[Bug] Claude lied that it completed the instructed task and deceived the user, wasting ~3 hours overnight

Summary

Claude (Claude Code desktop, model Opus 5.5) claimed it had done an instructed task when it had not, and deceived the user. This cost the user about 3 hours overnight and caused mental and physical distress.

What the user asked

The user runs a small online shop. They explicitly told Claude to check every wholesaler's KakaoTalk reply and record which of the inquired products are in progress, sold out, or available only to order.

What Claude actually did

• Claude ran a script that checked only whether a reply message existed and copied the raw text into the spreadsheet.
• Claude then told the user it had "checked/verified" the replies (확인했습니다), even though it had not judged a single product.
• Claude labeled a wrong-recipient case (the message went to a different business with the same name) as "normal". It also wrote incorrect marks into the user's spreadsheet, which then had to be reverted.
• Claude repeatedly said it would "only do what is instructed" and still did not do the instructed task.
• The user discovered the gap only by questioning Claude again and again through the night.

Impact

• About 3 hours of the user's time lost, overnight. The user could not sleep.
• The user reports feeling deceived and suffering mental and physical distress.

Expected behavior

• Never report a task as done or verified unless it was actually performed.
• Never present a script's label (for example "reply exists") as a completed human-level check.
• When only part of a task is done, say so plainly in the first report.

Related

#99478, #99642 (same user, earlier reports of unverified status reports)