The Agent Skills Directory

[SAFE]: The skill is professionally developed with significant attention to security, including detailed instructions for risk mitigation, secret hygiene, and error handling during the code review lifecycle.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from git repositories and sub-agent reports. It mitigates this risk by explicitly defining all review evidence as 'inert' data that cannot grant permissions or override the agent's core contract. It employs a closed structural schema (delegated_result_record) for external input and maintains strict write boundaries for all capabilities, including git and file system operations.
[DATA_EXFILTRATION]: The skill includes a robust secret hygiene protocol in references/review-workflow.md that mandates the identification and redaction of sensitive credentials like API keys, JWTs, and private keys using specific placeholders. This prevents sensitive data from being exposed in chat, saved files, or shared with other tools.
[COMMAND_EXECUTION]: Git and file system operations are limited to the verified review unit through 'Effect And Write Boundaries'. The skill prevents unauthorized command execution by restricting writes and prohibiting the agent from inventing host commands without proof.
[PRIVILEGE_ESCALATION]: Workflow-controlled data is stored in private directories with restricted permissions (mode 0700), adhering to the principle of least privilege and ensuring intermediate data is not exposed to other processes.

vibe-review