The Agent Skills Directory

[SAFE]: The skill requires explicit scope confirmation from the user (Phase 1), preventing unintended or resource-heavy repository-wide scans.
[SAFE]: All file operations are restricted to local code reading and writing audit reports to the project's internal '.codestable' directory.
[SAFE]: The 'Guardian Rules' enforce rigorous standards, such as capping the number of findings to avoid information overload and requiring specific 'file:line' evidence for every claim.
[SAFE]: While the skill ingests code that could contain indirect prompt injections (Category 8 surface), its capabilities are restricted to generating reports and suggesting tasks for other tools, with no risk of network exfiltration or direct code execution. Ingestion is performed via local file reads; no specific boundary markers or sanitization are defined, but the limited capability set and requirement for human-in-the-loop review of findings mitigate risks.

cs-audit