Antigravity SDK 0.1.18

[Improvements] Single compaction threshold dial: Simplified CompactionConfig to a single token_threshold property, deprecating legacy context token limits and interval dials. To configure context compaction, specify CompactionConfig(token_threshold=...).

[Improvements] Excluded ASK_QUESTION from Default Tools: Excluded BuiltinTools.ASK_QUESTION from BuiltinTools.default(). Default configurations run autonomously; agents in headless workflows will no longer attempt interactive user prompts. To re-enable interactive questions, explicitly pass BuiltinTools.ASK_QUESTION in enabled_tools.

[Patches] Removed Deprecated modified_arguments_json: Removed the legacy JSON string fallback in tool hook interception in favor of modified_args.

[Improvements] Antigravity SDK and local models support: Official local model support is now ready to use with first-class LiteRTAgentConfig and LocalOpenAIAgentConfig configurations. Developers can execute on-device models with automated lightweight presets or integrate with local OpenAI-compatible endpoints.

[Improvements] Standardized Evaluation Preset (AgentConfig.eval()): Adds a standardized, benchmark-ready preset that configures autonomous permissions, benchmark retry behavior, daemon execution in commands, and strips image generation and subagents to focus on core coding evaluations.

[Improvements] Custom subagent models: Allows developers to assign specific model targets to subagents independently of the root agent configuration.

[Improvements] Built-in schedule tool: Enables the schedule tool by default across standard tool groups, permitting agents and subagents to set timers and cron-based background jobs alongside task management.

[Improvements] Automatic lightweight presets for LiteRT: Instantiating LiteRTAgentConfig now automatically applies optimized lightweight presets—including reduced prompt overhead and synchronous context compaction—without requiring an explicit call to .lightweight().

[Improvements] Task management pairing: Automatically enables the manage_task tool whenever run_command or schedule is active, allowing background task lifecycle management.

[Improvements] Sandbox availability warning: Added a session startup warning when enable_sandbox=True is requested on an environment or OS backend where sandbox isolation cannot be enforced.

[Improvements] Optimized ToolRunner coercion: Improved type resolution for closure-scoped tools and forward-referenced ToolContext parameters, and introduced TypeAdapter caching to accelerate tool call execution.

[Improvements] Extended JSON Schema normalization: Added schema normalization support for OpenAPI and JSON Schema Draft 7 / 2020-12 keywords (such as multipleOf, prefixItems, and dependentSchemas) and prevented accidental mutation of uppercase sample values.

[Improvements] Top-Level ServiceTier Export: Re-exported ServiceTier at the root package namespace for easier import parity.

[Fixes] Service tier ingestion: Fixed an unhandled ValueError when connecting using gateways reporting unlisted backend service tiers (such as Vertex AI PROVISIONED_THROUGHPUT) by safely ignoring unknown tier values while preserving token counts.

[Improvements] LiteRT and local model examples: Added getting-started guides and end-to-end examples demonstrating on-device execution with LiteRTAgentConfig (Gemma 4 26B) and OpenAI-compatible local endpoints using LocalOpenAIAgentConfig.