The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage git worktrees, run project builds (npm ci, npm run build), and execute tests based on the PR content.
[INDIRECT_PROMPT_INJECTION]: The skill has a high attack surface for indirect prompt injection because it specifically processes untrusted content from GitHub Pull Requests.
Ingestion points: Reads PR diffs via gh pr diff, fetches PR branches into local worktrees, and reads CI logs using gh run view --log-failed.
Boundary markers: None observed in the prompt synthesis step. The skill instructs the LLM to "Read the review... automated test logs... and give a final recommendation."
Capability inventory: The skill can execute arbitrary project scripts (npm run), use the gemini CLI for further inference, and write/read files within the repository.
Sanitization: No sanitization is performed on the PR diff or log content before it is processed by the LLM in async-review.sh.
[DYNAMIC_EXECUTION]: The skill uses gemini headlessly (-p) to execute dynamic instructions generated during the review process, including a step that "physically exercises the newly changed code in the terminal."
[REMOTE_CODE_EXECUTION]: While it doesn't download and pipe scripts from the web, it fetches and executes code directly from a Pull Request (via git fetch and npm run), which is a form of remote code execution if the PR author is untrusted.

async-pr-review