The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes comprehensive defensive measures against indirect prompt injection. It explicitly instructs the agent that every telemetry field (group titles, error messages, stack frames, log lines, etc.) must be analyzed as inert data and never as instructions. It specifically forbids obeying imperative text, following links, or running commands found within evidence.
[DATA_EXFILTRATION]: The skill implements a sophisticated sanitization engine (scripts/sanitize-telemetry.py) that uses regex and key-matching to redact bearer tokens, GitHub PATs, private keys, AWS/GCP credentials, IP addresses, emails, and payment cards (with Luhn validation). All raw telemetry is stored in gitignored paths or operating system temporary directories to prevent accidental leakage.
[COMMAND_EXECUTION]: Static analysis identified subprocess.run usage in scripts/render-report.py. Technical review confirms this is a safe, internal implementation detail used to invoke the sibling sanitize-telemetry.py script on a controlled temporary file path to perform mandatory post-rendering validation.
[SAFE]: The skill enforces a strict read-only boundary for production and repository access. It requires cryptographic receipts (scripts/validate-receipt.py) to prove that every claimed telemetry query actually ran as intended, preventing data fabrication or misrepresentation.

woostack-respond