The Agent Skills Directory

[SAFE]: The skill is explicitly defined as read-only. It contains repeated, strict instructions to not modify files, artifacts, Git state, or remote systems, and to return all results within the conversation.
[SAFE]: It implements a robust defense against indirect prompt injection by instructing the agent to treat all natural language read from external sources (code, PR comments, CI logs) as evidence/data rather than authority, specifically warning against 'instruction-like text' within that data.
[SAFE]: The skill uses an independent review protocol ('Discovery Seats') to ensure that findings are verified across different simulated perspectives, reducing the risk of hallucinated or malicious findings.
[SAFE]: All external references and policies are locally contained within the skill's reference directory, and the skill includes specific guidance for identifying and reporting security vulnerabilities (like SSRF and XSS) in the code being reviewed.
[SAFE]: No obfuscation, malicious persistence mechanisms, or unauthorized remote code execution patterns were detected across the skill's files.

tk-review