The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data from repository source files and user requests to drive its orchestration logic, creating a surface for indirect prompt injection. This risk is mitigated by its architectural design, which requires a 'Sol root' model to perform final synthesis and verification, alongside a 'hostile' review phase where a separate agent attacks the implementation's correctness and security.
[COMMAND_EXECUTION]: The orchestrator performs shell-based operations for workspace discovery, code validation (linting/testing), and UI implementation. It specifically utilizes the Claude Code CLI (claude -p) with --permission-mode auto for visual tasks, allowing the agent to perform actions within the workspace without manual confirmation steps.
[SAFE]: The skill follows least-privilege principles by using read-only agents for exploration and uses well-known services (OpenAI, Anthropic). No malicious patterns such as credential harvesting, data exfiltration to unknown domains, or persistence mechanisms were detected.

sol-orchestrator