The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a self-improvement loop where it processes data from external web search results to discover container images and instructions the agent to update its own reference files with the results.
Ingestion points: Web search results for vendor-specific PyTorch images (Step 4.3 in SKILL.md and references/image-sources.md).
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands within search results are provided before processing the text.
Capability inventory: The skill possesses Edit and Write tool permissions to modify its own reference files, and can perform docker pull and docker run operations.
Sanitization: The risk is mitigated by a mandatory validation step using scripts/validate_pytorch.py, which ensures the discovered image correctly imports PyTorch and accesses the GPU before the reference files are updated.
[COMMAND_EXECUTION]: The skill executes various system commands through the subprocess module in its helper scripts to detect hardware state.
Evidence: scripts/detect_gpu.py uses subprocess.run(..., shell=True) to execute vendor-specific tools like nvidia-smi, rocm-smi, npu-smi, mx-smi, and ixsmi.
Evidence: scripts/find_data_disk.py executes df -h to find suitable mount points.
Context: All shell commands are hardcoded within the scripts and do not incorporate unvalidated user input, making the usage legitimate for the skill's purpose.
[EXTERNAL_DOWNLOADS]: The skill queries multiple external container registries to find appropriate images.
Registry Queries: The skill performs curl requests to NVIDIA NGC, BAAI Harbor, Metax Registry, Iluvatar Hub, and Docker Hub APIs to retrieve image tags.
Trusted Sources: Many of these sources (NVIDIA, Docker Hub) are well-known or trusted services. The use of BAAI Harbor (harbor.baai.ac.cn) for flagrelease projects is consistent with the vendor context (flagos-ai).

gpu-container-setup-flagos