The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from multiple external sources to influence autonomous agent behavior.
Ingestion points: External Neuro API messages handled by vrchat_neuro_handle_action, VRChat OSC events from vrchat_observation_from_osc, and multimodal observations (STT, vision, stream chat) via vrchat_observation_ingest.
Boundary markers: The instructions explicitly state that Neuro action data must be treated as untrusted input and require the use of a "Hermes safety gate" and dry-run profiles.
Capability inventory: The skill allows sending OSC commands to VRChat, synthesizing audio via VOICEVOX, and reading local system state (processes, windows, ports).
Sanitization: Operating rules require the rejection of invalid JSON, unknown actions, and unauthorized OSC payloads.
[DATA_EXFILTRATION]: The skill includes diagnostic tools that inspect local system metadata, exposing environmental state to the agent's context.
Evidence: scripts/vrchat_runtime_doctor.py is described as collecting "local port snapshots" and "relevant visible Windows desktop windows" for troubleshooting.
[EXTERNAL_DOWNLOADS]: The skill documentation identifies external dependencies and third-party code requirements.
Evidence: Documentation lists python-osc and websockets==15.0.1 as requirements, and requires a local clone of the neuro-sdk from https://github.com/VedalAI/neuro-sdk.

neuro-vrchat