The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/soup.sh helper script executes environment diagnostics, including Python version checks and hardware backend availability tests (CUDA/MPS) using subprocess calls to Python and the soup CLI.
[EXTERNAL_DOWNLOADS]: The skill documentation guides the user to install the soup-cli package and its training dependencies (torch, transformers, peft, trl) from the official Python Package Index (PyPI).
[PROMPT_INJECTION]: The skill processes untrusted external data through soup ingest and soup data commands. While necessary for its fine-tuning purpose, the lack of boundary markers and sanitization during ingestion from attacker-controlled sources (like datasets or third-party logs) creates an indirect prompt injection surface. The skill maintains significant capabilities, including shell execution and file modification, which could be targeted by such injections. Evidence chain: 1. Ingestion points: Dataset and log processing via soup data and soup ingest in references/commands.md. 2. Boundary markers: None identified. 3. Capabilities: Shell access and file writing. 4. Sanitization: Not specified.

soup