deepstream-dev — Claude Code skill for DeepStream video pipelines

What deepstream-dev does

The skill gives the agent a set of DeepStream reference documents. It has to read them from references/ before writing any code, because they hold the exact property names, the correct API calls and the usual failure points. Target stack: GStreamer 1.24.2, CUDA 13.1, TensorRT 10.14.1.48, NVIDIA driver 590+, Ubuntu 24.04 on x86_64 or ARM64/Jetson.

A pipeline runs Source → Stream Muxer → Inference → Tracker → OSD → Renderer. The tracker, secondary GIEs, analytics and nvmsgbroker stay out unless you ask for them, so the first result is a minimal working graph. Sources go through nvurisrcbin, which accepts RTSP, HTTP and file:// paths. Object and frame metadata come back as one-shot iterators from .frame_items and .object_items; counting them with len() is a dead end.

Some mistakes cost hours. Request pads use the sink_%u template, never a literal pad name. Every sink in a tee split or with dynamic sources needs async=0, otherwise the pipeline parks in PAUSED and nothing renders. An ONNX model with dynamic axes fails inside TensorRT with setDimensions: Error Code 3 unless the nvinfer config carries infer-dims=C;H;W. On WSL2 with Ubuntu 24 in Docker the skill drops display sinks entirely and writes an annotated MP4 instead, even when display was requested — a flag like --no-display does not switch it back.