October 30, 2026 · v2.0.9

v2.0.9 lets you steer a running reply, resumes after usage limits, and lets agents drive a browser

Queued messages can steer the agent mid-turn, threads resume on their own when a usage limit resets, and agents can drive the built-in browser and render HTML previews. Antigravity returns alongside Cursor's SDK and OpenCode 2, permission modes match each provider, models load instantly, and every tool step shares one compact row.

What's New

Steer a running reply: A message you queue while the agent works has a Steer button. Press it and Claude, Codex or OpenCode 2 takes the message into the current turn instead of waiting for it to finish. Your message appears above the reply it changed, on desktop and on the phone.
Resume after a usage limit: When a provider hits its limit, the thread shows when it will resume and picks up on its own shortly after the reset, even after a restart. Retry now and Cancel sit on the same banner, and queued messages wait instead of failing one by one.
Agents can drive a browser: Agents can open pages in the built-in browser, read them, take screenshots, click, type, scroll and wait for content. A soft glow around the page shows when the agent is in control, and your own input always wins.
HTML pages in chat: Agents can render reports and visualisations straight into the conversation, with Reload, Full screen and the raw result. Tools that return an app preview show it the same way. Works on the phone too.
Open subagent conversations: A task card can open the subagent's own conversation, with its instructions, replies and tools, read-only.
Math and diagrams: Formulas render properly and Mermaid diagrams render as diagrams, with a Show code toggle.
Images, audio, video and files in messages: Messages show image galleries, inline players and named files with their size.
Reopen closed tabs: Cmd/Ctrl+Shift+T brings back a closed browser, terminal, Review, Files or GitHub tab, and the whole panel layout (tabs, order, width, browser URLs) survives a restart.
Terminals keep their output: Terminal output is still there after you reopen a thread or restart Aiola, and an exited shell offers Restart.
Save a plan to Notebook: A proposed plan can be saved into the project's Notebook in one click.
Usage for every provider: Settings → Usage shows the limits each provider reports, including Antigravity's Gemini and Claude/GPT pools, OpenCode Go and, if you switch it on, Cursor account usage.

Providers

Antigravity is back: Install Google's agent from Settings → Providers, sign in with Google, and use its own models, slash commands, project skills, approvals and images. Each Aiola account gets its own profile.
Cursor runs on its official SDK: Sign in with Cursor inside Aiola. Models and options follow your Cursor account, the full transcript shows (commands, edits, searches, images, plans) and errors are explained. Older Cursor threads keep working.
OpenCode 2: Aiola now uses OpenCode 2, with its own model and skill lists, permission rules and plan agent. Dropped event streams reconnect and catch up. Older OpenCode versions get a clear upgrade message.
Claude: Adds Claude Sonnet 5.5, Haiku 5.5 and Opus 4.5. New chats default to Opus 5.5. Ultracode and Ultrathink effort where the model supports it. Large attachments are explained before they fail, and big edits no longer stall streaming.
Codex: Models and options come from Codex itself, each model uses its own default effort, Plan mode switches inside the same conversation, Codex questions and forms become cards, and token counts no longer inflate on resumed threads. Codex tools no longer fail to start with a database error.
Grok: Supervised, Auto and Full access with real approvals in Supervised, conversations survive a resume, and skills work as native commands.
Permission modes per provider: Supervised, Auto-accept edits, Auto and Full access each have a one-line description, and each provider only lists the modes it can really run. If a saved mode can't run, Aiola shows what it used instead. Each workspace keeps its own default mode, and Cmd/Ctrl+Shift+A cycles modes.

Improvements

Faster threads: Long threads load what's on screen first, big diffs and tool outputs load when you open them, and the sidebar and Central update in small pieces while other threads stream.
Clearer turns: A running turn shows one "Working for Ns" row. A finished turn folds its steps under "Worked for Ns" (or "You stopped after Ns") and keeps the answer and any warnings visible.
One look for every tool step: Commands, edits, searches, tasks, MCP calls and previews share one compact row that opens into a scrollable body. Searches show every query and grouped results. Task cards show their state and time. HTML and images open in a centred full-screen window.
Simpler approvals: Approvals offer Allow once, Allow for session and Deny, and can show the affected files and a coloured diff.
Simpler questions: Question cards are compact and wrap long answers. Fixed-choice questions have no free-text box.
Changed files: Long lists show five files with Show more, every file keeps undo and reapply, and Open diff jumps to that file in Review.
Sidebar shows what needs you: Chats waiting for approval or an answer, failed and running chats stay at the top, with clear status and hover cards.
Better chat formatting: Task checkboxes, footnotes, strikethrough, all heading levels, in-app thread links, skill chips and citation markers. Code and diffs scroll sideways instead of wrapping.
Plans: Plan cards show a Markdown preview with Implement, Implement in new chat, Copy and Save. Earlier revisions fold away.
Providers page: Shows each provider's real install path (shortened in the middle, full path on hover), what is wrong when a provider isn't ready, and no Sign out button.
Models appear instantly: Model lists are cached and refresh in the background, so a new thread never waits for providers, on desktop or phone.
Provider updates: Updating a CLI uses the installer it came from (Homebrew, npm, pnpm, bun, Volta, its own updater, Scoop or winget), checks the new version afterwards, and records failures in App Logs.
Message times: Hover a message to see when it was sent, with duration and cost for replies.
Setup prompts: Analytics, Stripe, App Logs and Feedback setup pages have a Copy prompt button.

Fixes

No surprise sign-outs: When the local server can't check your sign-in, Aiola shows a Retry banner instead of looping, small clock drift is tolerated, and paid accounts are no longer bounced to the paywall on a slow start.
Queues never get stuck: A finished, failed or stopped turn always releases the next queued message, and fast sends no longer leave ghost bubbles.
Stop stays stopped: Late output after Stop keeps steps Cancelled, and provider processes are cleaned up.
Accurate changed files: A turn only lists files the agent changed, not your own edits or pulled commits.
Approvals: A failed approval ends visibly, finished turns no longer look like they wait for approval, and approving one command never approves a different one.
Terminals stay live: A slow viewer no longer freezes the shell, and reconnects don't duplicate output.
Browser: A crashed page offers Reload, a hung page says so, and a page that redirected while loading no longer shows a red error.
GitHub: Pull request checks, reviews and merge state stay in sync.
Chat: New messages stay in view after sending, expanded steps stay open while scrolling, images from tools show once, and failed edits show their error text.
New chats: A new chat starts from the workspace defaults and always uses an installed provider.
Analytics: Relative ranges roll over at midnight in the workspace time zone, presets no longer share the wrong cache, Google's Gemini Notebook crawler is recognised, and bot icons are fixed.
Usage: Cards show loading, unavailable or failed states instead of a made-up zero, and times read in your local clock.

Phone

iPad and dual-screen: Native iPad support with a split view (chat list beside the chat), a Duo action rail, and every screen adapting to rotation and width.
Same features as desktop: Steer, usage-limit resume with Retry and Cancel, HTML pages, diagrams and math, attachments, subagent conversations, per-provider permission modes with short descriptions, and a new Usage page.
Cleaner thread: Finished work folds under its duration, tool cards match desktop, diffs scroll sideways with line numbers, and approvals show files and diffs with Allow once, Allow for session and Reject.
Chat list: Chats that need you come first, and the list stays visible while reconnecting.
Fixes: Locked-Keychain reads no longer look like a sign-out, notification taps open the right chat once, notification approvals wait for your computer, messages go to the right computer, large drafts are kept safely, the startup cache stays small, and sign-in fits short and landscape screens.