Focus mode hides substantive assistant messages, not just tool output

Focus mode hides all assistant text emitted between tool calls. Per the v2.1.97 release notes, the intent is to hide verbose tool logs and intermediate reasoning. The problem: the model doesn't distinguish "status update between tool calls" from "direct answer to the user," so substantive content gets suppressed the same way.

This happens frequently, not rarely. You can sometimes catch it in the grayed-out text that briefly appears and then disappears while Claude works in the background — messages you were meant to read, gone in a flash. There is no way to recover them short of toggling focus off mid-session.

Two failure modes

Visible: I ask a question. Claude answers between tool calls while doing follow-up work. The answer is hidden. I ask a related follow-up. Claude replies "I already answered that above." I never saw the answer.

Invisible (worse): Claude relays a warning, caveat, or finding between tool calls. It gets hidden. I have no idea it was ever said. Silent information loss with no way to know what I missed.

Severity is high because of the invisible case — the user can't report what they didn't see, and important context is lost without any signal.

Suggested fixes, ranked

1. Don't hide assistant text at all in focus mode. Only hide tool calls and tool results. This matches what focus mode is described as doing.
2. If hiding stays, show a persistent "N hidden messages" indicator the user can expand. The current brief gray flash is not discoverable.
3. Strengthen the system prompt so the model never emits substantive content between tool calls, only in the final message. Less reliable than a UI fix.

Environment

Claude Code v2.1.97+, focus view enabled.