---

name: build-and-maintenance description: "Use this sub-skill for Torch-TensorRT source builds, package variants, CI/test lane selection, contributor workflows, and safe maintainer diagnostics." metadata: disco-role: operating disable-model-invocation: true license: BSD 3-Clause

Torch-TensorRT Build and Maintenance

Use this sub-skill when the user is working on the repository itself: building from source, selecting wheel variants, diagnosing package/test setup, or choosing a safe maintainer test lane.

Route by maintenance task

User goal	Read/run
Build a wheel or install from source	references/build-and-test.md and references/package-variants.md
Understand package flags such as PYTHON_ONLY, NO_TORCHSCRIPT, USE_TRT_RTX, CU_VERSION, or JetPack/Windows ARM64 selectors	references/package-variants.md
Choose a safe local test lane	references/test-selection.md
Inspect CI suites or list available maintenance helpers	scripts/list_ci_suites.py --help
Probe source-build prerequisites without compiling everything	scripts/source_build_probe.py --help
Diagnose a source-build or test failure	references/troubleshooting.md

Maintenance workflow

1. Confirm whether the user is trying to build, test, package, or inspect the repository.
2. Decide whether the desired variant is standard, Python-only, no-TorchScript, TensorRT-RTX, JetPack, or Windows ARM64.
3. Check prerequisites before building: Python, PyTorch/CUDA/TensorRT family, Bazel or other build tools, and any target-specific SDKs.
4. Start from the smallest relevant lane or probe before recommending a broad test suite.
5. If a build is required, make the command match the selected package variant and platform.
6. If the task is only source inspection or CI understanding, use the bundled scripts and references rather than a full build.

Guardrails

• Do not run long builds, full test matrices, or release automation unless the user explicitly wants them.
• Do not hide package-variant choices behind a generic pip install . when flags change runtime features.
• Do not tell users to edit or run files from another checkout; the skill must stand on its bundled references/scripts.
• Do not treat a successful import as proof that build/test lanes are healthy.