Kirk Content Pipeline

What to consider before installing/using this skill: - Sandbox bypass: SKILL.md tells the agent to create a symlink from /Users/Shared/ksvc/pdfs into the project so subagents can read PDFs. This is an explicit workaround of permission controls — avoid blindly creating broad symlinks (they can expose many files). Instead copy only the PDFs you trust into a project-local directory, or run the skill in an isolated environment. - Untrusted pickle risk: build_extraction_cache.py loads RLM 'state.pkl' files with pickle.load (with fallback code). Unpickling arbitrary files can execute code. Only run this script on RLM state files you created or have verified; do not run it on pickles from untrusted sources. If possible, inspect or sandbox the script before use. - Data leakage risk: The cache JSON includes absolute pdf_path and extracted content; if you upload or share caches, you may leak filesystem paths and content. Inspect generated caches and scrub/remove sensitive metadata before exporting. - Network/web verification: The pipeline requires web cross-validation and multiple model checks (Gemini, other models). Confirm what endpoints are used and whether any API keys or credentials will be required or sent. The skill does not declare required credentials, so be cautious of implicit network calls triggered by higher-level subagents. - Mitigations: run the pipeline in a VM or disposable container, copy only the PDFs you want to process into a dedicated skill directory (avoid symlinking entire shared folders), review/execute the Python script under controlled conditions, and scan the generated JSON for sensitive content before sharing. If you need the capability but want lower risk, request a version of the skill that accepts explicit file paths (no auto-discovery), avoids unpickling or uses a safe deserialization format, and does not recommend symlink sandbox workarounds. Confidence: medium — the skill's intent and many behaviors are internally coherent for the stated purpose, but several explicit sandbox-workaround instructions and use of pickle/unverified local state are noteworthy security concerns that warrant caution.