The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Tailscale CLI and the DuoDuo agent CLI to configure networking routes, serve local services over HTTPS, and manage agent plugin states. All commands require owner approval and are scoped to the local environment and private tailnet.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a data ingestion path where user-supplied chat and voice transcripts from a mobile app or external 'cerebellum' service are processed by the agent. While this presents an attack surface for indirect prompt injection, the risk is mitigated by specific instructional prompts (e.g., the '--pocket' flag) designed to constrain agent output formats.
[DYNAMIC_EXECUTION]: The provided scripts/create-room.sh script uses node -e to dynamically parse and generate JSON-RPC payloads for communicating with the local agent daemon via a Unix socket. This execution is localized and includes regex-based validation for room identifiers and directory checks for workspace paths to prevent command injection.
[SAFE]: Implements proactive credential safety by instructing the owner to input the sensitive AMBIENT_CEREBELLUM_TOKEN directly into a local .env file via the terminal, ensuring the secret never enters the AI chat context, event logs, or history.
[SAFE]: Enforces a strict network security policy that permits service exposure only through tailscale serve (private tailnet access) and explicitly forbids the use of tailscale funnel to prevent unauthorized public access to the room and audio streams.

duoduo-ambient