The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several mechanisms to execute commands and scripts at runtime.
The scripts/generate_icon.py script invokes the codex CLI tool via subprocess.run to perform image generation. The command is structured to avoid shell injection and uses a sandbox flag (--sandbox workspace-write) to limit its scope.
The SKILL.md file and reference guides frequently use uvx and bunx to execute inline Python and JavaScript snippets for data visualization and PDF rendering.
[EXTERNAL_DOWNLOADS]: The skill dynamically fetches and runs packages from established registries to maintain a minimal local footprint.
It uses uvx and uv run --with to download well-known Python libraries from PyPI, including matplotlib, seaborn, numpy, pdf2image, cairosvg, and lifelines.
It uses bunx to retrieve @react-pdf/renderer and react from NPM for figure composition.
These downloads are restricted to reputable, official software repositories.
[PROMPT_INJECTION]: The skill implements a Visual QA process that involves an indirect prompt injection surface.
The agent is instructed to generate a preview image of the figure and then use the vision tool to inspect it for alignment and readability issues.
This process could be targeted by indirect prompt injection if malicious data were used to generate the figure elements (e.g., text in a plot label designed to trick a vision model), although the risk is consistent with standard vision-capable agent workflows.

scientific-figures