The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides numerous complex shell scripts to manage validation sessions, discover models, and track performance. These scripts utilize several system utilities including jq for JSON processing, bc for mathematical calculations, xxd for hex encoding, and date for timestamps.
[EXTERNAL_DOWNLOADS]: The skill relies on 'claudish' (available at https://github.com/MadAppGang/claudish), an external third-party CLI tool, to proxy requests to various AI models such as Grok, DeepSeek, and Gemini. It suggests using npx claudish or assumes the tool is pre-installed in the environment.
[DATA_EXFILTRATION]: The core functionality involves extracting local project data (e.g., via git diff) and transmitting it to external AI providers (OpenRouter, Google, OpenAI, X-ai) through the proxy tool. While this is the intended purpose for multi-model code review, it involves sending sensitive source code to external network endpoints.
[INDIRECT_PROMPT_INJECTION]: The validation pattern involves a consolidation phase where an agent reads and summarizes output files generated by external models (e.g., grok-review.md).
Ingestion points: The consolidation agent ingests text files residing in $SESSION_DIR/ that were authored by external, untrusted AI models.
Boundary markers: The skill describes using the --- delimiter to separate tool calls, but does not specify delimiters or 'ignore' instructions for the content of the external reviews being consolidated.
Capability inventory: The orchestration scripts possess the capability to execute shell commands (Bash) and perform subsequent AI tasks (Task).
Sanitization: The instructions do not mention sanitizing or escaping the content received from external models before processing it in the consolidation step.

multi-model-validation