The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes standard system utilities such as lscpu, free, nvidia-smi, ps, and taskset to discover hardware resources and manage the llama-server process. These operations are restricted to local system discovery and process management as part of the tuning workflow.
[EXTERNAL_DOWNLOADS]: The skill fetches model configuration data and README files from Hugging Face. These network operations target a well-known service and are used solely to derive optimal server parameters.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted prose from Hugging Face model cards to provide tuning hints.
Ingestion points: scripts/fetch_model_card.py (Hugging Face README prose).
Boundary markers: None for the raw text, but recommendations are clearly labeled as candidates in the report.
Capability inventory: Server process management (kill, nohup) and local network benchmarking.
Sanitization: Employs keyword-based regex filtering to extract only performance-relevant lines.
Mitigation: A mandatory human-in-the-loop approval gate is implemented in Step 5, ensuring no configuration changes are applied without user review.

llamacpp-tuner