[Bug] Anthropic API Error: Safety classifier incorrectly blocking legitimate marketplace account listing feature

Bug Description
I'm using Claude Code to build my own gaming-account marketplace (similar to FunPay). The task: a section for my own "World of Tanks" (Lesta) accounts. They are imported from my screenshot tool (a ready.json with the garage and stats) and become listings with vehicle-level filters. Buyers get the account through the site's existing auto-delivery.

In a single session the safety classifier stopped Opus 5.5 mid-response 5–6 times:

• on the service that writes an imported listing to the DB (the login/password goes into the auto-delivery field, the same way it does for every manual seller listing on the site);
• on unit tests for the ready.json parser, which used synthetic data;
• on a small change to the listing form (hiding the admin-only section from sellers);
• on a local browser check of regular catalog pages that have nothing to do with the feature.

Twice a file was left half-written and had to be deleted. After the first few stops the model declined to touch the feature at all, including design and code review, so it had to be finished by another model in a separate session.

It looks like after the first trigger everything related to the topic gets blocked, even harmless UI edits and checks. Please review whether these are false positives for a legitimate marketplace.

Environment Info

• Platform: win32
• Terminal: null
• Version: 2.1.280
• Feedback ID: 115cdbdd-1b2f-4dca-a502-f178b106c5b8

Errors

[]