Wink Pings

Claude Code Diagnosed Eight Database Incidents in Ten Days; Five Failed Because It Couldn't See the Scene

A paper reviews eight production database troubleshooting incidents handled by Claude Code: all five concurrency incidents had information-access gaps. Let the agent see the scene first; stronger reasoning comes later.

A new paper reviews how Claude Code diagnosed eight database incidents over ten days. The setting was a production message-dispatch engine, and all the problems were database-related. The authors compared, item by item, the evidence each incident required with the information the agent could actually see, and found that the most common bottleneck was access to production state.

image

image

image

All five incidents involving concurrent state had information-access gaps. The gaps included not being able to see live locks, truncated deadlock reports, lack of safe read-only query permissions, and no clear environmental context. One of them had only an access gap; the other four also involved reasoning errors, such as anchoring on an early cause, incomplete configuration exploration, and misreading the transaction semantics of the code. For the three incidents that did not depend on concurrency, the information the agent already had was sufficient.

One detail in the paper is crucial: the agent's incorrect assumptions were consistent with the evidence it could see; it did not fabricate things out of thin air. Unable to see the scene, it could only keep guessing along incomplete information, then ship a fix, deploy, observe, and roll back. Every wrong guess consumed a round of deployment and observation.

The authors' recommendation is to add the access needed for troubleshooting into existing security controls, creating a forensics layer: safe query APIs, lock and deadlock recorders, and production replay tools. In engineering terms, let the agent see the scene first, then talk about stronger reasoning. This judgment is not flashy, but it is very practical.

The study has boundaries. The paper covers only one system, and the proposal was not experimentally validated. It is more like a field observation than a complete answer.

This also serves as a reminder for evaluating troubleshooting agents: testing whether a model answers correctly only with complete logs and manually curated context is one thing; what the model can see in a production environment is another. Safe read-only queries, lock-wait and deadlock records, and replayable scenes may determine troubleshooting efficiency earlier than switching to a stronger model.

One online comment was blunt: if the evidence is gone, even stronger reasoning is useless.

Paper link: https://arxiv.org/abs/2611.17304

发布时间: 2026-11-18 17:26