The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to detect local provider availability (Ollama, Gemini CLI, Codex, Cursor) and to dispatch sub-agents. These operations are essential to the skill's multi-provider coordination purpose. The coordinator protocol uses temporary files and quoted heredocs to safely pass prompts to sub-agents, effectively preventing shell injection vulnerabilities.
[CREDENTIALS_UNSAFE]: The framework manages API keys for external providers like NVIDIA NIM. It adheres to security best practices by resolving these from environment variables at runtime and passing them to tools via process substitution (-H @<(...)), ensuring secrets never appear in the process argument list visible to other users.
[EXTERNAL_DOWNLOADS]: The detect-providers.sh script performs a reachability check to the official NVIDIA NIM API endpoint. This network operation is restricted to verifying API key validity and is documented as a core feature for enabling open-weight model support.
[DYNAMIC_EXECUTION]: The scripts/gen-star-history.py utility uses subprocess.run to call the GitHub CLI (gh). This script is a maintenance tool for generating star history SVGs for the documentation. The execution is limited to the specific gh api command and does not use a shell, minimizing risk.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input (the deliberation problem). It implements a 'Problem Restate Gate' and a multi-round, blind-first protocol which serves as a robust structural defense against immediate obedience to instructions embedded within the processed data.

council