The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to upload UI screenshots to catbox.moe, a public third-party file-sharing service. While this is intended to simplify PR previews, it transmits local visual data to an external server not controlled by the user or a trusted enterprise vendor.
[COMMAND_EXECUTION]: The skill executes multiple CLI tools including git, gh, trufflehog, and caffeinate to manage the development environment and PR lifecycle. It also runs a bundled script (scripts/autoreview) to manage the review loop.
[REMOTE_CODE_EXECUTION]: The automated review logic invokes external LLM engines (Codex, opencode, CodeRabbit) to analyze code. The skill specifically instructs these engines to run in isolated environments with plugins and shell access disabled to mitigate risks of unauthorized execution.
[PROMPT_INJECTION]: As a tool that processes untrusted code diffs and PR comments, the skill is subject to indirect prompt injection. Malicious instructions embedded in the repository content could attempt to influence the agent's code-fixing or review behavior.
[SAFE]: The repository contains a fixture file (tests/fixtures/typescript-sensitive-literals.ts) with numerous hardcoded credentials (e.g., API keys, tokens). These are clearly dummy values used exclusively for unit testing the secret-scanning functionality.
[SAFE]: Dynamic execution patterns found in the scripts/ directory, such as importlib and runpy calls, are restricted to the skill's own test harness and unit tests for internal validation.

send-it