The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio input via voice::transcribe and voice::dictation::push, converting it into text that is returned to the agent's context. This creates a surface for indirect instructions to be embedded in audio and executed by the agent.
Ingestion points: Audio is received through file paths or base64 PCM chunks in voice::transcribe and voice::dictation functions (SKILL.md).
Boundary markers: No specific boundary markers or instructions to ignore embedded commands in the resulting transcripts are documented.
Capability inventory: The system can transcribe audio, manage local model files, and execute system commands for speech processing.
Sanitization: No sanitization of the resulting transcript is mentioned before it reaches the agent context.
[COMMAND_EXECUTION]: The skill interfaces with local system binaries for speech synthesis and recognition.
Evidence: The skill configuration requires paths to executables such as whisper-cli and piper-tts (tts.piper.command). It also uses system-level tools like say or espeak in file-output mode.
[EXTERNAL_DOWNLOADS]: The skill includes functionality to manage and retrieve external assets required for its operation.
Evidence: The voice::models::download function is used to install local speech models. While the documentation states these are checksum-verified, it represents the programmatic retrieval of external binary data.

voice