菜单
Full GPU survey: VRAM allocation, process-level usage, compute utilization, thermals, power, and model inventory. Auto-run when: OOM errors appear, before loading new models, GPU >80% in session-context, or user asks about GPU usage. Invoke proactively without asking.
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
| name | gpu-audit |
| description | Full GPU survey: VRAM allocation, process-level usage, compute utilization, thermals, power, and model inventory. Auto-run when: OOM errors appear, before loading new models, GPU >80% in session-context, or user asks about GPU usage. Invoke proactively without asking. |
Comprehensive GPU breakdown.
nvidia-smi
nvidia-smi --query-gpu=name,driver_version,power.draw,power.limit,temperature.gpu,temperature.memory,fan.speed,clocks.gr,clocks.mem,utilization.gpu,utilization.memory,memory.used,memory.total,memory.free --format=csv,noheader
nvidia-smi --query-compute-apps=pid,name,used_gpu_memory --format=csv,noheader 2>/dev/null
curl -s http://localhost:11434/api/ps 2>/dev/null || echo "Ollama not running or no models loaded"
curl -s http://localhost:5000/v1/model 2>/dev/null || echo "TabbyAPI not running"
systemctl --user status hapax-daimonion studio-compositor visual-layer-aggregator 2>/dev/null | grep -E '(●|Active|Memory|CPU)'
cat /sys/class/drm/card*/device/pp_dpm_sclk 2>/dev/null; echo "---"; cat /proc/driver/nvidia/gpus/*/power 2>/dev/null | head -10
Present a ranked breakdown of VRAM consumers by process. Show compute vs memory utilization split. Compare current thermals/power to limits. Identify capacity for additional workloads (how many more models, what sizes). Flag anything thermally throttled or power-limited.
Background CI monitor for a PR. Auto-run when: PostToolUse detects a PR was just created (suggests it with PR number), user asks to watch CI, or after pushing to a branch with an open PR. Takes PR number as argument. Invoke proactively without asking.
Run the health monitor and report results. Auto-run when: session-context shows health is degraded or failed, after infrastructure changes (docker, systemd), when a service appears unreachable, or user asks about system health. Invoke proactively without asking.
Audit installed flatpak apps and find native alternatives. Auto-run when: storage-audit reveals large flatpak usage, operator mentions app performance issues related to sandboxing, or user asks about flatpaks. Invoke proactively without asking.
Audit branches across all repos and clean up. Auto-run when: session-context reports >3 non-main branches across repos, no-stale-branches hook blocks a branch create, after merging PRs, or user asks about branch cleanup. Invoke proactively without asking.
Scan for and resolve merge conflict markers. Auto-run when: conflict-marker-scan PostToolUse hook fires a warning, git merge/rebase/cherry-pick output shows CONFLICT (PostToolUse suggests it), or user asks to fix conflicts. Invoke proactively without asking.
Trace consent provenance. Auto-run when: work touches files containing personal data (profiles/, agents processing person names), after adding new data sources or storage, or user asks about consent or data protection. Invoke proactively without asking.