[BUG] Claude fails to accomplish a clearly linear task, compacts 4x in a row and produces ~20 lines of code.

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?

I have given Claude skills, context start is at 27%. Claude has a clearly delineated epic, story, task based workflow. Tasks are granular. I asked Claude to implement 6 unit tests. Claude, on your "highest thinking model" references completely unneeded files in wrong paths (reading Claude that are not needed for the task at hand). It has explicit paths to 6 unit tests it needs to write, implementation criteria, and acceptance criteria.

Claude is using skills that ensure it follows best jest unit testing practices, and compacts immediately before beginning work. It proceeds to compact 3 more times.

I adjust with deleting Claude files in sub directories, leaving one master Claude with clear standards, use pnpm for instance. This is referenced in user directory, GitHub repos directory, and project directory global Claude. It is referenced in the code. It is referenced in the task. It continues to use npx, wander around confused, compacting and burning my tokens.

I have downgraded models, and the behavior is the same. This release is awful.

Kudos to giving people 50$ to test out a non production ready release, but that is not enough with the quality of work you produced.

This workflow worked perfectly before your rushed 4.6 release. Since then, every single version you have given has resulted in compaction within ~25% of the usual code produced. Death spirals of compaction on the smartest model, medium model, and low model.

What Should Happen?

Claude should actually utilize the claude.md's and read specific instructions.
Claude should not spam explore every file in the directory that is remotely touching the task at hand (unit test) given instructions are clearly labeled in a story.md that is referenced at chat beginning. Claude should actually deliver the code (simple unit tests) instead of struggling to get off the ground.

Error Messages/Logs

Steps to Reproduce

Ask claude to do any form of task at hand in a front end code base, with clearly defined claude.md's, skills, and steps of work to deliver simple code delivery of unit tests

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.39

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

Other

Additional Information

_No response_