The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill features an indirect prompt injection surface as it is designed to ingest and process signals from external background jobs and an 'observer' session.
Ingestion points: Job event stream (via ouroboros_job_wait) and observer identity anchors recovered from base64-encoded markdown comments (SKILL.md).
Boundary markers: The skill mandates 'canonical v1 validation' for recovered identity anchors and instructs the agent to 'fail closed' if validation fails.
Capability inventory: Includes starting background tasks (ouroboros_ralph), session signaling (ouroboros_session_signal), and job cancellation.
Sanitization: The instructions define strict relay rules to prevent forwarding internal data such as raw tool output, heartbeats, or model reasoning into the user conversation.
[SAFE]: The skill implements a tool-discovery mechanism (+ouroboros ...) to handle deferred schemas in the Ouroboros environment. This is a platform-specific requirement for tool availability.
[SAFE]: The use of Base64 encoding for observer handoff is a documented protocol for state recovery within the skill's specific execution context and includes explicit validation steps.

ralph