The Agent Skills Directory

[DYNAMIC_EXECUTION]: The SKILL.md file implements a bootstrap mechanism that takes a base64-encoded string, decodes it, decompresses it using the DecompressionStream API, and executes the resulting JavaScript string using (0, eval)(code). This occurs during the initialization of the Mermaid Diagram Engine.
[OBFUSCATION]: The functional logic for the diagram engine is stored in vendor/beautiful-mermaid.min.js.gz.b64 as an opaque binary blob. Storing executable code in a compressed and encoded format prevents standard security tools from inspecting the library for malicious behavior or vulnerabilities before it is executed in the user's browser context.
[COMMAND_EXECUTION]: The skill requires the agent to execute several shell-based tools including python3 for multiple local utility scripts (e.g., fdk/tools/html_font.py), node for executing a visual audit script, and npx serve to host the generated content on a local port (8765).
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection. It ingests untrusted markdown and wiki files to generate complex HTML structures. The evidence chain is as follows:
Ingestion points: Reads source content from markdown files and wiki entries (SKILL.md).
Boundary markers: No explicit delimiters or sanitization steps are documented for the input text.
Capability inventory: The skill possesses file-writing capabilities to llmwiki/html/ and shell execution capabilities via python3 and node.
Sanitization: The logic includes a regex to strip @import from generated SVG styles, but lacks comprehensive sanitization for instructions that might be embedded in the user-provided markdown content.

docs-site-macos