The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches audit utilities and best-practice guidelines from trusted organizations and well-known registries. It downloads impeccable and react-doctor via npx from the public npm registry, and fetches security and database audit rubrics using the GitHub CLI from official repositories of OpenAI and Supabase.
[PROMPT_INJECTION]: The skill implements a secure pipeline for processing untrusted pull request data (code diffs and comments) with a well-defined evidence chain for mitigation.
Ingestion points: Untrusted PR comment bodies and diff content enter the agent context in scripts/load-prompt.sh and scripts/prefetch.sh.
Boundary markers: Present. The skill employs explicit XML-style tags and cryptographically random heredoc delimiters in scripts/load-prompt.sh to isolate untrusted data and prevent escape attempts.
Capability inventory: Orchestrator and sub-agents have access to shell execution, file system access, and GitHub API interactions, but risk is mitigated through isolated execution contexts.
Sanitization: Present. A sanitize_untrusted function in scripts/load-prompt.sh removes control characters and potential escape tags from user-supplied comments.

woostack-review