The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill performs audits by ingesting repository content via git, grep, and semantic search tools, which are then analyzed by subagents with file-modification capabilities. This setup presents an indirect prompt injection surface where malicious instructions embedded in the code being audited (e.g., in comments or literals) could attempt to manipulate the agents' logic or automated fixes.\n
Ingestion points: The skill reads codebase files through git diff, mcp__lumen__semantic_search, grep, glob, and find as described in SKILL.md and various reference documents.\n
Boundary markers: The protocol uses lane-specific reference documents and prompt templates to constrain subagent focus, but lacks explicit protection against adversarial instructions in data.\n
Capability inventory: The skill utilizes the Edit tool for auto-fixes in lanes 1-4, the Agent tool for dispatching sub-processes, and the git CLI for repository management.\n
Sanitization: No explicit sanitization or filtering of audited code content is performed before the data is processed by the AI models.\n- [COMMAND_EXECUTION]: The skill executes shell commands using the git utility to manage the auditing and fixing workflow.\n
git diff is used to identify changed files and establish the audit scope in quick and standard modes.\n
git commit is used to automatically persist changes made by the subagents to the repository history.

grove-audit