Assistant text between tool calls in a long agentic turn is never rendered in the terminal UI

Summary

Assistant text emitted between tool calls inside a single long-running turn is never rendered in the terminal UI. The user repeatedly asked a direct question mid-turn; the model answered twice as interim text blocks (text emitted between tool invocations while background subagents/tasks were running), and neither answer was ever shown. Only when the model made the answer the final message of a turn did it render.

Environment

• Claude Code 2.1.220 (CLI, terminal)
• macOS 26.5.2 (Darwin 25.5.0)
• Model: claude-fable-5
• Session shape when it reproduced: a very long agentic turn with many tool calls, multiple background subagents (Agent tool) and background Bash tasks, with <task-notification> events and mid-turn user messages interleaved

Steps to reproduce (as observed)

1. Start a task that produces a long turn: multiple background agents + background shell tasks, dozens of tool calls.
2. While the turn is running, send a mid-turn user message asking a question (it gets surfaced to the model alongside the next tool result).
3. The model responds with a text block between two tool calls and continues working.
4. Observe the terminal: the interleaved text block is not displayed (neither live nor after the turn completes). The user sees only tool activity and the turn's final message.

Happened twice in the same session (two separate mid-turn answers to the same question both invisible); a third ask as a normal top-of-turn message rendered fine.

Expected

Text blocks the assistant emits between tool calls should be rendered in order, as they are part of the visible conversation (they are present in the transcript/JSONL).

Actual

Mid-turn text blocks are silently dropped from the UI. The information is only recoverable from the session JSONL. This breaks mid-turn Q&A with a working agent: the user cannot tell whether the model ignored the question or answered invisibly.

Notes

• The dropped text was substantial (multi-paragraph with a markdown table), not just a one-liner status.
• Mid-turn messages from other sources (teammate-message blocks, task-notifications) displayed fine; only the assistant's own interleaved text was lost.