The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a metadata passthrough feature where arbitrary JSON keys/values are appended to the agent's prompt to steer its reasoning.
Ingestion points: Found in SKILL.md (Writes section) and references/metadata-passthrough.md. Data enters via the metadata field in POST payloads (REST/SSE) and WebSocket frames, as well as the MENTOR:CONTEXT_UPDATE message in the embedded iframe transport.
Boundary markers: The interpolated content is wrapped in <CONTEXT METADATA> and </END CONTEXT METADATA> markers as shown in references/metadata-passthrough.md.
Capability inventory: The skill facilitates chat turns, history exports, and session management with deployed AI agents.
Sanitization: The documentation does not specify if the values within the metadata object are sanitized to prevent them from containing the closing marker, which is a common vulnerability surface for prompt injection.
[COMMAND_EXECUTION]: The skill includes curl command examples to demonstrate how to interact with the API.
Evidence: Example shell commands in SKILL.md target the vendor's official domains (api.iblai.app and asgi.data.iblai.app) and use standard authorization headers.

iblai-api-agent-session