Context Usage Extension for Antigravity 2

by José M. Alarcón - Source code at GitHub

1. Description

AGY-ContextExt is a User Interface (UI) extension plugin for the Google Antigravity 2 desktop application (AGY).

This extension shows live data about AGY's current conversation context window usage and subscription rate limits:

The extension shows:

• The name of the active model.
• The total used tokens and the maximum token capacity.
• The total percentage of the used context window.
• The remaining free headroom tokens.
• A visual progress bar with color segments for each data category.
• A visual matrix with 200 colored blocks (each block represents 0.5% of the context window).
• A red square marker in the matrix that pinpoints the start of the "dumb zone" (~150k tokens), where model reasoning and code quality degrade sharply for programming tasks. Hover over it to view status details.
• A breakdown of tokens by category:
  • User messages
  • Agent responses
  • Tool calls and outputs
  • System instructions and cache
• Subscription rate limits by family of models:
  • 5-hour rolling refresh window usage and remaining tokens.
  • Weekly quota usage and remaining tokens.
  • Automatic light/dark mode based on the AGY theme.
• Configurable refresh interval and manual refresh mode:
  • Choose any interval from 1 to 60 seconds, or choose 0 for manual refresh mode. Default is 2s.
  • The indicator shows green "Live" during active polling, and switches to red "Manual" when the interval is 0.
  • A refresh button allows manual updates at any time.
• Model quotas information updates every 2 minutes (cached to prevent API rate limits). You can manually update it using the refresh button at the top-right corner.

2. Requirements

Before you install this extension, make sure that you have:

• Antigravity 2 desktop application installed on your computer.
• Read access to your user configuration directory:
  • Windows: %USERPROFILE%\.gemini\config\
  • macOS and Linux: ~/.gemini/config/
• Antigravity CLI installed if you want to see the current model usage quota information at the bottom of the extension UI.

3. Installation Procedure

Follow these steps to install the plugin into Antigravity 2:

1. IMPORTANT: Close the Antigravity 2 app.

2. Locate your global Antigravity plugins directory:

   • On Windows: %USERPROFILE%\.gemini\config\plugins\
   • On macOS and Linux: ~/.gemini/config/plugins/

   If the plugins folder does not exist, create it.

3. Create a folder named agy-context-ext inside the plugins directory.

4. Copy all files from the src directory of this repository into the agy-context-ext directory.

Or you can run the following commands from the root directory of this repository to achieve the same result:

• Windows (PowerShell):

  Copy-Item -Path ".\src\*" -Destination "$env:USERPROFILE\.gemini\config\plugins\agy-context-ext\" -Recurse -Force

• macOS and Linux (Terminal):

  mkdir -p ~/.gemini/config/plugins/agy-context-ext
  cp -R ./src/* ~/.gemini/config/plugins/agy-context-ext/

4. Activation Procedure

Follow these steps to make sure the extension is installed and enabled in Antigravity:

1. Start the Antigravity 2 desktop application. If Antigravity was open when you copied the files, restart the application.

2. In the left navigation bar, click Customizations.

3. Click the Installed tab.

4. Find Context Usage in the list.

5. Move the toggle switch to ON. Antigravity starts the background sidecar process.

5. Usage

Follow these steps to display the extension during a conversation:

1. Open any chat conversation in Antigravity, even a new empty one.

2. In the top-right corner of the conversation window, click the ⋮ (three dots) menu button.

3. Move your pointer over Extensions.

4. Click Context Usage.

The extension opens in the right side pane (Auxiliary Pane). The panel monitors your conversation and shows the live context usage in (almost) real time.

Note

In the Antigravity left sidebar, an item named agy-context-ext/panel displays automatically:

This is normal. Antigravity displays active background sidecar processes for user security and process visibility. You do not need to pay attention to it or perform any action on it. Just ignore it.

6. Color Reference

The extension uses the following color codes:

Color	Category	Meaning
Blue (#2563eb)	User Messages	Tokens from user inputs
Green (#059669)	Agent Responses	Tokens from agent thoughts and messages
Amber (#d97706)	Tool Calls	Tokens from tool execution calls and tool results
Purple (#7c3aed)	System & Cache	Tokens from system prompts, rules, and cache
Red (#ef4444)	Dumb Zone Limit	Practical review checkpoint for context degradation (~150k tokens)
Light Slate	Free Space	Available unused token capacity

7. Understanding the "Dumb Zone"

Modern Large Language Models (LLMs) support context windows of 1 to 2 million tokens. However, empirical studies show that code quality and complex reasoning degrade long before hitting physical capacity. In software development, this attention degradation is commonly called context rot or the "dumb zone".

Why Context Rot Occurs

The degradation does not come from raw token capacity limits, but from attention dynamics and cognitive noise:

• Information Interference: Previous failed attempts, discarded solutions, raw terminal logs, and conversational history compete directly with your current instructions.
• Lost in the Middle & Imperfect Retrieval: Transformer attention concentrates at the start (system prompt) and end of the prompt. Information stored in the middle of long sessions is frequently overlooked or forgotten.
• Cross-File Relational Complexity: While models can locate isolated facts in large contexts, their ability to reason across interconnected files and multi-step architectures drops significantly (ACL Anthology, 2025).
• Empirical Degradation: Benchmarks such as Context Rot Quantified (Boolean, 2026), LOCA-bench, and Google DeepMind's Gemini 4 Argon GraphWalks confirm performance drops between 15% and 27% across frontier models when context exceeds 128k–256k tokens.

Why 150k Tokens Is Marked (and How to Use It)

The red block at 150,000 tokens provides a practical review checkpoint, not an absolute limit:

• Not an absolute cliff: Exceeding 150k tokens does not automatically break the model. A clean 200k context containing coherent architecture specs and necessary files often performs better than a noisy 90k context filled with compiler stack traces and conflicting directions.
• Do not compact too early: Research by LangWatch (2026) and Microsoft Research (CompactionBench, 2026) demonstrates that aggressive resetting or compacting (/compact) can harm productivity. Critical decisions and project nuances are easily lost during summary compression, with developer corrections jumping up to 40% immediately following compaction.

Important

Behavioral Signal: Treat the red square as an observation cue. Continue working normally, but watch for symptoms of degradation (e.g., repeating discarded bug fixes, forgetting explicit architectural constraints, or editing unintended files). When symptoms appear, start a fresh session with a clear, concise handoff spec.

8. Troubleshooting

Problem: The panel is blank or shows an error

• Cause: The sidecar process did not start correctly.
• Solution: Open Customizations -> Installed. Turn the Context Usage switch to OFF, wait 2 seconds, and turn it to ON.

Problem: The panel does not show changes after an update

• Cause: The application cached the previous web view in memory.
• Solution: Close and restart the Antigravity desktop application.