[BUG] Claude Code session degradation + inability to debug visual/interactive issues

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've been using Claude Code to build a WordPress/Elementor site with GSAP scroll animations. Over the past week I've experienced the following consistent failures:

Visual access lies. Claude claims it can access a live site, attempts multiple methods (Chrome MCP, screencapture, qlmanage, ffmpeg, PyObjC), fails at all of them, and never clearly states it cannot see the site. It makes code changes to a live site while blind.

Unverified fixes presented as diagnoses. Claude reads JavaScript source and generates fixes it cannot verify. Each fix is presented confidently with guarantees ("this will not break anything"). The fixes break things.

Session degradation. Long sessions degrade into unusable states. Starting a new session means re-explaining everything from scratch. Memory notes carry forward inaccurate information from bad sessions.

Usage consumed by failure. A full session's worth of tokens spent on failed attempts to open a video file, failed fixes, and cleanup of those fixes — with zero net progress.

No escalation path. When Claude cannot do something (see a live site, run code on a device, inspect runtime state) it should say so immediately and stop. Instead it keeps trying, consuming usage, and blocking the user from finding help elsewhere.

What I needed: A way to debug a GSAP ScrollTrigger conflict on mobile. What I got was a week of broken code and a forest wallpaper screenshot.

What Should Happen?

Claude should be able to spot the conflict it's own code created and fix it.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce:

Open Claude Code with a live WordPress/Elementor site URL
Ask Claude to access the site
Claude attempts Chrome MCP navigation — fails silently
Claude attempts screencapture — captures desktop wallpaper
Claude continues attempting methods without disclosing failure
Ask Claude to fix a GSAP ScrollTrigger mobile scroll conflict
Claude reads page source, generates fix without being able to run or see the site
Paste fix into live site — behavior worsens
Claude generates another fix — no improvement
Session consumes full usage token budget
Start new session — Claude has no memory of what was tried, what broke, or current site state
Repeat from step 6
Expected: Claude discloses inability to visually access site after first failed attempt and redirects user to a viable debugging path (DevTools, screen share, console output)

Actual: Claude continues attempting visual access methods for 20+ minutes, never discloses failure, makes code changes to live site while blind, consumes session usage budget with no net progress

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

claude-sonnet-4-6

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_