Execution-Grounded Continual Learning for CLI-Based SONiC Network Operations

A recent paper on arXiv (2608.09184) presents a dual-path consequence-aware agent designed for command-line interface (CLI) operations within SONiC networks. This agent is capable of generating various complete actions, forecasting their outcomes, and ultimately choosing the best action through a process of utility- and risk-aware reranking. The proposal-side path distills reusable operational insights into accessible guidance, enhancing feasible-action coverage without altering the proposal LLM. Meanwhile, the selection-side path refines the consequence predictor via session-level LoRA updates informed by actual SSH feedback. This method overcomes the shortcomings of current techniques that either emphasize command generation or final configuration accuracy, neglecting the benefits of execution-grounded experience. The paper is classified as 'new' and was published on arXiv.