Changelog

What changed in each version. The desktop app (Mac / Windows) can check for and install updates itself; the Lom Agent phone app needs the new installer, which you install over the old one.

0.3.8 · 2026-11-06

Teams get a group chat view, plus fixes across team orchestration, workspaces, phone receipts and Settings; the built-in core is updated and supports signing in to MCP services.

• Team group chat: a new "Group chat" view at the top of a team. @ specific members or @everyone; with nobody selected, the message goes to the leader only. Leader assignments, members picking up and finishing tasks, and personal-assistant hand-offs all show up in the chat, with live typing / working / waiting status. Restarting or reconnecting the app does not resend anything. Phone members cannot be @-mentioned in the chat yet; their results still reach the leader through receipts.
• Team orchestration: the same task is not sent twice in quick succession; the leader and members (including members that have not started yet) can switch models; when a delegated task ends, the coordinator is reminded to check it, and an empty turn is no longer counted as done.
• Workspaces: a workspace must be chosen before a task starts; fixed files created by AI not opening and "Open workspace folder" failing; the Changes tab shows real file changes, this turn by default; preview failures and missing files are reported as they are.
• Phones: task receipts only count evidence from the current turn, so a previous turn's completion is no longer mistaken for this one; the phone page shows features based on what the phone supports.
• Server management: a new "Add machine" dialog with password or key login, groups and notes, and a connection test before saving; personal assistants can configure servers and run commands, with passwords and keys kept only in the system keychain.
• Settings: removing a machine and clearing an authorization ask for confirmation first; permission settings save more reliably; failures when signing out or copying diagnostics are reported honestly; English UI text is completed and shortcuts follow your OS.
• Windows: Lom's own title bar and window buttons, no blank strip at the top, even spacing around the workbench, and no console window on startup.
• Built-in core: sign-in for MCP services; built-in agent health checks explain why they failed; empty team member turns are reported as failures.

0.3.6 and 0.3.7 were internal test builds and were not released; everything in them is included in 0.3.8. You can check for and install updates from 0.3.5 (automatic updates work from 0.1.5). Install after the current task finishes; the app restarts. If macOS blocks the app the first time, right-click it and choose Open. The phone app is unchanged (1.14-lom.4).

0.3.5 · 2026-11-04

Settings now includes SSH server management and private-key import, personal assistants can hand team tasks directly to members or phones, and a batch of catalog, MCP and phone receipt fixes lands.

• SSH servers: Settings has a server list, health probes and import of existing private keys; keys stay only in the system keychain and a temporary private directory.
• Teams and phones: a personal assistant can hand team tasks to members or phones and read the phone's task receipts; permissions and stopping are covered too.
• Catalog and connectors: updated skill, plugin and connector catalogs, Activepieces app catalog, MCP OAuth connect and cancel, plus fixes for restoring selected apps and hiding internal phone services.
• Phone receipts and history: empty user bubbles are hidden and unconfirmed task receipts are restored after an app restart.
• Fixes: paths and recovery rules for personal-assistant file editing and command execution; the turtle-soup mode summarizes the stage every ten answers.

The desktop core is unchanged from 0.3.4 (delivery-048). You can check for and install updates from 0.3.4 (automatic updates work from 0.1.5). Install after the current task finishes; the app restarts. The phone app is unchanged (1.14-lom.4).

0.2.5 · 2026-10-24

Skills installed from Skill Hub are now found and used by the AI, and the default model no longer shows "Add model". Download the latest version

• Skill not found: before, the AI got a "Skill not found" error when it called an installed skill and had to work around it. The built-in agent's skill tool now sees every skill in the current chat directly. A newly installed skill takes effect in a new chat.
• Default model: when the model is briefly unavailable or not loaded yet, the composer shows the default Lom Pro instead of "Add model"; after you pick another model it shows that one.

The desktop core is updated (lom-core b9d39aa). You can check for and install updates from 0.2.4 (0.1.5 to 0.2.3 work too). Install after the current task finishes; the app restarts. Known limits: the skill fix has unit tests but has not been fully accepted against a real model; a chat that is already open does not pick up a newly installed skill, so start a new chat; the other known limits are the same as in 0.2.4. The phone app is unchanged (1.14-lom.4).

0.2.4 · 2026-10-24

A separate Feishu bot for each personal assistant, permission switching no longer blocked by running tasks, and Skill Hub and interface fixes. Download the latest version

• Multiple Feishu bots: each personal assistant can connect its own Feishu bot and stay online at the same time; registration, authorization, listening, approval and disconnecting do not affect one another. The existing account bot stays with the assistant it was bound to, keeps its authorization, and needs no new scan.
• Permissions: switching permissions in the composer is no longer refused because a task is running; built-in Lom "Full access" now takes effect correctly after a cold start.
• Skill Hub: installing a skill shows a loading animation, and installed skills are available in new chats.
• Turtle-soup questions: the card shows a looping loading animation while the AI prepares its next question.
• Workspace: new "Plan" and "Changes" tabs show the current task's steps and the files it changed; click a file to open it.
• Interface details: when a model is briefly unavailable the composer still shows its name instead of "Add model"; the main area reserves room for the scrollbar so the page no longer shifts sideways when you change a filter.

The desktop core is the same as 0.2.3. You can check for and install updates from 0.2.3 (0.1.5 to 0.2.2 work too). Install after the current task finishes; the app restarts. Known limits: real two-bot Feishu scanning, keychain authorization and end-to-end message delivery have not been accepted on real devices; the other known limits are the same as in 0.2.3. The phone app is unchanged (1.14-lom.4).

0.2.3 · 2026-10-24

Model switching no longer errors, errors are readable, and personal assistants can be deleted. Download the latest version

• Model changes: after the server back office changes the Lom models, creating a task and sending a message no longer fail (they used to be held up by the personal assistant's scheduled review task). When the Lom model a task was using is taken offline, Lom switches to an available model of the same tier and tells you; a bring-your-own key is never quietly swapped for a points-charging model.
• Switching while running: choosing a model or reasoning effort while a task is running no longer errors. It tells you it will switch after this turn and does so automatically; the later choice replaces the earlier one.
• Background sync: model upkeep for scheduled tasks now runs in its own background queue. A failure is shown only on the scheduled-tasks page and the personal assistant's task area, with a re-sync button, and no longer holds up sending.
• Readable errors: notices come in three tones (success, info, error); errors stay longer and long text wraps. Common core, network, permission and timeout errors are shown in plain language, and other technical detail reads "did not finish, please retry". A failed background sync no longer pops up a warning.
• Personal assistant: each assistant tab has its own close ×, which only collapses the tab and keeps the assistant, its sessions and memory. A new "Delete this assistant" deletes its scheduled tasks, archives its sessions (recoverable) and releases its Feishu binding; the documents folder stays on your computer. The permission selector and scrollbar display were also tidied.

The desktop core is the same as 0.2.2. You can check for and install updates from 0.2.2 (0.1.5 to 0.2.1 work too). Install after the current task finishes; the app restarts. Known limits: these changes have not been fully accepted against real devices, real models and a real Feishu account, and the delete confirmation and archive recovery still need to be checked in a real window; the known limits for the Feishu features, permission tiers, assistant delegation and learning are the same as in 0.2.2. The phone app is unchanged (1.14-lom.4).

0.2.2 · 2026-10-24

Personal assistant permission tiers, and Feishu approval, authorization and reading. Download the latest version

• Personal assistant permission tiers: each assistant can be set to "ask every time", "routine actions automatic" or "full access". Full access is wired to the core's fully automatic mode, so confirmations stop appearing one by one; "routine actions automatic" currently only auto-approves structured read-only actions and still asks before writing files. Changing the tier while the assistant is running takes effect after the turn; if a downgrade fails, the previous tier is kept.
• Feishu approval cards: when a turn started from Feishu needs your approval, a one-time card is pushed to your Feishu private chat; tap "Allow once" or "Deny" and it carries on, with no trip back to the computer. If the card cannot be tapped, reply "allow <code> / deny <code>". Answering on the computer first, an expired card or a repeated tap never submits twice. Full access runs directly and sends no approval.
• Feishu authorization: authorization now waits in the background and survives closing the panel; the link is pushed to your Feishu private chat so one tap on your phone finishes it; the result is checked against the real local state; the failure page only offers "retry authorization" and no longer rebuilds the bot, and revoking authorization does not delete the bot.
• The assistant can read your Feishu: nine new Feishu tools let the personal assistant search and read documents, list and search groups and private chats, read and search messages, view group members and find colleagues, and send text messages for you (confirmed first according to the permission tier; an unknown send result is never retried automatically). Complete "Authorize access to my Feishu" first.
• Feishu binding: Feishu is bound to one assistant, so switching pages or assistants no longer sends messages to a different assistant; other assistants show a truthful not-connected state.
• Creating an assistant: an empty name or a failed creation now shows the real reason instead of a misleading file-permission message.

The desktop core is the same as 0.2.1. You can check for and install updates from 0.2.1 (0.1.5 to 0.2.0 work too). Install after the current task finishes; the app restarts. Known limits: the Feishu features (authorization, approval cards, reading tools) and the permission tiers have not been fully accepted against a real Feishu account, real models and real devices; the card buttons need card callbacks enabled on the bot, otherwise approve by text reply; "routine actions automatic" does not yet auto-approve file writes, and "full access" lets the assistant run commands without asking, so use it only in folders you trust; personal assistant delegation and learning remain experimental. The phone app is unchanged (1.14-lom.4).

0.2.1 · 2026-10-23

Personal assistant delegation and automatic learning (experimental). Download the latest version

• Personal assistant delegation (experimental): the assistant can list your other sessions (desktop tasks, external CLIs, teams, phones), read their real progress (requests, replies, tool results, changed files and errors), and hand work to idle sessions, which are marked "dispatched by your personal assistant". Busy sessions are never interrupted. Existing assistants pick this up on their next message; no need to recreate them.
• Personal assistant learning (experimental): after each completed turn, the assistant summarizes once more with that turn's Lom model, writing your expression preferences into USER.md and recurring ways of working into AGENTS.md, effective from the next turn. Your explicit corrections take priority over the assistant's own reflections. This costs one extra model call per turn and rewrites those two documents automatically; ordinary CLI sessions never trigger it.
• Personal assistant: the default SOUL.md is now written in the assistant's own voice (its name and what it can do for you), and the assistant consistently speaks as Lom.
• Feishu (experimental): one QR scan now connects you, with no second authorization scan; an incoming message immediately gets a "typing" reaction that is removed when handling ends; fixed Feishu messages never arriving on Mac (the config directory path was too long for the message channel to start).
• Fixes: weekly scheduled tasks are saved with weekday abbreviations, fixing the default "weekly review" failing to be created and weekly tasks running a day early; after a continued task is archived or deleted, its imported history no longer comes back to the left list, and imported history can now be removed from the list.

The desktop core is updated with this release. This release is Mac-only for now: download it from the website. The Windows build and in-app automatic updates will follow. Known limits: the optional Feishu authorization may show a failure or quietly return; do not click "scan to reconnect" (it creates a new bot and voids the existing authorization) while this is being fixed; personal assistant delegation and learning have not been fully accepted against real models and devices; learning costs an extra model call per turn and rewrites USER.md and AGENTS.md automatically; the Feishu changes have not been verified against a real Feishu account. The phone app is unchanged (1.14-lom.4).

0.2.0 · 2026-10-23

Personal assistant and Feishu connection (experimental). Download the latest version

• Personal assistant (experimental): the first time an assistant opens its task area, three silent review tasks are created (daily 22:00, Sundays 03:00, the 1st of each month 09:00, in the local time zone). Permission is "ask every time", results stay in the assistant's session, and nothing is pushed out by default. Assistant documents are delivered by version and are not attached again while unchanged. Six Xiaolu avatars to choose from; No. 06 is the default.
• Feishu connection (experimental): bundles Feishu's official lark-cli 1.0.97 (MIT, checksum pinned); scan a QR code in the personal assistant panel to connect. Only messages from the authorized owner in a private chat with the bot are accepted; they go into the chosen assistant's session labelled "from Feishu". Replies stream into a card as they are written, and fall back to plain text if the card cannot be created. You authorize it yourself by scanning; this first release has not yet been tested long-term against a real Feishu account.
• Windows: closing the window now tucks Lom into the system tray and keeps it running, so scheduled tasks, the phone bridge and teams are not interrupted. Left-click the tray icon to open the window, right-click to quit; starting Lom again brings back the running window.
• Search: when CodeBuddy calls Lom search through DeferExecuteTool, the source bar now shows too; search results wrapped in a code fence are recognized as well.
• Fixes: archived tasks no longer come back to the left list; the Lom avatar in teams now sits on a round base.

You can check for and install updates from 0.1.9 (0.1.5 to 0.1.8 work too). Install after the current task finishes; the app restarts. Known limits: Feishu and personal-assistant self-evolution are experimental. Real Feishu scanning, sending and streaming cards have not been tested against a real Feishu account; the self-evolution write path exists only as an internal module, is not yet wired into conversations, and will not rewrite your documents on its own. The Windows tray and single-instance behaviour still need confirming on a real PC. The phone app is unchanged (1.14-lom.4).

0.1.9 · 2026-10-23

Personal Agent and team experience update. Download the latest version

• Personal Agent: the Personal Agent entry in the account menu now opens a floating panel inside the main window that you can drag, resize and minimize. Create several assistants, each with its own conversations, documents and scheduled tasks; they only use your Lom account models. WeChat, Feishu and calls still show "not connected".
• Teams: adding the same agent to a team several times gives each copy its own conversation; the mode switch shows "Side by side / Board / Single" in words; the team folder button opens the file panel on the right; the board shows the real conversations and run state of phone members.
• Search: when an external agent uses Lom web search, it now shows site icons and clickable source links too.
• Phones: the desktop can manage the sub-agents of real phones (read, create, edit, enable and disable). This needs Lom Agent 1.14-lom.4 on the phone.
• Fixes: changing the reasoning effort only says "applied" once it is confirmed; side-chat follow-ups now carry the main task's latest progress.

You can update from 0.1.8. Install when your current task is done; the app restarts. Known issues: the features above have not yet been fully tested on real devices, real CLIs and phones; Lom Agent 1.14-lom.4 has not been installed on a real phone.

0.1.8 · 2026-10-22

History import, web search and team board update. Download the latest version

• History import: scan and archive your local Codex and Claude Code conversations, sorted by source in the sidebar. Drag a conversation into another agent's chat box as an editable context attachment. It is never sent automatically and the original record is never modified.
• Web search: replies show the sources they used. Team members can search in parallel, and the leader summarizes the results and marks each as "agree", "conflict" or "one source only". External agents can use Lom search too.
• Continue old chats: imported Codex and Claude Code history can be continued in the matching CLI; conversations show as draggable reference cards and prefer their native titles.
• Send images to phones: a desktop agent can send images to phones you have authorized.
• The team page is now an activity board; click a member's card to open that member's conversation.
• Side chat can read the main task's full context (read-only) and refreshes live.
• Cleaner replies: one set of copy / regenerate buttons per turn; Lom and external agents share the same Markdown and thinking-block layout; nested lists keep their levels; IP addresses, version numbers and code are no longer rewritten; images keep their original aspect ratio.
• Input box: reasoning effort follows each agent's own levels; the placeholder types itself out and stops when your system asks for reduced motion.
• Phones: when a phone's model stops working, Lom no longer quietly switches to the computer's model; a scheduled task whose reply was lost no longer runs twice after a restart; phone conversations you haven't opened still sync their running state; after you confirm a send, the old draft is cleared and anything you're typing is kept.
• Other: a late response to a skill purchase no longer overwrites the credits of the account you switched to; choosing a work folder in an open task now saves it to that conversation; subscription model and tier mapping sync is fixed; Windows uses a custom title bar and hides the core's console window.
• Lom Agent 1.14-lom.3 (phone): fixes a crash when opening the permission dialog on some customized Android systems such as Flyme.

You can update from 0.1.7 (and from 0.1.5 / 0.1.6). Install when your current task is done; the app restarts.

Known issues: History import only supports the default folders of Codex and Claude Code for now; other sources have only been researched. Dragging, a real model reading the imported files, and upgrading across versions still need testing on real devices.

0.1.7 · 2026-10-19

Skillhub upload and review update.

• The upload menu offers "Let Lom upload" and manual upload, and lists the local skills you can actually reach.
• Choosing a skill opens the main Lom conversation with an editable prompt filled in. Your existing draft is kept and nothing is sent automatically.
• A Lom-specific skill upload tool, plus cloud-side review of the description, source and license.
• Verified open-source skills are always free and link to their official GitHub source.
• Stronger handling of duplicate uploads, version review and ZIP-security dependencies.

0.1.6 · 2026-10-18

Desktop experience and stability update.

• Recent tasks are grouped by agent. The history filter is independent of the current chat, so browsing another agent's records never switches your conversation or clears your draft.
• Group logos are larger and round, with the running state shown right on the logo's edge.
• Compact command and plus menus; the plus menu can open a side chat.
• Fixed the thinking and tool-call animations and smoothed the sidebar open and close transitions.
• Bundled stability fixes for stopping a response, switching accounts, permission prompts, disconnects and history recovery.
• A failing optional search setting no longer blocks the normal model from starting.

0.1.5 · 2026-10-17

Automatic updates for Mac and Windows.

• Mac (Apple silicon) and Windows (x64) can check for, download and install updates inside the app.
• The app checks at startup, and you can check by hand under Settings → About Lom → Check for updates.
• Shows the release notes and download progress, and verifies the signature after downloading. Nothing is installed until you click "Install and restart".
• Updates wait while a task is running or waiting for approval, so save anything you're editing first.

Versions 0.1.4 and earlier have no automatic updates, so install 0.1.5 or newer by hand once. The Mac app uses an ad-hoc signature and the Windows installer has no commercial code signature; the update signature does not replace system developer certification.