats header, newest first.

### Fixed

* **Terminal links wrapped by an agent open in full** - Claude Code and similar TUIs break a long URL at the right edge themselves and draw the rest on the next row (often indented under a bullet), so xterm never saw one wrapped line and clicking opened only the first row's part. The link detector now rejoins such rows while the text so far ends in an http(s) URL that reaches the last column, so every row of the link opens the complete URL.

## \[1.14.48] - 2026-10-06

### Added

* **Feature flags and Settings > Labs** - risky features ship behind release flags resolved per cluster: a node override in `[features]`, a cluster override or early access from Settings > Labs (also in F1), a remote rollout from `flags.json` next to the manifest, or the built-in default. The Company view is the first flag and starts off.

### Changed

* **One release track** - the semi-stable and unstable channels are gone. Every build publishes `manifest.json`, which every node, CLI and app follows; `manifest-unstable.json` stays as a byte-identical alias so nodes still configured for unstable reach this release. On start the daemon drops `[update] channel` from its config, and `unstable` turns on Settings > Labs early access for the cluster when Labs was never set. The Settings channel picker and the promote step are gone, `clusto update --channel` is ignored with a notice for one release, a `manifest_url` that names a `*.json` file is used verbatim, and `/app-unstable` redirects to `/app`. Risky changes now ship behind feature flags instead.

## \[1.14.38] - 2026-10-01

### Added

* **Clusto Assistant** - a built-in AI assistant that sees the view you have open. Open it from the sparkles button or F1 (Ask Clusto Assistant), pick Claude Code, Codex, or OpenRouter, and ask about setup, reports, failed task runs, cluster state, and configuration. It answers from the docs embedded in the installed CLI and read-only cluster tools, proposes changes as action cards you apply or decline, and can send feedback to the Clusto team. New `clusto docs list|search|read` command serves the same documentation to any agent with a shell.

## \[1.14.12] - 2026-08-20

### Fixed

* **Folders created from Files keep the expected Unix owner** - when a system `clustod` creates a folder through the whole-filesystem Files view, the new directory inherits the parent directory's user and group instead of remaining root-owned. A non-root daemon attempts this inheritance only when it owns the parent directory.

## \[1.14.9] - 2026-08-17

### Added

* **Built-in chat MCP tools** - every durable chat receives Clusto's built-in chat instructions and the permanent `hive-chat-context` MCP bridge. Claude, Codex, and OpenCode chats can call `hive_set_chat_title`, inspect cross-node attachments with `hive_list_context` and `hive_read_context_file`, and browse, search, read, create, or safely update Notes with `hive_list_notes`, `hive_search_notes`, `hive_read_note`, and `hive_write_note`. Existing notes require read-then-write with the returned SHA-256 to prevent concurrent overwrites. Native OpenRouter chats expose equivalent daemon-managed functions through the same bridge.

## \[1.14.5] - 2026-08-15

### Changed

* **Legacy downloads are now a one-release bridge** - `dl.clusto.app` is the only advancing release origin. `dl.vazac.dev` freezes its stable and unstable manifests at A1 1.14.4, with artifact URLs pointing to the canonical host, so an old client can install the identity-migration release and then take its next update from Clusto without continuing to ship new releases to the legacy domain.
* **Phase 4 public identity cutover** - build, release, installer, workflow, cloud, and download sources now publish Clusto names and `clusto.app` URLs. `CLUSTO_*` environment variables are canonical, with `HIVE_*` fallbacks retained through Phase 5; legacy artifact and domain aliases remain readable during the overlap.

## \[1.14.1] - 2026-08-14

### Added

* **Direct chat deletion everywhere** - right-click a chat row on desktop, use its touch/mobile delete control or active-chat actions menu, or search F1 for the chat by name. Every entry point opens the same destructive confirmation before deleting the chat and saved transcript.
* **Configurable navigation** - choose Full or Simple usage profiles, where Simple shows only Chats and Notes, reorder or hide individual feature destinations, and switch between Comfortable and Compact spacing. Connection, Settings, Documentation, and the navigation settings remain available as permanent escape hatches on desktop and mobile, with matching F1 commands.
* **Documentation inside the app** - open the public docs from the expanded or collapsed desktop menu, the mobile menu footer, or **F1 > Open Clusto documentation**.
* **Human and AI changelog** - root `CHANGELOG.md` is the single source for the public `/docs/changelog` page, raw Markdown, `llms.txt`, and `llms-full.txt`.
* **Task targeting and durable cross-node moves** - create and operate tasks on a selected cluster node from the app or CLI, filter the task list by node, and move a non-running task with its UUID, schedule, configuration, timestamps, and retained run history. Durable operation IDs make retries safe across disconnects and daemon restarts.
* **Preferred Notes serving node** - each device can pin Notes to one cluster node while connected elsewhere. All vault operations, settings, sync, and S3 tests route to that node; an offline preference stops instead of silently switching vaults.
* **Symlink-safe Agent Docs rollout** - discovery and status show symbolic-link targets. Clusto preserves safe links inside the managed root and blocks broken links, external targets, escaping parent links, and write-time link swaps with a visible reason.

### Fixed

* **Built-in chat MCP works with hostname-only TLS certificates** - when client TLS is enabled, Clusto derives the local MCP WebSocket endpoint from the configured certificate's DNS subject alternative name. It falls back to a local IP only when that exact IP is present in the certificate, and never disables hostname verification.
* **Windows Notes vault paths no longer fail their own root check** - Clusto normalizes Windows extended-length path prefixes before comparing the configured vault and resolved file path, preventing valid default vaults from reporting `path escapes vault root`.

## \[1.14.0] - 2026-08-13

### Fixed

* **Ctrl+C did nothing in Windows run-as terminals** - the service-side run-as helper received the ConPTY control event, but the PowerShell process created in the interactive user's Windows session did not. The helper now bridges the event to a target-session relay, which interrupts the foreground command while keeping the PowerShell session alive.

## \[1.13.89] - 2026-08-11

### Fixed

* **Windows agent chat died before the first reply** - the headless agent CLI was spawned through a `cmd /c` wrapper, and cmd drops the outer quotes of its command line whenever it sees more than two quote characters or a program with no executable extension. An install under `C:\Program Files\...` therefore exited 1 with `'C:\Program' is not recognized as an internal or external command`, surfacing in the chat as "The agent stopped without a reply (exit code 1)". The daemon now spawns the CLI directly (letting the standard library host `.cmd` shims with escaping that survives), prefers an executable extension when resolving a program on `PATH` instead of npm's extensionless shim, and runs batch-file CLIs through `cmd.exe /S /C` in the run-as helper, which `CreateProcessAsUserW` cannot launch on its own.

## \[1.13.70] - 2026-08-02

### Changed

* **Daemon updates now include the node CLI** - every manifest-based daemon update requires and stages the matching `clusto` artifact, installing it beside `clustod` on restart when missing and replacing it when stale. The desktop app no longer exposes separate local, per-node, or cluster-wide CLI update controls.

## \[1.12.10] - 2026-06-10

### Added

* **`clusto version`** - new command printing this CLI's version, its build platform, and the locally installed `clustod` version. `clusto --version` / `-V` print the CLI version alone.

### Changed

* **`clusto update` / `clusto update-node` default to the local node** - a bare `clusto update` already self-updates the CLI following the local node's `[update]` channel and manifest URL; `clusto update-node` (renamed from `update-daemon`, which remains as an alias) now likewise defaults to updating the **local node** with no arguments, instead of requiring `--from <url>`. Both still accept `--from`/`--channel`/`--remote-token` overrides.

## \[1.12.0] - 2026-06-05

### Added

* **Remote-URL updates** - daemons, the CLI, and the desktop / mobile app can self-update from an arbitrary HTTP(S) URL hosting a `RemoteUpdateManifest` JSON document. `clusto update --from <url>` replaces the running CLI, `clusto update-node --from <url> [--refresh-artifacts]` asks the connected daemon to self-update (and optionally bulk-refresh its bundled `clients/` and `daemons/` trees so subsequent peer pulls serve the new version), and the desktop app's Settings → **Updates** card drives both flows from the GUI. Optional Bearer token (`CLUSTO_UPDATE_TOKEN`) for private endpoints; every artifact is SHA-256 verified post-download. See `docs/remote-updates.md`.
* **Persistent + cluster-shared update settings** - the manifest URL, bearer token, auto-check toggle, and interval are now persisted across restarts. Stored on the daemon under a new `[update]` config section and replicated to every cluster peer via `PeerMessage::SyncRemoteUpdateConfig`, so isolated nodes can still auto-update on schedule. The desktop app mirrors the same settings to `localStorage` so it can keep checking and update itself even when no daemon is reachable. Daemons run a background auto-check task (6-hour default) that broadcasts `ServerMessage::RemoteUpdateAvailable` when a newer manifest version appears; install is always manually confirmed.

## \[1.4.0] - 2026-03-05

### Added

* **PTY sessions** - interactive terminal sessions via `portable-pty 0.8`; `SessionMode::Pty` spawns the process in a full PTY, forwarding raw bytes to clients. Clients send `SendPtyInput` for keystrokes and `ResizePty` for window resize events.
* **SQLite persistence** - session metadata and task records are persisted to SQLite (`rusqlite`) at the platform data directory. The daemon restores sessions on startup without re-spawning processes; status changes (Idle, Exited) are written back immediately.
* **Task scheduler** - `Scheduler` runs a 10-second tick loop and fires `Pending` tasks whose `next_run_at` has elapsed. Three schedule types: `OneShot` (immediate, runs once), `Interval { seconds }` (repeating), `Cron { expression }` (6-field cron via `cron 0.12`). Tasks are persisted to SQLite and restored on daemon restart.
* **Notifications** - `NotificationEngine` fires HTTP notifications on session completion, session error, task completion, and cluster role change. Three backends: generic webhook (POST/PUT with custom headers), ntfy (topic + optional bearer token), and Pushover. Configured via `[notifications]` in `config.toml`.
* **Projects** - `Project` type (UUID key, name, description, working directory, default model, extra args) with full CRUD via protocol messages and Tauri commands. Sessions can be associated with a project via `project_id`.
* **Process adoption** - `clusto adopt [--pid <pid>] [--name <name>] [--dir <dir>]` registers a running Claude Code process with the daemon without spawning it. Adopted sessions appear in `clusto ls` and can be killed via `clusto kill`. Dead adopted sessions are reaped automatically.
* **`clusto install`** - extracts the bundled `clustod` binary, runs an interactive config wizard (port, token, display name, cluster settings), and writes a systemd unit on Linux. Flags: `--non-interactive`, `--name <name>`, `--json`.
* **Node display names** - `display_name` field in `DaemonConfig` and `ClusterSection`. Names propagate via peer handshake (`PeerHello`/`PeerHelloAck`) and are shown in cluster status. Runtime rename via `clusto rename-node <id> <name>` or the app's pencil icon; change is broadcast to all peers and persisted to `config.toml`.
* **Daemon version + remote update** - `AuthResult` now carries `daemon_version` and `platform`. Version badges shown in ClusterStatusView and SettingsView. `clusto update-node --address <host:port>` and the app's Update button push a new `clustod` binary to a remote node via WebSocket and trigger a self-restart.
* **Cluster node remove** - `clusto` sends `RemoveNode`; the daemon removes the peer from its config, disconnects it, and broadcasts `PeerForgetNode` to remaining peers so they stop reconnecting.
* **`InitCluster` message** - allows the Tauri app to configure cluster settings on a daemon in standalone mode and optionally restart it to join the cluster.
* **Windows Service** - `clustod` registers with the Windows SCM; install/uninstall via the desktop app's Settings view or `sc` commands. The service uses `CREATE_NO_WINDOW` for all child processes.
* **PTY terminal view** - `PtyTerminal.vue` wraps xterm.js (`@xterm/xterm`) with `FitAddon` for automatic column/row fitting. Scrollback is buffered in the Pinia store (`MAX_PTY_SCROLLBACK = 2000` raw chunks) and replayed on component mount so navigating away and back preserves terminal history. A `ResizeObserver` keeps the terminal sized to its container; resize events are forwarded to the daemon via `resize_pty`. Theme-aware ANSI palettes (`DARK_THEME`/`LIGHT_THEME`) toggle reactively without re-creating the terminal.
* **Shell picker** - `NewSessionDialog.vue` shows a platform-filtered shell preset list for PTY sessions (PowerShell/cmd on Windows, bash/zsh/fish on Unix, plus a `Custom…` free-text entry). The selected command is passed to `new_session` as the optional `command` field; choosing the default `claude` preset sends `None` and lets the daemon decide.
* **App auto-update** - `check_update` Tauri command queries the update manifest and returns `Option<{ version, body }>`; `install_update` downloads and installs the update then restarts the app. Backed by `tauri-plugin-updater`.
* **Remote Terminals (#7)** - `TerminalKind` enum (`Claude` | `Shell { command, args }`) added to `clusto-protocol/src/session.rs`. `SessionInfo` gains a `kind` field (defaults to `Claude`, fully backward-compatible). New `OpenTerminal` client message and `TerminalOpened` server message. `rc open <node> [--shell powershell|bash|zsh|cmd]` CLI command opens a raw interactive shell session on any connected node. The PTY backend now accepts an arbitrary command rather than always launching `claude`.
* **Agent Teams (#9)** - `Team` and `TeamStatus` types added to `clusto-protocol/src/team.rs`. Six new `ClientMessage` variants: `CreateTeam`, `SpawnWorker`, `RelayTeamMessage`, `ListTeams`, `GetTeam`, `DisbandTeam`. Six new `ServerMessage` variants: `TeamCreated`, `WorkerSpawned`, `TeamMessage`, `TeamList`, `TeamUpdated`, `TeamDeleted`. `SessionInfo` gains `team_id: Option<Uuid>` and `is_team_leader: bool`. Teams are replicated across cluster nodes via `PeerMessage::TeamUpserted` / `TeamRemoved`, persisted in SQLite, and accessible via Tauri commands.
* **Desktop app** (Tauri 2.0) - Vue 3 + TypeScript frontend with Pinia stores. Views: Sessions, Cluster, Settings, Connect. Manages connection profiles, session lifecycle, cluster topology, local daemon install/start/stop/restart, and Windows Service registration.
* **Android app** - APK and AAB built via `cargo tauri android build`. Supports the same feature set as the desktop app.
* **CI workflow** - `ci.yml` runs `cargo test` for all Rust crates and Vitest for the frontend on every push.
* **Release workflow** - `release.yml` builds multi-platform binaries (Windows x64, Linux x64, Linux ARM64) and creates a GitHub Release with attached assets.

### Changed

* `SessionManager` event loop (`process_events`) now handles SDK process completion via an `mpsc` channel fed by per-session oneshot watcher tasks, replacing a polling approach.
* Session status transitions to `Idle` (not `Exited`) when an SDK process completes normally; `Exited` is reserved for killed sessions.
* Cluster state replication extended to include projects and their CRUD mutations.
* `ClusterNode.display_name` is `Arc<RwLock<Option<String>>>` to support runtime rename without daemon restart.
* `PeerHello` and `PeerHelloAck` carry `version` and `display_name` fields so joining nodes advertise their version immediately.
* `SyncClusterConfig` is broadcast to all peers after `UpdateClusterConfig` so peer lists stay consistent without restart.
* Follower nodes ban removed peers (via `PeerTransport` ban list) to prevent immediate reconnect after `RemoveNode`.

### Fixed

* Self-connections in `PeerTransport` are now rejected; a node no longer tries to connect to its own peer address.
* `remove-node` no longer triggers a reconnect loop when the removed peer attempts to rejoin.
* `update-node` relay correctly propagates auth failure back to the originating client.
* Version comparison in cluster upgrade flow uses semver ordering, not string comparison.
