The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands defined by the user for verification purposes. Evidence: scripts/run-verifier.sh uses bash -c to run the command string provided in the --verify argument or a goal.md file.
[PROMPT_INJECTION]: The audit subagent processes untrusted content from the repository, creating a surface for indirect prompt injection.
Ingestion points: scripts/spawn-audit-subagent.sh incorporates the goal text, verifier log tails, and git diff statistics into the prompt template.
Boundary markers: templates/audit-prompt.md uses Markdown headers and code blocks to delineate system instructions from external data.
Capability inventory: The subagent is restricted to a read-only set of tools (Read, Glob, Grep, and specific git/filesystem commands) via the --allowedTools flag in scripts/spawn-audit-subagent.sh.
Sanitization: Content is interpolated into the template using awk without a formal escaping mechanism for Markdown or control characters.
[EXTERNAL_DOWNLOADS]: The skill's documentation recommends the manual installation of external tools and libraries from public registries.
Evidence: references/troubleshooting.md suggests installing tiktoken via pip and @openai/codex via npm to support context probing and delegation features.

auto-loop