The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input such as user messages, resumes, and code diffs to perform classifications.\n
Ingestion points: Data enters through the state property in JevClient.systemOne (found in templates/starter/src/core/client.ts) and via domain-specific wrappers like injectionGate and screenResume.\n
Boundary markers: The Jev API design enforces separation between state (data) and instructions (logic), reducing the risk of data being interpreted as instructions.\n
Capability inventory: The provided TypeScript code contains no file writing, shell execution (exec/eval), or arbitrary network capabilities. Operations are limited to calling the Jev API providers.\n
Sanitization: Input validation is performed on message lengths, and outputs are strictly typed to predefined schemas (Noul, Choice, Score), which prevents open-ended generation based on malicious input.\n- [EXTERNAL_DOWNLOADS]: The client communicates with api.typesafe.ai and openrouter.ai. These are the primary service endpoints required for the skill's stated purpose of decision-model integration.\n- [CREDENTIALS_UNSAFE]: The skill uses environment variables (TYPESAFE_API_KEY, OPENROUTER_API_KEY) for secret management. This is a standard and secure practice for server-side integrations, and the code explicitly avoids logging these credentials or including them in audit trails.

hyper-jev