The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses git_diff and rejudge CLI commands to perform its core function. These are standard development tools. The use of temporary files in /tmp/ with unique identifiers is a common and safe practice for passing data between CLI tools.
[DATA_EXFILTRATION]: While the skill involves sending code diffs to an external multi-model panel (via the rejudge tool), this is the primary, documented purpose of the skill. The instructions explicitly state the review is read-only and targeted at the repository's code, which is expected behavior for a code review agent.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided strings for <PROBLEM>, <APPROACH>, and <DECISIONS>. However, it includes structural boundaries (EOF markers and specific tagging like [AGENT]/[USER]) that help mitigate accidental instruction following from within the data. As this is a vulnerability surface inherent to LLM-based code review, the risk is managed.
[PRIVILEGE_ESCALATION]: The skill instructions explicitly forbid high-privilege flags (e.g., --unsafe or --full), maintaining a restricted execution environment.

rejudge-diff