Skip to content

Google Antigravity 1.5.14

Released · Google

Security fix

What the release notes say

  • Added /thinking (/thinking [on|off], saved as "showThinking" in settings.json and off by default) to stream the model's full thinking live in the transcript while it thinks and keep finished thoughts expanded without pressing Ctrl+O.
  • Added a live Overview section to /swarm board on terminals at least 28 rows tall, showing the swarm's goal, overall task progress (Overview · N of M tasks done and any blocked count), and up to six workstream threads with status indicators for which workers are working, in review, waiting on a peer, failed, or done.
  • Added plan mode to --remote-agent sessions, so a remote agent can research and write a plan for your review before editing files or running commands: switch to it with Shift+Tab, start in it with --mode plan, or choose plan under Agent Mode on the Remote Agent tab of /config.
  • Added an Agent Mode setting (default, accept-edits, and plan) to the Remote Agent tab of /config in --remote-agent sessions: choosing a mode saves it as your default for new remote sessions and updates the current conversation, while Shift+Tab switches modes for the session without changing your saved default and --mode overrides a resumed conversation's saved mode.
  • Added a Switch to RTL text direction button to the header of Markdown (.md, .markdown) and .txt file tabs in the side panel of the Remote Control browser view, so right-to-left prose renders cleanly without paragraphs that start with English words snapping left, while code blocks and line numbers stay left-to-right.
  • Improved the Orchestrator row in /model on consumer plans in local sessions so you can pick both its Planner and Executor models and each one's thinking Effort (saved under "orchestrator" in settings.json and --settings): press Tab or Shift+Tab to move between Planner, Executor, and their Effort sliders, use ← and → to change the focused choice, and see the active planner next to Orchestrator in the status line.
  • Improved expanded tool groups in medium verbosity: expanding a group with Ctrl+O or by clicking now keeps the summary header's colors, including red status markers on failed runs, instead of dimming it to gray, and indents the expanded tool calls behind a │ gutter.
  • Improved permission prompts for creating or editing a file outside the workspace in the Remote Control browser view: the "always allow" options now suggest the enclosing repository or project folder (or the parent folder when there is none) instead of only that single file, while sensitive files such as .env or SSH keys and broad folders such as your home directory or /tmp stay limited to the exact file, and editing the target path in the prompt overrides the suggestion.
  • Improved turns that launch several subagents in a row in the Remote Control browser view: consecutive subagent launches now collapse into a single Started N subagents row (Starting N subagents while launching) that you can expand to see and open each subagent.
  • Changed Auto mode (the default Agent Mode) to allow reading files outside the workspace without asking for approval; previously Auto mode prompted before reading outside the workspace unless Non-Workspace Access was turned on in /config. Writing or editing files outside the workspace still asks for approval, and your ask and deny permission rules, protected configuration files such as mcp_config.json, other projects' conversations, and organization policies requiring outside-workspace file review still apply.
  • Fixed a security issue on Windows where network (UNC), device, and NT object paths in workspace permission rules, @ file mentions, plugins, sidecars.json, and custom agents could trigger a lookup that sent Windows credentials to a remote host; those paths are now rejected without being accessed.
  • Fixed model responses that error out, are canceled, or are cut off partway through running partial tool calls from the incomplete response or resending that broken attempt on retry, which could make later turns fail with thinking-signature validation errors.
  • Fixed wide Markdown tables in agent responses cutting off column headers with … and squeezing text columns until words broke mid-word on narrow terminals; headers now wrap and columns are sized to keep words intact while breaking long paths and URLs.
  • Fixed /plugin, agy plugin list, agy plugin enable, and agy plugin install listing the wrong skills, subagents, MCP servers, and hooks for plugins whose plugin.json declares an Agent Plugins $schema (such as reading mcp_config.json instead of mcp.json or reporting hooks.json as hooks named $schema and hooks), and fixed the agy plugin install summary cutting off subagent names at a ..
  • Fixed agy plugin install <plugin>@<marketplace> failing with unknown marketplace when you connect with a GEMINI_API_KEY instead of a saved sign-in.
  • Fixed plugin suggestions in agent responses telling you The <id> plugin needs you to sign in. Open /plugin to sign in. with a raw marketplace catalog ID even when the plugin was not installed; they now say Open /plugin to install or sign in to this plugin.
  • Fixed the status line permanently showing a red AI: Out of credits warning when AI credits are enabled and your credit balance is 0, even while normal model quota is still available; the status line now shows AI credits only when you have a positive balance.
  • Fixed --remote-agent sessions offering local-only and custom models that remote agents cannot run, and showing an unexplained error when a model is unavailable remotely: /model now marks models that remote agents don't support as not available for remote agents and omits local custom models, and requests with an unsupported model report <model> isn't available for remote conversations yet. Choose another model and try again.