[BUG] [VS Code Sidebar] Fable 5 doesn't show text responses in chat

Status Open
Reported on v2.1.205
Maintainer reply None cached
Activity 3 comments · opened Jul 9, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

I use VS Code: Sidebar for Claude Code
Most of the time Fabe 5 doesn't show text messages [explaining what it is doing or what it is going to do] in chat. I see only a list of Grep/Read/Thought for Xs/etc]. I find it major issue especially before AskUserQuestion event. It just throws questions to answer without explanation. I need to ask for temporary .md file to produce it.
Opus 4.8 doesn't have this issue.

What Should Happen?

It should include actual chat messages what it has done, what it is doing or what it is going to do, just like Opus.

Error Messages/Logs

Steps to Reproduce

Just run a session.

Claude Model

Fable 5

Is this a regression?

_No response_

Last Working Version

_No response_

Claude Code Version

latest 2.1.205

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code Sidebar

Additional Information

<img width="1195" height="1771" alt="Image" src="https://github.com/user-attachments/assets/0f52ae59-f7b5-4ab6-81d2-15ed40a2bffb" />

View original on GitHub ↗

3 Comments

kotok9 · 1 month ago

This looks like the same underlying bug I filed in #76032 (Desktop Code-session transcript). One clarification from controlled testing that may help triage: the assistant text is actually emitted — it's present in the local session .jsonl — but the transcript renderer drops it in spans that also contain an AskUserQuestion/ExitPlanMode call. So it's likely a renderer drop specific to Fable-5's output shape, not the model staying silent.

In a 2×2 (Fable-5 vs Opus-4.8 × text-and-dialog-same-turn vs separated), Opus-4.8 lost no text across ~11 trials while Fable-5 lost ~50% with the identical structure — matching your Fable-5-vs-Opus-4.8 observation. Cross-linking for the maintainers.

mdickinson · 1 month ago

#65620?

yanivhrs · 1 month ago

Same bug, with session-transcript evidence: the text is not only unrendered — it is also missing from the local session JSONL (VS Code extension, Windows, CLI 2.1.207, claude-fable-5, plan mode)

Environment

  • Claude Code v2.1.207, VS Code native extension (sidebar), Windows 11 Pro (build 26200)
  • Model: claude-fable-5, plan mode active for the whole session
  • Still reproduces on 2.1.207 (this issue was filed on 2.1.205)

What happened

Twice in one session the assistant emitted a multi-paragraph markdown text block (Hebrew/RTL) immediately followed by an AskUserQuestion call in the same turn. In both cases only the question dialog rendered; the explanatory text never appeared. The second occurrence was the worst case UX-wise: the dropped text was the explanation the user had explicitly just asked for, and the dialog then asked them to choose between options that only made sense given that explanation — an effectively context-free dialog. The user had to interrupt and ask where the answer went.

Transcript-level evidence

Inspecting the session JSONL (~/.claude/projects/<project>/<session-id>.jsonl) after the fact:

  • The entire session contains exactly one assistant text block — the very first assistant message of the conversation (which did render, and which itself preceded tool calls in the same turn).
  • Every subsequent mid-turn text block is gone. The affected spans are persisted as thinking (empty) → thinking (empty) → tool_use with no text record at all:
assistant thinking (empty)
assistant tool_use Agent              ← preceding status text missing
assistant thinking (empty) ×2
assistant tool_use AskUserQuestion    ← preceding ~150-word text block missing
assistant thinking (empty) ×2
assistant tool_use AskUserQuestion    ← preceding ~450-word, two-section markdown text missing

Data points for triage

  1. The drop is dual: not rendered and not persisted to the JSONL. This matches the mechanism documented in #67267 (terminal/tmux) — now confirmed on the VS Code extension surface.
  2. Not exclusive to AskUserQuestion: text emitted before an Agent (subagent) tool call was dropped the same way.
  3. Nondeterministic: the first turn's text also preceded tool calls in the same turn, yet persisted and rendered — consistent with #67426.
  4. Model correlation consistent with #76032 / #67166: observed on claude-fable-5; per those reports Opus 4.8 is unaffected.
  5. Because the text is also absent from the JSONL, --resume / transcript review cannot recover it; the content is simply lost to the user.

Session id withheld from the public thread for privacy; available to Anthropic staff on request (also submitted via the in-product /bug channel).