The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses global JavaScript evaluation to process user-supplied coordinate expressions.
Evidence: In template/kit.js, the window.renderSheet function uses (0, eval)(v) to parse coordinate strings provided via the --crop-at command-line argument. This allows for arbitrary JavaScript execution within the headless browser context if an agent is tricked into passing malicious strings to the renderer.
[COMMAND_EXECUTION]: The skill relies on spawning external processes to perform its core tasks, which are sensitive operations.
template/render.mjs uses node:child_process.spawn to execute ffmpeg for video encoding and launches a headless Chrome instance via puppeteer-core.
scripts/new_project.sh executes shell commands including npm install, rsync, cp, and sed for project scaffolding and dependency management.
[EXTERNAL_DOWNLOADS]: The skill initiates network requests to fetch external software and assets.
scripts/new_project.sh performs an npm install, downloading packages from the public NPM registry.
template/studio.html loads external fonts from Google Fonts (fonts.googleapis.com) to render subtitles and UI elements.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for data-driven attacks.
Ingestion points: The rendering engine processes command-line arguments in template/render.mjs and reads configuration from template/src/config.js and other project files.
Capability inventory: The skill has extensive file system access (reading and writing frames, project files), network access (via Puppeteer), and subprocess execution capabilities (ffmpeg).
The combination of reading external project files and performing dynamic execution based on CLI input provides a vector for influencing the agent's actions or the rendering process if the input data is compromised.

anime-cel