The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various shell commands for project discovery (e.g., find, ls) and runs third-party security analysis tools including bandit, pip-audit, osv-scanner, semgrep, and the cargo ecosystem. It also invokes project build scripts (e.g., make test) to validate remediations.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the local repository, such as source code and configuration files. This represents an attack surface where maliciously crafted files could attempt to influence the agent's analysis or remediation logic.
Ingestion points: Local repository files matched by globs in SKILL.md (e.g., .py, .rs, .toml, .yaml).
Boundary markers: Uses structured reports and requires mandatory NIST/CWE citations; findings are validated via an external script citation_verifier.py.
Capability inventory: Git commit execution, shell command execution via project build tools, and execution of local security scanners.
Sanitization: Proposals undergo blast-radius assessment and require explicit user approval (except when opt-in auto-apply is used for non-critical findings).
[DYNAMIC_EXECUTION]: The skill generates and applies code patches (git commits) to the local repository and verifies them by running executable test suites identified during the discovery phase.

harden