The Agent Skills Directory

[COMMAND_EXECUTION]: Employs standard developer tools including git and gh (GitHub CLI) for repository research. All operations are scoped to the NVIDIA/Megatron-LM repository, which is consistent with the vendor's own verified infrastructure.
[SAFE]: Incorporates a strict human-in-the-loop requirement. The instructions explicitly direct the agent to show the drafted response to the maintainer for review and forbid automatic posting to GitHub.
[PROMPT_INJECTION]: Identified as having a surface for indirect prompt injection because it processes untrusted data from external contributors. Ingestion points: Fetches issue title, body, and comments via the gh issue view command in SKILL.md. Boundary markers: No explicit delimiters are used to wrap the untrusted issue content during processing. Capability inventory: The agent has access to repository interaction tools including git, gh, and grep. Sanitization: There is no explicit sanitization of the fetched issue text. The security risk is assessed as safe because the agent only generates a draft for human approval rather than executing commands based on the external input.

respond-to-issue