[BUG] Instructions not passed to context of subagents

Environment

• Platform (select one):
• [x] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ ] Other: <!-- specify -->
• Claude CLI version: 1.0.81
• Operating System: Windows 11 (WSL)
• Terminal: <!-- e.g. iTerm2, Terminal App -->

Bug Description

When a subagent tries to perform an action and you choose "No, and tell Claude what to do differently", it switches back to the main context, and anything you tell it won't be passed along to the subagent's context, so it keeps making the same mistake.

Steps to Reproduce

<img width="2177" height="954" alt="Image" src="https://github.com/user-attachments/assets/17f0504c-8c7a-40fc-8c54-8414bb631314" />