The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple external sources, creating a surface for indirect prompt injection. Ingestion points: Processes GitHub pull request comments via gh pr view, review threads via gh api, and ticket descriptions from external issue trackers such as GitHub issues, Jira, and Kanopi. Boundary markers: Instructions in sub-skills/references/code-review-briefs.md establish a shared dispatch contract for sub-agents with axis-specific prompts, and sub-skills/references/gh-review.md restricts shell command usage (forbidding chaining characters like &&, ;, or |). Capability inventory: Performs extensive repository operations using git and gh, and executes local repository scripts (test, lint, typecheck) as part of verification workflows in sub-skills/create-pr.md and sub-skills/ci-fix.md. Sanitization: Uses structured JSON for API communication and requires explicit user confirmation before executing install/start commands for UI/UX browser testing of external pull request code.
[COMMAND_EXECUTION]: The skill dynamically constructs and executes shell commands for git and GitHub management. Evidence: Automates branch creation, committing, and pushing in sub-skills/commit.md. Executes gh pr create and gh run view for CI management. Mitigation: The skill implements a safety rule in sub-skills/references/gh-review.md stating that commands must be single calls without chaining characters or subshells, reducing the risk of command injection.

git