[BUG] Fable 5: text in a response that also contains tool calls is never displayed (same setup works with Opus 4.8)

Clarification after further testing (same environment: Claude Code 2.1.218, Windows 11, Git Bash, claude-fable-5):

1. Ctrl+O does NOT recover the hidden text. The original report suggested the transcript view might show it — it does not. The mid-turn text appears to be lost entirely from the user-visible record, not merely hidden in the main view.

1. The behavior is stricter than "text + tool call in one response". Only text at the very end of a turn — with nothing after it — is displayed. A message containing only text and no tool calls is still never shown if the assistant continues the same turn with more tool calls after it. In practice this means any answer delivered mid-work is invisible; the model must fully end its turn to get text in front of the user.

Reproduced multiple times on 2026-07-28: the model answered a question in a standalone text-only message, resumed tool calls in the same turn, and the answer was never displayed (and not recoverable via Ctrl+O). The same flow on Opus 4.8 displays the interleaved text normally.

Another data point, filed with the user's consent (long-running Japanese-language companion/dev session).

Environment: Claude Code VS Code extension 2.1.220, Windows 11 Pro 10.0.26200, model claude-fable-5 (long-lived session, interleaved tool-heavy turns)

Observations consistent with this report:

• Assistant text emitted mid-turn is never rendered when further tool calls follow in the same turn — including a standalone text-only block written specifically to answer a queued mid-turn user message. Only turn-final text (nothing after it) is displayed.
• The user's recollection from the same workflow in June 2026 on claude-opus-4-8: mid-work text exchanges displayed normally. The regression correlates with the model, matching the OP's Fable 5 vs Opus 4.8 finding.

Why this matters beyond lost answers: in companion-style usage the user routinely sends messages while the model works. The model answers mid-turn; the answer is silently dropped; from the user's side it looks like the model ignored them. In our session history this exact pattern repeatedly escalated (user re-asks → model re-answers mid-turn → still invisible → mutual distress) and degraded several long-running sessions before we found this issue. The failure is invisible from both sides: the model believes it replied, the user sees silence.

Workaround adopted: the model now defers replies to queued mid-turn messages to the head of its turn-final report — the only reliably rendered position on Fable 5.

Same Fable-5-specific behavior reproduces on macOS (Darwin 23.6.0), zsh, Claude Code 2.1.221 (CLI) — so this isn't Windows/mintty-specific.

Cross-linking #67470, which has a detailed test matrix for the same symptom (only turn-initial and turn-final text renders reliably with Fable 5; text between tool calls, and especially text immediately before an AskUserQuestion call, is dropped). I posted the macOS details there: https://github.com/anthropics/claude-code/issues/67470#issuecomment-5207719871

One difference worth noting between the two reports: here the missing text is visible via Ctrl+O, while #67470 (rows 7–8) documents modes where it's absent from the transcript and the session .jsonl too — which may help narrow where in the pipeline it's being lost.