The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes shell commands via the Bash or Shell tool to invoke local Python scripts and model runners. These commands include several parameters such as model names, timeouts, and output file paths. There is a potential risk of command injection if identifiers like session_id are not properly sanitized before being interpolated into these command strings.
[EXTERNAL_DOWNLOADS]: The skill relies on the presence of external CLI binaries including Anthropic's claude, Google's gemini, and the qwen transport tool. It performs smoke tests to verify these tools are present and authenticated in the host environment. These are documented dependencies required for the skill to communicate with third-party AI services.
[DATA_EXFILTRATION]: By design, the skill reads repository file contents and transmits them to various third-party AI providers (Anthropic, Google, OpenAI, etc.) to generate responses for the council rounds. This data sharing is the intended primary function of the skill and involves well-known technology vendors.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) as it processes arbitrary repository files and incorporates their content into prompts for the AI models. Maliciously crafted instructions within these files could potentially influence the consensus outcome. The skill employs stance rotation and multi-model synthesis to mitigate the impact of individual model bias or manipulation.

models-consensus