Desktop app collapses assistant text written between tool calls; substantive answers silently hidden under 'Ran N commands'

Summary

In the Claude Code desktop app, assistant text written between tool calls is collapsed under the "Ran N commands" fold along with the tool calls themselves. The user only reliably sees the final text block of a turn. When the model puts substantive content mid-turn (its natural behavior when narrating work), that content is silently lost to the reader — with no visual hint that collapsed prose exists.

Environment

• Claude Code 2.1.233, desktop app (macOS 26.5.2)
• Model: claude-fable-5

Two real repro cases from one session

Case 1 — numbered list appears to start at "4."
The assistant wrote steps 1–3 of a numbered procedure, then made a Bash tool call (printing an SSH public key needed for step 3), then continued with steps 4–5. Rendered result for the user: the earlier context, the "Ran 4 commands" fold, the key output, then a list beginning at item 4. Steps 1–3 were inside the fold. The user's literal reaction: "Your response to me started numeration at (4). what happened?"

Case 2 — an explicitly requested explanation collapsed entirely
User asked: "stop, kill the campaign and explain why you auto-jumped into that?" The assistant wrote the kill commands, then a multi-paragraph explanation, then two more tool calls (updating a memory file), then a short closing status. Rendered result: "Killed the ARIA campaign process >", "Ran 2 commands, used 2 tools >", and only the short closing status as visible text. The requested explanation was inside the fold. The user had to ask "can you answer my question" — the answer had been written and sent, just not rendered.

Why this matters

• The failure is silent on both sides: the model believes it answered; the user sees tool-call chrome and a fragment. Neither party gets a signal that content was hidden.
• It punishes exactly the behavior long agentic turns need — explaining decisions adjacent to the actions that implement them.
• Expanding the fold does surface the text, but users have no reason to expand "Ran 2 commands" hunting for prose.

Suggested directions

• Render assistant prose blocks between tool calls as normal message text (only fold the tool calls/results), or
• Give folds containing non-trivial prose a distinct affordance (e.g. "Ran 2 commands · 3 paragraphs"), or
• Surface mid-turn text blocks over some length threshold unconditionally.

🤖 Generated with Claude Code