The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The scripts/install_pandoc.sh script downloads the Pandoc binary from its official GitHub repository (jgm/pandoc). This is a well-known and widely trusted open-source technology service.
[COMMAND_EXECUTION]: The scripts/epub_to_markdown.py script invokes the pandoc binary using subprocess.run to convert XHTML content extracted from EPUBs into Markdown. The command uses a predefined set of safe arguments.
[COMMAND_EXECUTION]: The scripts/cache_clear.sh script uses rm -rf to delete local cache files in ~/.tsundoku/cache/markdown. It employs the -- delimiter to prevent argument injection and includes logic to verify paths before deletion.
[PROMPT_INJECTION]: The skill provides an indirect prompt injection surface because it processes untrusted EPUB data and outputs it to the agent's context. An attacker could embed malicious instructions within a book's text.
Ingestion points: EPUB files processed by scripts/epub_to_markdown.py.
Boundary markers: The output uses Markdown headers for chapters but does not include explicit delimiters to warn the agent about untrusted book content.
Capability inventory: Subprocess execution (pandoc, rm, curl) and file system access.
Sanitization: The converter strips potentially dangerous XHTML tags like <script> and <style>, though natural language instructions remain.

book-extract