The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a variety of system tools to detect hardware and driver status, including vendor-specific management interfaces like nvidia-smi, mthreads-gmi, npu-smi, and rocm-smi, as well as standard utilities like lspci.
[REMOTE_CODE_EXECUTION]: To establish the development environment, the skill executes remote code including the official uv installation script and the miniconda installer via shell commands.
[PRIVILEGE_ESCALATION]: The initialization process includes tasks that require root or sudo access, such as installing hardware drivers, managing kernel modules, and installing system-level packages.
[PERSISTENCE]: The skill modifies system and user shell configurations (e.g., ~/.bashrc, ~/.zshrc, ~/.condarc, and uv.toml) to set environment variables, path updates, and package repository mirrors.
[EXTERNAL_DOWNLOADS]: Downloads configuration files, installers, and library wheels from various external repositories and mirrors, including the Tsinghua University TUNA mirror, official PyTorch indices, and vendor-specific repositories for Moore Threads and Huawei.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external technical documentation and platform release matrices to generate deployment plans, which could lead to the ingestion of untrusted instructions from external web sources.
Ingestion points: Technical documentation, issue trackers, and compatibility matrices from vendor websites and GitHub repositories (Phase 3).
Boundary markers: Not explicitly present in the instructions for the document research phase.
Capability inventory: System-wide command execution, privilege escalation to root, file system modification, and network access.
Sanitization: The skill does not define specific sanitization routines, relying on the agent to interpret external technical data to formulate and execute shell commands.

accel-init