Google Antigravity

• Fixed a bug where the CLI would temporarily render skill commands without their slash prefix during optimistic updates by deferring prefix stripping to the serialization boundary, ensuring the UI always displays exactly what the user typed.
• Fixed a redundant CLI exit message by removing the "Resume in the same project" hint line, leaving only the standard resume command to simplify exit output.
• Resolved bugs during UI transitions (such as opening subagent details or logging out) by introducing a unified synchronization mechanism that prevents key lockups and ensures overlay panels like the /help view are properly reset.
• Improved command permission security by making "Always Approve" rule matching strict (non-regex) by default, while allowing users to explicitly opt-in to regex matching by prepending rules with regex:.
• Improved command permission usability by relaxing redirection checks, allowing safe commands with output redirection (e.g., tool > file) to match without requiring strict full-command approval.

0.1.5 June 25, 2026 Streaming tool calls, custom subagent delegates, and thread safety enhancements This release introduces native OpenTelemetry tracing support, declarative subagent configurations, improved type safety, and critical robustness and compatibility updates. Improvements (5) OpenTelemetry Tracing Support : Integrates OpenTelemetry tracing into the SDK to translate session, turn, step, and tool lifecycle events into standard GenAI-compliant semantic spans for advanced monitoring and performance debugging, with custom task-safe active span propagation for tool execution. Declarative Subagent Configurations : Added SubagentConfig and SubagentCapabilities in types.py to support constructing static subagents with declarative instructions and tools directly. Type Safety in AgentConfig : Type-annotated policies, hooks, and triggers parameters on AgentConfig and its subclasses to improve type safety and overall developer experience. Lifecycle Hook Routing : Shifted core orchestration of OnSessionStartHook , turn-level hooks ( PRE_TURN and POST_TURN ), and OnSessionEndHook to the connection layer, implementing the Python-side HookRouter for event routing. Public API Cleanup :

• Added support for --project and --new-project launch flags to allow users to explicitly set or create projects, and updated the project resolution logic to default regardless of the active workspace.
• Added a confirmation prompt when pressing Esc in comment mode with unsaved modifications to prevent users from accidentally discarding their work in review views.
• Added dynamic OSC8 terminal hyperlink support to render clickable links in supporting terminals, with automatic fallback stripping for backward compatibility.
• Introduced reverse diff cycling navigation mapped to shift+n in unified diff review mode to allow users to easily cycle backwards through diff blocks.
• Improved permission config merging priorities by ensuring project-specific configurations (located in ~/.gemini/config/projects/) take precedence over global settings in ~/.gemini/antigravity-cli/settings.json.
• Fixed a regression where ctrl+o scrollback clearing failed by restoring the use of cached fields rather than shared pointer comparisons for trajectory toggle detection.

• Added ctrl+c as an exit and interrupt key: the first press cancels active agent operations (like streaming responses), and a double-press triggers the exit flow. Also added a dynamic exit hint in the status line.
• Fixed ctrl+d behavior to act as a forward-delete when the input prompt contains text, only triggering the exit flow when the prompt is empty.
• Improved /resume loading performance by implementing a persistent metadata cache and parallel loader, eliminating severe latency with large conversation histories and preventing background loading log spam.
• Added an expanded AltScreen view for tool confirmations (accessible via ctrl+g), allowing users to view and edit the full command and associated permissions in a dedicated full-screen view, replacing the inline edit (e) key.
• Added the AGY_CLI_CMD_OUTPUT_PERCENTAGE environment variable, allowing users to customize the maximum height of command outputs in the TUI as a percentage of the terminal height.
• Added strict key name validation to the keybindings system to reject invalid key names (like typos) and suggest canonical alternatives, preventing "dead keys" from being registered.

2.1.1 June 22, 2026 Model quota screen and agent security fixes New model quota screen, agent permission and security fixes, and MCP stability improvements Improvements (1) Quota screen redesign: clearer, unambiguous view into “used” versus “remaining” credits in the Models tab of the Settings screen. Fixes (4) Fixed Agent Permission Issue: Fixed the bug where changes to the agent security mode were not persisted. Improved MCP server stability: increased resilience to unresponsive MCP servers and improved browser agent self-troubleshooting abilities when using the Chrome DevTools MCP server. MCP server schema compatibility: the mcp_config.json schema now accepts url in addition to serverUrl as a field. New entries in the sensitive paths list: .vscode and .cache are now recognized as sensitive paths that require explicit user confirmation before the agent can access them in strict mode. Patches (0) 2.0.4 June 2, 2026 Enterprise Authentication Fix Bug fixes Improvements (0) Fixes (1) Fixed the blank screen issue for Enterprise account authentication. Patches (0) 2.0.3 May 21, 2026 Migration flow from Antigravity 1.0 Added a migration flow that allows users to import customizations

• Improved compatibility with a broader set of ARM64 devices (e.g. raspberry pi 4b).
• Added antigravity_guide builtin skill to provide instant, in-context reference guides for the Antigravity 2.0, CLI, IDE, and SDK.
• Improved commit history navigation: scrolling now immediately loads and displays changed files and diffs.
• Improved Git integration by enabling ASCII node graphs (git log --graph) for visual parity with hg/jj.
• Improved commit hash matching to seamlessly resolve short (6-char) to long (64-char) hashes via prefix comparison.
• Added alert message type for system errors/warnings, separating them from standard command output.
• Added the CLI log file path to the /help menu for easy troubleshooting.
• Improved markdown rendering by upgrading glamour to v2.0.1 for cleaner headings and block padding.
• Improved authentication to automatically launch browser sign-in via rundll32.
• Fixed a bug where "ask" permissions were dropped during settings updates, ensuring settings.json preservation.
• Fixed permission engine matching bugs by escaping regex metacharacters (like $ or .) in saved rules, preventing infinite prompt loops.

0.1.4 June 18, 2026 Async tool execution buffers, memory state persistence, and error handling improvements This release introduces major architectural refactorings, public API standardizations, and key new capabilities centered on centralizing model configurations to natively support multi-model backends, exposing a new built-in Google Web Search capability, enabling environment variable passing for Model Context Protocol (MCP) servers, and simplifying the agent initialization flow by removing dynamic runtime registrations. Improvements (10) Built-in Web Search Tool : Exposes the SEARCH_WEB tool directly within the SDK, enabling agents to leverage Google Search for grounded real-time information retrieval, complete with new developer examples ( web_tools.py ). MCP Server Environment Variables : Added support for configuring and passing custom environment variables to launched stdio servers via the new env field in McpStdioServer . Base URL and HTTP Headers Support : Out-of-the-box support for setting custom base URLs and HTTP headers. Image Generation Aspect Ratio : Updated the SDK model config and wrapper to support specifying aspect_ratio within the image creation tool

• Added submodule support for plugins installation. External plugin installation now automatically resolves and initializes Git submodules.
• Optimized customizations permissions: Automatically grants read-only access to the builtin customizations directory, eliminating redundant permission prompts on startup.
• Improved glamour parser error handling (like nested checkboxes inside list emphasis) and preventing it from crashing the TUI, falling back to raw text with a warning banner.
• Updated bubbletea to v2.0.7: Resolves a potential TUI panic when terminal input is unavailable, fixes a data race in mouse handling within the Cursed Renderer, and corrects mouse release behavior under the Kitty Keyboard protocol.
• Hardened command execution permission checks by enforcing strict exact-match verification for PowerShell scripts, complex shell redirections ( > , 2>&1 ), and unparseable strings to prevent sandbox escapes.
• Hardened sandbox execution by adding .git to the core list of dangerous paths, preventing unauthorized or destructive repository modifications.

• Added support for capturing slash command history, allowing users to use the up arrow to replay previously entered slash commands.
• Redesigned the "Models & Quota" page (enabled by default, replacing the legacy usage page) to gracefully handle disabled quota buckets by displaying a dimmed "Disabled" status and omitting the progress bar.
• Added display of quota usage and execution mode in the status line.
• Improved /btw to be more token efficient and support streaming responses for a smoother user experience and fixed premature truncation.
• Fixed a bug where the /hooks command wrote configurations to ~/.gemini/antigravity-cli/hooks.json instead of the shared ~/.gemini/config/hooks.json, ensuring hooks remain synchronized between the TUI and the backend.
• Fixed a CPU compatibility issue (SIGILL on non-AES-NI CPUs), preventing immediate crashes on startup on older CPUs (like Intel Ivy Bridge) or VM environments that lack AES-NI support.
• Added a per-line guard against extremely long single-line pastes in the TUI prompt editor to prevent performance lag, replacing them with an expandable placeholder.

2.1.4 June 11, 2026 Quota Screen Redesign and PDF Support Quota screen redesign, PDF attachment support, new /btw slash command, and other bug fixes Improvements (7) Quota screen redesign: clearer, unambiguous view into “used” versus “remaining” credits in the Models tab of the Settings screen. Ask side questions via /btw: while in a conversation, type ‘/btw’ in the input and select the ‘btw’ option in the menu to write a message to an ephemeral, single-response agent that has the context of your current conversation. Simple conversation search functionality: cmd/ctrl+F to search for visible text in the conversation view. Support for PDF attachments to messages to Gemini models: drag and drop PDFs from your filesystem or add PDFs using the media option in the Add Context menu button in the input box. Breadcrumbs in file viewer pane header: quality-of-life UI to more easily see and navigate directories in your repo. Nested subagents in the overview pane: see all nested subagents belonging to the main conversation instead of only the subagents that are one level deep. Minor improvements to Projects UX: you can now specify a project name during the creation flow and sort

0.1.3 June 11, 2026 Initial subagent orchestration and standard Google AI model targets This release introduces per-server MCP timeout configurations and improves local connection error handling. Improvements (2) Per-Server MCP Timeout : Added configuration support to set custom timeouts (in seconds) for individual MCP servers ( BaseMcpServerConfig.timeout_seconds ). Terminal Error Propagation : The local connection now propagates terminal trajectory errors from the localharness binary as structured AntigravityExecutionError exceptions in the Python SDK during step collection. Fixes (0) Patches (0) 0.1.2 June 4, 2026 Google Antigravity SDK release 0.1.2 updates This release adds Windows platform support, introduces programmatic turn-level stream cancellation, simplifies safety policy configurations for the Model Context Protocol (MCP), and removes the deprecated MCP SSE transport. Improvements (3) Windows Platform Support : Native compatibility added for Windows x86_64 and ARM64 environments. Path and file URI resolution now correctly handles drive letters and directory separators under Windows. Programmatic Turn-Level Cancellation : Added programmatic stream cancellation via

• Added a configurable timeout for launching MCP servers, allowing users to specify a custom timeout or set it to -1 to disable the timeout completely.
• Revamped the artifact viewer gutter numbering and line mapping to accurately align terminal viewport lines with actual 1-based source file line numbers, including support for wrapped lines and collapsed Mermaid diagrams.
• Fixed a bug where the CLI could get stuck in a pending state (showing a transient spinner) after sending a message due to stale status updates.
• Fixed a bug where the wrong workspace directory was displayed in the header and /help menu when multiple workspaces were active.
• Fixed a desync bug in the agent state management where stale callbacks from previous runs could be used upon cache hits in new agent state.
• Fixed Windows-specific sandbox network proxy issues, resolving a hang during connection hijacking and correcting tunnel response protocols.
• Fixed a bug where the archival status timestamp was not correctly saved when archiving conversations.
• Fixed a potential stack overflow crash by introducing a non-recursive warning output mechanism for pre-conversation errors.

• Added shell-style path auto-completion for /open and /add-dir.
• Added optimistic rendering for user chat prompt submissions, injecting messages immediately into the viewport to eliminate perceived input lag.
• Added fuzzy and partial substring matching across slash commands. E.g. /el -> shows /help and /model while previous no suggested completions.
• Fixed a bug when suggestion was not triggered when @ is typed after (. Enabled unconditional typeahead suggestions whenever @ is typed without preceding whitespace, streamlining mention workflows.
• Skipped subagent conversations from /resume, keeping the standalone conversation picker focused purely on direct user initiated conversations.
• Added a stack_with_default flag to the statusLine configuration to render both the default Antigravity status line and custom status line output vertically stacked.
• Fixed a bug where entering a prompt immediately after pressing Esc (to interrupt an active agent stream) caused the newly typed input to be swallowed or rejected.

0.1.2 June 4, 2026 Google Antigravity SDK release 0.1.2 updates This release adds Windows platform support, introduces programmatic turn-level stream cancellation, simplifies safety policy configurations for the Model Context Protocol (MCP), and removes the deprecated MCP SSE transport. Improvements (3) Windows Platform Support : Native compatibility added for Windows x86_64 and ARM64 environments. Path and file URI resolution now correctly handles drive letters and directory separators under Windows. Programmatic Turn-Level Cancellation : Added programmatic stream cancellation via ChatResponse.cancel() . This programmatically aborts active generation turns directly from the client and raises AntigravityCancelledError (subclass of asyncio.CancelledError ) to cleanly signal cancellation in the async flow ( examples/getting_started/cancellation.py ). Direct MCP Safety Policy Configuration : Overloaded policy.allow , policy.deny , and policy.ask_user to accept server configurations ( BaseMcpServerConfig ) directly instead of typing namespaced string paths. Policy evaluation follows a 9-level precedence model (Specific > Prefix Wildcard > Global Wildcard) with longest-match prefix

• Added --model to set model when launching CLI. Also a new models subcommand to list available models.
• Added /permissions command which allows to add/edit/remove permissions rules for each of the three configs above directly inside the CLI.
• Allowed opening the Artifact Review panel (shortcut ctrl+r) while answering pending questions or tool permission confirmations, preserving your current progress when toggling back.
• Fixed a bug that metadata was written in the current directory as opposed to ~/.gemini/antigravity-cli/cache when running using -p.
• Improved statusline layout by merging active tip and artifact status on a single line and truncating with ellipsis on narrow terminals to prevent collisions.
• Improved customization support by allowing directories in the customization manager to be passed as workspace directories, enabling correct trajectory metadata population and /add-dir support.
• Added support for url in mcp_config.json to configure MCP servers directly via a URL.