The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Python subprocess module in scripts/llm_council.py to invoke various planning and judging CLI agents (Codex, Claude, Gemini, etc.). This execution is part of its core intended functionality and uses shlex.split for argument parsing.
[DATA_EXFILTRATION]: The orchestrator starts a local web server (using ThreadingHTTPServer on 127.0.0.1) to provide a visual interface. It mitigates risk by generating a cryptographically secure token (secrets.token_urlsafe) required for all non-static requests and by implementing a strict Content Security Policy (CSP) including X-Frame-Options: DENY and X-Content-Type-Options: nosniff.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-defined tasks and codebase information, which could potentially contain malicious instructions targeted at the planning agents. The skill implements defense-in-depth by:
Including explicit instructions in its prompt templates (references/prompts.md) for agents to treat all task and plan content as untrusted.
Anonymizing outputs using the anonymize_text function in scripts/llm_council.py, which redacts provider names and potential API keys (sk-...).
Validating the structural integrity of Markdown outputs before processing them further.

llm-council