The Agent Skills Directory

[DYNAMIC_EXECUTION]: The configuration loader in references/implementation-guide.md uses dynamic require calls to load environment-specific configuration files.
The code require(./speak.${env}.json) loads modules from a path computed using the NODE_ENV environment variable. While a standard pattern for managing development, staging, and production settings, it involves dynamic loading from computed paths.
[INDIRECT_PROMPT_INJECTION]: The architecture is designed to process untrusted external data (audio recordings and text) which serves as an entry point for potential indirect prompt injection attacks.
Ingestion points: Untrusted data enters the agent context through the SpeakClient.speech.recognize and SpeakClient.speech.score methods defined in src/speak/client.ts.
Boundary markers: The provided implementation snippets do not include explicit delimiter markers or specific instructions to the agent to ignore embedded commands within the processed data.
Capability inventory: The skill environment is granted significant tool access, including Read, Write, Edit, Bash, and Grep.
Sanitization: The reference code does not demonstrate sanitization or validation of the external content before it is interpolated or processed by the SDK.

speak-reference-architecture