The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses standard Git commands (git status, git diff, git diff --cached, git diff <base>...HEAD) to identify the scope of code changes for review. These are routine operations for a development-focused agent skill.
[REMOTE_CODE_EXECUTION]: The skill mentions running a local script node .claude/scripts/ai-signal-scan.cjs --json and node .claude/hooks/lib/review-receipt.cjs. These are internal project scripts located within the .claude directory, not remote code execution. No external scripts are downloaded and executed.
[DATA_EXPOSURE]: The skill reads project source code, documentation, and configuration files to perform its review. It also mentions reading from tmp/reports/ and tmp/workflow-runs/. These are internal project paths used for task tracking and reporting, which is standard for agentic workflows.
[PROMPT_INJECTION]: The skill instructions include clear goal definitions and rules for review. While it uses phrases like 'IMPORTANT MUST ATTENTION' to emphasize protocol steps, these are internal formatting markers to guide the agent's behavior during a review task, not attempts to bypass safety filters or override system instructions.
[DYNAMIC_EXECUTION]: The skill references a $skill-name invocation pattern for loading other skills (e.g., $performance-review, $why-review). This represents a modular architecture where the agent loads and follows additional local Markdown instructions for specific review domains. No untrusted data is being executed as code.

changes-review