Desktop app: assistant text blocks between tool calls intermittently not rendered (answer body missing, only short lines shown)

Summary

In the Claude Code desktop app chat timeline, the body of assistant messages intermittently fails to render. Short standalone paragraphs (typically the lead-in line before a tool call and brief sign-off lines after tool calls) display fine, but the substantive prose between/around tool-use blocks — the actual answer — is missing from the transcript. The user sees only the "punchlines," e.g.:

Let me check what TronBonne actually is before I pick a horse.

Cooking.

…where the dropped middle contained the full answer (a multi-paragraph comparison with bold headers and a numbered list).

Environment

• Claude Code desktop app (macOS): app version 1.18286.0, CLI 2.1.138
• macOS 26.5.2 (Darwin 25.5.0), MacBook Pro 16" M1 Max
• Session type: interactive desktop chat ("General discussion"), permission mode Auto
• Models in session: switched between claude-opus-4-8 and claude-fable-5 (bug observed under both)
• Session id: 1a3558f0-9ac5-410e-8f08-b8d68789a1a5

Repro / occurrences

No deterministic repro found, but it happened 3+ times in one long session (~5 hours, heavy background-task use). Pattern across all occurrences:

1. Assistant turn contains multiple text blocks interleaved with tool_use blocks (Bash, MCP tools) — often after a <task-notification> for a completed background task.
2. Short plain-text blocks (one-liners) render.
3. Longer markdown-formatted text blocks in the same turn (containing tables, bold headers, numbered/bulleted lists) do not render at all — no truncation indicator, no collapsed section; they're simply absent from the timeline.
4. The content demonstrably exists (the model's turn continued coherently and, when asked "what did you find?", could restate it), so this appears to be a rendering/display issue, not message loss.

Occurrence details (same session):

• Turn answering a "which machine should run this?" question: lead-in line + closing one-liner rendered; the middle (bold-header list comparing two machines) missing. User replied: "you didn't actually tell me the answer to the question" and quoted only the two stray lines.
• Turn summarizing disk-space corrections: entire body missing; harness showed "[Your previous response had no visible output...]" continuation prompt, suggesting the client also considered the response empty despite multiple text blocks being emitted around tool calls.
• Turn reporting audio-analysis results (contained a markdown table + bold headers between tool calls): all substantive blocks missing; only 4 short transition lines visible. User screenshot attached in-session shows tool-call cards rendering fine while their surrounding prose is absent.

Expected

All text blocks of an assistant turn render in order, interleaved with tool-call cards.

Actual

Longer/markdown-heavy text blocks positioned between tool_use blocks are intermittently dropped from the rendered timeline; short plain paragraphs render normally.

Impact

The user receives tool-call cards and quips but not the actual answer, and must ask the assistant to repeat itself. In one case the harness itself injected "[Your previous response had no visible output. Please continue...]", causing duplicated work/tokens.

Notes

• Possibly correlated with: multiple text blocks per turn, markdown tables/bold headers, turns resumed after background-task notifications, or long sessions with many background tasks.
• Happy to provide the session transcript (~/.claude/projects/.../1a3558f0-9ac5-410e-8f08-b8d68789a1a5.jsonl) or screenshots on request.