Version History
LIA functional updates version by version
v5.1
v5.1.0 — 🧭 The right mode, wherever you talkNovember 16, 2026
- 🧭 Automatic mode — Choose Automatic in the chat header: LIA picks Pipeline or ReAct for each request, favouring the loop when the work needs exploration or an action. When LIA chooses an engine, its badge appears below the answer as soon as it finishes and stays in the history, including when Telegram is active. Tickets, routines and Telegram offer the same choice; the menu says when the instance cannot choose automatically.
- 📎 Telegram reads more — Send photos, documents, albums or your location, and reply to a message to give LIA its context. Answers keep their headings, lists, tables and quotes; generated images and documents arrive as files. The chat's file limits still apply.
- ⏯️ Telegram keeps your place — /stop stops an answer; a text sent while LIA works waits for its turn. If the answer fails or is stopped, you can send or withdraw the waiting text. With spoken replies enabled, a voice note also gets a voice reply. Routine results reach Telegram too.
- 📲 Notifications through ntfy — Subscribe the ntfy app on your phone or computer to your instance's server, or another allowed server. A QR code and a test message help you connect. By default, only a notice leaves LIA; Full text is your choice for each subscription.
- ⏰ A reminder can wait — When it rings, postpone that occurrence from its message or notification, or confirm that you want to stop the series. The offered times follow your timezone, and pressing twice from two devices does not create two postponed reminders.
- 🖋️ Your words, your choice — Writing style is learned only when you ask, from Settings or a confirmed chat request. You can learn it again or forget it; it no longer renews itself. E-mails LIA writes use that style, while words you dictate explicitly stay as you gave them.
- 💬 Results return to their discussion — Calls, meetings started or imported from a discussion, and routines run manually from its chat return there, or to Main if that discussion was reset or deleted. A notification for another discussion names it and offers to open it.
- ✉️ Read the e-mail as it was sent — Open its original layout from the chat card, as in Tools, without a model call. Remote images wait for your click. When LIA summarises the mail or turns it into an event, it reads the words that layout actually shows.
- 💓 Health figures follow your day — Overlapping iPhone and Apple Watch step samples are counted once per time slice, including history already received. Charts, averages and requested periods follow your account's timezone; a day without samples is not shown as zero.
- 📅 Read across your lists — Availability, the home page and other calendar summaries read every calendar you own. Internal readers follow provider pages for calendars, task lists and mail history; a single tool request that reaches its limit says so.
- 🧠 A longer conversation keeps its bearings — Compaction preserves recent words you wrote and important identifiers. If a model refuses a conversation that no longer fits, LIA compacts it and tries once more only when the turn has not already carried out an action.
- 🧩 Restore through the chat — Attach a memory export and ask LIA to import it, or ask to restore a skill's previous version. LIA shows what will change and waits for confirmation, rechecking that the record has not changed meanwhile.
- 📇 Calls and texts on a contact's card — Recent calls and incoming SMS appear even for a contact with only a phone number. After a call, LIA distinguishes voicemail, no answer and a failed line, and says whether a voicemail message was left when the provider recorded it.
- 🔌 Stripe in Known MCPs — Add Stripe from the ready-made server gallery. OAuth discovery also reads the metadata published under a server's path, allowing connections that previously failed at sign-in.
- 📱 Small-screen fixes — Dialogs and long file or discussion names stay within the screen. The home page's unread-mail label describes the count actually shown, and an answer's mode badge wraps as a whole on narrow phones.
- 🌐 A clearer introduction — The public home page explains LIA's services, personal context and ways to act, with a more focused presentation and a sourced comparison that separates what is built in from what needs an extension.
- ⚙️ For administrators — The installer offers an optional ntfy server, off by default; its iPhone wake relay is a separate choice. Automatic mode has its own JEV configuration and states its fallback. Reading compressed documents is bounded by their expanded size; knowledge spaces explain when that limit is exceeded. Administrator diagnostics retain the processing details through the end of the answer.
v5.0
v5.0.0 — 💬 A discussion for every subjectNovember 14, 2026
- 💬 Several discussions — The pill in the middle of the chat header opens your discussions: Main, always there, and the side discussions you create for a subject, each with its own history, draft and cost. Rename, export, reset or delete them; reminders and routines keep arriving in Main, and Search all discussions finds your words everywhere.
- ⏳ Write while LIA answers — The field stays open: Send next keeps your message waiting, and it leaves by itself once the answer is done. Until then you can edit it, withdraw it or Send now. In ReAct mode, Add to the answer in progress lets LIA read it at its next step.
- 📡 Routines that react — A routine can start When another system calls it (a home automation box, a code forge, a form), through its own address and a secret shown once, or when A web page changes: LIA reads the public page itself and runs the routine when new lines appear, a price for instance.
- 🧩 Ready-made routines — The Template list fills the form: Morning briefing, Inbox triage, Weekly review, Overdue task. Only notify me when there is something new keeps the answer in the chat without disturbing you for nothing, and a routine now tells you when it failed, was skipped or missed its time.
- ✉️ Mail that knows it was answered — Chat cards, the home and Tools › Mail mark a message Answered, Reply drafted, not sent or Reply received, read from your mailbox. LIA stops offering to answer what you already answered, and notifications leave it out.
- 🖋️ E-mails in your voice — The first time LIA writes an e-mail for you, it learns how you write (greeting, sign-off, signature, tone) from your latest sent e-mails, never what you wrote about. See it, switch it off or choose Forget in Settings › Writing style.
- 📊 Richer documents — Spreadsheets with real formulas, totals, drop-down lists and charts; presentations with key figures, quotes, pictograms and editable charts; reports with a cover page and cited sources. Settings › Documents decides illustrations (each one is an image generation, counted in your costs) and the animations of a presentation.
- 🌐 Web search without a key — Where your instance runs its own search, LIA finds web links for you even with no search key: in the chat, in your interests and in the Tools. Your own Brave key keeps priority.
- 🧰 Known MCP servers — Known MCPs opens a gallery of servers ready to add (Notion, GitHub, Linear, Todoist, Canva…): pick one, the form opens already filled, and one you already have is marked Installed. You can also ask in the chat: « Add the Notion MCP server ».
- 🗂️ Categories on the workboard — The Categories button sorts your tickets into lanes, plus To categorize; deleting a category deletes no ticket. The Workboard is now in the menu, with The board at a glance at its top.
- ✅ All your tasks, all your files — The Tools open on All tasks, and LIA reads every list, naming the list of each task. In the chat you can now attach Word, Excel, PowerPoint, OpenDocument, EPUB, CSV, Markdown and other documents, no longer images and PDFs only.
- 🧠 Skills and memories — Each skill says how often you use it; those left unused are proposed with Switch them off or Keep them, and a replaced skill keeps one previous version to restore. Import brings an export of your memories back in.
- 🌦️ A livelier companion — It acknowledges your message, presents the answer, looks at what matters on the screen, moves its mouth on the voice you actually hear and plays the weather where you are. The theme button also reaches Automatic, and an interactive page can draw a map.
- 🔧 Fixed — An event moved in Tools › Calendar no longer lands an hour late. A task asked in a given list no longer goes to the default one, a mail you opened is no longer announced as unread, a failed web source is said instead of « nothing found », a list opened in a dialog no longer freezes the page, and the theme button no longer flickers between two open tabs.
- ⚙️ For administrators — The installer proposes the instance’s own web search (SearXNG) and now asks for an Anthropic and a DeepSeek key. A switch governs Routines started by a call, each account has a Writing style switch, and a capability you switch off disappears from every menu.
v4.6
v4.6.0 — 🧭 Everywhere, your wayNovember 10, 2026
- ✍️ Text menus everywhere — Select words anywhere in LIA, the chat included: Read, Question, Explain, Summarize, Simplify, Translate, Internet, or make a task, an event, a reminder, a ticket or a memory of them through the form you know, saved by your own click. While you write, a selection is corrected, formatted, shortened, made more formal or warmer.
- 💡 Proposals from your use — What you ask for on many days, whatever the hour, becomes a proposal in Alerts › Proposals: a routine ready before you ask, a routine moved, your notification hours widened, a source to connect. Accept does it at once, Later brings it back in two weeks, Never closes it.
- 🎭 Your personality in the voice — In Live, in Live direct, on a call with you and in the spoken comments, the voice speaks in the personality chosen in LIA Style, facts unchanged. A call to someone else stays cordial.
- 🧭 From a card to its tool — An e-mail, an event, a contact, a task, a file or a light shown in the chat opens in the Tools. On the home, a click opens the item in its tool, and More actions starts with the direct ones: mark as read, archive, done, postpone.
- 🎨 Your style — Choose your icons in Settings › Appearance: Lucide, Fluent Color or Streamline Flex. Automatic switches between light and absolute black at the hours you choose.
- ✉️ Cards shaped like their content — An envelope for a mail, an invitation for an event, a folder for files, the sources in the margin of a research, all in your palette. A location opens an interactive map once its cost is shown, and air quality and pollen join the weather.
- 🔎 Find and sort — Knowledge spaces are searched by their name or a file’s, sorted, and keep their favourites apart, right after Meetings in the menu. The activity timeline and every section of Alerts search, filter and sort, received SMS included. A meeting is deleted from its row.
- 👤 Contacts and connections — Write a contact with LIA, ask LIA to Call one of their numbers, and find the calls and texts in Latest exchanges. A message to a connection asked in words now leaves without a confirmation card.
- 🦾 Companions that move — Arms, wrists, legs and feet follow listening, thought and speech wherever a companion has them, and winter scarves fit every silhouette. Smiley leaves the catalogue: a saved choice moves to the default companion.
- 🔧 Fixed — A Microsoft To Do task due today no longer shows as late east of Greenwich. A call to someone else no longer takes your personality, an answering machine hears your portal’s message rather than one LIA composed, and visiting a public page no longer changes your language.
- ⚙️ For administrators — Each JEV decision threshold can be set within its bounds, with thirty days of counted judgements beside it, and mails can show Reply expected. A knowledge space can hold up to 10,000 documents.
v4.5
v4.5.0 — ✨ LIA, one click awayNovember 7, 2026
- ✨ LIA in the Tools — Yellow stars mark where LIA can help. It sums up, translates and reads a message aloud, sums up a conversation, a file or an attachment, proposes a reply, rephrases what you write and takes your dictation. Nothing leaves without your click, nothing is written at your provider, and the cost shows beside the answer.
- 🪄 One sentence is enough — Describe an event, a task or a search: LIA fills the form and you check it. A photo of a business card makes a contact, and a contact can be completed from their messages, each value with the message that says it. Sort weighs the messages of the page.
- ✉️ A reply can wait — A reply or a forward can be kept as a draft, in its conversation, with its quote and its files, on Gmail, Outlook and iCloud. LIA also attaches your files to the drafts it prepares.
- 🔔 Within reach of every item — Remind me, Ticket for LIA, Keep in a space, Tell me when they reply, ready questions for the chat, and on an event See the minutes and Record this meeting.
- 🧭 Screens that look alike — An open event, contact, file or task offers its acts in its bar, which always ends on Delete, then the stars. Every form opens and ends the same way, and every list looks like a field.
- 📱 From a phone to a large screen — On a large screen the mailbox opens on your newest message, which stays unread. On a phone each button of a page’s end takes its own line, and folders open as wide as the screen. On an iPhone, typing in a field no longer zooms the page.
- 💡 Lights and health — A room that is off stands apart from the page, a little darker in a light theme, a little lighter in a dark one. The Health card sets each measure on a dial, today and your average side by side.
- 🧠 Editing a memory — The form counts your characters and states the limit, and clearing a nuance or a topic really clears it. A memory past the limit no longer takes the whole list down.
- 🔧 Fixed — Reading LIA’s briefing aloud answers to your limits before a paid voice speaks. Write with LIA opens from an event’s description, a date field no longer passes the edge of its panel on an iPhone, and copying a text works on an instance over plain HTTP too.
- ⚙️ For administrators — The assists have their switch in the platform capabilities and their Tool Assist line in the LLM configuration. The sixteen domain-agent lines no turn called leave, Emails Agent stays. The catalogue gains GPT-6.1 Terra and GPT Image 2.6 Flare.
v4.4
v4.4.0 — 🛠️ All your work, by hand or by asking6 November 2026
- 🖥️ Pages that run in the chat — Ask for a game, a quiz or a calculator: LIA writes it and it runs in the conversation, with its .html file below it and its code behind a button. Asking for the code still gives you the code. The page stays in an isolated frame that reaches neither your account nor any site of its choosing.
- ☁️ OneDrive joins Google Drive — Tools, LIA, the home page’s Documents card, reading Word and Excel, knowledge-space folders: everything now works with OneDrive too. One drive is active at a time; switching keeps a linked folder’s documents, its sync paused.
- 📖 Your documents read in the page — A whole PDF, a Word file as its author laid it out, a Markdown, a spreadsheet or a CSV as a table, on all your devices. A plain text file changes from its detail, or by LIA after a card showing the whole new text.
- 🗂️ Your mail sorts — Archive, move, star, report spam, restore and undo, by hand or through LIA, on Gmail, Outlook and iCloud. The list reads itself again when mail arrives, and the rail says what is unread.
- ✍️ Write as in your mailbox — Attach a file from your device or your drive, format your words, change a reply’s recipients, keep your signature; a draft saves itself and Send leaves a moment to undo. A message shows its conversation, and LIA reads a whole exchange.
- 🏖️ Outlook: automatic replies and rules — As with Gmail, in Tools and through LIA. An older Outlook connection asks to be reconnected once.
- 📅 A complete calendar — Answer invitations, make an event repeat and change a whole series, on the three providers. Day view, drag and drop, search, an event’s own time zone, your own calendars. LIA shows a card before answering for you.
- ☑️ Tasks and contacts — Today, Overdue and Coming up gather all your lists; Microsoft To Do keeps its steps, importance and repetitions. A contact gains a photo, groups, a job, a birthday, vCard export and import, and your latest exchanges.
- 📁 Your files, in order — Folders, drag-and-drop uploads, rename, move, copy, share with someone or by a link, a trash with Undo. LIA organises your files too and shows a card before sharing one.
- 🔎 One search — Ctrl+K searches your mail, calendars, contacts and files at once, and says what it could not reach.
- 💡 Your lights — Hue zones, a whole room’s colour, scenes in motion and the lamps that are not answering. All off turns into All on once nothing is lit.
- 🪟 Room on a large screen — Tools takes the whole window, other pages and the header widen and the chat gains room, while the words of an answer keep a comfortable reading width. Phones and tablets are unchanged.
- 📱 Android app — A file downloads with your session and a link in a mail opens in your browser. This arrives with the next store update.
- 🔧 Fixed — Previewing a file no longer fails one time out of two. A forward works on Outlook and iCloud, copies included. A code block in an HTML answer stays whole, your own message shows what you typed, and a cancelled turn is no longer counted twice.
- ⚙️ For administrators — The timeout set for a model is the one its calls get, routines and tickets included. Pages have their own line in the LLM settings and their own switch. The Microsoft app gains the Files and MailboxSettings permissions and the OneDrive redirect address.
v4.3
v4.3.0 — 🧰 Your services, in your hands1 November 2026
- 🧰 Tools — Open your mail, calendar, tasks, contacts, files and Hue lights directly from the header. You act yourself, with no AI token cost.
- ✉️ Your mailbox — Read, search, write, reply, forward and keep drafts. The original layout opens in an isolated reader that blocks remote content; Gmail also offers labels, vacation replies and filters. Links in the original layout remain inactive.
- 📅 Your other services — Agenda, Week and Month, all-day events, task lists, contact details, Drive exports and lamp controls: each screen offers what your connected provider supports.
- 🪟 Room to work — Resize panes on a large screen; this device remembers their widths. On a phone, swipe the tool row. Contacts are suggested for both message recipients and event guests.
- 🛡️ Your MCP confirmations — Choose for your account, each server and each tool. “What it will do” explains when LIA asks in each mode; before you reduce confirmations, the page shows which questions disappear.
- 🔐 A choice you control — “Never ask” is suspended if a tool changes its declaration. The chat can make LIA ask more; reducing confirmations happens in Settings. Completed acts remain in the actions register.
- 🔄 A task continues after approval — A dedicated MCP agent asks once before its task when one of its tools needs your approval, then carries it through. LIA reports the result it obtained.
- 🎯 Which one did you mean? — Designate an item already shown by its sender, title or position. If several fit, LIA asks with numbered choices instead of picking the first; answer by number, words or a click.
- 📇 More faithful connected data — Contacts keep their multiple addresses and numbers, all-day dates stay consistent, and Drive full-text searches work across pages. Gmail reads reflect changes made in Tools. Contact lists are ordered by family name, then given name.
- 🗺️ Your maps stay active — With unchanged content, active maps and widgets retain their state during automatic display updates, without another activation or map load.
- ⚙️ For administrators and routine owners — Tune routine interval and daily-run bounds. Pre-approve selected MCP servers in a routine’s form; other calls needing approval remain refused when nobody can answer. New dashboard panels cover Tools.
v4.2
v4.2.0 — ⚡ Less waiting, and when you wait, you know why30 October 2026
- ⚡ Your MCP servers no longer hold up a message — LIA keeps what each of your MCP servers offers and reuses it from one message to the next, even after a restart, so a message usually waits for none of them.
- ⏳ A slow server is named — When your servers must be asked again, LIA prepares them in parallel. If one takes a while, a small notice names it; if it is still not ready after a short wait, your message is answered without its tools, and the server is ready for the next one.
- 🛡️ A tool is checked before it is used — When what LIA keeps of a server is no longer fresh, she first checks that the server still offers the tool in the same form. If not, the call is not sent and LIA tells you.
- 🏠 The home page fills in card by card — Your reminders, tickets and SMS appear at once; a slower card (weather, calendar, mails…) shows its outline, then fills in on its own. The greeting and the summary of the day come once every card has arrived, and reloading the page soon after reuses them.
- 🔁 An answer is asked for once — If an AI model’s answer is lost on the way, LIA no longer asks for it again behind your back: a request that may already have run, and been billed, is not sent a second time. A request the provider turned away, or never received, is still retried.
- 📅 No more duplicate events — Creating an event in Google, Outlook or Apple now survives a lost connection: the event carries an identifier that stops the calendar from creating a second one. Contacts, tasks and emails are sent only once.
- ☁️ Weather in half the requests — The weather card now asks for the day’s forecast in one go: three requests instead of six.
- 📇 A clearer action bar on a contact — On a person’s page in Relations, one button per kind — email, directions and, with telephony, call — opens a menu when there are several addresses or numbers. Address-book labels (home, work…) appear in your language, and the buttons are easier to tap on a phone.
- 🌍 A more refined LIA World — Materials, light and plants were reworked, the fountain, benches, easel, telescope and shelves look right again, and in a private conversation your own message now appears above your avatar.
- 📝 Public pages rewritten — The home page, Philosophy, Technical, Story and the blog articles were rewritten shorter and plainer in all six languages, and the home page now loads much lighter images, icons and texts.
- 📖 Help that matches the app — The help now says plainly that dictation works in your account’s language, one of the six languages of the interface, and when LIA asks you before calling on an expert.
- 🛠️ For administrators — New settings bound the preparation of MCP servers (wait ceiling, slow-server notice, parallel preparations), the progressive home page and a model’s retries after an error; Grafana shows the waits people actually see, and robots.txt gives every crawler, AI crawlers included, the same rules.
v4.1
v4.1.0 — 🕹️ A livelier world: play, gesture, keep the moment28 October 2026
- 🕹️ Play in the arcade — Three cabinets in LIA World offer tic-tac-toe, four-in-a-row and Reversi, with another person or against the computer. The server confirms every move, and playing calls no AI model.
- 👋 Say it with a gesture — Wave, thank, applaud, laugh, ask or celebrate: six gestures appear above your avatar. If one is refused, you are told why.
- 📸 Keep the moment — Take a private photo of your avatar, LIA and the scenery, never of other visitors. Download it, or share it where your device offers it.
- 🪑 Your dashboard at a table — Sit at a table and choose whether to open your full dashboard. Saying no keeps you seated and costs nothing.
- 💬 A quick conversation bar — Write to the nearest person, or to everyone when nobody is close. The destination is shown and stays put while you write.
- 🌿 A world that answers back — Familiar objects show your fired reminders, messages and tickets, LIA’s face follows what she is really doing, and seasons, the pond, the fountain, lamps and machines react. Ten starting appearances are ready to edit.
- 🛡️ Fewer interruptions — Coming back to the window or a brief reconnection no longer cuts LIA’s answer or ends a Live session, and accepting a player after a solo rematch no longer disconnects you.
- 🏳️ Flags beside languages — The language selector and the language settings show a small flag beside each language.
- 🔧 Steadier everyday screens — Retry in the Drive folder picker really reloads, reopening it lists your folders again, and the pricing tables no longer mix an old refresh into a newer search.
- 🍎 Photo sharing on iPhone — Saving a shared World photo to your library is properly declared to the system.
- 🔒 Your words stay yours — LIA’s internal tool registry now keeps only the length of a request, never an excerpt of it.
- 🛠️ For administrators — The arcade can be switched off and its waits and deadlines tuned, and every alert now names the host it comes from.
v4.0
v4.0.0 — 🌍 A world to explore, your assistant beside you27 October 2026
- 🌍 Step into LIA World — Explore a shared campus and garden with LIA, using keyboard or touch, a minimap and camera controls. The operator chooses whether this optional space is available.
- 🏠 Your own place — Choose your avatar and your home's colour palette. Your home preferences and assistant data stay private, even when you meet other people in the shared world.
- 🧭 Familiar functions, in their rooms — Open conversations, meetings, knowledge, Relations and other enabled activities in private native panels. Approval requests, permissions, costs and quotas follow the same rules.
- 🤝 Meet and write — Consent to displaying your account name, see admitted people, exchange direct text or join an explicitly public discussion. Accepted, unblocked connections can see your coarse World presence.
- 🌐 Translation by choice — Choose translation when you need it, with the configured model and spending controls. Private messages do not become assistant history. Direct voice between people is not part of this release.
- 🌙 A place with atmosphere — Day and night, optional cached weather, local sounds and separate volume controls accompany your visit. Decoration makes no model or weather-provider request.
- 🛠️ A controlled shared space — Administrators can enable World and set its capacity. A waiting queue and explicit tab control keep admission predictable; losing the connection pauses interaction until reconnection.
- 🧠 More model choices — Jev 1.14 and Gemini 4 Argon join the catalogue. Optional JEV assistance keeps its own controls, accounting and documented limits.
- 🛡️ Smoother uploads and stronger privacy — Large PDF and HEIC processing leave the main request loop available. Cancellation and transaction cleanup are more reliable, log addresses are better protected, and production startup allows more time for migrations.
- ⌨️ Focus returns to “+” — Dismiss a dialog opened from the attachment menu with Escape or Close, and keyboard focus returns to the “+” button.
- 📈 LIA’s evolution stays readable — In “LIA’s psyche”, history keeps every snapshot even when several share the same instant, without repeating time labels.
v3.7
v3.7.0 — 🔎 Found by first name, shared in plain sight25 October 2026
- 🔎 A first name is enough — Ask about a connection or a person you deal with by first name: LIA reaches the one it designates and names whom it read. When several people share that name, LIA asks which one — it never picks one for you.
- 🤝 Who you shared it with — Under the line of a skill, a plugin or an MCP server, « Shared with » lists the connections you gave it to, and « Received from » says who gave it to you. A copy already given is never sent twice.
- ♻️ Bring back what you kept — A bookmark, a generated image or a generated document comes back into any discussion, from the « + » or from its own card, as a copy: the original stays where it is.
- 🗓️ A timeline for each relationship — A Timeline section on the relationship card lines up commitments, calls, memories, relayed messages and text messages on one dated thread, the older ones a click away.
- 🎙️ Turn a recording into a meeting — A WAV, MP3 or M4A file you already have becomes a meeting with its minutes: « Import » now sits beside « Record » on the Meetings page and in the « + ». The file’s length is checked before anything is spent.
- 🔢 Each step says what it did — Under an answer, each step shows how many times a tool ran, how many results it found and read, and whether a call failed — while LIA works and afterwards.
- 💶 A phone call’s report shows its cost — The report of a call LIA made now shows what it cost, like any answer, and a voice session’s card updates itself in every open tab.
- 🏷️ E-mail and appointment labels read right — LIA no longer calls a still-valid label « expired », and its caution about an unverified e-mail no longer spills over onto your appointments.
- 🛠️ For administrators — Meeting imports have their own switch and limits (file size, chunk size, abandonment delay), the demonstrator’s configuration templates are organised by theme behind a table of contents, and a new metric counts imports by format and outcome.
- 🔍 The V15 audit is published — 7.9/10 across 24 areas on version 3.6.0, with its 17 open findings and the public register that follows each of them to its proof.
v3.6
v3.6.0 — 🛡️ Safer sign-ins, steadier connections24 October 2026
- 🔐 A password reset that protects your account — A reset link now opens only the account it was sent to, and only until your password changes; each account receives a few reset e-mails a day at most, whoever asks. If you sign in with Google and have no password, a reset request reminds you how you sign in — it never sends a link that would create a password.
- 🤖 An anti-robot check on sign-up, sign-in and reset — Your administrator can switch on Cloudflare Turnstile to keep scripts away from these three forms. Most people never see it, and if Cloudflare cannot be reached, you can still sign in.
- 💬 A discussion that waits an hour, not a day — When a discussion LIA proposed is still waiting for your reply, LIA now holds the next one back for an hour instead of a whole day, and the activity panel tells you it is waiting for you.
- 🔗 Connected services that stay connected — Access to Google, Microsoft and Philips Hue is renewed once, through a single door, even when several features ask at the same moment, and a service is marked disconnected only once the provider’s refusal is confirmed.
- 🛠️ For administrators — An administrator account receives no reset e-mail: a command on the server sets its password and signs every device out. New alerts warn of reset abuse, sign-up bursts and an unreachable anti-robot check; a notification counts as delivered only once it is archived, pushed or sent on a channel; alerts now reach the team receivers; the database bounds its own lock waits and idle transactions; and the configuration templates now carry the values production runs, while a new guided installation keeps writing a reply when its page is closed — the configuration of an instance already running is untouched.
- 📋 Every audit finding, followed to its proof — The technical audit’s findings now live in a public register, every closure or remediation with its proof: eighteen are remediated by this version and await the next audit’s verification.
v3.5
v3.5.1 — 🔁 Notifications that no longer repeat themselves23 October 2026
- 🔁 No more repeated notifications — Before writing to you, LIA again reads every notification it has already sent you: a birthday is announced once, not at every check.
- 📱 Your SMS, treated like your emails — A notification that drew on your SMS now shows its preview on your phone, can reach Telegram and the radio, and LIA remembers it — just like one drawn from your emails. A discussion that touches a sensitive topic stays private. To hide every preview on your lock screen, use your phone’s own setting.
v3.5.0 — ⚡ Lighter follow-ups, a simple start, clear hand-overs23 October 2026
- ⚡ Follow-ups that cost less — When your instance runs the ReAct mode on a Claude model, your next message reads the conversation back from the provider’s cache instead of sending it all again — and with frequent exchanges, that cache waits up to an hour for you. Each write is billed at its real price and counted in your usage.
- 🌤️ A simpler start — On an instance that provides the weather, a new account opens on the weather card alone, with a short checklist that asks for your home address first; switch the other cards on whenever you like with the Customize button above the grid. New accounts also start with occasional exchanges, which you can change in your usage preferences.
- 🔁 Who held the ticket, who holds it now — A ticket’s history reads “Held by: Me → LIA” for every hand-over, including when LIA hands a ticket back or a connection ends.
- 📱 Made for your phone — Tables, cards and code use the whole width of the conversation, an empty Home card shrinks to its sentence, and the buttons of a relationship card span the screen.
- 🗺️ Places and previews — On a wide screen, Position lists the places beside the map and shows the map’s cost right under its button, and a place’s details stay in view. In Relations, e-mail previews show their real characters, on up to four lines.
- 🛠️ For administrators — Two switches govern the Claude cache’s hour and the reuse of earlier reasoning, ReAct calls now appear in every model metric, an alert warns when a cache write arrives without its lifetime, and an overloaded provider is no longer read as a model error nor retried twice over.
- 🔍 The audit’s counter-review — The October technical audit was re-examined point by point against the current code: 7.8/10 over 24 areas instead of 8.0, with 12 new findings and a prioritized remediation plan, all published.
v3.4
v3.4.0 — 💬 LIA speaks first, your SMS join in, places show the way22 October 2026
- 💬 Proactive discussions — When you allow it, LIA can start a personal conversation of its own accord: take an interest in a plan, encourage you, ask an open question. You choose the hours, a daily maximum, the allowed sources and the sensitive topics, react in one click, and see — or correct — what LIA understood. Each run counts in your usage and quotas.
- 📱 Your SMS in LIA — An iPhone Shortcuts automation sends your SMS to a private, encrypted archive (Settings › SMS): search, sender names from your contacts, deletion. Recognised access codes are never stored, and LIA reads your SMS only if you give it access separately.
- 🗺️ Details and a route for each place — In Position, each place carries one number, the same in the list and on the map. Select it to see its address, rating, phone, website and opening hours without a new search, with a walking or driving route from where you are. Each route is billed and counted in your usage.
- 📇 Act from a relationship card — One button per email, address and number prepares the message, the directions or a call by LIA in the chat; on the contact card, an email opens your mail app and an address opens Google Maps.
- 🧭 Clearer screens — Alerts and Activity share one header, panels share one neutral line, relationship items are dated like emails, the Workboard card counts the tickets you hold and the journal’s last intervention details its cost.
- 🛡️ Stronger protection — For administrators: an address that scans the server or keeps failing to sign in is blocked for a while on every worker, every refusal for too many requests names its address in the logs, and the connectivity check tries again before reporting an outage.
- 🔧 Fixes — Handing a ticket to a connection no longer freezes your account, a skill shared with you installs on its first receipt, the Home SMS previews stay right after a deletion and the journal’s search hint is translated.
- 🔍 A fresh public technical audit — The code was re-examined area by area: 8.0/10 over 24 areas, security assessed separately. The report also publishes what remains to be done; the two fixes it asked for in the interface — the keyboard focus after deleting SMS, the settings search by a pane's full name — are in this version.
v3.3
v3.3.0 — 🔎 Find, share and locate19 October 2026
- 🔎 Search meeting minutes — Find a meeting by its title or the full content of its minutes. The list and page count follow your search.
- 🤝 Share useful capabilities — Send a skill, MCP server or plugin to an accepted connection after reviewing the package. They receive their own disabled copy, with your name and the sharing time. Credentials stay private.
- 🏷️ Clearer email and appointment labels — When enabled by the administrator, JEV can distinguish work, personal, spam and phishing/scam emails, plus one priority: normal, urgent or critical. Appointments can be work or personal. Uncertain labels stay hidden; the assistant receives the same accepted assessment.
- 🗺️ Places together on one map — Position comes before the place cards, including in HTML and Markdown. Activate the interactive map with a click after seeing its cost; unavailable maps keep a usable place list.
- 🧠 More useful JEV assessments — The questions now distinguish a reported incident from a verified one, and a security topic from an actual deceptive request. Critical cases and spam are better covered in the tested corpus without lowering alert thresholds. A missing badge still says nothing about safety.
- 🔧 One place to configure models — Model choices and generation budgets belong in LLM Configuration. Obsolete environment options and misleading troubleshooting advice are removed.
- 🛡️ Instance-specific provider settings — Fresh installations no longer inherit a maintainer’s OpenAI organization or Firebase project. Android push is offered only with a complete configuration.
v3.2
v3.2.0 — ❓ A FAQ for every feature, voices that react18 October 2026
- ❓ A FAQ for every feature — Ten new sections — the home page, Telegram, calculations and scripts, habits, generated files and kept answers, the personal radio, LIA’s psyche, the transparency registers, administration, self-hosting — and every other one completed: 491 questions, in six languages.
- 🎙️ Radio voices that react — In a discussion, a debate, the editorial, the opening or the sign-off, a chuckle, a laugh or a sigh may open a turn — never during the news, and only where the voice can perform it. A four-voice debate plays in smooth takes instead of falling back to line-by-line reading.
- 🗣️ Cues that fit — When LIA comments aloud, a cue colours the news, never how its own act went: a confirmation, a failure or a search that found nothing is said plainly. Every engine gets the expressions its vendor documents — pauses, a sentence’s tone, concern played as concern.
- 🎚️ A switched-off feature is no longer offered — The tools of a capability the administrator switches off leave both execution modes, Pipeline and ReAct, a delegated sub-agent and a skill’s ready-made plan included; the Workboard, LIA connections, MCP servers and Skills switches close only their pages. With bookmarks off, the bubble no longer offers to keep an answer; what you already kept stays readable.
- 🙂 The avatar and the voice follow your settings — With “Show emotional avatar” off, every answer returns to a neutral face, past ones included; the mood shapes the voice only when the psyche is on.
- 📏 A budget the model accepts — A role’s output budget is held to the cap its model accepts: a budget kept after a model change no longer makes document generation fail. In LLM Configuration, a badge and a warning show the administrator a budget above that cap.
- 🏠 Reconfigure without reinstalling — On a self-hosted instance,
./install.sh --reconfigurechanges the running installation: Enter keeps each answer in force, every setting the new answers would rewrite is shown before anything changes, and the previous state comes back if the new configuration does not start. The Google Gemini key becomes optional, and the final report names what still waits for a key. - 🔧 More reliable behind the scenes — A proactive notification is no longer lost to an unknown label; searches by meaning read the Gemini key entered in LLM Configuration, with no restart; an Ollama server that is off no longer slows every administration save.
- 📊 Supervision that tells the truth — The latency alert only measures the turns a person waits for, and an incident whose resolution was lost closes itself.
- 🌐 Accurate labels in every language — Chinese and German labels made precise, FAQ paths true to the screens, the last English labels translated, the informal register restored in French, German and Chinese, and nine messages that point to where their setting now lives.
- 🧮 The pricing workbook comes back in every language — A pricing workbook exported in German, Spanish or Italian imports back as it is: its true/false columns used to be refused.
v3.1
v3.1.1 — 📱 LIA's eyes stay where you put them17 October 2026
- 📱 On a phone, you place LIA's eyes — They no longer step aside or shrink to a dot on their own: drag them or minimize them, and they stay put while you scroll.
- 📻 Quieter music under the radio's voices — The station's music now sits lower under every programme, so the voices come through more clearly.
- 🔒 Image library patched — The library LIA uses to read your pictures, attachments and scans moves to a new version that fixes twenty-one security flaws, one of them critical.
v3.1.0 — 📝 Your drafts, five voices, an experience built on evidence17 October 2026
- 📮 Your mailbox drafts, from the conversation — On Gmail, Outlook or iCloud, LIA saves a draft, prepares a reply or a forward, lists, edits and trashes drafts without asking. Only sending asks for a confirmation, on a card showing the draft as your mailbox holds it.
- 🐟 Fish Audio, a fifth voice engine — Speech synthesis, transcription for voice mode and meetings; administrators clone or design voices in a library, after a consent attestation.
- 🗣️ You choose LIA's voices — A male and a female voice among those of the engine that speaks, kept when the administrator changes the model; every voice list can be narrowed to your language and the multilingual voices; every voice you pick, for LIA or for the radio, plays a short sample right away.
- 🎭 Voices that perform — Sighs, laughs or whispers where the engine can perform them; the radio plays its debates in one take, and you give each of its commentators a voice.
- 🧭 An experience learned from evidence — Per domain, LIA keeps what its answers truly established and the usefulness you gave them (👍/👎), separately. Where evidence is thin, it checks, asks one question or names its limit. See it in “LIA’s psyche”.
- 🏠 A home page that opens on your day — The date as the title, LIA's living face, a timeline agenda, an illustrated weather, the tickets held with your connections and your consumption in a single table.
- 🗓️ Clearer meetings, calls and alerts — Meetings and calls on a timeline by day and part of day; alerts in two groups, with two-button proposals; an error is never shown as an empty list again.
- 🎨 Redesigned settings — Panels that alternate their depth, buttons that keep their label while waiting, absolute black as a fourth choice, an avatar saved as you choose, a display-mode preview and generated documents that show their first page. The “Why does LIA think this?” block leaves the memory, interest and journal lists.
- 🧩 What a search found reaches the answer — In Pipeline mode, the durations, connections or next pages a tool read now reach the answer; a tool's refusal is never presented as a success.
- 🔧 More reliable behind the scenes — A delegated expert delivers even at its limit, a web search waits out a quota instead of failing, a routine never asks an invisible question nor repeats what it did, and an e-mail is sent only once.
- 🎭 Personalities get their character back — On instances where it had faded, each of the fourteen personalities regains its emotional profile; an account that chose nothing follows the default personality.
- 🔐 Security — Mail servers' certificates verified, third-party content neutralised in every spelling, no key echoed by a third-party service reaches an answer; signing a device out also stops its notifications.
v3.0
v3.0.1 — 🪞 A portrait that reads everything, lighter exchanges13 October 2026
- 🪞 A portrait fed by everything LIA keeps about you — Memories, interests, habits, debriefs, open loops, tickets, routines, reminders and kept answers are read whole; a source read only in part is flagged under the portrait.
- 🧠 A real analysis, a real summary — The full portrait reads you on seven facets instead of drawing up an inventory; the brief is a two- or three-sentence synthesis, not an extract.
- 🕰️ A portrait that no longer disappears — When your journals change, it stays on screen with a note that a recompilation is pending; deleting entries withdraws only what they supported.
- ⚡ Lighter ReAct turns — Fewer tools loaded on each turn, no more pointless recovery pass over a page a tool deliberately capped, smaller conversation checkpoints.
- 🧮 An instance that uses less memory — Each API process keeps its tool data in a compact form and reads MCP catalogues only once; the public demonstrator runs on two processes.
- 🔒 Dependencies fixed — Twenty-three security advisories handled, each fixed or accepted with a written reason.
- 🎬 Landing page and self-hosting — The historical films stay in the playlist; the API image keeps what its system tools need; the use-case times stay readable in the light theme.
v3.0.0 — 🚀 Practical commands, clear outcomes12 October 2026
- ⚙️ Commands for your settings and resources — Labels match the interface, with direct reads and confirmation for sensitive operations.
- 🔐 Private browser continuations — Files, account connections and permissions stay in their secure interface; authority and version are checked again before writing.
- ✅ Explicit outcomes — Recorded operation, waiting, partial result or conflict; typed references, contextual dates and declared weather units.
- ♻️ Reads reused within one request — ReAct reuses exact fresh reads within one request; changes invalidate affected observations.
- 🔌 Current MCP contracts and separate identities — Discovered schemas, structured results and delegation targets checked before admission.
- 🎨 Original native companions — Animated catalogue, contextual expressions and approximate volume, with reduced or static display.
- 🗂️ Clearer answers — Grouped trailing cards, details tied to their source and improved chip-icon contrast in light mode.
- 📻 Radio and voice — Priority for the first programme and sentence-tag cleanup; network approval restored for scheduled tasks.
- 📓 Journals and routines — Consent revisions and execution reservations; optional JEV decisions with bounded diagnostics and accounted fallbacks.
- 🎬 Presentation in French — Uses and faithful demonstration screens, original generated music and a dedicated player that pauses the background.
- 🏠 Self-hosting — SOPS secrets retain UTF-8 on Windows, and a sandbox is reported unavailable if Docker refuses execution.
v2.6
v2.6.1 — 🎛️ Continuous speech, a controllable avatar and an interface within reach6 October 2026 · 20:56 CEST
- 🎭 A visible stop control for the avatar — The window keeps its stop button even when the chat encounters an error. It shows whether closure is confirmed or still pending; an open session can consume your Simli plan, including during silence.
- 🎧 Continuous Live answers — Long answers keep their audio segments in order. The avatar follows playback without repeating a phrase after an interruption; if output fails, the next answer can return to local speech.
- 📱 Floating controls within reach — The avatar, animated companion and shortcuts adapt to the visible area when the keyboard opens, you rotate the screen or zoom. These temporary adjustments preserve your chosen position.
- 🖼️ The current interface in pictures — The gallery and documentation show current screens with fictional demonstration data. Updated screenshots load with their own revision to avoid displaying an old image. The expanded viewer fills the screen on desktop and mobile.
- 📻 A steadier start for Radio — Music playback starts more reliably. You can minimize or expand the player, resize the window while listening, and stop Radio using its usual controls.
v2.6.0 — 🗣️ A speaking face, faithful cards and every paid attempt accounted for5 October 2026 · 06:58 CEST
- 🎭 Speaking avatar (experimental beta) — Connect your own Simli account and explicitly enable an avatar in Settings. It accompanies voice comments, Live and Live direct; Radio keeps its own player. The floating window moves by dragging or arrow keys and offers three sizes. Speech-driven facial realism and physical Android/iOS behaviour still need qualification.
- 🔑 Your avatar, your plan — The Simli key uses the encrypted connector store. A connected avatar consumes your personal Simli plan even during silence between answers. Live standby closes its connection; waking opens a new one. Disable the avatar to end its connection.
- 🗑️ Delete an e-mail from its card — Prepare a deletion for the exact original message and mailbox, with LIA's normal confirmation. Replying and forwarding keep the same binding: a change of mail provider cannot redirect the action to another mailbox.
- 🌦️ Cards faithful to their sources — MCP results retain arrays and simple values, with the tool, method and source URL when supplied. Weather icons reflect the actual condition and day or night; dates and figures follow the display format. Shared, bounded cooldowns avoid repeated MCP calls after a server rate limit.
- 📌 Standing instructions kept in place — A newly stored instruction about how LIA should work is pinned immediately. Automatic edits and cleanup leave it alone; you can still edit, unpin or delete it yourself.
- 📻 Radio with less commercial noise — Advertising, sponsorships, offers and purchase pitches are filtered while real invoices, reservations and editorial, scientific or economic news remain useful sources. Raw PCM and μ-law voices are properly wrapped before mixing with the programme's audio.
- 🧾 Every paid attempt counted — Paid failed or cancelled model calls, background work and cache tokens reach the same usage ledger as successful replies. Prices are frozen for the operation so retries and optional decisions use the same rates.
- 🎛️ Two more optional JEV decisions (administrators) — Detect a reference requiring a memory lookup or a refusal of the whole proposed action. Each use starts off and keeps the ordinary fallback; JEV never approves an action. Decisions keep the exact authorised registry snapshot and accounted costs.
- ✨ A clearer introduction — The landing page's feature catalogue carries the examples without a duplicate use-case section. Help now explains avatar activation, personal billing, card actions and protected instructions in all six languages.
v2.5
v2.5.0 — 🌌 Cards that show everything, an app in the colours of the cosmos4 October 2026 · 04:23 CEST
- 🗂️ Cards that show everything the source supplied — E-mails, appointments, contacts, tasks, files, places, routes, weather, research, lights, tickets and MCP results: every card shows what was received, never inventing a missing value, and keeps the complete detail — an e-mail’s body, every organisation of a contact, conference links, subtasks, a place’s reviews and opening hours — behind a disclosure you can use from the keyboard. Opening a detail triggers no call and no cost.
- 🖼️ Galleries, previews and time slots — A place’s photos open in a full-screen gallery with their attributions, the weather can be compared slot by slot, and a generated document previews in place: a PDF’s first page, the first rows of a spreadsheet or a Word document, a presentation’s cover.
- 🗺️ Interactive routes — A route card can open as an interactive Google map: every alternative drawn, the one you pick highlighted, roadmap, satellite, hybrid or terrain background. The estimated cost is announced before it opens, then counted in your usage. (Administrators: the Maps JavaScript key and its price must be configured.)
- ✍️ Reply from a card — Replying to an e-mail, forwarding it or adjusting a reminder prepares the request in your message box, with a chip reminding you what it is about. Nothing is sent without you, and LIA checks again at the right moment that the account and the original message really are yours.
- ✨ The app in the colours of the cosmos — The landing page’s animated background now extends to the signed-in pages and to sign-in; the header and panels stay solid so they remain readable. Dialogs, menus, lists, tooltips and notifications take on a frosted glass that turns opaque again if your system asks for less transparency.
- 📋 A reorganised dashboard and settings — A new Workboard card shows what is waiting on you, what LIA is holding and what is overdue, each figure leading to the filtered board. Settings gain icons and folding sections: connectors, health, sandbox, MCP servers, calls, skills, knowledge spaces, habits and meetings.
- 🔖 A kept answer says what it cost — Every kept answer shows two costs: what it cost to produce and what it cost to index. The indexing cost of the documents in your knowledge spaces is now the one actually billed: it was underestimated about sevenfold.
- 🎬 Two videos on the landing page — The landing page now plays two videos in turn, each with its credit, and a “Play the next video” button skips from one to the other.
- 📈 Monitoring where everything answers each other (administrators) — In Grafana, a log line opens its trace, a trace its logs and metrics, a latency point the example that produced it. Every dashboard opens on seven days (logs and traces on four hours) with a shared menu, and an alert warns if logs get lost on the way.
- 🩹 Fixes — Your Google account photo shows in the chat again. A sentence with nothing to pronounce is no longer sent to speech synthesis, which billed it and then refused it.
v2.4
v2.4.0 — 🛡️ Updates that never go backwards, telemetry that names no one3 October 2026 · 03:30 CEST
- 🛡️ Updates that ripen before they come in — Every week, LIA re-reads the security advisories each of its dependencies publishes, including those no public database relays. A new version waits out a cooling-off period before it comes in, an update never moves anything backwards, and every building block of the server is pinned by its digest.
- 🔒 What LIA fetches stays bounded — An MCP server you add can no longer make it download an address of the server’s choosing, nor redirect it to another site. A compressed page is read under the same size limit as any other, and autonomous browsing runs on the system’s own Chromium, updated with it.
- 🕵️ No name in the technical measurements — The metrics and traces used to monitor LIA name the kind of page requested — “a relationship’s card” — never its content: a name or a search sitting in an address no longer gets into them.
- 🎙️ “Dis LIA” heard more often, “LIA, stop” to cut her voice — The wake word was retrained to miss a quickly spoken “Dis LIA” less often, at the cost of a rare trigger on a similar phrase. To cut LIA’s voice while she reads an answer, now say “LIA, stop”. Still in French, and in beta.
- 🎙️ Telegram voice notes: one limit, the right one — A note longer than the transcription accepts is refused with its maximum duration, in seconds, instead of an “I didn’t understand”. It is processed faster too.
- 💬 “Back to the bottom” always within reach — The button that takes you back to the latest answer sits just above your input box, instead of staying at the very bottom of the conversation, off screen.
- ✨ A home page that breathes, and pauses — An animated background links points of light to the rhythm of the video, the footer becomes a pane of glass resting on the planet, and the video melts into the page. A “Pause the animations” button, on every public page, freezes the whole backdrop, and your choice is remembered. Secondary text gains contrast too.
- 📈 Renewed monitoring (administrators) — Prometheus, Grafana, Loki, Tempo and their companions move to supported versions, pinned by digest. Grafana Alloy replaces Promtail for the logs: the same redaction, half the memory, and no usage statistics sent to its vendor. Checking the backups by a full restore works again.
- 🛠️ DevOps: its own model, effort and token (administrators) — The DevOps assistant uses the model and effort you set (
DEVOPS_CLAUDE_MODEL,DEVOPS_CLAUDE_EFFORT) and signs in with its own long-lived token: create it withclaude setup-tokenand setCLAUDE_CODE_OAUTH_TOKEN. The deployment script no longer copies your Claude session to the server. - 🩹 Fixes — Mathematical formulas are laid out correctly again. The personality picker offers a full touch target in Safari.
v2.3
v2.3.0 — 🎙️ A Live session that sleeps instead of ending, "Dis LIA" in your browser2 October 2026 · 01:15 CEST
- 💤 A Live session goes on standby — After a silence, when the page stays hidden or at your button, the session falls asleep instead of ending: the connection to the provider closes and nothing is billed. "Wake up" resumes it with LIA's context; only you decide to end it, and the duration and the meter count awake time alone.
- 🎙️ "Dis LIA", heard in your browser — The wake word is now a small model running on your device, iPhone included: no sound leaves before it is heard. Say "Stop" to cut LIA's voice while it reads an answer. It exists in French for now, in beta: its recognition is still improving. In another language, tapping the badge is still there.
- 🔔 No more daily notification limit — LIA evaluates every pass of your time window and only speaks when it is useful: an awaited message or the end of a meeting is no longer refused because a maximum was reached. The "per day" settings are gone; the window and the pauses between two messages stay.
- ✉️ Your contacts suggested in "Send by e-mail" — Type a last name, a first name or a phone number: LIA offers the matching contacts and inserts their address. An active contacts connector is required.
- 🔌 Reconnect everything from the alert — When two or more services of one Google or Microsoft account have expired, the connection alert offers to reconnect them all in a single authorization.
- 🛠️ A clearer user table — The administrator sees each switch of an account in its own column, with a legend, the sign-up date and sortable headers; the e-mail and the name stay visible while the table scrolls.
- 🎬 A landing page that moves, a faster blog — A video follows the visit from page to page and the titles beat to its music, unless you prefer less motion. The blog's illustrations arrive without waiting.
- 🩹 Fixes — The hands-free badge turns back on with a click right after being turned off (a message was covering it). An instance set above 20 MCP servers per person starts again.
v2.2
v2.2.0 — 🧩 A skill library, skills that act, kept apart30 September 2026 · 22:50 CEST
- 📚 Find skills in a public library — “Find skills” in Settings › Skills searches skills.sh or reads a GitHub address. You read a skill before installing it: its source, its exact version, each of its files and the published audits. The Installed tab tells you when its source has moved and shows you the update before you apply it.
- 🧰 Skills that run their commands — A skill can run a Python, Node or shell script in a throwaway copy of its folder, with the files attached to your message. What it writes comes back to you as generated files, checked before they reach you and kept in your gallery.
- 🌐 The network, only to permitted hosts — A command is offline by default. One that must install a package or fetch a repository names its hosts and goes through a dedicated proxy, once they are allowed. In ReAct mode LIA asks you; you can also allow a host in Settings › Sandbox network.
- ✅ A skill written by LIA, installed by you — When LIA creates or changes a skill in the conversation, it proposes it on a card: name, description, files and, for a replacement, what changes. Nothing enters your skills before you click Install.
- 🛡️ Skills from elsewhere stay apart — A skill from the library, an address or a plugin runs isolated: never your mail, your calendar or your connectors, and no images or web content in its answers. And a skill name is now unique for your account alone.
- 🤖 Two new models — Claude Sonnet 5.5 and GPT-6.1 Sol join the administrator's catalogue, with their reasoning settings and prices.
- 🔒 Better-protected exports — A Markdown export can no longer turn text into a booby-trapped link, and an event's description is cleaned more reliably.
v2.1
v2.1.1 — 🔠 Text at your size, readable exports, accurate weather alerts30 September 2026 · 03:19 CEST
- 🔠 Choose your text size — In Settings › Preferences › Personalization, set the text from 14 to 20 px with a live preview. Only the text changes: panels, spacing and icons keep their dimensions. Your choice follows you on every device.
- 📄 Readable downloaded answers — Download, Send by e-mail and a bookmark's export now produce real Markdown: headings, lists, tables, links and code blocks preserved, cards summed up as headings and bullets. Copying and sharing stay plain text.
- ✉️ Complete e-mail cards — In a list of e-mails, the cards show the message and its attachments again, even when LIA only needs the headers to answer you. A Gmail list also takes fewer calls.
- 🌦️ Weather routines that keep their word — A routine watching the weather always reads Google Weather and fires only for a change due within the next 4 hours, with more than a 50% chance of precipitation. LIA is told the hour, the day, the chance and the source. The weather card now only announces the next 24 hours.
- 🛡️ Safer web reading — When LIA reads a page, every redirect is checked before it is followed and the connection goes to the address that was checked: a page cannot lead it into your internal network. A booby-trapped text is formatted without slowing the server down, and a production instance refuses to start in debug mode.
- 📻 The radio follows your interests more closely — When a session starts, up to five of your interests are searched with your own key, and every story found can go on air. When you listen in company, nothing about your interests is broadcast.
v2.1.0 — 🧭 Clearer choices, a closer look at LIA, steadier speech29 September 2026 · 12:08 CEST
- 🧭 Explore LIA in action — The renewed home page and interactive demo show how research, approval and results fit together. Radio settings make programmes, sources, voices and spending easier to understand.
- 🎛️ Native decisions under admin control — Administrators can enable JEV separately for bounded tasks such as meeting formats, result previews and radio verification. Uncertain answers fall back to LIA's existing process; permissions and confirmations remain in place. Calls and costs appear in diagnostics.
- 🔊 More reliable speech — ElevenLabs synthesis shares a capacity limit across workers and voice features. Audio plays in order even when segments finish at different times.
v2.0
v2.0.0 — 📻 Your own radio, files you can keep, and a conversation that stays current28 September 2026 · 12:35 CEST
- 📻 Your personal radio — Start it when you want: a spoken programme about your day and cited news, in your language. Choose subjects, sources, voices and a verification level; see its spending as it plays. It stops on your timer and is available only when your instance enables it.
- 📌 Keep the files that matter — In My generated files, keep selected images, documents and screenshots beyond their usual deadline, within the account limits shown there. Stop keeping one and it receives a fresh deadline.
- ✉️ Send a file or answer by e-mail — Add a subject and optional note. A connected mailbox can address chosen recipients; without one, LIA can send only to your verified address. The form shows the size limit before sending.
- 🔄 A conversation that follows you — Messages from another tab, device, call or routine appear without reloading. Older messages and your reading position stay in place, and the connection catches up after an interruption.
- ⏱️ One clock per routine — Choose a schedule or a condition. LIA checks a condition on its own cadence and acts when it matches; the studio shows only the fields that apply.
- 🌍 Six languages throughout — Interface text, notifications and server messages use the account's declared language, with completeness checks across all six translations.
v1.47
v1.47.4 — 🧮 LIA computes instead of estimating, searches what it knows and shares your images25 September 2026 · 04:51 CEST
- 🧮 Exact figures, never estimated — a total, a percentage, VAT, a unit price: LIA computes them exactly instead of estimating, and says “about” when a result had to be rounded. Dates too: days until a deadline, the weekday of a date, working days (public holidays not deducted, and it says so), the time in another timezone. And a currency conversion uses the European Central Bank's reference rate, with the day of that rate — a bank applies its own.
- 🔎 LIA searches what it knows — beyond what your message evokes, it can search its memories and its own journal for a subject found along the way: the sender of an e-mail it just read, a project, an instruction you once gave it. On the phone, the same search finally answers the question asked instead of returning the latest memories.
- 🗂️ “What did you do for me this week?” — LIA answers from its own registers: the actions it took, their outcome, what it consulted, with exact totals even when the list is shortened. And “show me yesterday’s image again”: it finds the files it produced — images, documents, screenshots — and shows them again as cards, as long as they exist.
- ✉️ An e-mail to yourself, without confirmation — “e-mail me that” goes straight to your own mailbox, since you are the recipient; without a connected mailbox, LIA sends it to your account's verified address. An e-mail to anyone else still waits for your approval.
- 🎨 Better-written image descriptions, if you want — the “Enhance prompts automatically” option of the image settings lets LIA rewrite your description following the provider's own advice before generating: a text you put in quotes is kept, and at the slightest doubt your original description is sent as is. Your gallery keeps your own words.
- 🤝 Share an image with a connection — from an image's card, in the chat or the gallery: it lands in your connection's gallery and chat, with your comment quoted exactly as written — never turned into a link or a hidden image.
- 📤 Share and download in one click — under every answer, the Share and Download icons replace the “…” menu. Share opens your device's share sheet and, if you have connections, offers to pass the answer on to one of them; LIA prepares the message and you confirm it. And “Actions performed” no longer cuts a subject or a reminder after a few words.
- ⚙️ For administrators — an announcement addressed to some users no longer appears for everyone at their next sign-in, an empty selection is refused, and the history of sent broadcasts shows who each one reached and until when. The debug panel names the model each call really used and the role that made it, and resizes with the mouse or the keyboard.
- 🔏 Logs that keep facts, never your words — the server's technical logs no longer keep your searches, your interests or the names of your contacts or files: only counts, lengths and identifiers. When a database quotes the data it refuses, the quotation is removed before the line is written.
v1.47.3 — 🎚️ At your pace: you choose how LIA prepares your exchanges24 September 2026 · 14:59 CEST
- 🎚️ An exchange rhythm of your choice — a new Usage preferences section under Settings › Preferences › Personalization, for the ReAct mode. Frequent exchanges: LIA keeps every tool at hand and prepares each message so the next one re-reads what was already sent at a reduced rate — the first message costs a little more, the next ones much less. Occasional exchanges: it picks the tools each question needs, at a roughly constant cost. Your routines and Workboard tickets follow the same choice; until you make one, your instance's default applies.
- 💶 Long turns that stay economical — while a long answer or a routine chains its tools, the conversation history shown to the model no longer shifts mid-turn: what the model has already read keeps being re-read at a reduced rate instead of being billed again in full.
- 🔏 Your words stay out of the logs — the server's technical logs no longer record the content of a tool result or the answer you type to a confirmation question, only their length.
v1.47.2 — 🎯 What LIA sets out to find, it finds or tells you what is missing — and every cent at the real price24 September 2026 · 05:18 CEST
- 🎯 Answers judged on their result — in ReAct mode, LIA checks that each result really answers your question (the right day, the right place, a complete list), tries another source in front of an obstacle, and ends by saying what it could not obtain and what it tried; a fact found elsewhere carries its source. “Tomorrow” or “Saturday” become an exact date before a service is queried, in your language.
- 🧾 A failure is said, never guessed — when a service fails, LIA names which one and what it answered instead of inventing a “service not configured”, and a half-successful plan gives you back what worked. The actions it really performed — an image, a reminder — appear under the answer, and it no longer denies them.
- 🎨 Images: two providers — your administrator chooses OpenAI's GPT Image or Qwen Image 3.0, for creation and editing alike. Your quality and size preferences adapt to what the model can do (never a dearer option), and the format you chose — PNG, JPEG or WebP — is applied to every image.
- 🔌 Nothing to switch on for keyless services — Wikipedia, page reading, Google Places, Google Weather and air quality are provided by your instance: always on, they leave “My connectors”. An OpenWeatherMap key you add still takes precedence for the weather.
- 💶 Every call at the price the provider bills — tariffs re-read on the providers' pages, DeepSeek weekends at off-peak rates, a write to a model's cache at its real price, routes at the tier their request triggers; a price the administrator edits reaches every server at once. New models in the catalogue: Claude (Fable, Opus, Sonnet), GPT-6, Qwen 3.6 to 3.8, Gemini 3.8 Flash.
- 🔐 Safer and steadier — a Google sign-in can no longer activate a pending registration or unblock an account, and the real address of every call is recorded. No transaction stays open while a service answers: proactive notifications, reminders and Drive sync stay regular. In a Live session, an ElevenLabs agent goes through WebRTC on iPhone and iPad, and a Gemini key restricted by IP address is reported.
v1.47.1 — 🔗 One consent for your Google or Microsoft services, and a companion that follows real work22 September 2026 · 06:48 CEST
- 🔗 Connect several services in one go — choose the Google or Microsoft services you want, then authorize that account once. Reconnection follows the same path; the account is verified before LIA links any service.
- 🎛️ Keep control of each service — disconnect mail, calendar, contacts or tasks separately without cutting the others. Existing connections keep working until you choose to reconnect them.
- 🙂 A more attentive companion — its expressions follow actual work and the tone of LIA's delivered answer. Local time and already-cached weather add a quiet touch; reduced motion keeps it still.
- 🛠️ More reliable replies — tool and skill matching is more precise, and a prepared action no longer looks like a completed one. Administrators can follow the new tool metrics in the observability dashboards.
v1.47.0 — 🎧 Live mode: talk with LIA in real time, on your own key — and one policy per mode, on the phone as in the browserSeptember 2026
- 🎧 A Live session, voice to voice — the voice icon in the header now opens a real-time conversation with LIA on a live model you connect with your own key (Gemini Live, GPT-Live or an ElevenLabs agent). You talk, LIA answers, you can interrupt; captions and an indicative meter of what the provider is charging you sit in a band above the thread. Nothing the provider bills is recorded by the platform.
- 🧠 Two intelligences, one seam — the voice holds the conversation and hands every request for data or action to LIA, which runs it as an ordinary chat turn: the same confirmations, the same registers, the same quotas, drawn in the thread while you speak. A question LIA needs to ask you IS the answer the voice gives. The closing card adds up what LIA itself spent.
- 🎯 A direct session, and the phone's Live mode — a third entry of the voice menu, « Direct live session », lets the voice read your data itself (agenda, mails, contacts, places, memories — the domains you allow) without acting on anything; what you asked is relayed to the conversation at the end as a message from you. The same two modes now govern LIA's calls to you: Live (the default — each request is handled while you speak) or Live direct, chosen in Settings › Telephony · My identity.
- ⚙️ Settings › Live mode — the model (choosing it is choosing its provider), the voice with a sample on every change, the thinking level where the model offers one, how the conversation behaves (interruptions, end of speech, when LIA delivers an answer), the silence and the duration cap per model, an optional spend ceiling per session. Every model you connect keeps its own settings.
- 💶 Every euro the platform pays for you is counted — whatever the path — a map lookup during a call or a live session, the weather of the morning briefing, a geocoding, a photo LIA shows you: each one now lands in your usage and your cost tiles like a chat message does, and the phone bill and the live card show the Maps euros they used to count as mere requests. A voice synthesis is charged only for the sentences actually served.
- 🔧 Also fixed — the tool selection and the ReAct ranking had gone blind on an embedding cache written at the wrong dimension; the semantic validator had skipped every turn since a September change; an ambiguous answer to a confirmation was read as an edit; the phone's Live call nobody answered left no notification; an open tab held a database connection for its whole life; the dev certificate was renewed on its age rather than its expiry.
v1.46
v1.46.0 — 🧰 The sandbox becomes a toolbox: a script reaches the web through one door, and asks beforeSeptember 2026
- 🌐 A script LIA writes may reach the web — in ReAct mode, when a step needs a service no tool covers — checking that an API answers, reading a feed, calling an endpoint you use — the few lines of Python LIA writes may now reach it. Never directly: through one dedicated proxy, HTTPS only, to the hosts the script declares and nowhere else; a script that declares no host stays offline. It remains a complement — most questions never need it.
- 🔑 Your keys never enter the script — the services you connected with your own API key (web search, answer engine, weather) are reachable without a question: the script reads a temporary token, and the proxy swaps it for your real key on that host alone. The script never sees the key and cannot send it anywhere else.
- 🙋 An unknown host is asked, with three answers — when a script wants a host that is neither one of your connectors nor on your instance's list, LIA stops and asks: allow with the data of this turn, allow without it (the script then receives nothing of what LIA gathered), or refuse. Then the rest of your request goes on — a question is not the end of the answer — and your answer is remembered for next time.
- ⚙️ Settings › Sandbox network — a new section lists the hosts a script may reach without asking (your connectors, the hosts your administrator allowed) and every permission you gave, each with its data scope you can change and a revoke button. Past a number of permissions, a new one holds for its run only.
- 🧰 A toolbox, not only a calculator — the script now serves four roles: compute, check that a service answers, fill a gap no tool covers, transform data (a feed, a calendar file, an XML, a spreadsheet). Twenty-two libraries are declared and verified in the built image. Every network run is written to your register before it starts, and an administrator can switch the network off from the capability map — the calculator stays.
v1.45
v1.45.2 — 📎 A knowledge document joins a message, a mail attachment is read, a Drive folder is a treeSeptember 2026
- 📚 Attach a document of your knowledge spaces — the « + » of the composer now offers, beside the file picker, every document already indexed in your spaces, a paused space included: search it by name, tick it, and a copy joins the message like an upload — the space is never touched, and a conversation reset removes the copy alone. Fifteen formats, not only PDF and images.
- 📎 LIA reads the attachments of your e-mails — ask « read the attachment of this mail »: a document is read as text (PDF, Word, spreadsheets…), a picture or a scanned PDF goes to the vision model, a long text arrives in parts, and a refusal says why (too large, unreadable, unknown). What a stranger sent stays data, never an instruction.
- 📁 A Drive folder is synced with its sub-folders — a linked folder is now walked as a tree. Past a number of files you set (10 by default), Synchronise shows the exact count of files about to be indexed — new, modified, unchanged, unsupported — and waits for your confirmation. A folder inside an already linked one is refused rather than indexed twice.
- 🔍 Long web syntheses fold, code stays in its card — in the HTML-cards mode, a long search synthesis shows its lead and folds the rest behind « see more »; a code line wider than the card scrolls inside it, with a cue on the right edge.
- 📖 Every read of a space is in your register — linking a folder or a label, a background sync, a push: each now leaves its line in « What LIA consulted ». And iPhone photos (HEIC) are decoded on the server — as an upload, a mail attachment or a picture to edit.
v1.45.1 — 📚 What you keep is searched, your portrait says where it comes from, and every rich answer is a composed pageSeptember 2026
- 📚 Your kept answers are searched — every answer you bookmark is indexed in a knowledge space of your own, Kept answers, created the first time you keep one; weeks later LIA cites it when a question concerns it, and the bookmark card shows the indexing state and what it cost. Deleting the bookmark removes the document; the space itself stays yours to switch off.
- 🧭 Your portrait says its sources — under Settings › Journals, the portrait is now compiled from your memories, your interests, your learned habits and your relationship debriefs, each read under its own switch, and a provenance line under it says which sources were used, which were empty and which were off. A source that moved reopens the next compilation.
- 🎨 Rich HTML is a composed page — in the rich HTML display mode, every answer that carries data is laid out: a lead sentence, one section per facet in the component that fits it (figures as tiles, statuses as chips, procedures as steps, comparisons as columns or a table), and a closing callout when there is advice or a next step. A greeting stays one line. And the layout now applies to conversational turns too — unless your voice is on, since a spoken reply must stay plain.
- 📄 A document that fails to index tells you why — under a knowledge space, a document in error now carries an explanation in your language instead of a technical tooltip. A scanned PDF without a text layer is named as such, with its remedy: put it on Google Drive, open it with Google Docs (the text is recognised), then add that Google Doc to the space through its Drive source.
- ⚡ ReAct binds the tools your question needs — the ReAct mode used to bind every available tool and cut the excess in registration order, so the same tools — your own MCP servers among them — fell off on every turn. It now binds by relevance to the question, keeps every family reachable, and reads the same knowledge spaces the pipeline reads: turns are lighter and a kept answer is found from the loop too.
- 👀 A face that never plays the same thing twice — the brows sit on the eye, have weight, knit and lead each expression; the mouth speaks a generated speech that never loops; no two blinks, smiles or glances are alike; and the short scenes now wait for real resting time on a clock the page keeps across navigations, so you actually see them.
v1.45.0 — 📞 LIA calls you, and what you said comes back as your own messageSeptember 2026
- 📞 LIA calls you — declare your own number under Telephony · My identity, LIA calls it and reads a code you type back, and from then on “call me at eight with my agenda” places a real call to you with no confirmation card — including from a routine. The number is shown whole and never matched by name; a stranger picking up is told nothing.
- 🧠 On the phone, LIA knows what it knows in the chat — with the context switch on, the call carries your memories, agenda, reminders, open threads and latest exchanges; with live lookups enabled by your administrator, it reads everything the chat can read — mails, calendar, tasks, contacts, files, places, weather, memory — and acts on nothing. Each domain has its own switch in My identity, and the assistant speaks with the personality you configured.
- 💬 The call comes back as your message — when you hang up, what you said is replayed in the chat as your own turn, with a phone badge; a draft it produces waits for your confirmation there. If nobody picked up, if the line failed, or if someone else answered, the notification says exactly which.
- 💶 What a call costs LIA — the live lookups, the synthesis and the relayed turn are counted under one bill, shown on the relayed answer and on the calls list. What runs on your own ElevenLabs key — the voice, the line — is billed by ElevenLabs on your account and is never counted here.
v1.44
v1.44.7 — ✉️ Several drafts are reviewed one at a time, and an e-mail arrives clean at three levels of detailSeptember 2026
- ✉️ Several drafts in one request, reviewed one by one — “send an e-mail to X and another to Y” used to show one card and send both on the first click. Now each draft has its own card, its own question and its own edit, LIA says where you stand (“Draft 2 of 3”), nothing leaves before your last answer, and the report says to whom and what was sent — a draft you declined is reported, never lost. A bulk operation you already approved as a list keeps its single confirmation.
- 🗂️ Drafts and reports drawn as cards — in the chat, a draft to confirm and the report of what was done are now real cards, like an e-mail or an event: a header, fields with their icons, the body in a block. A sequence opens on the list of everything the request prepared, before the first card. On a ticket or on Telegram, the same content stays readable text.
- 📬 E-mails read whole and clean — quoted history and signatures leave the body, HTML becomes text for Gmail, Outlook and Apple alike, and a long message arrives in parts rather than cut mid-sentence. LIA now chooses how much to read: the headers to list, the whole text to read one message, or a short digest per message to reason over many — “summarise my unread mail”, “a synthesis of this week's newsletters”. A digest is computed once, kept thirty days and counted in your costs. In ReAct mode, a tool's results reach the model message by message, never cut in the middle of one.
- ⚙️ For administrators — the window a calendar listing covers when no end is named is a setting (
CALENDAR_TOOL_DEFAULT_DAYS_AHEAD, thirty days by default), the e-mail digest runs on its ownemail_digestLLM slot configured from the administration (output ceiling included), and two ReAct settings bound what one tool result may occupy in the model's context (REACT_TOOL_RESULT_MAX_TOKENS,REACT_TOOL_RESULT_WINDOW_FRACTION).
v1.44.6 — 🔔 A reminder arrives with its text, and a prompt states what the code enforcesSeptember 2026
- 🔔 Your reminders arrive with their text — on a model that thinks before answering, the thinking counted inside the message budget and a reminder could arrive with nothing but its bell. LIA now writes its short notifications without thinking, and if a sentence comes back empty or cut, it sends the plain reminder — never an empty notification.
- 🤖 For administrators — DeepSeek's current name —
deepseek-flashis configured from the administration with its full thinking ladder ("low" included), its off switch, its output ceiling and its time-window tariff; the former name is still accepted. An upgraded instance receives the model by migration, a fresh one by the seed. - 🧹 For administrators — a prompt states what the code enforces — the prompt audit corrected what the model was reading wrong: an empty context is no longer wrapped (nearly 250 fewer tokens per bare turn), Brave searches honour the bound they publish, the delegation block and the briefing windows read their numbers from the settings, and every sentence of a prompt lives in a versioned file.
v1.44.5 — 🧭 A demo that says what it offers, a server that says what it holdsSeptember 2026
- 🧭 The demonstrator says what it offers — the page that opens the public demo lists what is switched on there and what is not, under the limitations and before the button. That list is read from the demonstrator's own configuration, never kept by hand; if it does not answer, the page says so instead of showing an empty column.
- 📄 Your documents stay openable — reading, downloading and deleting your own files no longer depend on uploads: on an instance where uploads are off, a report LIA wrote still opens, and its expiry is still swept. On the demonstrator, document generation is now on.
- 🎛️ For administrators — every capability is decided — the demonstrator writes down its twenty-seven ceilings, on or off with the reason; the workboard and connections between accounts are open there. Every instance publishes the effective state of its capabilities in its public configuration, and the page that presents it relays it.
- 🧠 For administrators — memory, process by process — every API worker publishes what it holds in memory: two panels on the infrastructure dashboard and an alert that names the process. Speech transcription keeps a single resident model per worker, the supervisor no longer loads the application, and the database keeps half its memory for its cache. Self-hosters inherit the new sizing without changing anything.
v1.44.4 — 🔖 An answer you keep is yoursSeptember 2026
- 🔖 Keep an answer with one click — every assistant bubble carries a bookmark next to « copy ». Kept once, the answer stays with the request that produced it and its date, even after you reset the conversation: a bookmark is a copy, not a pointer. The icon fills when the answer is kept; a second click lets it go.
- 🗂️ A Bookmarks tab in « My generated files » — newest answer first, a search over answers and requests, the exact total against what your account may keep. Each card renders the answer with its formatting, shares it, downloads it as Markdown with your request quoted on top, or deletes it after asking.
- 🔌 New accounts start connected — Wikipedia, page browsing, Google Places, Google Weather and Google Environment ask nothing of you, so a new account starts with them active instead of five buttons to find. Existing accounts keep their choices; an instance without a platform Google key, or with a connector switched off, skips it.
- 🎛️ For administrators — a « Bookmarks » switch joins the capability panel: off, the bookmark action disappears while what was already kept stays readable and deletable. The capability map counts them, and the account export carries them.
- 📅 Apple Calendar keeps its name — a default calendar chosen by its name was never found on iCloud; it is now, through the same fix that repaired the calendar list.
v1.44.3 — 🧭 A learned habit holds everywhereSeptember 2026
- ⏸️ Pausing or blocking a learned window now holds everywhere — the status you set in the habits panel used to be read by nothing: proactive notifications, their scheduling and the assistant's context all kept following the blocked window. One reader serves all three now, so a paused habit is paused for real.
- 🗣️ A missed-routine offer names the request — "shall I run your usual e-mail review?" instead of a vague "shall I take care of it?". The habit row reads the same way: "Search · E-mails — every day ~08:30", because LIA now keeps what you usually ask for, not only when.
- 🌙 Learned habits have a life — the nightly job promotes a new recurring request, refreshes one that shifted, and demotes one you stopped — never on doubt, and never while you are silent for a day. A habit you resume comes back the same day; "Recompute now" does the same on demand.
- 🔒 One learning switch closes every door — off, nothing is recorded, nothing is suggested, nothing is seeded, and the keys you refused are never proposed again. The administrator's capability switches stop the acts too, while what was learned stays readable and editable: with proactive notifications switched off, their settings and history used to answer with an error.
- 🤫 Interest updates stand aside too — an interest notification now reads your calendar and your learned rhythm like the heartbeat does. A wake triggered by a new mail skips the rhythm but not a meeting, and the rhythm can never hide an imminent appointment.
- 🕒 Notifications read your clock — the decision is written at your local time, not the server's, and reading LIA counts as being there: presence is on by default, so a morning spent reading without writing is part of your rhythm.
- 🎛️ Thirteen sources, each saying whether it is connected — the proactive settings list every source in the server's order with a note when one is not available and why; habits and the workboard have their own icons, and every detector threshold is explained in your language.
- 📒 What LIA consulted, for proactive passes too — the consultation register had never received a single row from a proactive notification, so "what LIA read" was empty for every one of them. Each pass now files what it opened, calendar check included.
- 🔧 Also fixed — the mail-driven wake skipped every other pass; a calendar that failed to answer was recorded as a silence; switching off chat suggestions used to switch off learning; a notification merely labelled "habits" no longer spends the day's offer; and the Apple Calendar list no longer answers "Unable to load the list" — a calendar name was read through the wrong door.
v1.44.2 — ⏱️ Back at the instant that mattersSeptember 2026
- ⏱️ Anticipated moments — fifteen minutes after an important meeting ends, LIA asks how it went: one open question, at most two facts, never a judgement. The moment is re-checked before it is served, two back-to-back meetings earn one question, and the heartbeat's quota still applies.
- 🎛️ One switch per kind of moment — in the proactive notifications settings, beside the source switches. Nothing is offered that your account cannot produce: a kind waiting on a calendar you have not connected says so.
- 🤫 No interruption during a meeting — an ordinary proactive pass now reads your calendar first and stands aside while a meeting you attend is in progress. Being silent is exactly when the old activity rule believed you most available.
- 📅 A finished routine closes on its own — a routine whose series is over (end date reached, occurrences exhausted, single run consumed) used to stay "active" for ever, indistinguishable from one you had paused. It now reads "Finished"; extending its schedule restarts it.
- 👁️ « Watch » on a briefing e-mail — one gesture on a mail card posts a watch for that sender's reply, for two weeks. It refuses to create a second one for the same person, so an awaited reply is never announced twice.
- 🧭 Habit learning finally learns — the ledger of recurring requests had never recorded a single one, in five weeks: two readers looked up a field that did not exist. It records now, an alert would notice if it stopped, and the thresholds were recalibrated on a durable harness: a habit coming several times a week is named as such, a daily one is recognised after two weeks, and every bar you must clear is published in the "why no habit" explanation.
- 🖼️ Gallery thumbnails show the whole image — a portrait picture no longer arrives cropped, and a document card carries the mark of its type instead of a bare line of text.
- 🧭 A long conversation stopped acting after a few days of autonomous mode — the time spent inside tools was added up turn after turn for the whole life of the conversation instead of restarting on each one. After about six days of ordinary use the counter crossed its budget, and from then on every autonomous turn of that conversation stopped before running a single tool: a reminder asked three times, three times nothing, no error. The counter now restarts on every turn, in conversations, routines and tickets alike, and a guard refuses any future counter that would not.
- 🔧 Also fixed — a routine the system closed no longer reads "Paused"; a note about a missing connector no longer claims you switched it off; the heartbeat guide named a model and two wrong defaults, now corrected.
v1.44.1 — 🗂️ What LIA makes belongs to youSeptember 2026
- 🗂️ My generated files — the images, documents and screenshots LIA produced for you finally have a place of their own, in Settings. Three galleries, one per family, with a search on the name, two date filters and four sort orders. Before, the only route to yesterday's report was finding the conversation that wrote it.
- ⏳ You can see the expiry coming, and clearing a conversation no longer takes them with it — every file shows its deadline, whose tone warms six hours before. And "clear this conversation" now removes only what you put there: the images and reports LIA produced stay, including those from other conversations.
- 📌 Your standing instructions can no longer be deleted, only corrected — "call me by my first name", "always answer in French": that kind of rule is protected from deletion, by you and by the automatic housekeeping alike. You can still edit it, which is the real way to change your mind.
- 🎛️ An administrator can switch each feature off, one by one — the panel offered twelve while the product shipped twenty-five: the ticket board, journals, habits, long-term memory, spontaneous notifications and eight more could only be turned off by redeploying. Switching a capability off stops it; everything it already produced stays readable and deletable.
- 🧠 Context size is set model by model — one setting fixed it for every Ollama model at once, so a small model was asked for the same window as a large one. Each slot now carries its own, pre-filled with what your server says about the model.
- 🔬 The debug panel really tells the story of a turn — one line per model call with its provider, its parameters and its reasoning effort, the bounds actually applied to the autonomous loop, the slot's window, and what LIA consulted, decided and checked during the exchange.
- 🩹 Two outages fixed — an internal value was reaching five providers out of seven and killing every call whatever the configuration; and a model its provider had retired told you to "try again" instead of sending you to pick another one.
- 🖼️ Image thumbnails are back — an image LIA produced no longer showed in the thread or in the gallery, while opening it displayed it perfectly. The browser was asking for the thumbnail without your session, the API refused, and an authentication failure read as a display bug.
v1.44.0 — 🗂️ A board for what you are actually doingSeptember 2026
- 🗂️ The workboard — a ticket is a unit of work that has a life: a title, a description, a priority, a due date, sub-tickets, comments and a history. It crosses seven columns, from a plain idea to done. A task list has no lifecycle, a reminder rings once and leaves nothing behind, a routine never ends — this is the surface that was missing between them.
- 🤖 Hand a ticket to LIA and it works on it — LIA takes one ticket at a time, does it on its own, and comments what it found or produced. You get the result to validate, in its own column. If a run is postponed because a quota or a busy conversation got in the way, the ticket says so plainly: postponed is not failed, and the attempt is given back.
- 💬 When LIA needs your agreement, it asks on the ticket — the very card you would have seen in the chat is written there, with what it is about to do. Answer with one comment: yes, no, or what to change. Sending the comment is the answer — the ticket goes straight back to LIA. What you approved is exactly what runs: change one detail and LIA asks again rather than assuming.
- 🔔 It follows up, and Settings shows the board at a glance — an overdue ticket, one due soon, a question left waiting or a result nobody validated comes back in the chat, quoting what is actually waiting instead of paraphrasing it. Settings › Workboard states the totals, the exact count per column and what is on your plate, each one a link into the board already filtered.
- 📌 Pin the settings you actually use — pick up to five sections in Settings › My shortcuts and they follow you on every screen in a small floating capsule. Move it where you like, fold it into a dot, unfold it with one click. What is pinned follows your account from one browser to the next; where the capsule sits belongs to that device.
v1.43
v1.43.2 — 📊 Documents with the craft their format deservesSeptember 2026
- 📊 Documents with the craft their format deserves — ask for a deck and you get a real one: 16:9, a layout per intent — part opener, two-column comparison, native table — and a slide number on every slide. A Word file arrives with its contents, numbered headings and pagination, computed by Word itself. An Excel file arrives with typed columns, a filter and a frozen header. A PDF arrives with exact page numbers in its contents, and bookmarks.
- 📐 And nothing runs off the page — text is measured before it is placed: a crowded slide is shrunk, then split into “Title (2/3)” slides rather than clipped. An over-long bullet is cut at the end of a sentence, never mid-word. Checked by opening the files in Office: zero overflows across 69 slides.
- 🈶 In Chinese, Japanese and Korean too — an ideograph is twice as wide as a letter, which the measurement ignored: a dense Chinese slide overflowed by 124 points with nothing to report it. The running head and footer of a Chinese PDF printed as dots, for want of the right font.
- ✂️ A cut document is no longer delivered as a complete one — when the model hits its length limit mid-write, the result used to be closed silently and announced “generated successfully”, missing its last sections. LIA now tells you it was too long, names the limit, and offers to write something shorter or to split it into several documents.
- ⬇️ Downloads and images that need no relay — a generated document's link, a generated image, a screenshot, and the exports of your interests, journals and memories pointed at the website's address instead of the API's. On an install where the website does not relay the API, none of them opened. Those links now address the API directly.
v1.43.1 — 🤝 A written debrief for every relationshipSeptember 2026
- 🤝 A written debrief for every relationship — open someone's card and LIA now writes a synthesis at the top of it: where you stand with that person, what is still open, what to do next, what is worth remembering. It is written at most once a day, only when you open the card, and never during a conversation. If a refresh fails, the previous text stays under a line saying so: an empty panel would read as "there is nothing". You can switch it off from the Relations page.
- 💬 And the chat uses it when you name someone — no more announcing a search for facts already written down. The debrief arrives dated, with its age, and LIA knows to check any exact figure, count or status against her tools. If two people have similar names, nothing is injected at all: saying nothing beats handing you someone else's file.
- 🌙 A third tab: "On LIA's own initiative" — the registers showed what you had triggered; what LIA did by herself did not appear at all. Her morning sweeps, the interests she explores, the notifications she sends without being asked: all of it now has its own list, actions on one side, sources consulted on the other.
- 📤 Your extractions are no longer truncated — the files you download from the registers, and those an administrator exports, were capped at a number of rows. On a real account that left out 97.9 % of the detail of the model calls. They are now complete whatever their size, with the exact count stated in the header, and compressed along the way when your browser accepts it.
- ⚡ Your home page opens your sources only once — it loads in two pieces, the cards then the synthesis, and each was fetching your mail, your calendar and your weather on its own. Measured: 39 % of page loads opened everything twice. One act of reading now — and the consultation log stops counting double.
- 💶 The cost of a text LIA wrote is shown under it — today's synthesis, the home card and the relationship debrief all carry the same label: the model used, the tokens, the price. When nothing was spent, nothing is shown — "€0.00" would be a claim, not an absence.
- 🛡️ The ceilings apply to everything the platform pays for — an administrator sets a per-account quota and a daily spending ceiling for the instance. Five kinds of work slipped past them and spent where no ceiling could refuse. Every model call paid for by the instance is now subject to both: a refused request says so plainly and announces when it will be possible again, while a background task gives up quietly — a quota refusal is not a failure.
v1.43.0 — 🔁 Your reminders can now repeatSeptember 2026
- 🔁 A reminder can come back — until now a reminder rang once and vanished. You can now give it a rhythm: every day, Monday and Thursday, the 15th of each month, or every two hours between 9 and 5. A new "Reminders" section in settings lists them all, with the sentence describing their rhythm and the next dates. A reminder with no future left still disappears, exactly as before.
- 🗓️ Your routines can say far more — "every three days", "every other Tuesday", "the 2nd Tuesday of the month", "the last day of the month", "every 14 July", "every two hours between 9 and 5", "until 30 September" or "the next three times". Ask for it in conversation or set it by hand: it is the same vocabulary on both sides.
- ✏️ The forms breathe — fields are grouped under named headings ("what to do", "when", "how it runs"), and what you rarely change — an interval, an end rule, a start date — folds away behind "Advanced options". Five questions on screen instead of eight. The fold opens by itself when it already holds something, so it never hides a setting you made.
- 🐛 One day lost a year, in 73 timezones — at the spring clock change, when the offset moves at local midnight, the previous engine skipped the whole day: 142 runs lost a year, with nothing in the logs. Fixed, and the routine week now draws the 30 March cell again.
v1.42
v1.42.4 — 🖥️ Your local models, models like any otherSeptember 2026
- 🖥️ Any of LIA's brains can run on your own machine — a slot pointed at an Ollama model stayed stuck on “Generating the response…” and never finished. LIA now speaks to Ollama in its own language rather than through a compatibility layer: answers come back, the length limit is honoured, and a model that thinks obeys the depth you chose — including not thinking at all when you want a quick answer. Its thinking is shown separately, in the progress panel, and never eats into the answer.
- 🔎 LIA asks your server what each model can do — tools, vision, thinking and context size are read from the server itself instead of guessed from a name. In Settings › Administration › LLM Configuration the model list shows their real capabilities: the thinking control appears only for a model that thinks, and the sliders a local model does not accept disappear instead of being silently ignored.
- 📏 The context window is your machine's — LIA assumed a size unrelated to your server, which then trimmed the start of long conversations without saying so. It now asks for a window explicitly — the model's own maximum, or the `OLLAMA_NUM_CTX` value if you set one — and that same number decides when the history is summarised.
v1.42.3 — 🔎 A diagnosis that shows its evidenceSeptember 2026
- 📚 Your documents are back in every answer — for weeks every conversation quietly lost its document context: the embedding provider rejected each search on your knowledge spaces while the same text embedded fine one step earlier. The cause was a data type at the boundary of the provider's client, not your documents; it is fixed at the one place every embedding goes through.
- 🩺 The diagnostician reads its evidence before it speaks — an incident diagnosis used to conclude « insufficient evidence » while the answer sat in the metrics and the logs. It now receives a bounded evidence pack collected at the moment of the diagnosis — breakdown metrics, a sanitized log excerpt, the running build and its uptime — and Settings › Platform health shows « what the diagnostician read » under each diagnosis. A source it could not reach is stated, never mistaken for silence.
- 📖 The operations runbooks finally travel with the deployment — the diagnostician is grounded on a written procedure per alert; production had been running with an empty runbook folder and nothing said so. The bundle now ships them, the panel says when none is mounted, and the API warns at start-up.
v1.42.2 — 📅 Your week of routines, at a glanceSeptember 2026
- 📅 A week view above your routines — hours down, days across: every routine sits at its hour on its days, numbered in the order it fires, with the same number on its card. Each cell of the current week says what happened: white not executed, green executed, red failed, amber proposed and waiting for your approval, grey paused. A routine that checks a condition wears a ring; one running right now pulses. Click a number to jump to its card; the arrow keys walk the grid. The colours start over every Monday and clear when you change a routine's days or time.
- 🎙️ A meeting is never lost, its owner never blocked — a meeting whose minutes failed used to stay « processing » for hours. The job now resumes from what it already acquired (normalised audio, then the transcript) instead of paying the engine again; a meeting whose worker died is requeued, or marked failed with its reason, and you can delete it meanwhile. The detail shows the attempt of the budget and the previous attempt's reason.
- 🛠️ Routines that keep their word — the list no longer blinks away on every refresh (your open panels stay open); a routine paused during a time-zone move now wakes up on your new clock; and a routine scheduled at midnight in a zone whose clocks jump at midnight no longer skips that day.
v1.42.1 — 🎭 A face that lives between two answersSeptember 2026
- 🎭 Brows and a mouth that live — the avatar's eyes had a life of their own; its brows and mouth stood still between two answers. The brows now have an arch and stay faintly present at rest, one breath carries the whole face, the gaze lifts the brows and a blink dips them, and the mouth plays small mimics at an irregular, unhurried pace — a grin, a pout, a pucker, a held-back giggle. Every minute or two, on a resting face, a short sketch of three to five seconds: a fly to follow, a sneeze, a yawn and a stretch, hiccups, a dizzy spell, a double take. Any change of expression drops the sketch on the spot and leaves the face exactly where it was.
- 🏠 The avatar on the home page — LIA's face now greets visitors on the public home page, in the capsule look, sitting in a corner that stays put while the page scrolls and can be dragged anywhere. The spot it is dragged to belongs to the home page alone: it never moves the widget in the chat. No account is needed — everything the face does on its own plays on a resting expression. On a phone it is small and can be hidden to a restore dot.
- 🎚️ Calmer style previews — the six style cards in Settings › Personalization › Eyes style compare silhouettes, so they no longer sneeze or yawn mid-comparison: previews keep the breath and the resting hold, nothing else.
v1.42.0 — 📄 Your own unified recordSeptember 2026
- 📄 One file with everything recorded about you — the exports on each register cover one journal apiece. A new unified extraction, below the tabs of Registers, gathers all five records LIA keeps — your turns, the actions, the consultations, the model calls and any gap in the record itself — into a single machine-readable file, as a voluntary transparency record. It is the same file an administrator can produce, narrowed to your account alone: the route it calls takes no account parameter at all, so it can only ever return yours. And it holds no content, with every identifier replaced by a handle — yours included — so you can attach it to a portability request, a complaint or a bug report without editing anything out of it first.
- 🏅 The standards LIA follows, stated on the home page — under the security section, a strip now names them: GDPR, AI Act transparency principles, the Model Context Protocol, Agent Plugins, agentskills.io, OWASP Top 10, WebAuthn, OAuth 2.1 with PKCE, OpenTelemetry, Prometheus, Keep a Changelog, semantic versioning and WCAG 2.2 AA. An open standard is what makes a promise checkable against a text somebody else wrote.
- ✨ A new card on « More », and guides that stopped narrating releases — the card shows the small invisible decision inside the registers: the line is written before the action happens, and only turns green when the tool says so out loud. And the Story and Why guides no longer recount past releases one by one; they state what those releases established.
v1.41
v1.41.0 — 📋 Registers: what LIA did, and what it looked atSeptember 2026
- 📋 Two registers, kept automatically — the dashboard tile is now called Registers, and it answers two different questions that are never mixed together. Actions: everything LIA did on your behalf — an email sent, an event created, a file deleted — with its outcome and the confirmation you gave. Consultations: everything it looked at to answer you, recorded as the capability (« your calendar »), never what was searched. An action is written down before it happens and marked done only on an explicit result, so LIA can never report a success it has no proof of.
- 📊 A new « Overview » tab draws the period — ten charts built from those same registers: how your turns ended, actions by outcome, consultations by domain, calls and tokens per model, average latency per tool, and any gap in the record itself. Every chart carries the exact total of the whole period beside it, including what the top-12 folded into « other », so the bars can always be checked. Administrators get the same charts over one, several or every account.
- 📤 Exports in three formats, including the technical one — readable, CSV, and now the machine-readable format previously reserved for administrators. Your file is therefore complete enough to hand to someone else. Everything in these registers is yours: readable in the app, exportable, and deleted with your account.
- 🛠️ An export could cover the wrong period — when there were more rows than one file may carry, the export kept the oldest ones instead of the most recent, so it could show months-old activity while honestly announcing that it was capped. A capped export now always keeps the most recent window and says so.
- 🛠️ Two administrator screens were unreachable — the cross-account charts and the cross-account integrity check answered an error to every request. Both work now. In the same pass, a confirmed draft whose action was already on record without a result could be reported as done: it now says clearly that it was not retried, and asks you to check before trying again.
v1.40
v1.40.0 — 🔔 Three loops were running in silenceSeptember 2026
- 🧠 Starting a new conversation no longer erases what LIA learned — until now, the « new conversation » button cleared far more than the conversation: the rhythm and recurring requests LIA had observed, the mark telling it which mail it had already read, the thresholds it had adjusted for you. On a heavily used account that was every few days, so the learning never had time to conclude anything. Each stored value now says what it is, and a reset only clears what a conversation created. « Forget everything » and deleting your account still erase everything, as before.
- ⚡ LIA can now react to an important email or a fresh invitation in minutes — instead of waiting for its next scheduled pass. The rules are explicit and unchanged in spirit: it only looks at a message carrying the label you consider important, never at promotions, social or list mail; and for the calendar, an event starting soon, changed by someone else or still waiting for your answer. Everything else stays as it was — your notification window, your daily cap, the pauses between messages. A storm of arrivals is still a single wake-up.
- 👀 Reading LIA now counts as being there — the rhythm it learns used to see only what you typed. Opening the app counts, and so does a 👍 or 👎 on a notification; a notification it sends you never counts as your activity. If you read LIA more than you write to it, its idea of your day finally matches your day. Off by default, switched on by your administrator.
- 📬 A knowledge space can follow a Gmail label — put a label on the conversations that matter, and they become documents of the space, searchable weeks later. Remove the label in Gmail and the document leaves: the gesture you already know is the one that decides. Only labelled threads are copied — never your whole mailbox — and attachments contribute their names, not their content.
- 🕒 A proactive notification says what prompted it — the history now marks the ones that answered an email or an invitation, next to those that came from the regular pass. Useful to judge, at a glance, whether LIA spoke because something happened or because the clock said so.
- 🗂️ Synced sources are easier to handle — a linked Drive folder and a followed Gmail label share the same card: their state, their exact counts, and their two actions always visible rather than appearing on hover — on a phone they fold into a menu that names the source it belongs to.
v1.39
v1.39.1 — 🗂️ Thirty minutes formats, and LIA picks the right oneSeptember 2026
- 🗂️ A library of minutes formats — thirty ready-made templates, filed by use: meetings and teams, transcripts, conversation analysis, sales and consulting, technical, personal appointments, courses and training. Browse them, preview them, add the ones that speak to you to your own templates — one at a time or as a selection — then adapt them at home: rename, reorder the sections, rewrite the instructions, decline a template into several variants, delete what no longer serves.
- ✨ LIA picks the format by itself — with nothing to set up, it reads what was said and keeps the matching format: a project review, an interview, a medical appointment and a lecture do not share a structure. The minutes state which format was used, whether it came from you, from your default or from its own choice, and in one line why. You keep the wheel: pick your default in the settings, or the current meeting's format from the recording banner.
- 🔁 Change the format afterwards — minutes already written can be rewritten in another format from the stored transcript: either you replace the ones you have, or you get new minutes from the same meeting, keeping the link to their origin and joining the knowledge space as a document of their own. New along the way: a format can now hand back the full transcript, cleaned up and readable, rather than a summary.
- ✉️ The minutes by email, for everyone — they now leave from the application's own address to your account's, subject “Meeting minutes · Title”. No connected mailbox required any more: manual and automatic sending both work on any instance whose administrator configured email.
- 🗃️ Your documents move between your knowledge spaces — download a document, or a whole selection as a zip archive under the original names; move documents to another space, index and file follow; delete several at once. Every batch tells you what it did and what it left aside, with the reason: a document synced from Drive or written from a meeting stays where its owner put it.
- 🎙️ Recording has a place of its own — a dedicated button in the header on a computer, a “Record / Stop” entry in the logo menu on a phone whose trigger pulses red while capturing, and the banner that follows you on every page without ever pushing the typing area off screen. The composer's “+” goes back to what it was: attaching a file. And the meetings list finally accepts multiple selection to delete several at once.
v1.39.0 — 🎙️ Record a meeting, get the minutesSeptember 2026
- 🎙️ Record a meeting from the chat — the + button next to the message field now records a meeting with your phone or your computer as the microphone. A banner follows you on every page with the duration and what has reached the server; you keep chatting with LIA in the meantime, spoken answers paused so the microphone never hears the assistant.
- 📝 Minutes in your own structure — date, times, place and participants, then the sections you defined in Settings › Meetings: summary, topics, decisions, actions with owner and deadline, risks, open questions… editable, restorable, rebuildable with your current template. Download as PDF, send to your mailbox, or ask LIA later: every meeting joins a “Meetings” knowledge space.
- 🔁 Built for real life — a reload, a locked phone or a lost connection costs seconds, never the meeting: audio leaves in small segments, waits offline, and the banner offers resume, finalize or discard. After a long silence LIA asks whether you are still recording; a maximum duration finalizes by itself; a gap is stated in the minutes, never filled in.
- 🔀 A chain of engines, no dead end — the administrator's speech engine, then ElevenLabs or OpenAI with speaker separation, then the local Whisper — now able to transcribe any length, in speech windows. The engine and its price per hour are shown before you start, and a provider that refuses at processing time hands over to the next instead of failing your meeting.
- 💶 What it cost, stated everywhere — transcription and minutes as two amounts and a total, on the chat card (when you display costs), on the meeting page and in the list; counted in your consumption like any other exchange. A model without a configured price reads “not priced”, never zero.
- 🧭 A Meetings entry in the header between Relations and Alerts, only where the instance offers the feature. To make room, the language switch shows its flag alone at every width and the personality title waits for very wide screens. The composer's paperclip became a + — it opens actions, of which attaching a file is one — and gives the typing area back some width on phones.
v1.38
v1.38.6 — 🛡️ What an email says stays what an email saysSeptember 2026
- 🛡️ A message can no longer give orders — text written by someone else (an email body, an invitation description, a web page, a connected server's result) was already flagged as untrusted when LIA read it. But when a long conversation gets summarised to save room, that flag was lost: a sender's demand could resurface later as if it were an established fact. The summary now carries the provenance with it, and third-party requests are reported as what someone else wrote — never as something you decided.
- 🎯 Long tasks keep the thread — on a very long autonomous task, your original question could be pushed out of the assistant's working memory by the task's own intermediate results. It then finished without knowing what it had been asked. Your question is now pinned for the whole task, however long it runs.
- 🔍 The action record no longer forgets its beginning — the ⚙ panel under each answer kept only the most recent steps of a long turn and displayed that truncated number as the total. It now keeps the opening actions as well as the latest ones, states the real total, and says exactly how many intermediate steps are hidden.
- ✅ Fewer forgotten parts in a multi-part request — when you ask for three things at once, the internal reviewer checked that each planned step was correct, but never that nothing was missing. It now lists what your request asks for and checks that every part is covered, silently correcting the plan when one is not.
- 📊 Admins: context consumption is finally visible — two new metrics and a Grafana panel show how much context each autonomous iteration actually sends to the model, and how close that comes to the model's limit.
v1.38.5 — 🔑 Your servers know they speak for youSeptember 2026
- 🔑 “My” finally means you — a server connected with your own account (GitHub, bank…) now tells the assistant so, everywhere it reads that server's capabilities. “List my repositories” no longer ends in “what is your username?”: your connected account is the answer.
- 🔄 Connections that stay connected — a token-refresh defect silently broke a server's credentials on every second renewal, forcing you to reconnect over and over. Fixed at the root. And when a server does need reconnecting, the assistant now names it and points you to Settings instead of claiming the capability does not exist.
- ⏱️ Honest time budgets for autonomous turns — thinking time and tool-work time are now counted separately, each with its own bound, so one long delegation can no longer eat the whole turn. Administrators see both budget bars in the debug panel.
- 🧰 Your servers' data reaches the answer — results fetched by a connected server's sub-agent were sometimes summarized as “N items found” without the items. The actual rows now travel back into the answer, safely marked as external content.
v1.38.4 — 🔌 The tools that never showed upSeptember 2026
- 🔌 Your MCP servers stop losing tools — a connected server would announce forty tools and quietly deliver ten, so the assistant answered that it could not do something it was perfectly equipped to do. Every tool a server publishes is now read, including those written in the more advanced description styles that whole servers were being dropped for.
- 🏷️ Tools now carry their real names — the server list showed machine identifiers such as
accounts__list_financial_accounts. When a server publishes a readable name, that is what you see, with the technical identifier kept beside it for logs and rules. - 🛡️ A destructive tool always asks first — confirmation was a per-server setting, so switching it off for a server you merely query also switched it off for its few tools that delete something. A tool the server itself declares destructive now asks for your agreement in every case.
- 📊 Administrators can see a tool go missing — a tool the assistant fails to load simply vanished, with no counter and nothing a dashboard could show. It is now counted, charted, and traced back to the exact server and tool.
v1.38.3 — 🧠 The memory that was quietly going missingSeptember 2026
- 🧠 LIA stops quietly losing what it just learned — in one hour of real use, nearly half of the operations that store a conversation in memory, index a message or fetch a document context were being refused by the provider, and nothing said so: the answers arrived normally, only slightly less well informed. The background tasks now start at staggered times instead of all at once, wait their turn when needed, and try again on a temporary failure.
- 📊 Platform Health now watches this — nothing measured whether these operations actually succeeded, only how many were attempted, which says nothing once failures are retried. A dedicated indicator and an alert have been added, with a procedure telling an administrator which of the two levers to pull.
- 🌍 Incident diagnoses in your own language — an incident's analysis and recommended actions were always written in English, on a page where every other word is translated. They are now written in the language of each administrator who reads them.
- 🔍 Diagnoses that finally say something — they used to answer "insufficient evidence to determine the exact cause" almost every time. That was accurate: the analysis was given a number without the threshold it had crossed, and a number alone cannot be judged. It now receives the unit, the verdict and both alert levels.
v1.38.2 — 🙂 A face that answers the tone, not the habitSeptember 2026
- 🙂 LIA's face answers the TONE of her reply — after every message the avatar wore roughly the same pleased expression, including after an error or a long technical explanation. LIA now states the register of what she has just written — warm, assured, factual, apologetic, curious, concerned… — and the face plays it. Twelve distinct expressions instead of one, and only two of them smile: a technical answer keeps a focused face.
- 👄 An expressive mouth — the avatar now has a mouth that carries the expression: it opens, deepens, leans to one side and follows the face. It arrives at the corners first, the way a real smile does, instead of appearing all at once.
- 🛠️ Deployment fix — on some installations a monitoring container refused to start after an update, with a permission error on its start-up script. Fixed, and an automatic check now stops the case from coming back.
v1.38.1 — 🎨 Settings you can find at a glanceAugust 2026
- 🎨 Every settings family has its own colour, every setting its own icon — the page listed fifty-three sections that all drew the same blue icon on the same blue chip, and sixteen of them borrowed another section's drawing on top of that (one plug shape served four different settings). Colour cannot separate a repeated shape, so both were fixed: a distinct icon per setting, and one colour per family — twelve colours you learn, not fifty-three you decipher. It works the same way in the sidebar, which is the settings list on a phone.
- ♿ The colour never carries meaning on its own — the section you have open still stands out by its background, its weight and the accent colour, not by being a thirteenth hue; and whether a capability is on is still shown by a filled or hollow dot. Anyone who does not perceive these colours loses nothing. Each of them was measured against every theme — light, dark and OLED, across the five accent colours — before shipping.
v1.38.0 — 🎭 A full palette of moods, and a deployment that stops lyingAugust 2026
- 🎭 LIA is no longer stuck in a single register — her mood mechanics rested every personality on the assertive side of the scale, so five calmer moods stayed out of reach at rest, and an internal pulse crowned joy the dominant emotion regardless of what she had actually appraised: measured over 90 days of production, 31% of turns, and up to every single turn for a light user. Both settings are now active: half the personalities rest in a calmer register, and the emotion shown is the one LIA really appraised. Assertive characters stay assertive.
- 📊 No measurement stays invisible to administrators — a metric no dashboard displays is a metric nobody acts on: a context source failing silently once dropped health signals for a week with nothing to flag it. An automatic check now verifies that every measurement LIA produces reaches a dashboard, a rule or an alert, and two new Grafana panels cover heartbeat source health and skipped cache writes.
- 🚀 The deployment says what it knows — and stops leaving your secrets behind — for self-hosters: going to production used to end on an error you had to learn to disregard, because a dropped connection was indistinguishable from a real failure. The deployment now runs on the server, independently of your terminal: closing the laptop or losing Wi-Fi no longer interrupts it, and the closing message finally tells a real failure apart from a lost contact. Along the way, the decrypted secrets file is wiped from your machine on every exit path, errors included.
v1.37
v1.37.0 — 🧮 Answers that were computed, and a sandbox that finally has its toolsAugust 2026
- 🧮 LIA computes instead of guessing — when a question needs arithmetic over many rows, a comparison between two lists or durations across timezones, the assistant now writes a few lines of Python and runs them in a locked-down sandbox: no network, no database, nothing beyond the data already collected, and a throwaway container each time. The answer was computed, not estimated. It remains a complement: most questions never need it. Administrators can read every script that ran, with its purpose and output, in the debug panel.
- 🧩 The QR-code skill works again — and every other skill that needs a library. Scripts ran with the standard library only, because the sandbox account could not enter the directory holding the installed packages: the shipped QR-code skill could not load the library that draws the codes. Found by running a real script on the live server rather than by a test, and now guarded so it cannot come back.
v1.36
v1.36.0 — 🧮 LIA can now compute instead of guessingAugust 2026
- 🧮 A short script instead of mental arithmetic — an AI answers questions like "how long are my layovers in total", "which of these people appear in both lists" or "what does this add up to across timezones" plausibly, and nothing in the answer shows you it is wrong. When a step needs that kind of work, LIA now writes a few lines of Python and runs them in a locked-down sandbox — no network, no access to your data beyond what the turn already collected, a throwaway container each time — and answers with a result you could check. It stays a complement: most answers never need it, and it is deliberately told not to use it for anything it can answer directly. Administrators see every script that ran, in the debug panel.
v1.35
v1.35.0 — 🔎 A search that gets cut short now says soAugust 2026
- 🔎 No more “give me a minute, I'll get back to you” — when a long search hit its internal limit, the assistant sometimes ended the turn on the sentence it had written just before continuing, which read as a promise of work that never came: a turn ends when its answer is sent, so there was nothing to wait for. Now an interrupted investigation gives you what it DID find, says in one sentence that it was cut short and what is missing, and offers to continue — and it can no longer announce future work it will not do.
- 🧠 Your standing instructions now reach the autonomous mode too — in the step-by-step reasoning mode, long-term memory never reached the reasoning itself: the rules you had recorded about how the assistant should work could only influence the final wording, not the decisions. The same memory the other mode uses is now present where the decisions are actually made. Deep searches also get more room: the assistant keeps going as long as it keeps finding things, instead of stopping on a fixed count.
v1.34
v1.34.1 — 📌 Your standing instructions are remembered againAugust 2026
- 📌 “Rules & directives” memories can finally be created — long-term memory has a category for the standing instructions you give about how LIA should work (“answer me more briefly”, “always start with the conclusion”). The screen that shows them shipped, but nothing could ever fill it: the parser that turns an extraction into a memory rejected that exact category, and the rejection was logged at debug level — so the whole family was dropped in silence. The vocabulary now has a single source of truth, checked at startup against all three places that restate it, and a rejected memory is logged as a warning.
- 🩺 The platform can now say when it has lost the outside world — LIA watches its own health, and it met a real outage the day it shipped: the machine stopped routing traffic for its containers, so every call to an external service failed while the site itself stayed up and healthy. LIA reported the symptoms within six minutes, but nothing could name the cause. An administrator can now enable a small outbound-connectivity check that says it plainly, and each health figure is shown with the unit the server declares instead of one the screen guessed.
v1.34.0 — 🩺 LIA can diagnose itselfAugust 2026
- 🩺 Platform health, inside the product — administrators get a new Settings panel showing what LIA continuously measures about itself: per-check verdicts with exact values, firing alerts, degraded capabilities with their suggested fallbacks, and an incident history. When something critical opens, the administrators are notified in-app and by push.
- 🔎 “LIA, diagnose yourself” — in chat, an administrator can ask LIA for its own health, a curated platform metric, bounded service logs or an incident's stored diagnosis. Each incident can carry an automatic diagnosis grounded on the operations runbook for that alert, with a probable cause and recommended actions — proposals for a human, never executed on their own, and capped by a daily budget.
- 🛟 Honest answers under failure, for everyone — when a step fails during a request, LIA now tells you precisely what succeeded, what failed and why, and applies a known workaround when one exists (for example switching web-search provider when one of them is down) instead of failing on a timeout.
v1.33
v1.33.2 — 📘 Self-hosting: the upgrade path, written downAugust 2026
- 📘 Moving your own installation to a new version — self-hosting the assistant, you had no written way to move an existing installation to a newer release. There is one now: seven steps, with the database backup first because migrations apply the moment the new container starts and there is no way back, and a rollback that restores the previous version in seconds. LIA can walk you through it if you ask her.
- ✅ The documentation now has to be true — every version number and threshold written in the docs is recomputed from the code that owns it, and a mismatch stops the build. Six documents were quoting six different wrong values for the same setting; that class of error can no longer reach you. The figures published on the public pages follow the same rule: recomputed, never carried over.
v1.33.1 — 📍 Location settings in their own sectionAugust 2026
- 📍 A standalone "Location" section — your position sources (live geolocation, remembered last-known position, home address) leave the Google Places connector card for their own section under Preferences. This data always lived on your account, never on a connector — it now looks the way it works. Find it instantly by typing "address" or "GPS" in the settings search.
- 🧭 The FAQ points to the right place — several answers still directed you to "Settings > Connectors > Google Places" to configure your location; every language edition now points to the new section.
v1.33.0 — 📱 Native Android & iOS appsAugust 2026
- 📱 LIA in your pocket — one official app per store, working with any self-hosted LIA server: you type your server's address at first launch, and that's it. The interface is the same as the web app, so every improvement reaches the app instantly, with no store update.
- 🔐 Google sign-in works in the app — and fixed a real flaw on the way: signing in with Google used to skip the two-factor code entirely. The code step is now enforced everywhere, app included.
- 🔔 Notifications on both platforms — Android talks directly to your server's own notification project: nothing passes through anyone else. iPhone requires going through the app publisher (an Apple rule, not ours), so a minimal relay wakes the phone with a fixed neutral sentence — your actual messages never leave your server, and the relay never learns who was notified or why.
- ↩️ Connecting services returns to the app — linking Gmail, a calendar or an MCP server from the app now brings you back into the app when done, instead of leaving you stranded in the browser.
- 📴 A real offline screen — if your server is unreachable, the app shows its own screen with a retry button and, crucially, a way to change servers: a mistyped address no longer means reinstalling the app.
- 🧪 Verified on real devices — an automated bench drives the actual app on an emulator through ten scenarios, from first setup to forgetting a server. It caught three real bugs before anyone else could.
- 📖 A complete publication guide — for anyone who wants to publish their own build: accounts, signing, both stores' forms, step by step with no prior knowledge assumed.
v1.32
v1.32.0 — 🧠 One way to ask a model to thinkAugust 2026
- 🧠 Reasoning depth becomes one clear setting — every model now offers the same ladder, from “no thinking” to “maximum”, with the levels that particular model actually accepts. Before, the same instruction was stored two different ways depending on the provider, and the admin screen could offer a depth the model's API refused — the save went through and the call failed later.
- 🎛️ The dropdown only offers what works — the list of depths, the option to switch reasoning off, the token budget: all of it now comes from the same source the server validates against. A model that cannot stop thinking no longer pretends it can, and a depth a model does not accept is no longer selectable.
- 🔍 The model catalogue says where its numbers come from — each row now shows whether its capabilities were corroborated by the public registries, verified by a human, or never curated at all. Announced retirement dates appear next to the models concerned, with the source of the information.
- 🧹 Two obsolete settings removed from the catalogue form — they described how the reasoning control used to be drawn, and nothing had read them since. A field you can edit that changes nothing is worse than no field.
- 🎚️ Accepted depths are ticked, not typed — the LLM Pricing screen shows the depths the model's family actually offers, and you untick the ones this particular model refuses. Everything ticked means "no restriction". A free-text field used to invite writing a level that does not exist, and it was dropped in silence.
- 📗 The Excel workbook writes reasoning directly — two columns replace the "copy the shape of another model" template, which could remove depths without saying so when it came from another family. On import, a depth outside the family is refused while naming the ones that would have been accepted. Creating a model that does not reason no longer requires picking a reasoning template.
- 🔧 Fixed: a depth restriction can be widened again — re-ticking every depth did not restore the full ladder: the previous restriction silently survived.
v1.31
v1.31.3 — 🌑 A truly black black, and landmarks recoveredAugust 2026
- 🌑 OLED mode arrives — a perfectly black background, the one at which an OLED screen's pixels actually switch off: less battery drawn, and noticeably easier on the eyes at night. The header's theme button becomes circular: light, dark, absolute black, then light again. The switch is revealed by a circle opening from the button — skipped entirely if you have asked your system to reduce motion. Your five colour themes are untouched: only the neutral surfaces darken.
- 🎨 The display mode is now chosen in Settings — Appearance offers Light, Dark, System and an "Absolute black" switch. "System" follows your device and is reachable again: it is what every new account starts on, and the header button alone could never return to it. The OLED switch is only active in explicit dark mode, rather than changing your mode without telling you.
- ⌨️ Keyboard focus is visible again — three checkboxes (remember me, terms acceptance) showed no outline at all when reached with the keyboard, leaving no way to tell where you were in the form. They had no accent colour either. Fixed, and the outline now follows your theme.
- 👀 Two unreadable texts fixed — the badge marking that a skill answered was practically invisible in light mode, and its animated sheen stayed frozen over the word when animations are reduced. The quietest section headings — "try for example", the trace headers — were also too pale to read comfortably.
- ⚡ Lighter pages, and no more flicker — the nine selectable fonts were all downloaded on every visit, 618 KB for the single one displayed; only 162 KB remain, with identical rendering. And your colour theme is now applied before the first paint: the brief blue flash on load is gone. A missing page finally answers in your language.
v1.31.2 — 🔎 Document search finds its answers againAugust 2026
- 📄 Your documents answer again — search across your knowledge spaces was discarding one correct answer in three: ranking mixed meaning-based understanding with keyword matching, and the second signal could eliminate a perfectly relevant excerpt purely because it shared no word with your question. Asking "is my data encrypted?" returned nothing, while five passages answered it. Fixed.
- 🌍 In all six languages, Chinese most of all — the defect hit harder the further your question's language sat from your documents'. On Chinese-language documents, search found the right answer only 7% of the time; it now succeeds nearly 9 times out of 10. French, German, Spanish and Italian all improve markedly.
- 🤫 Fewer off-topic excerpts — when your message has nothing to do with your documents, LIA now slips fewer of them into its reasoning than before. The fix improves relevance and reduces noise at the same time; it does not trade one for the other.
- 🛡️ A startup that refuses an inconsistent configuration — if the semantic-search model is misconfigured, the application says so at boot instead of letting every write fail silently in production. The debug panel also shows the relevance threshold in force, drawn directly on each score bar.
v1.31.1 — 🌍 The Google ecosystem widensAugust 2026
- 👥 Your contact groups are recognised — "email the family group" now works: LIA finds your Google Contacts labels, expands them into addresses and shows the exact member count. It can also find someone you have exchanged with but never saved.
- 📅 Free slots and video calls — "find me 30 minutes on Thursday" computes real availability from your calendar, reading only your busy ranges and never the content of your meetings. And a meeting created as a video call carries its Meet link (or Teams, depending on your provider) — announced only if it was actually created.
- ✉️ Your Gmail mailbox becomes steerable — "is my auto-reply on?", "turn on my out-of-office from the 24th to the 30th with this message", "archive and label everything from this sender". As with every send, the preview shows the exact wording before anything is written, and nothing happens without your approval. You will be asked to reconnect Gmail once.
- 📊 Your spreadsheets and documents become readable… and writable — "read my budget and answer", "summarise these minutes", then "add this expense" or "append this note". The preview shows exactly the rows or text to be written; no new permission is requested.
- 🌿 Weather answers the question you did not ask — with the Google Environment connector on, a plain weather question also brings air quality and in-season pollen, on the chat card and on your home page. The index shown is your country's when one exists, with the category the service itself publishes — never an in-house interpretation.
- 🛡️ Safety and freshness — every web address LIA is about to open is checked against Google's threat lists, and a flagged page is refused. On freshness, calendar, files and messages can now tell LIA when they change instead of being polled periodically (an administration option, off by default).
v1.31.0 — ⚙️ The engine room moves to Python 3.14August 2026
- ⚙️ A modernized foundation — the whole assistant now runs on Python 3.14 and Debian 13, everywhere: your data is processed by a faster, current, long-supported runtime. Every one of the 229 underlying components was verified for the new platform before the switch — including on the little ARM server the public instance runs on.
- 🎙️ Telegram voice messages made future-proof — the audio layer that turns your voice notes into text relied on a piece Python removed in its newer versions; it was replaced before anything could break, and the exact audio conversion is now covered by automated tests so it can never regress silently.
- 🛡️ Guards that keep it that way — new permanent checks verify on every change that all parts of the system agree on the same Python version and that every native component loads correctly. A future upgrade that forgets a spot now fails loudly in development instead of quietly in production.
- ⚡ A fairness bug in the rate limiter is fixed — under simultaneous bursts, the shared request counter could let a few extra calls through its limit (two requests landing in the same microsecond collided). Reviving a long-dormant stress test surfaced it; the window now counts every request exactly, so quotas mean what they say.
- 🧭 The instance now reports its true version — the running server used to display a placeholder build identity in some deployments; the deployment script now reconciles it automatically, so what you see in diagnostics is exactly what was shipped.
- 🧹 Leaner and future-proof — five unused libraries were removed from the runtime (a smaller attack surface and faster installs), and every use of Python APIs scheduled for removal in future versions was modernized ahead of time. Test coverage grew again: the stress suite that found the limiter bug now actually runs.
v1.30
v1.30.16 — 👀 LIA looks back at youAugust 2026
- 👀 Expressive eyes in the chat — a small pair of cartoon eyes now lives on the chat page and reacts to everything: they watch your typing, squint while LIA thinks, scan while it searches, and react to each answer's emotional tone. Left alone they blink, look around, doze off progressively — and once in a while pull a genuine prank (they may even swap places).
- 🎨 Six looks to choose from — Cozmo glowing screens by default, plus Capsules, Marbles, Almond, Strokes and Rings with a moving pupil. Pick yours with live animated previews in Settings › Personalization › Eyes style.
- 🖐️ Fully yours — drag the eyes anywhere (arrow keys work too), resize them, or dismiss them to a tiny restore dot. By default they sit between the search field and the knowledge badge, and they always return exactly where they belong.
- 🧹 A lighter chat header — the amber “Processing…” pill and the little wait phrase are gone: the eyes now carry that state, and they say more (thinking, searching and answering are three different looks). The offline pill stays — that one tells you something the eyes cannot.
v1.30.15 — 🛡️ Sturdier in the real worldAugust 2026
- 🔔 Smarter notifications across tabs — the tab you are actually using always keeps its live connection; a long-forgotten tab can no longer pile up dead connections behind the scenes. Come back to a tab and it reconnects on its own.
- 🌍 Weather for any place name — an unknown or misspelled location now gets a clear "location not found" answer (with a second geocoder trying first), instead of a technical error.
- 🔎 Deep web research gets the time it needs — long syntheses are no longer cut off at 30 seconds, and an over-long search query is automatically shortened instead of failing.
- 🔒 Cleaner, more private technical logs — asking for your confirmation is no longer counted as an error, and draft content or recipient names never end up in the technical logs.
- 📊 For admins: honest showroom metrics — the demo funnel counts every mission again, clicks toward the free live instance are now measured, and a new alert watches for Redis connection-pool saturation.
v1.30.14 — 👁️ The invisible work becomes visibleAugust 2026
- 🕐 New Activity page — everything LIA did for you, as one chronological feed: notifications sent, interest discoveries, journal notes, detected habits, tracked and closed commitments, executed routines. With exact totals, never approximations.
- 💡 Proposals awaiting your go — the Alerts hub gains a “Proposals” section: routines LIA offered to handle that you haven't decided on yet. “I'm on it” opens the chat prefilled — nothing is sent without you. “Not now” teaches it to insist less.
- 📌 Say it once, it sticks — “answer shorter”, “stop suggesting that in the evening”: your instructions become durable rules, visible and editable in your memories (“Rules & directives” category). And when you correct LIA, it acknowledges once, without triple apologies — it simply changes.
- 🧠 A memory that dates its recollections — LIA can naturally say “last week you told me…”. And when a fact changes (you move, you switch jobs), the old one is no longer overwritten: it archives cleanly behind the new one and never resurfaces in searches.
- 🔊 Your briefing reads itself aloud — a Listen button on the morning synthesis, and a voice whose expressiveness follows the assistant's mood. If synthesis fails, you get a clear message — no more silent failures.
- 📈 Your habit streaks count — “9 days in a row, 7-day milestone reached, next: 30” in the Habits panel. And under chat suggestions, LIA now says why it proposes (“because you follow…”).
v1.30.13 — 🔄 The self-improvement loops closeAugust 2026
- 🔕 Notifications finally respect your ongoing conversations — the "don't interrupt an active user" rule existed in the configuration but never actually applied: it now genuinely checks your latest message before any proactive send. Mid-discussion, LIA waits its turn.
- 🎚️ The assistant tunes its own working memory, for you — the threshold deciding when a journal note deserves to enter a reply now adjusts per user, inside strict bounds, at most one small step per day, with a kill switch. Your vocabulary isn't someone else's: neither is the setting.
- 📓 The journal truly learns from your reactions — the "note → apply → observe your reaction → confirm or correct" loop is unblocked end to end: consolidation no longer stalls on small journals (the portrait stays fresh), and a note can no longer declare itself "validated" without real evidence.
- 🧭 Habits: the detection bar is shown in plain sight — when no activity window is detected, Settings now display the bar actually applied (it is stricter on weekends, where observed days are fewer). No more unexplained "no habit detected".
- ⚖️ Fair proactive distribution — account selection on each notification cycle is now randomized and filtered in the database: no account can be silently left behind the others anymore, and accounts that disabled the feature no longer consume a slot.
v1.30.12 — 🧱 A typed foundation under every conversationAugust 2026
- 🧱 The run context is now typed and verified — the "who, where, with which preferences" of every conversation used to travel in a 17-key bag with no contract; it now rides in a typed block validated at the graph's entrance: if it is missing, the application refuses instead of carrying on silently — including when resuming a pending confirmation.
- 🌍 A sub-assistant inherits the whole context — language, timezone, consented location: the hand-made copy that kept only 6 values out of 17 is replaced by full inheritance. A German-speaking profile can no longer get a sub-assistant answering in French by default.
- 🛡️ The tool contract facing the model is continuously guarded — every tool must prove it presents itself correctly to the model, with no internal parameter leaking; two new automatic guards lock that contract, and the orchestration libraries are up to date.
v1.30.11 — 📊 Model pricing is administered in ExcelAugust 2026
- 📊 The whole pricing grid exports, edits and re-imports in one go — the 124 models and their tariffs download as a structured Excel workbook: a translated usage notice, dropdown lists, cell validations and a referentials sheet. Fix the prices offline, re-import, and the application shows every change field by field before writing anything. No more 124 dialogs when a provider revises its grid.
- 🛡️ An import cannot break the catalogue — a row missing from the file never deletes anything: a forgotten Excel filter cannot empty your catalogue, and a removal is declared explicitly through an "active" column. The import is all-or-nothing: any problem is reported with the sheet and the cell to fix, and nothing is written. If a model changed between the preview and the confirmation, only the affected rows are refused — not the whole file.
- 💶 Two billing errors fixed — a dated model that owns its own tariff was billed at its base model's price (
gpt-4o-2024-05-13counted 2.50 instead of 5.00 on input), and two read paths could return different prices for the same model at the same instant. Both are repaired, and a database constraint stops the second from coming back. The workbook also states in words which models are billed zero for lack of an active tariff, and shows time-slot pricing on the row that carries the price. - 🔌 A deactivated model can be switched back on — deactivation had no inverse in the application. A retired model is now reactivated from the workbook, and a cached-input price can finally be cleared rather than endured.
v1.30.10 — 🧭 Your capabilities, up to date and on every cardAugust 2026
- ✨ The capability map finally shows everything LIA can do — image generation, document generation, plugins, learned habits, your MCP servers and phone calls were all missing from “Capabilities”: the page whose whole job is to be current had quietly fallen a few releases behind. It now maps nineteen capabilities instead of thirteen, and a capability an administrator switches off disappears from it immediately instead of being announced as available.
- 🧭 The settings home page now says where each section stands — “Memory — active, 128 items”, “Connectors — to set up”: every section that holds something says so on its card, before you open it. The whole page reads it in a single request, and stays silent rather than guessing while the answer is on its way or when a section has nothing to report.
- 🧩 MCP servers move in with the other extensions — “Application MCP” and “MCP servers” leave Preferences and join Features › Extensions & Data, beside Skills, Plugins and Spaces: a server that hands the assistant new tools extends it, where a connector links a personal account. Your existing links keep working. The FAQ also gains a Plugins section, and its settings instructions now describe the page as it actually is.
- 🆕 The public site shows what just shipped — the latest releases are now told in plain words on the home page, with the full history one click away. Visitors can see the product moving before they even create an account.
v1.30.9 — 🧭 Settings, redesigned: everything one click awayAugust 2026
- 🧭 A modern master-detail settings page — settings are no longer a wall of collapsed lists: a permanent side rail shows every section, grouped and iconed, and the chosen section opens fully in a panel beside it. With nothing selected, an overview presents every setting as a card with its description — finally visible before you open anything. On the phone, the list comes first and each section opens full screen with a back button.
- 🔎 Search now covers administration, and section links can be shared — administrators find every admin panel (users, pricing, LLM configuration…) through the same quick search, and every section has a stable address: reload or share the URL and you land exactly there. If a section isn't available on your account, the page says so plainly — and keeps watching, in case it simply took time to load.
- 📄 Generated documents open in the app — click a document card and it opens in a viewer tab: CSV as a real table, Markdown and text nicely rendered; Word, Excel and PowerPoint present a clean download panel, and PDFs keep opening straight in the browser.
v1.30.8 — 📄 Document generation: real files, on demandAugust 2026
- 📄 Ask for a real document, get a real file — LIA now writes complete documents on demand: CSV and Excel spreadsheets, Word reports, PowerPoint presentations, PDF, Markdown and plain text. Ask in the chat (“research the latest LLM models and turn it into a CSV”, “make a detailed presentation about Alsace”) and the finished file appears as a downloadable card below the answer — it can chain directly on a web research done in the same request. A dedicated writer model produces the content in your language; spreadsheets are safe by construction (no formula injection) and Excel-ready.
- 🕒 Files that say when they expire — every generated document card shows its filename, its size and its exact expiry deadline; PDFs open right in a browser tab, everything else downloads in one click. Generation is switchable instance-wide by the administrator and covered by the same usage limits as the rest of your AI usage.
v1.30.7 — 🧩 Agent Plugins: install portable pluginsAugust 2026
- 🧩 Install plugins from the whole AI ecosystem — LIA speaks the open Agent Plugins standard (agent-plugins.org, backed by AWS, Microsoft, OpenAI, Cursor and Vercel). A plugin packaged for ChatGPT, Cursor, GitHub Copilot or VS Code installs into LIA unchanged, from a zip file or an https link, and brings its skills and MCP servers in one step. Every install shows a full report — what was installed, updated, skipped (and why) or removed — and uninstalling removes everything the plugin brought, in one action. New “Plugins” section in Settings, in all six languages.
- 🧹 Account deletion now removes every skill and plugin file — deleting your account already purged your data from the database; it now also removes your imported skill files and installed plugin files from disk, completing the clean-slate guarantee.
v1.30.6 — 🔌 New-generation MCP protocolAugust 2026
- 🔌 LIA speaks both generations of the MCP protocol — the Model Context Protocol published a major revision that older clients cannot talk to. LIA now speaks the new stateless revision AND automatically falls back to the classic handshake: every MCP server you already configured keeps working unchanged, and new-generation servers plug in too. If a server requires a revision LIA does not speak, you now get a clear explanation instead of a cryptic connection error.
- 🔐 Hardened OAuth for MCP servers — and declining is no longer an error — the OAuth flow applies the new specification's security obligations: the authorization server's identity is verified before your authorization code is used, and stored client credentials are never sent to a different authorization server than the one that issued them (LIA re-registers automatically instead). If you decline the consent screen at the provider, you come back to your settings with a simple informational message — not an error page.
v1.30.5 — 🕐 Costs that follow the provider's clockAugust 2026
- 🕐 LLM costs now follow peak and off-peak hours — some providers (like DeepSeek) charge half price outside their peak UTC windows, and LIA used to count every call at the full rate. The pricing catalogue now supports time-based tariffs: each call is valued at the tariff of its exact moment, so your usage statistics and the instance spend ceiling reflect what the provider actually invoices. Admins define the windows in the LLM pricing settings — any provider, any number of windows.
- 📱 The pricing forms work on your phone again — the create/edit dialogs for LLM, Google API and image tariffs could be taller than the screen with their top out of reach, making them unusable on mobile. The scroll behaviour is fixed on all three: short forms stay centered, tall ones scroll from the very top.
v1.30.4 — ✅ Confirmations that always tell the truthAugust 2026
- ✅ A refused action is never again announced as done — a real incident showed it: an oversized relayed message was refused internally, yet the answer said “Message delivered”. Fixed at the root: every tool's verdict (success or failure) now travels unchanged all the way to the answer. If something could not be done, LIA tells you, with the actual reason.
- 📏 An oversized relayed message is explained with its numbers — messages relayed to your connections are limited to 2,000 characters (instance-tunable). If yours exceeds it, LIA now gives you the exact length and the limit, and offers to shorten or summarize it — instead of a generic question that never named the real problem.
- 🔢 Bulk-operation confirmations show the exact count — the confirmation card states the number of items actually measured (never the theoretical cap), with correct singular wording when there is only one.
v1.30.3 — 🧱 Foundations that no longer depend on someone else's weatherAugust 2026
- 🔤 Fonts now ship with the application — building a new version no longer downloads fonts from Google: the exact files (same families, same weights, identical rendering) live in the repository. Two updates had failed the day before because Google moved its files at the wrong moment — that failure class is gone, and an automated guard forbids its return.
- 📦 Self-host installation is qualified end to end — for administrators: every candidate version is now actually installed, on two architectures (PC and Raspberry Pi), in a disposable hermetic environment, before any publication. Seven defects in the qualification process were fixed one by one, each proven on a real run. This version ships the repaired end-to-end chain: build → qualification → publication.
v1.30.2 — ⚙️ An up-to-date agent engine and bounded spendingAugust 2026
- ⚙️ The agent engine catches up on five months of fixes — the LangChain/LangGraph ecosystem orchestrating every conversation moved to its latest versions, fully validated in an isolated environment first (8,489 tests executed before touching anything). In practice: fewer wasted calls, more reliable resumes after confirmations, and a deep fix to nested-agent persistence.
- 💸 Paid tools get a per-request ceiling — generating an image or querying Perplexity or Brave is now bounded per request: an agent looping on a billed API is stopped gracefully with an explanation instead of spending silently. Six tools covered, configurable, verified at startup.
- 🧠 History is no longer lost when its summary fails — when a long conversation is summarized to fit the model's memory, a failed summary used to mishandle the history; it is now fully preserved. And each model's memory window is read from the managed catalogue, not from a frozen table.
- 🔌 Closing the tab really stops the work in progress — an abandoned response stream now cancels the sub-agents it had launched, instead of letting them burn tokens on an answer nobody will read.
v1.30.1 — 🧾 Exact accounting and failures that name themselvesAugust 2026
- 📊 The “successful actions” counter finally works — it had shown 0 since forever: an internal classification defect counted every completed action as a plain answer, and “confirmed useful results” shrank to routines alone. Fixed and test-pinned — and your cycle's history has been reclassified, not just the future.
- 🧭 Repairable failures name themselves — rejected API key, exhausted provider credit, nonexistent model name: each case now has its precise message in all 6 languages, pointing at the exact screen where the repair lives. No more “service saturated, retry” on errors retrying will never fix.
- ⏱️ The timeout setting is actually applied — the admin “timeout” field existed from the screen down to the database, then nobody read it. It now reaches every provider, and the defaults were recalibrated against 30 days of real production latencies — without ever cutting a call that used to succeed.
- 🧾 Token accounting becomes a contract — every provider declares its accounting mode (verified at startup), the count is requested explicitly, and a paid call completing without usage triggers an alert: the spend ceiling can no longer go blind in silence.
- 🤖 A more resilient assistant on imperfect answers — truncated JSON, a trailing comma or prose around the payload are recovered instead of causing costly retries, and an empty answer is no longer memorized and replayed for five minutes.
v1.30.0 — 📍 On the move, LIA answers from where you areAugust 2026
- 📍 Your real position, everywhere — when your phone's live position is unavailable (an app left dormant does this), LIA now uses your last remembered position — if you enabled it — before falling back to your home address. This applies to chat, scheduled actions, the briefing and skills alike: 'restaurants nearby' while travelling stops answering from your living room.
- 🕐 A dated position says so — 'where am I?' without a live GPS used to answer 'enable geolocation'. It now answers from your remembered position with its age stated: 'based on your last known position at 9:30' — never presented as your current spot. And 'at home' is never resolved from a position captured on the road.
- 🔔 The mobile app wakes up by itself — on return to the app, your position refreshes silently when the permission is still granted; when the system dropped it (iOS does this after inactivity), a banner offers to re-enable it right when you open the chat — before you have to ask. The 'remember my last position' setting now lives on the Google Places connector, next to the home address, with the stored position, its date and a delete button.
- 📞 Honest phone-call errors — when the telephony provider rejects your key (ElevenLabs now requires keys starting with sk_), LIA says exactly that and what to do, instead of a misleading 'try again in a moment'.
- 🛠️ Four production fixes — the autonomous ReAct mode no longer fails on Gemini when an MCP tool declares a list parameter; the LLM admin dialog no longer loops on a 422 error when switching models; skill widgets keep their Docker access after every deployment; and a fresh multi-worker start boots in about a minute instead of respawning for twenty.
v1.29
v1.29.0 — 🏠 LIA installs on your machine, and an instance knows what it spendsAugust 2026
- 🏠 Install LIA on your own machine, guided —
./install.shasks a short questionnaire in your language, builds the images from the source you cloned, creates your administrator account and your provider keys without ever writing a secret on the command line, and then checks that the installation really works — not merely that the server answers. If a step fails,./install.sh --resumepicks up exactly where it stopped. - 💶 A daily spend ceiling for the whole instance — an administrator can now cap what the entire deployment spends in a day, in euros, and not only what each account consumes. First come, first served; afterwards LIA says come back tomorrow and gives the exact time it resets. The panel shows today's spend, the ceiling that really applies, and what is left.
- 🎛️ Switch a feature off in two seconds, without redeploying — dictation, speech synthesis, images, uploads, document spaces, web search, browsing, skills, MCP and telephony are now ten switches in the admin panel. Each row shows what the deployment allows, what the administrator chose, and what is actually applied — a capability the deployment forbids says so instead of pretending to be adjustable.
- 🎬 A guided demonstration, at /demo — six missions, one per mechanism that sets LIA apart: a crowded morning, a proactive alert, long-term memory, a phone booking, the daily briefing, and a tour of the settings. Approve, edit or refuse each prepared change, then read the final reply. Everything is labelled as a demonstration on synthetic data at every step, no external service is contacted, and a proof drawer links each element to its source.
- 📜 Terms of use accepted at sign-up — creating an account now requires an explicit acceptance, recorded with its date and version. The terms are written in the six languages, and this applies to every sign-up path.
- 🔒 Three faults the green test suite could not see — speech synthesis was billed but never counted against the spend ceiling; signing in with Google skipped the terms acceptance entirely; and eleven connector paths could link a real credential with no guard at all. All three are closed, and the new guards were deliberately broken to prove they go red.
v1.28
v1.28.0 — 🧭 LIA learns your habits, under your controlAugust 2026
- 🧭 LIA can learn your habits — if you turn it on — your activity rhythm (your usual weekday and weekend slots) and your recurring requests ("every Monday morning, the emails") are learned through transparent statistics, never an opaque model. Everything is off by default, and everything lives in Settings → Learn my habits. Nothing leaves your account.
- 📊 A panel that shows everything — a heat map of your 24 hours with an hour axis, your active-day percentage, your detected windows — and when there is no stable habit, the panel says so honestly instead of inventing one.
- ⏳ You always know where learning stands — a progress bar shows observation days acquired versus required, and recurring requests under observation appear with their progress ("3/4 distinct days") before they are even confirmed.
- 🔍 "Why this habit?" — every habit explains itself: the real days it was observed and the exact thresholds the detector applies. A displayed habit is proven, or it does not exist.
- 🕰️ Notifications at the right moment, without nagging — LIA can notice a missed routine (at most once a day, and it goes permanently quiet after two ignored remarks) and prefer your usual slots for its notifications — never exceeding the hours YOU set.
- 🧹 Total, immediate control — pause a habit, block it forever (never relearned), delete everything, or relaunch learning over your history in one click: the "Recompute now" button acts immediately and retroactively.
v1.27
v1.27.14 — 🛡️ Updates that leave the app runningAugust 2026
- 🛡️ An update no longer degrades the app while it installs — during a release, push notifications, the FAQ and the skills briefly stopped working: the update replaced the files underneath the running application. Updates now prepare everything aside and switch over in one move, so nothing is missing while the new version is being built.
- 💬 A conversation in progress is no longer cut off by an update — a chat being streamed was interrupted when a new version was installed. The application now finishes what it started before it stops.
- 💾 Database backups are actually kept — the scheduled dumps were stored inside the folder each update replaced, so they were erased every time. They now live outside its reach. If you administer an instance, check that your backup path points outside the deployment folder.
- 🌍 Messages from the assistant come back in your language — some replies and tool results fell back to their English source text whatever language you had chosen. The translation catalogues are now shipped with the application.
- ⏳ An account awaiting activation sees the page that explains the wait — a new sign-up whose account had not been activated yet landed on an application that looked broken, with every screen refusing to load. It now goes straight to the page describing the next step.
- 🔔 Push notifications stop chasing a device that no longer exists — when an app was uninstalled or its notification permit rotated, the assistant kept trying to reach it for days. Such a device is now retired on its first definitive refusal.
v1.27.13 — 🔁 An action never replays itselfAugust 2026
- 🔁 A clicked action no longer re-runs when you return to the chat — “Prepare a reply to the email…” could execute again when the browser resurrected the click's address (history, session restore). Every click now carries a one-shot id: a replayed address only prefills the composer, and you decide whether to send. “Run it now” links on proposed routines stay deliberately replayable — there, each click is a consent.
- 📥 Rich-HTML answers no longer crush their key-value lists — a long label (an email subject) squeezed the value column into an unreadable sliver overflowing the bubble. The layout is fixed, and answers already received display correctly again — nothing to do on your side.
- ⭕ The context gauge is visible on mobile — the pill showing how full the conversation is (with its totals tooltip on tap) joins the centre of the chat header, right of the knowledge-spaces indicator, at every screen size.
- 📊 The dashboard's Consumption opens expanded — messages, tokens, Google requests and cost are readable without an extra click; the section still folds if you prefer it quiet.
- 📤 Export is a visible button on phones — Long-term memory, Interests and Personal journals now show Export right in the toolbar instead of folding it into the “⋯” menu; the bar wraps cleanly on small screens.
v1.27.12 — 🔬 The debug panel reads like the runAugust 2026
- 🧭 The trace follows the execution order — seven numbered phases (request, router analysis, planning, execution, response context, background extraction, totals) replace an accumulated ordering where the query appeared after the routing decision.
- 🌊 A waterfall on the run's own timeline — every model call is positioned by its real start, shared between the pipeline and its background tasks: what ran in parallel and what serialized is read at a glance.
- 🔍 The invisible stages now have a section — the plan validator's verdict (informative: a rejected plan still executes), the ReAct loop against its published iteration cap, approval gates, context compaction, image-generation and voice-synthesis spend.
- ⚠️ Anomalies surface on the request header — failed steps, panic planning, critical token zones and payload drift are counted on each history entry, next to a scannable strip (route, engine, duration, tokens, cost): compare requests without unfolding them.
- 🧹 Idle sections fold away — on a conversational turn, two thirds of the sections carry no data; they collapse behind one per-phase disclosure instead of stacking N/A rows.
- 🎨 Readable in both themes — statuses take the design-system tokens of the five themes, pipeline nodes wear bi-theme identity families, and a stage that simply did not run shows neutral instead of a red FAIL.
v1.27.11 — 🧭 Settings that read like an indexAugust 2026
- 🧲 The gap between a label and its field finally renders — a defect in the label primitive meant every spacing setting was computed but never drawn: all forms lived with ~3 px whatever the code said. Fixed at the root and calibrated on real screenshots — every form now breathes the same way.
- ☰ One grammar for row actions — visible icons from tablet up, delete carrying its red at rest, and on phones a single named “⋮” menu. No more buttons revealed by hover that a keyboard user landed on without seeing anything.
- 📤 Export is back on phones and tablets for long-term memory and interests — it was simply removed below a desktop width — and “+ Add” keeps its label at every size.
- 🗂️ LIA Connections read as five foldable, icon-carrying sections whose badges say what waits inside — pending requests included — and both share directions display as mirrors: yours editable, theirs read-only.
- 📔 Personal journals fold their whole configuration behind one entry, move their buttons below the categories, and long entries clamp to three lines with “Show more”.
- ⏰ A routine says “Weekdays at 08:00” instead of naming five days, shows its next run and folds the rest — its card drops from about eight always-visible lines to three.
- 🔐 Authenticator apps (TOTP) now enroll from a switch, like every other feature — turning it on opens the ceremony, and the thumb only moves once the server confirms.
- 🧭 Connector groups are named “Google / Apple / Microsoft / external services”, and a data export finally explains its state: size and expiry when ready, generation and expiry dates plus what to do when expired.
- 🎛️ Every title carries an icon in the theme colour — notification frequency, travel weather (now foldable), each dashboard card in the customisation list — and grey badges are reserved for the inactive: “Enabled” is green, “pending” is blue.
- 🔔 Push notifications follow your global opt-in automatically — the per-feature duplicate switch that could silently mute proactive notifications is gone.
- 🫧 What comes from LIA stands out — in the thread, a proactive notification carries a light red wash that tells it apart from an answer; a pinned memory is framed in the theme colour, its pin set right before the text. On phones the action bars get compact — “Add”, “Clear”, “Import”, “URL” — to fit on one line, and a connection card frames “what I share” and “what they share” separately.
v1.27.10 — 🎨 Screens you can read at a glanceAugust 2026
- 🔴 A “high” notification no longer looks like a “medium” one — both rendered on a pale ground in two very close hues: on a real account, 89 high and 113 medium notifications were indistinguishable. Urgency is now carried by density, not hue alone: high fills, medium tints, low stays neutral. The difference holds in greyscale too.
- 🔢 The Alerts counts show before you open anything — the number that decides which section to open could only be had by opening it. The five totals now arrive with the page; the lists still wait for the fold.
- 🔘 Action buttons look alike everywhere — the relationship-card actions and the Alerts shortcuts adopt the outlined style used throughout the app.
- 🧠 A memory shown on a relationship card is now clickable — it was dead text: correcting it meant finding Settings, then the right section, by hand.
- 👤 A peer connection carries the app's own initials avatar, with its name in bold — a list of connections was told apart by reading, not by looking.
- 🎯 Contact-card icons take one colour per family — reach, dates, ties. The card already answered three questions; colour makes that structure visible.
- 📨 “Sent” and “received” each have their own colour — on relayed messages, on a relationship card's messages, and on exchanged emails. Only a small arrow separated them.
- 📅 Shared meetings and exchanged emails finally have a count, like their five neighbours — and every count in the app now wears the same colour.
- 📞 A call's objective is told apart from its summary — and a long objective no longer spills out of its tag: a tag is sized for a word, not a sentence.
- 📱 On a phone, two dense blocks are no longer shown — “Why does LIA think this?” and the interest-weight explanation pushed what you came to read off the screen. They stay whole on a wide screen.
v1.27.9 — 🔭 Why LIA thinks that, and what it can doAugust 2026
- 🔍 “Why does LIA think that?” under every memory and journal entry — a folded block shows the signals behind the conclusion, their date and their role, with a Correct button. What is kept is a reference, never a copy: your text stays where you wrote it. Delete the conversation and the reference empties while the dated row stays — the content never comes back. Settings → Memories, or Settings → Journals.
- 🔔 An “Alerts” destination for everything addressed to you — relayed peer messages, proactive notifications, interest notifications, pending reminders and scheduled routines, in one screen instead of five different settings panels. Each section arrives folded with its count already shown, and paginates by ten: you choose what to open without having to open it to find out. No list is fetched until something is unfolded. Your usual settings links keep working.
- 🌌 A star chart of your capabilities — each capability is a star: filled when active, outlined when waiting to be lit, its size saying what it holds. The figure joins only the live ones, so its shape is your configuration. No level, no percentage, no comparison with anyone. Home → Capabilities.
- 📈 An interest's weight explains itself — the originating signal, the last mention, the last notification and the calculation itself. Explaining uncertainty is more useful than turning it into a score. Settings → Interests, under each topic.
- 📞 No option raised on a call is accepted on your behalf — debrief follow-ups now prepare the message instead of sending it, and what the other party actually proposed (date, place, extra cost) is shown before the possible follow-ups.
- ✍️ Your card titles are no longer truncated — up to four action chips per row ate a third of the width. One trigger now opens the action menu, so every row of every card has the same text width. The full label appears on hover.
- 👍 Your feedback on an interest notification is finally recorded — the archived card and the recording route used two different identifiers. One vote also marked every other card in the conversation as judged: fixed in both directions.
- 🎨 Thirty-one places in the interface painted nothing, or painted black — health and personality chart tooltips, gradients, animations: a colour syntax inherited from an earlier version of the styling engine, invalid since the migration.
- 🧮 Your interest ranking and the displayed weight no longer disagree — two calculations each carried their own default value; there is now only one.
- 🔗 “LIA connection” is the last section of a relationship card on mobile — as on a wide screen, and only once in the page.
v1.27.8 — 🧭 What LIA produced, and what may interrupt youAugust 2026
- 👍 Your thumbs-up on a proactive notification is finally recorded — the archived card carried a fabricated identifier the recording route could not resolve, so every vote failed silently. No feedback had ever been recorded. Already-archived cards no longer show the button at all, rather than offering a gesture with no effect.
- ⏰ A routine no longer runs twice on the night the clocks change — the scheduled time exists twice that night, and the routine fired again sixty minutes later. 54 double runs measured across the 2026 transitions of seven time zones. Now at most one run per local day — without ever dropping today’s run when you test a routine ahead of time.
- 🔇 Eleven switches: being connected to a service is no longer being interrupted by it — to stop mail-driven nudges you had to disconnect the connector, losing the tool you ask with. Settings → Proactivity → Notification topics. Everything is on by default, and a switch that cannot produce anything says so (leave-by advice needs the calendar).
- 📜 The settings finally show what they produced — the last ten proactive notifications and, identically, the interest ones: date, content, sources, verdict already given. Until now you could tune a frequency without ever seeing the result. Interest notifications older than this version show without their text: it was not kept, and a reconstructed summary would be worse than an absence.
- 📁 Three blocks fold, closed by default — the eleven switches and the two histories. The panel becomes an index you open rather than a wall you scroll past, and a closed block costs no request. The number of refused topics stays on the header: folding must not hide a decision.
- ✨ The empty chat offers what LIA already knows about your day — a meeting to prepare, a batch of mail, a commitment waiting, a pending reminder. Nothing is fetched to produce them: no connector woken, no quota spent. With nothing to go on, the generic starters come back.
- 🏆 The dashboard opens on your results, not your consumption — results confirmed useful, successful actions, routines run, commitments closed. Messages, tokens and costs move under “Consumption”. “Time saved” was dropped rather than estimated: nothing here measures it.
- 🤝 Share an answer with a connected peer — under an assistant bubble, the “…” menu lists your accepted peers and prepares the request in the composer. Nothing is sent without your confirmation.
- ✅ A commitment’s actions are where you read it — mark done, dismiss and correct, straight from the “For you” card on the dashboard. The full ledger stays one click away.
- 🎨 Also — LIA’s portrait is chosen from two visible thumbnails; a routine duplicates without creating anything before you confirm; the preview of the next five runs shows the time in the routine’s own zone and flags clock changes; a reminder cancels from its card; the photo gallery is fully keyboard-navigable; three quick commands join the list; and a brief haptic accompanies gestures on mobile, set separately from reduced animations.
v1.27.7 — 🗣️ A real question rather than an error messageAugust 2026
- 🗣️ When LIA is stuck it asks a question — not its own diagnosis —
for_each pattern issue detected,Fabricated placeholder contact detail: step_2.to='[email protected]': that is what a French-speaking account received when LIA refused to run a doubtful plan. Jargon, a code path, and a made-up address you could believe came from your own address book. Now fifteen clear questions, in all six languages — and the app refuses to start if one is missing. - ✉️ An address you gave is never reinvented — you specify a recipient and, two turns later, LIA writes to
[email protected]. The value you gave is now carried over. A genuine change of mind on your side is never overwritten. - ✍️ Correcting a commitment LIA misheard — the ledger fills itself from your conversations, so it is sometimes wrong: an approximate wording, a "by Friday" read as the wrong Friday. The text and the deadline can now be edited. The other party and the meaning cannot: changing them would describe a different commitment, not correct this one.
- 🤝 A commitment's actions, on the person's card — done, follow up, no longer relevant, correct: those gestures existed only in the settings. They are now on the matching relationship's card as well — where you actually think of them.
- 🏷️ "Open loops" is now called "Commitments" — and the section moves up to Settings → Features, next to Interests. The settings search knows both words: if you knew it under the old name, you still find it.
- 📇 A relationship card reads better — the contact card always opens the left column (knowing who this is comes before everything else) and the LIA connection always closes it. The details now read as three blocks: how to reach, which dates matter, who the person is tied to.
- 🔁 An instruction is no longer lost when you answer a question — you answered a request for clarification and, with nothing to show for it, the instruction your answer was meant to trigger vanished when the conversation resumed. It survives now.
- 🧭 What LIA claims it can produce is checked before every delivery — thirteen tool descriptions promised data they do not return: weather, web search, Wikipedia, reminders, routes, task lists. LIA built plans on those promises and failed halfway through. An automated check now compares every promise to the real result.
- ⚙️ For the administrator: the skill-creation guide described steps that do not exist — it announced two step types the engine rejects and kept silent about two it accepts, so a skill written by following the documentation was rejected at execution. Fixed in all six languages, and a check now blocks any divergence between the guide and the engine.
v1.27.6 — 🔗 Every request leaves with the right nameAugust 2026
- 🔗 A request no longer leaves carrying the previous name — a 360° overview of someone, then another person's card opened and launched, sent the first one's sentence. The browser alone decides the address now: what you launch is what you asked for, including when the link only prefills the input box.
- 📞 A question about a call gets an answer, not a call — "when did I last call Marie?" could produce a plan to phone her and ask. LIA can now read your past calls, your open commitments and your relayed messages, instead of only having the means to act.
- 🤝 Merge two cards that are the same person — LIA groups what is written the same, accents and capitals aside. It cannot guess that a raw phone number and a name are one relationship: you tell it, once, from the bottom of the card. It is reversible, and you always see what was merged.
- ✉️ An excerpt of the email under its subject — "Re: Re: catch-up" says nothing about what the exchange was about. Up to two lines now appear under the subject, on the relationship card.
- 📇 Calling a number no longer creates a second card — calling "06 12 34 56 78" then "Marie Martin" built two relationships for one person. The number is recognised and attached to the right card.
- 📱 Buttons you can actually hit with a finger — six controls were too small on a phone screen, including the cross that minimises the companion: invisible without a mouse, yet clickable. They now meet the recommended size.
- ⚙️ For the administrator: LIA can know what it already knows about a connected contact — when a connected person is named in the conversation, their commitments, calls and your relayed messages can be handed to LIA directly rather than announced as a lookup. Off by default, enabled by
PEER_CONTEXT_INJECTION_ENABLED, and limited to what you ticked in the 360° scope.
v1.27.5 — 👥 The 360° overview of a person, complete and configurableAugust 2026
- 👥 The 360° overview of a contact works again — asking for a 360° overview of someone answered that LIA could not retrieve their recent interactions. The tool did exist, but was never loaded at startup: LIA picked it, then could not find it. It now finds the mail exchanged and the meetings shared, searching by address rather than by name — far more reliable.
- 📇 The contact card shows your whole address book — nickname, job title, birthday, note, postal addresses, family relations, links, important dates and messaging handles join the name, organization, emails and phone numbers. What your provider does not store simply does not appear, rather than showing up empty.
- 🎯 You choose what the 360° overview reads — a new section at the top of a relationship card lets you tick the sources to use, tell received exchanges from sent ones and meetings the person attends from those they organize, and set how many items each section contributes. Your choice is saved before the chat opens: LIA applies it, it does not guess it.
- 🔢 LIA no longer under-counts what it summarizes — the card shows "137 commitments" and lists only a few; the assistant received those few rows alone and answered "you have five". It now receives the exact count. And when a source could not be read, it says so instead of announcing there was nothing.
- 👤 The 360° overview covers the person on screen, guaranteed — pressing the button used to send a sentence LIA had to re-interpret; it could pick generic tools and miss the point. The button now carries the request itself: the 360° tool is in the plan, on the right person. Other useful lookups are kept.
- 💬 The assistant finally receives everything the 360° found — it could answer "I have no data at hand" right after retrieving open commitments, calls and relayed messages. The content stopped on the way, before reaching it. It now arrives in full.
- 🤝 A LIA contact who is not in your address book finally has their mail and meetings — exchanges are found by address, and a connected person you never saved had none. If they turned on address sharing, LIA uses it to find, in your mail and your calendar, what concerns you both.
- 📅 No more event of YOUR calendar presented as theirs — with no shared meeting to show, the assistant filled the gap with the next entry in your own calendar, one the person neither organised nor attended. An empty section is an answer; a meeting without them is an error.
v1.27.4 — 📧 Asking for 3 emails gives you 3 emailsJuly 2026
- 📧 LIA no longer claims it failed when it succeeded — asking for "my last 3 emails" sometimes answered that the retrieval had been blocked and that you should go check your email connector. The emails had been fetched, the connector was healthy, and the whole diagnosis was wrong. The answer now reports a failure only when something genuinely produced nothing.
- 🧩 A partly blocked request shows what it did get — when one capability fails while another works, LIA presents the results it obtained and names what was missing, instead of announcing a total failure.
- ⚙️ Per-service result limits are honoured to the letter — the limit each connector sets on how many items can be read at once is now handed to the planner, which sizes its requests accordingly. Any value above it is brought back within range before execution rather than failing the plan.
v1.27.3 — 🤝 Asking about a connected person now worksJuly 2026
- 🤝 « Is Jérôme free tomorrow at 10? » — asking about someone you are connected to now answers from THEIR calendar, the one they chose as their default, at the hours you read on your own clock. All-day entries (a birthday, a reminder) are told apart from real meetings, so a birthday no longer makes someone look busy all day.
- 🎯 LIA no longer guesses when it is blocked — when a request cannot run, the answer names exactly which capability was missing and stops there. No more sweeping « nothing is configured » about services that work perfectly well, and no more blaming you for a fault that may be on our side.
- 🔌 A disconnected service now stays visible — instead of a single alert you may have missed, a discreet banner remains for as long as the connection is broken, with a Reconnect button. It disappears by itself the moment the service works again.
- ✏️ Your /shortcuts are editable — you can now change the name or the text of a shortcut you already created, instead of deleting it and starting over.
- ✨ A few comfort touches — the chat composer takes the same frosted-glass look as the header, the settings search moves above the tabs, card titles line up with their icon, and the « How LIA works » introduction is one sentence again.
- 📧 « Summarise this email and draft a reply » now completes every time — this request used to fail, then go through minutes later with nothing changed. LIA could end up having to reply to a message without any way to read it. It now checks it has everything it needs before starting, instead of finding out midway.
v1.27.2 — 🌌 The public site gets its identityJuly 2026
- 🌌 A new visual identity — the home page and every public page (More, demo, Story, Philosophy, Technical, FAQ, blog) move to the cosmic theme: starfield background, a planetarium of features orbiting the live demo, and animations that play at the pace of your scroll.
- 🌓 Dark by default — public pages open in dark when you have never picked a theme; your theme button always wins, and your choice is remembered.
- ✅ Better readability — buttons, badges and pills all move back above the recommended contrast threshold, in light and dark alike.
- 🗓️ A more useful “For you” card — it now shows only the next upcoming automation, no more history. And a thumbs-up or thumbs-down on a proactive notification no longer pops an error message.
- 📱 Settings on narrow screens — the “My connectors” rows no longer overflow, the “Add a passkey” button moves below its text, “Recent calls” sticks to the last ten, and the skill-import buttons sit under their own separator.
v1.27.1 — ⭐ Your relations, front and centerJuly 2026
- ⭐ Relations in the navigation — the personal-CRM page gets its own entry in the bar (Knowledge stays one click away in the chat — its pill now always shows). Redesigned cards: initials avatars, colored signal pills, and a 360° sheet.
- ❤️ Persisted favorites — star anyone in one tap, without opening their card: your favorites keep their own band at the top, survive quiet periods, and a name filter appears once the list grows.
- 🤝 The LIA badge — relations who are also connected users of your instance carry a small badge, bridging your CRM and your connections.
- 🧭 Better findability — searching the settings for peers, social, discovery or loved ones now finds the Connections section; the « How LIA works » cards render their formatting properly.
v1.27.0 — 🤝 Connect with people you knowJuly 2026
- 🤝 Peer connections — two users of the same LIA instance can now connect: enable discovery in Settings, share your exact name, accept or decline requests in one click from the chat or the new Connections section. Nothing is shared by default.
- 💬 Assistant-to-assistant messages — ask LIA to tell something to a connection: you confirm before anything leaves, then the recipient's own assistant delivers it in its personality and language, naming you — and your assistant confirms delivery.
- 🗓 Chosen read-only shares — per connection, you can open your calendar availability (or details) and task titles, in both directions and visible to both sides; every access is recorded in a consultable log, and you can revoke at any time.
- ✨ Frosted-glass interface — dialogs, menus, toasts and the chat header now show a real glass effect (content slides beneath the blur), and native controls (dropdowns, scrollbars) finally follow dark mode.
v1.26
v1.26.4 — 📞 Your call reports, in your languageJuly 2026
- 📞 Phone calls · Reliable report — the report LIA sends you after a call now always arrives in your language, with the structured debrief (key points, commitments, open questions). Some AI model configurations silently degraded it to a raw English technical summary; that whole failure class is now closed.
- 🛡️ Admin · Configuration guardrail — saving an AI model configuration whose reasoning mode would starve the response budget is now rejected with an explicit message stating the required token floor — instead of failing silently on the next call.
v1.26.3 — 🔧 The product cockpit lights up for realJuly 2026
- 📊 Admin · Real numbers, right away — the product dashboard's gauges now compute their first values two minutes after startup (they previously waited a full hour that busy days never granted), and the routines, first-pass and time-to-first-value panels read exact SQL instead of placeholder notes.
- 🔧 Admin · Access repair — the read-only database role behind the product dashboard finally received its view permissions in production (panels showed “No data in response”), and the deployment now ships the maintenance script that keeps them converged.
v1.26.2 — 📊 LIA measures its own usefulnessJuly 2026
- 📊 Admin · Product cockpit — a 26th Grafana dashboard answers the only question that matters: how many people get a genuinely useful result? Validated outcomes (your explicit 👍, or an action left uncorrected), activation funnel, retention, cost per useful result in euros — computed at home, with no third-party analytics platform.
- 📞 Call debrief: the facts — after a call, the debrief now lists the concrete findings the call produced (opening hours, availability, someone's answer…), even when there is nothing to act on. What the call learned, not just what it did.
- 🕵️ Respectful product telemetry — an optional, operator-enabled measurement of the public pages and web performance: anonymous, strictly bounded event names, no identifier, no content, ever. Off by default.
- 🔧 Fixes — an empty call debrief is now truly absent (its scheduled purge really purges); database seed commands follow the configured credentials instead of hardcoded ones.
v1.26.1 — 🎨 Answers with real page designJuly 2026
- 🧩 Rich HTML mode gains real components — colored callouts with titles, badges with icons, collapsible sections, key-value lists, side-by-side columns, numbered steps and key-figure tiles. LIA uses them sparingly, when they genuinely help reading. Display mode "Rich HTML", under Settings > Personalization.
- 🌈 Code comes out highlighted — in Rich HTML mode, code blocks now get syntax highlighting and a copy button, just like in the other modes.
- 📋 Copying an answer gives clean text — copying, sharing or exporting a Rich-HTML answer now produces readable, structured text (and a rich rendering where pasting supports it), never raw HTML code again.
- 🔍 Polish — history search no longer breaks icon rendering; phantom line breaks in HTML answers are gone; inline code is clearly visible again in light theme.
v1.26.0 — 🎯 From reading to actingJuly 2026
- 🎬 Act straight from your dashboard cards — every card now offers immediate actions: summarize a document, draft a reply to a mail, mark a task done, reschedule. The action runs through the chat as usual, with the same approval step for anything that writes to the outside world.
- ✂️ Select a passage of an answer — a small action bar appears (a bottom sheet on mobile): explain, rephrase, translate, turn it into a task or a reminder, save it to memory, or ask a question about it.
- ⏰ Routines that wait for the right moment — a routine can trigger on a condition (a task is overdue, a matching mail arrives, the weather turns, a document is added, an event approaches) instead of a fixed time, with a « propose first » mode that notifies you rather than acting on its own.
- 👥 A personal CRM — see in one place everything LIA already knows about a person: open commitments, recent calls, remembered facts. Reachable from the dashboard's Quick Access.
- 📞 A richer call debrief — after a call: outcome, commitments made, tasks and reminders created, open questions, and a one-tap follow-up draft.
- 🕐 The dashboard tells the truth about freshness — « updated X ago », served from cache, last attempt when something goes wrong, and a one-click retry per card.
- 📚 Knowledge made easier to reach, plus shortcuts — Knowledge Spaces join the menu, the chat indicator is centered and discreet; installed on your home screen, LIA offers shortcuts (today's priority, quick add, push-to-talk), and an admin can create their own « / » commands.
v1.25
v1.25.33 — 📋 Paste, share, drive the demoJuly 2026
- 📋 Paste a screenshot straight into the chat — Ctrl+V (Cmd+V on Mac) with an image or PDF in the clipboard: the file is attached exactly as through the paperclip, same limits, same preview. Paste text along with a file and both arrive.
- 📤 Share or export an answer — a "…" menu at the bottom of every answer: send it to your apps (Messages, Mail, Notes…) when your device offers a share sheet, or download it as a dated Markdown file, accents preserved.
- 🎙️ The send button finally says what it does — empty field: a microphone shows when "hold to speak" is available; some text: the send icon; voice unavailable: a clearly disabled button. No more button inviting you to speak when it can't.
- 🎯 A real drop zone — drag an image or PDF towards the conversation: a clear "Drop your images or PDFs here" zone appears instead of a thin ring around the field.
- 🧩 "Customize" next to "Refresh all" — pick your dashboard cards and their order without digging through the settings: the button leads straight to the right section.
- 🎬 The demo is now yours to drive — on the public home page and /demo: pick the scene (She orchestrates, She anticipates, She calls, She extends herself), pause, replay. Without interaction, the loop keeps cycling as before.
- ✨ A small celebration — completing the starter checklist makes its check sparkle — once, and never if your system asks for reduced motion.
v1.25.32 — 🔍 Find a setting by its nameJuly 2026
- 🔍 A search field in settings — the page stacks some thirty collapsed sections across several tabs. Type the name of a setting: the search takes you straight there, switches tab if needed, opens the right section and places the cursor on it.
- 🗂️ Thirteen more settings become findable — Language, Appearance, Timezone, Font, Display mode, Strong authentication, My devices, Export my data, Image generation and a few others could only be reached by scrolling. They now have their own address and show up in the search.
- 🌍 The search understands your words, in all six languages — every setting carries a translated list of synonyms: "dark" finds Appearance, "2FA" and "TOTP" find Strong authentication, "GDPR" finds the data export, "cron" finds Scheduled actions.
- ✍️ Accents and apostrophes no longer block a search — typing "theme" finds "Thème", and "d'authentification" finally reaches the section whose text carried a typographic apostrophe. The fix also benefits the search in this FAQ.
- ⌨️ Everything works from the keyboard — arrows walk the results, Enter opens, Escape dismisses then clears. The number of matches is announced to screen readers, and the cursor lands on the section you picked.
- 💬 A missing setting says so — some sections only appear when the feature is active on your account. Rather than leaving you in front of a page that did not move, LIA tells you the section is not showing here.
- 📊 The figures on the "Directing an AI that codes" page were out of date — it claimed half the real number of tests and a third of the documented architecture decisions. They have been measured again.
v1.25.31 — 📌 Headers finally stay in placeJuly 2026
- 📌 The top bar no longer scrolls away — the header of the dashboard and of the legal pages was meant to stay pinned while you scroll. It never did: it drifted off with the page, on every screen, since the very beginning. It now stays where it belongs, so navigation is one click away wherever you are on a page.
- 🧭 Settings tabs stay with you — the settings page stacks about thirty sections. Once a few screens down, nothing told you which tab you were in, and switching meant scrolling all the way back up. The tabs now follow you down the page.
- 🍔 The logo menu no longer makes the header vanish — on a phone, opening the menu from the middle of a page pushed the top bar off-screen, and you had to scroll back to the very top before it would work again.
- 📱 Denser settings on a phone — sections give back some padding on small screens only, so you scroll noticeably less. Nothing is hidden or shortened: every description stays whole, and the desktop layout is untouched.
- ⚡ Help and Settings in one compact bar — the two large cards above your briefing became a single bar, freeing roughly 85 px on a phone for the content that matters. Both destinations keep their label and their description, and they are real links now: middle-click or open them in a new tab.
- 👍 Feedback in the same place on every message — the thumbs of proactive notifications and of interest notifications sat inside the bubble, under a full sentence. They now sit next to the copy button, exactly like on an ordinary answer. Once you have voted, your choice stays visible instead of disappearing — shown as chosen, and locked, because that verdict is final.
- ♿ Settings read correctly to a screen reader — each section announced its title twice, and the group headings were not headings at all. The page structure is now what it looks like, which makes it navigable by headings.
- 📖 A section of the technical guide is reachable again — the “Psyche Engine” chapter was missing from the guide's table of contents, in all six languages, since the day it was written.
v1.25.30 — 🛡️ An email can no longer give LIA ordersJuly 2026
- 🛡️ Content written by other people is now labelled as such — an email body, an invitation description authored by its organiser, a web page, a place listing: LIA reads these every day, and anyone can slip instructions meant for it inside them. Every piece of data now carries its provenance, so the assistant knows it is material to analyse, never an order to follow. An unknown data type is treated as external by default.
- 🔎 Manipulation attempts are spotted and named — seven families of trap are detected across the app's six languages: fake system message, "ignore previous instructions", identity switch, requests to send your data elsewhere, invisible characters, hidden HTML comment. Your content is never rewritten: an email stays exactly what its author wrote, with a short note beside it.
- ⏱️ The time you take to approve no longer counts against the assistant — in autonomous mode the agent's clock kept running while the approval request waited for you. Past two minutes of thinking, the work in progress was cut and the answer rebuilt from scratch, losing the steps already completed. The budget now measures actual compute time only.
- 🔁 The autonomous assistant no longer goes in circles — nothing stopped it from calling the same tool with exactly the same arguments forever. On the fourth identical call it is asked to change approach; on the fifth the turn ends cleanly instead of burning your budget in a loop.
- 💸 Autonomous mode costs less and works with more models — its internal instructions were re-sent in full every turn, then at every step of every turn. They now go out once, which lightens the bill and unblocks providers that refused the conversation from the second exchange on.
- 🧹 A setting advertised a feature that no longer existed — the technical panel showed a "hybrid search" switch for memory, while that engine had been gone for several versions. The display and the dead code are removed. Hybrid search is still very real, but where it actually lives: in your knowledge spaces, over your documents. Your memories are retrieved by a dual semantic search instead — over their content and over their keywords.
- ⚙️ Admins: the memory alert no longer cries wolf — it counted the disk cache, which the system reclaims on its own, so it fired permanently on a perfectly healthy service. It now measures the quantity that actually decides a forced shutdown, the metrics service limit was raised, and a diagnostic runbook comes with it. An alert that fires finally means something.
v1.25.29 — 🎯 Blocking a subject now means it never comes backJuly 2026
- 🚫 Blocking an interest finally sticks — you blocked a subject, and LIA re-created it twenty-five minutes later under a barely different name. Blocking was the strongest signal you had, and it was quietly undone. A blocked subject is now invisible to the whole learning loop: it can be neither re-created, nor renamed back into place, nor deleted out of the way.
- 🎯 Far fewer interests you never asked for — a subject you merely asked about became a lasting interest, so the proactive engine came back to you about things you had simply looked up once. Measured on real conversations: half of the exchanges that should have produced nothing were creating an interest; now none of them do, and the genuine ones are still caught.
- 🛡️ Your interest profile is protected from a runaway answer — one ordinary sentence was enough for LIA to propose deleting nineteen interests, that is your entire list. Deletions carried no confidence check of any kind. A single exchange can now remove at most a couple of subjects; beyond that the batch is refused and recorded.
- 👍 Proactive check-in notifications finally accept your feedback — the cards LIA sends on its own initiative had no thumbs at all, only the interest cards did. You can now tell it whether a check-in was useful, and it adjusts.
- 🗂️ Your verdicts now appear in the notification history — the feedback you gave was applied to the subject but left no trace in the history of what LIA sent you. Each verdict is now attached to the exact notification it came from.
- ⚙️ Administrators: the reasoning effort you pick is no longer erased — choosing an effort that happened to match the default silently stored nothing, and the three background analyses ran with no reasoning at all. Your choice is now kept whenever it fits the selected model.
v1.25.28 — 🧠 What never made it into memoryJuly 2026
- ✏️ Edit a skill by talking to LIA — until now LIA could create a skill but never adjust one. You can now ask it to tweak, enrich or fix one of yours: it re-reads the existing skill, applies your request, then shows you exactly what will be replaced and what will disappear before asking for your approval. Nothing is written until you confirm, and the skill's thumbnail is preserved.
- 💬 Your journals and LIA's mood now work from Telegram too — a conversation held on an external messenger fed neither your logbooks nor LIA's emotional state, even with both switched on. The setting was saved correctly, it simply was not passed along. A conversation is a conversation now, wherever you write from.
- 🧠 A request that triggers an action feeds memory again — when you asked, say, to send a message and slipped a fact about yourself into it, confirming the action lost that fact: neither memory, nor interests, nor the journal kept any trace of it. Those exchanges count again.
- 🔎 A contact with a common first name gets their memories back — LIA deliberately ignores contentless messages like ok or thanks. That filter also applied to people's names, and some of those words are real surnames: asking about a contact so named returned no memories at all, without any error message.
- 🌍 Fewer pointless computations across the six languages — a plain ja, gracias, grazie or 好的 triggered a full analysis while an ok or a thanks was correctly ignored. All six application languages are now treated equally.
- 🎯 More accurate interests — a subject you merely asked LIA to explain or look up could become an interest, though it committed you to nothing. LIA now records an interest only if you showed a lasting link to the subject — you love it, you practise it, you already know it, or you came back to it several times in the exchange — and it quotes your own words to justify it. Long-term memory follows the same demand: what you ask about is not what you are.
v1.25.27 — 🔁 Coming back whole after an updateJuly 2026
- 🔁 LIA comes back at full capacity after an update — right after an update the server rebuilds part of its internal index, and it was doing that work four times over at once. The provider refused, and two of the four processes serving you failed to start: for about a minute LIA answered on half its capacity, which showed up as slowness under load. That work is now done once, by one process, while the others simply wait for its result.
v1.25.26 — 🧠 What LIA knows about itselfJuly 2026
- 🧠 Answers about LIA are complete again — the built-in help LIA reads from had ended up holding three copies of every answer, so out of the five passages it consults, only two were actually different. It now holds exactly one copy of each, and if that ever drifts again the next restart repairs it on its own.
- 📚 Help updates reach you — refreshing the built-in help could fail silently at startup, leaving the version LIA quotes behind the product. Those failures are now visible and retried, and while one is being fixed the previous help keeps serving instead of disappearing.
- 🔄 The admin reindex button tells the truth — it announced a rebuild even when the help was already current, and even when another rebuild was already running. It now says which of the three actually happened.
v1.25.25 — 🩺 What was failing silentlyJuly 2026
- 🎨 High-quality images now make it through — a detailed image takes the provider over two minutes, but generation was cut off before that. You got nothing, with no explanation, and no setting could extend it. The time allowed for images is now based on their measured duration: high quality reaches the end.
- 🐈 Asking for an image stays a request for an image — "create a realistic image of a cat" sometimes went off to build a skill instead of drawing. The generation tool, correctly identified as the right one, was no longer even offered. The two now coexist: the requested action runs, and the detected skill adds to it rather than taking its place.
- 🧩 Importing a skill works again — every import failed, from the chat and from settings alike, because a storage folder had become unwritable for the application. The only visible sign was an error message in the middle of a reply.
- 🖼️ Skills finally show a preview — each skill's detail panel displayed an empty frame. All fourteen bundled skills now have their thumbnail: a sketch of what they produce — a calendar grid, a QR code, weather cards.
- 📱 No more replies stuck on your phone — when the phone suspends the tab, the connection dies silently: the screen stayed on "Generating the response…" forever, even though the answer was finished and saved. LIA now detects that silence and reconnects to the running answer on its own.
v1.25.24 — 🧭 Finding what LIA already knows how to doJuly 2026
- 📱 The logo becomes your menu on a phone — below a certain width the navigation bar disappeared and nothing replaced it: from the chat you could reach neither settings nor help without typing a URL. Tap the LIA logo now and the four destinations open, with the current page announced.
- 🎯 Links land you on the right section — the getting-started checklist sent six of its seven items to the top of the settings page, in front of about thirty collapsed blocks. Each item now opens its own section, unfolds it and scrolls to it. Seventeen sections are addressable this way, including from the home cards.
- 🏠 Unconfigured home cards are no longer holes — seven of the nine cards render nothing at all until their source is connected, so a fresh account met two empty cards and seven unexplained absences. One discreet line now names them and links to the setting that switches each on. Cards you hid yourself are never mentioned.
- ⚠️ A warning before the usage wall — until now nothing signalled that your quota was close: the interface simply blocked, mid-task. A banner now appears at 80 % of whichever limit binds first, names it, and gives the reset date when it is a per-period limit.
- 🖼️ Generated images announce their deletion date — they are kept for a limited time and then removed. Nothing said so: the image vanished from the history without warning. Each card now states how long it stays available, and switches to an alert tone in its final hours.
- 🔁 A failed turn replays in one click — when a turn failed you had to find your question and retype it. A Retry button replays exactly the request that failed, even if other messages arrived in between.
- 💬 The FAQ examples are clickable — the hundreds of ready-made phrases in this FAQ had to be copied by hand. One click now drops them into the chat composer, ready to edit before sending. Nothing is ever sent on your behalf.
- 📞 Calls: followed during, explained after — between your confirmation and the recap, LIA went completely silent. A banner now says who she is calling and where the call stands, and a settings section keeps the history. Above all: a call refused by the carrier (unverified source number, exhausted credit) used to block every further call for a quarter of an hour — the refusal is now detected and explained instead of being mistaken for a call in progress.
- 🗣️ LIA no longer asks for the same confirmation twice — for an email, an appointment or a call she warned that the action was irreversible and waited for your agreement… while a preview and an explicit confirmation were coming right after. You had to insist to be obeyed. That duplicate question is gone; the confirmation that matters, the one showing exactly what will be sent, is of course still there.
- ✅ In-app confirmations, and an interface that fits the screen — the most final admin actions went through the operating system's own dialog, which ignores the theme and labels its buttons in the computer's language. They now use LIA's dialogs. Along the way: on mobile the composer no longer hides under the browser's address bar, and on a tablet the logout button no longer sits off-screen in German.
v1.25.23 — 🔍 The tests that never ranJuly 2026
- 🔊 The voice no longer splits numbers in two — when LIA read out "it's 3.5 degrees", it said "it's three." and then, in a separate audio clip, "point five degrees". The dot of a decimal was taken for a full stop. Same for prices, version numbers and web addresses. A dot now ends a sentence only when a space follows it.
- ⏰ A reminder is no longer lost when your quota runs out at the wrong moment — if you had used up your usage budget in the exact minute a reminder was due, it was not postponed: it disappeared for good, with no error and no notification. It is now put back in the queue and reaches you on the next pass.
- 🌍 Notifications and follow-up questions in your language — reminder titles arrived in English for Chinese accounts, and whenever LIA asked you to clarify a change, the sentence was always in French whatever your language. All six languages are now served everywhere, including the wind direction on the weather card.
- ✏️ The value you correct stays in the conversation — when you answered "make it Jean instead" to a confirmation request, your correction replaced your original message with "run this tool with the modified parameters": the name you had just given vanished from the history. It is now preserved for every tool.
- ⚠️ Error messages finally tell you what happened — the detail returned by the server ("this model already exists", "invalid address") was replaced by a generic sentence, and a validation error displayed as "[object Object]". The real message now comes through everywhere. A server failure with an empty body is no longer mistaken for a silent success either.
- 🗑️ Confirmations name what they act on — a confirmation before deleting a label did not say which label, and deleting an email was presented to LIA as sending one, which switched off its caution rule for deletions. Every action is now announced for what it is.
- 🔧 On the admin side — the cost of each model call re-read the exchange rate from the database on every computation; it is now genuinely cached. The personality panel survives an expired session instead of freezing, and usage tracking shows the exact reason for a refusal rather than a generic message.
v1.25.22 — 🔒 A sign-out that leaves nothing behindJuly 2026
- 🔔 Signing out really does stop notifications on that device — the push registration outlived the sign-out. On a shared computer or phone, the next person saw your notifications arrive on the lock screen without ever having signed in to your account. Signing out now revokes that device before closing the session.
- 📍 Your location no longer follows the next account — the last known position and the geolocation permission stayed in the browser after a sign-out. On a shared profile, the next account inherited your position for weather and places, and above all inherited your consent: the browser kept collecting the location of someone who had never agreed to it. Both are now cleared.
- 🔒 Nobody else can switch off your notifications — revoking a device did not check who owned it, so any other signed-in account could silently disable notifications on a device that was not theirs. Ownership is now verified.
- 💬 Telegram no longer answers the same message twice — when Telegram redelivered a message it had not seen acknowledged, LIA processed it a second time: the assistant answered twice, and a linking code was consumed twice. Each message is now handled exactly once.
- 🖼️ Images in a conversation no longer give the conversation away — displaying an image hosted elsewhere (a photo in an email, a fetched web page) told that image's server the address of your LIA page, conversation id included. It no longer does.
- 🛡️ On the admin side: two protections that were not protecting — an alert now genuinely fires when the API is running with no request ceiling (it existed, but in a file the monitoring never read), and a deployment stops if the server's secrets are not properly locked down, instead of reporting success without having checked anything.
v1.25.21 — 📚 Document search stays on during upgradesJuly 2026
- 📚 Your documents stay searchable during an upgrade — when an administrator switches the model that indexes your document spaces, search used to go dark until the entire rebuild finished. The old index now keeps answering every query until the new one is fully ready, then switches over instantly, space by space — and resumes on its own if the rebuild is interrupted.
- 🛠️ One kind of upgrade still needs a short window — changing the size of the index (a rarer, deeper change) stays a planned maintenance moment, now clearly documented for administrators instead of silently returning nothing.
v1.25.20 — 🔒 What runs, runs in a sealed boxJuly 2026
- 📦 Skill scripts run in a sealed box — a skill's code now runs in a container created for that single run and destroyed right after: no network, no access to your files, no keys. You install a skill for what it produces, not for the trust you would have to extend to its author.
- 🖥️ Every server task asks for your approval — an administration request on a remote server is now prepared and submitted, never launched directly. That includes continuing an investigation already under way.
- 👁️ The confirmation shows everything that will be sent — the card displays the target server, the full task text and the instructions LIA wrote for itself. A summary you cannot read in full is not a confirmation.
- 📍 Your coordinates and tokens leave the logs — GPS positions join the masked parameters, and visited addresses finally pass through the privacy filter, including the ones that used to bypass it.
- 🧹 Switching accounts cleans the device — sensitive data kept locally is wiped on sign-out, and also when a different account signs in after a session has expired.
- 🗺️ Your maps belong to you alone — rendering maps and routes now requires your session and carries its own quota.
- 💬 The Telegram channel requires a secret — with no secret configured the webhook refuses every message instead of accepting them, and a leftover example value stops the application from starting.
- 🚦 The service stays available under a burst — a global ceiling now protects every route, sized far above real usage so it stays invisible in normal use.
- ⬆️ Oversized uploads are refused immediately — a file beyond the limit is rejected before being loaded, instead of occupying server memory only to end in an error.
v1.25.19 — 🛡️ LIA asks before it actsJuly 2026
- 🛡️ A bulk action always asks first — “reply to each of these emails” could send up to nine replies without ever asking you. Replies, forwards, label changes and reminder cancellations are now all confirmed from the very first item.
- ✋ LIA no longer acts on its own — after an exchange, LIA may complete its answer by checking adjacent domains. Seven tools that MODIFY things (labels, completed tasks, lights, scripts) were wrongly eligible for that: they are excluded now.
- 🗑️ The bin stays in the bin — searching for “deleted invoices” silently pulled trashed messages back into the results, presented as ordinary mail. Only an explicit request for the bin includes it now.
- 📅 “Reschedule my meeting” moves it instead of duplicating it — the word “reschedule” was read as a creation, so LIA created a second event rather than moving yours.
- 📨 “Reply to each of these emails” runs again — on some results the list to iterate over was seen as empty and the action ran on nothing, with no error message.
- 💶 Fairer conversation costs — cached tokens were counted twice on one path, inflating the displayed cost. Billing now reflects real usage.
- 🔒 Hardened display cards — a link or a colour supplied by an external service could inject code into a card. Everything is filtered at the source now.
- 🔗 More reliable Apple and Microsoft connectors — a network error could make a failed update look successful, or lose contact data without warning.
- 🔑 Password policy locked down — the rule is applied identically by the browser and by the server, and that consistency is now checked automatically.
v1.25.18 — 🎯 LIA says only what it checkedJuly 2026
- 🎯 No more approximate details — when an exchange calls no tool, LIA now anchors on the data actually retrieved moments earlier instead of rephrasing from memory. It announced an appointment at “4pm” when the calendar had returned 11:15am: fixed.
- 🙋 A plain “I don’t know” instead of a guess — if a piece of information was never retrieved, LIA says so plainly rather than filling the gap with a plausible value.
- 🕐 Weather in your local time — three-hour forecast slots were shown in UTC; they now follow your own timezone.
- 🌍 Weather in your language — weather summaries were written in French whatever your language. They are now translated across the 6 languages.
- 📅 Out-of-range forecast: a clear answer — asking for weather too far ahead used to return a silent approximation. LIA now tells you how far the forecast actually reaches.
- 📍 Two appointments, two forecasts — when several items were enriched on the same day at the same place, one of them silently lost its data. Each now keeps its own.
- 🗺️ No more weather for a town that does not exist — a parameter left empty became the word “null”, and LIA looked for a place by that name. It now falls back to your own location.
- ❤️ Your health signals stop vanishing — nearly one proactive notification in two was prepared without your activity data. The read is 50× lighter, and any drop is now traced.
v1.25.17 — 🔐 Your account grows a spineJuly 2026
- 🔑 Passkeys — sign in with Face ID, your fingerprint or your device code. Phishing-resistant, nothing to remember (Settings > Security).
- 📱 Authenticator app (TOTP) — add a 6-digit second step to password sign-in, with 10 single-use backup codes shown exactly once.
- 🛡️ Identity confirmation — sensitive actions (managing passkeys, exporting your data, signing out other devices) ask you to re-confirm who you are. Right after signing in, nothing to confirm for 5 minutes.
- 💻 My devices — see every session connected to your account (browser, system, approximate IP) and sign out one device or all the others.
- 🔔 New sign-in alerts — a push notification warns your devices when your account signs in from an unrecognized device. Can be turned off in My devices.
- 📦 Export my data — request a complete archive of your data (readable Markdown + JSON + your files); a notification tells you when it is ready to download.
- 📴 Offline mode — LIA now shows a proper offline page and loads faster thanks to smart caching. Your data is never cached on disk.
- 🎨 Reorganized Security settings — Strong authentication, My devices and Export my data are now three clean collapsible sections.
v1.25.16 — ✨ The conversation grows upJuly 2026
- 💬 Smarter input box — your draft survives a refresh or navigation, and ↑/↓ walk through your last 10 sent messages to re-edit them (like a terminal).
- 📌 Protected reading — while you scroll up to re-read, a streaming answer no longer drags you to the bottom. A floating button with a counter brings you back when YOU decide.
- 🔁 Follow-up suggestions — up to 3 tappable chips under the latest answer propose your likely next request; tapping fills the input, never sends.
- ⚡ Slash commands — type « / » in the chat: resume, briefing, agenda, search, plus everyday shortcuts (unread emails, weather, weekend weather, tasks, reminders, news) and your dialogue skills.
- 🧩 Your dashboard, your rules — hide or reorder the 9 briefing cards in Settings (a hidden card is not even fetched anymore); a starter checklist reveals what your instance can do; open commitments get their own consult/close section.
- 📲 LIA installs like an app — add it to your home screen (Android, iOS, desktop), share text or links straight into a chat draft, and use the new action row under every answer (copy, feedback, execution details).
- 🛍️ Skills gallery — browse skills as cards with previews and declared output channels, and install a skill directly from an https URL (server-validated, same strict pipeline as file import). Admins: the Telephony connector now appears in the connectors administration.
- 🌍 Your language everywhere — the assistant's journal and your compiled portrait are now always written in your configured language.
v1.25.15 — 📱 A showcase that fits your phoneJuly 2026
- 📱 The landing page finally looks right on phones — no more clipped text or buttons, and the layout no longer shifts during the demo animation. An automated test now replays the whole animation cycle on every release so it never comes back.
- ❓ Public FAQ redesigned — instant search, section navigation, and 51 questions instead of 13 (connectors, phone calls, voice mode, image generation, privacy). The “What can I ask LIA?” answer is reorganized into 12 expandable domains — same content, finally readable.
- 🎬 The demo animation has its own page — /demo, free of menus and distractions, ready to share on social networks in all 6 languages.
- ♿ Accessibility verified — the public pages pass the accessibility audit (contrast included), in light and dark themes.
v1.25.14 — 🧠 Foundations for more natural moodsJuly 2026
- 🧠 The mood engine, measured for real — a deep audit showed why the assistant's soft moods (tender, melancholic, content…) almost never appeared: its resting point was structurally locked on the assertive side. Two new settings prepare the fix (recentering + removing an artificial « joy » boost) — nothing changes by default: your admin activates them only after measuring.
- 📏 A measurement instrument for admins — a new read-only script replays the mood-quality battery against real usage (moods visited, variety, dominant emotions): activation decisions now rest on numbers, comparable before/after.
- 🛡️ Anti-regression guards — the desired mood properties are now enforced by automated tests (frozen resting points, a reachability oracle): a future change can no longer silently re-lock the palette.
v1.25.13 — 🔗 LIA connects its domains: commitments, automations, your documentsJuly 2026
- 🔁 Open loops — your commitments, tracked — say « I must call the plumber back » and LIA quietly keeps it as an open loop: a timely reminder at the right moment (never nagging — each loop has a cooldown), and it closes itself the day you say « done ». Also visible on the dashboard. Off by default — your admin enables it.
- ⏰ Create automations from the chat — « give me an AI press review every morning » now works as a sentence: LIA shows a confirmation card (days, time, instruction), and only creates the recurring automation once you confirm. List, pause or resume them in natural language too. And when you keep asking the same thing at the same hour, LIA can offer to automate it for you.
- 📄 Your documents become active — LIA can now search your uploaded document spaces on its own and combine them with other sources: « compare the PDF quote with Paul's email » is a real request.
- 👤 Person 360° view — « prepare my call with Marie »: one pass gathers her contact card, your recent emails, shared meetings and what LIA remembers about her — and says honestly which source was unavailable.
- 🗂️ Three new briefing cards (nine in total) — « For you » (your open loops + what your automations did, with the exact time of the next run), « Tasks » (overdue first, in red — a task due today is never counted late), « Documents » (latest modified Drive files: summarize in one click, or open in Drive).
- 🚗 Departure alert — « leave by 13:30 — 30 min drive »: for your next located meeting, LIA checks real traffic and tells you when to leave, weather included. Off by default.
- 🎂 Birthdays & 📞 call follow-up — LIA now warns you the very day (« Want me to draft a message, or call? ») and after a call placed on your behalf, it proposes to add the agreed appointment to your calendar.
- 🧭 Small touches — the chat search now sits on the left of the header (with the offline pill beside it); interest discoveries can be anchored to your city (« exhibition in Lyon this weekend », off by default).
v1.25.12 — ⚡ Six everyday quick wins: search, understand, actJuly 2026
- 🔍 History search finally complete — accents no longer matter (« reunion » finds « réunion »), matches are highlighted in the bubbles with a result counter, and a « Search entire history » button digs through everything ever said — click a result to jump right to that moment, then « Back to present ». The search is now available on mobile too.
- ⚙ The backstage survives a reload — the « N steps · X s » line under each response (what LIA actually did) used to vanish when the page closed. It is now saved with the message: you will find it after a reload and on your other devices.
- 🔌 « Reconnect » shows up whenever needed — when a connected service breaks, the repair banner appeared only at the exact moment of failure; later attempts just said « no connector ». The banner now returns every time the broken service is the missing one.
- 👍👎 Rate any response — two discreet thumbs next to Copy. Your verdict is remembered across devices and feeds LIA's understanding of you; a 👎 opens an optional « what went wrong? » line. LIA never regenerates an answer without you asking.
- 🗂️ The briefing becomes a cockpit — the dashboard cards are now clickable: a mail opens the chat prefilled with « Summarize this email and draft a reply », a meeting with « Prepare me for… », a birthday with a message suggestion. Nothing is sent until you press Enter.
- 🪞 « What LIA understands about you » surfaced — the compiled portrait gets a shortcut at the top of Identity & Memory in Settings, and a discreet dashboard line when it was freshly refined.
- 🤫 A quieter chat header — the permanently pulsing « Online » badge is gone (only « offline » and « processing » remain), and the token totals moved into the context gauge's tooltip, freeing a full line.
v1.25.11 — 🗺️ Widgets you can rely on: everywhere, durable, honest when they failJuly 2026
- 🗺️ The interactive map now works on iPhone — external widgets (like the map) were blocked by a security header only Chrome-based browsers could bypass. Every browser now displays them; on iOS the hands-free wake word gives way to push-to-talk.
- 📍 "Show me where I am" uses your real position — the assistant used to improvise (it could answer with a map of France). It now receives your geolocation when you share it, and the permission prompt appears the moment your message needs it.
- 💾 Widgets survive a reload — maps, games and diagrams reopen with the conversation, on any of your devices, instead of turning into grey "unavailable" boxes.
- 👻 No more duplicate or ghost widgets — a widget could appear twice, or an old one resurface under a new answer. Widgets are now placed by the application itself, never rewritten by the AI.
- 🩺 A widget that fails tells you why — instead of an endless blank square: an explanation, a Retry button, a link that opens the content in a new tab, and the technical detail to report.
- ✏️ Drawing requests go straight to the right tool — asking for a diagram could produce a map or an improvised file. These requests now reach the drawing app (Excalidraw) directly.
- 🔒 Security wave — 16 code-scanning alerts and 7 dependency advisories fixed at the source, and the dependency audit now blocks the pipeline instead of warning silently.
v1.25.10 — 🔔 Readable notifications, and a conversation that opens in the right placeJuly 2026
- 🔔 Your notifications no longer show code — a scheduled-action result could land on your lock screen as <div class="lia-response"><h2>… The text is now flattened before it is sent, on the push notification as well as the in-app toast.
- 🔁 The notification says what the conversation says — the text pushed to your phone was built before the final formatting, so it could differ from the message you opened seconds later in the chat. Both now come from the same source.
- 🗣️ The voice reads exactly what is written — two defects made it drift: it spoke the icon identifiers of data cards ("event", "mail") before the sentence, and it could swallow fragments of text containing comparisons such as "x<a and b>c". Both are fixed.
- ⬇️ A long conversation opens at your last message — history used to open midway, forcing you to scroll down to reach the end. Positioning is now immediate, however long the conversation.
- 🌏 Chinese interface: scheduled-action titles are translated — they were displayed in English while the rest of the interface was localised.
v1.25.9 — 🧪 The interface gets tested, and real bugs surfaceJuly 2026
- 🖼️ The image viewer becomes usable at the keyboard — opening a picture full screen now moves focus into the viewer, keeps the Tab key inside it, and hands focus back to the thumbnail you came from when you close. Screen readers announce it as a dialog named after the image. Before, a keyboard user was dropped at the top of the page and had to tab all the way back.
- 🔌 Your connector preferences stop being silently reverted — a preference you had just changed could be overwritten moments later by a background load that had started before your click, and the same setting could be requested twice. Both are fixed: what you choose now wins.
- 🔗 Connect all no longer skips a connector — when the authorization sequence resumed after a redirect while you had just started another one, the two queues overwrote each other and a connector could quietly drop out of the run. They can no longer overlap.
- 🧪 For admins — the interface is now genuinely tested — frontend coverage nearly doubled on functions (29 % → 55 %), with 40 new test suites over uploads, chat streaming, notifications, connectors, voice and push notifications. That work is exactly what surfaced the three fixes above: they were found by testing the real code instead of mocks.
- 📜 Seven versions reappear in this history — v1.15 and v1.20.17 through v1.20.22 had been written and translated, but were never added to the list this page renders: their entries existed and simply never displayed. A test now fails if any version is written without being wired, in either direction.
v1.25.8 — 🎯 Proactive notifications get varied and concreteJuly 2026
- 🔄 Topics finally rotate — proactive notifications used to circle back to the same interest almost every evening. They now draw from a varied sample (one interest per theme, least recently covered first), and a theme served by either notification flow steps aside for a while. One passion no longer monopolizes your feed.
- 🔗 Concrete suggestions, with sources — when a notification builds on one of your interests, LIA now looks up real, fresh facts and names them (a film, an announcement, an event) with clickable source links, instead of a vague "have a look at a recent release".
- 🧠 Anti-repetition that actually reads what was said — LIA re-reads the content of its last 10 notifications (7 days) and never proposes the same topic, activity or product again, even when it comes from a different source (an interest, a memory, a journal entry).
- ⚙️ For admins — five new .env knobs (sample size, anti-repetition window, enrichment toggle and timeout), a new Prometheus metric tracking enrichment outcomes, Brave-sourced notifications no longer mislabelled as "custom" in the audit trail, and the Telegram shutdown no longer logs a spurious error on every restart.
v1.25.7 — 🎛️ The agentic core becomes visible and controllableJuly 2026
- 🎛️ One-tap approval — when LIA is about to act for you (send an email, confirm a deletion…), a card appears with Confirm / Cancel buttons — and Modify for drafts. You can still reply by text or voice, exactly as before.
- 🔎 The backstage of every answer — under each message, a collapsed “⚙ N steps · X s” line shows what LIA actually did (routing, tools, reasoning), expandable in one click. The transparency of costs now extends to the flow of actions.
- 🔌 Reconnect in one click — if a connected service (Gmail, Calendar…) has expired, LIA tells you plainly in the chat with a Reconnect button that takes you straight to the right setting, instead of a vague apology.
- 🚀 Onboarding that takes action — the tutorial now lets you connect your services in one click, and each example command prefills the chat so you can try it right away.
- ✨ Experience fixes — cancelling a clarification no longer loops forever, the approval card disappears once the action is handled, and the tutorial no longer re-opens on every navigation.
v1.25.6 — 🎯 Interest notifications finally variedJuly 2026
- 🎯 Variety-first notifications — your interests are now grouped into subjects (several AI tools = one subject) and selection rotates them: recently covered subjects rest, rarely served ones get priority. One passion no longer monopolizes your notifications (measured: the dominant theme drops from ~50% to ~33% of notifications).
- 🔗 Clickable source links — every interest notification now ends with links to the original articles (Perplexity, Brave, Wikipedia), so you can read further with one tap. Push and Telegram get a readable text version.
- 🧹 Duplicate interests self-heal — case variants and near-identical topics ("Anthropic" vs "anthropic") are now merged automatically every night, keeping your interest list clean and your notification history intact.
- 🇨🇳 Chinese notification titles fixed — proactive notification titles were silently falling back to English for Chinese users; they now display correctly in 中文.
- ⚙️ For admins — the whole selection is tunable via .env (mode, subject cooldown, rarity exponents, source-link cap) with an instant uniform-draw fallback, and 4 new Prometheus metrics track selections, re-clustering and merges.
v1.25.5 — 📞 The phone assistant becomes trustworthyJuly 2026
- 🎙️ Natural calls end to end — the assistant greets the instant the line opens (no more seconds of silence), knows today's date ("tomorrow" resolves to the right weekday, spoken out loud), speaks with a crisp voice over the telephony-native audio format, and hangs up by itself once the goal is met.
- 🛡️ A strict mandate: never an unplanned expense — if the person offers an extra, an option or any commitment beyond your request (even a 3€ cheese topping), the assistant notes the offer and its price, says it cannot confirm by itself and that someone will call back. You decide — never the assistant.
- 📋 Complete, reliable call reports — the call summary always states every cost mentioned with its amount, flags anything left open and asks how you want to proceed; dates are always absolute ("Saturday July 18", never a "tomorrow" that goes stale).
- 📅 Faithful call follow-ups — "create the appointment" after a call now reuses exactly the date, time, place and title agreed on the phone. If an action plan still looks wrong after checks, LIA asks you instead of executing something false — and it will never invent an email address for an invitation.
- ⚙️ A self-healing connector — configuration fits one screen, settings changes apply automatically on the next call (no more disconnect/reconnect), a call wrongly stuck "in progress" frees itself, and call history survives disconnection.
- 🔧 For admins: 6 new telephony settings — the voice agent's language model (pinned to a model without "thinking aloud"), TTS model and voice, audio format, country dial code, guard grace window. The "Telephony Agent" LLM type (56th) is tunable from the admin panel.
v1.25.4 — 📊 Complete supervision: 25 dashboards, zero blind spotsJuly 2026
- 📈 Three new Grafana dashboards — journals and the user portrait, telephony (calls, durations, automatic recoveries) and the morning briefing each get a dedicated dashboard. Admins can finally watch these features live — they were running with no visibility at all.
- 🔭 No more invisible metrics — 100% of the app's 419 metrics are now covered by at least one panel (76 were orphaned). New dedicated sections: step-by-step response latency, the semantic guard, background runs, AI-cache savings, tool-registry health.
- 👁️ Panels you can finally read — when nothing happened, incident counters now show 0 instead of an ambiguous “No data” (you couldn't tell “all good” from “broken panel” before); every panel now carries an explanatory description, and views stay populated even across 7 days of history.
- 🔧 A performance measurement repaired — conversation load time was never actually measured (the probe listened on a path the engine never takes). The metric is now wired to the right entry point, with a test locking the fix in place.
v1.25.3 — 🧠 Sharper confirmations, leaner AI costsJuly 2026
- ✅ Confirmations that understand you better — when LIA asks you to confirm an action, your reply ("yes", "no, cancel", "send it to Marie instead") is now classified with its full set of examples restored — a long-standing silent bug kept them from ever reaching the AI. Edits and cancellations are recognized more reliably.
- 🎭 Confirmation questions in your assistant's voice — approval questions now follow the personality you chose (gentle, professional, playful…) instead of a fixed sarcastic register.
- 💰 Cheaper, faster AI calls — the fixed part of every internal instruction is now structured so providers can cache it: fewer billed tokens and lower latency on repeated requests, whatever model each function uses.
v1.25.2 — 🎬 The public landing page, reinventedJuly 2026
- 🎬 A demo that shows the real product — the home animation becomes a faithful miniature of the app: four real scenes (orchestration with your approval, a weather×calendar initiative carried through, a genuine phone call, a mini-app built in chat), and while LIA "thinks", a glass pane reveals the orchestration backstage with the live cost counter.
- 📖 A five-chapter story, zero information lost — the page now tells what makes LIA different ("LIA gets it done." · "Nothing leaves without you."…), each chapter opened by one of LIA’s own chat bubbles; every detailed feature card remains available in expandable catalogs, protected by an automatic zero-loss guard.
- 🔍 Transparency, front and center — a "LIA has nothing to hide" section (cost to the cent, public 8.3/10 audit, open source) and "A day with LIA": four tabbed sample days (freelancer, family, developer, administrator) to picture yourself in.
v1.25.1 — 🔒 Security hardeningJuly 2026
- 🔒 Your sensitive links stay private — verification and password-reset links, and the tokens they carry, are never written to the logs anymore. A central filter redacts credentials everywhere and fingerprints sign-in tokens, so activity can still be traced without ever exposing a secret.
- 🛡️ A trusted way to report a security issue — LIA now publishes a standard security.txt file and an updated security policy that routes vulnerability reports privately through GitHub Security Advisories.
- 🔐 Stronger transport and secrets — the site now enforces HTTPS strict transport (HSTS), and production secret files are locked down to owner-only access during deployment.
v1.25.0 — 🛡️ Audit V11: revised framework, cleaner than everJuly 2026
- 🛡️ The technical audit was fully revised — 24 normalized areas aligned with industry referentials (ISO/IEC 25010), every score backed by executed evidence: 8.3/10 with zero major and zero minor findings. The full report — method, scope and open worksites included — is public on GitHub.
- ♿ Accessibility proven on the real product — the automated accessibility journeys (dark mode, keyboard, 320 px mobile, 200% zoom) now run against a production build and pass 17/17. A reported dark-mode contrast defect turned out to be a test-harness artifact — the palette is clean.
- 🧪 A test campaign with zero blind spots — every previously skipped backend test was reactivated or realigned to current contracts: the integration suite now passes 579/579 with zero skipped tests, on Windows and Linux alike. More proof, fewer surprises.
v1.24
v1.24.0 — 📞 LIA calls on your behalfJuly 2026
- 📞 LIA can place a call for you — ask LIA to phone a contact ("call Marie and ask if she's free Tuesday"). It holds the conversation on your behalf, then reports back in the chat with a summary and, when it makes sense, a one-click follow-up.
- ✅ You stay in control — LIA always shows you who it will call and why, and waits for your confirmation before dialing. The summary comes back on its own, without blocking your conversation.
- 🔒 Private by design — during the call LIA can only share your free/busy availability, never your meeting titles, guests or locations. Calls are never recorded and the full conversation is never kept — only a short summary that expires.
- 🔌 Bring your own line — telephony is a per-user connector in Preferences → My Connectors: connect your own ElevenLabs number and key. Calls are billed on your own accounts.
v1.23
v1.23.13 — 🧸 LIA's floating companionJuly 2026
- 🧸 A companion that follows you — a LIA mini-avatar now floats on the dashboard pages (except chat, where the full avatar already lives). It shows the current mood, starts "thinking" when LIA works in the background, and carries a badge when notifications are waiting. A click opens chat; you can minimize it to a small dot.
- 🌅 A greeting that follows the hour — on an empty conversation, LIA welcomes you with ☕ in the morning, 👋 during the day, 🌛 in the evening, and 😴 at night (while it tidies its memories).
- 🔔 Notifications you can feel — when a proactive notification arrives, LIA's avatar gives a little attention wobble at the very moment it appears.
- ✨ Visible freshness on the dashboard — after refreshing a card, a discreet "updated ✨" confirms the data was just refreshed.
v1.23.12 — ✨ Wow effects: the psyche takes shapeJuly 2026
- ✨ Rare emotion particles — when a strong emotion takes hold of LIA (joy, tenderness, wonder, gratitude…), a few particles escape her avatar, once. It is rare — by design.
- 🎬 Gentle answer start — the working steps now fade into the first words of the reply instead of being abruptly replaced; and the dashboard weather card animates with the real sky (rain drops, snowflakes, sun halo).
- 🖼️ Product-level landing demo — the animated conversation on the presentation page now shows the real living steps and the writing caret, just like the app.
v1.23.11 — 🧱 Stronger foundationsJuly 2026
- 🧱 Internal architecture overhaul — the authentication domain now focuses purely on identity and sessions, while your profile and account lifecycle live together in one place. Nothing changes in how LIA looks or behaves: same features, same data, stronger foundations for what comes next.
v1.23.10 — ✨ The interface comes aliveJuly 2026
- 🎭 Living mood avatar — on the current reply, LIA's mood emoji is a real animated face; older messages keep their mood snapshot, which wakes up on hover. The colored ring pulses when LIA's mood shifts, and relationship milestones (Exploratory, Affective, Stable) are celebrated with a one-time notification.
- 🎪 Animated personalities — the active personality's emoji lives in the header, and each personality animates on hover in the selection menu.
- ⚙️ Living progress — while LIA works, past steps dim and the current one breathes; a caret blinks while the answer is being written; the waiting animation changes shape with every reply; the send button takes off; an animated 👋 welcomes you on an empty conversation.
- 🗂️ Living tab — if you switch to another tab while LIA works, the tab title alternates with “✦ LIA is writing…” until the reply is done.
- ♿ Your preferences respected — all these animations turn off with your system's “reduce motion” setting (animated files are then not even downloaded) and follow the psyche display setting.
v1.23.9 — 🧾 Confirmation cards made right (maintenance)July 2026
- 🌍 Confirmation previews fixed in all 6 languages — three display defects in the confirmation cards are fixed: deleting a subject-less email showed the French '(sans objet)' whatever your language (now localized in each of the 6 languages); forwarding an email without an added message displayed the literal text 'None'; deleting an empty reminder showed a blank card (now a '?' consistent with the other deletions).
- 🛡️ Card rendering consolidated (admin) — the internal machinery building these previews was reorganized into a dedicated component, locked by a test net that pins the display character-for-character (61 cases, 16 draft types, 6 languages), and the application now refuses to start if a draft type has no registered renderer. No action required: no new environment variable, no migration.
v1.23.8 — 🧰 Unified API errors (maintenance)July 2026
- ✅ Clearer error for inconsistent proactive-notification settings — setting a daily minimum higher than the daily maximum in the heartbeat settings used to fail with a generic 'internal error'. It now returns a clear validation message, displayed in your language.
- 🛡️ Sturdier, fully supervised API errors (admin) — every backend error now goes through LIA's central error mechanism: uniform responses, automatic structured logs and Prometheus metrics on each error path, with the API contract proven unchanged by a dedicated test net. A CI guard prevents regressions. No action required: no new environment variable, no migration.
v1.23.7 — 🚀 Tidier startup core (maintenance)July 2026
- 🚀 Startup sequence reorganized — the internal machinery that initializes LIA at boot (caches, agents, integrations, scheduled jobs, graceful shutdown) has been reorganized into dedicated modules, with the — critical — boot order now documented in black and white. Nothing changes in daily use: the boot was verified strictly identical, event by event, before/after. Future startup evolutions are much safer.
- 🛠️ No action required (admin) — no new environment variable, no migration; the technical guides (background jobs, MCP, caches) now point to the new locations of the initialization code.
v1.23.6 — 🧱 Sturdier voice core (maintenance)July 2026
- 🎙️ Voice engine consolidated — the internal machinery that speaks LIA's answers (progressive speech, voice-over after actions, cost attribution) has been reorganized into a dedicated, independently tested component. Nothing changes in what you see or hear — but every chat and voice event sequence is now locked by an automated safety net, making future voice evolutions much safer.
- 🛡️ Quality guards & self-hosting (admin) — the codebase now refuses oversized source files (they can only shrink) and blocks private infrastructure values before they are ever committed; the daily log-digest e-mail (Logwatch) is now configured through environment variables instead of a hand-edited config file.
v1.23.5 — 🔒 Safer production maintenance (admin)July 2026
- 🔒 Safer maintenance operations (admin) — the production reindexing script no longer asks the administrator to copy a session cookie into a terminal command; the operation now runs entirely inside the server. Your admin credentials no longer end up in the command history.
v1.23.4 — 🛡️ Cost guardrails for image generationJuly 2026
- 🖼️ Image generation protected — a technical rate limit (default: 10 images per 5 minutes per user) now blocks runaway loops before any cost is incurred. Normal use is unaffected; if the limit is reached, LIA tells you how long to wait.
- 🖥️ Remote server assistant capped (admin) — the DevOps tool is now limited to 5 tasks per 10 minutes per administrator: each task is a paid Claude run plus real actions on a server.
- 📊 Full monitoring (admin) — image generation now reports in the standard tool dashboards (usage, duration, blocked calls), and all four thresholds are tunable via environment variables without rebuilding.
v1.23.3 — ⚡ Faster, and your data finally connectsJuly 2026
- ⚡ Faster responses — LIA starts answering about 1 second earlier on every request: its internal preparation steps now run in parallel.
- 🗺️ "How do I get to my brother's?" is now reliable — LIA now systematically fetches the address from your contacts before computing the route. No more routes to a random location because a person's name was mistaken for an address.
- 🛡️ A guardrail before every action — if a plan tries to use a person's name where an address or an email is expected (route, invitation, sending mail), LIA catches it before acting and fetches the real information from your contacts first.
- 🔗 New natural chains — "email the attendees of this meeting", "reply to the sender", "what's the weather at my destination?", "add this restaurant's address to this contact": data from your tools (calendar, email, places, routes) now connects across domains.
- 🔌 Sturdier background generation — if the server restarts mid-response, the conversation now shows a clear message within a minute instead of waiting forever, and internal stream cleanup is guaranteed even after a hard crash.
v1.23.2 — 🚨 LIA now watches over itselfJuly 2026
- In case of an incident (application down, database or Redis unavailable, disk almost full, failed backup, public access cut off…), the administrator now receives an immediate email alert — no more waiting for the daily log digest to find out.
- 13 vital health checks run continuously, including an end-to-end probe of the public URL and a daily-backup watchdog: if last night's database backup failed, you know within minutes.
- Every alert email links to its step-by-step resolution guide (runbook), and a follow-up email confirms when the incident is resolved.
v1.23.1 — 🔍 The audit goes publicJuly 2026
- The technical audit score moves to 8.4/10 across 24 areas — and the full report is now published on GitHub: scorecard, method, remediation loop and open worksites included, with the commands to reproduce the measurements yourself.
- The "360° technical audit" tile on the landing page now opens the full report directly.
- Every audit figure displayed in the app (landing page, FAQ, Story page) has been refreshed, in all 6 languages.
v1.23.0 — 🧩 Skills that install themselvesJuly 2026
- 🧩 Created skills are installed instantly — Ask LIA to create a skill and it now lands directly in My Skills, active and ready to use: no more copying code blocks, rebuilding folders or uploading a zip. LIA announces the skill by its name once it's available.
- 💬 Guided creation remembers your answers — The Skill Generator's step-by-step dialogue (need analysis, archetype choice, generation) now keeps its context across turns: it no longer re-asks questions you already answered.
- 🔁 Iterate on your skills safely — Ask for a change and the skill is updated in place (even at your skill quota). A failed import can never corrupt the existing version: the previous skill is restored automatically.
- 🛡️ Clear name conflicts — Built-in skill names and names used by other users are reserved: an import reusing one is cleanly refused, and when generating, LIA simply picks a close variant.
- 🔒 Hardened skill imports for everyone — Manual uploads benefit too: strict name validation (a malicious file can no longer write outside its own skill folder), zip size and file-count guards, and one validated pipeline for every import path.
v1.22
v1.22.0 — 🔄 Answers that keep going without youJuly 2026
- 🔄 Your answers keep generating in the background — Leaving the page, switching apps or losing the connection no longer kills an in-progress answer: LIA finishes it server-side and it's waiting in your history when you come back — with its full cost accounting.
- 📡 Live resume when you return — Come back while LIA is still writing and the in-progress bubble rebuilds itself and continues streaming live, automatically. Works across tabs too — and only one answer runs at a time per conversation.
- ⏹️ A stop button, finally — Stop an in-progress answer with one click. The partial text is kept and marked "interrupted"; actions already performed (a sent email…) are not undone, and what was consumed is billed honestly.
- 🔊 Voice only when someone is listening — When an answer finishes in the background with nobody connected, LIA no longer synthesizes its audio: lower costs, same experience when you're there.
v1.21
v1.21.26 — 🛡️ A sturdier chat & voice coreJuly 2026
- 🎙️ Voice mode recovers from network drops — If the connection was lost while LIA transcribed your speech, the interface could stay stuck on "processing" forever. It now shows the error and goes right back to listening — and a rare background error that appeared after each wake-word recording is gone too.
- 🌍 Chat error messages in your language — Connection problems, expired sessions or usage limits used to appear with a fixed French prefix around an English message, whatever your language. These messages are now properly translated in all 6 languages.
- 🛡️ The chat engine is now regression-proof — The heart of LIA (message streaming, action approvals, voice mode) is covered by 434 automated tests with locked coverage thresholds: any regression in these critical areas is now blocked before it can reach you.
v1.21.25 — 🩺 Health checks that tell the truthJuly 2026
- 🩺 A real availability signal — LIA now exposes two health probes instead of one ambiguous endpoint: /health reports that the app is running, while the new /ready confirms that everything needed to serve you — database and cache — is actually up. Monitoring and deployments can finally tell a brief hiccup from a real outage.
- 🚦 Honest diagnostics for self-hosters — The one-command health check reports correct statuses again for every service (it had been failing silently since the development stack moved to HTTPS), and the documented startup behaviour now matches what really happens at boot.
- 🧱 Fresh foundations — The Redis client, the frontend test engine and about forty interface libraries were brought up to date, with the full test suite green.
v1.21.24 — 🛡️ Security hardening & sharper answersJuly 2026
- 📅 Your next appointments, without detours — Asking for your next appointments no longer triggers the full daily briefing: LIA now clearly distinguishes a specific question from a whole-day overview, and always reads your real calendar before answering.
- 🔒 Security updates across the board — Five third-party libraries were updated to their patched versions (including the SSH client and the Claude connector library), and our automated security scanning was repaired so every change is now checked continuously.
- ⚙️ Sounder foundations — Nearly 200 silent error-handling spots in the backend were made explicit, with a permanent CI guard preventing new ones, and the web platform moved to Node.js 24 LTS — issues get diagnosed faster and fixed sooner.
v1.21.23 — 🔌 Sturdier, cleaner connector errorsJuly 2026
- 🔁 Microsoft: clearer reconnection guidance — When your Outlook, Calendar or To Do access expires, the message now clearly tells you to reactivate the connector in Settings — exactly like Google.
- 🔒 More privacy in technical logs — GPS coordinates no longer appear in server logs during address lookups, and error responses from external services are trimmed so they never carry unnecessary content.
- 📊 Better connector monitoring — Every Google/Microsoft error (authentication, quota, outage) is now precisely classified and measured in our dashboards, so external-service incidents are detected faster — with zero change to how you use LIA.
v1.21.22 — 🛡️ Clear error messages and hardened reliabilityJuly 2026
- 🛠️ Forms explain again what's wrong — When an input is invalid (for example a too-short password at sign-up), LIA once again shows the precise message explaining the rule to follow, instead of a generic technical error. The defect affected every form with advanced validation rules.
- ✅ 450+ integration tests now watch over every update — The whole critical chain (accounts, connectors, conversations, pricing) is now automatically re-validated against a real database on every code change, and the test-coverage requirement was raised. In practice: less regression risk with each new version.
v1.21.21 — 🕐 The right time, in your timezoneJuly 2026
- 🕐 Voice mode announces the correct local time — When a voice comment mentions the current time, LIA now always resolves it in your preferred timezone. Previously, in rare cases, the spoken time could be the server's clock (UTC) instead of yours.
- 🌍 Date & time correctness, guaranteed for good — Every date LIA handles (displayed dates, birthday reminders, technical timestamps) is now guaranteed timezone-aware: an automated check permanently blocks the entire "server time vs your time" bug family from ever coming back.
- 📘 Installation guide fully rewritten (for self-hosters) — The Getting Started guide was rebuilt from scratch: production-proven default values throughout, step-by-step setup of the external platforms (Google, Microsoft, Apple…), a feature configuration reference and a final deployment checklist.
v1.21.20 — 🔒 Security hardening & reproducible updatesJuly 2026
- 🔒 17 vulnerable components updated — The automated security audit now inspects every building block LIA ships with (~240, including the indirect ones that were invisible to it before). It flagged 17 components with known vulnerabilities (25+ CVEs); all were updated, and the audit now reports zero known vulnerabilities.
- 📦 Every installation ships exactly what was tested — LIA's builds are now reproducible: each version locks the exact list of its components, verified by cryptographic fingerprint at installation. Two installations of the same version are guaranteed identical — no more invisible drift between them.
- 🛡️ For self-hosters — The software inventory (SBOM) published with each release now lists the exact versions actually shipped, transitive dependencies included. Updating a dependency is one command (task deps:lock), and the CI refuses any change that skips the lockfile.
v1.21.19 — ⚡ Faster under load & self-healing storageJuly 2026
- ⚡ Smoother when several conversations run at once — LIA saves the state of your conversation at every step so it can always pick up where it left off. Until now those saves went through a single database lane per worker and queued up behind each other; they now run in parallel, so response times stay steady even when several conversations are active at the same time.
- 🔄 Storage that heals itself — If the database restarts or a connection drops, LIA now detects and replaces it automatically. Before, this could silently freeze conversation saving until the application was restarted.
- 🛠️ For self-hosters — Pool sizes are tunable in .env (LANGGRAPH_CHECKPOINT_POOL_MIN/MAX_SIZE, LANGGRAPH_STORE_POOL_MIN/MAX_SIZE), with the full connection budget documented. Setting the max to 1 reproduces the previous behavior — a rollback that needs no redeploy.
v1.21.18 — 💾 Automatic database backupsJuly 2026
- 💾 Your data is now backed up automatically — Everything LIA keeps for you (conversations, memory, journals, settings) is saved every night, with 7 days, 4 weeks and 6 months of history kept. And the restore procedure is not theoretical: it is tested for real — an incident can no longer cost more than the last day.
- 🛠️ For self-hosters — A ready-to-use backup sidecar ships with the stack: schedule and retention configurable in .env, one command to back up (task backup:now), one to prove the restore works (task backup:verify), and a complete restore runbook.
v1.21.17 — ✨ New showcase & Story pageJuly 2026
- 📖 New “Story” page — Discover how LIA is built: nearly 100% of the code is written by an AI under human direction — written engineering rules, blocking automated checks, systematic review, adversarial audits. Accessible from the site menu (Story) and summarized on the home page with its key numbers.
- ✨ A showcase that shows the real product — The welcome page now opens on a live animated conversation (draft approval, weather card, multi-agent search), a verified-numbers band (agents, tools, tests, versions) and a clearer side-by-side view of the two execution modes. Every figure was re-checked against the code — including 7 AI providers, not 8 — in all six languages.
- 🔓 Public pages accessible again without an account — Logged-out visitors opening Philosophy, Technical, Blog or FAQ were silently sent back to the login page after a few seconds. Fixed — and a guard test now prevents this bug from ever coming back.
v1.21.16 — 🧹 Major internal cleanupJuly 2026
- 🧹 A leaner, safer engine — Around 13,600 lines of inactive code (old mechanisms that were never used) were removed after exhaustive verification, and the admin dashboards were cleared of their empty panels. LIA becomes simpler to maintain and evolve, with no visible change.
- 🌐 More robust web search and weather — The connections to Brave, weather and Perplexity now share a hardened common foundation: smart retries on transient errors, a circuit breaker that stops insisting when an external service is down, fairer request pacing, and a fix for a connection leak that could build up over time.
v1.21.15 — 🛠️ Sturdier under the hoodJuly 2026
- 🛠️ More reliable under the hood, with no visible change — The core that writes LIA's answers has been fully reorganized internally, with nothing changing in what you see, so it is more reliable and easier to evolve. The reorganization is locked in by around a hundred new automated tests that guarantee strictly identical behavior.
v1.21.14 — 🔒 Stronger privacy & smoother performanceJuly 2026
- 🔒 Stronger privacy in operational logs — LIA filters personal content from standard operational logs: uploaded file names, message text, event or task titles and reminders. Counters, event types and pseudonymous identifiers remain. Detailed diagnostic logs can still expose personal content; access and retention must be restricted.
- ⚡ A smoother app during image generation and uploads — Generating an image or importing a document into a knowledge space no longer briefly slows other actions down: these heavier operations now run in the background, so the app stays responsive for everyone.
v1.21.13 — 🔢 Mathematical formulas now renderJuly 2026
- 🔢 Readable math & science formulas — When LIA writes a mathematical or scientific formula, it now appears as a properly typeset equation — like in a textbook — instead of raw symbols and code. This works both for formulas inside a sentence and for standalone equations.
- 💶 Cleaner currency amounts — Prices such as “9$” inside a rich answer no longer show an unwanted backslash before the amount.
v1.21.12 — ✅ Clearer, more reliable action confirmationsJuly 2026
- 📧 Clearer batch confirmations — When LIA prepares several actions at once (for example an email to two people), the confirmation now shows the correct action and, for emails, who each one goes to — so you can check the recipients before confirming. A batch of sends no longer shows a misleading “confirm this deletion?” prompt.
- 🛡️ Reliable confirmations in autonomous mode — Asking LIA (in autonomous mode) to delete an email or event, or to run a connected-tool action, could freeze with no confirmation appearing. LIA now always shows a proper confirmation before any such action, exactly as in the standard mode.
- ✅ A consistent confirmation experience — Confirmations now behave identically across both execution modes (standard and autonomous), and nothing is ever carried out without your explicit approval.
v1.21.11 — 🛠️ Reliability & maintenanceJuly 2026
- 🧠 More reliable capture of your memories, journals and interests — Parsing of the assistant's internal outputs (from which it notes what it remembers about you) has been hardened: it now tolerates small formatting imperfections and silently fails far less often. As a result, what LIA should remember about you is captured more faithfully.
- ⚡ Leaner analysis of your requests — When your message carries no personal reference to resolve (such as “my wife” or “my brother”), a redundant AI analysis step is now skipped — slightly faster, leaner processing, with no change for you.
- 🛡️ Admin · bounded psyche history — LIA's emotional-evolution history (one snapshot per message) is now purged beyond a configurable retention window (90 days by default), keeping server storage growth bounded. Adjustable via configuration.
v1.21.10 — 🎭 LIA's personality finally comes throughJuly 2026
- 🎭 LIA's personality and mood finally show — Its tone, rhythm and word choice genuinely shift with its inner state, in every reply. LIA doesn't announce its mood — it lets you hear it. Ask it directly how it's doing and it can still tell you.
- 🌈 A richer, more changeable inner life — LIA's mood genuinely varies and evolves from one exchange to the next instead of staying stuck in a single register, and between conversations it gently drifts back toward neutral — the way a real mood settles.
- 💬 The same living LIA in its proactive messages — Reminders, voice comments and proactive nudges now carry the same evolving personality, so LIA speaks in one recognisable voice whether it answers you or reaches out first.
v1.21.9 — 🌍 The assistant speaks your language in approval flowsJuly 2026
- 🌍 Confirmations, edits and refusals now in your language — When LIA asks you to confirm, modify or cancel an action, its messages — and its understanding of your reply (“yes”, “no”, “rather send it to Marie”) — are now fully in your language, across all six languages. Non-French users no longer see any French in these exchanges.
- ⚡ Faster, more reliable follow-ups — References like “the restaurant of this meeting”, “directions to that address” or “show me the 2nd one” resolve directly again, without a needless extra planning step, and follow-ups on a file (“open the first one”) work reliably.
- 🛡️ Admin · cross-domain shortcut toggle — A new setting lets you enable or disable the LLM-free shortcut for cross-domain references (“the restaurant of this meeting”) in production without a redeploy.
v1.21.8 — 🗓️ More reliable calendar searchJuly 2026
- 🗓️ Searching your appointments by topic works reliably — Asking for “my medical appointments”, or for an event by a fuzzy or accented name (“hôtel particulier”), could return nothing. The assistant now lists your events over the period and filters the topic when answering, instead of relying on an unreliable calendar text search. Searching by person (attendee) stays direct.
- 🗓️ “My next appointments” no longer misses events — For an open request with no explicit date (“my next medical appointments”, “my next 3 appointments”), the assistant could wrongly narrow the search to too short a window. It now uses the right period, while still honouring your explicit dates (“on August 15”, “next week”).
- 🛡️ Admin · result volumetry enforced everywhere — The global ceiling on the number of items returned by external APIs (calendar, mail, contacts…) is now applied uniformly by every connector (Google, Apple, Microsoft), closing a bypass. Configurable.
v1.21.7 — 🎨 Regression fixes: voice, interactive widgets & diagramsJuly 2026
- 🎤 Voice input (push-to-talk) works again — A security tightening in v1.21.5 accidentally blocked voice capture, so holding the send button showed 'an error occurred during voice input'. Push-to-talk is fully restored.
- 🗺️ Interactive maps & 🎨 diagram widgets display again — The same tightening had stopped the interactive-map skill and third-party widgets (such as Excalidraw diagrams) from rendering. They work again, now running inside a dedicated secure sandbox that keeps the rest of the app strictly protected.
- 🎨 Complex diagrams no longer time out — Asking LIA to draw a detailed diagram could be cut off after two minutes, right before it finished. Diagram generation is now given the time it needs.
- 🤖 No more off-topic answers when something fails — When a requested action failed, LIA could reply with an unrelated result (for example a map for a diagram request). It now tells you honestly that the action did not succeed instead of masking it.
- ✅ The assistant checks its plan again — An internal safeguard that verifies the assistant's plan really matches your request had silently stopped working. It is fixed and made more robust across the whole assistant.
v1.21.6 — 🔒 Privacy, reliability & sandbox hardening (audit wave 4)July 2026
- 🔒 Complete erasure when you delete your account — Deleting your account now also permanently erases your health data (heart rate, steps) and your last known location, and any device that was sending health data to LIA can no longer do so afterward. Full alignment with your right to erasure (GDPR).
- 🤖 More reliable assistant agents — Fixed a problem that could make the specialized agents (calendar, email, contacts…) fail repeatedly on a request. They also always use the correct current date and time, even after the app has been running for a long time.
- 🔔 More complete proactive suggestions — Under heavy load, proactive notifications could silently miss part of your context (agenda, weather, tasks, interests…). Every source is now gathered reliably, so proactive messages reason on the full picture.
- 🛡️ Stronger isolation between simultaneous users — Closed a rare timing situation where, with many people using LIA at once, one person's private context (such as a personal journal excerpt) could momentarily influence another person's request. Your data stays strictly your own.
- 🧩 Admin · hardened skill sandbox — Scripts bundled with skills now run with reduced privileges and strict resource limits (CPU, memory, processes, file size), and can no longer reach the server's Docker socket. A runaway or malicious script is contained; the admin DevOps assistant is unaffected.
v1.21.5 — ⚡ Performance & hardening (audit wave 3)July 2026
- ⚡ A smoother, more responsive app — Heavy background operations (push notifications, image generation) no longer briefly freeze the interface, and email search returns results faster. Under the hood, blocking work was moved off the main thread and several sequential lookups now run in parallel.
- 🗓️ Correct date formats in every language — A regional-format bug caused dates to use an invalid locale for some languages (notably English and Chinese). Dates now display in the proper format for your language.
- 🛡️ Reinforced security — The personalized greeting on your dashboard can no longer run injected content coming from external data (for example a booby-trapped event title), and a strict browser security policy (Content-Security-Policy) was added as an extra layer against script injection.
- 🔒 Admin · internal services isolated from the network — In production, internal services (database, monitoring dashboards, metrics) are no longer reachable from the local network: only the web application is exposed, behind the secure tunnel. The firewall's interaction with Docker is now documented.
- 🌍 Admin · instant broadcasts & configurable translator — Translated broadcast announcements are now stored, so re-reading one costs nothing (previously re-translated on every view). The personality translator also appears in the admin LLM Configuration and honors your model overrides.
v1.21.4 — 🛡️ Reliability & privacy: systemic hardening (audit wave 2)July 2026
- 📊 Accurate usage statistics at the cycle rollover — At each new billing cycle, some of your dashboard counters (voice, images, requests) could keep part of the previous cycle's values depending on which event crossed the boundary (a message, a voice dictation, or simply opening the dashboard). All counters are now reset together: the usage and cost figures shown are reliable from the first day of the cycle.
- 🔒 Stronger privacy — no personal data in technical logs — Operational logs could contain personal information (home address and GPS coordinates, contact names and emails, email recipients and subjects, memory excerpts). This content is now removed (only technical counters and identifiers remain), with a safety net that automatically masks it even if a future oversight occurs. Compliant with data minimization (GDPR).
- 🇨🇳 Chinese labels for Chinese-speaking users — Two areas of the app (contact field types, data summaries) showed French instead of Chinese for accounts set to Simplified Chinese. They now display correctly in Chinese.
- 🌍 Fallback message in your language — When the assistant exceptionally cannot produce a reply, the short fallback message shown was always in French (without accents). It now appears in your language (6 languages).
- 🛡️ Admin · tool and connector robustness — A failing tool module can no longer vanish silently from the catalogue: the anomaly is now detected (a visible failure in continuous integration, a dedicated counter in production) rather than swallowed. In addition, the API-key "last used" timestamp and the error diagnostics of OAuth connectors are now correctly persisted to the database.
v1.21.3 — 🛠️ Reliability: audit fixes (calendars, tasks, labels, security)July 2026
- 🗓️ Connected services made reliable — Several fixes to your connected services: events created or edited in Apple Calendar now honor the requested timezone (a "10:00 in Paris" event is no longer saved at 10:00 UTC); Microsoft Calendar event search works with an apostrophe ("l'anniversaire" no longer breaks the query); and the Google Tasks "list my task lists" command, which failed on every call, works again.
- 🗺️ Correct route maps — An encoding bug distorted the geometry of multi-segment routes: route maps now render correctly.
- 🏷️ Ambiguous Gmail labels: it asks you — When a label you wanted to delete or apply matched several candidates, the assistant showed a generic error instead of asking which one. It now asks the disambiguation question.
- ⏹️ Stop really stops — Closing the conversation during an in-progress autonomous exploration (ReAct mode) now actually interrupts the work — previously the stop was swallowed and the loop kept going.
- 🌍 Connector messages in your language — "Service not activated" / "no service configured" connector messages now appear in your language (6 languages) instead of always in French.
- 🛡️ Admin · security and settings — Searching users by email is now restricted to administrators (it was reachable by any signed-in account — an account-enumeration risk); the planner budget and step count set in configuration are finally honored (one code path always served fixed values); and two internal settings (metrics sync interval, iterative MCP step timeout) become configurable.
v1.21.2 — 🔒 Hardened security: sanitized rendering and real IP addressesJuly 2026
- 🔒 Protection against malicious content (XSS) — The assistant's answers can relay content from elsewhere (an email body, a fetched web page, a tool result). That content is now sanitized before display: no embedded code can execute in your session anymore. All legitimate rendering (cards, action buttons, tables, interactive mini-apps) stays identical.
- 🧮 Math formulas finally reliable — Formulas now display correctly however the assistant writes them (inline, block, LaTeX code block…), dollar amounts (9$, 1.50$) are never swallowed by math rendering anymore, and a bug that froze the whole conversation after a formula or an interactive app (endless "Generating response…") is fixed.
- 🛡️ Admin · Real IP addresses and locked-down access — The production API is no longer reachable directly from the local network: Cloudflare is the single entry point. Per-IP rate limits now really apply per visitor (previously all clients shared the same internal IP — one aggressive client could block everyone), and both the logs and the Grafana geo map finally show real public IP addresses.
- ⚡ Leaner request processing — The internal layers every request passes through were rewritten in a lighter form: less work per request and per streamed response chunk, benefiting smoothness, especially on modest hardware.
- 🧰 Old conversations upgraded automatically — Conversations created with older versions are now silently migrated to the latest internal format when loaded. Nothing to do on your side; resuming old threads after updates becomes more reliable.
v1.21.1 — ✅ Trustworthy draft editing: what you approve is exactly what goes outJuly 2026
- ✅ What you approve is exactly what gets sent — When reviewing a draft (email, event…) before sending, your successive edits could be silently re-generated at confirmation time: the content actually sent could differ from the last version displayed. The mechanism has been rebuilt: each edit is applied exactly once, and the confirmed content is guaranteed identical to the last content you saw. The same guarantee applies to bulk operations: the list of items you confirm is exactly the list that gets processed, and refining it with a filter ("keep only this week's") no longer reloads your data at every step.
- 💬 Each draft version in its own bubble — When you requested a change, the new version overwrote the previous bubble in the conversation, making it look like the assistant wasn't responding. Every re-presented version now appears in a new bubble, and the reloaded history matches what you saw live.
- ✍️ "Sign with my name" actually works — The assistant didn't know your first name when writing emails: it left a placeholder ("[Your name]") or invented one. It now uses the first name from your profile, both at initial drafting and during edits. Nothing else changes: no signature is added unless you ask for one.
- ❓ Clarification questions are finally displayed — When your modification instruction was ambiguous, the assistant generated a clarification question… that was never shown. It now appears alongside the re-presented draft, so you know what to specify.
- 🕐 Times and dates in your timezone — The "today" opening hours on place cards, the age shown for contact birthdays and the default period of email searches are now computed in your timezone rather than the server's — including right after midnight.
- 🛡️ Reliability & monitoring (admin) — Closing the conversation during a bulk operation now actually cancels the in-flight work; the "stop on first error" mode really stops the batch; Grafana's HTTP metrics use bounded labels (route templates instead of raw identifiers — adjust your queries if they filtered on exact paths); and production startup no longer logs harmless errors (Telegram webhook, system embeddings).
v1.21.0 — ⚡ Faster answers and sharper proactive suggestionsJuly 2026
- ⚡ Noticeably faster answers — Major latency campaign: rich (HTML-mode) answers start 3–4 seconds earlier, the answer's context (memory, documents, journal) is now prepared in parallel with the proactive suggestions instead of waiting its turn, long answers render more smoothly, and connections to the AI models are reused between calls. Nothing to configure: it is all automatic.
- 💡 Sharper proactive suggestions — LIA now knows explicitly which data from your request can feed which tools (a contact's address can feed a route, a places search or the weather). Its follow-up initiatives after an answer become more precise and better targeted, whatever AI model is configured.
- 🛠️ No more data from the previous turn reappearing — In autonomous (ReAct) mode, asking for example for directions right after your upcoming events could re-display those events inside the new answer. Every conversation turn now starts from a clean slate.
- 🎯 Routes to your contacts' exact address — In autonomous mode, LIA now looks up the contact's precise address in your address book before computing a route, instead of settling for an approximation (such as the city) recalled from memory.
- 🏷️ Skill badge visible for everyone — The badge showing which skill generated an answer is now displayed for all users; it was previously limited to administrator accounts with the debug panel.
v1.20
v1.20.22 — 🛡️ Scheduled actions no longer add to your memory or interestsJune 2026
- 🛡️ Scheduled actions no longer add unwanted memories or interests — When LIA runs one of your scheduled actions (for example a daily briefing), the instruction it runs is now treated as an automated task, not as something you typed. As a result it no longer feeds LIA's long-term memory or your interest profile (nor the assistant's private journal or its mood) — only the messages you actually send do. Your memory and interests therefore stay faithful to what you really told LIA. Nothing to change on your side.
v1.20.21 — 🔒 Security hardening and accurate multi-worker metricsJune 2026
- 🔒 Security hardening — dependencies updated — 43 known vulnerabilities in third-party dependencies (frontend and backend, including the framework and AI-library stack) were resolved by updating them to patched versions. There is no user-visible change and nothing to do on your side; it keeps the platform's security posture fully up to date.
- 📊 Admin · Accurate metrics across all workers — In production LIA runs on several worker processes, but the monitoring (Prometheus / Grafana) only saw one of them, so the real-time rate and latency dashboards under-counted by roughly 4× and a few harmless warnings appeared on each deployment. Metrics are now correctly aggregated across all workers, so the admin Grafana dashboards are accurate. Nothing to change on your side.
v1.20.20 — 🔌 Your custom MCP servers now work in the autonomous (ReAct) mode tooJune 2026
- 🔌 Your custom MCP servers now work in the autonomous (⚡ ReAct) mode — Until now, the MCP servers you connect (GitHub, HuggingFace, your own services…) were only used by LIA in the standard mode; in the autonomous (⚡ ReAct) mode they were ignored and LIA fell back to a generic web search. They now work in both modes: LIA picks the most relevant tool from your MCP servers on its own, by description, without you having to ask for a specific one. For a server set to “Iterative mode”, the autonomous mode now uses its individual tools directly (it already reasons step by step), while servers with interactive widgets (e.g. Excalidraw) keep their dedicated agent. Admin: a new
REACT_MCP_EXPAND_ITERATIVE_ENABLEDsetting (on by default) lets you revert to the previous behaviour instantly. Nothing to change on your side. - 🔧 Fix — MCP tools with optional settings no longer fail on strict servers — Some MCP tools (e.g. GitHub repository search, which has optional “sort”/“order” settings) could fail when LIA left those settings empty, because certain servers reject an empty value with a type error. LIA now omits unset optional settings entirely, as the protocol expects, so these tools work reliably. Nothing to change on your side.
v1.20.19 — 📓 A sharper, more careful personal journal for the assistantJune 2026
- 📓 The assistant's personal journal is now far more careful — LIA keeps a private journal of behavioural notes about how best to help you. It was sometimes writing useless or even counter-productive notes from a single message — for instance deciding to always answer you briefly just because one of your messages happened to be short, or recording that it lacks access to your data and then refusing to even try. It now writes far more sparingly, only when a note is clearly grounded in something you actually said or did, and it never records false limitations about its own abilities. The result is steadier, better-judged behaviour over time. Nothing to change on your side.
- 🤖 Journal directives now also guide the autonomous mode — The behavioural directives LIA distills in its journal now steer its autonomous (ReAct) reasoning mode too, not only the standard mode — so it stays consistent however it tackles your request. Admin: a new
JOURNAL_REACT_CONTEXT_MAX_ENTRIESsetting controls how many directives are injected into that mode (default 3).
v1.20.18 — 🎨 Rich display stays consistent for the whole conversation, plus exact usage countingJune 2026
- 🎨 Rich (HTML) display stays consistent for the whole conversation — In the rich (HTML) display mode, the first replies displayed correctly, but as the conversation grew some drifted back to plain formatting without you changing any setting. This is fixed: every reply now keeps the rich HTML formatting throughout the conversation, in both execution modes. Nothing to change on your side.
- 📊 Exact token & cost counting for proactive messages (Admin · Debug Panel) — A proactive (“initiative”) message could appear twice in the Debug Panel — once correctly labelled and once as “unknown” — with the same tokens counted twice, which also inflated the per-message total, the usage logs and user statistics. Counting is now exact: each model call is recorded once. Nothing to change on your side.
v1.20.17 — 💭 Watch the assistant think, and richer reasoning settings for Claude modelsJune 2026
- 💭 See the assistant reason in real time — When a model with “reasoning” enabled is configured for an agent, its train of thought now streams live in the progress area while it works, then disappears as soon as the final answer arrives. It fills the wait with useful context and shows how the assistant approaches your request. It only appears when reasoning is actually active — nothing is shown for agents that don't reason.
- 🧠 Per-model thinking settings for Claude (Admin · Configuration LLM) — Each Claude model now exposes exactly the reasoning controls its API really supports: Opus 4.6 / Sonnet 4.6 use adaptive thinking with an effort level (off · low · medium · high · max); Opus 4.5 / Haiku 4.5 use a manual thinking budget (on/off + tokens). When thinking is on, temperature is locked automatically (the provider requires it), so you can no longer save an inconsistent configuration. Opus 4.5 also gains a separate global “effort” control.
- 🔧 Fix — Claude no longer fails when thinking is enabled — With extended thinking turned on for a Claude model, certain internal steps (such as request analysis) could fail and degrade the whole turn. This is fixed: those steps now use the only mode the provider allows with thinking, so turns complete normally. Nothing to change on your side.
v1.20.16 — 🔊 Voice mode no longer reads formatting aloud during conversationMay 2026
- 🔊 The voice no longer speaks formatting markup during conversation — When voice mode is on and you simply chat (without triggering a tool) with the rich (HTML) display mode, the voice sometimes read formatting elements aloud instead of the answer itself. Rich formatting is now reserved for replies that present data from tools; your conversational exchanges stay plain text, read naturally — and look identical on screen in every display mode. Nothing to change on your side.
v1.20.15 — 📊 Monitoring reliability: Grafana dashboards repaired and sub-agent metrics activatedMay 2026
- 📊 Grafana dashboards now read correctly — Many monitoring panels that stayed stuck on No data (because of outdated metric names, labels or queries) have been repaired, and the sub-agents monitoring dashboard now fills in: the execution metrics of sub-agents (how many run, their duration, their token usage and their errors) are finally emitted. This only concerns the self-hosting / admin monitoring stack — nothing changes in the assistant you talk to.
- 🧹 Metrics cleanup — 9 obsolete or never-emitted metrics were removed (lowering cardinality and dropping a label that contained a user identifier), and a small counter leak (the number of active sub-agents) was fixed.
v1.20.14 — 🔊 Voice mode reads a short commentary again (not the on-screen text), and 📜 the full history of long conversations becomes reachableMay 2026
- 🔊 Voice mode no longer reads the displayed text out loud when tools are used — When voice mode is on and your request needs a tool (checking emails, calendar, contacts, the weather, etc.), the voice is supposed to deliver a short spoken commentary, distinct from the on-screen reply (which can be long, structured, with cards or widgets). Recently the voice was reading the displayed text verbatim instead, which was both noisier and less natural. This came from a stale signal inherited from the previous turn of the conversation, which prematurely triggered the chat-mode direct read-aloud. Fixed — the spoken commentary is back on turns that use tools.
- 📜 Long conversations: scroll back to the beginning of the discussion — In a conversation that has grown to thousands of messages, the start was unreachable from the chat — only the latest ~50 messages were loaded with no way to scroll further back (the messages were always preserved in the database, just not displayed). You can now scroll up in the chat: older messages load progressively in batches, deduplicated, and your reading position is preserved (no visible jump). A small "Loading older messages…" indicator appears at the top during each fetch. Works in any conversation; the in-chat search still filters what's currently loaded.
v1.20.13 — 🔧 Responses no longer fail with the latest Gemini models (Gemini 3 series)May 2026
- 🔧 Fix — the latest Gemini 3.x models no longer trigger a generation error — With the most recent Gemini AI models (the Gemini 3 series, e.g. Gemini 3.5 Flash), some responses could fail with the message “A problem occurred while generating the response”, or degrade silently (occasionally unintelligible voice output, incomplete formatting). These models return their text in a different format that LIA didn't interpret everywhere. This is now fixed end to end, in both the default mode and ⚡ ReAct mode — including when you mix several AI models. Nothing to change on your side.
v1.20.12 — 🌙 Dormant interests stay visible and manageable, and your long-term memory warns you before forgettingMay 2026
- 🌙 Your dormant interests are no longer lost — When you stop mentioning a topic for a while, LIA puts the matching interest to sleep ("dormant") rather than acting on it. Until now, dormant interests simply vanished from your Settings even though they were still stored — you couldn't review, edit or remove them, and they were deleted automatically after 90 days. They now appear in a dedicated "Dormant" section, clearly marked, where you can reactivate, edit or delete them at any time. A reactivated interest starts fresh, as if newly detected. The counter now also shows the breakdown: active · dormant · blocked.
- ⚠️ Your memory now warns you before forgetting — LIA automatically forgets long-term memories that have become unimportant and unused (unless you've pinned them). Until now, nothing told you a memory was about to be forgotten. Memories close to automatic removal are now flagged in your Settings list with a "may be forgotten" warning, so you can pin the ones that matter before they disappear.
v1.20.11 — ⚡ ReAct mode becomes proactive, a home greeting that matches your real day, and several display and accuracy fixesMay 2026
- ⚡ ReAct mode is now proactive, like the default mode — In ReAct mode (the autonomous, step-by-step reasoning mode), after answering your request LIA can now run a few extra read-only checks in related areas and weave the result into its reply — for example the weather and the travel time for your next appointment, or the contact behind an email. This proactive "Initiative" already existed in the default (Pipeline) mode; both modes now offer it. It only triggers when genuinely useful, and only ever reads — it never creates or changes anything on its own.
- A home greeting that matches your real day — The one-line greeting at the top of your dashboard now reads your agenda correctly: it tells today from tomorrow, names the next appointment, and reacts to the shape of your day (a free day, a busy one, or a single thing coming up) — instead of vaguely (and sometimes wrongly) implying something was happening "today".
- Money amounts now display correctly — Amounts written with a dollar sign (e.g. "1.50$" … "9$") in the chat or in proactive notifications were sometimes mangled into a mathematical formula, with spaces removed and characters stacked. They now display as plain text everywhere.
- Admin — configurable home dashboard limits and accurate cost conversion — How much content each Today Briefing card shows (calendar, emails, birthdays, reminders, etc.) is now configurable via environment variables — the agenda now shows up to 10 upcoming events instead of 3 by default. Separately, the exchange-rate source used to convert LLM costs into euros was fixed (it had been falling back to a stored rate), so cost reporting is accurate again.
v1.20.10 — In ReAct mode, the reply no longer exposes the assistant's internal reasoning and no longer appears twiceMay 2026
- Cleaner replies in ⚡ ReAct mode — In ReAct mode (the autonomous, step-by-step reasoning mode), two glitches could spoil the final reply. (1) The assistant sometimes let its internal thinking steps show through — labels like "PLAN", "OBSERVATION", "CROSS-CHECK" followed by the "final answer" — instead of showing only the answer itself. (2) The reply text briefly appeared duplicated before collapsing to a single copy. Both are fixed: LIA now delivers only the final, clean, single reply, in ReAct mode just like in the default mode. Nothing changes on your side.
v1.20.9 — In ReAct mode, creating, modifying or deleting now asks for your confirmation before acting — no more false "Done!" with nothing actually happeningMay 2026
- ReAct mode now asks for confirmation before sensitive actions, just like the default mode — When you use the ⚡ ReAct mode (the autonomous, step-by-step reasoning mode) and asked LIA to create an appointment, send or reply to an email, add a contact, create a task or delete something, it would sometimes answer "Done! I've created the appointment…" while in reality nothing had been created. The action was only prepared internally and then silently dropped — the confirmation step that already exists in the default (Pipeline) mode was missing in ReAct mode. From this release both modes behave identically: a sensitive action (create / modify / reply / forward / delete an event, email, contact, task, file or label) now shows you the same confirmation card with the details, and the action is only really carried out after you confirm — you can also edit or cancel it. Reminders, which are designed to be created instantly, keep working without a confirmation step in both modes.
v1.20.8 — Conversations no longer freeze on long history, context-usage indicator in the chat, visible recovery notice when the summary failsMay 2026
- Conversations no longer get stuck when the history grows long — When your conversation accumulates many messages, LIA runs a background compaction step that summarises the older turns to free up context room. On May 16, 2026 a production user could no longer send any message: every turn re-triggered the summary, the language model answered once and then went silent for ~93 s until the connection was finally dropped at 125 s with "Erreur de connexion: network error". Root cause: no maximum delay on the LLM call, no retry on transient errors, no signal sent to the frontend during the wait. This release rebuilds end-to-end resilience: per-chunk timeout (35 s), global timeout (2 min), three attempts with progressive backoff, and — if all of that still fails — an explicit fallback that cleanly truncates the older history rather than leaving you with a silent useless summary. You'll literally see the note "The older conversation was truncated because the summary could not be generated" when that fallback fires, rather than a mute loss of information.
- Visible indicator in the chat while compaction runs — Before this release, when LIA was summarising your history you had no visual feedback: the send button froze for ~30 s without explanation. Now, the moment compaction starts, a notification (toast) appears at the top of the screen reading "Summarizing the conversation to free up context…" with a spinner; the message input is locked while the summary runs; when it's done the toast morphs into a confirmation "Conversation summarized — N tokens freed" (or a warning if the truncation fallback fired). The toast stays visible no matter where you've scrolled in the conversation.
- Context-usage pill in the chat header — A new discreet pill next to the search box continuously shows
tokens used / compaction thresholdas a coloured ring and a percentage. Green below 50 %, amber from 50 to 75 %, orange from 75 to 90 %, rose above 90 %. The displayed percentage is capped at 100 % to stay legible; in case of overshoot (typical right after a long answer), the hover tooltip explains "compaction will run on your next message". The pill persists across page refreshes — no need to send a message just to see it. - Admin: new operator settings, Grafana dashboard, transactional recovery script — Five new
.envsettings drive compaction without a rebuild:COMPACTION_PER_CHUNK_TIMEOUT_SECONDS(default 35),COMPACTION_GLOBAL_TIMEOUT_SECONDS(default 120),COMPACTION_MAX_RETRIES(default 3),COMPACTION_RETRY_BACKOFF_BASE_SECONDS(default 1),COMPACTION_INCLUDE_PREVIOUS_SUMMARIES(defaulttrue, consolidates prior summaries instead of stacking them turn after turn). New Grafana dashboard14-compaction.jsonwith 7 panels (strategy mix, latencies p50/p95/p99, per-chunk timeouts, global timeouts, errors by type, skip reasons, tokens saved). Five new Prometheus metrics includingcompaction_writer_unavailable_totalwhich makes any silent SSE-signal degradation visible. A transactional SQL scriptscripts/admin/reset_user_checkpoints.sqlresets a single user's checkpoints for stuck-conversation recovery, without touching messages or the application. Full operator documentation indocs/technical/COMPACTION_v2.md.
v1.20.7 — Bulk reminder deletion: the summary is finally readable, with each reminder's name and trigger timeMay 2026
- The summary after a bulk deletion is finally readable — When you said "delete all my reminders" and confirmed the bulk operation, LIA used to reply with a generic block "✅ 3/3 / ✅ Action completed successfully / ✅ Action completed successfully / ✅ Action completed successfully" — no reminder name, no date, just a counter and three identical lines. Each row now carries the deleted reminder and its trigger time, with a header that agrees grammatically with your language:
🔔 ✅ 3 reminders deleted / ✅ Buy groceries — Friday May 16, 2026 at 2:00 PM / ✅ Call Mom — Saturday May 17, 2026 at 9:00 AM / ✅ Doctor appointment — Sunday May 18, 2026 at 11:00 AM. The same format applies automatically to bulk deletions of emails, events, contacts, tasks, Drive files, and Gmail labels. - The confirmation panel before a bulk deletion finally looks the same regardless of the internal path — Before this release, asking for a bulk deletion could present two different looks depending on which code path was taken internally (FOR_EACH on one side, draft-critique batch on the other) — one showed
Reminder: Doctor on May 17, 2026 at 7:00 PM, the other showed🔔 Reminder: Doctor 🔔 Sunday May 17, 2026 at 7:00 PM(doubled emoji, broken formatting). You now see the same unified rendering on both sides:🔔 Reminder : Doctor - Sunday May 17, 2026 at 7:00 PM— one domain emoji, the category noun (Reminder / Email / Event / Task / Contact / File / Label) before the title, the weekday included in the date, and a dash separator between title and date. The dialog title also moves from "Confirmation required" to "Confirm deletion" for reminders, consistent with other deletion types.
v1.20.6 — More reliable web searches, end-to-end Gemini 3.x LLM support, 28 timeouts now tunable without rebuildMay 2026
- Web-search response enrichment (Brave) is more reliable — When LIA triggered a web search to enrich a response (fact-check, recent context, etc.), a misconfigured internal timeout (3 s) cut the request before the search engine had time to reply (5 s). Result: searches that looked like they ran but returned nothing, and answers thin on context. The wrapper timeout is now 8 s and stays consistent with the underlying HTTP call (internal rule: enclosing timeout ≥ HTTP × 1.5). No setting to change — it's fixed in the background.
- Gemini 3.x LLM: full pipeline (planner, memory, voice, HITL) now compatible — If you (or your administrator) had switched any LLM type to Gemini 3.1 Pro Preview (or any future Gemini 3.x) under LLM Configuration, several flows silently fell back to fallback behaviors or even displayed corrupted output. The cause: Gemini 3.x returns its content in a structured form (blocks) where every other provider uses a plain string. 33 places in the code read that string without protection. All of them have been migrated to the official
BaseMessage.textAPI (LangChain Core 1.2+) which handles both shapes. Concretely: the planner, memory reference resolution, HITL question generation (tool approval + plan approval), voice comments, broadcast notifications, and even the LLM cache keys now behave identically with Gemini 3.x or any other provider. - Admin: 28 timeouts previously hardcoded are now tunable via
.env— The full timeout plumbing (outbound HTTP to Brave / Perplexity / Wikipedia / OpenWeatherMap / Ollama, short/long connector operations, SSE polling, OAuth locks, scheduled-actions executor, WebSocket idle windows, image generation, browser agent, sub-agents, devops Claude…) is now exposed as operator-tunable settings — no more Docker rebuild to adjust a single timeout. A central referencedocs/technical/TIMEOUT_REGISTRY.mddocuments every parameter (role, default, range, consumer file, cascades) with three annexes: Annex A (LLM-call timeouts out-of-scope, managed via LLM Configuration), Annex B (mirror of frontend timing.ts), Annex C (known gaps deferred to a future wave). All defaults match the historical values — no behavior change by default. Three cleanups along the way: 2 dead constants removed, 1 duplicate (currency API) resolved, 1 orphan field (task_orchestrator_execution_timeout_seconds) actually wired as a soft cap on the total duration of multi-wave plans + new Prometheus counterlia_parallel_execution_global_timeout_total. - Preventive defense layer on the planner (in observe mode) — When you asked something like « my next two medical appointments », the planner could try to filter the word « medical » directly inside the search (which doesn't work for Google Calendar, Gmail, etc., which only index by literal text) — result: 0 hits or the wrong ones. A new validation layer (ADR-084) now detects these semantic-qualifier leaks (« urgent », « important », « medical »…) in text-search parameters of literal-search tools and reports them via a dedicated Prometheus counter. Current mode: observation only — no plan is mutated in this release, the time to validate frequency and absence of false positives. Phase 2 (auto-correction activation) is planned after a few weeks of telemetry. Controlled by
PLANNER_SEMANTIC_LEAK_MODE=observe|autocorrect|off.
v1.20.5 — Sub-agents: expert analyses finally substantive, expert voice preserved, 3 new admin settingsMay 2026
- Analyses delegated to a sub-agent are finally substantive — When you asked "answer me like a senior analyst, use a specialised sub-agent", the sub-agent used to go off exploring in every direction (up to 30+ successive searches), exhausted its budget before any synthesis, and returned either a truncated answer or a hollow blurb that the main assistant then drowned in its own conversational tone — the delegation added no real value. The sub-agent now follows a clear contract ("if an assistant could write the same thing without you, it's a failure"), calibrates the shape of its output to the signals in the request (deep analysis → sourced multi-section report; summary → faithful condensed version; comparison → parallel dimension-by-dimension structure), cites a URL for every figure, distinguishes facts from inferences, flags missing data instead of inventing it — and the main assistant relays its analysis verbatim (sections, sources, structure intact) without compressing or rewriting it in a sarcastic voice. Bonus: no more forced markdown tables in these answers — comparisons now use parallel lists or sections, which read better.
- Admin: 3 new settings to pilot the sub-agents —
SUBAGENT_TOOL_TIMEOUT_SECONDS(default 180 s, range 30-600) andSUBAGENT_TOOL_MAX_TIMEOUT_SECONDS(default 300 s, range 60-900) decouple the delegation time budget from the genericMAX_TOOL_TIMEOUT_SECONDSceiling (120 s, unchanged for other tools) — ReAct loops running on a slow reasoning model (deepseek high, gpt-5.2 high) can now legitimately last 90-150 s without being killed mid-synthesis.SUBAGENT_RESEARCH_TOOLS_WHITELIST(defaultbrave_search_tool,fetch_web_page_tool) refocuses the sub-agent on factual verification instead of exploring a catalogue of ~80 tools (which used to burn its entire budget on sterile exploration). Validation at load time: a misspelled name in.env(dashes instead of underscores, semicolons instead of commas) raises a clear error instead of a silently empty whitelist. - Admin: sub-agent iteration budget raised to 10 by default —
SUBAGENT_DEFAULT_MAX_ITERATIONSmoves from 5 to 10 (range widened to 1-30, up from 1-15). LangGraph counts every node visit as a superstep, so 5 barely supported 2 tool rounds + synthesis — at 10, the sub-agent has the headroom for 3-4 tool rounds and a complete synthesis without hittingGraphRecursionError. TuneSUBAGENT_DEFAULT_MAX_ITERATIONSin.envif you want to push further (up to 30 for very deep analyses, or drop to 5-6 to save tokens on simple tasks).
v1.20.4 — Reliable web browsing agent, Anthropic Claude 4.x fix, robust LLM config on model changeMay 2026
- The web browsing agent now completes multi-step tasks reliably — Asking LIA to "go to a website and do X" used to work for a quick lookup but fall apart on anything longer: the browsing loop was cut off after ~30 seconds, saw only a truncated view of the page, sometimes invented URLs with guessed parameters, and rarely reached the right form. Now the loop gets a real time budget (up to several minutes), sees far more of the page (accessibility-tree budget raised from 5 000 to 15 000 tokens), navigates through the site's own controls instead of fabricating URLs, handles autocomplete fields (it clicks the matching suggestion), re-checks the page after each action, detects when it's stuck and changes strategy, and — on transactional flows — advances up to (but not past) the login / payment / final-validation step, then stops and reports what it found and what you still need to do. A new admin setting
BROWSER_REACT_MAX_ITERATIONS(default 15) caps how many reason↔act iterations a single browsing task may run. - Admin: the Anthropic Claude 4.x models in the LLM Configuration dropdown now work — "Claude Opus 4.6 / 4.5", "Claude Sonnet 4.6" and "Claude Haiku 4.5" were registered in the model catalogue with a dotted id (
claude-opus-4.6) that the Anthropic API rejects with404 not_found_error— picking one of them for any LLM type made that feature fail immediately. The catalogue, themcp_app_react_agentdefault and the affected database rows are corrected to the dashed form the API expects (claude-opus-4-6, …); a one-shot migration rewrites existing configuration rows. - Admin: changing the model or provider of an LLM type can no longer break that type — In Configuration LLM, switching an LLM type from one model to another could leave behind a "reasoning effort" value whose shape no longer fit the newly selected model (for example a DeepSeek-style "off" left on a config now pointing at a Qwen model) — the feature then crashed on its next call. The reasoning-effort value is now automatically re-normalized to the new model when you change it (kept if compatible, otherwise cleared), the save endpoint rejects an incompatible combination with a clear error, and — as a final safety net — the runtime silently falls back to the model's default if a stale incompatible value ever slips through (logged for the admin).
v1.20.3 — Itinerary tool resilience: directions still work for events that have already startedMay 2026
- Directions to an in-progress meeting now work as expected — Asking LIA for the route to your next appointment when that appointment had already started (an all-day reservation, a meeting reviewed 30 minutes after kick-off, etc.) used to silently produce a chatty fallback with no map card and no ETA. The Google Routes API rejected the call because LIA passed the event's original start time as the target arrival time, which by then was in the past. Past arrival/departure times are now silently clamped to now + 60 seconds before reaching the API, while genuinely future timestamps remain untouched so traffic prediction stays accurate. Result: the HTML route card and ETA appear reliably, even when you ask after the meeting began.
v1.20.2 — Voice STT/TTS catalogue, per-message cost attribution, sentence streamingMay 2026
- New remote Speech-to-Text option (ElevenLabs Scribe) — Voice mode gains a per-user toggle in Settings → Voice mode: keep the free local Sherpa-onnx Whisper (default, audio never leaves the LIA server) or opt into ElevenLabs Scribe ($0.22/h, billed per audio duration) for higher accuracy on conversational speech. A discreet 🎤 badge on each user bubble shows the duration and EUR cost of the transcription. Cost flows into your usual Cost dashboard tile and usage limits — no separate quota.
- TTS becomes catalogue-driven: 3 providers, single admin form — Voice synthesis is now configured through Configuration LLM → Voice TTS with three providers seeded from day one: Edge (free, default, Microsoft neural voices), OpenAI (
tts-1,tts-1-hd) and ElevenLabs (eleven_multilingual_v2,eleven_turbo_v2_5,eleven_flash_v2_5). Per-provider tuning surfaces (Edge SSML rate/pitch/volume, OpenAI speed + format, ElevenLabs stability/similarity_boost/style/use_speaker_boost) and a live voice picker for ElevenLabs (custom + library voices from your account). - Per-message TTS cost on assistant bubbles — Paid TTS providers now surface a 🔊 badge on the assistant message showing the character count and the EUR cost of the synthesis (Edge free → no badge). Hover the badge for the provider, model and exact rate. Aggregated into the Cost dashboard tile and exportable via the new
tts-usageCSV (admin + user) and the extendedconsumption-summary. - Voice latency: first audio in ~1 second instead of ~5 seconds — A new sentence-level streamer dispatches TTS synthesis as soon as a complete sentence is produced by the LLM, in parallel. The first sentence starts playing while the assistant is still composing the rest of the answer — instead of waiting for the full text. Persistent HTTP keepalive on ElevenLabs saves another ~100–300 ms per sentence. Works in both chat and agent modes.
- Admin: STT remote kill switch + duration cap (cost-spike defence) — Two safety levers on the remote STT pipeline. The
ELEVENLABS_STT_ENABLEDenvironment flag now actually disables remote STT at runtime (instant fallback to local Sherpa for incident response or emergency quota). TheELEVENLABS_STT_MAX_AUDIO_DURATION_SECONDScap (default 5 min) is now enforced before each call — oversized clips are rejected with a clear error, no provider charge incurred. - Voice errors now show a precise reason — Failures across the three TTS providers (Edge / OpenAI / ElevenLabs) are now surfaced with a stable error code: provider timeout, rate limit reached (with retry-after delay when available), authentication issue, network error, empty audio response, generic HTTP error. The frontend can show a precise toast for each — instead of a generic "voice failed" message.
- New CSV exports for STT and TTS usage —
stt-usageandtts-usageare added under Settings → Consumption Export (per-user) and the admin export tab (per-user filter). They list every paid voice call with date, conversation, message ID, provider, duration / characters and EUR cost.consumption-summaryalso gets new STT and TTS columns so totals stay in one place.
v1.20.1 — Admin LLM: per-model sampling matrix and reusable reasoning shape templatesMay 2026
- Configuration LLM: sampling sliders only show what the API accepts — In the per-type LLM Configuration dialog, the four sampling inputs (
temperature,top_p,frequency_penalty,presence_penalty) now appear or hide per individual parameter based on what each model's API actually accepts. Anthropic 4.5+ surfaces only temperature; DeepSeek V4 surfaces all four; GPT-5 series surfaces none. The reasoning widget itself (enum select, integer budget, or toggle+budget) is also driven directly by the model's declared shape — the dialog never lets you pick a value the API would reject. - Pricing LLM: "Copy reasoning shape from…" template selector — When you add a new reasoning model in the Tarification LLM Texte admin form, you no longer have to type the reasoning widget / enum values / budget range by hand. A new dropdown lets you pick a representative model whose reasoning behavior is identical (e.g. "like gpt-5" for OpenAI o-series / GPT-5 reasoning, "like claude-opus-4-5" for Anthropic 4.5+, "like deepseek-v4-flash" for DeepSeek V4) and the four reasoning shape fields are snapshot-copied at creation time. The list deduplicates the catalogue automatically (~15 unique reasoning shapes today) and self-enriches: any model created in Custom (advanced) mode with a novel shape becomes available as a template afterwards.
- Toggle "Is reasoning model" gates the reasoning controls — The Pricing form now hides the template selector and the Custom block whenever the new model is non-reasoning. Switching the toggle off resets the widget to
noneand clears any pending template choice.kind(chat / image / audio / …), the four sampling toggles and the optional tooltip i18n key remain visible regardless and are saved per model — independent of the reasoning shape choice. - Audit log: reasoning template + kind + widget captured on every create / update — The admin audit trail (visible in the security log) now records, for every LLM model creation or edit, which template the admin picked (or null in Custom mode), the model kind and the resolved reasoning widget. Instant forensic visibility on which API contract a given row inherited from.
v1.20.0 — Stratified journal: levels, epistemic status, ambient user-model portraitMay 2026
- Personal journals get four abstraction levels (L0/L1/L2/L3) — Every entry now carries a level:
L0raw observation (weak signal worth tracking),L1operational directive (the WHEN→DO BECAUSE format),L2transversal pattern (synthesis of several convergent L1s),L3portrait facet. L2 and L3 are produced exclusively at consolidation, by topic clustering. The Settings panel now shows a level badge on each entry and a Group-by toggle (Theme / Level). - Epistemic status on every entry — Each journal entry exposes a
confidencedot (low / medium / high) plusevidenceandcontradictioncounters. LIA can now distinguish hypotheses still being tested from observations confirmed across many turns — and quietly demote entries the user keeps contradicting. Confidence and level are editable in the create/edit dialog. - Deferred self-evaluation (T → T+1) — At each turn, LIA looks at the directives it injected on the previous turn and reads your reaction in the current message. If you confirmed, the
evidencecounter ticks up; if you pushed back,contradictionticks up. The journal therefore measures its own usefulness — at zero added LLM cost (same extractor call, just enriched prompt). - New "How LIA sees you" section in the journal Settings — A compiled user-model portrait surfaces above the entries accordion, with two read-only tabs (Full / Brief), a compilation timestamp and a freshness badge. The full version (≈ 200 tokens) feeds the conversation and the planner; the brief version (≈ 60 tokens) is diffused everywhere else LIA speaks: ReAct, voice, reminders, heartbeat notifications, proactive interest pings, fallback responses. Same nuanced model of you, carried across every channel.
- Three corrective levers on the portrait — The portrait itself is a synthesis and is never directly editable. Three levers are exposed instead: (1) edit/delete the L3 source entries (standard CRUD), (2) 🚩 Signal a problem opens a free-text dialog and triggers a synchronous re-consolidation (~5–10 s) with the user signal pinned at the top of the prompt, (3) 🔄 Consolidate now runs the full consolidation cycle on demand, bypassing the cooldown.
- Smart filter "show only entries never used" — A new toggle at the top of the entries list surfaces entries with
uses=0and no last-injection date. These are the candidates for reformulation or deletion at the next consolidation pass — instant visibility on which directives never proved useful. - Better consolidation discipline — The consolidation prompt now performs an explicit pairwise dedup scan as STEP 1, audits theme misclassifications, actively clusters convergent L1s into L2 patterns, and compiles the user-model portrait — all in the same LLM call (no extra cost). Result: noisy duplicates collapse periodically, distinct entries actually emerge across themes and levels.
- GDPR: portrait scrubbed on account deletion — When an account is deleted, the three portrait columns (full, brief, compiled timestamp) are blanked out alongside the rest of the user data. The export endpoint also includes the compiled portrait under a
portraitkey for completeness.
v1.18
v1.18.1 — Today Briefing polish, fresh screenshots, enriched statisticsApril 2026
- LIA greeting now overlays the welcome image — The personalized AI-generated greeting now sits directly on top of the LIA avatar, replacing the random marketing taglines. More personal, with the tokens + EUR cost of the call still discreetly shown underneath.
- Refresh button always visible on mobile — On smartphones, the small sync icon on each dashboard card is now always shown (no need to hover — which doesn't exist on touch screens). On desktop, behavior is unchanged: visible on card hover with the +12° rotation affordance.
- Usage statistics: start date for the lifetime totals — Under each total counter (messages, tokens, Google API requests, cost), a small 'since DD/MM/YYYY' label shows your account creation date — the reference start of the lifetime totals is now explicit.
- Illustrative icons on dashboard sections — 'Mon dashboard' now sports a sunrise icon 🌅 (consistent with the daily-ritual nature of the briefing); 'Statistiques d'utilisation' gets a bar chart icon 📊. Both titles also share the exact same typography so the two main sections look like siblings.
- 12 fresh dashboard screenshots on the landing page — The landing carousel and README images are refreshed with the v1.18.x captures, and 4 new ones are added: chat debug panel, interactive skills, long-term memory, Psyche Engine. Cache is invalidated automatically so no browser stays on the old images.
- 'Skills' and 'Health data' descriptions trimmed on the home page — Both descriptions were 4× longer than the rest; rewritten to the same length as their neighbors in 6 languages, no information lost (the long-form details live in the FAQ and the docs).
v1.18.0 — Today Briefing: the home becomes a daily ritualApril 2026
- New 'Today' home page — The dashboard replaces static stats with a daily briefing: a personalized AI-generated greeting, a 2-3 sentence synthesis as a teleprompter, then a grid of 6 operational cards: enriched weather (with 5-day forecast, wind, rain probability), today's agenda, unread mails, upcoming birthdays, active reminders, health metrics.
- Non-blocking progressive rendering — The 6 cards appear as soon as the sources respond (≈ 1 s) without waiting for the AI. The synthesis arrives once the LLM has finished (≈ 1-3 s) with a discreet 'updated ✨' badge for 1.5 s. If a card temporarily fails (expired Google token, etc.), only that card shows its CTA — the others stay visible.
- Per-section cache with on-demand refresh — Each card has its own TTL adapted to its change rate (weather 1 h, agenda 10 min, mails 5 min, birthdays 24 h, reminders live, health 15 min). Per-card 'refresh' button visible on hover, 'refresh all' button at the top of the grid to re-fetch every source and regenerate the synthesis.
- Tokens and cost visible at a glance — Greeting and synthesis LLM calls display their actual consumption (token count + EUR cost) next to the timestamp, with a tooltip detailing the model used and the IN / OUT / CACHE breakdown. This consumption is also aggregated into your usual usage statistics, with no double-counting of cached tokens.
- Weather localized in 6 languages — Day abbreviations in the 5-day forecast strip now use the browser's date format (Mon, lun., Mo., lun, lun, 周一) instead of a fixed English label.
- Health card with aligned columns — The 'today' and '14-day average' values for each metric (step count, heart rate) are now laid out in a fixed grid: the vertical separators between today/average line up horizontally across both metrics, regardless of label length.
- Compact Quick Access — The previous 3 quick-access cards are replaced by 2 cards (Help + Settings) with refined icons and hover states. The 'Security' card with no actionable purpose was removed.
v1.17
v1.17.2 — Health data: the assistant dialogues with your measurements, generic ingestion APIApril 2026
- New 'Assistant' toggle in Settings → Health data — A single opt-in switch (off by default) authorizes the assistant to read your health data for 4 integrations at once: conversation, proactive notifications, memories, internal journals. While off, no health data ever reaches the LLM.
- The assistant answers your health questions — Ask in chat: 'How many steps today?', 'My heart rate this week?', 'Anything unusual in my health recently?'. LIA answers factually with totals, averages, and baseline comparisons. Never diagnoses — only reports numbers.
- Reliable answers over natural time ranges — 'This week', 'this month', 'the last 7 days' are interpreted correctly and the assistant surfaces the values directly in its reply (totals, per-day detail) without retreating behind 'no data'.
- Adaptive baseline (bootstrap → rolling) — LIA honestly qualifies its comparisons: the first 7 days are announced as 'based on N days of data only', then switches to a 28-day rolling median. Factual variation detection (streaks ≥ 3 days above 20% delta) and structural events (inactivity ≥ 3 days on steps).
- Proactive Heartbeat integration — When the toggle is on, proactive notifications can combine weather + calendar + health ('sunny + 4 days inactive → short walk?'). Injected signals only carry deltas / trends / events — never raw values.
- Generic ingestion API — The health API accepts measurements from any source (iOS Shortcuts, Android automation, third-party integration, IoT). When several sources cover the same period, the server merges them smartly (MAX for steps, average for heart rate). Landing and FAQ repositioned around this generic messaging.
- Extensibility for future measurement types — Architecture designed to accommodate new kinds (sleep, SpO2, calories…) without friction: ingestion, aggregation, baseline, Heartbeat, memory, and journal services adapt automatically.
v1.17.1 — Health data: daily batch ingestion, simultaneous Watch+iPhone handlingApril 2026
- Daily batch ingestion instead of hourly push — For your iPhone Shortcut to fire reliably, the iPhone no longer needs to be unlocked every hour. On unlock or at a fixed time, the Shortcut pushes the whole day's samples. Re-sending the same day is free: the server deduplicates automatically on the measurement interval.
- Two separate Shortcuts (steps + heart rate) — The v1.17.0 single hourly POST is replaced by two dedicated URLs:
/api/v1/ingest/health/stepsand/api/v1/ingest/health/heart_rate. Each measurement carries its own time range (ISO 8601 with timezone). Action required: reconfigure your iPhone Shortcut per the updated guide. - Smart fusion of simultaneous Watch + iPhone samples — When your Apple Watch and iPhone measure the same time interval, LIA now merges them intelligently: MAX for steps (each device counts a complementary subset of movement), rounded average for heart rate (two sensors aimed at the same signal). Fixes a 500 error that previously blocked longer-range uploads.
- Deletion by measurement type — Management actions under Settings → Health data → Management are revamped: delete all heart-rate measurements (step data kept), delete all step measurements (HR kept), or wipe everything. Each action truly deletes the affected rows rather than nulling columns.
- Chart UX improvements — Tooltips are now readable in dark mode (background and text follow the active theme). The Statistics section now displays the actual aggregation window (e.g. "Apr 21 18:00 → Apr 22 18:00") to defuse the "HR stats don't move when I change period" confusion — they're invariant when all measurements already fit inside the smallest window.
v1.17.0 — Health data: iPhone Shortcuts ingestion + chartsApril 2026
- New "Health data" section in Settings → Features — LIA can now receive heart rate and step counts from an iPhone Shortcut automation. Everything is visualized as a line chart (HR) and a bar chart (steps) with a period selector (hour / day / week / month / year) and dashed lines for the period average.
- Dedicated, revocable token auth — Generate a token (starts with
hm_…) from the app, copy it once (never shown again, only a display prefix remains) and paste it into your Shortcut. The token only authorizes health data ingestion — nothing else of your account. You can issue multiple tokens and revoke any of them independently. - Granular deletion — Three actions in the "Data management" sub-section: delete all heart rates (steps preserved), delete all steps (HR preserved), or wipe everything. GDPR-aware: raw HR and step values are never written to server logs.
- Step-by-step iPhone Shortcut setup guide — End-to-end documentation (steps A-E + hourly schedule) in
docs/guides/GUIDE_IPHONE_SHORTCUTS_HEALTH.mdwith payload example, troubleshooting and security notes. Default physiological bounds: HR ∈ [20, 250] bpm, steps ∈ [0, 15 000] per sample (configurable).
v1.16
v1.16.10 — Observability overhaul: 20 dashboards, full metric coverage, faster DBApril 2026
- Two new Grafana observability dashboards — A dashboard dedicated to sub-agents and skills (activity, token budget, rich outputs) plus another for ReAct mode and the browser (invocations, iterations, tool errors). Admins now have 20 dashboards covering 354+ panels, with no blind spots.
- 90+ Prometheus metrics revived — An audit surfaced that a significant share of metrics declared in code were never actually emitted in production (router, planner, HITL, OAuth, connectors, voice, drafts, initiative, data-registry, conversations). Instrumentation added everywhere, with defensive guards so metric failures never break the critical path.
- Faster DAU/WAU and connector activation queries — Three DB indexes added via migration (
obs_indexes_001):ix_conversations_updated_at,ix_conversations_created_at,ix_connectors_status. The gauge-updating background task drops from ~500ms (full scan) to <50ms on a populated DB. - No more Prometheus cardinality bomb on connectors — The
operationlabel onconnector_api_*metrics used to receive raw URL paths (with UUIDs, tokens, numeric IDs), blowing up Prometheus memory over time. A segment-by-segment sanitizer (UUID/id/hex_id/token → placeholders) is now applied upstream. - FastAPI validation errors (422) are tracked — Pydantic validation failures on incoming payloads are now counted per field + error_type (capped at 10 per request to bound cardinality), visible on dashboard 16 to surface bad-payload patterns.
v1.16.9 — LaTeX, code highlighting, history search and UX polishApril 2026
- One-click copy on messages and code — A copy button now appears on hover of every LIA reply, and on every code block. No more manual selection to grab an address, a sentence or a snippet.
- Math formulas rendered beautifully — When you ask LIA for an equation or a scientific calculation, formulas are now displayed with their mathematical notation (LaTeX). Examples:
$E = mc^2$inline,$$a^2 + b^2 = c^2$$for centered blocks. - Colored code with automatic language detection — Python, TypeScript, JSON, Bash, SQL... code blocks now render with syntax highlighting (25 languages supported). Light or dark theme based on your preferences.
- Search in your conversation history — A search bar has been added to the chat header. Type a word and LIA instantly filters matching messages among those loaded.
- Readable dates in the chat — Just the time for today's messages, "Yesterday 14:30" for yesterday's, the weekday for the current week, and the full date beyond. No more "Saturday, March 29, 2026" repeated under every recent message.
- Proactive feedback that sticks — When you click 👍 / 👎 / 🚫 on an interest suggestion, your choice is now persisted server-side and survives a reload or a device switch.
- Error messages in your language — Weather and Philips Hue smart-light responses are now translated to your preferred language instead of defaulting to French.
- Smarter mobile keyboard — On your phone, the chat input now enables auto-capitalization, autocorrect, and shows a dedicated "Send" button on the native keyboard instead of a plain return key.
v1.16.8 — Rich skills: cards, images, mini-apps right in the chatApril 2026
- Your skills can now render cards, images and interactive mini-apps — In addition to a text reply, a skill can now return an interactive HTML frame (Google Maps view, weather dashboard, monthly calendar, unit converter...) and/or an image (QR code, chart). Seven new built-in skills showcase this:
interactive-map,weather-dashboard,calendar-month,qr-code,pomodoro-timer,unit-converter,dice-roller. - Theme and language stay in sync with your frames — When you toggle dark mode or change the app language, the embedded mini-apps update live without a reload. Ergonomics stay consistent across the whole conversation.
- Frames now visible regardless of your display mode — Interactive mini-apps (skills, MCP apps, drafts) now show up in Rich HTML, Markdown and Cards display modes — previously they were invisible outside the Cards mode.
- End-to-end documentation for building rich skills — The Skills Guide now includes a complete "Localization, theming and runtime conventions" section in the Advanced tab, with copy-paste code examples (QR code with
segno, coin flip with animation and re-roll). The built-in skill-generator has been updated too, so it produces compliant skills on the first shot. - Skills settings: less visual clutter — The "Built-in skills" and "My imported skills" sections are now collapsed by default — when you've accumulated many skills, the Settings page stays readable.
v1.16.7 — Relevant weather alerts, even when you travelApril 2026
- No more nonsense temperature alerts at night — LIA no longer sends you "temperature dropping" notifications that just describe the natural day/night cycle. The detector now compares today's average temperature with tomorrow's — and alerts you in both directions (colder OR warmer) so you actually know how to dress for the day ahead.
- Weather alerts follow you when you're away (opt-in) — If you enable it, LIA can use your current browser position for proactive weather notifications when you're travelling more than 50 km from home. Your location is encrypted, never historized (each update overwrites the previous one), and wiped immediately if you turn the option off or remove your home address. The notification always tells you the city it's talking about, so you know where the forecast refers to.
- Full transparency on what's stored — A new "Currently stored" block under Proactive notifications shows the exact coordinates LIA has kept, when they were last updated, and a "Clear now" button to wipe them on demand. This is how privacy should be.
v1.16.6 — Sharper long-term memoryApril 2026
- Your factual corrections finally stick — When you say "actually, I changed jobs, I'm at Meta now" or "my daughter's name is Elsa, not Emma", LIA now updates the existing memory instead of creating a contradictory duplicate. Your history stays coherent.
- Pinned memories strictly protected — LIA never tries to modify or delete memories you've locked. You keep full control over what stays frozen.
- Less redundancy in what LIA remembers — A new daily cleanup automatically merges near-identical memories that accumulate over conversations. Your psychological profile stays crisp.
- Memories age more gracefully — LIA now better distinguishes what deserves to be remembered (durable facts useful for personalization) from what is transient (thanks, passing moods). Medium-importance preferences stop cluttering memory after a few weeks, while important information (family, identity, milestones) is preserved long-term.
v1.16.5 — Sharper conversational context trackingApril 2026
- "This meeting", "the first one", "delete it" — LIA finally keeps up — The assistant now remembers the last item you actually manipulated, searched for, or even just referenced. No more confusion between a meeting you just created and the one you're talking about now: when you say "delete this meeting" after asking "what was the first meeting returned?", LIA deletes the right one.
- Clearer HITL preview for updates — When LIA asks you to confirm a modification (meeting, contact, task), the preview is now structured as two clean blocks: "Changes" (only what actually changes, with old → new) and "Full details after update" (the complete post-update snapshot). No more contradictory "Other details unchanged" labels followed by values that just got updated.
- Consistency after create / update / delete — After an action is confirmed (e.g., a meeting created via HITL), the previously displayed list stays in sync automatically: updates propagate to search results, deletions remove the deleted item, and the conversational focus always tracks the latest handled item.
v1.16.4 — More reliable skill triggeringApril 2026
- Skills trigger more reliably on natural phrasing — LIA now understands your intent by reading each skill's description rather than relying on technical hints (detected domains). Variations like "I want my daily briefing", "Do my briefing", or "Summarize my day" now always activate the right skill, without misfiring.
- Custom skill authoring simplified — The
max_missing_domainsconfiguration field has been removed from the skill-editing guide. Skills trigger automatically when their description matches your request — no more fiddling with technical parameters for LIA to recognize them.
v1.16.3 — Enriched briefing, email fix & random phrasesApril 2026
- Daily briefing: emails and reminders now included — The daily briefing skill now consistently retrieves your recent emails and pending reminders, alongside calendar events, tasks, and weather. The issue where emails never appeared in the briefing is fixed.
- Skills compatible with all user profiles — If you haven't connected a service (e.g., Gmail), the corresponding steps are automatically skipped without blocking the rest of the briefing. You always get the maximum available information.
- Fun analyzing phrases are back — The witty waiting messages ("🔮 Consulting the crystal ball...", "🚀 Liftoff imminent...") are back during the initial analysis of your request, adding a touch of personality.
v1.16.2 — Progressive Execution Step DisplayApril 2026
- Real-time execution steps — You can now see each processing step as it happens: analysis, planning, validation, orchestration, and each tool call (emails, calendar, weather, contacts...). Steps accumulate visually and disappear when the response arrives.
- Tool names in ReAct mode — ReAct mode now shows which specific tool is being used (e.g., "Retrieving events...", "Searching contacts...") instead of a generic "Executing tool..." message. A reasoning snippet is also displayed.
- Pipeline steps now visible — Previously, only a random analyzing phrase was shown during Pipeline execution. You now see the actual progression: routing → planning → validation → task orchestration → tool execution → response.
- Improved tone consistency — LIA's personality and emotional state are now more reliably integrated into all responses, including proactive notifications, reminders, and voice comments. A safety guardrail ensures LIA never projects its own mood onto you.
- Reliable temporal memory — When you mention relative dates ("tomorrow", "next Saturday", "next week"), LIA now systematically converts them to absolute dates before storing them. Your memories stay reliable and relevant even weeks later.
v1.16.1 — LLM Configuration ReliabilityApril 2026
- Admin LLM overrides now fully enforced — Configuration changes made in the Administration panel (provider, model, temperature, reasoning mode, etc.) are now guaranteed to apply everywhere, including interest notifications and proactive content. Previously, some background tasks could ignore admin settings.
- Interest notification truncation fixed — Proactive interest notifications that ended with "..." mid-sentence are now complete. The issue was caused by an incorrect reasoning mode consuming output tokens.
v1.16.0 — ReAct Execution ModeApril 2026
- ReAct mode — New execution mode where the assistant reasons step by step: it calls a tool, analyzes the result, then decides autonomously what to do next. More autonomous but consumes 4 to 8× more tokens than Pipeline mode — best reserved for complex research or exploratory questions where the added value justifies the cost. Activate it via the 🔀/⚡ toggle in the chat header (🔀 = Pipeline, ⚡ = ReAct).
- Cross-domain initiative — In ReAct mode, the assistant proactively checks related information. Example: asking about Sunday's weather may also surface your calendar events for that day.
- Skills in ReAct — Your active skills are now available in ReAct mode. The assistant can autonomously activate and use them during its reasoning.
- Improved cost tracking — Token consumption and costs are now accurately tracked for all LLM providers, including in multi-step ReAct conversations.
- Execution steps in debug panel — ReAct reasoning steps (setup, reasoning, tool execution, finalization) now appear in the debug panel for transparency.
v1.15
v1.15.3 — Memory precision, interest deduplication, improved RAG searchApril 2026
- More precise memory retrieval — Irrelevant memories appear less often in conversation context, thanks to recalibrated similarity thresholds.
- Better targeted journal behavioral directives — Directives from your journals are now more relevant and less generic.
- Improved interest deduplication — Similar but distinct interests (e.g., Italian cuisine vs Japanese cuisine) are no longer incorrectly merged.
- Less repetitive content notifications — Proactive notifications now avoid sending you similar content repeatedly.
- More relevant RAG document search — Search results across your documents are more precise and better targeted.
- Optimized LLM models for extraction — Data extraction tasks now use more capable models for better results.
v1.15.2 — Psyche v2April 2026
- Psyche v2 — Expanded emotion palette to 22 emotions, graduated directives, serenity floor, narrative transitions, multi-emotion self-report, emotional resonance, proactive emotions, and enriched avatar tooltip.
v1.15.1 — MCP App Fix & ReAct TimeoutApril 2026
- MCP iterative timeout fix — Excalidraw diagrams and other iterative MCP tools were failing because the timeout was too short (60s) for the multi-step ReAct agent. The timeout is now automatically extended to at least 120s for all iterative MCP tools.
- MCP App widget display fix — Interactive widgets (like Excalidraw diagrams) could disappear from the response when the intelligent filtering did not recognize them as relevant results. MCP App widgets are now always preserved in the response.
v1.15 — Smart Memory Resolution, Optimization & SecurityApril 2026
- Iterative Mode for Your MCP Servers — Your personal MCP servers can now use iterative mode (ReAct agent). Enable it in settings so LIA reads the server documentation first, then executes tools step by step — ideal for complex APIs like Excalidraw
- 3-Phase Memory Resolution — When you say 'call my brother' or 'send an email to my wife', LIA now identifies personal references more accurately with targeted per-reference memory search. Fewer errors, less noise, more relevance
- Improved Error Messages — When issues occur (AI provider overload, content filter, timeout), error messages are now clear and translated in your language instead of showing technical codes
- Refined Proactive Initiative — Proactive suggestions after an action (weather before an appointment, contact related to an email) now focus only on different domains than what was already handled. Result: more relevant suggestions, less noise
- Safe Reminder Cancellation — Cancelling a reminder now requires confirmation before execution, preventing accidental deletions
- Localized Weather Dates — Weather forecasts now display dates in your language ('Monday, April 8' instead of raw ISO dates)
- Faster Responses — Internal prompts compressed by 30–70%, reducing response time and cost per query
- 'These emails / these appointments' Fix — Plural references like 'delete these emails' or 'modify these appointments' are now correctly detected to act on the right items
- Improved Contact Search — When the planner provides a name instead of a technical ID, LIA automatically searches for the contact by name instead of failing silently
- Iterative Mode Toggle Fix — Enabling/disabling iterative mode on personal MCP servers is now correctly saved and displayed
- Dedicated AI for MCP Apps — MCP servers with interactive widgets (like Excalidraw) now automatically use a more powerful AI model, configurable by administrators in the LLM Config panel under 'MCP App (ReAct)'
v1.14
v1.14 — LIA's Psyche: Dynamic Emotional IntelligenceApril 2026
- Response Display Mode — Choose between 3 display modes: HTML Cards (structured visual cards — default), Rich HTML (beautifully formatted responses) or Markdown (plain text). Configurable in Settings > Personalization
- LIA's Psyche — Your assistant develops a dynamic psychological state: mood (14 states), emotions (22 types), a deepening relationship, drives, and self-efficacy. Mood and emotions subtly influence the tone and style of every response
- Emotional Avatar — A mood emoji replaces the classic LIA logo on messages. The avatar reflects the emotional state in real time with a colored ring and detailed tooltip
- Emotional Dashboard — Visualize the evolution of mood (PAD), emotions, relationship, and drives across 4 interactive charts with time range selector (24h to 90d)
- Comprehensive Education Guide — 'Understanding the Psyche' section with 7 illustrated subsections: overview, personality traits, mood, emotions, relationship, drives, settings
- Customizable Expressiveness — Expressiveness (stoic → highly expressive) and stability (volatile → very stable) sliders to adjust emotional temperament. Mood refresh and full reset options
- Evolution Awareness — The assistant knows how its mood and emotions have shifted since the last message, providing narrative continuity to conversations
- Prompt Directives Overhaul — All context blocks injected into prompts now carry precise operational directives (how to use the data), improving response relevance
- Smart Skill Activation — Skills without a plan template are now automatically detected and activated. Skills with Python scripts run via a dedicated ReAct agent in an isolated loop; others inject passively with no extra LLM call
- Script Execution Fix — Skill script execution now works in all environments (Docker, Linux, Windows) with process isolation and robust error handling
- Faster Startup — Tool selection embeddings are now cached on disk. Server restart is faster and no longer consumes unnecessary API calls
- Proactive Suggestions Preserved — Suggestions generated by proactive initiative (weather, reminders, etc.) are no longer lost during intelligent response filtering
- Streamlined confirmations — Delete and modify operations now require only one confirmation instead of two or three. Confirmation titles are now explicit ("Confirm deletion", "Confirm sending", etc.)
- Anti-hallucination defense — The system automatically detects and strips parameters invented by the AI during planning, preventing silent failures on deletions and other operations
v1.13
v1.13 — Design System v4, Skills, Administration & Account LifecycleMarch 2026
- Account Lifecycle — Administrators can deactivate, delete (data purge with billing history preservation) and permanently erase (GDPR) user accounts in 3 secure steps
- Background Task Protection — Proactive notifications, reminders, scheduled actions and journal consolidation no longer consume AI tokens for deactivated or deleted accounts
- Visual Account Status — Admin panel visually distinguishes Active (green), Deactivated (red) and Deleted (gray strikethrough) accounts with state-appropriate action buttons
- Design System v4 — Complete visual redesign of all 14 HTML cards with reusable components, colored vignettes, chips, avatars and collapsible sections
- Skills Guide Redesigned — New comprehensive 3-tab guide (Fundamentals, Create, Advanced) with full agent, tool and parameter catalogue
- Enhanced Administration — 5 new counters in user management (Skills, MCP, Scheduled Actions, RAG Spaces, Blocked) and sortable Email/Blocked columns in limits
- Multi-provider — Calendar, Email, Contacts, Drive and Tasks tools now presented as compatible with Google, Microsoft and Apple
- DevOps Claude CLI — Administrators can diagnose servers via Claude Code in natural language directly from LIA
- Calendar DST Fix — Events created during daylight saving transitions now have the correct time
- Cross-domain context fix — "Show me the last email" after viewing calendar events now works correctly
- Route-to-calendar timezone fix — Travel time calculations to calendar events now use the correct local time (a 2-hour offset could cause arriving 2 hours too early)
v1.12
v1.12 — Image Generation & Web BrowsingMarch 2026
- AI Image Generation — Create images from text descriptions, with quality, size and format options
- Image Editing — Modify existing images using natural language instructions
- Image Preferences — Quality (low/medium/high), size (square/portrait/landscape) and format (PNG/WebP) configurable per user
- Live Screenshots — During web browsing, screenshots are displayed in real-time in the conversation
- Journal Deduplication — Similar entries are automatically merged to avoid duplicates
- Image Pricing — Admin panel to manage image generation pricing by model, quality and size
- Image Download — Download generated images and screenshots directly from the chat via a button or long-press on mobile
- Precise Hue Light Control — LIA now distinguishes a specific light (e.g., 'living room ceiling') from the entire room, thanks to automatic device discovery before each command
v1.11
v1.11 — Proactive Initiative & SecurityMarch 2026
- Proactive Initiative — After responding, LIA detects complementary signals and enriches its response (e.g., rainy forecast → checks for outdoor events)
- MCP Iterative Mode — Better Excalidraw diagram generation with a dedicated ReAct agent
- GPT-5.4 Support — New GPT-5.4 and GPT-5.4-mini models available
- Route Time Fix — Arrival and departure times now respect the user's timezone
- CC/BCC Modification — CC and BCC recipients can now be modified during draft review
- Enhanced Security — SSRF prevention, OAuth validation, parameter injection protection
v1.10
v1.10 — Voice & Push-to-TalkMarch 2026
- Mobile Push-to-Talk — Improved touch handling on smartphones (swipe cancellation, no context menu interference)
- Reduced Voice Latency — Parallelized and pre-warmed connections, faster end-of-speech detection
- Ready Chime — A short sound confirms LIA is ready to listen after the wake word
- Speech Language — Voice recognition uses the user's preferred language
- Safari iOS — Voice mode now works on Safari (iPhone/iPad)
- Rich HITL Review — Detailed draft display with emojis, before/after comparison and clickable links
v1.9
v1.9 — Usage Limits & ExportsMarch 2026
- Usage Limits — Per-user quotas (tokens, messages, cost) with monthly cycle and global limits
- Limits Dashboard — Admin interface to manage quotas and block/unblock users
- Smart Blocking — Chat is disabled with a clear message when quota is exceeded
- Consumption Export — Each user can export their usage (LLM tokens, Google API, summary) as CSV
- Journal Semantic Search — Migration to high-quality embeddings with search keywords
- Emotional Safety — Adaptive behavioral directives for sensitive topics detected in memory
- Email Fixes — Reply, forward, CC/BCC and recipients fixed across all providers (Google, Apple, Microsoft)
v1.8
v1.8 — Smart HomeMarch 2026
- Philips Hue Connector — Control your lights by voice: "turn off the living room", "set the office to blue"
- Hue Pairing — Bridge discovery and pairing wizard (local and cloud)
- 6 Smart Home Tools — List/control lights, rooms and scenes with fuzzy name matching
- Min/Max Temperature — Weather cards now show the day's temperature range
v1.7
v1.7 — Personal Journals & BlogMarch 2026
- Personal Journals — LIA maintains thematic journals (reflections, observations, analyses, learnings)
- Auto Consolidation — Periodic reorganization of notes with intelligent merging and synthesis
- Context Injection — Journals enrich both responses and planner reasoning
- Technical Blog — 20 detailed articles on architecture, integrations and features (6 languages)
- Self-Knowledge — LIA answers questions about itself through a built-in FAQ knowledge base
v1.6
v1.6 — Web Browsing & New ProvidersMarch 2026
- Web Browsing — LIA can navigate websites, fill forms and extract information
- Qwen Provider — Alibaba Cloud model support (qwen3-max, qwen3.5-plus, qwen3.5-flash)
- Ollama Models — Automatic discovery of models installed on your Ollama server
- 15 RAG Formats — PDF, DOCX, XLSX, PPTX, EPUB, CSV, HTML, RTF and more
- Google Drive Sync — Link a Drive folder to a RAG space for automatic vectorization
v1.5
v1.5 — Sub-Agents & SkillsMarch 2026
- Specialized Sub-Agents — Delegate complex tasks to experts (research, writing, data analysis)
- Skill Generator — Create your own custom skills using natural language
- Skill Management — Per-user and per-admin activation/deactivation
v1.4
v1.4 — RAG, Web Search & NotificationsMarch 2026
- RAG Knowledge Spaces — Import your documents to enrich LIA's responses
- Smart Compaction — Conversation history is automatically summarized to optimize costs
- Web Search Cache — Brave and Perplexity results are cached for faster responses
- External Content Protection — Web content is wrapped to prevent prompt injection
- Email Notifications — Proactive notifications now include your urgent unread emails
- /resume Command — Force history synthesis to free up context
v1.3
v1.3 — Stability & PerformanceMarch 2026
- FastAPI Update — Migration to FastAPI 0.135 for better streaming disconnection handling
- SSE Stability — Client disconnections during streaming are handled gracefully
v1.1
v1.1 — LAN AccessMarch 2026
- LAN Access — Configure access from other devices on your local network (mobile, tablet)
LIA is ready. And slightly impatient.
An assistant that acts, knows you and anticipates, while you stay in control of it.
Create my account