Hi Jiajun, >From the screenshot you shared it's not clear to me what you are trying to achieve specifically.
I have had good and consistent results with LLMs via [1], the most interesting ones being: - when a plan *isn't* retained as best, was it generated and discarded beacause of cost, or actually never generated? - OOMs and loops: detection and suggested RCA with involved rules - hygiene for rules attempts/application ratios, breakdown of what rules where applied etc (general characterization of the query at hand and its planning "profile") Best regards, Alessandro [1] https://calcite.apache.org/community/#debugging-planning-issues-using-calcites-built-in-loggers On Sun, May 31, 2026, 06:03 Jiajun Xie <[email protected]> wrote: > Hi everyone, > > When working with Calcite, I often need to keep track of changes to SqlNode > and RelNode. > > Standard logging cannot flexibly capture online issues, so I frequently use > Arthas to observe them. > > I want the AI agent to better understand Calcite’s runtime status, and I’m > currently building a skill for this purpose. > > Here is the link: > > https://github.com/JiajunBernoulli/ArthasClaw/blob/main/skills/calcite-rule-tracer/assets/calcite-rule-tracer.png > > Feel free to use it to help analyze online issues. >
