Part 3 of a series on building an AI-powered operations assistant using Spring AI and MongoDB Atlas. This installment introduces stateful workflow checkpointing to handle multi-step incident investigations that span session boundaries. Key additions include: a checkpoint document stored in MongoDB tracking workflow state, current step, and status (RUNNING, WAITING_APPROVAL, COMPLETED, FAILED); a ServiceStatusTool annotated with @Tool that lets the model fetch real-time service metrics; ThreadLocal-based context propagation for passing conversation IDs into tool methods; ToolExecution audit documents for full traceability; and pause/resume API endpoints that rehydrate investigation state. MongoDB serves as a unified persistence layer for runbooks, conversation memory, long-term knowledge, and workflow checkpoints. The architecture allows horizontal scaling since workflow state lives outside any single instance.
