Antigravity SDK 0.1.7

[Improvements] MCP and subagent support for local models (LiteRT and OpenAI): Developers using local Gemma (LiteRTAgentConfig) or local OpenAI-compatible endpoints (LocalOpenAIAgentConfig, such as Ollama or LM Studio) can now directly configure subagents and register Model Context Protocol (MCP) servers, enabling full local multi-agent and MCP tool workflows.

[Improvements] Multi-threaded hook and tool state handling: Developers can now safely read and mutate shared context variables across multi-threaded tools (asyncio.to_thread or ThreadPoolExecutor) using atomic updates and thread locking:

[Improvements] Custom subprocess environment variables: Developers can now pass custom environment variables directly to isolated agent instances using LocalAgentConfig, avoiding pollution of the global parent environment:

[Improvements] Extra high thinking severity support: Developers can now configure an "extra_high" thinking severity level for complex reasoning tasks without needing to specify or override the model name:

[Improvements] Default image generation model optimization: Updated the default image generation model (DEFAULT_IMAGE_GENERATION_MODEL) to "gemini-3.1-flash-lite-image". Previous default image models often took too long to run on average during standard agent execution loops; this lightweight model ensures dependable, high-speed image generation by default while remaining fully replaceable using explicit model configuration if higher fidelity is required.

[Improvements] Environment hydration: Hydrates GCP/Vertex parameters (project, location, routing) dynamically from standard GOOGLE_CLOUD environment variables when not explicitly passed during LocalAgentConfig initialization.

[Improvements] Local WebSocket connections: Retries socket connections by resolving to both "localhost" and "127.0.0.1", ensuring dependable harness discovery under containerized setups.

[Improvements] Tool runner public asyncness: Preserves original asynchronous and synchronous execution interfaces of tools when accessed using ToolRunner.get_public_callable.

[Fixes] Interactive console spinner clobbering: Fixed a bug in run_interactive_loop where background spinner animation frames (⠼ Reasoning...) continuously clobbered user confirmation prompts (async_input and ASK_USER policy checks) every 80ms. The active spinner is now explicitly cleared (\r\033[K) and paused when an interactive input prompt opens, keeping confirmation lines readable and resuming the spinner smoothly after input is submitted.

[Fixes] SDK idle transitions: Corrected premature agent shutdown situations by properly checking and clearing idling flags once a TrajectoryStateUpdate signals transition to running.

[Fixes] LiteRT connection engine: Rectified local engine initialization bugs, including a missing protobuf import, a token constructor argument mismatch, and OpenAITool inheritance.

[Fixes] Shutdown connection handshake: Added extra execution buffer to local connections during agent shutdown, ensuring safe persistent state storage actions.

[Fixes] Duplicate tool call events: Fixed client rendering issues by filtering out custom tool events from StepUpdate payloads to prevent duplicate event dispatches.

[Improvements] Config inheritance and gaps: Cleaned up redundant base overrides in LocalAgentConfig and corrected Pydantic validation failures arising from None initialization variables.