name: antigravity-platform description: > Google Antigravity platform knowledge — architecture, primitives, constraints, and best practices for building agent extensions on Antigravity. Use this skill PROACTIVELY whenever designing, creating, or modifying .agents/ files (skills, rules, workflows), when discussing Antigravity platform capabilities or limitations, when building MCP servers for Antigravity, or when structuring any agent Kit for distribution on the Antigravity platform. Also use when user mentions "antigravity", "agent kit", ".agents folder", "GEMINI.md", "skill structure", or asks about platform constraints and file limits. Related: seo (SEO orchestrator), skill-creator (skill building)

Google Antigravity Platform Knowledge Base

Platform Version: Google Antigravity 2.0 (Verified: 2026-05)
Source: antigravity.google

Welcome to the Google Antigravity platform knowledge base. This skill acts as a comprehensive, modular index detailing the architecture, capabilities, constraints, and customization interfaces of the Antigravity developer ecosystem.

🗺️ Modular Platform Documentation

To prevent context bloat and keep information highly scannable, the platform specifications are divided into modular guides under the docs/ subdirectory. Click on any section below to load its dedicated knowledge:

• 🏗️ Architecture & Project-Centric Model
  Learn about the 2.0 Project-centric model spanning multiple folders, isolated workspace folders, and Git worktrees support (Local vs. New Worktree mode).
• 🤖 Platform Primitives
  A deep dive into building blocks: Rules (globs, model decision, manual, 12k char limits), Workflows (/slash commands, chaining, 12k char limits), Skills, and Plugins namespaces.
• 🔄 Asynchronous Subagents & Orchestration
  How to manage parallel background execution via invoke_subagent and define_subagent, subagent lifecycle states (Running, Idle, Killed), communication protocols, the 10-level nesting depth limit, and the /teamwork-preview multi-agent framework.
• 🔒 Permissions & Unified Engine
  Specifications of the unified permission engine. Priority order (Deny > Ask > Allow), syntax format action(target) for system operations, write-implies-read implicit rules, path normalizations, and prompt card scope broadening.
• 🛡️ Terminal Sandboxing & Host Isolation
  Secure sandboxing mechanisms across nsjail (Linux), sandbox-exec (macOS), and AppContainer (Windows), outbound exfiltration prevention via network firewalls (AllowedDomains), and dynamic read-only/read-write mounts.
• ⚓ Hooks & Lifecycle Events Pipeline
  Run shell scripts at critical lifecycle points via .agents/hooks.json. Triggers (PreInvocation, PreToolUse, PostToolUse, PostInvocation, Stop), regex matcher filtering, and camelCase JSON stdin/stdout contracts.
• 🔌 Model Context Protocol (MCP) Server Configuration
  Connect external tools, databases, and schemas. Config file layout (mcp_config.json), stdio/websocket properties, Application Default Credentials (ADC), dynamic client registration, and browser-driven OAuth flow authentication.
• 🛠️ CLI agy Reference, Python SDK, & Google Bundles
  Keyboard-centric terminal agy TUI commands (/resume, /rewind, /permissions, /model, /keybindings, /tasks, /skills, /mcp, /agents, /open), programmatic python SDK harness, and pre-packaged extension bundles (Firebase, Android CLI, Science, Chrome DevTools).

⚡ Platform Quick-Reference Sheet

1. Unified Safety Priorities

Deny List > Ask List > Allow List

Any conflicting rule is resolved by strict precedence: Deny always overrides Ask, which always overrides Allow.

2. Customization Staging Directories

• Local Project level: .agents/ (e.g. .agents/rules/, .agents/workflows/, .agents/skills/, .agents/plugins/). Legacy .agents/ maintained as fallback.
• Global level (GUI): ~/.gemini/antigravity/
• Global level (CLI): ~/.gemini/antigravity-cli/
• Global plugins staging: ~/.gemini/config/plugins/ (or CLI ~/.gemini/antigravity-cli/plugins/)

3. Subagent Nesting Limit

Subagents can run in parallel and delegate recursively down to a maximum of 10 nested levels. Beyond this layer, the platform hard-blocks new subagent calls to prevent runaway resource exhaustion.

4. Interactive Permissions Broadening

When prompting under Ask mode, you can broaden targets directly inside interactive UI card prompts to prevent repeated popups (e.g. edit read_file(src/App.tsx) to read_file(src/) before clicking Allow).