Text block before a tool call sometimes not rendered in terminal — silent content loss in long sessions

Summary

Assistant text content that precedes a tool call within the same turn is sometimes not rendered in the terminal transcript — the user never sees it, even though the model generated it (the content demonstrably existed: it referenced a draft that was then written to an external system by the subsequent tool call).

The user reports this has happened multiple times in their sessions.

Environment

• Claude Code 2.1.219, macOS 26.2, TERM=xterm-256color
• Very long-running session (1M-context model, heavy tool use, session had been compacted earlier)
• Model: Fable 5 (also observed on other models per user recollection)

Repro pattern (best current hypothesis)

1. Long session with many tool calls.
2. Assistant emits a turn shaped: [substantial text block incl. a markdown blockquote] → [Bash tool call] → [tool result] → [closing text].
3. The pre-tool text block silently fails to render; the closing text after the tool result renders fine.
4. Net effect: drafted content (in our case, a ready-to-send reply draft inside a > blockquote) is invisible to the user, who reasonably concludes it was never produced.

Not deterministic — most identically-shaped turns render fine; frequency feels correlated with session length.

Impact

Silent content loss: the user acts on an incomplete transcript (e.g., "you never drafted X" when X was drafted). Worse than a crash because nothing signals the omission.

Notes

• The lost content was recoverable by re-emitting it in a later turn, which rendered normally — so this looks like a display/transcript issue, not a generation issue.
• Happy to provide timestamps privately; the session transcript contains customer business content so it isn't attached here.