The Agent Skills Directory

[COMMAND_EXECUTION]: Local shell scripts (scripts/e2e.sh, scripts/validate-terminal-example.sh, scripts/validate.sh) are used for testing and validating the skill's logic. These scripts execute project-local code and standard utilities like jq and python3.
[PROMPT_INJECTION]: The skill processes untrusted data (agent-generated synthesis and reviewer verdicts) but implements strong boundary markers and deterministic sanitization to prevent agent self-grading. Ingestion points: Reads CONTRACT.md, test.sh, and review JSONs. Boundary markers: Uses per-round nonces and mandatory cross-family review floors to prevent spoofing. Capability inventory: Logic is implemented in deterministic jq and shell scripts (close-gate.sh). Sanitization: Nonce validation and artifact separation ensure data integrity.
[SAFE]: No malicious patterns such as obfuscation, exfiltration, or persistence were found. The skill is focused on improving the security posture of agent orchestrations by providing a 'fail-closed' mechanism.

gc-membrane