The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation references the inclusion of the @daily-co/daily-js library via unpkg.com and the use of the daily Python package. These resources are associated with Daily.co, a well-known WebRTC infrastructure provider that powers the conversation platform.
[INDIRECT_PROMPT_INJECTION]: The protocol described in SKILL.md identifies mechanisms for injecting external data into the agent's conversational context, such as conversation.respond and conversation.append_context. This establishes a potential attack surface where data from external participants could influence agent behavior, which is an inherent property of interactive conversational agents.
Ingestion points: Data channel events processed by the client, such as conversation.utterance and conversation.tool_call found in SKILL.md.
Boundary markers: None identified in the provided documentation examples.
Capability inventory: Control over replica speech (echo), turn-taking (interrupt), and conversational state (overwrite_context).
Sanitization: Not addressed in the protocol message examples.

tavus-cvi-interactions