The Agent Skills Directory

[SAFE]: The skill provides standard C++ code templates for using the TensorRT Runtime API, including IRuntime, ICudaEngine, and the modern enqueueV3 API. These implementations align with official NVIDIA development practices.
[SAFE]: All referenced tools (trtexec), headers (NvInferRuntime.h), and libraries (libnvinfer.so, libcudart.so) are official resources provided by the vendor (NVIDIA) and are consistent with the skill's purpose for C++ inference deployment.
[SAFE]: The provided code snippets use standard C++ file I/O and CUDA memory management practices (cudaMalloc, cudaMemcpyAsync) to handle local model engine files as requested by the user.
[SAFE]: The skill defines a surface for ingesting binary model data from external .plan or .engine files, which is required for its primary purpose.
Ingestion points: SKILL.md (via std::ifstream in Step 2 and IStreamReaderV2 in Step 3).
Boundary markers: Not present, as engine files are binary blobs.
Capability inventory: File system read access and GPU execution via the TensorRT library.
Sanitization: The implementation relies on the official deserializeCudaEngine API for validating the integrity and compatibility of the engine file during loading.

trt-cpp-runtime-quickstart