Mid-turn user messages get buried between tool blocks and often end up unanswered

Summary

When the user sends a message while Claude is mid-turn (interrupting or queuing during tool execution), the message is rendered inline inside the running turn, wedged between tool-use blocks (e.g. right below a Bash permission prompt or tool result). Two problems compound:

1. Visually buried: the user's question sits between tool blocks and quickly scrolls away as subsequent tool calls stream in. It does not stand out as a conversation turn.
2. Often effectively unanswered: the model tends to address the interjected question in text between tool calls (which also scrolls away), then ends the turn with a summary of the original task's deliverables. From the user's perspective the question was silently ignored, and they have to re-ask it.

The user reports hitting this repeatedly.

Concrete example (real session)

1. Claude is mid-task (editing files, running tests via Bash).
2. User interrupts with an unrelated question ("why does cmd+click on this path resolve to a wrong doubled path?") attached to a permission denial.
3. Claude answers partially in text sandwiched between a memory-file Write and further edits, then ends the turn summarizing the original task only.
4. User's reaction (translated): "You didn't answer me — you just went and edited Memory." They re-sent the same question with the same screenshot.

Suggestion

• Render mid-turn user messages with distinct visual prominence (their own block style, not inline with tool results), so they read as a conversation turn rather than tool output.
• Strengthen the harness reminder that accompanies mid-turn messages so the model must address them in the final end-of-turn message, not only in passing between tool calls.

Environment

• Claude Code 2.1.212, macOS (darwin), iTerm2 3.6.11 + tmux