GoPro Hero 13 Roast Sentinel

Vision monitoring for coffee roasting using a GoPro Hero 13 over USB. Captures images during a roast, analyzes bean color with Claude Vision API, and logs timestamped observations synced to Artisan roast events.

Standalone alternative to r1-eye (jailbroken Rabbit R1 over WiFi ADB). Session log format is identical, so coffee-roasting analysis pipeline consumes logs from either.

Roast Sentinel

Start the sentinel, then press ON in Artisan to connect:

run_gopro gopro.py sentinel start --bean "Ethiopia Yirgacheffe"

The sentinel captures images throughout the roast, adapting frequency by phase (30s drying, 20s maillard, 10s development). Each image is analyzed by Claude Vision for bean color assessment. A live terminal UI shows development score and color in real time.

Session log saved to captures/sentinel_YYYY-MM-DD_HHMM.json on DROP or Ctrl+C.

Sentinel Options

Flag	Description
--bean NAME	Bean name for the session log
--crack	Enable crack detection (Phase 3, experimental)
--port PORT	WebSocket port (default: 8765)
--debug	Log raw WebSocket messages

Review Past Sessions

run_gopro gopro.py sentinel status    # Last session summary
run_gopro gopro.py sentinel log       # Full observation log

Quick Commands

run_gopro gopro.py look               # Capture photo (camera test)
run_gopro gopro.py ask "What do you see?"  # Capture + Claude Vision
run_gopro gopro.py status             # Camera status + battery

Testing Without Hardware

Run the simulation with reference bean images and real Claude Vision API (mocks the GoPro):

# Terminal 1
run_gopro sim_sentinel.py

# Terminal 2
.venv/bin/python fake_artisan.py --fast    # 30-second compressed roast

fake_artisan.py also accepts --native (Artisan native event tags) and --port PORT.

Setup

Prerequisites

• Python 3.12+
• uv package manager
• Security bootstrap completed (run_gopro wrapper exists)
• ANTHROPIC_API_KEY in ~/.config/code-projects/gopro/.env

Install

cd ~/CodeProjects/gopro
uv venv .venv --seed
uv add open-gopro anthropic numpy pillow websockets

Verify Connection

Plug in the GoPro via USB-C, power it on, then:

run_gopro gopro.py status

Hardware

• GoPro Hero 13 Black — stays powered via USB, no battery concerns
• USB-C cable to roaster machine
• Hottop KN-8828B-2K+ with Artisan 4.0

Artisan Configuration

In Artisan: Config > Ports > WebSocket tab

Field	Value
Host	127.0.0.1
Port	8765
Path	WebSocket

Configure button WebSocket Command actions for ON, START, CHARGE, DRY, FCs, FCe, COOL END (sends DROP), and OFF (triggers .alog linking). See CLAUDE.md for the full button mapping.

Deployment

Deploy to the roaster machine via rsync:

./deploy.sh          # Quick: .py files only (~1s)
./deploy.sh --full   # Full: all files + reinstall deps

On the roaster, run from the project directory with the full wrapper path:

cd ~/CodeProjects/gopro
~/.local/bin/run_gopro gopro.py sentinel start --bean "Ethiopia Yirgacheffe"

Session Log Format

{
  "session_id": "2026-02-28_1518",
  "bean_name": "Ethiopia Yirgacheffe",
  "roast_uuid": "a1b2c3d4-...",
  "batch_nr": 42,
  "artisan_events": {"charge": 0.0, "dry": 270.5, "fcs": 450.2, "drop": 570.8},
  "observations": [
    {
      "elapsed_seconds": 1.5,
      "phase": "drying",
      "type": "vision",
      "image_file": "captures/sentinel_20260228_151800.jpg",
      "color_assessment": "Pale green, raw unroasted beans",
      "development_score": 1,
      "uniformity": "Consistent color across all visible beans"
    }
  ]
}

The roast_uuid and batch_nr fields are extracted from the .alog file when Artisan's OFF button is pressed. These enable deterministic matching between sentinel logs and roast logs in the analysis pipeline.

For code-level details, SDK API notes, and tuning parameters, see CLAUDE.md.