The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/yomitan.mjs executes system binaries to handle dictionary archives.
The script uses execFileSync to call unzip for listing and reading files within ZIP archives (zipList, zipRead, cmdUnpack functions).
The script uses execFileSync to call zip when packaging dictionaries into new archives (cmdPack function).
Arguments are passed as an array to the child process, which mitigates standard shell injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (Yomitan dictionary JSON files and ZIP archives) that may contain content which is later presented to the AI agent.
Ingestion points: Untrusted data is ingested in scripts/yomitan.mjs via the loadDictionary and zipRead functions, which parse JSON content from dictionaries.
Boundary markers: The SKILL.md includes instructions for the agent to "summarize counts and samples" rather than outputting raw bank contents, providing a degree of separation between the raw data and the agent's response.
Capability inventory: The script has the capability to read/write files and execute command-line tools (zip, unzip).
Sanitization: While the script performs structural validation (e.g., checking entry field counts and types in checkEntry), it does not sanitize the semantic text content of definitions or terms for potential injection attacks.

yomitan-dictionary