Release Notes

Improvements

• Common: A rate limit on the model you chose is reported, not worked around: When a model you selected is rate-limited, overloaded, or unavailable, Caiioo now tells you which model and what happened, instead of silently switching you to a free one. Automatic switching still rescues you when Caiioo picked the model for you.
• Common: A shared page shows that it is loading: A published conversation now shows a loading indicator until the real interface arrives, and says so plainly if it cannot load. Previously it rendered a simplified stand-in that could stay on screen indefinitely, indistinguishable from the finished page.
• Common: Fewer duplicate entries in the skills menu: Skills that exist twice under different internal names are now listed once. Everything stays available, and any skill you have used keeps working.

Bug Fixes

• Common: Models no longer degenerate into unreadable output: A model switched to mid-conversation was given the previous model's reasoning data. Because that data is specific to the model that produced it, the replacement could deteriorate into unintelligible text — including inside tool calls, which left it unable to recover. Safeguards existed for two model families and exempted every other, which is why this appeared across many models at once.
• Common: A revoked account connection asks you to reconnect: When a provider permanently rejects a saved sign-in, Caiioo now records that and prompts you to reconnect. It previously retried the dead credential every few seconds indefinitely, without ever saying anything was wrong.
• Common: A provider rate limit no longer blames your API key: A rate limit or an overloaded provider reported itself as a problem with your Caiioo-managed key and told you to enter your own. It now names the model and the actual cause.
• Common: Published pages no longer carry your conversation's internal id: Shared pages included the identifier of the conversation on your own machine. It served no purpose on the page and has been removed.
• Common: Diagrams in documents finish rendering: A diagram written into a document while its renderer was still loading could be left showing a parse error for text that had since been corrected. Diagram labels containing brackets or punctuation are also generated correctly now.
• iOS: Background sync no longer risks an interrupted turn: Background work that talks to the app's web view now does so correctly, and a background task can no longer be completed twice.

New Features

• Common: Published conversations look like Caiioo: A shared page now renders in the real Caiioo interface — the same message layout, tool chain, model badge and formatting you see in the app. Link previews and browsers with JavaScript turned off still get a clean, readable page.
• Common: Choose how a conversation is published: Publishing opens a dialog where you set the public title, choose how long the page stays up — including until you delete it — and decide whether search engines may list it. Search listing is off by default, so a link you send to three people does not become a search result.
• Common: One public page per conversation: Publishing the same conversation again updates the page you already shared, so a link you have sent people keeps working. The share menu shows when a conversation is published and offers to take it down.
• Common: Shared pages: A new Settings screen lists everything you have published — what it is, when it comes down — and takes any of it down on the spot.
• Desktop: The assistant can see and open the document you are working on: The desktop apps now tell the assistant what is on screen, let it open a slate you have open, and find slates from earlier conversations by name or search term.

Improvements

• Common: Tool steps report their true outcome: A tool that reports a failure inside its own response is now shown as failed. Some steps that previously appeared to succeed will now appear as failed, in past conversations as well as new ones.
• Common: Editing a document section requires replacement text: Editing a section or line range without supplying replacement text is now refused rather than accepted. Clearing a section by omitting the text is no longer supported — use search and replace to remove content.
• Common: Long agent runs stop and explain themselves: When every tool call has failed for twenty iterations running, the run now ends with an explanation instead of continuing to the step limit.
• Desktop: Your sign-in renews before it lapses: The desktop apps refresh an ageing sign-in ahead of time, so the private relay stays connected.

Bug Fixes

• Common: An expired session asks you to sign in: When a session has fully expired the app now prompts you to sign in again, instead of leaving your devices greyed out and unreachable.
• Common: Background model selection stays with its provider: A helper model chosen under one provider is no longer sent to a different one, which produced "not a valid model ID" errors out of nowhere.
• Desktop: Sign in with ChatGPT completes: A repeating sign-in and model-refresh cycle no longer exhausts the shared rate limit.
• Desktop: Thread memory is kept: Archiving and searching within a conversation now persist, so "search this thread" returns results instead of coming back empty.

New Features

• Common: Publish a conversation or slate to the web: Turn any conversation or slate artifact into a shareable link. The published page shows your prompts, the sequence of tools the assistant used, and its final answers — but strips out the raw tool inputs, search queries, and results, so nothing it retrieved behind the scenes is exposed. Choose how long the link stays live, and revoke it at any time.
• Common: Choose which calendars send you reminders: In Agenda's show/hide calendars, pick which calendars are allowed to send reminders — including calendars shared with you — so a spouse's or family calendar can notify you while others stay silent. By default only calendars you own or deliberately add send reminders.

Improvements

• Desktop: Your computers can run each other's conversations: A conversation you send from one desktop to another now runs on the receiving computer when both are online and synced, completing the same cross-device hand-off already available to your phone and browser. Previously a second desktop could show as "not reachable" even while fully connected.

Bug Fixes

• Common: Only real calendar events become reminders: Free/busy blocks and untitled entries on calendars shared into yours no longer pile up as empty "(no title)" reminders.
• Common: Sign in with ChatGPT completes: Signing in with ChatGPT could get stuck repeating while the app refreshed its list of available models; it now finishes normally on every platform.

New Features

• Common: Manage your provider credentials in one place: Rename any connected account so "Work" and "Personal" are easy to tell apart, and when you have more than one credential saved for a provider, choose which one is active from the AI Providers screen without opening the credentials vault.

Improvements

• Common: Syncing between your devices is more dependable: Conversations, scheduled tasks, and your settings converge more reliably across your computers and phones, so a change made on one device reaches the others sooner.
• Common: A brief network hiccup at startup keeps you signed in: A momentary connection drop while the app is starting no longer returns you to the sign-in screen.

Bug Fixes

• Common: Past calendar events stop resurfacing as reminders: Calendar events that have already ended no longer accumulate as reminders or notify you about meetings that are already over.
• Common: Disconnecting an account stays disconnected everywhere: Removing a connected account now sticks across all your devices instead of returning on the next sync.
• Common: Conversations sync between your desktop computers: Two desktops signed in to the same Private Sync now exchange their conversations reliably.
• Common: Your data-retention setting syncs across devices: Changing how long conversations are kept on one device now carries to your others, and old attachments removed by retention stay removed everywhere.

New Features

• Android: Sign in with Apple: The Apple sign-in option now works on Android, so an account created with Apple on iPhone or Mac signs in everywhere. (Android previously hid the option.)
• Common: Private Sync is included in the Free plan: Cross-device sync — and everything built on it, including the Private Constellation — no longer requires a paid plan.
• Common: The app tells you when the constellation needs your passphrase: A device without your Private Sync passphrase used to show your other devices as selectable and then fail with "device not available". The device picker and settings now say exactly what is missing and where to enter it.
• Common: Tracked changes live inside your Word document: Edits and comments are stored in the .docx itself rather than alongside it, so they open as real tracked changes in Word, come back intact after a download and re-upload, and stay separate when a second author adds their own. Documents carrying the older side-by-side form are converted when you open them, wherever the conversion can be made exactly.
• Common: Your settings follow the conversation between devices: The model, mode and turn settings you picked apply wherever the conversation actually runs, including on a device that did not start it.
• Common: ElevenLabs adds a voice for real-time dialogue: Choose v3 Conversational in Voice settings for expressive, low-latency speech across more than 70 languages, with cost estimates that reflect its current half-rate pricing.

Improvements

• Common: A conversation flows both ways between your browser and your desktop: The desktop can now send turns to a conversation the browser extension hosts, prompts you type on either side appear on the other immediately, responses stream live from the first turn, and Stop works from whichever device you press it on.
• Common: Attached files travel the short way: A document attached on one device now reaches the device running the conversation directly over your own network, instead of timing out against the cloud when both devices are at home.
• Common: Web search is one tool with three ways to look: Ranked web results (Keenable) join the AI-synthesized answer and direct Google search inside the same search tool, so the assistant picks one approach per question instead of running two searches side by side.
• Desktop: The remote access toggle applies immediately: Turning LAN access on or off no longer asks you to restart the app.
• Common: Watch a conversation another device is running: A conversation you start on your phone now streams to your other devices while it happens, instead of appearing once it has finished.
• Extension: The browser extension joins your other devices: It can host a conversation you point at it and follow one another device is running, under the same identity as the rest of your devices.
• Common: Synced edits combine automatically: When two devices contribute to the same conversation, the sync indicator explains that their edits are being combined and clears as soon as the unified conversation is ready.

Security

• Common: A device runs only the conversations it is responsible for: Work sent from one of your devices to another is checked against what the receiving device actually hosts, and the settings carried with it are limited to those a conversation is allowed to change.
• Common: Attachments upload correctly whatever they are named: File names are escaped before an upload is assembled, so an unusual name cannot affect the structure of the request carrying it.

Bug Fixes

• Android: The top of the screen belongs to the app again: The header no longer renders under the clock and camera cutout, everything up there is tappable, and full-screen dialogs keep their close buttons out from under the status bar.
• Common: Ranked web search works: The included Keenable search failed for everyone; the service side is now deployed, configured, and self-repairing.
• Common: Provider keys save and stay saved: An API key added in settings (notably Unsloth) could fail to persist, while the settings screen refreshed itself in a loop.
• Common: Two devices editing one conversation no longer fight: Working on the same conversation from two devices could fork it — a "2/2" branch that could not be switched — and in the worst case the answer vanished from every device once sync settled.
• Common: Conversations containing a diagram or a chart open normally: Opening a conversation with a diagram or a chart in it could replace the whole app with an error screen, and restarting did not clear it — the app reopened the same document.
• Common: Diagram labels appear: Flowcharts and similar diagrams drew their boxes and connecting lines but left the text out of them.
• Common: The agent keeps working on a document you have open: Once a document had changed since the agent last read it, every further edit it attempted was declined and it could not recover — so it would report, repeatedly, that it was unable to save. It now re-reads the current document and works from that, and says clearly what happened when an edit cannot be combined with yours.
• Common: Your sign-in survives a hotel or office network: A captive portal or a corporate proxy answering on the network's behalf was treated as the service rejecting your sign-in, which ended the session — on hotel wifi, within about thirty seconds. Only a genuine rejection ends a session now.
• Common: Downloaded documents keep their accents, characters and emoji: A document containing anything outside plain English downloaded as unreadable characters.
• Extension: Editing a document in the browser keeps it a document: Saving a change to a slate could overwrite it with plain text while it still presented as a document — on Chrome, Edge and Safari this was the only behaviour.
• Common: The agent cannot damage a Word file it is editing: The protection that stops the agent writing text into a binary document was inactive for .docx files reached through the workspace, and did not cover two of the ways it writes at all.
• Common: A document keeps its formatting wherever it came from: Attachments arriving from different places were stored differently, so the same document could round-trip better or worse depending on its origin.
• Common: A Word document the assistant creates opens with its formatting: A document written from the assistant's own text opened in the editor as one unbroken block with the formatting marks showing as literal characters, though it downloaded correctly.
• Common: A downloaded document reads as the finished text: Exporting a document with tracked changes to Markdown or RTF used to include the struck-out text as if it were part of the document.
• Common: Tracked Word edits keep their layout and formatting: Downloading a document with tracked edits could shift tabs and numbered clauses, lose bold formatting or literal emphasis marks, and alter labels such as 1.10. The exported document now preserves those details while keeping the edits native to Word.
• Common: Word documents appear in Slate and open correctly from File Manager: Uploaded and AI-edited Word documents were hidden from Slate's document picker, and opening one from File Manager on a phone could use a browser-style window instead of the app's editor.
• Desktop: Copies of Word documents include their current tracked changes: Copying an edited Word document from the workspace could write the untouched upload to the destination folder. The copy now reconstructs the current document and keeps unedited imports byte-for-byte intact.
• Common: Text you are reading stays put while a response finishes: A live answer or an expanded tool result could disappear or jump when streaming reached its final state, even though the content returned a moment later. The same reading surface now stays in place through completion.
• Android: Android rejoins your device list: Android could not open the secure connection your devices use to reach one another, so it was missing from the device list everywhere else.
• Common: Choosing a file or a photo keeps your conversation running: On Android, opening the file or photo picker was read as putting the phone away, which handed the running conversation to another device mid-answer; and on iOS, saving a document now allows time to choose a name and a destination instead of giving up after fifteen seconds.
• Common: A computer registers with your other devices and stays registered: Registering was attempted once at start-up, before sign-in had finished, and sessions saved by older versions could outlive their real sign-in pass without renewing it. Registration now waits for sign-in and every session uses the pass's actual expiry date.
• Common: A finished answer reaches your other devices immediately: On computers, a completed answer waited for the next scheduled check — which stretches to minutes when little is changing — before travelling.
• Common: A computer that is behind asks the device running the conversation: Previously it could only wait for cloud storage, and once its records looked current it stopped trying altogether.
• Common: A merged conversation reads the same on every device: When two devices had both added to a conversation, the combined result could be ordered differently on each, so the same messages read as two different conversations.
• Common: Changes travel between devices without waiting for a nudge: A turn finished on your phone could sit there until you backgrounded the app, and a sync cycle that had to wait its turn could discard the change that started it. Syncing also spends far less time re-checking things that have not moved, so it stops reporting itself as busy when it is not.

Bug Fixes

• Common: A locked phone still sends your conversation: Changes made just before locking could stay on the phone indefinitely while it reported everything as synced. The app now finishes sending during the moment the system gives it at lock, and retries anything left over when you next open it.
• Common: Devices say what they are holding: Your devices now tell each other which conversations they have newer versions of — identifiers only, never content — so a device that is behind shows it immediately rather than appearing up to date.
• Common: Changes made during a sync are no longer missed: A change arriving from another device while a sync was already running could be skipped and never picked up again.

Bug Fixes

• Common: Device connections renew their own credentials: The connection between your devices relied on a sign-in pass with a 7-day life, renewed only when an app happened to restart. A device could show its companions as offline — and remote conversations could be unreachable — until the next relaunch. Every cross-device feature now renews the pass itself at the moment of use, on iPhone, Android, Mac, Windows, Linux, and the browser extension.

New Features

• Common: Local folder access for conversations: Grant a conversation access to folders on your computer from the composer, manage the granted folders in Settings, and the grant persists across turns and restarts.
• Common: Interject and stop from any device: While a conversation runs on another of your devices, your interjection or Stop reaches the device doing the work — the same controls as a local run.
• Common: Reopening a device catches up instantly: A phone or laptop that was closed while a conversation ran elsewhere now asks the host for the conversation's current state the moment it reopens, instead of waiting for a storage round-trip.
• macOS: Computer Use is discoverable: Asking for desktop control or a desktop screenshot now points directly at the Computer Use package in the Community Hub (tester tier), instead of reporting that no such tool exists.

Improvements

• Common: Changes sync the moment they happen: Reconnecting a provider account, adding an MCP server, tagging a file, editing a reminder, and dozens of other changes now trigger a sync within seconds on every platform, instead of waiting for the next scheduled cycle.
• Common: Reconnected sign-ins reach every device: Disconnecting and reconnecting a provider sign-in (like Grok) on one device now lands on your other devices — the newer sign-in always wins over the older disconnect, in both directions.
• Common: Photos keep their thumbnails across devices: A photo sent from another device renders as a picture on arrival, and its full-size content fetches directly from the device that has it.

Bug Fixes

• Common: AI document edits can no longer be silently overwritten: An open Slate editor's periodic save could restore stale content over an edit the AI had just written. Every save now verifies it is built on the document's current content; a save built on stale content is refused, the newer content stays, and the editor's local text is preserved in version history.
• Common: A streamed answer no longer vanishes at completion: On the device watching a remote conversation, the final answer persisted only after a later sync — it now persists the moment the turn ends.
• Common: Reruns respect the chosen device: Rerunning a response in a conversation hosted on another device ran locally instead, producing two divergent turns. Reruns now run on the conversation's chosen host.
• Common: A document that cannot be retrieved stops the turn: When a file attached from another device could not be fetched, the model answered without it. The turn now stops with a clear message naming the file instead of answering from nothing.
• Common: Deleted items stay deleted — and deletions never carry secrets: A device offline past the deletion-history window could revive deleted conversations, files, or credentials, and a deletion record could retain the credential it removed. Deletions now resolve by causality and are stored without secret material.
• iOS: The file manager list scrolls immediately: Drag-scrolling the file list required tapping a checkbox first on some documents. It scrolls on first touch now.
• Common: A spreadsheet-style table no longer changes a document's type: An ordinary Markdown table could make a document present itself as a spreadsheet, changing the guidance shown alongside it.

Improvements

• Common: Credentials sync across your devices, and a blank never erases: API keys, sign-ins, and saved provider accounts travel between your devices inside your own encrypted sync. A device that doesn't have a key can never wipe that key from your other devices — including entries in the provider account list that carry no secret yet. Signing out of one device no longer switches off the included Caiioo key on the others.

Security

• Common: A sign-in can never adopt another profile's data: On a device holding a profile with saved connections, API keys, apps, or conversations, signing in with a different account could hand all of it to the newly signed-in profile. A sign-in now receives a fresh profile unless the existing one is verifiably its own, and the data migrations that run at start-up only move data when its owner is unambiguous — otherwise the data stays put, readable by no one else.

Bug Fixes

• Common: A busy free model explains itself and the fallback stays armed: When a free model's servers were overloaded, the first message could end with a misleading message about privacy settings. The real cause is now named plainly with what to do, and the automatic switch to a working free model stays active for people using Caiioo's account key — including after pressing the "Use Caiioo's Account" button, which previously switched that protection off.

New Features

• Common: Poe joins the provider picker on paid plans: Choose Poe as your model provider with a single Poe API key. Costs show in Poe points and dollars, and the model picker carries benchmark medals for Poe's catalog.

Bug Fixes

• Common: New Google AI keys are accepted: A key created in Google AI Studio recently — the kind starting with "AQ." — was rejected as invalid at paste. Both current and older key formats now validate.
• Common: Poe costs count each request once: Conversation cost tracking could count one request several times when helpers ran in parallel, and could bill a request to its predecessor. Each request now counts exactly once.
• Common: Poe models show their benchmark medals: Tier medals and intelligence scores in the model picker did not appear for Poe models.
• Common: Poe thinking settings stay within what a model supports: A thinking level a Poe model does not offer was sent anyway and the whole request was refused. The setting is now left at the model's own default instead.
• Common: Switching to Poe fills the model list right away: The model dropdown could sit empty after switching providers, with no way to pick a model.
• Common: A new account's first conversation starts cleanly: For a brand-new user on free models, the very first message could fail with a confusing message from the model provider that pointed at the wrong fix, while the guided tour kept its highlight painted over the failure in the wrong place. The setup now verifies the included key before the first message, the error (in both the banner and the conversation) explains what to actually do, and the tour steps aside until the error is dismissed.
• Common: Tool permissions fit a narrow panel: In a narrow side panel, the tool-approval buttons could overflow sideways or put the main action below the fold, the permission-mode choices squeezed into unreadable columns, and a long tool name could push its Revoke button out of reach. All three lay out correctly at phone and side-panel widths now.

New Features

• Common: Anthropic joins the provider picker on paid plans: Choose Anthropic as your model provider and use your own API key, alongside the other bring-your-own-key providers.
• Common: Create an Anthropic API key from your Claude account: Sign in with your Claude account and an API key is created for you — approve access in the browser, paste the code shown on the callback page, and the key is saved. The key bills to your Anthropic Console account, and your Claude sign-in itself is not kept.

Bug Fixes

• Common: Switching profiles sticks: Selecting another profile could switch and then snap back a moment later, and the previous account's subscription details could briefly land on the newly selected profile. A background account refresh no longer moves you between profiles, and never writes to a profile that is not its own.
• Common: A streamed answer holds its place: During a response with several tool calls, text that had already appeared could vanish into the collapsed "Working" section each time a new tool started, dragging your reading position up with it. Once something has been shown below that section, it stays until the response completes.
• Common: Emphasized names read correctly with the Pseudonymizer on: A name the model wrote in italics or bold using underscores could reach the screen as its substitute instead of the real name, and a name written that way in a tool result could slip past masking on its way to the model. Emphasis marks now count as punctuation in both directions.
• Common: Quoted names are caught by the Pseudonymizer: A name inside quotation marks in an instruction — asking for code that uses it, asking for it to be echoed back — could pass to the model unmasked, even though the same name was caught in ordinary sentences. Quoted values now get a second, focused look before anything is sent.
• Common: Today's date stays true with the Pseudonymizer on: The date supplied to built-in skills could be treated as personal data and swapped for a substitute, so a planning request could start from the wrong day. Dates the app supplies itself are exempt; dates you write are still protected.
• Common: Your own conversation resumes right away after an interruption: After an interrupted response, the same device could refuse to continue that conversation for up to a minute and a half with a generic retry message. A device is never locked out of its own conversation, and when a conversation genuinely cannot start yet, the message says what is happening and how long it lasts.
• Common: Stop reaches the device doing the work: Pressing Stop on your phone could clear the spinner there while the computer actually running the conversation kept working. Stop now travels to the device running the response and cancels it there.
• Common: Syncing over the local network keeps flowing: After a round of syncing over your local network, further syncing could pause for up to a minute at a time. Rounds now hand off cleanly.
• iPhone & Android: A debug transcript lists each line once: The transcript attached to a support request repeated every line twice.

New Features

• Common: The conversation's host follows you instantly: Switching which device runs a conversation now travels to every device in about a second, in both directions, including while a response is being written. A rename of a device travels the same way.
• Common: Conversations work on weak cellular: When a network can hold short requests but not a steady connection — common on LTE — the app now carries the same conversation traffic over ordinary web requests instead, and quietly returns to the fast path when the network allows. Sending to your computer works either way.

Improvements

• Common: Live streaming to your phone travels light: Watching a conversation that runs on your computer now streams the response text itself, with bulky tool output arriving once at the end instead of riding along with every word. Long, tool-heavy turns stream smoothly to a phone.
• iPhone & Android: Typing always comes first: Background syncing now waits until you pause and keeps a respectful distance between runs, so the keyboard stays responsive no matter what your other devices are doing.
• Common: Reconnecting is a reflex: The connection re-establishes in about a second after unlocking your phone, switching networks, or returning to the app — and keeps trying for as long as it takes, instead of giving up after a few minutes away.
• Common: The connection indicator tells the truth: The device chip and presence dots now reflect whether this device can actually reach the others right now, turn amber the moment that stops being true, and name the machine that is actually the problem when a message cannot be sent.

Bug Fixes

• Common: A conversation no longer claims your computer is offline while it is running fine: Returning to the app after a lock or a network change could leave it convinced the other device was unreachable — indicators green, messages refused — until a restart. The connection now re-proves itself on every return, so what the app believes matches what the wire can do.
• iPhone: The keyboard no longer lags in waves: Bursts of background work — syncing, and oversized bookkeeping arriving from your computer — ran on the same thread as the keyboard. The bookkeeping now travels as a light notification and the syncing yields to your fingers.
• Common: Messages sent right after unlocking are not lost: A send in the first seconds after unlock could be refused while the connection was still coming back. The app now waits briefly for the connection instead of giving up with your message.
• Common: Device syncing is steadier between machines: Syncing between your devices resolves several situations where two machines could quietly disagree — about which device runs a conversation, about a renamed device, and about changes made while a response was streaming — and does less redundant work in the process.

Improvements

• Common: Devices keep up with each other: Syncing does considerably less redundant work each cycle, and it now asks your storage provider what has changed instead of re-reading the entire folder every time. An idle cycle finishes in around five seconds, and the saving grows with the size of your account rather than shrinking.
• macOS: Logs are where you would look for them: The app's logs now sit in Application Support alongside the rest of its data. Logs already written to the old location are still read, so nothing you have collected is lost.

Bug Fixes

• Common: A folder you add is ready to use straight away: A folder added from the composer, added from Settings, or granted when the agent asked for it was not made active in the conversation you added it from, so the agent still could not reach it until you selected it again. All three routes now activate it where you added it.
• Common: An attached document is read straight away: A document whose name did not end in .pdf was treated as a PDF anyway, which meant several minutes of repeated attempts to read it as a scanned page before giving up. What a file actually is now decides how it is read.
• Common: A conversation that cannot run elsewhere tells you why: Handing a conversation to another of your devices asked you to retry whatever had happened, including for reasons no amount of retrying would resolve — most often a conversation grown too large to send. It now says which it is.
• Common: Clicking something that opens a new page returns as soon as it has: A click that took the agent to a new page waited out its full timeout before reporting back, whatever the page had already done. It now notices the new page and returns in about a second.
• Safari: The agent can run scripts on a page: Running a script was available in every other browser and not in Safari, where it reported nothing useful. It goes through the same channel Safari already uses for the agent's other actions on a page.
• Common: Dismissing a cookie or consent banner reports what actually happened: The agent reported success without looking. It now checks the banner is gone, and says so plainly when one is still in the way.
• macOS: The browser list shows each browser once, with current activity: A restarted Safari could register itself afresh every second, filling the list with duplicate entries for the same browser, and a connected browser's "last seen" time was recorded when it connected and never moved after that.
• macOS: The Safari extension matches the app it came with: The extension inside the Mac app could be a version behind the app itself, because the release packaged whichever copy had last been built instead of building it as part of the release.

Bug Fixes

• Common: Grok account sign-in reaches xAI directly: Signed-in Grok went through a command-line proxy that refused any client whose version it did not recognise, reporting that a Grok CLI needed updating — there was no CLI to update. Both chat and the model list now use xAI's ordinary endpoint, which has no version check, so nothing here can go stale on someone else's release schedule.
• Common: The model picker lists Grok models: The picker sat spinning and then reported no models available. The request was never leaving the app: xAI's proxy declines the permission check a browser makes before sending, so the answer never arrived. Discovery now uses the endpoint that permits it.
• Common: A conversation's cost reflects the provider that served it: Costs for Anthropic, OpenAI, Google and xAI keys were calculated at a marketplace's published rates rather than the provider's own. Each route is now priced by whoever actually served the request, and where a rate cannot be established the figure is omitted rather than estimated.
• Common: Baseten deployments load: The deployment list came back empty because Baseten declines that same browser permission check. It is fetched where the check does not apply.
• macOS: Devices appear under their own names: Every device showed as its platform — "macOS app", "iOS app" — because the name arrives sealed and the key to open it lives in a different process from the list. The list now asks the process holding the key.
• Common: A conversation stays on screen while a device catches up: A large update is trimmed to fit the network, and a trimmed update no longer reads as though the conversation had been deleted, which had left the open conversation blank until the app was restarted.
• Common: Devices sync over the local network: Syncing between machines on the same network had been routed through the internet even when the desktop was reachable directly. The LAN path is restored, with the desktop writing through to the same single copy of your data so both routes always agree.
• Mobile: The browser button opens a browser: The Globe in the composer had nothing to open when the agent held no tabs of its own; it now opens a blank tab with the address bar focused.
• Common: Dialogs show their last line: A dialog's footer overlapped the final line of its content.

New Features

• Common: Conversations you can keep off the record: Any conversation can be marked temporary from the title bar. It stays on the device you started it on — Caiioo will not learn from it, sync it, or back it up — and you delete it like any other conversation. Turning it back off tells you plainly that the whole conversation becomes remembered from that point, including the messages already in it, so you can decide to start a fresh one instead.
• Common: Your devices have names, and you can remove them: The device list in Settings → Connectivity → Private Constellation now shows each machine's name rather than "iOS app" three times over, and every device except the one you are using has a Remove action. Removing a machine takes it out of the list and stops it being offered work, instead of only hiding it.
• Common: A device keeps its identity across a reinstall: The list reflects the machines you own rather than every installation you have ever made. On Android this survives an uninstall; on iPhone it is kept in the keychain.
• Common: The model picker says what each model can do on the provider serving it: The same model can behave differently depending on who serves it, so capabilities are now answered per provider rather than guessed from the model's name. Models that have not been measured yet say so and stay visible, so a release that shipped today is findable on the day it ships.
• Android: The status bar follows your theme: The clock and battery could disappear into a light background. They now switch to suit whatever the app is showing, including when a change of accent colour is what flipped it.

Improvements

• Common: Agents run until the work is done: A task now gets 100 steps instead of 60, and each helper agent you fan out to gets 25 of its own rather than 5 — ask for six helpers and you get six, each doing a whole job. What stops a run that is genuinely stuck is no longer the count: an agent that makes the same call, gets the same answer back, and produces nothing ten times over is stopped for making no progress, whatever budget it had left. Waiting on something slow — a build, a long job — is not that, and keeps running.
• Common: The Self Checker scores answers again: The ⚖️ check had stopped proposing the concrete pass/fail checks it once wrote — a figure that must appear, a value that must not — and returned only its written assessment. It writes them again, and when a turn genuinely has nothing checkable it now says so rather than leaving the slot blank.
• Common: The conversation header keeps its shape: The measurements beside a conversation's title used to wrap to a second line wherever they happened to run out of room, which moved as the title, language or font size changed. Identity is now always the first row and measurements always the last.

Security

• Common: Approval is required wherever a tool asks for it: A tool requiring approval in a context with no way to prompt — which is how every sub-agent runs — was carried out instead of refused. In Cautious mode that meant sub-agents could act without asking. Refusing is now an explicit outcome.
• Server: The local server always requires credentials: The local server treated "no agent currently running" as permission to accept local requests without credentials, and that is its state most of the day. It now accepts only credentials it issued for a run still in progress.
• Common: The sandboxed shell runs without Caiioo's credentials: A shell command started by a tool received the credentials the app uses to talk to itself, which would have let it reach features you had not approved for that tool. They are kept out of its environment now.
• Common: A patched image library: The component that shrinks large photos before attaching them is updated to a release fixing four reported vulnerabilities in its image decoder. It handles pictures from anywhere you attach them, so it is worth being current.
• macOS: Computer Use handles out-of-range instructions safely: An out-of-range number in a Computer Use instruction could take the whole desktop app down. Values are bounded now, on both sides of the request.

Bug Fixes

• Common: Helper agents finish their task: A helper agent sent off to do a piece of work would come back with "Iteration limit reached" having barely started. It was being told to wrap up and conclude on its very first step: the reminder to start winding down is meant to arrive near the end of a budget, and it was set to a fixed number that happened to equal a helper's entire allowance. Several helpers doing that in one request read as several of them failing. The reminder now scales to whatever budget an agent actually has, so it always lands near the end.
• Common: More models carry a full benchmark ranking: Model rankings are built from published measurements, and a model missing a speed figure was dropped from the ranking entirely — taking its quality and price data with it, even though most models publish quality without speed. Those models now rank on what has actually been measured.
• Common: Device syncing is more reliable and more efficient: Syncing between your devices does considerably less redundant work each cycle, and devices on the same network find each other's fast path more readily, so changes made on one device reach the others sooner.
• Common: A conversation runs on the device you picked for it: Choosing where a conversation runs sends the work to that machine directly, whether you start it straight away or come back to it later. Previously the conversation could quietly fail to start on the device you chose — sometimes with your message simply disappearing — if another of your computers had been more recently started, or if the two devices had been idle long enough to need a fresh secure connection. A conversation whose device you have since retired now says so and offers to continue here, instead of asking you to open a device that no longer exists.
• Common: Your device list loads, and each computer shows its own name: On a desktop the list could fail to load entirely, leaving you unable to see or choose any of your other devices. It now loads, and every machine shows the name you gave it when you set it up — including the one you are sitting at, which used to call itself "macOS app" while every other device knew it by name. A name you set in Settings still wins.
• Common: Renaming a conversation updates it everywhere at once: A conversation renamed from its title bar kept the old name in the sidebar, and one renamed from the sidebar could leave the header behind — on iPhone and Android the two stayed out of step until the app was reopened. Both now go through the same path and agree immediately.
• iPhone: The browser button can be tapped: It sat in the rounded corner of the screen, far enough into the curve to be hard to hit. It now sits clear of it.
• Common: The browser button stays available in simple mode: Turning on the simplified layout removed the button that opens the agent's browser, and on Android it was left out of the "+" menu as well, leaving no way to reach it. It is treated as a capability now rather than decoration, so the simplified layout keeps it.
• Common: Your location describes one place: When your position updated, a previously cached street or neighbourhood could stay attached to a new city, describing an address that does not exist. Only fields that belong to the current fix are used.
• Common: A Pro trial includes the shared credit pool: Starting a trial unlocked the Pro features but not the credit that makes them work, so the first request could come back unauthorised. Trials now keep the shared pool.

Security

• Server: Four authentication gaps closed: A sweep of the relay and server found four places where a signed-in account could reach something belonging to another account, including the ability to redirect another account's incoming messages. All four are closed and covered by tests. We have no evidence any of them were used.

Bug Fixes

• iPhone & Android: The message box stays at the bottom of the screen: On an empty conversation it floated part-way up the screen, and on a full one it sat below the bottom edge. The app was sizing itself to its own contents rather than to the screen. A long message you were still typing could also grow until it pushed its own send button out of view; the box now stops growing before that happens.
• iPhone & Android: The browser button opens a browser: On iPhone it opened a stripped-down window with no address bar and no way back, and on Android there was nothing behind it at all. It also sat under the home indicator on iPhone, and on narrow screens it disappeared from the row entirely. It now opens the in-app browser on both phones, sits clear of the home indicator, and when the row runs out of room the crowded-out controls move into the "+" menu rather than disappearing.
• Common: A conversation started on your phone reaches the device running your agent: Starting a brand-new conversation on your phone while the agent runs on another machine could end with nothing to show for it — the message left the box, nothing was saved, and nothing was reported. Your devices share conversations a few seconds after the fact, so one created moments earlier had not arrived yet and the other machine had no idea what was being asked about. The phone now sends the conversation along with the message.
• Common: Your device list shows which machines are awake: The availability dots in Settings were fixed at whatever was true the moment you opened the panel, so a device that went to sleep while you were looking still showed as available. They now refresh while you watch, age honestly between refreshes, and a list that briefly cannot be loaded no longer erases the devices you already knew about.
• Common: A folder you share with one conversation stays in that conversation: Granting the agent access to a folder said it applied to that conversation only, and then applied it everywhere. The limit was written but never consulted. It is now enforced where it counts.
• Common: The creativity slider takes effect on every provider: On six providers the slider had no effect, because the value was quietly restored to its default just after the app had deliberately removed it. A related change stops one rejected request from teaching the app the wrong thing about a model and repeating it for the rest of the session.
• Common: Reminders and scheduled tasks settle after syncing: Items created on one device could download again on every sync, indefinitely, reporting success each time — enough traffic to make everything else feel sluggish. Rows arriving from another device are now filed where that device can see them, which ends the cycle.

New Features

• Common: Settings reorganized into clearer categories — in your language: The settings panel's sections are regrouped into seven categories (Account, Connectivity, Data & Sync, Help, and more) and category names now translate in all 24 languages instead of rendering in English everywhere. The grouping is derived from the settings layout itself, so a section can no longer go missing from the category list. Settings search finds entries under the new names.
• Common: Your devices have a name — the Private Constellation: Multi-device continuity now lives under Settings → Connectivity → Private Constellation, and the guides use the same name throughout. Which of your machines the agent runs on is a connectivity question, so it sits beside your other connection settings rather than with backups. Same encrypted, single-tenant machinery — now findable by searching "Constellation".
• macOS: Press-to-talk dictation is built in — no companion app: System-wide dictation now ships inside the desktop app on macOS as it already did on Windows and Linux. Pro subscribers hold the hotkey anywhere and speak into whichever app has focus. The separate PrivateVoice download is retired.
• Common: Your device list says how long each device has been quiet: When you pick which of your devices hosts a conversation, the ones that have gone to sleep now show how long ago they were last heard from — "3m", "2h", "1d" — with the freshest first, so it's obvious which machine is actually there to take the work. Devices that stay silent long enough drop off the list instead of accumulating forever, and a preferred device that can't be reached now says so instead of looking available.

Bug Fixes

• Android: The app knows which version it is: The version shown in Settings read "dev" on every Android build ever released, and support tickets arrived carrying no version at all — which made it impossible to tell which build a report came from. The page asks for its build stamp with a request Android blocks for pages served from the app's own internal address, so it was never actually sent. The stamp is now handed to the page as it loads.

• iPhone & Android: The screen stops sliding around under your thumb: The phone apps could feel "slippery" — the whole interface scrolling like a web page, the composer bouncing, and the keyboard shoving everything out of place. Four separate causes were doing this at once, on both platforms: the app was sized to a height that doesn't shrink when the keyboard opens, the page underneath it could scroll and rubber-band freely, and nothing made room for the keyboard at all. The shell now sizes itself to the space the app actually has, stays put, and both platforms resize properly when the keyboard appears.

• iPhone & Android: Attaching a photo no longer risks closing the app: A picked or captured photo was decoded at full resolution before being attached — roughly 48 MB of memory for a 12-megapixel image, and more still when several were selected at once, which on a loaded phone could take the app down. Photos are now scaled down as they're read rather than after, which is invisible in the result: the size that was being discarded is larger than any model reads anyway.

• Remote access: Generated images now arrive in a browser session: When driving Caiioo from a browser over your private relay, an image the agent generated could simply never appear — anything over about a megabyte was dropped in transit with no error on either side, and even smaller ones failed to load because the browser was asking the wrong place for them. Large attachments are now sent in pieces and reassembled, and images load over the encrypted tunnel.

• Android: Calendar and notifications actually ask for permission: The app never requested either one — it only checked, found them missing, and told you to grant them in system Settings. Since nothing ever asked, they could never be granted the normal way, so every calendar request came back "access denied" and reminders and turn-finished notices never appeared at all. Both now prompt the first time they're needed, and editing an event asks for calendar write access rather than assuming read access covers it.

• Android: Your phone's own calendar works with the agent: Even with permission granted, asking Caiioo to use the device calendar quietly fell back to Google Calendar, and the agent was told the device calendar needed a Mac or iPhone. The Android calendar is now a first-class source: it can be asked for by name, it's included when your day is assembled from every calendar you have, and it's described accurately to the model.

• Android: Documents and code open again instead of reporting a failure: Opening a Slate document, code file, or drawing failed on Android — and worse, when the agent created one, the file was saved correctly but the agent was told the attempt had failed, so it often tried again. Slates now open, and the agent gets an accurate answer.

• Android: The dictation model list no longer offers one that cannot run: A faster speech-to-text engine was listed on capable phones, downloaded several hundred megabytes, then failed on every attempt because its final step is unfinished. It's hidden until it works.

• Android: "Continue with Apple" is no longer shown: The button appeared on the sign-in screen but could never complete, because the Android app has no Apple sign-in support behind it. Google and Microsoft sign-in are unaffected.

• Android: Links open in a browser instead of taking over the app: Tapping an external link — in an answer, or "Manage Plan" — replaced the whole app with that page rather than opening a browser, and signing in to Google from there was rejected outright. Links now open properly outside the app.

• iPhone & Android: You're told when a turn finishes in the background: A notice when a long answer finishes or fails while you're in another app was only ever sent on iPhone, even though Android has the same capability and the same problem.

• Remote access: A sleeping relay no longer interrupts a running turn: The private relay treats going idle as ordinary housekeeping, but on waking it mistook that for one of your devices having dropped out — and responded by re-securing the connection, which knocked every connected device into a fresh handshake mid-answer. Waking is no longer confused with a device change, and the connection is only re-secured when the hosting device genuinely moves.

• Common: Settings search finds what you type: A two-word search matched nothing unless a section happened to list that exact phrase — "ad blocker", "font size" and most other natural searches came back empty. Every word you type is now matched against a section's title, description and keywords, so search narrows as you type instead of falling off a cliff.

• Common: Ad blocking sits under a Privacy heading: It was a loose card among your profile settings; it is a privacy control and now reads as one.

• Common: Zero Data Retention now covers image inspection: With ZDR enforced, Caiioo already refused to send an image to a remote model for automatic captioning — but the agent's image-reading tool reached the same remote models by another route and asked no such question, so an image could leave the device when the agent chose to look at it. That route is now checked exactly like every other ZDR-governed path.

• Common: Models that can see no longer ask another model to look: A vision-capable model was calling the image-reading fallback tool to describe pictures already in front of it — three times in one session — costing a turn and a second model's tokens each time, and sometimes failing with an attachment error. The fallback is now offered only to models that genuinely cannot see images.

• Android: The app no longer draws under the status and navigation bars: On Android 15+ the edge-to-edge layout put content beneath the system bars, so the top of the conversation and the bottom controls could sit under them. The app now measures the real system insets and hands them to the web layer.

• Common: The Self Checker no longer marks correct time-zone conversions wrong: The quality judge reviewed answers without knowing the conversation's date and time zone, so a calendar event correctly converted from another zone could be scored as a mismatch — twice, confidently. The judge now receives the same date/time-zone grounding the agent had.

New Features

• Desktop: "Run it and tell me when it's done" now works: Start a build, a test suite, or any long command in the background and the agent wakes up on its own when the command exits — with the exit status and the tail of the output — instead of you having to ask "is it done yet?". Works alongside the live Watch control from the previous release.
• Common: Each device remembers its own text size: Display scale tuned for a 27-inch monitor is wrong on a phone. Font size is now stored per device — your Mac, phone, and tablet each keep their own setting — while still riding your encrypted private-sync backup, so a restored device comes back at its own size.

Bug Fixes

• Desktop: File search in large projects is ~200x faster: A broad search used to scan generated build output (one real project spent over three minutes reading a 2.3 GB build directory) and could look hung. Search now honors every .gitignore in the tree — including nested ones — skips binaries and oversized files at read time with a visible notice instead of silently, and reports progress while it works. The same broad search now completes in about a second.
• Common: Gemini tool calls work again: Every Gemini request with tools attached was failing outright with a schema error ("Unknown name prefixItems"). Google's API rejects a schema field other providers require; it is now stripped for Gemini only, with nothing lost — the equivalent constraint travels in a form Gemini accepts.
• Common: Claude 5 through gateway providers no longer fails on temperature: Gateways backed by Bedrock (such as CometAPI) reject the temperature parameter for Claude 5 models with a misleading server error. The request now retries once without the parameter and remembers the model for the session.
• Common: Composer buttons size to your actual selection — and stop vanishing: The top-row chips (mode, model, device, Folders, Skills) sized themselves against worst-case widths, so a short model name still wasted space and — on macOS — a browser-engine rounding quirk made the row believe it was overflowing and hide the Skills and Folders buttons entirely, at every window width. Chips now measure the real selection, compress only under genuine pressure, and the two buttons stay put.
• Mobile: The model picker updates on every switch, not just the first: Switching models a second time on iOS updated the running conversation but froze the picker's own display on the previous choice.

New Features

• Desktop: The agent can request access to another folder — you approve first: When work leads outside the folders you have shared, the agent can now ask for a specific directory by path. The request goes through the standard approval prompt showing the exact folder, and an approved grant applies to that conversation only — nothing becomes account-wide as a side effect. A folder that is already shared is reused rather than duplicated.
• Desktop: Watch a running command live, and nothing is ever lost: Long-running commands (builds, dev servers, test runs) now show a "Watch output" control that tails the live stream as it happens. Every command's complete output is also kept and addressable afterwards — the agent can ask for just the tail or just the error lines up front, and search or re-read the full record later instead of re-running the command. Oversized output is trimmed to the newest portion with a note saying exactly how to retrieve the rest.
• Common: The model picker shows each model's maker: Models now carry a colored maker mark — Anthropic, OpenAI, Google, xAI, Meta, Mistral, DeepSeek, Qwen, and more — in place of the one generic icon, so scanning a long model list reads at a glance. New vendors appear automatically with a neutral mark.

Bug Fixes

• Common: Sleeping devices stay visible when choosing a conversation's host: A phone goes quiet within a minute of being pocketed, and the host picker used to hide it entirely — which read as "my iPhone is not connected" and left the current host unnamed. Quiet devices now stay listed, greyed, with how long ago they were last seen. Automatic handoff still requires a live device.
• Common: Your model and provider choice rides along with a remote message: Sending into a conversation hosted on another device used to run with whatever settings that device had — a provider or model just picked on your phone could silently race the settings sync and lose. The choice now travels with the message itself. Credentials never travel; the host always authenticates with its own keys.
• Common: Very large conversations stream reliably between devices: A conversation past a size threshold silently stopped delivering its completion and prompt events between devices — the run looked like it never ended on the viewing device. Oversized events now automatically slim to the current turn, which merges safely, and the full history follows by ordinary sync.
• Mobile: The context-usage dial is tappable: The small ring showing how full the conversation's memory is was a hover-only target far below fingertip size. It now has a full-size touch area and opens on tap.

New Features

• Desktop & Mobile: Drive a conversation from your phone while it runs on your Mac: A conversation now has a host device, shown on a picker in the composer. Send a message from your phone into a Mac-hosted conversation and it executes on the Mac — with the Mac's folders and tools — while the answer streams live on both screens. Flipping the host from either device takes effect on the other within a couple of seconds.
• Mobile: Lock your phone mid-task and your desktop finishes it: With auto-continue enabled, locking the phone during a running turn hands the work to your preferred desktop, which completes it and syncs the finished answer back — visible on the phone when you return. The handed-off turn is clearly marked on the phone instead of spinning forever.

Bug Fixes

• Common: Sync between your devices actually delivers again: Encrypted sync uploads were being emptied in transit — every device reported successful syncs while conversation names, completed answers, and settings never arrived anywhere. The stored files were zero bytes across the board. Uploads now land intact, and a backlog of thousands of items that had been crawling under a self-imposed rate limit (the "Failed to write file: 429" error some testers saw) clears in seconds.
• Common: Changes synced from another device appear without reopening the app: A host flip, a renamed conversation, or a finished answer that arrived by sync was stored correctly but the open window was never told — it kept showing the old state until you switched away and back. Synced changes now refresh the open screen immediately.
• Mobile: "Use this device" and auto-continue off are honored: Turning auto-continue off, or explicitly keeping a conversation on the phone, did not actually keep the work there — locking the phone put the turn up for grabs and any desktop took it. A turn now stays on your device when you have said it should, resuming when you unlock; and a quick two-second app switch no longer hands your turn away at all.
• Mobile: A turn handed off can no longer also re-run on the phone: Returning to the phone after a handoff could silently re-execute work another device had already completed — including retry timers that fired the moment the app woke. Every path a released turn could sneak back through is now closed.
• Common: A message that cannot reach its host stays in the conversation: Sending into a conversation hosted on an unreachable device used to clear the composer and lose the message entirely, with no error. The message now stays in the conversation with a clear explanation of what to do.
• Common: The trial banner offers a current plan: The upgrade banner could offer a subscription tier that no longer exists, which made it disappear entirely for some users. It now always offers the current plan.
• Common: Token counts say what was measured: The "prompt tokens" figure is actually the total billed across every model call in a turn — it is now labeled that way, with the number of calls shown. Cache savings on long conversations are surfaced in the working steps as they happen.

New Features

• Testers — Sign in with ChatGPT instead of entering an API key: OpenAI can now be connected with the same web-account sign-in used by ChatGPT. The connection works directly from the browser extension, without installing or running the desktop app, and stays isolated as an explicitly enabled tester feature.

Bug Fixes

• Browser extension: Switching tabs updates the page Caiioo is working with: The extension could keep sending the previously active tab after you moved to another one, so a request about the page in front of you was answered from the wrong page. Tab changes now reach the relay immediately.
• Common: Agenda fits inside a narrow window: The Agenda dialog could grow wider than a phone or compact sidepanel, pushing controls and event details out of reach. Its controls now wrap and each list scrolls within the available space.
• Common: Reading during a long answer keeps your place: Content arriving above the current reading position could pull the conversation back toward the bottom even after you had deliberately scrolled up. Only an actual scroll gesture now re-enables automatic following.
• Documents: Reopening a Slate uses the copy that is already open: Opening the same document again could create a duplicate tab or fail to move to the requested section. Caiioo now finds the exact existing Slate, refreshes it, and activates it.
• Documents: Word exports keep links and tracked changes valid: Generated Word files could contain malformed tracked-change markup, duplicate change identifiers, or broken links inside tables. The exported document now remains well-formed and opens correctly in Word.
• Documents: Multi-part edits are saved atomically: A batch of Slate edits could partially save before a later edit failed, leaving a document in an unintended in-between state. The whole batch now succeeds once or leaves the document unchanged.
• Common: Mode-specific tools and settings no longer leak into the next mode: Switching modes could retain administrator defaults or enabled tools from the previous mode. Each mode now starts from its own defaults while preserving the settings that belong to you.

Bug Fixes

• Desktop: The long pause before Caiioo starts answering: Every turn began with roughly half a minute of nothing — the same delay on a two-message conversation as on a twenty-four-message one, which is what gave it away as a fixed toll rather than the conversation being slow. A local tool server that had failed to start was being recorded as started, so each turn opened by waiting out a full connection timeout against something that was never there. Two things were wrong and both are fixed: a failed start is now noticed at once instead of thirty seconds later, and tools installed through Homebrew are found when the app is opened from the Dock rather than from a terminal.

Bug Fixes

• Desktop: Reading a PDF or Word document from a workspace folder: Asking Caiioo to read a PDF or a .docx sitting in one of your own folders failed outright, with errors that pointed nowhere useful — the only way through was to shell out to a command-line extractor, which took several attempts to discover. Both now read directly. Extracted PDF text also comes back without the stray runs of spaces that used to sit between words.
• Desktop: Computer use works without pairing anything: Every computer-use action failed — first claiming no LAN connector was configured, then, once it could find the relay, being refused as not signed in — on machines whose relay was connected and the user signed in. The agent runs in its own process alongside the relay, and was both looking for a network pairing it never needed and offering a credential it could not produce. It now uses the connection and the credential it already has, and when there genuinely is no route the message says so instead of naming an unrelated setting.
• Desktop: You can see what computer use saw: A screenshot taken by the computer-use tool arrived as raw data that nothing on either end could display, so it reached neither you nor Caiioo. Asked to look at the screen, Caiioo would confidently report its exact dimensions — read from a line of text beside an image it had never seen. Screenshots now appear as pictures in the tool result and are shown to the model, the same way browser screenshots always have been.
• Common: The composer keeps its send button, whatever the width: The row of controls beside the message box could grow wider than the window, and the send arrow was pushed off the right edge with no way to reach it. The row now gives up space in order as it narrows — the tool-configuration and pop-out buttons first, then the words beside Settings and Agenda, leaving their icons. The send button is never one of the things given up.
• Common: Computer-use no longer breaks every turn on Gemini: With the computer-use tool switched on, any turn run through a Gemini model failed outright with a schema error from Google, taking every other tool down with it — the tool descriptions were valid for one vendor's schema rules and not another's. They now satisfy both.
• Common: Caiioo no longer reports work it did not do: A tool call could be lost between the model and the app — and when that happened the written part of the answer survived while the action did not, so you were handed a detailed, confident account of work that never ran. Nothing anywhere reported a problem; the only signal was you checking. Losing a call is now treated as the failure it is: the turn stops and says so, rather than answering. The same gap existed on every model provider and is closed on all of them, and the underlying causes — arguments arriving for a call the app had already dropped, and two calls in flight at once being mixed together — are fixed rather than papered over.
• Common: An MCP server that needs re-authorizing is not deleted: Caiioo checked its saved connections at startup, and read "this token needs attention" as "this server is gone" — so it deleted the server, and because your settings sync, the deletion travelled to every device you own. Eight servers disappeared this way from one account. A connection that stops working is now marked for reconnection and keeps everything you configured; a server is only removed when it genuinely no longer exists. Existing connections repair themselves at the next check, without re-authorizing.
• Common: Connecting an MCP server survives a restart mid-sign-in: Authorizing a server means leaving Caiioo, signing in, and approving — and if the app restarted anywhere in that window, the credential it had just registered was lost. The connection appeared to work and could never refresh afterwards, which is what eventually triggered the deletion above. The credential is now written down before you leave, so the round trip is safe to interrupt.
• Desktop & Extension: A click scrolls to what it is clicking: Clicking something below the fold on a long page did nothing and reported success, so anything built on top of it carried on as though the click had landed — occasionally writing the wrong value into a form as a result. Clicks now bring their target into view first, and a click that produces no observable effect is reported as a failure instead of a success.
• Common: Token counts and costs describe the request that was made: Some turns recorded a prompt of two tokens — impossible for a real conversation — because a later message from the provider was overwriting the real figure. Cached tokens were also billed at nothing at all, which on a long conversation understated a single turn by several dollars. Both are corrected going forward.

New Features

• Common: Keep several accounts or keys for the same provider: Each provider held exactly one credential, so a second account — or a work key alongside a personal one — meant overwriting the first and pasting it back later. Every provider now keeps a list, signed-in accounts and API keys alike, with one of them in use. Switching between them is a click, and removing one from Caiioo revokes nothing at the provider.
• Common: Run without a step limit: Long tasks — a large refactor, a multi-document review, a browser sequence with many steps — routinely hit the cap and stopped mid-task, which is an abrupt end rather than a finished job. Max Steps can now be set to Unlimited, both in Settings and for delegated sub-agents. An uncapped run still stops on the Stop button, on cost and context limits, and if it stops making progress.
• Common: Interrupt a run, and see whether your interjection landed: Typing while Caiioo is working queues your message for it to pick up at its next step — right while things are going well, wrong once they are not. There is now an Interrupt button beside it that stops the run and sends your message as an ordinary turn. Queued messages say so while they wait, and a message the run ended without ever reading is now marked as not delivered, instead of looking exactly like one that was acted on.

Bug Fixes

• Common: Configured MCP servers stop disappearing — and the ones already lost come back: Signing in can rename your profile behind the scenes, and only conversations and files were carried across. Everything else kept pointing at the old name and became invisible: still on disk, impossible to see. MCP servers are what got noticed; reminders, scheduled tasks and attachment folders went the same way and said even less about it, since a scheduled task simply stops running. All of them are carried across now, and anything stranded by an earlier version is adopted back on your next sign-in rather than needing to be set up again.
• Common: An approval dialog you can actually answer: A confirmation carrying a lot of content — a long proposal, a large set of tool arguments — grew past the window and took its Accept and Reject buttons off-screen, with no way to scroll to them and no way to dismiss it. The conversation was stuck until the app restarted. Dialogs are now bounded to the window, scroll their own content, and keep the buttons in view.
• Common: Connecting a remote MCP server works on the first attempt: The first authorization failed with "Missing required parameter: client_id" and only went through after a refresh. Caiioo was treating a blank credentials form as though you had filled it in, so it never registered with the server at all. It registers properly now — and if a server ever does return no client id, Caiioo says so itself rather than handing you the provider's error page.
• Desktop & Extension: Reading a form's checkboxes, and clicking what was just found: Inspecting a page returned form controls without saying whether they were ticked, so the only way to check a box's state was a screenshot per field. Clicking could also fail with "Element not found" on an element found a moment earlier, because menus that redraw themselves genuinely remove it for a frame. Element state is now reported directly, a click waits briefly for a redraw instead of giving up, and region screenshots work on pages with a strict security policy.
• Desktop & Extension: A failed Python run says what it got done first: When a script stopped part-way, the result was a bare error — nothing about which steps had already completed, and no way to tell an interpreter that had been restarted underneath you from an ordinary typo. Every run now reports whether it started with a fresh interpreter, how much of its tool-call budget is left, and what it printed before it failed. Screenshots taken from inside a script now say plainly that a script cannot see them.
• Common: A repeated tool result is reported as a new observation: When a result was identical to an earlier one, Caiioo replaced it with a pointer that read as "you already have this". For the common case — re-reading a page to check a change was saved — identical is the opposite of reassuring: it is evidence the change did not save. The note now says a fresh observation was made and that nothing had changed.
• Common: "Just file it" is taken as an instruction: Asked to do something directly, Caiioo would still confirm the details it had just been given, and look up a tool it had used successfully one turn earlier. Both now step out of the way when you have already been explicit.
• iOS: Deleting your account warns you about Apple billing first: Deleting a Caiioo account does not cancel a subscription bought through Apple — only you can do that, through Apple — so it was possible to delete the account and keep being charged, with no account left to explain the charge. Caiioo now says so before the deletion and offers to open Manage Subscriptions.
• iOS: A purchase the app could not report is reconciled on its own: If the confirmation to our servers failed after Apple had taken payment, the subscription worked on that device and nowhere else — a reinstall or a second device showed no subscription at all. Apple's own notification now completes the purchase server-side, so it follows you.

Bug Fixes

• Common: A long delegated task is no longer cut off after three minutes: Handing a big job to a sub-agent — deep research, a long summary — failed after 180 seconds with a timeout and no output at all, as though it had never started. It had started; it was thinking. Newer models reason silently before they write anything, and Caiioo was treating silence as a stall. It now recognises a working sub-agent and gives it up to twenty minutes, after which the old protection resumes so a genuinely stuck task still stops rather than hanging forever.
• iOS & Android: The folder button no longer opens an empty menu: Phones and tablets have no filesystem Caiioo can reach, so the workspace-folders menu correctly had nothing to show — but the button that opens it was still there, and tapping it opened a blank panel. The button is now hidden where folders cannot work.
• Common: CometAPI models show what they actually cost: Dated and specialised variants were shown at their model family's price rather than their own. Of 171 affected models 41 were wrong, in both directions — one shown at 500 times its real price, and, more seriously, others shown far cheaper than they actually are. Every model now shows its own rate.
• Common: A quality check that cannot run no longer counts as a failure: In the browser extension, one kind of self-evaluation check cannot execute at all, and it was being recorded as a failed check — quietly lowering an artifact's score and reading as a real quality problem. Checks that could not run are now reported as such and left out of the score entirely.

Bug Fixes

• Desktop & Extension: Browser automation works again: Clicking, filling fields, inspecting the page and taking scaled screenshots all failed with a Content Security Policy error, on every site — Wikipedia included. Chrome's extension rules forbid building code from text at runtime, and that was how each of these actions reached the page. They now call functions the extension already ships, so nothing is built on the fly. You need to update the Caiioo browser extension as well as the app — the extension is what performs these actions, and until it updates you will see a note asking you to.
• Desktop & Extension: A bad selector explains itself instead of failing the action: An unparseable CSS selector made the page script fail silently, and the action came back with an internal error that named nothing useful. Worse, the check runs before the action does, so one bad selector took out clicking, filling, typing, hovering and scrolling before any of them ran. Caiioo now says which operation failed and which selector the browser rejected. This also restores screenshot dimensions, which had been reported as 0x0.

New Features

• Desktop & Extension: Work with the files you already have: Point Caiioo at a folder and it reads and edits your files where they live, instead of making a second copy inside the app. Folders are set up once and can be switched on or off per conversation, and a new conversation starts with none of them active. Before overwriting a file you wrote, Caiioo saves a one-step undo — and it refuses outright to flatten a Word, Excel or PDF document into its own extracted text.
• Desktop: Run a project's own toolchain: The shell now starts in the right project directory, so npm test or cargo check runs against the folder you meant rather than wherever the app happened to be.
• Windows & Linux: Window placement controls: The sidepanel gains buttons to tile it left or right, float it, pin it on top, or tuck it below other windows.
• Common: Redact a single turn instead of deleting the rest: Removing something mid-conversation used to mean deleting that message and everything after it. You can now redact just that one turn — the prompt, the response and any tool calls go, the rest of the conversation stays, and the assistant only sees that a turn was redacted.
• Common: Caiioo can pick its own voice: When it reads something aloud it can now choose a voice, a speaking rate and a speech model to suit the moment, rather than always using your global voice setting.

Bug Fixes

• Common: Long tasks are no longer cut short at ten steps: One setting had four different defaults scattered across the code and the smallest silently won, so involved tasks stopped after ten steps regardless of your model or plan. There is now one default.
• Common: A failed turn shows its error straight away: An error appeared in the conversation only after you sent another message — and then attached itself to the earlier turn. It now appears when it happens.
• Common: Reading a spreadsheet reports where the data actually ends: Google's API trims empty rows from every read, so two reads of the same rows come back different lengths and a full page looks the same as the last page. Caiioo mistook one for the other and could write an update onto the wrong row. It now reads the sheet's real extent.
• Common: A key from the wrong provider is named on sight: Pasting, say, an OpenRouter key into the Poe field failed every request with a bare "Invalid API key". Caiioo now recognises the mismatch as you enter it.
• Messaging: Setup says what actually turns it on: You could follow every documented step and still receive nothing, because inbound messages need Remote Access switched on for the device — a separate, off-by-default setting that isn't synced between your devices. Settings and the guide now lead with it, and offer to turn it on.
• Common: The composer stops hiding its own buttons: On a narrow window the icon row was clipped, cutting off whichever buttons happened to sit last with no way to reach them. They now move into an overflow menu instead. The model selector also stayed hidden exactly when a provider failed to load — which is when you most need it to switch away.
• Common: A custom mode is no longer created by accident: Simply opening and closing the settings dialog created a permanent copy of a built-in mode you never edited, and synced it to your other devices.
• iOS & Android: The in-app browser recovers instead of stalling: Hitting the open-tab limit stopped research mid-task rather than reusing a tab, and PDFs would not open.
• Common: A whole turn no longer fails because of one tool's schema: Some providers check every available tool together and reject the entire request if one of them is malformed — so a single bad entry took down the turn, with an error naming a position rather than a cause.
• Common: CometAPI models load again: CometAPI changed the shape of one field in its catalog and the model list stopped loading entirely — the picker came up empty with no usable explanation. Caiioo now reads both the old and the new shape, and a check runs on every build so a change like this is caught before it ships rather than after.
• Common: Tool calls work through gateways that route to Claude: Asking for something that needed a tool could fail outright with a validation error from the provider. Gateways like CometAPI accept requests in one format and forward them to Claude, which is stricter — several of Caiioo's own tools were being sent in a form Claude rejects, and one bad tool takes down every other tool in the same request.

Bug Fixes

• Common: Asking about a tool no longer answers "unknown": Checking whether a tool was available — the calendar, for instance — could come back claiming the whole category did not exist. Caiioo would then go looking for something that was never missing, and in the worst case fell into a loop of empty approval dialogs instead of doing the work. It now answers with the tool, or says plainly that the tool was not found and how to list the ones that are.
• Common: An action Caiioo says it did actually happens: Some models wrote a tool call out as ordinary text rather than performing it, then reported a confident, invented result — including a bug-report number for a report that was never filed. A written-out intention is now carried out for real, and the next thing Caiioo tells you is the actual outcome.
• macOS & Desktop: Read-aloud works with ElevenLabs voices: Every ElevenLabs synthesis failed on the desktop app with an audio-decoding error, because the audio was requested in a format only a browser can decode. It now asks for a format the desktop can play.
• Common: The scroll-to-bottom arrow stops overlapping the suggestions bar: The arrow was positioned against the message box, so it landed on top of whatever sat above it and flickered as it appeared. It now sits against the conversation itself.

New Features

• Common: Caiioo can file bug reports and feedback for you: Ask it to report a problem and it opens the built-in support form pre-filled with an engineer-style write-up for you to review and send. Say "just file it" and — with your approval — it submits the ticket directly, attaching the conversation's technical transcript so the team can reproduce what you saw.

Bug Fixes

• Common: The view no longer jumps back when a reply finishes: The moment an answer completed, the suggestions panel appearing below the conversation silently shrank the reading area, knocking the view off the bottom — the end of the answer slid out of sight and the screen appeared to bounce back to your own message. The view now stays glued to the end of the reply through everything that appears at completion.
• macOS & Desktop: Reading during a reply holds your place: With the Working Steps panel open, tool calls and reasoning arriving above your reading position shoved the view around — worst in the Mac app, whose browser engine does no scroll correction of its own. The conversation now anchors whatever you are looking at, on every platform, and only ever follows the bottom when you are already there.
• Common: A streaming answer renders once, smoothly: One answer used to be drawn three different ways over its lifetime — a live copy, a delayed copy, and a final rebuilt copy — stitched together by guesswork that produced doubled text, blank flashes, and the layout switching shape mid-reply. There is now a single path from the model to the screen, with a steady layout from first token to final answer and a smooth typewriter reveal.
• iOS & Android: The finished answer always arrives: On phones, the fully composed final answer was silently never delivered — the screen kept whatever had streamed and lost anything assembled at the end. The final state now arrives on the same channel as the rest of the reply.
• Common: Asking for a tool by name always finds it: Saying "use the calendar tool" could come back with Caiioo unable to do it — the tool matcher ranked questions about tools poorly, even with the tool's own name in your message. A tool you name directly is now always made available.
• Common: Discovered tools can actually be used: When Caiioo looked up a missing tool mid-conversation, the lookup answered "call it by name and it will connect" — a promise some models can structurally never act on, leaving Caiioo stuck asking pointless permission dialogs instead of doing the work. Discovered tools are now connected immediately.
• Common: Images loading into a reply stop shoving the text: A generated image arriving used to expand from a one-line spinner to its full height, pushing what you were reading far down the page. Space is now reserved while it loads.

New Features

• Common: Caiioo can install what it needs: Asking for something Caiioo can't yet do meant being told which settings screen to open. It can now search the Community Hub, show you what a package does, and install it once you approve. It can also list your MCP servers, add one, and start a sign-in — including for servers stuck needing sign-in, which register no tools and so were previously invisible to it.

Bug Fixes

• Common: Edits to a PDF survive zooming: Changing text in a PDF and then zooming made the change disappear from the page. Worse, clicking that text and clicking away destroyed the edit for real, while the document still believed it was there — so what you saw and what would have been exported no longer agreed.
• Common: Syncing a Word, Excel or PDF file to Google Drive no longer replaces it: A file opened from Drive kept a working Sync button, and pressing it uploaded a plain-text rendering over your original document and reported success. Syncing now declines and points you at "Save to Drive", which writes a real file.
• Common: Comments are no longer deleted when you save: Comments Caiioo added, and every comment on a code file, were removed by the next save.
• Common: Deleting a document stops it saving over itself: On the app and on phones the editor stayed open and kept saving into the document you had just deleted, reporting "Saved" each time.
• Common: Undo can't reach into another document: After accepting a tracked change and then opening a different file, one press of Undo could replace that file with the previous document's contents.
• iOS & Android: Downloads actually save: Every export except PDF wrote nothing at all and still said "Downloaded".
• Common: Two changes in the same paragraph mark up correctly: The second crossed-out change landed in the middle of a nearby word.
• Common: Word's tracked changes can be accepted and rejected: On a document imported from Word the buttons did nothing — the change stayed and the review counter never moved.
• Common: Asking for several changes in a row keeps all of them: Each suggestion was measured against the original document rather than the one before it, so accepting the second undid the first.
• Common: Formulas survive exporting: A document containing maths exported to Word, web page or rich text with every formula missing.
• Common: Formatting no longer rewrites code and preformatted text: Pressing Format on a web page reflowed anything inside a code, script or style block, changing what it did.
• Common: Accented and non-Latin text in data views: Names and text with accents, Chinese, Japanese or emoji arrived scrambled.
• Common: Exporting a spreadsheet keeps every sheet: Two sheets whose names differed only by a punctuation mark produced an archive with one of them missing, while still reporting the full count.
• Common: A new conversation can't overwrite another one's document: Starting a document with a name already in use elsewhere replaced that document instead of making a new one.
• Common: An answer arriving is no longer interrupted by syncing: Sync starting while Caiioo was replying competed with the reply itself, so the answer visibly stalled and then landed all at once. Sync now waits until the answer is finished, on phones and on the Mac and Desktop apps alike. A Google Drive delete also left sync retrying forever.
• Common: A streaming answer no longer flickers as it arrives: Half-finished formatting was being drawn before it was complete — an unopened code block turned the rest of the message to monospace for a moment, a partly typed bold marker flipped everything after it, and a table appeared as a row of pipes before snapping into shape.
• Desktop & macOS: Long conversations no longer crash the assistant: A conversation with a very large amount of context could take the assistant down outright, with nothing in the log explaining why.
• Common: Security fixes in document previews: Text taken from a document — a 3D world file's contents, or the error shown for an invalid diagram — was inserted into the page without being escaped.

Bug Fixes

• Common: Long conversations on large-context models no longer crash: With models that accept very large conversations, Caiioo could misjudge how much a model would take — reading a much smaller limit in one place than another — and send far more than it had planned for. On desktop this could take the assistant down mid-answer. Attachments carried inside a message, such as a PDF, were also counted as costing nothing, so a large file could sit in a conversation unnoticed until it caused a failure.
• iOS & Android: Every full-screen view has a way back: Opening a cairn or an image filled the screen with no way to return to the chat, leaving a force-quit as the only option. The "← Chat" button now appears on every full-screen view.
• Common: Starting a new document no longer loses your edits: If you had just been working on a file you uploaded, everything typed into the next new document was saved to the wrong place and was gone when you reopened it.
• Common: Looking at a mode's instructions no longer makes your own copy of it: Opening a mode's prompt and saving — even without changing anything — turned that mode into a personal copy. It then showed up in the mode picker as yours and travelled to your other devices as a customized mode, with nothing to indicate it was never actually edited. A copy is now made only when you genuinely change the wording.
• Desktop & macOS: A crash reports what actually happened: When the assistant was stopped unexpectedly, the log recorded only that it had exited, with none of the detail explaining why. The reason is now captured and kept.

Bug Fixes

• Desktop & macOS: Saving a file into one of your folders works: Asking Caiioo to write a document out to a folder on your computer failed on macOS, Windows and Linux — often by reporting that it had no way to do it, and sometimes by saying it had succeeded when nothing appeared. Four separate causes are fixed: your attachments weren't reachable as a location to copy from, a generated Word document never made it out of the background process that created it (so no attachment appeared anywhere, despite a success message), two different checks disagreed about whether a file existed, and Caiioo often couldn't find the right tool for the request in the first place.

Improvements

• Desktop & macOS: A failed request leaves something useful in the log: When something went wrong, the app's log file often contained nothing from the assistant at all — the detail was only recorded if you had already switched on debug logging before the problem happened. Details from a failed request are now kept and written out automatically, so a problem can be diagnosed from the log without having to reproduce it first. Successful requests log no more than before.

New Features

• Desktop: Caiioo understands your code, not just the text of it: Searching for a name finds every place it appears — the definition, unrelated things that happen to share the name, and mentions in comments alike. Caiioo can now ask a real language server instead: jump to where something is defined, find everywhere it's genuinely used, read a symbol's type and documentation, list what a file defines, and see the errors and warnings your editor would show. Available on macOS, Windows and Linux, on the folders you've given Caiioo access to.
• Desktop: Long-running commands keep running: A command had to finish inside a single step, so starting a dev server, a watch build, or a full test suite either timed out or held everything up. Caiioo can now start a command in the background, get on with other work, and come back to read its output — and list or stop what's running. Background commands stay inside the same folder-only sandbox as before.
• Desktop & Extension: Search shows the code around a match: File search adds case-insensitive matching and the option to include the surrounding lines, so a result arrives with enough context to judge it instead of needing a second read of the file.

Bug Fixes

• Common: Editing an uploaded file no longer replaces it with a blob of text: Opening one of your own files in Slate and saving it overwrote the file with Slate's internal record of the edit. Reopening it then showed that record — a wall of braces and quotes — instead of your document. The original file is now kept as it was and edits are stored alongside it, matching how Word, Excel, PDF and rich-text files were already handled. Files already damaged this way recover the next time you open them.
• Common: The speaking-speed slider works on the newest ElevenLabs voice: On the v3 voice the slider did nothing — that voice can't change speed itself, and asked to, it returns normal-speed audio without reporting a problem. The speed is now applied while playing instead, keeping the pitch natural. One trade-off: with the slider away from normal, v3 waits for the whole clip before it starts rather than speaking as it goes. At normal speed it still starts right away.
• Common: The app's name is written Caiioo: The name appeared in lowercase under the icon on iPhone and Android, in the browser extensions, and in the Safari extension.

New Features

• Common: Your memories arrive on a new device ready to use: What Caiioo remembers from your conversations now syncs along with everything else. Previously a new device had to rebuild all of it from scratch before recall worked at all — slow on a phone or tablet, and it had to happen before your history was any use. Deleting a conversation on one device now also removes what was remembered from it everywhere, so an old conversation can't answer for you after you've deleted it.
• Common: Mark up a PDF: Caiioo can draw on a PDF — freehand ink to circle a clause or point an arrow at a figure, notes placed anywhere on the page, and a signature image stamped where you want it. The marks are written into the PDF itself, so they're still there when you download it and open it anywhere else. A stamped signature is a visual mark only: it carries no identity verification and is not a legally binding e-signature.
• Common: Precise file edits instead of whole-file rewrites: Caiioo can change several specific parts of a file in one go, and either all of the changes apply or none of them do. It can also show you exactly what a proposed rewrite would change before anything is written, and copy a file between folders without altering it — which is the only safe way to move images, PDFs and other non-text files.
• Common: Every ElevenLabs voice starts speaking right away: Only two voices played as they were generated; every other one — including the expressive v3 — went silent until the whole clip was ready. They all start speaking immediately now, and v3 is selectable in the voice picker.

Improvements

• Common: A faster, cheaper default voice: The default ElevenLabs voice is now Flash v2.5. ElevenLabs describes it as equivalent in quality to the previous default and recommends it for all uses, and it costs half as much per character. The cost estimate shown in settings has also been corrected — some voices were being quoted at the wrong rate.
• Common: An interrupted answer tells you what it may have already done: If Caiioo is cut off mid-answer — the app closes, the phone locks — any tool it had already started is no longer silently discarded. Anything that may have acted in the world, like sending an email, is named in the reply, so neither you nor Caiioo has to guess whether it happened, and it won't quietly do it a second time when you continue. Actions that are harmless to repeat, like reading a file, stay out of the way.

Bug Fixes

• Common: The app no longer stalls for minutes at a time: Syncing could get stuck re-downloading the same items over and over — roughly 50 seconds out of every 60 — and because settings and conversations are read through that same connection, everything felt slow while it ran. Three separate causes are fixed, and syncing settles instead of looping. Reminders, skills and other items that kept reappearing now stay put.
• Common: A dropped connection is retried again: On the browser extension, Mac, iPhone and Android, a connection that failed mid-answer was never retried — the check only recognized the kind of failure the server sees. Those failures are now recognized everywhere and retried, as long as nothing has been shown to you yet.
• Common: Some tools broke every message on a Grok connection: Turning on a tool with on-screen coordinates, such as Computer Use, made Grok reject every message in the conversation rather than just that tool. Their descriptions are now sent in the form Grok expects.
• Common: Reading a finished answer no longer snaps you to the bottom: If you scrolled up to read an answer as it finished, the view jumped back down to the end — twice, in quick succession.
• iOS: Answers appear as they're written on every connection: Some providers delivered the whole answer at once instead of word by word, because the list of which ones could stream had gone out of date.
• Windows: Shell commands run instead of hanging: A command sent to the workspace shell waited a full minute and then timed out, because it was being written in the wrong shell's language.
• Common: Long answers with reasoning no longer fail on Anthropic: A conversation that included the model's own reasoning could be rejected outright when it was sent back.

Bug Fixes

• Common: The assistant no longer breaks when it runs several tools at once: When a model asked for multiple tools in the same step, the streamed pieces of those requests were merged into one another — so the calls arrived garbled or ran with the wrong inputs. Each request is now tracked separately. If a request still arrives unreadable, the assistant is told so and can retry, instead of silently running the tool with no input at all.
• Common: Some tools broke every message on a direct Anthropic connection: Enabling a tool that has multiple actions — Workspace Files, Computer Use — made Anthropic reject every message in the conversation, not just uses of that tool. Their descriptions are now sent in the shape Anthropic requires.
• Common: Removing a folder asks first: The folders menu mixes two things — which folders this conversation can see (a checkbox) and removing a folder from Caiioo entirely. The second is not undoable, so it now confirms before removing.

New Features

• Desktop & Extension: Work with folders on your computer, without copying them in: Caiioo can now read, search, and write files directly in folders you designate — on macOS, Windows, Linux, and in the browser extension. Add one from the + menu next to the message box (Add folder…), and pick which folders each conversation can see. Files stay where they are: nothing is copied into Caiioo's file storage, so you don't end up with two versions of the same document drifting apart.
• Desktop & Extension: Point at a file without attaching it: + → Browse folder… lets you pick a file from one of your folders and drop it into the message as a reference. Caiioo opens it from disk when it needs it, so a large document costs you nothing until it's actually read.
• Windows: Shell commands, safely confined: The workspace shell now runs on Windows as well as macOS and Linux. Commands run in PowerShell inside a Windows AppContainer that can reach only the folder you pointed Caiioo at — not the rest of your disk, and not the network unless you turn that on for that folder.
• Windows & Linux: Dictate into any app: A global hotkey starts dictation anywhere in the system and types what you say into whichever app has focus — not just into Caiioo. Matches what the Mac app already did.

Improvements

• Desktop: Choose folders with the normal folder dialog: Adding a workspace folder used to mean typing an absolute path into a text box. macOS, Windows, and Linux now open the system folder chooser.
• Common: Workspace Files is available to everyone: It was previously hidden behind a tester-only add-on. It's now a standard part of Caiioo wherever there's a filesystem to reach (it stays unavailable on iOS and Android, which have no such access).
• Common: Folder syncing has been replaced by folder access: Syncing a folder used to import a copy of every file into Caiioo. Working with the real folder does the same job without the duplication, so syncing is retired. Existing synced folders are carried over automatically where possible, and the files already imported are left alone — delete them from the File Manager whenever you like. The one-off Import folder button is unchanged if you still want a copy.
• Windows & Linux: "Check for Updates…" now tells you what happened: The menu item quietly did nothing visible — it found updates but had no window to show them in. It now confirms when you're up to date, shows what's in the update, and asks before installing, like the Mac app.
• Windows & Linux: The app is called Caiioo again: The installer, Start menu, uninstall entry, window titles and tray all showed the name in lowercase.

Bug Fixes

• Windows: Only one tray icon: Two icons appeared in the system tray, and only one of them had a working menu.
• Windows & Linux: A sidepanel that opened off-screen can be recovered: If the window's saved position pointed at a monitor you no longer have — unplugged, or a resolution change — it opened where you couldn't reach it, and stayed there on every relaunch. It now returns to a visible screen.
• Windows & Linux: A panel that fails to load says why: A start-up failure left an empty window with nothing in the logs. You now get a readable error on screen, and the details are written to the app's log file.
• Common: Downloading local AI models works again: Model downloads failed with "Failed to fetch" after HuggingFace moved to new download servers that our security rules didn't recognize.
• Common: Grok and CometAPI are actually selectable: Both were announced as available to everyone, but the buttons to choose them were still hidden outside the tester program.
• Common: Signing out clears your Private Sync passphrase: The passphrase stayed on the device after sign-out. It's now cleared; your sync settings are kept, so signing back in just asks you to unlock.
• Common: Renaming a tool action no longer silently switches it back on: An action you had turned off could quietly re-enable itself after an update that renamed it internally.
• Windows & Linux: Google Workspace icons show up: The Gmail, Calendar and Drive icons in the credentials screen were missing from the installer.

New Features

• Common: Use Grok with your Grok account — no API key needed: xAI's Grok models are now available as a direct provider, and you can connect them either with an xAI API key or by simply signing in with your Grok account (Settings → AI Providers → xAI). Unlike some providers whose terms restrict account sign-in to their own apps, xAI's terms allow it — so this is available to everyone.

Improvements

• Common: Scheduled tasks now run on one device instead of every device: A task synced across your devices used to fire on each of them — three devices meant three runs, three threads, and three notifications. New tasks are now pinned to the device that created them, and a pin button in Settings → Scheduled Tasks moves a task to whichever device you want running it. If that device is offline when a run comes due, the run is skipped rather than fired late, and the task shows a "Missed" badge with the time it was due. Tasks created before this change keep their old run-everywhere behavior until you pin them.
• Common: Editing a scheduled task on one device no longer wipes its run history elsewhere: A task's run count, last-run status, and next-run time are now kept per device instead of being overwritten by whichever device synced last.
• macOS & Windows/Linux: The "No browser connected" banner can now be dismissed: The notice asking you to connect the Caiioo browser extension stayed pinned above the conversation even if you don't use browser control. It now has a ✕ to dismiss it; it returns only if a browser connects and later disconnects again.

Bug Fixes

• Common: Reasoning is visible again on the newest Claude models: On a direct Anthropic connection, the newest Claude models (Sonnet 5 and newer) produced answers with no visible thinking at all — the app wasn't asking for it in the form these models require. Their reasoning now streams and displays as before.
• Common: The "What's New" popup is no longer blank after an update: Right after updating, the release-notes popup could show nothing — the app was reading notes saved by the previous version, which of course had no entry for the new one. Notes are now always fetched for the version you're actually running.
• Common: Poe models keep their capabilities on desktop: Reasoning settings and image support for Poe models weren't applied when the assistant ran in the app's background process (macOS, Windows and Linux) — the model information now travels with each request, so Poe models behave the same everywhere.
• Common: Switching AI providers mid-conversation picks a valid model: Changing providers could quietly keep the previous provider's model selected, which the new provider then rejected. The composer now selects a model the active provider actually offers.
• Extension: Direct Anthropic connections work for accounts with strict CORS settings: Some Anthropic accounts reject any browser-originated request outright; the extension's requests carried a browser origin marker even though extensions are exempt from those rules. It's now removed, so those accounts connect normally.
• Common: Settings panels no longer overflow at narrow widths: Several settings sections (connections, sync, credentials, and more) pushed content off the edge when the side panel was narrow. They now fit.

Improvements

• Common: Deleting your data is now one clear, honest flow: Settings → Data Management had accumulated overlapping buttons — "Clear Server Cache", "Unlinked Conversations", "Erase Server Data", "Erase ALL" — that had been failing silently for months and used six different words for deleting things. They're gone. What remains is routine housekeeping (trash, retention, orphaned files) plus a single "Purge All Storage" action that now tells you everything it deletes — including settings, modes, and memory — and warns you when cloud sync is on that synced data will download again unless you also use "Delete All Cloud Data" in Cloud Sync settings.
• Common: "Keep API keys" during a purge now keeps all of them: Purging with "keep credentials" checked quietly dropped some credentials depending on platform — the extension lost Claude/Grok account sign-ins and newer provider keys (Perplexity, Baseten, Cloudflare, CometAPI), while desktop/mobile lost the Cartesia voice key. Every credential now survives on every platform, and a Cloudflare key you saved actually persists at all (it previously never saved anywhere).
• macOS & Windows/Linux: Storage usage is back in the menu bar / tray: The storage numbers shown by the macOS menu bar and the Windows/Linux tray had been silently failing to load since March. They work again.
• iPhone, iPad, Android, macOS & Windows/Linux: Automatic data retention cleanup actually runs: The "Data Retention" auto-cleanup schedule only ever ran in the browser extension — everywhere else it ran only when you pressed "Run now". Every platform now runs the daily cleanup.
• iPhone, iPad, Android & macOS: Data management results are visible again: Success and failure messages in Data Management used browser alert popups, which native apps never display. They now use the app's own dialogs, so you can see what happened.

Bug Fixes

• Common: The answer no longer appears twice: When the assistant's final answer restated something it had already written earlier in the same reply, both copies were shown. The duplicate is now dropped.
• Common: Your assistant no longer calls itself by the app's old name: Some longtime users — especially anyone who had customized a built-in persona before the rename — still got replies from an assistant introducing itself by the old brand, and "Reset to original" appeared to do nothing. All the places the old text could hide are now corrected, without touching anything you wrote yourself, and reset works again.
• Common: A rare message failure with Claude models is fixed: Certain replies could fail immediately because the app marked an empty piece of the conversation for caching, which Claude's provider rejects. It no longer does.

Bug Fixes

• Common: Deleting something now deletes it on your other devices: With cloud sync on, deleting a conversation (or other synced items) removed it only on the device you did it on — the deletion was never carried across, so the item quietly came back everywhere else. Deletions now propagate like every other change, and a build-time guard keeps this class of bug from shipping again.

New Features

• Common (early testing): The assistant can fan out tool calls from a Python script: For work like "read these 50 pages and pull out the pricing," the assistant previously made each call one at a time, writing out every request and reading back every full result. It can now write a short Python script that makes those calls concurrently, keeps the full results in its private workspace, and brings only the parts that matter into the conversation — faster, and far less clutter. Tools that need your approval can never be called from a script; the assistant still has to ask you directly, exactly as before. Available to testers on desktop and browser extensions.
• iPhone, iPad & Android: Continue an interrupted reply: When the system closed the app mid-reply, the half-written answer was quietly removed — the turn read as if it had vanished. An interrupted reply now shows a clear notice in its place, keeps the results of the steps that finished, and offers a Continue button to pick up where it left off — with guardrails so a reply that crashes the app on every launch can't retry forever.

Bug Fixes

• Common: Very long conversations no longer fail after trimming: When a conversation grew past what the model can read, trimming could separate a tool request from its result — which providers reject outright — and some unrelated errors were mistaken for an oversized conversation and "recovered" by discarding your history. Trimming now keeps requests and results together, unrelated errors no longer wipe anything, and the assistant is told exactly what was removed so it can re-fetch what it still needs instead of answering around a gap it can't see.
• Common: Granting one more Google permission no longer drops the ones you already had: Connecting an additional Google capability (say, Calendar after Gmail) could quietly revoke the access you'd granted earlier, so features that had been working started failing until you reconnected. New grants now add to the existing ones.
• Common: Your assistant no longer introduces itself by the app's old name: Longtime users who set up before the rename — including anyone who had customized a built-in persona back then — could still get replies from an assistant calling itself by the old brand. The built-in personas are now corrected everywhere they're stored, without touching anything you wrote yourself.
• Common (testers): Claude account sign-in is steadier: A temporary server hiccup during sign-in no longer flags your Claude account as needing reconnection — only a real rejection does — and an expired session now refreshes itself when the app lists your models instead of failing.

New Features

• iPhone, iPad & Android: Connect tool servers that ask you to sign in: Adding an MCP server that requires signing in — Notion, Linear and the like — has never worked on phones or tablets. The sign-in page opened in a window that couldn't hand your approval back, so the connection dead-ended with no explanation. Signing in now opens in Safari on iPhone and iPad and in a browser tab on Android, so you're already signed in to the service, and the approval is handed back to the app to finish the connection. If it does fail, the app now tells you why instead of leaving the dialog looking stuck.

Improvements

• iPhone, iPad, Android & Windows/Linux: Faster startup with an immediate loading indicator: The app showed a blank screen while it finished starting up. It now shows a loading spinner right away, and loads your saved conversation memory in parallel, so cold starts feel quicker.
• iPhone, iPad & Android: Startup no longer slows down as your history grows: The app loaded every message of every conversation before it could show you anything, and re-saved that whole pile each time a reply streamed in — so the more history you had, the longer the opening spinner sat there. Conversations now load one at a time, only when you open them, so startup and saving stay quick no matter how much history you've built up.
• iPhone, iPad & Android: A reply interrupted by the system is no longer left stuck: If your phone closed the app while the assistant was still working — a long lock, or switching away for a while — the whole turn's work was thrown away and you came back to a half-written reply that never finished. The app now saves its progress after each step, so at most the step in flight is lost, and it clears the stranded reply when you reopen it instead of leaving it hanging.
• Common: Sign-in errors say what actually went wrong: Every failed sign-in showed the same "Sign in was cancelled or failed" message regardless of the cause. The screen now shows the real reason, so a sign-in problem is something you can act on.
• Common: Settings tells you when you're using the key included with your account: An empty OpenRouter key field looked like something was missing even though paid models were working fine — you were running on the key included with Caiioo, which stays on the device you set it up on. Settings now says so instead of leaving an apparently blank required field.

Bug Fixes

• Common: Attached images and PDFs reach Claude and Gemini: On a direct Anthropic connection, images and PDFs you attached were dropped before the message was sent — so the model truthfully told you it couldn't see them. PDFs were dropped the same way on a direct Google connection. Both now arrive intact.
• Common: Follow-up messages to Claude's thinking models work: On a direct Anthropic connection, the second message in any conversation with a thinking-capable Claude model failed outright. The app now replays the model's reasoning in the form it expects.
• Common: A brief network hiccup no longer kills a reply: Replies occasionally died with a bare "fetch failed" when the connection to the model provider timed out while it was still being set up. The app now retries twice before giving up, and reports the actual network reason when it does.
• Common: Self Checker works with providers other than OpenRouter: Picking Anthropic, Google or OpenAI as the reviewer and leaving the model blank sent an OpenRouter-style model name to that provider, which answered "not found". Leaving it blank now picks a model that suits the provider you chose.
• Common: Switching away from OpenRouter no longer fails with a confusing provider error: If your default model was set while you were on OpenRouter's free tier, switching to another provider could still send OpenRouter's internal free-tier placeholder to it, and the rejection read like a provider outage. The app now catches this before sending and tells you to pick a model for the provider you're on.
• Common: Scheduled tasks keep using the model you picked: A scheduled task ran with whatever your current default model happened to be when it fired, so changing your default later — or a mismatch between your provider and model — could break the task with an "invalid model" error. Each task now remembers the model and provider you had when you created it and runs with those. If that model is ever unavailable, it falls back to your current default and tells you it did.
• Common: Newer Claude models (like Sonnet 5) work again: Sending a message to a newer Claude model failed with a "temperature is deprecated for this model" error, because the app always attached a creativity (temperature) setting these models no longer accept. The app now only sends creativity to models that support it — checked automatically per model — so newer models work again. The creativity slider is also disabled, with a short note, when the selected model manages its own creativity.
• Windows: The app starts reliably instead of failing to connect: On some machines the app's background service crashed the moment it started, leaving you with a "cannot connect" error and no side panel. It was picking up a stray copy of a shared component that happened to sit in your user folder. It now loads its components from its own install folder, where nothing else can shadow them.
• Windows: Signing in with Google, Microsoft or Apple works again: The sign-in page opened with most of its web address missing, so Google answered "Required parameter is missing: response_type" and Microsoft refused with an AADSTS900144 error. The app was handing the address to the Windows command prompt, which cut it off at the first "&". It now opens the page directly.
• Windows/Linux: Signing in no longer fails with a connection error: Even once the sign-in page opened correctly, email sign-in failed with "Failed to fetch" and Google, Microsoft and Apple sign-in failed right after you approved it — the app was blocking its own connections to Caiioo.ai, because its built-in security rules still listed only the old domain. Those rules are now derived from the address the app actually uses.
• iPhone, iPad & Android: Scheduled tasks and folders no longer disappear after you restart the app: Your scheduled tasks — and your file folders — could show up empty after you quit and reopened the app, only reappearing once Private Sync caught up. With sync turned off they could be lost for good on the next save. The app now reloads them from your device on startup, so they're there immediately.
• iPhone, iPad & macOS: Deleting a scheduled task works again: Tapping delete on a scheduled task did nothing at all — it was waiting on a confirmation prompt that never appeared, so the delete silently gave up. That section now uses the app's own confirmation dialogs, so deleting a task (and every other prompt there) works.
• iPhone & iPad: The app remembers you're signed in after a full restart: After quitting the app completely it sometimes reopened to the login screen — and signing in there didn't take; only force-quitting and reopening again worked. Your identity wasn't being restored from where it's actually saved on a cold start. It is now.
• iPhone & iPad: Signing in works without needing to restart the app: After signing in with an email and password or a license key, the app could fail your first message (or seem unable to reach the server) until you force-quit and reopened it, because your included Caiioo key wasn't set up during that kind of sign-in. It's now set up right after you sign in, so you can start using the app immediately.
• iPhone & iPad: Private Sync reconnects as soon as you sign in: If you already had Private Sync set up and opened the app signed out, sync stayed offline after you signed in until you restarted — so anything waiting on your other devices, like a saved provider key, never arrived. Sync now reconnects right after you sign in.

Improvements

• Common: Your location is turned into a place name entirely on your device: The app used to send your coordinates to a third-party service to name your city. It now does that lookup on-device from a built-in database, so your coordinates aren't shared with anyone. A new "Precise" option shares full GPS accuracy (neighborhood and coordinates) when you want it, and an optional network-based estimate — off by default — can fill in when GPS isn't available.

Bug Fixes

• Common: Local-network Private Sync no longer misses changes after the desktop app restarts: When syncing over your local network (or a shared folder), changes another device made just before you restarted the desktop app could be skipped — the app trusted an empty "nothing changed" answer and didn't reconcile until some later, unrelated change forced a full sync. After a restart the app now recognizes that its recent-change history is incomplete and does a full reconciliation, so changes from your other devices come through. A related fix keeps a sync lock held across a restart from being handed out twice.
• Common: Private Sync no longer asks for your key on every app restart: Your sync key was remembered only for the current session, so restarting the app (or your browser, for the extension) locked sync until you typed the key again. The key is now stored on your device — encrypted at rest — the first time you enter it, and sync resumes automatically on launch. You'll be asked one final time after updating; after that it sticks. Entering the key on each new device is still required, and turning sync off (or signing out) removes the stored key.
• Common: Browser automation works again when the assistant drives a Chrome or Edge browser from the desktop or web app: Typing into pages, pressing keys, and examining a page's controls failed with an internal "Unknown message type" error — the in-page helper wasn't set up to run the actions the app asked it to. This hit the macOS and Windows/Linux desktop apps and the web app whenever the assistant controlled a Chrome or Edge browser (the browser extension used on its own was unaffected). Those actions now run correctly.
• Common: Models that can hear audio now listen to your recording directly: When you picked a model that can process audio itself (like Gemini), your recording was still handed off to a separate helper model for transcription instead of letting your chosen model listen firsthand — losing the direct connection to your prompt. This happened everywhere except the browser extension, because only the extension checked whether the active model could hear. The app now checks the active model's real audio ability on every platform (iPhone, iPad, desktop) and sends the recording straight to it when it can listen; the helper model is only used for models that genuinely can't process audio.
• Common: The "speak" action works even without a voice provider set up: Asking the assistant to speak saves an audio clip, which needs a voice that produces sound files. If your text-to-speech was set to your browser's built-in voice (or off), that produced nothing and the action failed. It now falls back to an on-device voice so it still creates the clip.
• Common: Reading text from PDFs and images works with more providers: Extracting text no longer requires an OpenRouter key when you've set up a different AI provider for it; the app now uses whichever provider you configured.

Bug Fixes

• Common: A brand-new sign-in no longer fails its first message: On a fresh install, sending your first message before your included Caiioo key finished setting up in the background failed with an error that every provider had been skipped — and only recovered once you opened Settings and tapped "Use my Caiioo". The app now finishes setting up that included key on the spot for the first message, so it just works out of the box.
• Common: "Reset to original" on a mode's instructions actually restores them: If you had customized a built-in mode's instructions, "Reset to original" didn't take — it snapped straight back to your custom text (and on iPhone and iPad it did nothing at all). Resetting now clears your customization and restores the mode's original instructions.
• Common: Tables in web pages come through as readable tables: When the assistant read a web page, any table arrived as one run-together line with every cell mashed together (e.g. dates, amounts, and labels with no separation). Tables are now converted into proper rows and columns, so the assistant — and any document it saves — can actually read them. This applies both to normal pages and to restricted pages the assistant reads a different way.

Improvements

• Common: Default helper models upgraded to Gemini 3.5 Flash Lite: The built-in cheap helper models — the vision fallback that captions images for text-only models, the audio-analysis helper, the default sub-agent model, and the Self Checker's judge — now default to Gemini 3.5 Flash Lite (was 3.1 Flash Lite) for better quality at similar cost. Any model you've explicitly chosen in Settings is untouched.
• Common: Clearer guidance when the assistant misuses a multi-action tool: When the assistant called a multi-action tool (like Workspace Files) with a wrong or missing parameter, the error said only "Schema details unavailable", leaving it to guess the correct form. The error now spells out the exact parameters for the action it attempted, so it corrects itself on the first retry.

Bug Fixes

• Extension: Pasting a copied image into the composer works again: Right-clicking an image and choosing "Copy image", then pasting into the message box, did nothing — no attachment, no text, no error. The clipboard carried a hidden markup wrapper alongside the image, which made the composer treat it as a text paste with empty text and swallow it. Image pastes now attach the image whenever there's no actual text being pasted; copies from Word or Google Docs still paste as text.
• Common: The assistant no longer calls itself "Pebble Flow" on long-standing installs: If you had ever customized the default agent before the Caiioo rebrand, your saved copy of it — old name and old instructions included — silently overrode the current one on every update since. The app now recognizes those pre-rebrand leftovers and shows the current name and instructions, without touching modes you created and named yourself.
• Common: Fixed request rejections from stricter models (e.g. Kimi K3): Some models rejected our requests outright ("not a valid … json schema") because a few tool definitions contained internal schema shortcuts that not every provider understands. Tool definitions are now fully spelled out for every model.
• macOS: Opening a document you pasted from another conversation now works: Pasting a document's link that was created in a different conversation and asking the assistant to open it failed with the document not being found — even though clicking the same link opened it fine. The assistant now finds the pasted document no matter which conversation it originally came from. (The earlier fix covered links pasted within the same conversation; this extends it across conversations.)
• macOS: A disconnected or deleted workspace folder no longer breaks the rest of your workspace: If a folder you'd added to Workspace Files was deleted, renamed, or lived on an external or network drive that disconnected, every file action failed — including reading files in your other, still-available folders. A missing folder is now simply skipped so the rest of your workspace keeps working.
• Extension: Adding a local folder to Workspace Files now works via the folder picker: In the browser extension, adding a local folder by typing a path did nothing — the extension can't open a folder from a typed path the way the desktop apps can. Adding a folder now opens your browser's own folder picker and remembers the folder you choose. (On the desktop apps, typing a folder path still works as before.)

Bug Fixes

• macOS: The voice hotkey can no longer leave the microphone running: Releasing the push-to-talk key while the speech model was still loading could start a recording that nothing would ever stop — the microphone stayed hot until the app restarted. The start and stop are now strictly ordered, clicking the mic button reliably stops a hotkey-started recording, and closing the voice-model download dialog without choosing no longer switches the microphone on by itself.
• macOS: Workspace Files (file_workspace) actually works: Every file_workspace call on the desktop app failed with "WorkspaceProvider not initialized" — the agent runs in a background worker process that was never wired up with the workspace runtime, and the mounts you configured in Settings never reached it. The worker now receives your configured folder mounts and initializes the workspace on every turn, so read/write/search/glob/grep (and bash, on writable mounts) work as configured.
• Common: Folder mounts reject pasted URLs and relative paths: The "Folder path" field in Workspace Files settings accepted any text — a pasted link (e.g. a Slack URL) was saved silently as a "local folder" and simply never worked. Adding a mount now requires an absolute folder path and tells you immediately what's wrong instead of storing a broken mount.
• macOS/iOS: The trash icon on a folder mount removes it again: Removing a mount asked for confirmation via a browser-style popup that the app's embedded web view silently suppresses, so clicking the trash icon did nothing. It now uses the app's own confirmation dialog.
• Common: The assistant no longer sees tools it can't call: Opt-in tools that unlock via a Community Hub install (like Computer Use) appeared in the assistant's tool listing even when the package wasn't installed — the assistant would try to use one and hit "Tool not found". The listing now applies the same install gate as the tool runtime, so what's listed and what's callable agree.

New Features

• Common: Ask follow-up questions about any recording: When a recording is analyzed by the dedicated audio helper model, the assistant can now go back and re-listen with a targeted question — "which words were mispronounced?", "what's the tone of voice?" — via a new hearing tool, on the turn you attach it or any later one. Previously the assistant only ever saw a one-shot summary and would tell you it "cannot re-hear the audio".

Bug Fixes

• Common: Retrying a response no longer loses your uploaded photo: On the phone and tablet apps especially, hitting retry (or editing your message) regenerated the answer without the image you attached — the model would insist there was no photo, even though the first response saw it fine. The retried turn now carries the original message's attachments on every platform. (The desktop relay already re-gathered them itself, which is why this only bit phone/tablet apps.)
• Common: Recordings sent alongside transcribed speech no longer arrive as half a second of silence: On iPhone and iPad especially, with Audio Analysis mode on (used by the "for Languages" coach and the waveform composer button), the recording attached with your transcribed speech captured only a fraction of a second of hum or a beep — the analysis model then reported "no speech detected" even though your words were transcribed fine. The capture now starts reliably on iPhone/iPad and recovers automatically when the speech-recognition engine takes over the microphone session mid-recording, and the clip is packaged identically to a normal voice-memo recording.
• Common: Answers no longer disappear when the assistant keeps working after writing them: When the assistant wrote its answer and then made a few more tool calls (say, rendering a dashboard) before signing off with a short closing line, only the closing line was shown, read aloud, and copied — the real answer was buried inside the collapsed activity timeline. The full response is now recognized and kept, on screen and in copy/share/read-aloud.
• macOS: Fixed a crash when listing microphones (shipped early to macOS as 0.9.729a): Opening the voice input device picker could crash the whole app for setups where an audio device exposes more than one input stream — aggregate devices, some USB audio interfaces, or a Bluetooth headset mid-handoff. Reading such a device's channel layout wrote past the end of a fixed-size buffer and corrupted memory, which took the app down moments later. Device enumeration now sizes that buffer correctly.
• Common: Voice replies no longer go silent after a call, Siri, or an alarm (shipped early to macOS as 0.9.729a): On iPhone especially, read-aloud (including the on-device Kokoro voice) could stop producing sound — the speaker button showed "playing" but nothing played — after anything interrupted the phone's audio (a call, Siri, an alarm, another app taking over sound), and it stayed broken until the app restarted. Playback now recovers from that interrupted state automatically, and if the sound system genuinely can't be reclaimed you get a clear error instead of endless silence.

New Features

• iOS: Record audio from the attachment menu: The "Record audio" option in the composer's "+" menu — record a clip and attach it to your message — is now available on iPhone and iPad, matching the desktop app and browser extension. Pairs naturally with apps that listen to your recordings, like the "for Languages" pronunciation coach.

Bug Fixes

• Common: Test Runner reports can be exported after an app restart: A test run's results lived only in memory, so on iPhone especially — where the system quietly restarts the app's web view under memory pressure — asking for the CSV, transcript, or full report after a run answered "Run not found" and the results were gone for good. Runs are now saved as they progress: each test's result is kept the moment it completes, so even a run interrupted mid-way keeps everything it finished, and exports, status checks, and manual pass/fail verdicts all work after a restart on every platform.
• iOS/Android: Heavy activity no longer delays saving your threads indefinitely: The app batches disk writes to stay fast, but a steady stream of updates — several test-runner threads finishing together, or long busy responses — could keep postponing the write forever, so a sudden app termination lost everything since the last quiet moment. Work in progress now reaches disk at least every few seconds no matter how busy the app is.
• iOS/Android: Installing a hub app from caiioo.ai now works when the app has to start up: Tapping "Install" on the caiioo.ai hub hands the install to the app via a link — but if the app wasn't already running (or the system had quietly evicted it in the background), the hand-off arrived before the app finished starting and the install was silently lost: the app opened and nothing happened. The install now waits for startup to complete, and the app opens straight to Settings → Community Hub to run it.
• Web: The hub shows the right install button on phones: On iPhone and Android, the caiioo.ai hub offered "Browser extension" (which can't work on a phone and did nothing) and "Desktop app" (mislabeled — the link actually opens the mobile app). Phones now get a single "Install in the Caiioo app" button.

New Features

• Common: CometAPI is now a bring-your-own-key provider: Access 500+ models — GPT, Claude, Gemini, DeepSeek, and more — through a single OpenAI-compatible CometAPI key, added in Settings → AI Setup. The model picker shows benchmark-tier medals for its catalog and supports search and the Best / Value / ZDR quick filters, total CometAPI usage is tracked in Settings, and the internal helpers (thread titles, suggested replies, greetings) work with a CometAPI key too. When a task needs a capability your provider doesn't serve — web search, video or music generation, certain image models — the request routes through OpenRouter instead of failing with a confusing error: only when you have an OpenRouter key and have approved OpenRouter as a provider (with a clear message otherwise), so your prompt never goes to a provider you haven't OK'd. CometAPI usage costs are counted in the per-message and thread totals, and the model picker's CometAPI catalog is cleaner — correct prices for look-alike model names, and chat models that were wrongly hidden as "generators" now listed.

Improvements

• Common: MCP server cards can show the full error: A failing MCP server's error was clamped to four lines with only a hover tooltip, making long errors (chained connection failures, whole HTML error pages) unreadable and impossible to copy. A "Show full error" toggle now expands it into a scrollable box you can select and copy from.
• Common: Benchmarks run through the real app and show live progress: Test Runner benchmark and comparison runs now execute each turn through the same pipeline as your actual messages, so results reflect what the app really does rather than a parallel test harness. A floating panel shows the live run — suite name, per-test pass/fail, progress — with a cancel button, and failing tests produce a readable transcript of exactly what each tool was asked and returned.
• Common: Analytics choices take effect immediately: Turning the optional anonymous usage analytics on or off now applies at that moment. Previously, opting in mid-session could silently discard signals until the next scheduled send — up to a day later on the desktop app — and reliability signals could occasionally be delivered twice; sends are now confirmed exactly once.
• macOS: Window tiling tells you when it lacks permission: Tiling another app's window needs the Accessibility permission, and a missing or stale grant (common after app updates) made the feature fail with no signal at all. You now get a clear explanation matched to the actual problem, with a shortcut straight to the right System Settings pane.

Security

• Common: Account switching keeps each profile's connections strictly separate: When two accounts were used on the same installation, signing in or out at the wrong moment could file one account's service connections — including a Google connection and its access — into the other account's profile, and the Settings panel could keep showing the previous account's connections after a switch. All three underlying causes are fixed, and on start-up the app now automatically removes any connection that had already ended up in the wrong profile (deliberately-added secondary accounts are untouched).
• Common: Dependency security sweep: Cleared the one critical and all fixable high- and medium-severity advisories across the app's dependencies — including the libraries that process AI-generated text and HTML, which carried the highest real exposure. The few remaining advisories have no released fix yet and are tracked with active monitoring.

Bug Fixes

• iOS/Android: Locking the phone or switching apps no longer kills a running response: Locking the phone or hopping to another app while the assistant was working dropped the connection — the run ended with a raw "The network connection was lost." error and the answer never arrived. Short app-switches now usually continue uninterrupted, and when the connection does drop, the app automatically retries the interrupted request once you're back (an amber "Connection interrupted, retrying..." banner shows progress). If the network genuinely can't be reached after several attempts, you get a clear "Network connection lost" message instead of the raw error.
• macOS: Connected services no longer falsely expire on the desktop app: Background sign-in refreshes from the desktop app were sent without credentials and got rejected, so a healthy Google connection could be marked "needs re-authorization" within an hour of granting it. Refreshes now authenticate correctly and connections stay connected.
• Common: MCP servers with unusual sign-in replies now open the authorization flow: Some MCP servers answer an unauthenticated connection in a nonstandard way, and Caiioo surfaced a raw "Missing authorization token" error instead of opening the browser sign-in. Those servers now trigger the normal authorization flow.
• Common: The Agenda button's upcoming-reminders count is no longer cut off: The small number badge on the Agenda button in the message composer sat half-hidden under the edge of the text-entry box — most visible on iPhone. The full count now shows.
• macOS: Answers no longer vanish at the end of tool-heavy requests: On the desktop app, a request that used several tools could do all the work — the tools visibly ran and returned real data — and then never show the final answer; the run ended looking failed. The internal hand-off that delivered the finished answer could be cut off at the last moment on large responses. The answer now always arrives intact.
• Common: Deleted MCP servers stay deleted: Removing an MCP server marked it deleted without ever cleaning it up, so long-gone servers could reappear in your list ("a bunch of random MCPs are back"). Old deletions are now cleaned up for good, while active servers and recent deletes are untouched — and a sign-in left behind by a deleted server no longer triggers repeated authorization errors at start-up (only provably-orphaned connections are removed, and reconnecting restores one if you re-add the server).
• Common: Spreadsheet and document tool errors are no longer hidden as successes: When a Google Sheets or Docs action couldn't run — a missing required detail, an unknown action, an API failure — the error came back looking like a successful result, so the assistant could repeat the same broken call until it hit its step limit instead of correcting course. These now register as real errors the assistant reacts to.
• Common: Copied diagnostics never lose tool activity: When a debug transcript had to be shortened to fit size limits, the cut could remove tool calls and their results from the middle — making it look like a tool never ran or returned nothing, and pointing an investigation at the wrong cause. Shortening now trims reasoning text and long payloads instead; every tool call, its status, and any error always survive.
• Common: A tool that couldn't get authorization no longer shows a green check: When a tool's sign-in/authorization failed, the step was still marked successful — a misleading green check on a tool that never ran. It now shows as the error it is.
• Common: Multi-model benchmarks judge with the run's own model and key: The Test Runner's scoring judge used your active profile's model settings instead of the benchmark's own configuration, so on a free-tier profile the judge could quietly fail and zero out every score. The judge now uses the model and key the run was configured with, a misconfigured judge is reported up front, and a judge that produces no scores at all stops the run with a clear error instead of reporting every test as failed.
• Common: Tool calls no longer get stuck in a retry loop with certain streaming models: With some models, the very last fragment of a streamed tool call could be dropped when the provider closed the connection abruptly, leaving the tool's instructions incomplete — the assistant then failed to run the tool and kept retrying. The final fragment is now always processed.

Improvements

• Extension: Connecting to the desktop app is now an explicit opt-in switch: Settings → Connections has a new "Enable Caiioo Bridge" switch controlling whether the extension connects to the Caiioo desktop app on the same computer. It's off by default — the extension stays fully standalone and never probes for a local app until you turn it on — and flipping it takes effect immediately, no restart needed. While it's off, the Reconnect and port controls are grayed out instead of appearing to work.

Bug Fixes

• Extension: Reconnecting to the desktop app now sticks: Clicking Reconnect in Settings → Connections opened a connection that quietly died as soon as the browser put the extension to sleep — usually within a minute — and never came back until you clicked again. A manual reconnect now stays alive like the automatic connection does.
• macOS/iOS: Fixed a crash when opening certain links: Opening a link whose address contains percent signs — most visibly clicking "grant scopes" to give Caiioo access to your Google account — could crash the app instantly. These links now open normally.
• macOS: Python, PDF text extraction, and SVG rendering now work in the installed app: The distributed macOS app was missing several background components that development builds included, so running Python, reading text out of PDFs, rendering SVG documents, and the smart tool picker's ranking model silently failed only in the installed app. They now ship in the app, and a build check keeps them from being dropped again.
• Common: Editing an uploaded text file no longer garbles special characters: When the assistant edited a text or Markdown file you uploaded, characters like smart quotes, em-dashes, and accented letters (é) were corrupted throughout the file — including in parts the edit never touched. Files now round-trip byte-clean, and rejecting tracked changes no longer corrupts the document either.
• Common: Pointing the assistant at a document by pasting its link now works: Pasting a document's Slate link into the chat ("this is the file I want to update: http://…/slate.html?id=…") now resolves to the right document instead of failing with "No artifact found".
• Common: Bug reports keep the useful part of the diagnostics: Support tickets attached the assistant's debug transcript but cut it off mid-startup-logs when it was long, losing the tool calls and errors that actually explain the problem. Long transcripts now drop routine log noise first and keep the diagnostic sections.
• Common: Reviewing tracked changes in documents is more dependable: Several review problems are fixed in one pass — a suggested change ending at a paragraph boundary could be invisible when the document opened (the toolbar counted it, but nothing was highlighted); an accepted change could come back after closing and reopening the document; deciding on one change could silently discard a decision already made on an identical edit elsewhere in the document; and each finished review was saved twice into version history, pushing older versions out twice as fast. The accepted/rejected counts shown during review are now accurate too.
• Common: A document's version history stays its own — and restores completely: Switching between documents at the wrong moment could append one document's content onto another document's version history, so a later restore could put the wrong document's text in place. Restoring a version now also brings back the tracked changes that version carried, edits that mostly delete text no longer lose their redlines from the saved snapshot, the highlighted "current" entry is now determined by actual content, version comparisons after old versions are pruned diff against the nearest surviving predecessor instead of claiming there's nothing to compare, and the history preview no longer renders raw HTML embedded in document text.

New Features

• Common: Sign in with Microsoft: You can now sign in to your Caiioo account with a Microsoft account — personal as well as work or school accounts — on every platform, alongside the existing Google, Apple, and email options. If no Caiioo account exists for that email yet, you're asked to confirm before one is created.
• Common: LaTeX documents show typeset math: Opening a LaTeX document (Slate) now shows a preview with the math properly typeset, while the source stays safely editable in the code view — the preview never rewrites your file. The assistant also checks the math whenever it edits a LaTeX document and sees any errors immediately, so a broken equation gets corrected in the same turn instead of slipping through. Pasted LaTeX content is now recognized automatically.
• Common: New "for Grant Writers" app in the Hub: A grant-lifecycle copilot for nonprofit grant writers, now available to everyone. It prospects live federal funding opportunities (Grants.gov) fit-scored against your organization's profile, researches funders through their public 990 filings (ProPublica Nonprofit Explorer) and federal award history (USAspending), drafts LOIs, need statements, and program sections aligned to a funder's stated priorities, builds line-item budgets with justification, and writes interim/final grant reports — all grounded in open, keyless data sources. It never invents deadlines, eligibility, or outcomes data, and it never submits an application on your behalf.

Improvements

• Common: Settings shows where a tool works instead of hiding it: Tools that don't run on your current device used to be either missing from Settings → Tools entirely or shown as if usable. They now appear grayed out with a note saying which platforms they work on — and a tool that just needs the desktop app running says "Connect the Caiioo desktop app to use this tool" — so you can discover what's available elsewhere instead of never knowing it exists.
• Common: The Self Checker can verify claims by running real Python: When you score an answer with the ⚖ button, the judge can now write a small Python program to re-verify things no simple check can — recomputing a claimed statistic from the source data, or re-deriving a transformation — run locally in the same no-network sandbox as everything else, on the browser extension and desktop apps. Any check that can't run cleanly reports as failed rather than silently passing.
• Common: Accuracy safeguards now always apply: The built-in rules that keep the assistant grounded — stick to real sources, say so when unsure, cite where facts came from — used to be part of the default personality, so a custom instruction or an app with its own personality quietly replaced them. They're now a permanent layer underneath: your custom instructions and app personalities add to them instead of switching them off, and they can't be edited or deleted from settings.
• Common: Older documents with tracked changes get the clean redline: Documents saved with tracked changes by earlier versions — including restoring an older version that carried tracked changes — now display and navigate their changes with the same clean grouping as fresh edits, instead of the coarser marks the old format produced. The upgrade sticks after saving, so the old-style marks don't come back the next time you open the document.

Bug Fixes

• Common: Web browsing works inside dialogs and popups: When the assistant deliberately opened a dialog on a page — a search box, a filter panel — its next click or keystroke could be refused with a warning that something was in the way, sometimes stalling the whole task. That safety check now warns instead of refusing, correctly identifies what's actually clickable, and no longer suggests pressing Escape when that would close the very dialog the assistant just opened.
• Common: Page errors during browsing are reported instead of vanishing: In the browser extension, an error inside the page during a browsing action was silently discarded and looked like an empty result, leaving the assistant guessing at what went wrong. The real error message now comes back, matching how the desktop app already behaved.
• Common: Edits on different lines stay separate tracked changes: Two edits on neighboring lines or list items could merge into one oversized tracked change spanning both, making review and accept/reject clumsy. A tracked change now never spans a line break — edits on different lines are always separate, individually navigable changes — while normal grouping within a line is unchanged.
• Common: Apps that call external services work in the phone and desktop apps: In the iOS, Android, and Windows/Linux apps, tools that call external web services (for example an app searching a public database) failed instantly with a technical error before the request was even sent. These requests now go through on every platform.
• Common: Reasoning models keep their train of thought between turns: With some reasoning models — most visibly Gemini via OpenRouter — the model's stored thinking from earlier turns was sometimes wrongly discarded when building the next request, losing continuity and wasting re-thinking time. Stored reasoning is now recognized and carried forward correctly, and reasoning from an incompatible model family is still safely excluded.
• Common: Tool descriptions display correctly in Settings: Several tools showed the wrong description or one cut off mid-sentence in Settings → Tools and on Hub tool cards — for example, the document tool described itself as "List existing slates in thread" and the messaging tool as only "List connected messaging services and their status". Every tool and action now shows its full, correct description.
• Common: Tools that can't run on your device are no longer offered to the assistant: Platform-specific tools (for example, ones that need the Mac desktop app) could still be handed to the assistant on devices that don't support them, leading to failed attempts and confusing errors. They're now filtered out consistently everywhere — and if you ask for one by name, the assistant tells you which platforms it works on (or that it just needs the desktop app running) instead of claiming the tool doesn't exist.

Bug Fixes

• Common: Free-model requests recover when a provider is briefly unavailable: When a free model's provider reported it was temporarily out of capacity, the request used to fail outright. It now automatically switches to another free model and continues, with a short note telling you the switch happened — so a passing hiccup no longer interrupts your turn.
• Common: Document search finds phrases across line breaks and spacing: Searching a document (Slate) for a multi-word phrase could come back empty when the words were separated by a line break, a non-breaking space, or extra spaces — even though the phrase was clearly there. Search now matches phrases regardless of how the spacing falls, and if your query looks like a pattern (for example "Incident|Breach") with no matches, it suggests turning on regular-expression search.
• Common: Rewritten passages produce clean tracked changes: Substantially rewriting a paragraph could generate hundreds of tiny, scattered tracked-change marks that were impossible to read. The redline now collapses a rewritten region into a single clean change, while genuinely separate edits elsewhere stay distinct.

New Features

• Common: OpenAI is now a bring-your-own-key voice provider: OpenAI's gpt-4o-mini-tts joins the per-mode voice picker for text-to-speech. Its delivery is steerable — ask for an accent, tone, or pace in natural language (e.g. "read this with a warm Irish accent") and the voice follows it. Add your OpenAI API key in Settings → Voice.
• Common: Audio and video attachments now show a real preview: A recorded or attached audio clip renders its actual waveform on the attachment chip instead of a generic file icon, and a video attachment shows a still frame taken from the clip — so you can tell clips apart at a glance. The preview is generated once and reused.
• Common: Optional anonymous usage analytics — off by default in the EEA, UK, and Switzerland: You can choose to share anonymous, content-free signals that help us find and fix reliability problems — for example whether a turn succeeded or hit a particular kind of error, plus your explicit 👍/👎 feedback on a reply. It never includes your conversations, prompts, identity, IP address, or any identifier that follows you between sessions; it's first-party only (no third-party trackers); and events are stored with nothing that ties them back to you. In the EEA, UK, and Switzerland it stays off until you turn it on; elsewhere it's on with a one-tap off switch. Change it anytime in Settings → Privacy & Diagnostics, and there's no charge, discount, or feature difference either way.

Improvements

• Common: Settings → Voice warns when a voice provider isn't low-latency: Gemini and OpenAI finish generating the whole reply before any audio plays, so the first sound can lag a few seconds on longer responses — unlike ElevenLabs, Cartesia, and Resemble, which start speaking as the audio arrives. Picking one of the slower providers now shows a short note under the voice selector explaining the trade-off and pointing to the faster options, so the delay is expected rather than confusing.
• Common: The audio-analysis toggle is now hidden unless an app needs it: The waveform button next to the mic — which attaches your actual recording so a model that can hear can review it alongside your prompt — is a niche control that isn't useful for most tasks. It's now hidden by default and only appears in modes or apps that explicitly opt into it, so it no longer clutters the composer. If a recording was previously set to be captured, that capture also pauses while the toggle is hidden, so nothing is recorded without a visible off-switch.

Bug Fixes

• Common: Installed Hub apps no longer send every tool on each turn: In apps that pre-configure which tools the assistant reaches for (for Medicine, for Legal, and similar), that configuration was being ignored, so every turn quietly sent the full tool list — making turns slower and more expensive. The configuration is now honored, so each turn includes only the relevant tools.
• Common: Built-in skills show their real names: A default skill could appear with a placeholder name like "Skill 1781…" instead of its proper title; skills now show readable names (for example "Summarize Page"), and any already created with a placeholder are corrected automatically.
• Extension: Connected services no longer disconnect early when used through your private relay: Refreshing a connected service's sign-in through your private relay could drop the session token and be rejected, marking the connection expired before its time. The sign-in now refreshes correctly.
• Common: Fixed a freeze on launch after a subscription lapsed: If your Pro subscription expired or was cancelled, the app could get stuck repeating the free-tier downgrade while it loaded — freezing the interface before it finished opening. The downgrade now completes cleanly in a single pass, so the app starts normally and keeps every free-tier feature.
• Extension: Panels open reliably right after a background update: When the extension updated itself in the background while a window stayed open, opening Settings (or another section that loads on demand) could fail because the open window still referenced the previous version's files. The app now detects this and refreshes once to pick up the new version automatically.

New Features

• Common: Cartesia is now a BYOK voice provider: Cartesia Sonic for text-to-speech and Cartesia Ink for speech-to-text are now in the per-mode voice picker, the first-use TTS / STT dialogs, and the live-captions path. Add your Cartesia API key in Settings → Voice and pick Cartesia wherever a voice provider can be chosen.
• Common: Edit any forked Hub app inside Caiioo: Settings → Tools, Modes & Apps gains a per-primitive editor for personal apps (forked Hub apps and apps you've saved). Cards, views, workflows, skills, modes, and variables all get dedicated editors with safe defaults, so a malformed primitive in a fork can no longer break the editor. The agent can also snapshot a useful session into a draft personal app via the new save_session_as_app tool — say "save this as an app" (or similar) and it persists into your personal-apps store for you to refine.
• Common: Edit any mode's system prompt — Caiioo forks it for you: You can now edit the system prompt (or the entire definition) of any mode, including built-in and Hub-installed ones. The first edit automatically forks the mode's app into a personal copy, so your changes stick across mode switches and sync to your other devices like any personal app. A "Forked from … — your copy" banner shows whenever your copy is active, with a one-click Reset that restores the original. Existing custom modes migrate into this system automatically.
• Common: Caiioo for Medicine (beta): A new Community Hub-installable app mirroring the Caiioo for Legal pattern — a clinician decision-support factory that bundles differential-diagnosis, drug-interaction, and SOAP-note skills together with their reference materials. One-click install, forkable like any Hub app.
• Common: Use Caiioo from any browser through your own private relay: When you're signed in and your desktop relay is running, opening caiioo.ai in a browser without the extension now serves the same sidepanel UI through your private relay — so you can use Caiioo from a Chromebook, a public computer, or a tablet while every tool call still routes through your own machine.
• Common: Settings panel reorganized: The advanced settings layout is now seven user-intent categories — Account, Personalization & Privacy, AI Setup, Tools / Modes & Apps, Data & Sync, Connectivity, and Help — instead of the historical five. User Profile and Credentials Vault move into a dedicated Account category, LAN Relay / API Access / Messaging Gateway group together as Connectivity (the common thread is inbound/outbound network surfaces), Voice moves into AI Setup (it's AI behavior, not a tool), and Backup / Private Sync / Data Management form their own Data & Sync category. The simple-mode variant collapses the same controls into six categories with power-user controls hidden entirely.

Improvements

• Common: Document round-trip — second fidelity pass: Another sweep through Slate / DOCX / PDF / RTF / Google Docs closed a long list of round-trip findings. Highlights: generated tracked changes now emit Word-valid change ids and flatten nested change markers (so Word stops complaining about "corrupt" tracked changes on open); DOCX comments are deduplicated by id instead of double-counted; DOCX → Markdown export escapes literal | characters in table cells so the table doesn't dissolve; the RTF parser now consumes embedded binary segments (\binN) correctly so stray bytes no longer desync the rest of the document, skips \uc Unicode fallbacks properly, decodes \'hh hex via Windows-1252, preserves tracked revisions, and renders image placeholders; PDF ToUnicode CMaps decode surrogate pairs and bfrange array / multi-unit forms (non-BMP characters and emoji extract cleanly); track-changes prefix detection handles indented and nested list markers plus ) numbering; and the fidelity matrix now surfaces comment loss for docx → md/html/rtf and pdf → docx so you can see in advance what won't round-trip. RTF export also strips pending tracked deletions instead of inlining them as literal text.
• Common: Google Sheets formulas survive Slate sync: Editing a synced Google Sheet from Slate no longer rewrites formulas as their evaluated values. Tracked-change resolution in Slate is also corrected so accept / reject commits the right text when content already matches the target.
• Common: Google Docs sync ignores table-cell paragraphs: The sync scanner previously walked into table cells and produced wrong document offsets. Cells are now skipped before paragraph scanning, so edits land at the correct position in documents that contain tables.
• Common: Pseudonymizer protects machine-generated IDs end-to-end: Slate (and any tool that emits machine-generated identifiers) can now wrap them in a <!--no-pseudonymize--> marker that survives the entire pipeline — the pseudonymizer skips them, and the markers are stripped before they reach the screen. This closes the case where a tool-call's internal id got pseudonymized on one turn and dropped from the substitution map on the next, breaking follow-up tool calls that referenced it.
• Common: Self Checker now runs the judge on the server: Clicking the ⚖ button moves the verdict computation onto the relay instead of running it in the sidepanel, so the result completes even if you close the panel, returns faster, and stays consistent across devices. The judge call is also routed through the same thread-level pseudonymizer as the original turn, so a verdict on a pseudonymized message never leaks the real PII to the judge model.
• Common: Skill / Mode badges in the Settings inventory: Each installed skill in Settings → Tools, Modes & Apps now shows which mode(s) it belongs to, so you can see at a glance where the agent will actually reach for it.
• Common: Free-tier users on the provisioned key no longer 401 on web search: Search tools were resolving the OpenRouter key via a code path that bypassed the credential resolver, so users on the provisioned key (no BYOK) intermittently got 401s. The resolver now runs at every call site.
• Common: UI-context provider is noticeably snappier: The extension UI-context provider gained result caching, lazy mode loading, alias resolution, and tighter exclusions, so the model picker and mode picker don't stall on large workspaces.
• Common: Voice playback-speed slider now works for ElevenLabs and Cartesia: The speed slider had no effect with ElevenLabs or Cartesia — speech always played at normal rate regardless of the setting. Both providers now generate speech at your chosen speed. (ElevenLabs supports 0.7–1.2×, Cartesia 0.6–1.5×; the slider clamps to each provider's range. Resemble and Gemini don't offer a speed control and are unchanged.)
• Common: STT model loading is visible in the composer: When the on-device speech model is downloading or warming up, the composer now shows real download percentage, and concurrent load requests are serialized so two near-simultaneous "speak" actions don't kick off duplicate downloads.
• Common: Composer keeps the send button on-screen on narrow widths: The send button could previously clip out of the composer when the panel was very narrow. It now stays anchored regardless of width.
• Common: Remote sidepanel reconnects cleanly: For the new "Caiioo through a browser via your private relay" capability, the relay now addresses replies back to the remote browser correctly and primes it with an initial state snapshot on connect, so reconnects come up in the same state you left off.
• Common: UI-asset handlers echo request IDs: Responses now echo the originating request id back to the caller, fixing a class of stuck-spinner bugs where two asset fetches raced and the second result was discarded.
• Common: Caiioo for Legal is now just "for Legal": The "Caiioo for" prefix on the Legal app name was redundant once "for Medicine" landed without it. Existing installs continue to work via the legacy id.
• Common: Streaming reliability pass across every AI provider: A deep pass over how replies stream in from every provider. Reply fragments that arrived split across network packets could be dropped silently — breaking multi-turn extended thinking with Anthropic models, reasoning continuity with Gemini and GPT-5, and occasionally losing answer text outright with local MLX models — and non-English text or emoji could arrive corrupted into � characters in saved answers. Both are fixed everywhere. The Stop button now also cancels an in-flight Perplexity search instead of letting it finish (and bill) in the background, stopping a run now reaches any sub-agents still working, and images attached when chatting with Mistral models no longer get mangled in transit.
• Common: Skills now have proper names: Every skill carries an explicit display name, shown in Settings, the composer's skill picker, and to the agent itself — so skills whose prompts open with similar wording no longer collapse into indistinguishable rows. Typing "/" in the composer matches by name first, and publishing to the Hub now requires a name on every skill.
• Common: Hub-app modes keep their full configuration: Modes installed from the Community Hub (for Legal, for Medicine, …) could silently fall back to the general mode deep in the agent core, dropping their variables and tool configuration mid-run. They now resolve correctly everywhere, and a mode whose definition can't be resolved falls back to the standard Caiioo prompt instead of an empty one.
• Common: Provisioned-key (Caiioo-issued) account fixes: Re-issuing your Caiioo-provided AI key no longer wipes purchased credits — the remaining balance carries over to the new key. Revoking a key now actually revokes it with the provider before reporting success. Accounts with unlimited balances now display "Unlimited" instead of a number and no longer trigger automatic credit purchases. And a brief sign-in hiccup while restoring your key on a new device now retries instead of failing.
• Common: Free-tier model picks stay current: The free-tier model list no longer offers free models the provider has retired, which previously produced immediate errors when selected.
• Common: Video generation checks each model's real capabilities first: The video tool now reads each model's live capability sheet — supported aspect ratios, durations, resolutions, and whether it accepts reference frames — directly from the provider and validates your request before submitting, so an unsupported combination fails instantly with a clear message instead of after a long wait. Animating from reference images now sends them in a format every model accepts.
• Common: Pseudonymizer coverage — thread titles and helper calls: With the Pseudonymizer on, auto-generated conversation titles and the small internal AI helper calls now route through the same protection as your messages — on both the client and the relay — so a thread title can no longer carry a real name to the model. Sub-agent answers also display the real values on screen now instead of their substitute names.
• Common: GitHub sync handles non-English content and simultaneous edits: Files synced from GitHub containing accented or non-Latin characters no longer arrive garbled, and pushing a change to a file that moved on GitHub since your last sync now surfaces a conflict instead of silently overwriting the newer copy.
• Common: Transient server errors no longer sign you out: A temporary server error during the daily background sign-in refresh could clear your session and log you out. Only a genuine credential rejection signs you out now; anything transient keeps your session under a grace window.
• Common: Remote browser client boots cleanly: Opening caiioo.ai from a plain browser (through your private relay) no longer shows a long wall of connection errors while signing in, and after an update the UI always loads fresh instead of a stale cached copy.
• macOS, iOS, Android: Cold-start model prewarm: On-device voice and STT models now prewarm during app boot instead of lazy-loading on first use, hiding the multi-second first-call latency. The first tap of the mic now feels close to instant.
• iOS: App Store §3.1.1 sales-routing compliance: All external-payment surfaces (Stripe checkout buttons, "manage billing" links pointing off-app) are gated behind an iOS check, so the iOS client only ever offers in-app StoreKit purchases for digital subscriptions, matching Apple's anti-steering policy.
• iOS, macOS: Stale Safari extension registrations no longer freeze the relay path: When Apple's WebKit leaves multiple Safari extension registrations stale across app updates, Caiioo now evicts the superseded duplicates instead of freezing them dormant, so the relay-backed Safari extension keeps working without a manual reinstall.
• Android: Stale media permissions cleaned up: READ_MEDIA_IMAGES is gone from the manifest, and the legacy permissions implicitly added by the LiteRT GPU library are stripped, so the app's runtime permission prompt is now minimal — closer to what users actually consented to.

Security

• Common: Hardening sweep across imports, rendering, and logging: Crafted Word documents and conversation-import archives can no longer exhaust memory through decompression tricks (strict size caps and safer document parsing); a maliciously structured ChatGPT export can no longer hang the importer; three cross-site scripting risks in the document (Slate) page are closed; the server now refuses to fetch model-suggested URLs that point at internal or private network addresses; filter rules are rejected if their pattern could lock up the matcher; and a payment-webhook debug log no longer records secrets.

Bug Fixes

• Common: Voice provider API key appears in the right place: When a cloud provider was used only for speech-to-text (e.g. Cartesia Ink as your dictation engine), its API key field was stranded down in the Text-to-Speech (output) section instead of next to the speech-to-text picker — and stayed visible even after you switched your TTS voice to a different provider. Each provider's key field now renders under the selector that actually uses it (input vs. output), driven by a single voice-provider definition.
• Common: Auto-read now works with Cartesia and Gemini voices: The "read replies aloud" toggle silently never activated when your TTS voice was Cartesia or Google Gemini — both were mis-classified as on-device models awaiting a download. Auto-read now correctly turns on once the provider's API key (and voice, where required) is set.
• Common: Cartesia dictation works everywhere: Cartesia Ink as your speech-to-text engine is now correctly recognized as ready once its API key is set (instead of being treated like a local model awaiting download) — across the extension, native apps, and relay-backed setups (macOS, or the web client through your private relay), where the relay previously tried to load "cartesia" as an on-device model.
• Common: Gemini voice starts playing sooner: Google Gemini TTS reads replies aloud sentence-by-sentence now, so playback starts after the first sentence instead of waiting for the entire reply to be synthesized — matching how ElevenLabs, Cartesia, Resemble, and Kokoro feel. (Gemini's API generates a whole utterance at once, so the chunking is done on our side; the speed slider now applies to Gemini too.)
• Common: Voice playback errors are now shown, not silent: When reading a reply aloud fails (missing/invalid API key, an incompatible Resemble voice/model, a provider rejection, etc.), the reason now appears as an on-screen message instead of failing silently with only a console log. The messages are actionable (e.g. "The selected Resemble.ai voice doesn't support the 'chatterbox-turbo' model — choose a Chatterbox-compatible voice…").
• Common: Clearer Resemble.ai voice/model error: When a selected Resemble voice doesn't support the chosen model, the error now says exactly that and how to fix it, instead of surfacing a raw internal error payload.
• Common: HEIC/HEIF photos from modern iPhones convert again: Attaching a recent iPhone photo (HEIC/HEIF) failed to convert — it errored with "format not supported" and fell back to a plain file attachment instead of a viewable image. The built-in image converter was years out of date and couldn't read photos from current iPhones (full-resolution and HDR shots in particular). It's been replaced with an up-to-date decoder, so HEIC/HEIF attachments turn into JPEGs and display inline again.
• Common: Video and music tools always advertise the live model list: The video and music tools were only ever showing the LLM their three bundled fallback IDs — the per-turn background warm-up that fetched the live OpenRouter video/music registries inside the relay subprocess was racy and routinely missed the deadline before the tool's description was sent off. The cached registry state now primes the video and music model services on subprocess boot, so the tool always shows the current line-up (Kling, Hailuo, Wan, etc.), and a user-selected model resolves without a per-turn network round-trip.
• Common: Test Runner multi-model benchmarks now return meaningful scores: Running a "compare these models" benchmark with runLLMJudge = true used to silently produce zero scores and an input-order ranking, because the suite-level path checked only per-test evaluation settings and ignored the run-wide flag. The gate now honors either signal, and when the judge is on but no rubric was supplied a default rubric (factuality + completeness + clarity + helpfulness) is applied. The test_runner tool's get_result and export_transcript actions also no longer reject with runId and testId are required when only one of those was missing — the error now names the actual missing field.
• Common: Voice playback could be completely silent on iOS and Safari: Reading replies aloud sometimes produced no sound at all until the app restarted — the audio engine started in a suspended state and was never woken up. Playback now reliably produces sound.
• Common: Daily reminders fire at the time you set: A repeating reminder created for, say, 9:00 AM could drift and fire at the moment you created it each day instead. Recurrences are now anchored to the scheduled start time.
• macOS, Desktop: Scheduled-task notifications actually arrive: When a scheduled task finished and tried to notify you, the macOS and Windows/Linux desktop apps silently dropped the notification. It now appears as a normal system notification.
• Common: Sign-up failures are no longer silent: If the verification email can't be sent during sign-up, you now get a clear error right away instead of a sign-up that appears to succeed but never delivers the email.
• Common: The composer's "+" attach menu reappears on narrow panels: The earlier fix that kept the send button on-screen at very narrow widths inadvertently clipped the attach menu to nothing. Both now fit.
• Common: Claude model names work again for BYOK Anthropic users: Selecting certain Claude models with your own Anthropic key produced a "model not found" error because of an outdated internal model-name mapping. Model names now pass through to Anthropic as-is.
• Common: A failed step inside a multi-step app workflow now stops the workflow: A nested workflow that failed was reported to its parent as a success, so the workflow's error-handling branch never ran. Failures now propagate correctly.

New Features

• Common: The Pseudonymizer launches for everyone: Caiioo's on-device personal-data filter detects names, emails, phone numbers, addresses, IDs, organizations, cities, and other sensitive values in what you type and swaps them for realistic fakes before any of it reaches the model — then reverses the swap locally on the way back, so the conversation on your screen still shows the real values. The model never sees the real values; tool calls run on the real values after restoration. Turn it on with the new shield icon next to the send button — gray for off, blue for the Personal Data Filter, teal for PHI / Limited Data Set, emerald for PHI / Safe Harbor. While on, a thin ring colored to match the active mode wraps the composer as an ambient reminder, and a "🛡 pseudonymized — N substitutions" chip appears under each message that had values swapped; click it to see exactly which real → fake pairs were used, with category labels. Includes a strict PHI / Safe Harbor submode (HIPAA §164.514(b)(2): also strips dates beyond year, full geographic addresses, and ages over 89) for cases where you need a de-identified dataset you can share without a Data Use Agreement. (The feature was previously called "Anonymizer." We renamed it to "Pseudonymizer" because that's the technically correct word: GDPR Art. 4(5) reserves "anonymization" for one-way, irreversible transformations, and this one is reversible by design — your screen still shows real names.)
• Common: The Community Hub launches: A new one-click marketplace for installing the tools, modes, MCP servers, and complete apps that extend Caiioo. Browse it from inside the extension or from caiioo.ai/hub. The launch catalog includes Slack (sign in once with the Caiioo Slack app — no manual app config or developer mode), Notion, Linear, GitHub, Atlassian, PandaDoc, Microsoft 365, Legal Data Hunter (18M+ case-law, legislation, and doctrine documents across 110+ countries), and 18 additional self-registering MCP servers that authorize themselves through their own provider's sign-in — no copy-paste of secrets or developer-mode setup required from you. Each package shows a preview of exactly what it installs — tools, modes, skills, MCP servers, and variables — before you click Install.
• Common: Tool Network Access — explicit consent when a tool would leave your machine: When you pick a local chat provider (Ollama, MLX) you're making an intentional privacy choice — your conversation stays on the device. Tools that route through a remote provider (image generation, music, video, Perplexity web search) used to cross that line silently. Caiioo now stops on the first attempt, renders an Approve / Cancel card inline that names the tool and the provider it would send to, and remembers your decision per provider. Revoke at any time from Settings → Personalization & Privacy → Tool Network Access.
• Common: The Messaging Gateway launches (Pro): Caiioo can now answer messages on the platforms your customers and contacts actually use — WhatsApp Business, Messenger, Telegram, iMessage, Signal, Viber, and Beeper — through a single configurable gateway in Settings → Messaging. Each channel uses its provider's standard bot / business credentials, entered once per service. For every conversation you pick how the agent shows up: Auto (agent answers everyone; anything you type in the same thread is treated as coaching that shapes the next reply), Direct (agent still auto-answers, but anything you type goes through to the caller as you), or Takeover (agent steps back entirely and you take over). Image, audio, and video attachments flow through to the model in both directions, so you can ask "what's in this photo the customer just sent?" and answer in the same channel. Slack lives in the Community Hub as its own MCP package — install it from there rather than the Messaging Gateway.
• Common: My Day (Beta): A second Hub-installable app: a one-click morning briefing that pulls today's calendar, unread emails, and recent documents into a single dashboard. The composer button runs the brief; the agent renders into the same dashboard view every time, with follow-up skills for expanding any item, drafting a reply, or scheduling deep-work blocks.
• Common: Storybook Builder (beta): Another Hub-installable app: pick a style, audience, age range, and a freeform brief; the agent designs a cast, draws reference portraits to keep characters consistent, generates each page with prose and illustration, derives a cover from the best page, and renders the finished book into a readable scroll-layout viewer. Forkable like any Hub app, so you can customize the master prompt or swap the style options.
• Common: Free plan now includes the desktop apps: The macOS and Windows / Linux desktop apps used to require Pro. Free users can now sign in to the desktop apps on every platform — the platform itself is free, and Pro still differentiates on Pro-only capabilities (image generation, remote MCP servers, scheduled tasks, etc.).
• Common: Pro Mobile retired — one $9 Pro tier across web, desktop, and mobile: The $2.99 Pro Mobile in-app purchase (iOS and Android) is gone. The single $9/month Pro tier now unlocks every Caiioo client on every platform — Chrome, Edge, macOS, Windows, Linux, iOS, Android — with no separate mobile SKU and no cross-rail entitlement gymnastics. Existing Pro Mobile subscribers are grandfathered to Pro at no extra cost and keep every capability they had. New mobile installs see only the $9 Pro tier on the in-app paywall, matching the website.
• Common: Caiioo for Legal (beta): A new one-click install from the Community Hub activates fifteen cross-practice legal skills — contract markup, redline drafting, transactional drafting, memos & opinions, advocacy, batch playbook review, and more — backed by two comprehensive negotiation playbooks (buy-side and sell-side) with Preferred / Fall-back / Walk-away tiers across dozens of common clauses. Variables stay scoped to the thread you're working in, so each matter keeps its own client, counterparty, jurisdiction, and posture without bleeding into the next.
• Common: Self Checker — judge any answer with the new ⚖ button: Every assistant turn now has a ⚖ button in the action bar. Click it to score that answer against your request — the judge sees the full turn (your prompt, every tool call's inputs and outputs, attached images, and the assistant's reply), authors deterministic checks (exact match, contains, regex, number range, arithmetic), runs them, and renders a verdict card inline. Pick any provider you have a key for; the judge's LLM cost rolls into the conversation's running total so there are no hidden charges.
• Common: Test Runner — try Caiioo on your own list of examples (Pro): Hand Caiioo a list of prompts and a way to grade each answer — substrings that should appear, a pattern the response should match, specific tools that should get used, or a second model that scores the answer 1-10 across criteria you define. Caiioo runs each prompt in its own fresh conversation, captures the assistant's reply, what tools it called, how long it took, and how much it cost, then renders a pass/fail report you can export as a CSV. Useful for spot-checking that a new mode, model, or installed Hub app still behaves the way you expect.
• Common: Hub apps can ship ready-made reference materials: A Community Hub package can now bundle its own Slate templates, PDFs, and other reference files. Installing Caiioo for Legal, for example, drops the two negotiation playbooks straight into your library so the agent can mark up your contracts against them on the very first turn.
• Common: My Apps — fork any Hub app and edit it as your own: A new My Apps panel in Settings → Tools, Modes & Apps lets you fork any Community Hub app into a personal copy and edit any of its primitives — tools, skills, modes, cards, views, workflows, template attachments, and variables — with per-primitive editors. The agent can also snapshot a useful conversation into a draft personal app for you to refine.
• Common: Hub apps now sync across devices: Install a Community Hub package on your Mac and it shows up on your iPhone, and vice versa. Personal apps (apps you've forked or saved as your own) sync too. Per-package vector clocks mean no install ever overwrites a newer install from another device.
• Common: Hub Settings now lists every app primitive in its own section: Tools, Modes & Apps grew four new sections — Cards, Views, Workflows, and Template Attachments — alongside the existing Tool Configuration and Agent Modes. Each is a read-only inventory of what your installed apps actually shipped, with source attribution so you can see which Hub package brought in which skill, mode, or card.
• Common: Skills are now visible to the model: Previously skills were UI-only — clicking a skill chip pasted text into your message but the model itself had no awareness of them. Each mode now injects its available skills (name, description, prompt body) into the system prompt, so phrases like "use your contract analysis skill on this PDF" actually work.
• Common: Sub-agent cards render inline in the main chat: When a sub-agent emits a card (a Self Checker verdict, a generated chart, a structured result view), the card is reparented up to the parent conversation and rendered inline next to the sub-agent's text result — same as how sub-agent attachments already work.
• Common: Google Sheets — 14 new actions: Paste data (CSV / TSV / HTML), split text to columns, trim whitespace, remove duplicates, apply or clear toolbar filters with criteria and sort rules, move rows or columns, insert and delete cell ranges, protect ranges with editor permissions, define and update named ranges, attach developer metadata, fine-tune conditional formatting, and use the modern ColorStyle palette — all without leaving the chat.
• Common: Google Docs gets real comments, multi-tab support, and smart-chip awareness: Add, reply to, resolve, and delete native Docs comments that show up in the Docs UI for everyone on the document. Multi-tab documents now work correctly — the agent reads from and writes to the right tab instead of mashing every tab into one position space. Smart chips (people, links, equations, page breaks, date chips) are now recognized so search and edit operations land on the right character. Concurrent edits now fail loudly with a clear error instead of silently clobbering each other.
• macOS: Voice playback (Kokoro TTS) starts within a second: On-device voice was failing silently on macOS because the model wouldn't load inside the WebView. Voice now runs through the desktop app's helper process and streams sentence-by-sentence, so you hear the first sentence within about a second of clicking play, even on a busy machine.
• Web: Hub packages install via a caiioo:// link: Clicking Install on caiioo.ai now routes directly to whichever client you actually have — Chrome extension or native app — instead of firing both at once. If you have both installed, you'll see a picker. The Community Hub install modal also breaks open each package's payload — tools, modes, skills, MCP servers, and variables — so you can see exactly what gets installed before you click Install.
• Desktop: Linux AppImage registers the caiioo:// URL scheme: Linux users on the AppImage build can now install Hub apps from caiioo.ai with one click — the URL scheme registers on first launch without any system package install.

Improvements

• Common: Slate stability and round-trip fidelity overhaul: Closed roughly 40 individual bugs across the rich-text editor, the diff engine, version history, and Word import/export. Tracked changes no longer corrupt when an AI proposal lands on a document you've edited since; version-history snapshots are now true frozen copies; revisions resolve correctly when the content matches the target instead of stalling; AI proposals merge with existing redlines instead of overwriting other authors' edits; accepting or rejecting changes in a Word file persists to storage; rich-text exports handle emoji and other supplementary-plane characters; and the diff engine no longer confuses deletions and modifications when their text matches.
• Common: Slate — accept or reject every tracked change inside a selection: New ✓ Sel / ✗ Sel buttons in the Slate review toolbar mirror Word's "select a paragraph, accept all changes inside" behavior — highlight a region and one click resolves every tracked change that overlaps it. The diff engine also coalesces adjacent edits separated only by whitespace or punctuation into a single accept-or-reject unit, so reviewing an AI rewrite isn't a hundred individual clicks.
• Common: Word document import / export fidelity: A roughly 110-bug pass through the .docx parser fixed character formatting that ignored explicit "off" toggles, broken style inheritance, missing theme-color resolution, lost paragraph indentation, mishandled superscript / subscript / hidden text, wrong list numbering after headings, missing tab and line-break separators in extracted text, embedded images that weren't being extracted at all, and round-trip loss of embedded image references on export. Markdown export now uses CommonMark-correct list indentation.
• Common: PDF round-trip and rendering audit: An audit of the PDF pipeline preserves inline images, hex strings, and letter spacing on round-trip; resolves inherited page resources correctly; preserves transparency masks, decode, intent, and interpolation flags on image replacement; rescues special characters (Euro symbol, smart quotes, trademark) that Windows fonts can't natively encode; correctly handles emoji and other supplementary-plane characters in PDF-embedded fonts; and surfaces OCR errors instead of swallowing them. The Slate PDF viewer also opens noticeably faster on multi-page PDFs by rendering pages lazily, and large PDFs no longer fetch the file repeatedly when first opened.
• Common: PDF export preserves unencodable characters instead of failing: Special characters the chosen font can't encode are now passed through unchanged from the source PDF instead of aborting the entire export.
• Common: Pseudonymizer accuracy and coverage upgrades: The personal-data detector model was retrained with a locale-aware pipeline and now handles Chinese and a wider range of non-Latin scripts substantially better. Name spans now extend correctly across script boundaries — middle initials, leading honorifics, and contiguous Chinese, Japanese, Korean, and Arabic runs — so half-name leaks are closed. A new safety-net second pass catches misses before they're sent. Fragmented same-label spans are coalesced before substitution. Multi-language city and company-name detection is now in production. The calculator tool's numeric output is no longer mis-classified as a name.
• Common: Pseudonymizer hallucination inspector: Scans the assistant's reply for fake-shaped names that aren't in your session's substitution map — a hit means the model probably made up a name. Pairs with the existing leak inspector to give a complete view of what the model said about identity.
• Common: Pseudonymizer multilingual fakes: City names, company names, and personal names now generate locale-appropriate substitutes — a Spanish prompt gets Spanish-looking fakes, a Japanese prompt gets Japanese-looking fakes, and so on.
• Common: Pseudonymizer 30+ smaller correctness fixes: A multi-round audit cleared dozens of low, medium, and high-severity findings — script-coverage gaps, debug-log noise, leaks across sessions, URL trimming bugs, policy edge cases, audit-log privacy, restore robustness, Unicode handling in the user dictionary, structural-PII leaks under self-only mode, and more.
• Common: Telegram messages render with proper formatting: Messages sent through the Telegram bridge are now formatted using Telegram's native bold, italic, code, and link styling instead of showing raw asterisks, backticks, and broken "text (url)" syntax. Inbound images from Telegram also flow correctly to vision-capable models, and pre-formatted messages from the agent aren't re-formatted by the bridge.
• Common: Calendar sync covers every calendar in your account: Background sync iterated only your primary Google calendar; team and family calendars were silently absent. Every visible calendar is now synced. Event pagination is properly followed (so events past the first page no longer go missing), event timezones are preserved instead of being normalized to UTC, all-day events anchor correctly, and the Apple Calendar handling now routes Apple Reminders too.
• Common: Calendar / Agenda tool correctness: Closed a cluster of 19 bugs across calendar create / update / delete / list — most importantly, all-day event dates are now derived in your local timezone instead of UTC, so an event you set for Friday no longer lands on Thursday in eastern timezones.
• Common: Gmail tool reliability sweep: Fixed five bugs that were silently producing wrong-account results, broken reply threading, mangled "Doe, John" style recipient names, and hidden authentication failures. Replies now thread correctly in both Gmail and external mail clients (Outlook, Apple Mail, Thunderbird). Draft updates preserve the original conversation thread.
• Common: Gmail search by sender, category, age, attachment, and unread state: Asking the agent to find "unread emails from Bob from the last week with attachments" used to depend on the model remembering Gmail's exact search-operator syntax — and it would re-issue the same logical query with different wording until something worked. The Gmail tool now exposes first-class filter parameters (from, subject, label, hasAttachment, isUnread, isImportant, category, newerThan / olderThan like 7d/1m/1y, and after / before dates), so the agent picks the right filter on the first try.
• Common: Google Drive tool reliability sweep: Twelve fixes including refusing to read binary files as text, supporting shared-drive folder paths, surfacing pagination for large folder listings, mapping Slides to PPTX exports, removing the broken "owner" role from share options, and fixing destination-folder filtering on moves. The Drive transfer cache now expires public links after 24 hours instead of leaving them permanently public if cleanup fails.
• Common: Google Drive — full folder paths, shared drives, export, owner transfer, link discovery: Drive operations now accept human-readable folder paths (Engineering/Specs/Q3) instead of only opaque folder IDs, walk into Shared Drives as first-class destinations, export Docs / Sheets / Slides to specific MIME types (PDF, DOCX, XLSX, PPTX), transfer file ownership between users, and surface anyone-with-link and public links so the agent can answer "what's the shareable URL for this file?" without you copying it out manually.
• Common: Google Slides text edits land in the right place: The Slides tool used a 999999 magic number for "end of text," which the API rejected. It now looks up the actual text length and constructs proper ranges so partial-index edits (e.g. "style from character 5 onward") work as intended.
• Common: Google Sheets — 30+ smaller correctness fixes: Range parsing handles quoted sheet names with inner punctuation, unbounded references (A:A, 1:10), and columns past Z; image uploads write a real =IMAGE() formula instead of erasing the cell; HTML import decodes named, decimal, and hex entities including astral-plane characters; charts no longer crash the sheet info reader; search reports absolute column letters and surfaces per-sheet errors. Sheet diffs now emit both adds and deletes correctly.
• Common: Variables dialog now works for any Hub app: The "open variables" composer button used to be hardcoded for one specific app. It's now driven from the Hub manifest, so any app that ships a variables dialog gets its own button and label.
• Common: Personal apps surface across every reader: Personal apps (apps you've forked or saved) now contribute their skills, modes, MCP servers, tools, views, and variables through every place the agent reads them — not just the composer. They're a first-class app now, identical to Hub-installed apps.
• Common: Hub uninstall actually removes everything: Uninstalling a Hub app now also removes the package's modes, composer buttons, skills, template attachments, mode-variable patches, the cached system-disable list, and (if you were on it) the package's active mode. No more orphan modes lingering in the picker after uninstall.
• Common: Hub install warns about overlaps: When you install a package whose skills, modes, MCPs, tools, or views would duplicate something you already have, a toast surfaces the conflicts at install time so you can decide what to do, and per-row duplication indicators stay visible in the Hub install list and in your Settings inventory.
• Common: Per-thread variable overlays for matter-scoped work: Mode variables used to be sticky across every conversation, which is wrong for matter-scoped work (each legal matter, each client engagement, each project is a different context). Each thread can now carry its own variable overlay so the agent works on Matter A in one thread and Matter B in the next without mixing them up.
• Common: Hub install no longer pops a tab cascade for every required sign-in: Installing a multi-provider Hub app used to fire one sign-in tab per provider in sequence. Installs now complete fast and prompt for each remaining sign-in on demand, one at a time, instead of stacking tabs.
• Common: Cloud sync robustness: Team-sync key derivation now uses your organization ID and passphrase instead of your personal email, so every member of the same team derives the same key and can actually decrypt each other's items (this was previously broken). Sync timers, vector clocks, and manifest locking also tightened to prevent overlapping syncs from corrupting state.
• Common: "Use Caiioo's Account" button works for re-issued provisioned keys: The Settings button used to do nothing if your provisioned OpenRouter row had been deactivated by a previous switch to BYOK or a decrypt failure. It now mints a fresh row when the server says you don't have one, so the button always restores a working key.
• Common: OpenRouter key field stays in sync with Settings: The API key input now refreshes when the parent component pushes a new value (e.g. after clicking "Use Caiioo's Account") instead of holding onto the value it had when the page first rendered.
• Common: Composer buttons reload when mode settings change: Composer action buttons (Variables, etc.) now refresh immediately when a mode's settings change, instead of needing a chat reload.
• Common: Agent tools see your latest edits before they read a Slate: If you're still typing when an AI tool fires (Slate update, revision, tracked-changes resolve), the tool now waits for in-flight editor content to flush to storage so it operates on what you can actually see — not a stale snapshot from a second ago.
• Common: Inline cards size themselves to their content: The Self Checker verdict card and other inline cards now grow to fit their actual content instead of reserving a fixed slot, and pick up the parent app's theme (light or dark) instead of forcing a white background.
• Common: Floating action buttons default to the top-left corner: The floating ⚖ and 🛡 buttons (and any custom floating buttons) now land in the top-left of the composer by default instead of obstructing the send button on the right. You can still drag them anywhere.
• Common: User profile and mode-variable updates reject unknown fields: The agent could previously invent variable names like _clientName that silently went nowhere. Updates now require the key to exist in the schema and return a clear error otherwise.
• Common: Onboarding welcome screen simplified: First-run now shows three clear options — Free, Subscribe to Pro, or Bring Your Own API key with a trial — instead of the previous longer onboarding form. The first time you open the Composer, Settings, or Slate, a short interactive tour runs to point out the relevant controls.
• Common: Sign-in is one step: Caiioo used to require both a verified identity AND an active license check before letting you in. That redundant license probe is gone — once you've signed in, you're in. License state still gates Pro-only capabilities (image generation, etc.); it just no longer gates opening the app.
• Common: Settings search reveals advanced sections: Searching for a setting that lives under an "advanced" twist-down used to silently return no results because the section was collapsed. The search bar now reveals matching collapsed sections, and a new "Collapse all" button reverses it in one click.
• Common: Friendlier provider catalog: Newly released OpenRouter free-tier models are now prioritized in the model picker, and deprecated models are soft-removed instead of cluttering the list.
• Common: Host-language detection on every native platform: The macOS, iOS, Android, and Tauri shells now detect your OS interface language at launch and pass it through to the sidepanel, so first-run translations land in the right language without you having to set it manually. About 10,000 additional translation strings landed across 23 non-English locales.
• macOS: Single-instance enforcement: A second copy of the macOS app can no longer launch from a caiioo:// link when one is already running, even when macOS Launch Services has two registered copies (typically a leftover from a prior install).
• macOS: MCP servers installed via Homebrew now launch: macOS 15+ blocks notarized apps from running binaries that carry the "provenance" attribute, which Homebrew adds to everything it installs. Caiioo now auto-clears that attribute on permission-denied spawn, so local MCP servers (filesystem, memory, fetch, etc.) installed via Homebrew start working without any manual cleanup.
• macOS, iOS: Hub install via caiioo:// is reliable on cold launch: Tapping a caiioo:// install link on a freshly launched app could race the WebView's first load and silently drop the install intent. The intent is now persisted across the cold-launch race so the install completes once the app is ready.
• iOS: Apple's new "write-only" calendar permission is respected: iOS 17 introduced a third Calendar / Reminders permission — "write-only" — that lets Caiioo create events and reminders without seeing your existing ones. Caiioo previously treated this as "denied"; it now uses the write-only access correctly, so users who grant only that level can still ask the agent to schedule things.
• iOS: Clearer message when an in-app purchase can't be verified: Failed purchase verifications now surface a specific reason instead of a generic error, and the transaction is no longer silently finished, satisfying Apple's StoreKit guidance.
• Android: Saving a file no longer freezes the app: Writing a large file through the Android save-file picker used to run on the UI thread and could freeze the app for several seconds on slow storage. Writes now happen off the UI thread.
• Android: Streaming network responses deliver headers before chunks: Long-running streaming calls now deliver headers to the caller before any body chunks arrive, fixing a category of intermittent stream failures.
• Android: On-device transcription faster and more memory-efficient: The Whisper / Moonshine audio capture path now uses a primitive float buffer instead of a boxed list, reducing both memory and CPU. Loading, unloading, transcribing, and clearing the model cache also serialize through a shared lock now, so the app no longer occasionally crashes if you switch models mid-transcription.
• Desktop: Reliability and security audit of the Windows / Linux shell: A full pass through the Tauri shell — capability scope narrowed to the trusted sidepanel only (so an arbitrary visited page can't invoke privileged commands), HTML-escaping hardened on the caiioo:// Hub-install bridge, a graceful "Node.js not found" message instead of a silent crash, and a handful of robustness fixes.
• Desktop, macOS: On-device speech and voice model downloads now work: The desktop apps' WebView Content Security Policy was blocking downloads from huggingface.co, so the Kokoro voice and Whisper speech-to-text models couldn't load on first use. The policy now allows huggingface.co.
• Extension: Settings → Tools, Modes & Apps renamed and reorganized: The old "Tools & Capabilities" category is now "Tools, Modes & Apps" with a new "Connectivity" category split out from Advanced.
• Web: Pricing copy refined across 23 languages: An editorial pass on the website's Free / Pro feature bullets brought the non-English locales into line with the latest English source. Legal document "last updated" dates were refreshed.

Security

• Common: Sign-in hardening: Multiple findings closed from an audit of the sign-in and credential-vault pipeline. Google ID tokens are now only accepted if they were issued for Caiioo's own client IDs (closing a hole where any Google OAuth client's token could sign someone in). Apple sign-in now validates the nonce returned by Apple against the one Caiioo generated. Google and Apple sign-in refuse to auto-link an external identity onto an unverified-email account (closing a "shadow signup" takeover). The refresh-token endpoint now rejects revoked tokens, so a logged-out token can't be exchanged for a fresh one. The OAuth-token issuance endpoint now requires a live bearer token. The OAuth callback page itself was hardened against several script-injection vectors.
• Common: Per-profile isolation for all settings: The settings layer was sharing a single bucket across multiple sign-ins on the same install in certain code paths, which could leak a credential entered under one account into another account's view. Every settings read and write is now strictly per-profile, legacy shared buckets are wiped on first launch after upgrade, and cloud sync refuses to upload or download those legacy buckets.
• Extension: Local-bridge auto-discovery is now opt-in: The Chrome extension previously connected unconditionally to any Caiioo desktop app it found on localhost. It now only does so when you've explicitly turned on local-bridge access in Settings, so a desktop app installed by another user on a shared machine can't be silently bridged to your extension session.
• macOS, iOS: Hardened caiioo:// install bridge against injection: The hand-rolled string escaping on the JavaScript that processes caiioo://hub/install/<id> URLs only escaped single quotes — a maliciously crafted package ID could break out of the string literal and run arbitrary script in the WebView. Replaced with full JSON escaping on both platforms.
• iOS, macOS, Extension: Google sign-in flows now use the verified Caiioo OAuth app: Connecting Google for Private Sync, Calendar, Gmail, Drive, and the other Workspace tools used to show the "This app isn't verified" warning on iOS, macOS, and the Chrome / Edge extension because those platforms were still authenticating against an older, unverified Google Cloud project. Every platform now uses Caiioo's verified Google project end-to-end, so you see the proper Caiioo branding and the verified-app green check on the Google consent screen instead of the warning. One-time re-login: existing users on iOS and macOS will be signed out automatically and asked to sign in again the first time they open the app after this update — the previous sign-in tokens were issued by the older Google project and cannot be carried over.

Bug Fixes

• macOS: "Browser not connected" warning clears when Safari connects: Connecting the Safari extension to the Mac app left a stale warning saying browser control was unavailable — and the warning only named Chrome, even though Safari, Edge, and Firefox all work. The app now counts every connected browser (Safari connects over a different channel than the others, which is why it was being missed), so the warning clears the moment any browser connects, and its wording no longer singles out Chrome.
• Common: Newer image models you pick are actually usable: The image-model picker showed every model your account can generate with — including newly released ones like Recraft — but choosing one could fail with "Unknown image model" because the generator was checking your selection against a stale built-in list instead of the live one. The generator now refreshes the live model list before deciding, so anything visible in the picker works. The default model also changed from FLUX.2 [pro] to the faster, cheaper FLUX.2 [flex], which is a better all-round default.
• Common: Tracked changes inside Word documents now accept and reject correctly: Clicking accept or reject on a tracked change inside a .docx-backed Slate used the document position as the change ID, which never matched the real stored change ID — so the change visually flipped state but never persisted. Fixed.
• Common: Slate handles corrupt template installs: Hub-installed Slate templates that shipped without the proper wrapper used to hang the viewer on "Loading artifact" with no recovery. Templates now self-heal on load and the viewer surfaces a clear error if a template is genuinely corrupt.
• Common: Self Checker no longer narrates the verdict twice: When the Self Checker rendered its verdict card, the agent was also describing the scores in reply text — two copies of the same verdict on the same screen. Suppressed the prose recap when the card renders.
• Common: Self Checker reads image attachments correctly: The judge was decoding image attachments as text and feeding the resulting garbage into the evaluation. Image attachments are now forwarded to the judge as actual images, so any turn whose correctness depends on what's in an image scores correctly.
• Common: Self Checker shows only the latest verdict: Each ⚖ click was appending a fresh verdict card without removing the previous one, so the chat ended up with a stack of duplicates. Each turn now shows only the most recent verdict, with an × to clear it.
• Common: Reminders sync dialog renders Outlook events correctly: The internal "microsoft_calendar" source value was leaking into the platform-detect path in the reminders modal. The modal now falls back to its inference path so events render correctly regardless of where they came from.
• Common: Large Gmail attachments flow through reliably: Large Gmail attachments (a 25 MB file is roughly 33 MB after base64 encoding) were exceeding the inter-process message ceiling and silently dropping. The runtime now falls back to a temp file for oversized payloads so attachments make it through.
• Common: PandaDoc MCP install works again: PandaDoc moved their MCP endpoint to a new path. Caiioo's catalog entry now points at the new endpoint.
• Common: Workflow render steps no longer mis-resolve string outputs as attachments: Forwarding a string output from one workflow step into a View step used to fail because every string was treated as an attachment ID. Strings forwarded by reference are now distinguished from literal attachment IDs.
• Common: Subscription-status check no longer hangs in browser-only environments: The agenda tool was probing the desktop bridge on every cache-miss in environments that don't have one, throwing a "Failed to fetch" error every time. The probe is now gated on whether a desktop bridge is actually reachable.
• Common: Background process for the agent now exits cleanly: A change in a previous release left the agent's background process holding open after it finished, blocking the things that run after a turn — most visibly, the automatic conversation title. The process now exits cleanly so title generation and other post-turn steps fire as expected.
• Common: Lab and admin-only modes hidden from the mode picker: Two paths were still leaking the internal "lab" mode (and any other mode marked admin-only) into the user-facing mode picker. Fixed.
• Common: Duplicate Hub installs deduplicated on save: A race in the install pipeline could write two entries for the same package. New saves dedupe by package slug, and existing duplicates are healed on next read.
• Common: Custom mode IDs no longer collide with Hub installs: The storage key is now authoritative for a custom mode's ID, eliminating a class of collisions when a Hub-installed mode landed on a key that already had a user-edited copy.
• Common: Sub-agent attachments and cards reach the main chat: A sub-agent that generated an image, a chart, a Self Checker verdict, or any other attachment used to drop the result against an invisible sub-agent thread — and the parent conversation would sometimes invent a URL to fill the gap. Sub-agent attachments and cards now reparent to the main thread correctly across the agent runner's mid-run cleanup, so what the sub-agent produced shows up where you sent the request.
• Common: PDF reading on lazy-rendered pages: The "view original text" pop-up for a PDF chunk now works on pages that hadn't been rendered yet when the chunk was created.
• macOS: Slack, Notion, Linear and other Hub OAuth sign-ins now complete in the native app: Hub MCP servers that use HTTPS-only OAuth (Slack, Notion, Linear, etc.) couldn't redirect back to the local relay because their providers reject http:// callbacks. The macOS app now routes those flows through the Caiioo cloud relay's /oauth/callback, which then bounces the authorization code back to the local relay over the existing per-user channel — so sign-in for these providers from inside the native app now lands the same way it does in the extension. Includes a one-time migration that fixes existing installs whose stored profile wasn't yet wired to the local relay's identity provider.
• Common: "Ask the user" doesn't hang in sub-agent runs: An agent that called ask_user from inside a sub-process was hitting a shared in-memory singleton that wasn't reachable across processes, so the question never surfaced and the run stalled until you cancelled it. The collaboration controller is now per-thread, so the question shows up in the chat the way it does in the main agent loop.
• Common: Native macOS Calendar / Notes / Reminders helpers refreshed: Updated the helper binaries against current macOS SDKs to clear an issue where reminders sync intermittently saw the wrong items in 0.9.721.
• Extension: Self Checker verdict cards now render inside the extension: Inline scripts inside card templates were blocked by the extension's content-security policy. Cards now route through the extension's sandbox page so they render correctly.
• Common: Text shows up when you turn an SVG drawing into an image: Rendering an SVG — a logo, diagram, or chart — to an image used to drop all of its text (wordmarks, labels, captions) in the Chrome extension and the macOS app, leaving blank gaps where the words should be; only the mobile apps rendered them. Caiioo now ships fallback fonts (sans-serif, serif, and monospace, including bold), so text appears exactly as drawn — including symbols like &, ™, and accented characters — and any typeface the drawing asks for that isn't available falls back to a clean sans-serif instead of vanishing.

New Features

• Common: Pricing collapsed to Free and Pro, with a $5/mo OpenRouter credit bonus for Pro: We retired the separate Platform and Pro Mobile tiers and folded every Platform-only feature (Apple Calendar / Reminders / Notes, encrypted private relay, API access preview, native voice dictation, browser UI) into Pro. The website, account page, and in-extension pricing panel are now a clean two-card Free / Pro layout. As part of the change, every paid Pro month (Stripe or Apple) now also tops up your provisioned OpenRouter key with $5 of credit on top of your existing balance. Existing Platform subscribers are grandfathered into Pro at no extra cost and keep every capability they had. The OpenRouter credit-purchase floor was also lowered: end users no longer see the legacy $0.80 minimum-fee surcharge, just the 5.5% pass-through fee.
• Common: Slack is now a one-click Community Hub install: The standalone Slack integration is gone; Slack now installs from the Community Hub like Notion, Linear, GitHub, etc. Click Install on the Slack package in the Hub, sign in once with the Caiioo Slack app, and the official Slack MCP server is wired up immediately — no manual app config, no developer mode. Slash commands like /caiioo and /caiioo-help work from any channel where the bot is invited.
• Common: Workspace Files tool: A new sandboxed file tool lets the agent read, write, edit, and search files inside a workspace folder you point it at. The agent cannot escape that folder, and cannot reach the network through this tool. Reads auto-parse Office formats (docx/xlsx/pptx) and PDFs. Configure the folder in Settings → Tools → Workspace Files; the agent picks it up immediately.
• Common: PHI Safe Harbor submode for the Pseudonymizer: The PHI mode of the Pseudonymizer now offers two submodes — Limited Data Set (the existing 16-category strip, default) and Safe Harbor (the stricter HIPAA §164.514(b)(2) standard that also removes specific dates beyond year, full geographic addresses, and ages over 89). Pick the submode from the Pseudonymizer settings or the new in-chat toggle. Safe Harbor is the right choice when you need a de-identified dataset you can share without a Data Use Agreement.
• Common: Pseudonymizer is now generally available: The Pseudonymizer — which detects names, emails, addresses, IDs, and other sensitive values in what you type and swaps them for realistic fakes before any of it reaches the model — graduates out of the tester program with this release and is available on every plan, including Free. Turn it on in Settings → Privacy → Pseudonymizer, or with the new in-chat shield toggle. The model never sees your real values; substitutions are unmasked locally before tool calls run, so the output stays accurate.
• Common: Pseudonymizer in-chat controls: The Pseudonymizer no longer hides in Settings. A shield icon next to the send button (gray = off, blue = Personal Data Filter, teal = PHI/Limited Data Set, emerald = PHI/Safe Harbor) shows the active mode at a glance and opens a popover with the same mode picker you'd find in Settings. When the Pseudonymizer is on, a 1-pixel ring colored to match the mode wraps the composer as an ambient reminder. After each turn that had substitutions, a small "🛡 pseudonymized — N substitutions" chip appears under your message; click it to see exactly which real → fake pairs were swapped, with category labels.
• Common: Connect a browser on another device to your computer: A new pairing flow lets you safely use Caiioo in a browser on your phone, tablet, or another laptop, with everything still running on your main computer. In Settings → Connections → Caiioo Bridge → "Pair a device", generate a 6-character code that's good for 5 minutes and single-use. Open the same URL on the other device, enter the code, and that device stays paired from then on. Brute-force guesses are rate-limited to 10 attempts per IP per minute, and a server restart on your main computer re-pairs every device cleanly.

Improvements

• Common: Agenda items color-stripe by source calendar: Each event card in the Reminders / Agenda window now carries a colored stripe matching the calendar it came from, so you can tell at a glance whether an item is from your work, personal, or shared calendar. Google Calendar today; Apple and Microsoft follow once their data feeds expose a calendar color.
• Common: AI can read a Google Sheet without naming the exact cells first: The Google Sheets read tool used to refuse to run unless the agent specified an exact cell range like Sheet1!A1:D200. The model now can ask for a whole sheet, or every sheet in a workbook, and the tool figures out the layout itself. A built-in 1000-row ceiling keeps a giant spreadsheet from blowing past the model's context window — if a sheet gets capped, the response tells the model so it can ask for a smaller range or a higher row count next time.
• Common: Installing an OAuth tool from the Community Hub now opens sign-in for you: Adding a Hub package that needs sign-in (Notion, Linear, etc.) used to leave the card stuck in an "OAuth required" state with a "Sign in" button you had to click separately. The install now opens the sign-in tab for you automatically — your click on Install is treated as consent — and an "Opening sign-in for X…" line explains why a tab popped. The manual Sign in button stays as a backup.
• Common: Settings → Connections shows the real network address for other devices: The "Connect from other devices on your network" panel was showing 127.0.0.1:3847 (the loopback address that only works on this same computer) on the macOS app, so users were copy-pasting an address that couldn't actually be reached from another device. The panel now shows the real LAN IP and .local network name from the bridge itself. The address is also hidden when the new "Allow access from other devices" toggle is off, so you don't see an address that wouldn't work anyway.
• Common: Web search captures Google's AI Overview reliably: The web-browsing tool's Google scraper sometimes returned an empty AI Overview block, or pulled in noisy right-rail "Sources" text and inline scripts. We re-anchored extraction to Google's stable section markers and now strip script content and the right-rail before returning, so the agent sees only the readable answer.
• Common: Pseudonymizer protects company names too: The Pseudonymizer now treats organization names as protected entities, swapping in realistic fakes that preserve legal suffix style (Inc., LLC, GmbH) and "partners-vs-brand" structure. City and small-region names ("Springfield", "St. Albans") are also detected as a first-class category so they can be swapped or stripped per your active mode.
• Common: Pseudonymizer adds Spanish, French, German, and other multilingual coverage for cities and regions: The personal-data detector model was retrained with hand-authored multilingual data for city names across 23 languages, so users with non-English-language prompts now get the same level of protection as English speakers.
• Common: Pseudonymizer model download shows progress on iOS: The first-time PHI / Personal Data detector download could appear stuck on iOS because nothing told you it was still working. There's now a heartbeat progress log so you can see the download isn't frozen.
• Common: Subagents can search the model catalog by name: To keep the sub-agent tool description in budget, the catalog embedded in it is now the top 10 highest-value models per provider. When a sub-agent needs a model outside that list, it can call the new search_models action with a name fragment (e.g. "haiku") and get back exact catalog IDs, pricing, and capability flags — so the model can pick a concrete ID even for lesser-used choices.
• Common: Custom MCP servers show a friendly name in the credentials vault: When you signed into a custom MCP server, the credentials vault was labeling it with the raw connection URL, which was hard to scan. The vault now shows the friendly server name you gave it, falling back to the URL only when no name is set.
• Common: Community Hub now adds Legal Data Hunter: A new Community Hub package wires up Legal Data Hunter — 18M+ case-law, legislation, and doctrine documents across 110+ countries — as a one-click MCP install.
• Web: SOC 2 Type I badge on the trust page: Caiioo passed its SOC 2 Type I audit. The trust page on caiioo.ai now carries the AICPA SOC 2 badge alongside the existing security disclosures.
• iOS: Hide the keyboard toolbar that iOS adds to text fields: The prev/next arrows and Done button iOS attaches above the keyboard for web text fields ate noticeable vertical space in the composer. The bar is now hidden (using only Apple-public APIs, so this stays App Store safe), and stays hidden after page reloads.

Security

• Common: The local bridge stays on your computer by default; LAN access is opt-in: The local bridge that powers stdio MCPs and the desktop-only tools used to listen on every network interface, and on a public Wi-Fi network another device could request a session token and then call the privileged "run a shell command" endpoints. The bridge now only listens on your own computer by default. To use it from another device, turn on "Allow access from other devices on your network" in Settings → Connections, then pair the device with the new code flow (see above). Even when LAN access is on, session tokens are only handed out to requests coming from your own machine, so a stranger on the same Wi-Fi cannot get a foothold.
• Common: Pseudonymizer refuses to leak your real values to Perplexity: If the Pseudonymizer is on and the agent tries to run a Perplexity search whose query contains a pseudonymized identifier (a name, email, address it already swapped), Caiioo now blocks the search and surfaces a clear explanation instead of un-swapping the value and sending it to Perplexity. Perplexity is a third-party LLM service — the whole point of the Pseudonymizer is that values like that never reach an LLM service. To run the search, turn the Pseudonymizer off for that turn, or rephrase to avoid the protected value.
• Common: Spreadsheet parser swapped to a maintained library (GHSA-4r6h-8v6p-xvw6): The library Caiioo used to parse .xlsx attachments had an open prototype-pollution advisory and is no longer maintained. We replaced it with the well-maintained exceljs library. Spreadsheet uploads and the file-workspace's .xlsx reads return the same content as before.

Bug Fixes

• Common: Browser stuck on "Loading Caiioo… 0/0" after a server restart: When the host server didn't have an end-to-end-encrypted session with a connecting browser yet (because the server just restarted, or it's the browser's first connection), the browser was ignoring the server's request to start the key exchange — so encrypted messages never decrypted, and the UI got stuck loading. The browser now answers the key-exchange request correctly and the sidepanel loads.
• Common: Caiioo kept asking for Google permissions you'd already granted: When you granted a Google permission mid-conversation (from a sibling tool, the settings panel, or a sync from another device), the agent's in-memory copy of your account didn't notice, and it kept asking again for the same permission — sometimes several times in one run. Caiioo now re-reads your account once before complaining about a missing permission, so within-run grants are honored immediately.
• Common: 7 languages were being told the AI is "English": The internal map from locale code to language name covered 17 of the 24 supported languages. Users with their UI set to Arabic, Hebrew, Hindi, Bengali, Urdu, Turkish, or Dutch were getting a system prompt that said "respond in English" instead of their actual UI language. The map is now complete, and a test prevents this from regressing.
• Common: Reminders sync dialog showed raw text codes instead of translated labels: The Caiioo card in the Reminders / Agenda sync dialog was rendering literal placeholders like reminders.sync.CaiiooName instead of the translated text, in every language. Fixed.
• Common: Google sign-in stops ping-ponging when you grant an extra permission mid-conversation: When the agent needed a Google scope it didn't have yet (Drive, Sheets, etc.) and prompted you to grant it from a sidebar tab, the new sign-in could return the agent to a state where it asked for the SAME scope again. The OAuth flow now correctly merges the newly granted scope into your existing connection instead of overwriting it, so one approval is one approval.
• Common: Sub-agent attachments now show up in the main chat: When a sub-agent's tool created an image, file, or other attachment, it used to be stored against the sub-agent's invisible thread and the main conversation showed nothing — sometimes the parent model would invent a URL to fill the gap. Sub-agent attachments are now linked back to the parent thread automatically and render inline like any other tool result.
• Common: Pseudonymizer detector pass-2 catches names the first pass missed: The Pseudonymizer now runs a quick second detector pass that re-checks the message for any real values that should have been masked but weren't, before sending. Belt and suspenders for protected categories the model is most likely to miss.
• Common: Pseudonymizer no longer puts your real name in the My Identifiers placeholder: The Settings → Pseudonymizer → My Identifiers field was showing your account's real name as the example placeholder, which both looked like a leak and confused setup. Replaced with a generic placeholder.
• Common: Pseudonymizer pill text is readable in dark theme: The "🛡 pseudonymized" pill under user messages was using a light-mode text color in dark theme, making it nearly invisible. Fixed.
• Extension: Sidepanel "Get current location" dead-end fixed: The sidepanel's location request was returning "Permission denied" before the browser could even ask you, because the Chrome manifest was missing the location permission entirely. Permission added; the location request now reaches the browser prompt as expected. Existing users will see a one-time permission request on update.
• iOS: Subscribe page only sells Pro after the tier merge: The iOS in-app paywall briefly still showed the retired Pro Mobile and Platform tiers, which could leave users on a now-unsupported plan. The paywall now sells exactly the same Pro tier shown on the website. Existing Pro Mobile / Platform subscribers continue to be honored as Pro at no extra cost.
• iOS: Rare crash during navigation while a page was still loading: iOS could crash when a page navigation was cancelled mid-load (for example, tapping a link before the previous page finished). The fix routes every error path through the same already-guarded helper, so a cancelled load never tries to deliver a result on a closed page.
• Web: Sign in works inside in-app browsers (Slack / X / LinkedIn / Instagram) and on iOS Safari: The popup-style Google sign-in failed when caiioo.ai was opened from a link inside another app, because in-app browsers either block the popup or strip its connection back to the original page. On mobile and in-app browsers the site now uses a full-page redirect through your own browser session instead of a popup, so sign-in completes and returns you to the page you started on. Desktop popup sign-in is unchanged.
• Web: Community Hub polish on tablets and long names: The navigation bar on the website now switches to the hamburger menu at tablet widths (up to 1024px) instead of overflowing into the logo. Hub package modal titles no longer run under the close button. The "Coming Soon" pill wraps cleanly on narrow widths. Tool icons fall back to an emoji, then to a known logo from the company's website, then to a first-letter avatar — instead of letting a long internal slug (like "customerio") overflow out of the icon box.
• Web: Community Hub now shows the right author on each package: Every package in the Hub used to read "by caiioo" regardless of who actually built it. Authors now reflect the real maintainer — "Model Context Protocol", "oraios", or the vendor brand — and only fall back to "Caiioo" for tools and modes we built ourselves. Cards for integrations that aren't fully wired up yet are hidden from the Hub until they work, so you don't see installable cards that immediately error.
• Web: Sitemap stops triggering "page with redirect" warnings: Search Console was flagging every page on caiioo.ai as a redirect because the sitemap listed URLs without the trailing slash that the live site uses. The sitemap now matches the canonical URLs and includes per-language alternates, so search engines stop seeing the whole site as redirected.

Brand

• Common: PebbleFlow is now Caiioo: The product has been renamed to Caiioo — the same animal, the same app you've been using. You'll see the new name throughout the sidepanel, settings, the website, the macOS host app menu, and the Safari host app. Existing data, accounts, sign-ins, and sync are unchanged — only the display name moves. The marketing site lives at pebbleflow.ai for now and forwards-compatibly displays the new brand; canonical caiioo.ai routing follows in a later release.

New Features

• Common: Simple Mode is on by default for new users: First-run installs now land in Simple Mode — a calmer composer that hides per-message token and cost detail, the lossless-compression dropdown, and tab-context indicators. The model picker is still visible because choosing a model is a first-class action. Existing users keep whatever Simple Mode setting they had. A new eye-icon toggle in the composer reveals or hides the full detail in one tap.
• Common: Account deletion: You can now delete your Caiioo account from the website's Account page (also linked from the iOS app, per Apple's account-deletion requirement). Deletion removes your profile, sign-in credentials, AI credit balance, and the encrypted copy of your provisioned OpenRouter key; minimal compliance audit logs without account ID are retained as required by law. A new privacy-policy section spells out exactly what is removed and what is retained.
• iOS: Native on-device text-to-speech (Kokoro 82M): Kokoro voice synthesis now runs in the iOS host process via OnnxRuntime instead of inside the WKWebView, mirroring the e5-embeddings architecture from 0.9.719. The model gets the host process's increased-memory entitlement budget instead of competing with the sidepanel UI for WebKit's per-process cap, so on-device TTS no longer crashes the WebView under load on iPhone 13/14.

Improvements

• Common: Image generator works with slow streaming models: The image generation tool now opts into streaming for OpenRouter image models, fixing "Network error: Unable to connect to OpenRouter" on slow models like gpt-5.4-image-2 (~167 s end-to-end). The previous buffered path waited on ~1.8 MB of keepalive padding before the actual JSON arrived and the connect-layer timeout fired first.
• Common: Free-tier model selection picks a real model dynamically: The previous openrouter/free meta-router routed inside OpenRouter with no awareness of which downstream models supported tools or vision — Android users sending an image plus a tool call would hit "No endpoints found that support tool use". Caiioo now routes itself: onboarding, settings, and the upgrade-modal "use free models" CTA all pick a real free model that supports tools (and image input where available), and auto-swap on rate-limit or capability errors. Existing users on openrouter/free are migrated lazily on next launch.
• Common: Personal Intuition finds the right context on long messages: The retrieval query was sliced to the last 500 chars of your message and embedded as a single vector. On long pasted-then-asked messages or multi-topic turns, that either truncated before the actual question or blurred everything into a topic-flat centroid that scored badly. Caiioo now extracts a salient query — keeps the verbatim head and appends a deduped bag of content-bearing tokens drawn from up to the next 3000 chars — so retrieval stays on-topic even when your prompt is long.
• Common: Settings → "Minimal settings" toggle stops flipping its own title: The toggle alternated its label between "Minimal settings" (on) and "Advanced settings" (off), so the off state read as if checking the box would move you to advanced — the opposite of what happens. Title now stays constant; the description prefixes "On — " / "Off — " to make the current state unambiguous.
• Common: Voice-model download dialog handles indeterminate progress: The model-info card during download could overflow in the narrow sidepanel; layout is now anchored. When the server doesn't return a Content-Length, the progress block renders an indeterminate spinner instead of a frozen 0% bar.
• Common: Simple Mode shows a compact stat strip instead of hiding everything: Simple Mode previously hid the entire thread-stats summary. It now shows a slimmed strip — context-window ring, compact total tokens (e.g. "1k"), cost, and remaining credit — and the eye-toggle in the composer expands to full detail in one tap.
• Common: Subagent costs roll up correctly: The per-thread cost tally was missing subagent Perplexity / web search spend, helper costs, image/video/music generation costs, PDF OCR costs, and voice costs. All categories now roll up into the parent's sub_agent_cost line.
• Common: Settings → Tools selector overrides hidden default-off tools: Choosing "Always" or "Auto" on a tool that ships disabled by default was being silently ignored. Your dynamic-tool-config choice now overrides the default.
• Common: Local sidecar renamed to "Desktop app" in user-facing copy: All user-visible references to "PebbleFlow Relay" — the sidecar that backs local-stdio MCP servers and Desktop-only tools — now read "Desktop app", which is what users actually install.
• Common: Privacy copy tightened: Dropped redundant "telemetry" wording across legal policies, store listings, the website privacy page, and the in-app guide. "No analytics" already covers it; no factual change to what the apps do or do not collect.
• Common: Provider account view labels balance source: The provider-account panel now states which key each balance is reporting against (your BYOK key vs the Caiioo-provisioned key), so credits and remaining balance can no longer be misattributed at a glance.

Bug Fixes

• Common: Apple Sign-In after the brand transition: Apple Sign-In was failing for new sign-ins because the OAuth Services ID still pointed at the old com.sixcailloux.PebbleFlow.web. Switched to com.sixcailloux.Caiioo.web so "Sign in with Apple" works again on iOS, macOS, and the web account page.
• Common: "Sign in" button on OAuth-required MCP servers: Installing a remote OAuth-required MCP server (e.g. Notion) writes the server to storage even when the initial connect fails because OAuth is required, but no client gets registered. The "Sign in" button then sent a refresh request that returned a raw "MCP server <id> not found" error instead of opening the OAuth dialog. Refresh now falls back to a fresh addServer call when the client isn't registered, so the OAuth/DCR shape is surfaced and the sign-in dialog actually opens.
• Common: Bug reports submitted from native apps now record the correct app version: iOS, macOS, and Android shells now inject the installed app version into the WebView at boot, so submitted bug reports identify which build they came from instead of leaving the field blank or echoing the bundled sidepanel version.
• Android: Launch crash after the rebrand: An over-eager PebbleFlow → Caiioo rename had renamed the Application class and JNI symbols on the Kotlin side without renaming the corresponding native exports, so the app crashed at startup unable to resolve symbols. Reverted the renames; the app launches cleanly again.
• Common: Google Workspace re-authorization loop: Users were stuck re-authorizing Google Workspace every ~hour because their stored OAuth connection had no refresh_token. Six connected fixes guarantee a refresh token on every grant and preserve it through cloud sync, so Google connections survive the 1-hour access-token TTL without bouncing through the consent screen.
• Common: Recovering from "User not found" on OpenRouter: OpenRouter returns HTTP 401 "User not found" when a provisioned sub-key's underlying user record is gone but the key entity still exists. Caiioo previously surfaced this as a dead extension that only logout+login fixed. The provider now self-heals by swapping in a fresh provisioned key and retrying the request once.
• Common: Google Docs insert_component returned misleading errors: Inserting a component after a previously inserted table surfaced "Document not found" because the inserted table never got bound to its componentName. Tables now get a named range in the same insert phase, the position resolver respects non-default tabId, and app-level errors stop being misclassified as 404s.
• Common: Newly released OpenRouter models lost ZDR routing: When a model wasn't yet in Caiioo's intelligence database, the synthetic fallback record marked it as not-ZDR-capable, even when ZDR-only providers actually supported it. ZDR routing now uses the same provider-list fallback as the regular path.
• Common: API /v1/runs ignored attachments on the very first call: When the API endpoint kicked off an agent against a fresh thread, the attachment list on the user message was lost because the empty-thread branch pushed only text. Attachments now flow through correctly.
• Common: Native apps could read stale settings right after a model change: On memory-pressured Android, switching the model picker and immediately sending a message could let the agent read the previous model from disk because the 500 ms debounced flush hadn't fired yet. The send path now forces a state flush before spawning the agent.
• Common: Tester-bug triage (PF-260429 / PF-260430): Five fixes — managed-key (free-tier) users can now ingest documents through PDF OCR without typing their own key; macOS pins the Node sidecar's timezone to the host so dates resolve correctly under sandbox; scheduled tasks now persist on iOS/macOS/Tauri shells (the WebView storage stub was silently dropping writes); the configuration tool surfaces one-time and manual schedules as first-class options; and tool callsites recover from the OpenRouter 401 self-heal the same way the agent runner does.
• Common: Upgrade modal stops mixing tier subscription with credits/BYOK: "Add Credits" actually opened the subscription portal — the label lied. The modal now focuses on tier subscription only; iOS additionally hides any credit-purchase surface per Apple §3.1.1.
• Web: Delete Account section moved to the bottom of the account page: The destructive Delete Account block used to render inline between your identity card and the subscription/billing UI. It now lives at the very bottom of the page, after the FAQ. iOS deep-link behavior into the delete-only view is unchanged.
• iOS: Manage Plan now opens Apple's subscription sheet: Manage Plan on iOS previously fell through to the web account page for free, trial, Stripe, and unknown-source users, where Google OAuth in WKWebView would fail. Tapping Manage Plan on iOS now always opens Apple's StoreKit showManageSubscriptions sheet.
• iOS: Delete-account web view collapses to a delete-only page: When the iOS Delete Account button opens the website, the page now hides every billing surface (subscription, plans, credits, Stripe portal, FAQ) and shows only the deletion section, satisfying §3.1.1.
• iOS: ITMS-90208 framework-version validation fix: The onnxruntime framework's Info.plist is now patched at archive time to match the host app's deployment target, so App Store Connect stops rejecting builds with "framework does not support the minimum OS version specified in the Info.plist".
• Android: External links open in a Custom Tab so Google OAuth works: Tapping links like "Open pebbleflow.ai/account" used to open inside the Android WebView, where Google blocks OAuth with disallowed_useragent (Error 403). External links now route through the native bridge into a Chrome Custom Tab, which Google trusts.
• Android: Copy buttons under messages actually copy: The copy icon used navigator.clipboard.writeText directly, which silently no-ops in the Android WebView when user activation expires across the async boundary. Copy now routes through the native clipboard bridge.

Security

• Common: protobufjs CVE-2026-41242 patched (CVSS 9.8): Pinned protobufjs >= 7.5.5 (resolved to 8.0.1) to close a critical arbitrary-code-execution hole in Root.fromJSON. Caiioo doesn't import protobufjs directly — it's a transitive of onnxruntime-web — and risk-in-practice was low because we only feed bundled models, not user-supplied schemas. Patched anyway because the fix is trivial.
• Common: @xmldom/xmldom — 4 high-severity CVEs patched: Pinned @xmldom/xmldom >= 0.9.10. XML parsing is used in DOCX/XLSX redline pipelines and various extension code paths.
• Server: undici — 3 high-severity CVEs patched: Pinned undici >= 7.24.0 in cloud/relay. Affects the relay server's HTTP client only.
• Desktop: rustls-webpki + quinn-proto — 2 high-severity CVEs patched: Bumped these Tauri Rust dependencies. Affects the Windows + Linux desktop builds that go through the Tauri shell.

New Features

• Common: Video generation (Pro): Generate or animate short videos with Google Veo 3.1, OpenAI Sora 2 Pro, and ByteDance Seedance via OpenRouter. The tool picks valid durations and resolutions per model, polls until the job completes, and saves the result as a thread attachment.
• Common: Music generation (Pro): Generate songs and instrumental clips with Google's Lyria 3 Pro Preview via OpenRouter. Output is saved as an audio attachment that plays inline.
• Common: Dynamic video model catalog: The video generator fetches the current list of video-capable OpenRouter models at runtime, so new providers and models appear without a Caiioo update. A bundled snapshot keeps things working offline.
• Common: Dynamic music model catalog: Same for music generation — the tool picks up new music models as OpenRouter publishes them, with offline fallback.
• Common: Custom OAuth at Pro: Bring-Your-Own-Auth — the Google Workspace wizard, Microsoft 365 wizard, and the generic "Add Custom Provider" flow — is now visible to all Pro, Platform, Teams, and Enterprise users in Settings → Custom OAuth. Previously the tab and add buttons were hidden behind tester-only flags, so paying users couldn't reach BYOA setup.
• Common: Physics + Structural Analysis (Pro): The physics simulation tool (projectile motion, collisions, kinetic/potential energy, momentum, force, impulse, velocity-to-target) and structural analysis (beam loading, column buckling, material properties) is now available at Pro alongside the other creative and utility tools.
• Common: Seeing-Eye Dog — vision fallback for text-only LLMs: Text-only models like DeepSeek V4 Pro, Kimi K2.6, MiMo V2.5 Pro, and local Ollama models can now handle image attachments by routing them through a configured cheap vision model (default: Gemini 3.1 Flash Lite). Auto-captioning fires at message-build time and caches per attachment so subsequent turns don't repay; a dedicated vision({action: "inspect"}) tool gives the model targeted follow-up access. Settings → Tools → Vision Fallback Model chooses the helper.
• Common: XLSX cell-level tracked changes: Spreadsheet artifacts now support the same redlining UX as DOCX. AI proposals via propose_change(editMode: 'xlsx_cell') produce cell-located tracked changes anchored by cellRef + sheetName; user typed edits in track-changes mode bake cell-level diffs; cells with pending changes render <del>old</del><ins>new</ins> inline; the existing toolbar's next/prev/accept/reject works on cell changes; concurrent AI + user edits merge cell-by-cell with user-wins on same-cell conflicts.
• Common: Cost tracking for video and music generators: Generated videos (via OpenRouter /api/v1/videos) and music (via chat-completions) now roll their cost into thread totals just like image generation, with new video_gen_cost and music_gen_cost breakdown rows in the sidepanel cost dropdown.

Improvements

• Common: GPT-5 series stability: OpenAI's gpt-5, gpt-5.1, gpt-5.4, gpt-5.4-pro, and gpt-5.3-codex no longer produce silent empty responses on tool-heavy agentic conversations. A function-tool schema interaction was causing OpenAI's backend to terminate streams without producing any output; Caiioo now serializes tools in the shape these models require.
• Common: Better long-conversation behavior on GPT-5.4+: Multi-turn conversations with gpt-5.4, gpt-5.4-pro, and gpt-5.3-codex no longer early-stop on long tool-calling sequences — the phase marker OpenAI uses to distinguish intermediate commentary from final answers is now preserved across turns.
• Common: GPT-5.x reasoning visible in the thinking panel: gpt-5, gpt-5.1, gpt-5.4, gpt-5.4-pro, and gpt-5.3-codex now stream their reasoning summary into the in-chat thinking panel as they think, matching how Gemini already behaves. Previously the thinking panel stayed empty for these models even though OpenRouter was streaming reasoning text.
• Common: Transparent recovery from transient upstream outages: When OpenRouter routes a request to an upstream that returns a transient 5xx error before any content streams, Caiioo quietly retries on a different upstream instead of surfacing an empty reply.
• Common: Clearer errors when the model stream fails: Provider-side crashes, content-filter rejections, and other mid-stream failures now surface with a specific error message instead of the conversation just "stopping" with no explanation.
• Common: Generated videos and music render inline: Generated videos and music now appear in the conversation like generated images — a video or audio player surfaces above the final answer with a small badge showing which model produced it, instead of being tucked inside the collapsed tool section.
• Common: Default-model picker for video and music tools: Settings → Tools now includes a Default Model dropdown for the video and music generators, mirroring the existing image-generator picker. The agent uses your selected model by default; you can still override per request by naming a different model.
• Common: Personal Intuition indexed-status display: The Personalization settings panel now shows a live "Indexed: N threads · M chunks · K tokens · last indexed Xm ago" line so you can confirm the memory indexer is doing its job. The last backfill summary also stays visible after the run completes instead of vanishing.
• iOS: Native on-device embeddings (faster, less memory): Multilingual-e5-small inference now runs in the iOS host process via Apple's onnxruntime-objc instead of inside the WKWebView. This solves a per-process memory cap that was killing the WebView ~7 seconds after model load (the post-login crash). After the first encode of the tool catalog, subsequent tool-selection calls hit a per-text in-memory cache and complete in ~10 ms instead of ~5 s. The shared-pipeline refactor also stops Personal Intuition and on-device tool selection from each loading their own ~115 MB copy of the model.
• Common: Personal Intuition + on-device tool selection share one e5 pipeline: Both features previously instantiated their own multilingual-e5-small loader (~115 MB each, ~230 MB total in the renderer). They now delegate to a single shared pipeline that dedupes concurrent loads, halving cold-start cost.

Bug Fixes

• Common: Browser-session cookies no longer leak into API calls: If you'd visited openrouter.ai in a browser tab, the extension was unintentionally attaching your OpenRouter browser session (Clerk / Stripe / analytics cookies) to every API call. API calls are now strictly Bearer-authenticated.
• Common: GPT-5.4 responses no longer appear twice: Fixed a bug where gpt-5.4, gpt-5.4-pro, and gpt-5.3-codex responses showed the same text back-to-back — the final-answer payload was being accumulated alongside the token stream that had already rendered it.
• Common: Model picker refreshes after reload: The model list is now invalidated on version upgrades and honors a short in-memory TTL, so newly released OpenRouter models appear after a reload instead of being hidden behind a stale cache. Long-lived service workers no longer hold onto a frozen catalog.
• Common: Generated-media short links open the player: When a model writes a short link like [Listen](audio-…) after generating audio/image/video, clicking it now opens the slate player instead of failing as a broken external URL. Same fix applies to image and video links.
• Common: Generated media as markdown image embeds rendered broken: When the model wrote a video or audio attachment as ![Video 1](video-…) instead of [Video 1](video-…), it rendered as a broken <img> instead of opening the player. The leading ! is now stripped so the link opens the slate viewer.
• Extension: Sidepanel "Location permission denied" dead-end: Fixed a regression where every sidepanel location request returned "Location permission denied. Please allow location access in browser settings." regardless of what the user clicked. An attempt to route through the offscreen document hit PERMISSION_DENIED instantly because the offscreen context can't show a permission prompt and the manifest doesn't declare geolocation. The sidepanel now uses the original content-script-then-IP-geolocation chain that worked before.
• Common: Personal Intuition full rebuild left stale indexed counts: Fixed full-rebuild backfill skipping the per-thread index update, so Settings → Personalization showed inflated pre-rebuild totals after a rebuild. Threads scanned by full-rebuild but not previously indexed are also no longer invisible to retrieval.
• iOS: Post-login WKWebView crash loop: Three converged fixes for the post-login crash. (1) Native e5 embeddings now run chunked in batches of 8 with the ORT memory arena set to shrink between runs — peak working set stays under ~100 MB instead of spiking to ~3 GB and tripping iOS's per-process memory kill. (2) The on-device retriever now warms during requestIdleCallback instead of inline at boot, so it no longer competes with license sync, identity restore, UI render, and cloud-sync init. (3) The on-device retriever now re-provisions on identity / tier change, so fresh installs no longer silently fall back to cloud helpers because tier was undefined at first registration.
• iOS: Bogus "path traversal blocked" 403s: Fixed LocalFileSchemeHandler mis-flagging every 404 as a path-traversal attempt because NSString.standardizingPath only resolves /var → /private/var for files that exist on disk. Legitimate paths to non-existent resources (the iOS bundle excludes *.wasm, plus chrome-extension API paths like api/active-tab-context) returned 403 instead of 404, breaking transformers.js' wasm pre-fetch fallback. Replaced with a string-based .. / NUL check matching the Android handler.
• Common: PDF embedded images broke text-only models: Sending a PDF with embedded images to a text-only OpenRouter model (DeepSeek V4 Pro, Kimi K2.6, etc.) was hitting "No endpoints found that support image input" — the warning the new Seeing-Eye Dog routing was supposed to eliminate. The PDF delivery path now respects the same per-model supportsVision flag that image-block delivery already honored, so text-only models receive a text-only PDF and the agent can spawn a vision subagent if needed.
• Common: Ad-blocker level toggle didn't take effect: Toggling the ad-blocker level (Off / Standard / Aggressive) only changed the persisted value while the live DNR rules + static ruleset stayed in effect until the service worker happened to restart — so after toggling to Off, sites like ads.google.com remained blocked. The storage listener now watches the globalSharedSettings bucket where the setting is actually written, and the legacy adBlockerEnabled flag routes through the same bucket.
• Extension: Oversized images were dropped silently: When an image attachment exceeded the API cap, the extension service worker had no compress impl registered (only the server's sharp-backed one was) — so the defense in pushImageBlock caught the throw and dropped the image entirely. The SW now probes natural dimensions via createImageBitmap and routes the encode through the existing offscreen document, so extension-context callers get the compressed image instead of a dropped placeholder.
• Common: Generated images and screenshots rejected by providers: Anthropic caps base64 images at 5 MB; other providers have similar limits. Generated images from FLUX / Gemini / Seedream and large screenshots were being passed through at full size, producing 4xx errors that aborted the entire agent turn. Three layers of defense: image-generator compresses output before storage; a new pushImageBlock helper routes every image_url emission through compress-or-drop; and screenshot rebuild + live screenshot injection both go through the same helper. Conservative 4 MB cap that works on every provider.

New Features

• Common: Personal Intuition — associative cross-thread memory: Caiioo quietly remembers what you've talked about before — across every conversation, not just this one. Relevant memories come to mind as you chat; your agent might use one directly, or just let it color the response. Like how a scent can remind you of a place. Find it under Personalization → Personal Intuition; off by default. Includes Full-rebuild and Incremental backfill so you can index your existing threads.
• Common: Personal Intuition backup & restore: You can snapshot the full memory corpus to a JSON file under Backup & Restore → Personal Intuition, and restore it later if something goes wrong or you want to move it to another device.
• Common: Instant Tool Chooser default for every tier: The semantic tool chooser now runs locally on every device, on every tier — free included. Picks the right tools for each turn in ~10 ms, fully on your device. The picker UI in Settings → Tools clarifies the choice as "Instant Tool Chooser" vs "Quick Tasks LLM" (which uses whichever model you've marked with the lightning bolt in the model picker).
• iOS / Android: Instant Tool Chooser in the native apps: The same on-device tool chooser is now provisioned inside the iOS and Android apps' WebView, not just the browser extension and desktop.

Improvements

• Common: Slate sandbox allows HTTPS script CDNs: HTML slate artifacts can now load common libraries like Chart.js or D3 from reputable HTTPS CDNs (jsdelivr, unpkg, etc.). Data fetching is still locked to same-origin, so use bind_data to pipe attachment data into a slate.
• Common: Clearer Personal Intuition progress: Backfill now shows per-thread progress in the settings panel (e.g. "47/75 threads · 2,134 chunks") while it runs, instead of going silent until completion.
• Common: Faster tool selection after memory indexing: Indexing a large memory corpus no longer evicts the tool-catalog cache in the on-device retriever — the next tool call stays warm rather than paying a re-encode cost.

Bug Fixes

• Common: Concurrent sub-agents hit false tier-upgrade errors: Fixed a race where two sub-agents running in parallel could overwrite each other's active-thread context, causing Pro-gated actions (update_slate, etc.) to be rejected for the wrong sub-agent. Tier enforcement is now scoped to the specific thread of the dispatching call.
• Common: Slate CSV data bindings broke forEach: Fixed CSV-bound data arriving in the sandbox as a non-iterable object — data.forEach(...) inside sandbox JavaScript now works as expected.

New Features

• Common: Unified Pro Tier Across All Apps: Pro is now a single $9/mo subscription that unlocks every Caiioo app — Chrome, Desktop, and Mobile — instead of separate app-tier buckets. Platform ($14/mo) is repositioned as "Pro + infrastructure" (local server, API access preview, Messaging).
• Common: Pro Mobile Tier: A mobile-only Pro plan is available for $2.99/mo, sold directly in the App Store and Play Store. The in-app purchase itself is the entitlement — no separate license needed.
• Common: Ask-User Tool for Human-in-the-Loop: The agent can now pause mid-run and surface a four-way decision dialog (approve, approve with notes, reject, reject with notes). Your notes flow back to the model as plain-English guidance that overrides the proposed plan — no new cycle fires, the agent continues in place.
• Common: In-App Guide Search: Search the user guide directly from the Document menu. Results deep-link to caiioo.ai/guide, preserving the existing redirect flow.
• iOS: Monthly/Yearly Paywall Toggle: The iOS Subscribe sheet now lets you switch between Monthly and Yearly billing before purchase.
• Desktop: Auto-Updates on Windows and Linux: The Tauri desktop app now ships with the updater enabled, so Windows and Linux builds can receive updates in place instead of requiring a manual reinstall.
• Web: Tabbed Search Across Guide and Blog: The marketing site's guide and blog layouts now include a search bar with tabbed scope switching — title matches rank above body matches, and the active tab auto-switches to where the results are.
• Web: Dedicated Linux Install Page: The install page routes Linux users to /install/linux, which lays out AppImage, .deb, and .rpm choices with per-distro commands.
• Web: Stable Download Routes: /download/macos, /download/linux, and /download/windows are now stable, edge-worker-backed URLs that always resolve to the latest release — no website rebuild required per release.

Improvements

• Common: In-App Support Ticketing: Support requests now route through the in-app ticket flow as the primary path, with clearer credit-error messaging pointing you to the right place when something goes wrong.
• Common: Smarter Google OAuth Re-Auth: Google tools now verify the scopes actually granted by Google and trigger a just-in-time re-authorization prompt when a 403 indicates a missing scope, instead of silently failing.
• Common: Full Drive Scope for Google Writes: Write actions against Google Drive now request the full drive scope so edits to documents you didn't create succeed instead of hitting permission errors.
• Common: Image Generator Model Descriptions: The image generation tool surfaces a description for each available model so it's easier to pick the right one for the task.
• Common: More Accurate Slate Artifact Messages: When a tool creates or updates a Slate artifact, the model no longer claims the artifact is "displayed in the editor" — the wording now reflects what actually happens.
• Common: Better Dynamic Tool Selection from Tabs: The tab-context hint now emits tool IDs, so the dynamic tool selector can actually apply context-based selection rules that depend on which tools are available.
• Common: Voice Costs Attributed to Threads: Streaming TTS (ElevenLabs, Resemble) and STT (ElevenLabs Scribe) usage now rolls up as voice cost on the conversation where it happened, matching the existing one-shot synthesis path.
• Extension: Login Subtitle Emphasizes Local Agents: Small copy update on the login screen.

Bug Fixes

• Desktop: Linux File Picker Restored: Fixed the file picker failing on Linux in the Tauri desktop app.
• Desktop: Version Reporting Wrong: Fixed the Tauri desktop app reporting an incorrect version string in the UI and telemetry.
• Desktop: Quieter Linux Audio and Clearer Errors: Suppressed PipeWire log noise on Linux, surfaced OpenRouter errors instead of swallowing them, restored the API toggle read path, and fixed an incorrect parity banner.
• Common: Gmail Body Extraction Lost Links: Reversed the Gmail extraction precedence to try link-preserving markdown first, then fall back to plain text only when the markdown output is empty — emails no longer lose links in the common case.
• Common: Public API Settings Not Persisting: Fixed the Public API toggle and access token failing to save on some setups, and ensured these settings remain strictly local rather than syncing to the cloud.
• Common: Archive and Rename Felt Laggy: Delete, archive, unarchive, and rename now update the UI optimistically so the list reflects the change immediately instead of waiting for a server round-trip.
• Common: Archive Deletes Didn't Sync: Fixed archive deletions being applied directly to storage, bypassing the protocol path — they now propagate across clients like every other thread action.
• Common: Agenda Showed False "Missing Scopes": Fixed the reminders/agenda widget reading a stale tool-context profile right after you granted Google Calendar scopes, so it flagged missing scopes while tools worked fine. The UI now reads the fresh profile, and a refresh auto-runs the incremental-scopes consent flow sequentially across any connected Google accounts that still need it — no need to open the sync dialog to find "grant access".
• Common: Google Tool Calls Rejected with "Unknown name exclusiveMinimum": Fixed Gemini rejecting tool calls whose parameter schemas used numeric exclusiveMinimum / exclusiveMaximum bounds — these are now stripped before sending to Google's restricted OpenAPI subset.
• Common: Remote MCP Servers Without DCR Failed to Connect: Fixed adding remote MCP servers that don't support Dynamic Client Registration (e.g. Slack): the client now respects RFC 9728 protected-resource-metadata, follows authorization_servers to the AS host when it differs from the resource host, and stops fabricating a /register endpoint that doesn't exist.
• Server: Tester-Tier Users Rejected by Gated Endpoints: Fixed organizations, hub, admin, and provisioned-key endpoints rejecting users whose effective tier (from licenses) was higher than the stale tier snapshotted into their auth token. Gated endpoints now evaluate the effective tier consistently with /api/auth/me.
• Extension: Identity-Sync Ping-Pong Loop: Fixed conflicting identities between the extension and relay server looping forever on IDENTITY_SYNC. The extension now acks once against the same conflicting remote, warns that you must sign out on one side to resolve, and resets on disconnect / match / remote adoption.
• Common: Gemini Flash TTS Couldn't Be Selected or Configured: Fixed Gemini Flash TTS (voice) being unusable end-to-end — selecting it reverted on settings close, and the API key field appeared to not accept typing. The provider is now a first-class TTS option: selection persists, the API key + voice picker save correctly, and synthesis uses an SSE streaming endpoint so audio playback can begin before the full utterance finishes generating.
• iOS: New-Thread UI Didn't Switch: Fixed the iOS app failing to switch the active view to a newly created thread, and fixed a related modelName ReferenceError that could crash the composer.
• iOS: Startup Crash on Optional Storage Hook: Fixed an iOS crash when storage.primeSettingsCache wasn't available during startup — the call is now guarded.
• Common: Tier-Audit Discrepancies After Pricing Reshape: Fixed six pricing and feature inconsistencies surfaced by the tier audit, so entitlement gates, upgrade prompts, and feature flags now agree across the extension, website, and mobile apps.

iOS App Store Compliance

• iOS: External-Signup CTAs Hidden in Onboarding: Onboarding no longer shows calls-to-action that point users off-device for account creation or paid upgrades.
• iOS: Credit Purchase UI Hidden: Credit purchase surfaces are hidden on iOS per Apple's §3.1.1 rules; subscriptions route through StoreKit instead.
• iOS: Support and Credit-Error Wording Updated: The /support page and credit-error messaging have been reworded on iOS to meet Apple's guidelines on external purchase references.

New Features

• Common: Gemini 3.1 Flash TTS: New voice-output option powered by Google's Gemini 3.1 Flash TTS. 30 prebuilt voices, 70+ languages, and inline audio tags like [whispers] and [sighs] for expressive synthesis. Bring your own Gemini API key; get one free at aistudio.google.com. Audio is watermarked with SynthID.
• Common: Voice Cost in Conversation Total: Text-to-speech and speech-to-text API usage (Gemini, ElevenLabs, Resemble) now rolls into the per-thread cost summary alongside chat, image-gen, helper, and OCR costs — so the number you see is the real number.
• Common: SQL Database Tool: Attach SQLite databases to conversations and query them with SQL. Results can be rendered as live dashboards in Slate artifacts.
• Common: Calendar Invitations and Google Meet: Creating calendar events now sends invitation emails to attendees and can automatically add a Google Meet video link.
• Common: Per-Tool Provider Routing: Route individual tools (image generation, search, PDF OCR, etc.) to specific providers via Settings, instead of everything going through the default model.
• Common: Conversation Sharing: Share conversations as Markdown or plain text.
• Common: 14-Day Platform Trial: Start a 14-day Platform tier trial from the extension UI or the website account page. Pro subscribers are also eligible.
• Common: ZDR Quick-Filter in Model Picker: New button in the model selector dropdown to quickly filter for zero-data-retention models.
• Common: Simplified Settings: Settings now opens in a streamlined Simple view by default, showing just the essentials. Switch to Advanced mode anytime to see everything.
• Common: Free Google Gemini Access: Bring your own Google AI Studio key to use Gemini 2.5 Flash for free — Google's free tier is now available to all users, no paid plan needed.
• Common: Baseten Self-Hosted Models: Connect your own Baseten deployments to use self-hosted models directly in caiioo.
• iOS: WhisperKit On-Device Speech-to-Text: Dictation now runs fully on-device via WhisperKit — audio never leaves the phone.
• Android: On-Device Whisper STT: Speech-to-text via whisper.cpp runs fully on-device on Android with microphone permission flow.
• Android: GPU-Accelerated Moonshine STT: On GPU-capable Android devices, dictation can run on Moonshine with sub-second inference. English only; whisper.cpp remains the multilingual fallback.
• Web: Sign in with Apple on Account & Hub Pages: Apple Sign-In is now available on the website account and hub pages.

Improvements

• Common: Better PDF Export: Exporting a PDF from Slate now produces a true PDF file instead of opening the print dialog, so you get a clean document every time.
• Common: Google Slides Full Text: Requesting slide text without specifying a page now returns text from all slides in the presentation.
• Common: Smarter Web Search: Google search results are now extracted with better structure, capturing titles, snippets, and links more reliably after recent Google layout changes.
• Common: Kokoro Text-to-Speech on All Platforms: The Kokoro voice option now appears on iOS, macOS, and Android — previously it was hidden on native apps.
• Common: Cleaner Credit Balance Display: Account balance for prepaid and bring-your-own-key setups now shows your actual balance without confusing "Limit" framing.
• Common: Higher-Fidelity PDF to Word Conversion: PDF→DOCX export now uses a dedicated Document view in Slate, producing Word files that more faithfully preserve layout and structure from the source PDF.
• Web: Quieter Account and Auth Pages: Google Analytics and the cookie consent banner are suppressed on account and auth pages for a cleaner sign-in flow.

Bug Fixes

• Common: Settings Search Didn't Navigate: Fixed the settings search dropdown and deep links failing to jump to several sections (API Access, Private Sync, Backup & Restore, Voice, and others) — the target category stayed collapsed so nothing scrolled into view.
• Common: OAuth Prompts in Sub-Agents: Fixed OAuth authorization and tier-upgrade prompts being silently dropped when triggered from a sub-agent, which broke connect flows mid-run.
• Common: Ollama Model Not Saved After Onboarding: Fixed the Ollama model you selected during onboarding not persisting afterward.
• Common: UI Chunks Failed to Load During Rate Limiting: Fixed static UI assets being rate-limited alongside API requests, causing blank screens or missing panels until refresh.
• Common: Composer-Preprocessed Attachment Metadata Lost: Fixed attachment metadata being dropped on the server for documents preprocessed in the composer before being sent to the model.
• Common: Free Time Counted Attended Events as Busy: Fixed the agenda free-time calculation marking you busy during events you only attended — only events on your own calendar are now counted.
• Common: Google Search Results Broken: Fixed web search returning empty results after Google changed their search page layout.
• Desktop: Windows/Linux Login Out of Sync with Other Platforms: Fixed the Tauri desktop login flow diverging from Chrome/macOS/iOS — including incorrect tier detection and being unable to log back in after signing out.
• Desktop: Public API Blocked the Desktop UI: Fixed the desktop app's own UI being blocked by API authentication when the Public API setting was enabled.
• Common: Google Docs OAuth Errors Hidden: Fixed Google Docs operations silently failing when authentication expired instead of showing a clear error.
• Common: Settings Panel Crash: Fixed the settings panel crashing when restoring a provisioned API key.
• Common: Google Drive Sync Duplicating Documents: Fixed Google Docs being duplicated on every sync run instead of updating in place.
• Common: Google Drive Sync Inline Code Lost: Fixed inline code formatting being lost when syncing documents to Google Docs.
• Common: Gemini Thinking Extraction: Fixed thinking/reasoning content not being captured correctly from Gemini model responses.
• Common: Profile Switch Stale Data: Fixed switching between profiles sometimes showing conversations from the previous profile until a manual refresh.
• Common: Task Scheduler "Method Not Found" Error: Fixed scheduled tasks throwing errors every 60 seconds in certain setups.
• Common: Text Insertion Newlines in Google Docs: Fixed literal \n appearing in text insertions instead of actual newlines.
• Common: Google Account Hint on Tools: Fixed Google tools sometimes using the wrong Google account for API calls.
• Common: Settings and Thread Import Round-Trip: Fixed settings and thread exports not importing back correctly.
• Common: Sub-Agent Token Usage: Fixed sub-agent token usage not being counted in the parent conversation.
• Common: Messaging Bridge Reply During Retry: Fixed messaging bridge replies failing when the assistant retried a response.
• Common: Slate Redline Positioning: Fixed redline deletions landing at the wrong position when markdown was present.
• Common: Browser Connection Survives Sleep/Wake: Fixed "No browser connected" errors after the computer sleeps or sits idle for long periods.
• Common: Scheduled Tasks Survive Service Worker Sleep: Fixed recurring tasks stopping after long idle periods.
• macOS: Safari Google Search Failures: Fixed Google search not working in the Safari extension on macOS.
• iOS: Mobile App Access: Fixed the iOS app incorrectly requiring a paid plan — the mobile app is now available to all users.

New Features

• Common: Hub Delivery System: First-party tool definitions, mode configs, and MCP catalog are now prepared to be delivered from the cloud hub instead of being hardcoded in the bundle, with local caching and offline fallback in preparation for community launch.
• Common: Single-Tab Mode for Browser Tools: New per-mode browser tab policy reuses a single tab for all navigation in a conversation, preventing tab accumulation during messaging-heavy workflows like WhatsApp.
• Common: API Integration Binary Uploads: The api_integration tool now supports binary file uploads and direct local file uploads via presigned URLs, instead of always JSON-stringifying request bodies.
• Common: Credential Management via Tool: Save, list, and delete API credentials programmatically through the api_integration tool, previously only possible through the Settings UI.
• macOS: Audio Input Device Selection: Choose which microphone to use for voice input — lists all available audio devices and persists your selection.
• Android: OAuth Recovery on Low Memory: Android OAuth flows now survive activity recreation when the system kills the app for memory, recovering the auth session seamlessly.
• Common: 3-Layer Mobile CSS Architecture: New mobile-responsive CSS system with platform, layout, and component layers, ensuring consistent UI across phone, tablet, and desktop viewports.

Improvements

• Extension: Incremental OAuth via Tab: Incremental scope requests now open a full browser tab instead of a popup, fixing hangs on accounts that require interactive consent.
• Common: MCP Tool Schema Quality: All MCP tools now pass a quality lint test verifying complete schemas, proper descriptions, and consistent parameter definitions.
• Extension: Auto-Close Browser Tabs on Stop: Browser tabs opened by the assistant are now closed when a run is stopped, and OAuth Safari windows are closed on macOS.
• Android: Toolchain Upgrade: Android build upgraded to AGP 9.1, Kotlin 2.2.10, and Gradle 9.3.1.

Bug Fixes

• Common: Google Docs Table Positioning: Fixed mutations landing in wrong positions when tables precede the target text in Google Docs.
• Common: Page Content Footer Stripping: Fixed page content extraction incorrectly stripping footer elements from captured pages.
• Common: Web Browsing Pagination: Fixed pagination broken for page 2+ when no explicit maxLength was specified.
• Common: Content Script Fallback: Fixed page content extraction failing silently by falling back to executeScript when the content script is unavailable.
• Common: PDF Rendering for URL-Ingested Documents: Fixed PDF native rendering failing for documents ingested via URL.
• Common: Gemini Tool Call Parsing: Fixed double-quoted keys in LLM tool call arguments causing parse failures with Gemini models.
• Common: Amazon Nova Compatibility: Fixed tool_call messages rejected by Amazon Nova due to blank text fields.
• Common: Agent Run Not Terminating: Fixed agent runs not calling endRun reliably in service worker context, leaving browser tabs and state uncleaned.
• macOS: Voice Audio Resume: Fixed voice audio not resuming after macOS app sleep or dock minimize.
• macOS: Screenshots Not Appearing: Fixed macOS screenshots not appearing in the assistant UI due to incorrect context detection.
• macOS: Location Not Working: Fixed location broken in Safari and native macOS app by adding CoreLocation bridge and fixing IP fallback.
• macOS: Location Entitlement Missing: Fixed missing location entitlement and increased native location timeout.
• iOS: File Picker Broken: Fixed file picker not opening in iOS WKWebView by adding the runOpenPanel delegate.
• Extension: OAuth Scope Over-Granting: Removed include_granted_scopes from web popup and post-login OAuth paths to prevent unintended scope inflation.
• Extension: OAuth Consent Screen Skipped: Fixed OAuth scope elevation skipping the consent screen during incremental scope requests.
• Desktop: Console Window Visible on Windows: Fixed console window appearing when launching the Tauri sidecar on Windows.
• Desktop: Windows Build Broken: Fixed node binary path resolution and flattened node_modules for Windows Tauri builds.
• Server: Duplicate Webhook Processing: Fixed webhooks being processed multiple times when multiple clients were connected.
• Server: Promotion Codes on Credit Purchases: Disabled promotion codes for credit purchases in Stripe checkout.

New Features

• Common: Redesigned Onboarding: New onboarding screen presents 6 clear AI access paths, making it easier for new users to understand their options for connecting to AI providers.
• Common: PDF from Markdown: New documents created from scratch can now be exported as PDF, converting markdown content to formatted PDF output.
• Common: Unified Google OAuth: Google account connection now uses a single verified OAuth app with all scopes, replacing the previous multi-app setup. Simpler connection flow with fewer prompts.
• Common: Save to Slate Meta-Parameter: Tools can now output directly to a slate document via the _save_to_slate parameter, capturing structured results without copy-paste.
• Common: Slate Anchor Links: Chat messages with slate references now include clickable anchor links that scroll to the relevant section within the document.
• Common: Sandbox Data Bridge: HTML artifact sandboxes can now dynamically access slate data via a postMessage bridge, enabling live dashboards and interactive visualizations.
• Common: Tier Badges: Settings, tools, and automation features now display tier badges showing which access level is required, giving expanded-access users visibility into gating.
• macOS: Native Speech-to-Text (WhisperKit): On-device speech recognition via WhisperKit and CoreML, replacing browser-based STT with a faster, private, native implementation.
• macOS: Sparkle Auto-Update: Production-ready automatic updates for the macOS app and Safari extension via Sparkle, with appcast feed and delta updates.
• macOS: Native Tab Context: AppleScript-based tab context extraction provides the macOS app with awareness of frontmost browser tabs.
• macOS: Voice Hotkey Streaming: Voice hotkey now shows a real-time streaming overlay during recording with automatic WhisperKit model loading.
• Safari: Native Messaging: Safari extension now communicates with the macOS app via native messaging instead of HTTP polling, improving responsiveness and reliability.
• iOS/Android: Native Save Dialogs: JSON export now uses native save dialogs (NSSavePanel on macOS, SAF on Android) instead of broken blob downloads.
• Web: Teams Landing Page: New /teams page with team and enterprise messaging for prospective customers.
• Web: Blog Section: New blog section on the Caiioo website.
• Web: Free Trial Activation: Account page now supports free trial activation with aligned onboarding flow.
• Common: Shared Drive Support: Google Picker and all Drive API calls now support shared drives, enabling access to team-shared documents.

Improvements

• Common: Adaptive Greeting: Adaptive greeting now riffs on custom welcome messages instead of ignoring them, preserving the user's personalized tone.
• Common: OAuth Callback Cleanup: OAuth callback tabs now auto-close with a countdown timer instead of staying open.
• Common: OAuth Guidance: When a tool needs an account connection, the error message now includes actionable guidance on which account to connect.
• Common: Drive Sync Force-Rewrite: Drive Sync workflow now supports a force-rewrite option for re-syncing all files.
• macOS: Unified Debug Log Viewer: Debug log view now aggregates server, Swift, and UI logs into a single chronological stream.
• macOS: Bucketed Settings Storage: Server storage adapter refactored to bucketed architecture with robust migration, persistent markers, multi-profile atomicity, and first-writer-wins conflict resolution.
• macOS: Notarized Safari Extension: Safari extension is now signed with Developer ID and notarization-compatible entitlements, removing the "Allow Unsigned Extensions" requirement for distributed builds.
• iOS: Default to Apple Reminders: iOS and macOS now default to Apple Reminders and fetch them in the sync view automatically.
• Common: OAuth Scope Elevation: Incremental scope requests no longer display all 22 scopes — only the newly requested scopes are shown, with cleaner account selection prompts.
• Common: Google Account Auto-Retry: When a document returns 404, other connected Google accounts are automatically tried before surfacing an error.
• Common: Settings Deep-Link Navigation: All settings sections are now properly registered for search and deep-link navigation.

Bug Fixes

• Common: Tool Misrouting: Fixed tools being misrouted when models use colon-separated tool:action format in their responses.
• Common: Table Cell Formatting: Preserved formatting and links in table cells and cleared inherited list numbering that leaked between cells.
• Common: Slate Highlight Scroll: Fixed highlight scrolling in TipTap editor for background tabs, preview mode, and native views using reliable ProseMirror-based positioning.
• Common: Code Block Text Search: Fixed fallback to text search when structural block mapping fails for code blocks in slate documents.
• macOS: Settings Lost on Mode Switch: Global settings (API keys, connections) are no longer cleared when switching modes on macOS/iOS — an explicit allowlist now controls which settings reset.
• macOS: Storage Purge Parity: Full storage purge now clears all state fields consistently, matching chrome.storage.local.clear() behavior. Profile deletion and retention cleanup also route through buckets correctly.
• macOS: Settings Concurrency: Settings reads in applyModeDefaults now acquire the lock first, preventing race conditions. Storage stats read from buckets instead of stale flat cache.
• macOS: Migration Robustness: Storage migration retries on failure, uses persistent completion markers, and cleans up stale pre-migration data automatically.
• macOS: MCP Server Zombie Processes: MCP server child processes are now properly terminated when the macOS app quits.
• macOS: Calendar/Reminders in Agenda: Fixed Apple Calendar events and Reminders not appearing in the agenda modal.
• macOS: Google OAuth Flow: Fixed Google login failing in macOS WKWebView by routing OAuth through the relay server web flow.
• macOS: Sign-Out Navigation: Fixed sign-out not returning to the login screen on macOS.
• macOS: Remote Access Auth: Fixed 401 error on set-remote-access endpoint by removing incorrect auth requirement.
• Server: WhatsApp Reply Delivery: Fixed WhatsApp replies not posting back to WhatsApp and only staying in the PF chat thread.
• Server: WhatsApp Relay-Back Timeout: Extended WhatsApp relay-back timeout from 120s to 10 minutes and fixed resolution on generation errors.
• Server: Private Sync Reauth: Fixed private sync reauthentication flow, server-side OAuth guard, and sign-out cleanup on macOS.
• Server: Local Folder Sync Auth: Fixed 401 auth error on local/network folder sync operations.
• Server: Google Session Re-Auth: Unlock dialog now triggers re-authentication on Google session expiry instead of showing a dead-end error.
• Safari: Extension Signing: Fixed Safari extension appearing as unsigned by using release entitlements without get-task-allow for Developer ID signed builds.
• Common: OAuth Account Mismatch: Fixed OAuth creating mismatched connections by properly creating new connections instead of overwriting existing ones with different accounts.
• Common: OAuth Scope Inflation: Fixed mismatched connections inheriting the original account's scopes, causing unintended scope over-granting.
• Common: Thread List Wipe: Fixed STATE_UPDATE timeout/retry using wrong field name, which could wipe the thread list.
• Common: Tier Badge Accuracy: Fixed tier badges incorrectly showing "Tester" on free-tier features.
• Common: Settings Panel Overflow: Fixed flex overflow in settings panel card rows causing layout issues.
• Server: Private Sync API Key Loss: Fixed private sync losing API keys when syncing between devices.
• Server: Messaging Bridge: Fixed archived chats, stop button, and mode selection in the messaging bridge.
• macOS: Custom Mode Settings Lost: Fixed custom mode settings being silently lost on macOS/iOS due to a no-op save path.
• macOS: OpenRouter Headers Blocked: Fixed relay CORS blocking OpenRouter attribution headers on macOS/iOS.
• macOS: Tab Context Switching: Fixed tab context not updating when switching between browser tabs.

New Features

• Common: Sub-Agents: Full sub-agent system with persistent named agents, conversation history, parallel execution (committee pattern), sequential and interjection modes, abort cascade, and dedicated UI rendering with chronological status tracking.
• Common: MCP Tool Approval: Registered MCP tools now appear in the tool approval system, giving users visibility and control over which MCP tools agents can invoke.
• macOS: Sidepanel Push Mode: Sidepanel now supports push mode alongside overlay, with 425px default width and docked width persistence across sessions.
• Server: WhatsApp Rich Messages: WhatsApp connection now supports location sharing, contacts, reactions, and sticker messages in addition to text and media.
• Server: API Key Encryption at Rest: API keys stored in relay D1 storage are now encrypted at rest. Server-side OAuth keys are deleted after being saved locally.

Improvements

• Common: Model Alias Display: Sub-agent tool call arguments now show annotated model aliases for easier identification.
• Common: BYOA Connection Priority: Expanded-scope OAuth requests now prefer BYOA alternate connections over Basic connections, reducing unnecessary re-auth prompts.
• Common: Thought Signature Preservation: Thought/reasoning signatures are now preserved correctly for both OpenAI Responses API and Gemini multi-turn tool calling flows.
• macOS: Debug Log Propagation: Debug logging toggle now propagates to the Node.js server subprocess.
• Common: Bengali Language Support: Platform and website now support Bengali, plus trademark disclaimer and media post updates on the website.

Bug Fixes

• Common: Safari/WKWebView Streaming: Polyfilled ReadableStream async iteration for kokoro-js, fixing TTS streaming failures in Safari and WKWebView.
• Common: Custom Mode Creation: Fixed stale React closure causing custom mode creation to fail on first attempt.
• macOS: Agent Storage Mutations Lost: Agent subprocess storage mutations (mode creation, settings, skills) were silently lost on macOS due to missing persistence bridge.
• macOS: Clipboard Copy: Fixed clipboard copy failing silently in macOS WKWebView.
• macOS: WASM/WebGPU in WKWebView: Forced WASM backend for TTS/STT and local ONNX paths in macOS native app — CDN cross-origin imports and WebGPU JSEP module imports fail in localhost WKWebView.
• macOS: Transport Request Collisions: Fixed requestId collision breaking tool approval on macOS, plus improved MCP display names.
• macOS: WhatsApp Auto-Reply: Fixed end-to-end wiring for WhatsApp auto-reply on macOS relay server, including thread visibility.
• iOS: Voice Dictation Stuck: Fixed voice dictation stuck in recording state when stopRecording cleanup was bypassed.
• Safari: Browser Commands Timeout: Fixed Safari browser commands timing out due to hanging WebExtension APIs after service worker suspension.
• Safari: Heartbeat Stale Check: Skip heartbeat stale check for Safari HTTP-polling browsers to prevent false disconnects.
• Server: Private Relay Stability: Fixed Durable Object hibernation, idle timeout (1006), reconnect delays, stale socket handling, and zombie readyState issues. Eliminated split-brain between serverState and ctx.state.
• Server: Webhook Broadcast: Webhooks now broadcast to all server sockets after DO hibernation wake, working around zombie readyState.
• Server: MV3 Messaging Race: Fixed lazy-init messaging bridge race condition on webhook arrival preventing MV3 service worker crashes.
• Server: Webhook Verify CORS: Routed webhook verify test through server to avoid CORS rejection in browser.
• Extension: Logout Cleanup: Sign out now correctly clears local relay server session on macOS, detecting localhost context.
• Common: Security — Thread Isolation: State broadcasts now filter threads by current profile, preventing cross-profile data leakage.

New Features

• Common: WhatsApp Connection Diagnostics: Test Connection for WhatsApp now checks webhook subscription status and app secret validity in addition to API token, catching silent delivery failures from Meta.
• macOS: Native OAuth Flow: macOS app now uses ASWebAuthenticationSession for OAuth instead of browser redirects, with automatic upgrade of stale connections lacking refresh tokens.

Improvements

• Common: DOCX Search/Replace Robustness: Search and replace in DOCX documents now handles double spaces, non-breaking spaces, smart quotes, numeric entities, and case mismatches. Includes case-insensitive fallback when exact match fails and warnings when multiple instances are replaced.
• Common: DOCX List Numbering: DOCX export now generates proper Word list numbering (w:numPr) instead of literal bullet characters, so Word recognizes numbered and bulleted lists correctly.
• Common: Agent Startup Performance: Eliminated 30-50s agent startup delays caused by ghost MCP servers, HTTP readiness timeouts, and redundant model fetches. Model cache is now passed to agent subprocesses.
• Server: Security Hardening: Command injection fix in URL handler, timing-safe token comparison, Content-Security-Policy header on relay server, rate limiting on session endpoint, CORS restricted to known origins, and parallel E2E broadcast encryption.
• Server: Webhook Signature Verification: Webhook payloads now use base64-encoded raw body to prevent JSON round-trip corruption that invalidated HMAC signatures.
• Server: OAuth Token Persistence: Fixed split-brain between server state objects that caused OAuth connections (Google Drive, etc.) to be lost after app restart.
• macOS: Performance & Stability: Menu-driven tray polling (was unconditional 3s timer), debounced window state saves, non-blocking logging, WKWebView crash recovery with exponential backoff, and memory leak fixes for script message handlers.
• macOS: Centralized Logging: Relay server output and app lifecycle events now log to ~/Library/Logs/caiioo/ with 10MB auto-rotation, replacing silently dropped output.

Bug Fixes

• Common: File Picker Instructions: pick_files tool now returns clear instructions for the user instead of a dead UI action signal that silently failed.
• Server: Active Tab Poll Spam: Active tab polling in relay mode no longer generates endless timeout errors when no browser extension is connected. Uses exponential backoff (30s to 5min) on consecutive failures.
• Server: CORS for Native App: Fixed 127.0.0.1 origin blocked by CORS whitelist, breaking Google OAuth from macOS native app's WKWebView.
• Server: Orphaned Agent Processes: Agent subprocesses are now properly cleaned up on server shutdown and uncaught exceptions.
• macOS: Browser Orchestrator Log Spam: Fixed "Unhandled message type" warnings for browser protocol messages in extension connections.

New Features

• Common: Auto-Save Settings: Settings fields now persist as you edit with a 500ms debounce for text fields and immediate save for toggles/selects, preventing data loss on crash or navigation.

Improvements

• Common: DOCX Numbering Preservation: Tab characters in numbered DOCX sections (e.g., legal documents with "3.2.3 → Heading") are now preserved through the edit round-trip instead of being silently dropped.
• Common: Smarter Tool Selection: The dynamic tool selector now handles factual queries (business hours, prices, addresses) that need web verification, instead of only matching strict trigger words.
• Common: Agent Language Changes: Language changes initiated by the agent now apply immediately instead of requiring a manual settings toggle.
• Common: Multi-Account OAuth Fallback: When the primary Google account lacks required scopes, connected secondary accounts are checked before prompting re-auth. Fixes uncaught errors when secondary account tokens expire.
• Server: Mode Settings Validation: Server now validates tier permissions when saving settings, preventing free-tier users from persisting pro-tier settings.
• Common: Security Hardening: DOMPurify sanitization on reasoning preview output, restricted CORS to known origins, HTTPS-first IP geolocation lookup, and tightened web_accessible_resources.

Bug Fixes

• Common: Google Search Result URLs: Browser Google search now returns structured links with actual URLs. Previously, the agent needed 5 tool calls to extract URLs because AgentMarkdown lost link hrefs from Google's complex nested HTML — now links are extracted directly from the DOM and included in the first search result.
• Common: Private Sync Infinite Loop: Fixed sync loop caused by manifest backup file writes continuously triggering re-sync. Empty settings objects now propagate correctly across devices.
• Common: Messaging Bridge Updates: Inbound messages, agent replies, and relay-back responses now broadcast to the UI immediately instead of appearing only after the agent finishes responding.
• Extension: Private Relay Reconnect: Fixed relay staying disconnected after MV3 service worker restart by falling back to stored enabled state. Also fixed self-replacement race during enable() and identity change reconnect.
• Extension: Private Relay Race Conditions: Hardened enable/disable races, added unhandled rejection guards on messaging bridge callbacks, and fixed timer leak on settings panel unmount.

New Features

• Common: Zero-Day Risk (ZDR) Enforcement: Live ZDR data sync with an enforcement toggle for OpenRouter — view provider count breakdowns and block models flagged with zero-day risks.
• Common: Tool Call Healing: Weak models that emit XML tool calls inside reasoning blocks are now automatically detected and healed, improving reliability across non-frontier models.
• Common: Unified Token Lifecycle: Symmetric OAuth token refresh across all platforms (extension, server, macOS, iOS) via a unified TokenLifecycleManager, with proactive refresh before expiry.
• iOS: Composer Icon Scaling: Composer action icons now scale 1.18x on iOS for better tap targets, with a dedicated native-ios body class for platform-specific styling.

Improvements

• Common: Slate Proposed Changes: Navigation between proposed changes, hover-based accept/reject, and counter updates now work correctly in the Slate editor.
• Common: Slate Selector Dropdown: The slate type selector now uses an inline dropdown instead of a portaled popover, fixing z-index and positioning issues.
• Common: Propose Change Full Replacement: propose_change now allows full document replacement for small documents (500 characters or fewer) instead of requiring partial edits.
• Common: Voice Send Waits for Transcription: Pressing Send/Enter while recording now waits for the transcription to complete before sending, preventing empty or partial messages.
• Common: Private Sync Setup Flow: Passphrase dialog now auto-shows after OAuth during private sync setup, streamlining the onboarding flow.
• Common: BYOA Registry Refresh: BYOA provider registry now re-initializes after private sync downloads new credentials, ensuring imported connections are immediately usable.
• Common: Terminology Consistency: "Cloud Sync" renamed to "Private Sync" in user guide content with a terminology guard test to prevent regression.
• Server: Private Relay Stability: Compatibility date updated, Durable Object crash handling improved, and MV3 reconnect logic hardened for reliable long-lived connections.
• Server: Private Sync Settings Excluded: Private relay settings are no longer included in cloud sync payloads, preventing cross-device relay config conflicts.

Bug Fixes

• Common: SVG Image Handling: SVG files are no longer sent as image_url to providers, which caused HTTP 400 errors on models that don't support SVG format.
• Common: Private Sync Race Conditions: Fixed messaging relay failures caused by race conditions during private sync initialization. Sync loop and quarantine logic hardened.
• Common: Private Sync Data Deletion: Cloud data deletion and audit now work correctly after disabling private sync.
• Common: Model Matching Contamination: Fixed cross-family model matching in the auto-adjust system that could incorrectly map models across provider families.
• Common: Dynamic Tool Selector: Hardened the dynamic tool selector prompt to prevent agent-like reasoning that could cause tool selection failures.
• Common: Validation Error Messages: Required field validation now returns focused error messages instead of generic failures.
• Extension: Private Relay WebSocket Drops: Fixed WebSocket connections dropping after ~60 seconds due to MV3 service worker suspension. Keepalive and reconnect logic improved.
• Extension: Google OAuth Incremental Scopes: Fixed OAuth scope expansion using launchWebAuthFlow in extension context instead of failing silently.
• Extension: Relay Toggle Settings Reload: Suppressed unnecessary settings reload when toggling the relay, preventing UI flicker and diagnostic noise.
• Server: Durable Object Hibernation: Fixed WebSocket close handling after Durable Object hibernation that could leave connections in a stale state.
• Server: Messaging Credential Persistence: Messaging credentials now persist immediately on change instead of waiting for the next save cycle.
• iOS: Stale WebSocket Callbacks: WebSocket handlers are now cleared on close, preventing stale callbacks from firing after reconnection on iOS.
• Desktop: NSIS Installer Upload: Windows NSIS installer now uploads correctly via API to draft releases.

New Features

• Common: Slack Messaging Adapter: New Slack Events API adapter for the messaging bridge with bot signature verification, media download support, and settings UI configuration.
• Common: Slate File Roundtrip: Open and save local files directly in Slate with filesystem handle persistence and hash-based dirty tracking. New "Load File" button in the slate selector dialog preprocesses documents and opens them with the correct editor.
• Common: Slate File Menu Restructure: File type menu reorganized into Documents, Renderable (HTML/Vega/Mermaid), and Code & Data categories with auto-detection on rename.
• Common: Remote Browser App: Remote browser clients can now fetch sidepanel assets via the private relay, enabling browser-based access without the extension installed.
• Common: Device Identity in Private Relay: Relay clients now show a human-readable device name. When displaced by another device, the UI shows which device replaced you.
• Common: Composer Mic Chip Redesign: Microphone button redesigned as a visible chip/pill with clear recording state indicator, 44px minimum touch targets for mobile, and voice overlay rendered over the input area.
• Common: User Guide Link: User guide now linked from the Documentation & Legal settings section.
• iOS: BYOA Setup Gate: BYOA wizard on mobile now shows a message directing users to set up Private Apps on desktop, since the 8-step browser flow is unworkable on mobile.
• Server: Login Challenge: Email-based 6-digit MFA verification for login, with consent tracking and session revocation.

Improvements

• Common: Private Sync Settings-First: Settings now sync before conversations, ensuring tokens and config are available immediately while bulk thread sync continues in the background.
• Common: WebSocket Request Correlation: WebSocket and private relay transports now properly await server responses instead of resolving immediately, fixing 40+ UI operations on macOS sidepanel and relay clients (Drive audit, Ollama test, MCP operations, etc.).
• Server: Webhook Signature Verification: Webhook signature verification moved to route-level app secret for cleaner architecture.

Bug Fixes

• Common: WhatsApp Response Delivery: Fixed agent responses being silently swallowed instead of sent back to WhatsApp. The messaging relay-back path now logs diagnostic details when delivery fails, making future issues immediately diagnosable.
• Common: Thinking Block Signatures Across Models: Fixed "Invalid signature in thinking block" errors when switching from a non-Claude model to Claude with thinking mode. Reasoning format detection now defaults to 'unknown' instead of misidentifying as Anthropic format.
• Common: Slate Editability for New Documents: New DOCX, PDF, RTF, and XLSX slates now correctly open in their editors instead of rendering as static HTML.
• Common: Slate Export Fallback: New documents without original binary data now export gracefully — XLSX falls back to CSV, RTF to HTML conversion, PDF to browser print.
• Common: Sandbox Preview Rendering: Fixed blank HTML and Vega previews caused by sandbox origin mismatch in postMessage targeting.
• Common: Track Changes Column Offsets: Deletions in tracked changes now correctly map HTML-to-text offsets, fixing mispositioned changes in TipTap.
• Common: Private Sync Auth Recovery: Expired or revoked OAuth tokens during sync now auto-trigger re-authorization instead of requiring manual reconnect.
• Common: Voice Streaming Draft Indicator: Voice transcript overlay now shows "Preview" status with reduced opacity to signal the text is interim until recording stops.
• Common: MCP Tools on Native Platforms: Local MCP server tools are now properly registered on macOS, Windows, and Linux — previously the handler started the MCP process but never fetched or registered its tools.
• Common: Slate Context Chip Update: Active tab context chip now refreshes from storage when a slate tab's title changes after loading from Drive or GitHub. Dark mode title input text is no longer unreadable.
• Extension: Chrome Stub Polyfill: Chrome browser tabs at localhost with window.chrome but no extension APIs now get proper stub injection instead of being skipped.
• Extension: Relay Client Bundle Crash: Fixed import.meta.env crash in relay-client IIFE bundle by defining build-time environment variables.
• Extension: Remote App Login: Remote browser app login cleaned up with Google OAuth as the primary method and email/password as a collapsed secondary option.

Improvements

• Common: Generating Indicator Coordination: Per-message typing dots now coordinate with the thread-level fallback indicator, eliminating duplicate bouncing dots during generation. Dots now persist during reasoning and tool execution phases, only hiding once the final answer starts streaming.
• Common: Streaming Cursor: A blinking cursor now appears at the end of streaming text, providing a clear visual indicator that the response is still being generated.
• Common: Reasoning Preview Formatting: Collapsed reasoning blocks now render inline bold and italic formatting instead of raw markdown syntax.
• Common: Slate Document Listing: list_slates now correctly shows content size for DOCX, PDF, and XLSX files that were loaded via lightweight thread queries. Listings also include workflow guidance for editing DOCX and XLSX documents.
• Common: Private Sync Settings Section: Private Sync settings are now in their own dedicated section for clearer organization, separated from general settings.
• Extension: Connections Settings Restructured: The Private Relay section is reorganized into two clear subsections — Private Relay (remote access toggle, always visible) and Caiioo Bridge (local MCP servers and system tools). The remote access toggle no longer requires the Bridge to be running.
• Desktop: Tauri Unified Binary: Windows/Linux desktop app consolidated from compile-time variants into a single binary with runtime tier gating — sidepanel visibility determined by subscription tier at runtime.

Bug Fixes

• Extension: Private Relay Without Bridge: Private relay now connects independently without the Caiioo Bridge app. Previously, authentication was only initialized by the Bridge connection, leaving extension-only users (including Linux/Windows) unable to receive messaging webhooks or use remote access.
• Extension: Google OAuth Routing: Fixed OAuth popup failures on Chrome by skipping getAuthToken when unavailable and falling back to BYOA credentials. COOP popup resilience prevents blank windows on restrictive sites.
• Common: ElevenLabs Streaming STT: Fixed voice transcription dropping or failing when ElevenLabs streaming encounters connection interruptions. Native app voice input now falls back gracefully.
• iOS: External Links in WKWebView: Links that should open in Safari now correctly open externally instead of loading inside the app's WebView.
• Common: Ad-Blocker Orphaned Rules: Fixed dynamic DNR ad-blocking rules persisting after service worker restarts. disableAllBlocking now directly queries and removes orphaned rules even when the blocker instance is null.
• Common: License Sync on Profile Switch: Switching profiles now syncs the license from the server, ensuring tier-gated tools refresh immediately instead of requiring a restart.
• Common: Token Refresh Before Provisioning: Expired OAuth tokens are now refreshed before provisioning API keys, preventing silent failures. Fresh servers can now adopt existing identity connections.
• Web: Mobile Menu on iOS Safari: The hamburger menu on the marketing website is no longer transparent on iOS Safari.
• Common: React-18 Batching Race: Thread-level generating indicator added as a fallback for cases where React-18 state batching prevented per-message typing dots from appearing.

Improvements

• BYOA Reconfigure Pre-Fill: Reconfiguring an existing Google or Microsoft Private App now pre-fills the client secret and tenant ID, so you don't have to re-enter them.
• Model List Sorting: Models within each provider group now sort by release date (most recent first), ensuring consistent ordering across platforms.

Bug Fixes

• Private Sync Auth Recovery: OAuth token failures (expired sessions, revoked tokens, Chrome profile tokens) during private sync initialization now surface correctly to the UI instead of silently entering an error state.
• Slate Tracked Changes Reliability: 10 trust-critical fixes for tracked changes — revision merge race condition, diff base persistence across all update paths, DOCX export formatting preservation, HTML tag stripping for TipTap matching, accept/reject-all ordering, and convergence fixes for entity escaping and tag regex matching.
• Slate Data Integrity: Fixed pasted images writing to a dead local cache, cloud source metadata not persisting, and BroadcastChannel fallback for live AI updates in relay/mobile mode.
• Slate Security: Fixed XSS injection in image viewer and replaced wildcard postMessage origins with scoped extension URLs.

New Features

• Provider Error Banners: When an AI provider returns an error (402 payment required, missing API key, rate limit), an actionable banner now appears with clear instructions instead of a generic failure message.

Improvements

• DOCX Cross-Span Editing: The propose_change tool now correctly handles search/replace operations that span across multiple formatting runs (e.g., partially bold text). Intent is passed through so the agent can make contextual edits.
• PDF Save Performance: Native PDF save is significantly faster — redundant parsing eliminated, import modules cached, and unnecessary operator cleaning skipped.
• Mode Name in Messaging: The messaging settings mode selector now displays the mode's branding name instead of the internal ID.

Bug Fixes

• Google OAuth Token Refresh: Fixed "Unauthorized" errors when refreshing Google OAuth tokens that could block Calendar, Gmail, and Drive access.
• Provisioned Key Error Message: Users with provisioned API keys no longer see a misleading "Add funds" error — the message now explains the actual issue and next steps.
• Concurrent Thread Race Condition: Fixed a race where switching threads during generation could cause GENERATION_COMPLETE/ERROR events to apply to the wrong thread.
• Revision History Restore Button: The restore button in Slate's revision history now correctly appears when the current content has diverged from the selected revision.
• Double OAuth Popup: New Chrome extension users no longer see two OAuth popups when signing in for the first time.

New Features

• Settings Category Grouping: 19 settings sections are now organized into 5 collapsible categories for progressive disclosure. Agent Modes section shows cross-cutting indicator with navigation links to affected sections (Tool Config, Model Config, Appearance). API Provider section opens by default with model capability warning.

Bug Fixes

• Agenda Sync Duplicate Calendars: Google Calendar accounts no longer appear twice in the Sync tab when both a login (identity) connection and a Private Connection exist for the same email. Connections are now deduplicated before rendering.
• i18n Settings Categories: Settings category labels are now translated across all supported languages.

New Features

• Microsoft 365 Integration Foundation: Microsoft scope routing, Graph API client, and account service with BYOA support. Scope hierarchy definitions and tool-scope registry extensions for Microsoft provider. Microsoft added as a relay-proxied provider.
• Rich Inline Rendering (RIR): New codec architecture for rendering rich content inline in chat messages. Includes HTML, Markdown, and DOCX codecs with an extensible codec interface.
• PDF Structure Tree & Layout Analysis: Tagged PDF structure tree parser for semantic document understanding. Heuristic paragraph grouping for untagged PDFs provides fallback layout analysis. Image replacement/insertion and content overflow detection.
• Messaging Mode Selection: Messaging bridge settings now include a default mode selector, letting the agent respond in the right personality when handling inbound messages.

Improvements

• Provider Rate Limit Retry: All LLM providers now automatically retry on 429 (rate limit) and 529 (overloaded) responses with exponential backoff, instead of immediately failing. Up to 3 retries with jitter.
• Google OAuth Scope Superset Matching: Write scopes now satisfy read scope requirements (e.g., drive covers drive.readonly). Prevents unnecessary re-authorization when a broader scope is already granted.
• BYOA Full-Scope Initial Auth: When a Private Connection is needed, the initial auth flow now requests all necessary scopes upfront instead of prompting twice (once for basic, once for expanded).
• Google 404-to-Expanded-Access Promotion: When a BYOA connection gets a 404 on a Google Workspace file, the error is promoted to an expanded access prompt instead of a dead-end error message.
• Messaging Credentials Private Sync: Messaging bridge credentials now sync across devices via E2E encrypted private sync instead of being device-specific. Sync manifest bumped to v10.
• OAuth Wait Extended Timeout: BYOA and expanded access OAuth flows now get a 5-minute timeout (up from 2 minutes), giving users enough time to complete Google Cloud Console steps.
• Sparkle Framework Signing: macOS distribution builds now properly sign Sparkle framework nested binaries in inside-out order, fixing notarization failures.
• Tauri Node Binary Bundling: Tauri configs now bundle the Node.js binary as a resource for Windows/Linux sidecar execution.
• Messaging Settings Search: Messaging bridge section is now discoverable via the settings search bar with keywords like "whatsapp", "telegram", "webhook".
• Settings Relay Forwarding: Settings saves are now forwarded to the relay server via WebSocket bridge, ensuring messaging credentials and other config changes reach the server's state file.

Bug Fixes

• Slate Currency vs Math: Dollar amounts like $9/mo in markdown tables are no longer misinterpreted as LaTeX inline math delimiters. The math regex now respects escaped currency dollars.
• Conversation Page ToastProvider: Pop-out conversation tabs were missing ToastProvider, causing toast-dependent features to silently fail. Provider tree now matches sidepanel.

New Features

• Google Picker Integration: Full Google Drive file picker with popup/iframe hybrid, multi-view filters (Docs, Sheets, Slides, PDFs), and drive reference chips in the composer. Files picked through the Google Picker are automatically granted drive.file access.
• Google Sheets Native Sync: Open Google Sheets in Slate with full cell-level round-trip editing. The Sheets codec converts spreadsheet grid data to TipTap HTML tables and back, with conflict detection via Drive modifiedTime. Supports reading, batch cell updates, and structural changes (add/delete sheets, merge cells).
• Google Slides Read-Only View: Google Slides presentations render in Slate with slide thumbnails and extracted text content for searchability.
• Gmail Batch Fetch & Markdown Conversion: Gmail tool now fetches message metadata in batches (up to 100 at a time) instead of one-by-one, with a 5-minute label name cache and automatic HTML-to-markdown body conversion for cleaner LLM consumption.
• Selection Overlay: "Add to prompt" button appears when selecting text on any webpage, letting you quickly add selected content to the conversation.
• Relay Overlay Manager: Agent overlay commands now route through the browser extension, enabling agent-driven UI overlays on the active webpage.
• Inline Quick-Tasks Model Picker: Quick tasks model selection is now embedded directly in the model picker dropdown instead of a separate menu.
• Slate DOCX Export from Google Drive: Google Docs loaded in Slate now export as DOCX (via readFileContentAsBinary), activating the full TipTap visual editor with tracked changes instead of opening as plain markdown.
• Private Connection Setup Wizard: Improved BYOA wizard with a dedicated "Add Test User" step matching the current Google Cloud Console flow. Consent screen substeps now follow Google's actual 4-step accordion. Email instructions are personalized when the user's identity is known.
• Private Connection Promotion Banner: When the agent needs expanded Google scopes (Gmail, Calendar, etc.) and no Private Connection exists, an animated banner appears with a one-click path to the setup wizard. The agent stays paused while the user completes setup and automatically resumes when the connection is saved.
• Export Fidelity Warnings: Before exporting a document to a different format (e.g., PDF to DOCX), a fidelity dialog warns about potential formatting losses with severity-categorized items (info, warning, critical).
• PDF Annotation Extraction: PDF text annotations and markup are now extracted and preserved during document processing and reconstruction.
• Reminders Calendar Sync: Reminders modal now supports calendar sync with expanded access gating.

Improvements

• Google Docs Sync Fidelity: Improved formatting span extraction with proper tag matching (handles self-closing tags, mismatched nesting) and conflict detection for concurrent edits.
• Private Sync Account Mismatch Detection: The sync settings account dropdown now detects when the configured sync email doesn't match any connected Google account and shows a "not connected" indicator with the option to switch.
• Teams/Enterprise Effective Tier: License validation now uses effectiveTier from team/enterprise delegation, ensuring correct feature access when a user's tier is inherited from a team plan.
• "Private App" renamed to "Private Connection": All user-facing references now use "Private Connection" instead of "Private App" — clearer for non-developers who don't associate OAuth integrations with "apps".
• Google OAuth Scope Alignment (drive.file): All elevated Google Drive scopes (drive, drive.readonly, documents.readonly, spreadsheets.readonly, presentations.readonly) are now replaced with the narrower drive.file scope for non-BYOA users, matching Google's verified app requirements. BYOA users retain full scopes.
• Gmail Permissions Visible in Scope Selector: Gmail read, compose, and modify scopes are now visible to all users in the Google Permissions editor under "Highly sensitive permissions" instead of being hidden behind BYOA-only. These scopes are approved on the consent screen and requested via JIT when Gmail tools need them.
• Google Picker Consent Recovery: When Google permissions are revoked externally (e.g., from Google Account settings), the app now auto-detects stale scopes, invalidates them, re-authorizes, and retries — instead of silently failing with 403 errors.
• Agent Google Drive Guidance: Tool error messages and empty-result hints now instruct the agent to tell the user to click the + button in the composer and select Google Drive, instead of suggesting unavailable tool actions.
• Slate Drive Picker Simplified: Slate's "Load from Google Drive" now loads files directly, skipping the sidepanel-style "Work with in Drive" vs "Add to conversation" choice screen.
• Official Google Drive Logo: Drive chips and the attachment menu now use the official Google Drive logo instead of the generic green triangle.
• Unified Model Picker: Helper model selection falls through to the default model, with a single consolidated model dropdown.
• DOCX Non-Text Element Passthrough: DOCX reconstruction preserves non-text elements (images, charts, embedded objects) that aren't part of the text editing flow.
• PDF Multi-Segment Line Editing: PDF WYSIWYG editor handles lines split across multiple text segments.
• Auth Rate Limit Increased: Auth endpoint rate limit raised from 10/min to 30/min to accommodate rapid OAuth token exchanges during BYOA setup.
• OAuth Token Endpoint Reclassified: OAuth token exchange moved from auth rate limit bucket to general, preventing throttling during multi-scope authorization flows.

Bug Fixes

• iOS Modal Stacking: Fixed crash when presenting file pickers or slate overlays while another modal (e.g., browser overlay) was already showing. Pickers and overlays now present on the topmost view controller.
• Slate Line Range Validation: get_slate_content now returns a clear error when startLine is beyond the document length instead of silently returning empty content.
• Server Path Traversal: Storage keys are now validated against directory escapes, preventing path traversal attacks on the sync storage endpoint.
• Ad Blocker Regex Cap: DNR converter now caps regex rules at Chrome's 1000-rule limit and drops large bounded quantifiers that exceed RE2's memory limit, preventing extension install failures.
• OAuth Timeout on Expanded Access: Fixed 120-second timeout when the agent needs expanded Google scopes (Gmail, Calendar). The pending OAuth wait now extends to 5 minutes for BYOA setup, and completing the connection in settings automatically resumes the paused agent.
• Chat UI Freezes: Fixed overlapping async intervals and silent broadcast errors that could freeze the chat interface.
• Cloud Sync Download Batch: Fixed TypeScript null-check errors in the cloud sync download batch handler.
• Service Worker Dynamic Imports: Converted dynamic imports to static imports for Chrome MV3 service worker compatibility.
• BYOA-Only Scope Enforcement: convert_to_google_doc and convert_to_google_sheet actions now correctly require only drive.file instead of broad editor write scopes.
• Drive Chip Format: Three-segment format for drive reference chips with proper BYOA-only scope gating.
• Google Picker CORS: Picker now hosted on the private relay to resolve cross-origin issues with the popup scope flow.

Improvements

• Ollama Context Window Indicator: Context window usage circle now renders for Ollama models by querying the /api/show endpoint for num_ctx, instead of relying on OpenRouter's model list which doesn't include local models.
• Ollama Tool Calling: Fixed dynamic tool selection for Ollama — passes tool_choice through to the API (supported since Ollama v0.5.1), disables think mode when tools are active to prevent corrupted JSON, and falls back to a lightweight non-streaming helper model when the primary model doesn't support tool calling.
• OpenRouter Reasoning Cache: Fixed reasoning text being cleared prematurely by the streaming provider. Cache lifecycle is now owned by agent-runner at iteration boundaries, consistent with Anthropic and Google providers.
• PDF WYSIWYG Reliability: Global text alignment handles PDF.js splitting operators across multiple TextItems. Block editing disabled on unmatched spans (annotation/header text) to prevent export errors. CMap 2-byte decoding no longer corrupts Word-generated PDFs with 1-byte ASCII fonts. Operator matching bounds check prevents out-of-range indices on empty text items.
• Desktop Platform Detection: Dynamic tool selector uses localFolderSync capability instead of caiiooReminders for desktop detection, and appleScript/appleNotes instead of Calendar/Reminders for macOS detection (Calendar/Reminders are cross-Apple via EventKit).

New Features

• File Manager: Full file management system with nested folders, user tags, starring, and a dedicated File Manager UI. Organize attachments into folder hierarchies, filter by type (screenshots, AI images, photos), search, sort, and view files in grid or list mode. Thread auto-tagging links files to conversations. Private sync support via manifest v9.
• Local Folder Sync: Bidirectional sync between local filesystem folders and the File Manager. Mutation-driven resync automatically propagates deletes, bulk deletes, and moves to disk with a 2-second debounce. Subfolder move detection relocates files on disk when reorganized in caiioo. Navigating into any subfolder within a sync tree triggers auto-resync.
• Cross-Filetype Tracked Changes: Unified accept/reject workflow across DOCX, PDF, and Markdown. DOCX edits are now recorded as tracked changes with the same diff engine used by Markdown and code slates. PDF tracked changes integrate with the visual editor. resolve_tracked_changes action lets the agent programmatically accept or reject revisions.
• PDF WYSIWYG Editing: Direct text editing on rendered PDF pages with document-matched styling. Text layer links edits to source content-stream operators via textItemIndex for surgical reconstruction on export. Includes plainText baseline for change detection.
• DOCX Rendering Fidelity: Rich DOCX preview with paragraph alignment, hyperlinks, font color/size, line spacing, indentation, and table cell shading. Code view shows markdown conversion for token-efficient LLM consumption.
• Universal Messaging (internal testing only): Send and receive messages through WhatsApp, Telegram, and Slack with local-first privacy. The agent can compose and send messages, list conversations, and reply to threads — all routed through your own device.
• Google Meet Integration: Multi-action meeting recall tool with list_meetings (discover recent meetings with date range filtering) and get_meeting (fetch transcripts). Threaded through the full OAuth/JIT/scope-approval pipeline with credentials vault support.
• Gmail Send Email: Direct email sending via Gmail API, gated behind high-risk tool approval. Per-action risk level overrides allow send_email to require explicit confirmation while draft actions stay at medium risk.
• Microsoft BYOA: Bring Your Own App support for Microsoft 365. Register Azure AD app registrations for Microsoft OAuth connections. Multi-provider BYOA registry (Google + Microsoft) with tenant ID support and a 3-step Azure Portal setup wizard.
• Poe Provider: New LLM provider integration for Poe, with model listing, provider-grouped model selector, vision/tools/reasoning capability detection, and pricing display.
• What's New Dialog: API-served content pipeline delivers release notes and user guide pages. Async what's-new dialog shows version highlights on update.
• Private Sync Account Selector: Choose which Google account to use for private sync, instead of defaulting to the primary account.
• Native TipTap Diff Marks: Replaced the separate marked.js rendering overlay for AI revision previews with native ProseMirror marks (DiffAdded/DiffRemoved). Revisions now render as inline tracked changes with consistent typography, plus chunk-level accept/reject targeting.
• Mermaid Diagrams: Live Mermaid diagram rendering in code blocks within Slate's TipTap preview mode.
• Fuzzy Section Search in Help Tool: Help tool now supports fuzzy matching when searching for specific sections within documentation pages.

Improvements

• File Manager UX Polish: Folder sync icon indicators, delete confirmation dialogs, shift-click range selection, toolbar wraps instead of overflowing at narrow widths, and fixed tag persistence across reloads.
• Responsive Composer Layout: Three-zone flex layout prevents the new-thread button from being pushed off-screen at narrow widths. Mode and model selectors use staggered text-hiding breakpoints. Custom agent and AI chip icons replace generic defaults.
• Settings UX: Reordered sections for better flow (Personalization → Credentials → Tool Approvals → Agent Modes → Tool Configuration → Skills Library → Tool Servers → Document Processing → Voice).
• Revision State Integrity: Fixed corruption when deleting large content blocks during active AI revisions — sourceContent immutability, chunk relocation threshold raised to 0.7, stale chunks auto-rejected, and version history integrity checks.
• Agent Loop Reliability: Fixed abort controller race condition where old runs could delete new run's controller. Cancel signal now propagates into queued tools. Orphaned running states broadcast errors instead of leaving blank messages.
• Interjection Handling: Unified getActiveBranchMessages traversal on server and UI to skip interjections consistently. Fixed response disappearing after follow-up when streaming parentId pointed at the interjection instead of the user message.
• macOS Native Stability: Fixed restart race condition with process epoch tracking, URLSession leak on reconnect, and LineBuffer data race between pipe handler and stop.
• Google Sheets Fix: create_table now writes column header names instead of leaving the first row blank.
• Middleware Cleanup: Extracted requireAuth middleware from 10 inline auth checks across apple-routes, mcp-routes, and attachment-routes.
• Image Type Filtering: File Manager distinguishes screenshots, AI-generated images, and user photos with separate filter options and distinct icons.
• Mode/Model Picker Parity: Equalized font weight and icon stroke between mode picker and model picker.
• Beta Tier Gating: Credentials vault gates beta features behind tier checks.
• CIDFont Re-Encoding: PDF reconstruction supports CIDFont encoding with ToUnicode CMap tables and identity fallback for characters not in the map.
• PDF Export Renamed: "Download" renamed to "Export" across the PDF workflow for clarity.

Bug Fixes

• PDF Text Disappearing: Fixed PDF view text vanishing on blur and garbled content when exporting from WYSIWYG editing.
• PDF Tracked Changes Routing: Prevented PDF tracked changes from routing through the unified diff engine, which corrupted PDF-specific operator data.
• PDF Export Integrity: PDF download now fails loud with diagnostics instead of silently falling back to markdown export.
• Markdown Table Line Breaks: <br> tags in markdown tables now render as actual line breaks instead of literal text.
• DOCX Tracked Deletions Preserved: Fixed DOCX tracked deletions being stripped when toggling between visual and code view.
• Tracked Changes on Stored Markdown: Fixed propose_change on markdown slates showing no redlines when loaded from storage.
• Reject Reverts Content: resolve_tracked_changes with reject now properly reverts content for markdown and code slates.
• Slate Save Loop: Fixed infinite save loop triggered when track changes was active.
• Messaging Stability: Fixed routing, reply threading, profile handling, queue resilience, webhook signatures, sender sanitization, and payload limits.

New Features

• Bring Your Own App (BYOA): Use your own Google OAuth credentials instead of Caiioo's built-in app. Multi-app support lets teams share a single OAuth client across the organization while keeping individual API keys. BYOA connections auto-detect stale tokens and prompt reauth.
• Wikilink Navigation: Obsidian-style double-bracket wikilinks between Slate documents for building interconnected knowledge bases. Click a wikilink to navigate between Slates instantly.
• Skills Menu: Tab-to-use skill insertion, dynamic height dropdown, inline editing with variable chips, and a shared SkillForm component for full CRUD from the composer.
• DOCX Download: Download Slate markdown and text documents as DOCX files with tracked changes preserved.
• Table Auto-Populate: insert_component now auto-fills table cells from structured data, eliminating empty placeholder tables.
• Gemini 3.1 Flash: Added as an image generator model option alongside existing Gemini, FLUX, and Seedream models.
• Google Drive Save for All Accounts: Save to Google Drive is now available for any connected Google account, not just the primary one.
• Unified Diff Engine: Complete rewrite of Slate's change-tracking diff system (phases 0–7) with TOCTOU safety, caching, per-change accept/reject widgets, and mobile tap-to-toggle support.

Improvements

• Adaptive Private Sync: Sync polling rate adjusts dynamically based on activity. Download-only lock optimization reduces contention.
• OAuth Email Capture: Google OAuth always includes identity scopes so the user's email is captured on every connection, fixing blank emails on some accounts.
• Security Hardening: Hardened OAuth PKCE flows, agent subprocess boundaries, and per-account brute-force lockout for repeated failed authentication attempts.
• Browser Chip Layout: Browser selector chips are smaller and wrap instead of overflowing the container.
• Registry-Driven Image Providers: Replaced hardcoded image provider list with a dynamic registry, making it trivial to add new models.
• Content i18n Pipeline: Unified hash-tracked content sync for all documentation pages with incremental translation support.

Bug Fixes

• Math Expressions in Preview: Fixed math expressions being corrupted when switching from TipTap preview to code view.
• Wikilinks in Preview: Fixed wikilink syntax being stripped when toggling between TipTap preview and code view.
• Thread Messages Disappearing: Fixed thread messages being stripped by STATE_PATCH or vanishing when switching models mid-conversation.
• Interjection Orphans: Fixed orphaned parentIds after interjection message filtering causing render errors.
• Private Sync Infinite Loop: Resolved infinite sync loop caused by concurrent collection item clock conflicts and thread_data_ prefix pollution.
• BYOA Edge Cases: Fixed BYOA config not restoring on clear, async save races, extension using dead localhost popup instead of launchWebAuthFlow, and stale issuedByClientId precedence.
• Private Sync Scope Recovery: Fixed private sync unlock failing silently when Google Drive scopes were expired or missing.
• Duplicate Skills: Prevented duplicate skills from appearing after private sync merges.
• Tracked Changes Routing: Wired get_tracked_changes and get_comments into SlateTool action routing so the agent can read revision state.
• DOCX Download Source: Fixed downloadAsDocx using stale artifact contents instead of live editor state.
• Smart Model Name Abbreviation: Fixed long model names overflowing the compact picker with intelligent abbreviation.
• Markdown Link Rendering: Fixed chrome-extension: and Caiioo: URLs being blocked in rendered markdown links.
• Thread Creation Speed: Halved thread creation time by eliminating redundant settings cascade lookups.
• Chrome Tab Title: Restored dynamic Chrome tab title based on the active conversation thread.
• Adaptive Greeting Tokens: Bumped max_tokens for adaptive greetings to prevent truncation with reasoning models.

New Features

• Agent Interjection: Guide the AI mid-run by typing a message and pressing Enter while it's working. The agent sees your guidance at its next decision point and adjusts course — no need to cancel and restart. An amber-badged send button appears alongside the stop button when you have text to send.
• API Tool Management: The API integration tool now supports save_tool, list_tools, and delete_tool as first-class actions, making it easy to manage agent-discovered APIs directly from conversations. Free users see GET and list_tools; Pro users get the full set.
• Live Token Counter: Token and cost usage now updates in real time during streaming, displayed in each message footer. Specialized costs (image generation, web search) fold in incrementally as each tool completes.
• Auto-Expand Tool Cards: Tool cards automatically expand when they start receiving streaming arguments or need approval, and auto-collapse on successful completion. Errors and denied tools stay expanded for review.
• Monaco Diff Editor: Slate code revisions now use Monaco's native inline diff editor with per-change accept/reject buttons, replacing the previous custom decoration system.

Improvements

• Private Sync Auth Recovery: Private sync errors from expired or revoked Google tokens now show actionable "Reconnect Google" and "Grant Permissions" buttons instead of a passive error message. Reconnection requests the correct Drive permissions and pre-selects the right Google account.
• Private Sync Multi-Account: Private sync now authenticates against the Google account matching the config email instead of falling back to whichever account was connected first. Fixes syncing to the wrong Drive when multiple Google accounts are connected.
• Brand Refresh: New river blue and rock grey color palette derived from the Caiioo logo, replacing the previous purple theme across the entire app and website.
• OAuth Security: Google, GitHub, and Slack OAuth client secrets removed from the extension package. Token exchange now routes through a secure Cloudflare relay proxy, eliminating secrets from client code.
• Faster Startup: Attachment content (images, extracted text) is now loaded on demand instead of at boot, significantly reducing initial load time for conversations with many attachments.
• Faster Streaming: Context window usage now streams via the fast broadcast path instead of round-tripping through storage, giving real-time updates without async I/O per agent loop.
• Private Sync Speed: Google Drive file ID cache is persisted across restarts, eliminating a full file listing API call on first sync cycle.
• Unified Streaming Render Path: Streaming and final content now flow through a single render path, eliminating the flash when generation completes and preserving interjection messages.
• Cross-Model Switching: Switching between AI providers mid-thread (e.g., Gemini to Claude) no longer causes "Invalid signature in thinking block" errors. Foreign reasoning artifacts are automatically flattened to narrative text, preserving context without incompatible cryptographic signatures.
• Disabled Tool Guidance: When the AI tries to use a disabled tool, it now receives actionable suggestions (similar enabled tools, how to enable) instead of a generic "not found" error.
• Responsive Composer Layout: Composer toolbar uses staggered progressive disclosure — mode selector, model selector, settings, agenda, and help icons appear as space allows instead of being clipped.
• Settings Color Coding: Settings sections now use a 3-color system — accent tint for customization sections, alternating neutrals for the rest — for clearer visual grouping.
• Onboarding Errors: Free key provisioning now shows descriptive error messages below action buttons instead of silently failing.
• AI Provider Settings: Renamed "API & Provider" section to "AI Provider" for clarity.
• Extended Free Trial: Pro trial period extended from 7 days to 14 days, giving new users more time to explore Pro features.
• Optimistic Branch Switching: Switching between message branches and reloading messages now updates the UI instantly instead of waiting for storage round-trips.
• MCP Tool Configuration: Schema auto-correct and improved MCP tool configuration for more reliable tool setup.

Bug Fixes

• Per-Message Costs: Individual message costs now cumulatively match the thread header total. Helper, OCR, image gen, and search costs are attributed to individual messages instead of only the thread summary.
• Interjection Rendering: Interjection messages now render as compact amber annotations inline in the assistant timeline instead of being lost when generation completes.
• Loop Detection: Fixed false-positive repetition detection on citation-heavy responses where URLs naturally repeat 3+ times.
• Slate Diff Alignment: Fixed redline/greenline misalignment in preview mode for multi-line chunks where the overlay loop advanced by only 1 line regardless of chunk span.
• Slate Deletion Positions: Fixed double-counting of insertion shift causing deletions to render after their paired insertions instead of before.
• Private Sync Fixes: Fixed JIT scope request not firing across code-split chunks and infinite sync loop after uploads.
• Onboarding Flow: OAuth flow now completes onboarding properly, hyphenated mode IDs are supported in trigger detection, and 0 days remaining no longer coerces to null.
• Password Reset: Resetting password via email link now verifies the email address. Added resend-verification endpoint for users with expired verification emails.
• License Tier: Users with admin-granted tier but no license row are no longer stuck on the upgrade gate.
• Settings Scroll: Clicking the brain icon for a learned page now scrolls to the correct settings section instead of stopping short due to lazy-loaded sections above.
• Monaco Disposal: Fixed diff editor model disposal order to prevent dangling references and Monaco showing through TipTap preview.
• Walkthrough Targeting: Walkthrough steps now skip elements hidden by responsive container queries instead of spotlighting a 16x16 pixel at the top-left corner.
• Kokoro TTS: Fixed text-to-speech broken in the extension — ONNX runtime detection failed in offscreen documents, causing "no available backend found" errors.
• Custom API Tool Save: Fixed saving agent-discovered APIs as custom tools crashing in the extension due to forbidden dynamic import() in Service Worker context.
• Boot Performance: Fixed session migration running on every settings access (20+ times at boot), causing unnecessary disk writes and slower startup.
• Duplicate Storage Broadcasts: Suppressed double storage change notifications in relay/server mode that could cause UI flicker.
• Cairn Texture Recovery: Fixed texture generation failing when the in-memory world cache was evicted between load and render, and fixed renderer showing black for attachment-backed textures.
• Security Patches: Updated fast-xml-parser (DoS via DOCTYPE entity expansion) and tar (hardlink escape vulnerability).
• MV3 Compliance: Replaced CDN-loaded vega-embed with local vendor bundles, ensuring all chart rendering code passes Chrome Web Store remote code policy.
• Vega-Lite Charts: AI now generates Vega-Lite visualizations directly in Slate instead of CDN-loaded HTML, enabling interactive charts without remote script dependencies.
• Google Drive Save: Fixed 403 errors when saving to Google Drive from Slate caused by stale OAuth scope checks in non-agent Drive handlers.
• New User Onboarding: Fixed onboarding screen being skipped entirely for new users when an API key was auto-provisioned during login.
• OpenRouter OAuth Popup: Fixed "Connecting..." spinner stuck indefinitely when popup blockers silently killed the OAuth window.
• Trial Banner Copy: Trial banner no longer claims models are gated by tier — updated to accurately describe Pro features (write access, image generation, custom modes).
• DOCX Review Toolbar: Fixed toolbar flickering when showing document review controls by deferring scroll until initial visibility.
• Add to Prompt Overlay: Fixed overlay logo blocked by Content Security Policy and background opacity not matching theme.
• Message Edit/Reload State: Fixed in-memory state going stale after editing a message, reloading a response, or switching branches.
• Server Login Hang: Fixed WebSocket storage operations failing during the server login flow. Auth is now decoupled so the local session works immediately while server identity resolves in the background.
• WebSocket Connection Hang: Fixed a race condition where AUTH_STATE: ready arriving during initial WebSocket connection could orphan the auth promise, preventing the connection from ever completing.

New Features

• Slack Integration: Connect your Slack workspace to enable AI-powered Slack tools with full OAuth V2 authorization
• API Tool Credentials: Agent-discovered APIs can now be saved as custom MCP tools with stored credentials, enabling reuse across conversations
• Saved API Badge: Agent-created API tools display a "Saved API" badge in settings for easy identification
• Conversation Import: Import conversations from ChatGPT, Claude, Gemini, Perplexity, and Grok — upload your export file and Caiioo converts it with full message history, attachments, and metadata

Improvements

• Learned APIs Nesting: Learned API tools are now grouped under a collapsible section in settings for cleaner organization
• Caiioo Branding: Built-in tools now display the Caiioo icon for visual consistency
• Browser-Aware Links: URLs opened by the agent now launch in the browser you've selected in settings
• Slate Cloud Load: Loading a Slate document from Google Drive now triggers a JIT OAuth prompt instead of failing with a cryptic error

Bug Fixes

• DOCX Editing: Fixed paragraph run regex matching across nested spans, causing content corruption when editing formatted DOCX text
• Search/Replace Safety: Restored structural HTML pattern guard lost during refactoring — prevents edits from corrupting XLSX/DOCX internal markup
• Tool Approval Loop: Fixed race condition causing repeated approval prompts for the same tool within a single agent run
• Stale Identity: Fixed stale user identity not clearing on 401 during license sync, with logout escape on returning-user onboarding

New Features

• Minimum-Privilege Permissions: Google Workspace tools now request only the exact scopes needed for each action instead of broad access. Reading emails requests read-only; creating events requests only calendar write access.
• Enhanced Permission Dialog: The just-in-time authorization dialog shows exactly what permissions are being requested, why they're needed, and whether the operation is read-only or read-write
• Permission Selector: When manually connecting a Google account, you can choose which services (Calendar, Gmail, Drive, Docs, Sheets, Slides) and access levels (read-only vs read-write) to grant upfront. Edit Permissions button on existing connections.
• Image Results: Generated images and screenshots now render prominently between collapsed process steps and the final answer, eliminating the need to expand tool calls to see visual output

Improvements

• Private Sync Safety: Concurrent edits now download for merge instead of uploading, lock refresh verifies ownership, and manifest merge prevents data loss from premature trash operations
• P2P Sync Removed: Eliminated P2P real-time sync, removing 100+MB of wasteful WebSocket traffic and improving extension performance
• Settings Performance: Settings panel memoization and collapse context isolation for faster rendering
• DOCX Rendering: Improved paragraph rendering, numbering, and style support in the document parser

Bug Fixes

• Google Docs Scope: Fixed "insufficient authentication scopes" error when indexing Google Docs — index_document needed write scope for named ranges
• Google Docs Read: Fixed read_document and get_document_info failing when only Docs scopes were granted — these actions also need Drive read access
• Calendar Copy: Fixed copy_event action missing from the scope registry, preventing JIT permission requests
• API Key Persistence: Fixed BYOK API key being silently lost when saveCustomMode crashed the settings save path during Pro trial activation
• Settings Reload Race: Fixed unsaved API key being wiped when collapsing/expanding settings sections triggered a storage reload cycle
• Chrome Identity Cancellation: Fixed user cancellation during incremental scope requests not being distinguished from errors, and auth method now correctly transitions after obtaining a web OAuth refresh token
• Private Sync Overwrite: Fixed private sync overwriting recent turns when concurrent edits occurred
• Viewport Screenshots: Stopped surfacing viewport screenshots above final response text
• Agent Coordination: Fixed 7 issues across turn coordination, perception, and state management
• Track Changes: Fixed track changes mode broken after slate decomposition due to dual-state variable desync
• File Manager: Delete and download actions now available in all modes, not just orphaned files
• MCP Tool Names: Fixed tools using internal server IDs instead of human-readable names
• Agenda Tool: Fixed sync card always showing missing permissions, wired into Google OAuth connection system

Improvements

• New Document Creation: New DOCX, XLSX, PDF, and RTF files created from the "New File" dropdown are now fully editable with tracked changes support and proper
• Gemini Stability: Added reasoning repetition detection to automatically break Gemini thought loops and prevent leaked thinking text from appearing in chat
• Image Handling: DNG raw photo previews now respect EXIF orientation for correct display preventing distortion on uploaded images.

Bug Fixes

• New Document Editability: Fixed blank documents created via "New File" being rendered as static previews instead of editable rich text editors
• API Key Detection: Switching to your own API key (BYOK) now takes effect immediately without requiring a page refresh
• Screenshot Stability: Fixed attachment ID handling to prevent screenshot data from being lost or corrupted
• Tab Grouping: Fixed crash when browser had non-normal windows (e.g., devtools, popups)
• Console Noise: Removed unnecessary llms.txt probing that was spamming 404 errors
• Tool Timeouts: Removed fragile keepalive workarounds that could interfere with long-running tool calls like 4k image generation with Gemini
• Prompt Caching: Fixed multimodal content (images/screenshots) being dropped during prompt cache optimization

Internal

• Significant codebase quality improvements across error handling, type safety, and architectural layering
• Modularized core document parser into focused, maintainable modules

Improvements

• Model Selector Tier Medals: Replaced hardcoded recommendation badges with 5 API-driven tier medals (Best for Caiioo, Quality, Reliability, Speed, Value) showing gold/silver/bronze rankings directly from benchmark data
• Skills System: Skills are now have a restore defaults button to re-add built-in skills
• Prompt Caching: Further optimized prompt caching for cost savings

Bug Fixes

• Browser Compatibility: OAuth login and auto-connection now work in Vivaldi and other non-Chrome browsers
• Cost Tracking: Fixed race condition where cancelling a run could clobber cost and usage data; cost/usage data is now preserved on cancelled and errored runs
• Private Sync: Rewrote purgeAllStorage to nuclear clear, fixed sync listener race, consolidated private sync UI
• i18n: Added missing translation initialization to conversation.html popup page
• Security: Patched dependency vulnerabilities (qs, @casl/ability, axios, markdown-it)
• Benchmarks Page: Migrated benchmarks page backend to Cloudflare auth

Improvements

• MiniMax M2.5 Benchmarks: Full benchmark suite completed — tool accuracy 91%, trustworthiness 95% (rank #2), composite rank #5 with gold value tier
• Model Intelligence API: Increased default response limit from 100 to 500, ensuring all models with earned badges are visible to clients

Bug Fixes

• Missing Model Badges: Fixed 35 models with earned tier badges (gold/silver/bronze) not appearing in the extension model list — including Claude Opus 4.6 (gold composite), Claude Sonnet 4.5 (silver composite), and Claude Haiku 4.5 (bronze composite). Root cause: API defaulted to returning only 100 models sorted by trustworthiness, but tiers were computed from all 367 models. Models outside the top 100 by trustworthiness had their badges silently dropped.

Improvements

• Ollama Provider Parity: Full feature parity with OpenRouter — abort signal support, error handling, reasoning details, resolved model ID, and think-tag processing via streaming mixin
• AbortSignal Propagation: Subprocess SIGTERM/SIGINT now cancels in-flight LLM API calls
• Build-Time Schema Validation: Settings schema validation at build time catches missing SETTING_METADATA entries
• Platform Capabilities Caching: Cached for performance instead of recomputed on every access
• Website i18n: Trust page and pricing refactor translations synced across all 22 locales

Bug Fixes

• Website Authentication: Replaced legacy Supabase auth with direct Cloudflare Worker API calls for Google sign-in, email login, signup, and password reset
• Stripe Checkout Locale: Fixed "Invalid locale" error on checkout and portal by mapping navigator.language (e.g. en-US) to Stripe-supported locales with fallback to auto
• Password Minimum Length: Synced 12-character minimum across server signup, password reset, and website reset page
• Private Sync Profile Dedup: Login now detects and removes duplicate profiles caused by earlier sync bugs
• State Manager Init Race: Fixed initialization race condition in state-manager
• Agent Subprocess Cleanup: Zombie subprocess cleanup on agent termination
• Tab Group/Storage API Guards: Proper guards for tab group and storage APIs across platforms
• Geolocation Error Handling: Graceful handling of geolocation permission errors
• OAuth Refresh Locking: Prevents concurrent OAuth token refresh attempts
• Settings Save Mutex: Concurrent settings saves no longer clobber each other
• Agent JSON Parse Isolation: Malformed agent output no longer crashes the parser
• Thread Search Race: Fixed race condition in thread search results
• Streaming Version Staleness: Checks for stale version during streaming responses
• Content Script Timeouts: Added timeouts for content script message passing
• Context Pruning for Multimodal: Improved context pruning when multimodal content is present

Bug Fixes

• Private Sync Mode Variables: Fixed legacy settings migration running on every sync cycle, silently overwriting recent local edits (e.g., mode variable changes) with stale Drive data

New Features

OpenRouter OAuth One-Click Setup

• PKCE Key Creation: New one-click OpenRouter OAuth flow lets users create and link an API key without leaving Caiioo
• Redesigned Onboarding: Streamlined onboarding and upgrade flows with OpenRouter OAuth integration
• Privacy Warning: Free models onboarding option now displays a clear privacy/training data warning

Granular Private Sync

• Per-Item Sync: MCP servers, tool approvals, profiles, skills, modes, overrides, and reminders now sync at the individual item level instead of overwriting entire collections
• Per-Key Settings Sync: Settings sync granularly per key, preventing remote overwrites of unrelated local changes
• Deduplicated Reads: In-flight WebSocket storage reads are deduplicated to reduce unnecessary network traffic

Improvements

• Mode Settings UX: Overrides now auto-save, and Restore Defaults correctly restores deleted mode variables
• Responsive Composer: Skills button collapses to icon-only at narrow widths; removed layout spacer from Skills section
• Support Tickets: Build version and datetime are now included automatically
• Slate Search: New search_document action added to the Slate tool

Bug Fixes

• Google Slides/Sheets OAuth: Now throws proper OAuthAuthorizationRequiredError instead of returning setup text, enabling just-in-time authorization
• Configuration Tool UI Refresh: Agent mutations via the configuration tool now broadcast STORAGE_CHANGED so the UI updates immediately
• Tool Approval Modal: Fixed React hooks ordering violation (useMemo above early return)
• Localhost HTTPS: API integration tool now handles self-signed certificates for local HTTPS servers
• Mode Variables Persistence: Fixed i18n getter properties not materializing before storage, causing variables to vanish on reload
• Slate Thread Safety: Resolved race condition in Slate tools that caused duplicate slates during parallel AI turns
• Slate DOCX: Fixed deletion visibility and baseline corruption in tracked changes
• Private Sync Stability: Eliminated bouncing and vanishing edits from sync conflicts

Improvements

Slate Editor i18n

• Translated Toolbar & Menus: All Slate editor toolbar buttons, context menus, and dialog strings are now fully translated

Bug Fixes

• OAuth Fetch Timeouts: All OAuth token exchange and refresh requests now have a 15-second timeout, preventing infinite hangs on network stalls
• GitHub Private Email: Fixed GitHub connection failing when the user's profile email is private (now fetched from /user/emails API)
• OAuth Connection Dialog: Generalized OAuth connection dialog and fixed multiple token/connection bugs
• Password Length Consistency: Synchronized 12-character minimum password requirement across all signup and reset surfaces

New Features

Internationalization (20+ Languages)

• Full i18n Support: Caiioo is now available in 20+ languages including English, Spanish, French, German, Japanese, Korean, Chinese, Arabic, Hebrew, Hindi, and more
• RTL Language Support: Full right-to-left layout for Arabic, Hebrew, and Urdu
• UI Language Setting: Choose your preferred language in Settings — all UI elements, tool labels, and status messages are translated

Improvements

Settings Panel Performance

• Lazy-Loaded Sections: Settings panel sections now load on-demand, reducing initial render time
• Modular Storage: Settings are stored in granular per-section keys instead of a single monolithic blob, improving read/write performance
• Typed Getters: Internal settings access uses strongly-typed getters with change granularity tracking

Slate Track Changes Reliability

• 17 Revision Manager Fixes: Comprehensive hardening of the track changes system across diff computation, acceptance, rejection, and persistence
• Plain Text Diffing: Redline changes now diff plain text instead of raw markdown, producing cleaner and more accurate change highlights
• Persistent User Changes: User-made tracked changes now survive page refresh
• Race Condition Fix: Force-bake tracked changes before save to prevent data loss

Bug Fixes

• Service Worker Crashes: Replaced 112+ dynamic import() calls with static imports to prevent Chrome service worker crashes
• Mode Welcome Messages: Fixed language, provisioned key detection, and persistence issues in mode welcome messages
• Google Tool Account Selection: Account picker now dynamically reflects actually connected accounts
• DOCX Nested Lists: Fixed display markers and export corruption for nested list round-trips
• LaTeX Math Rendering: Fixed currency dollar escaping breaking LaTeX math expressions starting with numbers
• Tab Group Creation: Deferred lazy tab group creation until the web browsing tool is actually used
• Agenda OAuth: Re-throw OAuth errors so just-in-time authorization triggers correctly
• Settings Persistence: Added missing metadata entries for 6 settings that silently failed to save

Security Hardening

• Content Script Origin Validation: Messages from web pages to the extension are now restricted to caiioo.ai origins only, with strict same-origin checks preventing cross-origin message injection
• CSP Tightened: Removed development-only localhost script sources from the extension pages Content Security Policy
• Auth Response Scoping: Extension auth responses are now sent to the specific page origin instead of broadcasting to all frames

New Features

Multilingual Speech-to-Text

• Language Selection: Choose your STT language in Settings for accurate non-English voice input
• 90+ Languages Supported: Works with Whisper and ElevenLabs for comprehensive language coverage

Real-Time Voice Activity Detection

• Low-Latency VAD: New Silero VAD v5 integration detects speech in real-time with minimal delay
• Smarter Recording: Recording automatically starts and stops based on voice activity
• Reduced False Positives: Better distinction between speech and background noise

Suggestions Visibility Toggle

• Hide/Show Suggestions: New toggle to hide AI follow-up suggestions when you want a cleaner interface
• Persistent State: Your preference is saved and remembered across sessions

Improvements

Track Changes Auto-Bake

• Diff-Based Tracking: More efficient change tracking using diff algorithms
• Auto-Save: Changes are periodically saved to prevent data loss during long editing sessions
• Snapshot on Exit: Exiting track changes mode automatically bakes all pending changes

Enhanced Provider Streaming

• Gemini Improvements: Better streaming and reasoning capabilities for Google Gemini models
• Consistent Behavior: Unified streaming behavior across OpenRouter and native providers

New Features

Guided Onboarding Walkthrough

• Interactive UI Tour: New step-by-step walkthrough guides new users through every part of the interface after entering their API key
• Three Phases: Covers the composer (11 steps), settings panel (7 steps), and mode system (5 steps)
• Spotlight Effect: Each step highlights the relevant UI element with a focused spotlight
• Conversational Onboarding: AI-guided profile building to personalize your experience from the start

Prompt History

• Access Previous Prompts: Your recent prompts are saved and can be accessed in the composer
• Quick Reuse: Easily reuse or modify previous messages

Ad Blocker Levels

• Granular Control: Ad blocking now supports multiple levels instead of just on/off
• Choose Your Protection: Select the level of blocking that works for your browsing needs

DOCX Review Toolbar

• Track Changes Controls: New toolbar when viewing Word documents with tracked changes
• Accept/Reject Actions: Easily review and resolve document edits

Improvements

Thread Retention Settings

• Configurable Cleanup: Set how long to keep old threads before automatic cleanup
• Storage Management: Better control over your conversation history

Caiioo Animation

• Performance Optimizations: Smoother animation with improved rendering
• Wetness Effects: New visual overlay showing water saturation
• Direct Particle Rendering: Cleaner visuals with optimized particle drawing

Profile Switching

• Enhanced State Management: More reliable profile switching with improved protocol handling
• Better Sync: Profile changes sync correctly across the extension

Google Calendar

• Token Management: Improved access token handling for calendar operations
• More Reliable Sync: Better OAuth flow for calendar integration

New Features

Ad & Tracker Blocking

• Built-in Ad Blocker: Block ads and trackers using the Ghostery engine with MV3-compatible declarativeNetRequest
• Toggle in Settings: Enable or disable ad blocking from the Settings panel
• Filter List Updates: Automatic caching of filter lists for reliable blocking

Just-in-Time Google Permissions

• Incremental Authorization: Google tool permissions are now requested only when needed, not upfront
• Clearer Scope Management: Missing scopes trigger helpful error messages with options to grant access
• Better Privacy: Only request the specific Google scopes required for each tool

Improvements

Caiioo Animation

• Enhanced Physics: Improved particle dynamics with better elevation, speed, and density forces
• Smoother Flow: Particles now follow channel direction on spawn, reducing clumping
• Realistic Stacking: Particles stack naturally when blocked by pebbles or dams

Google OAuth Flow

• Faster Sign-In: Now prioritizes ID token retrieval for faster authentication
• Improved Reliability: Better token handling reduces auth failures on non-Chrome browsers

Onboarding & Settings

• Preview Mode: Test onboarding flows without clearing user data
• Cleaner Free Tier: Removed deprecated provisioned API key restoration UI

Internal

• Minigame System: New MinigameContainer infrastructure for interactive intro experiences
• Hidden Easter Egg: Minigame visibility state persisted across sessions

New Features

Interactive Water Simulation

• Caiioo Intro: New interactive fluid simulation on the loading screen - watch water flow through a dynamic S-shaped channel
• Draggable Pebbles: Move pebbles around to redirect water flow and create dams
• Sediment Dynamics: Realistic erosion and deposition - fast water picks up sediment, slow water deposits it
• Theme-Adaptive: Water colors automatically match your chosen theme tint

Vega/Vega-Lite Chart Support

• Interactive Visualizations: View and edit Vega and Vega-Lite charts directly in Slate
• Data Visualization: Create bar charts, line graphs, scatter plots, and complex multi-layer visualizations
• Spec Editing: Edit the JSON specification and see changes rendered in real-time

Mermaid Diagram Support

• Diagram Types: Create flowcharts, sequence diagrams, class diagrams, state diagrams, and more
• Live Preview: Edit Mermaid syntax with instant visual preview
• Export Options: Diagrams render as SVG for crisp output at any size

File Creation Templates

• Quick Create Menu: New dropdown menu when creating files in Slate with templates for common file types
• Template Categories: Markdown, code files, data formats, diagrams, and more
• One-Click Start: Jump straight into a new document with the right structure

Improvements

API Error Handling

• Visual Notifications: API errors now display as dismissible toast notifications
• Auto-Dismiss: Non-critical errors (like cancellations) automatically clear after a few seconds
• Clearer Messages: Better error messages help identify and resolve issues faster

Platform Capabilities

• Smart Feature Detection: Features that require specific platforms (like Apple Calendar on macOS) are now detected automatically
• Graceful Fallbacks: Tools adapt to your environment rather than failing silently
• Apple Tool Improvements: Apple Reminders now supports uncomplete and list actions

Safari & Non-Chrome Browsers

• Better OAuth Flow: Improved browser detection for Google sign-in
• Fallback Mechanism: Non-Chrome browsers now have a more reliable authentication path

New Features

Automatic Data Cleanup

• Storage Management: Old threads and attachments are automatically cleaned up based on your retention preferences
• Configurable Policies: Set how long to keep data before automatic cleanup

GitHub Sync

• Backup to GitHub: Sync your Caiioo data to a GitHub repository for backup and cross-device access
• Smart Conflict Resolution: Changes from multiple devices are automatically merged without data loss
• Selective Sync: Control which data is synced with .gitignore-style patterns

GitHub Tool

• AI GitHub Integration: The AI can now interact with GitHub on your behalf - create issues, browse repositories, manage pull requests, and more
• Repository Browsing: Ask the AI to explore codebases, find files, and understand project structure

Improvements

Browser Automation

• Enhanced Page Interaction: More reliable clicking, scrolling, and form filling on complex web pages
• Smarter Tool Selection: The AI now picks the right tools for each task more accurately

Document Handling

• Better Word Documents: Improved handling of tables, lists, and formatting in DOCX files
• Google Slides: More control over slide formatting, shapes, and layouts

Settings & UI

• Expanded Settings Panel: More configuration options with better organization
• Improved Thread List: Better sorting and filtering of your conversations
• Location Permission: Clearer flow when granting location access for location-based queries

New Features

Google Slides Integration

• AI-Powered Presentations: New Google Slides tool allows the AI to create, read, and modify slide presentations
• Full Slide Control: Create slides, add text boxes, images, shapes, and tables
• Template Support: Use built-in templates or work from blank presentations
• Collaborative Editing: Works with your existing Google account connection

Improvements

Code Quality

• TypeScript Fixes: Cleaned up type errors and unused imports across the codebase
• Test Coverage: Updated test fixtures to match current type definitions

New Features

Resemble.ai Text-to-Speech

• Professional Voice Synthesis: New Resemble.ai integration for high-quality AI voice generation
• Streaming Audio: Real-time audio streaming for responsive voice output
• Configurable Voices: Select from multiple professional voice options
• Truncation Warnings: Clear feedback when long text is truncated for synthesis

Calendar Sync Service

• Background Sync: Automatic calendar synchronization with Google Calendar
• Incremental Updates: Only changed events are synced for efficiency
• Alarm-Based Scheduling: Reliable sync scheduling using Chrome alarms

Improvements

Desktop App (Electron)

• Wake Detection: System now detects when your Mac wakes from sleep and refreshes OAuth tokens automatically
• Improved Reliability: OAuth connections stay fresh even after extended sleep periods

Rich Composer Input

• Enhanced Composition: Improved message input with better formatting support
• Tab Autocomplete: More responsive tab reference suggestions

New Features

Thread Search

• Search Your Conversations: New search box in the thread list to quickly find threads by title or content
• Instant Filtering: Type to filter - matching threads appear immediately with search highlighting
• Smart Debouncing: Search is optimized to not lag even with hundreds of threads

Physics Simulation Tool

• AI-Powered Physics: New physics tool lets the AI perform physics calculations and simulations
• Projectile Motion: Calculate trajectories, predict collisions, and solve motion problems
• Structural Analysis: Analyze stress, beam bending, buckling, and stability of structures
• Material Properties: Built-in database of common engineering materials (steel, aluminum, wood, concrete, etc.)
• Physics Formulas: Kinetic energy, momentum, force, impulse calculations

Cairn World Builder Enhancements

• Physics Engine: Full Rapier3D physics integration for realistic simulations
• Joints & Constraints: Create hinges, sliders, ball joints, and fixed connections between bodies
• Sensors & Triggers: Define sensor regions that detect when objects enter/exit
• Character Controller: First-person character with gravity, jumping, and collision response
• Game Mode: Real-time physics with fixed timestep for interactive exploration

Improvements

Private Sync v2

• Incremental Sync: Each thread and attachment is now synced individually rather than as one large file
• Faster Sync: Only changed items are uploaded, dramatically reducing sync time for large libraries
• Better Conflict Handling: Per-item vector clocks enable more precise merge resolution
• Reduced API Calls: Smart diffing means fewer Google Drive API requests

Sync Reliability

• Extension/Server Parity: Fixed attachment storage to properly track vector clocks on both platforms
• Tombstone Filtering: Deleted profiles are now correctly hidden on both extension and server

New Features

Private Sync (Free Tier)

• Cross-Device Sync: Sync your threads, settings, and attachments across all your devices via Google Drive
• End-to-End Encryption: All synced data is encrypted with your passphrase before leaving your device
• Automatic Background Sync: Changes sync automatically every 30 seconds with smart debouncing
• Conflict Resolution: CRDT-style vector clocks ensure changes merge correctly across devices

Improvements

Private Sync Efficiency

• Reduced Polling: Sync interval increased from 3s to 30s to reduce API calls
• Smart Debouncing: Waits 10 seconds after changes settle before syncing
• In-Flight Protection: Threads being actively processed by the AI are excluded from sync until complete

New Features

Slate Revision Navigation

• Navigate Between Changes: New navigation buttons to jump between pending revision chunks in the editor
• Change Counter: Visual indicator shows current position (e.g., "2 of 5") within pending revisions
• Keyboard Shortcuts: Use Alt+Up/Down to quickly navigate between chunks without leaving the keyboard

Improvements

Document Processing

• Remote OCR Fallback: When local PDF text extraction fails or produces poor results, documents are automatically processed via cloud OCR for improved accuracy
• Processing Status: Real-time feedback shows when documents are being processed remotely

Reasoning Model Support

• Multi-Turn Reasoning: Better caching of reasoning details across conversation turns, improving continuity for extended thinking models
• Cleaner Messages: Internal system notes are now stripped from rendered messages, preventing instruction leakage

New Features

Long-Term Memory

• Context Persistence: AI now maintains long-term memory across conversations, remembering important context about your preferences and workflows
• Usage Tracking: Enhanced tracking of token usage and costs with detailed logging

Model Intelligence

• Smart Model Selection: New model intelligence features help identify optimal models based on your usage patterns
• Ranking System: Models are ranked by performance and value metrics

Improvements

Storage Reliability

• Cross-Platform Storage: Model cache, tool approvals, and learned pages now work reliably across extension, server, and LAN modes
• Auth Timeout: Fixed potential hang when connecting to relay server - now times out gracefully after 5 seconds instead of waiting indefinitely
• Settings Load Speed: Settings and model picker now load faster after extension reload

Content Pagination

• Proper Page Sizing: Web page content is now split based on the actual model's context window (e.g., 131k tokens), not a hardcoded default. This means you see larger page chunks and fewer pages when using high-context models.

Bug Fixes

• API Key Persistence: Fixed issue where OpenRouter API key would be "forgotten" after briefly enabling then disabling LAN mode

New Features

Kokoro TTS - Local Neural Text-to-Speech

• High-Quality Voices: Kokoro is a local neural TTS model with natural-sounding speech across multiple voices
• No API Key Required: Runs entirely on-device using WebGPU/WASM - no cloud services or API keys needed
• Multiple Voices: Choose from American, British, and other accent options with male/female variants
• Long Text Support: Properly handles long text via streaming synthesis - no more cutoffs at 30 seconds
• Clean Speech: Automatically strips markdown formatting (bold, italics, etc.) before speaking

Improvements

Voice Settings

• Unified Voice Section: TTS and STT settings consolidated in a cleaner layout
• Voice Preview: Test selected voice before using it

Improvements

Settings Panel Search

• Filter Search: New search box at the top of Settings to quickly filter sections by keyword
• Instant Results: Type to filter - matching sections appear immediately
• Keyboard Friendly: Search is auto-focused when opening settings

MCP Server Reliability

• Startup Verification: MCP servers are now verified as running before returning success
• Better Error Messages: When MCP servers crash during startup, the actual error is shown instead of generic failure
• Fixed Examples: Corrected example package names to use @modelcontextprotocol/server-* (not @anthropic/mcp-server-*)

macOS App Authentication

• Self-validating Tokens: Fixed 401 errors when adding MCP servers before WebSocket state sync
• Faster Auth: HTTP endpoints no longer require waiting for WebSocket connection

New Features

ElevenLabs Voice Integration (BYOK)

• Cloud TTS: High-quality text-to-speech using ElevenLabs - choose from multiple voices and models
• Cloud STT: Scribe transcription with real-time streaming (~150ms latency) and 90+ language support
• Voice Selection: Browse and select from ElevenLabs voice library directly in Settings
• Model Options: Choose between Flash v2.5 (ultra-fast ~75ms), Turbo v2.5, or Multilingual v2 (best quality)
• Bring Your Own Key: Uses your ElevenLabs API key - no additional cost from Caiioo

Multilingual Whisper Model

• Whisper Tiny Multilingual: New local STT option supporting 99 languages (~39MB download)
• Same Size as English-only: Same compact 39MB size as Whisper Tiny English
• Language Detection: Automatically detects spoken language

Google Docs Enhanced Reading & Writing

• Markdown by Default: Text is now formatted as markdown by default when writing. Use useMarkdown: false for plain text insertions that preserve existing formatting.
• Rich Formatting: Converts markdown headings, bold, italic, strikethrough, and links to native Google Docs styles
• Lists & Tables: Supports ordered/unordered lists with nesting and markdown tables
• Suggestion Tracking: Pending suggestions shown with semantic tags: <ins>added text</ins> for insertions, <del>removed text</del> for deletions. Adjacent tags indicate replacements.
• Inline Comments: Comments appear inline with author attribution: <comment author="Name" on="quoted text">content</comment> with nested <reply> tags for threads

Improvements

Gemini Extended Thinking

• Improved Thought Signatures: Better handling of Gemini's thought_signature across streaming chunks, improving extended thinking continuity with multi-turn tool use

New Features

Google Sheets Cell Metadata

• Read Hyperlinks & Notes: Use includeMetadata: true to retrieve hyperlinks, notes, and data validation rules from cells
• Add Hyperlinks: New update_cell_metadata action to add clickable hyperlinks with custom display text
• Add Notes: Attach notes/comments to cells programmatically
• Data Validation: Create dropdowns, number ranges, text validation, and custom formula rules on cells

Safari Tiling (macOS)

• Smart Window Positioning: When opening links from the sidepanel, Safari windows automatically position next to the sidepanel for easy side-by-side browsing
• Screen Space Optimization: Tiling logic calculates optimal Safari placement based on available screen space
• Re-tile on Mode Change: Safari windows automatically reposition when the sidepanel changes modes

Copy/Paste in macOS Sidepanel

• Full Copy/Paste Support: Copy and paste now works reliably in the macOS sidepanel app
• System Keyboard Shortcuts: Standard ⌘C/⌘V shortcuts work as expected

Improvements

macOS Server Reliability

• Signal Handling: Improved handling of pipe signals to prevent unexpected app termination
• Restart Reliability: Server stop and restart operations are now more reliable with proper cleanup
• Connection Stability: Better handling of OAuth token refresh with retry logic and exponential backoff

Safari Extension

• Stable Browser IDs: Safari extension now generates stable browser IDs to prevent duplicate entries during reconnections
• Cleaner Reconnection: Server-side browser registration uses client-provided stable IDs for cleaner reconnection handling

Model Selector

• Scroll to Selected: When opening the model dropdown, it now automatically scrolls to the currently selected model

Tab References in macOS App

• Works in Sidepanel: Tab references and context now work in the native macOS sidepanel, not just the Chrome extension

New Features

Voice Output (Text-to-Speech)

• Read Aloud: AI responses can now be read aloud using Microsoft Edge TTS
• Auto-Play Option: Enable automatic reading of new AI responses in Settings
• Speed Control: Adjust playback speed from 0.5x to 2x
• Pause/Resume: Control playback with pause and resume buttons on each message

Local Speech-to-Text (Whisper)

• Whisper Upgrade: Opt into local Whisper transcription for more accurate voice input
• Model Download: Download the Whisper Tiny model (~40MB) for offline use
• Privacy: Audio processed locally, never sent to external servers
• Fallback: Falls back to Web Speech API if Whisper unavailable

Unified Agenda Tool

• Single Tool: New agenda tool consolidates calendar and reminder operations
• Multi-Provider: Works with Google Calendar, Apple Calendar, Apple Reminders, and Caiioo reminders
• Simpler for Agents: One tool interface for all scheduling needs

Improvements

Settings Panel

• Persistent Collapse State: Section open/closed states are now remembered across sessions
• Voice Settings: New section for configuring TTS and STT preferences

New Features

DOCX List Support

• Numbered Lists: Word documents with numbered lists now render correctly with proper formatting
• Bulleted Lists: Bullet point lists are preserved and displayed accurately
• Nested Lists: Multi-level list indentation is maintained in the HTML preview

CSV Export for Spreadsheets

• Export as CSV: XLSX files can now be exported as CSV for easy data extraction
• Format Selection: Choose between XLSX or CSV when downloading spreadsheet attachments

New Features

Tool Approval Workflow

• Interactive Approval: Certain tools now require user approval before execution - you'll see a modal asking to approve or deny the action
• Status Tracking: Tool executions now show 'pending approval' and 'denied' states in the timeline
• Safe by Default: Sensitive operations wait for explicit user consent before proceeding

Current Location Variable

• {{currentLocation}}: New variable for adding your current location context to prompts
• Geolocation Permission: Requires browser geolocation permission when first used
• Context Aware: Great for location-based queries like "restaurants near me" or travel planning

Batch Thread Management

• Multi-Select Mode: Toggle selection mode to pick multiple threads at once
• Batch Delete: Delete selected threads in a single action
• Batch Archive/Export: Archive or export multiple threads simultaneously

Image Viewer in Slate

• Dedicated Viewer: Images now open in a full-screen viewer within Slate
• Zoom Controls: Zoom in/out and pan around large images
• Download Option: Quick download button for saving images locally

Improvements

Reasoning Display

• Better Aggregation: Model thinking/reasoning blocks are now properly combined without duplication
• Cleaner Display: Reasoning content from extended thinking models displays more reliably

Ollama Integration

• Streaming Reasoning: Real-time streaming of reasoning/thinking content from local Ollama models
• Better Model Handling: Improved compatibility with Mistral models and strict message ordering requirements

Attachment Management

• Orphaned File Cleanup: New dialog in Settings to manage orphaned attachments that aren't linked to any thread
• Assign to Thread: Move orphaned attachments to existing threads
• Bulk Deletion: Clean up orphaned files to free up storage space

Slate Defaults

• Markdown by Default: When creating a new Slate without specifying type, markdown (.md) is now the default format

New Features

Wait Action for Browser Automation

• Discrete Wait Types: New wait action in browser automation with 4 specialized wait modes:
  • timeout - Simple delay (default 1000ms, max 30000ms) for basic timing
  • selector - Wait for element to appear or disappear (useful for spinners/loaders)
  • network_idle - Wait for fetch/XHR requests to settle (extension-only)
  • animation - Wait for CSS animations and transitions to complete
• Smart Element Visibility: Selector wait checks display, visibility, opacity, and offsetParent for accurate visibility detection

Improvements

Browser Automation Architecture

• Unified Script Execution: New executeInPage abstraction ensures consistent behavior across Chrome extension and relay/server contexts
• MAIN World Execution: Proper MAIN world script execution for operations that need to intercept page-level JavaScript (fetch, XHR)
• Graceful Degradation: Extension-only features now provide helpful error messages with alternatives when used in relay mode

Bug Fixes

OAuth Token Refresh

• Chrome Extension Token Refresh: Fixed stale token issue where Chrome's cached OAuth tokens weren't being refreshed properly. Tokens are now verified and stale tokens are cleared before retry.
• PKCE Authorization Flow: OAuth popup now uses authorization code flow with PKCE instead of implicit flow, enabling proper refresh token support without requiring a client secret.
• Direct Token Refresh: Connections with refresh tokens can now be refreshed directly using Google's token endpoint, without requiring a backend server.
• Token Validation: Added token verification step to catch revoked/invalid tokens early and trigger automatic re-authentication.

New Features

Apple Calendar Integration

• Native Calendar Access: New Apple Calendar tool provides fast, native access to your macOS calendars via EventKit
• Unified Agenda: Combined view of reminders and calendar events from all connected accounts (Google Calendar + Apple Calendar)
• Multi-Account Support: Pull events from multiple Google Calendar accounts and Apple calendars simultaneously

Learned Pages

• Smart Page Learning: Teach Caiioo about specific websites by capturing their structure
• URL Pattern Matching: Learned patterns automatically apply to similar pages on the same site
• DOM Snapshot Storage: Captured page structures help the AI better understand and interact with complex web apps

RTF Document Support

• Rich Text Editing: Upload and edit RTF (Rich Text Format) documents directly in Slate
• Bidirectional Conversion: Convert between RTF and HTML while preserving formatting
• Export Options: Download edited documents as RTF for use in Pages, Word, or other word processors

Improvements

Desktop App Security

• Relay Authentication: Secure HMAC-SHA256 authentication between extension and desktop server
• Per-User Tokens: Auth tokens are now tied to user identity for multi-user security
• Protected Endpoints: All sensitive API endpoints now require authentication

Performance

• Swift Helpers: Native Swift binaries for Apple Reminders, Calendar, and Notes provide 10x faster access than AppleScript
• Direct Callers: When running in desktop context, Apple tools bypass HTTP relay for lower latency
• Shared Utilities: Consolidated relay API client eliminates code duplication

Bug Fixes

• Calendar Event Deduplication: Events appearing in multiple calendars are now properly deduplicated in the unified view
• Prompt Caching: Fixed cache control markers being stripped during token estimation

New Features

• Image Deduplication: When you upload the same image multiple times in a conversation, the AI now recognizes it as a duplicate and references the original instead of processing it again. This saves context tokens and helps the AI understand you're referring to the same image.

Improvements

• Document Deduplication: Improved cross-source document matching - the same document content is now recognized whether it comes from a user upload, Gmail attachment, or web page ingestion.

Bug Fixes

• PDF Auth Errors: Fixed issue where authentication errors (401, 403) when fetching protected PDFs were being masked as generic extraction failures. The actual auth error is now properly surfaced with helpful guidance.

Bug Fixes

• Reload/Regenerate Button: Fixed issue where clicking reload on an assistant message would show the old response instead of the new one being generated. The UI now properly switches to the new branch during streaming.

Improvements

• Context Window Management: More conservative token estimation (3 chars/token) now used consistently across all pagination and context calculations, reducing the chance of context overflow errors with large web pages.

Improvements

Timeline & Reasoning Display

• Auto-Collapse on Completion: Reasoning blocks and process timeline now automatically collapse when generation finishes, keeping the conversation clean while still accessible
• Intermediate Thoughts Visible: Agent's "thinking out loud" between tool calls is now displayed inline in the timeline, giving better insight into the agent's decision-making process
• Tool Action Labels: Tool calls now show the specific action in the label (e.g., "Web Browse → click" instead of just "Web Browse")

Tool Results

• Screenshot Display Fix: Screenshots and images no longer auto-expand in tool results - only rendered text content (like search results) auto-expands

Bug Fixes

• Ollama CORS: Fixed connection issues when using Ollama from the Chrome extension - CORS headers are now automatically handled

Improvements

• Type Safety: Server storage adapter now uses proper TypeScript types instead of any for threads, skills, profiles, MCP servers, and license info
• Build System: Backup directory is now opt-in via caiioo_BACKUP_DIR environment variable (no longer hardcoded)

New Features

DNG/RAW Image Support

• Camera RAW Files: Upload DNG (Digital Negative) files directly from your camera or photo library
• Automatic Preview Extraction: Embedded JPEG previews are extracted from RAW files for fast processing
• Preserve Original Quality: Original RAW data is preserved while AI works with the high-quality preview

Improved Image Handling

• Server-Side Compression: Large images that exceed local compression limits are now processed by the desktop server
• Better Error Feedback: Visual error indicators (red border, alert icon) when image processing fails
• Graceful Fallbacks: Compression automatically falls back to server when offscreen document is unavailable

Improvements

• Shared Agent Architecture: Unified agent runner shared between extension and desktop server for consistent behavior
• Protocol Handler Consolidation: Storage and message handling now uses shared protocol handlers

New Features

Native Mobile Apps

• Android App: Native Kotlin app (android-app/) with WebView + native bridge channels (CalendarContract, SAF/photo picker, AlarmManager, OkHttp streaming)
• Shared UI: Mobile apps use the same React UI as the web extension for consistent experience
• iOS + Android: Native Swift (iOS) and native Kotlin (Android) with matching bridge APIs

Improvements

• Identity Synchronization: Improved sync of license and profile data between extension and desktop server
• Attachment Management: Better handling of attachments in LAN/relay mode
• Extension Client Tracking: Desktop server now tracks connected extension clients with timestamps

Removed

• Swift Relay App: macOS relay functionality now fully handled by native app server (introduced in 0.8.2)

New Features

Cross-Platform Desktop Server

• Caiioo Server: New Electron-based desktop application replaces the macOS-only Swift relay app
• Windows Support: Native Windows installer (NSIS) and portable executable
• Linux Support: AppImage and .deb packages for Linux distributions
• Menu Bar Integration: System tray/menu bar app with status indicators and quick controls

Platform-Specific Script Execution

• Unified Script API: New /api/script endpoint auto-detects platform and uses the appropriate script engine
• PowerShell on Windows: Execute PowerShell scripts for system automation on Windows
• Bash on Linux: Execute shell scripts (bash/sh/zsh) on Linux systems
• AppleScript on macOS: Existing AppleScript/JXA support preserved

Cross-Platform Helpers

• Desktop Notifications: Display notifications using native APIs on all platforms
• Clipboard Access: Read and write clipboard contents cross-platform
• Active Window Detection: Get the foreground application/window title
• System Info: Retrieve OS, CPU, and memory information

Improvements

• Automated Build Pipeline: Version, icons, and licenses sync from main project during build
• Smaller Distribution: Removed redundant mobile app codebases in preparation for unified builds

New Features

Flexible Sign-In Options

• Multiple Authentication Methods: Sign in with Google OAuth, email/password, or license key - choose what works best for you
• Account Linking: Link your Google account to an existing email/password account for seamless access across methods

Custom OAuth for MCP Servers

• Bring Your Own OAuth Credentials: For MCP servers that require pre-registered OAuth clients, you can now provide your own client ID and secret
• Dynamic Token Refresh: OAuth tokens are automatically refreshed, ensuring uninterrupted connections to MCP servers

Improvements

• Document Processing Indicators: Clear visual indicators when processing PDFs, Word documents, and Excel spreadsheets
• Settings Navigation by Tier: Settings panel now intelligently shows relevant options based on your subscription tier
• Better OAuth Error Handling: Improved error messages when MCP server OAuth discovery fails

New Features

MCP Server Browser & Management

• Browse MCP Servers: Search and install MCP servers from both the MCP Registry and npm directly within Settings
• Local MCP Servers: Run MCP servers locally on your machine via the relay app for enhanced privacy and control
• Credential Resolution: MCP servers can now pull credentials from the Credentials Vault for secure authentication
• Health Monitoring: Automatic health checks for connected MCP servers with status indicators

Profile Management

• Multiple Profiles: Create and switch between multiple user profiles
• Profile Switcher: Easy-to-access dropdown for switching profiles in the composer
• Profile Deletion: Remove profiles you no longer need

Google Account Selection

• Account Choice Dialog: When connecting Google services, choose between your Chrome profile account or add a different Google account
• Web OAuth Flow: Option to authenticate via web browser for accounts not signed into Chrome

Enhanced Browser Agent

• Go Back Action: Agent can now navigate back in browser history
• ARIA Snapshot: Capture accessibility tree snapshots for more efficient page analysis with fewer tokens

Mobile Apps

• iOS App: Native iOS client for accessing Caiioo on your local network
• Android App: Native Android client with server discovery and WebView caching

Multi-Device Sync

• Device Identity: Each connected device/extension is tracked with its own identity
• State Synchronization: Real-time state sync across multiple connected extensions
• Web Client Authentication: Secure authentication for web clients connecting to the relay

Improvements

• LLM Provider Key Management: Manage API keys for various LLM providers directly in the Credentials Vault
• MCP Tool Images: MCP tools that return images now display inline in the conversation
• PDF Text Detection: Improved detection of garbled/spaced-out text in PDF extraction with confidence assessment
• CORS Bypass for Images: Images are now fetched via offscreen document to bypass CORS restrictions

Bug Fixes

• Generation State Tracking: Improved final state capture to ensure generation completion is accurately detected

New Features

Excel Spreadsheet Support

• XLSX Import: Upload Excel spreadsheets (.xlsx) directly to your conversations
• Data Extraction: Spreadsheet content is parsed and made available to the AI for analysis
• Round-Trip Editing: Edit spreadsheets in Slate and export back to XLSX format

What's New Dialog

• Update Notifications: After updating Caiioo, a "What's New" dialog automatically shows release notes for the new version
• Version-Specific Notes: See exactly what changed in your update, with formatted feature lists and improvements
• Non-Intrusive: Dialog only appears once per update, and doesn't show on first install

New Features

Word Document Support with Tracked Changes

• DOCX Import: Upload Word documents (.docx) directly to your conversations
• Tracked Changes Visible: See insertions (green) and deletions (red strikethrough) with author and date on hover
• Comment Support: Comments are highlighted in yellow with tooltips showing comment text and author
• Slate Editing: View and edit DOCX content in the slate with full tracked changes styling
• Dark Mode Support: All tracked change and comment styles work in both light and dark themes

New Features

FLUX Image Generator

• AI Image Generation: Generate images from text descriptions using FLUX AI models via OpenRouter
• Image Editing: Edit existing images in your conversation by providing the attachment ID and editing instructions
• Multi-Reference Support: Combine elements from up to 10 images with flux.2-flex model
• Multiple Models: Choose from flux.2-pro (fast, default) or flux.2-flex (max quality)
• Automatic Storage: Generated images are saved to your conversation and displayed inline
• Cost Tracking: Image generation costs are tracked separately and added to thread totals

New Features

Enhanced Model Selector

• Unified Model Picker: Consistent model selection experience across composer and settings
• Privacy Indicators: Shield icon shows models with Zero Data Retention (ZDR) - your prompts won't be used for training
• Vision Support: Eye icon indicates models that can analyze images
• Recommended Models: Star icon highlights recommended choices (Claude Haiku 4.5, Claude Sonnet 4.5)
• Cost Transparency: See pricing per million tokens directly in the model list - easily spot free models
• Icon Legend: Quick reference in settings explains what each indicator means

New Features

PDF Document Ingestion

• Upload PDFs Directly: Attach PDF documents to your messages - they're automatically processed via Mistral OCR
• High-Quality Text Extraction: Tables, figures, equations, and formatting are preserved as Markdown
• Agent Document Ingestion: Agents can process PDF URLs they encounter while browsing using the new ingest_document action
• Persistent Storage: Extracted content is stored for future reference without re-processing costs

Bug Fixes

• Large Image Attachments: Images over 5MB are now automatically compressed before sending to LLM APIs, fixing "image exceeds 5 MB maximum" errors with providers like Google/Gemini

New Features

Multi-Thread Support

• Run Multiple Agents Simultaneously: You can now have up to 3 threads running at the same time
• Visual Running Indicators: Animated dots appear next to thread titles in the sidebar when that thread is actively generating
• Background Processing: Agents work in the background without stealing focus - browse freely while they work
• Thread Isolation: Each thread tracks its own tabs and state independently

Agent Non-Interference

• Stay in Control: When an agent opens tabs or navigates pages, it won't steal your focus if you've moved to a different tab
• Smart Tab Awareness: Agents only work with tabs they created or started with, never following you to new tabs mid-run

Apple Integration (macOS)

• Apple Notes: Read, create, search, and organize notes across folders - AI can help draft and edit notes directly
• Apple Reminders: Manage to-do lists and reminders - create, complete, and organize tasks with AI assistance

New Features

AI Follow-up Suggestions

• Smart Prompts: After the assistant responds, AI-generated follow-up suggestions appear to help continue the conversation
• Context-Aware: Suggestions are based on the conversation context and what you might want to do next

Improvements

LAN Relay Settings

• Display Network Address: When connected to the relay server, Settings now shows the actual IP address and .local hostname for easy mobile device connection
• Simplified Instructions: Connection info appears automatically once the server is running - no need to check the terminal

Bug Fixes

Slate Revision System

• Word-Level Accept/Reject: Individual word changes can now be reliably accepted or rejected without breaking subsequent changes
• Stable Change Tracking: Fixed issue where accepting a change would cause other pending changes to become unclickable
• Markdown Rendering in Preview: Change previews now render markdown formatting (bold, italic, etc.) instead of showing raw asterisks
• Infinite Loop Fix: Fixed browser hang when accepting pure text additions
• Complete Diff Display: Change preview now shows all deleted and added words, not just minimal differences

New Features

• Basic User Default Experience: New users now start as basic users instead of admin
• License Key Upgrade: Added prominent "Upgrade to Pro" button in Settings to unlock admin features with a license key
• Simplified Onboarding: Cleaner first-time user experience focused on core functionality

Bug Fixes

• Generation Timeout Recovery: Fixed silent failures during long Slate operations where the UI would show "generating" indefinitely
  • Backend now tracks activity during generation and times out after 2 minutes of no progress
  • Streaming chunks, tool execution, and agent decisions all reset the timeout
  • Automatic recovery when service worker restarts mid-generation (orphaned state detection)
  • Error message displayed to user when timeout occurs instead of silent hang

New Features

Thread Import/Export

• Export Selected Threads: Toggle selection mode to pick specific threads to export
• Export All Threads: One-click export of all conversations with attachments
• Import from File: Import threads from exported JSON files
• Selective Import: Choose which threads to import from a file
• Duplicate Detection: Automatically detects threads that already exist with option to skip or overwrite
• Attachment Support: All images, PDFs, and files are included in exports
• Archived Thread Support: Both active and archived threads can be exported/imported

Improved Prompt Caching

• Better cache hit rates for long conversations with Claude models
• Dynamic caching strategy adapts to conversation length

Google Workspace Integration

• Google Drive: Search, create, copy, move, share files and folders
• Gmail: Read emails, manage labels, create drafts (safety-first: drafts only, no auto-send)
• Google Calendar: List calendars, query events, create/update/delete events, find free time slots
• Inline Authorization: Authorize Google access directly from chat without visiting Settings

Web Browsing Improvements

• New click_coordinates action for clicking elements by screen position
• URLs now open in new tabs by default (prevents tab overwriting)
• Tabs automatically grouped per conversation thread
• Better CSS selector detection for reliable element clicking

Model Selection

• Model selector moved to composer area for quick switching
• Model persists globally across mode switches
• Vision-capable models marked with eye icon

Streaming UI

• Tool parameters display as they stream in
• Progressive display of reasoning/thinking blocks
• "Generating..." status indicator during response

Token Usage & Cost Tracking

• Shows input, output, reasoning, and cached tokens
• Displays cost per message and cumulative thread cost

UI Improvements

• Mode Selector Available to All Users: Non-admin users can now switch between modes in Settings
• Collapsible Settings Sections: Settings panel sections can be collapsed/expanded for easier navigation
  • AI & Model Configuration (LLM Provider, API Key, Model, Temperature, Max Iterations)
  • Personalization (Personal Instructions, Profile Variables)
  • Agent Mode Configuration (Mode Selector, Variables, Instructions, Branding, Tools, MCP Servers)
  • Appearance settings
  • Backup & Restore
  • Documentation
• Mode-Specific Theme Colors: Each built-in mode now has a default color scheme
  • Shopping Agent: Green
  • Travel Agent: Blue
  • Helper Agent: Pink
  • Caiioo (General): Purple
• Chromatic Mode: Enable color rotation to gradually cycle through the spectrum
• Theme Override Management:
  • Visual indicator shows when theme is customized (won't be lost on mode switch)
  • "Save as Mode Defaults" button (admin) persists theme as the mode's new default
  • "Reset to Mode Defaults" restores original mode theme colors
• New Caiioo branding and icon
• Mode selector moved to composer area
• New thread button shows mode selection dropdown
• Thread list toggle in composer top bar
• Vignette border effect on controlled browser tabs
• Floating stop button on controlled tabs
• Dropdown menus properly position near screen edges

Bug Fixes

• Model selector updates immediately after settings change
• New user onboarding now correctly launches Helper Agent
• Settings panel no longer crashes with malformed custom variables
• Tab group names update when thread title changes
• Slate accept/reject widgets positioned correctly
• Extended thinking works with more models (Haiku 4.5, Gemini, etc.)
• Fixed "maximum 4 cache_control blocks" error

New Features

Text Selection Context

• "Add to Prompt" button appears when selecting text on webpages
• Selected text shown as chips in composer with page context
• DOM location captured for precise agent interaction

Google Calendar Integration

• Full calendar management (list, create, update, delete events)
• Smart availability search across all calendars
• Natural language time parsing ("tomorrow", "next week")

Bug Fixes

Slate Mode

• Accept/Reject buttons no longer hidden behind editor content
• Word-level diff highlighting (not entire lines)
• Multiple propose_change calls now accumulate correctly
• View toggle no longer "bounces back" unexpectedly

Rebrand

• Renamed from "ContextFlow" to "Caiioo"

New Features

Mobile & LAN Access

• LAN Server: Access Caiioo from any device on your local network
• Conversation Sidebar: Open conversations in a browser tab for larger screen real estate
• Mobile Export: Export conversations as standalone HTML for offline viewing

Bug Fixes

• Fixed critical message branching logic bug
• Fixed archived thread operations

New Features

Extended Thinking Support

• Claude models now support extended thinking/reasoning blocks
• Collapsible reasoning UI to view model thought process
• Reasoning details cached for multi-turn continuity

New LLM Providers

• Ollama Integration: Run local LLMs through Ollama
• Provider selection in settings panel

Image Format Support

• HEIC/HEIF Support: Apple image formats automatically converted for compatibility

UI Improvements

• Tools menu auto-saves on close
• Better settings panel organization

Bug Fixes

• Variable autocomplete positioning improved

Core Features

• Multi-model AI chat (OpenRouter, Anthropic, OpenAI, Ollama)
• Browser automation and web scraping
• Slate for code and document editing
• MCP server integration
• Screenshot and vision capabilities
• Voice input
• Thread management and branching
• Profile and mode system