October 13 additions
Chat, files, and Spaces
- Completed streaming answers keep their place below the message they answer, including the first turn in a new chat and replies restored after switching chats.
- Chat sharing has a clipped preview, a clear public-link notice, and compact sharing targets. Library document and image viewers put Share beside Download; image sharing uses the original Library item.
- PDF, Word, presentation, and spreadsheet files open in their own viewer tabs. The first file selected in an empty Files tab fills that tab; later selections open additional tabs. Breadcrumbs, file options, zoom, download, and sharing controls have more consistent placement.
- Files includes Search text in all files with a flat, accent-colored result list, including when the folder is empty. Document overlays in Spaces include the Ask about this file composer.
- Pages created or moved by an agent appear in the page tree without a reload. The page chat dock can be dragged by its header grip, and markdown blocks the editor cannot render remain editable as plain text.
- Pinned pages and Spaces appear under Pinned without duplicates in the other sidebar lists. Meeting notes are filed in My Meetings and use a document-and-microphone icon.
- MFA sign-in keeps the verified session and finishes its redirect. New-chat sends retain the correct destination, and bot mentions refresh with the current roster.
Desktop and remote work
- Agent Monitor shows active or blocked coding-agent sessions from terminals and supported desktop apps, including Cursor. Background agent avatars appear before the chat actions, with approvals and questions available from their popovers.
- Cursor subscription models are available through the desktop agent integration. Provider artwork, page toolbar icons, and background agent chips are more consistent across themes.
- Tasks started from a phone are named from their first request, marked with their remote origin, and stream their queued turns live. They no longer wait on a desktop folder picker that the person sending cannot see.
- Connections refreshes when a listed device is revoked elsewhere, while unrelated device deletions do not reload the list. Remote environment selection and composer settings have updated controls.
- Computer Use starts its registered server when needed and reports startup failures clearly. Browser scrolling respects the requested behavior.
- Local execution preserves the original human restrictions across child continuations and resumes. Approved file writes also permit reading the approved target; unrelated paths retain their existing restrictions. Bot-host commands can write and reopen their own output without gaining access to the whole app folder.
- Native process supervision, cancellation, cleanup, and runtime-integrity checks are strengthened. Text patches preserve line endings, and local memory checks SQLite integrity before startup.
Agents, tools, and bots
- Work supports visible partial answers while a run continues, with preserved history through replay, interruption, reads, and forks. Response-speed choices stay with saved and queued runs; unsupported billing receipts remain pending instead of being charged as a verified tier.
- Task discovery supports repository exclusions, manual unread receipts keep their opening frontier, and MCP discovery includes callable schemas and resource lifecycles. Work's file panel shows compact Git details for checkouts and worktrees.
- Trusted HTTP hooks, scoped cancellation, completion-hook checks, and MCP destination rules preserve the task's permissions and credential boundaries during long runs and external-agent handoffs.
- Interactive interfaces have additional controls, variants, sizes, editable lists, and progressive rendering. Streamed drafts hand over to the committed interface without duplicate answers or losing input state. Support for ChatGPT's inline interface language is available behind its existing feature gate.
- Bot activity distinguishes self-created tasks, delegated results, and messages sent by a bot. Delegation instructions preserve the requested work. Bot notifications carry the sender identity and signed avatar; Mac communication notifications use the native integration with an Electron fallback.
- Template Creator is an always-on system plugin, with personal-template persistence supported locally and in the cloud once its pending deployment and package publication are complete.
- Work no longer fails when a loaded skill has no authority metadata. Latency improvements reduce redundant startup, memory, capability, goal, and network work while retaining the existing admission checks.
- Nano Banana 2.1 catalog metadata remains gated until provider availability is verified; saved model intent is preserved. Internal design-system specimens and regression coverage now include message parts, bot chat parts, and background coding agents.
Composer and chat
-
Predictions. After a reply finishes, the empty composer can suggest your next message as its placeholder. Press Tab to fill it in as an editable draft, or Escape to dismiss it. Nothing is sent until you send it, typing replaces it, and nothing is suggested when the next step is unclear.
-
Predictions replace the typed-prefix autocomplete rows. Turn them off in Settings → General → Composer → Show predictions. The Input Autocomplete setting is gone, and Suggested prompts keeps its own switch.
-
Chat, Work, Code, and Design each keep their own blank-chat draft and attachments, so text typed for Code no longer appears in Work.
-
Attached files keep their place while they upload and wrap onto new lines instead of scrolling sideways. Image and video thumbnails stay one size whether you attach one or several, and a file the browser can no longer read says "Can't read this file. Attach it again." instead of offering a Retry that cannot work.
-
Search finds Work and Code chats by title, and a multi-word search matches titles that contain every word in any order.
-
Opening a chat from the sidebar or search switches to it right away and shows a loading skeleton, instead of leaving the previous page on screen while the chat loads.
-
If your saved model is not available on your plan or credits, ZeroTwo switches to an included model when you send and tells you, instead of refusing the send again on every retry. Claude Haiku 5.5 is no longer marked premium, and GPT 5.6 Luna now is.
-
The greeting uses the nickname saved in Settings, the formatting toolbar stays hidden over an empty rich-text composer, words such as connectors/MCP are no longer highlighted as slash commands, and /cloud typed on a new chat applies to that chat only.
-
Composer action icons follow your UI font size, including Chat, Work, Code, and Design.
Interactive interfaces
- The assistant can build interactive interfaces inside a chat: inputs, sliders, selects, date pickers, tables, charts, and buttons. Changing a value or pressing Reset updates the interface on your device without asking the model. A button that does real work uses the usual tools, permissions, and approval prompts, and only runs when you click it.
- Interfaces appear where interactive widgets are turned on. Elsewhere, and on public shared pages, you see the same content as readable text. When the assistant updates an interface, what you already typed or selected is kept.
- When the chart widget is available, the assistant now draws charts as chart widgets by default rather than as an image or chart file, unless you ask for an image or a file. Map widgets check each place by name, street, and locality and reject results on the wrong street.
Memory
-
Import memory. In Settings → Personalization → Memory, choose Import, copy the prompt into ChatGPT, Claude, or Gemini, and paste its answer back. ZeroTwo saves each line as a memory and skips ones you already have, ones you asked it to forget, and lines that look like a password, key, or email address. One import saves up to 100 new memories; paste the same list again to bring in the rest.
-
Importing conversations from ChatGPT, Claude, or Grok now also queues memory learning from the most recent imported chats (up to 25), and the import message says so.
-
On desktop, notes added with Add or update are listed under "Waiting to be added to your memory" until the next consolidation. The summary stays on screen, dimmed, while it updates, and no longer regenerates and rewords itself each time you open it when nothing changed. Asking the agent to remember something saves a note directly instead of asking you to pick a memory folder.
-
Chat, Work, and Code now share the same memory manager. On desktop, the summary shows pending local notes without creating a second, separate Code memory view.
Work, Code, and agent runs
-
The plan bar shows "files changed" only after the agent has changed a file, counts only the agent's own edits rather than everything in the folder, and lists each diff when you hover or focus it. It goes away once the final answer is complete.
-
While the agent is thinking, the status shows the thought's own heading instead of a generic "Thinking…". The collapsed row for a question the agent asked now shows your answers, for example "Asked 2 questions: …".
-
Queued messages get a drag handle on every row, visible on touch screens too. A queued message that already reached the transcript no longer also shows a red Failed chip, and a task completion that arrives during a run appears as a compact result chip instead of a bubble from you.
-
Permission requests from the agent use the same compact Allow or Deny prompt as tool approvals, with the safest choice already selected. Skipping a Library approval counts as a no, so the agent does not ask the same question again.
-
A blocked goal says why in its banner, and a goal that hits a temporary error retries with growing delays instead of stalling at once.
-
A run you stopped is never resumed automatically.
-
A run parked for your computer now says why (offline, a window crash, or a lost connection) and whether it will pick up on its own. Pressing Continue from your computer on a task that moved to the cloud goes back to running there when it is online, and the Work location pill reads "Running in the cloud sandbox" while the task continues in the cloud.
-
Scheduled tasks started with Run now, queued follow-ups, and tool calls after a window reload no longer fail on desktop with an environment error. A run in a chat you are not viewing no longer opens a native folder picker over what you are doing.
-
Relative file links in answers, such as outputs/report.xlsx, open the file from the chat's working folder on desktop, and each file appears once instead of as both a link and a card. Links wrapped around a file card no longer leave stray brackets.
-
Running chats are lighter to draw and to keep current: sub-agent pills stop pulsing once a later step starts, shimmer text and sidebar orbs animate with less GPU work, and background chats update the open chat, composer, side panels, and desktop pet less often.
-
Saved sub-agents can continue with follow-up messages using their original model and reasoning level. Completed results are saved before a later lifecycle update or continuation can replace them, with bounded retries for temporary persistence failures.
-
A Work run that fails during startup now settles its admitted run and saved error reply rather than leaving the task marked as running. Temporary provider server errors use the existing bounded recovery path.
-
Generated files expose the same Review actions when a subscription model creates them.
Scheduled tasks
- A new one-time task starts on the next 9 AM that is still ahead of you. Picking a past time shows "That time has already passed" and blocks Create, and if saving fails a message says why instead of the dialog sitting there silently.
- Next-run labels show the time you set rather than the scheduler's delay of up to two minutes, and a local run that was left overdue reads "Missed run on …".
Library, Images, and Spaces
- Images opens faster: it loads the first page right away and applies the Creations and Uploads filters on the server instead of reading your whole library first. Tiles reserve their space from stored image sizes and the loading skeleton uses the same columns, so the grid no longer shifts as thumbnails arrive.
- Inside a Space, New now offers Design (where Design is enabled) and Image, and creates them there.
- Editing a Library file, including a design file, twice within about a minute no longer locks it against further edits, and Trash works on a file whose earlier edit was left unfinished.
- Uploads that never finish are settled automatically after a few hours, and brief storage hiccups when opening a file are retried.
Images, video, and generation
- Resizing an image to a new aspect ratio pads the original and fills the new border instead of cropping. Unmasked edits keep the source image's size instead of being capped near 1024 px, and very large upscales, such as 4x, are scaled to fit the file format limit and saved instead of failing.
- Image, video, and audio results now include their cost in credits, so the assistant can tell you what a generation cost.
- Video requests now match each model's supported sizes and formats, which fixes failed Seedance, Grok Imagine Video, and Kling runs. Wan 2.6 Flash lists only 720p.
- Repeated or failed image calls are contained: one run can make the same image request at most four times, a failed generation no longer ends with an invented image link, and Chat on a plan without video no longer offers the video tool.
Models and answers
- Chats and saved settings that point at retired models, such as DeepSeek V4 Flash, Kimi K2.5, and Llama 3.3 70B, continue on their current replacements.
- When a model run fails, the message names the cause (model unavailable, rate limited, overloaded, conversation too long, and others) instead of one generic line, and a provider that declines to answer ends as a normal reply suggesting you rephrase.
- Compacting a Work conversation, including
/compact, works again, and automatic compaction with Claude Sonnet 5.5 no longer fails.
- Oversized images are shrunk to fit Anthropic's and OpenAI's limits before sending, Grok handles PDFs through file search, mid-stream OpenAI server errors retry or fall back instead of ending the run, and YouTube videos longer than about an hour are analyzed at lower resolution instead of failing.
- Final answers are cleaned of stray drafting notes and broken file-link fragments, and a reply that continues after hitting the output limit starts its new section on its own line.
Subscription models and sources
- Grok, GPT 6 Luna, and Claude Haiku 5.5 subscription models use your connected provider account on desktop. Subscription OAuth remains local-only; cloud execution is blocked in both the picker and the API.
- Subscription models keep the selected model and reasoning setting through tools and sub-agent follow-ups. Model names and knowledge-cutoff metadata follow the selected upstream model.
- Grok subscription conversations support compaction, Stop and recovery, and collaboration with saved sub-agents. GPT 6 Luna tool calls keep a single identity while their arguments stream, preventing duplicate empty tool calls and question cards.
- GPT 6 Luna uses the shared reasoning choices, and Claude Haiku 5.5 uses adaptive reasoning with the five shared effort levels. Subscription context and output limits stay within their existing provider configuration.
- The Sources panel lists cited sources first, then other reviewed search and fetched sources under More, even when an answer cites only one source or no sources. Reviewed links and text sources open through the existing viewers.
- Fetched pages are retained in live and saved source lists without moving existing citation targets. Provider failures and refusals keep their proper status rather than appearing as a clean completion.
Billing and account
-
Accounts that sign in with a phone number alone can now check out, with Stripe collecting the email. A dropped connection while starting checkout says "Could not reach ZeroTwo. Check your connection and try again.", and returning from a stale or invalid checkout link no longer shows a false failure or upgrade dialog.
-
Adding a card waits until the payment form is ready, and turning on two-factor authentication for an account with no email explains why it cannot.
-
Deleting an account with a large history now clears its data in batches until nothing is left, instead of failing partway.
-
A Work run that is still writing after you delete its chat can no longer bring the chat back as an empty "New Chat".
-
Pro and Pro 2× include 1,500 Firecrawl credits and 10,000 ElevenLabs credits per month. Ultra and Ultra 2× include 3,000 and 20,000 respectively; annual plans receive the same monthly connector allowances. These allowances are tracked separately from ZeroTwo wallet credits.
-
Account usage and connector settings show the included ElevenLabs allowance. An exhausted allowance offers the existing option to connect your own account; your connected account takes precedence over the included connection.
-
Pricing cards use the same plan benefits and icons across Account and upgrade screens, including combined Work and Code access, deep research, Bots, and computer allowances.
Bots
- Bot instructions now treat an explicit request, such as sending an email, as authorization, so bots act on it instead of asking again, and saved routines keep the actions you requested rather than turning them into drafts.
- Handoffs between your bots are treated as work to do within your authority, a delegated task's result is labelled as a task result rather than something you said, and a bot reports a chain of handoffs as finished only when it can see the receipt.
Connectors and tools
-
Connector calls keep working more than five minutes into a long run, because each request now gets a fresh sign-in token.
-
GitHub errors caused by the ZeroTwo GitHub App lacking a repository permission now say so plainly, and the agent looks up your GitHub username instead of guessing it.
-
Sites rejects an invalid version, deployment, or domain id with a clear message. A Site's preview and live addresses follow its current name, saving a version from a local run fails clearly when your computer is not connected, and a Site's private access token is no longer repeated in answers.
-
The agent recovers better from malformed tool calls: renamed or wrapped tool names are repaired, placeholder values in optional fields are accepted, cut-off arguments are counted and explained, and image links passed to web fetch are refused right away with a clear reason.
-
The ElevenLabs starter plugin focuses on six creative workflows: music, sound effects, transcription, speech, voice changing, and voice isolation. Hosted audio tools support included plan credits or your own connected account.
-
The connection included with your plan cannot be edited or disconnected. Private voice and account-management tools require your own account.
Meetings and Design
- Meeting recordings from a meeting bot are compressed so hour-long calls fit within the upload limit. Retrying a meeting whose recording failed to import fetches the recording again instead of finishing with an empty transcript, and transcription requests that failed before leaving the server are fixed.
- Publishing a design as a Site while its HTTPS certificate is still being issued tells you to retry in about two minutes instead of reporting a server error.
Desktop
-
If the app window crashes, ZeroTwo reloads it on its own up to three times in five minutes, then asks whether to reload or quit.
-
Desktop control now names the app that will actually be controlled and ties your approval to it, so "Zero" cannot reach ZeroTwo itself. On Mac it no longer switches Spaces to reach a window, and says clearly when the window is on another Space.
-
Brief network drops are handled: request signing is retried, tool results are re-sent, concurrent security-token refreshes no longer cancel each other, a send that cannot go out keeps its draft, and queued follow-ups wait for you to sign in again instead of retrying into errors.
-
The browser tool keeps working on a tab after you leave the chat that opened it, uploads from outside the workspace through Chrome DevTools ask you per file, and a question from an extension names the other chat when it comes from a chat you are not viewing.
-
Settings → General adds Projectless task folder. Choose where new local Work and Code tasks outside projects create their dated folders; existing tasks keep their saved paths.
-
Settings → Storage adds This device, with measured local chat and worktree usage, Refresh, and cleanup by age or selected chats. Deletion requires a reviewed selection and confirmation, and preserves active or changed chats, cloud chats, archived chats, saved memories, and worktrees. Existing worktree and archived-chat management remain available there.
Polish and reliability
-
When a new version ships and your open tab can no longer load a file, ZeroTwo reloads once on its own if nothing would be lost (no stream running and no unsent text), instead of leaving an error on screen.
-
Dragging a chat over a collapsed project expands it instead of failing, the Plugins list keeps a stable alphabetical order when you toggle a plugin, wide notifications fit their column, and wide tables show a fade at the right edge while there is more to scroll. Images that cannot load show "could not be loaded" with a Retry instead of a blank tile.
-
Browsers before Chrome 116, Safari 17.4, and Firefox 124 no longer hit errors in Library and other background checks.
-
Backspacing a bullet no longer recreates it when a paragraph follows a list in Pages. In Canvas, image chips no longer flash another image's reference and the edit box uses the main composer's send button. The diff viewer wraps long lines and file names, and the branch pill shows a spinner instead of briefly saying "Detached HEAD".
-
Cloud environment setup and Design for Code have refreshed icons. The shared icon library also includes color filmstrip and speaker artwork.
-
Expanded developer file-preview samples and regression checks for document, spreadsheet, presentation, image, audio, and video formats, alongside desktop storage and subscription-model streaming, tools, and collaboration checks.
Source integration and Git publication do not verify production rollout. Pending database, package, push-function, and runtime stages remain tracked in the shared database update history. Native device acceptance and feature-gated capabilities remain subject to their recorded gates.
Some fixes here, such as automatic settling of unfinished uploads, deleted chats staying deleted, and included ElevenLabs usage, depend on database updates, service settings, and package publication that ship alongside this release. The subscription QA reports retain remaining native execution, terminal-delivery, source-selection, and visual acceptance gaps; this release note does not assert full model parity.