The Agent Skills Directory

EXTERNAL_DOWNLOADS (LOW): The skill relies on npx to fetch and run the Convex CLI and MCP server (npx -y convex@latest mcp start). While these are industry-standard tools, they involve runtime code execution from external sources (npm registry).
COMMAND_EXECUTION (LOW): The skill provides the agent with instructions to run powerful backend management commands (e.g., npx convex dev, npx convex run). This is essential for the skill's utility but constitutes a high-privilege capability tier.
CREDENTIALS_UNSAFE (LOW): The references/cli.md file identifies the location of sensitive authentication tokens at ~/.convex/config.json. While the skill does not explicitly exfiltrate these, their exposure in the documentation increases the risk if the agent is manipulated. The skill also provides tools for manipulating environment variables (convex_envSet).
PROMPT_INJECTION (LOW): The skill is susceptible to indirect prompt injection (Category 8). \n
Ingestion points: Data enters the agent's context through convex_logs, convex_data, and convex_runOneoffQuery. \n
Boundary markers: No specific delimiters or instructions to treat tool output as untrusted are present in the router or templates. \n
Capability inventory: The agent can execute functions (convex_run), modify environment variables (convex_envSet), and deploy code (npx convex deploy). \n
Sanitization: There is no evidence of sanitization or filtering of external data before it is processed by the agent.

convex