The Agent Skills Directory

[SAFE]: Implements a mandatory 'Corpus Approval Barrier' that requires explicit user approval of an immutable package snapshot's SHA-256 digest before any evaluation data can be materialized or evaluated.
[SAFE]: Uses isolated, capability-scoped variant roots (candidate vs baseline) for evaluation tasks, ensuring workers only have access to authorized workspace paths and no access to credentials, network resources, or the evaluator's private definitions.
[SAFE]: Employs robust filesystem security measures, including path containment checks via a dedicated contained utility and strict file identity validation using device/inode/mtime metadata and SHA-256 hashing to prevent TOCTOU and directory traversal attacks.
[SAFE]: Runtime evaluator scripts (validate.mjs, prepare.mjs, etc.) are offline-only, deterministic processors that are verified by an internal test suite (test-command-contract.sh) to ensure they do not import network modules or provider SDKs.
[SAFE]: Generated evaluation reports are hardened with a restrictive Content Security Policy (CSP) that disables all scripts, images, and network connections (default-src 'none'), protecting the user from XSS or data exfiltration when viewing evidence results.
[SAFE]: Utilizes safe command execution patterns by using node:child_process spawn and execFile without shell interpolation, and restricts executed commands to local Git and collection-native scripts.

woostack-eval