The Agent Skills Directory

[PROMPT_INJECTION]: The skill's architecture establishes a surface for indirect prompt injection by design, as the agent processes data from external HTTP tools and user-provided instructions.\n
Ingestion points: Data enters the agent's context through the instructions argument in convex/agent.ts and the body of requests handled by tool endpoints in convex/http.ts.\n
Boundary markers: The reference implementations do not demonstrate the use of delimiters or 'ignore' instructions to isolate untrusted content from the agent's primary system instructions.\n
Capability inventory: The agent can trigger state changes in the database via Convex mutations and perform external network operations through the defined tool calling interface.\n
Sanitization: No logic for input validation, escaping, or filtering is included in the provided code snippets to protect the agent context from adversarial data injection.

convex-subconscious