AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

get-available-resources

Detects host inventory and effective CPU, memory, disk, scheduler, container, and accelerator limits when a user asks for resource-aware planning or before a clearly resource-sensitive local workload. Produces a redacted JSON snapshot and conservative planning helpers without stress tests or assuming visible host hardware is usable.

Original source
K-Dense-AI
License
MIT
Pinned revision
5f2bc0341354fd5590aef76a533ada1e1e5b4e62

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read k-dense-ai--get-available-resources at revision 5f2bc0341354fd5590aef76a533ada1e1e5b4e62, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 4

references/sources.md

Official Sources and Verification

Reviewed 2026-10-01. This ledger covers all bundled commands, parsers, planning assumptions, and reference guidance. Living documentation is not a claim that every host, driver, or scheduler configuration was exercised.

Python and psutil

  • psutil 7.2.2 reference and release source verify cpu_count(logical=...), Process.cpu_affinity(), virtual_memory(), swap_memory(), disk semantics, and psutil.Error failures. The optional tested package remains 7.2.2, matching the current PyPI release metadata. Development/latest documentation can describe a newer version; it does not change this pin.
  • Python os reference verifies CPU affinity, statvfs, sysconf, and CPU-count overrides introduced in Python 3.13. The calling-thread/process count is not a cgroup quota detector.
  • CPython 3.14.8 shutil source confirms POSIX disk_usage().free uses f_bavail, not f_bfree. Both are read separately by the detector when available.
  • Python 3.14.8 concurrent.futures documentation source confirms process-aware defaults, the Windows ProcessPoolExecutor ceiling of 61, and the changed process start method.
  • Python 3.14.8 multiprocessing documentation source confirms the forkserver/spawn defaults and os.process_cpu_count() pool default. These are executor-specific constraints; the generic planner does not launch an executor or promise that every worker count fits every API.

Linux, containers, and disk

  • Kernel procfs documentation verifies MemAvailable, CPU topology/affinity fields, and the mountinfo root/mountpoint fields used to locate a cgroup2 mount.
  • Kernel cgroup v2 documentation verifies cpu.max, cpuset.cpus.effective, memory.current, memory.high, memory.max, hierarchical limits, and namespace-relative membership. Hidden ancestors cannot be measured from a restricted mount. memory.high remains a pressure boundary; remaining capacity is not a reservation.
  • Docker resource constraints verifies default unconstrained configuration, quota/period, cpusets, memory, and why host-visible swap is not necessarily container-usable swap.
  • OCI Runtime Specification 1.3.0 Linux configuration verifies CPU/memory/device controls independently of container markers.

NVIDIA

  • nvidia-smi manual verifies selective --query-gpu and --format=csv,noheader,nounits queries, unsupported N/A values, unstable indices, and output compatibility limits. The detector requests index/name/memory/driver/compute capability and retries without compute capability on a command error. No NVIDIA hardware was available for a driver-specific --help-query-gpu check in this review.
  • Container Toolkit configuration verifies NVIDIA_VISIBLE_DEVICES and driver-capability semantics. These are container-start configuration inputs; their presence in a running process does not prove kernel device isolation.
  • CUDA visibility and CUDA compatibility distinguish application enumeration from management visibility and runtime compatibility.
  • MIG device enumeration explains why physical management GPU counts are not framework-device counts.

AMD

  • AMD SMI CLI reference currently identifies 27.0.0. It verifies static --asic --vram --json, unavailable fields, and APU/partition differences. The narrower query avoids collecting every static field, including unrelated board and topology data.
  • AMD SMI command source verifies nested JSON vram.size: {value, unit}. The matching C++ implementation divides byte totals by 1024 squared despite labeling the result MB. The parser preserves that binary conversion. This pinned source check does not establish every installed AMD SMI release's output shape.
  • ROCm SMI CLI reference and official CLI source verify the retained --showproductname --showmeminfo vram --json fallback and flat byte-valued memory fields.
  • ROCm environment variables verifies visibility-variable scopes and recommends ROCR_VISIBLE_DEVICES on Linux and HIP_VISIBLE_DEVICES on Windows. The detector does not claim to cover every isolation variable or runtime. The old separate GPU-isolation URL was inaccessible; the current environment reference and container documentation support the bounded guidance retained here.

Apple, Windows, and Slurm

  • Apple system capabilities was read through its official documentation JSON representation; it verifies hw.logicalcpu/hw.physicalcpu and performance-level semantics. Archived sysctl manual cross-checks memory reporting.
  • Apple DTS system_profiler discussion explains integrated/SoC memory interpretation. The fixed sysctl -n and system_profiler SPDisplaysDataType -json probes were executed locally on macOS during this review. Intel/AMD display-vendor handling is synthetic, and no Metal framework workload was executed.
  • Windows processor groups, GetLogicalProcessorInformation, and GetLogicalProcessorInformationEx distinguish topology, group-limited observations, and Windows 11/Server 2022 defaults spanning groups. Windows behavior was covered by mocks only.
  • Slurm sbatch, srun, and CPU management verify every allowlisted allocation variable, compressed task lists, the current SLURM_NODEID index, memory units/scope, and site-dependent enforcement. Scheduler tests are synthetic; no Slurm allocation or accounting call was executed. Job/step and heterogeneous-component scope must still be checked in the workload's actual environment.

Verification boundary

The network-free suite exercises Linux cgroup hierarchy/delegated mounts, Slurm heterogeneous node counts, macOS/Windows probes, legacy/current AMD parsers, NVIDIA CSV, malformed snapshots, private output, and infeasible memory plans. All four CLI helpers were exercised on redacted local snapshots. No stress tests, large allocations, write-capacity probes, device changes, framework accelerator execution, or scheduler submissions were performed.