The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local utility and provider-specific bash scripts to perform its core functions.
Evidence: SKILL.md runs scripts/query-council.sh and scripts/validate-analysis.sh.
Evidence: agent-prompt-template.md runs scripts via the {SCRIPT_PATH} variable.
[DATA_EXFILTRATION]: User input and context are transmitted to external AI provider endpoints (such as OpenAI, Gemini, Grok, and Perplexity).
Evidence: Subagents query these external services to generate comparative analyses as part of the skill's primary workflow.
[DYNAMIC_EXECUTION]: The skill sources internal shell libraries at runtime to manage prompt roles and model fallback configurations.
Evidence: SKILL.md sources scripts/lib/prompts.sh and scripts/lib/roles.sh.
Evidence: agent-prompt-template.md sources scripts/lib/model_fallback.sh.
[INDIRECT_PROMPT_INJECTION]: The skill processes data returned from external AI models, creating a potential vector for indirect prompt injection attacks.
Ingestion points: Provider responses are collected and returned to the main agent context in Step 3 of SKILL.md.
Boundary markers: The implementation uses quoted heredocs (<<'EOF') to safely handle external text and requires all responses to pass JSON schema validation via validate-analysis.sh.
Capability inventory: The skill maintains capabilities to write files to the local .claude/council-cache directory and execute shell commands.
Sanitization: All external data is subjected to strict schema validation before being incorporated into the final synthesis, significantly reducing the risk of accidental instruction execution.

deep-execution