---
type: guidelines
---
# Vision guidelines

## 1. Camera-agnostic
- Any V4L2 capture device works. Never hard-code a node such as `/dev/video0`:
  numbering follows enumeration order. Take a `/dev/v4l/by-id/` path, or find
  the camera at runtime (`--camera auto`).  
- Use what the camera offers. MJPEG passes through untouched; anything else is
  encoded here. A requested size or rate the camera lacks falls back to its own
  default rather than failing.  
- UVC control names differ between cameras and kernels. Look them up on the
  device; reply that a control is missing rather than guessing.

## 2. One process owns the camera
- V4L2 capture is single-reader. This service opens the device; recording and
  inference take frames from it, never from the device.  
- UVC controls reset when the camera re-enumerates. Set them on open, not once.

## 3. The tablet link is the constraint
- Every socket to a display is paced (`SO_MAX_PACING_RATE`, project-yukart
  decision 28). Unpaced bursts stall the USB link for seconds.  
- Latest frame wins. A slow client skips frames; it never gets a queue of old ones.  
- Say when there is no camera. A client must be able to tell "unavailable" from
  "slow".

## 4. Python, on purpose
- Python 3.10 (the Orin's), standard library only; GStreamer and v4l-utils come
  from the host. The exception to Rust for Orin services is decision 29: camera
  pipelines and inference live in the Python ecosystem.  
- No pip installs on the car Orin until there is a reason; then a venv under
  `~/infra/svc-vision/`, never the system Python.

## 5. To be added