[DESIGN FLAW] Context compaction destroys workflow — no CLAUDE.md reload, no pause, auto-continues and breaks own work

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?

Problem: The fundamental design of context compaction is broken, not just implementation bugs.
Issues:

CLAUDE.md is NOT re-read after compaction — project memory is useless
Claude auto-continues immediately after compaction with partial context
Claude actively destroys work it just completed because it forgot doing it
No pause, no confirmation, no warning before continuing
User cannot react in time

Analogy: It's like having a brilliant employee who every 5 minutes becomes an imbecile and has to relearn everything — while actively destroying what they just built.
This is not a bug in compaction — compaction itself is the bug.
Requested:

Option to DISABLE auto-compaction entirely
PAUSE and ask user before continuing after compaction
Re-read CLAUDE.md after every compaction
Show what was lost in compaction

What Should Happen?

After context compaction:

1. Claude should PAUSE and ask user before continuing work
2. Claude should RE-READ CLAUDE.md to restore project context
3. Claude should NOT auto-continue and destroy work it just completed
4. User should have option to DISABLE auto-compaction entirely

Error Messages/Logs

No error message - this is a design flaw, not a crash.

The problem: compaction happens silently mid-task, Claude loses context, 
then immediately continues working and UNDOES or BREAKS what it just built 
because it forgot doing it.

Steps to Reproduce

1. Start a complex task that requires multiple steps
2. Have CLAUDE.md with project-specific instructions (server paths, workflows, file permissions rules)
3. Work until context compaction triggers automatically
4. Observe: Claude does NOT re-read CLAUDE.md
5. Observe: Claude immediately continues with partial memory
6. Observe: Claude "fixes" things by breaking what it just correctly built

Result: User must repeat instructions dozens of times.
It's like having a brilliant employee who every 5 minutes becomes an imbecile
and has to relearn everything — while actively destroying what they just built.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.25 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

I work on complex, multi-hour development sessions with intricate project dependencies.

Real-world impact of current compaction design:

• I have to manually type "read CLAUDE.md" dozens of times per session
• I have to explain the same project dependencies over and over again
• I have to revert changes that Claude made after compaction (because it "fixed" things it forgot it just built correctly)
• I have to restart from checkpoint because Claude destroyed hours of work
• I have to re-explain server paths, file permission workflows, database configurations repeatedly

CLAUDE.md was supposed to solve this — but it's NOT re-read after compaction, making it useless for long sessions.

This makes Claude Code extremely frustrating for professional use on complex projects.
The tool is brilliant, but compaction ruins it every single time.

Suggested solutions:

1. Option to completely disable auto-compaction
2. Always re-read CLAUDE.md after compaction
3. Pause and ask user before continuing after compaction
4. Show what context was lost so user can decide how to proceed