The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation outlines an update check that fetches version metadata from 'https://raw.githubusercontent.com/abpai/skills/main/versions.json'. This is a vendor-owned resource used for legitimate maintenance.
[COMMAND_EXECUTION]: The 'scripts/debate.sh' script executes external command-line tools '@anthropic-ai/claude-code' and '@openai/codex'. This is the core functionality of the skill and is documented as a prerequisite for architectural orchestration.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) because it ingests directory structures, dependency manifests, and code snippets from local repositories into its prompts. Ingestion points: The 'scripts/build-context.sh' script extracts data from local paths provided by the user. Boundary markers: The skill uses Markdown headers and code blocks to delimit the ingested content within the LLM prompt. Capability inventory: The 'scripts/debate.sh' script executes LLM CLI tools using the gathered context as input. Sanitization: No sanitization or escaping of the ingested code snippets is performed beyond basic line-count truncation.

arch-council