[BUG] Desktop app: assistant text in the same API message as a tool_use is silently hidden — no /focus or ctrl+o equivalent exists to reveal it

Preflight Checklist

• [x] I have searched existing issues — the closest reports are #50894 (CLI focus mode, which has TUI workarounds) and #24311 (closed as duplicate; terminal scrollback burying); neither covers this surface, addressed under "Not a duplicate" below
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code (2.1.173 via the desktop app's bundled runtime; the app auto-updates)

What's Wrong?

In the desktop app, assistant text emitted between tool calls within a single turn is silently hidden from the transcript. Text before a turn's first tool call renders; final end-of-turn text renders; everything in between — however substantive — does not, and nothing in the UI indicates content was hidden.

Per the community diagnosis on #50894, the collapse appears keyed on API message structure: assistant text sharing an API message with a tool_use block gets dropped together with the tool render. The terminal CLI has the same collapse in focus mode but offers /focus and ctrl+o to reveal the text; the desktop app exposes no equivalent control — the hidden text is unrecoverable in the UI.

The compounding case is AskUserQuestion: its answers return to the model as tool results, so a question → answer → follow-up exchange remains a single assistant turn. Any reasoning the model writes between questions is invisible. In practice the user is asked to choose between options whose entire justification was swallowed — in one of our sessions this happened twice in a row: the visible transcript was just two bare option lists, and the user's reply was "I see no reasoning."

Two failure modes, mirroring #50894's framing:

• Visible loss — explanations the user notices are missing and can ask for again
• Invisible loss (worse) — warnings, caveats, and reasoning the user never learns existed; the desktop transcript shows no "N messages hidden" indicator of the kind the CLI shows

What Should Happen?

Mid-turn assistant text should render in the desktop transcript — it is substantive model output, not tool noise. Failing that, the desktop app should provide what the CLI already has: an indication that messages are hidden, plus an affordance to reveal them.

Steps to Reproduce

1. Open a session in the desktop app
2. Prompt: Run 'echo ok' with Bash. After you see the result, write three short paragraphs explaining why the sky is blue, as plain text. Then use AskUserQuestion to ask me whether I'd like more detail.
3. Observe the transcript: the Bash call renders, the question renders — the three paragraphs never appear, and nothing indicates they exist

Control (same machine): the identical prompt in the terminal CLI renders the paragraphs (and ctrl+o additionally shows them in the detailed transcript), confirming the text is generated and transported — this is a desktop render decision, not transport loss.

Is this a regression?

Don't know — first systematically observed June 2026.

Claude Code Version

2.1.173 (reported by the desktop app's bundled runtime)

Platform

Anthropic API

Operating System

macOS 26.4.1

Terminal/Shell

N/A — the affected surface is the desktop app's transcript view; the terminal CLI on the same machine is the control case and behaves correctly

Additional Information

Not a duplicate of:

• #50894 — same underlying collapse, but scoped to CLI focus mode, where /focus and ctrl+o exist as workarounds; this report covers the desktop surface, which has no toggle and no reveal affordance
• #24311 (closed as duplicate of #23765) — terminal scrollback burying: text rendered but easy to miss; here the text is never rendered at all
• #51534 / #53447 — desktop hiding spinner verbs / tool-use summaries via a host-injected flag; here it is the assistant's own prose being hidden

Suggested fixes, in preference order (all at the render layer, consistent with the diagnosis endorsed on #50894):

1. Render mid-turn assistant text in the desktop transcript
2. Failing that, port the CLI's affordances — a hidden-messages indicator plus a reveal control
3. At minimum, the indicator alone — silent, unsignalled loss is the severity driver here