The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it transcribes live audio from phone calls and feeds the resulting text directly into the LLM conversation context.
Ingestion points: Incoming audio streams from Twilio are processed in scripts/server.py and scripts/server_realtime.py to produce text transcripts.
Boundary markers: There are no explicit delimiters or "ignore instructions" warnings surrounding the interpolated user transcript in the prompt.
Capability inventory: The agent has access to a web_search tool (utilizing the Brave Search API) and generates streaming audio output.
Sanitization: No sanitization or validation is performed on transcribed speech before it is processed by the LLM.
[COMMAND_EXECUTION]: The server script executes the ffmpeg utility as a subprocess to handle audio transcoding for Twilio compatibility.
Evidence: scripts/server.py uses asyncio.create_subprocess_exec to launch ffmpeg for converting audio streams to the required mu-law format.

phone-agent