Skip to content

Latest commit

Β 

History

5 Commits

Folders and files

NameName
Last commit message
Last commit date
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

ARGUS: Visual / Computer Interaction AI Worker & Vision Shield

High-Performance, On-Device Vision and Computer Interaction Middleware for Autonomous AI Coding Agents (Google Antigravity, Cursor, Cline). Slashes Multimodal Token Bleeding by 97.5% and Eliminates Infinite Retry Loops via Perceptual Circuit Breakers.


🌐 The Ecosystem: ARGUS + NEXUS + BUBU

ARGUS is the dedicated Visual & Computer Interaction Worker within the autonomous three-layer agent ecosystem:

ARGUS (Visual / Computer Interaction AI Worker)
  ↓
NEXUS (Always-On Host Coordinator)
  ↓
BUBU (General-Purpose AI Worker)
  • ARGUS: Visual and computer interaction worker utilizing local GPU to prevent multimodal token bleeding.
  • NEXUS: Always-on host coordinator managing task routing, provider quota (429/TPM) and rate limits, loop detection, and evidence aggregation.
  • BUBU: General-purpose AI worker offloading multi-file context analysis, architectural audits, and memory leak detection.

πŸ“‘ Table of Contents


🚨 The Crisis: Multimodal Token Bleeding

When state-of-the-art AI coding agents interact with desktop environments, browsers, or mobile emulators, they routinely capture and upload full-screen screenshots:

  1. Catastrophic Token Bleed:
    Every full-resolution image sent to remote multimodal LLMs burns 2,000 to 2,500 tokens ($$$/credits). A 40-step automation task incinerates 100,000 tokens within minutes.
  2. The Infinite Freeze Trap:
    If an application hangs, crashes, or a button fails to register, the agent remains blind to the freeze. It re-clicks and re-captures endlessly, burning developer credits until manually aborted.

πŸ›‘οΈ The Solution: ARGUS Shield

ARGUS sits locally between your operating system and your AI coding agent as an intelligent visual shield:

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                          ARGUS SHIELD ENGINE                           β”‚
β”‚                                                                        β”‚
β”‚  [Windows Desktop / BlueStacks]                                        β”‚
β”‚          β”‚                                                             β”‚
β”‚          β–Ό                                                             β”‚
β”‚  [Local Florence-2 Model (CUDA RTX ~680MB)]                            β”‚
β”‚          β”‚                                                             β”‚
β”‚          β”œβ”€β”€ Extracts Focused Cards & Coordinates                      β”‚
β”‚          β”œβ”€β”€ pHash Circuit Breaker (Halts after 3 static frames)       β”‚
β”‚          └── Bypasses Windows Desktop Isolation (OpenDesktopW)         β”‚
β”‚          β”‚                                                             β”‚
β”‚          β–Ό                                                             β”‚
β”‚  [Structured JSON Scene Graph] (<50 Tokens) ──► [Lead AI Agent]       β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
  • 97.5% Token Preservation: Replaces 2,000+ token raw images with high-density ~40 token JSON scene graphs.
  • 100% On-Device Privacy: Zero pixels leave your hardware. All vision embeddings run locally on your GPU.
  • Automated Safety Breaker: Trips immediately if the screen fails to change across 3 consecutive actions.
  • Non-Disruptive Background Operation: Controls Android emulators via ADB without stealing your mouse cursor or closing windows.

⚑ Quick Start (Zero-Prompt Auto-Onboarding)

Get up and running in under 30 seconds:

  1. Copy the Portable Package (Her proje/) into your project:
    Copy the contents of Her proje/ into your target repository root:
    • .agents/skills/argus/ (Vision shield engine and MCP tools)
    • .argus/logs/ (Log directories)
    • AGENTS.md (Zero-token on-demand onboarding rule)
    • .env & .gitignore
  2. Start coding normally: Open a new chat in Antigravity and prompt your agent with any task.
    • Antigravity's AGENTS.md rule automatically fulfills your coding task first.
    • At the end of the response, it asks for your ARGUS and BUBU mode preference once.
    • Zero-Token Guard: If you select direct (disabled), it writes the configuration immediately with zero token waste on reading skills or loading models.
    • If you select auto or local, it activates the shield on-demand and saves your setting permanently.

πŸ€– 1-Prompt Setup with an AI Coding Agent

If you are using Google Antigravity, Cursor, or Cline, paste this prompt into your conversation once:

Install and verify ARGUS for this project.

First, discover the ARGUS skill (.agents/skills/argus) and inspect project configuration.

If my ARGUS mode preference is not yet saved in .argus/config.json, present me with these choices once:
1. auto (Recommended - Hybrid on-device vision, 97.5% token preservation)
2. local (Pure local GPU Florence-2 - Zero cloud upload)
3. direct (Direct cloud vision / Standard screenshots)

If my preference is already saved, do not ask again.
Verify that the installation and hardware status are functional before executing GUI actions.

πŸ€– For AI Coding Agents

When operating in an ARGUS-equipped repository:

  1. Never Call Raw Screenshot Tools: Do not invoke windows-desktop-control:screenshot or view_file on display images. Always use argus.py --scan or smart_ui_scan.
  2. Source Disambiguation:
    • PC / Desktop requests ("ekranΔ±m / masaΓΌstΓΌm") $\rightarrow$ --source desktop.
    • Android / Game / BlueStacks requests $\rightarrow$ --source bluestacks.
  3. Circuit Breaker Discipline: If CIRCUIT_BREAKER_TRIGGERED is returned, stop immediately. Attempt at most 2 retries before escalating to the user (BLOCKED: LOCAL_GROUNDING_FAILED).

βš™οΈ Operational Modes

Configured in .argus/config.json:

Mode Visual Engine Cloud Upload When to Use
auto (Default) On-Device Florence-2 with intelligent fallback Zero (unless escalated) Recommended for all everyday automation and game agent loops.
local Strictly on-device Florence-2 GPU Zero High-security or offline environments where no pixels may leave the machine.
direct Standard Antigravity / Multimodal Cloud Screenshots Full PNGs Systems without dedicated GPU / CUDA where cloud vision is required.

πŸ—οΈ Architectural Highlights

1. Low-VRAM On-Device Vision (Florence-2)

ARGUS utilizes Microsoft's Florence-2-base foundation vision model in FP16 precision. Optimized on local hardware, it consumes only ~680 MB of VRAM, coexisting peacefully with active games and IDEs on modern GPUs (e.g. RTX 2060).

2. Perceptual Hash Circuit Breaker (Emniyet Şalteri)

ARGUS computes a 64-bit difference hash (dHash) after every GUI action. If the Hamming distance between frames is $\le 2$ across 3 consecutive steps, ARGUS halts execution, completely immunizing users from infinite token burnout.

3. Windows Desktop Session Isolation Bypass

Windows background agent processes are restricted by default to isolated desktop stations, causing OSError: screen grab failed. ARGUS re-binds calling threads directly to user32.OpenDesktopW("Default"), allowing direct inspection of active applications without minimizing or disrupting user windows.

4. Silent Headless Android & BlueStacks Automation

Using ADB standard stream pipes (screencap -p and input tap), ARGUS executes touch interactions, D-Pad navigation, and layout scraping completely in the background without commandeering the physical mouse pointer.


πŸ—ΊοΈ Repository & Documentation Map

ARGUS/
β”œβ”€β”€ README.md                          # Master product overview & quick start
β”œβ”€β”€ LICENSE                            # MIT License (Aethelion)
β”œβ”€β”€ INSTALL.md                         # Canonical installation & payload classification
β”œβ”€β”€ SECURITY.md                        # Threat model & privacy boundaries
β”œβ”€β”€ requirements.txt                   # Python dependencies
β”œβ”€β”€ toggle_shield.bat                  # One-click desktop status inspector
β”œβ”€β”€ .env.example                       # Environment template
β”œβ”€β”€ .gitignore                         # Runtime ignore rules
β”‚
β”œβ”€β”€ .agents/skills/argus/
β”‚   β”œβ”€β”€ SKILL.md                       # Antigravity skill interface definition
β”‚   └── scripts/
β”‚       β”œβ”€β”€ argus.py                   # Master CLI controller & mode engine
β”‚       β”œβ”€β”€ engine.py                  # Florence-2 vision engine & screen grabber
β”‚       β”œβ”€β”€ fallback_handler.py        # pHash circuit breaker & XML fallback
β”‚       └── server.py                  # FastMCP server implementation
β”‚
β”œβ”€β”€ .argus/
β”‚   β”œβ”€β”€ config.json.example            # Configuration template
β”‚   └── logs/                          # Runtime breaker logs
β”‚
└── studio/
    β”œβ”€β”€ README.md                      # Studio index
    β”œβ”€β”€ ARCHITECTURE.md                # Token preservation & isolation architecture
    β”œβ”€β”€ CIRCUIT_BREAKER.md             # dHash mathematical model & recovery rules
    └── PROTOCOL.md                    # Structured JSON scene graph contract

πŸ“„ License & Attribution

ARGUS is released as open-source software under the MIT License.
Copyright (c) 2026 Aethelion / aethelondev-stack. All rights reserved.

About

Universal On-Device Vision Shield & Token Guardian for Autonomous AI Coding Agents

Resources

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages