The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/test-review-harness.py script and its bash wrapper use subprocess.run to manage temporary Git repositories for testing. These executions are limited to standard Git operations and the invocation of the skill's own review scripts using safe, list-based argument passing.
[REMOTE_CODE_EXECUTION]: The skill coordinates remote code analysis through external LLM engines. However, the documentation and implementation specify rigorous isolation measures, such as using --safe-mode for Claude and --sandbox for Cursor, which disable project-local hooks, plugins, and filesystem access during the review process.
[CREDENTIALS_UNSAFE]: The repository includes a fixture file (tests/fixtures/typescript-sensitive-literals.ts) containing hardcoded API keys and tokens. These are explicitly marked as "FAKE" and are intended for testing the skill's integrated secret-scanning capabilities. The skill's operational instructions mandate a TruffleHog scan before any code is sent to an external engine.
[PROMPT_INJECTION]: The skill uses defensive prompt engineering by requiring the agent to distill a concise task contract from authoritative sources before invoking the reviewer. It also includes explicit instructions to block rather than infer missing requirements if sources are unavailable, reducing the risk of indirect injection from malicious ticket or PR descriptions.

autoreview