The Agent Skills Directory

[SAFE]: The skill implements a 'Zero-Action Policy' and 'Read-Only Policy' (SKILL.md), which strictly prohibits any GitHub mutations like commenting, closing issues, or merging PRs. This prevents the agent from making unauthorized changes to the target repository.
[SAFE]: Factual claims must be backed by GitHub permalinks using commit SHAs (rules/evidence.md). This requirement ensures that all reported findings are grounded in static, verifiable code evidence, preventing hallucinations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from GitHub Issue and PR titles and bodies (SKILL.md). While no explicit boundary markers or sanitization logic are defined, the risk is classified as safe because the skill's capability inventory is restricted to diagnostic reading (gh, git) and writing reports to /tmp, precluding any dangerous actions triggered by external content.
[SAFE]: All network operations are limited to authenticated GitHub API and Git operations. No attempts were found to access sensitive local files (e.g., .ssh, .aws) or exfiltrate data to third-party domains.

shen-shi