The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various performance-related tools and scripts, including sglang, vllm, tensorrt-llm, tokenspeed, nsight compute (ncu), and torch-profiler. These are standard tools for the stated purpose of LLM serving optimization.
[DATA_EXPOSURE]: The skill includes an explicit instruction to 'Never save Hugging Face tokens or other secrets in artifacts,' which is a positive security practice.
[REMOTE_CODE_EXECUTION]: The skill uses a Humanize RLCR loop triggered via setup-rlcr-loop.sh. This is a local script execution within the environment's resolved runtime roots (e.g., ~/.claude/skills/humanize) and is used for autonomous agentic workflows.
[PRIVILEGE_ESCALATION]: The skill contains explicit constraints against privilege abuse, specifically stating: 'Never kill another user's process or delete shared model caches.'
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect injection as it reads PR history and external documentation (Category 8c). However, it uses these only for decision support and includes mandatory verification steps (re-running benchmarks and profiles) that mitigate the impact of potentially poisoned instructions in external data. This risk is assessed as LOW and does not escalate the verdict.

sglang-sota-humanize-loop