The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions in components/osmo-cli/reference.md and components/osmo-cli/agents/workflow-expert.md that encourage the agent to execute shell commands autonomously after a high-level confirmation, specifically instructing the agent not to show raw commands to the user and to skip confirmation under certain conditions. This behavior reduces transparency and user oversight during infrastructure mutations.
[PROMPT_INJECTION]: Indirect Prompt Injection surface identified in components/osmo-cli/agents/logs-reader.md where untrusted data is processed.
Ingestion points: The subagent ingests data from workflow logs using the osmo workflow logs command.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the prompt instructions for the subagent.
Capability inventory: The main agent has high-privilege Shell and Read capabilities, allowing for modifications to Azure AKS and MicroK8s clusters and access to secrets in the .env file.
Sanitization: The instructions explicitly mandate returning error logs "verbatim," which provides a direct path for malicious workload output to be interpreted as instructions by the main agent.
[EXTERNAL_DOWNLOADS]: The skill performs several downloads and configuration fetches from well-known technology organizations and services.
HashiCorp: Installs Terraform from the official HashiCorp apt repository (apt.releases.hashicorp.com).
NVIDIA: Clones the OSMO repository from NVIDIA's official GitHub and adds helm charts from the NVIDIA Container Gallery (helm.ngc.nvidia.com).
Microsoft: Fetches modules from the official Physical AI toolchain repository on GitHub.
Python Packaging Authority: Installs the huggingface_hub package from PyPI.

physical-ai-infrastructure-setup-and-resilient-scaling