[Bug]: Claude replies disappear on updates and remain incomplete after expanding "Worked for" #17382

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/web

Observed in the packaged macOS desktop app's conversation timeline.

Steps to reproduce

Observed sequence in an existing, long-running Claude conversation; a deterministic minimal reproduction has not yet been established:

1. Open a Claude conversation with multiple assistant messages and tool calls within a turn.
2. Read an assistant response while the conversation receives new messages, or switch away to another tab and return.
3. The text being read disappears or becomes folded into the activity history.
4. Expand Worked for … / Worked and click the items underneath, including tool/activity disclosures.
5. The full original response is still not accessible in the UI.

The new-message and tab/focus triggers are reported from normal use. The complete live UI failure has not been independently reproduced during the investigation.

Expected behavior

Previously displayed assistant responses remain readable in full. If T3 folds them, expanding the relevant history restores all of their text. Explicitly expanded history should retain its state across new turns and tab/focus changes.

Actual behavior

Assistant text disappears while it is being read. Expanding Worked for and its child items still does not reveal the full response, so this is more than unexpected default folding.

Asking Claude to repeat the missed response is the only in-app workaround found so far.

Impact

Major degradation or frequent failure

The reporter repeatedly misses provider output and must request it again, interrupting work and consuming additional provider usage.

Version or commit

T3 Code (Alpha) 0.0.45, packaged build commit 6c8fed3.

Environment

• macOS 26.6.1 (build 25G76), Apple Silicon / arm64.
• Packaged T3 Code (Alpha) desktop app.
• Claude provider (claudeAgent); recent provider errors identify the model as claude-opus-5-5.
• Response streaming: Show finished paragraphs (inherited default, no project override).
• Observed in a long-running thread with 418 saved assistant messages, 491 reasoning messages, and 45 user messages at inspection time.

Logs or stack traces

No matching renderer exception or stack trace was established. Read-only comparison of retained provider completion logs against the local conversation database found:

Assistant completion events with nonempty detail checked: 108
Exact matches against saved assistant text:              108
Missing saved messages within that comparison:             0
Mismatched text within that comparison:                    0

This supports investigating the UI/history presentation, but does not establish that every historical or transient message was retained.

Relevant logs are under:

~/.t3/userdata/logs/provider/events.<thread-id>.log
~/.t3/userdata/logs/server.trace.ndjson
~/.t3/userdata/logs/desktop.trace.ndjson

Numbered siblings contain rotated history. Raw logs and transcripts are omitted because they contain private conversation text.

Screenshots, recordings, or supporting files

No screenshot or recording is attached.

Inspection of the installed apps/server/dist/client/assets/_chat-DWO1wYmt.js found:

• The terminal-message selection keeps the last assistant message per turn visible; earlier assistant messages are folded.
• Turn folding uses expandedTurnIds, while tool/activity disclosures use expandedWorkGroupIds.
• A React effect removes the previous turn ID from the expanded-turn set when the latest turn changes to another turn, clearing explicit expansion.

Running the exact installed folding functions against a fixture from the affected thread reproduced hiding two of three assistant messages in a settled turn. Opening the turn restored those messages at the folding-logic layer. Executing the exact installed expansion-reset effect also reproduced clearing previous-turn expansion on a new turn.

These findings do not explain the entire reported bug: the reporter has already expanded the turn and its children without recovering the full text. Disclosure-state restoration, history paging, or virtualized rendering may need further investigation. The tab-focus trigger remains unconfirmed, and no fix has been applied.

Related: #8879 describes a substantive answer hidden behind a trailing progress note with Grok. This report concerns Claude and, specifically, failure to recover the full response even after expanding the history.

Workaround

Ask Claude to repeat the missed response. The saved assistant text can also be recovered directly from the local database; this was used to export recent replies during investigation.

Expanding Worked for and its child items is not a working workaround for this incident. No setting was found in 0.0.45 that disables reply folding or forces full historical replies to remain visible. Wait for the full response only changes when new text is displayed and is not a verified fix.