The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill describes a pattern for fetching and executing a custom redaction scanner script from a private GitHub repository during a CI process.
Evidence: The CI configuration in Step 7 uses gh api to fetch script content and execute it: gh api repos/<org>/<private-repo>/contents/tools/check-redaction.sh --jq .content | base64 -d > check-redaction.sh followed by ./check-redaction.sh ..
[DYNAMIC_EXECUTION]: Decodes script content from the GitHub API using base64 -d to create an executable file at runtime within a CI environment.
Evidence: Step 7 CI workflow example demonstrates runtime decoding and execution of the check-redaction.sh script.
[COMMAND_EXECUTION]: Instructions guide the use of various shell commands for environment setup, repository management, and data synchronization.
Evidence: Use of sudo apt-get install, chmod +x, and git push --force for administrative and repository operations (Steps 3, 6, and 7).
[INDIRECT_PROMPT_INJECTION]: The skill establishes a pipeline for processing and publishing files from a private research repository which may contain untrusted data.
Ingestion points: Local files from the PRIVATE_ROOT directory referenced in tools/public-allowlist.txt (Step 5).
Boundary markers: Encourages fact categorization (Step 1) and the use of the check-redaction.sh script to block publishes containing sensitive patterns.
Capability inventory: File system operations (cp, mkdir, rm), shell script execution, and Git synchronization commands.
Sanitization: Implements regex-based scanning using ripgrep or grep to detect and block sensitive content patterns (Step 3).

redact-for-public-disclosure