The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads and installs software components from external sources.
It clones source code for vllm-plugin-FL, FlagGems, and FlagCX from the vendor's GitHub organization (flagos-ai).
It references the official vllm-project/vllm repository for version 0.13.0.
It installs the flagtree package from a vendor-specific PyPI index at https://resource.flagos.net/repository/flagos-pypi-hosted/simple.
[COMMAND_EXECUTION]: The skill executes multiple shell commands for system discovery and software installation.
Hardware detection is performed using tools such as nvidia-smi, npu-smi info, mthreads-gmi, and ixsmi.
Installation steps include directory management (mkdir -p), repository cloning (git clone), and Python package installation (pip install).
It performs a broad filesystem search using find / -maxdepth 5 -type d -name "<user_provided_model_name>" to locate model directories.
[INDIRECT_PROMPT_INJECTION]: The skill defines a process that ingests untrusted user input into sensitive shell commands, creating a vulnerability surface.
Ingestion points: In SKILL.md, the agent is instructed to ask the user for a model name, which is then used as a parameter in a find command and a Python script.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to sanitize or escape the user's input before it is interpolated into the shell command.
Capability inventory: The skill utilizes shell execution for filesystem searching, environment configuration, and repository management.
Sanitization: No evidence of validation or sanitization logic is present to prevent a malicious user from providing a string that alters the intended shell command logic (e.g., command injection via backticks or subshells).

vllm-plugin-fl-setup-flagos