The Agent Skills Directory

[SAFE]: No malicious patterns, data exfiltration, or obfuscation were identified. The skill follows its documented purpose and integrates with standard development tools.
[COMMAND_EXECUTION]: The skill uses git and forge CLIs (gh, glab, bb) to resolve review scopes and fetch PR metadata. These are standard operations for code review automation and are performed safely.
[COMMAND_EXECUTION]: The evals/run_evals.py script uses subprocess.run to execute the claude CLI for testing the skill's reviewer template against synthetic diffs. This execution is confined to the local testing environment.
[REMOTE_CODE_EXECUTION]: The skill provides an optional integration with an external_review MCP tool for a second opinion from external models. This functionality uses the platform's standard MCP server protocol.
[PROMPT_INJECTION]: To mitigate risks from processing untrusted code diffs and PR descriptions, the skill implements a 'freshness boundary' for subagents, a diff preflight classification stage, and explicit instructions for subagents to ignore commit history and focus on the diff content.

temper