The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads and executes an installation script using a piped-to-bash pattern: curl -fsSL https://openrouter.ai/labs/ori/install.sh | bash. This is documented as the standard installation method for the Ori tool authored by the same vendor (OpenRouterTeam).
[COMMAND_EXECUTION]: The skill executes several local commands including ori auth, ori eval, and bun to perform its primary function of model evaluation. It also handles PATH fallback by executing ~/.local/bin/ori directly if the binary is not in the environment path.
[EXTERNAL_DOWNLOADS]: Fetches the ori binary and project templates from openrouter.ai and raw.githubusercontent.com (implicit via standard tooling). These sources are official domains for the skill's author.
[PROMPT_INJECTION]: The automated scanner flagged a potential concealment attempt; however, analysis of SKILL.md shows this is a false positive. The instructions actually require the agent to increase transparency by narrating progress, disclosing costs, and showing internal tool outputs to the user in plain language. The 'concealment' refers to replacing internal developer jargon (like 'bakeoff') with user-friendly terms ('model comparison').
[DATA_EXFILTRATION]: While the skill accesses credentials.json and .env files to check for API keys, it contains strict instructions NEVER to print secret values or send them to external servers, only naming the key's location for the user's benefit.

spawn-ori-eval