The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from third-party merge request metadata and code diffs.\n
Ingestion points: The skill ingests untrusted third-party data using glab mr view, gh pr view, git diff, and git log commands as specified in SKILL.md (Steps 1 and 2).\n
Boundary markers: It implements strict boundary markers, including multi-layer fenced blocks and specific instructions for both the main agent and sub-agent to treat the content as data and ignore embedded instructions (Step 7).\n
Capability inventory: The skill can execute shell commands (git, glab, gh), write review files to a local scratch directory, and post comments to merge request platforms upon user approval (Steps 1, 9, and 10).\n
Sanitization: It includes logic to handle branch names securely by single-quoting inputs and converting them to commit SHAs to prevent shell command injection (Step 1).\n- [COMMAND_EXECUTION]: The skill generates shell commands using branch names, which are considered untrusted third-party input.\n
Evidence: SKILL.md (Step 1) demonstrates the construction of shell commands such as git fetch origin '<source-branch>'.\n
Context: The risk is mitigated by the skill's defensive programming, which includes wrapping inputs in single quotes and resolving them to static SHAs before use in most commands.

review-mr