The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external pull requests and has the capability to modify repository state.
Ingestion points: The agent ingests data from external sources via gh pr list (metadata), gh pr view (PR body/comments), and gh pr diff / git show (source code changes).
Boundary markers: Absent. The instructions do not explicitly direct the agent to ignore instructions embedded within the PR content or treat the diffs strictly as data.
Capability inventory: The skill possesses significant capabilities, including git push --force-with-lease to contributor branches, gh pr merge to the target repository, and gh pr close to terminate PRs.
Sanitization: No sanitization or validation is performed on the PR descriptions or commit messages before they are processed by the agent to classify PRs into triage tiers.
[COMMAND_EXECUTION]: The skill uses shell commands (git, gh) to perform actions based on variables retrieved from the network, such as PR numbers, author logins, and branch names. While these are used for intended triage workflows, the dynamic construction of commands (e.g., git push <author> HEAD:<branch>) relies on the integrity of the data provided by the GitHub API and the agent's interpretation of that data.

triage-prs