The Agent Skills Directory

[PROMPT_INJECTION]: The SKILL.md file includes a 'Guidelines for Claude' section that provides specific behavioral instructions and best practices for the AI model when interacting with the framework.
[COMMAND_EXECUTION]: The GitHubCLIToolset described in references/toolsets.md wraps the official GitHub CLI (gh), allowing the agent to execute system commands for repository management.
[REMOTE_CODE_EXECUTION]: The framework includes DSPy::CodeAct, a module that enables agents to synthesize and execute dynamic Ruby code to perform calculations or complex logic as part of a 'Think-Code-Observe' loop.
[EXTERNAL_DOWNLOADS]: The documentation references several external Ruby gems used as provider adapters and extensions, including dspy-openai, dspy-anthropic, dspy-gemini, and ruby_llm.
[DATA_EXFILTRATION]: The observability suite in references/observability.md is designed to export telemetry, traces, and evaluation scores to Langfuse (cloud.langfuse.com) for application monitoring.
[INDIRECT_PROMPT_INJECTION]: The framework's architecture for building agents that ingest external data (e.g., via GitHub CLI or web tools) presents an indirect injection surface.
Ingestion points: Data retrieved via GitHubCLIToolset (issue bodies, PR descriptions) or other tool observations entered into the agent's context.
Boundary markers: The framework utilizes structured data formats like TOON (which uses fences) and JSON Schema to delimit LLM inputs and outputs.
Capability inventory: High-privilege capabilities include CodeAct (Ruby execution) and GitHubCLIToolset (shell execution via CLI).
Sanitization: The framework mitigates risks by enforcing strict Sorbet type validation (T::Struct, T::Enum) and JSON schema compliance for all tool and signature interactions.

dspy-ruby