The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides specific system instructions to the model to ensure verbatim relay of transcribed speech to the agent. These instructions are designed to prevent semantic drift and hallucinations, which serves as a safety measure rather than a malicious injection.
[CREDENTIALS_UNSAFE]: The skill implements secure credential management by collecting the OpenAI API key via a secure environment prompt and storing it in a local, git-ignored .env file. The server script mints ephemeral client secrets for the browser, ensuring the long-lived API key remains protected server-side.
[COMMAND_EXECUTION]: The setup script (scripts/setup.py) uses subprocess.run to execute Node.js and Python test suites. This is a standard practice for verifying the integrity of the project during initialization and does not involve arbitrary command execution.
[EXTERNAL_DOWNLOADS]: The skill communicates with the official OpenAI Realtime API (api.openai.com) for WebRTC session management. This interaction is central to the skill's functionality and involves a well-known, trusted service.
[DATA_EXFILTRATION]: No unauthorized data exfiltration was detected. All network operations are limited to the user-configured Starchild runtime and the OpenAI API, as described in the documentation.
[REMOTE_CODE_EXECUTION]: The skill does not download or execute arbitrary scripts or binaries from remote servers. All functional code is provided within the bundled template files.

@2004/realtime-voice-preview