The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md direct the agent to execute the shell command git diff to gather changes for review. This is the primary intended behavior for a code review tool.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (source code and diffs) which is interpolated into the prompts for subagent reviewers in references/reviewer-prompt.md via the {DIFF_OR_FILES} placeholder.
Ingestion points: Files and diffs are collected in SKILL.md (Step 1) from the local repository or user input.
Boundary markers: The references/reviewer-prompt.md uses Markdown headers to separate the code content, but lacks robust delimiters (like XML tags) or specific warnings to ignore instructions embedded within the code under review.
Capability inventory: The primary agent can execute git commands and read files. The reviewer subagents are described as general-purpose and read-only.
Sanitization: The skill does not perform sanitization, escaping, or filtering of the code content before passing it to the subagents, allowing potential malicious instructions in the code to interact with the LLM's context.

interrogate