The Agent Skills Directory

[SAFE]: The skill is built upon a foundational 'user-always-decides' principle (Invariant 1). This invariant ensures that every routing decision, ensemble creation request, and workflow plan is presented to the user for explicit consent, preventing autonomous or unintended actions.\n- [DYNAMIC_EXECUTION]: The skill is designed to author and execute local Python and Bash scripts to handle non-LLM complexities such as file I/O, parsing, and tool execution. While this provides significant capability, it is documented as a core feature for building 'ensemble' systems and is strictly governed by the mandatory user approval process.\n- [EXTERNAL_DOWNLOADS]: The skill includes a bootstrap script (pull-models.sh) that installs standard Python libraries, including sentence-transformers, onnxruntime, numpy, and torch, which are well-known packages used for embedding-based verification and data processing. It also pulls models via the official Ollama interface.\n- [INDIRECT_PROMPT_INJECTION]: The skill possesses the capability to read and chunk project documents to extract code patterns. Although this exposes a surface for indirect prompt injection from processed data, the risk is mitigated by the agent's internal reasoning protocols and the requirement for human oversight on the resulting analysis.

conductor