The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes development tools and project gates as its primary function. It uses a custom wrapper (bounded_run.py) that implements several security and reliability features, including automated redaction of sensitive command-line options (tokens, secrets, API keys) and strict process group cleanup to prevent background execution.
[EXTERNAL_DOWNLOADS]: The skill downloads and executes development tools (such as react-doctor, fallow, and dart-decimate) using npx from the official npm registry. These downloads target specific or latest versions of well-known development utilities.
[PRIVILEGE_ESCALATION]: Several scripts (e.g., worktree.py, external_cli_restore_regression_check.py) manage file permissions using chmod. These operations are limited to setting appropriate execution bits for scripts or protecting internal configuration receipts (mode 0600), following the principle of least privilege for local configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill ingests repository metadata, such as git history, project manifests (hard-eng.gates.json), and documentation files (PRODUCT.md, DESIGN.md), to coordinate checks. It mitigates injection risks by validating manifest command syntax, forbidding shell interpreters (bash, sh) as primary executables in gates, and strictly limiting the scope of runtime execution to verified project roots.

deterministic-checks