Assistant answers get visually buried behind tool-call/wakeup blocks in long autonomous sessions

Summary

In long-running autonomous sessions (background tasks + scheduled wakeups), the assistant's user-facing prose answers get visually buried behind tool-call and wakeup blocks in the transcript UI. The practical effect: the user asks a direct question, the model genuinely answers it, but the answer renders collapsed between/behind tool blocks — so the user experiences it as being ignored, asks again, gets the same result, and concludes the model is refusing to answer.

Repro shape

1. Run a multi-hour autonomous workflow where the model monitors background Bash tasks and re-arms itself with scheduled wakeups each turn.
2. Mid-run, ask a direct question ("what's the status?", "what's left?").
3. The model replies with: brief text → status-check tool call(s) → more text → wakeup-scheduling tool call as the final block.
4. In the UI, the prose is easy to miss entirely; several times in a row the user believed no answer was given.

Why it matters

This didn't used to be a friction point — it has gotten noticeably worse in recent versions as agentic sessions lean harder on background tasks/wakeups. The failure mode is nasty because both sides are "right": the model did answer, the user never saw it. It erodes trust in exactly the long-autonomy workflows Claude Code is best at.

Suggestions

• Render the final text message of a turn with fixed prominence, even when tool-call blocks follow or surround it (e.g., pin it below the collapsed tool groups).
• Or: visually separate "answer prose" from "operational chatter" so collapsed tool groups can't swallow adjacent text.
• Or: when a turn contains a scheduled wakeup as its last block, surface the preceding text block as the turn summary.

Workaround we've adopted: instructing the model (via CLAUDE.md) to always end question-answering turns with prose and never with trailing tool blocks — it helps, but it's fighting the renderer.