菜单
Use when the user says "use pixi to install <some nvidia tool>" (or similar) and wants NVIDIA/CUDA/GPU packages installed via Pixi (no sudo/apt), e.g., CUDA toolkit pieces, cuDNN/NCCL, PyTorch CUDA builds, RAPIDS.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Prepare and run a user-confirmed interactive execution or test session for a source file, executable, script, demo, or directory. This skill must be invoked manually: use it only when the user explicitly invokes step-interactive-execution by name or clearly asks for this exact manual workflow. Read the target first without pausing, then switch to a strict step-by-step loop where the agent prints the exact CLI command to run next, waits for confirmation such as "continue", executes only that approved command, and reports the result before proposing the next step.
Manual invocation only. Drive a crashy, hanging, or half-broken system forward along a real production user path using real data. Two subskills: `prepare` to analyze the target and set up `<htt-home>/` with infrastructure dirs (logs, runs, issues); optionally creates `<htt-home>/autotest/` with automatic scripts and interactive guides only when the developer explicitly requests test-case generation. `run` drives testing — with or without autotest artifacts — patching forward through blockers. Run subskill operates in-place by default (stash + test on current branch) or in a disposable snapshot worktree when explicitly requested. Supports automatic and interactive driving. Default when ambiguous: both subskills, in-place, automatic. Not for CI-oriented unit, smoke, or mock-based integration tests.
Programmatically invoke Codex via the `codex` CLI for headless, non-interactive runs using `codex exec`, machine-readable JSONL via `--json`, final-message capture via `-o/--output-last-message`, and multi-turn continuation via `codex exec resume`. Use when you need scriptable Codex automation for repo reviews, issue investigation, implementation planning, or structured one-shot outputs.
Create and resume Codex CLI sessions with a persistent session-name-to-thread_id mapping plus last-used model stored in a workspace-scoped JSON file under system temp. Invoke only when the user explicitly names `$codex-cli-invoke-persist` or otherwise clearly requests this exact skill. Use it for deterministic, session-persistent Codex automation across turns and processes.
Manual invocation only. OpenSpec-specific hack-through-testing workflow targeting production-level end-to-end paths using real data and real user workflows — not CI smoke/unit/integration tests. Three subskills: `propose` to create an OpenSpec change with HTT-ready test cases (automatic scripts and interactive guides) by invoking `openspec-propose` or `openspec-ff-change`, `revise` to update an existing OpenSpec change so its artifacts support hack-through-testing-driven implementation and testing, and `run` to exercise an implemented OpenSpec change through the full hack-through-testing loop (in-place by default, or in a disposable snapshot worktree when requested). Use when the user explicitly asks for `openspec-ext-hack-through-test`, points to `openspec/changes/...` while asking to propose, revise, run, exercise, or prepare work under hack-through-testing principles, or wants OpenSpec work shaped for fast blocker discovery through patch-forward testing.
Review an OpenSpec change (or a single OpenSpec change artifact file) for completeness, coherence, and alignment with existing system design; capture actionable feedback plus open questions; write a review report under the change directory (review/review-YYYYMMDD-HHMMSS.md).
| name | pixi-install-nvidia |
| description | Use when the user says "use pixi to install <some nvidia tool>" (or similar) and wants NVIDIA/CUDA/GPU packages installed via Pixi (no sudo/apt), e.g., CUDA toolkit pieces, cuDNN/NCCL, PyTorch CUDA builds, RAPIDS. |
Use this skill when the user asks to install NVIDIA tooling via Pixi, especially in the form:
Examples:
This skill provides workflows for setting up No-Sudo, User-Space GPU environments. By installing CUDA toolkits and libraries via Pixi, you avoid modifying the host system (apt install) and ensure perfect reproducibility across different machines.
/usr/local/cuda.pixi.lock guarantees the exact same driver-compatible libraries everywhere.Ensure the correct channels are present in pyproject.toml. The order determines priority.
pixi project channel add nvidia (and pytorch if needed).nvidia MUST be prioritized over conda-forge for all NVIDIA tools and libraries (CUDA, cuDNN, NCCL, etc.), unless the user explicitly requests otherwise.Use specific versions to ensure compatibility between CUDA and the framework. Always prefer the nvidia channel for these packages.
pixi add pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
Instead of the massive cuda-toolkit, consider adding only what's needed:
pixi add cuda-compiler cuda-libraries-dev
After installation, verify GPU visibility:
pixi run python -c "import torch; print(torch.cuda.is_available())"
nvidia-smi on the host to ensure the installed pytorch-cuda version is supported by the host driver.libcuda.so or libcudart.so are not found, ensure the environment is activated (pixi shell).nvcc and cmake with the Pixi-managed toolkit.