The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes pull request review comments to drive automated code fixes, creating a surface where malicious comments could influence the agent's actions.
Ingestion points: The get_new_comments function in scripts/gl-snapshot.py retrieves review comments on the pull request that are not replies and were not authored by the agent itself.
Boundary markers: The agent's instructions in references/triage-process.md for evaluating comments do not include delimiters or specific instructions to ignore malicious directives embedded within the comment text.
Capability inventory: The agent has extensive capabilities including local file modification, git operations (commit, push), and repository interaction via the gh CLI.
Sanitization: The skill does not implement filtering or sanitization to distinguish between legitimate bot reviews and potentially malicious comments from other users.
[REMOTE_CODE_EXECUTION]: The feedback instructions in SKILL.md suggest using npx --yes hivenet@latest, which downloads and executes an external package at runtime without version pinning.
[COMMAND_EXECUTION]: The scripts/gl-snapshot.py script executes multiple commands using the gh CLI tool via subprocess.run to fetch PR data, CI logs, and interact with the GitHub API. While the script passes arguments as a list, it operates on data derived from external repository metadata.

greenlight-pr