ask_user: text written in the same response as an ask_user call is not shown #5460

Describe the bug

When one agent response contains explanatory text followed by an ask_user call, the user sees only the form: "Copilot needs information.", the form's message, and each field's title, description and options. The text the agent wrote before the call in that response never appears.

A one-line progress note the agent wrote earlier in the same turn, before a web-search call, was shown.

Impact:

• It happened three times in a row in one session. Each time the user saw only the option labels and asked for the explanation that had already been written. The agent, believing it had explained, wrote more text before the next form, which was hidden again.
• A decision can be made from option labels alone, without the evidence the agent meant to show.
• The agent's instructions tell it never to ask questions in plain text and to use ask_user instead. That makes "explain, then ask" the natural pattern, and this hides the explanation every time.
• Nothing tells the agent: the ask_user tool description does not mention it.

Affected version

GitHub Copilot CLI 1.0.109 (the session where it happened; copilot --version now reports 1.0.112, not retested)

Steps to reproduce the behavior

1. Ask the agent for a decision that needs evidence.
2. In one response, the agent writes several paragraphs (a table and bullet points) and then calls ask_user with a one-line message and a single-choice field.
3. The form appears; the paragraphs do not.

Expected behavior

The text the agent wrote before the call is visible before, or with, the form.

Additional context

• Windows (NT 10.0.26300), x86_64, Windows Terminal, PowerShell 7.6; interactive session; agent model Claude Opus 5.5.
• Workarounds: the agent puts all context inside the form's message, which is shown; or ends its turn with the explanation and asks in a later turn; or the user turns the form off with /settings askUser off (or starts with --no-ask-user), and the documentation says the agent "can still ask in its replies".
• Possible fixes: show the pending assistant text before opening the form, or inside it; or state the behaviour in the ask_user tool description, so agents put their context in message.
• The hidden text does not appear after the form is answered either.
• Logs are not attached: the session holds private repository content.