Keep Azure OpenAI connections intact through agent and team setup

Cache responses reliably across a wider range of agent inputs

We resolved a failure in cache key generation when the input contained types that are not directly JSON-serializable, ensuring caching works reliably across a broader range of agent inputs.

Get clean, deduplicated results from LanceDB hybrid search

We resolved an additional case where hybrid search could surface the same document more than once when it matched across multiple search indices.

Preserve tool names throughout LiteLLM streaming responses

We resolved an issue where empty string values in streamed LiteLLM responses could overwrite previously accumulated tool names, resulting in tool calls with missing identifiers.