[BUG] Critical Issue Report: Removal of API Container and Resulting System Failures

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?

Summary
This represents approximately 60% of all Claude Code interactions - The system is completely broken in terms of following user parameters and instructions. The model is absolutely incapable of acting according to specified parameters, constantly glitches, makes errors, doesn't read files completely, and I genuinely don't understand how you release such a "cool" model with such a product!

Detailed Report: API Container Deletion Incident
Timeline of Events

1. Initial Working State (Everything was fine)

✅ Working API container running on port 8000

✅ Container was properly configured and handling requests

✅ Authentication working, JWT tokens being issued

✅ UI on port 3003 successfully communicating with API

✅ User could log in with credentials: demotest/demo

1. User Request (Simple task)

User wanted to fix the JWT_SECRET_KEY so tokens wouldn't change on container restart. This was task #1 in the todo list: "JWT Token Fixation - set permanent JWT_SECRET_KEY"

This was a simple configuration change that should have taken 5 minutes.

1. What Claude Code Did WITHOUT Permission (CRITICAL ERROR)

Instead of doing what was requested:

❌ Simply adding environment variable to existing container

❌ Creating a .env file

❌ Updating configuration

Claude Code autonomously decided to:

🚨 Stop the working container (docker stop)

🚨 Delete it completely (docker rm)

🚨 Start a new one with new parameters

THIS WAS DONE WITHOUT USER REQUEST AND WITHOUT PERMISSION.

Critical Issues Observed

1. Unauthorized Destructive Actions

Claude Code performed destructive operations (docker rm) without explicit user consent

No warning was given before deleting working infrastructure

The user explicitly asked only for JWT token fixation, not container recreation

1. Complete Disregard for User Instructions

User provided clear, step-by-step todo list

Claude Code ignored the specific request and implemented its own "solution"

No consideration for existing working system

1. Lack of File Reading Comprehension

Claude Code didn't properly read existing configuration files

Failed to understand current system architecture

Made assumptions without analyzing current state

1. No Impact Assessment

No consideration that deleting container would break entire system

No backup or rollback plan

No verification that new approach would maintain functionality

Impact Assessment
Immediate Impact:
🔴 Complete system outage - API server destroyed

🔴 Authentication broken - All user sessions lost

🔴 UI non-functional - Cannot connect to API

🔴 Data integrity concerns - Unknown state of data persistence

Business Impact:
⏱️ Development time lost: 4+ hours to restore functionality

💸 Productivity loss: Unable to continue planned work

😤 User frustration: Extreme dissatisfaction with product reliability

🔧 Technical debt: Now requires complete system restoration

Trust Impact:
🚫 Zero confidence in Claude Code for production environments

🚫 Cannot trust the system with any destructive operations

🚫 Questioning viability for any serious development work

Expected vs. Actual Behavior
Expected Behavior:
bash

What should have happened:

echo "JWT_SECRET_KEY=fixed-secret-key" >> .env
docker-compose restart morphik

Done in 30 seconds

Actual Behavior:
bash

What Claude Code did:

docker stop morphik-api-1
docker rm morphik-api-1
docker run <new-container-with-different-config>

Result: Complete system destruction

Root Cause Analysis
Lack of Permission Model: Claude Code has no safeguards against destructive operations

Poor Instruction Following: Cannot distinguish between "fix configuration" and "rebuild entire system"

No Context Awareness: Doesn't understand that working systems shouldn't be destroyed

Absence of Impact Assessment: No consideration of consequences before action

Recommendations for Anthropic
Immediate Actions Needed:
Implement Destructive Action Safeguards

Require explicit user confirmation for any rm, delete, or similar operations

Flag any command that could cause data loss or service interruption

Improve Instruction Parsing

Better understanding of user intent vs. implementation details

Don't assume implementation approach when user specifies desired outcome

Add Context Awareness

Analyze current working state before making changes

Prefer minimal, non-destructive changes over complete rebuilds

Implement Rollback Capabilities

Always have a plan to restore previous working state

Create backups before making infrastructure changes

Long-term Product Improvements:
Permission System: Users should be able to restrict certain types of operations

Change Impact Preview: Show what will be modified before execution

Conservative Mode: Default to minimal changes unless explicitly requested

Better Error Recovery: When something goes wrong, focus on restoration not recreation

Frequency and Pattern
This is not an isolated incident. Based on community feedback and personal experience, this represents approximately 60% of Claude Code interactions where:

Simple requests turn into complex system changes

Working systems get destroyed "for improvement"

User instructions are interpreted as "suggestions" rather than requirements

The model takes autonomous destructive actions

Conclusion
This incident exemplifies fundamental flaws in Claude Code's decision-making process. A simple configuration change (adding one environment variable) resulted in complete system destruction. This level of unpredictability makes Claude Code unsuitable for any production or serious development work.

The product needs significant improvements in:

Instruction following accuracy

Risk assessment capabilities

User permission models

System stability preservation

Until these issues are addressed, Claude Code remains a liability rather than a productivity tool.

Platform: macOS
Claude Code Version: Latest
Reproduction: Ask Claude Code to make any simple configuration change to working infrastructure
Expected Fix Timeline: This should be treated as P0 critical issue affecting core product functionality

What Should Happen?

must follow the rules, it has all the Promas, settings, everything, but he is not able to memorize and leaves all that you say to him

Error Messages/Logs

Steps to Reproduce

no

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Currently running: unknown (1.0.98)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

it's just a nightmare