[BUG] Conext compressor entirely looses context sending ai to wrong path making technical debt

Environment

• Platform (select one):
• [ ] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ *] Other: <!-- specify --> "windows git bash"
• Claude CLI version: <!-- output of claude --version -->
• Operating System: <!-- e.g. macOS 14.3, Windows 11, Ubuntu 22.04 -->
• Terminal: <!-- e.g. iTerm2, Terminal App -->

Bug Description

<!-- A clear and concise description of the bug -->
When using claude code to make large technical revisions to classes, if it hits the context limit the compressor will crunch the conversation into a summary. That summary lacks all important context and considerations of current technical work. This leaves the ai with not enough context to continue where it left off. Causing the ai to go into loops of performing the wrong actions over and over again.

Example
I have being trying to modify a gguf_tokenizer i've been working on to use gpt-neoxfast for two days now, and every single time it gets to the context compression it gets confused, assumes i mean gpt2 tokenizer and starts undoing all the work it just did. I catch it doing it, tell it no we're doing the correct one, it admits it's confusion and start fixing it. But then the compressor hits again and it's off to the races reverting back to gpt2.

Steps to Reproduce

1. <!-- First step -->

start producing a large technical system that has similar competing versions of the same system but incompatible. Make sure it's a big script

1. <!-- Second step -->

have it work on it until compression

1. <!-- And so on... -->

compressor will nuke it's progress and context, leaving it confused and going the wrong direction creating technical debt

Expected Behavior

<!-- What you expected to happen -->
compressor should provide MORE contextual understanding, have working docs its creating to maintain proper true context, and keep the ai on task, without drift. Compressor should provide the last few actions as well as a launching point in order to prevent loss of context.

Actual Behavior

<!-- What actually happened -->
Ai compresses context, gets completely lost, makes assumsions instead of validating and then starts producing technical debt. Leaving me to rexplain everything over and over and over every 20 minutes. It's exhausting.

Additional Context

<!-- Add any other context about the problem here, such as screenshots, logs, etc. -->