The Agent Skills Directory

[SAFE]: The skill implements a comprehensive write-ahead authority protocol (R0-R4 action classes) to gate all effect-bearing operations, ensuring that outward effects (R3) always require human approval and governance changes (R4) are strictly controlled.
[SAFE]: Candidate experiments are isolated in external Git worktrees with strictly defined mutation surfaces (permitted globs), preventing candidates from modifying critical system components or accessing sensitive data like sealed promotion cases.
[SAFE]: All command execution is performed using argv arrays via an injected process runner, explicitly avoiding shell-based execution to mitigate command injection risks.
[SAFE]: The skill demonstrates high security awareness by treating all imported traces and external telemetry as untrusted data that cannot autonomously influence system state or bypass authority checks.
[SAFE]: A robust path resolution and validation system (anchored to the Git root) prevents directory traversal attacks and unauthorized escapes from the project directory.
[SAFE]: Sensitive data is protected through an unconditional ignore policy for the .gauntlet/secrets/ directory and a redaction system that scrubs telemetry and logs before they are persisted or exported.

gauntlet-loop