Rejection reason lost when parallel tool calls are batch-rejected

Bug

When Claude makes two (or more) parallel tool calls and the user rejects them individually:

1. The user rejects the first tool call with feedback ("tell Claude what to do instead")
2. The user rejects the second tool call without feedback (just declines it)
3. Claude does not process the user's feedback from the first rejection — the guidance is lost
4. The user's rejection message is not displayed in the conversation UI, so they can't even see what they typed

It appears that when one rejection has feedback and another doesn't, the model fails to pick up the feedback.

Expected behavior

• The user's rejection reason should be surfaced to Claude and acted upon
• The rejection message should be visible in the conversation history

Steps to reproduce

1. Trigger Claude to make 2+ parallel tool calls (e.g., two file edits)
2. When the permission prompt appears for the first call, reject it and type a reason explaining what to do instead
3. When the permission prompt appears for the second call, reject it without a reason
4. Observe that Claude ignores the guidance from step 2 and the message is not shown in the conversation

Environment

• Claude Code CLI on macOS
• Model: claude-opus-4-6