[BUG] sonnet4.5 ultrathink normal think, unbearable. Fix the context!

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude forgets the last prompt due to agent use and compacting of context. After that it wants to continue the next task instead of doing a check on last task causing it to lose those memories after the next prompt.

Also, non related to auto-compact as this happens also in the beginning and during a session with enough context left: It simply cant follow orders, it has to give its own spin to the cmds which are always completely off-base.

This happens today regularly, in contrast to yesterday and before that. A slow decrease untill now.

What Should Happen?

Its should act as can be reasonable expected. Not delete files, not remember last prompts, etc.

Error Messages/Logs

Steps to Reproduce

Ask it to do a codebase research with agents and see the context immediately decreasing to auto-compact and then forgetting 50% at least of what it should be doing or was doing.

<img width="1907" height="252" alt="Image" src="https://github.com/user-attachments/assets/cb6ef8c5-9a70-4eb2-a113-c6cfe5e9d288" />
<img width="1170" height="154" alt="Image" src="https://github.com/user-attachments/assets/a05a7a27-3810-4972-8153-92910e4dcf8a" />
<img width="1643" height="193" alt="Image" src="https://github.com/user-attachments/assets/45c75e73-873b-4178-98d1-b8153529918a" />

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

latest

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Xterm

Additional Information

_No response_