The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from GitHub pull requests, including titles, descriptions, and comments, and provides this content to an AI agent with write-access permissions to the repository's pull requests.
Ingestion points: PR metadata (title, body, labels) and linked issue content collected in scripts/main.py and scripts/worker.py.
Boundary markers: The prompt uses standard string blocks with headers and --- delimiters to isolate external content from system instructions.
Capability inventory: The AI agent is provided with terminal and file_editor tools, and a GitHub token with write access to pull requests and issues.
Sanitization: The skill includes strong negative constraints in the prompt (e.g., "Do not modify files, push commits, or merge the pull request") and implements a "Scope Gate" to validate the PR's purpose before detailed technical review.
[EXTERNAL_DOWNLOADS]: The automation downloads repository content as compressed tarballs from GitHub's official API to facilitate analysis.
Evidence: urllib.request is used in scripts/main.py to fetch current-head archives from api.github.com.
Safety Measures: The script explicitly validates the archive layout, rejects absolute or parent-directory (..) paths to prevent traversal, and skips device files and symlinks to prevent platform escape.
[COMMAND_EXECUTION]: The automation involves executing shell commands in the agent's workspace to handle repository setup and testing.
Evidence: The prompt instructs the agent to use gh (GitHub CLI) for metadata fetching and git for branch management within a profile-scoped environment.
Evidence: The setup workflow in SKILL.md provides shell snippets for users to package and upload the automation to the platform.

github-pr-reviewer