[MODEL] Claude code continued token spend even though issue was resolved, spending 350k+ estimated tokens over 24 hours

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

Asked claude to test and setup my meta ads pixels. In a 30 minute session, I exchanged observations and asked it to fix the missing pixel tags. Instruction was to: test the flow on local server and see if the pixels were firing. The issue got fixed and I confirmed, but Claude code continued to test things on local host endlessly for 24 hours. When I notices suspicious localhost activity on my posthog, I instructed claude code to switch off.

<img width="1920" height="1080" alt="Image" src="https://github.com/user-attachments/assets/af73403f-9a8e-46c1-bd98-b0b3d4c4d1a4" />

What Claude Actually Did

1. Claude continued to test meta pixel endlessly on local host
2. Only when I got to see activity on localhost on my posthog setup, I instructed claude code to kill local server
3. It went on to explain that I only instructed to test but then acknowledged the fault
4. Consumed 350k+ tokens as wastage

Expected Behavior

Claude should have:

1. Closed the local server if the issue was resolved
2. Come back honestly the first time if this was a miss
3. Have a clear estimate of token wastage

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

No, only happened once

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Shared in screenshot

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.145

Platform

Anthropic API

Additional Context

_No response_