[BUG] Model repeatedly begins implementation without explicit user approval

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 have a rule in my CLAUDE.md (global instructions) that states no implementation should happen until I explicitly approve it. The model consistently breaks this rule by deciding on its
own to implement something, then using file reads as the first step of that implementation.

Specific example from this session:

The user (me) was in proposal/design mode. Without any approval signal from me, the model said:

▎ "You're right. Let me stop drawing and just implement both fixes."

It then immediately began calling Read tools on source files as preparation for making code changes. When challenged, it acknowledged the violation but the pattern repeated multiple times
across the session.

What Should Happen?

The model should wait for an explicit action verb ("go ahead", "implement", "do it", "proceed") before taking any implementation steps, including preparatory reads.

Error Messages/Logs

Steps to Reproduce

The model should wait for an explicit action verb ("go ahead", "implement", "do it", "proceed") before taking any implementation steps, including preparatory reads.

Actual behavior:
The model decides to implement based on context (e.g. a frustration signal from the user, or a long back-and-forth) and begins without approval. It also repeatedly broke the same rule
after being corrected within the same session.

Session: claude-sonnet-4-6, Claude Code CLI, Windows 11

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.141

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

_No response_