The Agent Skills Directory

[SAFE]: The skill incorporates a robust security layer via scripts/ingest_guard.py, which proactively scans any ingested content for PII (including CPF/CNPJ checksum validation) and sensitive secrets such as API keys or private keys.
[SAFE]: Data integrity and agent safety are maintained through scripts/safe_facts.py, which acts as a structural barrier. It constrains the depth, size, and types of data objects passed to the agent's context, effectively mitigating potential indirect prompt injection attacks that could be hidden in source documents.
[SAFE]: The implementation is designed to be entirely offline and follows a 'least privilege' model. All scripts use only standard Python libraries, and the skill includes an automated audit tool (scripts/lint_bundle.py) to verify that the bundle remains isolated and free of dangerous commands or absolute paths.
[SAFE]: Defensive prompt engineering is integrated into the core SKILL.md, establishing strict behavioral gates and instructing the agent to treat all retrieved content as factual evidence rather than authoritative instructions. This is further reinforced by a suite of behavioral evaluations in the evals/ directory.

consultor-simples-nacional-by-rdd