The Archive · Android

Changelog

What changed in each version of TALOS for Android, newest first, from the changelog of the release in force.

Changelog of TALOS for Android 0.1.40 read from the release · released Oct 2, 2026

What changed in each version of TALOS for Android, as its changelog says — read from the release of 0.1.40. Newest first; every published version is listed on Versions.

0.1.40

Released Oct 2, 2026 · Release

Signing in to OpenRouter no longer fights itself.

Providers

“Sign in with OpenRouter” now spends the authorization code exactly once. Coming back from the browser, the app could exchange the same code a second time, which OpenRouter refuses (“Invalid code or code_verifier”): the sign-in showed an error even when it had worked, or failed when the second exchange won the race. Measured on the OnePlus Pad 3 with a made-up code: two exchanges before, one after. A finished, failed or abandoned sign-in now leaves nothing behind to be exchanged again, and the diagnostics report says whether a failed exchange came from the live sign-in or from resuming one after the app was recreated.
If OpenRouter rejects the saved key (“User not found.”), the error card is unchanged: it tells you to renew the key in Settings. Signing in again, or pasting a new key, replaces it.

0.1.39

Released Oct 2, 2026 · Release

Long conversations keep going on the phone’s own model, the Qualcomm NPU ships, and chats can be selected, archived and restored in bulk.

On-device models

The Qualcomm NPU ships in the app, built with Qualcomm’s official SDK. It turns on only after you accept Qualcomm’s software terms in the app, once, when you pick a local model; they stay available in Models. On the OnePlus Pad 3 the first word arrives in about 1.0 s on the NPU instead of about 9.5 s on the GPU.
llama.cpp b11312, with Q4_K_M models running on the NPU.
GPU programs are prepared for each installed model before you first use it, instead of compiling for about 80 seconds on the first message after every update.
A model is fingerprinted once when it arrives, instead of being re-read on every open (24–29 s each time before).
On-device models find the tools they need by search instead of reading a 65-line index.
The context the phone can give a model is now measured correctly while the model is open: the memory already holding the open conversation is counted back, so long chats are no longer refused as they grow (before: “needs 6312 tokens, the device can give 3584”). The model’s head size is read from the file as declared (256 for Spark-X2.5-4B instead of 160), measured on the device at about 147,000 bytes of cache per token.

Chats

Conversations are summarized automatically when they fill the context, right after a reply; a message you send meanwhile waits. A line in the chat says “Conversation summarized · X → Y tokens” with Undo, and “Compact now” in the chat menu asks first. The numbers are real tokens: the phone’s model counts them itself, and hosted models use the provider’s count. A summary is kept only if the next request gets at least 30% lighter. Measured on the Pad with Spark-X2.5-4B on a coding task: 8,427 → 4,264 tokens after seven turns, and the conversation went on.
Chats can be selected from a button next to the count, then archived, restored or deleted together. The selection bar stays at the top while you scroll, and Undo brings archived chats back.
The TALOS logo fills three quarters of the assistant’s circle.
When a reply says an action is done (“Saved the note ✅”) but no tool ran, TALOS asks the model once to really do it, and if it still does nothing it says so under the reply: “No tool ran: nothing was done.” Seen with Spark-X2.5-4B saving a note.

Codice

The Codice summarizes long sessions the same way, keeps the full history, and can undo.
The Codice stays inside its folder: it can no longer write into TALOS’s own files or outside the session folder without asking.
Every failure is explained in words with one next action; interactive visuals are back.
The session list no longer shows raw keys such as HARNESS.GROUPS.LAST7 as group titles; a finished run says “Finished” instead of “Stopped”; the approval card says “approved” or “denied” instead of staying “waiting”; the strip counts “tokens” instead of calling the sum of every step “context”.

Permissions and safety

Tool permissions are answered at once, and the Stop button is never covered.
The non-commercial “hey jarvis” wake-word model is no longer shipped.

0.1.38

Released Sep 25, 2026 · Release

The chat list works like the other stations, replies keep their reasoning and their line breaks, and the Codice’s automations run real requests on the phone’s own day.

Chats

The chat list follows Notes, Memory and Tasks: search, Options and New chat on one row; tabs for All, In progress, Queued, Paused and Finished, with counts; filters for model, period and content (attachments, generated documents); sorting by activity, creation or title. Archived chats stay at the bottom. The floating New chat button is gone, so there is only one.
Each row shows the status and the time together, the model of the last reply as its provider’s logo and a short name, icons for attachments and generated documents, and “new reply” until you open the chat. The sidebar shows the status and new replies too.
A chat whose queue is paused says “Paused” in the list and the sidebar, as it already did inside the chat. The turn queue fits one row on tablets and two rows on phones, with the app’s buttons.
Replies keep the model’s single line breaks: lists, addresses and verse no longer run together.
A reply that stops at the length limit says so and offers Continue, with OpenRouter, OpenAI, DeepSeek, Anthropic, Gemini and Ollama.
Models that always reason (111 of 458 on OpenRouter, counted on 24 September, GLM 5.3 Flash among them) always get a reasoning request, so their reasoning no longer spills into the reply.
File names such as Peloro.md or setup.sh in a reply are no longer links to websites (.md is Moldova’s domain). Web addresses, bare domains and emails still are.

Permissions

The permission sheet names the common fields in words (Title, Text, Format, Path, Address…), says which chat is asking in one short line, and closes the keyboard when it appears.
After you decide, the reply opens with “Permission granted: <tool>” or “Permission denied: <tool>” on one line, instead of a separate “handled” bubble. The reasoning the model wrote before asking moves into that reply.

Codice

Automations run a written request as a real Codice session, with the folder, model and reasoning level you chose. Each one remembers its last session and pauses itself, with the reason, when a start fails. The daily limit resets at midnight on the phone, not at 00:00 UTC.
The model picker lists the whole OpenRouter catalog, grouped by provider and searchable (460 models on 25 September). On touch screens it no longer opens the keyboard by itself.
The reasoning bar has seven levels, up to Maximum, and models that require reasoning get their required level on every call, automations and compaction included.
The Codice is in Italian throughout when the app is: side panel, command palette, sheets, session tree, about fifty notices and twenty accessibility labels. Demo data is no longer shown as real.
Updating the app no longer leaves the previous Codice server running in the background.

Everywhere

Android’s own checkboxes are replaced by the app’s, in its palette, in seven screens: chat deletion, backup, voice, personal voice, task schedules and model options.

0.1.37

Released Sep 15, 2026 · Release

Scrolling on station pages keeps Android’s native momentum again: the collapsing header added in v0.1.34 no longer takes over flings. Published for device testing of the scrolling regression reported on v0.1.36.

0.1.36

Released Sep 15, 2026 · Release

The Library loads more files as you scroll, with search and filters covering the whole collection. Item menus stay within the screen; phone headers and bottom spacing are corrected.
Download Center controls remain usable when opened from the navigation drawer.
Startup validation works with the existing strict Content Security Policy, without blocked dynamic code compilation.
Restored browser regression coverage for attachments, message actions, saved reasoning, chat persistence, appearance settings, and phone/tablet layouts.

0.1.35

The Library loads more items as you scroll, starting with 24. Search, filters and the total count still cover the whole collection; thumbnails and saved-link icons follow the displayed pages.
Item menus keep their full width inside the screen and open above their button when the actual menu height does not fit below it. Long menus can scroll.
Phone station headers put the title and add button on one row, with the subtitle underneath. The Library uses the same background as its enclosing sheet.
The last item in a station stays clear of the bottom edge. Model Lab categories and navigation labels follow the selected language.

0.1.34

Released Sep 14, 2026 · Release

Every station now has one toolbar, one options sheet and a menu on each item; each chat keeps its own draft; and “Run details” shows what a reply cost and how long it took.

Messages

Long-press a message, or tap its ⋯, for one “Message actions” sheet. A reply offers Share, Run details and Delete; your own message offers sending it again, Reuse prompt and Delete. Delete takes the question and its answer together, with Undo for five seconds, and files the reply created stay in the Library.
“Run details” lists the tools a reply used, what each one looked for and how it ended, the input and output tokens, the time to the first word and in total, and the cost: the figure OpenRouter reports, a marked estimate from its public price list when a provider reports none, and “Free, on this phone” for a model running on the device.

Drafts

Every chat keeps its own draft, attachments included. A new chat no longer inherits the text of the previous one, and a chat you started typing in stays in the history marked “Draft”.
The composer keeps its space in an empty chat too, so the last line no longer slides under it.

Stations

Library, Memory, Notes, Research and Tasks lose the fixed bar at the top. The large title starts the page and, as you scroll, folds into a slim bar that follows your finger; on a phone only the round menu button stays.
One toolbar row everywhere: search, list or grid, options, and the main action (“Add file”, “New note”, “New memory”, “New task”), an icon only on a phone.
Sorting moved into an options sheet drawn by the app, instead of Android’s own picker. In the Library it also holds “Group by chat”, and “Name (A–Z)” now sorts the files themselves, not only the groups.
Library: tabs All · Files · Images · Links. A card says the extension, or the site, and the chat only when grouping is on — no dates, no group headings, and “Excluded from context” only on a file that is. Every card has a ⋯ menu, links included: Open in browser, Attach, Save to phone, Delete. Long-press or right-click a file to start selecting. A Markdown file opens formatted, with Text one tap away. Preview titles stop at 35 characters and are cut at a word, so no line shows half its letters.
Research: four tabs — All, Running, Finished and To resume, which gathers the paused, the interrupted, the ones blocked by a permission and the ones that never concluded. No readable report, Cancelled and Failed are picked from the options sheet and show up as a fifth tab while chosen. Dates read “12 Sep”.
Tasks: a card’s ⋯ sits bottom right, as in the other stations.

The “+” sheet

Every entry has a bare amber icon on the same vertical line as the four categories above it: 28–29 px and 58 px from the sheet edge on tablet, phone and landscape. The entries used to sit 10 px and 23 px further in, inside a tinted tile.

Fixed

A page that stopped scrolling — the Library after starting a selection — kept the folded bar on and the large title hidden at the top.
“1 selected” in the Library read as a plural.

0.1.33

Released Sep 13, 2026 · Release

The composer no longer keeps an empty line under what you type.

The composer

With the field focused and empty, the box is one line tall instead of two. The placeholder used to sit half a line above the “+” and the send button; now it sits on the same line as they do.
Behind that was an arithmetic fault worth naming: the browser reports the field’s content height as a whole number, so a single 25.6 px line was reported as 26 — and rounding that up produced two lines. A line now has to be exceeded by more than a twentieth before the box grows, which is enough to absorb the rounding and never enough to hide text.
The box stops shrinking at about 128 px: two rows of 48 px — the controls and the tools beneath them — plus its padding. The field is no longer what decides the height.

0.1.32

Released Sep 13, 2026 · Release

The composer lines up with what you type, and the sheet behind “+” follows your finger instead of jumping.

The composer

The “+” and the button on the right sit on the middle line of the message field, whether the field holds one line or eight. They used to hang from the first line: 16.8 px above centre with the field empty, 39.7 px above it once a few lines were written, and at rest they overhung the top edge by 4 px.
The field is always a whole number of lines. It used to stop at 192 px with lines of 25.6 px — seven and a half — so the last line you could see was cut in half. It now holds seven whole lines and scrolls past them.
With the field empty the box is one line tall instead of three and a half. The taller floor is there to keep the box steady while it grows, and nothing grows when there is nothing in it.
The microphone and Send share a filled accent container again, with the same corner radius as the composer around them.

The “+” sheet

Dragging right to left across the list of actions moves between Attach, Create, Tools and Agent, and the actions underneath change with it.
The panel follows your finger as you drag, moving a little under half as far as your thumb and fading slightly on the way. Let go past the threshold and the new tab settles in over 230 ms; let go short of it and the panel springs back over 250 ms. At the first and last tab it gives only a little, and stays.
A drag up or down still scrolls the sheet and changes nothing, and a drag too short to be meant as one changes nothing either.

0.1.31

Released Sep 13, 2026 · Release

The composer has one button again, and it changes with what you are doing. The waiting mark is a single quiet orb.

The composer

The button beside the message field is the microphone while the field is empty, and becomes Send the moment you type. While TALOS is answering, or while you are dictating, it is Stop.
That button now sits next to the field itself, not in the row of tools underneath it.
The Library-context pill is a single icon, next to the model pill. Its mode and the number of sources it will use are still read out by screen readers, and tapping it still opens the sheet where you choose them.
Where dictation is unavailable the microphone stays where it is, switched off, with the line underneath that says why. A control that disappears explains nothing.
One thing is gone on purpose: the second microphone that appended dictation to text you had already written. With text in the field the button is Send. Dictation started on an empty field still appends to itself as before.

While TALOS is thinking

The three animated dots are gone. The wait shows only the TALOS orb with its turning ring, and it disappears at the first letter of the answer.

0.1.30

Released Sep 12, 2026 · Release

(Tagged as v0.1.30. A v0.1.29 tag exists on the repository but never produced a build: its CI run failed on a timing-dependent unit test, fixed here. There is no v0.1.29 release.)

The chat is now on the same calm design as the rest of the app, Settings keep the sidebar beside them on a tablet, and one search box reaches everything you have in TALOS. Your own recorded voice reads replies again.

The chat, redesigned

An empty chat opens on “What shall we do today?”, three prompt chips (presentation, analyse a file, run a research) and the two most recent conversations to pick up from, with the composer always docked at the bottom.
The composer is one shape: a round “+”, the model, a “Think” chip that switches reasoning on and off, the microphone and a round send button that becomes a stop button while a reply streams.
The “+” opens “What do you want to do?”: a search box and four categories (Attach, Create, Tools, Agent). Every entry does something real: attach a file, pick an image, take a photo, use the Library, improve the message, prefill a presentation, document or analysis, a new note or memory, phone control, Tool Forge, Models, the switch for the model’s tools, a task, a research, and Code where the phone bridge exists. The sheet springs up, the highlighted category slides, and the entries glide in, with the mockup’s timings.
Every reply carries a header with the TALOS orb (spinning while it works) and the model and time on the right. Your question is an accent bubble; the reply runs full width, or sits in a bubble if Appearance is set to bubbles. Actions live under each message: copy, edit, more on yours; copy, listen, regenerate, save to Library, more on the reply.
Edit a message: reopens your text in the composer and removes the turns after it, after a confirmation.
Stopping a reply keeps the text and marks it “Interrupted. The text is kept.” Reasoning expands in place.
Web browsing left the chat: no link pill, no “Browse mode” banner, no “/browse” command. The local browser stays available to the model’s tools. “Run a research” on the home opens Research, not a web search.

One search for everything

“Search Talos” in the sidebar opens a dialog that searches chats (titles and the text of messages), notes, tasks, memories, Library files and Settings categories, sixteen results at most, ranked. Arrow keys, Enter and Escape work; on a phone it is a sheet. The chats list now finds a conversation by a word that appears only in its messages, and shows the excerpt.

Tool Forge is back in the sidebar (it had dropped out of the redesign).
Press and hold, or right-click, a recent chat in the sidebar or a row in the chats list to get its menu: Open, Rename, Archive, Delete (and Select in the list). The Account button opens the Account page.
Recent chats and “Pick up where you left off” no longer list archived chats or Code sessions.
On a tablet the sidebar can be resized by dragging its edge, and a conversation shows a back arrow to the chat list.

Settings on the calm design

On a tablet the sidebar stays beside Settings; categories sit in a column with the detail next to them. “All settings” lists every category with a subtitle.
New controls, each tied to something real: “Approval duration” and “Revoke saved authorizations” under Agent Tools; a grid of the fourteen themes; frame-rate and pixel-ratio caps under Motion; three consents (Library access, memory writing, image attachments) under Privacy; Reminders links to Tasks and System to Doctor. Reading controls come before your personal voices. Two controls that no longer changed anything (composer shape and “+” surface, link suggestions) were removed. “Web search” is now called “Search engine”.

Small things that now move, and a few that now work

Ticking a checklist item on a note, from the card or the note itself, changes the note.
Research keeps its own list-or-cards choice; it no longer follows the Library’s.
Archiving a chat shows “Chat archived” with an Undo for eight seconds.
Confirmation dialogs grow out of the button that opened them (280 ms) and shrink back (210 ms); persisted messages slide in (220 ms); the groups under Agent Tools open and close by height (260 ms); the touch ripple follows the “Transition duration” slider. All of it respects “reduce motion”.

Your recorded voice reads replies again

Recording a voice worked, but playback in a chat was silent. Four causes in a row were fixed: the language was not passed, the real error was discarded, the start of speech was locked behind a check that never passed, and the audio buffer filled up and stopped. The recorded voice now follows the same speed slider as the synthetic voices, the first word arrives in about 2 seconds, and the long pause after a comma is gone.

Fixed

A model the app had fallen back to could be written into a chat as if you had chosen it; reopening that chat then showed the wrong model.
The draft of a brand-new chat was lost if you left before sending.
A tool call that GLM wrote inside its text showed up as raw tags; it is now executed like any other tool call.
The “Improve the message” entry no longer speaks of “prompt”.

0.1.28

Released Sep 12, 2026 · Release

The interface has been rebuilt on one calm, consistent design across the whole app — the sidebar, every station (Notes, Tasks, Memory, Library, Research) and the motion that ties them together — and each station now lets you create, open, edit and delete its things from the app itself.

The sidebar is the same on every screen: search, the eight sections, recent chats, then Diagnostics and Settings. On a tablet it stays open beside the content instead of sliding in; on a phone it slides in with a spring that settles in about half a second and follows the “Transition duration” slider in Settings (measured on a tablet: 525 ms at the default setting, 961 ms at double). Models moved from Settings into the sidebar. Each station opens with its own icon and title in the header, and a back arrow only once you are inside it.

Notes, Tasks, Memory: complete, not read-only

Notes can be pinned, edited, exported as text and deleted. Markdown renders as text, not as symbols.
Tasks have the states To do, In progress, Scheduled and Done, a checkbox that completes in one tap, a checklist you can tick inside the task, an edit screen, export, and a real pause for scheduled tasks: a paused task does not run until you resume it.
Memory filters by type (Preference, Project fact, Procedure, Policy note) plus a “To review” shelf for what the model proposed and you have not confirmed. Each memory can be edited, paused, exported and deleted, and shows where it came from.
Writing a note, a task or a memory yourself no longer rings the notification bell. Only things the model did do.

Library: real previews, and adding a file no longer sends it anywhere

Files in the Library show what they are: images as thumbnails, PDFs rendered from their first page on the device, text files as a typographic preview. Adding an image to the Library used to quietly queue it for your next chat message and ask the “this image leaves your phone” question at the wrong moment; adding is now just adding. The question is asked when you actually send an image to a remote model — including images that were already in your Library — and “No” keeps the message on the phone.

Research: a report is only “done” when it holds up

A research now ends in one of four honest states: concluded with a readable report; blocked by a permission; stopped before the report was written; or finished with a report that cannot be read back. A run that stopped because a step failed says so, with the step’s own error, instead of guessing. Cards use singular labels, say whether sources were already gathered, and update the moment a run finishes — no reopen needed. A quick research measured on a tablet with GLM 5.3 Flash writing and a local Qwen3-4B checking citations: 3 min 21 s, 8 sources, 26 claims verified.

Codice tells you when the phone bridge is off

Entering Codice with the phone bridge switched off now shows one message saying what does not work (the terminal and session commands), the two steps to switch it back on, and a button that opens Settings → Phone control. It appears once per visit and closes by itself when the bridge comes back.

Errors in plain words

Every error the chat can show now leads with a sentence about what happened and what to do; the technical code stays small, underneath, for diagnostics. Local models no longer show “unknown” in the model picker.

Plain words in the model picker and error cards

Local models in the picker say “On this device” instead of showing the file path, a status nobody translated is omitted rather than printed raw, and an error card for a local model names the model, not its path, and suggests what to do with a local model rather than “check the provider”.

Fixed

Row menus inside a station could open invisibly behind the sheet.
Tapping the dimmed area beside the sidebar did not close it.
A stopped research showed a running timer next to “Stopped”.
The sessions list in Codice showed a raw label instead of “Last 30 days”.
The “listen” action under a message had no accessible name.
Research plan chips read “5p” and “10 / 3m”; they now say “5 pages” and “10 pages · 3 min”.
Codice’s empty session and status strip spoke English inside an Italian app.
On a phone, the Codice bridge warning was drawn underneath the station.

0.1.27

Released Sep 6, 2026 · Release

The “+” menu is less cluttered: it no longer shows a second copy of the reasoning/effort control, and it no longer offers a web-browsing toggle that, on the default composer, had nowhere reliable left to live.

One fewer place to set reasoning effort

The “+” menu’s organized drawer showed its own reasoning/effort picker — the exact same control already reachable by tapping the model name, with no indication the two were connected. It’s gone from the “+” menu now; the model name stays the one place to set it.

Both the “+” menu’s organized drawer and its dropdown listed a “Browse the web” toggle. Removed from both. On the classic composer it still has its own icon button in the toolbar, unchanged. On the default composer (the newer, Claude-style one) there is currently no way to turn browsing on or off from the “+” menu — a deliberate simplification, not an oversight, and not yet replaced with another way to reach it from that composer.

0.1.26

Released Sep 4, 2026 · Release

Codice’s Terminal tab can now run a command itself, not just show one, its interface follows the app’s language, and web search works without a key.

Codice’s Terminal tab can run a command, not just show one

The Terminal tab showed a live log of commands the agent ran, but had no way for a person to type one in themselves — that ability existed only one screen over, reachable by typing !command into the chat. The Terminal tab now has its own composer that runs a command directly on the device, with an honest note when a session is still busy instead of a field that quietly does nothing.

Codice follows the app’s language

Codice’s interface — its tabs, its Running/Stopped state, the Terminal’s own messages — stayed in English regardless of what language the rest of the app was set to. It now follows the app’s language, falling back to English only where nothing has been translated yet. The Terminal’s new composer (above) was also restyled to match the pill-shaped, glass composer used everywhere else in the app, instead of standing out with a look of its own.

DuckDuckGo: a fifth web search option that needs no key

Every web search option until now — Tavily, Brave, SearXNG, a custom endpoint — needed a key or an address before the model could search the web at all. DuckDuckGo needs neither: it reads DuckDuckGo’s own public results page. It isn’t an official API, and under heavy use it can answer with a block instead of results — when that happens TALOS says so plainly, rather than reporting an empty web.

Fixed

The Terminal’s composer could end up sitting partly behind the bottom navigation bar on narrower screens, its send button cut in half.

0.1.25

Released Sep 3, 2026 · Release

The previous release corrected one bug by introducing its opposite: Codice stopped showing on a signed build entirely, instead of showing correctly. This release makes Codice real in the signed APK — visible, and able to run a command on the device it’s installed on.

Codice ships in the signed build

The native bridge behind Codice — the piece that lets a session run a sandboxed terminal, not just chat — lived in a part of the project that Android only compiles into development builds, since the day it shipped. Announcing Codice as a phone-hosted coding agent while that piece never reached the APK anyone could actually download was the root of both this release and the previous one. It now compiles into every build, signed release included, and was verified on one: opening Codice, starting a session, and running a real command all work on a production-shaped APK, not just a development one.

Fixed

A session hosted entirely on the phone — no computer involved — couldn’t run a single shell command: the run always failed as if no device were reachable, even though the device running it was the one right there. It now runs the command directly, the way it always should have on a session that was never talking to a second device in the first place.

0.1.24

Released Sep 3, 2026 · Release

The previous release could show “Codice” as available on a signed build — a debug-only feature with nothing behind it once you tapped in. This release makes sure that never happens again, and fixes the run-strip numbers, a composer draft that didn’t survive leaving the screen, and a few smaller Codice surfaces found while verifying it.

Codice never shows on a build that can’t run it

Codice needs a native bridge that only compiles into development builds — that has been true since it shipped, but nothing enforced it everywhere. The tablet layout decided whether to show Codice’s chrome by looking only at the current route’s name, and Vue Router restores whatever route was open last — so a phone that had ever run a debug build could boot straight into a Codice screen on a signed one, with the native side simply absent. A single guard, checked before every way of reaching the screen — a direct link, a route restored at boot, or in-app navigation — now redirects to chat instead whenever the build can’t run it. Codice is not, and has never been, in the signed release APK — v0.1.23’s notes described it without saying that; consider this the correction.

Fixed

When the on-device bridge a session depends on isn’t connected, the app now says so in plain language instead of leaving the run stuck or failing with a technical error.
The run strip’s step count, context size, and error count always showed a placeholder dash, even while a session was actively running with real numbers behind them.
A failed tool call didn’t say why in the run’s summary; it now shows the failure reason, and the warning icon on the affected step is read by a screen reader instead of hidden from one.
A prompt typed into the composer and left unsent used to disappear if you left the screen; it’s restored when you come back.
The button that opens Codice from the tablet sidebar was below the minimum touch-target size.
The executor-model list inside effort settings showed the same “Models” label as the primary model picker, with nothing to tell them apart.
A few remaining Italian strings in Codice’s menus and toasts are now in English, and the Hugging Face access status labels are sentence case like the rest of the app.

0.1.23

Released Sep 3, 2026 · Release

The coding agent moved onto the phone. Until this release “Codice” was a screen that showed what a session would look like; it now opens real sessions, runs real commands in a sandboxed terminal on the device itself, and talks to five model providers instead of one. The personal voice engine was rebuilt on Pocket, and the Model Lab stopped guessing at what a model costs.

Codice: a real coding agent, on the phone

Sessions are real and they persist. They survive closing the app and restarting it, and the list shows what each one is doing right now — running, waiting for an approval, finished, or finished with an error — instead of a static row.
A sandboxed terminal runs on the phone. No PC, no tunnel, no server on the other side of the room: commands execute on the device, through the ADB bridge, and their output comes back into the session.
Any model from OpenRouter can be picked live. OpenAI, DeepSeek and Ollama are wired to real sessions now, not just stored; Anthropic and Gemini are recognized and refused with an honest reason — their request shape isn’t translated yet — instead of either working silently or being hidden.
Tool calls collapse into groups with a per-file diff, so a long run reads as a list of changes instead of a wall of output.
The permissions pill and the model pill act on a live session instead of only looking like they do.
Long conversations are compacted, and now you can see it happen instead of wondering why the run paused.
The composer got the rest of its work: a Planner picker, a real “improve this prompt” that either rewrites it or says why it cannot, and an export that produces the actual session report.
Automations, Doctor, hooks, sub-agent delegation, the review centre and workspace files (drag and drop, copy, create file and folder) all came across from the desktop build rather than being written twice.
Deep research, Activities, Memory and the Library are connected to the kernel — read-only for research, on purpose.

Personal voice

The engine now runs on Pocket end to end: enrolment creates v2 profiles, older profiles are migrated instead of being abandoned, and installation is observable rather than silent when it fails.
A model is loaded on demand, when an intent needs it, instead of being held resident.
The Doctor screen reports performance signals from the engine, so a slow or stuttering read can be diagnosed from the report instead of guessed at.

Model Lab

The KV cache type (F16 or Q8_0) is a setting rather than whatever the file’s header happened to use, with the resolved type always shown.
A context-length slider with real reference points, and a provenance ledger that says where each number in a memory estimate comes from.
Every quantization is checked against your device as soon as the page opens, instead of waiting for a tap on each one.

Known, and not hidden

Local models cannot yet be used for a Codice session. The on-device engine is reachable only from the app’s web view, while the agent kernel runs in a separate process, and the bridge between the two does not exist yet. Choosing one is refused with that reason instead of failing halfway through an answer.
Anthropic and Gemini models are recognized in the picker but refused for Codice sessions: translating tool calls into their own request format isn’t done yet. Only OpenAI, DeepSeek, Ollama and OpenRouter route to a real call today.
The “Agents” tab in the context rail is still a mock-up. Delegation itself is real — the session tree reads actual child sessions — but that one panel is static markup that no code fills in yet, so the names in it are examples, not your sub-agents.

0.1.22

Released Aug 24, 2026 · Release

Browsing models on Hugging Face got a rebuild, and the Doctor screen can now see the personal voice engine instead of staying silent about it.

Browsing models on Hugging Face

The model detail page is now three tabs instead of one long scroll: quantizations, the full model card, and the raw file list. Every quantization is checked against your device automatically as soon as the page opens, instead of waiting for a tap on each one.

A resource ledger shows exactly where the memory estimate for a model comes from — weights and file size are exact, the KV cache size is exact once the header is read, compute and runtime overhead are the app’s own safety policy — instead of a single unexplained number.
The KV cache type (F16 or Q8_0) can be forced globally instead of only reading whatever the file’s header happened to use, and the resolved type is always shown next to the control, even on Automatic.
A direct link to the model’s Hugging Face page, and a button to copy it.
The context-length slider shows its 2K/32K/64K/128K reference points instead of a bare track.
A rounding bug that showed predicted speed as a fifteen-decimal number is fixed.

Fixed

The Doctor screen had no way to see the personal voice engine at all — only speech recognition was checked, never speech synthesis. It now shows whether the engine is installed, whether a usable voice profile exists, and a short log of the last few read requests, so a silent playback failure can be diagnosed from the report instead of guessed at.

0.1.19

Released Aug 23, 2026 · Release

The personal voice no longer stutters, and the local model runs faster.

The voice engine was rebuilt on Pocket TTS

The previous engine produced stutter and hardware underruns during real playback, and sentences that followed one another too closely would overlap or get their start clipped. The voice engine is now built on Kyutai’s Pocket TTS (MIT), run fully on-device through ONNX Runtime — no change to what stays local: your recordings never leave the phone.

Stutter and the measured hardware underruns are gone. Overlaps and clipped sentence starts between concatenated replies are fixed.
Sentence starts are stabilized without making the technical lead-in audible.
Measured on production streaming: -23.2 LUFS, true peak -0.8 dBFS, Pocket gain +12 dB with a limiter at -1 dBFS.
Time to first audio, warm: 449–537 ms; the engine now enforces a 600 ms honest cutoff rather than reporting a number it cannot back.
Italian speech recognition against 6 reference sentences: 6/6 exact, zero word error rate, no dropped first or last word.
The AudioTrack buffer grew to 1.205 s to absorb the timing variance that caused the underruns, without delaying playback start — the initial threshold that controls when audio starts did not change.
Existing voice profiles migrate to the new engine automatically; nothing needs to be re-recorded.

The local model runs faster

Continuing work on the on-device engine (llama.cpp):

A persistent OpenCL kernel cache removes a multi-second first-generation recompile that ran every time the app started.
The microbatch size for prompt processing moved from 192 to 512 tokens, measured to shorten prompt processing without changing output quality.
The static parts of the system prompt are now pre-computed once (AOT) instead of on every message, including a fix for models that need more than a system turn alone to accept it correctly.
A native thread pool replaces manual thread lifecycle management. CPU core affinity was measured explicitly (paired A/B campaign) and found to make no measurable difference on this hardware (≤1%, inside the ~2.4% run-to-run noise) — the mechanism ships as tested infrastructure, but default behavior is unchanged rather than switched on speculation.

Fixed

The instruction telling a local model which language to answer in was being skipped in one case, so a tool result in another language could pull the reply into that language even when everything else was in Italian. The instruction now always names the exception, in both the short local-model prompt and the full one, and stays within the prompt’s measured size budget.
The background scene now follows the selected motion complexity on its own, instead of being tied to the color theme.

0.1.18

Released Aug 22, 2026 · Release

The app can speak in your own voice. You record a set of phrases once, the device learns from them, and from then on a message can be read back in that voice instead of the stock one.

Teaching it your voice

The wizard walks you through twelve phrases in three styles, shows a real waveform that moves with your voice while you speak, and checks the recording quality before accepting it. The quiet-room screen before the wizard now shows a live microphone level too, so you can tell whether the room is actually quiet before you start. The level meter is on only on that screen and switches off the moment you leave it.

Your recordings never leave the device. They are encrypted with a key held in the phone’s hardware keystore, and once the voice profile is built the raw audio is destroyed.

It no longer dies while learning

The first real run with twelve phrases killed the app. Not “ran slowly” — Android’s low-memory killer terminated it, and said “process memory is leaking”: about 5.8 GB resident and 5.3 GB of swap.

The first attempt at a fix — turning off ONNX’s memory arena, which is the official recommendation for small models — brought it down to 3.5 GB and still died. That was reported as insufficient rather than shipped as a fix.

The real cause is that the encoder’s attention costs memory in proportion to the square of how much audio you hand it, which matches the memory climbing steadily throughout the encode rather than spiking at load. Twelve phrases concatenated is twenty to forty seconds of audio, and that is past what the model is built for: its own documentation puts the sweet spot at three to ten seconds and warns that clips beyond about fifteen “may introduce noise artifacts or degrade quality”.

So the reference is now built from the four normal-voice phrases rather than all twelve, with a hard twelve-second ceiling behind it that holds no matter what reaches it. A shorter reference is not a compromise for the sake of memory — it is the better recording. Measured after the change: the encode finishes in 5.4 seconds and the app survives.

The quality gate still runs on all twelve phrases. Only the audio that reaches the encoder changed.

If it crashes, you do not start over

Each accepted phrase is written to disk encrypted as you go. If the app is killed halfway through — or crashes — reopening the wizard picks up at the next missing phrase instead of asking for all twelve again, and it does not make you redo the consent screen or the microphone check. Verified twice: once with a forced stop, and once with a genuine crash that happened while chasing the bug above.

A personal voice can now actually be chosen — and it works

Recording a voice, and reading with it, used to be two separate stories. The voice showed up in “Voice personale” with rename and delete, and that was the end of it: nothing let you pick it as the voice that reads replies, and there was no way to hear it again once the wizard closed.

Both are fixed. A recorded voice now appears at the top of the same picker that lists the device’s own voices — pick it and it becomes the one that reads, exactly like picking any other voice. Every profile also gets its own “Listen” button, independent of which voice is currently active, and the one actually in use is marked so on its card. That memory jump alone turned out to prove only that synthesis had started, not that it finished — two of the fixes below were found this same way, by listening to what the device actually said rather than trusting the jump as success. With those closed, the chain now runs end to end for real.

A voice you record is now also tagged with the phone’s actual system language rather than whatever language the app’s interface happened to be set to at the time — those can differ, and only the system one is what should be on the label.

Fixes

Talking to it used the stock voice even with a personal voice chosen. When a reply is read aloud automatically because you spoke to it — the “you talk, it answers back in voice” path — the app always used the stock system voice, even with a personal voice selected. Deliberate at the time: that path speaks sentence by sentence as the reply streams in, queuing each one behind the last, and the personal engine had no real queue to speak into — a second sentence sent there would have cut the first one off instead of following it. It now waits for the reply to finish streaming and reads the whole thing in one call to the personal engine, the same way pressing the speaker icon on a finished message already did. You lose the word-by-word start; you gain that the voice you actually chose is the one you hear, which matters more.

A file name or code term with an underscore was read as “underscore”. documento_complesso came out as “documento underscore complesso” — a known symptom of handing a TTS engine raw text straight from a markdown-formatted reply. Underscores, **bold**, *italics*, and `inline code` markers are stripped before anything reaches either voice engine now; the words themselves are never touched.

A voice that had been renamed since you picked it read nothing. If the personal voice your settings pointed at no longer matched a saved profile — renamed, or replaced by a new recording — pressing the speaker icon produced only a generic error instead of reading the reply. It now falls back to the stock voice silently, the same way it already did when no personal voice had ever been chosen.

The device voice you actually heard could be a different one than the one you chose. Applying a chosen voice to the system engine can be refused — most often a network voice when the network is not reachable at that instant — and that refusal was never checked: reading went ahead anyway, on whichever voice the engine happened to have already. A refusal now retries once against a voice that does not depend on the network. Separately, the very first reading of a fresh app launch could run before the device’s own voice list had finished loading, silently skipping the choice altogether — reading always landed correctly from the second attempt in the same session, never the first. Both were real, found by listening to the device rather than trusting a memory measurement that had only confirmed the engine started, not that it finished successfully.

The voice engine’s files were showing up as chat models. After installing the voice engine, its two files sat forever in the same on-device folder the chat’s own model downloader uses, so both the settings’ “Local models” list and the model picker in the chat composer offered them alongside real language models — selecting one there could not have worked.

The voice wizard filled only half a tablet screen. On a real tablet the setup dialog stayed trapped inside the settings panel, with the category list still showing beside it and an empty strip below the footer. It looked correct on a phone, and on a tablet shrunk to phone size, which is why it survived this long.

The voice picker showed “Select an option” while a voice was selected. When network voices were available they crowded the list, and the local voice the app had actually chosen was not in it. The setting was right; the label was wrong.

Markdown documents are readable again. Files shared into a chat now render formatted instead of as raw text, and their card opens the document.

The voice you hear when you press play is the voice in your settings. The two could disagree; now they cannot.

The selection icon in the chat sidebar is gone. Press and hold already did the same thing.

What this release does not do yet

Read-aloud does not yet stream ahead of the text it is reading with a personal voice — it waits for the whole reply, where the stock voice can start mid- stream. The stock system voice remains the fallback and is unchanged: if you never set up a personal voice, nothing about the app sounds different.

Isolated to the generation step itself running behind real time, not to the playback buffer, batch size, or CPU thread count — all three were measured and ruled out one at a time. A short reply is unaffected. Fixing the generation step itself is bigger work than this release; it is not started yet.

0.1.17

Released Aug 21, 2026 · Release

The local engine can use the GPU.

Stop now stops

Pressing stop during a long prompt used to wait out the whole prefill. It was not slow; it did not work at all. The engine kept going and quit when it would have quit anyway: press at 1,500 ms and stop “took” 155; press at 200 ms and it took 1,458. The sum was 1,655 against 1,658 — three milliseconds apart across two opposite experiments, which is the signature of a stop that never happened.

The cause was a callback llama.cpp’s own header describes as CPU-only. Metal implements it, the OpenCL backend did not, and the generic per-backend lookup was quietly finding nothing.

1,425–1,440 ms → 32 ms, with a p95 of 36 and a worst case of 36 across nine runs. That p95 matches the CPU floor’s own 36 ms exactly: on this axis the GPU is no longer distinguishable from the processor. Stopping mid-answer costs 42 ms at the median, 50 at p95.

Generation pays nothing for it: 16.43 tokens per second before, 16.43 after. Prompt processing pays 6–7%.

The first message stops costing five seconds

Flash Attention was on because nothing had ever chosen it — the setting defaulted to automatic and nobody had measured what it did on this chip. Compiling its kernels cost 4.7 to 6.6 seconds on the first message of every session, and that cost was invisible because it was being discarded as a warm-up.

With it off on OpenCL targets the first message lands in 1.7 seconds instead of 6.2, and generation after a long prompt runs roughly twice as fast. The same sign held on three different model families. Nothing the model says changes: the semantic suite reads identically, 7 of 7.

The GPU is asked to prove itself, once

The engine ships an OpenCL backend now, with Adreno kernels compiled from source. But shipping it is not the same as using it, and a GPU is not automatically faster than the processor on a phone.

So the app runs one short qualification the first time you pick a local model — and it asks first, with a plain yes or no and an option never to ask again. Saying no is remembered and is not the end of it: there is a command in settings to run it later, or to change your mind. Saying no leaves the app exactly as it was.

The result is evidence, not an opinion. A backend is chosen only if it beats the processor’s time-to-first-token by a factor of two — a margin taken from the worst case actually measured, 43.2 seconds against 11.0. Without evidence, or when the device is running hot, the processor wins by default.

Under the hood

Every native library in the release build is aligned to 16 KB pages — 75 of 75, checked with llvm-readobj rather than trusted. The abort fix is carried as a patch file rather than a commit in the vendored engine, so a fresh clone still builds.

What this does not do yet

The qualification runs once and records what it saw. It does not re-run when the phone gets hotter or colder, and it does not yet compare a third backend.

0.1.16

Released Aug 20, 2026 · Release

The disagreement, in the open

When sources contradict each other, the passage for and the passage against now sit side by side on the report itself — no tap, no hunting. Nobody reading a report at 86% has a reason to open that one row out of twelve, and that row is the one that matters.

It could not happen before. The rule that marks a claim contested had tests and no caller: a run could never produce one, so neither the card nor the standing bar could ever show it. The GGUF report said in its own prose “the sources do not formally name a single maintainer” while the bar above it read 7 of 7 supported, 0 contested — the disagreement lived in the summary and never in the data.

Now the judge gets a second question: does this passage, from a different source, contradict the claim? One candidate per claim, chosen by word overlap and never by a model — a model picking what to judge is a second judgement paid for to move the first.

Three things are refused before the judge is paid, each found on the device:

An echo is not a disagreement. Two sites carrying one sentence had the judge saying “yes, it contradicts” about a sentence it had just approved.
A frame is not a difference. The next one had “says no” containing “says yes” plus a paragraph of context. The comparison runs in the shorter direction now.
A table is not a passage. An infobox scraped without spaces between cells — “0x46Developed byGeorgi Gerganov and communityInitial releaseAugust 22, 2023” — was called a contradiction while the text actually confirmed the date.

How a report holds over time

Every re-check has been writing its document to the Library for weeks, in prose: readable and impossible to compare. It now carries a block that reads back exactly, and the report lines the checks up with the drop between them.

Counted on the quoted passages, not on the pages. Reference decay has two halves — the address dies, or the page answers and no longer says what was cited — and among links that still resolve, published studies find only about 30% still contain the cited material. A page that changed elsewhere took nothing away from this report. We can ask that question because we kept the text; everyone else kept a URL.

Fixed

Reports were graded on verdicts the judge never gave. Asked to answer “SI | PARZIALE | NO — reason”, the on-device judge copied the menu instead of choosing from it, and the first word was taken as the verdict. Those 100% reports were inflated. The menu is gone from the question — three words, one per line, with an example — and coverage on the next run went from 25% to 100%, honestly this time.
Six pages counted as ten. Two branches of a plan can land on the same page; the report listed it twice. The model saw it as [1] and [6], so two claims “from different sources” could come from one; and “6 of 10 independent” had an inflated numerator and denominator. Same question on the Pad, before and after: 10 sources / 6 of 10 → 7 sources / 7 of 7.
Two sources, one name. Both rows read “GGUF”, both “same site as another source”. The domain is on the row now, where the mockup had it.
The raw protocol reached the screen: “| PARZIALE |” printed as the reason, under a verdict that said “contested”.
Two bells with the same badge on tablet, and the badge was clipped in landscape. The panel owns the hamburger, the bell and the downloads — not the chat options, which is why they had gone missing from the tablet pane in the first place.
“1 contese”. Italian agrees its adjectives and that line carries five in a row, so any value of 1 broke one. The counts are labelled now.
The list card left the contested count out, so a four-claim report read “75% · 3 supported · 0 in part · 0 contradicted” — the missing one was the one explaining the 75%.
“Deep Research V3” put an internal version number on screen.
A failure that named nothing. Asking for a search, granting the consent, and then reading “Did not work: reading a web page” — with no reason, right after a search that had in fact succeeded. The consent is not the cause: it is the gate, and the tool can only run, and fail, once it opens. Underneath was a page redirecting https → http, refused by the native client. TALOS did the right thing and said nothing about it. Failures now carry a stable code alongside the sentence, and the notice reads the code: “That page redirected to an unprotected connection, so TALOS did not follow it. The other sources it found still stand.” Eleven refusals have a sentence now.
A link in your own message was invisible. Measured on the device: link colour rgb(192, 139, 60), bubble background rgb(192, 139, 60) — identical, contrast 1:1. The address was there, underlined, and unreadable. Every address you type or paste.

Added

Sources export as BibTeX and RIS, for Zotero, Mendeley, EndNote or a LaTeX bibliography. Nothing about the research leaves in them — not the question, not the judge, not the identifiers: a bibliography file ends up in a shared folder, which is the least controlled place a personal detail can land.
The header line carries sources and tokens, and tokens only when an engine actually reported some.

Also, from before the tag

The report says what its own percentage is worth. The four benchmark measures now sit under the standing figure, each with the plain line that says what it means, and the score carries its date because cited pages die.
Contested claims reach the surface: listed in the standing line, counted as zero toward solidity like a contradiction, and a single one marks a run as worth a second look.
The accessibility-service description is readable in any language. It existed only in Italian, in the fallback file, so a phone set to German or Spanish showed an Italian sentence on the one screen that grants reading the whole display. English is the fallback now, Italian sits in its own file, and the Android lint gate — red since 19 August, including on the v0.1.14 push — is green again.

0.1.15

Released Aug 20, 2026 · Release

The honesty release: three things TALOS used to say with confidence that were not true, plus the groundwork for the deep-research rework.

Fixed

Location is precise, and when it cannot be, TALOS says so. The phone knew better — its own providers had the right position to 43 m — but TALOS held only approximate location permission and settled for it silently, producing output identical to a GPS fix. It now asks for precise location, and when that stays denied it still reads the approximate one but declares it and names the switch to turn on. Measured: ~2 km with the warning, 16 m without.
Local models call their tools instead of inventing the answer. Asked for the phone’s coordinates, Gemma 3 4B produced Milan, then “I am in a chat with a user”, then narrated “I am reading the phone’s location” without reading anything — three phrasings, no tool call. Four causes, each measured: the tool-lookup result looked like an answer, a correct direct call was discarded for skipping the lookup step, the protocol sat too far from where the model reads last, and an emptied code fence stayed in the reply. Gemma now calls on all three phrasings and returns the true coordinates on two; Qwen3-1.7B calls and answers correctly.
The answer comes back in the language of the interface. A local model handed an English tool result answered in English to an Italian question. The reminder now also follows the tool data, where the model reads last, instead of living only at the top of the prompt.
PDFs attach again. Attaching one failed every time with “TALOS could not examine this file”. pdf.js needs its worker and ours was never configured; in Node it silently fell back, so the tests said nothing about the phone.
A PDF opens inside TALOS instead of Android’s share sheet. It now opens in the built-in viewer, and the whole page fits in either orientation instead of running off the bottom.
The pending-authorization notice is said once. It appeared up to three times in one reply and outlived the request it described, because it was written into the message text as well as shown on its own card. Only the card carries it now, and it moves from pending to settled on its own.

Added

Independent sources are counted, not URLs. Three sites carrying the same press release are one piece of evidence, not three. Computed from the registrable domain and the origins a source itself cites, with no extra network request — and a source that cites several origins stays independent, because penalising that would punish the best sources.
Fidelity measures are computed, each with its date — coverage, citation faithfulness, claim groundedness and distinct proofs. A run nobody judged returns “unverified” rather than a percentage, because a percentage reads as a measurement and would be a measurement of nothing. They reach the report screen in the next version.
Citations export as BibTeX and RIS, for Zotero, Mendeley and EndNote. A citation describes a page, so the query, the judging model and the run id never leave with it.
A contested claim is now its own outcome, separate from partial. One source saying yes and another saying no is not “nearly supported”: it is the world disagreeing, and both passages are kept so they can be shown side by side.

0.1.14

Released Aug 19, 2026 · Release

Fixed

Qwen3-1.7B and Gemma 3 local chats now keep the same direct-turn contract as API providers. Exact replies, explicit no-tool requests and contextual last-word follow-ups no longer trigger a spurious tool call, leak an internal tool envelope or add a model-generated lead-in.
Local tool transport is selected from the GGUF template capabilities. Native templates with tool-call support use the embedded contract; templates without it use the AVM-owned prompt JSON protocol, including system-role compatibility and untrusted tool-result boundaries.
OpenRouter image generation remains fully executable. The progressive catalog gate is restricted to local models, so API providers still run generate_image, persist the generated binary and report storage failures correctly.

Changed

Doctor and parity diagnostics now expose the measured local template capabilities and selected tool transport, making GGUF compatibility inspectable instead of inferred from a model name.
Local parity probes cover native and prompt-JSON paths, with persistent fingerprints, failure stages and regression fixtures for reloads, tool results and system-role handling.
The mobile harness keeps the initial JavaScript chunk within its budget by loading non-bootstrap tool-result sanitization and direct-turn policy only when a chat send is already in progress.

Verification

Full Vitest suite: 5643 passed, 10 skipped.
Typecheck, web build, initial-chunk and parity gates: passed.
OK (9 tests) with only fixture-dependent assumptions skipped.
Real Gemma and Qwen prompts, reload persistence, portrait layout and final logcat were exercised on the reference tablet.

0.1.13

Released Aug 19, 2026 · Release

Fixed

Waking TALOS from a locked device no longer puts the assistant above the lock screen. Android keeps its normal authentication surface in front; TALOS stays silent and hidden until the keyguard has really disappeared. Immediately after unlock, the assistant is handed the microphone and enters listening mode. The cold and warm paths were exercised on the real Pad, including the short visual transition after authentication.
Barge-in now stops TALOS while it is speaking and starts listening again through the wake-word path, without opening a second recognition session.
Voice permission and accessibility state are represented by their real Android state. Non-actionable permissions are not presented as toggles in onboarding, while actionable settings remain available in Settings.
Local-model and permission flows no longer leak internal tool labels or crash while reading native permission metadata. The affected native and web paths have focused regression coverage and were rebuilt into the final APK.

0.1.12

Released Aug 18, 2026 · Release

Fixed

Granting a runtime permission no longer crashes the release build. R8 was removing the Capacitor permission metadata that TALOS reads when it asks for microphone, notifications, contacts, location, camera, calendar or mail access. The release now retains that contract, and the real Android artifact gate checks the aliases and platform permission strings before publication. The grant, denial, permanent-denial and system-settings paths were exercised on a OnePlus Pad running Android 16 without an application FATAL.

0.1.11

Released Aug 18, 2026 · Release

<!– v0.1.9 and v0.1.10 were tagged but never produced a signed APK: their builds failed (packaging, then a stale test delimiter). This is the release that ships the work, under a clean version. –>

<!– v0.1.9 was tagged but never produced a signed APK: its first build failed in packaging. This is the release that ships that work, under a clean version. –>

Fixed

Opening Diagnostics crashed the app. Every time. The diagnostics screen checks whether dictation has microphone access, and that check read a permission whose state the phone answered with “nothing” — the same class of bug as the OnePlus 13 startup crash in the last release, but this time inside the third-party dictation component, on a thread no JavaScript safety net can reach.
- Fixed at the native layer: TALOS now wraps that component with a safe version that reports unknown instead of falling over. Verified on a real device — Diagnostics opens, shows its report, and stays up.

Changed

The local model is no longer thrown away the moment you switch apps. Before, leaving TALOS for a second discarded the model even with memory to spare, and coming back paid the full cold-open cost — measured at four seconds warm, nearly nine cold, all of it in the loading. TALOS now keeps the model for a short grace period and only releases it under real memory pressure, so a quick glance at another app doesn’t cost you the wait when you return.
Long replies from the local model stream more smoothly. Watching a reply arrive used to recopy the whole answer from scratch on every refresh, so the longer the reply, the more work each update cost — growing with the square of the length. TALOS now passes only the newly written text, keeping long answers as light as short ones.

0.1.8

Released Aug 18, 2026 · Release

A crash-fix release.

Fixed

TALOS crashed on launch on the OnePlus 13. The app died before it could draw anything, every time. The cause was a permission-state read that Android is allowed to answer with “nothing” — and on this device it does. TALOS read that “nothing” as a value and fell over.
- It no longer does. A permission whose state cannot be read is now reported as unknown rather than crashing — because guessing “denied” would be inventing a fact, showing an “Allow” button for something perhaps already granted, or hiding a feature that is actually there.
- Verified on the OnePlus 13 itself: zero crashes, launch in 1.3 seconds, the screen drawn. Not deduced from an exit code — looked at.
The local model wasted its own warm-up. TALOS remembers the best thread setup for each phone and model, so it doesn’t re-measure every time. But it wasn’t noticing when the engine underneath changed — this release ships a new build of the local engine, and a setup measured with the old one would have been reused with the new. The memory now knows which engine it was measured with.

Changed

“Always allow” is back on web search. It had quietly disappeared: after a search, the safety chain would rise to its highest level and drop the “always” option — on the one action people use ten times a day. Searching the web is now an explicit exception, so the choice you make can hold for next time. The question still appears the first time, and still says why.
The local engine runs on the right cores. On Android, the code that pins work to specific CPU cores was being compiled out entirely — every thread ran everywhere. This release picks up the upstream fix.

Internal

Android, the native C++ engine and the browser tests now run on every proposed change. A whole class of defect — including a native crash and this release’s OnePlus 13 crash — lived in the layer the old checks never touched.

0.1.7

Released Aug 18, 2026 · Release

An external engineering review went through the whole codebase and found twenty-two issues. This release closes thirteen of them, plus one the review had missed. Every fix below was reproduced with a failing test before it was written, and the build was installed and used on a real tablet before this went out.

Fixed

A local model could have sent your conversation to a stranger. TALOS lets you point it at a model running on your own machine, over plain HTTP. It checked that address carefully — it even refuses host names, because a name is resolved later and by someone else. But it only checked the first hop. A server answering 302 Location: https://somewhere-else would have carried your prompt, and the authorization header with it, straight out of the network you had allowed.
- Redirects are now refused outright, on both the streaming path and the native bridge, and the boundary cannot be switched off by the code that calls it.
- The test for this stands up two real servers and checks whether the body reaches the second one. Before the fix, it did.
“I could not check” was being reported as “done”. When an action succeeded but the check that would confirm it failed, TALOS kept the success and dropped the doubt. For anything that changes the world — sending a message, writing a file — that is the wrong half to keep.
- There are now four outcomes instead of two, and the new one refuses to become either neighbour. Calling it a failure is worse than calling it a success: a failure is an instruction to try again, and trying again is how the same message gets sent twice.
The database migration wrote a full copy of your data in the clear. When an older installation moves onto a managed key, TALOS exports everything, writes it to disk, destroys the original and imports it back. That intermediate file was readable — for a few seconds, data that spends the rest of its life behind a PIN was plain text on the filesystem.
- It is now encrypted, and it is written to a temporary name, read back and compared, and only then renamed. The old file is destroyed after that, never before.
- Journals written by an older version are still readable. If your app was killed mid-migration and you update before reopening it, the new code finds the old file and resumes from it. Without that, an update meant to protect your data would have destroyed it.
Your PIN now protects the key with Argon2id. The previous protection was PBKDF2 at 210,000 iterations, which is no longer what current guidance recommends. Existing PINs keep working and are upgraded silently the next time you unlock — and if that upgrade fails for any reason, you still get in.
- The record now states which protection it uses, and that statement is authenticated: a record altered to claim weaker protection simply will not open.
A malformed backup file could freeze the app. Backup files declare the cost of the protection they were written with, and TALOS obeyed that declaration without limits. A file claiming four gigabytes of memory would have asked a phone for four gigabytes. The test that found this ran for five minutes before being killed; it now finishes in milliseconds.
- A file says what it was written with. It does not decide how much work your device is willing to do.
Any app on your phone could start TALOS’s internal profiler. A diagnostic tool, meant for hunting a startup problem, could be triggered from outside with a single command. It is now absent from release builds entirely — not disabled, absent, verified by reading the compiled bytecode.
A crash in the local engine, under the right timing. Preparing a prompt and generating a reply could touch the same piece of memory from two different threads, one freeing it while the other read it. Everything that touches the engine now runs in a single ordered sequence, and anything that breaks that order fails immediately with a message naming both threads instead of crashing silently.
- Calling the engine after closing it now reports a clear error too. Before, it handed the closed engine’s address back to the native layer.

New

A security policy for the app’s own web layer. TALOS’s interface runs in a web view that can reach the phone’s native features, so an injected script there would not steal a session — it would use the device. The app now declares what it is allowed to load and execute, and an injected script does not run.
- The policy is checked by loading the real app in a real browser and asking it whether it had to block anything. That test caught a first version of the policy that would have prevented everyone from unlocking their database.

Changed

The app is 6.4 MB smaller — 42.6 MB down to 36.2 MB, measured on the same build twice. Unused code and resources are now removed at build time.

Internal

Android, the C++ engine and the browser tests now run on every proposed change. They did not before, and that is how the engine crash reached this codebase.
- The browser suite turned out to be completely broken — all fifty-two tests — because a language screen loads late and covers the button the tests press. Nobody knew, because nothing ran them.
Build actions are pinned to exact commits rather than moving tags, the build toolchain verifies its own download against a published checksum, and automatic updates keep those pins current.

0.1.6

Released Aug 17, 2026 · Release

Fixed

“Hey TALOS” not answering on a tablet. Ten attempts in a real room, at normal speaking volume, produced zero activations — the highest score was 0.383 against a 0.50 threshold. The wake word now answers 10 out of 10 on the same device, in the same room, with the same voice.
- The cause was not the model, the microphone, the room noise or the volume. Each of those was ruled out by measurement: attenuating the audio by 18 dB still scored 0.951, and at the room’s real 16 dB signal-to-noise ratio the model still hit 0.967. It was the build that was installed.
- The threshold stays at 0.50, and that is a decision made from data rather than instinct. The ten real detections land between 0.544 and 0.949; the loudest thing in the room that is not the wake word reaches 0.319. Raising the threshold to 0.60 would lose two detections in ten and prevent nothing, because there is nothing to prevent.

New

TALOS no longer asks you to approve something that cannot work. Before a permission card appears, TALOS now checks whether what the action assumes actually exists — a contact, a file, an app. If it does not, nothing is asked and nothing runs, and TALOS tells you what is missing instead.
- It answers in three states, not two, and the third is the one that matters: “I could not tell”. A denied contacts permission means TALOS does not know whether your contact exists — so it proceeds rather than claiming the contact is not there. Treating “I don’t know” as “no” is the same mistake seen from the other side.

Internal

Two diagnostic switches for the wake word, off by default and turned on by creating a file: one writes the raw audio the model receives, the other swaps the microphone source. From now on “it doesn’t work on my phone” is settled by pulling the bytes and replaying them, rather than by guessing.

0.1.5

Released Aug 17, 2026 · Release

New

PDFs open inside TALOS. A document TALOS generates now shows its pages in the app — swipe through them, close, and the conversation is still where you left it. Until now the card showed the name and the size and tapping it did nothing.
- It renders through Android’s own PDF engine, so it adds no library to the app: 31 bytes to the startup path, against the ~16 MB of native code a PDF library would have brought.

Fixed

“Done.” said out loud to a task TALOS had just given up on. When TALOS drives the screen for you it ends the run by saying what happened — and it said “Done” whether it had finished or abandoned, because those were the same outcome internally. They are now two, and there is a third: if TALOS does not say which, you are told that too, rather than being told it worked.
A pending permission request swallowed the answer. When TALOS needed your approval mid-task, the notice was glued onto the reply and the whole thing was drawn inside the notice’s box — so the answer lost its formatting and the box filled the screen. The answer is now an answer, and the notice is a small chip under it.
“I have no web search” — said while the capability existed. With no search provider key set, the two web tools are not built at all, so the model could not see them and concluded it had none. TALOS now says the search key is missing and where to add it, instead of leaving you to conclude it cannot search.

Known limits

Unchanged from v0.1.4: verified on one device, a OnePlus Pad 3 running Android 16 with ColorOS.

0.1.4

Released Aug 17, 2026 · Release

Everything below is one story: TALOS presses “send” for you, and this release is about it never lying to you about whether it did.

Fixed

TALOS said “the message was sent ✓” and, in the next line, “it was not sent”. Both in the same answer, about a message still sitting in the WhatsApp input box. Four written rules told it not to, and none of them worked — because the sentence was not written after the send. It was written before the call, as a preamble, and the app glued the preamble and the conclusion together. A false opening now disappears when the tool reports it changed nothing. A true one stays: if the message really went, the line you already read stays where it was.
“The accessibility settings are already open” — while WhatsApp was on screen. Measured from the phone’s own activity log: TALOS opened them 37 milliseconds after opening WhatsApp, and WhatsApp went on launching four more windows over the next 850 milliseconds and buried them. The sentence was true for 37 thousandths of a second. TALOS no longer races: the card carries a button, and the screen changes when you tap it.
That button used to land on a page where TALOS was not listed. On OPPO/OnePlus phones the downloaded services live one level down, under “Downloaded apps”, so “find TALOS in the list” pointed at a list without TALOS. It now opens the page TALOS is actually on, and falls back to the general list everywhere else.
Sending a file had the same holes, and one more. It said “sent” on a single check where messages had required three since v0.1.1; it drew no card at all; and it told the model to open a settings page, which twice opened notification access instead — a permission that reads every notification from every app, offered for sending one message.
“That file is not in your Library” — about a file that was there, twice over. When TALOS could not read the Library it reported an empty one. Not being able to look and there being nothing to find are two different answers, and only one of them was ever true.
Two files with the same name were a dead end. TALOS correctly asked which one you meant, and then could not accept any answer you gave: it matched files by name, and two identical names stay identical. Answering “the first one” led back to the same question.
“Send file X to Y” often never called the send tool at all. TALOS searched your Library, described what was in the files, and stopped — and once claimed, with no basis, that it could only send files stored on the phone. Measured against the real set of tools: the send tool was chosen 8 times out of 12, and for “write to X on WhatsApp attaching Y” it was chosen 0 times out of 3. It is now 12 out of 12.

Changed, for new installs only

These are the values a phone gets when it installs TALOS for the first time. Nothing moves on a phone that already has the app: a default is what you get before you choose, and changing it is not permission to change your mind for you.

Interface text now starts at Default instead of Large. Every larger step is still one tap away in Settings, and the phone’s own text-size setting keeps applying underneath.
The composer starts as the compact bar.
The app icon follows the theme you have on, without being asked.
Choosing “Always” in the one-tap autonomy screen now also turns on “Let chats use the Library”. Saying yes to everything in one gesture and then finding your own Library disconnected reads as a fault, not a choice.

New

A card under the answer says whether it left. “WhatsApp — NOT sent · screen reading is off”, in words and not only in colour, with the reason when we know it. It is drawn by the app from what the phone reported, so it stands even when the sentence next to it says something else.
When two files share a name, you pick one with your finger. The list comes from your Library to the screen and each row is a button — no id to retype, no number to guess.

Known limits

Verified on one device: a OnePlus Pad 3 running Android 16, ColorOS. The settings page that contains TALOS is found by asking the phone, with a fallback everywhere else — but nobody has seen that fallback on a non-OPPO phone yet.
Unchanged from v0.1.1: opening an app that is already running resumes it where you left it, so a screen task can land inside the wrong screen. TALOS says it could not get there rather than claiming it did.

0.1.2

Released Aug 16, 2026 · Release

New

You can make TALOS your assistant without plugging anything in. Until now, if Android did not show the “become the assistant” dialog, the only way left was the ADB bridge — a cable or wireless debugging. And Android never shows that dialog: the assistant role is declared non-requestable in AOSP itself, so the request opens and closes in 53 milliseconds with nothing on screen. Measured on device. TALOS now takes you straight to the system page where you pick your assistant, with TALOS in the list. The bridge is still there, but last.
The button that used to promise “one tap and Android asks you” now says what actually happens. A prompt nobody will ever see is worse than no promise.

Fixed

Two of the ten permission rows never said whether they were granted. “Where you are” and “Files you choose” showed a dashed circle and no words, while the other eight said “Allowed” — so the absence of a word was doing the work of a word, on a page whose only job is answering “do you have it or not?”. Anyone using a screen reader had no way to tell at all.
- Location is a real permission and the app simply never asked the system about it. It does now — and it checks approximate location only, because Android 12 lets you grant approximate without precise, and checking both would have reported “not granted” to someone who had just granted it.
- Files has no permission behind it — you grant access by picking the file — so the row now says so instead of looking like something you forgot.

Known limits

Unchanged from v0.1.1: opening an app that is already running resumes it where you left it, so a screen task can land inside the wrong screen. TALOS says it could not get there rather than claiming it did.

0.1.1

Released Aug 16, 2026 · Release

New

The screen is read in the order you see it. Numbering used to follow the accessibility tree, which is the order Android built the screen in, not the order it is drawn. On three real screens, only 0 of 19, 1 of 18 and 2 of 32 indices matched what was actually on the glass. “Tap the first item” now means the first item from the top.
Controls with no text of their own get a name. An icon-only button used to be anonymous, so there was no way to refer to it. It now borrows the first text found inside it, with a bounded search. The sticker button in a keyboard is addressable; before, it was a blank.
Three more ways to act: long press, set a slider to a value, and open the recents screen — six actions in total, up from three.
Sliders are recognised by the range they expose, not by their class name. This is why the previous build could describe a volume slider and not move it: the control reports itself as neither clickable nor scrollable, so every name-based check missed it.
Setting a value is checked by reading it back. If the value did not move, TALOS says so — and distinguishes a control that refused, one that moved somewhere else, and one that moved partway.
First run says more and reads shorter. The opening story covers eight things instead of five, including the three that were true in the README but missing from the first screen anyone sees: that TALOS acts inside other apps, that “done” is a verified state, and that it hears you with the screen off. Each is a line you open if you want it, so the page fits one screen instead of being a wall to scroll.

Fixed

“Wi-Fi” and “wifi” did not match. The hyphen became a space, so the most common word on an Android settings screen could not be found by the way people actually type it.
Automatic saving of generated documents was off, and nothing could turn it on. A migration from a security review switched it off on every install — new ones included — while the default next to it still said on. A document the model made and you did not save scrolled away with the chat. It is on again, and turning it off on purpose now sticks.

Known limits

Opening an app that is already running resumes it where you left it, exactly as tapping its icon does. So “open Settings and tap the first item” can land inside whichever Settings screen was open last. TALOS notices and says it could not get there, rather than claiming it did — but it does not yet find its own way back to an app’s starting screen. Forcing the app to restart would fix the navigation and throw away your half-written message with it, so it doesn’t.

0.1.0

Released Aug 16, 2026 · Release

First public build. Signed APK, arm64-v8a, Android 8.0 or later, with build provenance you can verify from the release page.

0.1.40

Providers

0.1.39

On-device models

Chats

Codice

Permissions and safety

0.1.38

Chats

Permissions

Codice

Everywhere

0.1.37

0.1.36

0.1.35

0.1.34

Messages

Drafts

Stations

The “+” sheet

Fixed

0.1.33

The composer

0.1.32

The composer

The “+” sheet

0.1.31

The composer

While TALOS is thinking

0.1.30

The chat, redesigned

One search for everything

Sidebar and chats list

Settings on the calm design

Small things that now move, and a few that now work

Your recorded voice reads replies again

Fixed

0.1.28

One sidebar, everywhere

Notes, Tasks, Memory: complete, not read-only

Library: real previews, and adding a file no longer sends it anywhere

Research: a report is only “done” when it holds up

Codice tells you when the phone bridge is off

Errors in plain words

Plain words in the model picker and error cards

Fixed

0.1.27

One fewer place to set reasoning effort

The “+” menu no longer offers web browsing

0.1.26

Codice’s Terminal tab can run a command, not just show one

Codice follows the app’s language

DuckDuckGo: a fifth web search option that needs no key

Fixed

0.1.25

Codice ships in the signed build

Fixed

0.1.24

Codice never shows on a build that can’t run it

Fixed

0.1.23

Codice: a real coding agent, on the phone

Personal voice

Model Lab

Known, and not hidden

0.1.22

Browsing models on Hugging Face

Fixed

0.1.19

The voice engine was rebuilt on Pocket TTS

The local model runs faster

Fixed

0.1.18

Teaching it your voice

It no longer dies while learning

If it crashes, you do not start over

A personal voice can now actually be chosen — and it works

Fixes

What this release does not do yet

0.1.17