Claude does not suspect its own just-shipped change first when a user reports a regression

What happened

In a long session, Claude shipped two changes to one API endpoint of a web app. Both added per-request lookups to the endpoint. The user then reported that the page served by that endpoint had become very slow.

Claude began by timing an older upstream API call. It did not first check the lookups it had just added. The user had to interrupt and tell it to check the most recent change first.

Measuring later showed:

• the new lookups added only about 0.2 s;
• the real cause was in the deployment logs: rows stuck in a queued state were re-downloading large files on every page load.

So the investigation order was wrong, and the logs were not the first stop either.

Expected

When a user reports a regression right after Claude changed the affected code path, Claude should:

1. check its own most recent diff first, by measuring or reverting it;
2. check the deployment and runtime logs for that path;
3. only then look at older components.

Environment

Claude Code CLI with the Opus model. The session was long and its context had been compacted.