Assistant text between tool calls in a turn is not rendered to the user (mid-turn messages silently lost)

Summary

Assistant text emitted mid-turn — between tool calls — is not rendered to the user in the terminal transcript. The user only sees the final text block of the turn. The model, however, has no reliable indication of what rendered, so instructions it "posted" mid-turn silently vanish from the user's view while the model proceeds believing they were delivered.

Environment

• Claude Code 2.1.220, Windows 11 (win32), PowerShell host
• Long-running agentic session with many tool calls per turn

Repro (observed live today)

1. User asks the model for a set of steps to perform on another device.
2. The model replies in a turn shaped: text block (the steps) → Edit tool call → more tool calls → final short status text.
3. The terminal transcript shows the tool activity and the final status text. The steps block never appears.
4. User: "I don't see any steps posted above!" — the model believed they were posted.

A sibling class we've hit repeatedly: prose written in the same turn as an AskUserQuestion panel (before the panel) also fails to render — only the panel shows. Context placed inside the panel renders reliably; same-turn prose outside it can be lost.

Expected

Either (a) all assistant text blocks in a turn render in order interleaved with tool activity, or (b) the harness makes the contract explicit to the model so it never relies on mid-turn text (the system prompt's "may not be shown" hedge means the model cannot tell which side it's on).

Impact

• Instructions/answers addressed to the user are silently dropped; the user has to ask again.
• The model's mental state and the user's transcript diverge — in consent-sensitive flows (the model "told" the user something it functionally never said) this is worse than a cosmetic loss.

Notes

Deterministic from the model side: any turn where text is followed by further tool calls loses that text. Happy to provide fuller session context if useful.