The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text and voice design instructions which are interpolated into a JSON payload sent to the Xiaomi MiMo API. While this presents an attack surface where malicious text could influence the TTS model's behavior, the risk is mitigated by length constraints and the specialized nature of the target service.
Ingestion points: tool/worker.py (synthesize operation reading text, design, and instructions fields).
Boundary markers: Absent; content is directly assigned to JSON fields.
Capability inventory: Network communication via urllib.request and local file writing for WAV output.
Sanitization: Input length is capped at 8000 characters for text and 2000 for instructions.
[COMMAND_EXECUTION]: The tool/setup.sh script executes basic environment verification commands to ensure Python 3 and required standard libraries are available.
[DATA_EXFILTRATION]: User-provided text is transmitted to xiaomimimo.com subdomains for synthesis. This is the documented and primary purpose of the skill, and API keys are correctly handled via environment variables with redaction logic in place to prevent accidental leakage in logs.

narova-mimo