The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell commands via the Bash tool to interact with git, gh, mktemp, and the grok CLI. These are standard operations for a developer tool and are scoped to the project environment.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted code diffs and passes them to an external LLM engine (Grok). This represents an attack surface where malicious code comments or diff content could attempt to influence the agent's behavior.
Ingestion points: Reads DIFF and CHANGED_FILES from the local repository (SKILL.md Step 2).
Boundary markers: The skill explicitly instructs the engine to output JSON only and to focus strictly on the supplied diff, which acts as a structural boundary (SKILL.md Step 3).
Capability inventory: Uses Bash for git operations, file management, and CLI invocation. It does not possess direct eval or arbitrary network capabilities outside of the established CLIs.
Sanitization: The skill implements a robust 'Verification' step (SKILL.md Step 5) where every finding returned by the engine is manually checked against the source code by the agent before being reported as confirmed. It also explicitly forbids auto-applying any suggested fixes (SKILL.md Step 6).

grok-review