The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user content to determine routing, creating a surface for indirect prompt injection where malicious instructions in the task data could influence the routing decision.
Ingestion points: User-supplied content is interpolated into JSON payloads (such as task, requirements, and evidence fields) which are then processed by the jev-decide CLI tool as seen in SKILL.md and assets/example.json.
Boundary markers: The skill uses a structured JSON schema to separate data, but there are no explicit delimiters or instructions to the underlying model to ignore embedded commands within the free-text evidence fields.
Capability inventory: The skill executes the jev-decide CLI and recommends further actions to be performed by the host agent's primary tools (UI, file, mailbox).
Sanitization: The instructions do not specify any validation, filtering, or escaping of the user-provided data before it is sent to the routing provider.
[COMMAND_EXECUTION]: The skill directs the agent to interact with the local system via a CLI tool.
Evidence: The SKILL.md file provides instructions for executing the jev-decide command to validate and process JSON request files.

jev-route