The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions for building adapters that ingest and process external data, creating an attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through AgentConfig (provided in the payload of the start method) and AgentRunRequest.input (provided to the invoke methods) as described in SKILL.md.
Boundary markers: The instructions mandate validating untrusted values against the JSON Schemas included with the nemo-fabric-adapter-contract package to ensure data conforms to expected shapes.
Capability inventory: The skill provides a boilerplate TargetRuntime in SKILL.md which includes methods like invoke and invoke_openai_stream that pass external input directly to a target execution engine (self.target.run).
Sanitization: The skill promotes sanitization by instructing developers to map validated configuration into target-native constructs at the adapter boundary and to reject unsupported values with stable error codes.

nemo-fabric-build-adapter