LLM-driven control-data plane separation for zero-intrusion TCC transactions in legacy microservices

The proliferation of microservices introduces significant challenges in maintaining the strict consistency of distributed transactions , particularly when cloud-native workflows integrate with heterogeneous legacy systems lacking distributed coordination semantics. Traditional compensation protocols, such as Try-Confirm-Cancel (TCC), are highly invasive, compelling target services to implement complex state machines. To address this architectural impedance mismatch, we propose a zero-intrusion transactional architecture that logically and physically separates the control and data planes. In the offline control plane, a Large Language Model (LLM) functions as a semantic parser to analyze OpenAPI specifications and synthesize deterministic JSON mapping rules (DSL). In the online data plane, a TCC proxy operates as a Resource Manager for Apache Seata, transparently intercepting RESTful requests, caching the ``Try'' phase context, and deterministically executing the compiled DSL compensation rules during the ``Cancel'' phase. This solution ensures constrained \((O(1))\) routing latency and execution determinism by rigidly isolating the LLM from the real-time execution path, hence overcoming the unpredictability of generative AI agents. Comprehensive assessments in a medical scheduling context with 100 simultaneous global transactions illustrate the architecture's resilience. With an induced failure rate of 42%, the proxy effectively managed all enforced rollbacks, resulting in no phantom records and sustaining a complete 0.00% inconsistency rate.