Messages meant for the user are sometimes emitted as hidden thinking before a tool call

Summary

In Claude Code, a message the model intends for the user is sometimes emitted as a
thinking block instead of a text block. The user never sees it. It happens in replies
that also contain a tool call, and it is intermittent. Replies with no tool call have not
shown it.

What happens

The model writes a message addressed to the user (for example, the background for a question
it is about to ask) and then calls a tool in the same reply, often AskUserQuestion. In the
transcript, the message appears inside a non-empty thinking block, directly before the
tool_use block, and the reply has no text block at all. The user sees the tool's prompt
with none of the context written for it.

A made-up example of such a thinking block's content, in the shape we observed:

I have summarised the three options above, and now I ask which one to take.

The model writes it as if it has already told the user, in the first person, but the user
was told nothing.

Effect

• The user gets a question with no context, or misses a checklist or explanation entirely.
• The user cannot tell that anything is missing, and the model cannot either: in the

following turns it behaves as though the message was delivered.

• Written instructions do not help ("write the context as visible text before the question"),

because the model believes it did.

Measurements

Hand-checked sample (one session). 417 assistant rows: 75 text blocks and 113
thinking blocks. 101 of the thinking blocks were empty, holding only a signature. All 12
non-empty ones were messages meant for the user, each directly before a tool call in the
same reply.

Automatic count (1 day of sessions, 12 transcripts, one developer machine). A "reply" is
every transcript row sharing one message.id.

| | Count |
|---|---:|
| Replies with a tool call | 1371 |
| ...with non-empty thinking before the tool call and no text block | 117 (8.5%) |
| Replies with no tool call | 31 |
| ...with a text block | 31 (100%) |

The 117 is an upper bound. A count cannot tell a slipped message from ordinary reasoning
before a tool call. In the hand-checked sample, though, every such block was a slipped
message.

Trigger, as far as we can tell

• The reply contains a tool call, and the message was written before that tool call.
• Replies without a tool call have always delivered their text.
• It is intermittent: many messages written before a tool call do arrive as text.
• It is most noticeable with AskUserQuestion, where the missing text is the question's

context, but the shape is the same before any tool.

No minimal reproduction has been found. It appears within ordinary longer sessions.

Environment

• Model: claude-opus-5-5 (every reply in the measured window)
• Claude Code CLI 2.1.283, in the VSCode extension
• Windows 11

A second, smaller observation

AskUserQuestion options accept a preview field. The PreToolUse hook payload carries it,
but the VSCode extension did not display it: the user highlighted the first option and saw
only its label and description. We had tried to use the preview as a second, slip-proof
carrier for a question's context.

What we do meanwhile

Our PreToolUse hook on AskUserQuestion refuses a question when no visible text was written
since the user last spoke. After a refusal, the model writes the context as a reply with no
tool call, which has always arrived. A Stop hook then sends it back to ask the question in
the same turn.