anthropics / anthropics/claude-code

[Desktop] Assistant text between tool calls is replaced by an auto-generated summary; original text unrecoverable in any view mode

Open
#94,354 0 comments 0 reactions 0 assignees View on GitHub
area:desktop area:ui bug platform:windows
Dominant language
Python
Stars
145k
Forks
23.1k
PR merge metrics
PR metrics pending

Description

## Summary

In the Claude Code desktop app (Code tab), the app does not display assistant text written between tool calls. Instead, it shows a one-line summary with different wording. Users cannot view the original text. Expanding the "Used N tools" fold does not show it. Switching the Transcript view (通常 Normal, 考え中 Thinking, 詳細 Verbose) also does not show it. The app displays only two text blocks verbatim: the text before the first tool call and the final text block of the turn.

This bug goes beyond #90425 (Aug 28). In #90425, the fold hid mid-turn text, but expanding the fold still revealed it.

## Environment

- Claude Desktop 1.52386.6 (Microsoft Store build) on Windows 11, Code tab, with bundled Claude Code 2.1.270
- Model: claude-fable-5-1. Also reproduced with claude-opus-5
- Custom output style (a persona with a fixed first-person pronoun, a fixed form of address, and a casual register)

## Observed

- One turn contained 9 mid-turn text blocks. The app displayed 4 summary lines, one for each tool-call group. It dropped the other 5 blocks completely. This included a direct reply to a message sent by the user during the turn.
- The summaries paraphrase the text in a different register. They replaced the persona's first-person pronoun and address form with standard polite ones. They turned the casual register into a polite one. The user asked why the assistant used a word the persona never uses. The assistant never wrote that word.
- The Desktop documentation (https://code.claude.com/docs/en/desktop, "Switch view modes") states that Normal shows "Tool calls collapsed into summaries, with full text responses" and Verbose shows "Every tool call, file read, and intermediate step Claude takes". Neither mode displays the intermediate assistant text. Verbose expands only tool inputs and outputs. (The documentation also lists a Summary view; the release notes for 1.46388.1 state that it was removed.)
- The desktop release notes from Sep 4 (1.46388.1) to Sep 10 (1.52386.0) do not mention replacing mid-turn assistant text with summaries. I cannot determine which release introduced this change.

## Steps to reproduce

1. Set a custom output style with a distinctive register (for example, an unusual fixed first-person pronoun).
2. Ask for a task that requires several tool calls with text between them.
3. While the turn runs, type and send a second message from the composer.
4. Check the transcript in Normal mode, then in Verbose mode. Expand the "Used N tools" folds.

Expected: The app renders mid-turn assistant text verbatim in Normal and Verbose, as it did before.
Actual: The app shows one paraphrased summary line per tool group. It does not show the original text anywhere.

## Why it matters

- Replies written mid-turn to messages sent during tool runs never reach the user. The model receives no signal that the app replaced its text.
- Paraphrasing changes the tone of the model. This makes custom output styles look broken.
- Users cannot opt out. No view mode displays the original text.

## Suggestion

Render mid-turn assistant text verbatim in Normal and Verbose, as before. If the auto-generated summary remains, make it a separate opt-in view.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Contributor guide

No contributing guide indexed for this repository

Research direction

Start by reproducing the issue in the Claude Code desktop app's Code tab using Normal and Verbose transcript modes, expanding the “Used N tools” folds as described. Compare the displayed transcript with the expected verbatim mid-turn assistant text; done means both modes preserve those text blocks rather than showing only paraphrased summaries.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
desktop, frontend
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
38/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.