Starting and managing conversations
Starting and managing conversations
Starting a new conversation is now a proper setup screen instead of a blank chat that inherits whatever defaults happen to be active. Before anything begins, you can choose the agent, model, reasoning level, opening message or image, workspace, and the computer that should run it, then use groups, transcript search, status, and batch actions to manage the work later.Worktree conversations open immediately with the original message while naming, checkout, and setup continue. If a remote start fails, the submitted prompt and attachments stay visible with a same-session Retry or a read-only Check delivery action, and later messages wait until OpenClaw knows what happened instead of risking a duplicate or sending out of order.Supported Codex conversations can branch from an earlier message without changing the original, and explicit forks can use more than 100,000 tokens when the selected child model can safely accept that context. Automatic thread and subagent forks keep the 100,000-token limit.
Chat and history
Chat and history
Chat now feels much more alive while a Claw is working. Replies render as Markdown while they stream, older messages load without knocking you out of place, and recent conversations can appear before the Gateway finishes refreshing.The composer is easier to use during long-running work. You can change the model, reasoning, or speed before sending, edit and reorder queued follow-ups, and use dictation without accidentally sending the result.One of the more useful additions is
/btw, which opens a separate multi-turn conversation so you can ask a side question without derailing the work or filling the main history. Faster initial history uses a bounded transcript snapshot stored unencrypted in the browser profile.Follow the Work and Approve Actions
Follow the Work and Approve Actions
Following a Claw while it works is much easier. Tool calls now pair clearly with their results, command activity is easier to read, file changes show focused diffs, and larger screens can keep running background tasks beside the conversation with elapsed time and recent activity. People with permission to change the conversation can stop work there, while the full Tasks page covers work across conversations.Goals now begin from a normal composer mode and stay manageable through Edit, Pause, Resume, and Clear instead of filling the conversation with slash commands. Drafts survive reloads and session switches, and the card remains usable on a phone even when the objective is long. Older Gateways keep the textual /goal workflow rather than these structured controls. If you have used approval gates in ChatGPT or Claude Desktop, the approval flow should also feel familiar, with requests inside the conversation that triggered them and a quiet indicator elsewhere when something needs attention.Once a turn finishes, intermediate commentary and tools fold into a Worked for summary you can reopen. Live activity, search results, attachment-only replies, unresolved errors, and pending approvals stay visible instead of being tidied away before they are actually finished, and closing the approval queue leaves a request pending rather than deciding it for you.
Files, Git, browser, and terminal
Files, Git, browser, and terminal
OpenClaw can now keep the work beside the conversation instead of making you bounce between tools. On supported administrator connections, you can open and edit eligible existing text or Markdown files from Chat, and if a Claw changes the same file while you are working, the editor stops with a conflict instead of overwriting the newer version. The Changes panel shows branch commits and working-tree edits, while eligible GitHub checkouts add pull-request status, CI summaries, and a link to start a pull request on GitHub.A docked Browser panel can navigate, click, type, scroll, inspect elements, and mark up a screenshot before attaching that screenshot and page context to the conversation, but it requires advertised browser support and an administrator connection, and whatever appears in the page or capture may be sensitive.Chat, Details, and Discussion can sit in resizable columns or merge into tabs, while the dashboard, terminal, and desktop can open in focused views when you need more room. The file editor still cannot create or delete files, Changes stays read-only, and Create PR hands off to GitHub rather than submitting inside OpenClaw.
Settings and connected services
Settings and connected services
Settings is now easier to navigate. It has its own full-page workspace, searches both destinations and individual controls, separates common from advanced options, and keeps panel URLs bookmarkable. If a value is incomplete or invalid, the draft stays on screen so you can fix it, and settings the normal form cannot represent safely open in Raw mode instead of being silently lost.Major controls also have clearer homes. Connection details, agent identity and memory, Plugins, MCP servers, devices, channels, and pairing are organized around where you actually manage them, while Model Providers brings together the credential health, available models, quota, balance, budget, and spending data each provider reports. That provider view is read-only, and guided MCP management still lives in Plugins while Settings keeps diagnostics and raw editing.A browser with limited access can request administrator scope, but another administrator still has to approve the signed request from that device. Trusted dashboard links and the native pairing flow can help eligible administrators recover access without turning a shared token into a bypass.
Appearance, language, accessibility, and mobile
Appearance, language, accessibility, and mobile
The Control UI is easier to navigate without a mouse. Keyboard and focus behavior is more consistent across navigation, dialogs, tabs, Settings, plugins, skills, Usage, Workboards, the terminal, and the browser, with skip links, clearer assistive-technology labels, and focus returning to the right place after menus and dialogs close.On phones and tablets, text sizing now reaches navigation and sidebar content. Inputs no longer trigger iOS zoom, touch devices avoid desktop-style hover highlights, and Send or Stop responds to the first tap on the documented Android Chrome and installed-PWA layouts while desktop hover behavior stays intact.Appearance now includes seven built-in themes and separate Interface and Chat prose font pickers. You can mix any of nine self-hosted typefaces, or the system stack, with any theme while JetBrains Mono remains the code face. Authenticated profile choices can follow you across browsers, connections without a writable profile keep them browser-local, and all 20 maintained non-English Control UI locales load only when selected and fill previously missing strings.
Control UI Latency and Responsiveness
Control UI Latency and Responsiveness
The Control UI now reaches Chat faster. In a simulated default-chat test with a mocked Gateway and 50 millisecond HTTP/1.1 latency, JavaScript requests fell from 140 to 45 and startup fell from about 1.6 seconds to 575 milliseconds.It also stays lighter the longer you leave it open. Hidden panels stop fetching data they do not show, retained state is bounded, and returning to a conversation repeats less rendering.Reconnects reject stale responses and keep the right conversation selected as the connection changes. If the page bundle loads but never renders, OpenClaw now shows recovery guidance and a Try again action, while interrupted updates reconcile with the replacement Gateway and confirm success only after it reports the expected version.