The Agent Skills Directory

[SAFE]: The skill instructions demonstrate a strong focus on safety and verification. It explicitly requires matching repository identity and branch tips before deletion, and uses plumbing commands (git update-ref) and guarded pushes (--force-with-lease) to ensure atomicity. The presence of a comprehensive evaluation suite (evals/) containing simulation traces and assertion tests indicates a rigorous development and verification process.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from GitHub PRs and Git repositories which could theoretically contain malicious instructions. However, this is categorized as safe because the skill mitigates these risks through structural verification logic. Ingestion points: Data enters the context via gh PR history lookups and git branch/tip inspections. Boundary markers: None explicitly used for natural language interpolation. Capability inventory: The skill can perform local ref deletions via git update-ref and remote branch deletions via git push. Sanitization: The skill employs strict structural sanitization by requiring exact OID matches, ancestry verification against durable branches, and cross-repository PR identity validation.

cleanup-branches