The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses gh (GitHub CLI) and git commands to fetch PR data, view issues, and manage worktrees. These are standard operations for a code review tool and are used within the scope of its primary purpose.
[REMOTE_CODE_EXECUTION]: The skill involves dispatching parallel agents to review files. These agents are instructed to read files and provide analysis based on a strict rubric. No arbitrary remote script execution (like curl | bash) was detected.
[DATA_EXFILTRATION]: While the skill interacts with the GitHub API to post reviews, this is the intended functionality of the skill. It does not access sensitive local credentials like ~/.ssh or ~/.aws/credentials and instead uses the authenticated gh session.
[PRIVILEGE_ESCALATION]: No evidence of privilege escalation (e.g., sudo, chmod 777) was found. It uses standard temporary directories for worktrees.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code and issue descriptions. It addresses this by providing clear boundary markers in agent prompts and emphasizing that agents should only return findings based on the provided rubric. The severity of this surface is low as the LLM is used for analysis, not for generating executable actions from that data.

review-pr