The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to resolve the code to be reviewed, including the use of git diff --staged to identify currently staged changes.\n- [COMMAND_EXECUTION]: During the verification phase, the skill executes project-specific test commands. These commands are dynamically determined at runtime from project configuration files like CLAUDE.md or other local settings.\n- [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests and processes untrusted data from the project workspace (Ingestion points: source code files, CLAUDE.md). There are no explicit boundary markers (Boundary markers: absent) or input sanitization processes (Sanitization: absent) described in the workflow. However, the potential for malicious action is significantly mitigated by the requirement for user intervention and approval before any code modifications are committed (Capability inventory: file writing via apply_fixes, command execution via run_tests_if_available).

review-loop