May 2026 monthly summary for langchain4j/langchain4j. Key features delivered focus on streaming reliability and correctness of tool invocation in streaming mode: - Streaming path reliability: ensured per-tool ReturnBehavior is honored in the streaming path by using ToolServiceContext, aligning streaming behavior with the non-streaming path (commit 116e81aa08c1bbe78a09fcb75c125b6a6ebfa345). - Sentinel handling: added a guard to skip empty trailing tool_calls sentinel chunks emitted by some OpenAI-compatible providers to prevent ghost ToolExecutionRequests (commit 81d0a686ab4969764051adf7c505b9abe8705660). - Regression tests: introduced tests that validate correct ReturnBehavior handling in streaming and the sentinel-skip behavior; all related tests pass locally. Major bugs fixed: - Correctly read ReturnBehavior from per-invocation context in streaming responses, eliminating unintended behavior downgrade in streaming mode. - Prevent ghost ToolExecutionRequests by ignoring end-of-stream sentinel frames before building tool calls. - Comprehensive test coverage confirming positive and negative paths; local verification shows green results across module and core components. Overall impact and accomplishments: - Restored intended API semantics and streaming behavior, ensuring reliability and consistency across streaming and non-streaming paths. - Improved stability for OpenAI-compatible providers, reducing user-visible errors and malformed tool-execution chains. - Demonstrated end-to-end test-driven improvements with strong verification and no breaking changes to public API. Technologies/skills demonstrated: - Java-based streaming architecture, OpenAI/Tool streaming integration, and per-tool ReturnBehavior handling. - Guarded control-flow in streaming accumulators and per-tool context usage; robust regression testing and local verification. - Clear focus on business value: more reliable streaming results, fewer ghost tool invocations, and improved developer confidence in streaming paths.