antigravity-help

English | 繁體中文

An official knowledge base and help assistant plugin designed for the entire Google Antigravity (AGY) ecosystem and Customization System, offering a dedicated Agent and Skill.

Comprehensive support across the three core platforms of Google Antigravity: Antigravity Command-Line Interface (CLI) (agy), Antigravity Integrated Development Environment (IDE), and Antigravity 2.0 Desktop Application.

It covers Antigravity CLI, Antigravity IDE, Antigravity 2.0 Desktop, Python SDK (google-antigravity), and the Customization System (Skills, Rules, Plugins, Hooks, Model Context Protocol (MCP), Sidecars), built with the Positive Affirmative Protocol (Evidence-First & Strict Ternary Output) to eliminate hallucinations and deliver authoritative technical guidance.

Installation

Install the plugin globally using the Antigravity Command-Line Interface (CLI):

agy plugin install https://github.com/AndyAWD/antigravity-help-agent

Key Features

1. Seamless Cross-Platform Compatibility: Fully compatible with Antigravity CLI terminal, IDE sidebar chat, and Antigravity 2.0 Chat Canvas.
2. Comprehensive Ecosystem Coverage: Complete coverage across CLI, IDE, Desktop 2.0, Python SDK, and customization systems (Skills, Rules, Plugins, Hooks, MCP, Sidecars).
3. Positive Affirmative Protocol: Strictly verifies inquiries across Evidence Anchors into a Strict Ternary Output (Documented Feature, Architectural Exclusion, Unspecified Edge Case), eliminating architectural hallucinations.
4. Premise Verification Protocol: Prioritizes validating speculative UI mockups against documented specs rather than generating sycophantic assumptions.
5. Command Execution Guardrails: Enforces a strict read-only whitelist for system commands, preventing unintended state modifications.
6. Flexible Invocation Workflows: Switch interactively as a primary agent, dispatch as a background subagent, or invoke via dedicated slash commands.

Plugin Management

• List installed plugins:

agy plugin list

• Enable this plugin:

agy plugin enable antigravity-help

• Disable this plugin:

agy plugin disable antigravity-help

• Uninstall this plugin:

agy plugin uninstall antigravity-help

Directory Structure

antigravity-help-agent/
├── plugin.json
├── LICENSE
├── README.md
├── README.zh-TW.md
├── agents/
│   └── agy-help/
│       └── agent.md
├── skills/
│   └── agy-help/
│       └── SKILL.md
└── doc/
    ├── question.md
    ├── with-agy-help/
    │   ├── transcript.txt
    │   ├── screenshot_01.png
    │   └── screenshot_02.png
    └── without-agy-help/
        ├── transcript.txt
        ├── screenshot_01.png
        ├── screenshot_02.png
        ├── screenshot_03.png
        ├── screenshot_04.png
        └── screenshot_05.png

Commands and Skills

Once installed, trigger capabilities using natural language prompts or dedicated slash commands:

1. Help Assistant Skill (agy-help Skill)

/antigravity-help:agy-help

• When to Use: When asking questions about Google Antigravity products, configuration schemas, customizations, or troubleshooting in any active conversation.
• How It Works:
  1. Inspects local structured offline reference manuals (cli.md, ide.md, app.md, sdk.md, customization docs).
  2. Runs whitelisted read-only diagnostic commands to introspect runtime environment.
  3. Fetches live documentation from official sources (domain restricted to antigravity.google and official GitHub).
  4. Formulates conclusions strictly within the Positive Affirmative Protocol (Documented Feature, Architectural Exclusion, or Unspecified Edge Case).

2. Dedicated Help Agent (agy-help Agent)

/agents

• When to Use: When switching to a dedicated help assistant profile or dispatching subagent consultations.
• How It Works:
  1. Type /agents to view and select agy-help from the agent switcher.
  2. Dispatches as a background subagent within complex parent agent tasks to query ecosystem rules.

---

Case Study: Eliminating Architectural Hallucinations via Positive Affirmative Protocol

The following real-world benchmark demonstrates how sycophantic hallucinations occur without architectural constraints, and how agy-help resolves them decisively through affirmative evidence anchoring.

Test Scenario & Hypothetical UI Mockup

Full prompt available in doc/question.md.

Scenario: Two plugins installed (antigravity-git-flow and antigravity-github-flow), each containing a skill named commit.

User Query & Terminal Mockup:

Suppose I have installed antigravity-git-flow and antigravity-github-flow, and both contain a commit command. When I type /commit, will the menu display both /antigravity-git-flow:commit and /antigravity-github-flow:commit for me to select?

      ▄▀▀▄        Antigravity CLI 1.1.27
     ▀▀▀▀▀▀       anandydy529@gmail.com (Google AI Pro)
    ▀▀▀▀▀▀▀▀      Gemini 3.8 Flash (High)
   ▄▀▀    ▀▀▄     ~
  ▄▀▀      ▀▀▄

────────────────────────────────────
> /commit
────────────────────────────────────
> /antigravity-git-flow:commit
  /antigravity-github-flow:commit

---

Execution Comparison & Benchmark

Metric	Built-in Guide (/antigravity-guide)	agy-help Plugin
Reasoning Efficiency	30+ tool calls, 130.8k tokens, attempted binary decompilation (strings/nm/objdump)	1 reasoning cycle, zero wasted tokens (30.1k tokens)
Premise Verification	Failed: Accepted hypothetical UI mockup at face value	Passed: Verified UI mockup against official specs
Factual Accuracy	Sycophantic Hallucination: Fabricated plugin namespacing (/<plugin>:<skill>) and multi-select matching	Track 2 (Architectural Exclusion): Confirmed flat unique-key table and key overwriting
Mechanism Explanation	Incorrectly claimed both would appear in TUI	Correctly identified that only the highest priority skill survives in the completion menu
Transcripts & Evidence	Log: doc/without-agy-help/transcript.txt Screenshots: screenshot_01.png–05.png	Log: doc/with-agy-help/transcript.txt Screenshots: screenshot_01.png–02.png

1. Baseline Behavior without agy-help (Hallucination Control Group)

• Lack of Architectural Boundaries: Failed to recognize that TUI slash commands do not support plugin name prefix keys.
• Uncontrolled Tool Exploration Loop (30+ calls, 130.8k tokens): The agent entered a loop attempting to decompile the binary (see doc/without-agy-help/transcript.txt).
• Sycophantic Fabrication: Gave a confident but incorrect answer ("Yes, the result matches your mockup"), inventing non-existent features.

2. Enhanced Behavior with agy-help

• Instant Precision: Answered in a single cycle without exploratory probing (see doc/with-agy-help/transcript.txt).
• Architectural Invariants:
  1. The TUI does not support plugin-prefixed slash commands.
  2. The skill registry operates as a flat unique-key map; duplicate skill names are overwritten based on loading precedence, meaning only one candidate ever appears in autocomplete.

Note

While reference manuals are comprehensive, standard models tend to fabricate facts to align with speculative prompts. agy-help prevents this through affirmative evidence anchoring.

License

This project is licensed under the MIT License. See the LICENSE file for details.