The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill monitors and processes external, untrusted content from GitHub Pull Request comments and review summaries. This content is fed into a 'triage agent' to determine whether suggested changes should be applied.
Ingestion points: scripts/review-loop.sh fetches issue comments, inline review comments, and review summaries via the GitHub API.
Boundary markers: The triage process defined in references/review-loop.md utilizes an independent sub-agent with a skeptical prompt and 'Context:' headers to separate diff and comment data.
Capability inventory: The skill possesses significant capabilities, including executing shell commands (git commit, git push, gh pr merge), writing to the file system, and performing network operations via the GitHub CLI.
Sanitization: The primary mitigation is the use of an independent 'review-triage agent' instructed to be skeptical and to reject noise or harmful suggestions. However, the skill is designed to merge the PR automatically once checks pass, which increases the potential impact of a successful injection.
[COMMAND_EXECUTION]: The skill is designed to execute project-specific quality and security checks as defined in references/quality-validation.md.
Evidence: The skill instructs the agent to run commands such as npm test, pytest, black, ruff, and npm run build depending on the project type. While standard for CI/CD automation, this involves executing code defined within the repository being processed.

create-pr