The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from GitHub PRs, including titles, bodies, and comments. This information is used to classify PRs and suggest actions to the user. A malicious actor could potentially embed instructions in PR comments to influence the agent's behavior during triage or when performing actions.
Ingestion points: PR data (titles, bodies, reviews, and comments) is fetched in scripts/pr-detail.sh and scripts/fetch-prs.sh.
Boundary markers: Absent. The skill does not use specific boundary markers or instruction-ignoring wrappers when presenting PR content to the agent.
Capability inventory: The skill can merge PRs (scripts/merge-pr.sh), post comments, and request reviews (scripts/request-review.sh).
Sanitization: Absent. PR content is formatted via jq but is not sanitized for natural language instructions before being reported.
[COMMAND_EXECUTION]: The skill utilizes local shell scripts to interface with the GitHub CLI (gh). These scripts manage the workflow of fetching, enriching, and acting upon PR data, including sensitive operations like merging code and requesting reviews.

github-pr-triage