The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the rejudge command-line utility to perform multi-model code reviews. This is done locally to process files and diffs when the integrated tool is not available.
[DATA_EXFILTRATION]: The skill reads repository diffs and source code files to provide them to the rejudge tool for analysis. As the skill is intended for code review, this involves sending repository content to external models via the review tool, which is a standard part of its primary function.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it processes external data (source code changes via git_diff) and uses the resulting output of an LLM-based tool to propose and apply code fixes.
Ingestion points: Reads code differences from the repository using git_diff and the resulting review output from the rejudge tool.
Boundary markers: The instructions use structured prompt sections (e.g., <PROBLEM>, <APPROACH>) and specific output formatting rules to guide the agent.
Capability inventory: The skill has access to git_diff, the rejudge CLI, and the ability to write suggested fixes back to the filesystem.
Sanitization: There is no explicit sanitization or filtering of the diff content before it is passed to the review models.

rejudge-diff