The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection because it ingests untrusted data from an external source (Convex app events) and instructs the agent to perform high-privilege tasks like modifying the codebase based on that data.
Ingestion points: The wait_for_event tool retrieves external data payloads for convex_error, prod_error, and feature_request events from the Convex platform and Sentinel.
Boundary markers: There are no instructions or delimiters provided to help the agent distinguish between the event data and its own core instructions, increasing the risk that the agent will follow malicious instructions contained within the event payload.
Capability inventory: The skill explicitly directs the agent to "fix it" or "build it," implying file system write access and potentially the execution of build/test commands to verify changes.
Sanitization: No sanitization, validation, or filtering of the incoming event content is described in the workflow.

convex-monitor