LLeMbas-Desktop/CHANGELOG.md

Changelog

All notable changes to LLeMbas Desktop. The format follows Keep a Changelog, and the project uses Semantic Versioning.

[Unreleased]

Added

• Archived chats. Archive chat was one-way: an archived chat left the sidebar and could not be found again in the app. Archived, under Files in the sidebar, lists them newest first; each opens as any chat does, goes back into the sidebar with Restore chat, or is deleted. An archived chat that is open says so in its header, with Restore chat beside it and in its menu.
• An agent chat on another computer. A new agent chat could only work on this computer. Where the account has another computer linked, a chip before the folder chooses it (this computer is the default). Its folder is one its agent chats have worked in, or a path typed in; the Terminal before the first message opens there too.
• Language and theme in Settings. Under Appearance: the language (this computer's, or any of the app's nine; the window opens again in it) and the theme (the account's, or Moria or Shire). Both are kept in the app on this computer; the account's own stay as they are.

Fixed

• A terminal that ended could come back by itself. Shown again and ended by the server within one frame, the panel opened a new shell in its place; an ended shell is now only said.

[1.0.0-beta.6] — 2026-11-10

LLeMbas CLI 2.0.0-beta.6 inside, the look and words of LLeMbas 2.0.0-beta.8.

Fixed

• Approvals show again. An agent asking to run a command, edit a file or carry out a plan stopped with "Waiting for you to allow bash…" and no card to answer it: drawing the card's buttons failed, and the space for it stayed empty. Stopping and asking again gave the same. Now the card shows, with the command in the box to correct before it runs (it used to show the words "Run a command" there instead), what the model says it is for, why it is asking, and which subagent asks when one does. A card already on the screen is kept as it is while the reply goes on, so a command being corrected or an option picked is no longer lost when the server sends the list again; a card that cannot be drawn says why instead of leaving an empty space.
• Subagents after they are switched on. The app read which models a new chat is offered subagents on only when it signed in, so subagents switched on (or a model added) on the server while the app stayed open in the tray never showed on a new chat until it was restarted. A new chat now asks the server again. (An agent chat has no Subagents setting: its subagents run on the computer, and the server offers none for it.)

Added

• Terminal. An agent chat has a Terminal panel: a shell on the computer the chat works on, in its folder, beside the chat (over it in a narrow window). A new agent chat offers it as soon as the folder is chosen, before the first message -- and the chat that message makes carries on in the same shell. An open agent chat opens its own from the header, the same shell the web interface shows. Send to chat puts the selection, or the last command and its output, into the message box to read before sending; the panel copies them too, and steps between commands where the shell marks them. The terminal and the activity panel take turns beside the chat; closing either keeps the shell. Offered where the computer allows a terminal (needs the server's 2.0.0-beta.8). The app opens the connection itself, with its sign-in: the page never holds it. "Continue in a terminal" is now in the chat's menu only.
• What a reply is doing, the same way everywhere. The line under a reply being written now says it in the words every LLeMbas client uses (harness spec 3.4.0): queued, waiting for a model and why, loading it, reading the prompt (with how far), thinking, writing (with its speed), running a tool and what it runs, waiting for you, waiting for subagents, compacting, retrying (counting down), stopped or failed -- each with how long it has lasted, kept moving every second. From a server that sends no such line (before 2.0.0-beta.8), the app shows the server's own words as before. In every language the server has.
• Subagents, the whole tree (harness spec 3.4.0, needs the server's 2.0.0-beta.8). The activity panel shows the main agent and every subagent under the one it reports to: its name and id, its position, what it is doing right now, how full its context is, its state, model and mode (with what it may ask for), its messages in the room and the last one, how many slots are still open, and the ones that ended -- dimmed, with how they ended. Any subagent's conversation opens read-only, live while it works and after it ended, and so do its findings; a working one can be stopped (it and those under it). An agent chat's tree comes from its computer; while that computer is away the panel shows the tree as it last was and says the rest is kept there.
• Subagents for agent chats. A new agent chat offers the Subagents setting where the server says this computer can run them, and an open agent chat where the server offers it for that chat.
• A subagent's approval names it: "Subagent parser (s1), reviews the parser, asks:".
• tests/ui/: the page under a headless browser against a stand-in for the app's bridge, with the server's answers as a table (node tests/ui/run.mjs, needs Playwright and a Chromium). CI runs it on every push (the ui job), with Playwright and its Chromium installed for the run.

[1.0.0-beta.5] — 2026-11-09

Carries LLeMbas CLI 2.0.0-beta.5 and the look of LLeMbas 2.0.0-beta.7, which the Subagents setting and the panel's compaction need.

Changed

• Subagents, by that name everywhere. The chip under the message box says Subagents with the least and the most agents a chat runs ("Subagents 1–4"), and opens one slider with two handles, Min and Max, from 1 to the server's cap -- Min never passes Max, and both move with the arrow keys as well. The numbers count the chat's own agent: 1 is the agent alone, with no subagents; Max 4 is the agent and up to 3 subagents. The chip shows where the server offers subagents for the chat (for a new chat, for its model), and the most is the server's cap. Where the chat's model serves one request at a time, it says that subagents need another model, on another server, named in the chat. Subagents work on the chat's own model unless you name another one in the chat -- there is no setting for a subagent's model. A scheduled task's field is Subagents too. Needs LLeMbas 2.0.0-beta.7 for the server's word on all of this; with an older one the app guesses as before.
• The activity panel, reshaped. A button in the chat's header opens it at any time, and the line above the message box shows only while something runs ("3 running · 1 queued · 2 finished"). An agent chat's Environment (its folder, and where it is) comes first. Context says how far it is to the automatic compaction ("80.6k until auto-compact") beside a Compact button that summarises the earlier turns now (greyed while there is nothing to summarise). Background tasks lists subagents, a team's members and background commands under Running, each with how long it has run and a stop square, unfolding to what it is, its latest step, what it is writing and its conversation; the tasks under one fold beneath it ("Tasks: 2"), and the finished ones fold away under Finished. The team's Room and the Plan have sections of their own.

[1.0.0-beta.4] — 2026-11-09

Carries LLeMbas CLI 2.0.0-beta.3 and the look of LLeMbas 2.0.0-beta.6. Scheduled tasks need LLeMbas 2.0.0-beta.6 or later.

Fixed

• An agent that could not link now gets its link. Signing in asked the server for the agent's token only when the built-in LLeMbas CLI had no login for that server on file -- so a computer with an old login the server no longer takes got the app's own token alone, and the agent stayed refused. The sign-in now asks for the agent's token unless the agent's link works right now, and the CLI's new login replaces the old one. Link this computer (and Link again) is now its own approval for the agent's token, handed to the CLI the same way; before, it ran the CLI's own login, which for a server it was logged in to only refreshed the login it had.
• "Refused" means the server refused, and says the server's reason. A service that stopped for want of a login says not linked; any other reason, stopped with the reason.

Changed

• A new look, the shape of an AI desktop app. A slim sidebar: New chat, a search line, Scheduled, Files, then your chats of both kinds -- agent chats and chats on the server -- under Today and Earlier, an agent chat with its folder's name; Settings with your account at the foot. The sidebar folds away (its button in the title bar, or Ctrl+B) and is a drawer on a narrow window.
• A new chat is a chat page. A greeting, the choice of Agent (on this computer) or Chat (on the server), and the message box ready -- no form. What the chat starts with sits under the text as small chips, each opening a menu: the folder and the mode for an agent chat, the model with its picture, how hard it thinks (as the model offers it), and the team's size where a team is offered. An open chat has the same chips, and changing one changes the chat.
• The conversation in a centred column: your messages in bubbles on the right, with attached pictures as thumbnails and other files as tiles; replies full width; thinking and each tool call a compact row that unfolds. Send is a round button that Stop replaces while a reply is written.
• The activity panel opens from a chip above the message box, on every open chat; while agents work, the chip says what they are doing.
• The agent's status is a chip in the title bar that opens what the agent says and the one thing to do about it. When this computer is not linked -- or the server refused its link -- a new agent chat says so right above the message box, with Link this computer (or Link again) there, not only in Settings; while the approval waits, the code to compare is shown with Cancel.
• Files is a page of its own, with this computer's folders and your files on the server as two tabs. Settings is grouped -- Account, Agent, Project folders, Notifications, About -- with the groups listed beside them.

Added

• Scheduled tasks (with a server that has them, LLeMbas 2.0.0-beta.6 or later): the list with each task's next run, how its last run went and a switch to pause it, kept up to date as the server runs them; a task's page with what it does, when (and in which time zone), its model, effort, team and how it tells you, the next runs, and its history -- each run's status, time, whether it was late, caught up or run by hand, the model it ran on (and that it switched from a busy one), and its result to unfold in place, with its chat to open; earlier runs on request. Run now, Edit, pause and resume, and delete. A new task is said in your own words: the server reads them into a rule, a title and an instruction and shows the next five runs before you save; the title and instruction can be changed, and where the words give no time the days and the time can be set by hand. Each task's settings: when to tell you (when it matters, always, never), a new chat for each run or one that grows, its model (or the first free one of the scheduling models), its effort and its team. Models on the list chooses which models scheduled tasks may run on, each shown free or busy.
• A scheduled task's result comes as a system notification while the app runs, by the task's own setting: a result, never a run with nothing to report unless the task says always, nothing for never; a failure the first time, and on the third in a row the question whether to pause it. The app follows the server's event stream for this rather than asking every few seconds.

[1.0.0-beta.3] — 2026-11-09

Carries LLeMbas CLI 2.0.0-beta.3 and the look of LLeMbas 2.0.0-beta.5. Needs a LLeMbas 2.0 server (2.0.0-beta.5 or later for the activity panel, the Thinking switch and team sizes).

Added

• Attachments. A message can carry files: the attach button beside the composer (the system's file dialog), files dropped anywhere on the window, or an image pasted into the composer. Each is sent to the server at once and shown above the composer with its size, or with the server's reason when it refused it; one taken off before sending is deleted on the server. The app sends only files you chose, dropped or pasted -- the page never names a path. Agent chats get them too.
• Edit a message, regenerate a reply, copy either. Under your messages, Edit message rewrites one in place and starts the reply again from there (everything after it is dropped, as the app says before you save; on the agent's device too); under the last reply, Regenerate reply. Copy puts a message, or a reply's text as Markdown, on the clipboard.
• The chat's own menu (⋯ in its header, or a right click on it in the sidebar): rename it, open it in the web interface in your browser, archive it, delete it (asked first).
• Change a chat's model from its header, and how hard it thinks as the model offers it: a model with one level shows nothing to choose; one that can switch thinking off and has one effort shows a Think before answering switch; more levels keep the selector. A new chat starts at the model's own start level, and the control follows the model you pick.
• Agents, fewest and most, beside it where the chat can have a team (a chat on the server, on a model that serves at least two requests at once, with subagents on): how many agents the chat's team may have, its own agent included. Saved at once in the header; on a new chat, left empty, your default stands. A pair the server will not take is said in its words, and the fields go back.
• Model pictures: a model the server has a picture for shows it in the pickers and under each reply; the others show their initial on a colour of their own -- the same colour the web interface gives them.
• The activity panel beside a chat (the header's bolt button, the strip above the message box, or the palette): what is in the chat's context, by category, against the model's window; the files the chat and its agents wrote, with lines added and removed, each opening in the viewer; every background agent -- a team's members as a tree, helpers and commands -- with what it was asked, its state, how long it has run, its model, its latest step and what it is writing, its whole conversation to read and a button to stop it (a member after you confirm, for those under it stop with it); the team's room, who wrote to whom and which messages are reports; and the plan. It keeps up by itself while the chat is open. The strip above the message box says what the agents are doing ("2 agents working · 1 queued") and opens it. On a narrow window the panel covers the chat; at phone width it fills the window.
• Continue an agent chat in a terminal: its menu and its header show the command that opens the agent's session in a terminal (lembas -s …, in the chat's folder), to copy. On this computer it names the LLeMbas CLI as it starts here -- the one on your PATH, else the app's own.
• A command palette (Ctrl+K): new chats, settings, the agent and updates by name, and every chat of both kinds by its title (accents optional). Ctrl+N starts a new chat of the tab you are on; Ctrl+, opens the settings.

Changed

• The agent's status is the CLI's own word. The title bar and Settings no longer say running because a process runs: they say what the built-in LLeMbas CLI says of its link to the server -- connecting, running (linked), waiting and why (the server cannot be reached, the link dropped, a terminal on this computer holds the link), or refused and why. The same for a lembas service you run yourself. When the CLI says trying again will not help (the server refused this computer's token), the app no longer restarts it every second: it waits until you link the computer again. When the agent stops with an error, its last words are the reason shown, not just an exit status.

Fixed

• A Czech sentence in the viewer was in Slovak; "Server files" and the composer's placeholder were not translated.

[1.0.0-beta.2] — 2026-11-07

Carries LLeMbas CLI 2.0.0-beta.2 and the look of LLeMbas 2.0.0-beta.4.

Changed

• The agent keeps what its model has read (LLeMbas CLI 2.0.0-beta.2, harness spec 3.1.0): the system prompt is made once per session and sent unchanged, so a local model's server reuses the whole conversation instead of reading it again; the date and anything that changes during a session come as a short note on your next message; the tools stay the same for the session; compaction is the next turn of the same conversation.

Fixed

• The editor highlights code again. One of the editor's modules imported a file the app does not have, so every language that parses failed to load and files opened as plain text.

[1.0.0-beta.1] — 2026-11-03

LLeMbas Desktop 1.0 becomes an app of its own: the agent on this computer by default, your server's chats, two explorers and a viewer -- no longer a window onto the web interface. It needs a LLeMbas 2.0 server.

Carries LLeMbas CLI 2.0.0-beta.1. Needs a LLeMbas 2.0 server (2.0.0-beta.1 or later).

Changed

• The app's own interface replaces the server's web pages. The window shows the app's page (ui/, plain HTML, CSS and JavaScript modules, no build step): the chats, the transcript with its steps and tool calls, the reply as it is written, the questions and approvals that wait for an answer, and a composer. It talks to the server's /api/v2.
• Agent chats first. The sidebar opens on agent chats, those on this computer above those on your other devices; a new agent chat is started on this computer, in one of the agent's folders, with its mode and model. Chats on the server are the second tab.
• One approval signs the app in and links this computer. The app asks the server's device login for its own token (your chats, as the web interface reads and writes them, and dictation) and, in the same approval, for the agent's; the approval opens in your browser. The agent's token goes straight to the built-in LLeMbas CLI on its standard input (lembas login --token-stdin), which keeps it in its own configuration as lembas login does; the app never stores it. A computer linked already keeps its link, and Link again under Settings is still there. A server that cannot give both tokens from one approval gets the old two steps.
• The app's own token is not a device: the server lists it under Apps, and only the agent's link is among the devices an agent chat can start on.
• The look is still the server's: its design tokens, now copied from the 2.0 server, and its CodeMirror build for code (scripts/fetch-vendor.sh, pinned in vendor.lock.json). The window takes the account's theme.
• One window with one page: the second webview, and the Linux workaround it needed, are gone. The window still has the app's own title bar, and its edges resize it.
• The page's words, old and new, are in all nine of the server's languages.

Added

• Two explorers, kept apart: this computer's -- the open agent chat's folder, or one of the agent's folders -- read only inside the agent's folders, and the server's files, through /api/v2/files with your own access. A server file that is not text, or too large, says so in your language instead of failing.
• A viewer beside the chat: Markdown drawn, code highlighted (read-only CodeMirror), text, and HTML in a sealed frame on its own address, with no access to the app, no network and no way to navigate. A file the agent shows (show) opens in it: from this computer for an agent chat running here, otherwise as the server keeps it (/api/v2/chats/…/show), checked against the chat's access when it is opened. A file that is gone, too large or not text says so.
• Native notifications when a reply or an agent chat finishes while you are elsewhere (on by default; Settings, Notifications).
• A server older than LLeMbas 2.0 is refused with a message saying which side to update, and so is a server whose app interface is newer than this app knows. What counts is the link protocol (3) and the app interface (app_api 2), not the server's version number.
• Settings in the app: the server and the sign-in, the agent and its folders, notifications and updates. There are no admin pages.
• When the app cannot start, it says why in desktop.log in its log directory, not only on a terminal nobody sees. A missing tray no longer stops it from starting.
• With NVIDIA's own driver, WebKitGTK's DMA-BUF renderer is turned off before the window opens (unless you have set WEBKIT_DISABLE_DMABUF_RENDERER): that renderer leaves the window blank or ends the app there, on Wayland most of all.

Removed

• The server's web interface in the window, the microphone permission it needed, and the theme report the window read from it. The web interface stays in the browser, and as an installable web app.

Fixed

• A release can no longer be quietly replaced under the same version: installing the same version again does nothing, so people with the broken copy never got the fix. The release workflow now refuses a tag that has a release already, unless it is run by hand to replace it.
• A beta's .deb sorts below the release that follows it (1.0.0~beta.1 before 1.0.0), so apt does not keep a beta over its release.

[0.7.2] — 2026-10-25

Linking this computer again leaves one device, and LLeMbas CLI 1.5.2 inside.

Changed

• Carries LLeMbas CLI 1.5.2. Linking this computer again no longer leaves an offline copy of it on the server holding its agent chats: the old link is signed out, and a LLeMbas server 1.7.2 or later moves its chats to the new one. When the agent cannot connect, its log (agent.log in the app's log directory) now says why -- a certificate that is not trusted, for instance -- instead of only "could not reach the instance".

[0.7.1] — 2026-10-24

The microphone works, and LLeMbas CLI 1.5.1 inside.

Fixed

• Dictation said the microphone was blocked, with no way to allow it. The server's page asks for the microphone, and a webview grants it only if the app answers: WebKitGTK refuses whatever it is not answered on, and shows no prompt of its own. Desktop now allows the microphone -- the microphone only, and only for the server's page in the content view; a camera or a screen still gets the platform's refusal.

Changed

• Carries LLeMbas CLI 1.5.1: an image counts by its size, so a screenshot no longer makes the agent compact.

[0.7.0] — 2026-10-20

Carries LLeMbas CLI 1.5.0: harness spec 2.0.0 — background work as jobs the person can see, enter and stop, and subagents that run in the background and talk back.

Changed

• Carries LLeMbas CLI 1.5.0. The agent on this computer runs subagents in the background, keeps every background command and subagent as a job of its session, and reports them to a LLeMbas server 1.6.0, which shows them in the chat with Stop and a message box.

[0.6.0] — 2026-10-20

Carries LLeMbas CLI 1.4.0: harness spec 1.2.0, so the agent's memory, notes and skills are the same tools as the server's.

Changed

• Carries LLeMbas CLI 1.4.0. The agent on this computer and a LLeMbas server 1.5.0 now send a model the same library tools — memory, note_view, note_manage, skills_list, skill_view, skill_manage — whether the agent uses its own library or the server's.

[0.5.0] — 2026-10-18

Carries LLeMbas CLI 1.3.0: a plan in every agent chat, the instance's instructions and knowledge scope on the device, Edit and Regenerate on agent chats. The app's look comes from the server's tokens and follows the page's theme; its own pages speak nine languages.

Changed

• The app's own pages take their look from the server's tokens.css, copied by scripts/fetch-tokens.sh from a server tag and pinned in tokens.lock.json (the script with no argument verifies the copy). The hand copy had drifted: --danger, --control-h and --text-xl were not the server's. The pages name no colour, size or space of their own, and no font is bundled: the font rules are left out of the copy, so the title bar and the first-run page are set in the system's monospace face.
• The window's frame is painted in the theme's border colour read out of tokens.css at build time (build.rs), not a value written in Rust, and it changes with the theme.
• The first-run page and the title bar speak the system's language when it is one of the server's nine (Czech, German, Spanish, French, Hungarian, Italian, Polish, Slovak; English otherwise). ui/i18n.js is generated by scripts/gen-i18n.py: the sentences the server also has are translated as the server translates them, the app's own are in the script.

Fixed

• The title bar followed the desktop's light or dark setting, not the account's theme, so it could be dark over the parchment (Shire) page. It now follows the server page: a script in the content view reports the page's theme to the app, which paints the frame, the title bar and the first-run page in it and remembers it for the next start. The server's pages still get no IPC; the report is a navigation attempt the app reads and refuses.
• The close button's hover colour was a literal white; it is the theme's inverse ink.

[0.4.0] — 2026-10-14

Carries LLeMbas CLI 1.2.0, whose agent stays linked before any folder is allowed and follows changes to its folders without a restart.

Added

• Project folders: Settings… (in the title bar's and the tray's menu) lists the folders the agent may work in, with Add folder… (the system's folder dialog), Remove, the most an agent chat may do there, and whether the server may open a terminal. Turning the agent on with no folder chosen opens this page. Changes go through the bundled CLI (lembas remote).
• The server is told it is shown in LLeMbas Desktop and on which computer, so it lists this computer's device as Local, first.

Fixed

• The title bar's menu opens under the button, not in the middle of the window.
• The agent no longer shows a green light while it may open no folder: the title bar says agent: no folders.

[0.3.0] — 2026-10-14

Added

• Updates: the app looks for a new release at start and every six hours, shows an update button in the title bar when there is one, and installs it (the .deb with the system's password prompt, an AppImage in place) after checking its signature. The menu can also check by hand.

Fixed

• On Linux the server's page took only the lower half of the window, under an empty title bar area. The title bar is now its fixed height and the page takes the rest; the frame you resize the window by is drawn around both.

[0.2.0] — 2026-10-14

Added

• The window is LLeMbas's own: a title bar in the LLeMbas look with the server's address, the agent's state, a menu (the tray's items, plus Reload) and the window buttons. Closing still keeps the app in the tray. The server's pages stay in a view of their own, with no access to the app.

Fixed

• The app icon is the leaf alone, on a transparent background, instead of a dark square tile.
• The window shows the app's icon instead of a generic one (the window, the menu entry and the package now share one name, llembas-desktop).
• The package is called llembas-desktop (it was l-le-mbas) and replaces the old one.
• The bundled LLeMbas CLI is installed in the app's own directory instead of as /usr/bin/lembas, where it could hide, or be hidden by, a lembas installed by hand.
• SHA256SUMS on the GitHub release no longer lists itself.

[0.1.0] — 2026-10-14

The first alpha, for Linux. It carries LLeMbas CLI 1.1.0.

Added

• The app: the server's web UI in a window of its own, a first-run page to choose the server, and a tray menu (open, start and stop the agent, link this computer, change server, quit).
• The agent: the bundled LLeMbas CLI runs as lembas service while the app runs, restarted when it fails and stopped cleanly on quit. A CLI service already running on the computer is used instead of starting a second one.
• Linking: lembas login --json with the approval page shown in the window.
• Linux packages (.deb, AppImage) built by GitHub Actions from a tag.