Google Antigravity

• Improved /mcp panel navigation so up and down arrow keys strictly navigate between MCP servers and plugins, left and right arrow keys cycle available actions, and Esc exits action mode or the panel.
• Fixed duplicate agent entries appearing in the /agents picker when an agent with the same name is discovered across multiple configuration sources or plugins.
• Fixed tool initialization and agent startup failing when launched from an inaccessible or deleted working directory by using absolute URI schemes for in-memory parameter schemas.
• Fixed headless CLI invocations with piped standard output or standard error hanging on exit by setting FD_CLOEXEC on the preserved streams so child processes do not keep the caller's pipes open.
• Fixed MCP configuration parsing failing when mcp_config.json contains single-line (//) comments, multi-line (/* */) comments, or trailing commas.
• Fixed orphaned annotation files accumulating in local storage when deleting conversations.
• Fixed /btw side questions in conversations with an active /goal being forced to continue and making unintended tool calls.

• Improved /model <name> autocompletion to accept the proposed model name ghost text with Tab.
• Reduced subagent streaming overhead by sending subagent trajectory metadata once per subtrajectory instead of with every step.
• Fixed commands with subcommands (such as models or agents) hanging on an inherited, unclosed standard input pipe instead of executing immediately.
• Fixed CLI crashes caused by prompt hooks by catching hook panics and rejecting nil completion configurations in model requests.
• Fixed tool invocations and results omitting tool-call IDs when reconstructing request history for Gemini models.
• Fixed tool permission prompts for direct and MCP tool calls displaying generic prompt titles instead of their declared human-readable action descriptions.
• Fixed unconfigured or interrupted Google Cloud authentication and onboarding sessions dropping into broken chat sessions instead of prompting with the sign-in screen.
• Fixed transient authentication errors caused by token expiry clock skew by proactively refreshing browser and WIF OAuth tokens five minutes before expiration.

0.1.16 August 31, 2026 Default Gemini 3.8 Flash model upgrade, lightweight agent configuration, and Vertex AI Express mode The 0.1.16 release updates the default model for new agents to gemini-3.8-flash for enhanced reasoning capabilities, introduces a .lightweight() method on agent configs for small and local models, adds Vertex AI Express mode with API keys, and expands tool runner support for callable classes and dataclass functors. Improvements (8) Default Model Update to Gemini 3.8 Flash : The default model for new agents and configurations has been updated to gemini-3.8-flash , delivering higher reasoning quality and stronger task performance. Developers can override this default by specifying model in their configuration: Optimized Lightweight Agent Configuration : New .lightweight() method on agent configurations (including LocalAgentConfig and LiteRTAgentConfig ) applies preset optimizations for smaller, local-running models by configuring minimal tool sets, minimal prompting, and disabling background subagents: Vertex AI Express Mode (API Key Support) : Developers can now connect to Vertex AI using Express mode by providing an API key directly on

• Added a /model <name> argument that switches to a model by name, slug or label and saves it as your default in one step, with the rest of the first matching name offered as ghost text while you type; /model on its own still opens the picker, and an unrecognized name prints the valid ones.
• Improved the /effort hint so it completes what you have actually typed instead of always showing a fixed [low|medium|high] placeholder.
• Improved artifact handling in conversations that produce many files by coalescing bursts of filesystem events into a single rescan.
• Fixed selectable reasoning effort for Gemini 3.1 Pro and Gemini 3.5 Flash when you authenticate with a Gemini API key.
• Fixed the interface redrawing continuously while the tasks panel or a subagent detail panel was open with nothing running, which held process CPU near 32% instead of the roughly 10% it now settles at.
• Fixed a running subagent's elapsed timer freezing on screen whenever the parent agent was itself waiting.
• Fixed an HTTP 502 from the model endpoint ending the whole run instead of being retried, so a transient bad gateway is now treated as the temporary outage it is.

• Added /voice dictation, which transcribes speech straight into the prompt; press f5 or run /voice to start and stop it, and use the mic-serve subcommand to forward a local microphone to a CLI running on another machine over SSH. Only one live dictation stream is allowed per account, and a second one now says so instead of reporting an exhausted quota.
• Added a cost field to the status line data model, exposing the unrounded estimated cost of the current session so a custom status line can display running token spend.
• Improved the agent's code search by running an embedded ripgrep binary instead of shelling out to whatever the machine provides, so searches are faster and behave identically on systems with no ripgrep installed.
• Improved the always-proceed permission mode to auto-approve MCP tool calls and page reads as well, which kept prompting even though the mode exists to run without interruptions.
• Improved allow-always permission suggestions for script runners such as npm run, yarn, pnpm and cargo run by pinning the specific script name, so approving npm run dev no longer grants every script in the project.

• Added skill icon and visual branding support across the CLI, displaying emoji icons declared under metadata.icon in SKILL.md frontmatter across the /skills catalog list view, detail inspection headers, and slash command autocompletion popups, with proper multi-byte Unicode display width calculation to maintain terminal layout alignment.
• Improved @ file path autocompletion by indexing empty directories alongside files in ripgrep search results, allowing unpopulated and directory-only workspace structures to be discovered and traversed during path completion.
• Improved permission management by automatically granting workspace-scoped read access under the default review mode, eliminating repetitive approval prompts for reading or listing files within the workspace root while strictly maintaining confirmation prompts for file modifications and external access.
• Improved Git repository inspection performance by skipping recursive submodule worktree scans while continuing to track commit pointer updates to eliminate status latency in repositories with submodules.

0.1.15 August 25, 2026 Subagent-scoped custom tools, context compaction lifecycle hooks, and universal JSON schema normalization The 0.1.15 release introduces subagent-exclusive tool scoping to reduce context overhead, adds an on_compaction lifecycle hook for observing context checkpointing, expands platform compatibility with Alpine Linux musl wheels, resolves workspace path normalization and custom Vertex endpoint routing, and adds universal JSON Schema normalization for custom Python tools when targeting local OpenAI-compatible LLM endpoints. Improvements (9) Subagent-Scoped Custom Tools : Custom tools can now be registered directly on subagents without requiring registration on the root agent, strictly isolating them to the subagent's context and reducing root context tokens. Context Compaction Lifecycle Hook : Improved support for the on_compaction lifecycle hook to accurately capture compaction events and summaries when long-running conversations trigger checkpointing. Custom Base URL & Secure Vertex Proxy Support : VertexEndpoint now supports routing requests to custom base_url reverse proxies or enterprise gateways without leaking ambient Google Cloud Application Default

• Added support for passing a project name to --project, which previously accepted only a project ID and failed on anything else.
• Added item.rename and item.delete keybindings for the conversation picker's rename and delete actions, so f2 and f4 can be rebound in keybindings.json on keyboards without function keys.
• Improved @ file path completion with a typo-tolerant fallback, so a query with a transposed or mistyped character still finds the file; typo matches never outrank exact ones.
• Improved audio attachments by recognizing the standard formats Gemini models accept, including wav, mp3, m4a, aac, flac and opus, which were previously refused as unsupported.
• Improved keystroke responsiveness on Windows by discarding the key-release events the console reports and nothing in the CLI reads, which were roughly doubling the rendering work per keystroke.
• Improved the artifact viewer footer by moving the outline shortcut next to the scroll and page hints, so the three ways to move through a document read as one group.

0.1.14 August 21, 2026 Workspace path resolution against current working directory, compaction event deduplication, and Vertex auth overrides Resolves relative workspace paths against the current working directory, prevents duplicate compaction hook dispatches and compaction index inflation, and adds support for Vertex AI gateway authentication overrides. Improvements (1) Vertex Endpoint Authentication Overrides : Enhanced VertexEndpoint to support custom authentication headers and token overrides when routing inference requests through API gateways or proxy servers. Fixes (2) Workspace Path Resolution : Fixed relative workspace paths in LocalAgentConfig(workspaces=[...]) incorrectly resolving against app_data_dir instead of the process working directory. Compaction Hook Deduplication : Fixed on_compaction hook dispatching multiple times per compaction event and resolved over-counting in compaction_indices by adding terminal state deduplication. Patches (0) 0.1.13 August 18, 2026 Pre-tool argument modification, synchronous lifecycle hooks, RunCommandConfig timeouts, and step correlation IDs The 0.1.13 release introduces pre-tool argument modification capabilities in lifecycle