[BUG] VSCode harness mid-turn output swallowed from Fable but not from Opus

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?

Claude Fable's diagnosis:

Bug: mid-turn assistant text not rendered when the model streams interleaved thinking

Model-specific: occurs with claude-fable-5, does NOT occur with claude-opus-5 — verified in the same session, same pane, by switching models with /model. The trigger is the interleaved thinking blocks the fable model streams (opus emits none here), not the model's text itself.

Summary

In the Claude Code session pane (VS Code extension / remote app), assistant
text blocks that appear after a thinking block and before the next tool call
are silently dropped from the display. The pane shows the thinking indicator
("Thought for Ns") followed directly by the tool call; the text between them is
never rendered.

Because models configured with interleaved thinking attach a thinking block to
every post-tool-result continuation — including empty ones, shown as
"Thought for 0s" — this eats all mid-turn narration and, far worse, any
answers to questions the user sends mid-turn: the user sees the agent
acting on their question but never sees the reply, which reads as being
ignored. This happened repeatedly in a real session before it was diagnosed.

Turn-opening text and turn-final text render normally. Tool calls and results
render normally. Only interior text blocks are lost.

Environment

• Claude Code 2.1.177, VS Code extension (also reproduced via the remote app)
• VS Code 1.117.0, Linux (kernel 7.0.0-28-generic)
• Model: claude-fable-5 with interleaved thinking (reproduces)
• Model: claude-opus-5, no thinking blocks emitted (does NOT reproduce —

same session, same pane)

Historical correlation (~12 data points before the probes): every dropped
block was an answer to a user message; every rendered "mid-turn" block turned
out to sit immediately after a hidden turn boundary (background-task
completion notification, tool-load turn, interrupt), i.e. was effectively
turn-opening. Two turn-opening answers were also lost, but both directly
followed abnormal events (a session-limit reset and a /model interrupt).

Diagnosis

The renderer appears to special-case message segments that begin with a
thinking block: it renders the thinking indicator and the segment's tool call,
but skips a text block sitting between them. Segments without thinking blocks
([text, tool_use] — the claude-opus-5 shape) render fully, as does a
trailing text block with no tool call after it (the "final answer" path).

Impact / severity

Data loss in the primary communication channel. The agent's replies to
mid-turn user questions are the most common casualty, so from the user's side
the agent appears to read questions and refuse to answer them. Until
diagnosed, the only workaround is behavioural: the agent must repeat anything
important in the turn-final text block.

Notes

• The full repro (probe turns included) exists in a real session and can be

provided; the probes are trivial to re-run in any session on an affected
model.

• "Thought for 0s" rendering on empty thinking blocks may itself be worth a

look while in this code path.

What Should Happen?

All Fable's output directed at user communication should be visible to the user in the session pane.

Expected

Every assistant text block renders in order, interleaved with the tool calls
(this is exactly what happens with claude-opus-5 in the same session).

Actual

The pane shows: opening text → tool call → "Thought for Ns" → tool call →
… → final text. All text blocks between tool calls are missing.

Evidence (controlled probes, one session)

Probe 1 — four labeled blocks in one turn (model reasoning active):

| Block | Position | Rendered? |
|-------|----------|-----------|
| P1 | turn-opening text | yes |
| P2 | mid-turn text, short plain prose, no markdown | no |
| P3 | mid-turn text, markdown list + links | no |
| P4 | turn-final text | yes |

All three echo tool calls between them rendered. Content, length and
formatting made no difference — P2 was deliberately shaped like narration
lines that had rendered earlier in the session.

Probe 2 — the model attempted a turn with zero reasoning:
text A → bash → text B → bash → text C. Result: A and C rendered, B did
not — and the pane showed "Thought for 0s" before B. That is, the
stream inserts an empty thinking block on every post-tool-result continuation
even when the model produces no reasoning content, and the empty block is
sufficient to suppress the following text block.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce

1. Use a model that streams interleaved thinking (e.g. claude-fable-5).
2. Give it any task producing a multi-tool-call turn with narration between

calls, e.g.: "Run echo A, then tell me something, then run echo B,
then summarize."

1. Observe the session pane.

Claude Model

Other

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.177

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

VS Code integrated terminal

Additional Information

<img width="1144" height="878" alt="Image" src="https://github.com/user-attachments/assets/91073d2b-2c09-4563-a8dd-c2de85c9d125" />