[BUG] Plan mode auto-implements over-scoped plan without user approval of plan contents

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?

During plan mode, Claude identified 5 issues in a codebase. The user instructed Claude to fix only items 1, 2, and 3.
Claude wrote a plan that included all 5 issues plus additional derived work (new store state, new UI components, new
styles). The plan mode system then automatically transitioned to implementation by injecting the prompt "Implement the
following plan:" with the over-scoped plan document — without the user ever having reviewed or approved the plan's
contents.

The user never said "implement the following plan." That text was generated by the plan mode machinery. The user never
approved the plan. The system skipped the approval step and went straight to implementation of a plan that exceeded the
user's requested scope.

What Should Happen?

Expected Behavior

1. When the user says "fix 1, 2, 3" out of 5 identified issues, the plan should only contain items 1, 2, and 3.
2. Even if the plan is over-scoped, the user should have an opportunity to review and reject/edit the plan before

implementation begins.

1. The system should not auto-inject "Implement the following plan:" without the user having explicitly approved the plan

contents.

Actual Behavior

1. Claude wrote a plan covering all 5 issues plus additional work despite being told to fix only 3.
2. The plan mode system auto-transitioned to implementation without user approval of the plan contents.
3. Claude began implementing the full over-scoped plan, modifying files the user never asked to change.
4. The user had to interrupt execution mid-implementation.

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce

1. Enter plan mode on a codebase with multiple potential issues.
2. Claude identifies N issues (in this case, 5).
3. Tell Claude to fix only a subset — e.g., "fix 1, 2, 3."
4. Claude writes a plan document that includes all N issues plus additional derived work, ignoring the subset instruction.
5. Claude calls ExitPlanMode.
6. The system injects "Implement the following plan:" with the over-scoped plan into the next conversation turn — the user

never sees the plan contents or gets a chance to approve/reject before implementation begins.

1. Claude begins implementing the full over-scoped plan, modifying files the user never asked to change.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.50 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

_No response_