The Agent Skills Directory

[COMMAND_EXECUTION]: The skill launches external AI command-line tools, specifically codex, claude, opencode, and gemini, to process queries. It also executes python3 for processing configuration and handling API fallbacks.
[DATA_EXFILTRATION]: To perform its tasks, the skill reads local project files and accesses environment variables containing API keys (e.g., GEMINI_API_KEY). This data is sent to well-known AI service providers (Google, OpenAI, Anthropic, etc.) for analysis. The skill implements guardrails such as --sandbox read-only to ensure these external tools cannot modify the local environment.
[PROMPT_INJECTION]: The skill processes untrusted user input and source code, interpolating it into instructions for external models. It mitigates potential indirect prompt injection risks by using structured boundary markers and escaping techniques, such as wrapping code in CDATA sections and sanitizing closing tags.
Ingestion points: User prompts and standard input processed by consensus-query.sh and code-review.sh.
Boundary markers: Utilizes explicit text delimiters like --- Input --- and XML tags with CDATA blocks to isolate data from instructions.
Capability inventory: Includes read-only filesystem access, execution of local scripts, and network communication through authorized AI service CLIs.
Sanitization: Employs XML escaping and custom logic to escape CDATA termination sequences, preventing user-provided content from prematurely ending the data block.

agents-consilium