The Agent Skills Directory

[COMMAND_EXECUTION]: The Python script scripts/store_media.py executes ffmpeg and ffprobe via subprocess.run to perform media processing tasks. These calls use argument lists rather than shell strings, which prevents shell-based command injection.
[DYNAMIC_EXECUTION]: The Node.js utility scripts/drive_chrome.mjs allows the agent to execute JavaScript in a browser instance via an /eval endpoint. This is a functional requirement for the skill to interact with various game engines. Access is restricted to the local loopback interface (127.0.0.1) and protected by a unique, randomly generated security token to prevent unauthorized access from other local processes or web pages.
[EXTERNAL_DOWNLOADS]: The skill documentation recommends installing pillow via pip and playwright-core via npm. These are reputable, well-known libraries from standard public registries.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from web-based games (DOM content and canvas frames). While this presents a theoretical surface for indirect prompt injection if a game contains instructions intended to be interpreted by the AI's vision or OCR systems, the risk is mitigated by the skill's workflow, which centers on generating media for human review.
[SAFE]: The local server implementations in scripts/capture_server.mjs and scripts/drive_chrome.mjs include several defensive measures: they validate file names using regular expressions to block path traversal, enforce origin checks for CORS, and explicitly restrict navigation to http and https protocols to prevent the reading of local system files.

harmony-store-media