The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external GitHub pull requests, including filenames and code patches, which could contain instructions designed to manipulate the agent's behavior during classification.
Ingestion points: Fetches PR file data via gh api --paginate repos/<owner/repo>/pulls/<n>/files as described in SKILL.md.
Boundary markers: None explicitly defined to separate untrusted data from processing logic.
Capability inventory: Has the ability to write to the GitHub API using the markFileAsViewed mutation via gh api graphql.
Sanitization: The instructions advise using double quotes for shell variables ("<path>") to handle paths containing $, but there is no comprehensive sanitization for the content of the patches being classified.
[COMMAND_EXECUTION]: The skill uses gh (GitHub CLI) and jq for data processing and API interaction.
Evidence: Uses gh pr view, gh api --paginate, and gh api graphql to manage pull request state. This is legitimate and documented functionality for a triage skill.

pr-triage