Claude 5.x explanation written as a single thinking block before AskUserQuestion stays a collapsed Thought#27238

opened11d agobyJjacksonloughmiller-netizen

What happened

Sometimes text disappears when the AskUserQuestion tool gets used. I've on quite a few occasions had a question asking me about something the model claims to have explained, but that I never saw.

The missing text never shows up as a message, even after answering the question or reloading.

Diagnosis

On Claude 5.1+ models, when a reply has no text block and a single thinking block before the AskUserQuestion tool_use, that thinking block is usually the explanation the question refers to. T3 files it as a reasoning item, so it shows only as a collapsed, summarized "Thought" row and never as an assistant message.

The cause is the promotion rule from #18923 in apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts (v0.1.5):

:1755: // Claude 5.1+ emits at most one pre-tool text block as thinking; treat only a second consecutive thinking block as text.
:8250-8256: a thinking block is promoted to assistant text before a tool_use only if update.followsThinking.
:8281: followsThinking: previous !== null && previous.messageId === message.message.id. A lone thinking block is never promoted.

#18923 chose this rule from a sample across all tool calls, where lone pre-tool thinking blocks mostly read as reasoning. Right before AskUserQuestion, the opposite holds. The model has just written up what it is asking about, and a lone block is usually that write-up.

The server drops nothing. Every text block before a question was saved and shown. The text exists only as a reasoning summary.

main still has the same rule (claudeWritesUpdatesAsThinking / followsThinking), so no newer release fixes this.

Suggested fix: when a Claude 5.1+ assistant message has no text block and its tool_use is AskUserQuestion, promote the last non-empty thinking block to assistant text even without followsThinking. That keeps #18923's behaviour for ordinary tool calls.

Steps to reproduce

Not deterministic, because the model decides between text and thinking.

Claude provider, claude-opus-5-5, default thinking display (summarized), no launch-argument overrides.
Work in a thread where the agent reviews something and then asks you to choose via AskUserQuestion (e.g. "check this PR and ask me whether to merge it"). Repeat several times.
In roughly one in five question turns, the reply is thinking → tool_use AskUserQuestion with no text block. The question card appears with no visible explanation. The explanation appears only as a collapsed, reworded "Thought" row above it.
Confirm in the DB: orchestration_v2_projection_turn_items holds a single reasoning row …<msgId>:thinking:0 for that message, and no assistant_message row.

Version

0.1.5 (desktop app, local server reports 0.1.5)

Environment

macOS 26 (Darwin 25.6.0, arm64), Node v26.12.1, Claude Code 2.1.312, model claude-opus-5-5

Evidence

shell

1	# All AskUserQuestion calls in 60 provider logs, Oct 23-29 (141 calls), classified by the
2	# block before the tool_use and how the projection stored it:
3	text block before question 83 -> all saved as assistant_message, all shown
4	2 thinking blocks, no text 28 -> thinking:1 promoted to assistant_message
5	1 thinking block, no text 30 -> reasoning only, never an assistant_message
6
7	# Of the 30 lone blocks, about 26 are addressed to the user. Examples (names/numbers redacted):
8	"PR #NNNN bumps <app> to <version> with its release notes, all 22 checks pass, and merging starts no release ... Your options are to approve and merge at <sha>, hold it until <run> finishes, or run /code-review first."
9	"Two PRs are ready for a merge decision: #NNNN (vitest upgrade, all 19 checks pass, clears two CVEs) and #NNNN (doc-only plan update, no checks apply). I recommend merging both as-is ..."
10	# One the model wrote after noticing the problem itself:
11	"I realize the preceding explanatory text gets collapsed into a one-line summary before the question tool runs, so I need to fold the actual explanation ..."
12
13	# Same-thread contrast (one message each):
14	msg A: thinking:0 -> tool_use AskUserQuestion => turn_items: reasoning
15	msg B: thinking:0, thinking:1 -> tool_use AskUserQuestion => turn_items: reasoning, assistant_message
16
17	# Claude process args as launched by T3 0.1.5:
18	claude ... --thinking adaptive --thinking-display summarized --effort high --model claude-opus-5-5 ...

#18355 and #18923 introduced and narrowed the pre-tool thinking promotion. This issue is the case #18923's rule leaves out, a lone block before AskUserQuestion. #11353 and #14159 cover text blocks folded or hidden around questions, and they were closed as client-side fold issues. Here the text is never a message at all. #18674/#18680 are the reverse problem (reasoning promoted to text), which this fix avoids by limiting itself to AskUserQuestion.

Fix applied or workaround

Set Claude provider launch arguments to --thinking-display highlights, per a comment on #11353. Not yet verified on 0.1.5: the adapter also passes thinking: { display: "summarized" } to the SDK (ClaudeAdapterV2.ts:1548-1560), which may override the flag.

Filed by

Claude Code (claude-opus-5-5) via t3 triage