Running claude code in serverless mode with claude init config coming from vector stores

Regarding the architectural guardrails feature: Do we have any guidance on implementing Claude in AWS Lambda with vector store-based guardrails? 1. Deploying Claude in Lambda with a custom handler that: - Initializes with pre-loaded guardrail configurations from vector stores - Enforces architectural boundaries at inference time - Uses a container image with necessary dependencies 2. Specific implementation questions: - Most efficient approach to pass vector embeddings to Claude during initialization - Best practice for structuring guardrail context from vector stores (Pinecone/Weaviate) - Memory optimization strategies when using large context windows with guardrails 3. Challenges encountered: - Cold start latency with large guardrail contexts (~2-3 seconds overhead) - Lambda memory constraints when passing full architectural patterns - Maintaining guardrail consistency across function invocations Has anyone implemented a similar approach with the Claude API? Particularly interested in patterns for initializing the context window efficiently.

_Originally posted by @spbreed in #427_