The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingestion process for external Wiki documentation represents an indirect prompt injection surface where malicious instructions in a document could influence agent behavior.
Ingestion points: Documentation is ingested into the references/docs/ directory from an external staging source, as managed by the scripts/sync_wiki_docs.py script.
Boundary markers: SKILL.md provides clear operating instructions and output contracts to keep the agent focused on factual retrieval from the indexed files.
Capability inventory: The skill includes local execution of Python scripts for searching and syncing files, with file-write capabilities restricted to the skill's own directory.
Sanitization: The sync_wiki_docs.py script includes a validate_url function that filters for sensitive keywords (like api_key or token) in source URLs and uses safe_relative checks to prevent path traversal.
[COMMAND_EXECUTION]: The skill utilizes local Python scripts (search_docs.py, sync_wiki_docs.py, install_skill.py) to handle its core logic. These scripts are restricted to the Python standard library and do not invoke external shell commands or remote resources.
[SAFE]: The provided scripts demonstrate high security awareness, utilizing atomic write operations and strict path validation to ensure the integrity of the local environment and prevent common file system vulnerabilities.

baize