The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as CI logs, issue descriptions, and pull request diffs, creating a potential attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the context when the agent retrieves CI logs (references/ci-evidence.md), issue content (references/issues.md), or pull request diffs and comments (references/pull-requests.md).
Boundary markers: The skill does not define explicit LLM prompt delimiters but mitigates injection at the execution layer by mandating the use of file-based inputs (e.g., --body-file) and structured inputs to prevent untrusted text from being interpolated into shell syntax.
Capability inventory: The skill possesses the capability to execute gh and glab CLI commands for network communication and repository mutations, and can perform local file writes to store logs.
Sanitization: The instructions explicitly warn that logs and annotations are untrusted data and must not be followed as commands. Furthermore, it instructs the agent to redact secrets from excerpts and summarize error bodies narrowly to avoid exposing sensitive content.
[EXTERNAL_DOWNLOADS]: The skill references and utilizes official resources and tools from well-known service providers.
Documentation: It links to official technical documentation at docs.github.com, docs.gitlab.com, and cli.github.com for API contracts and CLI usage.
Tools: The workflows rely on the official GitHub CLI (gh) and GitLab CLI (glab) for performing repository operations and fetching data.

manage-hosted-repositories