System 1 Arcade 0.2.0

The built-in agent no longer needs Python.

• Laya runs inside the app through laya-go, a pure-Go
  port of the Laya package: no Python, PyTorch or virtual environment to set up. The first start
  downloads the model (about 800 MB) into the standard Hugging Face cache, with progress in the
  side panel; a model the Python package already downloaded is used as is. The model stays loaded
  while the agent plays and is freed when you stop it.
• Same answers as before. Over 343 recorded game requests the built-in agent gives the Python
  server's answers: every choice the same and 99.86% of numbers identical to 4 decimals (the rest
  differ by 0.0001). The answer cache (SYSTEM1_LAYA_CACHE) works as before.
• Runs on the CPU (Apple Accelerate on macOS). agents/laya_server.py still works as a custom
  agent and runs Laya on the GPU.
• Test connection for the built-in agent asks the loaded model, or says whether the model still
  has to be downloaded.
• The app keeps App Nap away while an agent plays. With the window hidden or the display
  asleep, macOS throttled the app, and the built-in agent's new sentences took up to 2.3 s instead
  of about 0.1 s: Frogger seed 1 ended at 3,250–4,900 points in the app against 20,420 headless.
• The agent asks at most once per game tick in realtime; with answers from the cache it used to
  ask thousands of times a second and keep a CPU core busy.
• cmd/headless -agent laya plays with the built-in agent, and -agent <url> with a custom one,
  for benchmarks without Python or a window.
• The app logs the model download's progress (every 10%) and, at the end of each game, how
  many decisions the agent made and how long they took (median, p99, slowest).
• SYSTEM1_AUTOSTOP_AFTER=<duration>, with SYSTEM1_AUTOSTART, stops the agent that long
  after launch, for unattended memory checks. Test-only and off unless set.
• scripts/run-agent-vm.sh --builtin plays with the built-in agent in a VM with no Python on the
  app's PATH, watching that no Python process ever starts; --fresh-download makes the app fetch
  the model itself, and --stop-after measures memory after the agent stops. Verified: all three
  games played in a VM with no Python, the app started no process, a fresh download matched the
  host's copy, and stopping the agent took the app from 1.7 GB to 186 MB. In a VM the model runs
  2.5–6 times slower than on the host, which costs Frogger points (and whole games when the host
  is busy); Tetris and Space Invaders played as on the host.
• SYSTEM1_SOUND=off mutes a session, for automated and autostarted runs, without changing
  the saved sound preference. scripts/run-agent-vm.sh sets it.
• scripts/build.sh --agent is now --python-tools (-PythonTools on Windows), and is only
  needed for the Python tools in agents/. SYSTEM1_PYTHON is gone. Building needs Go 1.27.

---

Install: download System1Arcade-0.2.0.dmg, open it and drag System 1 Arcade to
Applications. It's a universal app (Apple silicon and Intel), signed with Developer ID and
notarized by Apple. Requires macOS 11 or later.

Built-in agent: runs Laya inside the app, with no Python. On first start it downloads the
Laya model (about 800 MB) into the Hugging Face cache while the game waits.