Antigravity updates

Radar Ultra

Releases: 44

1.5.12

CLIGitHub

AI summary

  • Forking conversations with large artifact folders is much faster: files stay with the original conversation and the forked agent is pointed to them instead of copying them. Deleting the original conversation also deletes those files.
  • The Remote Control browser view now supports split screen, so you can keep conversations side by side. It isn’t available on phones.
  • The CLI now retries agent requests after brief connection problems, so a turn is less likely to stop with an error.
  • Remote Control scrolling now shows when older messages are loading and offers a Retry option if they fail to load, instead of retrying silently as you scroll.
  • With a Gemini API key, Gemini 4 Neon is now available in /model and --model, with Low, Medium, High, and XHigh settings.

Written by AI from the official notes. The original prevails.

1.5.11

CLIGitHub

AI summary

  • You can now upload attachments up to 20 MB in the Remote Control browser view; previously, files larger than a few megabytes could fail.
  • In the Remote Control browser view, clicking a quote in your message now loads and scrolls to it even if it is far back in a long conversation.
  • Remote Control file and artifact tabs now open with the sidebar closed, giving the content the panel’s full width. You can still open the sidebar with “Toggle sidebar.”
  • When using Vim mode, u now undoes one command at a time. Changes that enter Insert mode undo together with the text typed before Esc, and undoing or redoing operators such as dd or diw places the cursor at the start of the changed text.
  • You can validate a plugin before installing it from a local folder, a plugin ZIP file, an HTTPS link to a ZIP, or a GitHub repository URL. For a plugin in a subfolder, use the documented /tree/<branch>/<path> URL format.

Written by AI from the official notes. The original prevails.

1.5.10

CLIGitHub

AI summary

  • Remote-agent sessions now start in the default mode and ask before each workspace file edit. Press Shift+Tab to switch to accept-edits mode, which allows edits without asking.
  • Sandboxed commands can no longer read configuration and credential locations such as ~/.config, XDG_CONFIG_HOME, Windows AppData\Roaming, or ~/.git-credentials by default. Common developer-tool settings remain readable, and sandbox rules can explicitly allow other paths.
  • The CLI now waits and retries when per-minute rate or capacity limits require a delay of up to 4 minutes. Spend-cap, prepaid-balance, and weekly-quota errors stop the turn immediately.
  • Long sessions use less memory by keeping only recently used conversation steps in memory and loading older ones when needed; they also no longer load every earlier subagent’s full history at the start of each turn.
  • Remote Control’s browser view adds a prompt navigator, a Quote button, and an Overview Panel for conversation artifacts, modified files, subagents, and background tasks. Files and reviews now reopen in tabs after a page reload.

Written by AI from the official notes. The original prevails.

2.33.0

AppOfficial notes

AI summary

  • You can connect over SSH to a 64-bit Linux machine. Antigravity sets up its remote server automatically, and projects, agents, worktrees, and terminals run on that host and can keep working after you close the app.
  • GitHub pull request links in chats and URL cards now show live status pills and hover details. You can also see pull requests in the Overview panel and optionally in the conversation sidebar.
  • The built-in browser now supports Cmd/Ctrl+F page search; web apps with their own search continue to handle that shortcut.
  • Business and enterprise users can use /swarm and open UI plugin panels, when their organization’s admin settings allow it.

Written by AI from the official notes. The original prevails.

2.32.2

AppOfficial notes

AI summary

  • Side questions (/btw) now open in a dedicated tab, where you can ask follow-up questions over multiple turns. The agent can read files, search your workspace, and run read-only commands without interrupting the main conversation.
  • Fixed an issue that could truncate diagnostic logs.

Written by AI from the official notes. The original prevails.

1.5.9

CLIGitHub

AI summary

  • Added a workflow tool that lets the agent run a Python script to split work among subagents in parallel or stages and combine their results. It uses Python bundled with the CLI, so no separate Python installation is needed.
  • Organizations can now manage plugin access in admin settings even if they don’t use Gemini Enterprise. They can block all plugins, including built-ins, and prevent installing or enabling them; or disable the marketplace while leaving installed plugins working. Gemini Enterprise users remain subject to their organization’s Cloud Console policy.
  • Fixed a security issue: custom headers such as API keys for an HTTP MCP server are now sent only to that server’s own origin, not forwarded to another host after a redirect.
  • Improved resuming conversations: the picker no longer fetches every open conversation’s full history, avoiding stalls in long sessions. /resume <id> also accepts a prefix matching exactly one conversation and reports an error for an unknown ID instead of opening the picker.
  • Fixed several reliability issues, including cases where malformed terminal-size reports could freeze the CLI or use excessive memory, and where Windows commands or scheduled runs could behave incorrectly. The notes also describe fixes for authentication, settings, sign-in, clipboard, and other CLI issues.

Written by AI from the official notes. The original prevails.

1.5.8

CLIGitHub

AI summary

  • Security and permission checks are more reliable: fixes close a network-rule bypass, prevent unsafe Windows command approvals, and ensure MCP tools and Windows hooks follow their configured approval or blocking rules.
  • You can use Gemini 4 models with a Gemini API key, and Gemini 4 Hydrogen is also available with Application Default Credentials and through an enterprise model gateway. Which models are available depends on how you connect.
  • Enterprise administrators can control web fetching and Remote Control. With an enterprise model gateway, web pages are fetched from your machine, using proxy.address if configured; URL reading is disabled unless that proxy is configured when the relevant setting is used.
  • Starting a remote-agent conversation can be quicker: the CLI prepares a new conversation’s environment earlier, and starts a resumed conversation’s idle environment in the background when it opens.
  • Several reliability fixes help prevent interruptions, including retries after certain network failures, continued delivery of queued messages after an unreadable message, and fixes for Windows shell commands hanging or silently skipped hooks.

Written by AI from the official notes. The original prevails.

2.32.1

AppOfficial notes

AI summary

  • If you sign in with Google Cloud application default credentials, Gemini 4 Hydrogen is now available in the model selector, with low, medium, and high thinking levels.

Written by AI from the official notes. The original prevails.

2.32.0

AppOfficial notes

AI summary

  • Agents can now use Dynamic Workflows to coordinate parallel subagents and multi-step tasks in Python with the runtime included in the app.
  • Agent-controlled browser tabs now have stronger protections: rejected file-type downloads are reported to the agent and shown as blocked, and cross-site frames stay signed out and hidden from screenshots and page snapshots unless explicitly allowed and targeted.
  • If you use remote agents, you can clone Cloud GitHub repositories locally from the workspace picker and sign in to pull request links in the app.
  • You can choose Everforest Light or Dark themes, navigate to previous or next prompts with shortcuts, and optionally set a shortcut to submit a prompt without pressing Enter.
  • Other changes include sign-in alternatives when automatic browser redirect fails, clearer model-loading and plugin connection errors, and fixes for several issues such as conversation scrolling, saved drafts, and opening files.

Written by AI from the official notes. The original prevails.

1.5.7

CLIGitHub

AI summary

  • Agent sessions without a saved mode now start in Auto wherever it’s available, and in Default elsewhere. To keep the previous behavior, choose Default in /config. A strict Tool Permission setting still requires approval for file edits, even in Auto or accept-edits mode.
  • Fixed cases where commands or long output could make the agent seem stuck: long lines no longer freeze the interface, streamed steps keep appearing while tool output is re-rendered, and the agent is notified when a background command is waiting for input.
  • Approval rules now better respect saved preferences and organizational policies: saved command allow and deny rules apply when there is no OS sandbox, while unlisted commands still ask for approval. Commands that leave the terminal sandbox, commands held for review by admin policy, and certain file access outside the workspace still require approval.
  • Fixed browser sign-in on Windows sometimes failing when the browser and CLI used different network address types; the sign-in callback now listens on both IPv4 and IPv6.
  • Fixed Gemini 4 Carbon requests failing through some enterprise model gateways, and fixed formatting loss in background task details and unusable Remote Control links appearing at the end of some replies.

Written by AI from the official notes. The original prevails.

1.5.6

CLIGitHub

AI summary

  • Auto mode is now available for pay-as-you-go Google Cloud accounts that sign in with Application Default Credentials. Gemini Enterprise accounts still follow their organization’s permission settings and are not offered Auto mode.
  • Auto mode asks less often for some safe actions outside the terminal sandbox, including GitHub API reads, downloads into an existing clone, single-file downloads, and reading public SSH-related files. Private keys are still treated as secrets, and sending any .ssh file elsewhere is still blocked.
  • Plugin installs now include skills listed in Claude Code plugin manifests. Failed installs no longer leave a partly copied plugin enabled; unsafe plugin contents are refused, and some invalid command files are skipped with a warning.
  • Enterprise gateway models now receive native tool calls, helping them call tools reliably. Custom Gemini models also get a 65,535-token output limit by default instead of 8,192; an explicit admin limit still takes precedence.
  • Swarm boards now show bee indicators for members, with animated colored cards in terminals that support Kitty graphics. Other terminals show the queen as pixel art only when the window is at least 60×35.

Written by AI from the official notes. The original prevails.

2.31.2

AppOfficial notes

AI summary

  • When signed in with Google Cloud application default credentials, you can select Gemini 4 Carbon and choose a low, medium, or high thinking level.

Written by AI from the official notes. The original prevails.

2.31.1

AppOfficial notes

AI summary

  • On Windows, Gemini Enterprise sign-in with a Workforce Identity Federation account now keeps your selected license, preventing it from switching to Agent Platform and causing permission errors on requests.

Written by AI from the official notes. The original prevails.

1.5.5

CLIGitHub

AI summary

  • Swarm workers now start from a copy of your workspace that includes staged, unstaged, and untracked changes. Outside Git, an isolated workspace copy is available for folders up to 100,000 entries and 4 GiB.
  • Enterprise model gateways now require HTTPS for remote connections; only gateways on the same machine can use plain HTTP. This prevents prompts being sent unencrypted to a remote gateway.
  • Swarms are available as a preview: use /swarm <task> to have worker agents tackle a task in parallel. Each worker uses your quota, so the feature is labeled “Preview & Quota Heavy”; workers pause if their model runs out of quota.
  • You can monitor a running swarm in a live board with member states, models, and posts, or see its latest post and status in the active-items tray. The board opens with alt+s; press Enter on the tray row to open it.
  • Fixed several plugin-import problems: imported Claude Code commands now become properly named skills, and supported hooks are converted instead of being skipped. Hooks without an equivalent are listed as warnings.

Written by AI from the official notes. The original prevails.

1.5.4

CLIGitHub

AI summary

  • Auto mode is now available in agy acp sessions wherever the CLI offers it. It approves edits within the workspace and runs commands in the terminal sandbox; commands outside the sandbox still require approval unless Auto mode’s review clears them.
  • Fixed several agy acp issues: sessions now keep their permission mode when an editor reloads or resumes them, and turns no longer hang when a tool needs another approval—the editor receives a new permission request.
  • Auto mode’s review of commands outside the terminal sandbox now decides faster and times out less often, using a lighter-thinking Gemini Flash model.
  • Fixed subagent crashes, freezes, and questions closing while you answer them, including problems when the terminal is very narrow or its size is not yet known.
  • Fixed sign-in and connection problems: the remote-control service can recover if it starts before the network is available, and affected Windows and macOS sign-ins are repaired automatically when loaded. The missing-permission warning is also shown in --remote-agent sessions.

Written by AI from the official notes. The original prevails.

2.31.0

AppOfficial notes

AI summary

  • Browser downloads now pause until you approve them, and pages controlled by the browser agent can no longer lock your mouse or keyboard.
  • You can open full-page plugin views beside a conversation without navigating away. The window title bar and update button also stay visible on plugin views.
  • You can change the interface font size to 12, 14, or 16 pixels in Settings > Appearance.
  • Voice-call transcript cards now include a short summary. In multi-question prompts, spoken answers advance the question card, and answering on the card tells voice mode to move to the next question.
  • Pay-As-You-Go setup and license switching now suggest Google Cloud projects as you type, rather than requiring you to enter a project ID by hand.

Written by AI from the official notes. The original prevails.

2.30.0

AppOfficial notes

AI summary

  • You can now choose a simpler Knowledge Work layout for research, writing, and everyday tasks. It hides code changes and developer tools and uses plainer, more compact progress updates. This layout isn’t available on business or enterprise accounts.
  • On first sign-in, Antigravity asks what kind of work you do, sets up a matching layout, and suggests ways to get started. Choosing a suggestion can add the plugin it needs and fill in the prompt.
  • Voice mode is now available to everyone except business and enterprise accounts. Remote agents are available to Google AI Ultra subscribers and Google AI Pro subscribers on a paid plan.
  • You can ask Antigravity to bring MCP servers, skills, custom commands, subagents, and instruction files from tools such as Claude Code, Codex, and Cursor into a plugin you own.

Written by AI from the official notes. The original prevails.

1.5.3

CLIGitHub

AI summary

  • Agent turns that waited more than 60 seconds for a model to start responding are now retried, instead of failing with a timeout.
  • File tools now handle UTF-16 and legacy-encoded text instead of treating it as binary, preserve existing encodings when editing, and write PowerShell scripts with non-ASCII text in a format readable by Windows PowerShell 5.1.
  • In remote-agent sessions, files above the remote read limit now load in full up to 20 MiB; larger files report an error instead of appearing as incomplete content.
  • Plugin sign-in requests from the agent now appear in the CLI with instructions to open /plugin and sign in.
  • Fixed plugin management commands and /plugin uninstall for symlink-installed plugins and plugins with spaces or symbols in their names; agy plugin list also shows symlinked plugins again.

Written by AI from the official notes. The original prevails.

2.29.1

AppOfficial notes

AI summary

  • If your sign-in needs updating, voice mode now asks you to sign in again instead of showing “Couldn't connect.”
  • Voice mode now recognizes links and @ mentions in messages you type, including messages that contain only a link.
  • Voice mode is less likely to combine something you said earlier in a call with your latest words, which could prevent it from answering.
  • After a brief drop in a voice call, the transcript no longer shows parts of a reply you never heard.

Written by AI from the official notes. The original prevails.

1.5.2

CLIGitHub

AI summary

  • Plugins and skills from a workspace are now limited to files inside that workspace or its repository root. Unsafe plugin names and attempts to write outside the plugins folder or through symbolic links are also rejected, helping prevent malicious plugins from changing other configuration files.
  • With --dangerously-skip-permissions, system-denied commands remain blocked. In headless runs, commands or file access your Gemini Enterprise administrator requires to be reviewed are denied rather than automatically approved. Explicit deny rules still apply; other commands still run without asking.
  • Remote Control now retries brief network or server errors while setting up a connection instead of dropping the connection.
  • In --remote-agent sessions, skills from connected cloud plugins now appear in the / menu, can be entered at the start of a message, and are listed by plugin under Remote agent skills with /skills.
  • Several workflow fixes: remote projects now accept PDFs, other binary files, and files over about 8 MB; OAuth sign-in for reconnecting MCP servers keeps working with an already-open link; and queued messages no longer trigger a rapid retry loop after a failed turn.

Written by AI from the official notes. The original prevails.

Radar Ultra
An unofficial tool, not affiliated with OpenAI. Compiled from public posts; the original posts are authoritative.