The Agent Skills Directory

[SAFE]: The skill implements a robust 'two-channel' boundary security model (LSC-7). It explicitly identifies and isolates untrusted data (tool outputs, web content) from trusted instructions (user/assistant text). Untrusted content is sanitized and wrapped in <data> fences in the final context digest, significantly mitigating risks of indirect prompt injection.
[SAFE]: No malicious command execution, network exfiltration, or credential theft patterns were found. All scripts utilize the Python standard library only, avoiding supply chain risks from third-party dependencies.
[SAFE]: The automated detection of prompt injection patterns in references/capture.md is a false positive; the text describes an adversarial scenario used for validation of the skill's defensive mechanisms.
[SAFE]: The installation process for lifecycle hooks into .claude/settings.json is transparent, idempotent, and restricted to the scope defined by the user (project or global). Use of chmod +x is appropriate for the installation of the provided utility scripts.

context-hygiene-kit