[BUG] Background sessions on claude-fable-5 render no assistant text — only AskUserQuestion dialogs; transcript shows text blocks were generated

Preflight Checklist

• [x] I have searched existing issues — closest are #73561, #67880 and #65108; none covers this exact case (see "Related issues" below)
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code

What's Wrong?

In background sessions (jobs started via /bg), with model claude-fable-5, no assistant prose ever reaches the user. The only content that renders is AskUserQuestion dialogs. Turn-final text messages (no tool calls after them) — which the harness system prompt tells the model are the reliable channel — do not render either.

This is a rendering/extraction bug, not model silence. The session .jsonl transcript proves the text was generated:

• Session e9f93200-… (2026-06-11, fable-5 background job): assistant content blocks were 97 tool_use, 78 thinking, 9 text, totalling 8,606 chars of assistant text — including multi-paragraph turn-final recaps written specifically for the user. None of it rendered; I saw only AskUserQuestion dialogs (plus one stray one-liner immediately before a turn break).

A/B on the same machine, same settings.json (language pt-BR, outputStyle Explanatory, effortLevel xhigh, awaySummaryEnabled=false): background sessions on claude-opus-4-8 render assistant text normally; on claude-fable-5 they render none. Since the config is identical across both models, no user config field explains the delta.

Timeline: first observed on claude-fable-5 during its initial availability window (June 2026); still reproduces today on 2.1.199 (2026-07-02). It was absent while Fable was unavailable (sessions ran on Opus) and returned with Fable.

Environment: Claude Code CLI 2.1.199 · WSL2 (Linux 6.6.87.2-microsoft-standard-WSL2) · terminal TUI · model claude-fable-5 · background job sessions (/bg).

What Should Happen?

Assistant text produced during a background-session turn — at minimum the turn-final message — should render to the user exactly like in a foreground session, regardless of model. If the UI intentionally elides mid-turn text, it should say so visibly instead of silently dropping it.

Related issues

• #73561 — fable-5, wakeup-driven sessions: inconsistent mid-turn rendering (short lines render, long reports don't). Mine is stronger: nothing renders except AskUserQuestion.
• #67880 — fable-5, /loop: transcript-verified zero text blocks generated (model-side). Different root: in my transcripts the text blocks ARE present — the failure is on the rendering path.
• #65108 — opus-4-8, ScheduleWakeup turns: suspected rendering suppression (stale). Suggests the bg/wakeup rendering path has had issues before; fable-5 seems to hit it 100% of the time.

Error Messages/Logs

None. Sessions complete normally; the .jsonl is healthy and contains the missing text.

Steps to Reproduce

1. On Claude Code CLI 2.1.x (Linux/WSL2), start a background job (/bg) with model claude-fable-5.
2. Give it a task where the model writes prose (status updates, findings, a final summary) and at some point calls AskUserQuestion.
3. Observe the session UI: only the AskUserQuestion dialog renders; no assistant prose appears, including the turn-final message.
4. Open the session's .jsonl transcript and count assistant text blocks (e.g. jq -r 'select(.type=="assistant") | .message.content[]?.type' file.jsonl | sort | uniq -c): the text blocks are there.
5. Repeat the same flow with claude-opus-4-8: assistant text renders normally.

---

🤖 Drafted with Claude Code on the user's behalf (transcript analysis from the affected machine).